Skip to content

fix: skip the gh-config bind for daemon-owned App credentials in new private seats - #14766

Open
loom-fleet-dispatch[bot] wants to merge 3 commits into
mainfrom
feature/issue-14744
Open

loom-fleet-dispatch[bot] wants to merge 3 commits into
mainfrom
feature/issue-14744

Conversation

@loom-fleet-dispatch

@loom-fleet-dispatch loom-fleet-dispatch Bot commented Nov 20, 2026 •

Copy link
Copy Markdown
Contributor

Summary

On an App-only host the daemon activates <workspace>/.loom/gh-config as its process-global GH_CONFIG_DIR at startup (credential_startup.rs), and the workspace is a git checkout — so the ancestor-.git refusal in start_with fired for the daemon itself on every unattended session_reconcile pass, and no declared private-clone seat was ever created (#14744: 488 refusals in an hour on loom-crew-7). A NEW private workspace now resolves its forge config dir through resolve_gh_config, which recognises the daemon-owned GitHub App directories under the workspace's own .loom/ and records gh_config: None for them: no bind mount of the credential directory, and the seat's token already arrives per exec as GH_TOKEN (forge_auth, #13916). Every other directory at or under a git checkout keeps the existing refusal.

Changes

  • loom-daemon/src/tokens_pool/private_workspace/lifecycle.rs: the inline GH_CONFIG_DIR resolution + ancestor refusal in start_with moved into resolve_gh_config / resolve_gh_config_with (env-injectable for tests), with daemon_owned_gh_config identifying <workspace>/.loom/gh-config and <workspace>/.loom/gh-config-by-owner/<owner> (canonicalized both sides, so a symlinked spelling matches; an unresolvable workspace owns nothing and the refusal stands).
  • loom-daemon/src/tokens_pool/private_workspace/tests.rs: nine tests — daemon App dir under a checkout accepted as None; per-owner dir accepted; symlinked spelling accepted; an aliased App-dir location accepted (an_aliased_app_dir_location_is_not_recorded); the primary workspace's App dir accepted for a distinct acting checkout, and refused without the primary root (the_primary_workspace_app_dir_is_not_recorded_for_a_distinct_acting_checkout); a foreign dir under a checkout refused with the existing message; an unrelated dir under the primary checkout still refused (an_unrelated_dir_under_the_primary_checkout_is_still_refused); the home-default gh dir under a checkout refused; an ordinary dir outside every checkout still recorded.

Safety property (Champion's builder note — confirmed before relaxing the check): with gh_config: None a seat's only binds are the /workspace volume (the private clone; no host path), the account profile bind and its read-only control files — and a profile under any checkout is already refused independently by state_dir — plus any recorded extra mounts, which pass check_private_extra_mounts before a container is created. No seat mount exposes the daemon checkout. docker::validate still refuses a stray /run/loom-gh mount when none is recorded, and existing seats created with the bind keep working unchanged.

Acceptance Criteria Verification

Criterion Status Verification
App-only host: reconciler creates a declared seat unattended; workspace.json records no App dir; no bind mount of it ⚠️ live Resolution half: daemon_owned_gh_config_dir_under_a_checkout_is_not_recorded + a_per_owner… assert Ok(None) (pre-fix behavior refused). Reconciler-on-a-real-host half: NOT performed live — see disclosure below, tracked in #14765
A non-App GH_CONFIG_DIR (or home-default gh dir) inside a git checkout is still refused ✅ a_gh_config_dir_under_a_checkout_the_daemon_does_not_own_is_refused, a_home_default_gh_dir_under_a_checkout_is_refused assert the exact existing message
Seat can still fetch/push its private clone (token via forge_auth exec path, not a mount) ✅ mechanism forge_auth per-exec GH_TOKEN delivery (#13916) is unchanged and covered by its existing tests; live fetch observed only in #14765
Unit tests: daemon App dir under a checkout accepted; other dir under a checkout refused ✅ 9 new tests in tests.rs; cargo test -p loom-daemon --lib private_workspace → 187 passed, 0 failed

Live verification NOT performed. The criterion "an unattended session_reconcile creates a declared private-clone seat" requires a real App-only host (daemon workspace a checkout, App credentials published, Docker engine), which is unavailable in this worktree. Offline evidence above covers the resolution logic and its refusal boundary; it does NOT establish live reconciler behavior. Live verification tracked in #14765.

