upskill

Upskill Skill for ai_clients_dart

This skill automates the feedback loop from PR reviews to skill files. It extracts validated review findings (comments marked **Valid.** by the PR author), consolidates them into generalizable patterns, and updates the review checklists and implementation pattern files to prevent recurring issues.

Parse $ARGUMENTS for options:

• /upskill — Full apply mode (no flags). Runs the complete workflow (Steps 1–9), updating files and state after confirmation.
• --plan — Extract and analyze findings without editing any files. Safe to run anytime.
• --dry-run — Run the full apply workflow (extraction, consolidation, validation, proposed updates) but stop before actually editing files. Shows exactly what would change without changing anything.
• --from N — Start after PR number N instead of the config's last_checked_pr.

Step 1: Validate Environment

Perform all checks before proceeding. Fail fast with an actionable error message if any check fails.

1. CLI availability: git, gh, and jq must be on PATH.
2. GitHub auth: gh auth status must show authenticated.
3. Repository root: Use git rev-parse --show-toplevel to determine the repo root. All file paths are relative to this root.

Note: The Bash tool does not persist cd across calls. Prefix all shell commands with cd "$REPO_ROOT" && or use absolute paths.

1. Rate limit: Check remaining API calls:

   gh api rate_limit --jq '.rate.remaining'
   

   If remaining < 100, warn the user and stop — the extraction phase makes many API calls.

Step 2: Determine PR Range

1. Read .agents/skills/upskill/config/state.json:

   {
     "last_checked_pr": 122,
     "last_run_date": "2026-03-19"
   }
   

   Note: On the very first run, last_run_date will be null. In this case, rely solely on last_checked_pr to determine the PR range.

2. If the file does not exist or is unreadable, default last_checked_pr to 0 and warn: "No state file found — this is a first run, scanning all merged PRs."
3. If --from N was provided, override last_checked_pr with N.
4. Set START_PR to the effective starting PR number: use the --from N value if provided, otherwise last_checked_pr from state.
5. Record the starting PR number for the summary.