fix(coder): harden edit-apply pipeline against block duplication

Root cause: two proven corruption mechanisms — (M1) non-idempotent apply
stamped the same block N times when a quantized model re-emitted the same
edit_file call or a turn was retried; (M2) Levenshtein tier 4 was fail-open
with no uniqueness guard, silently splicing into the wrong location.

Fixes applied at every layer of the pipeline:

Matcher (fuzzy-match.ts): raise SIMILARITY_THRESHOLD 0.66 → 0.85; add
AMBIGUITY_EPSILON uniqueness guard — two windows within 0.05 of the top
score → ambiguous, not a guess; add block-anchor gate (≥3-line needles
require first+last line exact match before a window is scored).

Edit planner (pending_changes.ts): extract planEdit() as a pure function;
idempotency guards detect already-applied states (anchored insert re-stamp,
old-gone-but-new-present); findPendingDuplicate() collapses identical
pending rows at queue time so M1 never reaches applyOne.

Atomic writes (pending_changes.ts): temp-file + rename on the same
filesystem so a crash can't leave a half-written source file; realpath()
first so symlinks survive the rename.

Per-file mutex (pending_changes.ts): withFileLock() serializes concurrent
read-modify-write on the same path via a chained-Promise Map.

EOL preservation (pending_changes.ts): normalize CRLF → LF for matching,
restore native line ending on write so Windows-style files stay clean.

Context isolation (inference_context.ts): replace module-level singleton
with AsyncLocalStorage so concurrent inference runs (arena parallel
dispatch, dispatcher poll racing a user message) each get their own
scoped context with no clobbering.

Tests: plan-edit.test.ts (pure planEdit unit tests), extended fuzzy-match
and pending_changes_integration suites, ALS isolation test that proves
overlapping runs get correct session IDs.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

.gitignore

@@ -20,3 +20,4 @@ data/*
 !data/mcp.example.json
 !data/coder-providers.example.json
 codecontext/fork.tar.gz
+/Arena

CHANGELOG.md

@@ -2,6 +2,10 @@
 
 All notable changes per release tag. Most recent on top, ordered by tag creation date (which matches the git history). Tag names follow `vMAJOR.MINOR.PATCH-slug` — the slug describes what shipped, so the tag name alone is enough to recall the batch.
 
+## v2.7.20-arena-pane — 2026-06-06
+
+Adds the **Arena** pane for running the same prompt against 2–6 AI competitors simultaneously and picking the best result. A Battle is one Arena run: pick a battle type (Coding — backend+model with git worktrees producing diffs; or Q&A — BooChat persona+model producing text), write or generate a prompt, add contestants, and hit Start. Contestants are scheduled in two concurrent lanes — the local lane (llama-swap models, serial) and the cloud lane (Claude Code, OpenCode-on-cloud, parallel). The lane scheduler captures wall-clock duration for every contestant and tokens/sec for local models. When all contestants finish, a two-stage analysis (digest then judge) auto-runs on the DEFAULT_MODEL, writing `analysis.md` naming a winner; the user can override the winner per-row or trigger cross-examination. Results land in `/<project-root>/Arena/<dated-battle>/` with per-contestant `result.md`, diff patches for coding, and `manifest.json`. Replaces the old API-only `POST /api/arena` with dedicated `battles`/`contestants`/`cross_examinations` tables and full UI. Also adds a `DiffView` component with line-by-line colored unified diff and a per-row dropdown for winner override. Built on `v2.7.18-permission-modes`; pairs conceptually with the earlier `v2.7.17-orchestrator` multi-agent work (both share the pane kind pattern and `onTaskTerminal` hook).
+
 ## v2.7.18-permission-modes — 2026-06-05
 
 Adds a unified **permission picker** to the BooCoder composer — Plan / Ask Permission / Bypass — replacing the old raw per-agent mode dropdown that exposed each agent's full native vocabulary with inconsistent labels. The three options map generically onto every provider's existing mode metadata: the `plan`-id mode → Plan, the default mode → Ask, the `isUnattended` mode → Bypass (claude `bypassPermissions`, qwen `yolo`, opencode `full-access`); goose has no modes so it shows no picker, exactly as before. `modeId` stays the single wire field — the active unified mode is derived from it, so no contracts change was needed. Native BooCode gains its own mode set (registered in the manifest and exposed by the snapshot): **Ask** stages edits to the pending-changes queue as today, **Bypass** auto-applies the queue to disk after the turn (both the interactive messages path and the task-based dispatcher path), and **Plan** falls back to Ask — the shared `apps/server` inference engine is deliberately left untouched. A supporting fix preserves the `isUnattended` flag on live-probed ACP modes (`acp-derive.ts`) so opencode's bypass mode is still detectable from the wire. Coder 373 tests green, coder + web typecheck clean. Built on `v2.7.17-orchestrator`.

