RoBrain hooks for Codex CLI / IDE

The same three lifecycle hooks as the Claude Code plugin, running on Codex's native hooks system (Codex's hook contract is Claude-compatible: same stdin fields, same hookSpecificOutput.additionalContext output, same event names).

Hook	What it does
SessionStart	Injects the always-on project summary — top decisions with their rejected alternatives.
UserPromptSubmit	Two-tier pre-task veto: deterministic POST /veto-scan + semantic search. Also stashes the prompt for capture.
Stop	Ships the completed turn (stashed prompt + last_assistant_message from Codex's own stdin) to Perception for server-side extraction. No transcript-format assumptions.

Everything fails open: a dead Perception never blocks a Codex session.

Install

npx robrain@latest up          # start the backend, once
npx robrain install            # detects Codex; wires MCP + hooks into ~/.codex/config.toml

robrain install materializes the hook scripts into ~/.robrain/hooks/codex/ and writes the [hooks] entries inside RoBrain's managed block in ~/.codex/config.toml. Codex asks you to trust the hooks on first run.

Source layout

session-start.mjs, user-prompt-submit.mjs, and lib.mjs are shared verbatim with the Claude Code plugin (plugins/claude-code/hooks) — one implementation, two clients. Only stop.mjs is Codex-specific: Codex hands the assistant reply on stdin, so capture pairs it with the stashed prompt instead of parsing a transcript.

Not covered: the Codex cloud/web agent

The cloud agent runs in a sandboxed container with no local node or localhost Perception — hooks and the sensing_* MCP tools can't reach it (AGENTS.md instructions still can). Covering it requires a hosted remote MCP endpoint — tracked on the Rory Plans cloud roadmap, deliberately not part of the OSS path.