Codex Toolkit accepts small, testable improvements. An issue or pull request should state the user request it addresses and the behavior that should change.
- Keep each skill self-contained. Do not add a hidden shared runtime.
- Put trigger conditions in the
descriptionfield ofSKILL.md, not in a body-only "when to use" section. - Keep detailed or time-sensitive material in
references/and link it directly fromSKILL.md. - Prefer read-only helper scripts. If a script can write, its scope and safety checks must be explicit.
- Do not copy unlicensed prompts, prose, or code.
| Change | Expected evidence |
|---|---|
| Trigger or routing behavior | Positive, negative, and overlap cases where relevant |
| Workflow or stopping point | Matching workflow scenario and adversarial review note |
| Helper script | A smoke fixture that proves output and input preservation |
| Time-sensitive package claim | Primary source, checked date, and refresh boundary |
| New skill | SKILL.md, agents/openai.yaml, routing cases, workflow scenario, and any required notices |
python scripts/validate_skill_pack.py . --as-of YYYY-MM-DD
python scripts/run_smoke_tests.py . --as-of YYYY-MM-DD
npm pack --dry-runUse the date on which you checked time-sensitive sources. The pull request should explain any warning instead of hiding it with an older date.
Keep one behavioral change per pull request when practical. Describe what you ran, what passed, and what still requires a live Codex routing check.