bastion 0.58.0

.gitattributes

@@ -0,0 +1,3 @@
+# Test tooling for one test lab, built only with -tags lab; it is not part of
+# any release build, so it is left out of the published source.
+client-go/lab.go export-ignore

.gitea/workflows/ci.yml

@@ -16,7 +16,7 @@ on:
     branches: [main]
 
 env:
-  GO_IMAGE: golang@sha256:1ae0735f00daffa3aaf1363a5184c0d2dc55c78e3db4ec70241cdac97bf84b59
+  GO_IMAGE: golang@sha256:27f2d3b83cfc5803c8078f8e730af0f7c431bf067afaec8f60a36fb763187cba
   # The pull request's own head, or the pushed commit.
   WANT: ${{ gitea.event.pull_request.head.sha || gitea.sha }}
 
@@ -58,7 +58,10 @@ jobs:
             echo "== gofmt"; out="$(gofmt -l .)"; [ -z "$out" ] || { echo "$out"; exit 1; }
             for os in linux darwin windows; do echo "== go vet ($os)"; GOOS=$os go vet ./...; done
             echo "== go build"; go build -o /tmp/privatae-bastion .
-            echo "== go test"; go test -count=1 ./...'
+            if [ -f lab.go ]; then echo "== go vet -tags lab"; go vet -tags lab ./...; fi
+            echo "== go test"; go test -count=1 ./...
+            echo "== staticcheck"; sh ../deploy/staticcheck.sh
+            echo "== the GitHub release step, against a fake API"; python3 -I ../deploy/test_github_release.py'
       - name: the same for deploy/release-sign
         run: |
           docker run --rm --cpus 6 -e GOMAXPROCS=6 -v "$PWD":/src -w /src/deploy/release-sign \
@@ -80,9 +83,20 @@ jobs:
           test "$(git rev-parse HEAD)" = "$WANT"
       - name: no person's data in the tree
         # Reads files, runs nothing from the tree. The term check needs the
-        # maintainers' key, which a CI job never holds, so here it is skipped and
-        # the run says so; the entity-id, provenance, email and phone checks run.
-        run: python3 scripts/test_no_personal_data.py
+        # maintainers' key and the entity-id check the reserved ids, which a CI
+        # job never holds, so the job names files that do not exist: both
+        # classes are skipped and the run prints a NOTE for each; the
+        # provenance, email and phone checks run. Left at their defaults, a
+        # path that exists but cannot be read fails the run instead.
+        run: |
+          none="$PWD/.ci-holds-none/none"
+          NO_PERSONAL_DATA_KEY_FILE="$none" RESERVED_ENTITY_IDS_FILE="$none" python3 scripts/test_no_personal_data.py
+      - name: every test the claims table cites exists
+        # docs/claims.md names, for each security claim, the test that fails
+        # if the claim stops being true. A renamed or deleted test would leave
+        # the claim looking covered; this fails the run instead. Reads files
+        # and greps; runs nothing from the tree.
+        run: bash deploy/check-claims.sh
       - name: no Host source file names in the tree
         # The member repository publishes this tree as a release was built,
         # and its release step refuses a path into the Host's Python
@@ -117,8 +131,14 @@ jobs:
       - name: every fuzz target, 5 seconds each
         # A short pass on every commit. Run deploy/fuzz.sh with its default of
         # 20 seconds before changing a wall a host's input goes through.
+        # fuzz.sh does not fail on Go's own -fuzztime report, on the Go
+        # releases whose fuzzer reports its deadline as a failure, and fails on
+        # anything else; test_fuzz.sh checks both, with planted failing
+        # inputs, before the targets run, and fails once the image's Go no
+        # longer needs the workaround.
         run: |
           docker run --rm --cpus 6 -e GOMAXPROCS=6 -v "$PWD":/src -w /src \
             -e GOTOOLCHAIN=local -e CGO_ENABLED=0 "$GO_IMAGE" sh -euc '
             apk add -q --no-cache bash
+            bash deploy/test_fuzz.sh
             deploy/fuzz.sh 5'

.github/workflows/ci.yml

@@ -0,0 +1,248 @@
+# The checks of .gitea/workflows/ci.yml, for the copy of this repository on
+# GitHub. Gitea reads .gitea/workflows when it exists and then ignores this
+# directory, so each forge runs one of the two files, never both.
+#
+# The jobs, image and commands are the same as there. Every job runs in the Go
+# toolchain image deploy/build.sh builds releases with, pinned by the same
+# digest, so CI and the release use the same compiler. GitHub does not let a
+# job's container name come from env, so the digest is written out in each
+# job; it must match GO_IMAGE in deploy/build.sh and .gitea/workflows/ci.yml.
+# No action from the marketplace is used: the commit is fetched with git from
+# this repository's public URL, so nothing outside this file and the pinned
+# image runs. The job token can only read.
+name: ci
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+    branches: [main]
+
+permissions:
+  contents: read
+
+env:
+  # The pull request's own head, or the pushed commit.
+  WANT: ${{ github.event.pull_request.head.sha || github.sha }}
+  REPO_URL: ${{ github.server_url }}/${{ github.repository }}.git
+  GOTOOLCHAIN: local
+  CGO_ENABLED: "0"
+
+defaults:
+  run:
+    shell: sh -eu {0}
+
+jobs:
+  test:
+    runs-on: ubuntu-24.04
+    timeout-minutes: 30
+    container:
+      image: golang@sha256:27f2d3b83cfc5803c8078f8e730af0f7c431bf067afaec8f60a36fb763187cba
+    env:
+      PRIVATAE_REQUIRE_AGE: "1"
+    steps:
+      - name: tools
+        # pdftotext, python3, setpriv and git are what the extract and
+        # peer-identity tests call; without them those tests fail rather than
+        # skip. Alpine's git has no http-backend (it is in git-daemon), which
+        # the forge-token test serves its repository with.
+        run: apk add -q --no-cache poppler-utils python3 setpriv curl git git-daemon
+      - name: fetch the commit
+        # The checkout belongs to the runner's user and the job runs as root,
+        # so git is told the directory is safe; go build reads it to stamp the
+        # binary with the commit.
+        run: |
+          set -o pipefail
+          git config --global --add safe.directory "$PWD"
+          git init -q .
+          git fetch -q --depth=1 "$REPO_URL" "$WANT"
+          git checkout -q --detach FETCH_HEAD
+          test "$(git rev-parse HEAD)" = "$WANT"
+      - name: age
+        # The backup tests check this code against the age command itself: a
+        # pinned age release, its archive checked against the SHA-256 below,
+        # and PRIVATAE_REQUIRE_AGE makes those tests fail rather than skip if
+        # it is missing.
+        run: |
+          case "$(uname -m)" in
+            x86_64)  a=amd64; sum=cbe24006683f8eb669266162894b9a522a1af52f2665fbc63a4bb032ed26ac10 ;;
+            aarch64) a=arm64; sum=6b8dc4333c53a5a57c9e5834e3a48f92605d7154014cd07269ff3327db5d37f4 ;;
+            *) echo "no pinned age release for $(uname -m)"; exit 1 ;;
+          esac
+          curl -fsSL -o /tmp/age.tgz "https://github.com/FiloSottile/age/releases/download/v1.3.2/age-v1.3.2-linux-$a.tar.gz"
+          echo "$sum  /tmp/age.tgz" | sha256sum -c -
+          tar -xzf /tmp/age.tgz -C /tmp age/age
+          /tmp/age/age --version
+          echo "AGE_BIN=/tmp/age/age" >> "$GITHUB_ENV"
+      - name: gofmt, go vet on linux/darwin/windows, build, go test
+        working-directory: client-go
+        run: |
+          echo "== gofmt"; out="$(gofmt -l .)"; [ -z "$out" ] || { echo "$out"; exit 1; }
+          for os in linux darwin windows; do echo "== go vet ($os)"; GOOS=$os go vet ./...; done
+          echo "== go build"; go build -o /tmp/privatae-bastion .
+          echo "== go test"; go test -count=1 ./...
+      - name: the same for deploy/release-sign
+        working-directory: deploy/release-sign
+        run: |
+          out="$(gofmt -l .)"; [ -z "$out" ] || { echo "$out"; exit 1; }
+          for os in linux darwin windows; do GOOS=$os go vet ./...; done
+          go test -count=1 ./...
+      - name: staticcheck on linux, darwin and windows
+        # The pinned staticcheck release, built with this toolchain from the
+        # module proxy and checked against the checksum database.
+        run: sh deploy/staticcheck.sh
+      - name: the release workflow's GitHub step, against a fake API
+        run: python3 -I deploy/test_github_release.py
+
+  personal-data:
+    runs-on: ubuntu-24.04
+    timeout-minutes: 10
+    container:
+      image: golang@sha256:27f2d3b83cfc5803c8078f8e730af0f7c431bf067afaec8f60a36fb763187cba
+    steps:
+      - name: tools
+        run: apk add -q --no-cache bash git python3
+      - name: fetch the commit
+        run: |
+          set -o pipefail
+          git config --global --add safe.directory "$PWD"
+          git init -q .
+          git fetch -q --depth=1 "$REPO_URL" "$WANT"
+          git checkout -q --detach FETCH_HEAD
+          test "$(git rev-parse HEAD)" = "$WANT"
+      - name: this file's image is the image releases are built with
+        # Each job here names the digest itself, so a change to GO_IMAGE in
+        # deploy/build.sh that missed this file fails here.
+        run: |
+          want="$(sed -n 's/^GO_IMAGE="\(golang@sha256:[0-9a-f]\{64\}\)"$/\1/p' deploy/build.sh)"
+          test -n "$want"
+          test "$(grep -c "^      image: $want\$" .github/workflows/ci.yml)" -eq 3
+          test "$(grep -c '^      image: ' .github/workflows/ci.yml)" -eq 3
+          grep -q "^  GO_IMAGE: $want\$" .gitea/workflows/ci.yml
+          grep -q "^      GO_IMAGE: $want\$" .github/workflows/ci.yml
+      - name: no person's data in the tree
+        # Reads files, runs nothing from the tree. The term check needs the
+        # maintainers' key and the entity-id check the reserved ids, which
+        # this job never holds, so both are skipped and the run prints a NOTE
+        # for each; the provenance, email and phone checks run.
+        run: python3 -I scripts/test_no_personal_data.py
+      - name: every test the claims table cites exists
+        # docs/claims.md names, for each security claim, the test that fails
+        # if the claim stops being true. A renamed or deleted test would leave
+        # the claim looking covered; this fails the run instead. Reads files
+        # and greps; runs nothing from the tree.
+        run: bash deploy/check-claims.sh
+      - name: no Host source file names in the tree
+        # The member repository publishes this tree as a release was built,
+        # and its release step refuses a path into the Host's Python
+        # (server/... or brain/... .py): this repo has no such files, so the
+        # path can only name another program's source. git grep exits 1 when
+        # nothing matches and above 1 when it could not search; only 1 passes.
+        run: |
+          rc=0
+          git grep -nIE 'server/[a-z_/]+\.py|brain/[a-z_/]+\.py' || rc=$?
+          case "$rc" in
+            1) ;;
+            0) echo "these lines name another program's source file; describe it instead" >&2
+               exit 1 ;;
+            *) echo "git grep could not search the tree (exit $rc), so it was not checked" >&2
+               exit 1 ;;
+          esac
+
+  fuzz:
+    needs: test
+    runs-on: ubuntu-24.04
+    timeout-minutes: 20
+    container:
+      image: golang@sha256:27f2d3b83cfc5803c8078f8e730af0f7c431bf067afaec8f60a36fb763187cba
+    steps:
+      - name: tools
+        run: apk add -q --no-cache bash git
+      - name: fetch the commit
+        run: |
+          set -o pipefail
+          git config --global --add safe.directory "$PWD"
+          git init -q .
+          git fetch -q --depth=1 "$REPO_URL" "$WANT"
+          git checkout -q --detach FETCH_HEAD
+          test "$(git rev-parse HEAD)" = "$WANT"
+      - name: the fuzz runner fails on a failing input and on nothing else
+        # fuzz.sh does not fail on Go's own -fuzztime report, on the Go
+        # releases whose fuzzer reports its deadline as a failure, and fails on
+        # anything else; this checks both, with planted failing inputs, before
+        # the targets run, and fails once the image's Go no longer needs the
+        # workaround.
+        run: bash deploy/test_fuzz.sh
+      - name: every fuzz target, 5 seconds each
+        # A short pass on every commit. Run deploy/fuzz.sh with its default of
+        # 20 seconds before changing a wall a host's input goes through.
+        run: bash deploy/fuzz.sh 5
+
+  claims-on-a-machine:
+    # The claim tests a container cannot run: the op and action-op sandboxes
+    # (root, systemd units, overlayfs, a cgroup with the pids controller),
+    # the egress tests in private user and network namespaces, and the root
+    # command watch (nftables). They run here on the runner's own virtual
+    # machine, as root, because that is what they test: the bastion as a
+    # networked machine's system service. The machine is GitHub's and is
+    # discarded after the job. The toolchain is the one in the pinned image.
+    # Every test docs/claims.md cites is run, and the job fails if any of
+    # them skips.
+    needs: test
+    runs-on: ubuntu-24.04
+    timeout-minutes: 30
+    # This job runs on the runner itself, not in the image, and Ubuntu
+    # 24.04's sh is dash 0.5.12-6ubuntu5, which has no pipefail (Debian added
+    # it in 0.5.12-7): `set -o pipefail` there is a fatal error. bash has it.
+    defaults:
+      run:
+        shell: bash -eu {0}
+    env:
+      GO_IMAGE: golang@sha256:27f2d3b83cfc5803c8078f8e730af0f7c431bf067afaec8f60a36fb763187cba
+    steps:
+      - name: fetch the commit
+        run: |
+          set -o pipefail
+          git init -q .
+          git fetch -q --depth=1 "$REPO_URL" "$WANT"
+          git checkout -q --detach FETCH_HEAD
+          test "$(git rev-parse HEAD)" = "$WANT"
+      - name: tools, and the pinned toolchain on the machine
+        run: |
+          sudo apt-get update -qq
+          sudo DEBIAN_FRONTEND=noninteractive apt-get install -y -qq --no-install-recommends nftables iproute2 poppler-utils >/dev/null
+          # The forge-token claim test needs git 2.46 or later; Ubuntu 24.04's
+          # own git is 2.43.
+          if ! git version | awk '{split($3, v, "."); exit !(v[1] > 2 || (v[1] == 2 && v[2] >= 46))}'; then
+            sudo add-apt-repository -y ppa:git-core/ppa >/dev/null
+            sudo DEBIAN_FRONTEND=noninteractive apt-get install -y -qq git >/dev/null
+          fi
+          git version
+          docker create --name go "$GO_IMAGE" >/dev/null
+          sudo docker cp go:/usr/local/go /opt/go-pinned
+          docker rm go >/dev/null
+          /opt/go-pinned/bin/go version
+      - name: age
+        # The same pinned age release as the test job.
+        run: |
+          case "$(uname -m)" in
+            x86_64)  a=amd64; sum=cbe24006683f8eb669266162894b9a522a1af52f2665fbc63a4bb032ed26ac10 ;;
+            aarch64) a=arm64; sum=6b8dc4333c53a5a57c9e5834e3a48f92605d7154014cd07269ff3327db5d37f4 ;;
+            *) echo "no pinned age release for $(uname -m)"; exit 1 ;;
+          esac
+          curl -fsSL -o "$RUNNER_TEMP/age.tgz" "https://github.com/FiloSottile/age/releases/download/v1.3.2/age-v1.3.2-linux-$a.tar.gz"
+          echo "$sum  $RUNNER_TEMP/age.tgz" | sha256sum -c -
+          tar -xzf "$RUNNER_TEMP/age.tgz" -C "$RUNNER_TEMP" age/age
+          echo "AGE_BIN=$RUNNER_TEMP/age/age" >> "$GITHUB_ENV"
+      - name: every cited test, as root
+        working-directory: client-go
+        run: |
+          set -o pipefail
+          tests="$(grep -o '_test\.go`: `[A-Za-z0-9_]*`' ../docs/claims.md | sed 's/.*`: `//; s/`$//' | sort -u | paste -sd'|' -)"
+          test -n "$tests"
+          sudo env PATH="/opt/go-pinned/bin:$PATH" HOME="$HOME" AGE_BIN="$AGE_BIN" PRIVATAE_REQUIRE_AGE=1 GOTOOLCHAIN=local CGO_ENABLED=0 \
+            GOFLAGS=-mod=readonly GOMODCACHE="$HOME/go/pkg/mod" GOCACHE="$HOME/.cache/go-build" \
+            go test -count=1 -json -run "^($tests)\$" . > "$RUNNER_TEMP/claims.json" || {
+              grep -h '"Action":"fail"' "$RUNNER_TEMP/claims.json" | head -40; exit 1; }
+      - name: none of them skipped
+        run: bash deploy/check-claims.sh --skipped "$RUNNER_TEMP/claims.json"

.github/workflows/release.yml

@@ -0,0 +1,67 @@
+# Attaches each release's binaries, its signed manifest and the manifest's
+# signature to the release on GitHub, for the copy of this repository there.
+# Gitea reads .gitea/workflows when it exists and then ignores this directory.
+#
+# Nothing here is trusted to say what the release is. deploy/release-assets.sh
+# rebuilds the binaries from the release tag with the tag's own pinned build,
+# writes the release manifest from them, and fetches the signature published
+# for that manifest, which only the release key holder can make and which is
+# named by the manifest's digest. Unless the rebuilt bytes are exactly the
+# signed release, there is no signature by that name and nothing is attached.
+# The release key never comes near this workflow.
+#
+# A tag push runs it once the release is published. Run it by hand
+# (workflow_dispatch, from main) to attach the files to an earlier release
+# whose tag is here; its signature is still published, so that works for any
+# release from 0.53.2 on.
+#
+# No action from the marketplace is used: the commit is fetched with git from
+# this repository's public URL, and the build runs in the Go toolchain image
+# deploy/build.sh pins by digest. The job's token can write the release and
+# is passed only to the last step.
+name: release
+
+on:
+  push:
+    tags: ['v[0-9]+.[0-9]+.[0-9]+']
+  workflow_dispatch:
+    inputs:
+      tag:
+        description: the release tag, vX.Y.Z
+        required: true
+
+permissions:
+  contents: read
+
+defaults:
+  run:
+    shell: bash -euo pipefail {0}
+
+jobs:
+  assets:
+    runs-on: ubuntu-24.04
+    timeout-minutes: 40
+    permissions:
+      contents: write
+    env:
+      TAG: ${{ github.event_name == 'workflow_dispatch' && inputs.tag || github.ref_name }}
+      REPO_URL: ${{ github.server_url }}/${{ github.repository }}.git
+    steps:
+      - name: fetch this workflow's commit and the release tags
+        run: |
+          git init -q .
+          git fetch -q --tags "$REPO_URL" "$GITHUB_SHA"
+          git checkout -q --detach FETCH_HEAD
+          test "$(git rev-parse HEAD)" = "$GITHUB_SHA"
+      - name: rebuild the release and check it against its signature
+        run: deploy/release-assets.sh "$TAG" "$RUNNER_TEMP/assets"
+      - name: attach the files to the release
+        # The one step with the token. Its shell, git and Python are named by
+        # path, and Python runs isolated (-I), so nothing an earlier step put
+        # on PATH or in the environment chooses what runs here.
+        shell: /usr/bin/bash -euo pipefail {0}
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          /usr/bin/git show "refs/tags/$TAG:CHANGELOG.md" > "$RUNNER_TEMP/CHANGELOG.md"
+          /usr/bin/python3 -I deploy/github-release.py "$GITHUB_REPOSITORY" "$TAG" "$RUNNER_TEMP/assets" "$RUNNER_TEMP/CHANGELOG.md"

CHANGELOG.md

@@ -23,6 +23,613 @@ From 0.39.1 on, a heading marks a version *not published* or *held* once
 that is settled. The newest entries may not be published yet; the manifest
 says which version is. Older entries are not marked.
 
+## 0.58.0
+
+- **A Host answers four common questions about a machine itself, instead of
+  handing the person a command to run.** Each is a read-only observe op,
+  granted under `observe` like the rest of the family and listed in the ops
+  the machine advertises, so a grant can name it. Each value is checked
+  against a written rule before anything is read or run, and a refusal says
+  which rule it broke without repeating the value. Off Linux each answers
+  `unsupported`.
+  - `observe:folders <path> [depth]`: the 25 largest folders under an
+    absolute path, 1 to 3 levels down (2 when not given), and the total, on
+    that path's own filesystem only, counted the way `du -x` counts (disk
+    blocks, a file with several links once). Only folders are named, never a
+    file. What could not be read, and filesystems mounted under the path, are
+    listed against the folder that holds them, with a reason and a count.
+    `/proc`, `/sys`, `/dev`, a relative path, a path containing `..` and a
+    link that leads into those are refused. A walk that runs out of the
+    20-second budget answers with what it measured and says every size is at
+    least that. The walk follows no link: a folder swapped for one while it
+    reads, or reached through one put above it, is listed as changed and not
+    counted. A read that never returns (a network mount that stopped
+    answering) is answered without at the budget, and while four such reads
+    are still waiting no new walk starts.
+  - `observe:journal <unit> [minutes]`: one systemd unit's journal for the
+    last 1 to 60 minutes (15 when not given), up to 5000 lines and 200 KB,
+    newest kept. A name with no type is a service, as with `systemctl`. The
+    name must be in systemd's unit-name alphabet and one this machine has,
+    loaded now or installed; otherwise it is refused, with the units whose
+    names come close. Until now a Host could read logs only from a Loki,
+    which most machines do not run, so it asked the person to run
+    `journalctl` instead.
+  - `observe:logins [hours]`: sshd's failed and accepted logins over the last
+    1 to 168 hours (24 when not given). Failed attempts are counted and ranked
+    by source address and by user name, 20 of each; the last 20 accepted
+    logins come with user, source, method and time. A user name the other
+    side sent is cut to 32 characters, and anything outside the characters
+    account names use reads `?`. No raw line, key fingerprint or password is
+    returned. Only lines written by sshd running as root are read, so another
+    account on the machine cannot add login lines or bury the real ones under
+    its own.
+  - `observe:cert <port> [server name]`: the certificate this machine's own
+    listener on a TCP port serves: subject, names, issuer, dates, days left,
+    and whether the chain verifies against this machine's roots for that
+    name. The port must be one `ss` shows listening here. The handshake goes
+    to loopback first and otherwise only to an address of this machine the
+    listener is bound to, never to one another machine could hold.
+- **A journal the bastion cannot read is said, not shown as empty.**
+  `journalctl` run by an account outside the journal's groups shows only that
+  account's own lines, which would look like a quiet machine. `observe:journal`
+  and `observe:logins` first check that the bastion can read the system
+  journal (as root, which is how the networked install runs it, or as a member
+  of `systemd-journal`) and otherwise answer with the reason;
+  `observe:logins` then reports its counts as -1, never 0.
+- **A grant from now on can be held to a window and a rate, by the
+  machine.** An op grant's bounds may now carry `_window` (days, from, to and
+  a zone, read on the machine's own clock; a window may wrap midnight) and
+  `_rate` (at most so many applies in any 24 hours, and so many seconds
+  between two). An apply outside them does not run, spends no use and is no
+  strike toward the quarantine: the answer is `not_now`, exit 75, with the
+  bound that held it and when it next could run. A dry run is never held, and
+  an apply with arguments outside the op is still refused as before. The
+  zone database is in the binary, so a zone reads the same on every machine.
+  An older bastion refuses both keys as naming no param, so a grant carrying
+  them fails closed there; the enforcing report says `grant_run_bounds` on a
+  bastion that holds them.
+- **The machine keeps the runs, and who approved.** Each grant's record in
+  `ops.json` now holds the times of the applies that spent a use, the key
+  that approved it (the approval device's key, or the factor enrolled at this
+  console) and the machine's own seal over the record. The rate is counted
+  from that log, so it holds across a restart. A record edited outside the
+  bastion holds no grant until it is dropped and approved again, and so does
+  a grant installed by an older bastion (none was held when this shipped). A
+  grant whose approving key is no longer pinned is refused; `unpin-device`
+  drops every grant that key approved, so pinning it again brings none back.
+- **`op_grant_drop <grant | all>` takes a grant back from the Host.** It is
+  a tighten: no quarantine holds it and no approval is asked. A grant not
+  held here is already gone and is answered as dropped. At the machine,
+  `ops drop --all` takes back every op grant a Host holds, beside `ops drop
+  <grant>`.
+- **An action op may have no program.** With `interpreter: none` and no
+  `run` file, an action op's verbs are its whole work. Nothing the Host wrote
+  executes, so it needs no sandbox and runs on a machine with no overlay. Its
+  dry run is the verbs' own read of what each would move, and its apply the
+  verbs and the log checks. Such an op that ships a `run` file is refused,
+  as is one that names destinations, a builder, reads, or a credential no
+  verb uses. An op with a program is still refused where the sandbox is not.
+  A verb whose tool the machine lacks refuses the run as the machine's own
+  state, never a strike, and the enforcing report now says, for each verb,
+  whether this machine can run it (`verbs`).
+- **On your own machine, an action op reaches the internet, never your
+  network.** On a machine no fleet has anchored, an action op's destination
+  is never a private, ULA, CGNAT or reserved address (also inside NAT64 or
+  6to4): one written as an address is refused when the op arrives, and a
+  name is refused at apply if any address it resolves to is one. Another
+  computer of yours is reached through a bastion on that computer. An
+  anchored machine keeps private destinations, since a fleet's ship steps
+  reach its own private control planes. A verb's fetch is held the same way
+  (`tree.checkout` fetching its tree's repository, `k8s.set_image` fetching
+  its source to check ancestry): the repository's host is checked once and
+  git is held to that address, so an op with no program reaches no more than
+  one with a program. On an anchored machine too a verb's fetch is held to
+  its checked address and is never this machine, its loopback, a link-local
+  address or a cloud's metadata service. An image `k8s.set_image` would put
+  live, which the machine's container runtime pulls by name, is refused on
+  your own machine when its registry is on your network or this machine,
+  both when a program reports it and when the verb would set it.
+  `k8s.set_image` now says it needs git as well as kubectl, so a machine
+  without git is told before a run. What stops working: a ship step on your
+  own unanchored bastion that deploys to another computer on your LAN,
+  fetches from a forge on it or rolls an image from a registry on it; and a
+  verb's fetch needs git 2.37 or newer, which holds a fetch to one address.
+
+## 0.57.0
+
+- **Pairing with a Cloud Host is one pasted line.** A Host's "Set up a
+  bastion" card now gives a pairing code, `skp_…`, that carries the token,
+  the Host's enroll address and the SHA-256 of the Host's CA. `pair <code>`
+  (or the code on standard input after `pair -`) takes the CA from the enroll
+  port's own TLS chain, keeps it only under that fingerprint and only when the
+  port's certificate verifies under it for `host.cseries.internal`, then pairs
+  exactly as `--redeem-url` with `--host-ca` does. There is no certificate
+  file to save, so the step people skipped ("open host-ca.pem: The system
+  cannot find the file specified") is gone. `--host-ca` still works as before.
+- **The installers pair and start in one go.** `install.sh` with
+  `PRIVATAE_PAIR_CODE='<code>'` in its environment (the card's line is
+  `curl -fsSL https://privatae.ai/bastion/install.sh | PRIVATAE_PAIR_CODE='skp_…' sh`)
+  and `install.ps1 -Pair '<code>'` install or update the bastion as before,
+  hand the code to `pair -` on standard input, then start the bastion, or restart it when it
+  was running, and end with one plain sentence and `status`. On Windows that
+  is the logon task (every running copy is stopped first, so the new one gets
+  its control port), or the service on a networked machine; on a machine whose
+  bastion is already the Windows service, `-Pair` pairs that service from an
+  Administrator PowerShell instead of installing a second bastion. On Linux as
+  yourself it is a systemd user unit, with lingering asked for and the outcome
+  said; on macOS a LaunchAgent; as root, the system service. Without the code
+  both installers behave as they did.
+- **The code is in no process's arguments.** On macOS and Linux it comes
+  only from the environment: an argument would sit in the installer shell's
+  command line for the whole install, where any user of the machine can read
+  it with `ps`, and one at the same address could redeem it first. A
+  process's environment is readable only by its owner and root. `install.sh`
+  reads the variable before it starts any program and unsets it, so nothing
+  it starts inherits it. An empty or partial code is refused before the
+  download. `install.sh` now names no argument it refuses, so a code given in
+  the wrong place is never printed back. On Windows the code stays a
+  parameter of the script, inside the person's own PowerShell.
+- **On Linux, as yourself, the copy that serves is the one the installer
+  started.** Before the user unit starts, every copy of your bastion running
+  outside it (left in a terminal, or from another path) is stopped, as
+  install.ps1 does on Windows. Otherwise the unit's copy only waits behind the
+  old one's run lock, and the old copy never reads the new pairing. The
+  installer then checks for a few seconds that the unit is running and keeps
+  the same process. A unit that keeps restarting is reported as not running,
+  with its log command, and the install fails.
+- **A local bastion can serve a second Host, when you say so at the machine.**
+  A local bastion used to serve one Host and refuse any other, so a person
+  with two Hosts of their own (a Cloud machine beside a platform Host) could
+  not pair the second at all. `pair` now asks on this machine's terminal,
+  `/dev/tty` or the Windows console, never standard input, so a code piped
+  in by an installer still needs a person to answer: "This computer already
+  serves your Host … Add … as well? [y/N]". Only y or yes adds it; the first
+  pairing is untouched; with no terminal it is refused. Other callers
+  (`claim-anchor`, the pairing op) keep the refusal. On a local machine that
+  serves two Hosts, each changes only the servers it built and stops only the
+  coder runs it started. The name in the question is the token's, so anything
+  in it but a plain host name is shown quoted and escaped.
+- **`pair` exits 3 when the Host is already paired here.** Nothing changes, and
+  the installers report it as already connected. Over TCP, with a pairing
+  code or `--host-ca`, a Host of the same name under another CA (Cloud names
+  repeat) is a plain refusal, not exit 3, before the redeem and after it. A
+  pairing code that names a field twice is refused.
+- **An enroll port that does not answer says why within 10 seconds.** A Cloud
+  Host opens its pairing port only to the address the browser used when the
+  card was opened, for 15 minutes; from anywhere else the port does not
+  answer, and `pair` used to wait for the system's connect timeout. It now
+  stops after 10 seconds with that reason and what to do.
+- **The local-mode refusal names this system's way to a networked bastion**
+  (`install.ps1 -Networked` from an Administrator PowerShell on Windows,
+  `install.sh --networked` as root elsewhere) and says that the Host already
+  paired would have to be paired again with it. It used to name the Linux
+  script on Windows.
+
+## 0.56.0
+
+- **New: a Host with no platform can use the model servers on this
+  computer.** A Host running on this computer, or in a virtual machine on it,
+  can reach the model servers here (Ollama, a local language model server, a
+  personal-data filter) by name through the bastion, and is told which of
+  them it can use. It starts working once your Host supports it.
+- **Changed: for a Host with no platform, model management reaches only this
+  computer.** Listing, downloading and removing models for such a Host now
+  reach the model server set here only when that server is on this computer.
+  One set to another computer is refused, and the bastion's log says which
+  setting is not used and what to set. Such a Host is also told only of
+  the model servers it can use, and no longer of the service slots, or of
+  coder runs when it is in a virtual machine, since neither runs for it.
+- **Changed: on a local machine, the internet lane reaches only public
+  addresses.** A Host can no longer use a local machine's internet lane (the
+  egress lane) to reach devices on your home or office network at their
+  private addresses, such as 192.168.x.x or 10.x.x.x, or the private ranges
+  many VPNs use, as was already the case on an anchored machine. Your own
+  machines are reached through their approved lanes. A device that has a
+  public address, as many do over IPv6, is still reached at that address.
+- **Changed: plain-HTTP requests through the bastion are checked one per
+  connection.** On the internet lane and a coder run's web proxy, a
+  plain-HTTP (not HTTPS) connection now carries one request, which is checked
+  before it is sent on; anything sent after it closes the connection. Before,
+  later requests on the same connection reached the destination unchecked. A
+  malformed or ambiguous request is refused, and every refusal now says why.
+- **Safer: file access in a folder you shared stays inside it.** When a Host
+  lists, reads or moves files in a folder you shared with it, the bastion
+  stays inside that folder. If the folder, or any folder on its path, is later
+  swapped for a link to somewhere else, the bastion refuses until the folder
+  is shared again, and a file or folder inside it swapped for a link is never
+  followed out of it. Search, which goes through the folder's own search
+  agent, is not part of this change.
+- **Safer: the bastion limits how much it holds of what a Host sends.** On
+  the internet lane, the terminal and root commands, coder runs, the memory
+  mirror (the copy of your Host's database kept here) and the bastion's
+  check-in with its Host, a message from the Host is read only up to a set
+  size and refused past it, or passed along or dropped as it arrives. A Host
+  can no longer make the bastion hold a message of any size in memory. A
+  coder run's web proxy now carries at most 1024 connections at once, and the
+  memory mirror at most 8.
+- **Safer: asking to share a folder with a Host has limits.** Any user on the
+  computer can send that ask to the bastion, so each ask is now read only up
+  to a set size, and the number of asks open at once is capped, per user and
+  in total; past the cap, the ask is told the bastion is busy and to try
+  again.
+- **For contributors: the fuzz check in CI no longer fails on a timing bug in
+  Go's own fuzzer.** A real failure still fails it. The program itself is
+  unchanged by this.
+
+## 0.55.3
+
+- **A pairing change reaches every Host on the machine at once.** Each pairing
+  beats its Host every 60 seconds, and the beat is how every Host on the
+  machine learns that a pairing went active or was revoked. Approving or
+  revoking a pairing changed it on disk immediately but triggered no beat, so
+  every Host kept the old state for up to a minute: a connection request to a
+  Host the anchor had just approved was refused as not on a shared machine. Now
+  a grant that makes a pairing active, and a revoke, make every running
+  pairing beat right away. A refused grant, or a re-sync of a pairing that is
+  already active, beats nothing.
+
+## 0.55.2
+
+- **A search reply says how its text was sanitized.** Every successful reply
+  from the bastion agent (search, read range, catalog) was stamped
+  `sanitized: true`, while the agent the bastion deploys runs in `flag` mode,
+  which replaces e-mail addresses, phone numbers, card, bank and ID numbers,
+  IP and MAC addresses and keys, and leaves names, street addresses and
+  dates of birth in. A reply now carries `sanitize_mode` (the mode this
+  bastion wrote into the agent's manifest, or for an agent run by hand the
+  mode it reports, else `unknown`), `sanitized` true only for `sterilize`,
+  and `sanitize_note` saying what the mode leaves in. A reply cannot raise
+  the mode this bastion wrote.
+- **The README says what the bastion needs.** It said "no Docker, no Python".
+  The bastion itself needs neither, and the README now names the optional
+  features that do: coder runs, the service slot and the memory mirror need a
+  container runtime, folder search needs Python 3.8 or later, Python ops need
+  `/usr/bin/python3`, and the Linux accessibility reader needs python3 with
+  AT-SPI. It also said Go 1.25; `client-go/go.mod` needs 1.26.
+- **GitHub releases carry their files.** A new workflow,
+  `.github/workflows/release.yml`, runs on each release tag (and by hand for
+  an earlier one): `deploy/release-assets.sh` rebuilds the release from its
+  tag, writes the release manifest from the rebuilt binaries with
+  `deploy/manifest.py` (which `publish.sh` now uses too, so the two cannot
+  differ), and fetches the signature published for that manifest. Only when
+  the signature verifies are the six binaries, `version.json`, its signature
+  and `SHA256SUMS` attached to the release. Releases 0.53.2, 0.54.0 and
+  0.54.1 were rebuilt this way and each matched its signed manifest.
+- **The lab harness is out of the release build.** `mode=lab` drove one test
+  lab's own cluster and database, and its layout was in the published source.
+  It is now built only with `-tags lab` and is not in the published source; a
+  release bastion answers a lab run with exit 126, as a bastion without a lab
+  always did.
+- **staticcheck is clean** on linux, darwin and windows, for both Go modules,
+  and CI runs it (`deploy/staticcheck.sh`, a pinned staticcheck release, with
+  the per-OS runs merged so code used on one OS only is not reported unused).
+  Unused code is removed and error strings no longer end in punctuation; no
+  behaviour changes.
+- **The claim tests a container cannot run now run on GitHub.** 25 tests the
+  claims table cites (the op and action-op sandboxes, the egress tests in
+  private namespaces, the root command watch) skipped in CI, because a CI
+  container has no systemd, no namespaces of its own and no nftables. A new
+  job runs every cited test as root on GitHub's discarded virtual machine and
+  fails if any skips; `deploy/check-claims.sh --skipped` names them.
+
+## 0.55.1
+
+- **A root op's tasks are capped.** An op's tasks were capped at 64 with
+  `RLIMIT_NPROC`, which the kernel never applies to root, so a root op could
+  start as many tasks as it liked until its time limit. A root op now runs
+  in a systemd scope of its own in `privatae_ops.slice` with `TasksMax=96`,
+  which counts every task in it, root's included: the op's 64 and the
+  bastion's two processes that start it. Those two run with one CPU's worth
+  of Go runtime threads (`GOMAXPROCS=1`), so they take the same share of the
+  cap on a machine of any size: with the runtime set for 64 CPUs they took
+  44 of the 96, and set for 256 no root op could start. The scope's process
+  reads its own cgroup's `pids.max` before the op starts, and the op's
+  report carries `tasks_capped` only when that is the cap. Where no scope
+  can be made (no systemd, no pids controller, or `systemd-run` failing),
+  the op runs as before and its report's `tasks_note` says its tasks were
+  not capped, and why. A user op's tasks are held by `RLIMIT_NPROC` as
+  before, and its report says so. A scope is killed with the run at the
+  time limit, when the bastion stops, and by the startup sweep. Two ops can
+  run at once, so an action op's memory watch now reads each other op's
+  unit or scope in the slice as another program: it yields when a root op
+  beside it makes the machine wait on memory, as it did when root ops ran
+  in the bastion's own cgroup, and its own unit still never counts.
+
+## 0.55.0
+
+- **A local machine's egress lane never reaches the machine itself.** Before,
+  on a machine with no anchor the lane vetted no destination, so a host could
+  dial this machine's own services through it: `POST
+  http://127.0.0.1:11434/api/pull` reached ollama whatever the `models` switch
+  said. It now refuses `localhost` names, loopback and unspecified addresses
+  (`0.0.0.0/8` and `::` included), link-local addresses (the cloud metadata
+  address included), the rest of `64:ff9b::/32`, the deprecated IPv6 forms
+  below, and any address this machine takes as its own, as the anchored lane
+  checks them, also inside an IPv4-mapped, NAT64 or 6to4 address. The LAN
+  stays reachable: a private, CGNAT or reserved address that is not this
+  machine's is dialled. Every address a name resolves to is checked, and only
+  those addresses are dialled, never the name again.
+- **A container or VM on this machine is this machine, not the LAN.** On
+  Linux, an address the kernel routes straight onto a container's veth, or
+  onto a bridge whose every port is a veth, a tap or a dummy (the bridge
+  networks Docker, Podman, LXD and libvirt make for their containers and
+  VMs), is refused as this machine. A route through a gateway on such a link
+  leaves, and a link an IPv4 or IPv6 default route of the main table goes
+  out of (any nexthop's, for a multipath default) is never inside: a bastion
+  that is itself in a container reaches the internet through the machine it
+  runs on, and dials the network on its own veth, its LAN included, as
+  before. Before, the local egress lane dialled such an address as LAN,
+  so a host could reach a service in a container here, `172.17.0.2:11434`
+  for one, though the same service on loopback was refused. A bridge with a
+  network card, a bond or a VLAN among its ports is still the LAN. Every lane
+  reads the same check, so a coder run's proxy, the anchored lane and the
+  backup lanes now also refuse a public address on such a network; their
+  private addresses were refused already. macOS and Windows have no route
+  lookup here, so there such an address is still dialled.
+- **Every lane refuses the deprecated IPv6 forms.** IPv4-compatible `::/96`,
+  IPv4-translated `::ffff:0:0/96` and site-local `fec0::/10` are no
+  destination of their own, and where each is delivered depends on the
+  platform. A coder run's proxy and the backup lanes already refused them; the
+  egress lane now does too, anchored or local, by address and in any answer a
+  name resolves to. The anchored lane refuses nothing else it did not refuse
+  before. All the lanes now read one address classifier.
+- **Both egress lanes try each vetted address, as Happy Eyeballs does.** A
+  name with more than one address is dialled at each in order, the next
+  starting as soon as an attempt fails or when none has answered within
+  300 ms, and the first to answer is used; the others are cancelled, and one
+  that connects anyway is closed. Before, the anchored lane dialled only the
+  first address, and a coder run's proxy gave each address an even share of
+  the dial budget, so an address whose packets go nowhere held up the next
+  one for seconds. The next one now starts within 300 ms, about what the OS
+  dialler took when it was given the name.
+- **`pair` says to enrol a second factor after every pairing that has none.**
+  It said so only after it anchored a networked machine, so a local pairing,
+  one over TCP or a unix socket, and a later host on an anchored machine were
+  told nothing, and on a local machine a folder grant then installs on the
+  host's signed grant alone. The line names `--host` when more than one host
+  is paired, and `sudo -H` only when `pair` itself ran with machine authority,
+  so a local install is told to enrol as its own user. A pairing that keeps
+  its factor (a `--replace`), or a machine with an approval device pinned, is
+  not told.
+- **A root-command grant lasts at most 30 days.** It was a year. `root-exec
+  grant --for 31d` is refused and says the longest term; grant again before
+  it ends. A grant given under the longer term ends 30 days after it was
+  given, and `root-exec` shows that end. A grant file with no grant time is
+  refused, since it cannot be held to the term.
+- **SECURITY.md, the README's license section and `docs/host-on-this-machine.md`
+  name each kind of Host.** They described only a Host on Privatae's platform.
+  They now say that a C-Series Host runs on your own machine in its own Linux
+  virtual machine, is open source (Privatae/cseries, AGPL-3.0-only), needs no
+  Privatae account, and that `npx cseries` sets the bastion up networked,
+  paired, anchored and with a second factor on every pairing (the commands are
+  shown); that C-Series Cloud and a network membership are optional and paid;
+  and that a Host on Privatae's own platform is the case where the Host's
+  software is not open source.
+- **Security reports go to security@privatae.ai,** the address
+  privatae.ai/.well-known/security.txt names. SECURITY.md said info@.
+- **`docs/models.md` says what stops a host changing your models on a local
+  machine.** It told you to scope the pairing on your host, which a local
+  machine never reads. There, switching `models` off for the host is what
+  filters the ollama forward, and the egress lane is no way around it. On an
+  anchored machine the forward filters for every pairing, which the page now
+  says too.
+- **A tunnel endpoint the host sends must be one host and port.** The
+  pairing response's `endpoint` and `endpoints`, and the endpoints in a
+  heartbeat's rotate directive, were written into the WireGuard
+  configuration unchecked. That configuration is one setting per line, so an
+  endpoint with a newline in it added settings of its own:
+  `203.0.113.7:51820` followed by a newline and `allowed_ip=::/0` let the
+  tunnel carry every IPv6 address, not only the host's. An endpoint must now
+  be an IP address or a host name with a port from 1 to 65535, and nothing
+  else, no space or control character. A pairing response with any other is
+  refused and nothing is saved, and a rotate directive with one is not acted
+  on. A WireGuard key must be 32 bytes. The tunnel's peer is now configured in
+  one place, which checks every value it writes.
+- **A pairing file that holds such an endpoint is listed but never served.**
+  An older build may have saved one. Its tunnel and lanes never come up, and
+  `run` names the file and says why. Every other pairing still loads and
+  serves, `status` shows the pairing as `invalid` with the same reason, and
+  `unpair` removes it, as does pairing that host again with `pair --replace`.
+  One such file no longer stops `run`, `status` and `unpair` for the whole
+  machine.
+- **Such a pairing cannot hold the anchor or be approved, and no host is told
+  of it.** `claim-anchor` marked it active and anchored the machine to it; it
+  now refuses, says why, and changes nothing. An approval from the anchor
+  host marked it active and started it with the endpoint its file holds; it
+  is now refused. The heartbeat and the anchor host's pairing list leave it
+  out: a host knows only the states active, pending and revoked, and could
+  show a pairing in any other state as active.
+- **A harness run's agent credential is readable only by the harness.** The
+  credential file was readable by every user on the machine (0644), in a
+  folder anyone could open and write (0777), for the length of the run. A
+  bastion running as root now hands the folder (0700) and the file (0600) to
+  the user the harness image runs its agent as (uid 10005). One that is not
+  root cannot hand them over, so it leaves them open for the agent, and the
+  folder sits inside one that only the bastion's own user can enter.
+- **A harness run's tree is out of other users' reach.** In a worktree run the
+  tree mounted at `/work` was opened to every user (0777), its top folder
+  included, in the shared temporary folder. Any local user could read or
+  change the Host's tree during the run. It now sits in a folder only the
+  bastion's own user can enter, as a git run's tree already did, and a run
+  refuses a tree whose folder other users can enter. Opening the tree for the
+  agent no longer follows a link in it. Before, on a bastion running as root,
+  a link in the repository, to `/etc/shadow` for example, made the linked
+  file world-writable. A file is opened as 0666, or 0777 if it was
+  executable, so a git run no longer commits every file as executable.
+- **A worktree run returns only what is in its tree.** A file swapped for a
+  link while the tree was being packed, or a folder swapped for a link, was
+  read through the link, so a bastion running as root could return any file
+  on the machine to the Host. Each name is now opened inside the folder
+  already open above it, never through a link, and goes in only if it is the
+  plain file or folder that was looked at. On Linux a device or pipe swapped
+  in for a file in that gap is not opened either: opening some devices does
+  something by itself. Links, pipes, devices and sockets are left out
+  (before, each came back as an empty entry of its kind), and the run's
+  result names them in `worktree_skipped`, with `worktree_skipped_count`.
+- **A coding agent can no longer run code as the bastion through the
+  repository's git settings.** In a git run the bastion committed and pushed
+  in the agent's own folder, `.git` included, so a hook, a file system
+  monitor, a filter, a signing program or a changed remote the agent set ran
+  as the bastion's own user, which can be root, or sent the push, with its
+  token, to another server. The bastion now makes its own fresh clone, in a
+  folder the agent cannot reach, copies in the agent's plain files and
+  folders (never a link, never anything of its `.git`), and commits and
+  pushes only there. The branch gets one commit with the agent's final
+  files; the agent's own commits are not pushed one by one. A link the
+  agent made is left out and named in `worktree_skipped`; a link the
+  repository already holds stays as it is. Every git the bastion runs for a
+  harness or a mirror runs with no hooks, no file system monitor, no `ext::`
+  transport, and no system or user git settings of the machine.
+- **The agent in a git run no longer holds the push token, and no process
+  argument shows it.** The agent's clone named the repository with the
+  run's token in it, so the agent could push straight to the repository
+  during the run, past the bastion's copy, and every git the bastion ran
+  for the run had the token in its arguments, which any local user can
+  read. Every git call now names the repository without the token and gets
+  it from its environment, as an `Authorization` header sent to that
+  repository's address only; no clone's config holds it, and a run is
+  refused if the token is in any file of the agent's clone's `.git`.
+- **A user-privilege op is no longer refused because other programs run as
+  its user.** An op's tasks are capped at 64, and the cap counted every task
+  the machine ran as the op's user (`privatae-session`, or `nobody`, which
+  system daemons share), so on a busy machine a user op could not start or
+  could not fork at all. The op now runs in a user namespace of its own,
+  as the same user and seeing every file's real owner, where the cap counts
+  only its own tasks. Where the machine has no user namespaces, the op runs
+  as before.
+- **A harness run's git calls carry the harness identity.** git made one
+  from this machine's user and host name instead, which cost a host name
+  lookup per call (seconds where there is no resolver) and wrote this
+  machine's name into the clone handed to the agent.
+- **A run whose agent container cannot be removed uses nothing of its
+  tree.** A failed removal was ignored, so an agent still running could go
+  on writing the tree while the bastion read it. The run now fails, says
+  so, and copies, returns and pushes nothing.
+- **A git run whose commit fails pushes nothing and fails.** A failed add or
+  commit was ignored and the branch pushed anyway, and the run read as
+  successful. It now ends the run with an error and no push. A run in which
+  the agent changed no file pushes no branch and reports `changed: false`.
+- **Repositories that keep files in Git LFS work in git runs again.** No
+  user git settings are read any more, so the LFS filter `git lfs install`
+  sets there was lost. When `git-lfs` is installed, the bastion now supplies
+  the standard LFS filter itself on its own git commands, sends the large
+  files before the push, and sends them only to the repository's own LFS
+  address, whatever a `.lfsconfig` the agent wrote names. A filter of any
+  other name in a `.gitattributes` still runs nothing. Tested with git-lfs
+  3.7.1.
+- **Names that open as `.git` on Windows or macOS are left out of a git
+  run's copy.** Besides `.git` in any letter case, the copy now leaves out
+  `.git` followed by dots or spaces or a `:`, the short name `GIT~1`, and
+  `.git` with the characters HFS+ ignores, the names git itself refuses, so
+  no file the agent made can be written into the clean clone's own `.git`.
+- **A run's tree has a size limit, and a file is never read into memory
+  whole.** A file of any size in the tree, even one that takes no disk
+  space, was read whole into memory, so an agent could make the bastion run
+  out of memory. A worktree run now returns at most 9 MiB of tree, which is
+  what fits in one message to the Host beside the transcript; a larger tree
+  is refused with `worktree_too_large` and none of it is returned. A git run
+  copies at most 2 GiB, and refuses a larger tree with `tree_too_large`
+  before it pushes anything; each file and folder counts 512 bytes and its
+  name besides its content, so a tree of millions of empty files is refused
+  too, and the copy keeps nothing per file in memory. Of the entries left
+  out, the count is kept in full and the first 100 names. A folder is read
+  512 names at a time, each counted as it is read, so a folder of millions
+  of names is refused without being read whole first. A result too large for one message to the Host
+  is refused with `result_too_large` instead of being lost on the way.
+- **The local console shows a pairing that is not served as "invalid".** A
+  pairing whose file holds an endpoint this build refuses showed as active
+  on the console when its stored state was empty, as it is in older files.
+  It now reads as the setup page and `status` read it.
+- **Built with Go 1.26.10, golang.org/x/net 0.61.0, x/crypto 0.58.0 and x/sys
+  0.49.0.** `govulncheck` finds 13 known vulnerabilities this code reaches in
+  Go 1.25.14 and x/net 0.57.0, which 0.54.1 was built with: in crypto/x509
+  (intermediate certificates since found compromised were still trusted),
+  crypto/tls, net/http and HTTP/2, net/textproto, os and html/template. With
+  these versions it finds none. The
+  build image is still pinned by digest, and CI runs in the same one. The
+  module now needs Go 1.26, as x/net's fixed release does. A published release
+  is still rebuilt by `deploy/verify-release.sh` with the toolchain its own
+  tag pins.
+- **`go test ./...` passes for a contributor who is not root.** A root
+  command now starts in `/` when the process cannot enter root's home, as it
+  already did when that home is missing; the bastion runs these commands only
+  as root, so on a machine nothing changes, and the root-command tests now run
+  as any user. `TestTreeCheckoutMovesOnlyForward`, which needs a tree root
+  owns, skips as another user, and `TestClaimAnchorWithNoToken` checks the
+  refusal without machine authority and then skips the rest, where it passed
+  having checked nothing more.
+- **The tests name no folder of our own build machine.** The sandbox tests'
+  fixture folder is `/srv/privatae-optest` (`/tmp` cannot hold an op's
+  effects), and the action tests' run note asks for any folder outside
+  `/tmp`.
+- **How this repository is published (docs only).** From this release on,
+  GitHub receives each change merged since the last release as its own commit,
+  titled as the change was, instead of one commit per release. The README,
+  CONTRIBUTING.md and SECURITY.md say so, and say that releases from 0.53.2 on
+  can be verified from a clone, since the earlier tags are not public.
+- **The personal-data check reads what it used to miss.** A listed term
+  written right after a backslash escape (`\n`, `\t`, `\u00e9`) is found, and
+  so is any value inside a run of base64 or base64url that decodes to text.
+  The reserved entity ids are read from `/etc/privatae/reserved-entity-ids`
+  (one per line, `RESERVED_ENTITY_IDS_FILE` to point elsewhere) instead of
+  being written in the script; where that file is absent, as in public CI,
+  the entity-id check is skipped and the run prints a NOTE, as the term check
+  already does without its key. A file that holds no id, or a line that is
+  not one, fails the run.
+- **The former-name check reads comments and Markdown.** A kept identifier
+  is set aside only in Go code; in a comment, this changelog or any other
+  file it is a hit like the bare word, and the two entries that spelled the
+  device-fingerprint prefix out now name `fpSANPrefix` instead.
+- **CI on GitHub.** `.github/workflows/ci.yml` runs the jobs of
+  `.gitea/workflows/ci.yml` on GitHub Actions in the same Go image, pinned by
+  the same digest, with no action from the marketplace and a token that can
+  only read. A step fails if its digest stops matching `deploy/build.sh`.
+- Tests use documentation addresses (`2001:db8::/32`, `203.0.113.0/24`) and
+  neutral paths, and comments state their reason instead of citing a pull
+  request by number; a test keeps published source free of such numbers.
+
+## 0.54.1
+
+- **No change to what the bastion does.** The README is now a short front
+  page: install, quick start, local and networked modes, a summary of the
+  security model, how this repository is published, and how to build and
+  verify a release. The reference text moved into one page per topic under
+  `docs/`, listed in `docs/README.md`, with only its headings and links
+  changed. The README and CONTRIBUTING.md now say that issues and pull
+  requests are welcome on GitHub.
+- **The claims table is its own file, `docs/claims.md`,** and
+  `deploy/check-claims.sh` reads it there. Two rows that cited two tests each
+  now cite one test per row, so the check passes again.
+
+## 0.54.0
+
+- **Every audit row also goes to the system log.** The rows written to
+  `mgmt_audit.jsonl`, `pairing_audit.jsonl` and each pairing's `audit.jsonl`
+  (terminals, root commands, store ops, op runs) are also written to the
+  machine's system log, so a log collector can keep them where no host can
+  write: syslog facility `auth`, tag `privatae-bastion-audit` (the journal on
+  systemd) on Linux and macOS, and the Application event log, source
+  `PrivataeBastionAudit`, on Windows, registered when the system service is
+  installed and removed with it. Each line is the file's row with `audit`
+  (the file) and `pairing` added, and a root command's text left out (its
+  digest and length stay). The lines go through a bounded queue, so a slow or
+  stuck system log never holds up the audited op; a full queue drops lines and
+  the next line carries `system_log_dropped`. The file stays the record.
+- **A server is networked and anchored by default.** `install.sh` run as root
+  now installs the networked bastion (systemd on Linux, launchd on macOS);
+  `--local` keeps a root install local, a reinstall over a machine already
+  paired in local mode keeps it local unless `--networked` is passed, and run
+  as your own user it is local as before. On a networked machine with no anchor, the first `pair` run with
+  machine authority anchors it to that host. Before, such a machine refused
+  every op until `claim-anchor` was run by hand; now it is anchored at once:
+  every terminal and every write waits for the person's approval, and any
+  later host for the anchor's. `pair` then says to
+  enrol the second factor. `claim-anchor` with no token on a machine already
+  anchored to that host now succeeds and changes nothing (an installer that
+  anchors after pairing finds the work done); naming another host is still
+  refused without `--rotate`. The Windows installer is unchanged: `-Networked`
+  as before.
+
 ## 0.53.2
 
 - **No change to what the bastion does.** Two code comments named source files
@@ -183,13 +790,13 @@ says which version is. Older entries are not marked.
 - **The device fingerprint is sent under a second name to a Host with no
   platform.** A pairing with such a Host (`--redeem-url unix://…` or
   `https://…`) sends an operator certificate request that carries the
-  fingerprint as two URI SANs: first `urn:skema:hwfp:<fp>` (`fpSANPrefix`),
-  then `urn:privatae:hwfp:<fp>`, which that Host reads. A platform pairing's
-  request names it once, under `urn:skema:hwfp:`, as before: the platform
-  signs a request with that one name and refuses one with any other. A
-  signed accounting verifies with the fingerprint under either name, and a
-  signer certificate that names two different fingerprints is refused rather
-  than resolved.
+  fingerprint as two URI SANs: first under the original device-fingerprint
+  URN prefix (`fpSANPrefix` in pair.go), then `urn:privatae:hwfp:<fp>`, which
+  that Host reads. A platform pairing's request names it once, under
+  `fpSANPrefix`, as before: the platform signs a request with that one name
+  and refuses one with any other. A signed accounting verifies with the
+  fingerprint under either name, and a signer certificate that names two
+  different fingerprints is refused rather than resolved.
 
 ## 0.52.0
 

CONTRIBUTING.md

@@ -7,10 +7,15 @@ their approval on that machine.
 
 ## What you can do today
 
-Sign-in on this forge is not reachable from the internet, so you cannot file
-an issue or open a pull request here. You can read, clone, build and verify a
-release, and fork under the GPL. Report a bug or a security problem by email
-(see [SECURITY.md](SECURITY.md)).
+Issues and pull requests are welcome on GitHub, at
+[github.com/Privatae/bastion](https://github.com/Privatae/bastion).
+Development happens in a private repository, so an accepted change is carried
+into the development repository and arrives here with the next release, as
+its own commit. Sign-in on the git.privatae.ai forge is not reachable from
+the internet, so issues and pull requests cannot be opened there. You can
+also read, clone, build and verify a release, and fork under the GPL. Report
+a security problem by email, not in a public issue (see
+[SECURITY.md](SECURITY.md)).
 
