Skip to content

ai-memory 2.4.1: real Codex capture observed after the /hooks trust (receipt addendum) - #344

Merged
seathatflowsinourveins merged 6 commits into
mainfrom
claude/ai-memory-241-codex-capture-20260926
Sep 26, 2026
Merged

seathatflowsinourveins merged 6 commits into
mainfrom
claude/ai-memory-241-codex-capture-20260926

Conversation

@seathatflowsinourveins

@seathatflowsinourveins seathatflowsinourveins commented Sep 26, 2026 •

Copy link
Copy Markdown
Collaborator

What

This closes the one pending limitation on the ai-memory 2.4.1 cutover receipt for host nativestack-5975wx-20260925: Codex capture.

The 2.4.1 cutover (#307, receipt evidence/receipts/ai-memory-241-qualification-20260925.json) rewrote the binary path in 7 Codex hook commands. Codex skips new or changed hooks until they are trusted, so limitations[0] recorded that real Codex capture through 2.4.1 had not been observed. On 2026-09-26 the user reported trusting the commands in Codex /hooks. The trust action itself is not retained; what is retained is its observable result, the hooks running.

  • Addendum on the 2.4.1 receipt (addenda[0], dated 2026-09-26). The claim, the original limitations and data.cutover.pending stay as recorded, following the rtk 0.50.0 hook recipe: exclude git show REV:path in every global-option form and git branch #318 addenda style.
  • Retained probe in evidence/artifacts/ai-memory-241-codex-capture-20260926/:
    • capture.txt records ai-memory 2.4.1, codex-cli 0.157.1, and 7/7 Codex hook commands on tools/ai-memory-2.4.1.
    • It ran codex exec -s read-only with one echo command, which exited 0 with the expected output.
    • A read-only store count then found schema V67 and 1 Codex session, with one observation each of session-start, user-prompt, pre-tool-use, post-tool-use, stop and session-end.
    • codex-stderr-hook-lines.txt holds only the version banner and the hook: lines; session id, workdir and model lines are dropped.
    • codex_capture_count.py is the query that was used.
  • Failed first attempt kept. In capture-attempt1-untrusted-dir.txt, Codex exited 1 and 0 Codex sessions wrote observations. Its working directory and stderr were not retained.
  • Separate no-git-repo run. capture-attempt1-repro.txt ran from outside any git repository and keeps the complete stderr: "Not inside a trusted directory and --skip-git-repo-check was not specified.", with 0 sessions. It establishes only its own failure, not the cause of the first attempt's.
  • Hook config per event. codex-hook-events.txt shows one 2.4.1 command for each of the seven events, PreCompact included.
  • Dated update notes in docs/decisions/2026-09-25-workstation-sota-refresh.md and docs/decisions/2026-09-25-ai-memory-2-4-0-release-review.md.
  • manifests/evidence.json registers the 10 files. The component matrix and grand list were regenerated with no change.

Not covered

  • The seventh hook, PreCompact, is unverified: no PreCompact observation or hook line was recorded.
  • The probe counted observations by kind; it did not inspect their content.
  • This is not a requalification, and the ledger winner pin does not move.

Checks

  • uv run --no-project --python 3.13 … scripts/validate.py: {"components": 69, "hashed_files": 6591, "profiles": 4, "receipts": 158, "status": "passed"}. That covers integrity and scope checks only.
  • core.hooksPath is scripts/git-hooks, and the pre-commit hooks ran.

Review

A Codex cross-family review (codex exec -s read-only, gpt-6-astra, high) of the first head returned DISAGREE with five overclaims:

  1. the attempt-1 stderr was not retained;
  2. the trust action was not in the logs;
  3. the reason given for the missing PreCompact had no support;
  4. the prompt-allowlist behaviour was asserted but not verified;
  5. the Codex version history was not in the logs.

The fixes are in bbd9978b. A Codex recheck marked items 2, 4 and 5 FIXED. For 1 and 3 it found wording residues: the first attempt's cause was still implied, and the decision note said "unexercised". 2e5c7897 fixes both exactly as the recheck asked. The recheck also computed sha256 and byte counts for all 10 files and found they match manifests/evidence.json, and it found no personal paths, session ids or credential material. Under the bounded review loop (one review plus one repair round), there is no third round.

SOTA sources

🤖 Generated with Claude Code

Scout and others added 2 commits September 26, 2026 10:48
Adds a dated addendum to the 2.4.1 qualification receipt and a retained probe
(evidence/artifacts/ai-memory-241-codex-capture-20260926/): after the user trusted
the 7 rewritten Codex hook commands, one codex exec session wrote session-start,
user-prompt, pre-tool-use, post-tool-use, stop and session-end observations
through production 2.4.1 (store V67). The failed first attempt (outside a git
repository, Codex directory-trust refusal) is kept. The two decision records
get dated update notes; original limitations stay as recorded.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
@seathatflowsinourveins seathatflowsinourveins added the lane:foundation Foundation lane: Claude/Codex setup, hosts, memory, RAG, research, workers label Sep 26, 2026
@chatgpt-codex-connector

Copy link
Copy Markdown

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

Scout and others added 4 commits September 26, 2026 10:53
Cross-family review (Codex, DISAGREE) found five overclaims:
- the directory-trust error was not retained for attempt 1;
- the trust action is not in the logs;
- the PreCompact explanation had no support;
- the prompt-allowlist behaviour was not verified;
- the Codex version history was not in the logs.

Changes:
- attribute the trust to the user's report;
- add codex-hook-events.txt (per-event command counts) and capture-attempt1-repro.txt
  (the no-git-repo condition, complete stderr);
- state PreCompact as unverified;
- drop the allowlist and version-history statements.

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
…rified (recheck residuals)

Co-Authored-By: Claude Opus 5.5 (1M context) <[email protected]>
@seathatflowsinourveins

Copy link
Copy Markdown
Collaborator Author

Held for the v2026.09.26.1 release window (session 9a). scripts/release_due.py on this branch lists evidence/receipts/ai-memory-241-qualification-20260925.json under changed (it is referenced from the new-machine docs), so this merges after WINDOW DONE and after the re-pin PR merges. I'll re-run release_due first.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

lane:foundation Foundation lane: Claude/Codex setup, hosts, memory, RAG, research, workers

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant