fix: persist Codex review threads

[dinhnguyenminhhoang]

Summary

Persist Codex threads created by both review commands so their session IDs
remain resumable after the broker restarts.

Fixes #529

Changes

• Make native /codex:review threads non-ephemeral
• Enable thread persistence for /codex:adversarial-review
• Add regression assertions verifying both review paths start threads with
  ephemeral: false

Testing

• Targeted native review persistence test passes
• Targeted adversarial review persistence test passes
• Full test suite was also executed on Windows
• Existing Windows/environment-specific test failures are unrelated to this change

[urda]

Thanks, this matches what I asked for in #529. Two things before merge:

1. Please confirm a real review run now writes a rollout with token usage to ~/.codex/sessions (e.g. it shows up in ccusage codex). The tests only assert the fake fixture's flag.
2. Optional: pass a threadName for adversarial runs like the task path does, so persisted review threads aren't unnamed in the sessions tree.

[dinhnguyenminhhoang]

Thanks, this matches what I asked for in #529. Two things before merge:

1. Please confirm a real review run now writes a rollout with token usage to ~/.codex/sessions (e.g. it shows up in ccusage codex). The tests only assert the fake fixture's flag.
2. Optional: pass a threadName for adversarial runs like the task path does, so persisted review threads aren't unnamed in the sessions tree.

Thanks — I verified this using real Codex CLI 0.144.4 runs on this branch.

Native /codex:review:

• Completed successfully with a persistent thread ID.
• Wrote a non-empty rollout under ~/.codex/sessions.
• The rollout contains review lifecycle and message events, but no
  token_count events.

/codex:adversarial-review:

• Completed successfully with a persistent thread ID.
• Wrote a rollout under ~/.codex/sessions.
• The rollout contains 5 token_count events with non-zero usage
  (169597 input tokens, 940 output tokens, 170537 total tokens).

So thread persistence is working for both review paths, while local token
accounting is currently present only for the adversarial turn/start path.
The native review/start path does not appear to emit or persist
token_count events in Codex CLI 0.144.4.

Would you prefer this PR to remain scoped to thread persistence, with native
review usage accounting tracked separately, or should that be addressed
before merging?

[urda]

Keep it scoped to persistence, this delivers what #529 asked of the plugin. The missing token_count events on review/start look like Codex CLI behavior, not something this PR can fix. I'll track native review usage accounting separately. LGTM.

[dinhnguyenminhhoang]

Keep it scoped to persistence, this delivers what #529 asked of the plugin. The missing token_count events on review/start look like Codex CLI behavior, not something this PR can fix. I'll track native review usage accounting separately. LGTM.

Thanks for confirming and for the review. I'll keep the PR scoped to thread persistence.