Context
The structured-output protocol currently asks producer skills such as /requirements, /design, /breakdown, and /review-ai to append blocks like:
[/requirements] COMPLETE SKILL_OUTPUT:requirements
<!-- SKILL_OUTPUT:requirements
...
END_SKILL_OUTPUT -->
This is reasonable for downstream skill handoff: cross-verify can parse the metadata.
Codex integration problem
When these skills are used through Codex, the console/transcript exposes the raw HTML comment and the visible attribution line to the end user. Markdown comments are not hidden in that surface, so every requirements/review interaction can emit a large YAML-like internal metadata block.
This is not sensitive data by itself, but it is noisy, confuses users, and weakens the intended concise output of the Significance profile.
Reproduction
- Use
/requirements through Codex on a multi-file change.
- Observe the skill response in the Codex console/transcript.
- The
SKILL_OUTPUT:requirements attribution and HTML-comment payload are rendered as ordinary output instead of remaining invisible metadata.
Expected
- Normal Codex user-facing output should contain the concise human summary only.
- Structured handoff metadata should remain available to a real consumer such as
/cross-verify, but should not be printed when no consumer needs it.
Suggested compatibility-safe options
- Emit
SKILL_OUTPUT only when a consumer is explicitly requested or artifacts are persisted.
- Add a Codex-aware output mode that stores handoff metadata in a local artifact/session channel rather than response text.
- At minimum, make the visible attribution line optional and document the renderer limitation.
Please preserve the existing Claude/Markdown workflow where HTML comments are actually hidden. This looks like a cross-runtime output-contract mismatch rather than a defect in the requirements content itself.
Evidence
skills/requirements/SKILL.md requires the structured output block at the end of the result.
guides/structured-output-protocol.md says the HTML comment is invisible when rendered.
- In Codex console it is visibly rendered verbatim.
Context
The structured-output protocol currently asks producer skills such as
/requirements,/design,/breakdown, and/review-aito append blocks like:This is reasonable for downstream skill handoff:
cross-verifycan parse the metadata.Codex integration problem
When these skills are used through Codex, the console/transcript exposes the raw HTML comment and the visible attribution line to the end user. Markdown comments are not hidden in that surface, so every requirements/review interaction can emit a large YAML-like internal metadata block.
This is not sensitive data by itself, but it is noisy, confuses users, and weakens the intended concise output of the Significance profile.
Reproduction
/requirementsthrough Codex on a multi-file change.SKILL_OUTPUT:requirementsattribution and HTML-comment payload are rendered as ordinary output instead of remaining invisible metadata.Expected
/cross-verify, but should not be printed when no consumer needs it.Suggested compatibility-safe options
SKILL_OUTPUTonly when a consumer is explicitly requested or artifacts are persisted.Please preserve the existing Claude/Markdown workflow where HTML comments are actually hidden. This looks like a cross-runtime output-contract mismatch rather than a defect in the requirements content itself.
Evidence
skills/requirements/SKILL.mdrequires the structured output block at the end of the result.guides/structured-output-protocol.mdsays the HTML comment is invisible when rendered.