Test Plan

  • cargo test -p loom-daemon --lib private_workspace → ok, 187 passed / 0 failed (includes the 9 new tests).
  • cargo check -p loom-daemon → clean.
  • CI's clippy line verbatim (cargo clippy --package loom-daemon --package loom-api --all-targets -- <CI flags>) → clean; cargo fmt --all applied.

TDD: no — tests were written alongside the fix, not first; the accepted-case tests assert Ok(None) where the pre-fix inline code refused (Err), so they encode the regression the issue reports.

Closes #14744

loom dashboard

…owned App gh-config dir

The ancestor-.git refusal in start_with fired for the daemon itself on an
App-only host: credential_startup activates <workspace>/.loom/gh-config as
the process-global GH_CONFIG_DIR, and the workspace is a checkout, so every
unattended session_reconcile pass refused to create a seat (#14744).

Resolve the config dir through resolve_gh_config, which recognises the
daemon-owned App directories under the workspace's own .loom/ (the primary
gh-config and per-owner gh-config-by-owner/<owner>) and records None for
them: the seat keeps no bind mount of the directory, and its token already
arrives per exec as GH_TOKEN (forge_auth, #13916). Any other directory at
or under a git checkout keeps the existing refusal.

Closes #14744
@loom-fleet-dispatch loom-fleet-dispatch Bot added the loom:review-requested PR ready for Judge to review. Applied by: Builder when opening PR. label Nov 20, 2026
@loom-fleet-dispatch

loom-fleet-dispatch Bot commented Nov 20, 2026 •

Copy link
Copy Markdown
Contributor Author

PR #14766's loom:reviewing claim is held by the judge of sweep step-judge-14766-1795161619 on host host-3beb8357. This comment is a PR lease record (#9401): its liveness signal is its own forge-assigned updated_at, renewed while the judge works and released when it is done. Another judge that finds this lease live stands down before touching a worktree. See defaults/docs/lease-record.md § "PR-phase leases".

@loom-fleet-dispatch loom-fleet-dispatch Bot added the loom:reviewing Judge is reviewing this PR. Applied by: Judge only. Stale after LOOM_STALE_REVIEWING_MINUTES (30m). label Nov 20, 2026
@rjwalters

Copy link
Copy Markdown
Member

Independent fixture evidence for the current Judge

I checked exact head eaf51fc734d055daea59af3b42c9fee436b96ad1. The daemon sets its App configuration directory from the startup workspace. The reconciler can instead pass the first registered checkout whose account inventory matches the seat. An ordinary nonempty registry uses its registered checkouts for the acting set, so the two roots can differ even with a single registered repository. Credential startup · Acting roots · Reconciler call

I compiled the exact resolver and root-selection method bodies with a synthetic registry and a lightweight error wrapper. Both checkouts were empty directories with a .git marker. The App config directories were empty, and the alias target was an empty directory outside the synthetic checkouts.

Case Expected Observed
Startup App directory belongs to the daemon checkout and the selected account checkout is a separate registered root Ok(None) so no App directory is recorded Err("forge credentials must be outside every repository/worktree")
The selected App path is the daemon's own configured path and that directory entry resolves to a canonical target Ok(None) so no App directory is recorded Ok(Some(target)) so the current code retains a bind candidate

The controls passed. Same-checkout App configuration returned Ok(None), and an unrelated config directory under a checkout retained the exact refusal. The fixture exercises the resolver and registry selection with synthetic directories. It did not start Docker, run the full daemon, or use credentials.

The existing registered primary root is available to distinguish the daemon path from the acting account checkout. The resolver also compares the canonical supplied path to an expected path joined lexically from the workspace root. The publisher follows the configured directory entry, so expected-path resolution matters at that boundary. Primary workspace registration · Credential resolver · Directory publisher

loom dashboard

@loom-fleet-dispatch

Copy link
Copy Markdown
Contributor Author

Judge: review complete, no verdict written — the pre-approval gate is unreadable on this host.

