Repository navigation
Run both Atlas browser regressions with strict zero-skip evidence - #702
Conversation
Require exact two-test execution with zero skips, retain evidence, and self-test the runner. Actual hosted UI qualification remains pending.
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
|
FINITE PERSONAL ENGINEERING REVIEW PICKUP — Actual performer: OpenAI/Codex Architecture Task Sweep Exact object, paths and output
Review questions and obligations Exclusions and stops Authority/onboarding links: CBZ governance · owner mandate · standing delegation · workflow · review-family decision. Link delivery is not acknowledgement. |
d6g8k5htny-coder
left a comment
There was a problem hiding this comment.
ATS-CBZ702-ATLAS-BROWSER-REVIEW-20261029 — AMEND_SCOPED_ENGINEERING
Exact reviewed PR-reported base/head: b1d0a81d1223b4de5691e543163e1962aa5c7c79 / 625bfc89975efa1fd725076bc2b0961af2768279. Hosted tested merge: 6abed714e07816e74be07ec5cd3fd6f94e6d046e with those exact parents. Nonauthor same-family engineering review; organizational/scientific-independence credit 0; scientific effect NONE.
Blocking finding
ATS-CBZ702-F1 — P1: the focus regression chooses the unsupported hosted launcher, so the candidate does not execute both required tests.
The changed focus test defines CHROMIUM = shutil.which("chromium") or shutil.which("google-chrome"), while the already-working reader test prefers google-chrome. In exact run 40936994120, job 123670050520, the runner selected /usr/bin/chromium for the focus method. With the newly required chromium_sandbox=True, that launcher aborted with “No usable sandbox” in both normal and optimized modes. In each mode, the reader method passed and the focus method errored; the exact two-method step and therefore browser-smoke concluded failure.
This is not a reason to disable the sandbox. Bind both tests to the supported hosted browser selection (or otherwise configure a genuinely sandbox-capable launcher), preserve strict zero-skip/error semantics, and obtain a fresh successful hosted execution on the repaired exact head/current tested merge. The current failure and its evidence must remain retained.
Positive bounded evidence
- Fresh isolated source review found the fixed-ID runner fail-closed on inventory mismatch, skip/failure/error/subtest/expected-failure/unexpected-success, missing execution, reused output, source drift, and wrong exit/report state.
- Exact-head controls: 15/15 normal and 15/15 optimized passed; the six declared runner-source mutants were killed by their intended assertions with no test errors/skips.
git diff --check b1d0a81d…625bfc89is clean.- The workflow runs both real modes even after the first fails, preserves stdout/stderr and the always-uploaded browser artifact, and returns failure when either mode fails.
- Native readback: seven other checks completed successfully, but
browser-smokefailed; no blanket green claim is available. The configured Codex review completed without findings, but the native browser failure is controlling evidence. - Current Main is
c93cec70223281492709ff74ca84fed5dc6e40e8and changes onlydocs/RECON_NOVELTY_20260925.mdrelative to the reported base, with no path overlap. Nevertheless, the PR is behind and no current-Main tested merge or check rebinding is claimed.
Verdict remains AMEND until F1 is repaired and fresh exact-head hosted browser evidence succeeds. No source/ref/PR-body/workflow/protection/Ready/check-dispatch/merge/integration/canonical-root/status/proof/science/CBZ-0/scheduler mutation was performed. ONLY-OWN review role released.
Fresh reconstruction and qualification after workspace loss; preserve all UI assertions, sandbox requirements, and current Main changes.
|
FINITE PERSONAL FOLLOW-UP ENGINEERING REVIEW PICKUP — Actual performer: OpenAI/Codex Architecture Task Sweep Exact object, source and output
Review questions and required evidence Exclusions and stops Authority/onboarding links: CBZ governance · owner mandate · standing delegation · workflow · review-family decision. Link delivery is not acknowledgement. |
d6g8k5htny-coder
left a comment
There was a problem hiding this comment.
ATS-CBZ702-ATLAS-BROWSER-REPAIR-REVIEW-20261029B — PASS_SCOPED_ENGINEERING_REPAIR / INTEGRATION_CHECKS_NONTERMINAL
Exact reviewed base/head: c93cec70223281492709ff74ca84fed5dc6e40e8 / c0205ca9e8683b2847008f712582e183f723e186. Observed GitHub synthetic merge: 8bd88e5952e4a67ff394e1cab3c272e7ace2f961, tree c9e3a5656fc95feef3bd360af3ea393846afd1e8, parents exactly base/head above. Nonauthor/nonintegrator same-family engineering review; organizational/scientific-independence credit 0; scientific effect NONE.