CLAUDE.md

@@ -78,7 +78,7 @@ BooCoder at port 9502: `curl http://100.114.205.53:9502/api/health`. Runs as `bo
 
 - `FAST_MODEL` (optional) — cheaper model for titles, summaries, labeling (auto_name.ts, tool-summaries.ts). Falls back to session model or DEFAULT_MODEL. Set to a small llama-swap model (e.g. `nemotron-nano-4b`) to avoid loading the 35B for 20-token calls.
 - Qwen Code dispatch: `OPENAI_BASE_URL=http://100.101.41.16:8401/v1 OPENAI_API_KEY=dummy qwen -p "<task>" --output-format stream-json`. Install: `npm install -g @qwen-code/qwen-code@latest`. Node ≥22 on host (container stays Node 20; BooCoder dispatches via direct spawn on host). No `--yolo` flag — `-p` runs autonomously without prompts. ACP bridge is an HTTP daemon (not stdio); use PTY dispatch.
-- Arena: `POST /api/arena {project_id, input, contestants: [{agent?, model?}]}` dispatches the same task to N models/agents in parallel; each contestant gets its own task + worktree. `GET /api/arena/:id` for results; `POST /api/arena/:id/select/:task_id` picks a winner.
+- Arena: `POST /api/battles {project_id, battle_type, prompt, contestants}` starts a battle; `GET /api/battles/:id` returns battle + contestants + cross-examinations; `POST /api/battles/:id/stop` cancels; `POST /api/battles/:id/analyze` triggers/re-triggers two-stage digest→judge analysis; `GET /api/battles/:id/analysis` reads `analysis.md`; `POST /api/battles/:id/cross-examine {identity, model}` runs a cross-examination. All `/api/battles*` routes are served by `apps/coder` at port 9502 (proxied through `apps/server` as `/api/coder/battles*`).
 
 ## Workflow
 

CONTEXT.md

@@ -0,0 +1,67 @@
+# Context: BooCode
+
+Glossary of the domain language. Terms only — no implementation detail.
+
+## Workspace
+
+- **Pane** — one tile in the multi-pane workspace. Each pane has a *kind*:
+  Chat (BooChat), Coder (BooCoder), Terminal (BooTerm), Orchestrator, Arena,
+  plus artifact/settings kinds.
+
+- **Backend** — an AI engine a task is dispatched to: *native* (BooChat
+  inference on a local llama-swap model) or an *external* CLI agent (Claude Code,
+  OpenCode, Qwen, Goose). Code sometimes calls this the "agent" (`tasks.agent`).
+
+- **BooChat Agent** (a.k.a. *persona*) — a preset from the `data/AGENTS.md`
+  registry (e.g. "Code Reviewer", "Debugger"): a system prompt + tool whitelist +
+  sampling knobs that runs **on the native backend** with a chosen model.
+  Distinct from a Backend — this is the overloaded sense of "agent" the UI's
+  Agent picker selects.
+
+## Arena
+
+A way to run the **same prompt** against several AI competitors at once and pick
+the best result.
+
+- **Battle** — one Arena run. Dated. Produces a results folder at
+  `/<project-root>/Arena/<dated-battle>/`. (The earlier API-only feature called
+  this an "arena"; a Battle is one such run.)
+
+- **Battle Type** — what is being compared:
+  - *Coding* — Contestants change code; a result is the **diff** they produced
+    (plus their explanation). Each Contestant works in its own worktree.
+  - *Q&A* — Contestants answer a prompt; a result is the **text answer**. No
+    code changes.
+
+- **Contestant** — one competitor in a Battle, given the Battle's prompt. What
+  defines a Contestant depends on Battle Type:
+  - *Coding* — a **Backend + Model** (e.g. Claude Code + opus, native BooCode +
+    35b). Each works in its own isolated git **worktree** (a branched on-disk
+    copy of the project). Contestants do not see each other's work.
+  - *Q&A* — a **BooChat Agent (persona) + Model** (e.g. Debugger + 35b), running
+    on the native backend only. No worktree (no code changes).
+  The same model can appear under two Contestants, so a Contestant's identity is
+  the (backend-or-persona, model) pair, not the model alone.
+
+- **Benchmark** — per-Contestant performance captured during a Battle. Wall-clock
+  **duration** is recorded for every Contestant; **throughput** (tokens/sec) is
+  recorded only for local (llama-swap) models, which are the ones the speed
+  comparison is meaningful for.
+
+- **Arena results folder** (`/<project-root>/Arena/<dated-battle>/`) — where a
+  Battle's *results* are written (not the working copies — those stay in each
+  Contestant's worktree). Holds the per-Contestant result and the final
+  analysis.
+
+- **Lane** — how a Battle's Contestants are scheduled. The *local lane* holds
+  every llama-swap-backed Contestant and runs them strictly one at a time (the
+  local server can only load one model at a time, which also keeps their speed
+  Benchmark fair). The *cloud lane* holds cloud-backed Contestants (Claude Code,
+  OpenCode-on-cloud) and runs them all in parallel. The two lanes run
+  concurrently with each other.
+
+- **Analysis** — an end-of-Battle judgement of the Contestants' results,
+  produced by the default BooChat model, naming a **Winner**.
+
+- **Cross-examination** — an after-the-Battle step where a chosen model (from any
+  agent) is pointed at the Battle's results to interrogate / compare them.

