team-shinchan-fierce-compete - SKILL.md Agent Skill

name: team-shinchan:fierce-compete description: Deterministic competitive code tournament — N builders independently solve one task and return patches, an Action-Kamen judge scores them head-to-head, the winner is picked by score and applied. Opt-in main-loop Workflow tier. user-invocable: false

EXECUTE IMMEDIATELY

Fierce compete is a main-loop Workflow that runs a guaranteed competitive tournament: N builder agents independently solve the SAME task (read-only, each returning an apply-ready patch — no working-tree collisions, nothing to merge), an Action-Kamen judge scores them head-to-head on a rubric, and the winner is selected deterministically by total score. The Task path (/team-shinchan:debate competitive-code mode, Midori + worktrees) does this at the prompt level; this tier enforces the fan-out + scored judge with schema-validated output. Use for high-value, genuinely contested implementations where the best of N is worth the cost.

Step 0: Validate + opt-in

This skill calls the Workflow tool (main-loop only). Invoking /team-shinchan:fierce-compete IS the opt-in. Never delegate to a subagent — workflow() throws inside a Task child.
If args lack a clear implementation task: ask for it and STOP. Parse N (number of implementations, default 4, max 4 — full tournament unless the user asks for fewer) from "best of N"/"N개" if present. If > 2000 chars: truncate + warn.

Step 1: Resolve task + personas (main loop — the script can't read files)

Task + context: the implementation request; pass the relevant files (so builders read the right code).

Personas (DRY): resolve builder + judge personas from the agent files and pass them in args:

node ${CLAUDE_PLUGIN_ROOT}/src/workflow-personas.js bo actionkamen
node ${CLAUDE_PLUGIN_ROOT}/src/workflow-personas.js --learnings bo

Step 2: Run the fierce-compete Workflow

Workflow({
  scriptPath: "${CLAUDE_PLUGIN_ROOT}/skills/fierce-compete/fierce-compete.workflow.js",
  args: {
    task: "<the implementation request>",
    n: 4,                                  // 2..4 (default 4 when the user did not specify)
    files: ["<path>", "..."],              // context the builders should read
    builderPersona: "<workflow-personas.js bo>",
    builderLearnings: "<workflow-personas.js --learnings bo>",
    judgePersona: "<workflow-personas.js actionkamen>"
  }
})

Implement (N builders each emit a unified-diff patch, read-only) → Judge (Action Kamen scores each on correctness/completeness/quality, 1-5). Returns { task, n, implementations, scores, winner, rationale, dissent }. winner is chosen deterministically (max total, tie → higher correctness) and carries the winning patch. Note: builders are read-only, so patches are untested until applied — Step 4 is where the winner is applied and verified.

Step 3: Record the tournament artifact

Compute the next index NNN from .shinchan-docs/tournaments/COMPETE-*.json (zero-padded, start 001) and write the full return value plus an ISO timestamp to .shinchan-docs/tournaments/COMPETE-{NNN}.json. Also append a decision entry to .shinchan-docs/debate-decisions.md (same ledger competitive-code uses), recording the winner and the dissent.

Step 4: Present, apply the winner, confirm

Present the score table (per impl: correctness/completeness/quality/total), the winner, the rationale, and the surviving dissent. Then apply the winner:

git apply <winner.patch>     # write winner.patch to a temp file first

If git apply fails (LLM diffs can drift), re-create the change from the winner's approach + patch as a precise spec (a normal edit), then run the tests/build to verify (the patch was untested at generation time). Never finalize without confirmation — ask the user to accept the winner, pick a different implementation, or re-run. Surface the test result as evidence.