Repository navigation
Conversation
|
Note Reviews pausedIt looks like this branch is under active development. To avoid overwhelming you with review comments due to an influx of new commits, CodeRabbit has automatically paused this review. You can configure this behavior by changing the
Use the following commands to manage reviews:
Use the checkboxes below for quick actions:
No actionable comments were generated in the recent review. Walkthrough
Merge Risk:
|
|
✅ Deterministic PR hygiene checks passed. |
|
@coderabbitai review |
Action not completed
|
|
@coderabbitai review Please review the latest head |
Action performed
|
Preserve POSIX recorded-path trust checks and current Windows census and packed-command behavior. Preserve both sides of the test-layout registry without changing its established formatting. Validation: focused synthetic/loopback coverage, typecheck, structure and privacy checks; file-size gate passed after retaining the upstream registry format. Independent conflict-resolution review found no blocker.
|
No description provided. |
|
No description provided. |
Keep the 1007 and 4005 metadata-format assertions without requiring root-only sticky regular-file permissions on Darwin. Restore owner access to the directory before fixture cleanup. Focused discovery and trust tests: 74 pass, 18 platform skips, 393 assertions. Typecheck, privacy, size and whitespace checks passed; independently reviewed.
|
Review-thread audit: no unresolved Devin Review, CodeRabbit, or Codex connector threads remain on this PR. Fixed on head: the macOS ACL trust-boundary finding, the census fixture trust-tool finding, the BSD |
|
@coderabbitai review |
Action performed
|
Pack each version's two local packages in one real npm invocation and reuse only Node's compiled code within the fixture-owned root. Keep fresh per-command npm config and package caches, the separate real installation, existing readback assertions, and CI budgets unchanged.
Keep all five shards, test deadlines, assertions and required gates intact. Raise only the full Linux job ceiling to 18 minutes after healthy complete batch output was cancelled at finalization. Preserve historical release evidence policies and pin the reviewed workflow snapshot.
|
@coderabbitai review |
Action performed
|
|
Closing this branch in favor of two focused upstream reviews on
Both upstream PRs remain open for review. |
Summary
Remote Link treats a recorded runtime path as a discovery hint. Previously its first version probe could execute the recorded binary or package entrypoint before checking who controls that path.
lscan silently treat an ACL read failure as an absent ACL. Reusing macOS hints needs a native ACL reader that distinguishes those outcomes along the entire path. Existing Desktop, system and PATH discovery continues.Addresses the Daybreak finding and the macOS ACL review finding.
Historical reviewed checkpoint: base
c36793a9336b70cfe5bc3b7f6e6f751e730fb6fc, headb689549a97d0298db567aeb1cfce7c436fc9c83c, independently reviewed tree487f1b2e00fccaf74574ac5825744347d821a43c. The latest integrated head and its evidence are listed below.The earlier implementation and verification below used base
e87572127d7804216ebfcf0feecaf4c22c181748and tree8beec34186cc08bd34d214a233425c5dcfcc5d15.Compatibility and limits
A macOS installation available only through its runtime record is no longer selected or detected. Add it to the SSH account's PATH or install it in an existing discovery location. Full-install census, including the
--require-ephemeralcheck, can only detect the remaining known install locations on macOS; it may proceed when a record-only full install exists. This change does not execute or remove that undiscovered installation.This change removes the extra execution authority provided by arbitrary census paths. Existing Desktop/system locations and the SSH account's PATH retain their prior executable/version checks; this PR does not add ownership or ACL validation to those fallback locations. A path rejected through the census can still execute if it is independently selected through one of those locations. A controlled writable-Desktop fixture produces the same calls across all four consumers with and without a census record, confirming that this is an existing fallback policy rather than authority granted by the record.
The checks do not claim protection against root or unrestricted same-account mutation. The shell checks do not bind an opened executable descriptor through execution.
Verification
Current integration — 2026-11-19
Current head:
48a43573170eb6df3d0f0d9f0f28f90ed6e26a3d. It includes integration commit7f21d442813c8b4779b9367126bb2c505f513849, which normally merges the prior PR head374867f6a068c50beaa2ec4e17e57876c88c87ccwith current dev80e8952357d11ccf1421b4aa5620e781625dcdb0. The published GitHub tree exactly matches the locally verified tree709c9f66e746dafba75a253cbc17fac654320f1a.The actual test-layout conflict was resolved by retaining all 3,340 dev mappings without value changes, plus this PR's one trust-test mapping. Independent source review confirmed that both changed runtime files and all five related regression files are byte-identical to the preceding PR head. The automatically integrated English/Korean documentation retains the same trust contract.