apps/coder/src/index.ts

@@ -13,7 +13,7 @@ import type { WsFrame } from '@boocode/contracts/ws-frames';
 // v2.0.0 Phase 2C: write tools + adapter for BooChat ToolDef compatibility.
 import { WRITE_TOOLS } from './services/tools/index.js';
 import { adaptWriteTool } from './services/tools/adapter.js';
-import { setInferenceContext, clearInferenceContext } from './services/tools/inference_context.js';
+import { runWithInferenceContext } from './services/tools/inference_context.js';
 // Routes
 import { registerMessageRoutes } from './routes/messages.js';
 import { registerSkillRoutes } from './routes/skills.js';
@@ -23,8 +23,8 @@ import { registerAgentSessionRoutes } from './routes/agent-sessions.js';
 import { registerTaskRoutes } from './routes/tasks.js';
 import { registerInboxRoutes } from './routes/inbox.js';
 import { registerStatsRoutes } from './routes/stats.js';
-import { registerArenaRoutes } from './routes/arena.js';
 import { registerRunsRoutes } from './routes/runs.js';
+import { registerArenaRoutes } from './routes/arena.js';
 import { registerProviderRoutes } from './routes/providers.js';
 import { registerWorktreeSafetyRoutes } from './routes/worktree-safety.js';
 import { registerLifecycleRoutes } from './routes/lifecycle.js';
@@ -34,10 +34,13 @@ import { createDispatcher } from './services/dispatcher.js';
 // Orchestrator (Phase 2): DB-backed flow-runner; advances on the dispatcher's
 // onTaskTerminal hook.
 import { createFlowRunner } from './services/flow-runner.js';
+// Arena: DB-backed battle-runner; also advances on the onTaskTerminal hook.
+import { createBattleRunner, type DispatchContestantFn } from './services/arena-runner.js';
+import { createAnalyzer } from './services/arena-analyzer.js';
 import { agentPool } from './services/agent-pool.js';
 import { createOrphanWorktreeReaper } from './services/orphan-worktree-reaper.js';
 import { probeAgents } from './services/agent-probe.js';
-import { getProviderSnapshot, persistProbedModels } from './services/provider-snapshot.js';
+import { getProviderSnapshot, persistProbedModels, fetchLlamaSwapModels } from './services/provider-snapshot.js';
 import { setPermissionHooks } from './services/permission-waiter.js';
 import { publishAgentStatus } from './services/agent-status-publish.js';
 import { homedir } from 'node:os';
@@ -171,22 +174,27 @@ async function main() {
     }
   );
 
