orchestrate

/orchestrate

Claude Code Native Adaptation

Spawn .claude/agents/ subagents via Task tool instead of CLI (oh-my-ag agent:spawn). Task tool returns synchronously, so no polling needed.

Step 1: Load Plan

• Check if .agents/plan.json exists
• If not: Guide user to run /plan first

Step 2: Initialize Session

1. Load .agents/config/user-preferences.yaml
2. Response language follows language setting in .agents/config/user-preferences.yaml
3. Generate session ID (format: session-YYYYMMDD-HHMMSS)
4. Show agent routing (based on skill-routing.md)

Step 3: Spawn Agents (Parallel Task tool calls)

Spawn agents by priority tier:

• Multiple Task tool calls in same message = true parallel execution
• Each agent: Use .claude/agents/{agent}.md definition
• Include in prompt: Task description, API contract, context
• Include API contracts from .agents/skills/_shared/api-contracts/ if they exist
• Load only task-relevant context (check codebase structure around affected domains)

Agent mapping:

Domain	Subagent File
backend	.claude/agents/backend-impl.md
frontend	.claude/agents/frontend-impl.md
mobile	.claude/agents/mobile-impl.md
db	.claude/agents/db-impl.md
qa	.claude/agents/qa-reviewer.md
debug	.claude/agents/debug-investigator.md
pm	.claude/agents/pm-planner.md

Step 4: Monitoring

Task tool returns results directly → No polling needed. Check status, files changed, issues in each agent result.

Step 5: Agent-to-Agent Review Loop (Native Loop)

Main agent directly controls this loop. Maintain iteration counter.

iteration = 0
MAX_SELF = 3, MAX_CROSS = 2, MAX_TOTAL = 5

LOOP:
  iteration += 1
  if iteration > MAX_TOTAL → FORCE_COMPLETE (include quality warning)

  [1] Self-Review:
      Check self-review section in implementation agent results
      PASS → Proceed to [2]
      FAIL (self_count < MAX_SELF) → Re-spawn Task tool (with feedback) → LOOP
      FAIL (self_count >= MAX_SELF) → Force proceed to [2]

  [2] Automated Verification:
      Run lint/type-check/tests via Bash tool
      PASS → Proceed to [3]
      FAIL → Feed output back to agent as correction context (max 2 retries) → re-run [2]
      FAIL (retries exhausted) → Proceed to [3] with verification failure noted

  [3] Cross-Review:
      Spawn `qa-reviewer` subagent via Task tool
      Parse QA results: PASS / FAIL
      PASS → ACCEPT
      FAIL (cross_count < MAX_CROSS) → Feedback format:
        ## Review Feedback (iteration {n}/{MAX_TOTAL})
        **Reviewer**: qa-reviewer
        **Verdict**: FAIL
        **Issues**: [Specific file:line references]
        **Fix instruction**: [How to fix]
      → Re-spawn implementation agent Task tool (include feedback) → LOOP
      FAIL (cross_count >= MAX_CROSS) → Report to user with review history

Step 6: Collect Results

After all agents complete:

• Collect .agents/results/result-{agent}.md
• Organize completed/failed tasks, changed files, remaining issues

Step 7: Final Report

Session summary:

• Completed tasks
• Failed tasks (if failed after retries, include error details)
• Next step suggestions: Manual fix, re-run specific agent, /review QA

$ARGUMENTS