Skip to content

Commit 0db1f65

Browse files
committed
fixup! docs(audit): inventory the transcript capture producers (#716, slice 1)
1 parent 238a405 commit 0db1f65

1 file changed

Lines changed: 16 additions & 7 deletions

File tree

‎docs/audits/transcript-capture-producers-2026-11-05.md‎

Lines changed: 16 additions & 7 deletions
Original file line numberDiff line numberDiff line change
@@ -47,10 +47,18 @@ not executed.
4747
file stem, falling back to `session_id`
4848
(`session_lifecycle.py:240-241`, `mcp_server/shared/injection_markers.py:41-47`); the
4949
profile update is keyed by `sha256(session_id)`
50-
(`session_lifecycle.py:359-372`). *Inferred:* a Codex session continued
51-
into a second rollout file gets two log rows but one profile update, so
52-
the second file's activity never reaches the profile — acceptance item 1
53-
of #716 fails on the profile path.
50+
(`session_lifecycle.py:359-372`). A Codex session continued into a
51+
second rollout file is lost at two places, depending on the path:
52+
- *Queued Codex delivery* (the hook path): `enqueue` keys the job by
53+
`sha256(session_id)` alone, returns when that job's completed receipt
54+
exists, and `os.link` keeps the first payload of a job still pending
55+
(`plugins/hypermnesia-mcp-codex/scripts/session_queue.py:89-111`). The
56+
second SessionEnd never reaches the lifecycle handler: no second log
57+
row, no second profile update. This intake dedup is the first blocker
58+
for acceptance item 1 of #716.
59+
- *Direct lifecycle call* (no queue): *inferred* from the two keys above,
60+
two log rows but one profile update, so the second file's activity
61+
never reaches the profile.
5462
- **Duplicate delivery.** Log row skipped when its `sessionId` is already
5563
present (`session_lifecycle.py:257-258`), but only the last 1000 rows are
5664
kept (`:259-261`), so a replay older than that window appends again
@@ -113,9 +121,10 @@ not executed.
113121
## What a later slice of #716 can reuse
114122

115123
- ADR-1084's queue-then-receipt pattern (P1) is the only replay-safe
116-
delivery in the tree; it is keyed by `session_id`, not by file, which is
117-
the right key for a multi-file Codex session but is not applied to the
118-
session-log row.
124+
delivery in the tree. It is keyed by `session_id` alone, so it treats a
125+
second rollout of the same session as a duplicate and drops it: reusing
126+
it for multi-file capture needs a delivery identity that also names the
127+
file (or the turns it carries), not only the session.
119128
- The exact-replay dedup (P2) absorbs identical replays only; it cannot
120129
express "turn N of session S was replaced".
121130
- Nothing here implements a durable cursor, prefix-rewrite detection or

0 commit comments

Comments
 (0)