docs(audit): inventory the transcript capture producers (#716, slice 1) by cdeust · Pull Request #874 · cdeust/Cortex

cdeust · 2026-11-05T01:31:45Z

Summary

First slice of #716 (resumable incremental transcript capture). The issue's own first step is "audit all capture producers first; reuse any existing cursor/idempotency mechanism". This PR is that audit, with no code change: docs/audits/transcript-capture-producers-2026-11-05.md.

Findings (each cited to a line at main 2126d2a):

No producer keeps a cursor into a transcript (no offset, line count or last-seen turn). Existing progress markers: a 64 KiB head hash (backfill), transcript stem / sha256(session_id) (SessionEnd), spool and queue files, exact-content dedup in remember.
Nothing discovers Codex rollouts; a rollout is read only when SessionEnd hands its transcript_path, and only tool names and turn counts are extracted.
Codex SessionEnd intake (session_queue.py) keys its job by sha256(session_id) alone and drops a second SessionEnd of the same session (receipt exists, or os.link keeps the first queued payload): a second rollout file reaches neither the session log nor the profile — the first blocker for acceptance item 1 of feat: specify and verify resumable incremental transcript capture across Codex rollouts #716. The stem / session_id key split in session_lifecycle only matters for a direct (unqueued) call.
The resident capture worker acknowledges before storing (inferred: loss, not duplicate, on crash after ack).

Refs #716 — not Closes.

Type of change

Documentation (audit)

Test plan

python scripts/check_doc_claims.py → OK; python scripts/check_craftsmanship.py → OK.
Every cited line number re-read with grep -n/sed -n against the tree before commit.
No test added: the PR changes no behaviour. The acceptance tests of feat: specify and verify resumable incremental transcript capture across Codex rollouts #716 belong to the later slices.

Not verified here: the claims marked inferred in the audit were derived from the cited code, not executed.

Audit notes

Remaining slices of #716, after the owner's product decision (build a cursor, or document the limits and close, per the issue's Alternatives considered):

Test: one session over two rollout files reaches the log and the profile with both files' activity; the fix starts at the Codex intake key (sha256(session_id) in session_queue.py), which must also name the file or its turns, then the stem / session_id split for direct calls.
Repeat delivery + lost acknowledgement test on the capture worker path (ack-before-store).
Prefix-rewrite detection, or an explicit documented limit.
Native event receipts per advertised host, separate from synthetic hook tests.

Breaking changes

None.

Generated by Claude Code

No producer keeps a cursor into a transcript, none discovers Codex rollouts, and the only rollout reader extracts tool names and turn counts. The audit cites each session key, dedup and replay path so the owner can decide between a durable cursor and documented limits. Refs #716 Co-Authored-By: Claude Opus 5.5 <[email protected]> Claude-Session: https://claude.ai/code/session_018kyZYthhBeHYUUVXweieTP

chatgpt-codex-connector · 2026-11-05T18:47:55Z

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-11-05T18:51:28.764765Z	`0db1f65`	New commits
🔒 Security Review	✅ Completed 2026-11-05T18:52:00.889948Z	`0db1f65`	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

claude · 2026-11-05T18:48:25Z

Claude finished @cdeust's task in 2m 59s —— View job

PR #874 review (1 of 3)

Read the PR diff (238a405, docs only)
Ran five reviews: documentation accuracy, code quality/CI, security, performance, test coverage
Checked the Blocker against plugins/hypermnesia-mcp-codex/scripts/session_queue.py:95-111
Posted the review: docs(audit): inventory the transcript capture producers (#716, slice 1) #874 (comment)

Blocker (1): the audit's P1 claim at docs/audits/transcript-capture-producers-2026-11-05.md:50-53 is false for Codex. It says a Codex session continued into a second rollout file "gets two log rows but one profile update". In fact, Codex intake keys each job by sha256(session_id). A second SessionEnd with the same session_id is dropped at intake, either by the existing receipt or by the swallowed os.link error. So the second file reaches neither the log nor the profile, and the planned "stem / session_id key split" fix would not cover Codex.

Notes (3, no action before merge):

The inventory misses a producer: record_session_end.
Some cited lines are missing or off by one.
Two head/tail-read caveats are missing from P3/P4.

CI gates are unaffected and no security problems were found.

Verdict: block

chatgpt-codex-connector

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 238a40585d

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

chatgpt-codex-connector · 2026-11-05T18:49:30Z

+  (`session_lifecycle.py:359-372`). *Inferred:* a Codex session continued
+  into a second rollout file gets two log rows but one profile update, so
+  the second file's activity never reaches the profile — acceptance item 1
+  of #716 fails on the profile path.


