Repository navigation
Make verdict-review-gate apply the landscape validator's withheld-key policy - #144
Merged
Merged
Conversation
…tead of a drifted copy The gate's own withheld_packet_keys predated #124 and lacked the top-level, copy-disposition and withheld[] label checks; importing scripts/landscape.withheld_packet_keys from the base makes a key the tooling owner adds (gap_receipts, blind-lanes round 2) apply here with no gate change. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
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. |
…ct the record's wording Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
seathatflowsinourveins
enabled auto-merge (squash)
September 23, 2026 19:47
seathatflowsinourveins
deleted the
claude/gate-shared-withheld-policy
branch
September 23, 2026 19:52
This was referenced Sep 23, 2026
seathatflowsinourveins
added a commit
that referenced
this pull request
Oct 11, 2026
…s host (correction #144) On the current host (ai-memory 2.6.0) the capture-mode file the installer writes under --apply was absent, so lifecycle capture ran in the default denylist mode and captured every directory, including scratch and test worktrees and GPT read snapshots. The maintenance doc said the allowlist installers had been applied. It now records the measurement, the vendor source lines, and the repair under this stack's untracked per-checkout marker convention (.gitignore:73): the main checkout carries the marker, the trading repository carries none, read snapshots strip copied markers, and install-hooks re-applies allowlist mode; plus how to verify it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
seathatflowsinourveins
added a commit
that referenced
this pull request
Oct 11, 2026
…ection #144) (#988) * docs: record that ai-memory allowlist capture was not in force on this host (correction #144) On the current host (ai-memory 2.6.0) the capture-mode file the installer writes under --apply was absent, so lifecycle capture ran in the default denylist mode and captured every directory, including scratch and test worktrees and GPT read snapshots. The maintenance doc said the allowlist installers had been applied. It now records the measurement, the vendor source lines, and the repair under this stack's untracked per-checkout marker convention (.gitignore:73): the main checkout carries the marker, the trading repository carries none, read snapshots strip copied markers, and install-hooks re-applies allowlist mode; plus how to verify it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * docs: state ai-memory enrollment by the documented design and give the working check The GPT read of #988 found two P2s. The prescribed `ai-memory hook --check-capture` cannot run without --event, --agent and --server-url; the doc now gives the documented form (recipes/README.md:418) and what to expect for enrolled and excluded directories, plus the inverse. And "lane worktrees carry none and are not captured" was not exact: a worktree below the enrolled checkout, or one that receives the marker through .worktreeinclude, is enrolled by design (recipes/README.md, "Worktrees and the capture marker"). The paragraph now states that design, that read snapshots and blind lanes must never carry or inherit the marker (the read kit aborts on any ancestor marker), and the measured state: 2 of 404 worktrees carry a marker, none nested under the main checkout. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * test: pin the ai-memory capture correction's two statements; fix the read's P3 citations The GPT micro of 77e95b7 (P2) found no regression for the two corrected statements, so the correction rule (AGENTS.md, Corrections) was not met. Two doc-contract tests fail on the first version's text (8690d10: 2 failures) and pass at this head: - the documented check carries hook --event --agent --server-url --check-capture --data-dir (HookArgs, ai-memory v2.6.0 cli.rs:2957-2966; the bare form exits 2); - enrollment follows recipes/README.md "Worktrees and the capture marker" (marker walk to $HOME, marker.rs:938-1025; .worktreeinclude), without the false "lane worktrees carry none". The independent Claude read's P3s: the per-event read is hook.rs:720-726 and the v2.6.3 reader sits unchanged at :1383-1399; the read kit strips markers, then aborts on a remaining or ancestor marker; the trading-research bank is cited (foundation-finalization :70,157) instead of the unsourced "exposure tags"; the check lines are recipes/README.md:419-421. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
seathatflowsinourveins
added a commit
that referenced
this pull request
Oct 12, 2026
