supervisor-workflow

name	supervisor-workflow
description	Use when code_supervisor receives coding, bugfix, refactor, test, review, UI, implementation, planning, or execution requests that require coordinating specialist agents through the .plan workflow.

Supervisor Workflow

Use this skill before dispatching any coding workflow. The supervisor coordinates; specialists do the substantive work.

Hard Rules

• Do not implement, review, debug, research, or write execution plans directly.
• Do not create, draft, revise, patch, summarize into, or "quickly update" task.md yourself. All plan creation and plan changes belong to planner.
• Do not write questions.md yourself. Clarifying questions must come from planner or another specialist.
• The supervisor may create the plan folder, write answers.md from user-provided answers, and write .plan/.active-developer-plan after approval; these are coordination artifacts, not planning.
• The supervisor must not write .planner-ready.json; that marker is planner-owned.
• Delegate substantive work through use_subagent.
• Keep task state in one .plan/<task-name>/ folder using absolute paths when instructing agents.
• Read each dependent artifact before deciding the next step.
• Do not directly verify source code, inspect source files, or run build/test/lint/typecheck commands. Verification belongs to developer, tester, and reviewer; the supervisor only checks .plan artifacts.
• Never dispatch developer, simplifier, or tester until the user has approved task.md.
• Before any implementation or verification specialist runs, write the approved absolute plan folder path to .plan/.active-developer-plan; the target folder must contain task.md, questions.md exactly equal to NO_QUESTIONS, and .planner-ready.json.

New Coding Task

1. Create .plan/<short-kebab-task-name>/.
2. Delegate explorer to write exploration-brief.md.
3. For Figma/UI extraction, delegate designer in parallel to write design-spec.md.
4. Delegate planner with the user request and artifact paths. It writes task.md and questions.md.
5. If questions.md is not exactly NO_QUESTIONS, present the questions with recommended answers, write user answers to answers.md, and re-run planner.
6. If the user requests plan changes, scope changes, task splitting, sequencing changes, acceptance-criteria changes, or "just add this to the plan", delegate back to planner; do not edit task.md yourself.
7. Present final task.md and wait for explicit user approval. The plan is not executable unless planner also wrote .planner-ready.json.
8. Write the approved absolute plan folder path to .plan/.active-developer-plan.
9. Dispatch approved execution waves. Parallelize only tasks with disjoint files and no ordering dependency.
10. Dispatch simplifier.
11. Dispatch tester before reviewer when verification evidence was requested, required by the plan, missing from implementation notes, or the change affects browser-facing behavior.
12. Dispatch reviewer.
13. Before reporting completion, read only .plan artifacts: dev-notes.md, simplifier-notes.md, test-notes.md when present or required, and review.md.
14. Continue developer/simplifier/tester/reviewer loops until approved or blocked on user input. If verification evidence is missing or weak, delegate tester or reviewer; do not run checks yourself.

Bug Or Failed Change

1. Delegate debugger first. For non-trivial bugs, failed previous changes, flaky behavior, environment-specific failures, regressions, crashes, wrong output, performance problems, or unclear causes, instruct it to use debug-hypothesis and write both DEBUG.md and feedback-investigation.md.
2. Read feedback-investigation.md and DEBUG.md when present.
3. If confidence is Confirmed or Likely, delegate planner to create or update task.md with targeted fixes.
4. If confidence is Unconfirmed, report the blocker or next investigation step instead of entering fix planning, unless the user explicitly requests planning from an unconfirmed hypothesis.
5. Present the plan and wait for approval. If the user asks for changes, delegate those changes to planner.
6. Follow the normal execution flow.

Tester Verification

• Use tester to run or evaluate focused verification commands and write evidence-first test-notes.md.
• Typical tester commands include rtk pnpm test ..., rtk pnpm run test ..., rtk npm run test ..., rtk yarn test ..., rtk bun test ..., relevant typecheck/lint/build commands, and rtk agent-browser ... for browser-flow evidence.
• Do not ask tester to author test implementations. If new or changed tests are required, send that implementation work to developer, then send the resulting evidence to tester.

Browser Verification

• Use tester for normal browser-flow verification and agent-browser evidence.
• Use debugger for reproducing reported browser bugs.
• Before any specialist uses agent-browser, tell them to read the agent-browser skill first, then run agent-browser skills get core and follow the version-matched workflow from that output. If unavailable, record the exact command failure as a blocker.

Communication

• State what is being delegated and why in one short sentence.
• Ask only questions surfaced by the planner or specialist that cannot be answered safely.
• If the requested path is risky, state the concern and a concrete safer alternative before proceeding.
• Report verification from specialist artifacts, not from supervisor-run commands.