tribunal

/tribunal — Multi-Mode AI Debate

Two AI buddies engage on a codebase question using one of six modes. Every claim requires FILE:LINE evidence.

Modes

| Mode | Flag | AIs do | Claude's role | Best for | |------|------|--------|---------------|----------| | adversarial | (default) | Argue FOR vs AGAINST | Judge — picks winner | Binary decisions, should/shouldn't | | socratic | --socratic | Probe assumptions with questions | Synthesizer — surfaces insights | Early exploration, unclear framing | | steelman | --steelman | Argue the OTHER side's strongest case | Calibrator — shows true strength of each side | Avoiding confirmation bias | | red-team | --red-team | Attack from different angles, no defense | Risk assessor — prioritized vulnerability table | Poking holes in designs/plans | | synthesis | --synthesis | Each proposes a solution, then hybridize | Merger — evaluates proposals + hybrid | Finding a third option | | postmortem | --postmortem | Investigate failure from different angles | Investigator — unified timeline + root cause | Bug investigation, incident analysis |

How to invoke

/tribunal "Should we refactor the auth middleware?"
/tribunal --socratic "Is our error handling resilient enough?"
/tribunal --steelman "Should we migrate to microservices?"
/tribunal --red-team "Review our new payment flow"
/tribunal --synthesis "How should we restructure the data layer?"
/tribunal --postmortem "Why did the deploy fail yesterday?"
/tribunal --mode red-team "Audit the new API endpoints"

Step-by-step workflow

Phase 0: Setup

1. Parse the question and mode. Check for mode flags (--socratic, --steelman, --red-team, --synthesis, --postmortem, or --mode X). Default: adversarial.
2. Detect available buddies:

source "${CLAUDE_PLUGIN_ROOT}/hooks/lib.sh"
AVAILABLE=$(ai_buddies_available_buddies)

1. Require at least 2 buddies.
2. Tell the user which buddies will participate, the mode, and round count.

Phase 1: Dispatch

Before dispatching: Use your judgment — if the conversation contains context that would help the debaters (constraints, failed approaches, user preferences), summarize it in $CONVERSATION_CONTEXT. If the question is self-contained, leave it empty.

IMPORTANT: Set the Bash tool's timeout parameter to 600000 (10 minutes). The default Bash timeout (120s) will kill slow engines mid-debate.

MANIFEST_PATH=$(bash "${CLAUDE_PLUGIN_ROOT}/scripts/tribunal-run.sh" \
  --question "THE_QUESTION" \
  --cwd "$(pwd)" \
  --mode MODE_NAME \
  --rounds 2 \
  --timeout 600 \
  --conversation-context "$CONVERSATION_CONTEXT")

Phase 2: Read results

Read $MANIFEST_PATH (tribunal-manifest.json):

{
  "question": "...",
  "mode": "adversarial|socratic|steelman|red-team|synthesis|postmortem",
  "rounds": 2,
  "debaters": ["codex", "gemini"],
  "arguments": { ... }
}

Phase 3: Evaluate

Read the mode-specific guide for how to evaluate results and format output:

${CLAUDE_PLUGIN_ROOT}/skills/tribunal/modes/{mode}.md

Read the file matching the mode from the manifest, then follow its judging/synthesis instructions and output format.

Round constraints

| Mode | Rounds | Reason | |------|--------|--------| | adversarial | Flexible (default 2) | More rounds = deeper rebuttals | | steelman | Flexible (default 2) | Like adversarial but with reversed positions | | socratic | Fixed 2 | Round 1 asks, Round 2 answers | | red-team | Fixed 2 | Round 1 attacks, Round 2 chains attacks | | synthesis | Fixed 2 | Round 1 proposes, Round 2 hybridizes | | postmortem | Fixed 2 | Round 1 investigates, Round 2 cross-examines |

Auto-triggers

1. Forge close call: When forge scores are within 3 points
2. Review disagreement: When /codex-review and /gemini-review conflict

Configuration

| Key | Default | Description | |-----|---------|-------------| | tribunal_rounds | 2 | Default round count | | tribunal_max_buddies | 3 | Max debaters |

Rules

• Evidence over eloquence. Every claim needs file:line evidence.
• Verify citations. Always read referenced files to confirm.
• No-evidence = zero. Enforce in all modes.
• Follow the mode guide. Each mode has different judging criteria and output format.
• Always clean up worktrees after the session.