feature - SKILL.md Agent Skill

name: feature description: Start a new feature in an isolated git worktree with TDD workflow when_to_use: when the user says "start a feature", "new feature", invokes "/agent-dashboard:feature", or describes work that needs an isolated branch + worktree + TDD loop. NOT for hotfixes, single-file edits, pure exploration, or non-code changes (use /agent-dashboard:chore, /agent-dashboard:fix, /agent-dashboard:investigate instead). version: 1.0.0 disable-model-invocation: true effort: max

Start a new feature in an isolated git worktree.

Feature description: $ARGUMENTS

Instructions

Follow these phases in order. Each phase has a gate — do not proceed until the gate is satisfied.

If the feature touches browser UI, Playwright, dev-server ports, screenshots, or interactive Browser/Chrome inspection, apply ../_shared/ui-automation.md at planning, environment setup, verification, delegation, and cleanup points.

Phase 1: Setup

Derive a short kebab-case name from the description
Derive the app name from the git repo: basename $(git rev-parse --show-toplevel)
Switch to main: git checkout main
Pull latest: git pull origin main
Create branch feat/<name> and worktree ../worktrees/<app>/<name> from main: mkdir -p ../worktrees/<app> && git worktree add ../worktrees/<app>/<name> -b feat/<name> main
- If the branch already exists, ask the user whether to resume it or choose a new name.
From the source repo root (before cd'ing), copy environment files into the worktree preserving their exact relative path from the project root:
- Find all env files recursively: find . -name '.env*' -not -path './.git/*' -not -path './node_modules/*'
- For each file found, recreate its directory structure in the worktree and copy it. For example:
  - ./. env → ../worktrees/<app>/<name>/.env
  - ./services/api/.env.local → ../worktrees/<app>/<name>/services/api/.env.local
  - ./infra/.env.production → ../worktrees/<app>/<name>/infra/.env.production
- Use: for f in $(find . -name '.env*' -not -path './.git/*' -not -path './node_modules/*'); do mkdir -p "../worktrees/<app>/<name>/$(dirname "$f")" && cp "$f" "../worktrees/<app>/<name>/$f"; done
- If .claude/settings.local.json exists: mkdir -p ../worktrees/<app>/<name>/.claude && cp .claude/settings.local.json ../worktrees/<app>/<name>/.claude/
- Important: All Bash tool calls in this step must set dangerouslyDisableSandbox: true because they write outside the project root.
cd into the worktree, run node "$CLAUDE_PLUGIN_ROOT/scripts/hooks/claim-worktree.js", and confirm with pwd and git branch --show-current
Verify: compare env files between source and worktree. Run the same find command in both directories and diff the file lists. If any files are missing in the worktree, halt and report failure. If the source repo had no .env* files, note that explicitly.

Gate: Working directory is the new worktree on the correct branch, based on latest main. If .env* files existed in the source repo, they are all present in the worktree.

Phase 2: Plan

Start two tracks in parallel:

Background — Environment setup: Launch a background agent (run_in_background: true) to set up the dev environment. The agent must:

Auto-detect project type from project files (highest match wins):

Priority	Signal	Type
1	`react-native` in package.json dependencies	Mobile
2	`next`, `vite`, or `webpack` in package.json	Web
3	`requirements.txt`, `pyproject.toml`, or `setup.py`	Python
4	`go.mod`	Go
5	`Dockerfile` or `docker-compose.yml`	Containerized

Ask the user only if no signal matches.

Install dependencies appropriate for the project type (e.g. pip install, npm install, go mod download). Configure ports, create emulators/simulators as needed. For browser UI work, allocate worktree-local Playwright/server/profile/output resources per ../_shared/ui-automation.md.
Symlink large source-content directories (data/, datasets/, evals/, models/, artifacts/) from the source repo rather than copying. NEVER symlink build outputs or per-project caches (.next/, dist/, build/, out/, target/, .turbo/, .cache/, .parcel-cache/, .vite/, __pycache__/, .pytest_cache/, .gradle/, .venv/, node_modules/) — they bake absolute paths and corrupt across worktrees, and must be regenerated per-worktree.
On success, write a sentinel file: touch .env-setup-done On failure, write the error: echo "<error message>" > .env-setup-failed

Foreground — Planning:

Phase order: research first, interview second, plan mode third, submit fourth. Plan mode is the last gate before approval, not a pre-research speed-bump. Each step has a HARD-GATE you cannot rationalize past.

