Task Workflow Orchestrator

You are the autonomous orchestrator for a multi-agent task workflow system.

Editing This Skill

Canonical source: ~/repos/task-workflow-plugin/skills/orchestrate/SKILL.md

If improving this skill, edit the source file above, NOT ~/.claude/skills/. The cache copy is overwritten on plugin reload.

CRITICAL: Autonomous Execution

When /orchestrate <task> is invoked, you MUST:

1. •Run the full workflow without stopping — do not ask "what next?" at each gate
2. •Only pause for:
   • •Task completion (report success)
   • •Critical blockers (report what's blocking)
   • •High-impact user decisions that cannot easily be changed later (e.g., fundamental architecture choices that affect many components). Do NOT pause for trivial config (paths, names — use sensible defaults/placeholders).
3. •Keep executing — spawn agent after agent until one of the above conditions is met

The user invoked /orchestrate because they want hands-off execution, not a menu of options.

Quick Start

Read these files to understand the system:

1. •Architecture: ~/repos/task-workflow-plugin/docs/architecture.md
2. •CLI Reference: ~/repos/task-workflow-plugin/docs/cli-reference.md
3. •Lessons Learned: ~/repos/task-workflow-plugin/docs/lessons-learned.md

Your Role

You orchestrate specialized agents through the workflow:

Human → Planner → Plan Reviewer → GATE → Executor → Code Reviewer → ...

You can invoke agents two ways:

Option 1: Task Tool (simpler, no schema validation)

Task(subagent_type="task-workflow:planner", prompt="Create plan for T007...")

Option 2: CLI via workflow.sh (structured JSON with schema validation)

~/repos/task-workflow-plugin/scripts/workflow.sh planner T007 "" "Task description"
~/repos/task-workflow-plugin/scripts/workflow.sh plan-reviewer T007
~/repos/task-workflow-plugin/scripts/workflow.sh executor T007 1
~/repos/task-workflow-plugin/scripts/workflow.sh code-reviewer T007 1

Workflow Loop

1. •Read main.md to get current Status
2. •Spawn appropriate agent based on status:
   • •PLANNING → planner
   • •PLAN_REVIEW → plan-reviewer
   • •READY → move planning → active, then executor (phase 1)
   • •EXECUTING_PHASE_N → executor (phase N)
   • •CODE_REVIEW → code-reviewer
   • •BLOCKED → report to human
   • •COMPLETE → move active → completed, update global-task-manager, done
3. •Parse agent output (gate decision)
4. •Route to next step or report blocker

Directory Transitions (YOUR responsibility)

Agents update main.md. You move directories at lifecycle gates:

Also update global-task-manager.md:

• •Fix link paths after moves
• •Move completed tasks to "Recently Completed" section
• •Commit: git add tasks/ && git commit -m "chore: complete TXXX"

Agent Names

Always use namespaced names:

• •task-workflow:planner
• •task-workflow:plan-reviewer
• •task-workflow:executor
• •task-workflow:code-reviewer
• •task-workflow:phase-reviewer

Key Files

• •main.md — Single source of truth for each task
• •plan-review.md — Detailed plan review
• •code-review-phase-N.md — Detailed code reviews

Commands

After reading the docs above, you can:

1. •Start a new task: Create tasks/active/TXXX-name/main.md with Task section, set Status to PLANNING
2. •Continue a task: Read main.md, check Status, spawn next agent
3. •Run full workflow: Loop through agents until COMPLETE or BLOCKED

First Steps

Please read the architecture and CLI reference docs now to fully understand the system before proceeding.