fleet-rebuttal

Files (1)

• SKILL.md

SKILL.md (first lines)

# Fleet Rebuttal

The cross-machine rebuttal protocol is the multi-machine analogue of 3-Surgeons cross-examination. Instead of 3 LLMs on one machine, 3 machines each run local surgery, produce structured verdicts, then critique each other's work. The Chief synthesizes everything into a final decision with preserved dissent.

This is the core of the 3x3x3 vision: same hospital, same protocol, different operating rooms, same patient.

## The 4-Phase Protocol

### Phase 1: Chief Assigns Task Brief

The Chief (human or coordinator node) sends an identical task brief to all machines. Each machine receives the same objective but works independently in its own worktree/branch.

**Task Brief Schema:**

```json
{
  "protocol": "3s-lan/v1",
  "message_type": "task_brief",
  "task_id": "task-abc123",
  "from": "chief",
  "objective": "Improve orchestration reliability without increasing complexity",
  "scope": ["orchestrator/", "agents/", "tests/"],
  "constraints": [
    "Must not break existing 71 tests",
    "No new external dependencies",
    "Performance budget: <5ms added latency"
  ],
  "context": "Current orchestrator has 3 known retry bugs. See issue #42.",
  "deadline_utc": "2026-04-05T18:00:00Z",
  "expected_output": "verdict"
}
```

**How to send the task brief:**

```bash
# Broadcast task brief to all fleet nodes
TASK_ID="task-$(date +%s | tail -c 7)"
curl -sf -X POST http://127.0.0.1:8855/message \
  -H "Content-Type: application/json" \
  -d "{
    \"type\": \"task_brief\",
    \"from\": \"chief\",
    \"payload\": {
      \"subject\": \"Task brief: ${TASK_ID}\",
      \"body\": {
        \"protocol\": \"3s-lan/v1\",
        \"message_type\": \"task_brief\",
        \"task_id\": \"${TASK_ID}\",
        \"from\": \"chief\",
        \"objective\": \"<objective here>\",
        \"scope\": [\"<dirs>\"],
        \"constraints\": [\"<list>\"],
        \"context\": \"<background>\",
        \"deadline_utc\": \"<ISO 8601>\",
        \"expected_output\": \"verdict\"
      }
    }
  }"
```

### Phase 2: Local Surgery + Verdict

Each machine receives the task brief and runs its full local 3-Surgeons protocol:

1. Head Surgeon triages the brief
2. Local surgeons investigate independently in their own worktree
3. Cross-examination within the local cell
4. Head Surgeon produces a **verdict packet** (see fleet-verdict skill)

Each machine works in its own branch:
- mac1: `feature/${TASK_ID}-mac1`
- mac2: `feature/${TASK_ID}-mac2`
- mac3: `feature/${TASK_ID}-mac3`

When local surgery is complete, each machine sends its verdict to the Chief (or broadcasts to all nodes).

**Timeline:** Machines work in parallel. Chief sets a deadline in the task brief. Verdicts arriving after deadline are accepted but marked `late: true`.

### Phase 3: Cross-Machine Critique

This is where the 3x3x3 protocol diverges from single-machine surgery. The Chief sends each machine's verdict to the OTHER machines for critique.

**Critique routing (round-robin):**
- mac1's verdict goes to mac2 and mac3
- mac2's verdict goes to mac1 and mac3
- mac3's verdict goes to mac1 and mac2

**Critique Request Schema:**

```json
{
  "protocol": "3s-lan/v1",
  "message_type": "critique_request",
  "task_id": "task-abc123",
  "from": "chief",
  "to": "mac2",
  "verdict_under_review": {
    "machine": "mac1",
    "summary": "...",
    "proposed_changes": ["..."],
    "assumptions": ["..."],
    "risks": ["..."],
    "confidence": 0.82,
    "evidence": ["..."]
  },
  "instructions": "Review this verdict. Challenge assumptions. Identify risks the author missed. State whether you agree, disagree, or partially agree. Be specific."
}
```

**How to send critique requests:**

```bash
# Chief sends mac1's verdict to mac2 for critique
curl -sf -X POST http://127.0.0.1:8855/message \
  -H "Content-Type: application/json" \
  -d "{
    \"type\": \"critique_request\",
    \"from\": \"chief\",

From supportersimulator/multi-fleet, MIT.

More from supportersimulator/multi-fleet