Execute Plan

When NOT invoked via command (skill loaded programmatically or by another skill)

1. Load the plan from .cursor/plans/<file>.md (user path or @-reference) or from a pasted body if no file.
2. Confirm a ## User approval to execute section exists and does not contain PENDING — do not execute.
3. Confirm there is an Approved: yes line (or equivalent explicit approval recorded per reference.md).
4. If any check fails, do not spawn implementation tasks; instruct the user to finish Create Plan approval or fix the section.

Parse

• Tasks are sections matching heading ## Task <TaskID> (level-2 markdown heading).
• Read Dependencies inside each task; build a directed graph. Integration validation lives under ## Integration validation and runs only after all tasks PASS.

Execution waves

1. Compute ready set: tasks whose dependencies are all PASS or none.
2. For each wave, spawn one Task subagent per ready task in the same assistant turn (parallel calls).
3. Wait for all wave results before computing the next wave.
4. Do not start a task until its dependencies are PASS.

Subagent mandate (each task)

Give each subagent:

• Repo root path.
• The full markdown for that task (from ## Task ... through the end of that section).
• The Executor prompt from reference.md including max correction rounds = 10.
• readonly: false for normal execution (implement + run commands).
• Project context: If .cursor/project.md exists in the repo root, read it and include its full contents as the <PROJECT_CONTEXT> value in the executor prompt. If the file does not exist, pass the string 'None' for <PROJECT_CONTEXT>.
• Test writing requirement: For tasks that modify source code, if the project has an existing test directory or test configuration (e.g. package.json test script, *.sln, pytest.ini, go.mod with test files): write tests before implementing the feature (TDD). Place tests following the project's existing test conventions. Focus on happy paths; keep test complexity low. Write the minimum tests that verify the task's stated behavior. Tests must pass before the task is PASS. If no test structure exists, or if the task only modifies documentation, configuration, or skill/plan files, note "No tests applicable" in the result.

Plans may use ### Code only, ### Implementation (explicit_instructions) only, or both. The executor must follow ### Steps, ### Contracts, and whichever implementation sections exist—no improvising names, paths, or public surfaces that the plan does not specify.

Per-task outcome

• PASS — All validations in the task passed after fixes AND all newly-written tests pass; subagent returns evidence (commands + key output + test results).
• FAIL — After 10 correction rounds a validation still fails, OR newly-written tests still fail after 10 rounds; return logs, last commands, files touched, and which validation or test failed.

Per-wave code review

After all tasks in a wave return PASS, spawn a wave-code-reviewer subagent (readonly: true). Load the prompt from .cursor/skills/wave-code-reviewer/SKILL.md. Provide:

• The task markdown for every task in the wave.
• The combined list of files touched across all wave tasks.

The reviewer returns PASS or REVISE. If REVISE:

1. Spawn the wave executor(s) again with the reviewer's REQUIRED change requests appended to their task markdown.
2. Spawn the wave-code-reviewer again after the executor(s) complete.
3. Repeat up to 3 reviewer–executor rounds per wave. If still REVISE after 3 rounds, note the remaining issues and proceed to the next wave.

The wave is fully PASS only when the reviewer returns PASS (or the cap is reached).

Aggregation

• If any task FAIL, optionally spawn one recovery Task with a narrow brief (or escalate to user). Do not claim completion.
• When every task PASS (including wave code review), run Integration validation commands from the plan verbatim; same validate–fix loop applies (orchestrator may run fixes if integration spans tasks—prefer a dedicated subagent).
• Final step — test-plan: Once all tasks and integration validation PASS, spawn the test-plan subagent (readonly: false). Load the skill from .cursor/skills/test-plan/SKILL.md. Provide the plan file path. The subagent will discover and run the project's full test suite and assess scenario coverage against the plan's manual validation task.
• Summarize: task IDs, PASS/FAIL, wave code review rounds per wave, test-plan result, commands run, remaining risks.

Edge cases

See reference.md for overlap, ambiguous scopes, and parsing notes.

Additional resources

• reference.md — Executor prompt template, defaults, examples