docs: ai-memory --capture-assistant was not in force on this host (correction #149) - #997

Scope

• Correction Record the regular-session no-signal run and the first extended-hours paper run #149 in docs/memory-landscape-maintenance.md: the documented --capture-assistant opt-in was not in force on this host (0 of 9 Claude and 0 of 7 Codex hook commands carried it on 2026-10-11; the 0 of 7 is the native Codex profile's and the Desktop profile's seven were not measured; the counts, the Claude re-stage and the pending Codex trust are CC host measurements, recorded as correction Record the regular-session no-signal run and the first extended-hours paper run #149 in coordination/lessons-20261005/cc-corrections-20261005.md, not re-read by this lane). Claude was re-staged with it on 2026-10-11 and only its Stop hook carries the flag; native Codex is pending a trust refresh. Follow-ups from the docs: ai-memory allowlist capture was not in force on this host (correction #144) #988 read: the HookArgs v2.6.3 range in the doc-contract test is :2961-2970, test 1 checks every --check-capture sh block, and the read kit's "repositories commit the marker" comment is recorded as a host note against .gitignore:73 (the kit lives in the co-op's host state directory, not this repository).
• Review-thread fixes (one commit, 8b8c8f54): the 2026-09-21 qualification prose after the correction is dated and points at correction Record the regular-session no-signal run and the first extended-hours paper run #149; the host state is labelled as CC measurements; the cites are pinned to ai-memory v2.6.3 = 385b81ea; the Codex trust-hash mechanism is corrected to the computed current hash versus the stored trusted_hash.
• Locator correction (one commit, 03fc660b, CC correction Document interim foundation/trading lane ownership (docs/lanes.md) #180): the ledger sits at coordination/lessons-20261005/cc-corrections-20261005.md (state-root relative, host state), not under command-center/. The doc text and the test pin take the true form, and the test rejects the old one.
• Terminal fix (one commit, 07f373ec; the CC's adjudication of the three Codex threads on 03fc660b): the installer lead at docs/memory-landscape-maintenance.md:111-118 is dated and cites the retained 2026-09-21 record (native-commands.json:295-393: the three installers ran with --capture-assistant, exit 0 each); correction Record the regular-session no-signal run and the first extended-hours paper run #149 no longer says "the earlier hook applications ran without the flag" (the same record contradicts it), names the two later flagless applications and dates "not in force" to 2026-10-11; its Codex count and Codex conclusions are bounded to the native profile (hook-trust.json: seven hook commands in each of two profiles, native :35-77, Desktop :279-321), and the Desktop profile is stated as not measured.
• Base commit: 3a8a8461ee60bd499efcb78156ce4c3c052430c5
• Lane: lane:foundation
• Owned paths touched: docs/memory-landscape-maintenance.md, tests/test_memory_landscape_capture_doc.py, manifests/evidence.json (re-registered hashes only, in each of the five commits)

Codex threads resolved

Each thread has a reply with the fix line and the commit. The first four are resolved at 8b8c8f54; the three Codex raised on 03fc660b are resolved at 07f373ec (two fixed, one not accepted).

