issue-flow（问题流程）

Issue Flow — AI-Native Issue-Driven Development

You are orchestrating a complete development cycle from a GitHub Issue to a merged PR. Follow the phases below strictly. Every major decision requires human confirmation.

Announce at start: "I'm using the issue-flow skill to implement this GitHub Issue."

Initial request: $ARGUMENTS

Phase 0: Preflight

Goal: Validate environment and resolve the target Issue.

Actions:

1. Environment check: Run gh auth status to verify GitHub CLI is authenticated. If it fails, tell the user to run gh auth login and stop.
2. Parse arguments:

- #123 or just 123 → extract issue number, detect repo from gh repo view --json nameWithOwner - https://github.com/org/repo/issues/123 → extract owner, repo, and number - Empty → run gh issue list --state open --limit 20 and use AskUserQuestion to let the user pick an issue

1. Fetch Issue details: gh issue view <N> --json number,title,body,labels,comments,assignees,state Store: ISSUE_NUMBER, ISSUE_TITLE, ISSUE_BODY, ISSUE_LABELS.
2. Check for existing work:

- Search for branch issue/<N>-* via git branch -a --list '*issue/<N>*' - If found, AskUserQuestion: resume existing branch / start fresh / cancel

1. Detect project features (used later for team composition):

- Has test framework? (check for jest.config*, vitest.config*, pytest.ini, *test* dirs) - Has CI? (check .github/workflows/, .gitlab-ci.yml, etc.) - Has linter? (check .eslintrc*, biome.json, .prettierrc*) - Primary language(s) from file extensions

Phase 1: Worktree Setup

Goal: Create an isolated workspace.

Actions:

1. Use the EnterWorktree tool with name issue-<N> to create an isolated worktree.
2. Create and switch to branch issue/<N>-<slugified-title>:

- Slugify: lowercase, replace spaces/special chars with -, truncate to 50 chars - Example: Issue #42 "Add OAuth2 Login Support" → issue/42-add-oauth2-login-support - Run: git checkout -b issue/<N>-<slug>

Phase 2: Technical Planning

Goal: Deep codebase exploration → technical plan → user approval.

Step 2a: Codebase Exploration

Launch 2-3 code-explorer agents in parallel using the Task tool (subagent_type: code-explorer). Tailor each agent's focus to the Issue:

• Agent 1: Explore existing implementations and patterns directly related to the Issue's requirements. Return a list of 5-10 key files.
• Agent 2: Analyze architecture, dependencies, and extension points of affected modules. Return a list of 5-10 key files.
• Agent 3 (optional, for complex issues): Investigate test patterns, CI configuration, and related toolchain. Return a list of 5-10 key files.

After agents return, read all key files they identified to build deep understanding.

Step 2b: Design Technical Plan

Based on the exploration, design a technical plan with:

1. Summary: 2-3 sentence overview of the approach
2. Files to modify/create: List with brief description of changes
3. Implementation steps: Ordered, concrete steps (each step = one logical unit of work)
4. Test strategy: What to test, how to test
5. Risk assessment: Potential issues and mitigations

Format the plan according to rules/plan-format.md.

Step 2c: Publish Plan to Issue

Post the plan as a comment on the Issue using gh issue comment:

gh issue comment <N> --body "$(cat <<'EOF'
<plan content formatted per rules/plan-format.md>
EOF
)"

Use the HTML comment marker <!-- issue-flow-plan --> at the top so the plan can be identified and updated idempotently.

Step 2d: User Confirmation

Use AskUserQuestion with options:

• Approve — proceed with implementation
• Modify — user provides feedback, return to Step 2b
• Cancel — abort the workflow

Do NOT proceed without explicit approval.

Phase 3: Team Execution

Goal: Implement the plan — either directly or with an agent team.