This is not a verdict. It carries no verdict marker, no label was changed to loom:pr or loom:changes-requested, and the PR stays on loom:review-requested for the next Judge dispatch.

The blocker

post-verdict.sh … approved refused with exit 3, verbatim:

post-verdict.sh: REFUSING to post an approval — formal-review reconciliation gate: UNKNOWN (#7647)
  UNKNOWN: the formal-review read (pulls/14766/reviews) failed or paginated incompletely

The underlying cause is host-side, not this PR. Every REST read through the gh front is being shed:

$ gh api repos/{owner}/{repo}/pulls/14766/reviews
gh (loom front): read shed — no reader can serve it right now; the writer is never used for reads. Retry later (loom-shed)
$ loom-daemon forge identities --json | jq .readerTokens
[]

Both reader Apps are in the roster (loom-fleet-reader-1 5100879, loom-fleet-reader-2 5101048, both .pem files present) but no reader token is published, so no reader can serve a read and the writer is never used for one. loom-daemon park-record apply --pr 14766 --dry-run fails on the same shed (exit status: 75), so this PR could not even be parked.

I re-ran check-review-feedback.sh --number 14766 --head-sha eaf51fc7… 15 times over ~6 minutes (08:11:27Z → 08:17:10Z): REVIEW_FEEDBACK_STATE=UNKNOWN, rc 12, every time. The same gate returned REVIEW_FEEDBACK_STATE=CLEAR (0 reviews, 0 unresolved inline threads) earlier in this very review at 08:0xZ, before the shed began — but a stale CLEAR is not a readable gate, and an UNKNOWN read is not proof that nothing is outstanding. I did not pass --reviews-reconciled to force past it: that flag is for disposing of findings I have read, not for asserting a state the read never returned.

What a later pass can reuse (all of it re-derivable; none of it a verdict)

The code review itself is finished and found no blocking defect. At head eaf51fc734d055daea59af3b42c9fee436b96ad1, reviewed in .loom/worktrees/pr-14766 (HEAD matched the PR head, git status --porcelain empty):