 ## Checks a change has to pass
 
@@ -20,8 +25,15 @@ gofmt -l .            # must print nothing
 go vet ./...
 CGO_ENABLED=0 go build .
 go test ./...
+cd .. && deploy/staticcheck.sh   # must find nothing, on linux, darwin and windows
 ```
 
+Some claim tests need a real machine and root: the op sandboxes, the
+namespace tests and the root command watch skip in a container and as a
+user. GitHub's CI runs every test `docs/claims.md` cites as root on its own
+virtual machine and fails if one skips; to run them yourself, use a machine
+you can throw away.
+
 The backup tests also check this code against the age command itself when
 one is on PATH or named by `AGE_BIN`; CI installs a pinned age release and
 sets `PRIVATAE_REQUIRE_AGE`, so there they fail rather than skip.

SECURITY.md

@@ -4,7 +4,7 @@
 
 Please do not open a public issue for a security problem.
 
-Email **info@privatae.ai** with:
+Email **security@privatae.ai** with:
 
 - the version (`privatae-bastion version`) and platform,
 - what you found and how to reproduce it,
@@ -22,13 +22,15 @@ notes if you wish.
   machine do more than the grant and the local approval allow: reading files
   outside a granted store, deleting where only archiving is allowed, running a
   program that is not in the fixed table, bypassing the second factor, opening
-  a listener that is not documented in the README's security model.
+  a listener that is not documented in the security model
+  ([docs/security-model.md](docs/security-model.md)).
 
 ## What is out of scope
 
-- The Privatae platform the bastion connects to. It is separate software with
-  its own reporting path; reports about it sent to the address above are
-  forwarded.
+- The Host the bastion connects to. It is separate software: a C-Series Host
+  is [Privatae/cseries](https://github.com/Privatae/cseries), and a Host on
+  Privatae's own platform is software that is not open source. Reports about
+  either sent to the address above are forwarded.
 - Findings that require root on the machine already. The bastion's root ops
   assume the machine's administrator is trusted.
 
@@ -100,7 +102,7 @@ one installed.
 **Checking our work.** Release builds are reproducible.
 `deploy/verify-release.sh` fetches the published manifest, checks its
 signature, checks out the release's source, rebuilds it, and compares every
-binary's digest. From 0.43.0 on, each public snapshot in this repository
+binary's digest. From 0.53.2 on, each published release in this repository
 carries the annotated tag `v<version>`, whose message records the release's
 `built_from`. That field names the commit in the repository the release was
 built in, which is not a commit here, so the script uses the tag and prints
@@ -113,9 +115,30 @@ the folder you granted. The agent is a separate program in its own repository
 (`privatae/bastion-agent`), and it reads your files in order to index them. What
 leaves your machine is what its manifest and your grant allow. Its extraction
 and PII redaction happen inside it, so check that repository, not this one,
-for how well those work.
+for how well those work. The bastion runs it in `flag` mode, which does not
+find names, street addresses or dates of birth, and every search reply says so
+([DCN stores](docs/context.md#distributed-context-network-dcn-stores)).
 
-**The platform.** Your host runs on our platform, which is not open source. The
-bastion keeps the host to your grant and your approvals whatever the platform
-does. What the host does with what you share with it is governed by software
-you cannot read here.
+**The Host.** The bastion holds your Host to your grant and your approvals.
+What the Host does with what you share with it is the Host's own software, and
+which software that is depends on where the Host runs.
+
+- **A C-Series Host** runs on your own machine, in its own Linux virtual
+  machine. Its source is [Privatae/cseries](https://github.com/Privatae/cseries),
+  under AGPL-3.0-only; the bastion and the bastion agent are GPL-3.0-or-later.
+  `npx cseries` installs the bastion beside it as the machine's system service
+  and, on every pairing, runs `set-mode networked`, then
+  `pair --redeem-url <the Host's https enroll address> --host-ca <its CA> -`
+  with the token on standard input, then `claim-anchor` and
+  `stepup-enroll --console-only --host <the Host>`: the machine is anchored to
+  that Host and you approve its writes here with your own authenticator. No
+  Privatae account is needed
+  ([A Host in a VM on this machine](docs/host-on-this-machine.md#a-host-in-a-vm-on-this-machine)).
+- **Optional and paid, from Privatae:** C-Series Cloud, which runs a Host on a
+  server of its own that Privatae sets up and keeps patched, and a Privatae
+  network membership, which joins a Host on your own machine to the Privatae
+  network.
+- **A Host on Privatae's own platform** is the case where the Host's software
+  is not open source. There, the bastion keeps the Host to your grant and your
+  approvals whatever the platform does, and what the Host does with what you
+  share with it is governed by software you cannot read here.

client-go/agentowner.go

@@ -0,0 +1,433 @@
+package main
+
+import (
+	"bytes"
+	"errors"
+	"os/exec"
+	"strconv"
+	"strings"
+)
+
+// WHO RUNS AN AGENT, NOT ONLY WHICH ACCOUNT.
+//
+// The process scan names the account an agent runs as (procUser on every
+// platform, winTokenUser on Windows) and says nothing else about it. The
+// question an assessment asks is about the account: is it an administrator, is
+// it a person's own account or one that exists for a service, and is that
+// person logged in right now. "An agent runs as a person, holds administrator
+// rights, and nobody is at the keyboard" is a different finding from "an agent
+// runs under a service account with no rights", and the account name alone
+// cannot tell them apart.
+//
+// Three facts, each with an honest third value:
+//
+//	owner_class        domain_user, local_user, service, system, or unknown
+//	owner_admin        yes, no, or unknown
+//	owner_interactive  yes, no, or unknown
+//
+// UNKNOWN IS NEVER FOLDED INTO NO. A box that could not read the group list
+// has not learned that the account is unprivileged, and saying "no" there is
+// the confident-wrong answer this plane refuses everywhere else
+// (baseline_identity.go draws the same line for sudo rules it could not read).
+//
+// Every decision is in this file, with no build tag, so the Linux test run
+// exercises the macOS and Windows rules too. The platform files
+// (agents_<goos>.go) only read: /proc and /etc on Linux, `id`, `dscl` and
+// `who` on macOS, the process token and the session list on Windows.
+//
+// These ride on agentObs, so they travel only where `agents` does: inside the
+// heartbeat's `observing` gate (agents.go).
+
+const (
+	ownerClassDomain  = "domain_user"
+	ownerClassLocal   = "local_user"
+	ownerClassService = "service"
+	ownerClassSystem  = "system"
+	ownerUnknown      = "unknown"
+	ownerYes          = "yes"
+	ownerNo           = "no"
+)
+
+// ownerFacts is the three answers for one account.
+type ownerFacts struct {
+	Class       string
+	Admin       string
+	Interactive string
+}
+
+// unknownOwner is the answer when nothing could be read: a process whose
+// owner this bastion may not inspect, or a build with no reader.
+var unknownOwner = ownerFacts{Class: ownerUnknown, Admin: ownerUnknown, Interactive: ownerUnknown}
+
+func yesNo(b bool) string {
+	if b {
+		return ownerYes
+	}
+	return ownerNo
+}
+
+// ── Linux and macOS ─────────────────────────────────────────────────────────
+
+// unixOwner is what a Linux or macOS reader collected about one account.
+// Every field that can be unreadable carries its own "known" flag.
+type unixOwner struct {
+	UID  int // -1 when the uid could not be read
+	Name string
+	// LocalRecord: "yes" when the account is in this machine's own account
+	// database (/etc/passwd; the macOS local directory node), "no" when it is
+	// not, "unknown" when that could not be read.
+	LocalRecord string
+	// DirectorySource: this machine resolves accounts from a directory
+	// (nsswitch names sss, ldap, winbind or nis; a macOS mobile account was
+	// copied from one). Only then does "not in the local database" mean a
+	// domain account rather than a uid with no account at all.
+	DirectorySource bool
+	// SystemUIDMax is the highest uid the platform hands to system accounts.
+	SystemUIDMax int
+	Groups       []string // group names the account or process holds
+	GroupsKnown  bool
+	AdminGroups  map[string]bool
+	// SudoApplies: on this platform a sudo rule can grant administration to
+	// any user or group (Linux), so "in no admin group" is only a "no" once
+	// the rules were read as well. macOS leaves it false and decides on the
+	// admin group, which is what its stock sudoers grants.
+	// SudoNamed / SudoKnown: a sudo rule names the account directly or
+	// through a group (sudoPrincipals), and whether the rules were readable.
+	SudoApplies bool
+	SudoNamed   bool
+	SudoKnown   bool
+	// Sessions is the set of account names with a login session now
+	// (logind, utmp, `who`); SessionsKnown is false when none could be read.
+	Sessions      map[string]bool
+	SessionsKnown bool
+}
+
+// linuxSystemUIDMax is shadow-utils' default SYS_UID_MAX: accounts below
+// UID_MIN (1000 by default) are system accounts. login.defs can move both;
+// linuxUIDLimits reads it.
+const linuxSystemUIDMax = 999
+
+// darwinSystemUIDMax: macOS creates people's accounts from uid 501 up, and
+// keeps the ids below for its own daemons (_www, _spotlight and the rest).
+const darwinSystemUIDMax = 500
+
+func classifyUnixOwner(o unixOwner) ownerFacts {
+	f := ownerFacts{Class: ownerUnknown, Admin: ownerUnknown, Interactive: ownerUnknown}
+	switch {
+	case o.UID == 0:
+		f.Class = ownerClassSystem
+	case o.UID == -1:
+		// Unread uid: the class cannot be told from a name alone. Other
+		// negative ids are real: macOS's directory gives nobody -2.
+	case o.UID == 65534 || o.UID <= o.SystemUIDMax:
+		// nobody, and every uid in the system range, is an account that exists
+		// for a program rather than a person.
+		f.Class = ownerClassService
+	case strings.ContainsAny(o.Name, `@\`):
+		// user@realm (SSSD) and DOMAIN\user (winbind) are directory names.
+		f.Class = ownerClassDomain
+	case o.LocalRecord == ownerYes:
+		f.Class = ownerClassLocal
+	case o.LocalRecord == ownerNo && o.DirectorySource:
+		f.Class = ownerClassDomain
+	}
+
+	switch {
+	case o.UID == 0:
+		f.Admin = ownerYes
+	case o.GroupsKnown && hasAnyGroup(o.Groups, o.AdminGroups):
+		f.Admin = ownerYes
+	case o.SudoKnown && o.SudoNamed:
+		f.Admin = ownerYes
+	case o.GroupsKnown && (o.SudoKnown || !o.SudoApplies):
+		f.Admin = ownerNo
+	}
+
+	if o.SessionsKnown && o.Name != "" {
+		f.Interactive = yesNo(o.Sessions[o.Name])
+	}
+	return f
+}
+
+func hasAnyGroup(groups []string, want map[string]bool) bool {
+	for _, g := range groups {
+		if want[g] {
+			return true
+		}
+	}
+	return false
+}
+
+// linuxAdminGroups: the groups a stock distribution's sudoers grants
+// administration to (sudo on Debian and Ubuntu, wheel on Fedora, RHEL and
+// Arch, admin on older Ubuntu).
+var linuxAdminGroups = map[string]bool{"sudo": true, "wheel": true, "admin": true}
+
+// darwinAdminGroups: macOS administrators are members of `admin`.
+var darwinAdminGroups = map[string]bool{"admin": true}
+
+// parseStatusIDs reads the real and effective uid and the supplementary gid
+// list out of /proc/<pid>/status. ok is false when there is no Uid line.
+func parseStatusIDs(status string) (realUID, effUID int, gids []int, ok bool) {
+	realUID, effUID = -1, -1
+	for _, line := range strings.Split(status, "\n") {
+		switch {
+		case strings.HasPrefix(line, "Uid:"):
+			f := strings.Fields(line)
+			if len(f) >= 3 {
+				r, err1 := strconv.Atoi(f[1])
+				e, err2 := strconv.Atoi(f[2])
+				if err1 == nil && err2 == nil {
+					realUID, effUID, ok = r, e, true
+				}
+			}
+		case strings.HasPrefix(line, "Groups:"):
+			for _, g := range strings.Fields(strings.TrimPrefix(line, "Groups:")) {
+				if n, err := strconv.Atoi(g); err == nil {
+					gids = append(gids, n)
+				}
+			}
+		}
+	}
+	return realUID, effUID, gids, ok
+}
+
+// parseGroupNames maps gid -> name from an /etc/group body.
+func parseGroupNames(body string) map[int]string {
+	out := map[int]string{}
+	for _, line := range strings.Split(body, "\n") {
+		p := strings.Split(strings.TrimSpace(line), ":")
+		if len(p) < 3 || strings.HasPrefix(p[0], "#") {
+			continue
+		}
+		if n, err := strconv.Atoi(p[2]); err == nil {
+			out[n] = p[0]
+		}
+	}
+	return out
+}
+
+// parsePasswdUIDs is the set of uids /etc/passwd itself lists.
+func parsePasswdUIDs(body string) map[int]bool {
+	out := map[int]bool{}
+	for _, line := range strings.Split(body, "\n") {
+		p := strings.Split(strings.TrimSpace(line), ":")
+		if len(p) < 3 || strings.HasPrefix(p[0], "#") {
+			continue
+		}
+		if n, err := strconv.Atoi(p[2]); err == nil {
+			out[n] = true
+		}
+	}
+	return out
+}
+
+// parseLoginDefsSysMax returns the highest system uid login.defs declares:
+// SYS_UID_MAX when present, else UID_MIN-1, else the shadow-utils default.
+func parseLoginDefsSysMax(body string) int {
+	sysMax, uidMin := -1, -1
+	for _, line := range strings.Split(body, "\n") {
+		f := strings.Fields(strings.TrimSpace(line))
+		if len(f) < 2 || strings.HasPrefix(f[0], "#") {
+			continue
+		}
+		n, err := strconv.Atoi(f[1])
+		if err != nil {
+			continue
+		}
+		switch f[0] {
+		case "SYS_UID_MAX":
+			sysMax = n
+		case "UID_MIN":
+			uidMin = n
+		}
+	}
+	switch {
+	case sysMax >= 0:
+		return sysMax
+	case uidMin > 0:
+		return uidMin - 1
+	}
+	return linuxSystemUIDMax
+}
+
+// nsswitchHasDirectory reports whether the passwd line of nsswitch.conf names
+// a directory source, so an account missing from /etc/passwd is resolved from
+// one rather than simply absent.
+func nsswitchHasDirectory(body string) bool {
+	for _, line := range strings.Split(body, "\n") {
+		line = strings.TrimSpace(line)
+		if !strings.HasPrefix(line, "passwd:") {
+			continue
+		}
+		for _, src := range strings.Fields(strings.TrimPrefix(line, "passwd:")) {
+			switch src {
+			case "sss", "ldap", "winbind", "nis", "nisplus":
+				return true
+			}
+		}
+	}
+	return false
+}
+
+// parseLogindUserState reads systemd-logind's per-user state file
+// (/run/systemd/users/<uid>). active and online mean the user has a session
+// now; lingering means services run for them with no session.
+func parseLogindUserState(body string) (interactive bool, ok bool) {
+	for _, line := range strings.Split(body, "\n") {
+		if v, found := strings.CutPrefix(strings.TrimSpace(line), "STATE="); found {
+			switch v {
+			case "active", "online":
+				return true, true
+			case "lingering", "closing", "offline":
+				return false, true
+			}
+			return false, false
+		}
+	}
+	return false, false
+}
+
+// utmp record layout on glibc Linux (bits/utmp.h): 384 bytes, ut_type a
+// 16-bit value at offset 0, ut_user 32 bytes at offset 44. USER_PROCESS (7)
+// is a login session.
+const (
+	utmpRecordSize  = 384
+	utmpUserOffset  = 44
+	utmpUserSize    = 32
+	utmpUserProcess = 7
+)
+
+// parseUtmpUsers is the set of account names with a USER_PROCESS record.
+func parseUtmpUsers(b []byte) map[string]bool {
+	out := map[string]bool{}
+	for off := 0; off+utmpRecordSize <= len(b); off += utmpRecordSize {
+		rec := b[off : off+utmpRecordSize]
+		if int(rec[0])|int(rec[1])<<8 != utmpUserProcess {
+			continue
+		}
+		name := rec[utmpUserOffset : utmpUserOffset+utmpUserSize]
+		if i := bytes.IndexByte(name, 0); i >= 0 {
+			name = name[:i]
+		}
+		if len(name) > 0 {
+			out[string(name)] = true
+		}
+	}
+	return out
+}
+
+// parseWhoUsers is the set of account names `who` lists (macOS): the first
+// field of every line.
+func parseWhoUsers(out string) map[string]bool {
+	users := map[string]bool{}
+	for _, line := range strings.Split(out, "\n") {
+		if f := strings.Fields(line); len(f) > 0 {
+			users[f[0]] = true
+		}
+	}
+	return users
+}
+
+// parseIDGroups reads `id -Gn <user>`: group names separated by spaces.
+func parseIDGroups(out string) []string { return strings.Fields(out) }
+
+// dsclRecordMissing: dscl ran to the end and exited non-zero, which is how it
+// says the local node has no such record. A dscl the timeout killed has an
+// exit code of -1 and has said nothing about the account.
+func dsclRecordMissing(err error) bool {
+	var ee *exec.ExitError
+	return errors.As(err, &ee) && ee.ExitCode() > 0
+}
+
+// parseDsclUser reads `dscl . -read /Users/<name> UniqueID OriginalNodeName`.
+// A mobile account (cached from a directory) carries OriginalNodeName.
+func parseDsclUser(out string) (uid int, mobile bool, ok bool) {
+	uid = -1
+	for _, line := range strings.Split(out, "\n") {
+		line = strings.TrimSpace(line)
+		if v, found := strings.CutPrefix(line, "UniqueID:"); found {
+			if n, err := strconv.Atoi(strings.TrimSpace(v)); err == nil {
+				uid, ok = n, true
+			}
+		}
+		if strings.HasPrefix(line, "OriginalNodeName:") {
+			mobile = true
+		}
+	}
+	return uid, mobile, ok
+}
+
+// ── Windows ─────────────────────────────────────────────────────────────────
+
+// Well-known SIDs, from Microsoft's tables:
+// https://learn.microsoft.com/en-us/windows/win32/secauthz/well-known-sids and
+// https://learn.microsoft.com/en-us/windows-server/identity/ad-ds/manage/understand-security-identifiers
+const (
+	sidLocalSystem       = "S-1-5-18"
+	sidLocalService      = "S-1-5-19"
+	sidNetworkService    = "S-1-5-20"
+	sidAdministrators    = "S-1-5-32-544"
+	sidInteractive       = "S-1-5-4"
+	sidRemoteInteractive = "S-1-5-14"
+	sidLocalAccount      = "S-1-5-113"
+	sidNTServicePrefix   = "S-1-5-80-"
+	sidVirtualMachinePfx = "S-1-5-83-"
+)
+
+// winOwner is what the Windows reader collected from one process token.
+type winOwner struct {
+	SID      string
+	Domain   string // the account's domain as LookupAccount names it
+	Computer string // this machine's NetBIOS name
+	// Groups are the token's group SIDs, deny-only ones included: an
+	// administrator running without elevation holds Administrators as a
+	// deny-only group, and is still an administrator account.
+	Groups      map[string]bool
+	GroupsKnown bool
+	// SessionSIDs: accounts with a Remote Desktop Services session now
+	// (WTSEnumerateSessions + WTSQueryUserToken, which needs LocalSystem).
+	SessionSIDs   map[string]bool
+	SessionsKnown bool
+}
+
+func classifyWinOwner(o winOwner) ownerFacts {
+	f := ownerFacts{Class: ownerUnknown, Admin: ownerUnknown, Interactive: ownerUnknown}
+	switch {
+	case o.SID == "":
+	case o.SID == sidLocalSystem:
+		f.Class = ownerClassSystem
+	case o.SID == sidLocalService || o.SID == sidNetworkService ||
+		strings.HasPrefix(o.SID, sidNTServicePrefix) || strings.HasPrefix(o.SID, sidVirtualMachinePfx):
+		f.Class = ownerClassService
+	case o.GroupsKnown && o.Groups[sidLocalAccount]:
+		f.Class = ownerClassLocal
+	case o.Domain == "":
+	case strings.EqualFold(o.Domain, o.Computer):
+		f.Class = ownerClassLocal
+	case strings.HasPrefix(o.SID, "S-1-5-21-") || strings.EqualFold(o.Domain, "AzureAD"):
+		// A machine or domain account whose domain is not this computer:
+		// an Active Directory or Entra ID account.
+		f.Class = ownerClassDomain
+	}
+
+	switch {
+	case o.SID == sidLocalSystem:
+		// "System is a hidden member of Administrators" (the second page above).
+		f.Admin = ownerYes
+	case o.GroupsKnown:
+		f.Admin = yesNo(o.Groups[sidAdministrators])
+	}
+
+	switch {
+	case o.SessionsKnown && o.SID != "":
+		f.Interactive = yesNo(o.SessionSIDs[o.SID])
+	case o.GroupsKnown && (o.Groups[sidInteractive] || o.Groups[sidRemoteInteractive]):
+		// No session list (the bastion is not LocalSystem), but this very
+		// process was started by an interactive logon, so the account has
+		// one. The converse proves nothing: a scheduled task's token is not
+		// interactive while its owner may well be logged in.
+		f.Interactive = ownerYes
+	}
+	return f
+}

client-go/agentowner_linux_test.go

@@ -0,0 +1,83 @@
+//go:build linux
+
+package main
+
+import (
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+// The Linux reader end to end over an invented machine: a /proc with two
+// processes and an /etc and /run that say who they belong to. Nothing here is
+// read from the box the test runs on.
+func TestLinuxOwnerFactsReaderOverAFixtureMachine(t *testing.T) {
+	root := t.TempDir()
+	write := func(rel, body string) {
+		t.Helper()
+		p := filepath.Join(root, rel)
+		if err := os.MkdirAll(filepath.Dir(p), 0o755); err != nil {
+			t.Fatal(err)
+		}
+		if err := os.WriteFile(p, []byte(body), 0o600); err != nil {
+			t.Fatal(err)
+		}
+	}
+	write("etc/passwd", "root:x:0:0::/root:/bin/sh\npat:x:1001:1001::/home/pat:/bin/bash\n"+
+		"kim:x:1002:1002::/home/kim:/bin/bash\nsvcdemo:x:120:120::/nonexistent:/usr/sbin/nologin\n")
+	write("etc/group", "root:x:0:\nsudo:x:27:pat\npat:x:1001:\nkim:x:1002:\n")
+	write("etc/sudoers", "root ALL=(ALL:ALL) ALL\n%sudo ALL=(ALL:ALL) ALL\n")
+	write("etc/nsswitch.conf", "passwd: files\n")
+	write("run/systemd/users/1001", "UID=1001\nSTATE=active\n")
+	write("proc/400/status", "Name:\tdemo\nUid:\t1001\t1001\t1001\t1001\nGroups:\t27 1001\n")
+	write("proc/401/status", "Name:\tdemo\nUid:\t1002\t1002\t1002\t1002\nGroups:\t1002\n")
+	write("proc/402/status", "Name:\tdemo\nUid:\t120\t120\t120\t120\nGroups:\t\n")
+	write("proc/403/status", "Name:\tdemo\nUid:\t1002\t0\t1002\t1002\nGroups:\t1002\n")
+
+	oldID, oldProc := identityRoot, procRoot
+	identityRoot, procRoot = root, filepath.Join(root, "proc")
+	t.Cleanup(func() { identityRoot, procRoot = oldID, oldProc })
+
+	read := ownerFactsReader()
+	for _, c := range []struct {
+		pid  int
+		user string
+		want ownerFacts
+	}{
+		{400, "pat", ownerFacts{ownerClassLocal, ownerYes, ownerYes}},
+		{401, "kim", ownerFacts{ownerClassLocal, ownerNo, ownerNo}},
+		{402, "svcdemo", ownerFacts{ownerClassService, ownerNo, ownerNo}},
+		// Effective root under kim's account: administrator, class unchanged.
+		{403, "kim", ownerFacts{ownerClassLocal, ownerYes, ownerNo}},
+		// A pid with no status file: everything unknown, nothing guessed.
+		{404, "ghost", unknownOwner},
+	} {
+		if got := read(c.pid, c.user); got != c.want {
+			t.Errorf("pid %d (%s): got %+v, want %+v", c.pid, c.user, got, c.want)
+		}
+	}
+}
+
+// Every observation scanAgents returns carries the owner facts: the wiring,
+// not only the classifier. The one process is this test itself, so the facts
+// come from the real /proc and are only checked for being filled in.
+func TestScanAgentsCarriesOwnerFacts(t *testing.T) {
+	prev := scanProcessesFn
+	t.Cleanup(func() { scanProcessesFn = prev })
+	self := os.Getpid()
+	scanProcessesFn = func() (map[int]agentObs, error) {
+		return map[int]agentObs{self: {PID: self, User: procUser(self), Signature: "claude-code",
+			Cmdline: "claude", raw: "claude"}}, nil
+	}
+	got, err := scanAgents()
+	if err != nil || len(got) == 0 {
+		t.Fatalf("scan: %d observations, %v", len(got), err)
+	}
+	// The network signal may add this box's own model clients; every
+	// observation, however it was caught, must carry the facts.
+	for _, a := range got {
+		if a.OwnerClass == "" || a.OwnerAdmin == "" || a.OwnerInteractive == "" {
+			t.Fatalf("owner facts missing from pid %d", a.PID)
+		}
+	}
+}

client-go/agentowner_test.go

@@ -0,0 +1,209 @@
+package main
+
+import (
+	"encoding/binary"
+	"testing"
+)
+
+// Every value below is invented. No account, group or SID here belongs to a
+// real machine.
+
+func TestClassifyUnixOwner(t *testing.T) {
+	linux := func(o unixOwner) unixOwner {
+		o.SystemUIDMax = linuxSystemUIDMax
+		o.AdminGroups = linuxAdminGroups
+		o.SudoApplies = true
+		return o
+	}
+	cases := []struct {
+		name string
+		in   unixOwner
+		want ownerFacts
+	}{
+		{"root is the system and an administrator",
+			linux(unixOwner{UID: 0, Name: "root", LocalRecord: ownerYes, GroupsKnown: true, SudoKnown: true,
+				Sessions: map[string]bool{}, SessionsKnown: true}),
+			ownerFacts{ownerClassSystem, ownerYes, ownerNo}},
+		{"a daemon account in the system range",
+			linux(unixOwner{UID: 115, Name: "svc-demo", LocalRecord: ownerYes, GroupsKnown: true, SudoKnown: true,
+				Sessions: map[string]bool{}, SessionsKnown: true}),
+			ownerFacts{ownerClassService, ownerNo, ownerNo}},
+		{"nobody is a service account",
+			linux(unixOwner{UID: 65534, Name: "nobody", LocalRecord: ownerYes}),
+			ownerFacts{ownerClassService, ownerUnknown, ownerUnknown}},
+		{"a person in the sudo group, logged in",
+			linux(unixOwner{UID: 1001, Name: "pat", LocalRecord: ownerYes, Groups: []string{"pat", "sudo"},
+				GroupsKnown: true, Sessions: map[string]bool{"pat": true}, SessionsKnown: true}),
+			ownerFacts{ownerClassLocal, ownerYes, ownerYes}},
+		{"a person a sudo rule names directly",
+			linux(unixOwner{UID: 1002, Name: "lee", LocalRecord: ownerYes, Groups: []string{"lee"},
+				GroupsKnown: true, SudoNamed: true, SudoKnown: true}),
+			ownerFacts{ownerClassLocal, ownerYes, ownerUnknown}},
+		{"no admin group and the sudo rules read: no",
+			linux(unixOwner{UID: 1003, Name: "kim", LocalRecord: ownerYes, Groups: []string{"kim"},
+				GroupsKnown: true, SudoKnown: true, Sessions: map[string]bool{"pat": true}, SessionsKnown: true}),
+			ownerFacts{ownerClassLocal, ownerNo, ownerNo}},
+		{"no admin group but the sudo rules unreadable: unknown, never no",
+			linux(unixOwner{UID: 1003, Name: "kim", LocalRecord: ownerYes, Groups: []string{"kim"}, GroupsKnown: true}),
+			ownerFacts{ownerClassLocal, ownerUnknown, ownerUnknown}},
+		{"groups unreadable: unknown, never no",
+			linux(unixOwner{UID: 1004, Name: "ash", LocalRecord: ownerYes, SudoKnown: true}),
+			ownerFacts{ownerClassLocal, ownerUnknown, ownerUnknown}},
+		{"not in passwd on a machine that resolves from a directory",
+			linux(unixOwner{UID: 501234, Name: "501234", LocalRecord: ownerNo, DirectorySource: true}),
+			ownerFacts{ownerClassDomain, ownerUnknown, ownerUnknown}},
+		{"not in passwd and no directory: a uid with no account, not a domain user",
+			linux(unixOwner{UID: 501234, Name: "501234", LocalRecord: ownerNo}),
+			ownerFacts{ownerUnknown, ownerUnknown, ownerUnknown}},
+		{"a realm-qualified name is a directory account",
+			linux(unixOwner{UID: 70001, Name: "pat@corp.example", LocalRecord: ownerUnknown}),
+			ownerFacts{ownerClassDomain, ownerUnknown, ownerUnknown}},
+		{"uid unread",
+			linux(unixOwner{UID: -1, Name: ""}),
+			ownerFacts{ownerUnknown, ownerUnknown, ownerUnknown}},
+		{"macOS admin group member",
+			unixOwner{UID: 501, Name: "sam", LocalRecord: ownerYes, Groups: []string{"staff", "admin"},
+				GroupsKnown: true, SystemUIDMax: darwinSystemUIDMax, AdminGroups: darwinAdminGroups,
+				Sessions: map[string]bool{"sam": true}, SessionsKnown: true},
+			ownerFacts{ownerClassLocal, ownerYes, ownerYes}},
+		{"macOS standard user: the admin group decides",
+			unixOwner{UID: 502, Name: "jo", LocalRecord: ownerYes, Groups: []string{"staff"},
+				GroupsKnown: true, SystemUIDMax: darwinSystemUIDMax, AdminGroups: darwinAdminGroups},
+			ownerFacts{ownerClassLocal, ownerNo, ownerUnknown}},
+		{"macOS daemon",
+			unixOwner{UID: 88, Name: "_demo", LocalRecord: ownerYes, SystemUIDMax: darwinSystemUIDMax,
+				AdminGroups: darwinAdminGroups},
+			ownerFacts{ownerClassService, ownerUnknown, ownerUnknown}},
+		{"macOS mobile account",
+			unixOwner{UID: 1204567, Name: "rio", LocalRecord: ownerNo, DirectorySource: true,
+				SystemUIDMax: darwinSystemUIDMax, AdminGroups: darwinAdminGroups},
+			ownerFacts{ownerClassDomain, ownerUnknown, ownerUnknown}},
+	}
+	for _, c := range cases {
+		if got := classifyUnixOwner(c.in); got != c.want {
+			t.Errorf("%s: got %+v, want %+v", c.name, got, c.want)
+		}
+	}
+}
+
+func TestClassifyWinOwner(t *testing.T) {
+	const box = "DESK-DEMO"
+	user := "S-1-5-21-1111111111-2222222222-3333333333-1001"
+	cases := []struct {
+		name string
+		in   winOwner
+		want ownerFacts
+	}{
+		{"LocalSystem",
+			winOwner{SID: sidLocalSystem, Domain: "NT AUTHORITY", Computer: box},
+			ownerFacts{ownerClassSystem, ownerYes, ownerUnknown}},
+		{"NetworkService",
+			winOwner{SID: sidNetworkService, Domain: "NT AUTHORITY", Computer: box,
+				Groups: map[string]bool{}, GroupsKnown: true},
+			ownerFacts{ownerClassService, ownerNo, ownerUnknown}},
+		{"a virtual service account",
+			winOwner{SID: "S-1-5-80-1-2-3-4-5", Domain: "NT SERVICE", Computer: box},
+			ownerFacts{ownerClassService, ownerUnknown, ownerUnknown}},
+		{"local administrator, unelevated, at the desktop",
+			winOwner{SID: user, Domain: box, Computer: box,
+				Groups:      map[string]bool{sidAdministrators: true, sidInteractive: true},
+				GroupsKnown: true, SessionSIDs: map[string]bool{user: true}, SessionsKnown: true},
+			ownerFacts{ownerClassLocal, ownerYes, ownerYes}},
+		{"domain user, not an administrator, no session now",
+			winOwner{SID: user, Domain: "CORPDEMO", Computer: box,
+				Groups: map[string]bool{"S-1-5-3": true}, GroupsKnown: true,
+				SessionSIDs: map[string]bool{}, SessionsKnown: true},
+			ownerFacts{ownerClassDomain, ownerNo, ownerNo}},
+		{"the local-account group decides even with no domain name",
+			winOwner{SID: user, Groups: map[string]bool{sidLocalAccount: true}, GroupsKnown: true},
+			ownerFacts{ownerClassLocal, ownerNo, ownerUnknown}},
+		{"no session list, but this process is an interactive logon",
+			winOwner{SID: user, Domain: box, Computer: box,
+				Groups: map[string]bool{sidRemoteInteractive: true, sidInteractive: true}, GroupsKnown: true},
+			ownerFacts{ownerClassLocal, ownerNo, ownerYes}},
+		{"no session list and a batch token: unknown, never no",
+			winOwner{SID: user, Domain: box, Computer: box,
+				Groups: map[string]bool{"S-1-5-3": true}, GroupsKnown: true},
+			ownerFacts{ownerClassLocal, ownerNo, ownerUnknown}},
+		{"token unreadable",
+			winOwner{},
+			ownerFacts{ownerUnknown, ownerUnknown, ownerUnknown}},
+	}
+	for _, c := range cases {
+		if got := classifyWinOwner(c.in); got != c.want {
+			t.Errorf("%s: got %+v, want %+v", c.name, got, c.want)
+		}
+	}
+}
+
+func TestOwnerFileParsers(t *testing.T) {
+	status := "Name:\tdemo\nUid:\t1001\t0\t1001\t1001\nGid:\t1001\t1001\t1001\t1001\nGroups:\t27 1001 \n"
+	r, e, gids, ok := parseStatusIDs(status)
+	if !ok || r != 1001 || e != 0 || len(gids) != 2 || gids[0] != 27 {
+		t.Fatalf("status: %d %d %v %v", r, e, gids, ok)
+	}
+	if _, _, _, ok := parseStatusIDs("Name:\tdemo\n"); ok {
+		t.Fatal("a status with no Uid line must not read as uid 0")
+	}
+	names := parseGroupNames("root:x:0:\nsudo:x:27:pat\n# comment\nbad\n")
+	if names[27] != "sudo" || names[0] != "root" || len(names) != 2 {
+		t.Fatalf("groups: %v", names)
+	}
+	uids := parsePasswdUIDs("root:x:0:0::/root:/bin/sh\npat:x:1001:1001::/home/pat:/bin/bash\n")
+	if !uids[1001] || !uids[0] || uids[5] {
+		t.Fatalf("passwd: %v", uids)
+	}
+	for body, want := range map[string]int{
+		"":                                  linuxSystemUIDMax,
+		"UID_MIN 500\n":                     499,
+		"UID_MIN 1000\nSYS_UID_MAX   899\n": 899,
+		"# SYS_UID_MAX 10\n":                linuxSystemUIDMax,
+	} {
+		if got := parseLoginDefsSysMax(body); got != want {
+			t.Errorf("login.defs %q: %d, want %d", body, got, want)
+		}
+	}
+	if !nsswitchHasDirectory("group: files\npasswd:  files sss systemd\n") ||
+		nsswitchHasDirectory("passwd: files systemd\n") {
+		t.Fatal("nsswitch directory detection")
+	}
+	for body, want := range map[string][2]bool{
+		"UID=1001\nSTATE=active\nSESSIONS=3\n": {true, true},
+		"STATE=lingering\n":                    {false, true},
+		"STATE=something-new\n":                {false, false},
+		"UID=1001\n":                           {false, false},
+	} {
+		got, ok := parseLogindUserState(body)
+		if got != want[0] || ok != want[1] {
+			t.Errorf("logind %q: %v %v", body, got, ok)
+		}
+	}
+	if u := parseWhoUsers("sam      console  Oct 20 09:00\njo       ttys001  Oct 20 09:05\n"); !u["sam"] || !u["jo"] || len(u) != 2 {
+		t.Fatalf("who: %v", u)
+	}
+	uid, mobile, ok := parseDsclUser("OriginalNodeName:\n /Active Directory/DEMO\nUniqueID: 1204567\n")
+	if !ok || uid != 1204567 || !mobile {
+		t.Fatalf("dscl: %d %v %v", uid, mobile, ok)
+	}
+	if g := parseIDGroups("staff admin everyone\n"); len(g) != 3 || g[1] != "admin" {
+		t.Fatalf("id: %v", g)
+	}
+}
+
+func TestParseUtmpUsersReadsOnlyLoginSessions(t *testing.T) {
+	rec := func(typ uint16, user string) []byte {
+		b := make([]byte, utmpRecordSize)
+		binary.LittleEndian.PutUint16(b, typ)
+		copy(b[utmpUserOffset:], user)
+		return b
+	}
+	var blob []byte
+	blob = append(blob, rec(utmpUserProcess, "pat")...)
+	blob = append(blob, rec(8, "gone")...) // DEAD_PROCESS: a session that ended
+	blob = append(blob, rec(utmpUserProcess, "lee")...)
+	blob = append(blob, 1, 2, 3) // a torn tail is ignored, not misread
+	got := parseUtmpUsers(blob)
+	if !got["pat"] || !got["lee"] || got["gone"] || len(got) != 2 {
+		t.Fatalf("utmp: %v", got)
+	}
+}

client-go/agents.go

@@ -214,6 +214,14 @@ type agentObs struct {
 	// (baseline.go). Empty where the platform does not supply one: darwin and
 	// windows never populate procFacts.Unit, so those fall back to the origin.
 	Unit string `json:"unit,omitempty"`
+	// The account behind User (agentowner.go): domain_user, local_user,
+	// service, system or unknown; whether it administers this machine; and
+	// whether it has a login session now. yes, no or unknown each — unknown
+	// is what an unreadable token or group list says, never no. Empty only on
+	// an observation made outside scanAgents.
+	OwnerClass       string     `json:"owner_class,omitempty"`
+	OwnerAdmin       string     `json:"owner_admin,omitempty"`
+	OwnerInteractive string     `json:"owner_interactive,omitempty"`
 	Evidence         []evidence `json:"evidence,omitempty"`
 	raw              string     // the UNTRUNCATED command line, for classification only; never reported
 	// automation: the signature came from the automation tables, so the
@@ -304,6 +312,7 @@ func scanAgents() ([]agentObs, error) {
 	}
 	extra := originExtras()
 	facts := procFactsReader()
+	owners := ownerFactsReader()
 	out := make([]agentObs, 0, len(byPid))
 	for pid, a := range byPid {
 		if faIsOwnChild(pid) {
@@ -317,6 +326,11 @@ func scanAgents() ([]agentObs, error) {
 		if f, ok := facts(pid); ok {
 			a.Unit = f.Unit
 		}
+		// Who the owning account is, beside who it is called. Read here
+		// rather than in each platform's walk so a process the network
+		// signal alone caught (mergeConnections) gets the same answer.
+		o := owners(pid, a.User)
+		a.OwnerClass, a.OwnerAdmin, a.OwnerInteractive = o.Class, o.Admin, o.Interactive
 		a.Evidence = append([]evidence{signatureEvidence(a)}, endpointEvidence(a.LLMEndpoints)...)
 		if a.DelegationToken != "" {
 			a.Evidence = append(a.Evidence, evidence{Signal: "delegation-tag", Tier: tierKnown, Value: "present"})
@@ -327,6 +341,133 @@ func scanAgents() ([]agentObs, error) {
 	return out, nil
 }
 
+// scanMCPProcessesFn is the seam a test uses to hand the correlation a
+// process list; production never reassigns it.
+var scanMCPProcessesFn = scanMCPProcesses
+
+// mcpServersReport reads the declared tool servers and correlates them with
+// the process list. It lives here, not in mcpconfig.go, because reading the
+// process list runs `ps` on macOS: the config reader itself starts nothing,
+// and the claims test holds it to that.
+func mcpServersReport() ([]mcpServerObs, mcpConfigScan) {
+	declared, scan, read := cachedMCPConfigs()
+	if !scan.Supported {
+		scan.ProcessScan = "unsupported"
+		return []mcpServerObs{}, scan
+	}
+	procs, err := scanMCPProcessesFn(mcpWantedBases(declared))
+	switch {
+	case err == nil:
+		scan.ProcessScan = "ok"
+	case agentScanSupported:
+		scan.ProcessScan = "failed"
+	default:
+		scan.ProcessScan = "unsupported"
+	}
+	if !read {
+		// No read of the files has finished yet (timed_out says so). With
+		// nothing known to be declared, every tool-server-looking process
+		// would read as running_not_declared, which nobody has checked.
+		return []mcpServerObs{}, scan
+	}
+	out := correlateMCP(declared, procs, err == nil)
+	if out == nil {
+		out = []mcpServerObs{}
+	}
+	if len(out) >= mcpMaxServers {
+		scan.Truncated = true
+	}
+	return out, scan
+}
+
+// mcpCache holds the last complete read of the config files. The heartbeat
+// is every 60 s and the read is up to a thousand files across every account;
+// like the scheduler inventory (schedCacheFor) it is redone every ten
+// minutes, while the process correlation above runs on every beat. One read
+// runs at a time, and the beats wait for it at most mcpReadTimeout IN TOTAL,
+// counted from when it started: a home on a network mount that stopped
+// answering holds that one goroutine, delays one beat by at most that much,
+// never the beats after it, and never starts a second read.
+var mcpCache struct {
+	sync.Mutex
+	at       time.Time
+	have     bool
+	declared []mcpServerObs
+	scan     mcpConfigScan
+	inflight chan struct{}
+	started  time.Time // when the read in flight began
+	gen      int
+}
+
+const mcpCacheFor = 10 * time.Minute
+
+// mcpReadTimeout is how long one beat waits for a read in progress.
+var mcpReadTimeout = 20 * time.Second
+
+// cachedMCPConfigs is the last complete read. read is false when no read has
+// finished yet, so the list says nothing about what is declared.
+func cachedMCPConfigs() (declared []mcpServerObs, scan mcpConfigScan, read bool) {
+	mcpCache.Lock()
+	if mcpCache.have && time.Since(mcpCache.at) < mcpCacheFor {
+		d, s := mcpCache.declared, mcpCache.scan
+		mcpCache.Unlock()
+		return d, s, true
+	}
+	done := mcpCache.inflight
+	if done == nil {
+		done = make(chan struct{})
+		mcpCache.inflight, mcpCache.started = done, time.Now()
+		gen := mcpCache.gen
+		go func() {
+			d, s := readMCPConfigs()
+			mcpCache.Lock()
+			if mcpCache.gen == gen {
+				mcpCache.declared, mcpCache.scan, mcpCache.at, mcpCache.have = d, s, time.Now(), true
+				mcpCache.inflight = nil
+			}
+			mcpCache.Unlock()
+			close(done)
+		}()
+	}
+	wait := mcpReadTimeout - time.Since(mcpCache.started)
+	mcpCache.Unlock()
+	timedOut := false
+	if wait <= 0 {
+		// An earlier beat already waited the whole time for this read.
+		select {
+		case <-done:
+		default:
+			timedOut = true
+		}
+	} else {
+		t := time.NewTimer(wait)
+		select {
+		case <-done:
+		case <-t.C:
+			timedOut = true
+		}
+		t.Stop()
+	}
+	mcpCache.Lock()
+	defer mcpCache.Unlock()
+	if !mcpCache.have {
+		// Nothing read yet: say so, rather than "no servers".
+		return nil, mcpConfigScan{Supported: true, TimedOut: true}, false
+	}
+	s := mcpCache.scan
+	s.TimedOut = timedOut
+	return mcpCache.declared, s, true
+}
+
+// resetMCPConfigCache forgets the last read; a read still running is
+// ignored when it finishes. Tests use it to start clean.
+func resetMCPConfigCache() {
+	mcpCache.Lock()
+	defer mcpCache.Unlock()
+	mcpCache.gen++
+	mcpCache.have, mcpCache.declared, mcpCache.scan, mcpCache.inflight = false, nil, mcpConfigScan{}, nil
+}
+
 var (
 	llmIPCache   map[string]string
 	llmIPCacheAt time.Time
@@ -371,6 +512,30 @@ func resolveLLMIPs() map[string]string {
 // to the anchor host, because approving is the anchor's job. "" = no pairing
 // identity (tests, tools): mode, anchor and unapproved attempts still travel.
 func buildHeartbeatBody(self, hostname, platform, version, wgPub, fingerprint, tunnelIP, endpoint string,
+	savePending bool, mirror *mirrorReport, enforcing map[string]any, peers []peerAnnounce) string {
+	return buildHeartbeatBodyOn(beatLanes{}, self, hostname, platform, version, wgPub, fingerprint, tunnelIP, endpoint,
+		savePending, mirror, enforcing, peers)
+}
+
+// beatLanes is which of a pairing's lanes run, so its beat advertises those
+// and no others. The zero value is a tunnel pairing: the forwards listen on
+// the tunnel's stack and the service slots are reachable there.
+type beatLanes struct {
+	// local: a Host with no platform (run.go runUnixPairing). The forwards and
+	// the service slots are absent; the model servers ride the serve lane
+	// (servelane.go), and the beat says which in lane_serves.
+	local bool
+	// noCoders: a tcp pairing, whose coder runs have no control plane to use
+	// (pair_tcp.go coderTransportRefusal), so it is never offered one.
+	noCoders bool
+}
+
+func beatLanesFor(cfg *Config) beatLanes {
+	return beatLanes{local: cfg.localTransport(), noCoders: cfg.tcpTransport()}
+}
+
+// buildHeartbeatBodyOn is buildHeartbeatBody for a pairing whose lanes are lanes.
+func buildHeartbeatBodyOn(lanes beatLanes, self, hostname, platform, version, wgPub, fingerprint, tunnelIP, endpoint string,
 	savePending bool, mirror *mirrorReport, enforcing map[string]any, peers []peerAnnounce) string {
 	caps := []string{}
 	// WHY SOME CAPABILITIES ARE WITHHELD HERE AND OTHERS ARE NOT — the asymmetry
@@ -408,7 +573,17 @@ func buildHeartbeatBody(self, hostname, platform, version, wgPub, fingerprint, t
 	// gets exactly what it got before: scope.filterModels is the identity then.
 	scope := pairingModelScope(self)
 	reachable, models := false, []string(nil)
-	if inference {
+	// A local pairing has no forward to advertise: its model servers ride the
+	// serve lane, and each lane below is advertised only when lane_serves
+	// carries its service (the target set and this machine's, the server
+	// answering, the family held). Its probes are of the address the lane
+	// dials, never the tunnel's guesses.
+	var served laneServeReport
+	if lanes.local {
+		served = laneServes(self)
+		reachable = containsStr(served.served, "ollama")
+		models = served.models
+	} else if inference {
 		reachable, models = ollamaModels(tunnelIP)
 		models = scope.filterModels(models)
 		// Reachable but nothing shared: advertising `embed` would put this box
@@ -431,7 +606,15 @@ func buildHeartbeatBody(self, hostname, platform, version, wgPub, fingerprint, t
 	// box running both.
 	llmModels := []string{}
 	llmPort := llmForwardPort
-	if !inference {
+	if lanes.local {
+		// No port listens for it here: the Host reaches the server through the
+		// serve lane, so llm_port is the "no port" value.
+		llmPort = 0
+		if containsStr(served.served, "llm") {
+			caps = append(caps, "llm-openai")
+			llmModels = served.llmModels
+		}
+	} else if !inference {
 		// The host stores 0 as NULL; with llm-openai withheld and
 		// llm_models empty nothing on the host resolves a port for this row.
 		llmPort = 0
@@ -457,7 +640,9 @@ func buildHeartbeatBody(self, hostname, platform, version, wgPub, fingerprint, t
 	// probe: an up forwarder with a dead upstream is a distinct state the
 	// host wants before committing a fail-closed write path to it.
 	piiAdvertised, piiOK := false, false
-	if inference && scope.allowsPII() {
+	if lanes.local {
+		piiAdvertised, piiOK = containsStr(served.served, "pii"), served.piiOK
+	} else if inference && scope.allowsPII() {
 		piiAdvertised, piiOK = piiProbe()
 	}
 	if piiAdvertised {
@@ -497,7 +682,7 @@ func buildHeartbeatBody(self, hostname, platform, version, wgPub, fingerprint, t
 	// "build a server on this machine" on it, and an offer the box would refuse
 	// is a control that fails after the user chose it. The slot list rides with
 	// it, withheld (absent, not empty) from a pairing outside the family.
-	servicesOn := servicesCapable(self)
+	servicesOn := !lanes.local && servicesCapable(self)
 	if servicesOn {
 		caps = append(caps, serviceCapability)
 	}
@@ -505,7 +690,10 @@ func buildHeartbeatBody(self, hostname, platform, version, wgPub, fingerprint, t
 	// only when this machine would accept a run-up from this pairing — Linux,
 	// root, coders on, a runtime, the family, and (anchored) a standing coder
 	// grant. Absent, never empty, otherwise.
-	coders := coderHeartbeat(self)
+	var coders map[string]any
+	if !lanes.noCoders {
+		coders = coderHeartbeat(self)
+	}
 	if coders != nil {
 		caps = append(caps, coderCapability)
 	}
@@ -554,11 +742,21 @@ func buildHeartbeatBody(self, hostname, platform, version, wgPub, fingerprint, t
 	// done for nothing, on the user's own hardware.
 	var schedEntries []schedEntry
 	var schedScan schedScanReport
+	// Declared tool servers (mcpconfig.go): the agent clients' own config
+	// files, correlated with the process list. Same gate and same
+	// quiet-vs-blind rule as the schedulers: read only for a pairing entitled
+	// to observe, and the capability travels only from a build that can.
+	var mcpServers []mcpServerObs
+	var mcpScan mcpConfigScan
 	if observing {
 		schedEntries, schedScan = scanSchedulers()
 		if schedScan.Supported {
 			caps = append(caps, "scheduler-observe")
 		}
+		mcpServers, mcpScan = mcpServersReport()
+		if mcpScan.Supported {
+			caps = append(caps, "tool-server-observe")
+		}
 	}
 	// Context preparation (contextprep.go): the tools on this box that
 	// turn a user's bytes into their content — a PDF's text, an image's OCR.
@@ -623,6 +821,13 @@ func buildHeartbeatBody(self, hostname, platform, version, wgPub, fingerprint, t
 		// absence as a client too old to say.
 		"endpoint_save_pending": savePending,
 	}
+	// What this pairing serves over the serve lane (servelane.go), on a local
+	// pairing only: ["ollama","llm","pii"] or fewer, [] when nothing is. Absent
+	// on a tunnel pairing, which has the forwards instead, and from a bastion
+	// that predates the lane.
+	if lanes.local {
+		body["lane_serves"] = served.served
+	}
 	// The service slots this pairing may reach, with the tunnel port for each
 	// (service.go). Absent when the capability is withheld, never empty-for-no.
 	if servicesOn {
@@ -674,6 +879,13 @@ func buildHeartbeatBody(self, hostname, platform, version, wgPub, fingerprint, t
 		// that read 0 crontabs is visibly different from one that read 40.
 		body["schedulers"] = schedEntries
 		body["scheduler_scan"] = schedScan
+		// The tool servers each agent client declares, and which are running
+		// (or running with no declaration). Paths, server names and command
+		// lines, so inside this gate with the agent list; env values and
+		// header values never leave mcpconfig.go, only their names and
+		// whether one looked like a secret.
+		body["tool_servers"] = mcpServers
+		body["tool_server_scan"] = mcpScan
 		// Where this machine keeps its own record — the findings journal and
 		// the signed accounting (machinestore.go, baselinesign.go). The host
 		// drafts a machine store at this root for the user to approve, once;
@@ -846,7 +1058,11 @@ func buildHeartbeatBody(self, hostname, platform, version, wgPub, fingerprint, t
 		}
 	}
 	if piiAdvertised {
+		// pii_port is the tunnel's port; a local pairing has none and reaches
+		// the sterilizer through the serve lane.
+		if !lanes.local {
 			body["pii_port"] = piiForwardPort
+		}
 		body["pii_ok"] = piiOK
 	}
 	if faAdvertised {

client-go/agents_darwin.go

@@ -101,6 +101,79 @@ func scanProcesses() (map[int]agentObs, error) {
 	return out, nil
 }
 
+// ownerFactsReader answers the three owner facts agentowner.go decides on.
+// macOS keeps accounts in Open Directory, not /etc/passwd, so it asks the
+// system's own tools, each a fixed argv under runTimed's timeout: `dscl .`
+// (the LOCAL node only) for the uid and whether the account is a mobile copy
+// of a directory account, `id -Gn` for its groups (nested groups resolved by
+// the system), and `who` once per scan for who is logged in. Answers are kept
+// per account for the scan, so ten agents of one person cost three commands.
+func ownerFactsReader() func(pid int, user string) ownerFacts {
+	var sessions map[string]bool
+	sessionsOK, sessionsRead := false, false
+	cache := map[string]ownerFacts{}
+	return func(pid int, user string) ownerFacts {
+		if user == "" || strings.ContainsAny(user, "/ \t\n") {
+			return unknownOwner
+		}
+		if f, ok := cache[user]; ok {
+			return f
+		}
+		if !sessionsRead {
+			sessionsRead = true
+			if out, err := runTimed("who"); err == nil {
+				sessions, sessionsOK = parseWhoUsers(string(out)), true
+			}
+		}
+		o := unixOwner{UID: -1, Name: user, LocalRecord: ownerUnknown,
+			SystemUIDMax: darwinSystemUIDMax, AdminGroups: darwinAdminGroups,
+			Sessions: sessions, SessionsKnown: sessionsOK}
+		out, err := runTimed("dscl", ".", "-read", "/Users/"+user, "UniqueID", "OriginalNodeName")
+		if err == nil {
+			if uid, mobile, ok := parseDsclUser(string(out)); ok {
+				o.UID = uid
+				o.LocalRecord = ownerYes
+				if mobile {
+					o.LocalRecord, o.DirectorySource = ownerNo, true
+				}
+			}
+		} else if dsclRecordMissing(err) {
+			// dscl ran and has no such record in the local node: the account
+			// comes from a directory. The uid then has to come from id.
+			o.LocalRecord, o.DirectorySource = ownerNo, true
+			if out, err := runTimed("id", "-u", "--", user); err == nil {
+				if n, err := strconv.Atoi(strings.TrimSpace(string(out))); err == nil {
+					o.UID = n
+				}
+			}
+		}
+		if out, err := runTimed("id", "-Gn", "--", user); err == nil {
+			o.Groups, o.GroupsKnown = parseIDGroups(string(out)), true
+		}
+		f := classifyUnixOwner(o)
+		cache[user] = f
+		return f
+	}
+}
+
+// scanMCPProcesses is the process list mcpconfig.go correlates declared tool
+// servers against: one `ps` sweep, every process with its parent and
+// arguments. An error is a sweep that did not happen.
+func scanMCPProcesses(want map[string]bool) ([]mcpProc, error) {
+	_, argvs, parents, err := psSweep()
+	if err != nil {
+		return nil, err
+	}
+	self := os.Getpid()
+	out := make([]mcpProc, 0, len(argvs))
+	for pid, argv := range argvs {
+		if pid != self {
+			out = append(out, mcpProc{PID: pid, PPID: parents[pid], Argv: argv})
+		}
+	}
+	return out, nil
+}
+
 // procParent answers from the sweep scanProcesses just took.
 func procParent(pid int) int {
 	psCache.Lock()

client-go/agents_linux.go

@@ -64,6 +64,35 @@ func scanProcesses() (map[int]agentObs, error) {
 	return out, nil
 }
 
+// scanMCPProcesses is the process list mcpconfig.go correlates declared tool
+// servers against: every process, with its parent and command line, from one
+// /proc walk. The want set only matters on Windows, where reading a command
+// line costs a handle per process. An error is a walk that did not happen,
+// and then no running state is guessed.
+func scanMCPProcesses(want map[string]bool) ([]mcpProc, error) {
+	entries, err := os.ReadDir(procRoot)
+	if err != nil {
+		return nil, fmt.Errorf("read %s: %w", procRoot, err)
+	}
+	self := os.Getpid()
+	var out []mcpProc
+	for _, e := range entries {
+		pid, err := strconv.Atoi(e.Name())
+		if err != nil || pid == self || !e.IsDir() {
+			continue
+		}
+		p := mcpProc{PID: pid}
+		if raw, err := os.ReadFile(procRoot + "/" + e.Name() + "/stat"); err == nil {
+			p.PPID = parseProcStatPPID(string(raw))
+		}
+		if raw, err := os.ReadFile(procRoot + "/" + e.Name() + "/cmdline"); err == nil {
+			p.Argv = splitNul(raw)
+		}
+		out = append(out, p)
+	}
+	return out, nil
+}
+
 // procParent reads one process's parent pid from /proc/<pid>/stat.
 func procParent(pid int) int {
 	raw, err := os.ReadFile("/proc/" + strconv.Itoa(pid) + "/stat")
@@ -126,6 +155,109 @@ func procUser(pid int) string {
 	return ""
 }
 
+// ownerFactsReader answers, per process, the three facts about its owner that
+// agentowner.go decides on. The machine-wide tables (/etc/passwd, /etc/group,
+// login.defs, nsswitch.conf, the sudo rules, who is logged in) are read once
+// per scan, on first use, and every process looks itself up in them.
+//
+// The uid and groups come from /proc/<pid>/status, so they are what the
+// PROCESS holds, the same source procUser names it from.
+func ownerFactsReader() func(pid int, user string) ownerFacts {
+	var (
+		loaded      bool
+		groupNames  map[int]string
+		groupsOK    bool
+		passwd      map[int]bool
+		passwdOK    bool
+		sysMax      int
+		directory   bool
+		sudoers     map[string]bool
+		sudoKnown   bool
+		sessions    map[string]bool
+		sessionsOK  bool
+		logindUsers = identityPath("/run/systemd/users")
+	)
+	load := func() {
+		loaded = true
+		if b, err := os.ReadFile(identityPath("/etc/group")); err == nil {
+			groupNames, groupsOK = parseGroupNames(string(b)), true
+		}
+		if b, err := os.ReadFile(identityPath("/etc/passwd")); err == nil {
+			passwd, passwdOK = parsePasswdUIDs(string(b)), true
+		}
+		sysMax = linuxSystemUIDMax
+		if b, err := os.ReadFile(identityPath("/etc/login.defs")); err == nil {
+			sysMax = parseLoginDefsSysMax(string(b))
+		}
+		if b, err := os.ReadFile(identityPath("/etc/nsswitch.conf")); err == nil {
+			directory = nsswitchHasDirectory(string(b))
+		}
+		var partial string
+		sudoers, partial = sudoPrincipals()
+		sudoKnown = partial == ""
+		// Who has a session: logind's own record when it runs, else utmp.
+		// Neither readable leaves the answer unknown rather than "nobody".
+		if _, err := os.Stat(logindUsers); err != nil {
+			if b, err := os.ReadFile(identityPath("/var/run/utmp")); err == nil {
+				sessions, sessionsOK = parseUtmpUsers(b), true
+			}
+		}
+	}
+	return func(pid int, user string) ownerFacts {
+		if !loaded {
+			load()
+		}
+		raw, err := os.ReadFile(procRoot + "/" + strconv.Itoa(pid) + "/status")
+		if err != nil {
+			return unknownOwner
+		}
+		realUID, effUID, gids, ok := parseStatusIDs(string(raw))
+		if !ok {
+			return unknownOwner
+		}
+		o := unixOwner{UID: realUID, Name: user, LocalRecord: ownerUnknown, DirectorySource: directory,
+			SystemUIDMax: sysMax, AdminGroups: linuxAdminGroups, SudoApplies: true,
+			SudoNamed: sudoers[user], SudoKnown: sudoKnown}
+		if passwdOK {
+			o.LocalRecord = yesNo(passwd[realUID])
+		}
+		if groupsOK {
+			o.GroupsKnown = true
+			for _, g := range gids {
+				if n, ok := groupNames[g]; ok {
+					o.Groups = append(o.Groups, n)
+				}
+			}
+		}
+		switch {
+		case sessionsOK:
+			o.Sessions, o.SessionsKnown = sessions, true
+		default:
+			b, err := os.ReadFile(logindUsers + "/" + strconv.Itoa(realUID))
+			switch {
+			case err == nil:
+				if live, ok := parseLogindUserState(string(b)); ok {
+					o.Sessions, o.SessionsKnown = map[string]bool{user: live}, true
+				}
+			case os.IsNotExist(err):
+				// logind keeps a file for every user it tracks; none means
+				// this account has no session and no lingering services.
+				if _, derr := os.Stat(logindUsers); derr == nil {
+					o.Sessions, o.SessionsKnown = map[string]bool{}, true
+				}
+			}
+		}
+		f := classifyUnixOwner(o)
+		if effUID == 0 {
+			// A process running with effective uid 0 holds root whatever
+			// account started it: that is the administrator answer that
+			// matters, while the class stays the account's own.
+			f.Admin = ownerYes
+		}
+		return f
+	}
+}
+
 func procCmdline(pid int) string {
 	raw, err := os.ReadFile("/proc/" + strconv.Itoa(pid) + "/cmdline")
 	if err != nil {

client-go/agents_other.go

@@ -34,3 +34,11 @@ func procEnvToken(pid int) string              { return "" }
 func procExe(pid int) string      { return "" }
 func procParent(pid int) int      { return 0 }
 func processParents() map[int]int { return map[int]int{} }
+
+// ownerFactsReader: no reader here, so every owner fact is unknown.
+func ownerFactsReader() func(pid int, user string) ownerFacts {
+	return func(int, string) ownerFacts { return unknownOwner }
+}
+
+// scanMCPProcesses: no enumerator, so running state is never measured here.
+func scanMCPProcesses(map[string]bool) ([]mcpProc, error) { return nil, errNoEnumerator }

client-go/agents_windows.go

@@ -306,6 +306,7 @@ func processParents() map[int]int {
 // scanProcesses: an error is a snapshot that did not happen (Toolhelp can
 // fail transiently), never an empty box; the heartbeat withholds `agents`.
 func scanProcesses() (map[int]agentObs, error) {
+	winBeginSIDScan()
 	out := map[int]agentObs{}
 	snap, err := windows.CreateToolhelp32Snapshot(windows.TH32CS_SNAPPROCESS, 0)
 	if err != nil {
@@ -390,6 +391,50 @@ func scanProcesses() (map[int]agentObs, error) {
 	return out, nil
 }
 
+// scanMCPProcesses is the process list mcpconfig.go correlates declared tool
+// servers against. Every process comes from one Toolhelp snapshot with its
+// parent; the command line, which costs a handle and a memory read, is read
+// only for a program in want (interpreters, package runners, and every
+// program a declared server names) or whose own name carries the word mcp.
+// The rest keep their parent link with no command line, which is all the
+// correlation needs from them. An error is a snapshot that did not happen.
+func scanMCPProcesses(want map[string]bool) ([]mcpProc, error) {
+	snap, err := windows.CreateToolhelp32Snapshot(windows.TH32CS_SNAPPROCESS, 0)
+	if err != nil {
+		return nil, fmt.Errorf("process snapshot: %w", err)
