Codex Performance Goal

Optimize performance only against an explicit evaluator. This is the native Codex adaptation of OMX $performance-goal: keep evaluator-gated progress, goal honesty, small reversible patches, and completion audit; replace .omx/goals/performance with repo-local benchmark artifacts or task/work-trace files when durable state is needed.

Contract

• Do not optimize before a baseline and pass/fail contract exist.
• Define the metric: latency, throughput, memory, cost, bundle size, accuracy/latency tradeoff, or another measurable target.
• Prefer existing project benchmarks and profiling scripts.
• Keep changes small and reversible.
• Record every measurement with command, environment, inputs, and result.
• Do not mark a goal complete until the evaluator passes and regression checks are clean.

Workflow

1. Clarify objective and metric.
2. Find or create the evaluator: command, benchmark script, profiling procedure, or reproducible manual measurement.
3. Capture baseline before edits.
4. Identify bottleneck hypotheses from evidence.
5. Apply one optimization at a time.
6. Rerun evaluator and regression tests.
7. Keep, revert, or revise based on evidence.
8. Finish with before/after results and residual risk.

Goal Mode

When Codex goal tools are active, treat the active goal as focus/accounting only. Call update_goal({status:"complete"}) only after evaluator pass plus completion audit proves the whole objective.

Output Shape

• Objective and metric
• Evaluator contract
• Baseline
• Changes tried
• Final measurement
• Regression checks
• Kept/reverted decisions
• Residual risk and follow-up

References

Read references/performance-evidence-checklist.html for measurement quality gates.