Prior finding disposition
ATS-CBZ702-F1 is repaired for this exact source. Both real test modules now select ATLAS_BROWSER_EXECUTABLE first, then google-chrome, then chromium; each passes that selected variable directly to Playwright and retains chromium_sandbox=True. The runner binds runtime evidence separately to the launcher selected by each module. The original failed run at 625bfc89975efa1fd725076bc2b0961af2768279 remains documented and was not relabeled or erased.
Exact hosted run 40950918974, job 123718493908, completed browser-smoke successfully. The complete log—not only the green badge—shows in both normal and optimized modes:
tests.test_atlas_focus_geometry.AtlasListFocusGeometry.test_keyboard_focus_ring_fits_inside_the_list_scroll_boundary ... oktests.test_atlas_reader_scroll.AtlasReaderScroll.test_new_selection_after_a_long_read_keeps_its_heading_visible ... okRan 2 tests/OK
The artifact public-shop-browser-40950918974 was preserved as artifact ID 12931733641, digest sha256:2bcd2bd3ceec0641c56f52bab6159e141e825a10042d9276bd716e7f12c1df36. I did not decode or claim unobserved fields from that binary artifact.
Independent bounded verification
- Isolated exact-head controls: 17/17 normal and 17/17 optimized, zero failures/errors/skips. The six declared runner-source mutants were killed by their designated assertions in each mode.
- Source inspection confirmed strict fixed two-ID inventory; refusal of skips/failures/errors/subtest failures/expected failures/unexpected successes/missing execution; new-output exclusivity; tracked-source identity before/after; runtime launcher path/version/hash; both workflow modes; retained stdout/stderr/artifacts; and nonzero workflow result if either mode fails.
git diff --check c93cec70…c0205ca9is clean.- The focused command reports 17 unittest methods per mode; I did not treat the documentation's parenthetical “32 combined browser-contract controls” as a separate executed 32-test batch.
Current limits
This is a scoped repair PASS, not PR integration qualification. At final readback, architecture / research-architecture, formal / formal-evidence, public-shop, one public-intake, and browser-smoke were successful; navigation remained in progress, landing-checks queued, and a second public-intake queued. The PR remained open/blocked. Missing or later checks are not rebound or inferred green.
No source/ref/PR-body/workflow/protection/Ready/check-dispatch/merge/integration/canonical-root/status/proof/science/CBZ-0/scheduler mutation was performed. ONLY-OWN follow-up review role released.
…observed UI evidence
|
FINITE PERSONAL CURRENT-COMPOSITION ENGINEERING REVIEW PICKUP — Actual performer: OpenAI/Codex Architecture Task Sweep Exact object, paths and finite output
Questions and required evidence
Exclusions / dependencies / stops
Authority retained: Owner mandate · Standing delegation · Workflow · Current live board #619 · Current task-local record #684. Link delivery is not acknowledgement. |
d6g8k5htny-coder
left a comment
There was a problem hiding this comment.
ATS-CBZ702-COMPOSITION-EVIDENCE-REVIEW-20261029C — PASS_SCOPED_COMPOSITION_EVIDENCE / CURRENT_HOSTED_QUALIFICATION_PENDING
Exact reviewed base/head: 1e2be44c69e08e47472041ac329e1b27f7dec793 / 1390112fe58093d906c878043d6b985067610002. Current GitHub synthetic merge: 56191560f8618e325497732ceeb54ce01ed1d124, tree 6786ec405cef50e0fef73c7fcd9ca3945c555230, parents exactly base/head above. The head itself is a two-parent composition of previously reviewed repaired source c0205ca9e8683b2847008f712582e183f723e186 and this Main base. Nonauthor/nonintegrator same-family engineering/evidence review; independence credit 0; scientific effect NONE.
Scoped composition/evidence result
No blocking composition or retained-evidence defect was found in this finite scope.
- Relative to current Main, the head changes exactly the 24 PR paths: the original four executable/workflow paths, the bounded documentation path, and the 19-file retained-evidence prefix. No inherited Main path is changed outside that list.
- The four executable/workflow blobs are byte-identical to repaired head
c0205ca9:.github/workflows/public-shop.ymle75bc07d…,tests/test_atlas_browser_regressions_runner.py7b974214…,tests/test_atlas_focus_geometry.pyf9eeaaa0…, andtools/run_atlas_browser_regressions.py3bf3a61a…. This review therefore does not relabel earlier checks; it establishes exact source preservation on the new object. - Fresh isolated exact-head controls passed 17/17 normal and 17/17 optimized. The declared six runner-source mutants were assertion-killed per mode; zero errors/skips occurred.