… to the native profile Fixes the two upheld Codex threads on #997 at 03fc660 (4245318057, 4245318059). The third, 4245318064 (P1, layer-resolution record), is not accepted: the layer-resolution protocol (#1008) governs layer adoption and verdict decisions, and this change adds no layer verdict. It corrects a host-state claim under AGENTS.md "Corrections", which the pinned sources already cited (ai-memory v2.6.3 = 385b81ea, Codex rust-v0.162.1 = 092d3acd) and the fail-before tests below satisfy. 1. The installer lead (4245318057). The retained 2026-09-21 record, evidence/artifacts/memory-landscape-20260921/native-commands.json:295-393 at 03fc660, shows the three installers run with --capture-assistant, exit 0 each, so the lead was true but undated; what it contradicted was correction #149's undated "The earlier hook applications ran without the flag". The lead now opens "At the 2026-09-21 qualification", cites that record and says it is not the hook state now. Correction #149 says the qualification's installers carried the flag, so the flag was lost before the 2026-10-11 measurement (the loss path is host history not reconstructed here), and names the two later flagless applications, the #144 repair and the 2.6.3 re-stage of 2026-10-10. "Not in force" is dated 2026-10-11. 2. The second Codex profile (4245318059). hook-trust.json at 03fc660 lists seven hook commands for the native profile (:35-77) and seven for the Desktop profile (:279-321), 14 in all, so the CC's 0 of 7 is one profile's. Correction #149 and the qualification prose after it now bound the count and the Codex conclusions (trust pending, capture off) to the native profile and say the Desktop profile was not measured. tests/test_memory_landscape_capture_doc.py has 16 tests (10 kept, 6 new). On the 03fc660 doc text the module gives 11 OK and 5 failures, 0 errors; on this text 16 OK. The new tests read the artifacts: the line ranges the doc cites, the flag and exit code of the three installer records, and seven commands per profile. manifests/evidence.json: hashes of the doc and the test re-registered only. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
4 tasks done
seathatflowsinourveins
added a commit
that referenced
this pull request
Oct 12, 2026
…rrection #149) (#997) docs: ai-memory --capture-assistant was not in force on this host (correction #149) * docs: ai-memory --capture-assistant was not in force on this host (correction #149) The maintenance doc said the allowlist installers carried --capture-assistant. 0 of 9 Claude and 0 of 7 Codex hook commands did: ai-memory bakes the flag only into the Stop command and only preserves an opt-in already present (install_hooks.rs:689-701, 1104-1109 @ 89bd8ded; render_shared.rs:1425-1427), so applications without the flag never added it. Claude was re-staged with it on 2026-10-11; Codex waits for a trust refresh (trusted_hash changes). Also: the HookArgs v2.6.3 range in the doc-contract test becomes :2961-2970, test 1 checks every --check-capture block, a new test fails on the previous text, and the read kit's "repositories commit the marker" comment is listed as a host note against .gitignore:73. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> * test: name the missing correction heading instead of raising ValueError section_from() used str.index, so a document lacking the #149 heading (the 16f0a1f text) gave 3 test errors instead of contract failures. It now raises AssertionError naming the heading, and the first #149 test no longer asserts on the whole document before extracting the section: the module gives 6 OK on the head text and 4 assertion failures, 0 errors, on the 16f0a1f text. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> * docs: reconcile correction #149 with the dated qualification prose and pin it to the installed releases Fixes the four Codex review threads on #997 (each verified valid at 805aca0): 1. Contradiction. The two paragraphs after the correction said, in the present tense, that both clients' assistant Stop excerpts are opted in and that Codex trust ended in a captured outcome. They now open "At the 2026-09-21 qualification" (git blame 54896e7 and every artifact under evidence/artifacts/memory-landscape-20260921 carry that date) and point at correction #149. A new test reads that prose, which correction_149_section stops before. 2. Unverified host state. The 0 of 9 and 0 of 7 counts, the earlier hook applications, the Claude re-stage and the pending Codex trust are labelled CC host measurements, cited to correction #149 in command-center/lessons-20261005/cc-corrections-20261005.md and "not re-read by this change". 