+	}
+	defer windows.CloseHandle(snap)
+	var e windows.ProcessEntry32
+	e.Size = uint32(unsafe.Sizeof(e))
+	self := os.Getpid()
+	var out []mcpProc
+	for err = windows.Process32First(snap, &e); err == nil; err = windows.Process32Next(snap, &e) {
+		pid := int(e.ProcessID)
+		if pid <= 0 || pid == self {
+			continue
+		}
+		p := mcpProc{PID: pid, PPID: int(e.ParentProcessID)}
+		exe := windows.UTF16ToString(e.ExeFile[:])
+		if base := binBase(exe); want[base] || mcpWord(base) {
+			cmdline := ""
+			if h, herr := winOpenForRead(pid); herr == nil {
+				cmdline = winCmdline(h)
+				windows.CloseHandle(h)
+			}
+			if argv, derr := windows.DecomposeCommandLine(cmdline); derr == nil && len(argv) > 0 {
+				p.Argv = argv
+			} else {
+				p.Argv = []string{exe}
+			}
+		}
+		out = append(out, p)
+	}
+	if !errors.Is(err, windows.ERROR_NO_MORE_FILES) {
+		return nil, fmt.Errorf("process walk: %w", err)
+	}
+	return out, nil
+}
+
 // ── connections: GetExtendedTcpTable ────────────────────────────────────────
 
 var (
@@ -463,14 +508,133 @@ func winTokenUser(h windows.Handle) string {
 	if err != nil || tu.User.Sid == nil {
 		return ""
 	}
-	account, domain, _, err := tu.User.Sid.LookupAccount("")
+	return winAccountLabel(tu.User.Sid.String(), winScanSIDNames())
+}
+
+// winSIDScan is the current scan's account-name cache (sidnames.go), shared
+// by winTokenUser and ownerFactsReader so a SID is looked up once per scan
+// and the scan's lookups stay within sidLookupBudget. scanProcesses starts a
+// new one; the readers that run after it in scanAgents use that one.
+var winSIDScan struct {
+	sync.Mutex
+	names *sidNames
+}
+
+func winBeginSIDScan() *sidNames {
+	n := newSIDNames(winLookupSID, sidLookupBudget, sidLookupSlots)
+	winSIDScan.Lock()
+	winSIDScan.names = n
+	winSIDScan.Unlock()
+	return n
+}
+
+func winScanSIDNames() *sidNames {
+	winSIDScan.Lock()
+	n := winSIDScan.names
+	winSIDScan.Unlock()
+	if n == nil {
+		return winBeginSIDScan()
+	}
+	return n
+}
+
+// winLookupSID is LookupAccountSid on this machine for one SID string.
+func winLookupSID(s string) (string, string, error) {
+	a, err := winLookupSIDAccount(s)
+	return a.Account, a.Domain, err
+}
+
+// winLookupSIDAccount is LookupAccountSid with the SID's type. Unbounded:
+// callers go through sidNames or winAccounts.
+func winLookupSIDAccount(s string) (sidAccount, error) {
+	sid, err := windows.StringToSid(s)
 	if err != nil {
-		return tu.User.Sid.String()
+		return sidAccount{}, err
 	}
-	if domain != "" {
-		return domain + `\` + account
+	account, domain, typ, err := sid.LookupAccount("")
+	return sidAccount{Account: account, Domain: domain, Type: typ}, err
+}
+
+// winAccounts is the process-wide account cache for lookups made outside a
+// scan: the approvals pipe and the folder agent's drop (sidnames.go).
+var winAccounts = newSIDDirectory(winLookupSIDAccount)
+
+// ownerFactsReader answers the three owner facts agentowner.go decides on,
+// from each process's own token: its user SID and that SID's domain, and its
+// group SIDs (deny-only ones included, which is how an administrator running
+// without elevation still reads as an administrator account). The session
+// list is read once per scan and needs LocalSystem, which the system service
+// is; elsewhere it is left unknown and the token's own logon type answers
+// what it can (classifyWinOwner). The domain comes from the scan's shared
+// account-name cache (winScanSIDNames), so a SID winTokenUser already looked
+// up is not asked again, and an unreachable domain controller leaves the
+// class unknown instead of holding the scan.
+func ownerFactsReader() func(pid int, user string) ownerFacts {
+	computer, _ := windows.ComputerName()
+	names := winScanSIDNames()
+	var sessions map[string]bool
+	sessionsOK, sessionsRead := false, false
+	return func(pid int, _ string) ownerFacts {
+		if !sessionsRead {
+			sessionsRead = true
+			sessions, sessionsOK = winSessionSIDs()
 		}
-	return account
+		h, err := windows.OpenProcess(windows.PROCESS_QUERY_LIMITED_INFORMATION, false, uint32(pid))
+		if err != nil {
+			return unknownOwner
+		}
+		defer windows.CloseHandle(h)
+		var tok windows.Token
+		if err := windows.OpenProcessToken(h, windows.TOKEN_QUERY, &tok); err != nil {
+			return unknownOwner
+		}
+		defer tok.Close()
+		o := winOwner{Computer: computer, SessionSIDs: sessions, SessionsKnown: sessionsOK}
+		if tu, err := tok.GetTokenUser(); err == nil && tu.User.Sid != nil {
+			o.SID = tu.User.Sid.String()
+			o.Domain = winAccountDomain(o.SID, names)
+		}
+		if tg, err := tok.GetTokenGroups(); err == nil {
+			o.Groups, o.GroupsKnown = map[string]bool{}, true
+			for _, g := range tg.AllGroups() {
+				if g.Sid != nil {
+					o.Groups[g.Sid.String()] = true
+				}
+			}
+		}
+		return classifyWinOwner(o)
+	}
+}
+
+// winSessionSIDs is the set of user SIDs with an active or disconnected
+// Remote Desktop Services session (the console is one). ok is false when the
+// list could not be read: WTSQueryUserToken needs LocalSystem's privilege.
+func winSessionSIDs() (map[string]bool, bool) {
+	var info *windows.WTS_SESSION_INFO
+	var count uint32
+	if err := windows.WTSEnumerateSessions(0, 0, 1, &info, &count); err != nil {
+		return nil, false
+	}
+	defer windows.WTSFreeMemory(uintptr(unsafe.Pointer(info)))
+	list := unsafe.Slice(info, count)
+	out := map[string]bool{}
+	for _, s := range list {
+		if s.State != windows.WTSActive && s.State != windows.WTSDisconnected {
+			continue
+		}
+		var tok windows.Token
+		if err := windows.WTSQueryUserToken(s.SessionID, &tok); err != nil {
+			if errors.Is(err, windows.ERROR_PRIVILEGE_NOT_HELD) || errors.Is(err, windows.ERROR_ACCESS_DENIED) {
+				return nil, false
+			}
+			continue // a session with no user (the services session, a login screen)
+		}
+		if tu, err := tok.GetTokenUser(); err == nil && tu.User.Sid != nil {
+			out[tu.User.Sid.String()] = true
+		}
+		tok.Close()
+	}
+	return out, true
 }
 
 func procUser(pid int) string {

client-go/anchor.go

@@ -574,9 +574,10 @@ var allCaps = []string{"dcn", "egress", "harness", "inference", "models", "obser
 // writes a pairing PENDING when an anchor exists (pair.go), and state "" reads
 // active (config.go), so B comes straight up with full lanes.
 //
-// LOCAL is exempt and must stay so: a local bastion serves one host and cannot
-// be anchored at all (cmdClaimAnchor refuses outside networked mode), so the
-// pairing IS the authorization there. Most bastions are local.
+// LOCAL is exempt and must stay so: a local bastion serves the one host it was
+// paired with, and a second only on a yes typed at the machine (pair_terminal.go),
+// and it cannot be anchored at all (cmdClaimAnchor refuses outside networked
+// mode), so the pairing IS the authorization there. Most bastions are local.
 //
 // self-update is deliberately INSIDE the refusal. Directing a binary swap is the
 // most dangerous thing on the list, and an unanchored networked box cannot say
@@ -810,6 +811,11 @@ type pairingEntry struct {
 // --replace, so a pending host re-presenting a token keeps its original age
 // (the 10-minute unapproved clock does not restart). nil when the directory
 // cannot be read — the host keeps what it has, as with peers.
+//
+// A pairing whose file this build refuses (Config.invalid) is left out: it is
+// never served, so to a host it is not on this box. A host knows the states
+// active, pending and revoked and may read any other as active, so "invalid"
+// is never sent. The machine's own status and setup page still show it.
 func pairingsReport() []pairingEntry {
 	cfgs, err := loadPairings()
 	if err != nil {
@@ -817,7 +823,7 @@ func pairingsReport() []pairingEntry {
 	}
 	out := make([]pairingEntry, 0, len(cfgs))
 	for _, c := range cfgs {
-		if c.HostName == "" {
+		if c.HostName == "" || c.invalid != "" {
 			continue
 		}
 		out = append(out, pairingEntry{HostName: c.HostName, Handle: handleOf(c.HostName),
@@ -849,7 +855,7 @@ func unapprovedAttempts(now time.Time) []peerAnnounce {
 		if err != nil || now.Sub(since) < unapprovedAfter {
 			continue
 		}
-		out = append(out, peerAnnounce{HostName: e.HostName, Handle: e.Handle, State: e.State, Since: e.Since})
+		out = append(out, peerAnnounce(e))
 	}
 	return out
 }
@@ -978,6 +984,20 @@ func cmdRepinAnchor() error {
 	return nil
 }
 
+// anchorAlreadyHolds: the anchor names an active pairing here, and host
+// (claim-anchor --host, "" for none) names that same pairing.
+func anchorAlreadyHolds(cfgs []*Config, prev *anchorClaim, host string) bool {
+	c := pairingFor(cfgs, prev.AnchorHost)
+	if c == nil || !c.active() {
+		return false
+	}
+	if host == "" {
+		return true
+	}
+	named, err := pickPairing(cfgs, host)
+	return err == nil && named.HostName == c.HostName
+}
+
 // pickPairing resolves --host against this box's pairings. It accepts the full
 // host_name the box reports, or just the handle in front of it, because the
 // name the user is looking at depends on which surface they read it from —
@@ -1014,13 +1034,24 @@ func cmdClaimAnchor(token string, rotate bool, owner, host string) error {
 		return err
 	}
 	if boxMode() != modeNetworked {
-		return fmt.Errorf("claim-anchor needs networked mode — a local bastion serves one host and needs no anchor (%s, then `privatae-bastion set-mode networked`)", adminHowTo())
+		return fmt.Errorf("claim-anchor needs networked mode — a local bastion serves only the Hosts its person said yes to and needs no anchor (%s, then `privatae-bastion set-mode networked`)", adminHowTo())
 	}
 	cfgs, err := loadPairings()
 	if err != nil {
 		return err
 	}
 	prev := currentAnchor()
+	// Already anchored to the host this asks for (no token, no other host
+	// named), and that host's pairing is active: done, not an error. `pair`
+	// anchors a networked box to its first host by itself, and an installer
+	// that runs claim-anchor after pairing must find that a success. An anchor
+	// whose pairing is pending or revoked is not done, and is refused below as
+	// before. --host takes the short handle as well as the full name
+	// (pickPairing), so the same host named either way is the same request.
+	if prev != nil && !rotate && token == "" && anchorAlreadyHolds(cfgs, prev, host) {
+		fmt.Printf("already anchored to %s (owner %q)\n", prev.AnchorHost, prev.Owner)
+		return nil
+	}
 	if prev != nil && !rotate {
 		return fmt.Errorf("this box is already anchored to %s — `claim-anchor --rotate <token>` moves the anchor (needs the current anchor's consent, or %d days of silence from it)", prev.AnchorHost, int(anchorDeadAfter.Hours()/24))
 	}
@@ -1088,9 +1119,16 @@ func cmdClaimAnchor(token string, rotate bool, owner, host string) error {
 		case len(cfgs) == 0:
 			return fmt.Errorf("this box has no pairing yet — pair it to your host first (`privatae-bastion pair <token>`, token from Network → Set up a bastion), then `privatae-bastion claim-anchor` with no token")
 		default:
-			return fmt.Errorf("this box serves %d hosts — say which one should hold the anchor:\n%s\n  e.g. `privatae-bastion claim-anchor --host %s`\nThe anchor host approves every pairing made here from then on; the other pairings keep the lanes they already have.",
+			return fmt.Errorf("this box serves %d hosts — say which one should hold the anchor:\n%s\n  e.g. `privatae-bastion claim-anchor --host %s`\nThe anchor host approves every pairing made here from then on; the other pairings keep the lanes they already have",
 				len(cfgs), "    "+strings.Join(hostNames(cfgs), "\n    "), hostNames(cfgs)[0])
 		}
+		// A pairing whose file this build refuses is never served, so the box
+		// could not reach the host it would be anchored to, and that host would
+		// approve every pairing made here. Refused with the reason, before
+		// anything is written.
+		if cfg.invalid != "" {
+			return fmt.Errorf("%s cannot hold the anchor: %s", cfg.HostName, cfg.invalid)
+		}
 		// REVOKED is not "waiting", it is a decision, and it must not be undone by
 		// a command the user ran to do something else. pair.go holds the same line
 		// for a re-pair (pending again, never straight back to active),

client-go/anchor_test.go

@@ -606,7 +606,7 @@ func TestClaimAnchorWithNoToken(t *testing.T) {
 		if err := cmdClaimAnchor("", false, "", ""); err == nil || !contains(err.Error(), "machine authority") {
 			t.Fatalf("claim-anchor must need machine authority whether or not a token is given: %v", err)
 		}
-		return
+		t.Skip("the rest needs root: without machine authority claim-anchor stops at the refusal checked above")
 	}
 
 	t.Run("local mode refuses — the token is not what gates this", func(t *testing.T) {
@@ -658,14 +658,58 @@ func TestClaimAnchorWithNoToken(t *testing.T) {
 		}
 	})
 
-	t.Run("already anchored — still refused without --rotate", func(t *testing.T) {
+	t.Run("already anchored to that host — success, and nothing changes", func(t *testing.T) {
+		// `pair` anchors a networked box by itself, and an installer that runs
+		// claim-anchor afterwards must find that done, not an error.
 		freshBox(t, true)
 		writePairing(t, tMember)
 		if err := writeAnchor("Acme", tMember); err != nil {
 			t.Fatal(err)
 		}
-		if err := cmdClaimAnchor("", false, "", ""); err == nil || !contains(err.Error(), "already anchored") {
-			t.Fatalf("dropping the token must not let a second claim through: %v", err)
+		if err := cmdClaimAnchor("", false, "", ""); err != nil {
+			t.Fatalf("claim-anchor on a box already anchored to its host must succeed: %v", err)
+		}
+		if err := cmdClaimAnchor("", false, "", tMember); err != nil {
+			t.Fatalf("naming the anchor host itself must succeed: %v", err)
+		}
+		if err := cmdClaimAnchor("", false, "", handleOf(tMember)); err != nil {
+			t.Fatalf("naming the anchor host by its handle (as --host accepts) must succeed: %v", err)
+		}
+		if a := currentAnchor(); a == nil || a.AnchorHost != tMember || a.Owner != "Acme" {
+			t.Fatalf("the anchor must be untouched: %+v", a)
+		}
+	})
+
+	t.Run("anchored to a host whose pairing is pending or revoked — not reported done", func(t *testing.T) {
+		// A success here would tell an installer the anchor holds while its
+		// host has no active lanes on this box.
+		for _, st := range []string{pairingPending, pairingRevoked} {
+			freshBox(t, true)
+			writePairingState(t, tMember, st)
+			if err := writeAnchor("Acme", tMember); err != nil {
+				t.Fatal(err)
+			}
+			if err := cmdClaimAnchor("", false, "", ""); err == nil {
+				t.Fatalf("anchor pairing %s: claim-anchor must not report the anchor done", st)
+			}
+			if got := loadState(t, tMember); got != st {
+				t.Fatalf("anchor pairing %s: state changed to %q", st, got)
+			}
+		}
+	})
+
+	t.Run("already anchored — another host is still refused without --rotate", func(t *testing.T) {
+		freshBox(t, true)
+		writePairing(t, tMember)
+		writePairing(t, tAnchor)
+		if err := writeAnchor("Acme", tMember); err != nil {
+			t.Fatal(err)
+		}
+		if err := cmdClaimAnchor("", false, "", tAnchor); err == nil || !contains(err.Error(), "already anchored") {
+			t.Fatalf("naming a different host must not move the anchor without --rotate: %v", err)
+		}
+		if a := currentAnchor(); a == nil || a.AnchorHost != tMember {
+			t.Fatalf("the anchor must be untouched: %+v", a)
 		}
 	})
 

client-go/approvalpipe.go

@@ -138,6 +138,30 @@ type pipeClient struct {
 	// "" when nobody is.
 	SessionUserSID string
 	SessionActive  bool // WTSActive: the session someone is in front of
+	// NameTimedOut: the account directory did not answer for UserSID within
+	// pipeSIDLane's budget, so UserName is empty and nobody can say whose
+	// account approved. pipeApprove refuses; list and deny go on, by SID.
+	NameTimedOut bool
+}
+
+// pipeSIDLane bounds the client's account-name lookup (sidnames.go). The tray
+// waits 2*pipeIOTimeout for its answer, and the name is looked up before the
+// request is read, so the lookup gets well under that: a directory that does
+// not answer costs one connection 2 s, and holds at most one goroutine for
+// the pipe however many connections come.
+var pipeSIDLane = sidLane{slots: make(chan struct{}, 1), budget: 2 * time.Second}
+
+// pipeClientName fills c.UserName from the account directory, or sets
+// c.NameTimedOut when the directory did not answer in time. A SID with no
+// account leaves UserName empty: the log shows the SID, as before.
+func pipeClientName(c *pipeClient, dir *sidDirectory, lane sidLane) {
+	acct, outcome := dir.resolve(c.UserSID, lane)
+	switch outcome {
+	case sidResolved:
+		c.UserName = sidAccountLabel(acct)
+	case sidTimedOut:
+		c.NameTimedOut = true
+	}
 }
 
 // purposeDigest is the sha256 of a purpose as the tray carries it back.
@@ -330,6 +354,14 @@ func pipeTarget(req pipeRequest) (dir string, r stepupRequest, err error) {
 var pipeApproveMu sync.Mutex
 
 func pipeApprove(req pipeRequest, c pipeClient, now time.Time) pipeResponse {
+	// FAILS CLOSED ON A SLOW DIRECTORY. The rules above decide by SID, but an
+	// approval's audit line names the account that gave it, and when the
+	// directory could not say whose account this is, nothing is approved.
+	// Before the throttle and the code, so this counts as no wrong code.
+	if c.NameTimedOut {
+		return pipeResponse{Error: "not approved: this machine's account directory did not answer for " + c.UserSID +
+			" in time; try again in a minute, or approve from an Administrator terminal"}
+	}
 	pipeApproveMu.Lock()
 	defer pipeApproveMu.Unlock()
 	code := strings.TrimSpace(req.Code)

client-go/approvalpipe_windows.go

@@ -311,12 +311,9 @@ func pipeClientOf(h windows.Handle) (pipeClient, error) {
 		return c, fmt.Errorf("cannot read the connecting process's account (pid %d): %w", c.PID, err)
 	}
 	c.UserSID = u.User.Sid.String()
-	if name, dom, _, err := u.User.Sid.LookupAccount(""); err == nil {
-		c.UserName = name
-		if dom != "" {
-			c.UserName = dom + `\` + name
-		}
-	}
+	// Bounded and cached (sidnames.go): a directory that does not answer
+	// leaves NameTimedOut set, and pipeApprove refuses.
+	pipeClientName(&c, winAccounts, pipeSIDLane)
 	var sess, n uint32
 	if err := windows.GetTokenInformation(tok, windows.TokenSessionId, (*byte)(unsafe.Pointer(&sess)), uint32(unsafe.Sizeof(sess)), &n); err != nil {
 		return c, fmt.Errorf("cannot read the connecting process's session (pid %d): %w", c.PID, err)

client-go/auditlog.go

@@ -0,0 +1,179 @@
+package main
+
+// Audit rows to the system log, as well as to their files.
+//
+// Every audit file on this box (mgmt_audit.jsonl, pairing_audit.jsonl, and a
+// pairing's audit.jsonl for terminals, root commands, store ops and op runs)
+// lives here, on this machine. The host can fetch them over the tunnel but
+// never edit them (mgmt_audit.go handleAuditFetch). That still leaves the host
+// as the only central view of a fleet, and the host is the agent the audit is
+// about. So each row is also written to the machine's own system log, where
+// whatever already collects that log (journald or syslog forwarding, a log
+// shipper, a SIEM) takes it to a store the host cannot write:
+//
+//	Linux, macOS  syslog, facility auth, tag privatae-bastion-audit
+//	              (journald on a systemd machine: SYSLOG_IDENTIFIER=privatae-bastion-audit)
+//	Windows       the Application event log, source PrivataeBastionAudit, event id 1
+//
+// The system-log line is the file's row with two fields added in front of it:
+// `audit`, which file the row went to, and `pairing`, the pairing it belongs to
+// ("" for a box-level file). The file stays the record: a system log that is
+// missing or refuses a write costs that copy and nothing else, and is never a
+// reason to refuse or slow the op being audited.
+//
+// A root command's text is the one thing the file has that the system-log line
+// does not: it is whatever the host typed, a password on a command line
+// included, and the system log is readable by more accounts than the 0600 file
+// and forwarded off the machine by design. The line keeps command_sha256 and
+// command_bytes, so a collector can still match it to the file and the
+// transcript and see if either was changed.
+//
+// The write itself never runs on the audited op's goroutine. The syslog socket
+// is a blocking write: a journald or syslog daemon that is stuck or slow would
+// otherwise hold every audited op, and every other audit writer behind the
+// lock, for as long as it stays stuck. Lines go into a bounded queue that one
+// goroutine drains; a full queue drops the line and counts it, and the next
+// line queued after the gap carries the count (auditDroppedField).
+
+import (
+	"bytes"
+	"encoding/json"
+	"path/filepath"
+	"sync"
+	"sync/atomic"
+	"time"
+)
+
+const (
+	auditSyslogTag    = "privatae-bastion-audit"
+	auditEventSource  = "PrivataeBastionAudit"
+	auditEventID      = 1
+	auditFilePairing  = "pairing_audit.jsonl"
+	auditFilePairingD = "audit.jsonl"
+)
+
+// auditSink takes one line for the system log. A variable so a test can watch
+// the lines; the default queues it for the platform's writer and never blocks.
+var auditSink = auditEnqueue
+
+const (
+	// auditQueueMax lines wait for the system log; beyond that they are
+	// dropped and counted. A burst of audited ops far larger than any lane
+	// produces fits; a system log stuck for good holds no more than this.
+	auditQueueMax = 4096
+	// auditDroppedField is added to the first line queued after a drop: how
+	// many lines the system log lost just before it, so the gap shows where.
+	auditDroppedField = "system_log_dropped"
+	// auditFlushOnExit bounds how long a finished command waits for its rows
+	// to reach a slow system log.
+	auditFlushOnExit = 2 * time.Second
+)
+
+// auditQueue is the bounded queue in front of a system-log writer, drained by
+// one goroutine so lines keep their order.
+type auditQueue struct {
+	ch      chan []byte
+	write   func([]byte)
+	once    sync.Once
+	dropped atomic.Int64
+	pending atomic.Int64 // queued or being written
+}
+
+func newAuditQueue(n int, write func([]byte)) *auditQueue {
+	return &auditQueue{ch: make(chan []byte, n), write: write}
+}
+
+// The box's one queue, in front of systemAuditWrite (auditlog_*.go).
+var auditQ = newAuditQueue(auditQueueMax, func(l []byte) { systemAuditWrite(l) })
+
+func auditEnqueue(line []byte) { auditQ.enqueue(line) }
+
+// auditFlush waits, at most d, for queued lines to reach the system log. A
+// command that audits and then exits (pair, claim-anchor, set-mode) calls it
+// on the way out (main.go), so its rows are not lost with the process; a
+// stuck system log costs that wait and nothing more.
+func auditFlush(d time.Duration) { auditQ.flush(d) }
+
+// enqueue hands one line to the drain goroutine, or drops it and counts it
+// when the queue is full. It never waits on the writer.
+func (q *auditQueue) enqueue(line []byte) {
+	q.once.Do(func() { go q.drain() })
+	n := q.dropped.Swap(0)
+	if n > 0 {
+		line = auditWithDropped(line, n)
+	}
+	q.pending.Add(1)
+	select {
+	case q.ch <- line:
+	default:
+		q.pending.Add(-1)
+		q.dropped.Add(n + 1)
+	}
+}
+
+func (q *auditQueue) drain() {
+	for line := range q.ch {
+		q.write(line)
+		q.pending.Add(-1)
+	}
+}
+
+func (q *auditQueue) flush(d time.Duration) {
+	deadline := time.Now().Add(d)
+	for q.pending.Load() > 0 && time.Now().Before(deadline) {
+		time.Sleep(5 * time.Millisecond)
+	}
+}
+
+// auditWithDropped adds the drop count to a line (every line here is the
+// JSON object auditToSystemLog made).
+func auditWithDropped(line []byte, n int64) []byte {
+	fields := map[string]any{}
+	dec := json.NewDecoder(bytes.NewReader(line))
+	dec.UseNumber()
+	if err := dec.Decode(&fields); err != nil {
+		return line
+	}
+	fields[auditDroppedField] = n
+	out, err := json.Marshal(fields)
+	if err != nil {
+		return line
+	}
+	return out
+}
+
+// auditToSystemLog sends one audit row, as it was written to its file, to the
+// system log, with the file's name and the pairing in front.
+func auditToSystemLog(file, pairing string, rec []byte) {
+	fields := map[string]any{}
+	dec := json.NewDecoder(bytes.NewReader(rec))
+	dec.UseNumber() // numbers go out exactly as the file has them
+	if err := dec.Decode(&fields); err != nil {
+		return
+	}
+	if cmd, ok := fields["command"].(string); ok {
+		// The text stays in the file; see the top of this file.
+		delete(fields, "command")
+		if _, ok := fields["command_bytes"]; !ok {
+			fields["command_bytes"] = len(cmd)
+		}
+	}
+	fields["audit"] = file
+	fields["pairing"] = pairing
+	line, err := json.Marshal(fields)
+	if err != nil {
+		return
+	}
+	auditSink(line)
+}
+
+// auditPairingOfDir names the pairing of a per-pairing audit file from its
+// directory: the last path element, which is the pairing's host name
+// (Config.dcnDir: dcn/<host_name>), except for a legacy pairing that kept the
+// old shared directory, where it is that directory's name.
+func auditPairingOfDir(dir string) string {
+	if dir == "" {
+		return ""
+	}
+	return filepath.Base(filepath.Clean(dir))
+}

client-go/auditlog_notwindows.go

@@ -0,0 +1,34 @@
+//go:build !windows
+
+package main
+
+import (
+	"log/syslog"
+	"sync"
+)
+
+var (
+	auditSyslogMu sync.Mutex
+	auditSyslog   *syslog.Writer
+)
+
+// systemAuditWrite sends one line to syslog (journald on a systemd machine),
+// facility auth, at info. Called only from auditQ's drain (auditlog.go), never on
+// an audited op's own goroutine: the write blocks while the daemon is stuck. The connection is opened on first use and dropped
+// after a failed write, so the next row tries again; a machine with no syslog
+// socket costs the copy, never the row in the file.
+func systemAuditWrite(line []byte) {
+	auditSyslogMu.Lock()
+	defer auditSyslogMu.Unlock()
+	if auditSyslog == nil {
+		w, err := syslog.New(syslog.LOG_AUTH|syslog.LOG_INFO, auditSyslogTag)
+		if err != nil {
+			return
+		}
+		auditSyslog = w
+	}
+	if err := auditSyslog.Info(string(line)); err != nil {
+		_ = auditSyslog.Close()
+		auditSyslog = nil
+	}
+}

client-go/auditlog_queue_test.go

@@ -0,0 +1,95 @@
+package main
+
+import (
+	"encoding/json"
+	"reflect"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+)
+
+// The sink the binary ships with, read before auditlog_test.go's init swaps
+// it out (package variables are set before any init runs).
+var auditSinkShipped = auditSink
+
+// TestAuditSinkIsTheQueue: the shipped sink is the queue, never the
+// platform's writer called on the audited op's own goroutine.
+func TestAuditSinkIsTheQueue(t *testing.T) {
+	if reflect.ValueOf(auditSinkShipped).Pointer() != reflect.ValueOf(auditEnqueue).Pointer() {
+		t.Fatal("the system-log sink must be auditEnqueue: a direct write blocks the audited op while syslog is stuck")
+	}
+}
+
+// TestAuditQueueNeverWaitsOnAStuckWriter: with a writer that never returns,
+// far more rows than the queue holds are all taken at once; the ones past the
+// bound are dropped and counted, and the count rides on the next line that
+// gets into the queue.
+func TestAuditQueueNeverWaitsOnAStuckWriter(t *testing.T) {
+	release := make(chan struct{})
+	var mu sync.Mutex
+	var got []string
+	q := newAuditQueue(8, func(l []byte) {
+		<-release
+		mu.Lock()
+		got = append(got, string(l))
+		mu.Unlock()
+	})
+	done := make(chan struct{})
+	go func() {
+		for i := 0; i < 1000; i++ {
+			q.enqueue([]byte(`{"n":1}`))
+		}
+		close(done)
+	}()
+	select {
+	case <-done:
+	case <-time.After(5 * time.Second):
+		t.Fatal("audit rows waited on a stuck system log")
+	}
+	close(release)
+	q.flush(5 * time.Second)
+	// Nothing lost without a count: every one of the 1001 rows is either a
+	// line written or counted on a line written after the gap.
+	q.enqueue([]byte(`{"n":2}`))
+	q.flush(5 * time.Second)
+	mu.Lock()
+	defer mu.Unlock()
+	var counted float64
+	for _, l := range got {
+		m := map[string]any{}
+		if err := json.Unmarshal([]byte(l), &m); err != nil {
+			t.Fatal(err)
+		}
+		if n, ok := m[auditDroppedField].(float64); ok {
+			counted += n
+		}
+	}
+	if len(got) > 10 || counted == 0 || float64(len(got))+counted != 1001 {
+		t.Fatalf("lines written %d (at most 10), dropped and counted %v: want 1001 in all", len(got), counted)
+	}
+}
+
+// TestAuditSystemLogOmitsTheRootCommandText: a root command's text stays in
+// the 0600 file; the system-log line carries its digest and length.
+func TestAuditSystemLogOmitsTheRootCommandText(t *testing.T) {
+	var line string
+	prev := auditSink
+	auditSink = func(l []byte) { line = string(l) }
+	t.Cleanup(func() { auditSink = prev })
+	auditToSystemLog(auditFilePairingD, "p", []byte(`{"event":"start","command":"PGPASSWORD=hunter2 psql","command_sha256":"ab"}`))
+	m := map[string]any{}
+	if err := json.Unmarshal([]byte(line), &m); err != nil {
+		t.Fatal(err)
+	}
+	if strings.Contains(line, "hunter2") || m["command"] != nil {
+		t.Fatalf("the command text reached the system log: %s", line)
+	}
+	if m["command_sha256"] != "ab" || m["command_bytes"] != float64(len("PGPASSWORD=hunter2 psql")) {
+		t.Fatalf("the line must keep the digest and the length: %s", line)
+	}
+	auditToSystemLog(auditFilePairingD, "p", []byte(`{"command":"x","command_bytes":70000}`))
+	if !strings.Contains(line, `"command_bytes":70000`) {
+		t.Fatalf("a row's own command_bytes (the full length) is kept: %s", line)
+	}
+}

client-go/auditlog_syslog_test.go

@@ -0,0 +1,69 @@
+//go:build !windows
+
+package main
+
+import (
+	"log/syslog"
+	"net"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+)
+
+// TestAuditRowsDoNotWaitOnAStuckSyslogSocket: a real syslog writer on a
+// datagram socket nobody reads blocks once the socket's queue is full, as
+// /dev/log does while journald is stuck. Audit rows behind the queue all
+// return at once; the same writer called inline (as 0.54.0 first did) does not.
+func TestAuditRowsDoNotWaitOnAStuckSyslogSocket(t *testing.T) {
+	dir, err := os.MkdirTemp("", "al")
+	if err != nil {
+		t.Fatal(err)
+	}
+	t.Cleanup(func() { os.RemoveAll(dir) })
+	sock := filepath.Join(dir, "s")
+	l, err := net.ListenUnixgram("unixgram", &net.UnixAddr{Name: sock, Net: "unixgram"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	// Closing the reading end is what lets the stuck writes below return
+	// (refused) so their goroutines end; the writer itself is never closed,
+	// since its Close would wait behind a write that is stuck.
+	t.Cleanup(func() { l.Close() })
+	w, err := syslog.Dial("unixgram", sock, syslog.LOG_AUTH|syslog.LOG_INFO, auditSyslogTag)
+	if err != nil {
+		t.Fatal(err)
+	}
+	write := func(line []byte) { _ = w.Info(string(line)) }
+
+	// The inline writer does block: without this the test proves nothing.
+	inline := make(chan struct{})
+	go func() {
+		for i := 0; i < 5000; i++ {
+			write([]byte(`{"op":"x"}`))
+		}
+		close(inline)
+	}()
+	select {
+	case <-inline:
+		t.Skip("this kernel never blocked a datagram writer; nothing to bound")
+	case <-time.After(2 * time.Second):
+	}
+
+	q := newAuditQueue(64, write)
+	done := make(chan struct{})
+	go func() {
+		for i := 0; i < 5000; i++ {
+			q.enqueue([]byte(`{"op":"y"}`))
+		}
+		close(done)
+	}()
+	select {
+	case <-done:
+	case <-time.After(5 * time.Second):
+		t.Fatal("audit rows waited on a stuck syslog socket")
+	}
+	if q.dropped.Load() == 0 {
+		t.Fatal("rows past the bound must be counted as dropped")
+	}
+}

client-go/auditlog_test.go

@@ -0,0 +1,154 @@
+package main
+
+import (
+	"bufio"
+	"encoding/json"
+	"os"
+	"path/filepath"
+	"strings"
+	"sync"
+	"testing"
+)
+
+// The test binary never writes to the real system log: the suite's audit
+// rows would otherwise land in the journal of whatever machine runs it.
+func init() { auditSink = func([]byte) {} }
+
+// auditSinkCapture swaps the system-log sink for one that keeps the lines.
+func auditSinkCapture(t *testing.T) func() []map[string]any {
+	t.Helper()
+	var mu sync.Mutex
+	var lines [][]byte
+	prev := auditSink
+	auditSink = func(line []byte) {
+		mu.Lock()
+		defer mu.Unlock()
+		lines = append(lines, append([]byte(nil), line...))
+	}
+	t.Cleanup(func() { auditSink = prev })
+	return func() []map[string]any {
+		mu.Lock()
+		defer mu.Unlock()
+		var out []map[string]any
+		for _, l := range lines {
+			m := map[string]any{}
+			if err := json.Unmarshal(l, &m); err != nil {
+				t.Fatalf("system-log line is not JSON: %s", l)
+			}
+			out = append(out, m)
+		}
+		return out
+	}
+}
+
+func auditFileRows(t *testing.T, path string) []map[string]any {
+	t.Helper()
+	fh, err := os.Open(path)
+	if err != nil {
+		t.Fatalf("audit file %s: %v", path, err)
+	}
+	defer fh.Close()
+	var out []map[string]any
+	sc := bufio.NewScanner(fh)
+	for sc.Scan() {
+		m := map[string]any{}
+		if err := json.Unmarshal(sc.Bytes(), &m); err != nil {
+			t.Fatalf("audit row is not JSON: %s", sc.Text())
+		}
+		out = append(out, m)
+	}
+	return out
+}
+
+// auditSameRow: the system-log line is the file's row, every field equal,
+// with `audit` and `pairing` added.
+func auditSameRow(t *testing.T, what string, file, logged map[string]any, wantFile, wantPairing string) {
+	t.Helper()
+	if logged["audit"] != wantFile || logged["pairing"] != wantPairing {
+		t.Fatalf("%s: system-log line names audit=%v pairing=%v, want %q %q", what, logged["audit"], logged["pairing"], wantFile, wantPairing)
+	}
+	for k, v := range file {
+		if lv, ok := logged[k]; !ok || !jsonEq(lv, v) {
+			t.Fatalf("%s: field %q is %v in the file and %v in the system log", what, k, v, logged[k])
+		}
+	}
+	if len(logged) != len(file)+2 {
+		t.Fatalf("%s: system-log line has %d fields, the file row %d (+2 expected)", what, len(logged), len(file))
+	}
+}
+
+func jsonEq(a, b any) bool {
+	ja, _ := json.Marshal(a)
+	jb, _ := json.Marshal(b)
+	return string(ja) == string(jb)
+}
+
+// TestAuditRowsReachTheSystemLog: every audit writer on this box sends the row
+// it writes to its file to the system log as well, unchanged, naming the file
+// and the pairing: the management lane (mgmt_audit.jsonl), pairing changes
+// (pairing_audit.jsonl), and a pairing's audit.jsonl for store ops and op runs,
+// terminals and root commands.
+func TestAuditRowsReachTheSystemLog(t *testing.T) {
+	t.Setenv("XDG_CONFIG_HOME", t.TempDir())
+	t.Setenv("HOME", t.TempDir())
+	got := auditSinkCapture(t)
+	cfgDir, err := configDir()
+	if err != nil {
+		t.Fatal(err)
+	}
+	pairDir := filepath.Join(t.TempDir(), "acme.host.example")
+	if err := os.MkdirAll(pairDir, 0o700); err != nil {
+		t.Fatal(err)
+	}
+
+	mgmtAudit("acme.host.example", "uptime", "observe", "allowed", "", 0, []string{"a"}, "")
+	pairingAudit("pairing_add", "other.host.example", "ok")
+	dcnAudit(pairDir, "dcn_read", 7, "/srv/notes.txt", "allowed")
+	sessionAudit(pairDir, "session", "s-1", "ops", "start", 0, "opened")
+	rootExecAudit(pairDir, map[string]any{"session_id": "s-2", "event": "stop", "exit": 0})
+
+	lines := got()
+	if len(lines) != 5 {
+		t.Fatalf("got %d system-log lines, want 5: %v", len(lines), lines)
+	}
+	mgmt := auditFileRows(t, filepath.Join(cfgDir, mgmtAuditFile))
+	pair := auditFileRows(t, filepath.Join(cfgDir, auditFilePairing))
+	per := auditFileRows(t, filepath.Join(pairDir, auditFilePairingD))
+	if len(mgmt) != 1 || len(pair) != 1 || len(per) != 3 {
+		t.Fatalf("file rows: mgmt %d, pairing %d, per-pairing %d; want 1, 1, 3", len(mgmt), len(pair), len(per))
+	}
+	auditSameRow(t, "management lane", mgmt[0], lines[0], mgmtAuditFile, "acme.host.example")
+	auditSameRow(t, "pairing change", pair[0], lines[1], auditFilePairing, "")
+	auditSameRow(t, "store op", per[0], lines[2], auditFilePairingD, "acme.host.example")
+	auditSameRow(t, "terminal", per[1], lines[3], auditFilePairingD, "acme.host.example")
+	auditSameRow(t, "root command", per[2], lines[4], auditFilePairingD, "acme.host.example")
+}
+
+// TestAuditFileRowStandsWithoutASystemLog: a system log that takes nothing
+// costs the copy only; the row is still written to its file.
+func TestAuditFileRowStandsWithoutASystemLog(t *testing.T) {
+	prev := auditSink
+	auditSink = func([]byte) {}
+	t.Cleanup(func() { auditSink = prev })
+	pairDir := filepath.Join(t.TempDir(), "acme.host.example")
+	if err := os.MkdirAll(pairDir, 0o700); err != nil {
+		t.Fatal(err)
+	}
+	dcnAudit(pairDir, "dcn_read", 1, "/x", "allowed")
+	if rows := auditFileRows(t, filepath.Join(pairDir, auditFilePairingD)); len(rows) != 1 {
+		t.Fatalf("file rows = %d, want 1", len(rows))
+	}
+}
+
+// TestAuditSystemLogKeepsNumbersExact: a large id goes to the system log as
+// the digits the file has, not as a float.
+func TestAuditSystemLogKeepsNumbersExact(t *testing.T) {
+	var line string
+	prev := auditSink
+	auditSink = func(l []byte) { line = string(l) }
+	t.Cleanup(func() { auditSink = prev })
+	auditToSystemLog(auditFilePairingD, "p", []byte(`{"store_id":9007199254740993}`))
+	if !json.Valid([]byte(line)) || !strings.Contains(line, `"store_id":9007199254740993`) {
+		t.Fatalf("system-log line %s does not carry the id exactly", line)
+	}
+}

client-go/auditlog_windows.go

@@ -0,0 +1,54 @@
+//go:build windows
+
+package main
+
+import (
+	"strings"
+	"sync"
+
+	"golang.org/x/sys/windows/svc/eventlog"
+)
+
+var (
+	auditEventMu  sync.Mutex
+	auditEventLog *eventlog.Log
+)
+
+// systemAuditWrite sends one line to the Application event log under the
+// source PrivataeBastionAudit, event id 1, at information level. The source
+// is registered when the system service is installed
+// (registerAuditEventSource); before that, or if the log refuses, the copy is
+// lost and the row in the file is not. Called only from auditQ's drain
+// (auditlog.go): the event log service can be slow to answer.
+func systemAuditWrite(line []byte) {
+	auditEventMu.Lock()
+	defer auditEventMu.Unlock()
+	if auditEventLog == nil {
+		l, err := eventlog.Open(auditEventSource)
+		if err != nil {
+			return
+		}
+		auditEventLog = l
+	}
+	if err := auditEventLog.Info(auditEventID, string(line)); err != nil {
+		_ = auditEventLog.Close()
+		auditEventLog = nil
+	}
+}
+
+// registerAuditEventSource registers PrivataeBastionAudit as an Application
+// event log source, so the rows show with their own source name. Already
+// registered is success.
+func registerAuditEventSource() error {
+	err := eventlog.InstallAsEventCreate(auditEventSource, eventlog.Info|eventlog.Warning|eventlog.Error)
+	if err != nil && strings.Contains(err.Error(), "already exists") {
+		return nil
+	}
+	return err
+}
+
+// removeAuditEventSource unregisters the source when the service is removed.
+// Events already written stay in the log.
+func removeAuditEventSource() {
+	_ = eventlog.Remove(auditEventSource)
+}

client-go/backupfetch_test.go

@@ -16,7 +16,6 @@ import (
 	"path/filepath"
 	"strings"
 	"testing"
-	"time"
 
 	"github.com/gorilla/websocket"
 )
@@ -64,48 +63,6 @@ func drive(t *testing.T, backupRoot string, req backupFetchReq) (string, string,
 	}
 }
 
-func awaitJob(t *testing.T, backupRoot, id string) backupJob {
-	t.Helper()
-	deadline := time.Now().Add(10 * time.Second)
-	for time.Now().Before(deadline) {
-		out, _, code := drive(t, backupRoot, backupFetchReq{Op: "status", JobID: id})
-		if code != 0 {
-			t.Fatalf("status exited %d", code)
-		}
-		var j backupJob
-		if err := json.Unmarshal([]byte(out), &j); err != nil {
-			t.Fatalf("status is not JSON: %v (%s)", err, out)
-		}
-		if j.State != "running" {
-			return j
-		}
-		time.Sleep(25 * time.Millisecond)
-	}
-	t.Fatal("job never left running")
-	return backupJob{}
-}
-
-func startAndAwait(t *testing.T, root string, req backupFetchReq) backupJob {
-	t.Helper()
-	out, errOut, code := drive(t, root, req)
-	if code != 0 {
-		t.Fatalf("start exited %d: %s", code, errOut)
-	}
-	var started struct {
-		ID    string `json:"id"`
-		State string `json:"state"`
-	}
-	if err := json.Unmarshal([]byte(out), &started); err != nil {
-		t.Fatalf("start reply is not JSON: %v (%s)", err, out)
-	}
-	if started.State == "done" { // already-present short circuit
-		var j backupJob
-		_ = json.Unmarshal([]byte(out), &j)
-		return j
-	}
-	return awaitJob(t, root, started.ID)
-}
-
 func artifactServer(t *testing.T, body []byte) (*httptest.Server, string) {
 	t.Helper()
 	srv := httptest.NewTLSServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {

client-go/backupreseal.go

@@ -655,6 +655,9 @@ func cmdBackupDecide(host, prefix string, approve bool) error {
 			}
 		}
 		r, err = resealApprove(dir, who, id, code)
+		if err != nil {
+			return err
+		}
 	} else {
 		r, err = resealRefuse(dir, resealWho{Name: cfg.HostName}, id, "")
 	}

client-go/backupreseal_bind_test.go

@@ -10,6 +10,7 @@ package main
 import (
 	"encoding/json"
 	"fmt"
+	"os"
 	"path/filepath"
 	"strings"
 	"testing"
@@ -149,3 +150,36 @@ func TestHostNumberComesFromTheTunnelAddress(t *testing.T) {
 		}
 	}
 }
+
+// With a second factor enrolled, a wrong code is refused by resealApprove,
+// and cmdBackupDecide returns that refusal. It once checked the outer err
+// while the approval's error went to the one its `who, err :=` declared, so
+// a wrong code fell through to r.State on a nil request and panicked.
+func TestResealDecideWithAWrongCodeReturnsTheRefusal(t *testing.T) {
+	dirs := withFixturePairings(t, map[string]int64{fixtureHost: fixtureEntity})
+	fx := newBackupFixture(t, "")
+	fx.installAsBastionKey(t)
+	own, _ := age.GenerateX25519Identity()
+	id := resealStartAs(t, fixtureHost, dirs[fixtureHost], resealReq{Op: "start", Sidecar: string(fx.sidecar(t)),
+		To: own.Recipient().String(), Stamp: fixtureStamp, HostLabel: "B"})
+	if err := stepupSave(dirs[fixtureHost], &stepupFile{SecretB32: "GEZDGNBVGY3TQOJQGEZDGNBVGY3TQOJQ",
+		EnrolledAt: "2026-08-29T00:00:00Z", LastCounter: -1}); err != nil {
+		t.Fatal(err)
+	}
+	r, w, err := os.Pipe()
+	if err != nil {
+		t.Fatal(err)
+	}
+	_, _ = w.WriteString("not-a-code\n")
+	_ = w.Close()
+	realStdin := os.Stdin
+	os.Stdin = r
+	t.Cleanup(func() { os.Stdin = realStdin; _ = r.Close() })
+
+	if err := cmdBackupDecide(fixtureHost, id, true); err == nil {
+		t.Fatal("a wrong second-factor code approved the request")
+	}
+	if st := resealStatusAs(t, fixtureHost, dirs[fixtureHost], id); st["state"] == "done" || st["sidecar"] != "" {
+		t.Fatalf("a wrong code handed the backup key over: %v", st)
+	}
+}

client-go/beatnow.go

@@ -0,0 +1,80 @@
+package main
+
+import (
+	"context"
+	"sync"
+	"time"
+)
+
+// A beat out of turn, for every pairing on this box.
+//
+// Each running pairing beats its own Host on a 60-second ticker, and every
+// beat carries the box's pairings as `peers` with the state of each (peers.go).
+// That is how every Host on the machine learns that a pairing went active or was
+// revoked. A pairing_grant_sync or pairing_grant_drop changes that state on disk
+// at once, but nothing beat: every Host kept the old state for up to a minute.
+// So a pairing approved a second after a beat still read as pending for the
+// next minute, and a Host asking to connect to it in that time was refused. The
+// anchor Host can read its own confirmed grant, but no other Host on the
+// machine has one, and a revoke reached the other Hosts no sooner.
+//
+// So a change to a pairing's state asks every running pairing's beat loop to beat
+// now. The signal carries nothing: the beat reads the pairings from disk as it
+// always does, so it reports the state the change wrote. One pending signal per
+// loop is kept and the rest are dropped, so a burst of changes is one extra beat
+// per Host, never one per change.
+
+var beatNow = struct {
+	sync.Mutex
+	subs map[chan struct{}]struct{}
+}{subs: map[chan struct{}]struct{}{}}
+
+// subscribeBeatNow registers one beat loop. The channel holds at most one
+// pending signal; stop unregisters it.
+func subscribeBeatNow() (<-chan struct{}, func()) {
+	ch := make(chan struct{}, 1)
+	beatNow.Lock()
+	beatNow.subs[ch] = struct{}{}
+	beatNow.Unlock()
+	return ch, func() {
+		beatNow.Lock()
+		delete(beatNow.subs, ch)
+		beatNow.Unlock()
+	}
+}
+
+// kickBeats asks every registered beat loop to beat now. It never blocks: a
+// loop that already has a signal pending, or is mid-beat, beats once more.
+func kickBeats() {
+	beatNow.Lock()
+	defer beatNow.Unlock()
+	for ch := range beatNow.subs {
+		select {
+		case ch <- struct{}{}:
+		default:
+		}
+	}
+}
+
+// runBeatLoop beats every `every`, and also whenever kickBeats asks. beforeTick
+// runs before a ticked beat only (the tunnel's path failover, which belongs to
+// the clock, not to a pairing change); it may be nil. It returns when ctx ends.
+func runBeatLoop(ctx context.Context, every time.Duration, beforeTick, beat func()) {
+	kick, stop := subscribeBeatNow()
+	defer stop()
+	t := time.NewTicker(every)
+	defer t.Stop()
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-t.C:
+			if beforeTick != nil {
+				beforeTick()
+			}
+			beat()
+		case <-kick:
+			beat()
+		}
+	}
+}

client-go/beatnow_pairing_test.go

@@ -0,0 +1,66 @@
+//go:build !windows
+
+// Tagged !windows like anchor_test.go, whose box helpers (setupBox, driveOp,
+// signTerms) it uses; beatnow_test.go runs the signal and the loop everywhere.
+
+package main
+
+import (
+	"crypto/ed25519"
+	"os"
+	"testing"
+	"time"
+)
+
+// optest v0.16.145: an approval left every Host on the machine reading
+// "pending" until the next 60-second beat. A grant that flips a pairing active,
+// and a revoke, ask every pairing to beat now; a refused grant, and a re-sync
+// of a pairing that is already active, change no state and ask nothing.
+func TestPairingStateChangeBeatsEveryHost(t *testing.T) {
+	priv := setupBox(t)
+	writePairingState(t, tMember, pairingPending)
+	hn, _ := os.Hostname()
+	future := time.Now().Add(time.Hour).UTC().Format(time.RFC3339)
+	kick, stop := subscribeBeatNow()
+	defer stop()
+
+	_, otherPriv, _ := ed25519.GenerateKey(nil)
+	terms, sig := signTerms(t, otherPriv, grantTerms(tMember, []string{"observe"}, future))
+	if _, _, code := driveOp(t, tAnchor, "pairing_grant_sync", []string{terms, sig}); code != 126 {
+		t.Fatalf("a grant under another key must be refused (126): %d", code)
+	}
+	if signalled(kick, 100*time.Millisecond) {
+		t.Fatal("a refused grant asked for a beat")
+	}
+
+	terms, sig = signTerms(t, priv, grantTerms(tMember, []string{"observe"}, future))
+	if _, errOut, code := driveOp(t, tAnchor, "pairing_grant_sync", []string{terms, sig}); code != 0 {
+		t.Fatalf("a good grant must be accepted: exit %d, %s", code, errOut)
+	}
+	if loadState(t, tMember) != pairingActive {
+		t.Fatal("grant_sync must flip the pairing active")
+	}
+	if !signalled(kick, time.Second) {
+		t.Fatal("a grant that flipped a pairing active did not ask every pairing to beat now")
+	}
+
+	terms, sig = signTerms(t, priv, grantTerms(tMember, []string{"observe", "dcn"}, future))
+	if _, errOut, code := driveOp(t, tAnchor, "pairing_grant_sync", []string{terms, sig}); code != 0 {
+		t.Fatalf("a re-sync of an active pairing must be accepted: exit %d, %s", code, errOut)
+	}
+	if signalled(kick, 100*time.Millisecond) {
+		t.Fatal("a re-sync that changed no pairing state asked for a beat")
+	}
+
+	terms, sig = signTerms(t, priv, map[string]any{"anchor": tAnchor, "issued_at": time.Now().UTC().Format(time.RFC3339),
+		"kind": "pairing_revoke", "server": hn, "subject": tMember})
+	if _, errOut, code := driveOp(t, tAnchor, "pairing_grant_drop", []string{tMember, terms, sig}); code != 0 {
+		t.Fatalf("a good revoke must be accepted: %d %s", code, errOut)
+	}
+	if loadState(t, tMember) != pairingRevoked {
+		t.Fatal("grant_drop must mark the pairing revoked")
+	}
+	if !signalled(kick, time.Second) {
+		t.Fatal("a revoke did not ask every pairing to beat now")
+	}
+}

client-go/beatnow_test.go

@@ -0,0 +1,79 @@
+package main
+
+import (
+	"context"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+// signalled reports whether ch carries a signal within d.
+func signalled(ch <-chan struct{}, d time.Duration) bool {
+	select {
+	case <-ch:
+		return true
+	case <-time.After(d):
+		return false
+	}
+}
+
+// One signal reaches every registered loop, a burst is one pending signal per
+// loop, and a loop that stopped is not signalled.
+func TestKickBeatsReachesEveryLoopOnce(t *testing.T) {
+	a, stopA := subscribeBeatNow()
+	defer stopA()
+	b, stopB := subscribeBeatNow()
+	gone, stopGone := subscribeBeatNow()
+	stopGone()
+	kickBeats()
+	kickBeats()
+	kickBeats()
+	stopB()
+	for name, ch := range map[string]<-chan struct{}{"a": a, "b": b} {
+		if !signalled(ch, time.Second) {
+			t.Fatalf("loop %s was not asked to beat", name)
+		}
+		if signalled(ch, 50*time.Millisecond) {
+			t.Fatalf("loop %s holds more than one pending beat after a burst", name)
+		}
+	}
+	if signalled(gone, 50*time.Millisecond) {
+		t.Fatal("a loop that unregistered was still signalled")
+	}
+}
+
+// The beat loop beats on a kick without waiting for its tick, and the
+// tick-only step (path failover) does not run for it.
+func TestRunBeatLoopBeatsOnKick(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	var beats, ticks atomic.Int32
+	done := make(chan struct{})
+	go func() {
+		defer close(done)
+		runBeatLoop(ctx, time.Hour, func() { ticks.Add(1) }, func() { beats.Add(1) })
+	}()
+	// The loop registers inside its goroutine; keep asking until it hears.
+	deadline := time.Now().Add(3 * time.Second)
+	for beats.Load() == 0 && time.Now().Before(deadline) {
+		kickBeats()
+		time.Sleep(10 * time.Millisecond)
+	}
+	if beats.Load() == 0 {
+		t.Fatal("a kicked beat loop did not beat before its 1-hour tick")
+	}
+	if ticks.Load() != 0 {
+		t.Fatal("a kicked beat ran the tick-only step")
+	}
+	cancel()
+	select {
+	case <-done:
+	case <-time.After(2 * time.Second):
+		t.Fatal("the beat loop did not return when its context ended")
+	}
+	beatNow.Lock()
+	left := len(beatNow.subs)
+	beatNow.Unlock()
+	if left != 0 {
+		t.Fatalf("a returned beat loop stayed registered (%d left)", left)
+	}
+}

client-go/ciimage_test.go

@@ -0,0 +1,53 @@
+package main
+
+import (
+	"os"
+	"path/filepath"
+	"regexp"
+	"testing"
+)
+
+// The Go toolchain image is pinned by digest in deploy/build.sh, which builds
+// releases, and again in each forge's CI, so CI tests with the compiler the
+// release is built with. GitHub's workflow checks that the three agree, but
+// only GitHub runs that file, and only once a commit is published there. This
+// test runs wherever go test runs, so a digest changed in one place and not the
+// others fails before the change is merged.
+var (
+	releaseImageRe = regexp.MustCompile(`(?m)^GO_IMAGE="(golang@sha256:[0-9a-f]{64})"$`)
+	giteaImageRe   = regexp.MustCompile(`(?m)^  GO_IMAGE: (\S+)$`)
+	githubImageRe  = regexp.MustCompile(`(?m)^      image: (\S+)$`)
+)
+
+func TestCIUsesTheReleaseGoImage(t *testing.T) {
+	read := func(rel string) string {
+		data, err := os.ReadFile(filepath.Join("..", rel))
+		if err != nil {
+			t.Fatal(err)
+		}
+		return string(data)
+	}
+	release := releaseImageRe.FindAllStringSubmatch(read("deploy/build.sh"), -1)
+	if len(release) != 1 {
+		t.Fatalf(`deploy/build.sh: want one GO_IMAGE="golang@sha256:<64 hex>" line, found %d`, len(release))
+	}
+	want := release[0][1]
+	for _, c := range []struct {
+		rel string
+		re  *regexp.Regexp
+		n   int
+	}{
+		{".gitea/workflows/ci.yml", giteaImageRe, 1},
+		{".github/workflows/ci.yml", githubImageRe, 3},
+	} {
+		got := c.re.FindAllStringSubmatch(read(c.rel), -1)
+		if len(got) != c.n {
+			t.Errorf("%s: want %d image line(s), found %d", c.rel, c.n, len(got))
+		}
+		for _, g := range got {
+			if g[1] != want {
+				t.Errorf("%s names %s; deploy/build.sh builds releases with %s", c.rel, g[1], want)
+			}
+		}
+	}
+}

client-go/claims_static_test.go

@@ -11,6 +11,7 @@ package main
 // runs) is driven in claims_unix_test.go and beside each lane's own tests.
 
 import (
+	"fmt"
 	"go/ast"
 	"go/parser"
 	"go/token"
@@ -49,6 +50,18 @@ func parseSources(t *testing.T) *srcPkg {
 		if err != nil {
 			t.Fatalf("parse %s: %v", n, err)
 		}
+		p.addFile(n, f)
+	}
+	if len(p.files) < 50 {
+		t.Fatalf("parsed only %d source files — is the test running in the package directory?", len(p.files))
+	}
+	return p
+}
+
+// addFile registers one parsed file and its functions. A test that proves a
+// check can fail adds a synthetic file through here, so it is indexed exactly
+// as a real one would be.
+func (p *srcPkg) addFile(n string, f *ast.File) {
 	p.files[n] = f
 	for _, d := range f.Decls {
 		fd, ok := d.(*ast.FuncDecl)
@@ -66,11 +79,23 @@ func parseSources(t *testing.T) *srcPkg {
 			p.methods[fd.Name.Name] = append(p.methods[fd.Name.Name], key)
 		}
 	}
+}
+
+// funcKeysIn is the p.funcs key of every function declared in file n.
+func (p *srcPkg) funcKeysIn(n string) []string {
+	var out []string
+	for _, d := range p.files[n].Decls {
+		fd, ok := d.(*ast.FuncDecl)
+		if !ok {
+			continue
 		}
-	if len(p.files) < 50 {
-		t.Fatalf("parsed only %d source files — is the test running in the package directory?", len(p.files))
+		key := funcKey(fd)
+		if p.funcs[key] != fd {
+			key = key + "@" + n
 		}
-	return p
+		out = append(out, key)
+	}
+	return out
 }
 
 // funcKey is "name" for a function and "Type.name" for a method.
@@ -488,10 +513,30 @@ func TestListenAddressesAreOnlyEverAllocatedOnLoopback(t *testing.T) {
 	}
 }
 
-// No MCP endpoint. Every HTTP route the binary registers, on any listener,
-// is in this table; none of them, and no op name on the management lane,
-// names MCP; no string in the code addresses an /mcp path or speaks JSON-RPC,
-// the protocol MCP is carried on; and no dependency is an MCP library.
+// No MCP endpoint: the bastion never speaks the protocol. Every HTTP route
+// the binary registers, on any listener, is in this table; none of them, and
+// no op name on the management lane, names MCP; no string in the code is a
+// URL or route with an /mcp path, or speaks JSON-RPC, the protocol MCP is
+// carried on; and no dependency is an MCP library.
+//
+// WHAT CHANGED, AND WHY THE GUARANTEE DID NOT. This test used to refuse any
+// string containing "/mcp" at all. The connection-config reader
+// (mcpconfig.go) has to name the files agent clients declare their tool
+// servers in — .cursor/mcp.json, claude_desktop_config.json, an mcpServers
+// key — and READING A FILE SPEAKS NO PROTOCOL. So the check now targets the
+// protocol instead of the word (mcpReaderViolations):
+//
+//   - no literal anywhere is a URL or route with an mcp path segment;
+//   - the config basenames and keys may appear in mcpconfig.go and nowhere
+//     else, so a second file that starts reading (or writing) them is a
+//     change a reviewer sees here;
+//   - nothing declared in a config-reader file (mcpconfig*.go) imports, or
+//     can reach through this package's call graph, net.Dial or net.Listen,
+//     net/http, os/exec, TLS dialling or the websocket client. The reader is
+//     handed its process list by agents.go; it starts and dials nothing.
+//
+// TestMCPReaderCheckCatchesADialAndAStrayBasename proves each of those can
+// fail.
 func TestNoRouteOpOrDependencySpeaksMCP(t *testing.T) {
 	p := parseSources(t)
 	// A route served by more than one surface names each file, "a.go|b.go":
@@ -547,14 +592,10 @@ func TestNoRouteOpOrDependencySpeaksMCP(t *testing.T) {
 		}
 	}
 
+	for _, v := range mcpReaderViolations(p) {
+		t.Error(v)
+	}
 	for name, f := range p.files {
-		for _, s := range stringLits(f) {
-			l := strings.ToLower(s)
-			if strings.Contains(l, "/mcp") || strings.Contains(l, "jsonrpc") ||
-				strings.Contains(l, "modelcontextprotocol") || strings.Contains(l, "tools/call") {
-				t.Errorf("%s carries %q: nothing in the bastion addresses an MCP endpoint or speaks JSON-RPC", name, s)
-			}
-		}
 		for _, imp := range f.Imports {
 			if strings.Contains(strings.ToLower(imp.Path.Value), "mcp") {
 				t.Errorf("%s imports %s", name, imp.Path.Value)
@@ -575,6 +616,346 @@ func TestNoRouteOpOrDependencySpeaksMCP(t *testing.T) {
 	}
 }
 
+// mcpConfigNames are the file basenames and keys agent clients declare tool
+// servers under. Only mcpconfig.go may name them. settings.json and
+// config.toml are left out on purpose: many programs use those names (the
+// coder lane reads the NVIDIA runtime's config.toml), so they say nothing
+// about tool servers; the keys beside them do.
+var mcpConfigNames = []string{"claude_desktop_config", "mcp.json", "mcp_config.json", ".claude.json",
+	"mcpservers", "context_servers", "mcp_servers",
+	// The clients' own directories, which name a settings.json or
+	// config.toml as surely as the basename would.
+	".cursor", ".codex", ".zed", ".codeium", ".vscode"}
+
+// mcpURLOrRoute reports whether a string literal is a URL or route that
+// addresses an mcp path. Outside a URL, every "/mcp" counts (/mcp, /mcp/x,
+// /mcp-proxy) except a client's config FILE name: /mcp.json and
+// /mcp_config.json.
+func mcpURLOrRoute(l string) bool {
+	l = strings.ToLower(l)
+	if i := strings.Index(l, "://"); i >= 0 {
