Verification and troubleshooting
Use this page when setup fails or you need to assess a validation report.
Troubleshoot setup
Section titled “Troubleshoot setup”| Symptom | Action | Check before continuing |
|---|---|---|
npm cannot find package.json or a script | Change to the installed creator directory. | Its manifest declares skills:install and preflight. |
| A source cannot be downloaded | Restore source access, then repeat npm run skills:install. | Run preflight after every install attempt; a failed install can leave partial sources. |
| Preflight reports content mismatch | Preserve changed files as evidence, then restore frozen sources with npm run skills:install. | Preflight returns installed-content-verified. Do not edit hashes to accept unexpected content. |
| A support reference is missing | Restore the complete locked set. | Preflight passes without substituting global skills. |
Python reports No module named 'yaml' | Provide PyYAML in a task-local environment. | The named Codex validator runs. Alternate YAML parsing is a different check. |
Freshness is unknown | Check source access and retain the dated result. | A locked run can remain reproducible while freshness is unknown. |
| The agent returns a plan without files | Restate the destination and acceptance check, then inspect the files. | The artifact exists and affected behavior has been tried or explicitly remains untested. |
Create a task-local validator environment outside locked source directories:
python3 -m venv /tmp/skill-creator-validation/tmp/skill-creator-validation/bin/python -m pip install PyYAML==6.0.3From the installed creator directory, validate a generated skill:
PYTHONDONTWRITEBYTECODE=1 /tmp/skill-creator-validation/bin/python \ .runtime/skills/codex-creator/scripts/quick_validate.py \ /absolute/path/to/generated-skillReplace the last path with your generated skill directory. A successful structural check reports Skill is valid!. Keep trial outputs outside locked dependency directories.
Recorded release evidence
Section titled “Recorded release evidence”The October 12 evaluation record reports these checks:
| Check | Recorded outcome | Limit |
|---|---|---|
| Three creator tasks and three matching baseline tasks | Both variants satisfied artifact assertions. | No demonstrated superiority, token saving, or latency improvement. |
| Structural validation | Creator and generated structures passed the named Codex validator. | Structure does not establish behavior. |
| Seven integrity tests | Passed receipt, tampering, local drift, missing source, missing reference, escaping symlink, and manifest mismatch cases. | Integrity does not establish instruction quality or freshness. |
| Clean-room public installation | Skills CLI 1.7.2 discovered the intended entrypoint and restored six dependencies. | Dated installation evidence. |
Automatic discovery through each runtime and generated skills acting on live business systems were not established. Description-only trigger judgments are proxies.
Check a new skill before relying on it
Section titled “Check a new skill before relying on it”Choose a representative task and a nearby request outside scope. Preserve raw inputs, the artifact, receipt, and observed results. Test helpers against real temporary fixtures, including a relevant failure.
Use independent review for consequential workflows. Compare against the previous skill or no skill before claiming an improvement. Report unavailable telemetry and missing execution as gaps.
See validation choices for the canonical guidance.