-  // Wrap the inference runner to set/clear the write-tool context around each run.
-  // The inference runner calls enqueue() which fires asynchronously — we hook
-  // into the enqueue to set context before the run starts.
+  // Wrap the inference runner to bind the write-tool context around each run.
+  // enqueue() starts its async loop synchronously, so wrapping the call in
+  // runWithInferenceContext propagates the per-run context (sql, sessionId, the
+  // Plan/Ask/Bypass gate) through every awaited tool execution — and concurrent
+  // runs (a user message racing a dispatcher-polled native task) each get their
+  // own, instead of clobbering a shared global.
   const inferenceApi = {
-    enqueue: (sessionId: string, chatId: string, assistantId: string, user: string) => {
-      // Set the inference context so write tools can access sql + sessionId.
-      // The context persists for the duration of the inference run. Since
-      // BooCoder is single-user and runs one inference at a time per session,
-      // this module-level state is safe.
-      setInferenceContext({ sql, sessionId, taskId: null });
-      inference.enqueue(sessionId, chatId, assistantId, user);
+    enqueue: (
+      sessionId: string,
+      chatId: string,
+      assistantId: string,
+      user: string,
+      permissionMode?: 'plan' | 'ask' | 'bypass',
+    ) => {
+      runWithInferenceContext({ sql, sessionId, taskId: null, permissionMode }, () => {
+        inference.enqueue(sessionId, chatId, assistantId, user);
+      });
     },
     cancel: async (sessionId: string, chatId: string) => {
-      const result = await inference.cancel(sessionId, chatId);
-      clearInferenceContext();
-      return result;
+      // No context to clear — AsyncLocalStorage scopes it to each run's own chain.
+      return inference.cancel(sessionId, chatId);
     },
     hasActive: (chatId: string) => inference.hasActive(chatId),
   };
@@ -220,31 +228,119 @@ async function main() {
 
   // Orchestrator (Phase 2): the flow-runner reacts to the dispatcher's
   // onTaskTerminal hook to advance flow_runs. Created before the dispatcher so its
-  // terminal callback can be wired in. Its launch() is driven by the runs route
-  // (a later phase); resume on startup is a later phase too.
+  // terminal callback can be wired in.
   const flowRunner = createFlowRunner({ sql, broker, log: app.log, config });
 