- All 18 archived text members match the committed manifest's size and SHA-256 exactly, and the prefix contains exactly those members plus the manifest. Corrected normal/optimized result records each preserve two named successes, zero failures/errors/skips, and tested source
8bd88e5952e4a67ff394e1cab3c272e7ace2f961; original records preserve the one-error failure. The manifest retains the artifact digests independently observed in the prior exact-source review: corrected2bcd2bd3…, originale6ac7e32…. - Documentation explicitly binds the green browser observation to historical corrected source/tested merge and says the later composition still needs its own hosted execution. It does not promote mathematical, kernel, scientific-register, full-regression, or CBZ-0 status.
- No unresolved review thread exists; the PR remained open, Ready, mergeable, and at the unchanged reviewed base/head during the terminal recheck.
Currentness boundary
Current-head public-shop run 40959777574, including browser-smoke job 123749203084, remains queued. The current navigation, public-shop, formal-evidence, and architecture jobs are likewise queued. Consequently this verdict is not a current hosted-browser PASS, tested-merge CI completion, readiness/independent-gate receipt, branch-protection satisfaction, integration/merge authorization, or scientific acceptance. The historical green job remains historical only.
Root B-003 retains canonical/integration custody. Author/publisher/integrator roles remain separate. This finite review pickup is released after its terminal mirror; any head/base move or current hosted result is a new object/evidence event.
|
COMPLETION / ONLY-OWN RELEASE — Exact-head COMMENT review: #702 (review) Terminal verdict: PASS_SCOPED_COMPOSITION_EVIDENCE / CURRENT_HOSTED_QUALIFICATION_PENDING at unchanged base/head Verified: four executable/workflow blobs preserved exactly from repaired head Current-head public-shop/browser-smoke and the four other observed jobs remain queued. No old check is rebound, and no hosted-CI, readiness, protection, merge/integration, canonical, proof, science, full-regression or CBZ-0 conclusion is claimed. Released only this invocation's personal review role. Root B-003 and author/publisher/integrator custody remain unchanged. Owner mandate · Standing delegation. Link delivery is not acknowledgement. |
|
FINITE PERSONAL CURRENT-COMPOSITION ENGINEERING/EVIDENCE REVIEW PICKUP — Actual performer: OpenAI/Codex Architecture Task Sweep Repository and exact object
Exact finite affected-path set Expected output Exclusions and custody Dependencies, checks and stop conditions Owner mandate · Standing delegation · Workflow. Link delivery is not reader acknowledgement. |
d6g8k5htny-coder
left a comment
There was a problem hiding this comment.
ATS-CBZ702-COMPOSITION-HARNESS-REVIEW-20261029D — PASS_SCOPED_ENGINEERING_REPAIR / CURRENT_BROWSER_EVIDENCE_PASS / FULL_CHECKSET_NONTERMINAL
Exact reviewed base/head: 1546e8bd6626b76c070d634fd9e4d18727239985 / 55b8758aec1b0dae63c89c3ef6a536c53721dfa9. Current GitHub synthetic merge: 6d42fc363069695ba97ec29238e2d0ad466402fd, tree b82b434d28a514ceeaac5545081cc6fc05c7890d, parents exactly base/head above. Nonauthor/nonintegrator same-family engineering/evidence review; organizational/scientific-independence credit 0; scientific effect NONE.
Scoped repair/composition result
No blocking defect was found in the assertion-origin/code-object-identity repair or its composition with merged PR #708.
- The repaired oracle admits only one exact synchronous selected body execution and exact code-object identities for the escaping unittest assertion-helper frames. Fixture, cleanup, helper, callback, forged-filename, repeated/missing execution, deferred return, terminal-accounting, error/skip/expected-failure and structurally-equal cloned-code cases are refused.
- Fresh isolated review on CPython 3.12.14 passed 31/31 focused controls and 46/46 combined browser-contract controls in each of normal and optimized modes. The pristine classifier killed all seven grader safeguard mutants per mode; every reported mutant had the intended assertion failure(s) and zero errors. The deliberate deferred-body mutant emitted its documented unawaited-coroutine warning; that warning was not treated as the kill.
