[dem-extra1]

Why

Claude's PR-activity harness reads .claude/skills/steward/SKILL.md from a PR's head branch, and that file overrides the harness defaults. The lab's rules live in ai-config's steward skill (Morrison-Lab/ai-config#4101).

What

• A new .claude/skills/steward/SKILL.md that points to skills/steward/SKILL.md in Morrison-Lab/ai-config. Add a steward pointer for the PR-activity harness pds#31 is the reference PR.

🤖 Generated with Claude Code

https://claude.ai/code/session_014YzWJF4E24Edb6RKXbkWaH

---

Generated by Claude Code

[github-actions]

Caution

Claude review did not run: the configured API credential is unusable.

No review ran at all, so this PR has not been reviewed. This comment exists so that fact is visible from the PR itself rather than only from the run log (gha#543).

The pre-flight check found that every configured API credential secret carries whitespace inside its value, so it cannot be sent as an HTTP Authorization header. The request would be rejected at the door, so the review was not attempted and no money was spent.

This is a repository configuration defect rather than anything about the diff. Nothing about this PR needs to change, and re-running the workflow will fail identically until the secret is repaired.

A repository admin fixes it under Settings -> Secrets and variables -> Actions by re-setting CLAUDE_CODE_OAUTH_TOKEN (or ANTHROPIC_API_KEY) to a single-line value with no spaces or line breaks. The usual cause is pasting a multi-line block -- a PEM key, a JSON credential file, a wrapped terminal copy -- where the token belongs. claude setup-token prints a fresh token to copy.

The Pre-flight credential shape check step in the linked run names which secret is affected and where the whitespace falls, without printing the value itself.

Until an external verdict lands, this PR is not clean. The red require-review is correct rather than a flake, and a self-review stands in for the missing verdict without replacing it.

View run

[dem-extra1]

review / claude-review, require-review and require-clean-verdict are red because the review never ran. The workflow's pre-flight check found whitespace inside the repo's CLAUDE_CODE_OAUTH_TOKEN (or ANTHROPIC_API_KEY) secret, the same problem as on #82. It's a repo configuration issue that this one-file change can't fix, and a re-run fails the same way. A repo admin needs to re-set the secret as a single line; a re-run of the review workflow will then review this PR.

---

Generated by Claude Code