+		rest := l[i+3:]
+		if j := strings.IndexByte(rest, '/'); j >= 0 {
+			return strings.Contains(rest[j:], "/mcp")
+		}
+		return false
+	}
+	for i := 0; ; {
+		j := strings.Index(l[i:], "/mcp")
+		if j < 0 {
+			return false
+		}
+		rest := l[i+j+len("/mcp"):]
+		if !strings.HasPrefix(rest, ".json") && !strings.HasPrefix(rest, "_config.json") {
+			return true
+		}
+		i += j + len("/mcp")
+	}
+}
+
+// mcpReaderSink: a package-level function or type that opens a connection,
+// resolves a name, or starts a process — named as a call or as a value.
+func mcpReaderSink(recv, sel string) bool {
+	switch {
+	case recv == "net" && (strings.HasPrefix(sel, "Dial") || strings.HasPrefix(sel, "Listen") ||
+		strings.HasPrefix(sel, "Lookup") || strings.HasPrefix(sel, "Resolve") || sel == "Dialer" || sel == "Resolver"):
+		return true
+	case strings.Contains(recv, "net.Dialer") || strings.Contains(recv, "net.Resolver"):
+		return true
+	case recv == "http" || recv == "exec" || recv == "websocket" || recv == "rpc" || recv == "jsonrpc":
+		return true
+	case recv == "tls" && strings.HasPrefix(sel, "Dial"):
+		return true
+	case (recv == "os" || recv == "syscall" || recv == "unix" || recv == "windows") &&
+		(sel == "StartProcess" || sel == "ForkExec" || sel == "Exec" || strings.HasPrefix(sel, "CreateProcess") ||
+			strings.HasPrefix(sel, "ShellExecute") || sel == "WinExec"):
+		return true
+	case (recv == "syscall" || recv == "unix" || recv == "windows") && (sel == "Socket" || sel == "Connect" || sel == "Bind" ||
+		sel == "Listen" || sel == "Sendto" || sel == "Sendmsg" || sel == "SendmsgN" || strings.HasPrefix(sel, "WSA") ||
+		sel == "ConnectEx" || sel == "GetAddrInfoW" || sel == "GetHostByName" || strings.HasPrefix(sel, "DnsQuery")):
+		return true
+	case (recv == "syscall" || recv == "windows") && (strings.HasPrefix(sel, "LoadLibrary") || sel == "LoadDLL" ||
+		sel == "MustLoadDLL" || sel == "NewLazyDLL" || sel == "NewLazySystemDLL"):
+		// A DLL loaded by name reaches any function in it, past every check here.
+		return true
+	case recv == "registry" && sel == "OpenRemoteKey":
+		return true
+	}
+	return false
+}
+
+// mcpReaderMethodSink: a METHOD call that connects, resolves or starts,
+// whatever the receiver is called — d.Dial on a net.Dialer variable,
+// c.Get on an http.Client, cmd.Run on an exec.Cmd. Over-approximate on
+// purpose: it only applies to code the reader can reach.
+var mcpReaderMethodSink = map[string]bool{"Dial": true, "DialContext": true, "DialTCP": true, "DialUDP": true,
+	"DialUnix": true, "DialIP": true, "Listen": true, "ListenPacket": true, "LookupHost": true, "LookupIP": true,
+	"LookupIPAddr": true, "LookupAddr": true, "LookupNetIP": true, "Do": true, "Get": true, "Post": true,
+	"PostForm": true, "Head": true, "RoundTrip": true, "Start": true, "Run": true, "Output": true,
+	"CombinedOutput": true, "Connect": true}
+
+// mcpReaderAllowedImports is every package a config-reader file may import:
+// parsing, strings, and opening files. Anything else is a change a reviewer
+// has to see here first. x/sys/windows and its registry package are the
+// Windows open (CreateFile and the handle's own type, attributes and owner)
+// and the profile list it checks owners against; each can also start a
+// process, open a socket, load a DLL or open a remote registry, and those
+// calls are sinks in mcpReaderSink like their unix equivalents.
+var mcpReaderAllowedImports = map[string]bool{"bytes": true, "encoding/json": true, "errors": true, "io": true,
+	"io/fs": true, "net": true, "net/url": true, "os": true, "path/filepath": true, "regexp": true,
+	"runtime": true, "sort": true, "strconv": true, "strings": true, "syscall": true, "unicode": true,
+	"unicode/utf8": true, "golang.org/x/sys/unix": true, "golang.org/x/sys/windows": true,
+	"golang.org/x/sys/windows/registry": true}
+
+// mcpValueSink: a sink named as a VALUE rather than called (f := net.Dial).
+// The http, exec and websocket packages also export constants and types
+// (http.StatusOK, websocket.BinaryMessage), which name no connection; only
+// their functions count here.
+func mcpValueSink(recv, sel string) bool {
+	switch recv {
+	case "http", "exec", "websocket", "rpc", "jsonrpc":
+		switch sel {
+		case "Get", "Post", "Head", "PostForm", "Command", "CommandContext", "Dial", "DialContext",
+			"DialHTTP", "DefaultClient", "DefaultDialer", "DefaultTransport", "NewClient":
+			return true
+		}
+		return false
+	}
+	return mcpReaderSink(recv, sel)
+}
+
+// mcpSinksIn lists the network and process calls inside one declaration.
+func mcpSinksIn(n ast.Node) []string {
+	var out []string
+	called := map[*ast.SelectorExpr]bool{}
+	ast.Inspect(n, func(x ast.Node) bool {
+		if c, ok := x.(*ast.CallExpr); ok {
+			if sel, ok := c.Fun.(*ast.SelectorExpr); ok {
+				called[sel] = true
+			}
+		}
+		return true
+	})
+	ast.Inspect(n, func(x ast.Node) bool {
+		switch v := x.(type) {
+		case *ast.SelectorExpr:
+			recv := exprText(v.X)
+			if (called[v] && mcpReaderSink(recv, v.Sel.Name)) || (!called[v] && mcpValueSink(recv, v.Sel.Name)) {
+				out = append(out, recv+"."+v.Sel.Name)
+			}
+		case *ast.CallExpr:
+			if sel, ok := v.Fun.(*ast.SelectorExpr); ok && mcpReaderMethodSink[sel.Sel.Name] {
+				out = append(out, exprText(sel.X)+"."+sel.Sel.Name+"()")
+			}
+		}
+		return true
+	})
+	return out
+}
+
+// mcpReaderViolations is every way the package breaks the guarantee above.
+func mcpReaderViolations(p *srcPkg) []string {
+	var out []string
+	names := make([]string, 0, len(p.files))
+	for n := range p.files {
+		names = append(names, n)
+	}
+	sort.Strings(names)
+	var readerFiles []string
+	for _, name := range names {
+		if strings.HasPrefix(name, "mcpconfig") {
+			readerFiles = append(readerFiles, name)
+		}
+		for _, s := range stringLits(p.files[name]) {
+			l := strings.ToLower(s)
+			if mcpURLOrRoute(s) || strings.Contains(l, "jsonrpc") || strings.Contains(l, "tools/call") {
+				out = append(out, fmt.Sprintf("%s carries %q: nothing in the bastion addresses an MCP endpoint or speaks JSON-RPC", name, s))
+			}
+			// The protocol's namespace. The reader may name the npm scope
+			// its reference servers are published under, to recognise one
+			// running; nothing else may name it at all.
+			if strings.Contains(l, "modelcontextprotocol") && !(name == "mcpconfig.go" && l == "@modelcontextprotocol/") {
+				out = append(out, fmt.Sprintf("%s carries %q: only the config reader may name the reference servers' npm scope", name, s))
+			}
+			if name == "mcpconfig.go" {
+				continue
+			}
+			for _, b := range mcpConfigNames {
+				if strings.Contains(l, b) {
+					out = append(out, fmt.Sprintf("%s names %q (%s): only mcpconfig.go reads agent clients' tool-server config", name, s, b))
+				}
+			}
+		}
+	}
+	if len(readerFiles) == 0 {
+		return append(out, "no mcpconfig*.go file: the reader moved, so this check is no longer checking it")
+	}
+	var roots []string
+	for _, name := range readerFiles {
+		for _, imp := range p.files[name].Imports {
+			path, _ := strconv.Unquote(imp.Path.Value)
+			if !mcpReaderAllowedImports[path] {
+				out = append(out, fmt.Sprintf("%s imports %s: the config reader reads files and nothing else", name, path))
+			}
+		}
+		roots = append(roots, p.funcKeysIn(name)...)
+	}
+	// Package-level variables are nodes too ("var:name"): a function held in
+	// one (var f = net.Dial, var g = func() { ... }) is reached by whoever
+	// reads the variable, and a reader file's own variables are roots, since
+	// their initialisers run with the package.
+	g := p.callGraph()
+	vars := map[string][]ast.Expr{}
+	for _, name := range names {
+		for _, d := range p.files[name].Decls {
+			gd, ok := d.(*ast.GenDecl)
+			if !ok || gd.Tok != token.VAR {
+				continue
+			}
+			for _, sp := range gd.Specs {
+				vs := sp.(*ast.ValueSpec)
+				for _, id := range vs.Names {
+					key := "var:" + id.Name
+					vars[key] = append(vars[key], vs.Values...)
+					if strings.HasPrefix(name, "mcpconfig") {
+						roots = append(roots, key)
+					}
+				}
+			}
+		}
+	}
+	refs := func(n ast.Node) []string {
+		var out []string
+		ast.Inspect(n, func(x ast.Node) bool {
+			if id, ok := x.(*ast.Ident); ok {
+				if _, isVar := vars["var:"+id.Name]; isVar {
+					out = append(out, "var:"+id.Name)
+				}
+				if _, isFn := p.funcs[id.Name]; isFn {
+					out = append(out, id.Name)
+				}
+			}
+			return true
+		})
+		return out
+	}
+	sinks := map[string][]string{}
+	for key, fd := range p.funcs {
+		g[key] = append(g[key], refs(fd)...)
+		if s := mcpSinksIn(fd); len(s) > 0 {
+			sinks[key] = s
+		}
+	}
+	for key, vals := range vars {
+		for _, v := range vals {
+			g[key] = append(g[key], refs(v)...)
+			sinks[key] = append(sinks[key], mcpSinksIn(v)...)
+		}
+	}
+	seen := reach(g, roots, nil)
+	for fn, path := range seen {
+		if what := sinks[fn]; len(what) > 0 {
+			out = append(out, fmt.Sprintf("the config reader can reach %s, which reaches a network or process call (%s): %s",
+				fn, strings.Join(what, ", "), path))
+		}
+	}
+	sort.Strings(out)
+	return out
+}
+
+// THE CHECK ABOVE CAN FAIL. A reader that starts a process, imports a
+// network package, reaches a dialler through another function, or a second
+// file that names a client's config file or an /mcp URL: each one is added
+// as a synthetic source and must be reported. The real tree must report
+// nothing, and the scan must have found sinks and roots to compare, so a
+// broken extraction cannot pass by finding nothing.
+func TestMCPReaderCheckCatchesADialAndAStrayBasename(t *testing.T) {
+	clean := parseSources(t)
+	if v := mcpReaderViolations(clean); len(v) != 0 {
+		t.Fatalf("the real tree must be clean: %v", v)
+	}
+	if len(clean.funcKeysIn("mcpconfig.go")) < 10 {
+		t.Fatal("found almost no functions in mcpconfig.go; the root extraction is broken")
+	}
+	add := func(p *srcPkg, name, src string) {
+		t.Helper()
+		f, err := parser.ParseFile(p.fset, name, src, parser.SkipObjectResolution)
+		if err != nil {
+			t.Fatal(err)
+		}
+		p.addFile(name, f)
+	}
+	cases := []struct {
+		name, want string
+		files      map[string]string
+	}{
+		{"the reader execs", "can reach mcpSneakRun", map[string]string{
+			"mcpconfig_sneak.go": "package main\nfunc mcpSneakRun() { _ = exec.Command(\"x\") }\n"}},
+		{"the reader dials through a helper elsewhere", "can reach sneakDialHelper", map[string]string{
+			"mcpconfig_sneak.go": "package main\nfunc mcpSneakCall() { sneakDialHelper() }\n",
+			"sneak_helper.go":    "package main\nfunc sneakDialHelper() { net.Dial(\"tcp\", \"x\") }\n"}},
+		{"the reader imports net/http", "imports net/http", map[string]string{
+			"mcpconfig_sneak.go": "package main\nimport _ \"net/http\"\n"}},
+		{"another file names a config file", "only mcpconfig.go reads", map[string]string{
+			"sneak_other.go": "package main\nvar sneakPath = \".cursor/mcp.json\"\n"}},
+		{"another file names the servers key", "only mcpconfig.go reads", map[string]string{
+			"sneak_other.go": "package main\nvar sneakKey = \"mcpServers\"\n"}},
+		{"a URL with an mcp path", "addresses an MCP endpoint", map[string]string{
+			"sneak_other.go": "package main\nvar sneakURL = \"https://h.example.com/v1/mcp\"\n"}},
+		{"a route", "addresses an MCP endpoint", map[string]string{
+			"sneak_other.go": "package main\nvar sneakRoute = \"/mcp\"\n"}},
+		{"a route with a suffix", "addresses an MCP endpoint", map[string]string{
+			"sneak_other.go": "package main\nvar sneakRoute = \"/mcp-proxy\"\n"}},
+		{"the reader dials through a Dialer variable", "reaches a network or process call", map[string]string{
+			"mcpconfig_sneak.go": "package main\nfunc mcpSneakD() { var d net.Dialer; d.Dial(\"tcp\", \"x\") }\n"}},
+		{"the reader dials through a function value", "reaches a network or process call", map[string]string{
+			"mcpconfig_sneak.go": "package main\nfunc mcpSneakF() { f := net.Dial; f(\"tcp\", \"x\") }\n"}},
+		{"the reader dials from a package-level function literal", "reaches a network or process call", map[string]string{
+			"mcpconfig_sneak.go": "package main\nvar mcpSneakV = func() { net.Dial(\"tcp\", \"x\") }\n"}},
+		{"the reader opens a raw socket", "reaches a network or process call", map[string]string{
+			"mcpconfig_sneak.go": "package main\nfunc mcpSneakS() { syscall.Socket(2, 1, 0) }\n"}},
+		{"the reader resolves a name", "reaches a network or process call", map[string]string{
+			"mcpconfig_sneak.go": "package main\nfunc mcpSneakL() { net.LookupHost(\"x\") }\n"}},
+		{"the reader imports a websocket client not on a deny list", "imports golang.org/x/net/websocket", map[string]string{
+			"mcpconfig_sneak.go": "package main\nimport ws \"golang.org/x/net/websocket\"\nfunc mcpSneakW() { ws.Dial(\"x\", \"\", \"y\") }\n"}},
+		{"the reader calls a dialler held in another file's variable", "reaches a network or process call", map[string]string{
+			"mcpconfig_sneak.go": "package main\nfunc mcpSneakH() { sneakDialer(\"tcp\", \"x\") }\n",
+			"sneak_helper.go":    "package main\nvar sneakDialer = net.Dial\n"}},
+		{"another file names a client's config directory", "only mcpconfig.go reads", map[string]string{
+			"sneak_other.go": "package main\nvar sneakDir = \".codex/\"\n"}},
+		{"the Windows reader starts a process", "reaches a network or process call", map[string]string{
+			"mcpconfig_sneak.go": "package main\nimport \"golang.org/x/sys/windows\"\nfunc mcpSneakCP() { windows.CreateProcess(nil, nil, nil, nil, false, 0, nil, nil, nil, nil) }\n"}},
+		{"the Windows reader opens a socket", "reaches a network or process call", map[string]string{
+			"mcpconfig_sneak.go": "package main\nimport \"golang.org/x/sys/windows\"\nfunc mcpSneakWS() { windows.WSASocket(2, 1, 0, nil, 0, 0) }\n"}},
+		{"the Windows reader loads a DLL by name", "reaches a network or process call", map[string]string{
+			"mcpconfig_sneak.go": "package main\nimport \"golang.org/x/sys/windows\"\nvar mcpSneakDLL = windows.NewLazySystemDLL(\"ws2_32.dll\")\n"}},
+		{"the Windows reader opens a remote registry", "reaches a network or process call", map[string]string{
+			"mcpconfig_sneak.go": "package main\nimport \"golang.org/x/sys/windows/registry\"\nfunc mcpSneakRR() { registry.OpenRemoteKey(\"h\", registry.LOCAL_MACHINE) }\n"}},
+	}
+	for _, c := range cases {
+		p := parseSources(t)
+		for n, src := range c.files {
+			add(p, n, src)
+		}
+		found := false
+		for _, v := range mcpReaderViolations(p) {
+			if strings.Contains(v, c.want) {
+				found = true
+			}
+		}
+		if !found {
+			t.Errorf("%s: not reported (want %q in %v)", c.name, c.want, mcpReaderViolations(p))
+		}
+	}
+	// A config file name is not a URL path, and a file in the reader may
+	// name one.
+	for _, ok := range []string{".cursor/mcp.json", "mcp_config.json", ".codeium/windsurf/mcp_config.json", "managed-mcp.json"} {
+		if mcpURLOrRoute(ok) {
+			t.Errorf("%q read as an MCP URL or route", ok)
+		}
+	}
+}
+
 // callGraph maps each function to what it calls or refers to, by name.
 // A method call through a value is resolved to every method of that name in
 // the package, so the graph over-approximates: a path it cannot rule out is
@@ -582,6 +963,14 @@ func TestNoRouteOpOrDependencySpeaksMCP(t *testing.T) {
 // followed; none of the handlers below does that.
 func (p *srcPkg) callGraph() map[string][]string {
 	g := map[string][]string{}
+	// The per-file duplicates ("name@file"), by name, built once: looking
+	// them up by prefix for every identifier made this graph cost seconds.
+	dups := map[string][]string{}
+	for k := range p.funcs {
+		if i := strings.IndexByte(k, '@'); i >= 0 {
+			dups[k[:i]] = append(dups[k[:i]], k)
+		}
+	}
 	for key, fd := range p.funcs {
 		if fd.Body == nil {
 			continue
@@ -593,11 +982,7 @@ func (p *srcPkg) callGraph() map[string][]string {
 				if _, ok := p.funcs[v.Name]; ok {
 					out = append(out, v.Name)
 				}
-				for k := range p.funcs {
-					if strings.HasPrefix(k, v.Name+"@") {
-						out = append(out, k)
-					}
-				}
+				out = append(out, dups[v.Name]...)
 			case *ast.CallExpr:
 				if sel, ok := v.Fun.(*ast.SelectorExpr); ok {
 					out = append(out, p.methods[sel.Sel.Name]...)
@@ -794,8 +1179,8 @@ func TestObserveOpsWriteNothing(t *testing.T) {
 			roots = append(roots, fn)
 		}
 	}
-	if len(roots) != 14 {
-		t.Fatalf("expected the 11 observe and 3 computer handlers, found %d", len(roots))
+	if len(roots) != 18 {
+		t.Fatalf("expected the 15 observe and 3 computer handlers, found %d", len(roots))
 	}
 	stop := map[string]bool{"sendMgmt": true, "sendMgmtExit": true, "refuseMgmt": true,
 		"storeRecordObserve": true}
@@ -871,6 +1256,10 @@ func TestObserveOpsWriteNothing(t *testing.T) {
 		"dmesg -T",
 		"iptables -S",
 		"journalctl -n1 -o short-unix --no-pager",
+		// observe:logins: sshd's lines over a window of whole hours.
+		"journalctl <sinceArg> -o short-iso --no-pager -q --lines=50000 _COMM=sshd _COMM=sshd-session _COMM=sshd-auth _UID=0 + _SYSTEMD_UNIT=ssh.service _SYSTEMD_UNIT=sshd.service _UID=0",
+		// observe:journal: one unit this machine's own lists name, over minutes.
+		"journalctl <unitArg> <sinceArg> -o short-iso --no-pager -q --lines=5000",
 		"kubectl --kubeconfig <observeKubeconfig> get nodes -o json",
 		"kubectl --kubeconfig <observeKubeconfig> get pods -A -o json",
 		"nft list chain inet pubwall input",
@@ -885,6 +1274,8 @@ func TestObserveOpsWriteNothing(t *testing.T) {
 		"system_profiler SPApplicationsDataType -json",
 		"systemctl is-active fail2ban",
 		"systemctl list-timers --all --no-legend --plain",
+		"systemctl list-unit-files --no-legend --plain", // observe:journal's unit check
+		"systemctl list-units --all --no-legend --plain",
 		"systemctl list-units --failed --no-legend --plain",
 		"ufw status",
 		"ufw status verbose",

client-go/claims_tunnel_test.go

@@ -90,6 +90,70 @@ func TestATunnelAddressThatIsNotOneHostIsRefused(t *testing.T) {
 	}
 }
 
+// Tunnel endpoints as the host may send them. The first is the hostile
+// shape: the configuration text is one setting per line, so a newline ends
+// the endpoint and starts a setting of the sender's choosing.
+var (
+	badTunnelEndpoints = []string{
+		"203.0.113.7:51820\nallowed_ip=::/0", "[fe80::1%x\nallowed_ip=::/0]:51820",
+		"203.0.113.7:51820\r", "203.0.113.7:51820\t", " 203.0.113.7:51820", "203.0.113.7:51820 ",
+		"relay\x00.example.net:51820", "relay a.example.net:51820", "", "203.0.113.7",
+		"203.0.113.7:0", "203.0.113.7:65536", "203.0.113.7:+80", "relay.example.net:port",
+		"-relay.example.net:51820", "relay..example.net:51820", "[relay.example.net]:51820",
+		"[fe80::1%eth0]:51820", "0.0.0.0:51820", "[::]:51820", "2001:db8::7:51820",
+	}
+	goodTunnelEndpoints = []string{"203.0.113.7:51820", "[2001:db8::7]:51820", "relay-a.example.net:51820"}
+)
+
+// An endpoint that is not one host:port is refused at every place the peer
+// is configured, and nothing is set on the device.
+func TestATunnelEndpointThatIsNotHostPortIsRefused(t *testing.T) {
+	for _, bad := range badTunnelEndpoints {
+		cfg := &Config{WGPrivateKey: testWGKey(t), EndpointPubKey: testWGKey(t),
+			Endpoint: bad, HostTunnelIP: "2001:db8::e:3"}
+		dev := testWGDevice(t, "2001:db8::a:1")
+		if err := configureTunnelPeer(dev, cfg); err == nil {
+			t.Errorf("%q: configureTunnelPeer must refuse", bad)
+		}
+		if err := swapPeer(cfg, dev, bad, testWGKey(t)); err == nil {
+			t.Errorf("%q: swapPeer must refuse", bad)
+		}
+		if peers := wgPeers(t, dev); len(peers) != 0 {
+			t.Errorf("%q: a refused endpoint must leave the device with no peer, it holds %v", bad, peers)
+		}
+	}
+}
+
+// uapiRecorder keeps the configuration text it is given, in place of a device.
+type uapiRecorder struct{ texts []string }
+
+func (r *uapiRecorder) IpcSet(s string) error { r.texts = append(r.texts, s); return nil }
+
+// Whatever endpoint the pairing holds, the text the device is given names
+// one peer with exactly one allowed address, the host's alias, or the
+// pairing is refused before any text is built.
+func TestThePeerConfigurationHasExactlyOneAllowedAddress(t *testing.T) {
+	for _, ep := range append(append([]string{}, goodTunnelEndpoints...), badTunnelEndpoints...) {
+		rec := &uapiRecorder{}
+		cfg := &Config{WGPrivateKey: testWGKey(t), EndpointPubKey: testWGKey(t),
+			Endpoint: ep, HostTunnelIP: "2001:db8::e:3"}
+		if err := configureTunnelPeer(rec, cfg); err != nil {
+			continue
+		}
+		for _, text := range rec.texts {
+			seen := map[string]int{}
+			for _, line := range strings.Split(strings.TrimSuffix(text, "\n"), "\n") {
+				k, _, _ := strings.Cut(line, "=")
+				seen[k]++
+			}
+			if seen["allowed_ip"] != 1 || seen["public_key"] != 1 || seen["endpoint"] != 1 ||
+				!strings.Contains(text, "\nallowed_ip=2001:db8::e:3/128\n") {
+				t.Errorf("%q: the configuration must name one peer with one allowed address, got %q", ep, text)
+			}
+		}
+	}
+}
+
 func TestTheTunnelReachesOnlyThePairedHost(t *testing.T) {
 	for _, hostIP := range []string{"2001:db8::e:3", "2001:db8::e:ffff", "2001:db8::1"} {
 		cfg := &Config{

client-go/coder_egress.go

@@ -10,8 +10,8 @@ package main
 // over the machine's own connection.
 //
 // THE VETTING APPLIES ALWAYS, anchored box or not. The egress lane
-// (egress.go) only vets on an anchored box, because on a local box the host
-// using it is the person's own. A coding agent is not the person: it runs
+// (egress.go) keeps a local box's LAN reachable, because on a local box the
+// host using it is the person's own. A coding agent is not the person: it runs
 // whatever a prompt or a repository talks it into, so this proxy reaches the
 // public internet and nothing else — not this machine, not its LAN, not the
 // cloud metadata address, nor any address this machine takes as its own (a
@@ -23,7 +23,7 @@ package main
 import (
 	"bufio"
 	"context"
-	"errors"
+	"io"
 	"net"
 	"os"
 	"path/filepath"
@@ -50,35 +50,12 @@ var (
 	coderDialOut egressDialer = (&net.Dialer{Timeout: egressDialTimeout}).DialContext
 )
 
-// coderInternalIP is internalIP (egress.go) plus the IPv6 forms that carry an
-// IPv4 address and are not covered there: IPv4-compatible ::a.b.c.d (::/96),
-// IPv4-translated ::ffff:0:a.b.c.d, and the deprecated site-local fec0::/10.
-// IPv4-mapped ::ffff:a.b.c.d is covered by internalIP through To4.
+// coderInternalIP is internalIP (egress.go): a coder run's proxy and the
+// backup lanes refuse the same address classes as the egress lane, from the
+// one classifier (classifyIP), the IPv4-compatible, IPv4-translated and
+// site-local fec0::/10 forms included.
 func coderInternalIP(ip net.IP) bool {
-	if internalIP(ip) {
-		return true
-	}
-	if ip.To4() != nil || len(ip) != net.IPv6len {
-		return false
-	}
-	zero := true
-	for _, b := range ip[:12] {
-		if b != 0 {
-			zero = false
-			break
-		}
-	}
-	if zero {
-		return true // ::/96 — never a public host
-	}
-	if ip[0] == 0 && ip[1] == 0 && ip[2] == 0 && ip[3] == 0 && ip[4] == 0 && ip[5] == 0 &&
-		ip[6] == 0 && ip[7] == 0 && ip[8] == 0xff && ip[9] == 0xff && ip[10] == 0 && ip[11] == 0 {
-		return true // ::ffff:0:0/96
-	}
-	if ip[0] == 0xfe && ip[1]&0xc0 == 0xc0 {
-		return true // fec0::/10
-	}
-	return false
+	return internalIP(ip)
 }
 
 // coderLANAllowed: the owner's allow_lan names this requested host (an entry
@@ -174,17 +151,32 @@ func coderVettedAddrs(ctx context.Context, hostport string, allowLAN []string) (
 }
 
 func coderRefuse(c net.Conn, why string) {
-	_, _ = c.Write([]byte("HTTP/1.1 403 Forbidden\r\nContent-Length: 0\r\nConnection: close\r\nX-Privatae-Refused: " +
-		strings.NewReplacer("\r", " ", "\n", " ").Replace(why) + "\r\n\r\n"))
+	writeEgressRefusal(c, "403 Forbidden", why, why)
 }
 
 // handleCoderEgressConn serves one connection from a run: HTTP CONNECT or an
 // absolute-form plain HTTP request, the same parsing as handleEgressConn.
 // Destinations are never logged; refusals are counted on the run.
+// coderEgressMaxHeadBytes bounds a request's line and headers together.
+const coderEgressMaxHeadBytes = 64 << 10
+
+// coderEgressMaxConns bounds the connections one run's proxy carries at once.
+// A whole Chromium profile opens at most 256 sockets at a time (its socket
+// pool limit), and package managers far fewer, so four times the browser's
+// figure never refuses real work; past it a connection is answered 503 and
+// closed, and each one held costs this machine at most its head and two copy
+// buffers.
+const coderEgressMaxConns = 1024
+
 func handleCoderEgressConn(client net.Conn, allowLAN []string, refused *int64) {
 	defer client.Close()
 	_ = client.SetReadDeadline(time.Now().Add(30 * time.Second))
-	br := bufio.NewReader(client)
+	// The request line and headers are read under one bound, the head limit
+	// below has always named: the run's own processes send them, and a line
+	// that never ends must not grow in this machine's memory before that
+	// limit is looked at.
+	headLim := &io.LimitedReader{R: client, N: coderEgressMaxHeadBytes}
+	br := bufio.NewReader(headLim)
 	line, err := br.ReadString('\n')
 	if err != nil {
 		return
@@ -196,7 +188,6 @@ func handleCoderEgressConn(client net.Conn, allowLAN []string, refused *int64) {
 	method, target := parts[0], parts[1]
 	var head strings.Builder
 	head.WriteString(line)
-	hostHdr := ""
 	for {
 		h, err := br.ReadString('\n')
 		if err != nil {
@@ -206,38 +197,41 @@ func handleCoderEgressConn(client net.Conn, allowLAN []string, refused *int64) {
 		if h == "\r\n" || h == "\n" {
 			break
 		}
-		if head.Len() > 64<<10 {
+		if head.Len() > coderEgressMaxHeadBytes {
 			return
 		}
-		if len(h) >= 5 && strings.EqualFold(h[:5], "host:") {
-			hostHdr = strings.TrimSpace(h[5:])
-		}
 	}
 	_ = client.SetReadDeadline(time.Time{})
+	// The head is in; what follows is the stream itself. What the reader
+	// already buffered past the head stays first in line.
+	headLim.N = 1<<63 - 1
 	ctx, cancel := context.WithTimeout(context.Background(), egressDialTimeout)
 	defer cancel()
 
+	// Read the way the bastion's own egress lane reads it (egress_http.go): a
+	// CONNECT target through parseEgressAuthority, and plain HTTP as ONE
+	// request, rebuilt and sent with `Connection: close`, so the run's later
+	// requests on a keep-alive connection never reach the destination unvetted.
 	connect := strings.EqualFold(method, "CONNECT")
 	hp := ""
+	var req egressRequest
 	if connect {
-		hp = ensurePort(target, "443")
-	} else {
-		host := hostHdr
-		if i := strings.Index(target, "://"); i >= 0 {
-			rest := target[i+3:]
-			if s := strings.IndexAny(rest, "/?#"); s >= 0 {
-				host = rest[:s]
-			} else {
-				host = rest
-			}
-			if at := strings.LastIndexByte(host, '@'); at >= 0 {
-				host = host[at+1:]
-			}
-		}
-		if host == "" {
+		host, port, ok := parseEgressAuthority(target, 443)
+		if !ok {
+			atomic.AddInt64(refused, 1)
+			writeEgressRefusal(client, "400 Bad Request", egressNoHost, egressNoHost)
 			return
 		}
-		hp = ensurePort(host, "80")
+		hp = egressHostPort(host, port)
+	} else {
+		r, bad := parseEgressRequest(head.String())
+		if bad != nil {
+			atomic.AddInt64(refused, 1)
+			writeEgressRefusal(client, bad.status, bad.body, bad.body)
+			return
+		}
+		req = r
+		hp = egressHostPort(req.host, req.port)
 	}
 	dialAddrs, why := coderVettedAddrs(ctx, hp, allowLAN)
 	if why != "" {
@@ -245,36 +239,19 @@ func handleCoderEgressConn(client net.Conn, allowLAN []string, refused *int64) {
 		coderRefuse(client, why)
 		return
 	}
-	var upstream net.Conn
-	err = errors.New("no address")
-	for i, a := range dialAddrs {
-		// Each address gets an even share of what is left of the dial budget,
-		// so one that black-holes cannot use it all.
-		actx, acancel := ctx, context.CancelFunc(func() {})
-		if left := len(dialAddrs) - i; left > 1 {
-			if dl, ok := ctx.Deadline(); ok {
-				actx, acancel = context.WithTimeout(ctx, time.Until(dl)/time.Duration(left))
-			}
-		}
-		upstream, err = coderDialOut(actx, "tcp", a)
-		acancel()
-		if err == nil {
-			break
-		}
-	}
+	upstream, err := dialEach(ctx, coderDialOut, dialAddrs)
 	if err != nil {
-		_, _ = client.Write([]byte("HTTP/1.1 502 Bad Gateway\r\nContent-Length: 0\r\nConnection: close\r\n\r\n"))
+		writeEgressRefusal(client, "502 Bad Gateway", egressUnreachable, "")
+		return
+	}
+	if !connect {
+		oneEgressExchange(br, client, upstream, req)
 		return
 	}
-	if connect {
 	if _, err := client.Write([]byte("HTTP/1.1 200 Connection established\r\n\r\n")); err != nil {
 		_ = upstream.Close()
 		return
 	}
-	} else if _, err := upstream.Write([]byte(head.String())); err != nil {
-		_ = upstream.Close()
-		return
-	}
 	splice(br, client, upstream)
 }
 
@@ -317,6 +294,7 @@ func coderListenEgress(runID, dir string, uid int) error {
 	}
 	run := &coderEgressRun{ln: ln}
 	coderEgress[runID] = run
+	slots := make(chan struct{}, coderEgressMaxConns)
 	go func() {
 		for {
 			c, err := ln.Accept()
@@ -325,7 +303,20 @@ func coderListenEgress(runID, dir string, uid int) error {
 			}
 			// allow_lan is read per connection: the owner's edit applies to
 			// running runs at once.
-			go handleCoderEgressConn(c, loadCoderConfig().AllowLAN, &run.refused)
+			select {
+			case slots <- struct{}{}:
+				go func() {
+					defer func() { <-slots }()
+					handleCoderEgressConn(c, loadCoderConfig().AllowLAN, &run.refused)
+				}()
+			default:
+				go func() {
+					defer c.Close()
+					_ = c.SetWriteDeadline(time.Now().Add(time.Second))
+					why := "this run already has " + strconv.Itoa(coderEgressMaxConns) + " connections open"
+					writeEgressRefusal(c, "503 Service Unavailable", why, why)
+				}()
+			}
 		}
 	}()
 	return nil

client-go/coders.go

@@ -1229,10 +1229,17 @@ func coderContainerGone(rt, cname string) bool {
 // anchor pairing on an anchored box (serviceOwnerRefusal's rule).
 // An unknown owner proves nothing about self, so it gets the anchor-only rule
 // too: every run this bastion starts records and labels its owner.
+//
+// A local machine has no anchor, so anchorOnly lets every pairing through
+// there; once it serves several Hosts (localServesSeveralHosts), each stops
+// only the runs it started, as with the servers each built.
 func coderOwnerRefusal(self, owner string) string {
 	if owner != "" && owner == self {
 		return ""
 	}
+	if localServesSeveralHosts() {
+		return "refused: that run was started by " + orUnknown(owner) + "; on a local machine that serves several Hosts each stops only the runs it started"
+	}
 	if why := anchorOnly(self, "stopping a run another host started"); why != "" {
 		return "refused: that run was started by " + orUnknown(owner) + "; " + why
 	}

client-go/coders_linux_test.go

@@ -474,8 +474,14 @@ func TestCoderEgressProxyOverTheSocket(t *testing.T) {
 	} {
 		head, c := ask(req)
 		c.Close()
-		if !strings.HasPrefix(head, "HTTP/1.1 403") || !strings.Contains(head, "X-Privatae-Refused:") {
-			t.Fatalf("must be refused with 403 + X-Privatae-Refused: %q -> %q", req, head)
+		want := "HTTP/1.1 403"
+		if strings.Contains(req, "user@") {
+			// Userinfo is refused outright, before any lookup, as the Host's
+			// egress gate refuses it (parseEgressAuthority); still counted.
+			want = "HTTP/1.1 400"
+		}
+		if !strings.HasPrefix(head, want) || !strings.Contains(head, "X-Privatae-Refused:") {
+			t.Fatalf("must be refused with %s + X-Privatae-Refused: %q -> %q", want, req, head)
 		}
 	}
 	head, c := ask("CONNECT public.example:443 HTTP/1.1\r\n\r\n")

client-go/computer_tree_test.go

@@ -47,7 +47,7 @@ func TestBoundTreeRecountsAndCaps(t *testing.T) {
 		t.Fatalf("full tree: count=%d trunc=%v %+v", count, trunc, out)
 	}
 	// the cap holds on THIS side even if a helper lied about truncation
-	out, count, trunc = boundTree([]*computerNode{win}, 3)
+	_, count, trunc = boundTree([]*computerNode{win}, 3)
 	if count != 3 || !trunc {
 		t.Fatalf("capped: count=%d trunc=%v", count, trunc)
 	}

client-go/config.go

@@ -76,12 +76,20 @@ type Config struct {
 	// that its file does not hold yet (rotate.go, setUnsaved; guarded by
 	// rotation.mu). Unexported: never in the file.
 	savePending bool
+	// invalid is why loadPairings found this pairing's file unusable (an
+	// endpoint that is not one host:port), naming the file; "" for a usable
+	// one. Such a pairing is listed, so status shows it and unpair removes it,
+	// but it never serves: active() is false. Unexported: never in the file.
+	invalid string
 }
 
 const (
 	pairingActive  = "active"
 	pairingPending = "pending"
 	pairingRevoked = "revoked"
+	// pairingInvalid is reported, never stored: the file holds a value this
+	// build refuses (Config.invalid).
+	pairingInvalid = "invalid"
 
 	// transportUnix is Config.Transport for a Host on this machine.
 	transportUnix = "unix"
@@ -103,11 +111,15 @@ func (c *Config) localTransport() bool { return c.unixTransport() || c.tcpTransp
 // active is the ONE place the state field is interpreted: "" and "active" serve,
 // anything else does not. run.go, the heartbeat and the mgmt lane all ask this.
 func (c *Config) active() bool {
-	return c.State == "" || c.State == pairingActive
+	return c.invalid == "" && (c.State == "" || c.State == pairingActive)
 }
 
-// stateName is the state as reported: the empty legacy value reads "active".
+// stateName is the state as reported: the empty legacy value reads "active",
+// and a pairing whose file is unusable reads "invalid" whatever its state.
 func (c *Config) stateName() string {
+	if c.invalid != "" {
+		return pairingInvalid
+	}
 	if c.State == "" {
 		return pairingActive
 	}
@@ -316,17 +328,36 @@ func (c *Config) updateOnDiskIf(apply func(disk *Config) error) error {
 // updatePairingOnDisk is updateOnDisk for a caller that only has the host
 // name (the mgmt handlers). Returns the pairing as written, or nil with
 // errPairingGone when no such file exists.
+//
+// The copy is checked as loadPairings checks a file, so a caller that starts
+// the pairing from it never serves one whose file this build refuses.
 func updatePairingOnDisk(host string, apply func(disk *Config)) (*Config, error) {
+	pd, err := pairingsDir()
+	if err != nil {
+		return nil, err
+	}
 	c := &Config{HostName: host}
 	var out *Config
-	err := c.updateOnDisk(func(d *Config) {
+	err = c.updateOnDisk(func(d *Config) {
 		apply(d)
 		cp := *d
+		cp.invalid = invalidReason(&cp, filepath.Join(pd, host+".json"))
 		out = &cp
 	})
 	return out, err
 }
 
+// invalidReason is why the pairing c, read from file, is never served, naming
+// the file, or "" when it may be served. It is the check the pairing response
+// and the rotate directive get, for a file an older build saved from either.
+func invalidReason(c *Config, file string) string {
+	if err := checkEndpoints(append([]string{c.Endpoint}, c.Endpoints...)...); err != nil {
+		return fmt.Sprintf("pairing %s: %v — it is not served; pair it again (`privatae-bastion pair --replace <token>`) or remove it (`privatae-bastion unpair %s`)",
+			file, err, c.HostName)
+	}
+	return ""
+}
+
 // unpairHost tears down and deletes one pairing: its file and its grant file.
 // The context store and DCN dir are the USER's data and are left in place —
 // deleting a pairing must not delete what the user put on this machine.
@@ -495,31 +526,15 @@ func loadPairings() ([]*Config, error) {
 		if err := json.Unmarshal(raw, &c); err != nil {
 			return nil, fmt.Errorf("pairing %s: %w", e.Name(), err)
 		}
+		// One such file must not stop every other pairing, status or unpair:
+		// it loads, never serves.
+		c.invalid = invalidReason(&c, filepath.Join(pd, e.Name()))
 		cfgs = append(cfgs, &c)
 	}
 	sort.Slice(cfgs, func(i, j int) bool { return cfgs[i].HostName < cfgs[j].HostName })
 	return cfgs, nil
 }
 
-// loadConfig keeps single-pairing call sites working: with exactly one pairing it
-// returns it. With several it REFUSES — a straggler caller must name a host and
-// fail safe, not silently act on an arbitrary one (that ambiguity is the same
-// family of bug as the single-config unpair).
-func loadConfig() (*Config, error) {
-	cfgs, err := loadPairings()
-	if err != nil {
-		return nil, err
-	}
-	switch len(cfgs) {
-	case 0:
-		return nil, fmt.Errorf("no pairings")
-	case 1:
-		return cfgs[0], nil
-	default:
-		return nil, fmt.Errorf("%d hosts are paired — this operation must name one (see --host)", len(cfgs))
-	}
-}
-
 // hostNames lists the paired host names (for error messages).
 func hostNames(cfgs []*Config) []string {
 	names := make([]string, 0, len(cfgs))
@@ -600,7 +615,7 @@ func removeStalePairingFile(hostName string) error {
 func checkPairConflict(cfgs []*Config, handle string, replace bool) (*Config, error) {
 	existing := pairingFor(cfgs, handle)
 	if existing != nil && !replace {
-		return nil, fmt.Errorf("already paired with %s — this daemon serves multiple hosts, so pairing never overwrites; to rotate that host's credentials run `privatae-bastion pair --replace <token>`", existing.HostName)
+		return nil, &alreadyPairedError{host: existing.HostName, msg: fmt.Sprintf("already paired with %s — this daemon serves multiple hosts, so pairing never overwrites; to rotate that host's credentials run `privatae-bastion pair --replace <token>`", existing.HostName)}
 	}
 	if existing == nil && replace {
 		return nil, fmt.Errorf("pair --replace: no existing pairing matches '%s' — pair it normally first", handle)

client-go/console_hostops.go

@@ -142,6 +142,14 @@ func cmdOpsDrop(host, prefix string) error {
 	}
 	dir := cfg.dcnDir()
 	prefix = strings.TrimSpace(prefix)
+	if prefix == "--all" {
+		dropped, err := opsDropWhere(dir, "dropped at this machine's console", func(string, *opGrantRec) bool { return true })
+		if err != nil {
+			return err
+		}
+		fmt.Printf("%d op grant(s) dropped — the host cannot run under any of them again\n", len(dropped))
+		return nil
+	}
 	var hits []string
 	for nonce, rec := range opsLoad(dir).Grants {
 		if rec.Dropped == "" && strings.HasPrefix(nonce, prefix) {

client-go/contain.go

@@ -135,7 +135,7 @@ func maybeContain(pairing, dcnDir string, heartbeatResponse []byte) {
 		// already passed the agent checks above, and an agent whose own
 		// fingerprint is in the accounting (one running as a service, say)
 		// would otherwise be spared in silence — killTree returns 0 for a
-		// spared root and a zero kill is not recorded (review of #100).
+		// spared root and a zero kill is not recorded.
 		root := pid
 		killed := killTree(pid, func(p int) bool {
 			if p == root {

client-go/containfinding.go

@@ -104,7 +104,7 @@ func recordContainResult(pairing string, o findingOrder, outcome string, stopped
 // "Carried out once" held only while this process lived: a bastion that
 // restarted after a kill and before the beat carried the answer (a self-update,
 // a crash) would receive the order again, run it again, and report
-// not_a_finding — or kill a restarted copy (review of #100). So every result
+// not_a_finding — or kill a restarted copy. So every result
 // is written to the config dir as it is recorded and read back on first use.
 
 var containResultsLoaded bool
@@ -165,7 +165,7 @@ func containResultsSaveLocked() {
 	}
 	if b, err := json.Marshal(disk); err == nil {
 		// Atomic: a crash mid-write must not corrupt the record and make the
-		// box forget answers it gave (review of #100).
+		// box forget answers it gave.
 		_ = atomicWrite(p, b)
 	}
 }
@@ -231,7 +231,7 @@ type findingView struct {
 	protected map[int]bool
 	// identity: each finding pid's identity read straight after the walk. The
 	// kill checks it before touching the root, so a pid reused between the walk
-	// and the kill is never signalled (review of #100).
+	// and the kill is never signalled.
 	identity map[int]string
 }
 
@@ -441,9 +441,9 @@ func maybeContainFindings(pairing, dcnDir string, heartbeatResponse []byte) {
 // process in the tree, not only the ones an order names — an accounting entry
 // (accountingPIDs), an app the user is running, anything this bastion started,
 // init, the bastion. killTree does not signal a protected process or walk
-// beneath it (review of #100: stopping a replaced sshd would otherwise have
-// taken every ssh session with it). It also returns the bastion's ANCESTORS —
-// its launch chain — which the finding path refuses by name: if any process in
+// beneath it: without that, stopping a replaced sshd would take every ssh
+// session with it. It also returns the bastion's ANCESTORS — its launch
+// chain — which the finding path refuses by name: if any process in
 // a tree were an ancestor of the bastion, the tree's root would be one too, so
 // the named check always fires first and a tree-level check could never be
 // reached.

client-go/containfinding_test.go

@@ -290,7 +290,7 @@ func TestProcessFindingsFromTheRealComparison(t *testing.T) {
 	}
 }
 
-// THE GUARDS HOLD ACROSS THE WHOLE TREE (review of #100, blocking). A finding
+// THE GUARDS HOLD ACROSS THE WHOLE TREE. A finding
 // whose tree holds an accounting entry, an app, or something the bastion
 // started: those processes and everything beneath them are spared, the rest
 // of the tree is stopped.
@@ -334,8 +334,8 @@ func TestContainFindingRefusesTheBastionsOwnAncestors(t *testing.T) {
 }
 
 // THE REFUSALS liveProcessFindings MAKES, through the real function with only
-// the machine replaced (review of #100: every other test replaced the whole
-// function, so none of these was tested).
+// the machine replaced. Every other test replaces the whole function, so
+// this is the one that tests these refusals.
 //
 // MUTATION: drop the root check.   RESULT: red (no_authority not returned).
 // MUTATION: drop !digestOK.         RESULT: red (an edited accounting is used).
@@ -375,7 +375,7 @@ func TestLiveProcessFindingsRefusesWhatItCannotTrust(t *testing.T) {
 	}
 }
 
-// RESULTS SURVIVE A RESTART (review of #100): an answer recorded before the
+// RESULTS SURVIVE A RESTART: an answer recorded before the
 // process died is still there — and still stops a re-run — after it starts.
 //
 // MUTATION: skip containResultsSaveLocked.  RESULT: red.
@@ -413,7 +413,7 @@ func TestContainFindingRunsAnOrderOnce(t *testing.T) {
 	}
 }
 
-// THE REAL FILTER IS WIRED IN (review of #100): with containFilter left at its
+// THE REAL FILTER IS WIRED IN: with containFilter left at its
 // default, a new process the bastion itself started is filtered by the same
 // deltaFilter the pass uses, and so is never containable; the same process
 // started by anyone else is.
@@ -422,7 +422,7 @@ func TestContainFindingRunsAnOrderOnce(t *testing.T) {
 func TestContainFilterIsTheDeltaPassFilter(t *testing.T) {
 	// The process's AGE is supplied (an hour), not read from a real process:
 	// a real pid's age depends on where the suite runs — pid 1 in a fresh PID
-	// namespace is seconds old and reads as short-lived (review of #100).
+	// namespace is seconds old and reads as short-lived.
 	self := os.Getpid()
 	prevTree, prevAge := processTree, deltaAgeFn
 	t.Cleanup(func() { processTree, deltaAgeFn = prevTree, prevAge })
@@ -446,7 +446,7 @@ func TestContainFilterIsTheDeltaPassFilter(t *testing.T) {
 	}
 }
 
-// THE ROOT'S IDENTITY IS CARRIED FROM THE WALK (review of #100): a pid that
+// THE ROOT'S IDENTITY IS CARRIED FROM THE WALK: a pid that
 // changed hands between the finding walk and the kill is not signalled at all.
 //
 // MUTATION: killTree(pid, protected) instead of killTreeExpect with the view's
@@ -471,7 +471,7 @@ func TestContainFindingDoesNotKillARootThatChangedHandsAfterTheWalk(t *testing.T
 	}
 }
 
-// THE AGENT PATH HAS THE SAME TREE GUARD (review of #100): an accounting entry
+// THE AGENT PATH HAS THE SAME TREE GUARD: an accounting entry
 // or an app inside an agent's tree is spared with its subtree.
 //
 // MUTATION: killTree(pid, spareFromContain) in maybeContain.  RESULT: red.
@@ -497,9 +497,9 @@ func TestAgentContainSparesProtectedProcessesInTheTree(t *testing.T) {
 	}
 }
 
-// AN AGENT WHOSE OWN FINGERPRINT IS IN THE ACCOUNTING IS STILL STOPPED (review
-// of #100, round 3): the tree guard applies BENEATH the agent, never to the
-// agent the host named. Before this, such an agent was spared in silence.
+// AN AGENT WHOSE OWN FINGERPRINT IS IN THE ACCOUNTING IS STILL STOPPED: the
+// tree guard applies BENEATH the agent, never to the agent the host named.
+// Before this, such an agent was spared in silence.
 //
 // MUTATION: killTree(pid, guard) — the guard on the root too.  RESULT: red.
 func TestAgentContainStopsAnAgentThatIsItselfInTheAccounting(t *testing.T) {

client-go/contextsock.go

@@ -43,9 +43,11 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"io"
 	"net"
 	"os"
 	"path/filepath"
+	"sync"
 	"time"
 )
 
@@ -126,16 +128,79 @@ func startContextSock(done <-chan struct{}) {
 		_ = ln.Close()
 		_ = os.Remove(p)
 	}()
-	go func() {
+	go contextSockAccept(ln)
+	uilogf("context socket listening on %s\n", p)
+}
+
+// The socket is 0666 and the daemon runs as root (or SYSTEM), so every user on
+// the machine can make it read. These bound what they can make it hold: an ask
+// is a folder path and a display name, far under contextAskMaxBytes, and no
+// more than contextSockMaxInFlight asks are read at once. Past that a
+// connection is told the daemon is busy and closed; the asker tries again.
+//
+// One person may hold at most contextSockMaxPerUser of those slots, keyed on
+// the identity the kernel gives (never on anything the asker says): without
+// it, one user opening contextSockMaxInFlight idle connections and dialling
+// again as each deadline ran out kept every other user's asks "busy" for as
+// long as they liked. A person adds a folder at a time; four at once is far
+// past that.
+const (
+	contextAskMaxBytes     = 64 << 10
+	contextSockMaxInFlight = 32
+)
+
+// contextSockMaxPerUser is a var only so the test of the overall cap can lift
+// it (every test connection is the same user).
+var contextSockMaxPerUser = 4
+
+var contextSockUsers = struct {
+	sync.Mutex
+	held map[string]int
+}{held: map[string]int{}}
+
+// contextSockTakeUser takes one of user's slots, or reports there is none.
+func contextSockTakeUser(user string) bool {
+	contextSockUsers.Lock()
+	defer contextSockUsers.Unlock()
+	if contextSockUsers.held[user] >= contextSockMaxPerUser {
+		return false
+	}
+	contextSockUsers.held[user]++
+	return true
+}
+
+func contextSockReleaseUser(user string) {
+	contextSockUsers.Lock()
+	defer contextSockUsers.Unlock()
+	if contextSockUsers.held[user]--; contextSockUsers.held[user] <= 0 {
+		delete(contextSockUsers.held, user)
+	}
+}
+
+// contextSockAccept serves ln until it is closed, at most
+// contextSockMaxInFlight connections at a time.
+func contextSockAccept(ln net.Listener) {
+	slots := make(chan struct{}, contextSockMaxInFlight)
 	for {
 		c, err := ln.Accept()
 		if err != nil {
 			return
 		}
-			go contextSockServe(c)
-		}
+		select {
+		case slots <- struct{}{}:
+			go func() {
+				defer func() { <-slots }()
+				contextSockServe(c)
 			}()
-	uilogf("context socket listening on %s\n", p)
+		default:
+			go func() {
+				defer c.Close()
+				_ = c.SetDeadline(time.Now().Add(time.Second))
+				_ = json.NewEncoder(c).Encode(contextAskReply{Why: "the bastion is busy " +
+					"answering other asks on this machine; try again in a moment"})
+			}()
+		}
+	}
 }
 
 func contextSockServe(c net.Conn) {
@@ -154,8 +219,15 @@ func contextSockServe(c net.Conn) {
 		return
 	}
 
+	if !contextSockTakeUser(user) {
+		_ = json.NewEncoder(c).Encode(contextAskReply{User: user, Why: "the bastion is busy " +
+			"answering other asks of yours; try again in a moment"})
+		return
+	}
+	defer contextSockReleaseUser(user)
+
 	var req contextAskRequest
-	if err := json.NewDecoder(c).Decode(&req); err != nil {
+	if err := json.NewDecoder(io.LimitReader(c, contextAskMaxBytes)).Decode(&req); err != nil {
 		_ = json.NewEncoder(c).Encode(contextAskReply{User: user, Why: "the request could not be read"})
 		return
 	}
@@ -200,10 +272,10 @@ func contextAskViaSock(root, name string) (sent bool, host, note string, availab
 	return true, rep.Host, "", true
 }
 
-// contextSockUnsupported is the note used where the platform has no
+// errContextSockUnsupported is the note used where the platform has no
 // implementation yet, so the reason a person sees names the platform rather
 // than pretending the feature is merely broken.
-var contextSockUnsupported = errors.New("peer identity over a local socket is not " +
+var errContextSockUnsupported = errors.New("peer identity over a local socket is not " +
 	"implemented for this platform yet")
 
 func contextSockFmtUser(uid uint32, name string) string {

client-go/contextsock_bounds_test.go

@@ -0,0 +1,130 @@
+package main
+
+// The context socket is 0666: every user on the machine may connect, and the
+// daemon behind it runs as root (or SYSTEM). So what one connection can make
+// the daemon hold must be bounded by the daemon, not by the asker.
+
+import (
+	"net"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"testing"
+	"time"
+)
+
+// An ask is a folder path and a display name. A client that streams a request
+// with no end gets the daemon to read only a bounded amount of it, then a
+// reply and a close; before, the decoder kept every byte until the 30 s
+// deadline, and nothing capped how many such connections were open.
+func TestAContextAskWithNoEndIsReadOnlyToABound(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("no unix-socket peer credentials on windows; contextsock_windows.go says why")
+	}
+	sock := filepath.Join(t.TempDir(), "s.sock")
+	ln, err := net.Listen("unix", sock)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer ln.Close()
+	go func() {
+		c, err := ln.Accept()
+		if err != nil {
+			return
+		}
+		contextSockServe(c)
+	}()
+
+	c, err := net.Dial("unix", sock)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer c.Close()
+	const total = 16 << 20
+	chunk := []byte(strings.Repeat("a", 64<<10))
+	sent := 0
+	if _, err := c.Write([]byte(`{"root":"`)); err != nil {
+		t.Fatal(err)
+	}
+	for sent < total {
+		_ = c.SetWriteDeadline(time.Now().Add(2 * time.Second))
+		n, err := c.Write(chunk)
+		sent += n
+		if err != nil {
+			break
+		}
+	}
+	// The kernel's socket buffers hold a few hundred KiB the daemon never
+	// read; anything near the whole stream means the daemon read it all.
+	if sent >= 4<<20 {
+		t.Fatalf("the daemon accepted %d MiB of one ask with no end; it must stop at a bound", sent>>20)
+	}
+}
+
+// Asks in flight are capped: a connection past the cap is closed at once, so
+// many idle or endless connections cannot multiply the bound above.
+func TestContextAsksInFlightAreCapped(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("no unix-socket peer credentials on windows; contextsock_windows.go says why")
+	}
+	sock := filepath.Join(t.TempDir(), "s.sock")
+	ln, err := net.Listen("unix", sock)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer ln.Close()
+	// Every connection here is the same user; lift the per-user share so the
+	// overall cap is what this test meets (TestOneUserCannotHoldEveryContextSockSlot
+	// holds the share).
+	oldShare := contextSockMaxPerUser
+	contextSockMaxPerUser = contextSockMaxInFlight
+	t.Cleanup(func() { contextSockMaxPerUser = oldShare })
+	go contextSockAccept(ln)
+
+	var held []net.Conn
+	defer func() {
+		for _, h := range held {
+			_ = h.Close()
+		}
+	}()
+	// Each of these sends nothing and holds its slot until its deadline.
+	for i := 0; i < contextSockMaxInFlight; i++ {
+		h, err := net.Dial("unix", sock)
+		if err != nil {
+			t.Fatal(err)
+		}
+		held = append(held, h)
+	}
+	// Give the daemon's accept loop time to take every held connection.
+	time.Sleep(200 * time.Millisecond)
+
+	extra, err := net.Dial("unix", sock)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer extra.Close()
+	_ = extra.SetReadDeadline(time.Now().Add(2 * time.Second))
+	buf := make([]byte, 512)
+	n, rerr := extra.Read(buf)
+	if ne, ok := rerr.(net.Error); ok && ne.Timeout() {
+		t.Fatalf("a connection past the cap of %d was left open; it must be refused", contextSockMaxInFlight)
+	}
+	if n > 0 && !strings.Contains(string(buf[:n]), "busy") {
+		t.Fatalf("a connection past the cap got %q; it must be told the daemon is busy", buf[:n])
+	}
+
+	// Once a held connection goes, a new ask gets a slot again.
+	_ = held[0].Close()
+	held = held[1:]
+	time.Sleep(200 * time.Millisecond)
+	again, err := net.Dial("unix", sock)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer again.Close()
+	_ = again.SetReadDeadline(time.Now().Add(500 * time.Millisecond))
+	n, rerr = again.Read(buf)
+	if ne, ok := rerr.(net.Error); !ok || !ne.Timeout() {
+		t.Fatalf("after a slot was freed, a new ask was not served (read %q, %v)", buf[:n], rerr)
+	}
+}

client-go/contextsock_other.go

@@ -0,0 +1,10 @@
+//go:build !linux && !darwin && !windows
+
+package main
+
+// The release builds linux, darwin and windows (deploy/build.sh). Any other
+// unix (freebsd, openbsd) still builds, with the local context socket unable
+// to tell who is on the other end: contextPeerUser then fails, and
+// contextSockServe answers every connection with that reason and forwards
+// nothing. An ask that cannot be attributed is refused, never let through.
+func contextPeerUID(uintptr) (uint32, error) { return 0, errContextSockUnsupported }

client-go/contextsock_peruser_test.go

@@ -0,0 +1,86 @@
+package main
+
+import (
+	"net"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"testing"
+	"time"
+)
+
+// One person cannot take every slot: past their own share they are told the
+// daemon is busy at once, while slots remain for everyone else. Before, one
+// user's idle connections, re-dialled as each deadline ran out, held all of
+// them and every other user's ask was refused.
+func TestOneUserCannotHoldEveryContextSockSlot(t *testing.T) {
+	if runtime.GOOS == "windows" {
+		t.Skip("no unix-socket peer credentials on windows; contextsock_windows.go says why")
+	}
+	const share = 4 // contextSockMaxPerUser
+	// A test before this one may still be releasing its slots.
+	if n := contextSockSlotsHeld(t, 0); n != 0 {
+		t.Fatalf("%d slots still held from an earlier test", n)
+	}
+	sock := filepath.Join(t.TempDir(), "s.sock")
+	ln, err := net.Listen("unix", sock)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer ln.Close()
+	go contextSockAccept(ln)
+
+	var held []net.Conn
+	defer func() {
+		for _, h := range held {
+			_ = h.Close()
+		}
+	}()
+	for i := 0; i < share; i++ {
+		h, err := net.Dial("unix", sock)
+		if err != nil {
+			t.Fatal(err)
+		}
+		held = append(held, h)
+	}
+	// Wait until the daemon has given those four their slots (a fixed sleep
+	// lost this race on a loaded machine and handed the 5th one first).
+	if n := contextSockSlotsHeld(t, share); n != share {
+		t.Fatalf("the daemon gave %d slots to %d idle connections of one user", n, share)
+	}
+
+	extra, err := net.Dial("unix", sock)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer extra.Close()
+	_ = extra.SetReadDeadline(time.Now().Add(2 * time.Second))
+	buf := make([]byte, 512)
+	n, rerr := extra.Read(buf)
+	if ne, ok := rerr.(net.Error); ok && ne.Timeout() {
+		t.Fatalf("one user held %d idle connections and a %dth was still given a slot: "+
+			"one user can take all %d", share, share+1, contextSockMaxInFlight)
+	}
+	if !strings.Contains(string(buf[:n]), "busy") {
+		t.Fatalf("past one user's share the ask got %q; it must be told the daemon is busy", buf[:n])
+	}
+}
+
+// contextSockSlotsHeld waits up to 10 s for the per-user slots held across
+// all users to reach want, and returns the count it last saw.
+func contextSockSlotsHeld(t *testing.T, want int) int {
+	t.Helper()
+	n := -1
+	for deadline := time.Now().Add(10 * time.Second); time.Now().Before(deadline); time.Sleep(10 * time.Millisecond) {
+		contextSockUsers.Lock()
+		n = 0
+		for _, k := range contextSockUsers.held {
+			n += k
+		}
+		contextSockUsers.Unlock()
+		if n == want {
+			break
+		}
+	}
+	return n
+}

client-go/contextsock_windows.go

@@ -33,11 +33,11 @@ const contextSockDir = `C:\ProgramData\privatae-bastion`
 func contextSockDialable(string) string { return "" }
 
 func contextSockListen(path string) (net.Listener, error) {
-	return nil, contextSockUnsupported
+	return nil, errContextSockUnsupported
 }
 
 func contextPeerUser(c net.Conn) (string, error) {
-	return "", contextSockUnsupported
+	return "", errContextSockUnsupported
 }
 
 var _ = os.Remove

client-go/contextstt.go

@@ -859,10 +859,7 @@ func sttIsNonSpeechMark(text string) bool {
 	// ONE group, not two. "(yes) and (no)" ends in a bracket and is speech; the
 	// opening bracket must be closed only at the very end.
 	inner := t[1 : len(t)-1]
-	if strings.IndexByte(inner, closer) >= 0 {
-		return false
-	}
-	return true
+	return strings.IndexByte(inner, closer) < 0
 }
 
 type sttSegment struct {

client-go/ctxcap_measure_test.go