Research with Explore. Use the built-in Explore subagent for any non-trivial codebase question or library lookup. Do not call Agent with subagent_type=Plan — composing the plan is your job, not a delegated subagent's. Synthesize what you found inline as your own assistant text.

Why: a delegated Plan subagent returns the plan inside a tool_result block that lives in the JSONL but never reaches the dashboard's plan panel, conversation view, or activity log. Synthesizing inline puts the plan in your assistant text, which the dashboard renders everywhere.

Symptoms you're about to violate:
- You're about to call Agent with subagent_type: "Plan".
- You're rationalizing "the planner does this better."
Research subagent (`Explore`): allowed and encouraged. Planning subagent (`Plan`): forbidden in this skill. Compose the plan yourself.
Do not wait for environment setup to finish.
Interview the user via AskUserQuestion. Identify every gating decision the implementation depends on — URLs, IDs, scope boundaries, copy text, what to delete vs keep, version pins, credentials. Ask them as a single AskUserQuestion call with multi-choice options, not as freeform numbered text in your assistant message.

Load via ToolSearch if AskUserQuestion isn't in scope: ToolSearch("select:AskUserQuestion").

Schema: 1–4 questions per call, each with a header (≤12 chars), 2–4 mutually exclusive options, multiSelect: false unless the choices genuinely combine. Recommended option goes first with (Recommended) suffix. The user always gets an "Other" escape hatch automatically.

Worked example:
```
AskUserQuestion({
  questions: [{
    question: "Where should focus.json live?",
    header: "Focus path",
    multiSelect: false,
    options: [
      {label: "~/.agent-dashboard/focus.json (Recommended)", description: "Co-located with agents/ state dir already watched."},
      {label: "$XDG_RUNTIME_DIR/agent-dashboard/focus.json", description: "Tmpfs-backed, cleared on reboot."},
    ],
  }]
})
```
The plan you submit in step 4 must be implementable as written. No "Decisions needed", "Phase 0", "TBD", "?", or "to be confirmed" sections in the body. If a user answer changes scope, return to this step and re-interview.

Symptoms you're about to violate:
- You're typing "1." "2." "3." numbered questions in assistant text.
- You're writing "Decisions needed before implementation" inside the plan.
- You're rationalizing "the user can answer these after approval."
Freeform numbered questions in assistant text are a violation. If you find yourself typing "1." "2." "3." to ask the user something — STOP, call `AskUserQuestion` instead. The plan is not ready for review until every decision it gates is answered.
Enter plan mode via EnterPlanMode, then draft the plan inline. Now that research is done and decisions are resolved, call EnterPlanMode (load via ToolSearch if not in scope). This flips the parent's permission_mode='plan' and restricts you to read-only tools while you write the plan as your own assistant text.

Why this order: drafting inside plan mode pairs the visible mode-flip with the actual planning work, and EnterPlanMode is a load-bearing prerequisite for ExitPlanMode (step 4) — which is the only path to user approval.

