fix(hooks): ship hooks.codex.json so Codex session start gets a JSON envelope

[STiFLeR7]

Fixes #156.

What changed and why

Codex requires SessionStart hook output to be either empty stdout or a single JSON envelope. It rejects the plain-text output that always-on.mjs writes, which Claude Code accepts directly, and silently drops the entire ruleset with no indication of which hook failed.

Codex reads hooks/hooks.codex.json in preference to hooks/hooks.json when present, the same mechanism the issue points to in remember@claude-plugins-official. Added hooks/codex-session-start.mjs, a thin wrapper that runs the existing always-on.mjs unchanged, captures its stdout, and re-emits it as a hookSpecificOutput envelope when non-empty. hooks.json and always-on.mjs are untouched, so Claude Code keeps receiving plain text exactly as before.

Observable behavior before and after

Before: a Codex session with the always-on flag set shows "Hook failed: hook returned invalid session start JSON output" and the ruleset never reaches the model.
After: the same session receives a valid hookSpecificOutput envelope containing the same ruleset text, with no hook failure.

Safety and compatibility

No change to always-on.mjs, hooks.json, or the sh/PowerShell fallbacks. The new wrapper only intercepts stdout for the duration of the import and restores it immediately after, matching the existing opt-in/fail-safe/no-network behavior of the hook it wraps. If always-on.mjs exits early with no output (no opt-in flag, missing skill file, or an internal error), the wrapper also emits nothing, preserving the existing silent no-op behavior.

Verification performed

Ran the full hooks test suite locally (Node available, sh/PowerShell not on this machine): python -m pytest tests/test_always_on_hooks.py -v, 8 passed. Updated run_codex_hook to read hooks.codex.json instead of hooks.json, added test_codex_command_wraps_the_hook_output_in_the_session_start_envelope asserting the parsed JSON envelope shape and hookEventName, and added test_codex_hook_uses_a_launcher alongside the renamed test_claude_hook_uses_a_launcher so both launcher shapes are covered independently. The two existing silent-output tests (no opt-in flag, missing plugin) needed no changes since the wrapper already emits nothing when always-on.mjs produces no text.

Authorship and provenance

Autonomous agent-authored. I read the issue, the existing hooks.json launcher pattern, and always-on.mjs, then implemented and tested the fix with no human writing code. A human reviewed the diff and the test run before this PR was opened. No failed checks to report; the full existing test file plus the new/updated cases all pass locally.

[ayghri]

I was not able to reproduce this issue. Official documentation (https://learn.chatgpt.com/docs/hooks#sessionstart) says plain-text output is accepted.

[STiFLeR7]

Checked the docs you linked: they do confirm plain text on stdout is accepted for SessionStart, which contradicts what I assumed. I can't explain the reporter's exact error (codex-cli 0.154.0, 'invalid session start JSON output') from that alone, maybe a version-specific behavior or another installed hook's output getting merged with this one. Could use the reporter's help narrowing that down before deciding whether this PR is still needed.