Account for intake deduplication in the multi-rollout conclusion

For the Codex hook path described above, a second rollout with the same session_id never reaches the lifecycle handler: plugins/hypermnesia-mcp-codex/scripts/session_queue.py::enqueue keys the job solely by sha256(session_id), returns immediately when its completed receipt exists, and preserves the first queued payload when the job is still pending. Consequently, this path produces neither a second log row nor a second profile update. Please distinguish direct lifecycle calls from queued Codex delivery and identify the intake deduplication as an additional blocker; otherwise the audit directs the later multi-file fix toward the log/profile key split while overlooking where the second file is actually discarded.

Useful? React with 👍 / 👎.

Confirmé et corrigé dans 0db1f65. enqueue clé le job par sha256(session_id) seul, rend la main dès qu'un reçu completed/ existe, et os.link conserve la première charge d'un job encore en attente (session_queue.py:89-111). Sur le chemin Codex en file d'attente, le second rollout n'atteint donc jamais session_lifecycle. L'audit distingue maintenant la livraison en file d'attente (rejet à l'entrée, premier obstacle au critère 1 de #716) de l'appel direct (la séparation des clés journal/profil, toujours marquée inferred). La section « What a later slice can reuse » ne présente plus la clé session_id comme la bonne pour une session sur plusieurs fichiers.

Generated by Claude Code

… slice 1)

claude · 2026-11-05T18:51:03Z

Claude finished @cdeust's task in 1m 32s —— View job

Review 1 of 3: no Blockers. Verdict: approve

Read the diff against 2126d2a (the audit doc and a CHANGELOG entry; no code changed)
Ran five reviews: documentation accuracy, code quality, security, performance, test coverage
Posted the review: docs(audit): inventory the transcript capture producers (#716, slice 1) #874 (comment)

The cited lines match the code, and the fixup 0db1f65 addresses Codex's comment about the intake dedup. I didn't run the doc-claims or craftsmanship checks; I read the scripts instead. Neither can fail here: one doesn't scan docs/audits/ or CHANGELOG.md, and the other only checks Python. The /review-pr skill failed to load, so I ran the five reviews directly.

Notes, no action before merge

The capture_client.py:39 citation is off by one; the docstring it backs is at :40.
In P2, "at-least-once" for the spool is too broad. A failed or timed-out store is renamed *.rejected and never retried.
P2, P4 and P5 cite whole files with no line numbers, unlike the rest of the doc.

claude · 2026-11-05T18:51:23Z

PR #874 review (1 of 3), head `238a405`

I ran five reviews: documentation accuracy, code quality/CI, security, performance and test coverage. All but one of the cited lines match the tree. CI gates are unaffected: check_doc_claims.py doesn't scan docs/audits/ or CHANGELOG.md, and the craftsmanship gate only checks Python files. Nothing leaks.

Blocker

The P1 claim about Codex sessions is false (docs/audits/transcript-capture-producers-2026-11-05.md:50-53, repeated at :115-118).

The audit says a Codex session continued into a second rollout file "gets two log rows but one profile update". That can't happen on the Codex path:

Codex SessionEnd reaches session_lifecycle only through session_queue.py intake (plugins/hypermnesia-mcp-codex/hooks/hooks.json:92).
Intake keys the job by sha256(session_id) (session_queue.py:95-96).
If the first job already finished, its receipt exists, so the second SessionEnd returns at :97-98. Receipts are never deleted (the only writer is :203).
If the first job is still queued, os.link raises FileExistsError, which is swallowed (:108-111). The second payload and its transcript_path are discarded.

So Codex writes one log row and one profile update, both from the first file. The second file's activity reaches neither, so acceptance item 1 of #716 fails on both paths, not just the profile path. The stem / session_id split only applies to the direct Claude hook.

This matters because the remaining-slices plan aims at "the fix of the stem / session_id key split". On Codex that fix would change nothing: the drop happens at intake. Your own :57-58 already describes this short-circuit, so the audit contradicts itself.

Fix this →

Notes, no action before merge

The inventory misses a producer: the record_session_end MCP tool writes memories through remember (mcp_server/handlers/record_session_end_memory.py:46-81). The "five producers" count in CHANGELOG would change.
Some citations don't meet the "every claim cites a line" promise. P2's memory_exact_sql.py/remember.py/capture_drain.py (capture_drain.py:65-72) and P4/P5 have no line numbers, and capture_client.py:39 is the def line (the docstring is at :40).
P3/P4 should qualify two points. Both readers see only head and tail, so a file's middle is never captured. And "a long transcript that grows is never re-read" holds only for files over 64 KiB.