- The repaired test source SHA-256 is
08ca3dbdaf40e7a48caf658c431d4242c3ec95cf0657bf318357ea17fbf2ec30. Every one of the nine files enumerated bybrowser-code-identity-20261029/SHA256SUMS.jsonmatches its committed digest. The retained RED log demonstrates the parent equality-based false acceptance. Protected production/workflow blobs are unchanged across both repair commits: runner3bf3a61a…, focus testf9eeaaa0…, reader test82cc2a7f…, public-shop workflowe75bc07d…. - All five intake paths introduced by merged PR #708 are blob-identical between this head and current Main (workflow
1fd332ca…, test82737d5d…, documentation7b91500d…, checkerc5785c3b…, registry119fcac4…). This review does not re-review or take custody of PR #708.
Current native hosted evidence
Exact current public-shop run 40967962663, browser-smoke job 123777172200, completed successfully on synthetic merge 6d42fc36… / tree b82b434d… using CPython 3.12.15. The complete log, not merely the green badge, shows 31/31 focused controls in each mode and both required real browser methods passing in each mode.
Artifact public-shop-browser-40967962663, ID 12938804062, independently downloaded at SHA-256 320af61a6c2abd0c1283543718e50a041221f4a32092f5e8e2038083ccc03997. Its normal and optimized result.json each contain exactly the two declared IDs, tests_run=2, zero failures/errors/skips/expected failures/unexpected successes, optimize flags 0/1, and unchanged source before/after at synthetic merge 6d42fc36… / tree b82b434d…. Both selected Chrome 156.0.8078.47 at SHA-256 aea09d69….
Current Research navigation run 40967962649 also completed successfully. Workspace-landing run 40967962703 remains in progress, so this is not a full current-checkset PASS or integration qualification. The local environment did not contain the documented pinned CPython 3.11.16; I verified the committed pinned-run receipts by digest and replayed on 3.12.14, while the current hosted run used 3.12.15. No 3.11 result is invented or rebound.
Limits
This COMMENT review establishes only the scoped engineering repair and exact current browser evidence. It is not an APPROVE, readiness/protection/ruleset satisfaction, full regression, merge/integration authorization, canonical publication, hostile-process attestation, universal unittest soundness, mathematical/proof acceptance, scientific-register change or CBZ-0 conclusion. Root B-003 and author/publisher/integrator custody remain unchanged. Any later head/base move is a new object. ONLY-OWN review role released after terminal readback.
|
COMPLETION / ONLY-OWN RELEASE — Exact-head COMMENT review: #702 (review) Terminal verdict: PASS_SCOPED_ENGINEERING_REPAIR / CURRENT_BROWSER_EVIDENCE_PASS / FULL_CHECKSET_NONTERMINAL at unchanged base/head Verified: exact-identity assertion-origin repair; 31/31 focused and 46/46 combined controls in both modes locally; seven/7 safeguard mutants per mode killed with zero errors; all nine committed code-identity receipt hashes match; five inherited intake paths are byte-identical to merged #708/Main. Exact current public-shop/browser-smoke run Research navigation is successful; workspace-landing Released only this invocation's personal review role. Root B-003 and author/publisher/integrator custody remain unchanged. Owner mandate · Standing delegation. Link delivery is not acknowledgement. |
|
FINITE PICKUP — landed-evidence readback for merged PR #702 Performer: OpenAI/Codex Repository/object boundary:
Expected output: verify the merge identity/tree and read complete current landed job evidence, rerun the exact bounded runner/control suites on the landed bytes in normal and optimized modes, and publish one terminal PASS/HOLD record that names exact selection/count/skip/error results and any remaining admission limits. Exclusions: no source edit, branch, merge, deployment, workflow rerun, canonical registry/status change, theorem/proof acceptance, scientific inference, or takeover of #712 / another worker's custody. Another lane performed the merge; this pickup claims only this fresh readback. Dependencies/protections: owner mandate, standing delegation, current |
|
TERMINAL — PASS_SCOPED_LANDED_ENGINEERING / EFFECTIVE_ON_MAIN_ENGINEERING_BASELINE_ONLY Performer: OpenAI/Codex Identity/currentness:
Complete native push-run readback on that merge:
Executed evidence:
Negative-attempt preservation: the first local attempt ran from the wrong directory (import/file-not-found; nonqualifying), and an archive-only retry was refused because it lacked Git identity. The detached-worktree control result above is the valid local result. No negative outcome is rebound as PASS. Disposition: the earlier #702 full-checkset hold is closed for this landed engineering scope by exact-merge push evidence. The change is merged and presently effective as Main engineering behavior. This grants no theorem/proof acceptance, scientific or canonical-register status, cross-repository admission, #709 qualification, #712 qualification, or future-head validity. No source, branch, workflow rerun, deployment, merge, registry/status, or scheduler mutation was made in this pickup. Finite custody released. Governing onboarding references retained: owner mandate and standing delegation; link delivery is not acknowledgement. |
Current composition with merged intake repair
Head
55b8758aec1b0dae63c89c3ef6a536c53721dfa9, treeb82b434d28a514ceeaac5545081cc6fc05c7890d, parents3154eaf54922ee512b584038d7dad28ebd843bd8and Main1546e8bd6626b76c070d634fd9e4d18727239985. All45 browser paths and five intake paths are byte-identical to their qualified parents. Pinned Python3.11.16 passed154 combined controls per mode and seven browser grader-source mutants per mode. Current composed-head hosted qualification is pending; all following source observations remain historical.Parent3154's actual hosted browser artifact12939420228 was downloaded and SHA256875ed91842923af50e48bd9d94984a8676559d9786f0b206aecd25cfceb8d719 verified. Both modes have exactly two successful required IDs and zero errors/failures/skips;31 runner controls per mode also passed. That run tested the prior1e2 composition and is not new-head green.
