Repository navigation
Safe-output comments cannot carry hidden machine-readable state because every HTML comment is stripped #67174
Description
Summary
sanitize_content_core.cjs runs removeXmlComments on every safe-output body (add_comment and the others), so an agent or a post-agent step cannot store hidden, machine-readable state in a comment. gh-aw appends its own markers (<!-- gh-aw-agentic-workflow: ... -->, <!-- gh-aw-workflow-id: ... -->) after sanitization, so hidden markers are an accepted pattern internally but not available to workflow authors.
Use case
An on-demand review that supports -increment needs the commit the previous review covered. We ended every review comment with <!-- gh-aw--state: {"xxx_sha":"..."} --> and read it back from the PR's comment history (the cache-memory alternative does not persist for issue_comment-triggered runs). The marker was present in the agent's safe-output item and in the post-agent step's output, but absent from every comment GitHub stored, so every -increment request failed with "no baseline".
Reproduction
Running sanitizeContent (v0.91.1) on a body that ends with a hidden marker <!-- gh-aw-w-state: {"xxx_sha":"<sha>"} -->:
| Marker form | Result after sanitization |
|---|---|
<!-- ... --> |
removed |
[//]: # (...) |
[//]: # (content dropped) |
[label]: <https://...?x=<sha>> |
URL replaced with a redacted placeholder |
| Visible line with the SHA in backticks | kept |
Expected
One of:
- a safe-output option that preserves HTML comments whose text starts with an allow-listed prefix (for example
allow-html-comments: ["gh-aw-state:"]), still stripping everything else, or - a documented, supported way to attach hidden metadata to a safe-output comment.
If neither is wanted, the safe-outputs documentation should say that hidden markers are removed so authors do not rely on them.
Workaround today
Append a visible line and parse that.
Activity
pelikhan commentedon Oct 11, 2026
/scout research how to use safe-outputs data to store structured data in safe outputs
github-actions commentedon Oct 11, 2026
github-actions commentedon Oct 12, 2026
test body
pelikhan commentedon Oct 16, 2026
@copilot determine if the "data" feature in safe-outputs provide a solution for this.
pelikhan commentedon Oct 16, 2026
Current-state refresher (2026-10-16): Partially addressed — supported structured state exists, but it is not hidden metadata.
The maintainer's latest question has a concrete answer:
safe-outputs.datacan carry the previous-review SHA, provided the workflow enables data (optionally with a schema) and the safe-output item supplies adataobject. The current documentation describes this channel; the validator appends it as fenced JSON so it survives sanitization. Regression tests cover this behavior.However,
add-commentrenders a visible “Structured data” JSON block, not an HTML comment. The sanitizer still removes HTML comments. Thus data addresses persistence/machine readability, not the exact hidden-state requirement.#69437 is open as a draft with no changed files at inspection. Remaining decision: accept/document the visible structured-data approach or design a separate reviewed hidden-metadata mechanism. Static inspection; tests were read, not run.