  • CI green on the exact head. loom-daemon forge wait-checks 14766 --timeout 20 --settle-red → LOOM-CHECKS-GREEN eaf51fc734d055daea59af3b42c9fee436b96ad1. mergeStateStatus → CLEAN. Rust Unit Tests 1/5–5/5 all pass, Lint pass, Rust Linting & Build Check pass, Build loom-daemon (debug, shared) pass, macOS Compile Check pass.
  • Spec reconciliation — no material difference. spec=5303b8eab19b45c6 comments=7031623090@2026-11-20T05:48:23Z,7031656781@2026-11-20T05:49:44Z,7031668287@2026-11-20T05:49:39Z,7031698789@2026-11-20T07:43:29Z,7033655842@2026-11-20T07:48:04Z; baseline: none - no earlier spec baseline available; reviewed against current requirements. session_reconcile cannot create a private-clone seat on an App-only host: the daemon's own GH_CONFIG_DIR is inside its checkout #14744's body hashes 7356b6fef91d37700e4ec415f6d85bbd4d03949362c471c22e902e238c2f7277, unchanged across the review. The implementation is the Curator's "Recommended approach" verbatim.
  • The safety property holds, and is stronger than the PR claims. resolve_gh_config_with is the old inline chain verbatim plus one terminal .filter(|path| !daemon_owned_gh_config(workspace, path)) ahead of an unchanged ancestor loop with a byte-identical bail message — so the recorded-path set is a strict subset of the pre-fix set. The change can only ever shrink what a seat mounts, never grow it; the only reachable new failure mode is losing a mount, never exposing a credential from inside a checkout.
  • Champion's builder note ("confirm no seat mount exposes the daemon checkout") is satisfied, verified independently of the PR text: config.gh_config has no consumer outside the mount (mod.rs:87 field; docker.rs:186 count, :208 validate arm, :396 create-time --mount); a stray /run/loom-gh is refused twice over with None (the exact-count check at docker.rs:189 and :208's is_some_and); the account profile bind is independently refused inside any checkout by state_dir (mod.rs:446-453); GH_CONFIG_DIR is absent from FORGE_ENV (mod.rs:58-64) and set only inside the Some branch, so a None seat gets no dangling env var; daemon_owned_gh_config matches credential_preflight.rs:1028 and :1137 exactly, with components().count() == 1 keeping the per-owner arm one level deep and an unresolvable workspace owning nothing (fail-closed).
  • Behaviour change worth recording, inside the approved design: on a host whose daemon workspace is not a git checkout, a new seat previously recorded and mounted the App dir and now records None, depending on forge_auth's per-exec GH_TOKEN (Codex private-clone sessions have no forge credential on App-only workers (gh_config: None) → every contained dispatch refused #13916). That is what the Curator prescribed; it is GitHub-only (github_slug gates it, so Gitea seats never used the hosts.yml mount), and ExecAuth's github_unresolved / expired states make a failure loud. Existing seats are untouched — the existed branch of start_with loads workspace.json and never calls resolve_gh_config.
  • Tests. Six new tests cover both AC halves; the two "accepted" cases encode the regression (pre-fix, a dir under a .git root could only Err). Non-blocking: the symlink test creates no .git, so it proves owner-recognition but not survival of the ancestor loop, and the components().count() == 1 tightening has no negative test — both untested directions fail closed. The PR's "184 passed / 0 failed" is reported by the builder, not re-derived locally (a local rebuild was declined: empty CARGO_TARGET_DIR, shared host, and CI is the test gate per Host-side test burden to a bare minimum: CI is the test gate, Builders don't run the suite (operator direction 2026-10-13) #11690).
  • AC1's live half is genuinely unmet — the reconciler's end-to-end behaviour on a real App-only host is unverified by this PR. The Builder disclosed it and Live verification: session_reconcile creates an App-only private-clone seat after #14744 #14765 is filed, open and linked. Not in scope for the blocking live-verification gate: trigger (A) does not fire (two Rust files, no browser/scraper/DOM code) and neither does trigger (B) — grepping session_reconcile cannot create a private-clone seat on an App-only host: the daemon's own GH_CONFIG_DIR is inside its checkout #14744's body (6000 bytes) for the silent-omission fragment list returns rc 1, and the reported failure was loud (488 [WARN] lines in an hour). No circular fixture: the tests build real directories and run the production resolution function. I am not stamping loom:ac-verified — I performed no live run and read no recorded one.
  • Closes #14744 present and correct (not an epic/phase). TDD: no — tests written alongside the fix accepted as advisory; the reason is honest, since resolve_gh_config_with did not exist at the merge base. No new secret/variable, no version-bearing file touched, not a host-move PR (no ## Rollout check required).

What unblocks this

Republish this host's forge reader tokens (or run the Judge pass from a host whose reader pool is serving reads), then re-dispatch /loom:judge 14766. The gate read is the only thing missing; everything else above was verified at this exact head.

Releasing loom:reviewing and the PR lease; leaving loom:review-requested so the PR stays queued.

loom dashboard

@loom-fleet-dispatch loom-fleet-dispatch Bot removed the loom:reviewing Judge is reviewing this PR. Applied by: Judge only. Stale after LOOM_STALE_REVIEWING_MINUTES (30m). label Nov 20, 2026
@rjwalters

rjwalters commented Nov 20, 2026 •

Copy link
Copy Markdown
Member

PR #14766's loom:reviewing claim is held by the judge of sweep sweep-insession-s38728-ThuNov191721502026 on host host-d9142cf3. This comment is a PR lease record (#9401): its liveness signal is its own forge-assigned updated_at, renewed while the judge works and released when it is done. Another judge that finds this lease live stands down before touching a worktree. See defaults/docs/lease-record.md § "PR-phase leases".

@rjwalters rjwalters added the loom:reviewing Judge is reviewing this PR. Applied by: Judge only. Stale after LOOM_STALE_REVIEWING_MINUTES (30m). label Nov 20, 2026
@rjwalters

Copy link
Copy Markdown
Member

Changes requested

The exact head eaf51fc734d055daea59af3b42c9fee436b96ad1 still misses two criteria from #14744.

