Read-only review of 8d9ec2f3c0b5e6f468063f52759b348dc29ddb89 against 7c3b3b3605.
Native executable: codex-cli 0.163.0, Linux x86_64.
Actual head CodexAdapterV2 with isolated CODEX_HOME, mock Responses HTTP endpoint, empty mock MCP endpoint.
Mock model gpt-5.4; auto-compact limit 100; context window 20000; post-turn compaction disabled.
Mock Responses usage: 10000 input tokens, 10 output tokens. All response texts are Done.

Regression test against base adapter (supplied via Vite read-only load hook): FAIL
Observed item/completed -> thread/inject_items -> turn/completed.
Same regression test at head: PASS. Full CodexAdapterV2.test.ts: 86/86 PASS.
Read-only additional replay variants: interrupted, failed, repeated compaction, other-thread completion: 4/4 PASS.

Native sequence: seed turn, automatic compaction turn, immediately following turn, manual compaction.
T3 next turn starts immediately on terminal event without waiting for restore acknowledgement.
IN reply id=1 OK
IN reply id=2 OK
OUT turn/start id=3
IN reply id=3 OK
IN turn/completed completed
OUT turn/start id=4
IN reply id=4 OK
IN item/completed contextCompaction
IN turn/completed completed
OUT thread/inject_items id=5
OUT turn/start id=6
IN reply id=5 OK
IN reply id=6 OK
IN item/completed contextCompaction
IN turn/completed completed
OUT thread/inject_items id=7
OUT thread/compact/start id=8
IN reply id=7 OK
IN reply id=8 OK
IN item/completed contextCompaction
IN turn/completed completed
OUT thread/inject_items id=9
{"result": "terminal", "counts": {"responses": 6, "compactions": 3, "answers": 3, "injects": 3, "terminals": 4}}

All native subprocesses exited via their tracked Effect scopes; mock HTTP servers closed in finally.
No real provider credentials or external model calls were used.
Limit: no macOS/Windows execution, full T3 delegate-parent wake integration, or native autonomous goal continuation.