autopilot-run

<user_request> $ARGUMENTS </user_request>

  **You MUST:**
  - Use Task tool to delegate to task-executor agent
  - Track progress via Tasks
  - Enforce state transitions
  - Generate proof artifacts

  **You MUST NOT:**
  - Execute task work directly
  - Skip proof generation
  - Bypass state machine
</orchestrator_role>

<todowrite_requirement>
  Track execution phases:
  1. Fetch task from Linear
  2. Validate prerequisites
  3. Classify and route
  4. Execute via task-executor
  5. Generate proof
  6. Validate and transition state
  7. Report completion
</todowrite_requirement>

</critical_constraints>

<phase number="2" name="Classify and Route">
  <steps>
    <step>
      Extract tags from issue labels
    </step>
    <step>
      Select agent/command based on tag mapping:
      - Read config from .claude/autopilot.local.md
      - Apply tag precedence rules
      - Fall back to default if no match
    </step>
    <step>
      Transition state: Todo -> In Progress
      ```bash
      bun run ${CLAUDE_PLUGIN_ROOT}/scripts/linear-client.ts transition \
        --id "$ISSUE_ID" \
        --state "In Progress"
      ```
    </step>
  </steps>
</phase>

<phase number="3" name="Execute Task">
  <steps>
    <step>
      Create session directory:
      ```bash
      SESSION_PATH="ai-docs/sessions/autopilot-${ISSUE_ID}-$(date +%Y%m%d-%H%M%S)"
      mkdir -p "${SESSION_PATH}/proof"
      ```
    </step>
    <step>
      Write task context:
      ```bash
      cat > "${SESSION_PATH}/task-context.md" << EOF
      # Task: ${ISSUE_ID}

      **Title**: ${title}
      **Description**: ${description}
      **Tags**: ${tags}
      **Acceptance Criteria**:
      ${acceptance_criteria}
      EOF
      ```
    </step>
    <step>
      Delegate to task-executor agent:
      ```
      Task: autopilot:task-executor
        Prompt: "SESSION_PATH: ${SESSION_PATH}

                 Execute task from ${SESSION_PATH}/task-context.md

                 Command to run: {selected_command}
                 Skills to load: {selected_skills}

                 Write execution log to ${SESSION_PATH}/execution-log.md
                 Return brief summary when complete."
      ```
    </step>
  </steps>
</phase>

<phase number="4" name="Generate Proof">
  <steps>
    <step>
      Delegate to proof-generator agent:
      ```
      Task: autopilot:proof-generator
        Prompt: "SESSION_PATH: ${SESSION_PATH}
                 TASK_TYPE: {task_type}
                 ISSUE_ID: ${ISSUE_ID}

                 Generate proof-of-work artifacts.
                 Write proof summary to ${SESSION_PATH}/proof/summary.md
                 Return confidence score."
      ```
    </step>
    <step>
      Read confidence score from response
    </step>
  </steps>
</phase>