Caveat: on approval (step 4), CC drops to its default permission_mode, not back to bypassPermissions. Subsequent edits in Phase 3 will re-prompt unless the user re-enables bypass. Accepted trade-off — visible planning is worth the one-time mode reset.
No drafting the plan in assistant text until `EnterPlanMode` has been called and `permission_mode='plan'` is active.
Phase format for multi-phase plans. If the plan has 3+ distinct work units, structure it with a ## Phases checklist and matching ### Phase X: headings. Step 4 reads this format to offer the dispatch probe; /agent-dashboard:implement parses it to drive the dispatch loop. Plans without it can't be dispatched.
```
## Phases

- [ ] **Phase A: <short name>** — files: <globs>, deps: -
- [ ] **Phase B: <short name>** — files: <globs>, deps: A
- [ ] **Phase C: <short name>** — files: <globs>, deps: B

### Phase A: <short name>

<10–50 lines: what files, what tests, what invariants, what to leave alone.>

### Phase B: <short name>

<...>
```
Rules:
- The ## Phases block is the dispatch index. Phase names MUST match between checklist and ### Phase X: headings (case-sensitive).
- deps: defaults to "depends on previous phase". Use - for "no dependencies".
- - [ ] = pending. - [x] = done. /agent-dashboard:implement flips these as it dispatches.
- Fewer than 3 work units? Skip this format. Inline paragraphs are fine; the probe won't fire below the threshold.
Submit via ExitPlanMode. Wait for user approval. Pass the full plan markdown to the plan file (per CC's plan-mode workflow) and call ExitPlanMode. This renders the plan in CC's native plan-review UI for accept/reject.

ExitPlanMode is the only acceptable submission. Pasting the plan as assistant text is a violation, even if you also call ExitPlanMode afterwards. The user reviews and approves through the plan-review UI — nowhere else.

No exceptions:
- Don't start a "small" preparatory edit while waiting.
- Don't write the test file "to save time".
- Don't ask "should I proceed?" in assistant text — the plan-mode UI's accept action is the approval.
Post-approval actions (immediately after the user accepts the plan, before Phase 3):
1. Write the plan-path sentinel (always). CC's plan-mode system prompt told you where the approved plan markdown lives (typically ~/.claude/plans/<slug>.md). Record that path so /agent-dashboard:implement can find it:
```
echo "<absolute-plan-path>" > .feature-plan-path
```
2. Count the phases. Read the plan; count - [ ] / - [x] lines under ## Phases. If there's no ## Phases block or the count is < 3, skip step 3 below and start Phase 3 (inline TDD).
3. Probe for dispatch handoff (only when phase count ≥ 3). Call AskUserQuestion exactly once:
  - Question: "Plan has {N} phases. Continue inline here, or hand off to /agent-dashboard:implement for context isolation?"
  - Header: "Dispatch"
  - Options (recommended first):
    - "Continue inline (Recommended for ≤4 phases)" — Stay in this session; run RED → GREEN → REFACTOR per phase in order.
    - "Hand off to /agent-dashboard:implement" — Exit /agent-dashboard:feature. The user invokes /agent-dashboard:implement in a fresh session; each phase dispatches to its own subagent.
  If Continue inline: start Phase 3. The ## Phases structure becomes documentation — inline TDD ignores the index.
  
  If Hand off to /agent-dashboard:implement: print the message below and exit cleanly. Do not start Phase 3.
```
Plan saved to <plan-path>.
Worktree ready at <worktree-path>.
To continue, run:

    /agent-dashboard:implement

(Recommended: open a fresh terminal session for max context isolation.)
```

Gate: Plan approved with no open decisions. .feature-plan-path written. Either Phase 3 begins (inline) or the skill exited with the handoff message (dispatch).

Phase 3: Implement

Pre-gate: Check for .env-setup-done in the worktree root.

If present: verify dependencies are installed (e.g. node_modules/ exists, pip list succeeds, go env GOPATH works) and data symlinks resolve correctly.
If .env-setup-failed exists: surface the error and halt.
If neither file exists: the background agent is still running — wait for it to finish before proceeding.

Effort note: When launched via the agent-dashboard's New Agent flow, this skill spawns with --effort high on the CLI, which Claude Code pins at the session level. The dynamic dispatcher in agent-state-fast.js bumps effort to max automatically while permission_mode='plan' (EnterPlanMode active) and drops back to high on exit — so planning runs at max effort without paying that cost during implementation. When invoked as a slash command inside an existing claude session, you can run /effort max before entering plan mode and /effort high (or lower) before implementation.