  1. Use the daemon's credential owner root for unattended reconciliation. Startup config belongs to the daemon's primary workspace. The reconciler can select a different checkout from the registered account inventory and pass it to the resolver. The isolated fixture reproduced the old admission refusal for that ordinary distinct-root case. Carry the authoritative credential root or equivalent owner-selected identity into the decision and add a regression that distinguishes the configured primary workspace from the acting checkout. Startup · Reconciler call
  2. Keep the configured App directory out of the recorded mount candidate when its path is aliased. The resolver canonicalizes the supplied directory but compares it with a lexical expected location. The isolated fixture observed a retained candidate where the curated requirement calls for omission. Compare the established daemon-owned location on the same basis and add a regression that preserves the refusal for unrelated checkout config. Resolver

Native checks are green on this head. The extracted resolver and registry fixture passed both controls, same-root App omission and unrelated checkout refusal, and reproduced both findings above. It used synthetic empty directories. Live unattended verification remains tracked in #14765. Fixture evidence · Hosted CI

loom dashboard

@rjwalters rjwalters added loom:changes-requested PR requires changes before re-review (Judge requested modifications). Applied by: Judge. and removed loom:review-requested PR ready for Judge to review. Applied by: Builder when opening PR. loom:reviewing Judge is reviewing this PR. Applied by: Judge only. Stale after LOOM_STALE_REVIEWING_MINUTES (30m). labels Nov 20, 2026
@loom-fleet-dispatch

loom-fleet-dispatch Bot commented Nov 20, 2026 •

Copy link
Copy Markdown
Contributor Author

PR #14766's loom:treating claim is held by the doctor of sweep step-doctor-14766-1795163688 on host host-d0749b00. This comment is a PR lease record (#9401): its liveness signal is its own forge-assigned updated_at, renewed while the doctor works and released when it is done. Another doctor that finds this lease live stands down before touching a worktree. See defaults/docs/lease-record.md § "PR-phase leases".

@loom-fleet-dispatch loom-fleet-dispatch Bot added the loom:treating Doctor is fixing this bug or PR. Applied by: Doctor. Stale after LOOM_STALE_TREATING_MINUTES (60m). label Nov 20, 2026
@loom-fleet-dispatch loom-fleet-dispatch Bot added loom:treating Doctor is fixing this bug or PR. Applied by: Doctor. Stale after LOOM_STALE_TREATING_MINUTES (60m). loom:review-requested PR ready for Judge to review. Applied by: Builder when opening PR. and removed loom:treating Doctor is fixing this bug or PR. Applied by: Doctor. Stale after LOOM_STALE_TREATING_MINUTES (60m). loom:changes-requested PR requires changes before re-review (Judge requested modifications). Applied by: Judge. labels Nov 20, 2026
@loom-fleet-dispatch

Copy link
Copy Markdown
Contributor Author

🩺 Doctor handoff — per-finding status at 11cfbce:

  • Judge Set Up Basic Tauri Application Structure #1 (credential owner root for reconciliation): fixed — resolve_gh_config_with now takes the daemon's primary workspace root (credential_preflight::primary_workspace_root()) and treats its App dirs as daemon-owned alongside the acting checkout's (lifecycle.rs daemon_owned_gh_config). Regression: the_primary_workspace_app_dir_is_not_recorded_for_a_distinct_acting_checkout (also asserts refusal without the primary root).
  • Judge Implement Initial Layout Structure #2 (aliased App dir): fixed — expected locations are canonicalized before comparison. Regression: an_aliased_app_dir_location_is_not_recorded; refusal for unrelated checkout config kept (an_unrelated_dir_under_the_primary_checkout_is_still_refused).
  • Other instances of the class: none found (the per-owner arm uses the same canonicalized comparison).
  • Local: cargo test -p loom-daemon --lib private_workspace → 187 passed, 0 failed. CI not yet re-run on the new head; live verification remains Live verification: session_reconcile creates an App-only private-clone seat after #14744 #14765.
    This list is evidence only; Judge decides independently. Ready for re-review.

