Execute Task

Required references

Load these files before running:

• skills/shared/references/execution/task-file-contract.md
• skills/shared/references/execution/task-management.md
• skills/shared/references/review/review-protocol.md

Workflow

1. Resolve artifact files from trigger using task-file-contract.md.
2. Require PRD, TDD, and tasks-plan before doing any execution work.
3. Run kickoff branch setup from review-protocol.md Step 1 rules, carrying and committing the required planning artifacts on the new branch when they are the only uncommitted kickoff files.
4. Execute according to mode:
   • Standard mode: implement the requested task/sub-task in the main agent.
   • One-shot mode: treat the entire unchecked remainder of the task plan as the execution scope; if kickoff carried uncommitted planning artifacts, commit them before the first implementation sub-task, then for each sub-task in file order the main agent spawns one worker subagent, waits for completion, integrates the result, spawns one fresh review subagent for that review round with PRD, TDD, tasks-plan, the temp sub-task contract when it exists, and the exact review scope, applies findings, and owns task updates and commit before moving to the next sub-task.
5. For each completed sub-task:

• create/update tasks/tmp/plan-task-<task-id>.md with the sub-task contract before coding, then keep it current as implementation and review findings refine the slice
• identify repo-local implementation, test, and validation reference patterns before coding, record the chosen pattern in the sub-task contract, and justify any deliberate deviation
• write or update the targeted test first when the slice is practically testable
• run the targeted test command and confirm the intended failure before implementation begins
• implement the change using PRD + TDD + tasks-plan + exact sub-task block + the chosen local pattern
• rerun the targeted verification until the slice passes, then expand to any additional relevant repo-defined checks for the touched surface such as lint, format-check, typecheck, test, or build
• run one sub-task review round automatically using the active prompt profile from review-protocol.md, in a fresh review subagent when subagents are available; in one-shot mode, defer Prompt G to the final full-branch review
• apply fixes from review findings and rerun relevant tests
• keep temp review artifacts under tasks/tmp/ when --preserve-review-artifacts is enabled; otherwise clean them up per protocol
• mark checklist updates in tasks/tasks-plan-<plan-key>.md
• create a dedicated commit for the sub-task
• in one-shot mode, immediately re-open tasks/tasks-plan-<plan-key>.md after the commit, identify the next unchecked sub-task in file order, and continue directly when one exists

6. In one-shot mode, after all sub-tasks are complete, run one final full-branch review round automatically using the active prompt profile from review-protocol.md in one fresh review subagent before finalization. This is the review round that must satisfy Prompt G for frontend-facing work.
7. Run finalization from review-protocol.md Step 9 rules only after all unchecked sub-tasks in the task plan are complete, and do not treat one-shot execution as finished until the feature branch is pushed and the PR has been created.
8. Before any terminal handoff in one-shot mode, re-open tasks/tasks-plan-<plan-key>.md and confirm there are no remaining unchecked sub-tasks anywhere in the file. If any remain, continue execution instead of handing off unless a real blocker prevents further progress.
9. In one-shot mode, do not stop after an intermediate sub-task merely to report status, preserve a clean commit boundary, or hand off remaining work. Only stop early for a real blocker that cannot be resolved inside the current run.
10. In one-shot mode, do not emit user-visible mid-run progress updates while unchecked sub-tasks remain. Keep executing silently across sub-task boundaries instead of reporting intermediate completion, verification, or “continuing into X” status.
11. In one-shot mode, do not expose internal worker handoffs as user-visible content. Message shapes like “Completed and committed now: ...”, “Latest passing verifies: ...”, “I've already started X ...”, or “Remaining unchecked work is ...” are invalid while unchecked sub-tasks remain, even if they say execution is still in progress.
12. Assume any user-visible one-shot message before finalization can be treated as the run's terminal output. Only send a user-visible message early when a real blocker prevents continuation and the exact required user action is stated.

Mode behavior

• Standard mode: single-agent execution, pause for approval between sub-tasks.
• One-shot mode: sequential worker-subagent loop across the entire remaining unchecked task file, no approval pauses between sub-tasks, one main-agent integration plus fresh-review-subagent cycle per sub-task, defer frontend browser evidence to the final branch-wide review, then run one final full-branch review in a fresh review subagent before finalization. The run is terminal only after finalization or an explicit unresolved blocker.

Execution defaults

• Treat a failing-test-first red/green loop as the default for code-bearing, practically testable slices.
• A test-first exception is allowed only when a failing-first loop is not practical for that exact slice; record the exception reason in tasks/tmp/plan-task-<task-id>.md before implementation starts.
• Before coding, search the repo for similar implementations and tests, follow an existing local pattern when it is a good fit, and record why any new pattern or deviation is necessary.
• Before coding, inspect the repo's actual validation commands and config rather than assuming standard names exist.
• Do not invent lint, format, hook, or similar repo tooling mid-execution unless the planned task explicitly includes introducing that tooling or the user asks for it.

Review relationship

Task execution includes automatic review rounds. Manual review runs through explicit $review-chain activation, optionally scoped to a specific <task-id>.