Current exact-identity correction
Head
3154eaf54922ee512b584038d7dad28ebd843bd8, tree82489ff334d784e2874565b84ed2daf049a0b2ae, parent8d24c37ff52c012294d71bbcbf08bd24fae4e22a.A new regression demonstrated that Python code objects can compare structurally equal without being the same object. The earlier set-membership implementation therefore did not enforce its stated exact-identity contract. Allowed objects are now retained in a tuple and checked with explicit identity comparisons. A cloned approved function with a changed filename is rejected. Original failing regressions and earlier claims are preserved with this correction; the finding does not imply changed mathematical outputs or blanket invalidity of historical finite kills.
Final pinned Python3.11.16 qualification passed 46 controls per normal/optimized mode, zero skips. Independently inspected and rerun controls include 31 affected methods and seven grader-source mutations per mode. All original test methods and production bytes remain unchanged. Exact evidence and registration updates are retained under the existing origin-repair evidence directories.
Current-head hosted/full qualification remains pending. The previous sections below are historical source-bound records, not qualification of this new head.
Current source and qualification
Head
8d24c37ff52c012294d71bbcbf08bd24fae4e22a, treeeda39a4cb82dc9e2193e5564c4c1449039ac5260, parent1390112fe58093d906c878043d6b985067610002. This strengthens only the mutation test harness and its evidence/docs. Production browser tests, fixed-ID runner and workflow remain unchanged. All 17 original named runner tests remain; 13 origin/lifecycle controls were added.A reproduced setup-only AssertionError false acceptance exposed the previous mutation meta-test. The repair captures raw traceback and actual registered method code identities, checks lifecycle/population/terminal accounting, and requires a passing baseline. Subject/helper/callback and fixture assertions cannot substitute for an intended oracle. Deferred or unsupported bodies refuse.
Fresh local qualification: 45 combined controls per mode passed, including all six original production mutants. Six additional grader-source mutants per mode were rejected using the pristine structured oracle, not formatted error text. Root independently inspected the source, ran 30 affected methods per mode and reran the six grader-source cases under normal/optimized Python. The deliberate disabled-lazy-guard mutation emits a RuntimeWarning; it is preserved and not counted as the kill. Portable replay and original false-acceptance evidence are in docs/cbz/evidence/browser-origin-20261029/.
Current-head hosted qualification is pending. This is bounded engineering evidence, not arbitrary hostile-process attestation or mathematical certification.
Actual historical browser evidence
On parent1390112, the current sandboxed browser regressions actually passed: artifact12937636079 SHA2563271bcae587a3c5aacec1256efb8a069e6671e919ae254d1b22316709b42db9f was independently downloaded and verified. Both mode records show exactly two required successful IDs and zero errors, failures or skips; tested merge56191560f8618e325497732ceeb54ce01ed1d124 has tree6786ec405cef50e0fef73c7fcd9ca3945c555230. This is parent-source evidence, not new-head green.
Earlier corrected c0205ca execution and the original Chromium sandbox failure remain preserved in docs/cbz/evidence/atlas-browser-20261029/. Chrome-first selection keeps sandbox enabled and all original UI assertions intact. No security setting, dependency pin or workflow permission was weakened; no denied local launch was retried.
Integration limits
The composition preserves Main1e2be44 and its377 added paths. Canonical registration addenda are in docs/cbz/ATLAS_BROWSER_REGRESSIONS.md. Five real-record gaps, whole-federation regression, final kernel/auditor qualification and CBZ-0 remain separate and unestablished. No PR approval is submitted by this lane.