@@ -26,9 +26,7 @@ import (
 	"github.com/gorilla/websocket"
 )
 
-// uvicornWSMaxSize is the host's WebSocket frame ceiling: the server default,
-// which nothing on the host overrides, so this is the live value.
-const uvicornWSMaxSize = 16 << 20
+// uvicornWSMaxSize (harness.go) is the host's WebSocket frame ceiling.
 
 func TestMeasureCtxPutCapacity(t *testing.T) {
 	if os.Getenv("MEASURE_CTX_PUT") == "" {

client-go/dcn.go

@@ -26,6 +26,7 @@ import (
 	"encoding/base64"
 	"encoding/json"
 	"fmt"
+	"io/fs"
 	"os"
 	"path/filepath"
 	"runtime"
@@ -74,6 +75,10 @@ type dcnGrantRec struct {
 	// the signed bytes (a public key; the terms stay exactly what was
 	// signed). The managed local agent seals the store's catalog to it.
 	SealRecipient string `json:"seal_recipient,omitempty"`
+	// RootReal is where the granted folder led when this box installed the
+	// grant: its real path, pinned here (dcn_store.go), outside the signed
+	// bytes. An op whose root leads anywhere else is refused.
+	RootReal string `json:"root_real,omitempty"`
 }
 
 // dcnTerms is parsed FROM the verified canonical bytes — never from anything else.
@@ -189,6 +194,7 @@ func dcnAudit(dir, op string, storeID int64, path, decision string) {
 		"ts": time.Now().UTC().Format(time.RFC3339), "op": op,
 		"store_id": storeID, "path": path, "decision": decision,
 	})
+	auditToSystemLog(auditFilePairingD, auditPairingOfDir(dir), rec)
 	fh, err := os.OpenFile(filepath.Join(dir, "audit.jsonl"),
 		os.O_APPEND|os.O_CREATE|os.O_WRONLY, 0o600)
 	if err != nil {
@@ -523,6 +529,17 @@ func handleDcnGrantSync(ws *websocket.Conn, dir string, args []string) {
 		dcnRefuse(ws, dir, "grant_sync", payload.StoreID, "", "grant does not verify: "+err.Error())
 		return
 	}
+	// The folder this grant names is pinned as it resolves now; a re-sync or a
+	// renewal of the same root keeps the pin it already has (dcn_store.go).
+	if prev, held := f.Stores[fmt.Sprint(payload.StoreID)]; held {
+		prevRoot := ""
+		if pt, perr := dcnVerifyRec(f.VerifyPubkey, prev); perr == nil {
+			prevRoot = pt.Root
+		}
+		rec.RootReal = dcnPinFor(&prev, prevRoot, terms.Root)
+	} else {
+		rec.RootReal = dcnPinFor(nil, "", terms.Root)
+	}
 	if hn, _ := os.Hostname(); hn != "" && terms.Server != "" && terms.Server != hn {
 		dcnRefuse(ws, dir, "grant_sync", payload.StoreID, "",
 			fmt.Sprintf("grant is for server '%s', this box is '%s'", terms.Server, hn))
@@ -657,15 +674,26 @@ func handleDcnList(ws *websocket.Conn, dir string, args []string) {
 	if terms == nil {
 		return
 	}
-	base := terms.Root
+	base := "."
 	if a.Prefix != "" {
 		p, ok := dcnResolvePath(terms.Root, a.Prefix)
 		if !ok {
 			dcnRefuse(ws, dir, "list", storeID, a.Prefix, "prefix escapes the store root")
 			return
 		}
-		base = p
+		rel, err := dcnRel(terms.Root, p)
+		if err != nil {
+			dcnRefuse(ws, dir, "list", storeID, a.Prefix, "prefix escapes the store root")
+			return
 		}
+		base = filepath.ToSlash(rel)
+	}
+	r, err := dcnOpenStore(dir, storeID, terms)
+	if err != nil {
+		dcnRefuse(ws, dir, "list", storeID, a.Prefix, err.Error())
+		return
+	}
+	defer r.Close()
 	limit := a.Limit
 	if limit <= 0 || limit > dcnMaxListEntries {
 		limit = dcnMaxListEntries
@@ -677,12 +705,14 @@ func handleDcnList(ws *websocket.Conn, dir string, args []string) {
 	}
 	var files []entry
 	truncated := false
-	_ = filepath.WalkDir(base, func(p string, d os.DirEntry, err error) error {
+	// Walked through the store's handle: a link that leads out of it is not
+	// followed, whatever the names do during the walk.
+	_ = fs.WalkDir(r.FS(), base, func(p string, d fs.DirEntry, err error) error {
 		if err != nil {
 			return nil // unreadable subtree: skip, never abort the listing
 		}
-		rel, rerr := filepath.Rel(terms.Root, p)
-		if rerr != nil || rel == "." {
+		rel := filepath.FromSlash(p)
+		if rel == "." {
 			return nil
 		}
 		if dcnExcluded(terms.Excludes, rel) {
@@ -726,8 +756,23 @@ func handleDcnStat(ws *websocket.Conn, dir string, args []string) {
 		dcnRefuse(ws, dir, "stat", storeID, a.Path, "path escapes the store root")
 		return
 	}
-	info, err := os.Stat(abs)
+	rel, err := dcnRel(terms.Root, abs)
 	if err != nil {
+		dcnRefuse(ws, dir, "stat", storeID, a.Path, err.Error())
+		return
+	}
+	r, err := dcnOpenStore(dir, storeID, terms)
+	if err != nil {
+		dcnRefuse(ws, dir, "stat", storeID, a.Path, err.Error())
+		return
+	}
+	defer r.Close()
+	info, err := r.Stat(rel)
+	if err != nil {
+		if dcnEscaped(err) {
+			dcnRefuse(ws, dir, "stat", storeID, a.Path, "path escapes the store root")
+			return
+		}
 		if os.IsNotExist(err) {
 			dcnResult(ws, dir, "stat", storeID, a.Path, map[string]any{
 				"state": "gone", "path": a.Path})
@@ -770,8 +815,25 @@ func handleDcnReadRange(ws *websocket.Conn, dir string, args []string) {
 		dcnRefuse(ws, dir, "read_range", storeID, a.Path, "path escapes the store root")
 		return
 	}
-	fh, err := os.Open(abs)
+	rel, err := dcnRel(terms.Root, abs)
 	if err != nil {
+		dcnRefuse(ws, dir, "read_range", storeID, a.Path, err.Error())
+		return
+	}
+	r, err := dcnOpenStore(dir, storeID, terms)
+	if err != nil {
+		dcnRefuse(ws, dir, "read_range", storeID, a.Path, err.Error())
+		return
+	}
+	defer r.Close()
+	// Opened through the store's handle: a name swapped for a link to a file
+	// outside the store after the check above is refused, never followed.
+	fh, err := r.Open(rel)
+	if err != nil {
+		if dcnEscaped(err) {
+			dcnRefuse(ws, dir, "read_range", storeID, a.Path, "path escapes the store root")
+			return
+		}
 		if os.IsNotExist(err) {
 			dcnResult(ws, dir, "read_range", storeID, a.Path, map[string]any{
 				"state": "gone", "path": a.Path})

client-go/dcn_plan.go

@@ -191,10 +191,28 @@ func dcnReportPath(root string, planID int64) string {
 	return p
 }
 
-// dcnApplyPlan runs the plan: check, rename in order (each re-checked just
-// before it runs; rolling back on the first failure), verify, report. Every
-// path it touches has passed dcnCheckPlan.
+// dcnApplyPlan is dcnApplyPlanIn on root opened where it leads now (the op
+// handler opens the PINNED folder instead, dcn_store.go).
 func dcnApplyPlan(root string, excludes []string, plan *dcnPlanTerms) map[string]any {
+	real, err := dcnRealRoot(root)
+	if err != nil {
+		return map[string]any{"state": "refused", "mode": "apply", "plan_id": plan.PlanID,
+			"reason": "the store's folder cannot be opened: " + err.Error()}
+	}
+	r, err := dcnOpenRootAt(real)
+	if err != nil {
+		return map[string]any{"state": "refused", "mode": "apply", "plan_id": plan.PlanID, "reason": err.Error()}
+	}
+	defer r.Close()
+	return dcnApplyPlanIn(r, root, excludes, plan)
+}
+
+// dcnApplyPlanIn runs the plan: check, rename in order (each re-checked just
+// before it runs; rolling back on the first failure), verify, report. Every
+// path it touches has passed dcnCheckPlan, and every change is made through
+// r, the store's handle, so a folder that becomes a link to somewhere else
+// between the check and the rename is refused rather than followed.
+func dcnApplyPlanIn(r *os.Root, root string, excludes []string, plan *dcnPlanTerms) map[string]any {
 	problems := dcnCheckPlan(root, excludes, plan.Moves)
 	if len(problems) > 0 {
 		return map[string]any{"state": "refused", "mode": "apply", "plan_id": plan.PlanID,
@@ -207,11 +225,11 @@ func dcnApplyPlan(root string, excludes []string, plan *dcnPlanTerms) map[string
 	// and what is NOT (so the user knows exactly what is still displaced).
 	rollback := func() (back, stuck []string) {
 		for i := len(applied) - 1; i >= 0; i-- {
-			if _, err := os.Lstat(applied[i].src); err == nil {
+			if _, err := r.Lstat(applied[i].src); err == nil {
 				stuck = append(stuck, applied[i].from) // something now sits at the source; never replace it
 				continue
 			}
-			if err := os.Rename(applied[i].dst, applied[i].src); err == nil {
+			if err := r.Rename(applied[i].dst, applied[i].src); err == nil {
 				back = append(back, applied[i].from)
 			} else {
 				stuck = append(stuck, applied[i].from)
@@ -230,24 +248,30 @@ func dcnApplyPlan(root string, excludes []string, plan *dcnPlanTerms) map[string
 		return rep
 	}
 	for i, m := range plan.Moves {
-		src, _ := dcnResolvePath(root, m.From)
-		dst, _ := cleanRel(root, m.To)
-		if err := os.MkdirAll(filepath.Dir(dst), 0o755); err != nil {
+		srcAbs, _ := dcnResolvePath(root, m.From)
+		dstAbs, _ := cleanRel(root, m.To)
+		src, serr := dcnRel(root, srcAbs)
+		dst, derr := dcnRel(root, dstAbs)
+		if serr != nil || derr != nil {
+			return fail(i, m, "path escapes the store root")
+		}
+		if d := filepath.Dir(dst); d != "." {
+			if err := r.MkdirAll(d, 0o755); err != nil {
 				return fail(i, m, "could not create the target folder: "+err.Error())
 			}
+		}
 		// Re-check the instant before the rename: the world may have changed since
 		// the pre-check (a file appeared at the target; a parent became a symlink).
 		// rename(2) would replace silently — this lane never lets it.
 		if _, ok := dcnResolvePath(root, m.To); !ok {
 			return fail(i, m, "target's parent no longer resolves inside the store root")
 		}
-		if _, err := os.Lstat(dst); err == nil {
+		if _, err := r.Lstat(dst); err == nil {
 			return fail(i, m, "target appeared since the check (a plan never overwrites)")
 		}
-		if err := os.Rename(src, dst); err != nil {
+		if err := r.Rename(src, dst); err != nil {
 			why := err.Error()
-			var le *os.LinkError
-			if errors.As(err, &le) && errors.Is(le.Err, syscall.EXDEV) {
+			if errors.Is(err, syscall.EXDEV) {
 				why = "target is on a different filesystem than the source (a plan only renames; it never copies and removes)"
 			}
 			return fail(i, m, why)
@@ -257,14 +281,16 @@ func dcnApplyPlan(root string, excludes []string, plan *dcnPlanTerms) map[string
 	// Verify pass: every target present, every source gone, declared sizes hold.
 	var mismatches []dcnPlanProblem
 	for i, m := range plan.Moves {
-		src, _ := dcnResolvePath(root, m.From)
-		dst, _ := cleanRel(root, m.To)
-		info, err := os.Lstat(dst)
+		srcAbs, _ := dcnResolvePath(root, m.From)
+		dstAbs, _ := cleanRel(root, m.To)
+		src, _ := dcnRel(root, srcAbs)
+		dst, _ := dcnRel(root, dstAbs)
+		info, err := r.Lstat(dst)
 		if err != nil {
 			mismatches = append(mismatches, dcnPlanProblem{i, m.From, m.To, "target missing after move"})
 			continue
 		}
-		if _, err := os.Lstat(src); err == nil {
+		if _, err := r.Lstat(src); err == nil {
 			mismatches = append(mismatches, dcnPlanProblem{i, m.From, m.To, "source still present after move"})
 		}
 		if m.SizeBytes != nil && !info.IsDir() && info.Size() != *m.SizeBytes {
@@ -372,9 +398,19 @@ func handleDcnPlanApply(ws *websocket.Conn, dir string, args []string) {
 	// a retry (a lost reply, a host re-run) gets the earlier report back BEFORE
 	// any gate can turn it into a refusal. (The signed expiry is the guard that
 	// does not depend on a file the user can remove.)
+	store, err := dcnOpenStore(dir, storeID, terms)
+	if err != nil {
+		dcnRefuse(ws, dir, op, storeID, planKey, err.Error())
+		return
+	}
+	defer store.Close()
 	repPath := dcnReportPath(terms.Root, plan.PlanID)
+	repRel := ""
 	if repPath != "" {
-		if prev, err := os.ReadFile(repPath); err == nil {
+		repRel, _ = dcnRel(terms.Root, repPath)
+	}
+	if repRel != "" {
+		if prev, err := store.ReadFile(repRel); err == nil {
 			var old map[string]any
 			if json.Unmarshal(prev, &old) == nil && old["state"] == "ok" {
 				old["note"] = "already applied on this machine; this is the earlier report"
@@ -398,13 +434,13 @@ func handleDcnPlanApply(ws *websocket.Conn, dir string, args []string) {
 			return
 		}
 	}
-	rep := dcnApplyPlan(terms.Root, terms.Excludes, plan)
+	rep := dcnApplyPlanIn(store, terms.Root, terms.Excludes, plan)
 	if rep["state"] == "ok" && plan.Approval != "" {
 		markApprovalUsed(dir, plan.Approval) // single use: the files have moved
 	}
-	if repPath != "" {
+	if repRel != "" {
 		if raw, err := json.MarshalIndent(rep, "", "  "); err == nil {
-			_ = atomicWrite(repPath, raw) // the machine's own record; best-effort
+			_ = dcnRootAtomicWrite(store, repRel, raw) // the machine's own record; best-effort
 		}
 	}
 	dcnResult(ws, dir, op, storeID, planKey, rep)

client-go/dcn_rootpin_test.go

@@ -0,0 +1,322 @@
+package main
+
+// A machine store is the folder the person granted, for as long as the grant
+// lives. The bastion often runs as root and the folder belongs to someone who
+// is not, so whoever can write in it, or above it, must not be able to point
+// a read or a move somewhere else: by turning the granted folder into a link
+// after the grant (no race needed), or by swapping a file for a link between
+// the check and the open.
+
+import (
+	"crypto/ed25519"
+	"encoding/base64"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"github.com/gorilla/websocket"
+)
+
+// dcnRunOp runs one DCN handler on the far side of a real WebSocket and
+// returns what it sent: the O payload, the R text and the exit code.
+func dcnRunOp(t *testing.T, h func(ws *websocket.Conn)) (out, errText string, exit int) {
+	t.Helper()
+	up := websocket.Upgrader{}
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		ws, err := up.Upgrade(w, r, nil)
+		if err != nil {
+			return
+		}
+		defer ws.Close()
+		h(ws)
+	}))
+	defer srv.Close()
+	ws, _, err := websocket.DefaultDialer.Dial("ws"+strings.TrimPrefix(srv.URL, "http"), nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer ws.Close()
+	_ = ws.SetReadDeadline(time.Now().Add(10 * time.Second))
+	exit = -1
+	for {
+		_, data, err := ws.ReadMessage()
+		if err != nil || len(data) == 0 {
+			return
+		}
+		switch data[0] {
+		case 'O':
+			out += string(data[1:])
+		case 'R':
+			errText += string(data[1:])
+		case 'X':
+			exit, _ = strconv.Atoi(string(data[1:]))
+			return
+		}
+	}
+}
+
+type dcnStoreFixture struct {
+	dir   string // the pairing's dcn dir
+	base  string
+	store string // the granted folder
+	other string // a folder the grant never named
+	priv  ed25519.PrivateKey
+	pub   ed25519.PublicKey
+}
+
+func newDcnStoreFixture(t *testing.T) *dcnStoreFixture {
+	t.Helper()
+	fx := &dcnStoreFixture{dir: t.TempDir(), base: t.TempDir()}
+	fx.store = filepath.Join(fx.base, "store")
+	fx.other = filepath.Join(fx.base, "other")
+	for _, d := range []string{fx.store, fx.other, filepath.Join(fx.store, "docs")} {
+		if err := os.MkdirAll(d, 0o755); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if err := os.WriteFile(filepath.Join(fx.store, "a.txt"), []byte("MINE"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(fx.other, "a.txt"), []byte("OTHER"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	fx.pub, fx.priv, _ = ed25519.GenerateKey(nil)
+	return fx
+}
+
+// sync installs a read grant for the store through the real grant_sync op.
+func (fx *dcnStoreFixture) sync(t *testing.T, mask ...string) {
+	t.Helper()
+	if len(mask) == 0 {
+		mask = []string{"read"}
+	}
+	rec := signedRec(t, fx.priv, map[string]any{"kind": "dcn_store_grant", "class": "machine",
+		"root": fx.store, "mask": mask,
+		"expires": time.Now().Add(24 * time.Hour).UTC().Format(time.RFC3339)})
+	payload, _ := json.Marshal(map[string]any{"store_id": 1, "terms_canonical": rec.TermsCanonical,
+		"signature": rec.Signature, "verify_pubkey": base64.StdEncoding.EncodeToString(fx.pub)})
+	_, why, exit := dcnRunOp(t, func(ws *websocket.Conn) {
+		handleDcnGrantSync(ws, fx.dir, []string{base64.StdEncoding.EncodeToString(payload)})
+	})
+	if exit != 0 {
+		t.Fatalf("grant_sync: exit %d: %s", exit, why)
+	}
+}
+
+func (fx *dcnStoreFixture) op(t *testing.T, h func(*websocket.Conn, string, []string), args map[string]any) (map[string]any, string, int) {
+	t.Helper()
+	raw, _ := json.Marshal(args)
+	out, why, exit := dcnRunOp(t, func(ws *websocket.Conn) {
+		h(ws, fx.dir, []string{"1", string(raw)})
+	})
+	var obj map[string]any
+	_ = json.Unmarshal([]byte(out), &obj)
+	return obj, why, exit
+}
+
+func (fx *dcnStoreFixture) read(t *testing.T, path string) (string, string, int) {
+	t.Helper()
+	obj, why, exit := fx.op(t, handleDcnReadRange, map[string]any{"path": path})
+	b, _ := base64.StdEncoding.DecodeString(func() string { s, _ := obj["content_b64"].(string); return s }())
+	return string(b), why, exit
+}
+
+// repoint turns the granted folder into a link to the other one.
+func (fx *dcnStoreFixture) repoint(t *testing.T) {
+	t.Helper()
+	if err := os.Rename(fx.store, fx.store+".was"); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.Symlink(fx.other, fx.store); err != nil {
+		t.Skipf("symlinks unavailable: %v", err)
+	}
+}
+
+func TestDcnStoreRepointedAfterTheGrantIsRefused(t *testing.T) {
+	fx := newDcnStoreFixture(t)
+	fx.sync(t)
+	if got, why, exit := fx.read(t, "a.txt"); exit != 0 || got != "MINE" {
+		t.Fatalf("the granted file must read: exit %d %q %s", exit, got, why)
+	}
+	fx.repoint(t)
+	if got, _, exit := fx.read(t, "a.txt"); exit == 0 || got == "OTHER" {
+		t.Fatalf("after the granted folder became a link elsewhere, read_range returned %q (exit %d); it must refuse", got, exit)
+	}
+	if obj, _, exit := fx.op(t, handleDcnStat, map[string]any{"path": "a.txt"}); exit == 0 {
+		t.Fatalf("stat through the repointed folder must refuse, got %v", obj)
+	}
+	if obj, _, exit := fx.op(t, handleDcnList, map[string]any{}); exit == 0 {
+		t.Fatalf("list through the repointed folder must refuse, got %v", obj)
+	}
+}
+
+// A folder inside the store that becomes a link to somewhere else after the
+// grant is the same escape one level down.
+func TestDcnSubfolderRepointedIsRefused(t *testing.T) {
+	fx := newDcnStoreFixture(t)
+	fx.sync(t)
+	if err := os.RemoveAll(filepath.Join(fx.store, "docs")); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.Symlink(fx.other, filepath.Join(fx.store, "docs")); err != nil {
+		t.Skipf("symlinks unavailable: %v", err)
+	}
+	if got, _, exit := fx.read(t, "docs/a.txt"); exit == 0 || got == "OTHER" {
+		t.Fatalf("a link inside the store to a folder outside it read %q (exit %d)", got, exit)
+	}
+}
+
+// The race: a file swapped for a link to a file outside the store between the
+// check and the open. Run for long enough that the old check-then-open loses.
+func TestDcnReadRangeSwapRaceNeverLeavesTheStore(t *testing.T) {
+	fx := newDcnStoreFixture(t)
+	fx.sync(t)
+	target := filepath.Join(fx.store, "a.txt")
+	realTmp := filepath.Join(fx.store, ".real")
+	linkTmp := filepath.Join(fx.store, ".link")
+	var stop atomic.Bool
+	done := make(chan struct{})
+	go func() {
+		defer close(done)
+		for !stop.Load() {
+			_ = os.Symlink(filepath.Join(fx.other, "a.txt"), linkTmp)
+			_ = os.Rename(linkTmp, target)
+			_ = os.WriteFile(realTmp, []byte("MINE"), 0o644)
+			_ = os.Rename(realTmp, target)
+		}
+	}()
+	deadline := time.Now().Add(3 * time.Second)
+	n := 0
+	for time.Now().Before(deadline) {
+		n++
+		if got, _, _ := fx.read(t, "a.txt"); got == "OTHER" {
+			stop.Store(true)
+			<-done
+			t.Fatalf("read %d returned the file outside the store through a swapped link", n)
+		}
+	}
+	stop.Store(true)
+	<-done
+}
+
+// A move plan applies through the store too: a target folder that leads
+// outside the store for an instant between the re-check and the rename must
+// not receive the file.
+func TestDcnPlanMoveRaceNeverLeavesTheStore(t *testing.T) {
+	fx := newDcnStoreFixture(t)
+	if err := os.Remove(filepath.Join(fx.other, "a.txt")); err != nil {
+		t.Fatal(err)
+	}
+	docs := filepath.Join(fx.store, "docs")
+	var stop atomic.Bool
+	done := make(chan struct{})
+	go func() {
+		defer close(done)
+		for !stop.Load() {
+			_ = os.Rename(docs, docs+".real")
+			_ = os.Symlink(fx.other, docs+".link")
+			_ = os.Rename(docs+".link", docs)
+			_ = os.Rename(docs, docs+".link")
+			_ = os.Remove(docs + ".link")
+			_ = os.Rename(docs+".real", docs)
+		}
+	}()
+	defer func() { stop.Store(true); <-done }()
+	deadline := time.Now().Add(3 * time.Second)
+	for i := 0; time.Now().Before(deadline); i++ {
+		_ = os.WriteFile(filepath.Join(fx.store, "a.txt"), []byte("MINE"), 0o644)
+		plan := &dcnPlanTerms{PlanID: int64(i), StoreID: 1, Root: fx.store,
+			Moves: []dcnMove{{From: "a.txt", To: "docs/a.txt"}}}
+		_ = dcnApplyPlan(fx.store, nil, plan)
+		if _, err := os.Lstat(filepath.Join(fx.other, "a.txt")); err == nil {
+			t.Fatalf("move %d landed outside the store, through a folder that was a link for an instant", i)
+		}
+		_ = os.Remove(filepath.Join(fx.store, "docs.real", "a.txt"))
+		_ = os.Remove(filepath.Join(docs, "a.txt"))
+	}
+}
+
+// The plan report is written beside its name and renamed into place. Whoever
+// can write in the store could plant a link at that side name; the write must
+// never follow it, or the root bastion truncates whatever file the link names
+// (an excluded .ssh/authorized_keys included) and fills it with the report.
+func TestDcnReportWriteNeverFollowsAPlantedSideName(t *testing.T) {
+	store := t.TempDir()
+	victim := filepath.Join(store, ".ssh", "authorized_keys")
+	if err := os.MkdirAll(filepath.Dir(victim), 0o700); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(victim, []byte("ssh-ed25519 AAAAexample user@example\n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	plans := filepath.Join(store, ".agent", "plans")
+	if err := os.MkdirAll(plans, 0o700); err != nil {
+		t.Fatal(err)
+	}
+	rel := filepath.Join(".agent", "plans", "7.json")
+	// Every side name a writer could guess: this process's pid, and a spread
+	// of others, all pointing at the victim.
+	for _, pid := range []int{os.Getpid(), 1, 2, 100, 1000} {
+		side := filepath.Join(store, rel+".tmp-"+strconv.Itoa(pid))
+		if err := os.Symlink(filepath.Join("..", "..", ".ssh", "authorized_keys"), side); err != nil {
+			t.Fatal(err)
+		}
+	}
+	r, err := os.OpenRoot(store)
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer r.Close()
+	if err := dcnRootAtomicWrite(r, rel, []byte(`{"state":"ok"}`)); err != nil {
+		t.Fatalf("report write: %v", err)
+	}
+	if got, _ := os.ReadFile(victim); string(got) != "ssh-ed25519 AAAAexample user@example\n" {
+		t.Fatalf("the report write followed a planted link: the victim now holds %q", got)
+	}
+	if got, err := os.ReadFile(filepath.Join(store, rel)); err != nil || string(got) != `{"state":"ok"}` {
+		t.Fatalf("report not in place: %q %v", got, err)
+	}
+	if fi, err := os.Lstat(filepath.Join(store, rel)); err != nil || fi.Mode()&os.ModeSymlink != 0 || fi.Mode().Perm() != 0o600 {
+		t.Fatalf("report is not a 0600 regular file: %v %v", fi, err)
+	}
+}
+
+// A grant installed before pins existed has no root_real. It still works,
+// pins where its folder leads at first use, keeps that pin, and a repoint
+// after that is refused like any other.
+func TestDcnGrantFromBeforePinsPinsAtFirstUse(t *testing.T) {
+	fx := newDcnStoreFixture(t)
+	fx.sync(t)
+	f, err := dcnLoadFile(fx.dir)
+	if err != nil {
+		t.Fatal(err)
+	}
+	rec := f.Stores["1"]
+	rec.RootReal = ""
+	f.Stores["1"] = rec
+	if err := dcnSaveFile(fx.dir, f); err != nil {
+		t.Fatal(err)
+	}
+	if got, why, exit := fx.read(t, "a.txt"); exit != 0 || got != "MINE" {
+		t.Fatalf("a grant from before pins no longer reads: exit %d %q %s", exit, got, why)
+	}
+	f, err = dcnLoadFile(fx.dir)
+	if err != nil {
+		t.Fatal(err)
+	}
+	want, _ := filepath.EvalSymlinks(fx.store)
+	if got := f.Stores["1"].RootReal; got != want {
+		t.Fatalf("first use pinned %q, want %q", got, want)
+	}
+	fx.repoint(t)
+	if got, _, exit := fx.read(t, "a.txt"); exit == 0 || got == "OTHER" {
+		t.Fatalf("after the first-use pin a repointed folder was read: exit %d %q", exit, got)
+	}
+}

client-go/dcn_store.go

@@ -0,0 +1,199 @@
+package main
+
+// dcn_store.go — every byte a machine-store op touches is reached through ONE
+// handle on the folder the person granted.
+//
+// WHY. The bastion is often root (a networked install, the macOS and Windows
+// system service), and the granted folder is often someone else's. Checking a
+// path and then opening it by name leaves two ways out of the store for
+// whoever can write in it or above it:
+//
+//   - the granted folder itself turned into a link to another folder after the
+//     grant. The old check resolved the root again on every op and judged the
+//     path against wherever it led NOW, so no race was needed;
+//   - a file or folder inside the store swapped for a link between the check
+//     and the open (or the rename, for a move plan).
+//
+// So: the folder's real path is PINNED at grant sync (rec.RootReal, this box's
+// own field, outside the signed bytes), an op whose root now leads elsewhere
+// is refused, and the op opens the pinned folder as an os.Root, which never
+// follows a link out of it, whatever happens to the names afterwards. That
+// handle is then checked to be the folder at the pinned path reached with no
+// link at all (dcnNoLinkStat: an open that refuses a link in every component
+// on Linux and macOS), so a swap of a folder above the store between the two
+// opens is refused too.
+
+import (
+	"crypto/rand"
+	"encoding/hex"
+	"errors"
+	"fmt"
+	"io/fs"
+	"os"
+	"path/filepath"
+	"strconv"
+)
+
+// errDcnRepointed is the refusal for a granted folder that leads elsewhere now.
+var errDcnRepointed = errors.New("the granted folder now leads somewhere other than the folder that was granted; " +
+	"a new grant for it is the repair")
+
+// dcnRealRoot is where the grant's root leads now: its links resolved.
+func dcnRealRoot(root string) (string, error) {
+	real, err := filepath.EvalSymlinks(filepath.Clean(root))
+	if err != nil {
+		return "", err
+	}
+	if !filepath.IsAbs(real) {
+		return "", errors.New("the granted folder is not an absolute path")
+	}
+	return filepath.Clean(real), nil
+}
+
+// dcnPinFor is the real root a sync records: the one already pinned for the
+// same root while the store is held (a re-sync, a renewed grant), else where
+// the root leads now ("" when it cannot be resolved yet; pinned at first use).
+func dcnPinFor(prev *dcnGrantRec, prevRoot, root string) string {
+	if prev != nil && prev.RootReal != "" && filepath.Clean(prevRoot) == filepath.Clean(root) {
+		return prev.RootReal
+	}
+	real, err := dcnRealRoot(root)
+	if err != nil {
+		return ""
+	}
+	return real
+}
+
+// dcnStorePin is the pinned real root of a held store, pinning it now (and
+// saving it) for a grant synced before pins existed.
+func dcnStorePin(dir string, storeID int64, root string) (string, error) {
+	grantsMu.Lock()
+	defer grantsMu.Unlock()
+	f, err := dcnLoadFile(dir)
+	if err != nil {
+		return "", err
+	}
+	key := strconv.FormatInt(storeID, 10)
+	rec, ok := f.Stores[key]
+	if !ok {
+		return "", errors.New("no grant for this store on this box")
+	}
+	if rec.RootReal != "" {
+		return rec.RootReal, nil
+	}
+	real, err := dcnRealRoot(root)
+	if err != nil {
+		return "", err
+	}
+	rec.RootReal = real
+	f.Stores[key] = rec
+	if err := dcnSaveFile(dir, f); err != nil {
+		return "", err
+	}
+	return real, nil
+}
+
+// dcnOpenStore opens the store a verified grant names, as the folder that was
+// granted, or refuses.
+func dcnOpenStore(dir string, storeID int64, terms *dcnTerms) (*os.Root, error) {
+	pin, err := dcnStorePin(dir, storeID, terms.Root)
+	if err != nil {
+		return nil, fmt.Errorf("the granted folder cannot be opened: %w", err)
+	}
+	real, err := dcnRealRoot(terms.Root)
+	if err != nil {
+		return nil, fmt.Errorf("the granted folder cannot be opened: %w", err)
+	}
+	if real != pin {
+		return nil, errDcnRepointed
+	}
+	return dcnOpenRootAt(pin)
+}
+
+// dcnOpenRootAt opens real (a resolved, absolute path) as an os.Root and
+// checks that the handle is the folder reached at real with no link in any
+// component, so nothing above the folder was swapped while it was opened.
+func dcnOpenRootAt(real string) (*os.Root, error) {
+	r, err := os.OpenRoot(real)
+	if err != nil {
+		return nil, fmt.Errorf("the granted folder cannot be opened: %w", err)
+	}
+	rfi, err := r.Stat(".")
+	if err != nil {
+		_ = r.Close()
+		return nil, fmt.Errorf("the granted folder cannot be opened: %w", err)
+	}
+	nfi, err := dcnNoLinkStat(real)
+	if err != nil || !nfi.IsDir() || !os.SameFile(rfi, nfi) {
+		_ = r.Close()
+		return nil, errDcnRepointed
+	}
+	return r, nil
+}
+
+// dcnRel is a path the sandbox accepted (cleanRel/dcnResolvePath's joined
+// form) as a name inside the root handle.
+func dcnRel(root, joined string) (string, error) {
+	rel, err := filepath.Rel(filepath.Clean(root), joined)
+	if err != nil || rel == ".." || filepath.IsAbs(rel) || len(rel) >= 3 && rel[:3] == ".."+string(os.PathSeparator) {
+		return "", errors.New("path escapes the store root")
+	}
+	return rel, nil
+}
+
+// dcnEscaped: the root handle refused a name because it leads out of the store.
+func dcnEscaped(err error) bool {
+	var pe *os.PathError
+	return errors.As(err, &pe) && pe.Err != nil && pe.Err.Error() == "path escapes from parent"
+}
+
+// dcnRootAtomicWrite is atomicWrite inside the store's handle: folders 0700,
+// the file 0600, written beside its name and renamed into place.
+func dcnRootAtomicWrite(r *os.Root, rel string, data []byte) error {
+	if d := filepath.Dir(rel); d != "." {
+		if err := r.MkdirAll(d, 0o700); err != nil {
+			return err
+		}
+	}
+	// The side name is unguessable and opened O_EXCL, as atomicWrite's
+	// CreateTemp is: whoever can write in the store could otherwise plant a
+	// link at a guessable name, and an open that follows it would truncate
+	// and fill whatever file in the store the link names. O_EXCL refuses any
+	// existing name, a link included, without following it.
+	var f *os.File
+	var tmp string
+	for try := 0; ; try++ {
+		var nb [12]byte
+		if _, err := rand.Read(nb[:]); err != nil {
+			return err
+		}
+		tmp = filepath.Join(filepath.Dir(rel), ".dcn-"+hex.EncodeToString(nb[:])+".tmp")
+		var err error
+		f, err = r.OpenFile(tmp, os.O_WRONLY|os.O_CREATE|os.O_EXCL, 0o600)
+		if err == nil {
+			break
+		}
+		if !errors.Is(err, fs.ErrExist) || try >= 8 {
+			return err
+		}
+	}
+	if _, err := f.Write(data); err != nil {
+		_ = f.Close()
+		_ = r.Remove(tmp)
+		return err
+	}
+	if err := f.Sync(); err != nil {
+		_ = f.Close()
+		_ = r.Remove(tmp)
+		return err
+	}
+	if err := f.Close(); err != nil {
+		_ = r.Remove(tmp)
+		return err
+	}
+	if err := r.Rename(tmp, rel); err != nil {
+		_ = r.Remove(tmp)
+		return err
+	}
+	return nil
+}

client-go/dcn_store_darwin.go

@@ -0,0 +1,21 @@
+//go:build darwin
+
+package main
+
+import (
+	"os"
+
+	"golang.org/x/sys/unix"
+)
+
+// dcnNoLinkStat describes the directory at p opened with O_NOFOLLOW_ANY, which
+// fails on a link in any component (macOS 11 and later).
+func dcnNoLinkStat(p string) (os.FileInfo, error) {
+	fd, err := unix.Open(p, unix.O_RDONLY|unix.O_DIRECTORY|unix.O_NOFOLLOW_ANY|unix.O_CLOEXEC, 0)
+	if err != nil {
+		return nil, err
+	}
+	f := os.NewFile(uintptr(fd), p)
+	defer f.Close()
+	return f.Stat()
+}

client-go/dcn_store_linux.go

@@ -0,0 +1,48 @@
+//go:build linux
+
+package main
+
+import (
+	"errors"
+	"os"
+	"strings"
+
+	"golang.org/x/sys/unix"
+)
+
+// dcnNoLinkStat describes the directory at p (absolute, clean) reached one
+// component at a time with O_NOFOLLOW, so a link anywhere in the path fails
+// instead of being followed. O_PATH: a component the bastion may traverse but
+// not list (a home at 0711) still opens.
+func dcnNoLinkStat(p string) (os.FileInfo, error) {
+	if !strings.HasPrefix(p, "/") {
+		return nil, errors.New("not an absolute path")
+	}
+	fd, err := unix.Open("/", unix.O_PATH|unix.O_DIRECTORY|unix.O_CLOEXEC, 0)
+	if err != nil {
+		return nil, err
+	}
+	for _, c := range strings.Split(p, "/") {
+		if c == "" {
+			continue
+		}
+		if c == "." || c == ".." {
+			_ = unix.Close(fd)
+			return nil, errors.New("not a clean path")
+		}
+		next, err := unix.Openat(fd, c, unix.O_PATH|unix.O_NOFOLLOW|unix.O_DIRECTORY|unix.O_CLOEXEC, 0)
+		_ = unix.Close(fd)
+		if err != nil {
+			return nil, err
+		}
+		var st unix.Stat_t
+		if err := unix.Fstat(next, &st); err != nil || st.Mode&unix.S_IFMT != unix.S_IFDIR {
+			_ = unix.Close(next)
+			return nil, errors.New("a component is not a directory")
+		}
+		fd = next
+	}
+	f := os.NewFile(uintptr(fd), p)
+	defer f.Close()
+	return f.Stat()
+}

client-go/dcn_store_other.go

@@ -0,0 +1,14 @@
+//go:build !linux && !darwin
+
+package main
+
+import "os"
+
+// dcnNoLinkStat on a platform with no open that refuses a link in every
+// component (Windows): the final component is judged without following it.
+// The store is still reached only through its os.Root and only while its real
+// path matches the pin; what this leaves is a swap of a folder above the
+// store, by someone who can rename it, timed between the two opens.
+func dcnNoLinkStat(p string) (os.FileInfo, error) {
+	return os.Lstat(p)
+}

client-go/device.go

@@ -296,13 +296,6 @@ func nonceSave(m map[string]string) error {
 	return atomicWrite(p, raw)
 }
 
-func nonceSeen(nonce string) bool {
-	deviceMu.Lock()
-	defer deviceMu.Unlock()
-	_, ok := nonceLoad()[nonce]
-	return ok
-}
-
 // burnNonce records the nonce until its approval's expiry (after which a
 // replay fails on expiry anyway).
 func burnNonce(a *deviceApproval) {
@@ -331,17 +324,6 @@ func approvalOK(dir, approval, purpose string) error {
 	return stepupCheckApproval(dir, approval, purpose)
 }
 
-// approvalKind names how an approval was given, for the audit line.
-func approvalKind(approval string) string {
-	switch {
-	case strings.TrimSpace(approval) == "":
-		return ""
-	case isDeviceApproval(approval):
-		return "device"
-	}
-	return "console"
-}
-
 // approvalRequired: does a change on this box need the person's approval?
 // Always on an anchored box (writes.go); on a local box only when the person
 // enrolled a factor here — the rule the DCN lanes have had since stepup.go.
@@ -1013,6 +995,11 @@ func cmdPinDevice(key string, unpin bool) error {
 			return err
 		}
 		fmt.Println("device key unpinned; approvals signed by it are refused from now on")
+		// Every op grant it approved dies with it (ops_bounds.go), dropped
+		// now so pinning the key again does not bring one back.
+		if n := opsDropApprovedBy("device:" + strings.TrimSpace(key)); n > 0 {
+			fmt.Printf("%d op grant(s) it approved dropped\n", n)
+		}
 		return nil
 	}
 	if err := pinDeviceKey(key); err != nil {

client-go/egress.go

@@ -3,8 +3,10 @@ package main
 import (
 	"bufio"
 	"context"
+	"errors"
 	"io"
 	"net"
+	"net/http"
 	"strings"
 	"time"
 )
@@ -23,94 +25,110 @@ type egressDialer func(ctx context.Context, network, addr string) (net.Conn, err
 // egress is refused (403, before any dial) unless its pairing grant carries
 // the "egress" cap: this lane makes requests leave THIS machine's connection
 // under its address, and nothing in the mgmt op table gates it.
-func handleEgressConn(client net.Conn, dialOut egressDialer, self string) {
+//
+// It reports whether it took the job: true once the request passed every check
+// and was dialled out, whatever the destination then answered; false when the
+// worker closed before a request, or the request was malformed or refused. The
+// worker pool (runEgressWorkers) parks again at once after a job it took, and
+// waits after one it did not.
+func handleEgressConn(client net.Conn, dialOut egressDialer, self string) bool {
 	defer client.Close()
-	br := bufio.NewReader(client)
+	// The request line and headers are read under a bound, the one Go's own HTTP
+	// server uses (http.DefaultMaxHeaderBytes, 1 MiB): the host sends them, and a
+	// head that never ends must not grow in this machine's memory.
+	headLim := &io.LimitedReader{R: client, N: egressMaxHeadBytes}
+	br := bufio.NewReader(headLim)
 	line, err := br.ReadString('\n')
 	if err != nil {
-		return
+		return false
 	}
 	parts := strings.Fields(strings.TrimSpace(line))
 	if len(parts) < 2 {
-		return
+		return false
 	}
 	method, target := parts[0], parts[1]
 
-	// Buffer the head so a plain-HTTP request can be replayed verbatim; capture Host.
+	// Buffer the head: a plain-HTTP request is parsed from it and rebuilt
+	// (parseEgressRequest), never replayed verbatim.
 	var head strings.Builder
 	head.WriteString(line)
-	hostHdr := ""
 	for {
 		h, err := br.ReadString('\n')
 		if err != nil {
-			return
+			return false
 		}
 		head.WriteString(h)
 		if h == "\r\n" || h == "\n" {
 			break
 		}
-		if len(h) >= 5 && strings.EqualFold(h[:5], "host:") {
-			hostHdr = strings.TrimSpace(h[5:])
-		}
 	}
 
+	// The head is in; what follows is the stream itself, unbounded. What the
+	// reader already buffered past the head stays first in line.
+	headLim.N = 1<<63 - 1
+
 	if why := capsAllow(self, "egress"); why != "" {
-		_, _ = client.Write([]byte("HTTP/1.1 403 Forbidden\r\nContent-Length: 0\r\nX-Privatae-Refused: " + why + "\r\n\r\n"))
-		return
+		writeEgressRefusal(client, "403 Forbidden", why, why)
+		return false
 	}
 
 	ctx, cancel := context.WithTimeout(context.Background(), egressDialTimeout)
 	defer cancel()
 
 	if strings.EqualFold(method, "CONNECT") {
-		dialAddr, why := egressVettedAddr(ctx, ensurePort(target, "443"))
-		if why != "" {
-			_, _ = client.Write([]byte("HTTP/1.1 403 Forbidden\r\nContent-Length: 0\r\nX-Privatae-Refused: " + why + "\r\n\r\n"))
-			return
+		// The target is read the way the Host's egress gate reads it
+		// (parseEgressAuthority): userinfo, an unbracketed IPv6 literal or a
+		// malformed port is refused, never guessed at.
+		host, port, ok := parseEgressAuthority(target, 443)
+		if !ok {
+			writeEgressRefusal(client, "400 Bad Request", egressNoHost, egressNoHost)
+			return false
 		}
-		upstream, err := dialOut(ctx, "tcp", dialAddr)
+		dialAddrs, why := egressVettedAddrs(ctx, egressHostPort(host, port))
+		if why != "" {
+			writeEgressRefusal(client, "403 Forbidden", why, why)
+			return false
+		}
+		upstream, err := dialEach(ctx, dialOut, dialAddrs)
 		if err != nil {
-			_, _ = client.Write([]byte("HTTP/1.1 502 Bad Gateway\r\nContent-Length: 0\r\n\r\n"))
-			return
+			writeEgressRefusal(client, "502 Bad Gateway", egressUnreachable, "")
+			return true
 		}
 		if _, err := client.Write([]byte("HTTP/1.1 200 Connection established\r\n\r\n")); err != nil {
 			_ = upstream.Close()
-			return
+			return true
 		}
 		splice(br, client, upstream)
-		return
+		return true
 	}
 
-	// Plain HTTP, absolute-form: "GET http://host/path HTTP/1.1".
-	host := hostHdr
-	if i := strings.Index(target, "://"); i >= 0 {
-		rest := target[i+3:]
-		if s := strings.IndexByte(rest, '/'); s >= 0 {
-			host = rest[:s]
-		} else {
-			host = rest
+	// Plain HTTP: ONE request per connection, parsed and rebuilt so the
+	// destination reads the host, port and body the lane vetted, sent with
+	// `Connection: close` (egress_http.go). Anything the Host sends after that
+	// request's body closes the connection unforwarded: a keep-alive connection
+	// can no longer carry later requests past the vetting.
+	req, refused := parseEgressRequest(head.String())
+	if refused != nil {
+		writeEgressRefusal(client, refused.status, refused.body, refused.body)
+		return false
 	}
-	}
-	if host == "" {
-		return
-	}
-	dialAddr, why := egressVettedAddr(ctx, ensurePort(host, "80"))
+	dialAddrs, why := egressVettedAddrs(ctx, egressHostPort(req.host, req.port))
 	if why != "" {
-		_, _ = client.Write([]byte("HTTP/1.1 403 Forbidden\r\nContent-Length: 0\r\nX-Privatae-Refused: " + why + "\r\n\r\n"))
-		return
+		writeEgressRefusal(client, "403 Forbidden", why, why)
+		return false
 	}
-	upstream, err := dialOut(ctx, "tcp", dialAddr)
+	upstream, err := dialEach(ctx, dialOut, dialAddrs)
 	if err != nil {
-		_, _ = client.Write([]byte("HTTP/1.1 502 Bad Gateway\r\nContent-Length: 0\r\n\r\n"))
-		return
+		writeEgressRefusal(client, "502 Bad Gateway", egressUnreachable, "")
+		return true
 	}
-	if _, err := upstream.Write([]byte(head.String())); err != nil {
-		_ = upstream.Close()
-		return
-	}
-	splice(br, client, upstream)
+	oneEgressExchange(br, client, upstream, req)
+	return true
 }
 
+// egressMaxHeadBytes bounds an egress request's line and headers together.
+const egressMaxHeadBytes = http.DefaultMaxHeaderBytes
+
 // ensurePort returns host:port, adding defPort when target has no explicit port.
 func ensurePort(target, defPort string) string {
 	target = strings.TrimSpace(target)
@@ -145,47 +163,68 @@ func splice(clientReader io.Reader, client, upstream net.Conn) {
 	_ = client.Close()
 }
 
-// egressInternalRefusal: on an ANCHORED box the egress lane reaches the public
-// internet and nothing else. Loopback, link-local and private addresses — this
-// machine's own services, the LAN it sits on, the mesh — are refused before
-// any dial, whether named by address or by a name that resolves to one. The
-// lane is USE of the machine's connection (writes.go), and that holds only
+// egressInternalRefusal is the refusal egressVettedAddrs gives, or "".
+//
+// The lane is USE of the machine's connection (writes.go), and that holds only
 // while it cannot be turned inward: `POST http://127.0.0.1:11434/api/pull`
-// down this lane would reach ollama around the forward's filter, and anything
-// on the LAN around every other gate. So is any address the machine takes as
-// its own (ownaddr.go): a public one on its NIC, or a range a `local` route
-// delivers here. A local box keeps the lane it had.
-// Resolution failures refuse too — a name this box cannot resolve is not one
+// down this lane would reach ollama around the forward's filter and around the
+// `models` switch, and this machine's other services around theirs. So on
+// every box the lane refuses this machine: `localhost` names, loopback and
+// unspecified addresses, link-local ones (the cloud metadata address
+// included), and any address the machine takes as its own (ownaddr.go), such
+// as a public one on its NIC or a range a `local` route delivers here.
+//
+// The network around it is refused too, on every box: the lane reaches the
+// public internet and nothing else, so private, CGNAT and reserved addresses
+// (the LAN, the mesh) are refused, by address and by any address a name
+// resolves to. A LOCAL box used to keep its LAN reachable here; it no longer
+// does (2026-10-26): the Host's egress gate holds every dial it makes to the
+// same rule, and a person's own machines
+// are reached through their approved lanes, never through egress. Only the
+// wording differs between the two kinds of box.
+//
+// Resolution failures refuse too: a name this box cannot resolve is not one
 // it can vouch for.
 //
-// THE ADDRESS THE CHECK VETTED IS THE ADDRESS DIALLED. egressVettedAddr
-// returns "ip:port" on an anchored box, and the dial takes that, never the
-// name: a name resolved once for the check and again for the dial is a
-// rebinding hole (a public answer for the check, 127.0.0.1 for the dial).
+// THE ADDRESS THE CHECK VETTED IS THE ADDRESS DIALLED. egressVettedAddrs
+// returns "ip:port" forms, and the dial takes those, never the name: a name
+// resolved once for the check and again for the dial is a rebinding hole (a
+// public answer for the check, 127.0.0.1 for the dial).
 func egressInternalRefusal(ctx context.Context, hostport string) string {
-	_, why := egressVettedAddr(ctx, hostport)
+	_, why := egressVettedAddrs(ctx, hostport)
 	return why
 }
 
-// egressResolve is how the anchored lane resolves a name. A seam so a test can
-// answer for a name without DNS.
+// egressResolve is how the lane resolves a name. A seam so a test can answer
+// for a name without DNS.
 var egressResolve = net.DefaultResolver.LookupIPAddr
 
-// egressVettedAddr: (the address to dial, "") or ("", the refusal). On a
-// local box the input is returned as it came (the OS resolves at the dial,
-// as before); on an anchored box every address the name resolves to is
-// checked and the FIRST one is what the dial gets, as "ip:port".
+// egressVettedAddr: (the first address to dial, "") or ("", the refusal).
 func egressVettedAddr(ctx context.Context, hostport string) (string, string) {
-	if currentAnchor() == nil {
-		return hostport, ""
+	all, why := egressVettedAddrs(ctx, hostport)
+	if why != "" {
+		return "", why
 	}
+	return all[0], ""
+}
+
+// egressVettedAddrs: every address the name resolved to, as "ip:port" in the
+// resolver's order, or (nil, the refusal). Any one refused address refuses the
+// name. The dial tries them in order (dialEach), so a dual-stack site still
+// answers from a machine whose IPv6 goes nowhere, as it did when the OS dialled
+// the name.
+func egressVettedAddrs(ctx context.Context, hostport string) ([]string, string) {
+	anchored := currentAnchor() != nil
 	host, port, err := net.SplitHostPort(hostport)
 	if err != nil {
 		host, port = hostport, ""
 	}
 	host = strings.Trim(strings.TrimSpace(host), "[]")
 	if strings.EqualFold(host, "localhost") || strings.HasSuffix(strings.ToLower(host), ".localhost") {
-		return "", "refused: egress to " + host + " — this lane reaches the public internet only on an anchored machine"
+		if anchored {
+			return nil, "refused: egress to " + host + " — this lane reaches the public internet only on an anchored machine"
+		}
+		return nil, "refused: egress to " + host + " — this lane never reaches this machine"
 	}
 	var addrs []net.IP
 	if ip := net.ParseIP(host); ip != nil {
@@ -193,50 +232,163 @@ func egressVettedAddr(ctx context.Context, hostport string) (string, string) {
 	} else {
 		resolved, err := egressResolve(ctx, host)
 		if err != nil || len(resolved) == 0 {
-			return "", "refused: egress to " + host + " — could not resolve it on this machine"
+			return nil, "refused: egress to " + host + " — could not resolve it on this machine"
 		}
 		for _, r := range resolved {
 			addrs = append(addrs, r.IP)
 		}
 	}
 	for _, ip := range addrs {
-		if internalIP(ip) {
-			return "", "refused: egress to " + host + " (" + ip.String() + ") — this lane reaches the public " +
+		if anchored && internalIP(ip) {
+			return nil, "refused: egress to " + host + " (" + ip.String() + ") — this lane reaches the public " +
 				"internet only on an anchored machine, never this machine, its network or the mesh"
 		}
+		if !anchored && thisMachineOrLinkIP(ip) {
+			return nil, "refused: egress to " + host + " (" + ip.String() + ") — this lane never reaches " +
+				"this machine or a link-local address"
 		}
-	// This machine is not only loopback: a public address on its NIC, or a
-	// range the kernel delivers locally through a `local` route, reaches its
-	// 0.0.0.0 services too (ownaddr.go). This lane has no allow list, so the
-	// refusal has no exception; a local box returned above and is unchanged.
+	}
+	// This machine is not only loopback: a public address on its NIC, a LAN
+	// address on it, or a range the kernel delivers locally through a `local`
+	// route, reaches its 0.0.0.0 services too, and a container's or VM's
+	// address on a network made here reaches a service on it (ownaddr.go). This
+	// lane has no allow list, so the refusal has no exception on either kind of
+	// box.
 	own, err := ownDestination(addrs)
 	if err != nil {
-		return "", "refused: egress to " + host + " — this machine could not tell whether that is itself, so it cannot be vetted"
+		return nil, "refused: egress to " + host + " — this machine could not tell whether that is itself, so it cannot be vetted"
 	}
 	if own {
-		return "", "refused: egress to " + host + " — that is this machine; this lane reaches the public " +
+		if anchored {
+			return nil, "refused: egress to " + host + " — that is this machine; this lane reaches the public " +
 				"internet only on an anchored machine, never this machine, its network or the mesh"
 		}
-	return net.JoinHostPort(addrs[0].String(), port), ""
+		return nil, "refused: egress to " + host + " — that is this machine; this lane never reaches this machine"
+	}
+	// The network around this machine, on a local box as on an anchored one
+	// (above): any one address of the name that is not the public internet
+	// refuses it. Checked after this machine's own addresses, so those keep
+	// saying "that is this machine".
+	if !anchored {
+		for _, ip := range addrs {
+			if internalIP(ip) {
+				return nil, "refused: egress to " + host + " (" + ip.String() + ") — this lane reaches the " +
+					"public internet only, never this machine's network or the mesh; reach your own " +
+					"machines through their approved lanes"
+			}
+		}
+	}
+	out := make([]string, 0, len(addrs))
+	for _, ip := range addrs {
+		out = append(out, net.JoinHostPort(ip.String(), port))
+	}
+	return out, ""
 }
 
-// internalIP: loopback, link-local, private (RFC 1918, ULA), CGNAT, unspecified.
-func internalIP(ip net.IP) bool {
-	if ip.IsLoopback() || ip.IsLinkLocalUnicast() || ip.IsLinkLocalMulticast() ||
-		ip.IsPrivate() || ip.IsUnspecified() || ip.IsMulticast() {
-		return true
+// dialEachStagger is how long an attempt has to answer before the next
+// address is tried beside it: the connection attempt delay of Happy Eyeballs
+// (RFC 8305 recommends 250 ms), which the OS dialler used when it was given
+// the name.
+var dialEachStagger = 300 * time.Millisecond
+
+// dialEach dials vetted addresses, in order, and returns the first that
+// answers. The next address starts as soon as an attempt fails, or when the
+// attempts in flight have not answered within dialEachStagger, so an address
+// whose packets go nowhere costs a fraction of a second, not the dial budget.
+// Only these addresses are ever dialled. The attempts still in flight when one
+// answers are cancelled, and any that connects anyway is closed.
+func dialEach(ctx context.Context, dialOut egressDialer, addrs []string) (net.Conn, error) {
+	if len(addrs) == 0 {
+		return nil, errors.New("no address")
+	}
+	ctx, cancel := context.WithCancel(ctx)
+	defer cancel()
+	type attempt struct {
+		c   net.Conn
+		err error
+	}
+	results := make(chan attempt, len(addrs))
+	next, inFlight := 0, 0
+	start := func() {
+		a := addrs[next]
+		next++
+		inFlight++
+		go func() {
+			c, err := dialOut(ctx, "tcp", a)
+			results <- attempt{c, err}
+		}()
+	}
+	stagger := time.NewTimer(dialEachStagger)
+	defer stagger.Stop()
+	start()
+	var err error
+	for inFlight > 0 {
+		var tick <-chan time.Time
+		if next < len(addrs) {
+			tick = stagger.C
+		}
+		select {
+		case r := <-results:
+			inFlight--
+			if r.err == nil {
+				go func(n int) {
+					for ; n > 0; n-- {
+						if late := <-results; late.c != nil {
+							_ = late.c.Close()
+						}
+					}
+				}(inFlight)
+				return r.c, nil
+			}
+			err = r.err
+			if next < len(addrs) {
+				start()
+				stagger.Reset(dialEachStagger)
+			}
+		case <-tick:
+			start()
+			stagger.Reset(dialEachStagger)
+		}
+	}
+	return nil, err
+}
+
+// ipClass is what an address reaches, read by every lane that dials for
+// someone else. The address rules live here once: internalIP is "anything but
+// public", and the local egress lane refuses the classes that are this machine
+// or its link.
+type ipClass int
+
+const (
+	ipPublic      ipClass = iota // the public internet
+	ipThisMachine                // loopback, unspecified, 0.0.0.0/8
+	ipLink                       // link-local unicast and multicast (the cloud metadata address)
+	ipNetwork                    // private (RFC 1918, ULA), CGNAT, multicast, benchmarking, reserved
+	ipUnreadable                 // a translated form whose IPv4 address this box cannot read
+	ipObsolete                   // IPv4-compatible ::/96, IPv4-translated ::ffff:0:0/96, site-local fec0::/10
+)
+
+func classifyIP(ip net.IP) ipClass {
+	switch {
+	case ip.IsLoopback(), ip.IsUnspecified():
+		return ipThisMachine
+	case ip.IsLinkLocalUnicast(), ip.IsLinkLocalMulticast():
+		return ipLink
+	case ip.IsPrivate(), ip.IsMulticast():
+		return ipNetwork
 	}
 	if v4 := ip.To4(); v4 != nil {
-		// 100.64.0.0/10 (carrier NAT) is as internal as 10/8 for this purpose.
-		if v4[0] == 100 && v4[1]&0xc0 == 64 {
-			return true
+		// 0.0.0.0/8 is "this host on this network".
+		if v4[0] == 0 {
+			return ipThisMachine
 		}
-		// 0.0.0.0/8 ("this network"), 198.18.0.0/15 (benchmarking) and
-		// 240.0.0.0/4 (reserved) are never a public host.
-		if v4[0] == 0 || (v4[0] == 198 && v4[1]&0xfe == 18) || v4[0] >= 240 {
-			return true
+		// 100.64.0.0/10 (carrier NAT) is as internal as 10/8 for this purpose;
+		// 198.18.0.0/15 (benchmarking) and 240.0.0.0/4 (reserved) are never a
+		// public host.
+		if (v4[0] == 100 && v4[1]&0xc0 == 64) || (v4[0] == 198 && v4[1]&0xfe == 18) || v4[0] >= 240 {
+			return ipNetwork
 		}
-		return false
+		return ipPublic
 	}
 	// IPv6 forms that carry an IPv4 address inside them reach that address, so
 	// they are judged by it: NAT64 64:ff9b::/96 (the last four bytes — how an
@@ -245,15 +397,51 @@ func internalIP(ip net.IP) bool {
 	if len(ip) == net.IPv6len {
 		if ip[0] == 0x00 && ip[1] == 0x64 && ip[2] == 0xff && ip[3] == 0x9b {
 			if ip[4]|ip[5]|ip[6]|ip[7]|ip[8]|ip[9]|ip[10]|ip[11] == 0 {
-				return internalIP(net.IPv4(ip[12], ip[13], ip[14], ip[15]))
+				return classifyIP(net.IPv4(ip[12], ip[13], ip[14], ip[15]))
 			}
 			// The rest of 64:ff9b::/32 (local-use translators, RFC 8215) maps
-			// to IPv4 in ways this box cannot read: refused.
-			return true
+			// to IPv4 in ways this box cannot read.
+			return ipUnreadable
 		}
 		if ip[0] == 0x20 && ip[1] == 0x02 {
-			return internalIP(net.IPv4(ip[2], ip[3], ip[4], ip[5]))
+			return classifyIP(net.IPv4(ip[2], ip[3], ip[4], ip[5]))
 		}
+		// Deprecated forms that are no destination of their own: IPv4-compatible
+		// ::a.b.c.d (RFC 4291 2.5.5.1), IPv4-translated ::ffff:0:a.b.c.d (SIIT)
+		// and site-local fec0::/10 (RFC 3879). Where each is delivered depends on
+		// the platform, so no lane dials them. (:: and ::1 are matched above.)
+		if isZeros(ip[0:12]) ||
+			(isZeros(ip[0:8]) && ip[8] == 0xff && ip[9] == 0xff && ip[10] == 0 && ip[11] == 0) ||
+			(ip[0] == 0xfe && ip[1]&0xc0 == 0xc0) {
+			return ipObsolete
+		}
+	}
+	return ipPublic
+}
+
+func isZeros(b []byte) bool {
+	for _, c := range b {
+		if c != 0 {
+			return false
+		}
+	}
+	return true
+}
+
+// internalIP: loopback, link-local, private (RFC 1918, ULA), CGNAT,
+// unspecified, multicast and reserved, also inside a NAT64 or 6to4 address,
+// and the deprecated IPv4-compatible, IPv4-translated and site-local forms.
+func internalIP(ip net.IP) bool {
+	return classifyIP(ip) != ipPublic
+}
+
+// thisMachineOrLinkIP: what the local egress lane refuses by address class.
+// An address it cannot read is refused, since it could be either, and so is a
+// deprecated form, which is no destination of its own.
+func thisMachineOrLinkIP(ip net.IP) bool {
+	switch classifyIP(ip) {
+	case ipThisMachine, ipLink, ipUnreadable, ipObsolete:
+		return true
 	}
 	return false
 }

client-go/egress_dial_test.go

@@ -0,0 +1,122 @@
+package main
+
+// dialEach, which both egress lanes dial through: an address whose packets go
+// nowhere must not hold up the next one for its share of the dial budget, the
+// way the OS dialler's Happy Eyeballs never did when it was given the name.
+
+import (
+	"context"
+	"errors"
+	"net"
+	"reflect"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+type closeRecordingConn struct {
+	net.Conn
+	closed atomic.Bool
+}
+
+func (c *closeRecordingConn) Close() error {
+	c.closed.Store(true)
+	return c.Conn.Close()
+}
+
+// MUTATION: dial the addresses one after another, each with an even share of
+// what is left of the budget.
+// RESULT: red — the second address waits out the first one's share (2 s here).
+func TestDialEachDoesNotWaitOutAnAddressThatGoesNowhere(t *testing.T) {
+	ctx, cancel := context.WithTimeout(context.Background(), 4*time.Second)
+	defer cancel()
+	var mu sync.Mutex
+	var dialled []string
+	abandoned := make(chan struct{})
+	dial := func(ctx context.Context, _, addr string) (net.Conn, error) {
+		mu.Lock()
+		dialled = append(dialled, addr)
+		mu.Unlock()
+		if addr == "[2001:db8::30]:443" {
+			<-ctx.Done() // no answer and no error: the packets go nowhere
+			close(abandoned)
+			return nil, ctx.Err()
+		}
+		a, b := net.Pipe()
+		go func() { _ = b.Close() }()
+		return a, nil
+	}
+	began := time.Now()
+	c, err := dialEach(ctx, dial, []string{"[2001:db8::30]:443", "198.51.100.30:443"})
+	took := time.Since(began)
+	if err != nil {
+		t.Fatalf("the second address answers: %v", err)
+	}
+	_ = c.Close()
+	if took > 1500*time.Millisecond {
+		t.Fatalf("the second address waited %v for the first", took)
+	}
+	select {
+	case <-abandoned:
+	case <-time.After(2 * time.Second):
+		t.Fatal("the attempt left in flight must be cancelled once another answers")
+	}
+	mu.Lock()
+	defer mu.Unlock()
+	if want := []string{"[2001:db8::30]:443", "198.51.100.30:443"}; !reflect.DeepEqual(dialled, want) {
+		t.Fatalf("only the vetted addresses, in order: got %v, want %v", dialled, want)
+	}
+}
+
+// A connection that answers after another already won is closed, not leaked;