• 4240862885, contradiction: the prose after the correction (docs/memory-landscape-maintenance.md:136-145 and :147-155) opens "At the 2026-09-21 qualification" and points at correction Record the regular-session no-signal run and the first extended-hours paper run #149; OlderQualificationStatementTests reads it (correction_149_section stopped before it).
• 4240862887, unverified host state: the counts, the later hook applications, the Claude re-stage and the pending Codex trust are labelled CC host measurements with a repo-safe locator; the hook files were not re-read, the ledger entry was (it matches). The reply quotes the locator as first written; 03fc660b corrects it.
• 4240862892, pin: the cites are at ai-memory v2.6.3 = 385b81ea and byte-identical at v2.6.0 = 89bd8ded; the test requires both pins beside each cite.
• 4240862894, trust hash: re-rendering changes the computed current hash, the stored trusted_hash is unchanged, and the /hooks review writes the reviewed hash; cited at Codex rust-v0.162.1 = 092d3acd and in the 2026-10-04 decision record.
• 4245318057 (P2), the lead still said the installers were run with the flag: reply 4245477987. The retained 2026-09-21 record (native-commands.json:295-393) shows they were, exit 0 each, so the lead was true but undated; the sentence it contradicted was correction Record the regular-session no-signal run and the first extended-hours paper run #149's undated "The earlier hook applications ran without the flag". The lead opens "At the 2026-09-21 qualification", cites the record and says it is not the hook state now (:111-118); Record the regular-session no-signal run and the first extended-hours paper run #149 (:134) says the qualification's installers carried the flag, the flag was lost before the 2026-10-11 measurement, the loss path is host history not reconstructed here, and two later applications (the Make verdict-review-gate apply the landscape validator's withheld-key policy #144 repair, the 2.6.3 re-stage of 2026-10-10) ran without it. The reading of the lead as a false claim is corrected by this artifact; the lead was not rewritten to say flagless. Two InstallerLeadTests tests fail on 03fc660b (the third guards the artifact and passes either way).
• 4245318059 (P2), the second Codex profile: reply 4245478146. hook-trust.json lists seven hook commands per profile, native :35-77 and Desktop :279-321 (14 in all), so the CC's 0 of 7 is one profile's. Record the regular-session no-signal run and the first extended-hours paper run #149 states the 0 of 7 as the native profile's, says the Desktop profile's seven were not measured (no claim about their flag or trust state) and bounds its Codex conclusions to the native profile; the qualification prose after the corrections (:153-155) does the same. CodexProfileScopeTests (2 tests) and one OlderQualificationStatementTests test fail on 03fc660b.
• 4245318064 (P1), record layer resolution before upstream research: not accepted, reply 4245478240, resolved as not applicable. The layer-resolution protocol (W6 governance: preserve survivors and resolve layers before research #1008) governs layer adoption and verdict decisions; this PR changes no layer verdict and is a correction of a host-state claim under AGENTS.md "Corrections", satisfied by the pinned upstream sources (ai-memory v2.6.3 = 385b81ea, Codex rust-v0.162.1 = 092d3acd) and the fail-before tests.

SOTA sources

• ai-memory v2.6.3 = 385b81ea95f7b73ede7146d0ae606c4caa7b43c1 (akitaonrails/ai-memory; the tag peels to this commit on the upstream remote and ai-memory --version on the host prints 2.6.3), read from the pinned local upstream mirror: crates/ai-memory-cli/src/commands/install_hooks.rs:689-701 (an existing opt-in is preserved on a bare re-apply, never added), :1104-1109 (existing_capture_assistant_opt_in) and :1142 (baked_capture_assistant scans the installed ai-memory commands for the flag); crates/ai-memory-cli/src/commands/render_shared.rs:1425-1427 (flag baked only when event == "stop"); crates/ai-memory-cli/src/cli.rs:2961-2970 (HookArgs, required --event, --agent, --server-url).
• ai-memory v2.6.0 = 89bd8ded3c1ab8b769cf99417d038ed0364403c8: the same install_hooks.rs and render_shared.rs ranges are byte-identical at the same lines (install_hooks.rs 660-705 and 1100-1150, render_shared.rs 1385-1432, diffed between the mirrors; the first differing hunks are install_hooks.rs:4316 and render_shared.rs:2179, plus same-count replacements at render_shared.rs:336 and :341); crates/ai-memory-cli/src/cli.rs:2957-2966 (HookArgs).
• openai/codex rust-v0.162.1 = 092d3acd6bec3e3a14bdc7e7a2810ab628ab759d (the tag peels to this commit on the upstream remote; the installed codex-cli is 0.162.1), read from the pinned local upstream mirror: codex-rs/hooks/src/engine/discovery.rs:796-813 (hook_trust_status: a stored hash that differs from the current hash is Modified), :777-794 (hook_hash) and :715-721 (a hook runs only when Managed or Trusted unless bypass_hook_trust); codex-rs/tui/src/hooks_rpc.rs:58-92 (write_hook_trusts writes trusted_hash = the reviewed current hash). First-party: docs/decisions/2026-10-04-codex-rtk-hook-qualified.md Decision 1 states the same mechanism at rust-v0.159.3.
• First-party artifacts at 03fc660b (unchanged by this PR; the new tests re-derive every cited line range from the files): evidence/artifacts/memory-landscape-20260921/native-commands.json:295-393 (hook-install-claude :295, hook-install-codex :328, hook-install-desktop :361, each with --capture-assistant in argv and exit_code 0) and evidence/artifacts/memory-landscape-20260921/hook-trust.json:35-77 and :279-321 (seven hook commands for each of the native and Desktop profiles, the Stop command modified before the review and trusted after, in both).
• Not authenticated by this lane: the host hook counts and the Claude re-stage (hook settings files are out of bounds). Both are the CC's measurements, recorded as correction Record the regular-session no-signal run and the first extended-hours paper run #149 in coordination/lessons-20261005/cc-corrections-20261005.md; that entry was read and matches the doc. The entry gives "0 of 7 Codex" without naming a hooks.json; the native scope is the CC's option in its adjudication, and the Desktop profile's flag and trust state stay unclaimed.