Decision: Direct vs Team

Direct implementation (no team) when ALL of these are true:

• Plan involves 1-2 files
• No complex cross-module changes
• No separate test/review/docs work needed

Team execution when ANY of these are true:

• Plan involves 3+ files across multiple modules
• Tests need to be written or updated
• Security-sensitive changes
• Documentation updates required
• Frontend + backend changes together

Path A: Direct Implementation

1. Implement changes following the plan step by step
2. Run existing tests if available
3. Skip to Phase 4

Path B: Team Execution

1. Create team: Use TeamCreate with name issue-<N>.
2. Decide team composition: Based on Issue characteristics, select roles from the candidate pool defined in rules/team-roles.md. Consider:

- Issue labels (e.g., frontend, security, docs) - File types in the plan (.tsx → frontend, .sql → backend, etc.) - Whether tests exist and need updating - Risk level of the changes

1. Spawn teammates: Use the Task tool with team_name parameter for each role. Give each teammate:

- The technical plan - Their specific tasks - Context about the codebase patterns discovered in Phase 2

1. Create tasks: Use TaskCreate for each implementation step from the plan. Set up dependencies with addBlockedBy where steps depend on each other.
2. Assign and coordinate: Assign tasks to teammates via TaskUpdate. Monitor progress, resolve blockers, and coordinate between teammates.
3. Iteration limit: Each task gets at most 2 fix iterations. If a task still fails after 2 rounds:

- Log the issue - AskUserQuestion: fix manually / skip / abort

1. Shutdown team: After all tasks complete, send shutdown_request to each teammate, then TeamDelete.

Phase 4: PR & CI

Goal: Commit, push, create PR, handle CI.

Step 4a: Commit & Push

1. Stage all changes: review with git status and git diff
2. Commit with a descriptive message referencing the Issue: feat: <summary from plan> (#<N>)
3. Push the branch: git push -u origin issue/<N>-<slug>

Step 4b: Create PR

Create PR using gh pr create with the template from rules/pr-template.md:

gh pr create --title "<title>" --body "$(cat <<'EOF'
<PR body per rules/pr-template.md, includes Closes #N>
EOF
)"

Step 4c: Post Implementation Summary to Issue

Add a comment to the Issue summarizing what was implemented:

gh issue comment <N> --body "$(cat <<'EOF'
<!-- issue-flow-impl -->
## Implementation Complete

- PR: #<PR_NUMBER>
- <brief summary of changes>
EOF
)"

Step 4d: CI Check

1. Wait briefly, then check CI status: gh pr checks <PR_NUMBER> --watch --fail-fast
2. If CI fails: AskUserQuestion with options:

- Auto-fix — attempt to diagnose and fix CI failures (max 2 iterations) - Manual — user will fix manually - Abort — close the PR

1. If CI passes (or no CI configured): AskUserQuestion with options:

- Merge — merge the PR now - Keep open — leave PR open for human review - Request review — assign reviewer via gh pr edit --add-reviewer

Phase 5: Cleanup

Goal: Clean up resources after merge.

Actions (only if PR was merged):

1. Check Issue status: Verify if Closes #N auto-closed the Issue. If not: gh issue close <N> --comment "Closed via PR #<PR_NUMBER>"
2. Report: Output a completion summary: ## Issue Flow Complete - Issue: #<N> <title> - Branch: issue/<N>-<slug> - PR: #<PR_NUMBER> (merged) - Files changed: <count> - Implementation: <1-2 sentence summary>

Error Recovery

If the workflow is interrupted at any point, the user can re-run /issue-flow #<N> to resume. See references/recovery-guide.md for detailed recovery scenarios.

Key Principles

• Human-in-the-loop: Every major decision (plan approval, merge, CI failure handling) requires explicit user confirmation
• Idempotent comments: Issue comments use HTML markers (<!-- issue-flow-plan -->, <!-- issue-flow-impl -->) so re-runs update rather than duplicate
• Isolation: All work happens in a worktree — the main branch is never touched until merge
• Proportional response: Simple changes skip the team overhead; complex changes get full team coordination