ir-exec - SKILL.md Agent Skill

name: ir:exec description: Investigate a GitHub issue and produce an implementation plan. Use at the start of a new task in a fresh worktree. Triggers on "/ir:exec", "plan issue", "plan this issue", "investigate issue", or when the user provides an issue number/URL and asks for a plan.

Goal: read the issue, understand the affected code, then propose an implementation plan and enter plan mode.

Issue numer is provided from the git branch and/or worktree name. If there is no branch or worktree ask the user for a ticket.

Confirm the worktree. Run pwd and git status -sb. If not in a clean worktree, point it out — don't block.
Fetch the issue (and its comments — they often contain the real spec):
```
gh issue view <N> --comments
```
For cross-repo: gh issue view <N> --repo <owner/repo> --comments.
Investigate the code. Delegate to one or more Explore subagents with thoroughness "medium" — one prompt that names the issue's key terms (file names, symbols, error strings, feature names) and asks: where does this live, what touches it, what's the current behavior. Don't grep manually first; the subagent handles it and protects the main context.
Synthesize a plan. Distill into:
- Problem — one sentence, what's broken or missing.
- Approach — the chosen direction in 2–4 bullets, naming files/functions you'll touch.
- Steps — ordered, concrete edits. Each step is one logical change.
- Risks / unknowns — anything you'd want the user to weigh in on before coding.
Present the plan for approval by calling ExitPlanMode. Do not start editing until approved.

Keep the plan tight. If it's longer than ~30 lines, you're over-planning — collapse steps.
If the issue is ambiguous or under-specified, surface that in Risks / unknowns rather than guessing silently.
Skip steps only when clearly redundant (e.g. user already pasted the issue body).