This head also applies the measured-duration refresh from upstream PR #8940. The previous timing table assigned 60 unrecorded completed files only 8.58 seconds although their observed total was about 170.5 seconds. The refreshed table uses completed hosted-file samples and preserves all 20 newer dev timing rows exactly, including the split Link stop tests. Final TSV SHA-256:
2d56880fe64338d36b424bb73178b025078ca1482da5dd4be6e73c6792759be5; Git blob:a07f90e6098082e1452ea5cc075c801271321eca. That data-only refresh preserved the shard count, job limits, assertions, required gates and test selection. The later bounded outer-job budget change is documented separately below.bun test tests/ci-workflows/ci-shard-balance.test.ts tests/ci-workflows/ci-crash-disposition.test.ts tests/clients/link-served-runtime-trust.test.ts tests/clients/link-child-update-launcher.test.ts tests/clients/link-gated-ocx-argv.test.ts tests/clients/link-pinned-sync-argv.test.ts tests/clients/link-ssh-argv-discovery.test.ts tests/test-layout.test.ts tests/test-layout-tooling.test.ts: 244 passed / 19 Windows-related skips / 0 failed / 7,924 assertions.bun run typecheck: 2,625 files, no errors.node .github/scripts/release-evidence.cjs policy-checkand the exact CI commandnode --test .github/scripts/release-dispatch-guard.test.cjs .github/scripts/release-evidence.test.cjs .github/scripts/release-history.test.cjs .github/scripts/release-resume.test.cjs .github/workflows/deploy-docs.test.cjs: 69 passed / 0 failed / 0 skipped, including the fixed historical Git trees. The original production-pin mismatch was reproduced before the change.bun run structure:check,bun run privacy:scan, and whitespace checking against dev passed.Runs used the normal repository preload, its unmodified network guard, isolated temporary filesystems and local fixtures. No live provider, live SSH endpoint or installed application was exercised. The documented focused-validation exception applies because concurrent worktrees share resources. Full local tests, native macOS/Windows acceptance and live-provider acceptance are not claimed.
The earlier policy-only follow-up appends the documented release-evidence fingerprint snapshot for the new TSV. All 26 historical approval records remain unchanged. The new record changes only the TSV hash; the other 26 pins, version and required platform jobs are identical. Independent security review approved policy blob
51c1f898f2bdd69b7a012bd191ac51acb3c1779b. That policy-only change preserved runtime, tests and documentation; the later fixture-only and outer-job budget changes are documented separately below.Fixture preparation cost follow-up —
99bdab2bBoth attempts of run 44463273114 for preceding head
93eb2c1bfailed to complete. In attempt 1, shards 2/5 and 3/5 reached the unchanged 15-minute job limit in batches 60/63 and 57/65. One evidence-based failed-job-only retry was allowed; it again timed out in batches 60/63 and 64/65, with no failed assertion before cancellation. No further retry of that head was started. The release-contract gate and the other three test shards passed, but those successes do not make either attempt green.Commit
99bdab2baddresses repeated fixture startup cost in one test helper,tests/helpers/codex-cli-apply-fixture.ts. It does not change production runtime, regression assertions, test selection, the TSV, job counts, limits or release policies.A controlled offline A/B used three independent fixtures per variant on Node 24.22.3. Baseline setup took 4,277 / 3,597 / 3,699 ms, and the new helper took 1,889 / 1,744 / 1,830 ms (median 3,699 → 1,830 ms). All four npm-created archive hashes, the installed version, empty installer trace and independent command workspaces were checked: 30 assertions passed. This measures local preparation cost; it is not a guarantee about hosted job duration.
Focused existing regression command:
bun test tests/codex-integration/codex-cli-update-apply.test.ts tests/codex-integration/codex-cli-update-process-fixture.test.ts tests/codex-integration/codex-cli-update-plan.test.ts tests/codex-integration/codex-cli-update-artifacts.test.tsreported 141 passed / 1 platform skip / 1 failed / 800 assertions. The failed management lost-response case stops atwaitForDurableAdmission; the same single case was reproduced with the original helper in the same local sandbox, failing at the same point after the same eight assertions. It is not counted as passing. Hosted run 44468943739 at99bdab2balso passed that same case with the new helper in 11,535.38 ms; that observed result does not erase the local sandbox failure or the hosted job cancellation described below. The actual installer, lease, integrity/readback, argument and production cache-isolation cases passed locally.The integrated tree also passed
bun run typecheck(2,625 files), privacy scanning and whitespace checks. Independent source/security review checked the actual npm implementation's output order, distinct archive names including Windows, all path consumers, cache and cleanup ownership, and unchanged assertions. Reviewed helper blob:5cf138bef3281035b161f196966044efe8df685c; reviewed patch SHA-256:407d0022838a1b2a87e85a1c0eb4fbb58490d108961bdfd01af1b51811ba1056.Bounded outer-job budget follow-up —
48a43573Run 44468943739 at
99bdab2bremained unsuccessful at the 15-minute outer job limit, even though both cancelled shards emitted their complete test results:The follow-up changes exactly three files. The full Linux job's bounded outer allowance increases 15 → 18 minutes, including checkout, dependency setup, GUI build and cleanup. All five shards, the default 120-second batch deadline, existing longer per-file deadlines including 300-second conformance cases, assertions, selection and required gates remain intact. No retry-to-green mechanism or trigger, permission, runner-trust or platform-coverage change is introduced. The workflow test still asserts the exact job-owned limit, now 18.