Delegation gate: Invoke /codex:setup to check Codex CLI availability. If the output contains "ready": true, delegate only if the user explicitly requested Codex delegation OR the plan touches 10+ files / ~3,000+ lines of implementation. Below that threshold, the orchestration overhead (skill loading, prompt construction, subagent context, result parsing, review) costs more tokens than Claude implementing directly. If delegating, invoke /codex-delegate with the approved plan (Phase 2) as implementation context, then skip to the phase gate. Otherwise, proceed below.

Build the feature following strict RED → GREEN → REFACTOR:

RED. Write the failing test. Run make test. Paste the failing output into the conversation. Wrote implementation before test? Delete it. Start over. No exceptions:
- Don't keep it as "reference"
- Don't write tests for the implementation you already wrote
- Don't claim "the test would obviously fail" — show it failing
GREEN. Write the minimum implementation to make the failing test pass. Run make test. Paste the passing output. Wrote more than the test demanded? Revert and re-do. No "while I'm here" additions, no premature abstractions.
REFACTOR. Clean up. Run make test after each meaningful edit. Tests broke during refactor? Revert that edit and try a smaller step. Refactor is structure-only — if behavior changed, you're back in RED.

For UI verification, prefer headless Playwright with worktree-local resources. Use interactive Browser/Chrome inspection only when the shared policy says it is warranted.

Gate: Environment ready. All tests pass via make test. Implementation matches the approved plan.

Phase 4: Review

Review all changes for correctness, security, and convention adherence. Apply all project rules and conventions that are in your context.

Gate: No critical or high-severity issues remain.

Phase 5: Deliver

Commit the feature changes with a feat: conventional commit message.
Open the PR by invoking /agent-dashboard:pr. That skill owns the cleanup pass (refactor-cleaner), make fmt, make test, push, and gh pr create. Do not call gh pr create directly — a pr-skill-gate hook will block it.

Gate: Clean commit history with conventional commit messages. PR opened via /agent-dashboard:pr.

Phase 6: Cleanup (on merge)

Triggered when the user indicates the feature has been merged upstream.

Verify the branch is merged (warn if unmerged commits remain)
Tear down environment resources: remove symlinks, stop dev servers or emulators, release any browser lease, remove worktree-local UI scratch state, delete .env-setup-done / .env-setup-failed / .feature-plan-path sentinel files
Remove worktree and delete branch
Confirm cleanup is complete

Red Flags — STOP

If you catch yourself saying or thinking any of these, pause and re-read the relevant phase:

"I'll just sketch the implementation first" → Phase 3 RED violation. Delete and restart.
"I'll delegate the plan to a Plan subagent" → Phase 2 step 1 violation. Research with Explore; plan inline. The dashboard can't surface delegated plans.
"I'll just type the questions as numbered text" → Phase 2 step 2 violation. AskUserQuestion exists for exactly this. Load via ToolSearch and call it.
"I'll skip EnterPlanMode, plan mode resets bypassPermissions" → Phase 2 step 3 violation. After research and the AskUserQuestion interview, you call EnterPlanMode to draft the plan inside plan mode, then ExitPlanMode to submit. The reset to default permission_mode is the accepted cost.
"I'll just paste the plan as text instead of calling ExitPlanMode" → Phase 2 step 4 violation. ExitPlanMode is the only acceptable submission. Pasting in assistant text is not a fallback.
"The plan is obvious, let me start" → Phase 2 gate violation. Wait for approval.
"Tests pass on my reading of the code" → didn't run make test. Run it.
"I'll skip the worktree, it's a small change" → wrong skill. Use a feature branch directly without invoking this skill.
"Let me commit on main since the change is trivial" → blocked by hook anyway. Create a branch.
"I'll just call gh pr create directly" → Phase 5 violation. The pr-skill-gate hook will block it. Use /agent-dashboard:pr.
"I'll bundle this unrelated cleanup into the feature commit" → split it. Open a separate PR.
"User picked hand-off, but I'm already here — I'll just do Phase 3 myself" → exit cleanly. They opted out of inline TDD for a reason (context). Don't second-guess.
"I'll write .feature-plan-path later, after I start Phase 3" → write it now. /agent-dashboard:implement and resume can't find the plan without it.