Mission#

Build the core capability that lets Kelly do her job: watch a list of repos, see what changed since her last review, summarise the diff, and update or create docs in docs/ to reflect reality.

What to build#

A new skill at ~/.openclaw/workspaces/kelly/skills/docs-review/ (Kelly's local skills dir, isolated from the global skill registry).

SKILL.md#

Description: "Run a documentation review pass: scan watched repos for changes since the last review, identify which docs need updates, and write the updates."

Inputs: optional repos array (default: all repos in kelly/state/watched-repos.json).

Outputs:

• Updated docs/*.md files in each scanned repo (committed or staged depending on config)
• A summary at ~/.openclaw/workspaces/kelly/memory/<date>.md listing what was reviewed and what changed
• An entry appended to ~/.openclaw/workspace/docs/INDEX.md if a new project was added

Data files#

• kelly/state/watched-repos.json — list of { path, name, lastReviewedSha, since } objects
• kelly/state/last-review.json — timestamp + repo SHAs from the most recent run

Algorithm (rough)#

1. For each watched repo:
   1. git log --since="<last review timestamp>" --oneline → list of commits
   2. For each commit, git show --stat <sha> to see touched files
   3. If any of these are touched → docs may need an update:
      • app/api/**/*.ts (API routes)
      • lib/*.ts (server actions / shared logic)
      • components/**/*.tsx (UI components)
      • *.md files outside docs/ (config docs)
      • package.json dependencies changed
   4. For each affected area, identify which docs/*.md covers it. If none, propose creating a new one.
2. For each doc to update: read the current doc, read the relevant code, generate the updated doc.
3. Write the doc, append to memory log, optionally commit.
4. Update lastReviewedSha for each repo.

Acceptance criteria#

• Skill files exist and load without error
• kelly/state/watched-repos.json populated with at least 3 repos: ~/projects/mission-control, ~/.openclaw/workspace, ~/.openclaw/workspaces/kelly (meta — she watches her own home)
• Manual trigger works: openclaw agent --agent kelly --message "Run a docs review pass" produces real diffs in docs/ files
• No false invents: if Kelly can't tell what a piece of code does, she writes a TODO + opens a task assigned to Chris instead of guessing
• Commits land with message format docs(<project>): <one-line summary>

Notes#

• Don't burn cycles diffing things outside scope. Keep watched-repos small at first; grow it as she proves reliable.
• Kelly should NOT touch code, ever. Only docs/, README.md if the repo allows, and her own state files.