[blafourcade]

🎯 What & why

Codex calls were measured, but no Codex session was ever asked for its task: each hook returned early because its guard accepted only Claude Code's payload. The same three hooks now serve Codex too, and Claude Code behaves exactly as before.

🛠️ How it works

• Which tool sent the payload. hooks/lib/tool.cjs checks for Claude Code first, using the unchanged isClaudePayload. Otherwise the payload is Codex when it has a session_id and its transcript_path is a Codex session file. The path decides, not CLAUDE_CODE_SESSION_ID, because a Codex started inside Claude Code inherits that variable. The shared fixture section is now hookTool (claude | codex | null).
• Whether a person can see a block. Codex shows a block's reason only in its TUI, and a hook receives nothing that names the surface. The session file's first line does: the hook reads up to 1 MiB of it. The largest first line measured was 48 KiB. A prompt is held only when all of these hold:
  • the first line is a session_meta with source: cli and originator: codex-tui;
  • the payload has no agent_id;
  • permission_mode is default.
    Anything else lets the prompt through (presence.cjs). A change in Codex can only stop the asking, never block work.
• The session a declaration binds. aidd telemetry task binds AIDD_TELEMETRY_SESSION_ID, else CLAUDE_CODE_SESSION_ID (declaringSessionOf, runtime/wiring/telemetry.ts). Only the Codex hook sets the first one, to its payload's session_id. A sub-agent's calls carry their root's session, so they take its task.
• What a Codex block returns. Only {decision, reason}. Codex was never seen to accept suppressOriginalPrompt, so the typed declaration stays visible there; it is still not sent to the model.
• Session start. For Codex, session-start.cjs does nothing, not even the consent check, because Codex gives a hook no process id. catch-up.cjs ingests for both tools.
• The merged .codex/hooks.json. It kept only type, command, timeout and statusMessage, so async was lost. Each entry is now kept whole, as the plugin tree (renameCodexHookEvents) already does (flat-hooks-merge.ts). Across plugins/*/hooks/hooks.json the entry keys are only type, command, timeout and async. One existing test pinned a timeout of the wrong type being dropped; it now expects that value to pass through, as it does on the plugin route.

🧪 How to verify

• node scripts/check-tests-leave-git-alone.js -- node --test 'scripts/__tests__/**/*.test.js', sandboxed: 476 tests pass. The hook rows of the presence rule are in codex presence: a prompt is held when, and only when, every condition holds.
• cd cli && pnpm test: includes the new end-to-end test asks a Codex TUI session, binds its declaration to that session, and attributes its calls and its sub-agent's.
• The red runs, the hook mutations done by hand, and every gate's output are in evidence.md.

⚠️ Heads-up

• Not observed yet; real run pending. The real run with the Codex TUI has not been done: trust the hooks with /hooks, check the TUI holds a prompt, check codex exec never does. The trust and async effects below come from reading the openai/codex source.
• Trust again after this change. Codex trusts a hook by its entry. A project whose merged .codex/hooks.json gains async gets a changed entry, so Codex asks for that hook to be trusted again. The plugin's own hooks.json is unchanged.
• Limits, written into the usage contract and the README:
  • Nothing asks until the hooks are trusted.
  • codex exec, the IDE extension, the desktop app and full-access mode are never asked.
  • A TUI thread resumed by codex exec with an auto-review approver would be held with nobody to see why. Not observed.
  • A Codex /clear or fork is asked again.

🔗 Linked issue

None. Follows #1057: Codex is measured, now it is asked for its task.

✅ I certify

• I DO CERTIFY I READ EACH LINE OF THE PULL REQUEST BECAUSE I AM A SOFTWARE ENGINEER, NOT A AI PUPPY.

🤖 Generated with Claude Code

https://claude.ai/code/session_01TnCgcujM2xi6T6dkmVogVw