/issue-pipeline: End-to-End Issue Pipeline

You orchestrate the full lifecycle of GitHub issues through parallel sub-agents. Each phase runs all issues concurrently in isolated worktrees.

Arguments

Parse $ARGUMENTS to get issue numbers and flags.

Examples:

• /issue-pipeline 33 34 35 — full pipeline (plan → implement → review → merge)
• /issue-pipeline 33 34 --skip-plan — issues already have plans, skip to implement
• /issue-pipeline 33 34 --skip-review --no-merge — implement only, no review or merge
• /issue-pipeline 33 — single issue, full pipeline

If no issue numbers provided, ask: "Which issues? Provide issue numbers separated by spaces."

Flags:

• --skip-plan — Skip plan phase (issues already have planned label + plan comments)
• --skip-review — Skip review phase, merge after implement
• --no-merge — Stop after review, don't auto-merge
• --dry-run — Show what would happen without executing

Pre-Flight Check

Before starting, verify all issues exist and check their state:

for num in {issue_numbers}; do
  gh issue view $num --json number,title,labels,state --jq '"\(.number) [\(.labels | map(.name) | join(","))] \(.state) — \(.title)"'
done

Report findings:

• Issues with planned label → can skip plan phase
• Closed issues → skip entirely with warning
• Issues without planned label → need plan phase

Pipeline Phases

Phase 1: Plan (parallel)

Skip if: --skip-plan flag OR issue already has planned label.

For each issue needing a plan, launch a sub-agent:

Agent tool parameters:
  description: "Plan issue #{number}"
  isolation: "worktree"
  run_in_background: true
  prompt: |
    You are working on GitHub issue #{number} for gstack-game at {repo_root}.

    Execute the /issue-plan workflow:
    1. Read the skill at {repo_root}/.claude/skills/issue-plan/SKILL.md
    2. Follow its instructions for issue #{number}

    This means: Research → Innovate → Plan, posting all 3 phases to the issue.

    IMPORTANT:
    - You are in a worktree — do NOT modify source code, planning only
    - Post all findings to the GitHub issue as comments
    - Add the "planned" label when done

Wait for ALL plan agents to complete. Report status table.

Phase 2: Implement (parallel)

For each issue with a plan, launch a sub-agent:

Agent tool parameters:
  description: "Implement issue #{number}"
  isolation: "worktree"
  run_in_background: true
  prompt: |
    You are implementing GitHub issue #{number} for gstack-game at {repo_root}.

    Execute the /issue-action workflow:
    1. Read the plan from the issue comments: `gh issue view {number} --json comments`
    2. Also check for local artifacts: `.tmp/deep-dive/issue-{number}/plan.md`
    3. Read the skill at {repo_root}/.claude/skills/issue-action/SKILL.md
    4. Follow its instructions for issue #{number}

    IMPORTANT:
    - You are in a worktree — create a feature branch, make changes, commit
    - Edit .tmpl files, never generated .md directly
    - Run `bun run build` after template changes
    - Run `bun test` — all tests must pass before creating PR
    - Create a PR with `gh pr create` — include "Closes #{number}" in the body
    - Do NOT merge the PR

Wait for ALL implement agents to complete. Collect PR URLs/numbers. Report status table.

Phase 3: Review (parallel)

Skip if: --skip-review flag.

For each PR created in Phase 2, launch a sub-agent:

Agent tool parameters:
  description: "Review PR #{pr_number}"
  isolation: "worktree"
  run_in_background: true
  prompt: |
    You are reviewing PR #{pr_number} for gstack-game at {repo_root}.

    Execute the /pr-review-loop workflow:
    1. Read the skill at {repo_root}/.claude/skills/pr-review-loop/SKILL.md
    2. Follow its instructions for PR #{pr_number}

    IMPORTANT:
    - You are in a worktree
    - If fixes are needed, check out the PR branch, fix, commit, and push
    - The review loop continues until LGTM or max 3 iterations

Wait for ALL review agents to complete. Report verdicts.

Phase 4: Merge

Skip if: --no-merge flag.

For each PR with LGTM verdict:

gh pr merge {pr_number} --merge --delete-branch

For PRs without LGTM: report them as needing manual review.

After all merges, sync main:

git checkout main && git pull

Report final status table.

Execution Rules

1. All sub-agents use isolation: "worktree" — never work on main directly
2. All sub-agents use run_in_background: true — maximize parallelism
3. Launch ALL agents for a phase in a SINGLE message (one Agent tool call per issue, all in parallel)
4. Wait for ALL agents in a phase before starting the next phase
5. Report a status table after each phase — issue number, status, key details
6. If any agent fails, report it but continue with the rest — don't block the pipeline
7. Between phases, sync main — git checkout main && git pull to pick up merged changes

gstack-game Specific Rules

• Template workflow: Agents edit .tmpl files, run bun run build to regenerate .md, run bun test (14 tests)
• Preamble changes: If an issue changes skills/shared/preamble-*.md, ALL 28 SKILL.md files are regenerated
• Commit style: Conventional Commits (feat: / fix: / test: / docs:)
• PR merges use --merge (not squash) to preserve commit history

Status Table Format

After each phase:

| Issue | Phase     | Status | Details |
|-------|-----------|--------|---------|
| #33   | plan      | ✅     | Plan posted to issue |
| #34   | plan      | ⏭️     | Already planned |
| #35   | implement | ✅     | PR #38 created |
| #36   | implement | ❌     | bun test failed |

After merge:

| Issue | PR  | Plan | Implement | Review | Merge | Result |
|-------|-----|------|-----------|--------|-------|--------|
| #33   | #38 | ✅   | ✅        | ✅     | ✅    | DONE   |
| #34   | #39 | ⏭️   | ✅        | ✅     | ✅    | DONE   |
| #35   | #40 | ✅   | ✅        | ⚠️     | ❌    | NEEDS REVIEW |

Error Handling

• Agent fails: Mark as ❌, continue with other issues. Report at end.
• PR creation fails: Try to identify branch, suggest manual fix.
• Review finds issues: pr-review-loop handles fix + re-review automatically (max 3 rounds).
• Merge conflict: Report conflict, skip merge for that PR, suggest manual resolution.
• bun test fails: Agent should NOT create PR. Report test failures.

Anti-Sycophancy

• ❌ "All issues completed successfully!" (unless they actually all did)
• ❌ "Easy batch!" (some issues may have unexpected complexity)
• ✅ Report exact counts: "4/5 issues completed. #36 failed at implement phase: bun test returned 2 failures."
• ✅ If all agents fail, say so plainly: "0/3 issues completed. All agents encountered errors."