test(server): record Codex multi-agent V2 subagents live on 0.156.1

[juliusmarminge]

The subagent_v2 and subagent_v2_nested Codex transcripts were hand-built, with synthetic thread ids and hand-ordered frames. This records both live against Codex 0.156.1 on gpt-5.6-sol, which runs multi-agent v2.

Stacked on #13505, which adds the recorder changes this uses: the adapter's turn/start defaults, --model, and <workspace>.

What changed

• Recorder:
  • adds the subagent_v2 and subagent_v2_nested scenarios;
  • adds a per-run thread/start config override. The nested scenario sets agents.max_depth: 3, because by default Codex only lets the root agent spawn. Replay ignores thread/start config, so the transcript still matches the adapter's frames.
• Transcripts: both are straight recordings, with no hand edits.
• Prompts: the new prompts ask for the spawns explicitly. On gpt-5.6-sol, "just say hello" answers directly and spawns nothing.
• Model: each fixture's Codex variant uses its own modelSelection (gpt-5.6-sol). The shared selection is unchanged.
• Assertions: they now use the agent paths and answers the model actually chose (/root/hello; /root/relay_one/relay_two/hello_child). The CodexReplayFixtures expectations use the recorder's method labels instead of the hand-written role labels.

What the live recordings showed

• In both recordings, and at every nesting level, Codex emits:

  1. the child's thread/status/changed (idle),
  2. the parent's subAgentActivity(started),
  3. the child's thread/status/changed (active) and turn/started.

  So the adapter registers the child before its first turn. This matches the 0.145 capture in codexMultiAgentWire.json. The "child turn starts before the parent's activity" race did not happen, so this PR contains no adapter change.

• Children finish first. Each parent then emits subAgentActivity(completed), a wait collabAgentToolCall, and its own answer. The adapter projects all four nested threads with the right lineage and results.

Verification

• vp test run src/orchestration-v2/testkit/OrchestratorReplayFixtures.integration.test.ts -t "subagent_v2|thread_rollback": 4 passed.
• vp test run on CodexReplayFixtures, OrchestratorReplayFixtures.contract, CodexAdapterV2.test.ts, ThreadFork, ThreadMergeBack and scripts/codexReplayRecordingRecords.test.ts: 139 passed.
• vp exec tsc --noEmit -p . in apps/server: no errors or warnings. vp run knip:check passes. vp lint on the touched files reports only an unused assertUserMessagesExclude in fixtures/shared.ts, which is already on the base branch.
• Recorded with T3_CODEX_BIN=<codex 0.156.1> node scripts/record-codex-app-server-replay-fixture.ts --scenario subagent_v2 --model gpt-5.6-sol, and the same with --scenario subagent_v2_nested.
• Not addressed here: plan_questions/codex fails on the base branch too. Its 0.120 transcript has no isBlocking on item/tool/requestUserInput, so the 0.156 decode rejects the request. Replay then stalls for 60 seconds instead of reporting the error.

Model: Claude Opus 5.5 (Claude Code)

🤖 Generated with Claude Code

---

[macroscopeapp]

Approvability

A prior Macroscope approval was dismissed. Re-evaluating e79ede0…

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

ℹ️ No successful main baseline artifact is available yet. This run establishes the initial measurement.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	—	4.9 KiB	—	6.8 KiB	✅
Codex	Thread snapshot wire	—	3.7 KiB	—	4.9 KiB	✅
Codex	Live turn WebSocket wire	—	1.1 KiB	—	2.0 KiB	✅
Codex	Live turn WebSocket decoded	—	20.4 KiB	—	29.3 KiB	✅
Codex	Live turn messages	—	1	—	8	✅
Claude	Total thread wire	—	4.9 KiB	—	6.8 KiB	✅
Claude	Thread snapshot wire	—	3.7 KiB	—	4.9 KiB	✅
Claude	Live turn WebSocket wire	—	1.2 KiB	—	2.0 KiB	✅
Claude	Live turn WebSocket decoded	—	20.8 KiB	—	29.3 KiB	✅
Claude	Live turn messages	—	2	—	8	✅

Baseline: unavailable · PR result: a955a2c · Source CI: failure

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 106.1 KiB
• Claude decoded thread snapshot: 106.4 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

Dismissing prior approval to re-evaluate a955a2c

Dismissing prior approval to re-evaluate e79ede0