Best possible solution:

Merge the focused regression test after exact-head CI finishes, and handle the WebChat overwrite behavior in the linked implementation issue or PR.

Do we have a high-confidence way to reproduce the issue?

Yes for the PR coverage path: the added test performs two mirror writes with distinct scopes and inspects the resulting JSONL branch. I did not run local tests because this review is read-only, but the source path and PR-head checks support the scenario.

Is this the best way to solve the issue?

Yes: for a test-only PR, the updated approach is boundary-clean and narrowly covers consecutive Codex mirror turns without adding a new public API or reaching into core internals.

What I checked:

• Updated PR diff: The final PR head adds only one test in the Codex plugin transcript mirror test file and removes the earlier direct import from src/gateway/session-utils.fs.js. (extensions/codex/src/app-server/transcript-mirror.test.ts:122, f58124c5f38a)
• Boundary policy: The scoped extension guide treats bundled plugins like third-party plugins and forbids importing core internals from src/**; the updated test imports only plugin SDK helpers and the local transcript mirror module. (extensions/AGENTS.md:3, f58124c5f38a)
• Current implementation path: Current main mirrors app-server result snapshots with a per-turn idempotency scope and the mirror helper appends through the shared session transcript append path under a session write lock. (extensions/codex/src/app-server/run-attempt.ts:1778, 9dc3271efbd7)
• Existing append behavior: appendSessionTranscriptMessage reads the current transcript leaf and appends the next message with parentId set to the prior leaf, which is the behavior the new regression test locks down for consecutive mirror calls. (src/config/sessions/transcript-append.ts:188, 9dc3271efbd7)
• CI status: The exact PR head shows the previously failing check-additional-boundaries-c now successful; one unrelated broader check was still in progress at review time. (f58124c5f38a)
• Related bug context: The PR references the open WebChat transcript-overwrite regression report, but it is regression coverage only and should not close that issue by itself.

Likely related people:

• steipete: Recent GitHub path history shows repeated work on transcript-mirror.ts, transcript-append.ts, and the async session transcript changes that this regression test exercises. (role: recent maintainer and transcript mirror/append owner; confidence: high; commits: c02bf2f399db, f7ed29e11812, 6147e1b91d3e; files: extensions/codex/src/app-server/transcript-mirror.ts, src/config/sessions/transcript-append.ts)
• vincentkoc: Recent history shows work on the Codex run-attempt caller path and gateway transcript reader behavior adjacent to the reported session-history loss. (role: recent adjacent maintainer; confidence: medium; commits: e9ca63cf06ec, eb1a0aa574b5, d122a3492d17; files: extensions/codex/src/app-server/run-attempt.ts, src/gateway/session-utils.fs.ts)

Remaining risk / open question:

• This is regression coverage only; the broader WebChat overwrite report remains open and should be fixed or resolved separately.
• One exact-head CI check was still in progress during this read-only review, although the prior boundary failure is now green.

Codex review notes: model gpt-5.5, reasoning high; reviewed against 90d25d59c610.