| name | gauntlet-orchestrate |
| description | Run the 3-agent Gauntlet debate. Codex (devils advocate) vs Reactor (positive advocate), Scribe mediates. Two rounds, then user decisions. |
| tags | ["workshop","gauntlet","debate","multi-agent"] |
| version | 1.0.0 |
gauntlet-orchestrate — Multi-Agent Debate
When to use
When an idea has all 10 fields filled and is ready for the Gauntlet stage.
Triggered by: s:gauntlet:run in Tap, or manual dispatch.
The Debate (2 rounds, 3 agents)
Setup
- Read idea from ideas-registry.json — all 10 fields
- Build idea context string (title, description, capabilities, constraints, success_test, etc)
- Set idea stage to "gauntlet"
Round 1
Run these as 3 ops_insert_task calls (Codex tasks can run back-to-back):
Task A — Devils Advocate (Codex):
Agent: spec-dev
Prompt: "You are the Devils Advocate. Challenge this idea on 6 dimensions:
1. Problem reality 2. Beneficiary honesty 3. Cost awareness
4. Success test rigor 5. Overlap check 6. Urgency reality
Be specific. Max 300 words.
IDEA: {context}"
Task B — Positive Advocate (Codex, blocked_by A):
Agent: spec-dev
Prompt: "You are the Positive Advocate. Counter each concern with a concrete fix.
IDEA: {context}
DEVILS ADVOCATE SAID: {output from Task A}
For each point: Valid/Invalid because X. Fix: Y. Max 300 words."
Task C — Mediator (Scribe, blocked_by B):
Agent: spec-projects
Prompt: "You are the Mediator. Summarize: which concerns resolved, which open.
IDEA: {context}
DEVILS ADVOCATE: {Task A output}
POSITIVE ADVOCATE: {Task B output}
List open decisions the user must make. Max 200 words."
Round 2 (blocked_by Round 1 mediator)
Same structure, but each agent responds to the full Round 1 transcript.
Codex pushes harder on unresolved items. Reactor proposes fixes. Scribe produces final scorecard.
User Decisions
After Round 2, Scribe's output becomes a decision menu:
- Each unresolved issue = one decision with button options
- Post to both Telegram topic (via Tap) and Bridge/Feedback
- User resolves each decision via buttons or typed answers
Completion
When all decisions are resolved:
- Mark Gauntlet as PASSED
- Update idea stage to "greenlight"
- Run workshop-dispatch skill to split into tasks
Cost
~4 Codex calls + 2 Scribe calls per Gauntlet.
At 3000/week Codex cap: sustainable for 2-3 Gauntlets/day.
Spec
Full spec: docs/gauntlet-v2-debate.md
Tools
- ops_insert_task with blocked_by for sequencing
- inbox_check to get debate outputs
- bearings_ask for user decisions
- chart_search for overlap check dimension
