Codex Goals — persistent multi-turn objectives with evidence-based completion

Codex Goals (/goal) are persistent objectives that let Codex work toward a defined outcome across many turns in a thread under a completion contract — it keeps going until the objective is verified complete against evidence (tests, benchmarks, artifacts) rather than stopping after each turn. Built for multi-turn tasks where the next step depends on the last result: performance optimization, flaky-test investigation, dependency migrations, research reproduction.

Without a goal, you re-state the objective after every result and the agent stops at each turn. A goal makes the agent continue autonomously toward a measurable finish line while idle and within budget — "you define the outcome, Codex works against the evidence in the thread." The Claude Code counterpart is the Stop-hook outer-loop pattern (see B-130); Goals is the Codex-native, first-class version with its own command surface.

1. Install / update Codex: npm install -g @openai/codex@latest (requires Codex ≥ 0.128.0).
2. Create a goal: /goal <outcome>. Inspect with /goal; control with /goal pause, /goal resume, /goal clear.
3. Write the goal to the structure template: <desired end state> verified by <specific evidence> while preserving <constraints>. Use <allowed inputs>. Between iterations, <iteration policy>. If blocked, <stop condition>.
4. Let it run. Codex continues only while threads are idle and budget remains; completion requires evidence-based verification; hitting the budget triggers a progress-and-blockers summary, not a false success.

Weak vs strong goal — the strong one names the outcome, the verification surface, and the constraint:

# weak
/goal Improve performance

# strong
/goal Reduce p95 latency below 120 ms on checkout benchmark while keeping correctness suite green

• Goals scope to a single thread, not global project state.
• Skip goals for single edits, vague objectives, or tasks with no auditable finish line — a normal prompt is better there.
• See also B-130 (Outer-loop autonomy / Ralph) — the Claude Code equivalent of the same "work autonomously until done" idea.