The release evidence policy retains all 27 historical records and appends one reviewed snapshot. Only the workflow fingerprint changes; the other 26 pins, version and 14 Windows jobs stay identical. New policy blob:
7df235ec7904dd19a4b2e588ecfba1da627e9972. The exact three-file common candidate passed 251 CI-contract tests / 9,573 assertions, 69 Node release-contract tests, type checking and policy-check. Independent manual security review verified the complete parsed YAML is otherwise identical, the runner bytes and fail-closed evidence requirements are unchanged, and the historical records remain intact. Review patch SHA-256:3e18ab649f9766f74859c7eb9ae2f400f1eddbc8d8c4496f77a22fff5f675f7b. No automatic Cloud scan is claimed.The fixture helper
5cf138be, duration tablea07f90e6and all runtime/security regressions are unchanged by this follow-up. The actual CI merge07ba0c931f68af0147e324145b2379210621fc35has the same tree709c9f66e746dafba75a253cbc17fac654320f1aas the reviewed branch.Current exact-head main CI 44472490973 and Catalog 44472490959 both completed successfully. The main run passed on attempt 1, with no retry at this head. All five full Linux shards, gates, documentation, structure, storage, API usage, npm, keyring, Docker and the Linux desktop packaging/E2E job succeeded. The final
ciaggregator confirmed every job requested by this PR event succeeded at 2026-11-19 06:50:19 UTC.This execution verifies the published tree. All five jobs happened to finish below 15 minutes on these hosts, so the result is not attributed solely to the revised 18-minute ceiling and does not guarantee future runner speed. The earlier failed runs remain recorded above. Native macOS/Windows acceptance and maintainer security approval remain separate requirements; skipped platform jobs are not claimed as executed.
Intermediate hosted evidence —
7f21d442813c8b4779b9367126bb2c505f513849Run 44461327832 executed the measured-duration table on actual hosted runners. All 3,359 general files / 327 batches completed across the five Ubuntu test jobs, reporting 71,580 passed / 0 failed in aggregate.
Every job finished within the unchanged 15-minute limit. This is observed execution evidence for the intermediate head, not a guarantee about future runner speeds. Docker, documentation, structure, storage, API-usage and npm-global jobs also passed.
The overall run was not green: its release-contract gate failed because the production policy still pinned the old TSV blob. The final follow-up above fixes that mismatch through the documented snapshot append, preserving the original assertion. This intermediate run was allowed to finish so its actual test outcomes were retained before the new commit was pushed.
Earlier exact-head CI and investigation
At preceding head
374867f6a068c50beaa2ec4e17e57876c88c87cc, main CI 44448708478 concludedcancelledand the aggregate check failed. Shard 2/5 reached the 15-minute job limit in batch 56/61; its first 55 batches reported0 fail, with no failed assertion before cancellation. Its unfinished batches are not claimed as tested. Shard 3/5 completed all 65/65 batches with0 fail, but the job was cancelled at finalization. Shards 1/5, 4/5 and 5/5, and Catalog async contracts, passed.The slow-file investigation found stale timing inputs and runner variation. It did not establish a product timer leak or an unnecessary wait that could safely be removed. The intentionally bounded Lab, scheduler and configuration-lock cases retain their existing contracts. No blind retry of this head was started. Timing replay is an allocation estimate and leaves setup, process-boundary overhead and runner variation to actual CI.
The still earlier head
b689549a97d0298db567aeb1cfce7c436fc9c83chad incomplete main CI, attempt 2: shard 2/5 reached the 15-minute job limit in batch 60/63; the other four shards passed. A failed-job retry had already been used. The original 20-case security regression produced 17 failures / 3 compatibility controls on unfixed code, and the subsequent platform correction produced 10 new failures while retaining 20 Linux controls on its preceding version. The local platform checks use Linux fixtures and BSD-format / Darwin shims, not native macOS or Windows execution.Checklist
The last item records implementation and independent source review, not maintainer approval.
Summary by CodeRabbit
PATH.