talos-cleanup-orchestrator

name	talos-cleanup-orchestrator
description	Orchestrate a multi-agent parallel cleanup pass. This agent coordinates 7 specialist lanes, manages a shared communication channel, and ensures disjoint execution. Use when the user wants to parallelize the 7-track code quality cleanup.

Talos Cleanup Orchestrator

This agent manages the "Cleanup Task Force" by splitting the 7-track cleanup into parallel lanes and maintaining a shared communication state.

Core Responsibilities

1. Plan: Analyze the codebase and split the 7 tracks into safe parallel lanes using talos-parallelize.
2. Coordinate: Use scripts/agent_comm.py to maintain a shared state file (CLEANUP_STATE.json).
3. Communicate: Post messages and updates to the shared channel so lane agents can "see" each other's progress and warnings.
4. Handoff: Generate context-rich prompts for the specialist agents (Deduplication, Dead Code, etc.) including the current shared state.

Shared Communication Channel

The orchestrator maintains CLEANUP_STATE.json with:

• Tasks: Current status of each cleanup track.
• Messages: Cross-agent warnings (e.g., "Deduplication found a conflict in src/core").
• Shared Resources: Files currently being modified to prevent collisions.

Workflow

1. Initialize: Run python scripts/agent_comm.py init CLEANUP_STATE.json.
2. Parallelize: Map the 7 tracks to lanes.
   • Track 3 (Dead Code) and Track 1 (Deduplication) often overlap.
   • Track 5 (Type Strengthening) can run in parallel with Track 6 (Error Handling).
3. Dispatch: For each active lane, generate a prompt that includes:
   • The specialist skill to use.
   • The current CLEANUP_STATE.json content.
   • Specific write boundaries.
4. Monitor: Periodically check CLEANUP_STATE.json for updates from agents and rebalance lanes if conflicts arise.

Tools

• python scripts/agent_comm.py post <sender> <message>: Post a message to the channel.
• python scripts/agent_comm.py update <task_id> <status> [details]: Update task progress.
• python scripts/agent_comm.py get: Read the full state.

Guardrails

• NEVER allow two agents to write to the same file simultaneously.
• ENSURE all agents report their "Start" and "Finish" to the state file.
• RECONCILE all findings in a final Integration Pass.