Dev Intel: Codex goals and lifecycle events

Codex 0.133.0 の goals / permission profiles / lifecycle events を、OpenClaw heartbeat の次の設計材料として圧縮したメモ。

Generated: 2026-05-22T21:11:00+09:00

Lane: 開発ネタ発掘

Why this is useful:

Codex 0.133.0 is not just a feature bump. The notable shift is that goals now persist across active turns, permission profiles are inspectable/refreshable, plugin discovery is more visible, and extensions can observe subagent/tool/turn/approval lifecycle events. That is close to the shape OpenClaw heartbeat has been trying to build manually: a run is not “one chat reply”, it is a stateful work loop with goals, permissions, events, and handoff points.

What I made/changed:

I made this source-backed note and rendered it as a View. The practical steal for OpenClaw is a small run_lifecycle_event envelope:

• goal_id: what the run is trying to finish
• permission_profile: what the run may do without asking
• event_type: turn started, tool used, approval requested, subagent started, output shipped, run paused
• evidence_uri: the artifact, log, View, diff, or test proving progress
• owner_next_action: only present when 健人くん has a real decision to make

Sources/Evidence:

• OpenAI Codex changelog, 2026-05-21: goals enabled by default, remote-control readiness/status, permission profile APIs, plugin discovery visibility, lifecycle events for subagents/tools/turn metadata/approval processing. https://developers.openai.com/codex/changelog
• Community codex-cli-hooks: hook surface expanded from SessionStart/Stop to UserPromptSubmit, Pre/PostToolUse, PermissionRequest, Pre/PostCompact. https://github.com/shanraisshan/codex-cli-hooks
• Codex hooks discussion: users wanted completion/feedback hooks first, then broader event hooks for interruption and follow-up. https://github.com/openai/codex/discussions/2150

Prediction:

If OpenClaw records heartbeat/task work as lifecycle events instead of prose reports, owner notifications should get shorter and more useful: “what changed, proof, next decision” can be generated from state instead of rewritten every time.

Verify by:

• Check whether future heartbeat artifacts can point to one goal_id and one owner_next_action, or intentionally stay silent.
• Check whether repeated Telegram/heartbeat issues become event patterns instead of fresh one-off explanations.

Observed:

The current heartbeat already has the pieces scattered across guard state, task ledger, View artifacts, and creative state. The missing piece is a single event vocabulary that ties them together.

Next safe action:

Add a local-only proposal for run_lifecycle_event to the heartbeat/task docs, then instrument one low-risk script path later. Do not change external delivery behavior from heartbeat.

Notify:

yes — this is source-backed and directly affects how 健人くん can make OpenClaw/ひめの less chatty and more inspectable.