+// an address that fails at once hands over to the next without waiting; and
+// when every address fails the last error is returned.
+func TestDialEachClosesALateWinnerAndReturnsTheLastError(t *testing.T) {
+	late := make(chan *closeRecordingConn, 1)
+	dial := func(ctx context.Context, _, addr string) (net.Conn, error) {
+		switch addr {
+		case "198.51.100.31:443":
+			time.Sleep(2 * dialEachStagger) // answers, but after the second address did
+			a, b := net.Pipe()
+			go func() { _ = b.Close() }()
+			rc := &closeRecordingConn{Conn: a}
+			late <- rc
+			return rc, nil
+		case "198.51.100.32:443":
+			a, b := net.Pipe()
+			go func() { _ = b.Close() }()
+			return a, nil
+		}
+		return nil, errors.New("refused: " + addr)
+	}
+	c, err := dialEach(context.Background(), dial, []string{"198.51.100.31:443", "198.51.100.32:443"})
+	if err != nil {
+		t.Fatal(err)
+	}
+	_ = c.Close()
+	select {
+	case rc := <-late:
+		deadline := time.Now().Add(2 * time.Second)
+		for !rc.closed.Load() && time.Now().Before(deadline) {
+			time.Sleep(10 * time.Millisecond)
+		}
+		if !rc.closed.Load() {
+			t.Fatal("a connection that answered after the winner must be closed")
+		}
+	case <-time.After(3 * time.Second):
+		t.Fatal("the slow attempt never answered")
+	}
+
+	began := time.Now()
+	_, err = dialEach(context.Background(), dial, []string{"198.51.100.33:443", "198.51.100.34:443"})
+	if err == nil || err.Error() != "refused: 198.51.100.34:443" {
+		t.Fatalf("every address failed, so the last error: %v", err)
+	}
+	if took := time.Since(began); took >= dialEachStagger {
+		t.Fatalf("a failed address hands over at once, not after the stagger: %v", took)
+	}
+	if _, err := dialEach(context.Background(), dial, nil); err == nil {
+		t.Fatal("no address is an error")
+	}
+}

client-go/egress_http.go

@@ -0,0 +1,420 @@
+package main
+
+import (
+	"bufio"
+	"errors"
+	"io"
+	"net"
+	"net/http/httputil"
+	"strconv"
+	"strings"
+)
+
+// ONE plain-HTTP request per egress connection, read the way the Host's egress
+// gate reads it (its strict request parse and its authority parse: the same
+// refusals, the same normalised host, the same rebuilt head).
+//
+// Everything the lane writes to a destination is what it checked, and nothing
+// more:
+//   - the head is parsed strictly and REBUILT, so the destination reads the same
+//     host, port and body length the lane vetted: never a second reading of the
+//     Host's bytes (a Host header naming another site, a query or userinfo glued
+//     to the host, a padded Transfer-Encoding);
+//   - the request goes out with `Connection: close`, its body is framed by
+//     Content-Length or chunked encoding, and a byte the Host sends after the
+//     body closes the connection with nothing of it forwarded. Before this the
+//     lane wrote the first head and then spliced the whole connection, so every
+//     later request on a keep-alive connection reached the destination unread.
+//
+// An honest client sends its next request only after the response, which the
+// destination closes behind `Connection: close`, so it opens a new connection
+// and meets the vetting again.
+
+// egressHostChars are the characters a host name may have here (lower-cased).
+const egressHostChars = "abcdefghijklmnopqrstuvwxyz0123456789-._"
+
+// egressTokenChars are RFC 9110 token characters: methods and header names.
+const egressTokenChars = "!#$%&'*+-.^_`|~0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz"
+
+// egressHopByHop are headers that belong to one hop. The lane speaks for itself
+// on the next one, and a proxy credential never goes on to the destination.
+var egressHopByHop = map[string]bool{
+	"connection": true, "proxy-connection": true, "keep-alive": true, "upgrade": true,
+	"te": true, "trailer": true, "proxy-authorization": true, "proxy-authenticate": true,
+}
+
+// egressTrailerLimit bounds the trailer section of a chunked body, which is
+// read to its end and dropped (the gate's limit).
+const egressTrailerLimit = 8192
+
+// egressRefusal is a request the lane refuses before it dials anything.
+type egressRefusal struct{ status, body string }
+
+func egressBad(body string) *egressRefusal {
+	return &egressRefusal{"400 Bad Request", body}
+}
+
+const egressNoHost = "The bastion could not read where this request is going (no host in the " +
+	"request line or in a Host header), so it sent nothing."
+
+// egressUnreachable is the body of a 502: the destination passed the vetting,
+// and no address of it answered from this machine.
+const egressUnreachable = "The bastion could not reach that destination from this machine's own " +
+	"connection, so nothing was sent."
+
+// egressHostPort is host and port as the vetting takes them.
+func egressHostPort(host string, port int) string {
+	return net.JoinHostPort(host, strconv.Itoa(port))
+}
+
+// writeEgressRefusal answers the Host with a status and a short plain-text
+// body saying why, and closes the exchange. why, when set, also goes in
+// X-Privatae-Refused, the header the Host already reads.
+func writeEgressRefusal(w io.Writer, status, body, why string) {
+	body += "\n"
+	var b strings.Builder
+	b.WriteString("HTTP/1.1 " + status + "\r\n")
+	b.WriteString("Content-Type: text/plain; charset=utf-8\r\n")
+	b.WriteString("Content-Length: " + strconv.Itoa(len(body)) + "\r\n")
+	b.WriteString("Connection: close\r\n")
+	if why != "" {
+		b.WriteString("X-Privatae-Refused: " + strings.NewReplacer("\r", " ", "\n", " ").Replace(why) + "\r\n")
+	}
+	b.WriteString("\r\n")
+	b.WriteString(body)
+	_, _ = io.WriteString(w, b.String())
+}
+
+// parseEgressAuthority: (host, port, true) from "host", "host:port", "[v6]" or
+// "[v6]:port", or false when the lane can't read it exactly as the destination
+// would. Refused rather than guessed: userinfo ("a@b"), an unbracketed IPv6
+// literal, characters no host name has, a non-numeric or out-of-range port.
+// The host is lower-cased and loses one trailing dot.
+func parseEgressAuthority(s string, defaultPort int) (string, int, bool) {
+	s = strings.TrimSpace(s)
+	if s == "" || strings.Contains(s, "@") {
+		return "", 0, false
+	}
+	var host string
+	port := defaultPort
+	digits := func(p string) bool {
+		if len(p) == 0 || len(p) > 5 {
+			return false
+		}
+		for _, c := range p {
+			if c < '0' || c > '9' {
+				return false
+			}
+		}
+		return true
+	}
+	if strings.HasPrefix(s, "[") {
+		end := strings.IndexByte(s, ']')
+		if end < 0 {
+			return "", 0, false
+		}
+		// An IPv6 literal, IPv4-mapped forms included (the vetting judges the
+		// address they map to); a bare IPv4 address in brackets or a zone is
+		// not one.
+		inner := s[1:end]
+		ip := net.ParseIP(inner)
+		if ip == nil || !strings.Contains(inner, ":") || strings.Contains(inner, "%") {
+			return "", 0, false
+		}
+		host = ip.String()
+		rest := s[end+1:]
+		switch {
+		case rest == "":
+		case strings.HasPrefix(rest, ":") && digits(rest[1:]):
+			port, _ = strconv.Atoi(rest[1:])
+		default:
+			return "", 0, false
+		}
+	} else {
+		if strings.Count(s, ":") > 1 {
+			return "", 0, false
+		}
+		h, p, hasPort := strings.Cut(s, ":")
+		if hasPort {
+			if !digits(p) {
+				return "", 0, false
+			}
+			port, _ = strconv.Atoi(p)
+		}
+		h = strings.ToLower(h)
+		h = strings.TrimSuffix(h, ".")
+		if h == "" || strings.HasPrefix(h, ".") || strings.Contains(h, "..") {
+			return "", 0, false
+		}
+		for _, c := range h {
+			if !strings.ContainsRune(egressHostChars, c) {
+				return "", 0, false
+			}
+		}
+		host = h
+	}
+	if port <= 0 || port >= 65536 {
+		return "", 0, false
+	}
+	return host, port, true
+}
+
+// fmtEgressAuthority writes host[:port] back, the port left out when it is the
+// scheme's default.
+func fmtEgressAuthority(host string, port, defaultPort int) string {
+	h := host
+	if strings.Contains(host, ":") {
+		h = "[" + host + "]"
+	}
+	if port == defaultPort {
+		return h
+	}
+	return h + ":" + strconv.Itoa(port)
+}
+
+// egressRequest is the one plain-HTTP request the lane forwards.
+type egressRequest struct {
+	host    string
+	port    int
+	head    []byte // rebuilt, CRLF, ends at its blank line
+	chunked bool
+	length  int64
+}
+
+func egressTokenOK(s string) bool {
+	if s == "" {
+		return false
+	}
+	for _, c := range s {
+		if !strings.ContainsRune(egressTokenChars, c) {
+			return false
+		}
+	}
+	return true
+}
+
+// parseEgressRequest reads the one request in head (the request line and
+// headers through their blank line, as handleEgressConn read them) and
+// rebuilds it, or refuses it. CONNECT never comes here.
+func parseEgressRequest(head string) (egressRequest, *egressRefusal) {
+	lines := strings.Split(head, "\n")
+	// The read stops at the first blank line, so the head ends "\n", blank, "\n":
+	// drop the empty tail after the last "\n" and the blank line itself.
+	if len(lines) < 3 {
+		return egressRequest{}, egressBad("The request ended before its headers did, so the bastion sent nothing.")
+	}
+	lines = lines[:len(lines)-2]
+	for i, ln := range lines {
+		lines[i] = strings.TrimSuffix(ln, "\r")
+		if strings.ContainsAny(lines[i], "\r\x00") {
+			return egressRequest{}, egressBad("The request's headers carry a stray carriage return or NUL, so the bastion sent nothing.")
+		}
+	}
+	parts := strings.Split(lines[0], " ")
+	if len(parts) != 3 || parts[0] == "" || parts[1] == "" || parts[2] == "" || !egressTokenOK(parts[0]) {
+		return egressRequest{}, egressBad("The request line is not 'METHOD target HTTP/1.x', so the bastion sent nothing.")
+	}
+	method, rawTarget, version := strings.ToUpper(parts[0]), parts[1], parts[2]
+	if version != "HTTP/1.1" && version != "HTTP/1.0" {
+		return egressRequest{}, egressBad("The bastion forwards HTTP/1.0 and HTTP/1.1 only, so it sent nothing.")
+	}
+	type header struct{ name, value string }
+	var headers []header
+	for _, ln := range lines[1:] {
+		if strings.HasPrefix(ln, " ") || strings.HasPrefix(ln, "\t") {
+			return egressRequest{}, egressBad("A header is folded onto a second line (obsolete line folding), so the bastion sent nothing.")
+		}
+		name, value, ok := strings.Cut(ln, ":")
+		if !ok || !egressTokenOK(name) {
+			return egressRequest{}, egressBad("A header line is not 'Name: value' (no space may come before the colon), so the bastion sent nothing.")
+		}
+		headers = append(headers, header{name, strings.Trim(value, " \t")})
+	}
+	values := func(n string) []string {
+		var out []string
+		for _, h := range headers {
+			if strings.EqualFold(h.name, n) {
+				out = append(out, h.value)
+			}
+		}
+		return out
+	}
+
+	hosts := values("host")
+	if len(hosts) > 1 {
+		return egressRequest{}, egressBad("The request carries more than one Host header, so the bastion sent nothing.")
+	}
+	var host, target, authority string
+	var port int
+	if scheme, after, abs := strings.Cut(rawTarget, "://"); abs {
+		if strings.ToLower(scheme) != "http" {
+			// https:// in absolute form asks the lane to make the TLS connection,
+			// which it does not do: it used to send the request in plain text to
+			// port 80. Every client sends https as a CONNECT.
+			return egressRequest{}, egressBad("The bastion forwards plain http:// requests only; send https as a CONNECT. Nothing was sent.")
+		}
+		cut := len(after)
+		for _, sep := range []string{"/", "?", "#"} {
+			if i := strings.Index(after, sep); i >= 0 && i < cut {
+				cut = i
+			}
+		}
+		var ok bool
+		if host, port, ok = parseEgressAuthority(after[:cut], 80); !ok {
+			return egressRequest{}, egressBad(egressNoHost)
+		}
+		path, _, _ := strings.Cut(after[cut:], "#")
+		if !strings.HasPrefix(path, "/") {
+			path = "/" + path
+		}
+		authority = fmtEgressAuthority(host, port, 80)
+		target = "http://" + authority + path
+	} else {
+		if !strings.HasPrefix(rawTarget, "/") && rawTarget != "*" {
+			return egressRequest{}, egressBad(egressNoHost)
+		}
+		hostHdr := ""
+		if len(hosts) == 1 {
+			hostHdr = hosts[0]
+		}
+		var ok bool
+		if host, port, ok = parseEgressAuthority(hostHdr, 80); !ok {
+			return egressRequest{}, egressBad(egressNoHost)
+		}
+		authority = fmtEgressAuthority(host, port, 80)
+		target = rawTarget
+	}
+
+	// Body framing (RFC 9112 section 6). Every reading two parsers could disagree
+	// on is refused: both length headers, an unknown or doubled coding, a coding
+	// on HTTP/1.0, Content-Length values that differ.
+	req := egressRequest{host: host, port: port}
+	te, cl := values("transfer-encoding"), values("content-length")
+	if len(te) > 0 {
+		var codings []string
+		for _, v := range te {
+			for _, c := range strings.Split(v, ",") {
+				codings = append(codings, strings.ToLower(strings.TrimSpace(c)))
+			}
+		}
+		if len(cl) > 0 || version == "HTTP/1.0" || len(codings) != 1 || codings[0] != "chunked" {
+			return egressRequest{}, egressBad("The request's body length is ambiguous (Transfer-Encoding with " +
+				"Content-Length, on HTTP/1.0, or a coding other than chunked), so the bastion sent nothing.")
+		}
+		req.chunked = true
+	} else if len(cl) > 0 {
+		seen := map[string]bool{}
+		for _, v := range cl {
+			for _, c := range strings.Split(v, ",") {
+				seen[strings.TrimSpace(c)] = true
+			}
+		}
+		var one string
+		for k := range seen {
+			one = k
+		}
+		n, err := strconv.ParseInt(one, 10, 64)
+		if len(seen) != 1 || one == "" || len(one) > 18 || err != nil || n < 0 || strings.TrimLeft(one, "0123456789") != "" {
+			return egressRequest{}, egressBad("The request's Content-Length is not one number, so the bastion sent nothing.")
+		}
+		req.length = n
+	}
+
+	named := map[string]bool{}
+	for _, v := range values("connection") {
+		for _, t := range strings.Split(v, ",") {
+			named[strings.ToLower(strings.TrimSpace(t))] = true
+		}
+	}
+	var b strings.Builder
+	b.WriteString(method + " " + target + " " + version + "\r\n")
+	b.WriteString("Host: " + authority + "\r\n")
+	for _, h := range headers {
+		lk := strings.ToLower(h.name)
+		if egressHopByHop[lk] || named[lk] || lk == "host" || lk == "content-length" || lk == "transfer-encoding" {
+			continue
+		}
+		b.WriteString(h.name + ": " + h.value + "\r\n")
+	}
+	if req.chunked {
+		b.WriteString("Transfer-Encoding: chunked\r\n")
+	} else if len(cl) > 0 {
+		b.WriteString("Content-Length: " + strconv.FormatInt(req.length, 10) + "\r\n")
+	}
+	b.WriteString("Connection: close\r\n\r\n")
+	req.head = []byte(b.String())
+	return req, nil
+}
+
+// forwardEgressBody copies the one request body from src to dst, framed as the
+// rebuilt head says. A chunked body goes out re-encoded (plain sizes, no
+// extensions, trailers read and dropped), so the destination parses exactly the
+// body this parser found.
+func forwardEgressBody(dst io.Writer, src *bufio.Reader, req egressRequest) error {
+	if !req.chunked {
+		if req.length == 0 {
+			return nil
+		}
+		_, err := io.CopyN(dst, src, req.length)
+		return err
+	}
+	cw := httputil.NewChunkedWriter(dst)
+	if _, err := io.Copy(cw, httputil.NewChunkedReader(src)); err != nil {
+		return err
+	}
+	// The trailer section, read to its blank line within a bound, and dropped.
+	read := 0
+	for {
+		line, err := src.ReadSlice('\n')
+		read += len(line)
+		if err != nil || read > egressTrailerLimit {
+			return errors.New("chunked body: bad trailer section")
+		}
+		if string(line) == "\r\n" || string(line) == "\n" {
+			break
+		}
+	}
+	if err := cw.Close(); err != nil {
+		return err
+	}
+	_, err := io.WriteString(dst, "\r\n")
+	return err
+}
+
+// oneEgressExchange writes the one rebuilt request and its body to upstream,
+// and carries the destination's answer back until the destination closes.
+// Nothing the Host sends after the body is forwarded: its first byte closes
+// both sides.
+func oneEgressExchange(clientReader *bufio.Reader, client, upstream net.Conn, req egressRequest) {
+	answered := make(chan struct{})
+	go func() {
+		_, _ = io.Copy(client, upstream)
+		close(answered)
+		_ = client.Close()
+	}()
+	defer func() {
+		_ = upstream.Close()
+		<-answered
+	}()
+	if _, err := upstream.Write(req.head); err != nil {
+		return
+	}
+	if err := forwardEgressBody(upstream, clientReader, req); err != nil {
+		return
+	}
+	// The request is complete. A further byte is a second request on this
+	// connection: never forwarded, and the connection closes. End of stream
+	// from the Host is passed on, as the splice did.
+	go func() {
+		var one [1]byte
+		if n, _ := clientReader.Read(one[:]); n > 0 {
+			_ = upstream.Close()
+			_ = client.Close()
+			return
+		}
+		if cw, ok := upstream.(interface{ CloseWrite() error }); ok {
+			_ = cw.CloseWrite()
+		}
+	}()
+	<-answered
+}

client-go/egress_local_test.go

@@ -0,0 +1,229 @@
+//go:build !windows
+
+// Tagged !windows: it borrows setupBox (anchor_test.go) and stubOwnAddrs
+// (ownaddr_test.go).
+
+package main
+
+// The egress lane on a local (unanchored) box. It refuses this machine and
+// its link-local addresses, so a host cannot reach this machine's services
+// (ollama above all) around the switches that govern them, and, since
+// 2026-10-26, the network around it as well: the LAN and the mesh
+// (egress_network_test.go). An anchored box is unchanged: the public internet
+// only.
+
+import (
+	"bufio"
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"net"
+	"reflect"
+	"strings"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+// answerEgressNames answers the egress lane's name lookups from m.
+func answerEgressNames(t *testing.T, m map[string][]string) {
+	t.Helper()
+	old := egressResolve
+	t.Cleanup(func() { egressResolve = old })
+	egressResolve = func(_ context.Context, host string) ([]net.IPAddr, error) {
+		ips, ok := m[strings.ToLower(host)]
+		if !ok {
+			return nil, errors.New("no such host")
+		}
+		var out []net.IPAddr
+		for _, s := range ips {
+			out = append(out, net.IPAddr{IP: net.ParseIP(s)})
+		}
+		return out, nil
+	}
+}
+
+// A machine with a LAN address of its own and an AnyIP range, and names that
+// answer with this machine, with the LAN and with the public internet.
+func localEgressFixture(t *testing.T) {
+	t.Helper()
+	t.Setenv("XDG_CONFIG_HOME", t.TempDir())
+	if currentAnchor() != nil || boxMode() != modeLocal {
+		t.Fatal("the fixture must be a local, unanchored box")
+	}
+	stubOwnAddrs(t, []string{"127.0.0.1/8", "::1/128", "192.168.50.7/24", "fe80::7/64"}, localRange("203.0.113.0/24"))
+	answerEgressNames(t, map[string][]string{
+		"ollama.example.test": {"127.0.0.1"},
+		"mixed.example.test":  {"198.51.100.8", "127.0.0.1"},
+		"self.example.test":   {"192.168.50.7"},
+		"nas.example.test":    {"192.168.50.30"},
+		"dual.example.test":   {"2001:db8::8", "198.51.100.8"},
+		"site.example.test":   {"198.51.100.9"},
+	})
+}
+
+// MUTATION: restore the early return for an unanchored box in egressVettedAddrs.
+// RESULT: red — 127.0.0.1:11434 is handed to the dial, and a name is dialled
+// as a name instead of the address that was vetted.
+func TestLocalEgressRefusesThisMachineAndTheLAN(t *testing.T) {
+	localEgressFixture(t)
+	ctx := context.Background()
+	for _, hp := range []string{
+		"127.0.0.1:11434", "127.9.9.9:80", "localhost:11434", "ollama.localhost:80", "[::1]:11434",
+		"0.0.0.0:11434", "0.1.2.3:80", "[::]:80",
+		"[::ffff:127.0.0.1]:11434", "[64:ff9b::7f00:1]:11434", "[2002:7f00:1::]:80", "[64:ff9b:1::5]:80",
+		"169.254.169.254:80", "[fe80::1]:80",
+		"192.168.50.7:11434", // this machine's own LAN address
+		"203.0.113.5:80",     // a range a `local` route delivers here
+		"ollama.example.test:11434", "mixed.example.test:443", "self.example.test:11434",
+		"unknown.example.test:443",
+		// the network around this machine (egress_network_test.go has the rest)
+		"192.168.50.20:8080", "10.1.2.3:443", "172.16.4.4:22", "[fd12:3456::9]:80", "100.64.0.9:80",
+		"nas.example.test:443",
+	} {
+		if addr, why := egressVettedAddr(ctx, hp); why == "" || addr != "" {
+			t.Errorf("%s must be refused on a local box, got dial %q", hp, addr)
+		}
+	}
+	for hp, want := range map[string][]string{
+		"198.51.100.8:443":      {"198.51.100.8:443"},
+		"site.example.test:443": {"198.51.100.9:443"},
+		"dual.example.test:443": {"[2001:db8::8]:443", "198.51.100.8:443"},
+	} {
+		got, why := egressVettedAddrs(ctx, hp)
+		if why != "" || !reflect.DeepEqual(got, want) {
+			t.Errorf("%s is the internet, so a local box dials %v: got %v %q", hp, want, got, why)
+		}
+	}
+}
+
+// The anchored lane keeps its own wording and still refuses the LAN that a
+// local box reaches.
+func TestAnchoredEgressStillRefusesTheLAN(t *testing.T) {
+	setupBox(t)
+	stubOwnAddrs(t, []string{"127.0.0.1/8", "::1/128", "192.168.50.7/24"}, localRange("203.0.113.0/24"))
+	ctx := context.Background()
+	for _, hp := range []string{"192.168.50.20:8080", "10.1.2.3:443", "[fd12:3456::9]:80", "100.64.0.9:80", "127.0.0.1:11434"} {
+		if addr, why := egressVettedAddr(ctx, hp); addr != "" || !strings.Contains(why, "only on an anchored machine") {
+			t.Errorf("%s must be refused on an anchored box as before, got %q %q", hp, addr, why)
+		}
+	}
+	if addr, why := egressVettedAddr(ctx, "198.51.100.8:443"); why != "" || addr != "198.51.100.8:443" {
+		t.Fatalf("the public internet passes: %q %q", addr, why)
+	}
+}
+
+// Through the lane itself, with a real dialer: a request for this machine is
+// answered 403 and never reaches the listener; a public name is dialled at
+// its vetted address; a dual-stack name falls back to its second address.
+func TestLocalEgressLaneNeverDialsThisMachine(t *testing.T) {
+	localEgressFixture(t)
+	ln, err := net.Listen("tcp", "127.0.0.1:0")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer ln.Close()
+	var hits atomic.Int32
+	go func() {
+		for {
+			c, err := ln.Accept()
+			if err != nil {
+				return
+			}
+			hits.Add(1)
+			c.Close()
+		}
+	}()
+	ask := func(dial egressDialer, req string) string {
+		srv, cli := net.Pipe()
+		go handleEgressConn(srv, dial, "home.host.example")
+		_ = cli.SetDeadline(time.Now().Add(5 * time.Second))
+		_, _ = io.WriteString(cli, req)
+		line, _ := bufio.NewReader(cli).ReadString('\n')
+		cli.Close()
+		return line
+	}
+	netDial := (&net.Dialer{Timeout: 2 * time.Second}).DialContext
+	target := ln.Addr().String()
+	for _, req := range []string{
+		"CONNECT " + target + " HTTP/1.1\r\n\r\n",
+		"GET http://" + target + "/api/tags HTTP/1.1\r\nHost: " + target + "\r\n\r\n",
+		"CONNECT localhost:" + strings.Split(target, ":")[1] + " HTTP/1.1\r\n\r\n",
+	} {
+		if line := ask(netDial, req); !strings.HasPrefix(line, "HTTP/1.1 403") {
+			t.Errorf("a local box must refuse %q, answered %q", req, line)
+		}
+	}
+	time.Sleep(50 * time.Millisecond)
+	if n := hits.Load(); n != 0 {
+		t.Fatalf("the lane reached this machine's listener %d times", n)
+	}
+
+	var asked []string
+	recording := func(_ context.Context, _, addr string) (net.Conn, error) {
+		asked = append(asked, addr)
+		return nil, errors.New("stop here")
+	}
+	if line := ask(recording, "CONNECT site.example.test:443 HTTP/1.1\r\n\r\n"); !strings.HasPrefix(line, "HTTP/1.1 502") {
+		t.Fatalf("a public destination is dialled (and this dial fails): %q", line)
+	}
+	if !reflect.DeepEqual(asked, []string{"198.51.100.9:443"}) {
+		t.Fatalf("the vetted address, never the name: %v", asked)
+	}
+
+	asked = nil
+	fallback := func(_ context.Context, _, addr string) (net.Conn, error) {
+		asked = append(asked, addr)
+		if strings.HasPrefix(addr, "[") {
+			return nil, errors.New("network is unreachable")
+		}
+		a, b := net.Pipe()
+		go func() { _, _ = io.Copy(b, b) }()
+		return a, nil
+	}
+	if line := ask(fallback, "CONNECT dual.example.test:443 HTTP/1.1\r\n\r\n"); !strings.HasPrefix(line, "HTTP/1.1 200") {
+		t.Fatalf("a dual-stack name answers from its second address: %q", line)
+	}
+	if want := []string{"[2001:db8::8]:443", "198.51.100.8:443"}; !reflect.DeepEqual(asked, want) {
+		t.Fatalf("each vetted address in turn: got %v, want %v", asked, want)
+	}
+}
+
+// What docs/models.md tells a person with a local box to do: switch `models`
+// off for the host. The forward then filters store-changing requests, and the
+// egress lane is no way around it, since it never reaches this machine's
+// ollama, on loopback or on the machine's own LAN address.
+func TestLocalModelsSwitchLeavesNoEgressRouteToOllama(t *testing.T) {
+	localEgressFixture(t)
+	const host = "home.host.example"
+	writePairingState(t, host, pairingActive)
+	if inferenceFilterNeeded(host, laneOllama, pairingModelScope(host)) {
+		t.Fatal("an unscoped local pairing is a raw splice before the switch, so this test would prove nothing")
+	}
+	if err := localDenySet(host, "models", true); err != nil {
+		t.Fatal(err)
+	}
+	if !inferenceFilterNeeded(host, laneOllama, pairingModelScope(host)) {
+		t.Fatal("with `models` off the ollama forward must filter /api/pull")
+	}
+	var dialled []string
+	dial := func(_ context.Context, _, addr string) (net.Conn, error) {
+		dialled = append(dialled, addr)
+		return nil, errors.New("stop here")
+	}
+	for _, dest := range []string{"127.0.0.1:11434", "localhost:11434", "192.168.50.7:11434", "[::1]:11434"} {
+		srv, cli := net.Pipe()
+		go handleEgressConn(srv, dial, host)
+		_ = cli.SetDeadline(time.Now().Add(5 * time.Second))
+		fmt.Fprintf(cli, "POST http://%s/api/pull HTTP/1.1\r\nHost: %s\r\n\r\n", dest, dest)
+		line, _ := bufio.NewReader(cli).ReadString('\n')
+		cli.Close()
+		if !strings.HasPrefix(line, "HTTP/1.1 403") {
+			t.Errorf("egress to %s must be refused: %q", dest, line)
+		}
+	}
+	if len(dialled) != 0 {
+		t.Fatalf("nothing for this machine may be dialled: %v", dialled)
+	}
+}

client-go/egress_network_test.go

@@ -0,0 +1,76 @@
+//go:build !windows
+
+// Tagged !windows: it borrows localEgressFixture and answerEgressNames
+// (egress_local_test.go).
+
+package main
+
+// The egress lane on a local (unanchored) box refuses the network around the
+// machine as well as the machine itself: the LAN, the mesh and every other
+// address that is not the public internet, by address and by any address a
+// name resolves to. The Host's egress gate holds every dial to the same rule;
+// a person's own machines are reached through their approved lanes.
+
+import (
+	"bufio"
+	"context"
+	"errors"
+	"io"
+	"net"
+	"reflect"
+	"strings"
+	"testing"
+	"time"
+)
+
+// MUTATION: drop the local-box internalIP loop at the end of egressVettedAddrs.
+// RESULT: red — 192.168.50.20, nas.example.test and the rest are handed to the
+// dial, and the lane dials 192.168.50.30.
+func TestLocalEgressRefusesTheLANAndTheMesh(t *testing.T) {
+	localEgressFixture(t)
+	answerEgressNames(t, map[string][]string{
+		"nas.example.test":  {"192.168.50.30"},
+		"half.example.test": {"198.51.100.8", "10.1.2.3"},
+		"mesh.example.test": {"fd12:3456::9"},
+		"nat.example.test":  {"64:ff9b::a01:203"}, // NAT64 for 10.1.2.3
+		"pub.example.test":  {"198.51.100.8"},
+	})
+	ctx := context.Background()
+	for _, hp := range []string{
+		"192.168.50.20:8080", "10.1.2.3:443", "172.16.4.4:22", "[fd12:3456::9]:80", "100.64.0.9:80",
+		"198.18.0.1:80", "[64:ff9b::a01:203]:443",
+		"nas.example.test:443", "half.example.test:443", "mesh.example.test:80", "nat.example.test:443",
+	} {
+		if addrs, why := egressVettedAddrs(ctx, hp); why == "" || addrs != nil || !strings.Contains(why, "approved lanes") {
+			t.Errorf("%s is not the public internet and must be refused on a local box, got dial %v %q", hp, addrs, why)
+		}
+	}
+	if got, why := egressVettedAddrs(ctx, "pub.example.test:443"); why != "" || !reflect.DeepEqual(got, []string{"198.51.100.8:443"}) {
+		t.Fatalf("the public internet passes at its vetted address: %v %q", got, why)
+	}
+
+	// Through the lane itself: 403, and nothing is dialled.
+	var asked []string
+	recording := func(_ context.Context, _, addr string) (net.Conn, error) {
+		asked = append(asked, addr)
+		return nil, errors.New("stop here")
+	}
+	for _, req := range []string{
+		"CONNECT nas.example.test:443 HTTP/1.1\r\n\r\n",
+		"GET http://10.1.2.3/ HTTP/1.1\r\nHost: 10.1.2.3\r\n\r\n",
+		"CONNECT half.example.test:443 HTTP/1.1\r\n\r\n",
+	} {
+		srv, cli := net.Pipe()
+		go handleEgressConn(srv, recording, "home.host.example")
+		_ = cli.SetDeadline(time.Now().Add(5 * time.Second))
+		_, _ = io.WriteString(cli, req)
+		line, _ := bufio.NewReader(cli).ReadString('\n')
+		cli.Close()
+		if !strings.HasPrefix(line, "HTTP/1.1 403") {
+			t.Errorf("a local box must refuse %q, answered %q", req, line)
+		}
+	}
+	if len(asked) != 0 {
+		t.Fatalf("nothing on the network around this machine is dialled: %v", asked)
+	}
+}

client-go/egress_obsolete_test.go

@@ -0,0 +1,80 @@
+//go:build !windows
+
+// Tagged !windows: it borrows setupBox (anchor_test.go) and stubOwnAddrs
+// (ownaddr_test.go).
+
+package main
+
+// The deprecated IPv6 forms are no destination of their own, and where each
+// is delivered depends on the platform, so every lane refuses them: the
+// anchored egress lane, the local one, and a coder run's proxy, which refused
+// them first.
+
+import (
+	"context"
+	"testing"
+)
+
+var obsoleteForms = []string{
+	"[::c633:6408]:443",        // IPv4-compatible ::198.51.100.8
+	"[::7f00:1]:11434",         // IPv4-compatible ::127.0.0.1
+	"[::c0a8:3214]:80",         // IPv4-compatible ::192.168.50.20, a LAN address
+	"[::ffff:0:c633:6408]:443", // IPv4-translated ::ffff:0:198.51.100.8
+	"[::ffff:0:7f00:1]:80",     // IPv4-translated ::ffff:0:127.0.0.1
+	"[fec0::5]:80",             // site-local
+	"[feff:1::9]:443",          // the top of fec0::/10
+	"compat.example.test:443",  // a name with one such answer among public ones
+	"sitelocal.example.test:80",
+}
+
+func answerObsoleteNames(t *testing.T) {
+	answerEgressNames(t, map[string][]string{
+		"compat.example.test":    {"198.51.100.8", "::c633:6408"},
+		"sitelocal.example.test": {"fec0::5"},
+		"public.example.test":    {"198.51.100.8", "2001:db8::8"},
+	})
+}
+
+// MUTATION: classify the three forms as public in classifyIP.
+// RESULT: red — both egress lanes hand them to the dial.
+func TestEveryLaneRefusesTheDeprecatedIPv6Forms(t *testing.T) {
+	ctx := context.Background()
+	refusedOn := func(t *testing.T, lane string) {
+		t.Helper()
+		for _, hp := range obsoleteForms {
+			if addrs, why := egressVettedAddrs(ctx, hp); why == "" || addrs != nil {
+				t.Errorf("%s must be refused on the %s egress lane, got dial %v", hp, lane, addrs)
+			}
+		}
+		for _, hp := range []string{"198.51.100.8:443", "public.example.test:443"} {
+			if _, why := egressVettedAddrs(ctx, hp); why != "" {
+				t.Errorf("%s is the public internet and passes the %s egress lane: %s", hp, lane, why)
+			}
+		}
+	}
+
+	t.Run("anchored", func(t *testing.T) {
+		setupBox(t)
+		stubOwnAddrs(t, loopbackOnly, localRange())
+		answerObsoleteNames(t)
+		refusedOn(t, "anchored")
+	})
+
+	t.Run("local", func(t *testing.T) {
+		localEgressFixture(t)
+		answerObsoleteNames(t)
+		refusedOn(t, "local")
+	})
+
+	t.Run("a coder run's proxy agrees", func(t *testing.T) {
+		stubOwnAddrs(t, loopbackOnly, localRange())
+		for _, hp := range obsoleteForms[:7] {
+			if addrs, why := coderVettedAddrs(ctx, hp, nil); why == "" || addrs != nil {
+				t.Errorf("%s must be refused on a coder run's proxy, got dial %v", hp, addrs)
+			}
+		}
+		if _, why := coderVettedAddrs(ctx, "198.51.100.8:443", nil); why != "" {
+			t.Errorf("the public internet passes a coder run's proxy: %s", why)
+		}
+	})
+}

client-go/egress_one_request_test.go

@@ -0,0 +1,393 @@
+//go:build !windows
+
+package main
+
+import (
+	"bufio"
+	"bytes"
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"net"
+	"net/http"
+	"strings"
+	"sync"
+	"testing"
+	"time"
+)
+
+// The bastion's egress lanes carry ONE plain-HTTP request per connection, read
+// the way the Host's egress gate reads it, and every refusal says
+// why in its body.
+//
+// MUTATIONS, each red:
+//   - handleEgressConn writes head.String() to upstream and splices, as before:
+//     TestEgressPlainHTTPCarriesOneRequestPerConnection, ...SendsTheHostItChecked;
+//   - oneEgressExchange's watcher forwards instead of closing: ...OneRequest...;
+//   - parseEgressAuthority accepts "@" or cuts the authority at "/" only:
+//     TestEgressPlainHTTPReadsTheHostAsTheGateDoes;
+//   - writeEgressRefusal writes Content-Length: 0: TestEgressRefusalsCarryABody.
+
+// recordingOrigin stands in for a destination that keeps a connection open
+// after answering (it ignores Connection: close, as a careless server may),
+// and records every byte one connection carried to it.
+type recordingOrigin struct {
+	ln    net.Listener
+	conns chan string
+	// silent: never answer, so the lane never holds an unread answer when it
+	// closes (a close with unread data is a reset, which can cost the origin
+	// bytes it has not read yet).
+	silent bool
+}
+
+func newSilentOrigin(t *testing.T) *recordingOrigin {
+	o := newRecordingOrigin(t)
+	o.silent = true
+	return o
+}
+
+func newRecordingOrigin(t *testing.T) *recordingOrigin {
+	t.Helper()
+	ln, err := net.Listen("tcp", "127.0.0.1:0")
+	if err != nil {
+		t.Fatal(err)
+	}
+	o := &recordingOrigin{ln: ln, conns: make(chan string, 16)}
+	t.Cleanup(func() { _ = ln.Close() })
+	go func() {
+		for {
+			c, err := ln.Accept()
+			if err != nil {
+				return
+			}
+			go o.serve(c)
+		}
+	}()
+	return o
+}
+
+func (o *recordingOrigin) serve(c net.Conn) {
+	defer c.Close()
+	var mu sync.Mutex
+	var got bytes.Buffer
+	answered := false
+	buf := make([]byte, 4096)
+	for {
+		_ = c.SetReadDeadline(time.Now().Add(800 * time.Millisecond))
+		n, err := c.Read(buf)
+		mu.Lock()
+		got.Write(buf[:n])
+		if !o.silent && !answered && bytes.Contains(got.Bytes(), []byte("\r\n\r\n")) {
+			answered = true
+			_, _ = io.WriteString(c, "HTTP/1.1 200 OK\r\nContent-Length: 2\r\n\r\nok")
+		}
+		mu.Unlock()
+		if err != nil {
+			break
+		}
+	}
+	o.conns <- got.String()
+}
+
+// next is everything the origin's next connection carried, once it ended.
+func (o *recordingOrigin) next(t *testing.T) string {
+	t.Helper()
+	s, ok := o.nextWithin(8 * time.Second)
+	if !ok {
+		t.Fatal("the destination never saw its connection end")
+	}
+	return s
+}
+
+func (o *recordingOrigin) nextWithin(d time.Duration) (string, bool) {
+	select {
+	case s := <-o.conns:
+		return s, true
+	case <-time.After(d):
+		return "", false
+	}
+}
+
+func (o *recordingOrigin) addr() string { return o.ln.Addr().String() }
+
+// A keep-alive connection carries one request to the destination, never two:
+// neither a request sent after the first answer (a client reusing the
+// connection) nor one pipelined behind the first.
+func TestEgressPlainHTTPCarriesOneRequestPerConnection(t *testing.T) {
+	first := "GET http://198.51.100.30/first HTTP/1.1\r\nHost: 198.51.100.30\r\n\r\n"
+	second := "GET http://198.51.100.30/second HTTP/1.1\r\nHost: 198.51.100.30\r\n\r\n"
+
+	t.Run("reused after the answer", func(t *testing.T) {
+		o := newRecordingOrigin(t)
+		var asked []string
+		cli, cleanup := dialHandler(t, redirectDial(o.addr(), &asked))
+		defer cleanup()
+		_, _ = io.WriteString(cli, first)
+		br := bufio.NewReader(cli)
+		resp, err := http.ReadResponse(br, nil)
+		if err != nil {
+			t.Fatalf("the first request got no answer: %v", err)
+		}
+		body, _ := io.ReadAll(resp.Body)
+		if string(body) != "ok" {
+			t.Fatalf("the first answer was %q", body)
+		}
+		_, _ = io.WriteString(cli, second)
+		got := o.next(t)
+		if strings.Contains(got, "/second") {
+			t.Fatalf("a second request on one connection reached the destination unvetted:\n%s", got)
+		}
+		if !strings.Contains(got, "\r\nConnection: close\r\n") {
+			t.Fatalf("the request went out without Connection: close:\n%s", got)
+		}
+		_ = cli.SetReadDeadline(time.Now().Add(3 * time.Second))
+		if rest, _ := io.ReadAll(br); len(rest) != 0 {
+			t.Fatalf("the lane answered the second request: %q", rest)
+		}
+		if len(asked) != 1 {
+			t.Fatalf("one connection, one dial: %v", asked)
+		}
+	})
+
+	t.Run("pipelined", func(t *testing.T) {
+		o := newRecordingOrigin(t)
+		var asked []string
+		cli, cleanup := dialHandler(t, redirectDial(o.addr(), &asked))
+		defer cleanup()
+		_, _ = io.WriteString(cli, first+second)
+		got := o.next(t)
+		if !strings.Contains(got, "GET http://198.51.100.30/first HTTP/1.1\r\n") {
+			t.Fatalf("the first request did not reach the destination:\n%s", got)
+		}
+		if strings.Contains(got, "/second") {
+			t.Fatalf("a pipelined second request reached the destination unvetted:\n%s", got)
+		}
+	})
+}
+
+// The destination reads the host the lane vetted, in the one Host header; hop
+// headers, those Connection names and a proxy credential stay on this hop.
+func TestEgressPlainHTTPSendsTheHostItChecked(t *testing.T) {
+	o := newRecordingOrigin(t)
+	var asked []string
+	cli, cleanup := dialHandler(t, redirectDial(o.addr(), &asked))
+	defer cleanup()
+	_, _ = io.WriteString(cli, "GET http://198.51.100.31/p?q=1 HTTP/1.1\r\n"+
+		"Host: inside.example.test\r\nProxy-Authorization: Basic eDp5\r\n"+
+		"Connection: keep-alive, X-Hop\r\nX-Hop: 1\r\nX-Keep: 2\r\n\r\n")
+	got := o.next(t)
+	head, _, _ := strings.Cut(got, "\r\n\r\n")
+	want := "GET http://198.51.100.31/p?q=1 HTTP/1.1\r\nHost: 198.51.100.31\r\nX-Keep: 2\r\nConnection: close"
+	if head != want {
+		t.Fatalf("the destination got\n%q\nwant\n%q", head, want)
+	}
+	if len(asked) != 1 || asked[0] != "198.51.100.31:80" {
+		t.Fatalf("the lane must dial the address it vetted: %v", asked)
+	}
+}
+
+// stubEgressNames answers names without DNS and records each name exactly as
+// the lane asked for it.
+func stubEgressNames(t *testing.T, m map[string]string) *[]string {
+	t.Helper()
+	var mu sync.Mutex
+	var asked []string
+	old := egressResolve
+	t.Cleanup(func() { egressResolve = old })
+	egressResolve = func(_ context.Context, host string) ([]net.IPAddr, error) {
+		mu.Lock()
+		asked = append(asked, host)
+		mu.Unlock()
+		ip, ok := m[host]
+		if !ok {
+			return nil, errors.New("no such host")
+		}
+		return []net.IPAddr{{IP: net.ParseIP(ip)}}, nil
+	}
+	return &asked
+}
+
+// The host is read as the gate reads it: the authority ends at the first "/",
+// "?" or "#", is lower-cased and loses one trailing dot; userinfo, an
+// unbracketed IPv6 literal, a bad port, https in absolute form, two Host
+// headers and an ambiguous body length are refused with 400 and a body, before
+// anything is resolved or dialled.
+func TestEgressPlainHTTPReadsTheHostAsTheGateDoes(t *testing.T) {
+	for _, c := range []struct {
+		req, wantName, wantLine string
+	}{
+		{"GET http://public.example.test?@10.0.0.1/ HTTP/1.1\r\n\r\n", "public.example.test",
+			"GET http://public.example.test/?@10.0.0.1/ HTTP/1.1"},
+		{"GET http://public.example.test#@10.0.0.1/ HTTP/1.1\r\n\r\n", "public.example.test",
+			"GET http://public.example.test/ HTTP/1.1"},
+		{"GET http://Public.Example.Test./x HTTP/1.1\r\n\r\n", "public.example.test",
+			"GET http://public.example.test/x HTTP/1.1"},
+		{"GET /x HTTP/1.1\r\nHost: Public.Example.Test:8080\r\n\r\n", "public.example.test",
+			"GET /x HTTP/1.1"},
+	} {
+		o := newRecordingOrigin(t)
+		names := stubEgressNames(t, map[string]string{"public.example.test": "198.51.100.40"})
+		var asked []string
+		cli, cleanup := dialHandler(t, redirectDial(o.addr(), &asked))
+		_, _ = io.WriteString(cli, c.req)
+		got, _ := o.nextWithin(4 * time.Second)
+		cleanup()
+		if len(*names) != 1 || (*names)[0] != c.wantName {
+			t.Errorf("%q: the lane resolved %q; want only %q", c.req, *names, c.wantName)
+		}
+		if line, _, _ := strings.Cut(got, "\r\n"); line != c.wantLine {
+			t.Errorf("%q: the destination got the request line %q; want %q", c.req, line, c.wantLine)
+		}
+	}
+
+	for _, req := range []string{
+		"GET http://user:pw@public.example.test/ HTTP/1.1\r\n\r\n",
+		"GET http://198.51.100.41:80@public.example.test/ HTTP/1.1\r\n\r\n",
+		"GET http://2001:db8::1/ HTTP/1.1\r\n\r\n",
+		"GET http://public.example.test:80x/ HTTP/1.1\r\n\r\n",
+		"GET https://public.example.test/ HTTP/1.1\r\n\r\n",
+		"GET / HTTP/1.1\r\nHost: public.example.test\r\nHost: 198.51.100.41\r\n\r\n",
+		"GET / HTTP/1.1\r\n\r\n",
+		"POST http://public.example.test/ HTTP/1.1\r\nContent-Length: 3\r\nTransfer-Encoding: chunked\r\n\r\n",
+		"POST http://public.example.test/ HTTP/1.1\r\nContent-Length: 3\r\nContent-Length: 4\r\n\r\n",
+		"POST http://public.example.test/ HTTP/1.1\r\nTransfer-Encoding: gzip, chunked\r\n\r\n",
+		"GET http://public.example.test/ HTTP/1.1\r\n X-Folded: 1\r\n\r\n",
+		"CONNECT user@public.example.test:443 HTTP/1.1\r\n\r\n",
+		"CONNECT 2001:db8::1:443 HTTP/1.1\r\n\r\n",
+		"CONNECT [198.51.100.41]:443 HTTP/1.1\r\n\r\n",
+	} {
+		names := stubEgressNames(t, map[string]string{"public.example.test": "198.51.100.40"})
+		var asked []string
+		cli, cleanup := dialHandler(t, redirectDial("127.0.0.1:9", &asked))
+		_, _ = io.WriteString(cli, req)
+		resp, err := http.ReadResponse(bufio.NewReader(cli), nil)
+		if err != nil {
+			t.Errorf("%q: no answer: %v", req, err)
+			cleanup()
+			continue
+		}
+		body, _ := io.ReadAll(resp.Body)
+		cleanup()
+		if resp.StatusCode != 400 || len(body) == 0 || int64(len(body)) != resp.ContentLength {
+			t.Errorf("%q: answered %d with body %q; want 400 with a body saying why", req, resp.StatusCode, body)
+		}
+		if len(*names) != 0 || len(asked) != 0 {
+			t.Errorf("%q: refused requests resolve and dial nothing; resolved %q, dialled %q", req, *names, asked)
+		}
+	}
+}
+
+// A request body goes out framed as its head says, chunked bodies re-encoded,
+// and nothing after it.
+func TestEgressPlainHTTPBodyIsFramed(t *testing.T) {
+	for _, c := range []struct{ req, wantBody string }{
+		{"POST http://198.51.100.42/u HTTP/1.1\r\nHost: 198.51.100.42\r\nContent-Length: 5\r\n\r\nhello" +
+			"GET http://198.51.100.42/smuggled HTTP/1.1\r\n\r\n", "hello"},
+		{"POST http://198.51.100.42/u HTTP/1.1\r\nHost: 198.51.100.42\r\nTransfer-Encoding: chunked\r\n\r\n" +
+			"5;ext=1\r\nhello\r\n0\r\nX-Trailer: 1\r\n\r\nGET /smuggled HTTP/1.1\r\n\r\n", "5\r\nhello\r\n0\r\n\r\n"},
+	} {
+		o := newSilentOrigin(t)
+		var asked []string
+		cli, cleanup := dialHandler(t, redirectDial(o.addr(), &asked))
+		_, _ = io.WriteString(cli, c.req)
+		got := o.next(t)
+		cleanup()
+		_, body, _ := strings.Cut(got, "\r\n\r\n")
+		if body != c.wantBody {
+			t.Errorf("the destination got the body %q; want %q", body, c.wantBody)
+		}
+	}
+}
+
+// A refusal says why in its body, not only in a header: a 403 from the vetting
+// and a 502 from a dial that found no answer, on both request forms, and the
+// coder-run proxy's refusals too.
+func TestEgressRefusalsCarryABody(t *testing.T) {
+	failDial := func(ctx context.Context, network, addr string) (net.Conn, error) {
+		return nil, fmt.Errorf("no route")
+	}
+	okDial := func(ctx context.Context, network, addr string) (net.Conn, error) {
+		return nil, fmt.Errorf("must not dial %s", addr)
+	}
+	check := func(name string, resp *http.Response, err error, want int) {
+		t.Helper()
+		if err != nil {
+			t.Errorf("%s: no answer: %v", name, err)
+			return
+		}
+		body, _ := io.ReadAll(resp.Body)
+		if resp.StatusCode != want || len(body) < 10 || int64(len(body)) != resp.ContentLength {
+			t.Errorf("%s: answered %d with body %q (Content-Length %d); want %d with a short body saying why",
+				name, resp.StatusCode, body, resp.ContentLength, want)
+		}
+	}
+	for _, c := range []struct {
+		name string
+		dial egressDialer
+		req  string
+		want int
+	}{
+		{"plain HTTP to this machine", okDial, "GET http://127.0.0.1:9/ HTTP/1.1\r\n\r\n", 403},
+		{"CONNECT to this machine", okDial, "CONNECT 127.0.0.1:9 HTTP/1.1\r\n\r\n", 403},
+		{"plain HTTP, no answer", failDial, "GET http://198.51.100.43/ HTTP/1.1\r\n\r\n", 502},
+		{"CONNECT, no answer", failDial, "CONNECT 198.51.100.43:443 HTTP/1.1\r\n\r\n", 502},
+	} {
+		cli, cleanup := dialHandler(t, c.dial)
+		_, _ = io.WriteString(cli, c.req)
+		resp, err := http.ReadResponse(bufio.NewReader(cli), nil)
+		check(c.name, resp, err, c.want)
+		cleanup()
+	}
+
+	oldDial := coderDialOut
+	t.Cleanup(func() { coderDialOut = oldDial })
+	coderDialOut = failDial
+	for _, c := range []struct {
+		name, req string
+		want      int
+	}{
+		{"coder run to a private address", "GET http://10.0.0.1/ HTTP/1.1\r\nHost: 10.0.0.1\r\n\r\n", 403},
+		{"coder run, no answer", "GET http://93.184.216.34/ HTTP/1.1\r\nHost: 93.184.216.34\r\n\r\n", 502},
+	} {
+		boxEnd, runEnd := net.Pipe()
+		var refused int64
+		go handleCoderEgressConn(boxEnd, nil, &refused)
+		_ = runEnd.SetDeadline(time.Now().Add(5 * time.Second))
+		_, _ = io.WriteString(runEnd, c.req)
+		resp, err := http.ReadResponse(bufio.NewReader(runEnd), nil)
+		check(c.name, resp, err, c.want)
+		_ = runEnd.Close()
+	}
+}
+
+// The coder-run proxy carries one request per connection too: the run's own
+// processes are the sender there.
+func TestCoderEgressCarriesOneRequestPerConnection(t *testing.T) {
+	o := newRecordingOrigin(t)
+	oldDial := coderDialOut
+	t.Cleanup(func() { coderDialOut = oldDial })
+	coderDialOut = func(ctx context.Context, network, addr string) (net.Conn, error) {
+		return (&net.Dialer{}).DialContext(ctx, network, o.addr())
+	}
+	boxEnd, runEnd := net.Pipe()
+	var refused int64
+	go handleCoderEgressConn(boxEnd, nil, &refused)
+	_ = runEnd.SetDeadline(time.Now().Add(8 * time.Second))
+	_, _ = io.WriteString(runEnd, "GET http://93.184.216.34/first HTTP/1.1\r\nHost: 93.184.216.34\r\n\r\n")
+	br := bufio.NewReader(runEnd)
+	resp, err := http.ReadResponse(br, nil)
+	if err != nil {
+		t.Fatalf("the first request got no answer: %v", err)
+	}
+	_, _ = io.ReadAll(resp.Body)
+	_, _ = io.WriteString(runEnd, "GET http://93.184.216.34/second HTTP/1.1\r\nHost: inside.example.test\r\n\r\n")
+	got := o.next(t)
+	_ = runEnd.Close()
+	if strings.Contains(got, "/second") || strings.Contains(got, "inside.example.test") {
+		t.Fatalf("a second request on one connection reached the destination unvetted:\n%s", got)
+	}
+	if !strings.Contains(got, "\r\nConnection: close\r\n") {
+		t.Fatalf("the request went out without Connection: close:\n%s", got)
+	}
+}

client-go/egress_relay.go

@@ -14,35 +14,60 @@ import (
 
 const egressWorkerCount = 4
 
+// egressRefusedWait is how long a worker waits before it parks again after a
+// job it did not take (malformed, refused) or after its WebSocket closed while
+// parked. A job it took parks the worker again at once, so the person's own
+// traffic is never slowed by it; a Host that answers every worker with a job
+// this machine refuses costs it at most egressWorkerCount dials, handshakes
+// and refusals a second, not as many as the Host can send. The serve lane's
+// workers wait the same (serveRefusedWait).
+var egressRefusedWait = time.Second
+
 // wsConn adapts a worker WebSocket into a net.Conn so the existing handleEgressConn drives
 // it unchanged. The host frames the proxied client stream as binary messages: 'D'+payload =
 // data, 'E' = end-of-stream. Reads surface the host's bytes; writes go back as 'D' frames;
 // Close/CloseWrite send 'E'.
+//
+// A 'D' frame is STREAMED into the caller's buffer, never read whole: the host picks a
+// frame's length, and one frame of any length must cost this machine no more memory
+// than the read it is asked for.
 type wsConn struct {
 	ws  *websocket.Conn
-	rbuf []byte
+	cur io.Reader // the rest of the 'D' frame being read, or nil
 	wmu sync.Mutex
 }
 
 func (c *wsConn) Read(p []byte) (int, error) {
-	for len(c.rbuf) == 0 {
-		mt, data, err := c.ws.ReadMessage()
+	for {
+		if c.cur != nil {
+			n, err := c.cur.Read(p)
+			if err == io.EOF {
+				c.cur = nil
+				err = nil
+			}
+			if n > 0 || err != nil {
+				return n, err
+			}
+			continue
+		}
+		mt, r, err := c.ws.NextReader()
 		if err != nil {
 			return 0, err
 		}
-		if mt != websocket.BinaryMessage || len(data) == 0 {
-			continue
+		if mt != websocket.BinaryMessage {
+			continue // NextReader discards what is left of it
 		}
-		switch data[0] {
+		var tag [1]byte
+		if _, err := io.ReadFull(r, tag[:]); err != nil {
+			continue // an empty frame
+		}
+		switch tag[0] {
 		case 'D':
-			c.rbuf = data[1:]
+			c.cur = r
 		case 'E':
 			return 0, io.EOF
 		}
 	}
-	n := copy(p, c.rbuf)
-	c.rbuf = c.rbuf[n:]
-	return n, nil
 }
 
 func (c *wsConn) Write(p []byte) (int, error) {
@@ -80,19 +105,25 @@ func (c *wsConn) SetWriteDeadline(t time.Time) error { return c.ws.SetWriteDeadl
 // bastion's tunnel IP) and serves one egress job per worker, reconnecting after each. Each
 // worker authenticates with the bastion's bearer, then blocks until the host dispatches a
 // request; handleEgressConn then dials it out the USER's connection (dialOut) and relays.
+// A worker reconnects at once after a job it took, and after egressRefusedWait after one
+// it refused or a close while parked. The func returned waits until every worker has
+// exited, which is once pctx is done and each one's job has ended.
 //
 // pctx is the pairing's context (run.go): when the anchor revokes a pairing the
 // workers must exit, not re-dial a closed tunnel every 3s for the process's life.
 func runEgressWorkers(pctx context.Context, self string, wsURL string, tlsCfg *tls.Config,
 	dialHost func(ctx context.Context) (net.Conn, error), bearer string,
-	dialOut egressDialer, n int) {
+	dialOut egressDialer, n int) (wait func()) {
 	dialer := &websocket.Dialer{
 		NetDialContext:   func(ctx context.Context, network, addr string) (net.Conn, error) { return dialHost(ctx) },
 		TLSClientConfig:  tlsCfg,
 		HandshakeTimeout: 20 * time.Second,
 	}
+	var wg sync.WaitGroup
 	for i := 0; i < n; i++ {
+		wg.Add(1)
 		go func() {
+			defer wg.Done()
 			for pctx.Err() == nil {
 				ctx, cancel := context.WithTimeout(pctx, 25*time.Second)
 				ws, _, err := dialer.DialContext(ctx, wsURL, http.Header{})
@@ -108,11 +139,15 @@ func runEgressWorkers(pctx context.Context, self string, wsURL string, tlsCfg *t
 				}
 				stopKA := armWSKeepalive(ws) // ping/read-deadline: a dead tunnel unblocks the park
 				unpark := context.AfterFunc(pctx, func() { _ = ws.Close() })
-				handleEgressConn(&wsConn{ws: ws}, dialOut, self) // blocks until a job, serves it, then returns
+				took := handleEgressConn(&wsConn{ws: ws}, dialOut, self) // blocks until a job, serves it, then returns
 				unpark()
 				stopKA()
 				_ = ws.Close()
+				if !took {
+					sleepCtx(pctx, egressRefusedWait)
+				}
 			}
 		}()
 	}
+	return wg.Wait
 }

client-go/egress_relay_test.go

@@ -0,0 +1,166 @@
+//go:build !windows
+
+// The egress worker pool (egress_relay.go: runEgressWorkers) against a
+// stand-in Host's /bastion-egress: how fast workers park again after a job
+// they took and after one they refused. Uses localEgressFixture
+// (egress_local_test.go), so it carries the same build tag.
+
+package main
+
+import (
+	"context"
+	"io"
+	"log"
+	"net"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"github.com/gorilla/websocket"
+)
+
+// egressHost is a Host's /bastion-egress as the pool meets it: every worker
+// that parks is counted and handed the job job(n) (n is its place in line),
+// and what the bastion answered is read to its end.
+type egressHost struct {
+	parked atomic.Int32
+	ok     atomic.Int32 // answers that began "HTTP/1.1 200"
+}
+
+func startEgressHost(t *testing.T, job func(n int32) string) (*egressHost, string, func(context.Context) (net.Conn, error)) {
+	t.Helper()
+	h := &egressHost{}
+	up := websocket.Upgrader{}
+	srv := httptest.NewUnstartedServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		c, err := up.Upgrade(w, r, nil)
+		if err != nil {
+			return
+		}
+		defer c.Close()
+		if _, _, err := c.ReadMessage(); err != nil { // BEARER
+			return
+		}
+		n := h.parked.Add(1)
+		if err := c.WriteMessage(websocket.BinaryMessage, append([]byte{'D'}, job(n)...)); err != nil {
+			return
+		}
+		_ = c.SetReadDeadline(time.Now().Add(5 * time.Second))
+		var answer strings.Builder
+		for {
+			_, m, err := c.ReadMessage()
+			if err != nil || len(m) == 0 || m[0] == 'E' {
+				break
+			}
+			answer.Write(m[1:])
+		}
+		if strings.HasPrefix(answer.String(), "HTTP/1.1 200") {
+			h.ok.Add(1)
+		}
+	}))
+	srv.Config.ErrorLog = log.New(io.Discard, "", 0)
+	srv.Start()
+	t.Cleanup(srv.Close)
+	addr := srv.Listener.Addr().String()
+	dial := func(ctx context.Context) (net.Conn, error) {
+		return (&net.Dialer{}).DialContext(ctx, "tcp", addr)
+	}
+	return h, "ws://" + addr + "/bastion-egress", dial
+}
+
+// runEgressPool runs egressWorkerCount workers for the test's life, and at its
+// end stops them and waits until each has exited: a worker still vetting a job
+// would otherwise read the fixture's stubs while the test's cleanup puts them
+// back.
+func runEgressPool(t *testing.T, wsURL string, dialHost func(context.Context) (net.Conn, error), dialOut egressDialer) {
+	t.Helper()
+	ctx, cancel := context.WithCancel(context.Background())
+	wait := runEgressWorkers(ctx, "home.host.example", wsURL, nil, dialHost, "test-bearer", dialOut, egressWorkerCount)
+	t.Cleanup(func() {
+		cancel()
+		done := make(chan struct{})
+		go func() { wait(); close(done) }()
+		select {
+		case <-done:
+		case <-time.After(10 * time.Second):
+			t.Error("the egress workers did not exit 10 s after the pool stopped")
+		}
+	})
+}
+
+// SECURITY (this machine's resources): a Host that answers every parked egress
+// worker with a job this machine refuses (a request for this machine, a line
+// that is no request) cannot make it dial, handshake and refuse in a loop as
+// fast as the Host answers. Each worker that refused waits before it parks
+// again, so the pool costs at most egressWorkerCount dials a second.
+//
+// MUTATION: in runEgressWorkers, park again at once whatever handleEgressConn
+// reports (as the pool did).
+// RESULT: red — hundreds of workers park in the 2 s window.
+func TestARefusingHostCannotSpinTheEgressWorkers(t *testing.T) {
+	localEgressFixture(t)
+	var dialled atomic.Int32
+	dialOut := func(ctx context.Context, network, addr string) (net.Conn, error) {
+		dialled.Add(1)
+		return nil, io.EOF
+	}
+	h, wsURL, dialHost := startEgressHost(t, func(n int32) string {
+		if n%2 == 0 {
+			return "NOT-A-REQUEST\r\n"
+		}
+		return "CONNECT localhost:443 HTTP/1.1\r\nHost: localhost:443\r\n\r\n"
+	})
+	runEgressPool(t, wsURL, dialHost, dialOut)
+	time.Sleep(2 * time.Second)
+	// egressWorkerCount park at the start, then each worker once a second:
+	// about 12 in 2 s. A pool that parks again at once makes hundreds.
+	n := h.parked.Load()
+	t.Logf("%d workers parked in 2 s", n)
+	if limit := int32(egressWorkerCount * 4); n > limit {
+		t.Fatalf("%d egress workers parked in 2 s on a Host that refuses every job (at most %d): the pool re-dials in a loop", n, limit)
+	}
+	if n < egressWorkerCount {
+		t.Fatalf("only %d workers parked: the pool did not run, so the count proves nothing", n)
+	}
+	if d := dialled.Load(); d != 0 {
+		t.Fatalf("a refused job was dialled out %d times", d)
+	}
+}
+
+// The person's own traffic is not slowed: a worker that took a job parks again
+// at once, so 40 jobs through 4 workers finish in a fraction of the time one
+// wait per job would take (40 / 4 x egressRefusedWait = 10 s).
+//
+// MUTATION: in runEgressWorkers, wait egressRefusedWait after every job.
+// RESULT: red — 40 jobs need about 10 s.
+func TestEgressWorkersParkAgainAtOnceAfterAJobTheyTook(t *testing.T) {
+	localEgressFixture(t)
+	dialOut := func(ctx context.Context, network, addr string) (net.Conn, error) {
+		if addr != "198.51.100.9:443" {
+			t.Errorf("dialled %q, not the address the request named", addr)
+		}
+		ours, theirs := net.Pipe()
+		_ = theirs.Close() // the destination accepts and ends at once
+		return ours, nil
+	}
+	h, wsURL, dialHost := startEgressHost(t, func(int32) string {
+		return "CONNECT 198.51.100.9:443 HTTP/1.1\r\nHost: 198.51.100.9:443\r\n\r\n"
+	})
+	start := time.Now()
+	runEgressPool(t, wsURL, dialHost, dialOut)
+	const jobs = 40
+	deadline := start.Add(10 * time.Second)
+	for h.ok.Load() < jobs && time.Now().Before(deadline) {
+		time.Sleep(10 * time.Millisecond)
+	}
+	took := time.Since(start)
+	t.Logf("%d accepted jobs answered in %v", h.ok.Load(), took)
+	if got := h.ok.Load(); got < jobs {
+		t.Fatalf("only %d of %d accepted jobs were answered 200 in %v", got, jobs, took)
+	}
+	if limit := 3 * egressRefusedWait; took > limit {
+		t.Fatalf("%d accepted jobs took %v (over %v): a worker that took a job waited before it parked again", jobs, took, limit)
+	}
+}

client-go/egress_test.go

@@ -38,6 +38,16 @@ func dialHandler(t *testing.T, dialOut egressDialer) (net.Conn, func()) {
 	return cli, func() { cli.Close(); ln.Close() }
 }
 
+// redirectDial stands in for the user's internet: the lane asks for a public
+// destination, and this dials the loopback backend instead, recording what the
+// lane asked for. The lane never reaches loopback itself.
+func redirectDial(backend string, asked *[]string) egressDialer {
+	return func(ctx context.Context, network, addr string) (net.Conn, error) {
+		*asked = append(*asked, addr)
