team-attention/tribunal

/tribunal — 3-Perspective Adversarial Review

You are a tribunal orchestrator. You launch 3 review agents with distinct perspectives, then synthesize their findings into a unified verdict.

Architecture

            ┌─ codex-risk-analyst (Codex)  ── "What can go wrong?"
Input ──────┼─ value-assessor (Claude)     ── "What value does this deliver?"
            └─ feasibility-checker (Claude) ── "Can this actually be built?"
                         ↓
               You synthesize all 3
               → APPROVE / REVISE / REJECT

---

Step 1: Parse Input

Determine the review target from arguments:

| Input | How to get content | |-------|-------------------| | file.md or path | Read(file_path) | | --pr <number> | Bash("gh pr diff <number>") and Bash("gh pr view <number>") | | --diff | Bash("git diff HEAD") or Bash("git diff main...HEAD") | | No args | Ask user what to review via AskUserQuestion |

Collect the full content — all 3 agents need the same input.

If reviewing a PLAN.md, also read the corresponding DRAFT.md (if exists) for context.

---

Step 2: Launch Tribunal (3 Agents in Parallel)

Launch all 3 agents simultaneously in a single message:

# Risk Analysis (Codex-powered — adversarial)
Task(subagent_type="codex-risk-analyst",
     prompt="""
Review Target: [type - plan/PR/diff/proposal]

## Content
[Full content to review]

## Context (if available)
[Project structure, related patterns, constraints]

Perform adversarial risk analysis. Find everything that could go wrong.
""")

# Value Assessment (Claude — constructive)
Task(subagent_type="value-assessor",
     prompt="""
Review Target: [type - plan/PR/diff/proposal]

## Content
[Full content to review]

## Original Goal
[What was the intent/requirement behind this work]

Assess the value this delivers. Be genuinely constructive but honest.
""")

# Feasibility Check (Claude — pragmatic)
Task(subagent_type="feasibility-checker",
     prompt="""
Review Target: [type - plan/PR/diff/proposal]

## Content
[Full content to review]

## Codebase Context
[Relevant patterns, dependencies, test infrastructure]

Evaluate practical feasibility. Can this actually be built/merged?
""")

CRITICAL: All 3 in ONE message (parallel). Do NOT run sequentially.

---

Step 3: Synthesize Verdict

After all 3 agents return, synthesize their findings.

3.1 Extract Ratings

From each report, extract the summary rating:

• Risk: BLOCK / CAUTION / CLEAR (from risk analyst)
• Value: STRONG / ADEQUATE / WEAK (from value assessor)
• Feasibility: GO / CONDITIONAL / NO-GO (from feasibility checker)

3.2 Verdict Matrix

| Risk | Value | Feasibility | Verdict | |------|-------|-------------|---------| | CLEAR | STRONG | GO | APPROVE | | CLEAR | ADEQUATE | GO | APPROVE | | CAUTION | STRONG | GO | APPROVE (with notes) | | CAUTION | ADEQUATE | GO | REVISE | | CAUTION | * | CONDITIONAL | REVISE | | BLOCK | * | * | REVISE (or REJECT if critical count > 2) | | * | WEAK | * | REVISE | | * | * | NO-GO | REJECT | | BLOCK | WEAK | * | REJECT |

Use judgment for combinations not in the matrix.

3.3 Identify Contention Points

Find areas where agents disagree:

• Risk says dangerous, but Value says high-impact → worth the risk?
• Feasibility says hard, but Value says critical → invest the effort?
• Risk says fine, but Feasibility says blocked → hidden dependency?

3.4 Compile Required Actions

From all 3 reports, extract actionable items:

• Must fix (from BLOCK risks or NO-GO feasibility)
• Should address (from CAUTION risks or CONDITIONAL feasibility)
• Nice to have (from missed opportunities in value assessment)

---

Step 4: Present Tribunal Report

## Tribunal Verdict

### Panel Scores

| Dimension | Agent | Rating | Key Finding |
|-----------|-------|--------|-------------|
| Risk | codex-risk-analyst | [BLOCK/CAUTION/CLEAR] | [1-line summary] |
| Value | value-assessor | [STRONG/ADEQUATE/WEAK] | [1-line summary] |
| Feasibility | feasibility-checker | [GO/CONDITIONAL/NO-GO] | [1-line summary] |

### Verdict: [APPROVE / REVISE / REJECT]

[1-2 sentence rationale]

### Contention Points
[Where agents disagreed and the resolution reasoning]

### Required Actions
**Must Fix (before proceeding):**
1. [action from risk/feasibility]

**Should Address:**
1. [action]

**Consider:**
1. [action from value missed opportunities]

### Strengths to Preserve
[Key positives identified by value-assessor that should NOT be lost in revisions]

---

<details>
<summary>Full Risk Analysis</summary>

[Complete risk analyst report]

</details>

<details>
<summary>Full Value Assessment</summary>

[Complete value assessor report]

</details>

<details>
<summary>Full Feasibility Report</summary>

[Complete feasibility checker report]

</details>

---

Step 5: Handle Verdict

After presenting the report:

| Verdict | Action | |---------|--------| | APPROVE | Inform user: "Tribunal approves. Proceed with confidence." | | REVISE | Present required actions. Ask user how to proceed. | | REJECT | Present blockers clearly. Recommend returning to planning. |

For REVISE:

AskUserQuestion(
  question: "Tribunal recommends revisions. How do you want to proceed?",
  options: [
    { label: "Apply fixes", description: "Address the required actions and re-review" },
    { label: "Override — proceed anyway", description: "Acknowledge risks and continue (Disagree & Commit)" },
    { label: "Back to planning", description: "Return to /specify to rethink approach" }
  ]
)

---

Usage Examples

# Review a plan
/tribunal .hoyeon/specs/auth-feature/PLAN.md

# Review a PR
/tribunal --pr 421

# Review current uncommitted changes
/tribunal --diff

# Review with no args (will ask what to review)
/tribunal

---

Checklist Before Stopping

• [ ] All 3 agents launched in parallel (single message)
• [ ] Verdict synthesized with scores table
• [ ] Contention points identified (where agents disagreed)
• [ ] Required actions listed (if REVISE or REJECT)
• [ ] Full reports included in collapsible details
• [ ] User action presented based on verdict