_arena-protocol

Arena Mode Protocol

Shared module for competitive agent evaluation. When /go --arena is active, multiple agents compete on the same story and gate-keeper selects the winner.

---

When Arena Mode is Active

1. Spawn Contestants

For each story, spawn 2-3 independent agent runs:

Contestant	Agent	Model	Rationale
A	coder	opus	Maximum quality
B	coder	sonnet	Cost-efficient alternative
C	coder	opus	Different perspective

The orchestrator determines contestants based on story complexity and cost-routing config.

2. Isolate Solutions

Each contestant works in complete isolation:

• Separate git worktrees (.arena/contestant-{a,b,c})
• No access to other contestants' output
• Same story file, same acceptance criteria
• Same project context (read-only access to shared codebase)

# Worktree setup (managed by arena-evaluate.sh)
git worktree add .arena/contestant-a -b arena/${story}-a
git worktree add .arena/contestant-b -b arena/${story}-b

3. Collect Solutions

Wait for all contestants to complete (or timeout):

• Default timeout: 2x normal story timeout
• If only 1 contestant completes, use that solution (graceful degradation)
• If 0 complete, fail the story with detailed error

4. Evaluate

Gate-keeper scores each solution on weighted criteria:

Criterion	Weight	What's Measured
Correctness	40%	Meets acceptance criteria, tests pass
Quality	25%	Code clarity, patterns, maintainability
Security	20%	No vulnerabilities, proper auth, input validation
Performance	15%	Efficient algorithms, no N+1 queries

Scoring scale: 0.0 to 1.0 per criterion.

5. Select Winner

• Highest weighted total score wins
• Winner's code is cherry-picked to the main branch
• Loser branches are deleted
• All scores logged to .claude/arena-results.jsonl

---

Resource Management

• Arena mode uses 2-3x resources (tokens, time)
• User confirmation required before arena starts
• Timeout per contestant: configurable, default 2x normal
• Graceful degradation: partial results are better than no results