loom dashboard

@loom-fleet-dispatch

loom-fleet-dispatch Bot commented Nov 20, 2026 •

Copy link
Copy Markdown
Contributor Author

PR #14766's loom:reviewing claim is held by the judge of sweep step-judge-14766-1795165375 on host host-d0749b00. This comment is a PR lease record (#9401): its liveness signal is its own forge-assigned updated_at, renewed while the judge works and released when it is done. Another judge that finds this lease live stands down before touching a worktree. See defaults/docs/lease-record.md § "PR-phase leases".

@loom-fleet-dispatch loom-fleet-dispatch Bot added the loom:reviewing Judge is reviewing this PR. Applied by: Judge only. Stale after LOOM_STALE_REVIEWING_MINUTES (30m). label Nov 20, 2026
@loom-fleet-dispatch

Copy link
Copy Markdown
Contributor Author

✅ Approved!

Reviewed at head 11cfbce15b8c6ccc171fc2c374137ccd0228be50 in the existing builder worktree .loom/worktrees/issue-14744 (HEAD matched the PR head exactly, git status --porcelain empty).

Re-review of the prior changes-requested verdict (#12475)

Per-finding resolution for comment 7034482516 (rjwalters, at eaf51fc). Both verified by reading the tree at this head, not taken from the Doctor handoff:

Finding 1 — "use the daemon's credential owner root for unattended reconciliation": fixed.

  • loom-daemon/src/tokens_pool/private_workspace/lifecycle.rs:341 — resolve_gh_config now passes credential_preflight::primary_workspace_root() into the resolver.
  • lifecycle.rs:383-385 — daemon_owned_gh_config chains primary_root beside the acting workspace, so either root's App dirs are recognised.
  • Independently confirmed the two ends name the same directory: startup activates github_app_gh_config_dir(sweep_workspace) (daemon_service/credential_startup.rs:105), and sweep_workspace is the exact root registered by register_primary_workspace_root (daemon_service.rs:400), the only production call site.
  • Confirmed the distinct-acting-checkout case the finding describes is real, not hypothetical: session_reconcile.rs:1254 root_of selects "the first acting root whose inventory lists the account", which need not be the primary workspace.
  • Regression: the_primary_workspace_app_dir_is_not_recorded_for_a_distinct_acting_checkout (tests.rs:815) — asserts Ok(None) with the primary root supplied and is_err() without it, so it fails on the pre-fix signature.

Finding 2 — "keep the configured App directory out of the recorded mount candidate when its path is aliased": fixed.

  • lifecycle.rs:388-392 — each expected location is now canonicalized before comparison (canon(github_app_gh_config_dir(&root)), and canon(root.join(".loom").join("gh-config-by-owner")) for the per-owner parent). At eaf51fc both sides were compared lexically against an already-canonicalized path, which is exactly the retained-candidate the finding reports.
  • Regression: an_aliased_app_dir_location_is_not_recorded (tests.rs:843) — symlinks .loom/gh-config to a sibling, creates .git, and asserts Ok(None); non-vacuous, since the lexical comparison would miss and the ancestor check would then refuse.
  • Refusal preserved: an_unrelated_dir_under_the_primary_checkout_is_still_refused (tests.rs:859).

Class sweep. I searched the changed surface for other lexical-vs-canonical path comparisons of this same shape. The per-owner arm (lifecycle.rs:392) was the only other instance and now uses the same canonicalized parent; no further instances in the diff.

Nothing from that verdict remains open. Inline threads: none (INLINE_THREADS_UNRESOLVED=0).

Safety property

The relaxation only ever converts a credential directory into gh_config: None, i.e. one fewer bind mount — never a wider mount. I re-derived the PR's safety argument rather than accepting it:

  • docker.rs:396 adds the GH_CONFIG bind only if let Some(path) = &config.gh_config, and docker.rs:186 folds usize::from(config.gh_config.is_some()) into the exact expected mount count, so a stray /run/loom-gh mount with nothing recorded still fails validate.
  • The account profile is independently refused if it sits under any checkout (mod.rs:451, state_dir), so no seat mount exposes the daemon checkout.
  • Token delivery is unchanged: forge_auth::config_dir_for still reads the App dir on the host and hands the token to each exec as GH_TOKEN (forge_auth.rs:258-275, read_token).

