knowledge-repo

name	knowledge-repo
description	Scan merged PRs and propose updates to AI context files (CLAUDE.md, AGENTS.md) and skill files (SKILL.md, prompts) as a git-apply-able patch. Non-interactive.
user-invocable	true
allowed-tools	Bash, Read, Write, Agent, Glob

knowledge-repo

Scan merged PRs from the last N days, extract knowledge relevant to AI agent context, and propose updates to context and skill files as a git apply-able patch.

Arguments

Parse $ARGUMENTS for:

• --days N (default: 7) — how far back to scan merged PRs

Model Selection

• Extract agents (Phase 3): use sonnet — mechanical data extraction, one PR per agent
• Synthesize, Review, Revise agents (Phases 4-6): use opus — requires judgment, cross-PR reasoning, and style matching

Pipeline

Execute these phases in order. Exit early where indicated.

Phase 1: Setup

1. Read the repo context file provided by the CI runner:

   cat artifacts/repo-context.json
   

   This contains forge (github/gitlab) and slug (owner/repo). If it does not exist, print an error and STOP.

2. Create the artifacts directory structure:

   mkdir -p artifacts/pr-data artifacts/pr-extractions
   

3. Run bash $SKILL_DIR/scripts/list-context-files.sh artifacts/context-file-list.txt

   • If exit code is non-zero (no context files found), write this run report and STOP:

     echo '{"early_exit":"no_context_files","prs_scanned":0}' > artifacts/run-report.json
     

Phase 2: Verify PR Data

PR data is pre-fetched by the CI runner and placed in artifacts/pr-data/.

1. Count the PR data files:

   ls artifacts/pr-data/*.json 2>/dev/null | wc -l
   

   If zero, write this run report and STOP:

   echo '{"early_exit":"no_prs","prs_scanned":0}' > artifacts/run-report.json
   

Phase 3: Extract

For each .json file in artifacts/pr-data/:

1. Extract the PR ID from the filename (e.g., 123.json → 123).
2. Read the file $SKILL_DIR/prompts/extract-agent.md.
3. In the prompt text, replace every {ID} with the actual PR ID.
4. Dispatch a background Agent with the constructed prompt, using model sonnet.

Dispatch in waves of up to 10 agents. After dispatching a wave:

• Poll every 30 seconds for the expected extraction files (artifacts/pr-extractions/{id}.md)
• Timeout after 5 minutes per wave
• Log any timed-out PR IDs and continue to the next wave

After all waves complete, check if any extraction file contains knowledge items. Read each artifacts/pr-extractions/*.md file and look for content under the ## Knowledge Items heading beyond just the heading itself.

If NO extraction file contains any knowledge items, write this run report and STOP:

echo '{"early_exit":"no_knowledge","prs_scanned":N,"prs_with_knowledge":0}' > artifacts/run-report.json

(Replace N with the actual count of PR data files.)

Phase 4: Synthesize

1. Read $SKILL_DIR/prompts/synthesize-agent.md.

2. Dispatch a foreground Agent with the prompt, using model opus.

3. After the agent completes, check for changes:

   git diff --stat
   

   If the diff is empty (no tracked files were modified), write this run report and STOP:

   echo '{"early_exit":"no_changes","prs_scanned":N,"prs_with_knowledge":M,"changes_proposed":0}' > artifacts/run-report.json
   

4. Save the diff for the review agent:

   git diff > artifacts/proposed-diff.txt
   

Phase 5: Review

1. Read $SKILL_DIR/prompts/review-agent.md.
2. Dispatch a foreground Agent with the prompt, using model opus.
3. After the agent completes, read artifacts/review.md and parse the verdict from the YAML frontmatter.

Phase 6: Revise (conditional)

If the review verdict is PASS, skip this phase entirely.

If the verdict is REVISE:

1. Read $SKILL_DIR/prompts/revise-agent.md.
2. Dispatch a foreground Agent with the prompt, using model opus.

Phase 7: Artifacts

1. Capture the final patch:

   git diff > artifacts/proposed-updates.patch
   

2. Reset the working tree:

   git checkout -- .
   

3. Write artifacts/run-report.json with these fields:

   • forge: the detected forge type
   • repo: the owner/repo slug
   • date_range: {"from": "YYYY-MM-DD", "to": "YYYY-MM-DD"}
   • prs_scanned: count of files in artifacts/pr-data/
   • prs_with_knowledge: count of extraction files with non-empty knowledge items
   • knowledge_items: {"high": N, "medium": N, "low": N} counts from extractions
   • changes_proposed: count of changes listed in artifacts/changes-summary.md
   • review_verdict: the verdict from artifacts/review.md
   • patch_file: "artifacts/proposed-updates.patch"

   Build these counts by reading the extraction and summary files.

4. Print: "Knowledge sync complete. Artifacts written to artifacts/"