+		return (&net.Dialer{}).DialContext(ctx, network, backend)
+	}
+}
+
 // CONNECT establishes a raw tunnel; bytes flow both ways through it.
 func TestEgressConnConnectTunnel(t *testing.T) {
 	backend, err := net.Listen("tcp", "127.0.0.1:0")
@@ -54,9 +64,10 @@ func TestEgressConnConnectTunnel(t *testing.T) {
 		_, _ = io.Copy(c, c) // echo
 	}()
 
-	cli, cleanup := dialHandler(t, (&net.Dialer{}).DialContext)
+	var asked []string
+	cli, cleanup := dialHandler(t, redirectDial(backend.Addr().String(), &asked))
 	defer cleanup()
-	fmt.Fprintf(cli, "CONNECT %s HTTP/1.1\r\nHost: %s\r\n\r\n", backend.Addr(), backend.Addr())
+	fmt.Fprint(cli, "CONNECT 198.51.100.20:443 HTTP/1.1\r\nHost: 198.51.100.20:443\r\n\r\n")
 	br := bufio.NewReader(cli)
 	status, err := br.ReadString('\n')
 	if err != nil || !strings.Contains(status, "200") {
@@ -82,6 +93,9 @@ func TestEgressConnConnectTunnel(t *testing.T) {
 	if string(buf) != want {
 		t.Fatalf("tunnel echo mismatch: got %q", string(buf))
 	}
+	if len(asked) != 1 || asked[0] != "198.51.100.20:443" {
+		t.Fatalf("the lane must dial the address it vetted: %v", asked)
+	}
 }
 
 func TestEgressConnPlainHTTP(t *testing.T) {
@@ -89,15 +103,18 @@ func TestEgressConnPlainHTTP(t *testing.T) {
 		fmt.Fprint(w, "plain-http-ok")
 	}))
 	defer backend.Close()
-	host := strings.TrimPrefix(backend.URL, "http://")
 
-	cli, cleanup := dialHandler(t, (&net.Dialer{}).DialContext)
+	var asked []string
+	cli, cleanup := dialHandler(t, redirectDial(strings.TrimPrefix(backend.URL, "http://"), &asked))
 	defer cleanup()
-	fmt.Fprintf(cli, "GET http://%s/ HTTP/1.1\r\nHost: %s\r\nConnection: close\r\n\r\n", host, host)
+	fmt.Fprint(cli, "GET http://198.51.100.21/ HTTP/1.1\r\nHost: 198.51.100.21\r\nConnection: close\r\n\r\n")
 	body, _ := io.ReadAll(cli)
 	if !strings.Contains(string(body), "plain-http-ok") {
 		t.Fatalf("plain HTTP forward: got %q", string(body))
 	}
+	if len(asked) != 1 || asked[0] != "198.51.100.21:80" {
+		t.Fatalf("the lane must dial the address it vetted: %v", asked)
+	}
 }
 
 func TestEgressConnDialFailure(t *testing.T) {
@@ -106,7 +123,7 @@ func TestEgressConnDialFailure(t *testing.T) {
 	}
 	cli, cleanup := dialHandler(t, failDial)
 	defer cleanup()
-	fmt.Fprint(cli, "CONNECT unreachable.example:443 HTTP/1.1\r\n\r\n")
+	fmt.Fprint(cli, "CONNECT 198.51.100.22:443 HTTP/1.1\r\n\r\n")
 	status, _ := bufio.NewReader(cli).ReadString('\n')
 	if !strings.Contains(status, "502") {
 		t.Fatalf("expected 502 on dial failure, got %q", status)

client-go/endpoint_invalid_test.go

@@ -0,0 +1,230 @@
+//go:build !windows
+
+// Uses fakeToken from pairing_redeem_test.go and setupBox/driveOp from
+// anchor_test.go, both unix-only.
+
+package main
+
+// A pairing whose file holds an endpoint this build refuses loads as invalid
+// and is never served (endpoint_test.go). These are the paths that could
+// still act on it: anchoring the box to it, an approval from the anchor host,
+// what the hosts are told about it, and the two ways a person clears it.
+
+import (
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+)
+
+// unservedEndpoint is one the endpoint check refuses (badTunnelEndpoints).
+const unservedEndpoint = "203.0.113.7:51820\nallowed_ip=::/0"
+
+// writeUnservedPairing writes a pairing file in the given state whose
+// endpoint this build refuses, and returns its path.
+func writeUnservedPairing(t *testing.T, host, state string) string {
+	t.Helper()
+	pd, err := pairingsDir()
+	if err != nil {
+		t.Fatal(err)
+	}
+	b, _ := json.Marshal(Config{HostName: host, WGPrivateKey: "k", BastionTunnelIP: "fd12:9::a:1",
+		State: state, Endpoint: unservedEndpoint})
+	p := filepath.Join(pd, host+".json")
+	if err := os.WriteFile(p, b, 0o600); err != nil {
+		t.Fatal(err)
+	}
+	return p
+}
+
+// fileState is the state field as written in host's pairing file.
+func fileState(t *testing.T, host string) string {
+	t.Helper()
+	pd, _ := pairingsDir()
+	raw, err := os.ReadFile(filepath.Join(pd, host+".json"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	var c Config
+	if err := json.Unmarshal(raw, &c); err != nil {
+		t.Fatal(err)
+	}
+	return c.State
+}
+
+// claim-anchor names the reason and writes nothing: the pairing keeps its
+// state, and the box is neither anchored nor claimed. As the sole pairing and
+// named with --host among others, in the legacy active state and pending.
+func TestClaimAnchorRefusesAPairingThatIsNotServed(t *testing.T) {
+	if os.Geteuid() != 0 {
+		t.Skip("claim-anchor needs machine authority, which needs root")
+	}
+	for _, c := range []struct {
+		name   string
+		others []string
+		host   string
+	}{{"sole pairing", nil, ""}, {"named with --host", []string{tAnchor}, tMember}} {
+		for _, st := range []string{"", pairingPending} {
+			t.Run(c.name+"/"+st, func(t *testing.T) {
+				t.Setenv("XDG_CONFIG_HOME", t.TempDir())
+				withMarkerAt(t, filepath.Join(t.TempDir(), "machine.json"))
+				if err := setBoxMode(modeNetworked); err != nil {
+					t.Fatal(err)
+				}
+				for _, o := range c.others {
+					writePairing(t, o)
+				}
+				p := writeUnservedPairing(t, tMember, st)
+				err := cmdClaimAnchor("", false, "", c.host)
+				if err == nil || !contains(err.Error(), "not a host:port") || !contains(err.Error(), p) {
+					t.Fatalf("claim-anchor must refuse the pairing, naming its file and why: %v", err)
+				}
+				if a := currentAnchor(); a != nil {
+					t.Fatalf("a refused claim anchored the box: %+v", a)
+				}
+				if oc := currentOwnerClaim(); oc != nil {
+					t.Fatalf("a refused claim wrote the owner claim: %+v", oc)
+				}
+				if got := fileState(t, tMember); got != st {
+					t.Fatalf("a refused claim changed the pairing's state from %q to %q", st, got)
+				}
+			})
+		}
+	}
+}
+
+// The anchor host's approval of such a pairing is refused: it is not made
+// active, no grant is kept, and the refusal names no file of this machine.
+// The copy updatePairingOnDisk hands back carries the same mark, so a caller
+// that starts a pairing from it starts nothing.
+func TestAnApprovalOfAPairingThatIsNotServedIsRefused(t *testing.T) {
+	priv := setupBox(t)
+	writeUnservedPairing(t, tMember, pairingPending)
+	future := time.Now().Add(time.Hour).UTC().Format(time.RFC3339)
+	terms, sig := signTerms(t, priv, grantTerms(tMember, []string{"observe"}, future))
+	_, errOut, code := driveOp(t, tAnchor, "pairing_grant_sync", []string{terms, sig})
+	if code != 126 || !contains(errOut, "not served") {
+		t.Fatalf("the approval must be refused (126, not served): %d %s", code, errOut)
+	}
+	pd, _ := pairingsDir()
+	if contains(errOut, pd) {
+		t.Fatalf("the refusal sent to the host names a file of this machine: %s", errOut)
+	}
+	if got := fileState(t, tMember); got != pairingPending {
+		t.Fatalf("a refused approval changed the state to %q", got)
+	}
+	if _, err := loadPairingGrant(tMember); err == nil {
+		t.Fatal("a refused approval kept the grant")
+	}
+
+	c, err := updatePairingOnDisk(tMember, func(d *Config) { d.State = pairingActive })
+	if err != nil {
+		t.Fatal(err)
+	}
+	if c.invalid == "" || c.active() {
+		t.Fatalf("the copy written back must still be invalid and never active: invalid=%q active=%v", c.invalid, c.active())
+	}
+	if startPairing(c) {
+		t.Fatal("a pairing that is not served was started")
+	}
+}
+
+// What the hosts hear: neither the peer list every host gets, nor the
+// pairing list the anchor host gets (in the heartbeat and from pairing_list),
+// names such a pairing, and every state they do carry is one a host knows.
+// The machine's own status still shows it invalid (endpoint_test.go).
+func TestNoHostIsToldOfAPairingThatIsNotServed(t *testing.T) {
+	setupBox(t)
+	const beta, pend = "beta.host.privatae.ai", "pend.host.privatae.ai"
+	writePairing(t, beta)
+	writeUnservedPairing(t, tMember, "")
+	writeUnservedPairing(t, pend, pairingPending)
+	known := map[string]bool{pairingActive: true, pairingPending: true, pairingRevoked: true}
+
+	peers := peersFor(tAnchor)
+	if len(peers) != 1 || peers[0].HostName != beta {
+		t.Fatalf("the peers must be only %s, got %+v", beta, peers)
+	}
+	for _, p := range peers {
+		if !known[p.State] {
+			t.Fatalf("a peer was sent with a state a host does not know: %+v", p)
+		}
+	}
+	list := pairingsReport()
+	var names []string
+	for _, e := range list {
+		names = append(names, e.HostName)
+		if !known[e.State] {
+			t.Fatalf("a pairing was sent with a state a host does not know: %+v", e)
+		}
+	}
+	if got := strings.Join(names, ","); got != tAnchor+","+beta {
+		t.Fatalf("the anchor host's list must be %s,%s, got %s", tAnchor, beta, got)
+	}
+	out, errOut, code := driveOp(t, tAnchor, "pairing_list", nil)
+	if code != 0 || contains(out, tMember) || contains(out, pend) || contains(out, pairingInvalid) {
+		t.Fatalf("pairing_list must leave the pairing out: %d %s %s", code, out, errOut)
+	}
+	if u := unapprovedAttempts(time.Now().Add(24 * time.Hour)); len(u) != 0 {
+		t.Fatalf("a pairing that is not served is no unapproved attempt: %+v", u)
+	}
+	cfgs, err := loadPairings()
+	if err != nil {
+		t.Fatal(err)
+	}
+	if c := pairingFor(cfgs, tMember); c == nil || c.stateName() != pairingInvalid {
+		t.Fatalf("this machine still lists it invalid: %+v", c)
+	}
+}
+
+// pair --replace over such a pairing pairs it again, served; unpair removes
+// one. Both through the commands a person runs.
+func TestPairReplaceAndUnpairClearAPairingThatIsNotServed(t *testing.T) {
+	uiTestConfigDir(t)
+	withMarkerAt(t, filepath.Join(t.TempDir(), "machine.json"))
+	prev := unpairNeedsRoot
+	unpairNeedsRoot = func() bool { return false }
+	t.Cleanup(func() { unpairNeedsRoot = prev })
+	const host, good = "acme.host.privatae.ai", "203.0.113.7:51820"
+	redeem := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_ = json.NewEncoder(w).Encode(map[string]any{
+			"operator_crt": "CERT", "ca_crt": "CA", "bearer": "b",
+			"endpoint": good, "endpoints": []string{good},
+			"endpoint_pubkey": testWGKey(t), "bastion_tunnel_ip": "2001:db8::a:1",
+			"host_tunnel_ip": "2001:db8::e:3", "host_name": host,
+		})
+	}))
+	defer redeem.Close()
+	old := redeemURLBase
+	redeemURLBase = redeem.URL
+	t.Cleanup(func() { redeemURLBase = old })
+
+	writeUnservedPairing(t, host, "")
+	if err := cmdPair(fakeToken("acme"), "", "", true); err != nil {
+		t.Fatalf("pair --replace over a pairing that is not served: %v", err)
+	}
+	cfgs, err := loadPairings()
+	if err != nil || len(cfgs) != 1 {
+		t.Fatalf("after pair --replace one pairing loads: %v %v", cfgs, err)
+	}
+	if c := cfgs[0]; c.HostName != host || c.invalid != "" || !c.active() || c.Endpoint != good {
+		t.Fatalf("pair --replace must leave a served pairing: %q invalid=%q active=%v endpoint=%q",
+			c.HostName, c.invalid, c.active(), c.Endpoint)
+	}
+
+	writeUnservedPairing(t, host, "")
+	if err := cmdUnpair(host); err != nil {
+		t.Fatalf("unpair of a pairing that is not served: %v", err)
+	}
+	if cfgs, err := loadPairings(); err != nil || len(cfgs) != 0 {
+		t.Fatalf("after unpair nothing is left: %v %v", cfgs, err)
+	}
+	pd, _ := pairingsDir()
+	if _, err := os.Stat(filepath.Join(pd, host+".json")); !os.IsNotExist(err) {
+		t.Fatalf("the pairing file is still there: %v", err)
+	}
+}

client-go/endpoint_test.go

@@ -0,0 +1,207 @@
+//go:build !windows
+
+// Uses fakeToken from pairing_redeem_test.go, unix-only.
+
+package main
+
+// A tunnel endpoint arrives from the host three ways: in the pairing
+// response, in the rotate directive on a heartbeat answer, and from the
+// pairing file those were saved to. Each one refuses an endpoint that is
+// not one host:port, the file by never serving that pairing (the
+// device-level checks are in claims_tunnel_test.go).
+
+import (
+	"encoding/json"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// A pairing response whose endpoint, or any entry of its endpoints list, is
+// not one host:port is refused and nothing is saved. An IPv4, an IPv6 and a
+// host name endpoint pair.
+func TestAPairingResponseWithAnUnsafeEndpointSavesNothing(t *testing.T) {
+	cfgHome := uiTestConfigDir(t)
+	withMarkerAt(t, filepath.Join(t.TempDir(), "machine.json"))
+	var endpoint, hostName string
+	var endpoints []string
+	redeem := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_ = json.NewEncoder(w).Encode(map[string]any{
+			"operator_crt": "CERT", "ca_crt": "CA", "bearer": "b",
+			"endpoint": endpoint, "endpoints": endpoints,
+			"endpoint_pubkey": testWGKey(t), "bastion_tunnel_ip": "2001:db8::a:1",
+			"host_tunnel_ip": "2001:db8::e:3", "host_name": hostName,
+		})
+	}))
+	defer redeem.Close()
+	old := redeemURLBase
+	redeemURLBase = redeem.URL
+	t.Cleanup(func() { redeemURLBase = old })
+
+	hostName = "acme.host.privatae.ai"
+	for _, bad := range badTunnelEndpoints {
+		if bad == "" {
+			continue // a missing endpoint has its own refusal
+		}
+		for _, c := range []struct {
+			endpoint string
+			list     []string
+		}{{bad, nil}, {"203.0.113.7:51820", []string{"[2001:db8::7]:51820", bad}}} {
+			endpoint, endpoints = c.endpoint, c.list
+			if _, _, err := pairWithToken(fakeToken("acme"), false, false); err == nil || !strings.Contains(err.Error(), "not a host:port") {
+				t.Errorf("%q in %+v: the pairing must be refused for its endpoint, got %v", bad, c, err)
+			}
+		}
+	}
+	_ = filepath.Walk(cfgHome, func(path string, fi os.FileInfo, err error) error {
+		if err == nil && !fi.IsDir() {
+			t.Errorf("a refused pairing left a file behind: %s", path)
+		}
+		return nil
+	})
+
+	for _, good := range goodTunnelEndpoints {
+		uiTestConfigDir(t) // a local bastion serves one host: a fresh box each
+		withMarkerAt(t, filepath.Join(t.TempDir(), "machine.json"))
+		endpoint, endpoints = good, []string{good}
+		cfg, _, err := pairWithToken(fakeToken("acme"), false, false)
+		if err != nil {
+			t.Fatalf("%q must pair: %v", good, err)
+		}
+		if cfg.Endpoint != good || len(cfg.Endpoints) != 1 || cfg.Endpoints[0] != good {
+			t.Fatalf("%q: saved %q %q", good, cfg.Endpoint, cfg.Endpoints)
+		}
+	}
+}
+
+// A pairing file that already holds an unsafe endpoint (written by an older
+// build, or by hand) loads as invalid, naming the file and why: it is never
+// served, while every other pairing still loads, status shows it, and unpair
+// removes it. A machine that upgrades is never stopped by one such file.
+func TestAPairingFileWithAnUnsafeEndpointLoadsButIsNeverServed(t *testing.T) {
+	uiTestConfigDir(t)
+	withMarkerAt(t, filepath.Join(t.TempDir(), "machine.json"))
+	supervised(t)
+	prev := unpairNeedsRoot
+	unpairNeedsRoot = func() bool { return false }
+	t.Cleanup(func() { unpairNeedsRoot = prev })
+	pd, err := pairingsDir()
+	if err != nil {
+		t.Fatal(err)
+	}
+	write := func(c *Config) string {
+		t.Helper()
+		b, _ := json.Marshal(c)
+		p := filepath.Join(pd, c.HostName+".json")
+		if err := os.WriteFile(p, b, 0o600); err != nil {
+			t.Fatal(err)
+		}
+		return p
+	}
+	const host, other = "acme.host.privatae.ai", "other.host.privatae.ai"
+	write(&Config{HostName: other, Endpoint: goodTunnelEndpoints[0]})
+	for _, bad := range badTunnelEndpoints {
+		if bad == "" {
+			continue // a pairing over a socket has no endpoint
+		}
+		for _, c := range []*Config{
+			{HostName: host, Endpoint: bad},
+			{HostName: host, Endpoint: "203.0.113.7:51820", Endpoints: []string{bad}},
+		} {
+			p := write(c)
+			cfgs, err := loadPairings()
+			if err != nil || len(cfgs) != 2 {
+				t.Fatalf("%q: one bad file must not fail the load: %v %v", bad, cfgs, err)
+			}
+			got := pairingFor(cfgs, host)
+			if got.active() || got.stateName() != "invalid" {
+				t.Errorf("%q: the pairing must load invalid: active=%v %q", bad, got.active(), got.stateName())
+			}
+			if out := statusOutput(t); !strings.Contains(out, host+"  [invalid]") ||
+				!strings.Contains(out, "not a host:port") || !strings.Contains(out, p) {
+				t.Errorf("%q: status must show the pairing invalid, naming its file and why:\n%s", bad, out)
+			}
+			if !pairingFor(cfgs, other).active() {
+				t.Errorf("%q: the other pairing must still be served", bad)
+			}
+			if startPairing(got) {
+				t.Errorf("%q: an invalid pairing was started", bad)
+			}
+		}
+	}
+	// The supervisor starts nothing for it either (the other pairing is
+	// removed first: starting it would raise a real tunnel).
+	if err := os.Remove(filepath.Join(pd, other+".json")); err != nil {
+		t.Fatal(err)
+	}
+	superviseOnce()
+	supervisor.Lock()
+	running := len(supervisor.runs)
+	supervisor.Unlock()
+	if running != 0 {
+		t.Fatalf("the supervisor started %d run(s) for an invalid pairing", running)
+	}
+
+	if h, err := unpairHost(host); err != nil || h != host {
+		t.Fatalf("unpair must remove an invalid pairing: %q %v", h, err)
+	}
+	if cfgs, err := loadPairings(); err != nil || len(cfgs) != 0 {
+		t.Fatalf("after unpair nothing is left: %v %v", cfgs, err)
+	}
+
+	for _, good := range goodTunnelEndpoints {
+		write(&Config{HostName: host, Endpoint: good, Endpoints: goodTunnelEndpoints})
+		if cfgs, err := loadPairings(); err != nil || len(cfgs) != 1 || cfgs[0].Endpoint != good || !cfgs[0].active() {
+			t.Fatalf("%q must load and be served: %v %v", good, cfgs, err)
+		}
+	}
+	write(&Config{HostName: host, Transport: transportUnix})
+	if cfgs, err := loadPairings(); err != nil || len(cfgs) != 1 || !cfgs[0].active() {
+		t.Fatalf("a pairing with no endpoint must load and be served: %v %v", cfgs, err)
+	}
+}
+
+// statusOutput is what `status` prints.
+func statusOutput(t *testing.T) string {
+	t.Helper()
+	r, w, err := os.Pipe()
+	if err != nil {
+		t.Fatal(err)
+	}
+	stdout := os.Stdout
+	os.Stdout = w
+	serr := cmdStatus()
+	os.Stdout = stdout
+	_ = w.Close()
+	out, _ := io.ReadAll(r)
+	if serr != nil {
+		t.Fatalf("status: %v", serr)
+	}
+	return string(out)
+}
+
+// A rotate directive that names any unsafe endpoint is not acted on; one
+// that names only host:port entries is.
+func TestARotateDirectiveWithAnUnsafeEndpointIsIgnored(t *testing.T) {
+	for _, bad := range badTunnelEndpoints {
+		if bad == "" {
+			continue // an empty entry is no entry
+		}
+		for _, d := range [][]byte{
+			directive(bad, nil, "PK="),
+			directive("203.0.113.7:51820", []string{bad, "[2001:db8::7]:51820"}, "PK="),
+		} {
+			if c, pk := parseRotateTarget(d); len(c) != 0 || pk != "" {
+				t.Errorf("%q: the directive must be ignored, got %q %q", bad, c, pk)
+			}
+		}
+	}
+	c, pk := parseRotateTarget(directive("203.0.113.7:51820", goodTunnelEndpoints, "PK="))
+	if pk != "PK=" || !sameList(c, goodTunnelEndpoints) {
+		t.Fatalf("a directive of host:port entries must be taken: %q %q", c, pk)
+	}
+}

client-go/enforcing.go

@@ -193,6 +193,14 @@ func buildEnforcingReport(dir string, allowed []string) map[string]any {
 		if !aok {
 			rep["action_ops_reason"] = awhy
 		}
+		// Each verb, and whether this machine has the tool it runs: an op
+		// with no program (interpreter: none) runs here when its verbs do,
+		// with or without the sandbox above.
+		rep["verbs"] = opVerbsReport()
+		// An op grant's bounds may hold its runs to a window and a rate
+		// (ops_bounds.go). Said by the box, so the host sends neither to one
+		// that would refuse them as naming no param.
+		rep["grant_run_bounds"] = true
 	}
 	// This machine's encryption key, signed by its own key: what the host
 	// seals an action op's credentials to (ops_seal.go).

client-go/freeagent.go

@@ -60,6 +60,27 @@ type freeagentConfig struct {
 	Cert string // PEM: the client identity this box presents
 	Key  string
 	CA   string // PEM: the root the local agent's server cert must chain to
+	// Mode is the sanitize mode this bastion wrote into the agent's manifest
+	// (faManagedSanitizeMode for every agent it deploys), or "" for an agent
+	// someone runs by hand, whose manifest this bastion did not write.
+	Mode string
+}
+
+// faManagedSanitizeMode is the sanitize mode the bastion writes into every
+// manifest it deploys (faManifestText). Flag mode is the agent's rules pass
+// only: it replaces e-mail addresses, phone numbers, card, bank and ID
+// numbers, IP and MAC addresses and keys, and does not find names, street
+// addresses or dates of birth.
+const faManagedSanitizeMode = "flag"
+
+// faSanitizeNotes says, for each mode a reply can carry, what personal data
+// may still be in the text, in the words the host relays. A mode with no entry
+// here is reported as not known to remove personal data.
+var faSanitizeNotes = map[string]string{
+	"flag": "rules only: e-mail addresses, phone numbers, card, bank and ID numbers, IP and MAC addresses " +
+		"and keys were replaced; names, street addresses and dates of birth were not",
+	"sterilize": "",
+	"none":      "the agent's manifest has no sanitize stage, so nothing in the text was replaced",
 }
 
 func freeagentEnv() (freeagentConfig, bool) {
@@ -252,6 +273,16 @@ func freeagentProbe() (bool, bool) {
 // local agent's own 403 (grant/mask) and 503 (manifest not
 // serving) are refusals with the reason it gave; 404 on read_range is gone.
 func freeagentMap(op string, status int, body []byte, err error) (map[string]any, string, bool) {
+	return freeagentMapMode(op, "", status, body, err)
+}
+
+// freeagentMapMode is freeagentMap for an agent whose manifest this bastion
+// wrote with sanitize mode `mode` ("" when it did not write it). A successful
+// reply says how its text was sanitized: `sanitize_mode` is the mode this
+// bastion wrote, else the one the agent reports, else "unknown"; `sanitized`
+// is true only for "sterilize", the mode that looks for names; and
+// `sanitize_note` says what the mode leaves in.
+func freeagentMapMode(op, mode string, status int, body []byte, err error) (map[string]any, string, bool) {
 	if err != nil {
 		return map[string]any{"state": "offline", "detail": "freeagent: " + err.Error()}, "", true
 	}
@@ -267,7 +298,7 @@ func freeagentMap(op string, status int, body []byte, err error) (map[string]any
 		return nil, "freeagent: malformed reply", false
 	case status == http.StatusOK:
 		payload["state"] = "ok"
-		payload["sanitized"] = true
+		faStampSanitize(payload, mode)
 		return payload, "", true
 	case status == http.StatusNotFound && op == "read_range":
 		return map[string]any{"state": "gone", "detail": reason}, "", true
@@ -286,10 +317,33 @@ func freeagentMap(op string, status int, body []byte, err error) (map[string]any
 	}
 }
 
+// faStampSanitize writes the three sanitize fields onto a successful reply.
+// The mode this bastion wrote wins over the one the reply carries: a reply
+// cannot make text it never sterilized read as sterilized.
+func faStampSanitize(payload map[string]any, mode string) {
+	if mode == "" {
+		mode, _ = payload["sanitize_mode"].(string)
+	}
+	if mode == "" {
+		mode = "unknown"
+	}
+	note, known := faSanitizeNotes[mode]
+	if !known {
+		note = "the local agent did not say how this text was sanitized; it may contain personal data"
+	}
+	payload["sanitize_mode"] = mode
+	payload["sanitized"] = mode == "sterilize"
+	delete(payload, "sanitize_note")
+	if note != "" {
+		payload["sanitize_note"] = note
+	}
+}
+
 // handleDcnSearch — op `dcn_search`: args = <store_id> <json {query, k}>. Gated
 // by the synced grant with the 'search' mask, then forwarded to the
-// local agent's /search. Hits are sanitized windows with document ids the host
-// feeds back to read_range.
+// local agent's /search. Hits are windows of the agent's sanitized text, with
+// document ids the host feeds back to read_range; faStampSanitize says which
+// sanitize mode produced them.
 func handleDcnSearch(ws *websocket.Conn, dir string, args []string) {
 	storeID, a, ok := dcnParseArgs(ws, args)
 	if !ok {
@@ -319,7 +373,7 @@ func handleDcnSearch(ws *websocket.Conn, dir string, args []string) {
 	}
 	status, body, err := freeagentCall(c, http.MethodPost, "/search",
 		map[string]any{"query": q, "k": k})
-	payload, why, ok := freeagentMap("search", status, body, err)
+	payload, why, ok := freeagentMapMode("search", c.Mode, status, body, err)
 	if !ok {
 		dcnRefuse(ws, dir, "search", storeID, "", why)
 		return
@@ -394,7 +448,7 @@ func handleDcnCatalog(ws *websocket.Conn, dir string, args []string) {
 			"state": "offline", "detail": "no catalog yet — the local agent's first sweep has not finished; try again later"})
 		return
 	}
-	payload, why, ok := freeagentMap("catalog", status, body, err)
+	payload, why, ok := freeagentMapMode("catalog", c.Mode, status, body, err)
 	if !ok {
 		dcnRefuse(ws, dir, "catalog", storeID, "kind:"+kind, why)
 		return
@@ -427,7 +481,7 @@ func handleDcnReadRangeDoc(ws *websocket.Conn, dir string, storeID int64, a *dcn
 	}
 	status, body, err := freeagentCall(c, http.MethodPost, "/read_range",
 		map[string]any{"doc": a.Doc, "start": start, "end": end})
-	payload, why, ok := freeagentMap("read_range", status, body, err)
+	payload, why, ok := freeagentMapMode("read_range", c.Mode, status, body, err)
 	if !ok {
 		dcnRefuse(ws, dir, "read_range", storeID, "doc:"+a.Doc, why)
 		return

client-go/freeagent_drop.go

@@ -81,11 +81,19 @@ type faOwner struct {
 	gid  uint32
 	name string
 	home string
-	// plat is what this platform needs beyond the above: nothing on unix, the
+	// plat is what this platform needs beyond the above: nothing on unix (an
+	// owner there is fully described by its uid, gid, name and home), the
 	// owner's SID on Windows (freeagent_drop_windows.go).
 	plat faOwnerPlat
 }
 
+// faOwnerPlat holds the owner's SID, as text, and their bare user name. Only
+// Windows fills it; it stays zero on unix.
+type faOwnerPlat struct {
+	sid  string
+	user string
+}
+
 // faResolveOwner decides the identity for a store's agent. Where this
 // platform drops and the daemon holds machine authority it is the granted
 // folder's owner, refused for the machine's own accounts and for anyone who

client-go/freeagent_drop_unix.go

@@ -20,9 +20,6 @@ func faDropSupported() bool { return true }
 // own 0700 leaf.
 func faRootStateBase() string { return "/var/lib/privatae/freeagent" }
 
-// faOwnerPlat: a unix owner is fully described by its uid, gid, name and home.
-type faOwnerPlat struct{}
-
 // faHandsOverByChown: on unix every file root writes for the child is handed
 // over by changing its owner.
 const faHandsOverByChown = true

client-go/freeagent_drop_windows.go

@@ -71,12 +71,6 @@ func faRootStateBase() string {
 	return filepath.Join(d, "freeagent-state")
 }
 
-// faOwnerPlat holds the owner's SID, as text, and their bare user name.
-type faOwnerPlat struct {
-	sid  string
-	user string
-}
-
 // faHandsOverByChown: on Windows the hand-over is a permission entry on the
 // state leaf, inherited by everything later created in it.
 const faHandsOverByChown = false
@@ -98,15 +92,13 @@ func faResolveDropOwner(root, stateDir string) (*faOwner, error) {
 	if err != nil || sid == nil {
 		return nil, fmt.Errorf("cannot read the folder's owner: %v", err)
 	}
-	name, domain, typ, err := sid.LookupAccount("")
+	// Bounded and cached (sidnames.go); no answer in time refuses the start.
+	acct, err := faWinDropOwnerAccount(sid.String(), winAccounts, faDropSIDLane)
 	if err != nil {
-		return nil, fmt.Errorf("refusing to run the agent: the granted folder's owner (%s) has no account on this machine", sid.String())
+		return nil, err
 	}
-	full := name
-	if domain != "" {
-		full = domain + `\` + name
-	}
-	if why := faWinOwnerRefusal(sid.String(), typ, full); why != "" {
+	name, full := acct.Account, sidAccountLabel(acct)
+	if why := faWinOwnerRefusal(sid.String(), acct.Type, full); why != "" {
 		return nil, errors.New(why)
 	}
 	home := faWinProfileDir(sid.String())
@@ -432,12 +424,6 @@ func (c *winChild) Wait() error {
 	return nil
 }
 
-func faPathOwner(string) (uint32, uint32, error) {
-	return 0, 0, errors.New("ownership is not read as a uid on windows")
-}
-
-func faChown(string, uint32, uint32) error { return nil }
-
 // sysProcAttr: the Windows drop does not go through a credential; faStartChild
 // starts the child under the owner's token itself.
 func (o *faOwner) sysProcAttr() *syscall.SysProcAttr { return nil }
@@ -459,5 +445,4 @@ func faFdSane(f *os.File) error {
 	return nil
 }
 
-func faInfoOwner(os.FileInfo) (uint32, uint32, bool) { return 0, 0, false }
 func faInfoLinks(os.FileInfo) (uint64, bool) { return 0, false }

client-go/freeagent_drop_winrules.go

@@ -5,12 +5,38 @@ package main
 // that gather their inputs are in the tagged file.
 
 import (
+	"fmt"
 	"sort"
 	"strconv"
 	"strings"
+	"time"
 	"unicode/utf16"
 )
 
+// faDropSIDLane bounds the lookup of the granted folder's owner (sidnames.go).
+// The start runs on the store's own goroutine and nothing waits on it, so the
+// budget is generous enough for a slow directory that does answer. It is a
+// bound all the same: the store is marked as ensuring for as long as the start
+// runs, and every reconcile skips it while it is, so a lookup that never
+// returned would hold the store as "installing" for good. Timed out, the store
+// is failed with the reason and the reconcile tries again
+// (faReconcileEvery), by when a late answer is in the cache.
+var faDropSIDLane = sidLane{slots: make(chan struct{}, 1), budget: 10 * time.Second}
+
+// faWinDropOwnerAccount is the folder owner's account from the directory. It
+// fails closed: no account, or no answer in time, and no agent is started.
+func faWinDropOwnerAccount(sid string, dir *sidDirectory, lane sidLane) (sidAccount, error) {
+	acct, outcome := dir.resolve(sid, lane)
+	switch outcome {
+	case sidTimedOut:
+		return sidAccount{}, fmt.Errorf("refusing to run the agent for now: this machine's account directory did not answer "+
+			"for the granted folder's owner (%s) within %v; it is tried again at the next check", sid, lane.budget)
+	case sidUnmapped:
+		return sidAccount{}, fmt.Errorf("refusing to run the agent: the granted folder's owner (%s) has no account on this machine", sid)
+	}
+	return acct, nil
+}
+
 // winSidTypeUser is SidTypeUser: the SID names a person's account rather than
 // a group, an alias or a well-known identity.
 const winSidTypeUser = 1

client-go/freeagent_managed.go

@@ -500,7 +500,7 @@ func faManifestText(id, cn string, terms *dcnTerms, sweep, recipient, embedModel
 		fmt.Fprintf(&b, "  exclude: [%s]\n", strings.Join(ex, ", "))
 	}
 	fmt.Fprintf(&b, "mask: [%s]\nstages: [%s]\n", strings.Join(mask, ", "), strings.Join(stages, ", "))
-	b.WriteString("sanitize:\n  mode: flag\n")
+	fmt.Fprintf(&b, "sanitize:\n  mode: %s\n", faManagedSanitizeMode)
 	if containsStr(stages, "embed") {
 		fmt.Fprintf(&b, "embed:\n  model: %s\n  dim: %d\n  chunk: {size_tokens: 512, overlap_tokens: 64}\n", yamlStr(embedModel), embedDim)
 	}
@@ -677,7 +677,8 @@ func faManagedConfig(storeID int64) (freeagentConfig, bool) {
 	if err != nil || port <= 0 {
 		return freeagentConfig{}, false
 	}
-	return freeagentConfig{Addr: "127.0.0.1:" + strconv.Itoa(port), Cert: inst.Cert, Key: inst.Key, CA: inst.Cert}, true
+	return freeagentConfig{Addr: "127.0.0.1:" + strconv.Itoa(port), Cert: inst.Cert, Key: inst.Key, CA: inst.Cert,
+		Mode: faManagedSanitizeMode}, true
 }
 
 // ── ensure / reconcile ──────────────────────────────────────────────────────

client-go/freeagent_test.go

@@ -190,8 +190,11 @@ func TestFreeagentProbeAndEnv(t *testing.T) {
 
 func TestFreeagentMap(t *testing.T) {
 	m, why, ok := freeagentMap("search", 200, []byte(`{"hits":[]}`), nil)
-	if !ok || m["state"] != "ok" || m["sanitized"] != true || why != "" {
-		t.Fatalf("200 → ok/sanitized, got %v %q", m, why)
+	if !ok || m["state"] != "ok" || why != "" {
+		t.Fatalf("200 → ok, got %v %q", m, why)
+	}
+	if m["sanitized"] != false || m["sanitize_mode"] != "unknown" || m["sanitize_note"] == nil {
+		t.Fatalf("a reply that names no mode is not called sanitized, got %v", m)
 	}
 	m, _, ok = freeagentMap("read_range", 404, []byte(`{"error":"no such document"}`), nil)
 	if !ok || m["state"] != "gone" {
@@ -298,3 +301,45 @@ func TestDcnOpArgsCarryFreeagentFields(t *testing.T) {
 		t.Fatalf("args not carried: %+v", a)
 	}
 }
+
+// The sanitize fields say what was done to the text. An agent this bastion
+// deployed runs in flag mode, which leaves names in: its replies are never
+// called sanitized, whatever the reply itself claims. An agent run by hand is
+// reported in the mode it names, and only sterilize counts as sanitized.
+func TestASanitizeStampNamesTheModeThatRan(t *testing.T) {
+	for _, c := range []struct {
+		name, mode, body, wantMode string
+		wantSanitized, wantNote    bool
+	}{
+		{"deployed, reply silent", faManagedSanitizeMode, `{"hits":[]}`, "flag", false, true},
+		{"deployed, reply claims sterilize", faManagedSanitizeMode, `{"hits":[],"sanitize_mode":"sterilize","sanitized":true}`, "flag", false, true},
+		{"by hand, flag", "", `{"hits":[],"sanitize_mode":"flag"}`, "flag", false, true},
+		{"by hand, sterilize", "", `{"hits":[],"sanitize_mode":"sterilize","sanitize_note":"x"}`, "sterilize", true, false},
+		{"by hand, silent", "", `{"hits":[],"sanitized":true}`, "unknown", false, true},
+		{"by hand, no sanitize stage", "", `{"hits":[],"sanitize_mode":"none"}`, "none", false, true},
+		{"by hand, a mode this build does not know", "", `{"hits":[],"sanitize_mode":"other"}`, "other", false, true},
+	} {
+		for _, op := range []string{"search", "read_range", "catalog"} {
+			m, why, ok := freeagentMapMode(op, c.mode, 200, []byte(c.body), nil)
+			if !ok || why != "" || m["state"] != "ok" {
+				t.Fatalf("%s/%s: %v %q", c.name, op, m, why)
+			}
+			note, hasNote := m["sanitize_note"].(string)
+			if m["sanitize_mode"] != c.wantMode || m["sanitized"] != c.wantSanitized || hasNote != c.wantNote || (hasNote && note == "") {
+				t.Errorf("%s/%s: got mode %v sanitized %v note %q", c.name, op, m["sanitize_mode"], m["sanitized"], note)
+			}
+		}
+	}
+	if !strings.Contains(faSanitizeNotes["flag"], "names") {
+		t.Fatal("the flag note must say names are left in")
+	}
+}
+
+// The mode the stamp reports for a deployed agent is the mode the bastion
+// writes into its manifest.
+func TestTheDeployedManifestUsesTheStampedMode(t *testing.T) {
+	text := faManifestText("s1", "cn", &dcnTerms{Root: "/tmp/x", Mask: []string{"search"}}, "", "", "m", 8)
+	if !strings.Contains(text, "sanitize:\n  mode: "+faManagedSanitizeMode+"\n") {
+		t.Fatalf("manifest sanitize block:\n%s", text)
+	}
+}

client-go/fuzz_grants_test.go

@@ -345,3 +345,51 @@ func FuzzADeviceApprovalIsReadOneWayOnly(f *testing.F) {
 		}
 	})
 }
+
+// FuzzAGrantsRunBoundsHoldWhateverTheHostSends: whatever bounds text a host
+// sends, a run bound the machine accepts is one it can hold — a window has a
+// length, a zone and a day, a rate names a limit inside the ceilings — and
+// judging a run against it at any moment answers, never panics, and never
+// lets a run inside an interval or past a day's count.
+func FuzzAGrantsRunBoundsHoldWhateverTheHostSends(f *testing.F) {
+	m := &opManifest{Params: map[string]opParam{"count": {Type: "integer", Min: ptr64(1), Max: ptr64(5)}}}
+	for _, s := range []string{
+		`{"_window":{"days":["mon"],"from":"22:00","to":"06:00","tz":"Europe/Berlin"}}`,
+		`{"_rate":{"per_day":2,"min_interval_s":600},"count":{"max":2}}`,
+		`{"_window":null,"_rate":null}`, `{"_x":1}`, `{"_window":{"from":"02:00","to":"02:00"}}`,
+	} {
+		f.Add(s, int64(1761739200), int64(300))
+	}
+	f.Fuzz(func(t *testing.T, raw string, at, gap int64) {
+		b, err := parseGrantBounds(m, raw)
+		if err != nil {
+			return
+		}
+		if w := b.Window; w != nil {
+			if w.from == w.to || w.loc == nil || w.day == [7]bool{} {
+				t.Fatalf("accepted a window that is no window: %s", raw)
+			}
+		}
+		if r := b.Rate; r != nil {
+			if (r.PerDay == nil && r.MinInterval == nil) || (r.PerDay != nil && (*r.PerDay < 1 || *r.PerDay > opGrantUsesMax)) ||
+				(r.MinInterval != nil && (*r.MinInterval < 1 || *r.MinInterval > opRateMinInterval)) {
+				t.Fatalf("accepted a rate outside the ceilings: %s", raw)
+			}
+		}
+		now := time.Unix(at%(1<<40), 0)
+		last := now.Add(-time.Duration(gap%(1<<30)) * time.Second)
+		if last.After(now) {
+			last = now
+		}
+		why, _, _ := opNotNowWhy(b, []time.Time{last}, now)
+		if b.Window == nil && b.Rate != nil && b.Rate.MinInterval != nil &&
+			now.Sub(last) < time.Duration(*b.Rate.MinInterval)*time.Second && why == "" {
+			t.Fatalf("a run inside the interval was let through: %s", raw)
+		}
+		if b.Window == nil && b.Rate != nil && b.Rate.PerDay != nil && *b.Rate.PerDay == 1 && now.Sub(last) < 24*time.Hour && why == "" {
+			t.Fatalf("a second run in a day was let through: %s", raw)
+		}
+	})
+}
+
+func ptr64(v int64) *int64 { return &v }

client-go/fuzz_observe_test.go

@@ -0,0 +1,123 @@
+package main
+
+// Fuzz the walls a host's value meets in observe:folders, observe:journal,
+// observe:cert and observe:logins. Whatever the input, a path that passes is
+// absolute, clean and outside the kernel's trees; a unit that passes is in
+// systemd's alphabet and shaped like no flag; a server name that passes is a
+// DNS name; and the sshd parser names only addresses as sources and bounded
+// names as users, whatever a remote side put in a user name.
+
+import (
+	"net/netip"
+	"path"
+	"strings"
+	"testing"
+)
+
+func FuzzFolderPathIsAbsoluteCleanAndOffTheKernelTrees(f *testing.F) {
+	for _, s := range []string{"/", "/var/lib", "/var//lib/", "/proc", "//proc/1", "/a/../proc", "/a..b",
+		"rel", "/dev/shm", "/x\x00y", "/\xff", "/srv/./data", "/sys/", "/devices"} {
+		f.Add(s)
+	}
+	f.Fuzz(func(t *testing.T, in string) {
+		p, err := checkFolderPath(in)
+		if err != nil {
+			return
+		}
+		if !strings.HasPrefix(p, "/") || path.Clean(p) != p || strings.Contains(p, "..") {
+			t.Fatalf("%q passed as %q", in, p)
+		}
+		for _, r := range p {
+			if r < 0x20 || r == 0x7f {
+				t.Fatalf("%q passed with a control character", in)
+			}
+		}
+		for _, root := range folderRefusedRoots {
+			if p == root || strings.HasPrefix(p, root+"/") {
+				t.Fatalf("%q passed into %s", in, root)
+			}
+		}
+	})
+}
+
+func FuzzUnitNameIsSystemdsAlphabet(f *testing.F) {
+	for _, s := range []string{"example", "getty@tty1", `dev-disk-by\x2dlabel-data.device`, "-.mount", "--unit=x",
+		"worker@.service", "a@b@c", "x;id", "data.mount", ".service", "@x"} {
+		f.Add(s)
+	}
+	f.Fuzz(func(t *testing.T, in string) {
+		u, err := normalizeUnitName(in)
+		if err != nil {
+			return
+		}
+		if !unitNameRe.MatchString(u) || len(u) > unitNameMax || !hasUnitSuffix(u) {
+			t.Fatalf("%q passed as %q", in, u)
+		}
+		if strings.HasPrefix(u, "-") || strings.HasPrefix(u, ".") || strings.HasPrefix(u, "@") {
+			t.Fatalf("%q passed as a flag-shaped %q", in, u)
+		}
+		if u != in && u != in+".service" {
+			t.Fatalf("%q became %q: the name is kept or given .service, nothing else", in, u)
+		}
+	})
+}
+
+func FuzzCertArgsAreAPortAndADNSName(f *testing.F) {
+	for _, s := range [][2]string{{"443", "box-a.example"}, {"0", ""}, {"65536", "x"}, {"443", "-x"},
+		{"443", "a..b"}, {"8443", "192.0.2.10"}, {" 443 ", "A.EXAMPLE"}, {"443", "a_b"}} {
+		f.Add(s[0], s[1])
+	}
+	f.Fuzz(func(t *testing.T, port, name string) {
+		p, n, err := parseCertArgs([]string{port, name})
+		if err != nil {
+			return
+		}
+		if p < 1 || p > 65535 || !serverNameRe.MatchString(n) || n != strings.ToLower(n) ||
+			strings.HasPrefix(n, "-") || strings.HasPrefix(n, ".") || strings.Contains(n, "..") {
+			t.Fatalf("(%q, %q) passed as (%d, %q)", port, name, p, n)
+		}
+	})
+}
+
+func FuzzSSHLoginsNameOnlyAddressesAndBoundedNames(f *testing.F) {
+	f.Add(fxSSHD)
+	f.Add("2026-10-29T10:00:01+00:00 h sshd[1]: Invalid user a from 1.2.3.4 port 1 from 203.0.113.5 port 2\n")
+	f.Add("2026-10-29T10:00:01+00:00 h sshd[1]: Failed password for x from evil/host port 1 ssh2\n")
+	f.Add("x y z: Accepted publickey for \x00 from ::1 port 9 ssh2: K\n")
+	f.Fuzz(func(t *testing.T, text string) {
+		tally := parseSSHLogins(text)
+		var rep loginsReport
+		tally.fill(&rep)
+		sum := 0
+		for src, n := range tally.bySource {
+			sum += n
+			if _, err := netip.ParseAddr(src); err != nil && !loginHostRe.MatchString(src) {
+				t.Fatalf("source %q is not an address or a host name", src)
+			}
+		}
+		if sum != rep.FailedTotal {
+			t.Fatalf("sources count %d, total %d", sum, rep.FailedTotal)
+		}
+		users := 0
+		for name, u := range tally.byUser {
+			users += u.Count
+			if len(name) > loginNameMax || loginName(name) != name {
+				t.Fatalf("user %q is not a bounded name", name)
+			}
+		}
+		if users != rep.FailedTotal {
+			t.Fatalf("users count %d, total %d", users, rep.FailedTotal)
+		}
+		if len(rep.Accepted) > loginsTopN || rep.AcceptedTotal < len(rep.Accepted) {
+			t.Fatalf("accepted %d of %d", len(rep.Accepted), rep.AcceptedTotal)
+		}
+		for _, a := range rep.Accepted {
+			if _, err := netip.ParseAddr(a.Source); err != nil && !loginHostRe.MatchString(a.Source) {
+				t.Fatalf("accepted source %q", a.Source)
+			}
+			if loginName(a.User) != a.User || loginMethod(a.Method) != a.Method {
+				t.Fatalf("accepted %+v", a)
+			}
+		}
+	})
+}

client-go/go.mod

@@ -1,14 +1,14 @@
 module git.privatae.ai/privatae/privatae-bastion
 
-go 1.25.0
+go 1.26.0
 
 require (
 	filippo.io/age v1.3.2
 	fyne.io/systray v1.12.2
 	github.com/creack/pty v1.1.24
 	github.com/gorilla/websocket v1.5.3
-	golang.org/x/crypto v0.55.0
-	golang.org/x/sys v0.47.0
+	golang.org/x/crypto v0.58.0
+	golang.org/x/sys v0.49.0
 	golang.zx2c4.com/wireguard v0.0.0-20260522210424-ecfc5a8d5446
 	gopkg.in/yaml.v3 v3.0.1
 	rsc.io/qr v0.2.0
@@ -18,7 +18,7 @@ require (
 	filippo.io/hpke v0.4.0 // indirect
 	github.com/godbus/dbus/v5 v5.1.0 // indirect
 	github.com/google/btree v1.1.2 // indirect
-	golang.org/x/net v0.57.0 // indirect
+	golang.org/x/net v0.61.0 // indirect
 	golang.org/x/time v0.7.0 // indirect
 	golang.zx2c4.com/wintun v0.0.0-20230126152724-0fa3db229ce2 // indirect
 	gvisor.dev/gvisor v0.0.0-20250503011706-39ed1f5ac29c // indirect

client-go/go.sum

@@ -14,14 +14,14 @@ github.com/google/btree v1.1.2 h1:xf4v41cLI2Z6FxbKm+8Bu+m8ifhj15JuZ9sa0jZCMUU=
 github.com/google/btree v1.1.2/go.mod h1:qOPhT0dTNdNzV6Z/lhRX0YXUafgPLFUh+gZMl761Gm4=
 github.com/gorilla/websocket v1.5.3 h1:saDtZ6Pbx/0u+bgYQ3q96pZgCzfhKXGPqt7kZ72aNNg=
 github.com/gorilla/websocket v1.5.3/go.mod h1:YR8l580nyteQvAITg2hZ9XVh4b55+EU/adAjf1fMHhE=
-golang.org/x/crypto v0.55.0 h1:+KWHjbgOaAQ66dh/YlkZKHlz9ZUlq61AFirAR9ntP8M=
-golang.org/x/crypto v0.55.0/go.mod h1:uq0V9dE/fzQuJtbnL+2EhWOE63vo164FY8xqEnV9xis=
-golang.org/x/net v0.57.0 h1:K5+3DljvIuDG9/Jv9rvyMywYNFCQ9RSUY6OOTTkT+tE=
-golang.org/x/net v0.57.0/go.mod h1:KpXc8iv+r3XplLAG/f7Jsf9RPszJzdR0f58q9vGOuEU=
-golang.org/x/sys v0.47.0 h1:o7XGOvZQCADBQQ4Y7VNq2dRWQR7JmOUW8Kxx4ZsNgWs=
-golang.org/x/sys v0.47.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
-golang.org/x/term v0.45.0 h1:NwWyBmoJCbfTHpxrWoZ9C6/VxOf7ic219I8xZZFdrf0=
-golang.org/x/term v0.45.0/go.mod h1:9aqxs0blBcrm/n0L9QW0aRVD+ktan8ssZromtqJC43w=
+golang.org/x/crypto v0.58.0 h1:COkYLr4k7nDI4r6QtMML2AfzeIswWHAJthLS6K/M/54=
+golang.org/x/crypto v0.58.0/go.mod h1:Xh+kl5A+M0gMiWDytH2Hhr5Z7WbNCpv/eB/EFgfSk+w=
+golang.org/x/net v0.61.0 h1:RR+6j/BTBCrggOLugxa+66R8FvoN6MHlzDlZIbN9YFw=
+golang.org/x/net v0.61.0/go.mod h1:WgDCOTH8iwtB66T4/GfI94AHo/hhdb7Hah9NDseQfiI=
+golang.org/x/sys v0.49.0 h1:XbzkgYJHdqh/8m2Uu0W/dQv8nktxx4BFHp1M0gROTXA=
+golang.org/x/sys v0.49.0/go.mod h1:hNLxWAXmnKAxqDtdwIYC4bM9oQPEecfsnNMuSxOs3og=
+golang.org/x/term v0.47.0 h1:cpyOwQkr2Nzxe2p6KnrTG7Ei2wtzjsuSjhpXXepZ/BM=
+golang.org/x/term v0.47.0/go.mod h1:efAfAuG6RKABvg1ArGebRkvX6LXul14SM67o2e4mUCY=
 golang.org/x/time v0.7.0 h1:ntUhktv3OPE6TgYxXWv9vKvUSJyIFJlyohwbkEwPrKQ=
 golang.org/x/time v0.7.0/go.mod h1:3BpzKBy/shNhVucY/MWOyx10tF3SFh9QdLuxbVysPQM=
 golang.zx2c4.com/wintun v0.0.0-20230126152724-0fa3db229ce2 h1:B82qJJgjvYKsXS9jeunTOisW56dUokqW/FOteYJJ/yg=

client-go/goosbuild_test.go

@@ -0,0 +1,104 @@
+package main
+
+import (
+	"go/ast"
+	"go/build"
+	"go/parser"
+	"go/token"
+	"os"
+	"sort"
+	"strings"
+	"testing"
+)
+
+// EVERY PLATFORM FUNCTION EXISTS ON EVERY TARGET THAT BUILDS THIS PACKAGE.
+//
+// A function with one file per platform (contextPeerUID in contextsock_linux.go
+// and contextsock_darwin.go) and a caller in a file every unix builds
+// (contextsock_unix.go) leaves any other unix with an undefined name, and
+// nothing in CI cross-builds one: `GOOS=freebsd go build` failed that way.
+// The release targets are linux, darwin and windows (deploy/build.sh);
+// freebsd and openbsd are not released, but they build, with what has no
+// implementation there refusing (contextsock_other.go and the other
+// *_other.go files).
+//
+// This asks the toolchain which files each target compiles (go/build's
+// MatchFile) and checks that every package function called from those files
+// is declared in them. It is not a type check, and it does not compile; it
+// catches the missing-platform-file mistake, which is the one that happened,
+// on the platform the tests already run on.
+func TestEveryCalledPlatformFunctionIsDeclaredOnEveryTarget(t *testing.T) {
+	entries, err := os.ReadDir(".")
+	if err != nil {
+		t.Fatal(err)
+	}
+	type parsed struct {
+		name string
+		file *ast.File
+	}
+	var files []parsed
+	declaredAnywhere := map[string]bool{}
+	fset := token.NewFileSet()
+	for _, e := range entries {
+		n := e.Name()
+		if e.IsDir() || !strings.HasSuffix(n, ".go") || strings.HasSuffix(n, "_test.go") {
+			continue
+		}
+		f, err := parser.ParseFile(fset, n, nil, parser.SkipObjectResolution)
+		if err != nil {
+			t.Fatalf("%s: %v", n, err)
+		}
+		files = append(files, parsed{n, f})
+		for _, d := range f.Decls {
+			if fd, ok := d.(*ast.FuncDecl); ok && fd.Recv == nil {
+				declaredAnywhere[fd.Name.Name] = true
+			}
+		}
+	}
+	if len(files) < 50 {
+		t.Fatalf("parsed only %d files; is the test running in the package directory?", len(files))
+	}
+	for _, goos := range []string{"linux", "darwin", "windows", "freebsd", "openbsd"} {
+		ctx := build.Default
+		ctx.GOOS = goos
+		ctx.CgoEnabled = false
+		declared := map[string]bool{}
+		var matched []parsed
+		for _, p := range files {
+			ok, err := ctx.MatchFile(".", p.name)
+			if err != nil {
+				t.Fatalf("%s: %v", p.name, err)
+			}
+			if !ok {
+				continue
+			}
+			matched = append(matched, p)
+			for _, d := range p.file.Decls {
+				if fd, ok := d.(*ast.FuncDecl); ok && fd.Recv == nil {
+					declared[fd.Name.Name] = true
+				}
+			}
+		}
+		missing := map[string]string{}
+		for _, p := range matched {
+			ast.Inspect(p.file, func(x ast.Node) bool {
+				c, ok := x.(*ast.CallExpr)
+				if !ok {
+					return true
+				}
+				if id, ok := c.Fun.(*ast.Ident); ok && declaredAnywhere[id.Name] && !declared[id.Name] {
+					missing[id.Name] = p.name
+				}
+				return true
+			})
+		}
+		var out []string
+		for fn, where := range missing {
+			out = append(out, fn+" (called in "+where+")")
+		}
+		sort.Strings(out)
+		if len(out) > 0 {
+			t.Errorf("GOOS=%s: called but not declared for this target: %s", goos, strings.Join(out, ", "))
+		}
+	}
+}

client-go/harness_creds_linux_test.go

@@ -0,0 +1,44 @@
+package main
+
+// The agent's credential for a harness run is a file in a folder mounted into
+// the container. No other local user may reach either: as root the bastion
+// hands both to the container's user, owner-only; as anyone else it opens them
+// for that user, inside a folder only the bastion can enter.
+
+import (
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"testing"
+)
+
+func TestTheHarnessCredentialIsReadableOnlyByTheContainersUser(t *testing.T) {
+	// A stand-in runtime: on `run` it records the mode and owner of the folder
+	// the credential folder sits in, of that folder and of the file, then
+	// fails, so the run stops there.
+	bin := t.TempDir()
+	record := filepath.Join(bin, "record")
+	rt := filepath.Join(bin, "runtime")
+	script := "#!/bin/sh\n[ \"$1\" = run ] || exit 1\nfor a; do case \"$a\" in *:/home/privatae/.claude) d=${a%:/home/privatae/.claude};; esac; done\n" +
+		"[ -n \"$d\" ] && stat -c '%a %u %g' \"$(dirname \"$d\")\" \"$d\" \"$d/.credentials.json\" > " + record + "\nexit 1\n"
+	if err := os.WriteFile(rt, []byte(script), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	spec := harnessSpec{HarnessID: "h1", Image: "harness:test", AgentCreds: `{"claudeAiOauth":{}}`}
+	if _, err := driveHarnessAgent(rt, t.TempDir(), spec, 1); err == nil {
+		t.Fatal("the stand-in runtime fails the start; the run must report it")
+	}
+	got, err := os.ReadFile(record)
+	if err != nil {
+		t.Fatalf("the credential folder was not mounted: %v", err)
+	}
+	self := strconv.Itoa(os.Geteuid()) + " " + strconv.Itoa(os.Getegid())
+	want := "700 " + self + "\n777 " + self + "\n666 " + self
+	if os.Geteuid() == 0 {
+		want = "700 " + self + "\n700 10005 10005\n600 10005 10005" // the harness image's user
+	}
+	if strings.TrimSpace(string(got)) != want {
+		t.Fatalf("enclosing folder, folder, then file: want %q, got %q", want, strings.TrimSpace(string(got)))
+	}
+}

client-go/harness_gitrun_bounds_linux_test.go

@@ -0,0 +1,154 @@
+package main
+
+// A git run's copy is bounded by how many entries a tree holds as well as by
+// its bytes; it never hands git the run's token in an argument or a config
+// file; and it uses nothing of the agent's tree while the agent's container
+// may still be running.
+
+import (
+	"encoding/base64"
+	"net/http"
+	"net/http/cgi"
+	"net/http/httptest"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// Many empty files cost nothing in bytes; each still counts against the
+// copy, so a tree of them is refused and nothing is pushed.
+func TestAGitRunRefusesATreeOfTooManyEntries(t *testing.T) {
+	f := newGitRunFixture(t)
+	f.setAgent(t, `cd "$1"
+i=0; while [ $i -lt 40 ]; do mkdir d$i; j=0; while [ $j -lt 100 ]; do : > d$i/f$j; j=$((j+1)); done; i=$((i+1)); done
+`)
+	saved := maxCopiedTreeBytes
+	maxCopiedTreeBytes = 1 << 20 // 4000 entries at 512 bytes each are 2 MB
+	t.Cleanup(func() { maxCopiedTreeBytes = saved })
+	res, code := f.run(t)
+	if code == 0 || res["ok"] != false {
+		t.Fatalf("a tree of 4000 empty files was copied under a 1 MiB limit: %v", res)
+	}
+	if e, _ := res["error"].(string); !strings.Contains(e, "too large") {
+		t.Errorf("the error does not say the tree is too large: %q", e)
+	}
+	if branchPushed(f) {
+		t.Error("the branch was pushed")
+	}
+}
+
+// A container the runtime does not confirm removed fails the run before any
+// of its tree is copied.
+func TestAGitRunFailsWhenTheAgentsContainerIsNotRemoved(t *testing.T) {
+	f := newGitRunFixture(t)
+	f.setAgent(t, "cd \"$1\" && echo new > edit.txt\n")
+	docker := filepath.Join(filepath.Dir(f.agentFile), "docker")
+	b, err := os.ReadFile(docker)
+	if err != nil {
+		t.Fatal(err)
+	}
+	failing := strings.Replace(string(b), "ps|pull|rm) exit 0;;", "ps|pull) exit 0;;\nrm) echo 'Error: removal stuck' >&2; exit 1;;", 1)
+	if failing == string(b) {
+		t.Fatal("the stand-in runtime changed shape")
+	}
+	if err := os.WriteFile(docker, []byte(failing), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	res, code := f.run(t)
+	if code == 0 || res["ok"] != false {
+		t.Fatalf("a run whose container was not removed reads ok: %v", res)
+	}
+	if e, _ := res["error"].(string); !strings.Contains(e, "could not be removed") || !strings.Contains(e, "removal stuck") {
+		t.Errorf("the error does not say the container was not removed: %q", e)
+	}
+	if branchPushed(f) {
+		t.Error("the branch was pushed")
+	}
+}
+
+// Against a server that needs the token, the run clones and pushes with the
+// token in no git argument and in no clone's config.
+func TestAGitRunKeepsTheTokenOutOfEveryArgumentAndConfig(t *testing.T) {
+	f := newGitRunFixture(t)
+	f.setAgent(t, "cd \"$1\" && echo new > edit.txt\n")
+	realGit, err := exec.LookPath("git")
+	if err != nil {
+		t.Skip("no git")
+	}
+	want := "Basic " + base64.StdEncoding.EncodeToString([]byte("harness:"+gitRunToken))
+	backend := &cgi.Handler{Path: realGit, Args: []string{"http-backend"}, Root: "/git",
+		Env: []string{"GIT_PROJECT_ROOT=" + filepath.Dir(f.origin), "GIT_HTTP_EXPORT_ALL=1",
+			"REMOTE_USER=harness", "GIT_CONFIG_NOSYSTEM=1", "HOME=" + t.TempDir()}}
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.Header.Get("Authorization") != want {
+			w.Header().Set("WWW-Authenticate", `Basic realm="test"`)
+			http.Error(w, "auth", http.StatusUnauthorized)
+			return
+		}
+		backend.ServeHTTP(w, r)
+	}))
+	t.Cleanup(srv.Close)
+
+	logs := t.TempDir()
+	argvLog, cfgLog := filepath.Join(logs, "argv"), filepath.Join(logs, "config")
+	bin := t.TempDir()
+	wrapper := "#!/bin/sh\nprintf '%s\\n' \"$*\" >> " + argvLog + "\n" +
+		"prev=; for a; do if [ \"$prev\" = -C ] && [ -f \"$a/.git/config\" ]; then cat \"$a/.git/config\" >> " + cfgLog + "; fi; prev=$a; done\n" +
+		"exec " + realGit + " \"$@\"\n"
+	if err := os.WriteFile(filepath.Join(bin, "git"), []byte(wrapper), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	t.Setenv("PATH", bin+string(os.PathListSeparator)+os.Getenv("PATH"))
+
+	repoURL := strings.Replace(srv.URL, "http://", "http://harness:"+gitRunToken+"@", 1) + "/git/" + filepath.Base(f.origin)
+	b, code := runGit(harnessSpec{Mode: "git", HarnessID: "h1", Image: "harness:test",
+		RepoURL: repoURL, Branch: "agent-work", MaxMinutes: 1})
+	if code != 0 {
+		t.Fatalf("the run failed: %s", b)
+	}
+	if !branchPushed(f) {
+		t.Fatal("the branch is not in the origin")
+	}
+	argv, _ := os.ReadFile(argvLog)
+	if !strings.Contains(string(argv), "push origin agent-work") {
+		t.Fatalf("the stand-in git did not see the push:\n%s", argv)
+	}
+	cfg, _ := os.ReadFile(cfgLog)
+	if len(cfg) == 0 {
+		t.Fatal("no clone's config was read")
+	}
+	for name, body := range map[string][]byte{"an argument": argv, "a clone's config": cfg} {
+		for _, secret := range []string{gitRunToken, strings.TrimPrefix(want, "Basic ")} {
+			if strings.Contains(string(body), secret) {
+				t.Errorf("the token is in %s:\n%s", name, body)
+			}
+		}
+	}
+}
+
+// Every git call carries the harness identity: nothing of this machine's user
+// or host name reaches the clone the agent is handed, and the commit is the
+// harness's.
+func TestAGitRunsGitCarriesTheHarnessIdentity(t *testing.T) {
+	f := newGitRunFixture(t)
+	reflog := filepath.Join(t.TempDir(), "reflog")
+	f.setAgent(t, "cp \"$1/.git/logs/HEAD\" "+reflog+" && cd \"$1\" && echo new > edit.txt\n")
+	res, code := f.run(t)
+	if code != 0 {
+		t.Fatalf("the run failed: %v", res)
+	}
+	b, err := os.ReadFile(reflog)
+	if err != nil {
+		t.Fatal(err)
+	}
+	host, _ := os.Hostname()
+	if !strings.Contains(string(b), " harness <harness@privatae.ai> ") || (host != "" && strings.Contains(string(b), host)) {
+		t.Errorf("the agent's clone records another identity:\n%s", b)
+	}
+	who := strings.TrimSpace(testGit(t, f.origin, "log", "-1", "--format=%an <%ae> / %cn <%ce>", "agent-work"))
+	if who != "harness <harness@privatae.ai> / harness <harness@privatae.ai>" {
+		t.Errorf("the commit is by %q", who)
+	}
+}

client-go/harness_gitrun_fixes_linux_test.go

@@ -0,0 +1,360 @@
+package main
+
+// What a git run hands the agent, how it ends when its commit cannot be made,
+// and the git-lfs filter it supplies itself.
+
+import (
+	"bufio"
+	"bytes"
+	"fmt"
+	"io"
+	"io/fs"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"testing"
+)
+
+const gitRunToken = "tok3n0f7hisrun5a1e"
+
+// The agent's clone names the repository without the run's token, and the
+// token is in no file of the tree it is handed; the run still pushes.
+func TestAGitRunHandsTheAgentNoToken(t *testing.T) {
+	f := newGitRunFixture(t)
+	snap := filepath.Join(t.TempDir(), "agent-saw")
+	f.setAgent(t, "cp -a \"$1\" "+snap+" && cd \"$1\" && echo new > edit.txt\n")
+	b, code := runGit(harnessSpec{Mode: "git", HarnessID: "h1", Image: "harness:test",
+		RepoURL: "file://harness:" + gitRunToken + "@" + f.origin, Branch: "agent-work", MaxMinutes: 1})
+	if code != 0 {
+		t.Fatalf("the run failed: %s", b)
+	}
+	if bytes.Contains(b, []byte(gitRunToken)) {
+		t.Errorf("the result carries the token: %s", b)
+	}
+	n := 0
+	err := filepath.WalkDir(snap, func(p string, d fs.DirEntry, err error) error {
+		if err != nil || !d.Type().IsRegular() {
+			return err
+		}
+		n++
+		body, err := os.ReadFile(p)
+		if err == nil && bytes.Contains(body, []byte(gitRunToken)) {
+			t.Errorf("the token is in the agent's tree, in %s", p[len(snap):])
+		}
+		return err
+	})
+	if err != nil || n == 0 {
+		t.Fatalf("the agent's tree was not read (%d files): %v", n, err)
+	}
+	if got := testGit(t, f.origin, "show", "agent-work:edit.txt"); got != "new\n" {
+		t.Errorf("edit.txt = %q", got)
+	}
+}
+
+func TestRepoURLPartsAndLFSPins(t *testing.T) {
+	bare, secret, err := repoURLParts("https://host-user:" + gitRunToken + "@git.example.invalid/o/r.git")
+	if err != nil || bare != "https://git.example.invalid/o/r.git" || secret != gitRunToken {
+		t.Errorf("parts = %q, %q, %v", bare, secret, err)
+	}
+	if bare, secret, err := repoURLParts("/srv/x/origin.git"); err != nil || bare != "/srv/x/origin.git" || secret != "" {