3. Pin. The host runs ai-memory 2.6.3 (385b81ea). install_hooks.rs:689-701, :1104-1109 and :1142 and render_shared.rs:1425-1427 are byte-identical at the same lines in v2.6.3 and v2.6.0 (89bd8ded), so both pins are cited and the test requires the v2.6.3 pin beside each cite. 4. Trust hash. Re-rendering changes the computed current hash; the stored trusted_hash is unchanged, the mismatch marks the hook modified, and the /hooks review writes the reviewed hash into trusted_hash (docs/decisions/2026-10-04-codex-rtk-hook-qualified.md Decision 1 at rust-v0.159.3; re-read at rust-v0.162.1 = 092d3acd, the installed codex-cli: discovery.rs:796-813 and :715-721, hooks_rpc.rs:58-92). The test rejects "re-rendering changes ... trusted_hash". tests/test_memory_landscape_capture_doc.py now has 10 tests: on the 805aca0 doc text the module gives 6 OK and 4 assertion failures, 0 errors (one per thread); on this text 10 OK. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> * docs: cite correction #149's ledger by its true locator, coordination/lessons-20261005/... The #149 correction named the CC's ledger as command-center/lessons-20261005/cc-corrections-20261005.md, the form the brief prescribed. The CC recorded that as its correction #180: the ledger sits at coordination/lessons-20261005/cc-corrections-20261005.md (state-root relative), a sibling of command-center/, not under it. The doc keeps "(host state, not in this repository)"; the test now pins the true locator and rejects the old form. Found by looking up the file name under the host state root; the CC confirmed it in #180. Fail-before: with the new pin the module gives 9 OK and 1 assertion failure, 0 errors, on the 8b8c8f5 doc text; 10 OK after. validate.py and evidence_manifest.py --check pass. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com> * docs: date the installer lead and bound correction #149's Codex count to the native profile Fixes the two upheld Codex threads on #997 at 03fc660 (4245318057, 4245318059). The third, 4245318064 (P1, layer-resolution record), is not accepted: the layer-resolution protocol (#1008) governs layer adoption and verdict decisions, and this change adds no layer verdict. It corrects a host-state claim under AGENTS.md "Corrections", which the pinned sources already cited (ai-memory v2.6.3 = 385b81ea, Codex rust-v0.162.1 = 092d3acd) and the fail-before tests below satisfy. 1. The installer lead (4245318057). The retained 2026-09-21 record, evidence/artifacts/memory-landscape-20260921/native-commands.json:295-393 at 03fc660, shows the three installers run with --capture-assistant, exit 0 each, so the lead was true but undated; what it contradicted was correction #149's undated "The earlier hook applications ran without the flag". The lead now opens "At the 2026-09-21 qualification", cites that record and says it is not the hook state now. Correction #149 says the qualification's installers carried the flag, so the flag was lost before the 2026-10-11 measurement (the loss path is host history not reconstructed here), and names the two later flagless applications, the #144 repair and the 2.6.3 re-stage of 2026-10-10. "Not in force" is dated 2026-10-11. 2. The second Codex profile (4245318059). hook-trust.json at 03fc660 lists seven hook commands for the native profile (:35-77) and seven for the Desktop profile (:279-321), 14 in all, so the CC's 0 of 7 is one profile's. Correction #149 and the qualification prose after it now bound the count and the Codex conclusions (trust pending, capture off) to the native profile and say the Desktop profile was not measured. tests/test_memory_landscape_capture_doc.py has 16 tests (10 kept, 6 new). On the 03fc660 doc text the module gives 11 OK and 5 failures, 0 errors; on this text 16 OK. The new tests read the artifacts: the line ranges the doc cites, the flag and exit code of the three installer records, and seven commands per profile. manifests/evidence.json: hashes of the doc and the test re-registered only. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The gate kept its own copy of
withheld_packet_keys, written before #124 merged, and that copy had drifted fromscripts/landscape.py. It lacked:current_choice,decisionandrationalekeys;review_status;archived/licensekeys;withheld[]policy-label check.This was not a bypass in CI, because every verdict-changing PR also runs
landscape.pyas a validator at the base's copy. But the gate's own message under-reported. It also would not have picked up blind-lanes round 2's newgap_receipts/gap_receipts_notekeys without a second edit. That gap was raised by agent-lab-17.Change
scripts/landscape.withheld_packet_keys, so there is one policy.--withhold-labelspacket does.Acceptance (local, raw)
validate.py,evidence_manifest --check,landscape.pyandbuild_verdicts --check.scripts/verdict_review_gate.pyonly and no verdict data.An independent review is running; auto-merge will be enabled after it.
🤖 Generated with Claude Code