execute

/orc:execute — Execute Tasks

Input

$ARGUMENTS

Precondition

Tasks MUST already exist (created via /orc:tasks or manually). If no tasks exist, stop and tell the user to run /orc:tasks first.

Exception — direct execution: If $ARGUMENTS contains a concrete, self-contained task description (not a reference to a plan), create a single task with acceptance criteria inferred from the description, then proceed.

---

Convention Detection

Run these checks once before entering the execution loop.

Step 0: Detect commit conventions

Check these sources in priority order. Use the first match found:

1. Project CLAUDE.md — Read the project's CLAUDE.md. Look for a section with heading containing "Commit" (e.g., "Commit Messages", "Commit Convention", "Commits"). If found, follow that convention exactly. Skip remaining checks.

2. Commitlint config — Check for these files at the repo root (in this order):

   • .commitlintrc.json
   • .commitlintrc.yml / .commitlintrc.yaml
   • .commitlintrc.js / .commitlintrc.cjs / .commitlintrc.mjs
   • commitlint.config.js / commitlint.config.cjs / commitlint.config.mjs / commitlint.config.ts

   If found and the config is a static JSON/YAML file or a JS/TS file with a plain object export:

   • extends: ["@commitlint/config-conventional"] or extends: ["@commitlint/config-angular"] → confirms the preset format. Apply any custom rules (e.g., type-enum, scope-enum, subject-case).
   • Other static configs → extract the format rules and apply them.
   • Complex configs (dynamic rules, conditional logic, plugin imports) → fall back to default.

3. Default — Conventional Commits: type(scope): description

Precedence rule: Detected conventions replace format rules (message structure, type list, scope constraints). They do NOT replace process rules (logical grouping, gate requirements, no unrelated changes staged).

---

Execution Loop

For each task, follow these steps in exact order. Do NOT skip ahead. Each step has a gate.

Step 1: Pick up the task

• TaskGet to read the full description and acceptance criteria.
• Confirm all blockedBy tasks are complete. If blocked, skip to the next unblocked task.
• TaskUpdate(status='in_progress').

Verification Tasks

If the task has metadata.type === "verification", skip the main loop (Step 2 Implement, Step 3 Review + Validate, Step 4 Commit) and follow this path instead:

• Spawn the validator agent with the verification scenarios from the task description.
• PASS → proceed to Step 5 (mark complete).
• FAIL → report failing scenarios. Spawn worker to fix the failing areas. Re-run validator. Repeat until PASS.

Step 2: Implement — delegate to worker

• Spawn the worker agent with the task description, acceptance criteria, and file paths.
• NEVER edit implementation files yourself. NEVER substitute with direct tool calls.
• Wait for the worker to return.

⛔ STOP. The worker is done. Do NOT commit. Three mandatory gates remain.

Step 3: Validate

Spawn the validator agent:

• Run tests (existing + any new ones) and type checking.
• Return explicit pass/fail.

Handling results:

• PASS: proceed to Step 4.
• FAIL: delegate fixes to worker, then re-run validator. Repeat until PASS.

Step 4: Commit

You may ONLY reach this step with:

• ✅ validator PASS (this cycle)
• ✅ No unrelated changes staged

Apply the commit convention detected in Step 0.

• Imperative mood (unless the detected convention specifies otherwise)
• Logically grouped changes — not everything in one commit, not every line separately

Step 5: Complete

• TaskUpdate(status='completed').
• Check TaskList for the next unblocked task. Return to Step 1.

---

Rules

• NEVER execute if no tasks exist (unless direct execution via arguments)
• NEVER edit implementation files yourself — delegate to worker
• NEVER run tests yourself — delegate to validator
• NEVER skip the validator gate
• NEVER commit without validator PASS in the current cycle
• Verification tasks skip worker/reviewers/commit — validator only. Verification FAIL blocks completion.
• ALWAYS detect commit conventions before committing — Step 0 is mandatory
• If blocked by human-required input (credentials, decisions, access), stop and report