Implement Issues

Implement Issues

Turn a set of GitHub issues into open PRs via parallel, worktree-isolated lanes. One lane per issue: its own worktree, its own branch, its own PR. This skill orchestrates and composes branch, commit, and pr — it never merges. Hand-off to babysit-pr / babysit-pr --merge for shepherding and merge.

Invocation

• /implement-issues #12 #14 #17 — implement exactly these issues.
• /implement-issues (bare) — triage the backlog (gh issue list), prefer self-contained, file-disjoint issues, propose a set and confirm before spawning any lane.
• --sequential — run one lane at a time (still one worktree/branch/PR each).

Inputs & readiness

• Load repo context first. At start, from the target repo checkout root, read AGENTS.md and/or CLAUDE.md if present. Precedence: AGENTS.md is authoritative when both exist — do not also apply conflicting CLAUDE.md instructions; if only one exists, use that file; if both conflict in a blocking way, stop and ask. Record which filename is authoritative (e.g. AGENTS.md or CLAUDE.md) and treat its house standards, operating agreement, and standing constraints as binding for every lane. Pass that chosen filename into each lane prompt so lanes do not re-pick. See the stand-general skill's Per-repo agent context section for the expected file shape. Missing the file is fine — fall back to chat instructions and stand-* skills.
• Take issues from the args, or select from gh issue list favouring self-contained and file-disjoint work.
• Every issue must carry an AI Implementation Prompt comment (per the issue skill). Missing it means the issue is not fan-out-ready — write one first, or run that issue as a single lane yourself. Do not hand a bare issue to a sub-agent.
• Reproduce against current main before implementing — specs can be stale. If the reported problem no longer exists, report back rather than fabricating a change.

Conflict grouping

• Build a file-overlap matrix from each issue's declared file targets.
• Disjoint file sets → run in parallel lanes.
• Overlapping file sets → sequence them, or bundle into a single lane only when the issues genuinely must share files.
• Cap at ~10 lanes — beyond that, batch across sessions.

Lane claiming

Before spawning, claim each lane — assign the issue to yourself and/or add a working label. This prevents a second session from opening a duplicate PR for the same issue (a race that has bitten this workflow before).

Per lane

1. Orchestrator creates the worktree per the branch -w convention: a sibling directory, branch type/NN-slug cut from a freshly fetched origin/main. Never place a worktree under .claude/ — lint configs commonly exclude that path and will silently scan nothing.
2. Sub-agent prompt is pointers, not content. Instruct the lane to:
   • Work only inside the lane worktree path the orchestrator created.
   • Read issue #N and its AI Implementation Prompt comment.
   • From the lane worktree root (absolute path), read the authoritative context file the parent already chose (AGENTS.md or CLAUDE.md) when present. Do not resolve context from the parent checkout or any other directory; do not re-pick between the two files. Treat its standards/constraints/contract as binding; follow applicable stand-* skills.
   • Run the full lint and full test suite before every commit — never a subset.
   • Use the commit skill, then the pr skill with Closes #N.
   • Never merge. Report the PR URL on completion.
3. Orchestrator verifies each PR exists once the lane reports done.

Guardrails

• Signing pre-flight before spawning. Confirm commit signing works now — a locked signing key must be surfaced to the user before you go AFK, not discovered mid-run.
• A lane that cannot stay green STOPS and reports. Never weaken or delete tests, never open a knowingly broken PR.
• Final report lists every PR URL and every blocked lane with the reason.

Hand-off

Shepherding and merging are not this skill. Once PRs are open, hand off to babysit-pr (drive to merge-ready) or babysit-pr --merge.

See also backlog: the interactive routing entry point that composes this skill with babysit-pr. It does not auto-invoke either.

Harness mechanics (Claude Code)

• Run lanes as background sub-agents so the main thread stays free.
• Model per lane by risk:
  • docs / config / mechanical → sonnet
  • logic / refactor / security / auth / CI-touching → opus
  • never haiku for code.
  • An explicit user instruction overrides these defaults.
• Elsewhere (harnesses without sub-agent spawning): run the lanes sequentially using the same worktree-per-lane layout.