Verdict: block

claude · 2026-11-05T18:52:41Z

Review 1 of 3 — PR #874 (head `0db1f65`)

Five reviews were run: documentation accuracy, code quality, security, performance and test coverage. Every cited file:line was checked against the tree at 2126d2a, and the PR changes no code. No Blockers.

The cited ranges match the code:
- session_lifecycle.py:233-238/240-241/257-261/359-372
- session_queue.py:89-111/201-204
- capture_server.py:56-58
- backfill_helpers.py:79-84
- backfill_memories.py:339-345
- transcript_rollout.py:1-14
The Codex P2 comment on 238a405, that the intake dedup drops the second rollout, is addressed in the fixup 0db1f65.
check_doc_claims.py does not scan docs/audits/ or CHANGELOG.md, and check_craftsmanship.py only checks Python, so neither gate can fail on this PR.
The test-coverage claims hold:
- no test covers a session over two rollout files or a rewritten prefix;
- no import_sessions idempotency test;
- no compaction replay test.

Notes, no action before merge

capture_client.py:39 points at ) -> None:. The "never replayed" docstring is at :40; cite :37-42.
P2 spool: a failed or timed-out store is retired as *.rejected and never retried (capture_drain.py:41-45,64-71). "At-least-once" holds only for a crash during the store.
P2, P4 and P5 cite whole files with no line numbers, which falls short of the doc's "every claim cites the line read".

Verdict: approve

cdeust mentioned this pull request Nov 5, 2026

feat: specify and verify resumable incremental transcript capture across Codex rollouts #716

Open

4 tasks

cdeust marked this pull request as ready for review November 5, 2026 18:47

chatgpt-codex-connector Bot reviewed Nov 5, 2026

View reviewed changes

fixup! docs(audit): inventory the transcript capture producers (#716,…

0db1f65

… slice 1)

cdeust merged commit 2d6082d into main Nov 5, 2026
18 checks passed

cdeust deleted the fix/716-rollout-capture branch November 5, 2026 21:04

cdeust mentioned this pull request Nov 5, 2026

ci: run only the mutants of the changed functions in the mutation gate (ADR-1122) #898

Merged

10 tasks done

Navigation Menu

Repository navigation

docs(audit): inventory the transcript capture producers (#716, slice 1) - #874

docs(audit): inventory the transcript capture producers (#716, slice 1)#874
cdeust merged 2 commits into
mainfrom
fix/716-rollout-capture

cdeust commented Nov 5, 2026 •

edited

Loading

Uh oh!

chatgpt-codex-connector Bot commented Nov 5, 2026 •

edited

Loading

Uh oh!

claude Bot commented Nov 5, 2026 •

edited

Loading

Uh oh!

chatgpt-codex-connector Bot left a comment

Uh oh!

chatgpt-codex-connector Bot Nov 5, 2026

Uh oh!

cdeust Nov 5, 2026

Uh oh!

claude Bot commented Nov 5, 2026 •

edited

Loading

Uh oh!

claude Bot commented Nov 5, 2026

Uh oh!

claude Bot commented Nov 5, 2026

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

cdeust commented Nov 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Type of change

Test plan

Audit notes

Breaking changes

Uh oh!

chatgpt-codex-connector Bot commented Nov 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Codex Review Summary

Uh oh!

claude Bot commented Nov 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

PR #874 review (1 of 3)

Uh oh!

chatgpt-codex-connector Bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

chatgpt-codex-connector Bot Nov 5, 2026

Choose a reason for hiding this comment

Uh oh!

cdeust Nov 5, 2026

Choose a reason for hiding this comment

Uh oh!

claude Bot commented Nov 5, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Review 1 of 3: no Blockers. Verdict: approve

Uh oh!

claude Bot commented Nov 5, 2026

PR #874 review (1 of 3), head 238a405

Blocker

Notes, no action before merge

Uh oh!

claude Bot commented Nov 5, 2026

Review 1 of 3 — PR #874 (head 0db1f65)

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

cdeust commented Nov 5, 2026 •

edited

Loading

chatgpt-codex-connector Bot commented Nov 5, 2026 •

edited

Loading

claude Bot commented Nov 5, 2026 •

edited

Loading

claude Bot commented Nov 5, 2026 •

edited

Loading

PR #874 review (1 of 3), head `238a405`

Review 1 of 3 — PR #874 (head `0db1f65`)