Execution Phase Details

This document contains detailed execution logic for the executing-plans skill.

Step 2: Execute Batch

Default: First 3 tasks (or use dependency graph for collab workflow)

Standard Execution (No Dependency Graph)

For each task:

1. Mark as in_progress
2. Follow each step exactly (plan has bite-sized steps)
3. Run verifications as specified
4. Mark as completed

Step 2.1: Per-Task Execution with Item Type Routing

Before executing each task, determine its item type and follow the appropriate execution path.

Determine Item Type:

1. Read design doc
2. Find the work item that this task belongs to
3. Check the

   Type:

   field from the work item
4. Map to execution flow

Routing Logic:

FUNCTION executeTask(task, itemType):
  IF itemType == "task":
    # Skip TDD for operational tasks
    EXECUTE task steps directly (from task-planning Prerequisites/Steps/Verification)
    RUN verification checks
    MARK task as complete if verification passes
  ELSE IF itemType IN ["code", "bugfix"]:
    # Normal TDD flow
    INVOKE test-driven-development skill
    WRITE failing test
    IMPLEMENT code per design spec
    VERIFY test passes
    MARK task as complete
  ELSE:
    STOP - unknown item type, ask user

For Task Type Items:

1. Execute prerequisites validation (verify all prerequisites exist)
2. Execute each step in order (run commands/actions)
3. Run verification checks (confirm success)
4. Mark task as complete

For Code/Bugfix Type Items:

1. Invoke test-driven-development skill
2. Work through red-green-refactor cycle
3. Mark task as complete

Dependency-Aware Execution (Collab Workflow)

When a task dependency graph is present, use intelligent parallel dispatch:

Find Ready Tasks:

ready_tasks = tasks where:
  - status is "pending"
  - all depends-on tasks are in "completed"

Parallel Dispatch Logic:

1. From ready tasks, identify parallel-safe group:
   • Tasks explicitly marked

     parallel: true

   • OR tasks with no file overlap and no shared dependencies
2. If multiple parallel-safe tasks exist:
   • Update task diagram: set all parallel tasks to "executing"
   • REQUIRED: Spawn Task agents in parallel (single message, multiple tool calls)
   • Each Task agent MUST invoke

     mermaid-collab:subagent-driven-development:implementer-prompt

     skill
   • Task prompt includes: task ID, files, description, relevant pseudocode
   • Wait for all agents to complete
   • Update task diagram: completed → "completed", failed → "failed"
3. If only sequential tasks remain:
   • Execute one at a time in topological order
   • Update diagram before/after each task

RED FLAG - INLINE IMPLEMENTATION: If you find yourself using Edit/Write tools directly on source files instead of spawning Task agents, you are violating the subagent requirement. STOP and use Task tool instead.

Task Agent Prompt Template (Collab Workflow)

You are implementing a task from the collab workflow.

## Design Document Location
Collab Session: .collab/<session-name>

**Per-item documents (read these for your task's context):**
- Interface: .collab/<session-name>/documents/interface-item-<N>.md
- Pseudocode: .collab/<session-name>/documents/pseudocode-item-<N>.md
- Skeleton: .collab/<session-name>/documents/skeleton-item-<N>.md

## REQUIRED: Read Per-Item Documents First
Before implementing, read the per-item documents for your work item:
1. Read interface-item-<N>.md → function signatures, types, file paths
2. Read pseudocode-item-<N>.md → step-by-step logic for your task
3. Read skeleton-item-<N>.md → task graph and planned files

These documents are the SOURCE OF TRUTH. Follow them exactly.

## Task Details
Task ID: <task-id>
Item Number: <item-number>
Files: <task-files>
Description: <task-description>

## Your Task's Design Spec
Interface:
<paste from interface-item-<N>.md>

Pseudocode:
<paste from pseudocode-item-<N>.md>

## Instructions
1. Read the per-item documents above
2. Implement EXACTLY as specified - no interpretation
3. Write tests
4. Report what you implemented

Task Prompt with Test Patterns

When dispatching tasks, include the

tests

field in the prompt:

## Targeted Tests

During TDD (RED-GREEN-REFACTOR), run ONLY these tests:
{task.tests}

Command: npm run test:ci -- {tests joined by space}

Do NOT run the full test suite during TDD cycles.

Task Completion Handling

When a task completes:

1. Move task from

   in_progress

   to

   completed

2. Check what tasks are now unblocked (their

   depends-on

   all satisfied)
3. Add newly unblocked tasks to the ready queue
4. Repeat until all tasks done

Wave Completion Checkpoint

After all tasks in a wave complete:

1. Run full test suite:

   npm run test:ci
   

2. If tests fail:

   Full test suite failed after wave completion.
   Investigate failures before proceeding to next wave.
   

   STOP and report.

3. If tests pass:

   Full test suite passed. Proceeding to next wave.
   

Example Execution Flow:

Wave 1: [auth-types, utils] (parallel: true) → dispatch together
  ↓ both complete
Wave 2: [auth-service] (depends-on: auth-types) → dispatch
  ↓ complete
Wave 3: [auth-middleware] (depends-on: auth-service) → dispatch