Agentpack Resolve

Use when the user invokes $agentpack-resolve.

Examples: $agentpack-resolve pr 123 or $agentpack-resolve PR #123 focus on backward compatibility.

The skill invocation owns the full loop; the internal --check and --reply commands are agent steps, not extra user commands.

1. Run agentpack resolve "$ARGUMENTS"; use --pr <number> when the PR is explicit.
2. Read .agentpack/resolve.prompt.md and .agentpack/resolve-comments.json completely.
3. Write every comment disposition to the declared plan TOON with exact source evidence citations.
4. Run agentpack resolve --check; do not edit code until the plan passes.
5. Apply all validated fixes in one pass, run targeted and relevant project checks, and write one cited reply record per comment.
6. Run agentpack resolve --reply; it refuses missing citations, invalid replies, and a changed PR head.
7. Start a fresh resolve pass after posting and repeat until no actionable unresolved comment remains or the iteration limit is reached.
8. Use concise inline replies: a short bold outcome, one explanation, exact path:line citations, Suggested fix: when relevant, and validation status. Do not mark a thread resolved merely because a reply was posted.

If a comment cannot be verified or fixed, record blocked with evidence and explain the blocker. Do not silently defer actionable comments.