+		t.Errorf("a path = %q, %q, %v", bare, secret, err)
+	}
+	pins := strings.Join(lfsEndpointPins("https://u:"+gitRunToken+"@git.example.invalid/o/r"), " ")
+	for _, k := range []string{"lfs.url", "lfs.pushurl", "remote.origin.lfsurl", "remote.origin.lfspushurl"} {
+		if !strings.Contains(pins, k+"=https://u:"+gitRunToken+"@git.example.invalid/o/r.git/info/lfs") {
+			t.Errorf("%s is not pinned to the repository's own LFS address: %s", k, pins)
+		}
+	}
+	if p := lfsEndpointPins("/srv/x/origin.git"); p != nil {
+		t.Errorf("a path got pins: %v", p)
+	}
+}
+
+// standInGit puts a git on PATH that fails on the subcommand fail and is the
+// real git otherwise.
+func standInGit(t *testing.T, fail string) {
+	t.Helper()
+	real, err := exec.LookPath("git")
+	if err != nil {
+		t.Skip("no git")
+	}
+	bin := t.TempDir()
+	script := "#!/bin/sh\nfor a; do [ \"$a\" = " + fail + " ] && { echo 'stand-in: " + fail + " failed' >&2; exit 128; }; done\nexec " + real + " \"$@\"\n"
+	if err := os.WriteFile(filepath.Join(bin, "git"), []byte(script), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	t.Setenv("PATH", bin+string(os.PathListSeparator)+os.Getenv("PATH"))
+}
+
+func branchPushed(f gitRunFixture) bool {
+	return exec.Command("git", "--git-dir", f.origin, "rev-parse", "--verify", "-q", "agent-work").Run() == nil
+}
+
+// A commit that fails is the run failing: nothing is pushed.
+func TestAGitRunWhoseCommitFailsPushesNothing(t *testing.T) {
+	f := newGitRunFixture(t)
+	f.setAgent(t, "cd \"$1\" && echo new > edit.txt\n")
+	standInGit(t, "commit")
+	res, code := f.run(t)
+	if code == 0 || res["ok"] != false {
+		t.Errorf("a run whose commit failed reads ok: code %d, %v", code, res)
+	}
+	if e, _ := res["error"].(string); !strings.Contains(e, "commit") || !strings.Contains(e, "nothing was pushed") {
+		t.Errorf("the error does not say the commit failed: %q", e)
+	}
+	if branchPushed(f) {
+		t.Error("the branch was pushed")
+	}
+}
+
+// A run that changed nothing is ok, says so, and pushes no branch.
+func TestAGitRunThatChangesNothingPushesNothing(t *testing.T) {
+	f := newGitRunFixture(t)
+	f.setAgent(t, "true\n")
+	res, code := f.run(t)
+	if code != 0 || res["ok"] != true || res["changed"] != false {
+		t.Errorf("want ok with changed false: code %d, %v", code, res)
+	}
+	if _, ok := res["branch"]; ok {
+		t.Errorf("the result names a branch: %v", res["branch"])
+	}
+	if branchPushed(f) {
+		t.Error("a branch was pushed for a run that changed nothing")
+	}
+}
+
+// Names a file system may open as .git are left out of the copy, and the
+// rest of the run goes on.
+func TestAGitRunLeavesOutEveryNameThatOpensAsDotGit(t *testing.T) {
+	f := newGitRunFixture(t)
+	names := []string{".git.", ".GIT  ", "git~1", "GIT~1. ", ".g\u200cit", ".git::$INDEX_ALLOCATION"}
+	script := "cd \"$1\" && echo new > edit.txt\n"
+	for _, n := range names {
+		script += "mkdir '" + n + "' && echo '[core]' > '" + n + "/config'\n"
+	}
+	f.setAgent(t, script)
+	res, code := f.run(t)
+	if code != 0 || res["ok"] != true {
+		t.Fatalf("the run failed: code %d, %v", code, res)
+	}
+	files := testGit(t, f.origin, "ls-tree", "-r", "--name-only", "-z", "agent-work")
+	for _, n := range names {
+		if strings.Contains(files, n+"/") {
+			t.Errorf("%q is in the commit", n)
+		}
+	}
+	skipped, _ := res["worktree_skipped"].([]any)
+	listed := map[string]bool{}
+	for _, s := range skipped {
+		listed[fmt.Sprint(s)] = true
+	}
+	for _, n := range names {
+		if !listed[n] {
+			t.Errorf("%q is not listed as left out: %v", n, skipped)
+		}
+	}
+	for _, n := range []string{".git", ".Git", ".git.", ".git . .", "GiT~1", ".gi\ufefft", ".git:x"} {
+		if !dotGitName(n) {
+			t.Errorf("dotGitName(%q) = false", n)
+		}
+	}
+	for _, n := range []string{".gitignore", ".gitattributes", "git", ".github", "git~2", "x.git"} {
+		if dotGitName(n) {
+			t.Errorf("dotGitName(%q) = true", n)
+		}
+	}
+}
+
+// With git-lfs installed, the bastion's own git calls clean and smudge
+// through it and push its objects; the agent's .git cannot replace it, and a
+// filter of another name the agent's .gitattributes names runs nothing.
+func TestAGitRunUsesTheLFSFilterOnlyFromThisProcess(t *testing.T) {
+	f := newGitRunFixture(t)
+	seed := filepath.Join(t.TempDir(), "s")
+	testGit(t, "", "clone", "-q", f.origin, seed)
+	if err := os.WriteFile(filepath.Join(seed, ".gitattributes"), []byte("*.bin filter=lfs -text\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(seed, "big.bin"), []byte("v1\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	testGit(t, seed, "add", "-A")
+	testGit(t, seed, "commit", "-q", "-m", "lfs")
+	testGit(t, seed, "push", "-q", "origin", "main")
+
+	bin := t.TempDir()
+	lfsLog := filepath.Join(t.TempDir(), "lfs.log")
+	wrapper := "#!/bin/sh\nGO_HELPER_GITLFS=1 GO_HELPER_GITLFS_LOG=" + lfsLog + " exec " + os.Args[0] +
+		" -test.run='^TestHelperGitLFS$' -- \"$@\"\n"
+	if err := os.WriteFile(filepath.Join(bin, "git-lfs"), []byte(wrapper), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	t.Setenv("PATH", bin+string(os.PathListSeparator)+os.Getenv("PATH"))
+
+	mark := func(name string) string { return "touch " + filepath.Join(f.marks, name) }
+	f.setAgent(t, `set -e
+cd "$1"
+git config filter.lfs.process '`+mark("agent-lfs-process")+`; exit 1'
+git config filter.lfs.clean '`+mark("agent-lfs-clean")+`; cat'
+git config filter.evil.clean '`+mark("evil")+`; cat'
+printf '*.bin filter=lfs -text\n*.txt filter=evil\n' > .gitattributes
+echo v2 > big.bin
+echo new > edit.txt
+`)
+	res, code := f.run(t)
+	if code != 0 || res["ok"] != true {
+		t.Fatalf("the run failed: code %d, %v", code, res)
+	}
+	ents, _ := os.ReadDir(f.marks)
+	for _, e := range ents {
+		t.Errorf("the agent's %s ran as this process", e.Name())
+	}
+	logged, _ := os.ReadFile(lfsLog)
+	for _, want := range []string{"smudge big.bin", "clean big.bin", "push origin agent-work"} {
+		if !strings.Contains(string(logged), want+"\n") {
+			t.Errorf("git-lfs did not %s; it logged:\n%s", want, logged)
+		}
+	}
+	for _, line := range strings.Split(string(logged), "\n") {
+		if strings.HasSuffix(line, ".txt") {
+			t.Errorf("git-lfs ran for a file of another filter: %s", line)
+		}
+	}
+	if got := testGit(t, f.origin, "show", "agent-work:big.bin"); got != "v2\n" {
+		t.Errorf("big.bin = %q", got)
+	}
+}
+
+// TestHelperGitLFS is the stand-in git-lfs: run by the wrapper above, it is
+// an identity filter that logs what it was asked to do.
+func TestHelperGitLFS(t *testing.T) {
+	if os.Getenv("GO_HELPER_GITLFS") != "1" {
+		return
+	}
+	args := os.Args
+	for i, a := range args {
+		if a == "--" {
+			args = args[i+1:]
+			break
+		}
+	}
+	logf, err := os.OpenFile(os.Getenv("GO_HELPER_GITLFS_LOG"), os.O_WRONLY|os.O_APPEND|os.O_CREATE, 0o644)
+	if err != nil {
+		os.Exit(2)
+	}
+	switch {
+	case len(args) > 0 && args[0] == "filter-process":
+		if err := helperFilterProcess(logf); err != nil {
+			fmt.Fprintln(os.Stderr, "stand-in git-lfs:", err)
+			os.Exit(1)
+		}
+	case len(args) > 0 && (args[0] == "clean" || args[0] == "smudge"):
+		fmt.Fprintf(logf, "%s %s\n", args[0], args[len(args)-1])
+		_, _ = io.Copy(os.Stdout, os.Stdin)
+	default:
+		fmt.Fprintln(logf, strings.Join(args, " "))
+	}
+	logf.Close()
+	os.Exit(0)
+}
+
+func pktRead(r *bufio.Reader) ([]byte, bool, error) {
+	var h [4]byte
+	if _, err := io.ReadFull(r, h[:]); err != nil {
+		return nil, false, err
+	}
+	n, err := strconv.ParseUint(string(h[:]), 16, 16)
+	if err != nil {
+		return nil, false, err
+	}
+	if n == 0 {
+		return nil, true, nil
+	}
+	b := make([]byte, n-4)
+	_, err = io.ReadFull(r, b)
+	return b, false, err
+}
+
+func pktList(r *bufio.Reader) ([]string, error) {
+	var out []string
+	for {
+		b, flush, err := pktRead(r)
+		if err != nil || flush {
+			return out, err
+		}
+		out = append(out, strings.TrimSuffix(string(b), "\n"))
+	}
+}
+
+func pktWrite(w io.Writer, lines ...string) {
+	for _, l := range lines {
+		fmt.Fprintf(w, "%04x%s", len(l)+4, l)
+	}
+	_, _ = w.Write([]byte("0000"))
+}
+
+// helperFilterProcess speaks git's long-running filter protocol, version 2,
+// returning every file as it came.
+func helperFilterProcess(logf io.Writer) error {
+	r, w := bufio.NewReader(os.Stdin), bufio.NewWriter(os.Stdout)
+	if _, err := pktList(r); err != nil {
+		return err
+	}
+	pktWrite(w, "git-filter-server\n", "version=2\n")
+	if err := w.Flush(); err != nil {
+		return err
+	}
+	if _, err := pktList(r); err != nil {
+		return err
+	}
+	pktWrite(w, "capability=clean\n", "capability=smudge\n")
+	if err := w.Flush(); err != nil {
+		return err
+	}
+	for {
+		head, err := pktList(r)
+		if err == io.EOF {
+			return nil
+		}
+		if err != nil {
+			return err
+		}
+		cmd, path := "", ""
+		for _, kv := range head {
+			k, v, _ := strings.Cut(kv, "=")
+			switch k {
+			case "command":
+				cmd = v
+			case "pathname":
+				path = v
+			}
+		}
+		var body bytes.Buffer
+		for {
+			b, flush, err := pktRead(r)
+			if err != nil {
+				return err
+			}
+			if flush {
+				break
+			}
+			body.Write(b)
+		}
+		fmt.Fprintf(logf, "%s %s\n", cmd, path)
+		pktWrite(w, "status=success\n")
+		data := body.Bytes()
+		for len(data) > 0 {
+			n := min(len(data), 65516)
+			fmt.Fprintf(w, "%04x", n+4)
+			_, _ = w.Write(data[:n])
+			data = data[n:]
+		}
+		_, _ = w.Write([]byte("0000"))
+		pktWrite(w)
+		if err := w.Flush(); err != nil {
+			return err
+		}
+	}
+}

client-go/harness_gitrun_linux_test.go

@@ -0,0 +1,355 @@
+package main
+
+// A git run's tree is the agent's, .git included. These tests stand in for an
+// agent that uses that .git against the bastion: hooks, a file system
+// monitor, a filter, a signing program, a changed remote and a link to a file
+// outside the tree, besides ordinary edits. The container runtime is a
+// stand-in script that does the agent's work on the mounted tree; the sidecar
+// it would start is a local HTTP server on the sidecar's port.
+
+import (
+	"encoding/json"
+	"net"
+	"net/http"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"sort"
+	"strings"
+	"testing"
+	"time"
+)
+
+const gitRunOutsideSecret = "outside-the-tree-7c41"
+
+// gitRunFixture is an origin with one commit, a decoy repository, a file
+// outside the tree, a folder the agent's code marks when it runs, and a
+// stand-in runtime that runs agent (a shell script, given the tree's path as
+// $1) when the run starts its container.
+type gitRunFixture struct {
+	origin, decoy, outside, marks, agentFile string
+}
+
+// setAgent writes the script the stand-in runtime runs as the agent.
+func (f gitRunFixture) setAgent(t *testing.T, script string) {
+	t.Helper()
+	if err := os.WriteFile(f.agentFile, []byte(script), 0o755); err != nil {
+		t.Fatal(err)
+	}
+}
+
+func testGit(t *testing.T, dir string, args ...string) string {
+	t.Helper()
+	cmd := exec.Command("git", args...)
+	cmd.Dir = dir
+	out, err := cmd.CombinedOutput()
+	if err != nil {
+		t.Fatalf("git %v: %v: %s", args, err, out)
+	}
+	return string(out)
+}
+
+func newGitRunFixture(t *testing.T) gitRunFixture {
+	t.Helper()
+	if _, err := exec.LookPath("git"); err != nil {
+		t.Skip("no git")
+	}
+	tmp := t.TempDir()
+	// The test's own git, and git as the code before this change ran it, read
+	// no config of this machine's.
+	home := filepath.Join(tmp, "home")
+	if err := os.Mkdir(home, 0o700); err != nil {
+		t.Fatal(err)
+	}
+	t.Setenv("HOME", home)
+	t.Setenv("XDG_CONFIG_HOME", home)
+	t.Setenv("GIT_CONFIG_NOSYSTEM", "1")
+	t.Setenv("GIT_AUTHOR_NAME", "seed")
+	t.Setenv("GIT_AUTHOR_EMAIL", "seed@example.invalid")
+	t.Setenv("GIT_COMMITTER_NAME", "seed")
+	t.Setenv("GIT_COMMITTER_EMAIL", "seed@example.invalid")
+	f := gitRunFixture{
+		origin:  filepath.Join(tmp, "origin.git"),
+		decoy:   filepath.Join(tmp, "decoy.git"),
+		outside: filepath.Join(tmp, "outside", "secret"),
+		marks:   filepath.Join(tmp, "marks"),
+	}
+	for _, d := range []string{filepath.Dir(f.outside), f.marks} {
+		if err := os.Mkdir(d, 0o755); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if err := os.WriteFile(f.outside, []byte(gitRunOutsideSecret), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	testGit(t, tmp, "init", "-q", "--bare", "-b", "main", f.origin)
+	testGit(t, tmp, "init", "-q", "--bare", "-b", "main", f.decoy)
+	seed := filepath.Join(tmp, "seed")
+	testGit(t, tmp, "init", "-q", "-b", "main", seed)
+	for name, body := range map[string]string{"keep.txt": "keep\n", "edit.txt": "old\n", "gone.txt": "bye\n"} {
+		if err := os.WriteFile(filepath.Join(seed, name), []byte(body), 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if err := os.WriteFile(filepath.Join(seed, "tool.sh"), []byte("#!/bin/sh\n"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	testGit(t, seed, "add", "-A")
+	testGit(t, seed, "commit", "-q", "-m", "seed")
+	testGit(t, seed, "push", "-q", f.origin, "main")
+
+	bin := t.TempDir()
+	f.agentFile = filepath.Join(bin, "agent.sh")
+	agentFile := f.agentFile
+	runtime := "#!/bin/sh\ncase \"$1\" in\nps|pull|rm) exit 0;;\ninspect) echo 127.0.0.1; exit 0;;\n" +
+		"run) for a; do case \"$a\" in *:/work) d=${a%:/work};; esac; done\n" +
+		"  /bin/sh " + agentFile + " \"$d\" > " + filepath.Join(bin, "agent.log") + " 2>&1 || exit 1\n" +
+		"  echo standin0123456789; exit 0;;\nesac\nexit 1\n"
+	if err := os.WriteFile(filepath.Join(bin, "docker"), []byte(runtime), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	t.Setenv("PATH", bin+string(os.PathListSeparator)+os.Getenv("PATH"))
+	t.Cleanup(func() {
+		if t.Failed() {
+			if b, err := os.ReadFile(filepath.Join(bin, "agent.log")); err == nil {
+				t.Logf("agent log:\n%s", b)
+			}
+		}
+	})
+
+	// The sidecar the run drives: it answers, and /run reports success.
+	ln, err := net.Listen("tcp", "127.0.0.1:8844")
+	if err != nil {
+		t.Fatalf("the sidecar's port is taken: %v", err)
+	}
+	srv := &http.Server{Handler: http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path == "/run" {
+			_, _ = w.Write([]byte(`{"ok":true,"transcript":"done"}`))
+		}
+	})}
+	go func() { _ = srv.Serve(ln) }()
+	t.Cleanup(func() { _ = srv.Close() })
+	return f
+}
+
+func (f gitRunFixture) run(t *testing.T) (map[string]any, int) {
+	t.Helper()
+	b, code := runGit(harnessSpec{Mode: "git", HarnessID: "h1", Image: "harness:test",
+		RepoURL: f.origin, Branch: "agent-work", MaxMinutes: 1})
+	var res map[string]any
+	if err := json.Unmarshal(b, &res); err != nil {
+		t.Fatalf("result: %v: %s", err, b)
+	}
+	return res, code
+}
+
+// theAgent turns its tree's .git against whoever runs git in it next, and
+// makes ordinary changes to the files.
+func theAgent(f gitRunFixture) string {
+	mark := func(name string) string { return "touch " + filepath.Join(f.marks, name) }
+	return `set -e
+d="$1"; cd "$d"
+g="$d/.git"
+printf '#!/bin/sh\n` + mark("pre-commit") + `\n' > "$g/hooks/pre-commit"
+printf '#!/bin/sh\n` + mark("post-commit") + `\n' > "$g/hooks/post-commit"
+chmod +x "$g/hooks/pre-commit" "$g/hooks/post-commit"
+printf '#!/bin/sh\n` + mark("fsmonitor") + `\nexit 1\n' > "$g/fsmon.sh"; chmod +x "$g/fsmon.sh"
+git config core.fsmonitor "$g/fsmon.sh"
+printf '* filter=evil\n' > .gitattributes
+git config filter.evil.clean '` + mark("filter") + `; cat ` + f.outside + `'
+git config filter.evil.smudge cat
+printf '#!/bin/sh\n` + mark("gpg") + `\necho >&2; echo "[GNUPG:] SIG_CREATED D 1 8 00 0 X" >&2\necho "-----BEGIN PGP SIGNATURE-----"; echo x; echo "-----END PGP SIGNATURE-----"\n' > "$g/gpg.sh"
+chmod +x "$g/gpg.sh"
+git config commit.gpgsign true
+git config gpg.program "$g/gpg.sh"
+git config remote.origin.url ` + f.decoy + `
+ln -s ` + f.outside + ` link-to-outside
+echo new > edit.txt
+echo added > added.txt
+rm gone.txt
+mkdir sub; echo deep > sub/deep.txt
+`
+}
+
+func TestAGitRunNeverRunsGitInTheAgentsTree(t *testing.T) {
+	f := newGitRunFixture(t)
+	f.setAgent(t, theAgent(f))
+	res, code := f.run(t)
+	if code != 0 || res["ok"] != true {
+		t.Errorf("the run failed: code %d, %v", code, res)
+	}
+
+	// Nothing the agent set in its .git ran.
+	ents, _ := os.ReadDir(f.marks)
+	for _, e := range ents {
+		t.Errorf("the agent's %s ran as this process", e.Name())
+	}
+	// The push went to the repository the run was given, not the one the
+	// agent named.
+	if refs := strings.TrimSpace(testGit(t, f.decoy, "for-each-ref")); refs != "" {
+		t.Errorf("the push went to the agent's remote: %s", refs)
+	}
+	out, err := exec.Command("git", "--git-dir", f.origin, "rev-parse", "--verify", "-q", "agent-work").Output()
+	if err != nil {
+		t.Fatalf("the branch is not in the origin: %v", err)
+	}
+	tip := strings.TrimSpace(string(out))
+	if n := strings.TrimSpace(testGit(t, f.origin, "rev-list", "--count", "main.."+tip)); n != "1" {
+		t.Errorf("the branch carries %s commits on main, want 1", n)
+	}
+	// Exactly the agent's file changes: the link is not there, and the outside
+	// file's content is in no file of the commit.
+	files := strings.Fields(testGit(t, f.origin, "ls-tree", "-r", "--format=%(objectmode) %(path)", tip))
+	got := map[string]string{}
+	for i := 0; i+1 < len(files); i += 2 {
+		got[files[i+1]] = files[i]
+	}
+	want := map[string]string{".gitattributes": "100644", "added.txt": "100644", "edit.txt": "100644",
+		"keep.txt": "100644", "sub/deep.txt": "100644", "tool.sh": "100755"}
+	if !sameMap(got, want) {
+		t.Errorf("the commit holds %v, want %v", got, want)
+	}
+	for path := range got {
+		body := testGit(t, f.origin, "show", tip+":"+path)
+		if strings.Contains(body, gitRunOutsideSecret) {
+			t.Errorf("%s in the commit holds the file outside the tree", path)
+		}
+	}
+	for path, body := range map[string]string{"edit.txt": "new\n", "added.txt": "added\n", "keep.txt": "keep\n"} {
+		if _, ok := got[path]; ok {
+			if b := testGit(t, f.origin, "show", tip+":"+path); b != body {
+				t.Errorf("%s = %q, want %q", path, b, body)
+			}
+		}
+	}
+	if s, _ := res["worktree_skipped"].([]any); len(s) != 1 || s[0] != "link-to-outside" {
+		t.Errorf("left out = %v, want [link-to-outside]", res["worktree_skipped"])
+	}
+	for _, k := range []string{"branch", "commit", "commit_out"} {
+		if _, ok := res[k]; !ok {
+			t.Errorf("the result has no %s", k)
+		}
+	}
+	if res["commit"] != tip {
+		t.Errorf("the result names commit %v, the branch is at %s", res["commit"], tip)
+	}
+}
+
+// A link the repository itself holds stays as it is when the agent leaves it
+// alone: a run that cannot copy a link must not delete it.
+func TestAGitRunKeepsALinkTheRepositoryHolds(t *testing.T) {
+	f := newGitRunFixture(t)
+	seed := filepath.Join(t.TempDir(), "s")
+	testGit(t, "", "clone", "-q", f.origin, seed)
+	if err := os.Symlink("keep.txt", filepath.Join(seed, "alias")); err != nil {
+		t.Fatal(err)
+	}
+	testGit(t, seed, "add", "-A")
+	testGit(t, seed, "commit", "-q", "-m", "a link")
+	testGit(t, seed, "push", "-q", "origin", "main")
+	f.setAgent(t, "cd \"$1\" && echo new > edit.txt\n")
+	res, code := f.run(t)
+	if code != 0 {
+		t.Fatalf("the run failed: %v", res)
+	}
+	if mode := strings.Fields(testGit(t, f.origin, "ls-tree", "agent-work", "alias")); len(mode) == 0 || mode[0] != "120000" {
+		t.Errorf("the repository's link is gone from the branch: %v", mode)
+	}
+	if diff := strings.Fields(testGit(t, f.origin, "diff", "--name-only", "main", "agent-work")); len(diff) != 1 || diff[0] != "edit.txt" {
+		t.Errorf("the branch changes %v, want [edit.txt]", diff)
+	}
+}
+
+// A file far larger than the copy may take (sparse, so it costs nothing to
+// make) is refused from its size, before a byte of it is read: the run ends
+// at once, pushes nothing, and says why.
+func TestAGitRunRefusesATreeLargerThanItCopies(t *testing.T) {
+	f := newGitRunFixture(t)
+	f.setAgent(t, "cd \"$1\" && truncate -s 64G huge.bin && echo new > edit.txt\n")
+	type out struct {
+		res  map[string]any
+		code int
+	}
+	ch := make(chan out, 1)
+	start := time.Now()
+	go func() { r, c := f.run(t); ch <- out{r, c} }()
+	var o out
+	select {
+	case o = <-ch:
+	case <-time.After(30 * time.Second):
+		t.Fatal("a run with a 64 GiB file was still going after 30 s")
+	}
+	t.Logf("returned in %v", time.Since(start))
+	if o.code == 0 || o.res["ok"] != false {
+		t.Fatalf("a tree over the cap was pushed: %v", o.res)
+	}
+	if e, _ := o.res["error"].(string); !strings.Contains(e, "huge.bin") || !strings.Contains(e, "too large") {
+		t.Errorf("the error does not say what was too large: %q", e)
+	}
+	if err := exec.Command("git", "--git-dir", f.origin, "rev-parse", "--verify", "-q", "agent-work").Run(); err == nil {
+		t.Error("the branch was pushed")
+	}
+}
+
+// The tree a worktree run returns is refused the same way, and no part of
+// it is returned.
+func TestAWorktreeRunRefusesATreeLargerThanTheLaneCarries(t *testing.T) {
+	f := newGitRunFixture(t)
+	f.setAgent(t, "cd \"$1\" && truncate -s 64G huge.bin\n")
+	ch := make(chan []byte, 1)
+	start := time.Now()
+	go func() {
+		b, _ := runWorktree(harnessSpec{Mode: "worktree", HarnessID: "h1", Image: "harness:test", MaxMinutes: 1})
+		ch <- b
+	}()
+	var b []byte
+	select {
+	case b = <-ch:
+	case <-time.After(30 * time.Second):
+		t.Fatal("a run with a 64 GiB file was still going after 30 s")
+	}
+	t.Logf("returned in %v", time.Since(start))
+	var res map[string]any
+	if err := json.Unmarshal(b, &res); err != nil {
+		t.Fatal(err)
+	}
+	if res["ok"] != false || res["worktree_b64"] != nil {
+		t.Fatalf("a tree over the cap was returned: ok=%v, %d bytes of tree", res["ok"], len(b))
+	}
+	if e, _ := res["error"].(string); !strings.Contains(e, "huge.bin") || !strings.Contains(e, "too large") {
+		t.Errorf("the error does not say what was too large: %q", e)
+	}
+}
+
+func sameMap(a, b map[string]string) bool {
+	if len(a) != len(b) {
+		return false
+	}
+	ka := make([]string, 0, len(a))
+	for k := range a {
+		ka = append(ka, k)
+	}
+	sort.Strings(ka)
+	for _, k := range ka {
+		if b[k] != a[k] {
+			return false
+		}
+	}
+	return true
+}
+
+// A repository with no commit yet gets the run's commit as its first.
+func TestAGitRunIntoAnEmptyRepositoryPushesItsFirstCommit(t *testing.T) {
+	f := newGitRunFixture(t)
+	empty := filepath.Join(t.TempDir(), "empty.git")
+	testGit(t, "", "init", "-q", "--bare", "-b", "main", empty)
+	f.setAgent(t, "cd \"$1\" && echo first > first.txt\n")
+	b, code := runGit(harnessSpec{Mode: "git", HarnessID: "h1", Image: "harness:test",
+		RepoURL: empty, Branch: "agent-work", MaxMinutes: 1})
+	if code != 0 {
+		t.Fatalf("the run failed: %s", b)
+	}
+	if got := testGit(t, empty, "show", "agent-work:first.txt"); got != "first\n" {
+		t.Errorf("first.txt = %q", got)
+	}
+}

client-go/harness_gitrun_reallfs_linux_test.go

@@ -0,0 +1,143 @@
+package main
+
+// A git run against a repository that keeps a file in Git LFS, with the real
+// git-lfs. It runs only where git-lfs is installed; every server it talks to
+// is in this test, on loopback.
+
+import (
+	"crypto/sha256"
+	"encoding/base64"
+	"encoding/hex"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"net/http/cgi"
+	"net/http/httptest"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"testing"
+)
+
+func lfsPointer(body string) (oid, pointer string) {
+	sum := sha256.Sum256([]byte(body))
+	oid = hex.EncodeToString(sum[:])
+	return oid, fmt.Sprintf("version https://git-lfs.github.com/spec/v1\noid sha256:%s\nsize %d\n", oid, len(body))
+}
+
+func TestAGitRunWithTheRealGitLFS(t *testing.T) {
+	if out, err := exec.Command("git-lfs", "version").Output(); err != nil || !strings.HasPrefix(string(out), "git-lfs/") {
+		t.Skip("git-lfs is not installed")
+	}
+	f := newGitRunFixture(t)
+	realGit, _ := exec.LookPath("git")
+	want := "Basic " + base64.StdEncoding.EncodeToString([]byte("harness:"+gitRunToken))
+
+	var mu sync.Mutex
+	store := map[string][]byte{}
+	v1oid, v1ptr := lfsPointer("v1\n")
+	store[v1oid] = []byte("v1\n")
+	var srvURL string
+	backend := &cgi.Handler{Path: realGit, Args: []string{"http-backend"}, Root: "/git",
+		Env: []string{"GIT_PROJECT_ROOT=" + filepath.Dir(f.origin), "GIT_HTTP_EXPORT_ALL=1",
+			"REMOTE_USER=harness", "GIT_CONFIG_NOSYSTEM=1", "HOME=" + t.TempDir()}}
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if strings.HasPrefix(r.URL.Path, "/store/") { // the object hrefs the batch answer gives
+			oid := strings.TrimPrefix(r.URL.Path, "/store/")
+			mu.Lock()
+			defer mu.Unlock()
+			if r.Method == http.MethodPut {
+				b, _ := io.ReadAll(r.Body)
+				store[oid] = b
+				return
+			}
+			_, _ = w.Write(store[oid])
+			return
+		}
+		if r.Header.Get("Authorization") != want {
+			w.Header().Set("WWW-Authenticate", `Basic realm="test"`)
+			http.Error(w, "auth", http.StatusUnauthorized)
+			return
+		}
+		if !strings.Contains(r.URL.Path, "/info/lfs/") {
+			backend.ServeHTTP(w, r)
+			return
+		}
+		if !strings.HasSuffix(r.URL.Path, "/objects/batch") {
+			http.NotFound(w, r) // locks: not supported here
+			return
+		}
+		var req struct {
+			Operation string `json:"operation"`
+			Objects   []struct {
+				Oid  string `json:"oid"`
+				Size int64  `json:"size"`
+			} `json:"objects"`
+		}
+		_ = json.NewDecoder(r.Body).Decode(&req)
+		var objs []map[string]any
+		mu.Lock()
+		for _, o := range req.Objects {
+			obj := map[string]any{"oid": o.Oid, "size": o.Size, "authenticated": true}
+			_, have := store[o.Oid]
+			if req.Operation == "download" || !have {
+				obj["actions"] = map[string]any{req.Operation: map[string]any{"href": srvURL + "/store/" + o.Oid}}
+			}
+			objs = append(objs, obj)
+		}
+		mu.Unlock()
+		w.Header().Set("Content-Type", "application/vnd.git-lfs+json")
+		_ = json.NewEncoder(w).Encode(map[string]any{"transfer": "basic", "objects": objs})
+	}))
+	t.Cleanup(srv.Close)
+	srvURL = srv.URL
+	var foreignHits atomic.Int64
+	foreign := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		foreignHits.Add(1)
+		http.Error(w, "no", http.StatusForbidden)
+	}))
+	t.Cleanup(foreign.Close)
+
+	seed := filepath.Join(t.TempDir(), "s")
+	testGit(t, "", "clone", "-q", f.origin, seed)
+	for name, body := range map[string]string{".gitattributes": "*.bin filter=lfs diff=lfs merge=lfs -text\n", "big.bin": v1ptr} {
+		if err := os.WriteFile(filepath.Join(seed, name), []byte(body), 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+	testGit(t, seed, "add", "-A")
+	testGit(t, seed, "commit", "-q", "-m", "an lfs file")
+	testGit(t, seed, "push", "-q", "origin", "main")
+
+	// The agent changes the large file and names a foreign LFS server.
+	f.setAgent(t, `set -e
+cd "$1"
+grep -qx v1 big.bin
+echo v2 > big.bin
+printf '[lfs]\n\turl = `+foreign.URL+`/x.git/info/lfs\n\tpushurl = `+foreign.URL+`/x.git/info/lfs\n[remote "origin"]\n\tlfsurl = `+foreign.URL+`/y\n' > .lfsconfig
+`)
+	repoURL := strings.Replace(srv.URL, "http://", "http://harness:"+gitRunToken+"@", 1) + "/git/" + filepath.Base(f.origin)
+	b, code := runGit(harnessSpec{Mode: "git", HarnessID: "h1", Image: "harness:test",
+		RepoURL: repoURL, Branch: "agent-work", MaxMinutes: 1})
+	t.Logf("result: %s", b)
+	if code != 0 {
+		t.Fatalf("the run failed")
+	}
+	v2oid, v2ptr := lfsPointer("v2\n")
+	if got := testGit(t, f.origin, "show", "agent-work:big.bin"); got != v2ptr {
+		t.Errorf("big.bin on the branch = %q, want the pointer to v2", got)
+	}
+	mu.Lock()
+	got := string(store[v2oid])
+	mu.Unlock()
+	if got != "v2\n" {
+		t.Errorf("the LFS object for v2 was not sent to the repository's own LFS server (%q)", got)
+	}
+	if n := foreignHits.Load(); n != 0 {
+		t.Errorf("the foreign LFS server named in .lfsconfig got %d requests", n)
+	}
+}

client-go/harness_tar_linux.go

@@ -0,0 +1,43 @@
+package main
+
+import (
+	"errors"
+	"strconv"
+
+	"golang.org/x/sys/unix"
+)
+
+// openTreeEntry opens the entry name in the folder dfd with flags, for
+// packFolder.
+//
+// A folder is opened directly: with O_DIRECTORY the kernel refuses anything
+// that is not a folder before it opens it. Anything else is first opened as a
+// path only (O_PATH), which runs no driver code, so a device, pipe or terminal
+// swapped in since the look is never opened: opening some devices does
+// something by itself (a watchdog arms, a tape rewinds on close). Only a plain
+// file is then opened for reading, through /proc/self/fd, which reaches the
+// file the path fd holds and no other.
+func openTreeEntry(dfd int, name string, flags int) (int, error) {
+	if flags&unix.O_DIRECTORY != 0 {
+		return unix.Openat(dfd, name, flags, 0)
+	}
+	pfd, err := unix.Openat(dfd, name, unix.O_PATH|unix.O_NOFOLLOW|unix.O_CLOEXEC, 0)
+	if err != nil {
+		return -1, err
+	}
+	defer unix.Close(pfd)
+	var st unix.Stat_t
+	if err := unix.Fstat(pfd, &st); err != nil {
+		return -1, err
+	}
+	if st.Mode&unix.S_IFMT != unix.S_IFREG {
+		return -1, errNotPlainFile
+	}
+	fd, err := unix.Open("/proc/self/fd/"+strconv.Itoa(pfd), flags&^unix.O_NOFOLLOW, 0)
+	if errors.Is(err, unix.ENOENT) {
+		// No /proc mounted: open by name as other systems do. packOpened
+		// still checks that what was opened is the file that was looked at.
+		return unix.Openat(dfd, name, flags, 0)
+	}
+	return fd, err
+}

client-go/harness_tar_linux_test.go

@@ -0,0 +1,70 @@
+package main
+
+import (
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"golang.org/x/sys/unix"
+)
+
+// A name swapped for a special file between the walk's look and its open must
+// never be opened: opening a device runs its driver (a watchdog arms, a tape
+// rewinds on close, a serial line raises DTR), whatever is read after. A named
+// pipe stands in for any of them, since anyone can make one, and inotify
+// reports every open of it except one made only as a path (O_PATH), which runs
+// no driver code. Both a name looked at as a plain file and one looked at as
+// a folder are swapped.
+func TestMakeTarNeverOpensASpecialFileSwappedInBeforeTheOpen(t *testing.T) {
+	for _, swap := range []string{"a.txt", "sub"} {
+		t.Run(swap, func(t *testing.T) {
+			_, _, tree := outsideAndTree(t)
+			in, err := unix.InotifyInit1(unix.IN_NONBLOCK | unix.IN_CLOEXEC)
+			if err != nil {
+				t.Skipf("no inotify: %v", err)
+			}
+			defer unix.Close(in)
+			watched := false
+			packTreeBeforeOpen = func(rel string) {
+				if rel != swap {
+					return
+				}
+				p := filepath.Join(tree, swap)
+				if err := os.RemoveAll(p); err != nil {
+					t.Errorf("remove %s: %v", p, err)
+					return
+				}
+				if err := unix.Mkfifo(p, 0o644); err != nil {
+					t.Errorf("mkfifo %s: %v", p, err)
+					return
+				}
+				if _, err := unix.InotifyAddWatch(in, p, unix.IN_OPEN); err != nil {
+					t.Errorf("watch %s: %v", p, err)
+					return
+				}
+				watched = true
+			}
+			defer func() { packTreeBeforeOpen = nil }()
+			b, skipped, err := packWorktree(tree)
+			if err != nil {
+				t.Fatalf("packWorktree: %v", err)
+			}
+			if !watched {
+				t.Fatal("the swap never ran")
+			}
+			buf := make([]byte, 4096)
+			if n, err := unix.Read(in, buf); n > 0 {
+				t.Fatalf("the special file swapped in for %s was opened (%d bytes of inotify events)", swap, n)
+			} else if err != unix.EAGAIN {
+				t.Fatalf("read inotify: %d %v", n, err)
+			}
+			if _, ok := tarEntries(t, b)[swap]; ok {
+				t.Fatalf("%s was a named pipe when it was opened and must be left out", swap)
+			}
+			if !strings.Contains(","+strings.Join(skipped, ",")+",", ","+swap+",") {
+				t.Fatalf("%s must be named as left out, got %v", swap, skipped)
+			}
+		})
+	}
+}

client-go/harness_tar_other.go

@@ -0,0 +1,13 @@
+//go:build !windows && !linux
+
+package main
+
+import "golang.org/x/sys/unix"
+
+// openTreeEntry opens the entry name in the folder dfd with flags, for
+// packFolder. These systems have no open that runs no driver code, so a
+// special file swapped in since the look is opened (O_NONBLOCK, O_NOCTTY),
+// then closed unread by packOpened.
+func openTreeEntry(dfd int, name string, flags int) (int, error) {
+	return unix.Openat(dfd, name, flags, 0)
+}

client-go/harness_tar_unix.go

@@ -0,0 +1,114 @@
+//go:build !windows
+
+package main
+
+import (
+	"errors"
+	"fmt"
+	"os"
+
+	"golang.org/x/sys/unix"
+)
+
+// errNotPlainFile is openTreeEntry's refusal of a name that is no longer a
+// plain file when it is opened.
+var errNotPlainFile = errors.New("not a plain file")
+
+// packTree hands the tree under root to sink, walking it by open folder
+// (packWorktree says why). Each name is looked at with fstatat and opened
+// relative to the folder above it (openTreeEntry, per system), never
+// following a link, and it goes in only if the opened file is still the plain
+// file or folder that was looked at (same type, device and inode). A pipe or
+// device seen as one is never opened, so the walk cannot wait on one; on
+// Linux one swapped in between is not opened either, and elsewhere O_NONBLOCK
+// and O_NOCTTY cover it. Each left-out entry goes to sink.skip.
+func packTree(sink treeSink, root string) error {
+	fd, err := unix.Open(root, unix.O_RDONLY|unix.O_DIRECTORY|unix.O_NOFOLLOW|unix.O_CLOEXEC, 0)
+	if err != nil {
+		return &os.PathError{Op: "open", Path: root, Err: err}
+	}
+	d := os.NewFile(uintptr(fd), root)
+	defer d.Close()
+	return packFolder(sink, d, fd, "", 0)
+}
+
+func packFolder(sink treeSink, d *os.File, dfd int, rel string, depth int) error {
+	names, err := readTreeNames(sink, d, rel)
+	if err != nil {
+		return err
+	}
+	if err := sink.folder(rel, names); err != nil {
+		return err
+	}
+	for _, name := range names {
+		p := name
+		if rel != "" {
+			p = rel + "/" + name
+		}
+		var st unix.Stat_t
+		if err := unix.Fstatat(dfd, name, &st, unix.AT_SYMLINK_NOFOLLOW); err != nil {
+			sink.skip(p) // gone since the folder was read
+			continue
+		}
+		kind := st.Mode & unix.S_IFMT
+		if kind == unix.S_IFDIR && name == ".git" {
+			continue
+		}
+		if (kind != unix.S_IFREG && kind != unix.S_IFDIR) || (kind == unix.S_IFDIR && depth >= maxTreeDepth) {
+			sink.skip(p)
+			continue
+		}
+		if packTreeBeforeOpen != nil {
+			packTreeBeforeOpen(p)
+		}
+		flags := unix.O_RDONLY | unix.O_NOFOLLOW | unix.O_CLOEXEC | unix.O_NONBLOCK | unix.O_NOCTTY
+		if kind == unix.S_IFDIR {
+			flags |= unix.O_DIRECTORY
+		}
+		cfd, err := openTreeEntry(dfd, name, flags)
+		if err != nil {
+			sink.skip(p) // a link or special file now, or gone
+			continue
+		}
+		err = packOpened(sink, cfd, name, st, p, depth)
+		if err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+// packOpened hands sink the entry opened as fd, and for a folder what is in it,
+// if it is still the entry fstatat saw; it closes fd.
+func packOpened(sink treeSink, fd int, name string, seen unix.Stat_t, p string, depth int) error {
+	var got unix.Stat_t
+	if err := unix.Fstat(fd, &got); err != nil {
+		unix.Close(fd)
+		return fmt.Errorf("stat %q: %w", p, err)
+	}
+	kind := got.Mode & unix.S_IFMT
+	if kind != seen.Mode&unix.S_IFMT || got.Dev != seen.Dev || got.Ino != seen.Ino {
+		unix.Close(fd)
+		sink.skip(p) // swapped between the look and the open
+		return nil
+	}
+	// A plain file or folder now: back to blocking reads, so os.File reads it
+	// as it reads any file it opened itself.
+	if err := unix.SetNonblock(fd, false); err != nil {
+		unix.Close(fd)
+		return fmt.Errorf("open %q: %w", p, err)
+	}
+	f := os.NewFile(uintptr(fd), name)
+	defer f.Close()
+	fi, err := f.Stat()
+	if err != nil {
+		return fmt.Errorf("stat %q: %w", p, err)
+	}
+	if err := sink.put(f, fi, p); err != nil {
+		return err
+	}
+	if kind == unix.S_IFDIR {
+		return packFolder(sink, f, fd, p, depth+1)
+	}
+	return nil
+}

client-go/harness_tar_unix_test.go

@@ -0,0 +1,204 @@
+//go:build !windows
+
+package main
+
+import (
+	"archive/tar"
+	"bytes"
+	"io"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"time"
+
+	"golang.org/x/sys/unix"
+)
+
+const tarOutsideSecret = "outside-the-tree-3f9a"
+
+// tarInsideA is as long as the outside file: a writer who knows a file's size
+// can make the one it swaps out the same size, so the tar's header fits.
+const tarInsideA = "inside-the-tree-a-000"
+
+// tarEntries reads back what makeTar produced: name -> content.
+func tarEntries(t *testing.T, b []byte) map[string]string {
+	t.Helper()
+	out := map[string]string{}
+	tr := tar.NewReader(bytes.NewReader(b))
+	for {
+		hdr, err := tr.Next()
+		if err == io.EOF {
+			return out
+		}
+		if err != nil {
+			t.Fatalf("read tar: %v", err)
+		}
+		data, err := io.ReadAll(tr)
+		if err != nil {
+			t.Fatalf("read %s: %v", hdr.Name, err)
+		}
+		out[hdr.Name] = string(data)
+	}
+}
+
+// outsideAndTree makes a file outside the tree that a run must never return,
+// and a tree with one file and one folder.
+func outsideAndTree(t *testing.T) (outsideDir, secret, tree string) {
+	t.Helper()
+	outsideDir = t.TempDir()
+	secret = filepath.Join(outsideDir, "secret")
+	if err := os.WriteFile(secret, []byte(tarOutsideSecret), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	tree = t.TempDir()
+	if err := os.WriteFile(filepath.Join(tree, "a.txt"), []byte(tarInsideA), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.Mkdir(filepath.Join(tree, "sub"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	// Named as the outside file is: a writer who knows the names in a folder
+	// outside can fill a folder of the tree with the same names, then swap it.
+	if err := os.WriteFile(filepath.Join(tree, "sub", "secret"), []byte("inside b"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	return outsideDir, secret, tree
+}
+
+func assertNoOutside(t *testing.T, ents map[string]string) {
+	t.Helper()
+	for name, data := range ents {
+		if strings.Contains(data, tarOutsideSecret) {
+			t.Fatalf("the tar holds a file from outside the tree, as %q", name)
+		}
+	}
+}
+
+// The tree's writer swaps a file for a link, and a folder for a link, after
+// the walk has seen each one as a plain file or folder and before it opens it.
+// The hook stands exactly in that gap, so the swap lands there every time.
+func TestMakeTarNeverReadsThroughALinkSwappedInBeforeTheOpen(t *testing.T) {
+	for _, swap := range []string{"a.txt", "sub"} {
+		t.Run(swap, func(t *testing.T) {
+			outsideDir, secret, tree := outsideAndTree(t)
+			packTreeBeforeOpen = func(rel string) {
+				if rel != swap {
+					return
+				}
+				p := filepath.Join(tree, swap)
+				_ = os.RemoveAll(p)
+				if swap == "sub" {
+					_ = os.Symlink(outsideDir, p)
+				} else {
+					_ = os.Symlink(secret, p)
+				}
+			}
+			defer func() { packTreeBeforeOpen = nil }()
+			b, err := makeTar(tree)
+			if err != nil {
+				t.Fatalf("makeTar: %v", err)
+			}
+			ents := tarEntries(t, b)
+			assertNoOutside(t, ents)
+			if swap == "a.txt" {
+				if _, ok := ents["a.txt"]; ok {
+					t.Fatalf("a.txt was a link when it was opened and must be left out, got it in the tar")
+				}
+			}
+		})
+	}
+}
+
+// No hook: a writer flips one name between a plain file and a link to a file
+// outside the tree as fast as it can while the tree is packed again and again.
+func TestMakeTarNeverReadsThroughALinkUnderAConcurrentWriter(t *testing.T) {
+	_, secret, tree := outsideAndTree(t)
+	name := filepath.Join(tree, "a.txt")
+	tmpLink, tmpFile := filepath.Join(tree, ".l"), filepath.Join(tree, ".f")
+	stop := make(chan struct{})
+	done := make(chan struct{})
+	go func() {
+		defer close(done)
+		for {
+			select {
+			case <-stop:
+				return
+			default:
+			}
+			_ = os.WriteFile(tmpFile, []byte(tarInsideA), 0o644)
+			_ = os.Rename(tmpFile, name)
+			_ = os.Symlink(secret, tmpLink)
+			_ = os.Rename(tmpLink, name)
+		}
+	}()
+	defer func() { close(stop); <-done }()
+	deadline := time.Now().Add(3 * time.Second)
+	passes := 0
+	defer func() { t.Logf("%d passes", passes) }()
+	for i := 0; time.Now().Before(deadline); i++ {
+		passes = i + 1
+		b, err := makeTar(tree)
+		if err != nil {
+			continue // a name that vanished mid-walk; the next pass tries again
+		}
+		for n, data := range tarEntries(t, b) {
+			if strings.Contains(data, tarOutsideSecret) {
+				t.Fatalf("pass %d: the tar holds the file outside the tree, as %q", i, n)
+			}
+		}
+	}
+}
+
+// A link, a named pipe and a socket in the tree are left out and named as
+// left out; the pipe is never opened, so the walk does not wait on it.
+func TestMakeTarLeavesOutAndNamesWhatIsNotAPlainFileOrFolder(t *testing.T) {
+	_, secret, tree := outsideAndTree(t)
+	if err := os.Symlink(secret, filepath.Join(tree, "link")); err != nil {
+		t.Fatal(err)
+	}
+	if err := unix.Mkfifo(filepath.Join(tree, "sub", "pipe"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.Mkdir(filepath.Join(tree, ".git"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(tree, ".git", "config"), []byte("x"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	type res struct {
+		b       []byte
+		skipped []string
+		err     error
+	}
+	ch := make(chan res, 1)
+	go func() { b, s, err := packWorktree(tree); ch <- res{b, s, err} }()
+	var r res
+	select {
+	case r = <-ch:
+	case <-time.After(10 * time.Second):
+		t.Fatal("packing the tree waited on the named pipe")
+	}
+	if r.err != nil {
+		t.Fatalf("packWorktree: %v", r.err)
+	}
+	ents := tarEntries(t, r.b)
+	assertNoOutside(t, ents)
+	for _, want := range []string{"a.txt", "sub", "sub/secret"} {
+		if _, ok := ents[want]; !ok {
+			t.Errorf("%s missing from the tar (got %v)", want, ents)
+		}
+	}
+	if ents["a.txt"] != tarInsideA || ents["sub/secret"] != "inside b" {
+		t.Errorf("file contents changed in the tar: %v", ents)
+	}
+	for _, gone := range []string{"link", "sub/pipe", ".git", ".git/config"} {
+		if _, ok := ents[gone]; ok {
+			t.Errorf("%s is in the tar and must not be", gone)
+		}
+	}
+	got := strings.Join(r.skipped, ",")
+	if got != "link,sub/pipe" {
+		t.Errorf("left out = %q, want link,sub/pipe", got)
+	}
+}

client-go/harness_tar_windows.go

@@ -0,0 +1,83 @@
+//go:build windows
+
+package main
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+)
+
+// packTree hands the tree under root to sink (packWorktree says why it is
+// walked this way). Every name is opened through an os.Root at the tree, so
+// no link, junction or name swapped in while it is packed can open a file
+// outside the tree; an entry goes in only if the opened file is the plain
+// file or folder that was looked at. Each left-out entry goes to sink.skip.
+func packTree(sink treeSink, root string) error {
+	r, err := os.OpenRoot(root)
+	if err != nil {
+		return err
+	}
+	defer r.Close()
+	return packRootFolder(sink, r, "", 0)
+}
+
+func packRootFolder(sink treeSink, r *os.Root, rel string, depth int) error {
+	open := rel
+	if open == "" {
+		open = "."
+	}
+	d, err := r.Open(open)
+	if err != nil {
+		return fmt.Errorf("open folder %q: %w", rel, err)
+	}
+	names, err := readTreeNames(sink, d, filepath.ToSlash(rel))
+	d.Close()
+	if err != nil {
+		return err
+	}
+	if err := sink.folder(filepath.ToSlash(rel), names); err != nil {
+		return err
+	}
+	for _, name := range names {
+		p := filepath.Join(rel, name)
+		tp := filepath.ToSlash(p)
+		seen, err := r.Lstat(p)
+		if err != nil {
+			sink.skip(tp)
+			continue
+		}
+		if seen.IsDir() && name == ".git" {
+			continue
+		}
+		if (!seen.Mode().IsRegular() && !seen.IsDir()) || (seen.IsDir() && depth >= maxTreeDepth) {
+			sink.skip(tp)
+			continue
+		}
+		if packTreeBeforeOpen != nil {
+			packTreeBeforeOpen(tp)
+		}
+		f, err := r.Open(p)
+		if err != nil {
+			sink.skip(tp)
+			continue
+		}
+		fi, err := f.Stat()
+		if err != nil || fi.Mode().Type() != seen.Mode().Type() {
+			f.Close()
+			sink.skip(tp)
+			continue
+		}
+		err = sink.put(f, fi, tp)
+		f.Close()
+		if err != nil {
+			return err
+		}
+		if fi.IsDir() {
+			if err := packRootFolder(sink, r, p, depth+1); err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}

client-go/harness_tree_linux_test.go

@@ -0,0 +1,125 @@
+package main
+
+// A harness run's tree is bind-mounted into the container at /work and opened
+// to every user, because the agent runs as a uid of its own. What keeps the
+// other users of this machine out of it is the folder it sits in, which only
+// this process can enter. Opening the tree follows no link in it.
+
+import (
+	"archive/tar"
+	"bytes"
+	"encoding/base64"
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"testing"
+)
+
+// harnessStandIn writes a stand-in container runtime named name into a new
+// folder. On `ps` and `pull` it succeeds; on `run` it records, into the
+// returned file, the mode and owner of the folder the /work tree sits in, then
+// fails, so the run stops there.
+func harnessStandIn(t *testing.T, name string) (bin, record string) {
+	t.Helper()
+	bin = t.TempDir()
+	record = filepath.Join(bin, "record")
+	script := "#!/bin/sh\ncase \"$1\" in ps|pull) exit 0;; run) ;; *) exit 1;; esac\n" +
+		"for a; do case \"$a\" in *:/work) d=${a%:/work};; esac; done\n" +
+		"[ -n \"$d\" ] && stat -c '%a %u' \"$(dirname \"$d\")\" > " + record + "\nexit 1\n"
+	if err := os.WriteFile(filepath.Join(bin, name), []byte(script), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	return bin, record
+}
+
+func TestAWorktreeRunsTreeSitsInAFolderOnlyTheBastionCanEnter(t *testing.T) {
+	bin, record := harnessStandIn(t, "docker")
+	t.Setenv("PATH", bin+string(os.PathListSeparator)+os.Getenv("PATH"))
+	// The temporary folder every user may write in, as /tmp is.
+	shared := t.TempDir()
+	if err := os.Chmod(shared, os.ModeSticky|0o777); err != nil {
+		t.Fatal(err)
+	}
+	t.Setenv("TMPDIR", shared)
+
+	var tb bytes.Buffer
+	tw := tar.NewWriter(&tb)
+	_ = tw.WriteHeader(&tar.Header{Name: "src/", Typeflag: tar.TypeDir, Mode: 0o755})
+	_ = tw.WriteHeader(&tar.Header{Name: "src/a.txt", Typeflag: tar.TypeReg, Mode: 0o644, Size: 2})
+	_, _ = tw.Write([]byte("a\n"))
+	_ = tw.Close()
+	spec := harnessSpec{Mode: "worktree", HarnessID: "h1", Image: "harness:test",
+		WorktreeB64: base64.StdEncoding.EncodeToString(tb.Bytes())}
+	if _, code := runWorktree(spec); code == 0 {
+		t.Fatal("the stand-in runtime fails the start; the run must report it")
+	}
+	got, err := os.ReadFile(record)
+	if err != nil {
+		t.Fatalf("the tree was not mounted: %v", err)
+	}
+	if want := "700 " + strconv.Itoa(os.Geteuid()); strings.TrimSpace(string(got)) != want {
+		t.Fatalf("the folder the tree sits in: want %q, got %q", want, strings.TrimSpace(string(got)))
+	}
+}
+
+func TestOpeningARunsTreeFollowsNoLinkAndKeepsWhatIsExecutable(t *testing.T) {
+	bin, _ := harnessStandIn(t, "runtime")
+	rt := filepath.Join(bin, "runtime")
+	outside := filepath.Join(t.TempDir(), "shadow")
+	if err := os.WriteFile(outside, []byte("x\n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	tree := t.TempDir()
+	files := map[string]os.FileMode{"run.sh": 0o755, "notes.txt": 0o644}
+	for name, mode := range files {
+		if err := os.WriteFile(filepath.Join(tree, name), nil, mode); err != nil {
+			t.Fatal(err)
+		}
+	}
+	if err := os.Mkdir(filepath.Join(tree, "sub"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.Symlink(outside, filepath.Join(tree, "sub", "link")); err != nil {
+		t.Fatal(err)
+	}
+	if _, err := driveHarnessAgent(rt, tree, harnessSpec{HarnessID: "h1", Image: "harness:test"}, 1); err == nil {
+		t.Fatal("the stand-in runtime fails the start; the run must report it")
+	}
+	for p, want := range map[string]os.FileMode{
+		tree: 0o777, filepath.Join(tree, "sub"): 0o777,
+		filepath.Join(tree, "run.sh"): 0o777, filepath.Join(tree, "notes.txt"): 0o666,
+		outside: 0o600,
+	} {
+		fi, err := os.Stat(p)
+		if err != nil {
+			t.Fatal(err)
+		}
+		if fi.Mode().Perm() != want {
+			t.Errorf("%s: want %04o, got %04o", p, want, fi.Mode().Perm())
+		}
+	}
+}
+
+func TestARunsTreeInAFolderOtherUsersCanEnterIsRefused(t *testing.T) {
+	bin, record := harnessStandIn(t, "runtime")
+	rt := filepath.Join(bin, "runtime")
+	open := t.TempDir()
+	if err := os.Chmod(open, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	tree := filepath.Join(open, "tree")
+	if err := os.Mkdir(tree, 0o755); err != nil {
+		t.Fatal(err)
+	}
+	_, err := driveHarnessAgent(rt, tree, harnessSpec{HarnessID: "h1", Image: "harness:test"}, 1)
+	if err == nil || !strings.Contains(err.Error(), "open to other users") {
+		t.Fatalf("want the tree refused for its folder, got %v", err)
+	}
+	if _, err := os.Stat(record); err == nil {
+		t.Fatal("a refused tree must start no container")
+	}
+	if fi, _ := os.Stat(tree); fi.Mode().Perm() != 0o755 {
+		t.Fatalf("a refused tree must not be opened: %04o", fi.Mode().Perm())
+	}
+}

client-go/harness_treebounds_unix_test.go

@@ -0,0 +1,119 @@
+//go:build !windows
+
+package main
+
+import (
+	"errors"
+	"os"
+	"path/filepath"
+	"runtime"
+	"strconv"
+	"strings"
+	"testing"
+)
+
+// A copy of a tree of very many empty files stops at its limit, and what it
+// holds in memory does not grow with the number of files.
+func TestACopyOfVeryManyEmptyFilesIsRefusedWithBoundedMemory(t *testing.T) {
+	src, dst := t.TempDir(), t.TempDir()
+	for i := 0; i < 50; i++ {
+		d := filepath.Join(src, "d"+strconv.Itoa(i))
+		if err := os.Mkdir(d, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		for j := 0; j < 400; j++ {
+			if err := os.WriteFile(filepath.Join(d, "f"+strconv.Itoa(j)), nil, 0o644); err != nil {
+				t.Fatal(err)
+			}
+		}
+	}
+	var before, after runtime.MemStats
+	runtime.GC()
+	runtime.ReadMemStats(&before)
+	_, err := mirrorTree(src, dst, 1<<20)
+	runtime.ReadMemStats(&after)
+	if !errors.Is(err, errTreeTooLarge) {
+		t.Fatalf("20000 empty files under a 1 MiB limit: %v", err)
+	}
+	got := after.TotalAlloc - before.TotalAlloc
+	t.Logf("refused: %v; %d KiB allocated", err, got>>10)
+	if got > 16<<20 {
+		t.Errorf("the copy allocated %d MiB", got>>20)
+	}
+	n := 0
+	_ = filepath.WalkDir(dst, func(string, os.DirEntry, error) error { n++; return nil })
+	if n > (1<<20)/512+1 {
+		t.Errorf("%d entries were written under a limit of %d", n, (1<<20)/512)
+	}
+}
+
+// Left-out entries are counted in full, and only the first
+// maxSkippedReported names are kept.
+func TestLeftOutEntriesAreCountedAndTheirNamesBounded(t *testing.T) {
+	src, dst := t.TempDir(), t.TempDir()
+	for i := 0; i < 3*maxSkippedReported; i++ {
+		if err := os.Symlink("x", filepath.Join(src, "l"+strconv.Itoa(i))); err != nil {
+			t.Fatal(err)
+		}
+	}
+	skips, err := mirrorTree(src, dst, maxCopiedTreeBytes)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if skips.count != 3*maxSkippedReported || len(skips.names) != maxSkippedReported {
+		t.Errorf("count %d, %d names kept", skips.count, len(skips.names))
+	}
+	res := map[string]interface{}{}
+	skips.report(res)
+	if res["worktree_skipped_count"] != 3*maxSkippedReported {
+		t.Errorf("reported %v", res)
+	}
+}
+
+func TestRepoGitAuthPutsTheTokenOnlyInTheEnvironment(t *testing.T) {
+	bare, env, err := repoGitAuth("https://host-user:tok3n0f7hisrun5a1e@git.example.invalid/o/r.git")
+	if err != nil || bare != "https://git.example.invalid/o/r.git" {
+		t.Fatalf("bare = %q, %v", bare, err)
+	}
+	joined := strings.Join(env, "\n")
+	if !strings.Contains(joined, "GIT_CONFIG_KEY_0=http.https://git.example.invalid/o/r.git.extraHeader") ||
+		!strings.Contains(joined, "GIT_CONFIG_VALUE_0=Authorization: Basic ") {
+		t.Errorf("env = %q", env)
+	}
+	if _, env, _ := repoGitAuth("/srv/x/origin.git"); env != nil {
+		t.Errorf("a path got credentials: %v", env)
+	}
+}
+
+// One folder holding very many names is refused once the names read pass
+// the limit, without the whole folder being read first.
+func TestAFolderOfVeryManyNamesIsRefusedWithoutReadingItWhole(t *testing.T) {
+	src := t.TempDir()
+	long := strings.Repeat("n", 190)
+	for i := 0; i < 30000; i++ {
+		if err := os.WriteFile(filepath.Join(src, long+strconv.Itoa(i)), nil, 0o644); err != nil {
+			t.Fatal(err)
+		}
+	}
+	saved := maxReturnedTreeBytes
+	maxReturnedTreeBytes = 1 << 20
+	t.Cleanup(func() { maxReturnedTreeBytes = saved })
+	for name, walk := range map[string]func() error{
+		"copy": func() error { _, err := mirrorTree(src, t.TempDir(), 1<<20); return err },
+		"tar":  func() error { _, _, err := packWorktree(src); return err },
+	} {
+		var before, after runtime.MemStats
+		runtime.GC()
+		runtime.ReadMemStats(&before)
+		err := walk()
+		runtime.ReadMemStats(&after)
+		got := after.TotalAlloc - before.TotalAlloc
+		t.Logf("%s: %v; %d KiB allocated", name, err, got>>10)
+		if !errors.Is(err, errTreeTooLarge) {
+			t.Errorf("%s: a folder of 30000 names under a 1 MiB limit: %v", name, err)
+		}
+		if got > 4<<20 {
+			t.Errorf("%s: %d KiB allocated; the 30000 names alone are about 6 MiB", name, got>>10)
+		}
+	}
+}

client-go/harness_treecap_unix_test.go

@@ -0,0 +1,71 @@
+//go:build !windows
+
+package main
+
+import (
+	"os"
+	"path/filepath"
+	"runtime"
+	"strings"
+	"testing"
+	"time"
+)
+
+// A file far larger than the tree a run may return (sparse, so it costs
+// nothing to make) is refused from its size: packing returns at once, with
+// an error and no tar, having read none of it.
+func TestPackingRefusesAFileLargerThanTheLaneCarriesWithoutReadingIt(t *testing.T) {
+	tree := t.TempDir()
+	if err := os.WriteFile(filepath.Join(tree, "a.txt"), []byte("a\n"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	f, err := os.Create(filepath.Join(tree, "huge.bin"))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if err := f.Truncate(64 << 30); err != nil {
+		t.Skipf("no sparse file here: %v", err)
+	}
+	f.Close()
+	type res struct {
+		b   []byte
+		err error
+	}
+	var before, after runtime.MemStats
+	runtime.ReadMemStats(&before)
+	ch := make(chan res, 1)
+	start := time.Now()
+	go func() { b, _, err := packWorktree(tree); ch <- res{b, err} }()
+	var r res
+	select {
+	case r = <-ch:
+	case <-time.After(20 * time.Second):
+		t.Fatal("packing a tree with a 64 GiB file was still going after 20 s")
+	}
+	runtime.ReadMemStats(&after)
+	t.Logf("returned in %v, %d MiB allocated", time.Since(start), (after.TotalAlloc-before.TotalAlloc)>>20)
+	if r.err == nil || !strings.Contains(r.err.Error(), "huge.bin") || !strings.Contains(r.err.Error(), "too large") {
+		t.Fatalf("want the tree refused for huge.bin, got %v", r.err)
+	}
+	if r.b != nil {
+		t.Fatal("a tar came back with the refusal")
+	}
+	if got := after.TotalAlloc - before.TotalAlloc; got > 64<<20 {
+		t.Fatalf("packing allocated %d MiB", got>>20)
+	}
+}
+
+// A result larger than one message on the lane is replaced by a refusal that
+// says so, never sent to be dropped by the host; one that fits goes as it is.
+func TestAResultTheLaneCannotCarryIsRefusedNotSent(t *testing.T) {
+	small := []byte(`{"ok":true,"mode":"worktree"}`)
+	if b, code := fitLane(small, 0); string(b) != string(small) || code != 0 {
+		t.Fatalf("a result that fits was changed: %s %d", b, code)
+	}
+	big := append([]byte(`{"ok":true,"mode":"worktree","worktree_b64":"`), make([]byte, uvicornWSMaxSize)...)
+	b, code := fitLane(big, 0)
+	if code == 0 || len(b) > 1024 || !strings.Contains(string(b), `"result_too_large":true`) ||
+		!strings.Contains(string(b), `"ok":false`) {
+		t.Fatalf("an oversized result was not refused: %d bytes, code %d: %.200s", len(b), code, b)
+	}
+}

client-go/hostops_test.go

@@ -484,7 +484,7 @@ func TestRepeatedRefusedWritesQuarantineThePairing(t *testing.T) {
 	for i := 0; i < quarantineStrikes; i++ {
 		driveJob(t, tAnchor, putJob(""), ctx, "")
 	}
-	if u := quarantinedUntil(tAnchor, time.Now()); u.IsZero() || u.Sub(time.Now()) > quarantineBase+time.Minute {
+	if u := quarantinedUntil(tAnchor, time.Now()); u.IsZero() || time.Until(u) > quarantineBase+time.Minute {
 		t.Fatalf("a cleared hold is forgiven — the next one is the base length again: %v", u)
 	}
 	// Strikes double when the hold lapses on its own.
@@ -494,7 +494,7 @@ func TestRepeatedRefusedWritesQuarantineThePairing(t *testing.T) {
 	for i := 0; i < quarantineStrikes; i++ {
 		driveJob(t, tAnchor, putJob(""), ctx, "")
 	}
-	if u := quarantinedUntil(tAnchor, time.Now()); u.IsZero() || u.Sub(time.Now()) < 2*quarantineBase-time.Minute {
+	if u := quarantinedUntil(tAnchor, time.Now()); u.IsZero() || time.Until(u) < 2*quarantineBase-time.Minute {
 		t.Fatalf("a repeat hold doubles: %v", u)
 	}
 }
@@ -1010,12 +1010,15 @@ func TestAnchoredEgressRefusesLoopbackAndPrivateDestinations(t *testing.T) {
 	if addr, why := egressVettedAddr(ctx, "localhost:80"); why == "" || addr != "" {
 		t.Fatalf("a refused name yields no address: %q %q", addr, why)
 	}
+	// A local box refuses this machine too, and the network around it
+	// (egress_local_test.go and egress_network_test.go have the rest).
 	t.Setenv("XDG_CONFIG_HOME", t.TempDir())
-	if why := egressInternalRefusal(ctx, "127.0.0.1:11434"); why != "" {
-		t.Fatalf("a local box keeps its lane: %s", why)
+	stubOwnAddrs(t, loopbackOnly, localRange())
+	if why := egressInternalRefusal(ctx, "127.0.0.1:11434"); why == "" {
+		t.Fatal("a local box must refuse its own ollama on the egress lane")
 	}
-	if addr, _ := egressVettedAddr(ctx, "example.com:443"); addr != "example.com:443" {
-		t.Fatalf("a local box dials the name it was given: %q", addr)
+	if addr, why := egressVettedAddr(ctx, "192.168.1.1:80"); why == "" || addr != "" {
+		t.Fatalf("a local box refuses its LAN on the egress lane: %q %q", addr, why)
 	}
 }