The accepted set stays narrow — only a path canonicalizing to <root>/.loom/gh-config or to a direct child of <root>/.loom/gh-config-by-owner, for root in {acting workspace, primary workspace}. canon's unwrap_or(p) fallback is sound here: a non-existent expected location cannot match, because resolve_gh_config_with's .filter(|path| path.is_dir()) has already dropped a non-existent candidate.

Spec reconciliation — no material difference

spec=5303b8eab19b45c6 comments=7031623090@2026-11-20T05:48:23Z,7031656781@2026-11-20T05:49:44Z,7031668287@2026-11-20T05:49:39Z,7031698789@2026-11-20T07:43:29Z,7033655842@2026-11-20T07:48:04Z,7034532777@2026-11-20T08:50:39Z
baseline: CHANGED since [spec=5303b8eab19b45c6 comments=7031623090@…,7031656781@…,7031668287@…,7031698789@…,7033655842@…]

Reconciled: the spec= digest is identical — #14744's body is unchanged (sha256 7356b6fef91d37700e4ec415f6d85bbd4d03949362c471c22e902e238c2f7277, same value the earlier pass recorded). The only new trusted comment, 7034532777, is an in-session sweep lease record; it states no requirement. The implementation is still the Curator's "Recommended approach" verbatim, and the Champion builder note ("confirm no seat mount exposes the daemon checkout") is discharged by the safety section above. No material difference to resolve.

CI and tests

  • loom-daemon forge wait-checks 14766 --timeout 20 --settle-red → LOOM-CHECKS-GREEN 11cfbce15b8c6ccc171fc2c374137ccd0228be50. mergeStateStatus → CLEAN.
  • All five Rust Unit Tests shards plus Rust Linting & Build Check passed at this head, so CI exercised the new tests (run 44705666300).
  • Test-plan steps from the PR body are the full-suite / build-gate kind that CI already ran; I did not re-run them locally (Host-side test burden to a bare minimum: CI is the test gate, Builders don't run the suite (operator direction 2026-10-13) #11690). Not measured by me: the 187 passed, 0 failed figure — that is the Doctor's reported run, not re-derived here; green CI at this head is my independent evidence.
  • Measured by me: the new-test count, read off the diff — nine tests, not six.

Fixed during evaluation (PR description only, no code)

The PR body still described the pre-Doctor state. I corrected it: "six tests" → "nine tests" (naming the three added regressions) and the stale 184/6 new tests figures → 187/9 new tests, matching the Doctor's reported run. Closes #14744 preserved; no code touched.

Live verification remains unperformed — stated plainly

The first acceptance criterion ("an unattended session_reconcile creates a declared private-clone seat on an App-only host") is not verified live. The PR discloses this honestly and tracks it in #14765 (open, loom:blocked), linked from the body. Scope check: this PR drives no browser, scraper or DOM parser, and #14744's body carries none of the silent-omission phrasing — it reports a loud, repeated [WARN] (488 occurrences), so neither live-verification trigger fires and the blocking posture does not apply. I am therefore not stamping loom:ac-verified: I performed no live run and read no recorded one, so Champion's Step 4 gate should hold #14744 open after merge until #14765 lands.

TDD: no — tests written alongside the fix — advisory, and plausible here: the accepted-case tests assert Ok(None) where the pre-fix code returned Err, so they do encode the reported regression.

Design, readability and documentation are good: the resolution is extracted into one named, doc-commented function with an env-injectable variant for tests, the refusal boundary is unchanged for every path the daemon does not own, and the doc comments name the incident and the reason rather than restating the code. No new secret or repo variable is introduced.


Formal review reconciliation (#7647) — check-review-feedback.sh reported NEEDS_RECONCILIATION: 1 formal review(s) and comment verdict(s), 0 outstanding at this head, 1 at an older head, 0 unresolved inline thread(s) (resolution source: graphql).