Evidence-class table

Local commands run

$ python3 -E -s -B tests/test_memory_landscape_capture_doc.py -v   (final tests, doc text of 805aca01)
Ran 10 tests ... FAILED (failures=4)   # one per thread, 0 errors
$ python3 -E -s -B tests/test_memory_landscape_capture_doc.py -v   (locator pin, doc text of 8b8c8f54)
Ran 10 tests ... FAILED (failures=1)   # the true-locator pin, 0 errors
$ python3 -E -s -B tests/test_memory_landscape_capture_doc.py -v   (at 03fc660b)
Ran 10 tests ... OK
$ python3 -E -s -B scripts/validate.py
exit 0 (status passed, 70 components, 239 receipts, 11703 hashed files)
$ python3 -E -s -B scripts/evidence_manifest.py --check
exit 0 ({"files": 11703, "status": "passed"})
$ gitleaks git --log-opts="3a8a8461..03fc660b"
4 commits scanned, no leaks found
$ python3 -E -s -B tests/test_memory_landscape_capture_doc.py -v   (committed test file, doc text of 03fc660b)
Ran 16 tests ... FAILED (failures=5)   # 11 OK, 0 errors, rc 1; the five are the new regressions
$ python3 -E -s -B tests/test_memory_landscape_capture_doc.py -v   (at 07f373ec)
Ran 16 tests ... OK   # rc 0
$ python3 scripts/validate.py   (without -I)
exit 0 ({"components": 70, "hashed_files": 11703, "profiles": 4, "receipts": 239, "status": "passed"})
$ python3 scripts/evidence_manifest.py --check   (without -I)
exit 0 ({"files": 11703, "status": "passed"})
$ gitleaks git --log-opts="3a8a8461..07f373ec" --config .gitleaks.toml
5 commits scanned, no leaks found (rc 0)
$ git merge-tree --write-tree origin/main 07f373ec   (origin/main 2c4fdf86, 7 commits ahead of the base)
rc 0, clean (tree feecf72b); no rebase, so the push was the fast-forward 03fc660b..07f373ec
pre-push gate (host-name scan, three registry tests, published-digest guard): passed
privacy scan of the diff, commit message, thread replies and this body (names_checked=3): 0 hits, positive control ok

Decision record

None; a documentation correction with a regression test.

Host evidence

Not applicable (no evidence/hosts/ change).

Follow-ups

• Open, host side and outside this PR: the Desktop profile's seven Codex hook commands (hook-trust.json:279-321) were not measured on 2026-10-11, so their flag and trust state are unclaimed; measuring them, and re-staging with the trust refresh if they lack the flag, is a CC host action. Three notes are recorded here so they are not lost:
  • The lead and the retained record: the installers did run with --capture-assistant on 2026-09-21 (native-commands.json:295-393). The doc now says that, dates it, and no longer says the earlier hook applications ran without the flag; how the flag was lost by 2026-10-11 is host history not reconstructed here.
  • The message of 8b8c8f54 says every artifact under evidence/artifacts/memory-landscape-20260921/ carries 2026-09-21. Precisely, four run artifacts carry 2026-09-21 timestamps (hook-trust.json, native-client-results.json, native-commands.json, state-observation.json), sources.json carries it among older upstream dates, and qualification.json carries no date. The doc claims only the date of the qualification prose.
  • The CC ledger entry for correction Record the regular-session no-signal run and the first extended-hours paper run #149 carried a loose trusted_hash sentence, and the brief gave a locator that did not resolve. The CC recorded both as its correction Document interim foundation/trading lane ownership (docs/lanes.md) #180; this PR takes the true locator in 03fc660b and states the trust mechanism as the decision record does.

Checklist

• No workflow changes.
• No secrets printed, logged or committed.
• No new paid surface.
• Peer-owned files and worktrees preserved.

🤖 Generated with Claude Code

Follow-ups after landing (CC, 2026-10-12T08:53:46Z)

The final reads' P3s, carried here because the landing did not move the head:

• GPT family (fix3 at 07f373ec): the carried unittest diagnostic citation is still missing; keep ordinary Stop rendering distinct from recovery after manual config edits; inherited hostname labels; the cited captures are absent from extra/; gaps in the merged-verification environment.
• Claude family (micro at 07f373ec): the body says the tests "re-derive every cited line range", but correction Record the regular-session no-signal run and the first extended-hours paper run #149's repeat of 295-393 is a literal; "0 of 9 / 0 of 7 hook commands" counts ai-memory commands only (the live totals are 30 and 14); mutants M14 and M15 survive and M4 and M7 are carried; the kit quote is carried.