Skip to content

SPEC.md

Latest commit

 

History

History
122 lines (95 loc) · 4.64 KB

SPEC.md

File metadata and controls

122 lines (95 loc) · 4.64 KB

Agent Teams Enhancement Spec

Goal

Enhance the existing team-tasks skill at /home/ubuntu/clawd/skills/team-tasks/ to support two new modes that replicate Claude Code's Agent Teams capabilities:

  1. Shared workspace mode — multiple agents work on the same codebase
  2. Debate mode — send the same question to N agents, collect responses, cross-evaluate

Current State

  • scripts/task_manager.py already supports linear and dag modes
  • Data files in /home/ubuntu/clawd/data/team-tasks/<project>.json
  • Agents are dispatched via OpenClaw's sessions_send and results tracked in JSON

Requirements

1. New mode: debate

Add a new mode --mode debate to task_manager.py init.

Debate workflow:

  1. init <project> --mode debate -g "question" — creates project with a question/topic
  2. add-debater <project> <agent-id> [--role "perspective"] — add agents as debaters, each with an optional role/perspective
  3. round <project> start — generates dispatch info for round 1 (each debater gets the question + their role)
  4. After collecting responses: round <project> collect <agent-id> "response" — save each debater's response
  5. round <project> cross-review — generates cross-review prompts (each debater reviews others' responses)
  6. After cross-review: round <project> collect <agent-id> "review" — save reviews
  7. round <project> synthesize — outputs all positions + reviews for final synthesis
  8. status <project> — shows debate state, rounds, and positions

Data structure for debate mode:

{
  "project": "security-review",
  "mode": "debate",
  "goal": "Review security of auth module",
  "debaters": {
    "agent-1": { "role": "security expert", "responses": [] },
    "agent-2": { "role": "performance analyst", "responses": [] }
  },
  "rounds": [
    {
      "type": "initial",
      "status": "done",
      "responses": { "agent-1": "...", "agent-2": "..." }
    },
    {
      "type": "cross-review", 
      "status": "in-progress",
      "responses": {}
    }
  ],
  "currentRound": 1,
  "status": "active"
}

2. Shared workspace tracking

Add --workspace <path> to init for all modes. This records the shared workspace path in the project JSON so all agents work in the same directory.

For dag and linear modes, add the workspace path to ready/next output so the dispatcher knows where to point agents.

3. Cross-review prompt generation

The round <project> cross-review command should output structured prompts for each debater:

Agent: agent-1 (security expert)
Your previous response: <agent-1's response>

Other debaters' responses:
- agent-2 (performance analyst): <agent-2's response>
- agent-3 (testing specialist): <agent-3's response>

Task: Review the other responses. Do you agree or disagree? What did they miss? Update your position if needed.

Constraints

  • Keep backward compatibility with existing linear and dag modes
  • Same CLI pattern as existing commands
  • Same data directory: /home/ubuntu/clawd/data/team-tasks/
  • Python 3.12+, no external dependencies beyond stdlib
  • Keep code clean and readable — this will be used by AI agents

Files to modify

  • /home/ubuntu/clawd/skills/team-tasks/scripts/task_manager.py — add debate mode commands

Testing

After implementing, verify with this test scenario:

TM="python3 /home/ubuntu/clawd/skills/team-tasks/scripts/task_manager.py"

# Create debate
$TM init security-debate --mode debate -g "Review the auth module at /tmp/auth.py for security vulnerabilities"

# Add 3 debaters
$TM add-debater security-debate code-agent --role "security expert focused on injection attacks"
$TM add-debater security-debate test-agent --role "QA engineer focused on edge cases"
$TM add-debater security-debate monitor-bot --role "ops engineer focused on deployment risks"

# Start round 1
$TM round security-debate start

# Collect responses
$TM round security-debate collect code-agent "Found SQL injection in login()"
$TM round security-debate collect test-agent "Missing input validation on email field"
$TM round security-debate collect monitor-bot "No rate limiting on auth endpoints"

# Generate cross-review prompts
$TM round security-debate cross-review

# Collect cross-reviews
$TM round security-debate collect code-agent "Agree with test-agent on validation. monitor-bot's rate limiting is critical."
$TM round security-debate collect test-agent "code-agent's SQL injection is most severe. Adding rate limit tests."
$TM round security-debate collect monitor-bot "Both findings are valid. Recommending WAF as additional layer."

# Synthesize
$TM round security-debate synthesize

# Check status throughout
$TM status security-debate