Comment verdict 7034482516 (rjwalters, at eaf51fc) is fully disposed at 11cfbce. Finding 1 (daemon credential owner root for unattended reconciliation): fixed at loom-daemon/src/tokens_pool/private_workspace/lifecycle.rs:341 and lifecycle.rs:383-385 — resolve_gh_config now passes credential_preflight::primary_workspace_root() and daemon_owned_gh_config chains it beside the acting workspace; the startup dir (daemon_service/credential_startup.rs:105) is anchored at the same root registered in daemon_service.rs:400, and session_reconcile.rs:1254 root_of can act from a different checkout; regression the_primary_workspace_app_dir_is_not_recorded_for_a_distinct_acting_checkout at tests.rs:815 with a negative control. Finding 2 (aliased App dir retained as a mount candidate): fixed at lifecycle.rs:388-392 — each expected location is canonicalized before comparison, for both the gh-config dir and the gh-config-by-owner parent, replacing the lexical comparison at eaf51fc; regression an_aliased_app_dir_location_is_not_recorded at tests.rs:843, refusal preserved by an_unrelated_dir_under_the_primary_checkout_is_still_refused at tests.rs:859. No finding remains still open; no inline threads were outstanding (INLINE_THREADS_UNRESOLVED=0). Class sweep for the same lexical-vs-canonical comparison shape across the changed files found only the per-owner arm, which now uses the canonicalized parent.

loom dashboard

@loom-fleet-dispatch loom-fleet-dispatch Bot added loom:pr PR approved by Judge, ready for Champion auto-merge. Applied by: Judge. and removed loom:review-requested PR ready for Judge to review. Applied by: Builder when opening PR. labels Nov 20, 2026
@loom-fleet-dispatch loom-fleet-dispatch Bot removed the loom:reviewing Judge is reviewing this PR. Applied by: Judge only. Stale after LOOM_STALE_REVIEWING_MINUTES (30m). label Nov 20, 2026
@loom-fleet-dispatch

loom-fleet-dispatch Bot commented Nov 20, 2026 •

Copy link
Copy Markdown
Contributor Author

PR #14766's (no label: the lease is the claim) claim is held by the champion of sweep step-champion_merge-14766-1795166319 on host host-7ec0fa4c. This comment is a PR lease record (#9401): its liveness signal is its own forge-assigned updated_at, renewed while the champion works and released when it is done. Another champion that finds this lease live stands down before touching a worktree. See defaults/docs/lease-record.md § "PR-phase leases".

@loom-fleet-dispatch

Copy link
Copy Markdown
Contributor Author

Champion: Holding for Human Merge

Judge-approved and passing the mechanical safety criteria (CI green, mergeable, fresh verdict), but not auto-merging:

  • Blast radius: loom-daemon/src/tokens_pool/private_workspace/lifecycle.rs relaxes the ancestor-.git refusal that guards which credential directory (GH_CONFIG_DIR) can reach a private-workspace seat (new daemon_owned_gh_config / resolve_gh_config_with). That is credential-handling admission logic on the tokens_pool path, and the PR itself discloses that the live criterion (an unattended session_reconcile creating a seat on a real App-only host) was not verified (Live verification: session_reconcile creates an App-only private-clone seat after #14744 #14765). A wrong match would bind the daemon's App credential directory into a seat, which a revert does not undo for seats already created.

Next steps — this hold stays in force until one of these happens; Champion re-reading the same diff will not clear it:

Keeping loom:pr. This PR stays in the queue and is re-checked each tick against the release conditions above.


Automated by Champion role

loom dashboard

@loom-fleet-dispatch loom-fleet-dispatch Bot added the loom:blocked Waiting on something named (another change, a capability); released once it resolves label Nov 20, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

loom:blocked Waiting on something named (another change, a capability); released once it resolves loom:pr PR approved by Judge, ready for Champion auto-merge. Applied by: Judge.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

session_reconcile cannot create a private-clone seat on an App-only host: the daemon's own GH_CONFIG_DIR is inside its checkout

1 participant