-  // Phase 4: dispatcher — polls tasks table and runs inference. onTaskTerminal
-  // notifies the flow-runner when a step's task settles (D-2).
+  // Arena SEAM (a): build the local-model set from the live llama-swap model list.
+  // Both bare IDs ('qwen3.6-35b') and prefixed IDs ('llama-swap/qwen3.6-35b') are
+  // included so opencode-style prefixed contestants and native-style bare contestants
+  // both classify correctly as local.
+  const localModelsList = await fetchLlamaSwapModels(config).catch(() => []);
+  const localModels = new Set([
+    ...localModelsList.map((m) => m.id),
+    ...localModelsList.map((m) => `llama-swap/${m.id}`),
+  ]);
+
+  // Arena dispatch function — Phase 4 SEAM (b).
+  // Coding: insert a tasks row with agent=identity (null for native/boocode);
+  //   the dispatcher creates a worktree and runs the external agent (or native).
+  // Q&A: pre-create a session with agent_id stamped to the persona slug so native
+  //   inference loads the persona's system_prompt + tools from AGENTS.md;
+  //   task.session_id is pre-set so runNativeInference reuses the session.
+  const dispatchContestant: DispatchContestantFn = async ({
+    projectId,
+    prompt,
+    identity,
+    model,
+    battleType,
+  }) => {
+    if (battleType === 'qa') {
+      const sessionName = `Arena Q&A [${identity}]: ${prompt.slice(0, 30)}`;
+      const [session] = await sql<{ id: string }[]>`
+        INSERT INTO sessions (project_id, name, model, agent_id, status)
+        VALUES (${projectId}, ${sessionName}, ${model}, ${identity}, 'open')
+        RETURNING id
+      `;
+      const [task] = await sql<{ id: string }[]>`
+        INSERT INTO tasks (project_id, input, model, session_id)
+        VALUES (${projectId}, ${prompt}, ${model}, ${session!.id})
+        RETURNING id
+      `;
+      return { taskId: task!.id, sessionId: session!.id };
+    }
+    // Coding: boocode = native inference (no external agent); any other identity
+    // is an external agent name (claude, opencode, qwen, goose) that maps to
+    // available_agents and gets its own per-task worktree via runExternalAgent.
+    // Session is created lazily by the dispatcher, so sessionId is unknown here.
+    const agentName = identity === 'boocode' ? null : identity;
+    const [task] = await sql<{ id: string }[]>`
+      INSERT INTO tasks (project_id, input, agent, model)
+      VALUES (${projectId}, ${prompt}, ${agentName}, ${model})
+      RETURNING id
+    `;
+    return { taskId: task!.id, sessionId: null };
+  };
+
+  // Arena analyzer: two-stage digest→judge (v1). Pluggable seam — a v2 Han
+  // Orchestrator flow can replace this without schema changes.
+  const analyzer = createAnalyzer({
+    sql,
+    broker,
+    log: app.log,
+    config,
+    localModels,
+  });
+
+  // Arena battle-runner: notified on the same onTaskTerminal hook as the flow-runner.
+  const battleRunner = createBattleRunner({
+    sql,
+    broker,
+    log: app.log,
+    dispatch: dispatchContestant,
+    onBattleComplete: (battleId) => {
+      void analyzer.analyze(battleId);
+    },
+    onCrossExamStart: ({ battleId, crossExamId, identity, model }) => {
+      void analyzer.crossExamine(battleId, crossExamId, { identity, model });
+    },
+    localModels,
+  });
+
+  // Compose onTaskTerminal: both flow-runner and battle-runner are notified.
+  // Each ignores tasks it doesn't own (flow-runner checks flow_steps.task_id;
+  // battle-runner checks contestants.task_id).
+  const onTaskTerminal = (taskId: string, state: string): void => {
+    flowRunner.handleTaskTerminal(taskId, state);
+    battleRunner.handleTaskTerminal(taskId, state);
+  };
+
+  // Phase 4: dispatcher — polls tasks table and runs inference. The composed
+  // onTaskTerminal hook notifies both the flow-runner and the battle-runner when
+  // any task settles.
   const dispatcher = createDispatcher({
     sql,
     inference: inferenceApi,
     broker,
     log: app.log,
     config,
-    onTaskTerminal: flowRunner.handleTaskTerminal,
+    onTaskTerminal,
   });
   dispatcher.start();
 
-  // Phase 5: re-advance any flow_runs that were 'running' when the service last
-  // stopped (D-9). Runs AFTER dispatcher.start() so re-dispatched 'pending' tasks
-  // are picked up by the dispatcher's startup poll.
+  // Re-advance in-flight flow_runs and battles after a coder restart. Both run
+  // AFTER dispatcher.start() so re-dispatched 'pending' tasks are picked up.
   void flowRunner.initResume().catch((err) => {
     app.log.error(
       { err: err instanceof Error ? err.message : String(err) },
       'flow-runner: initResume failed',
     );
   });
+  void battleRunner.initResume().catch((err) => {
+    app.log.error(
+      { err: err instanceof Error ? err.message : String(err) },
+      'arena: initResume failed',
+    );
+  });
 
   // v2.6 Phase 3: configure + start the agent-pool lifecycle sweep (idle-TTL +
   // LRU-cap eviction of warm backends, plus each backend's proactive health probe)
@@ -281,8 +377,8 @@ async function main() {
   registerTaskRoutes(app, sql, inferenceApi, dispatcher.cancelExternalTask);
   registerInboxRoutes(app, sql);
   registerStatsRoutes(app, sql);
-  registerArenaRoutes(app, sql);
   registerRunsRoutes(app, sql, flowRunner, dispatcher.cancelExternalTask);
+  registerArenaRoutes(app, sql, battleRunner, dispatcher.cancelExternalTask, config);
   registerProviderRoutes(app, sql, config);
   registerWorktreeSafetyRoutes(app, sql);
   registerLifecycleRoutes(app, sql);

apps/coder/src/routes/arena.ts

@@ -1,136 +1,412 @@
 /**
- * v2.0.5: Arena routes — competitive dispatch of the same task to multiple agents.
+ * Arena routes — HTTP surface for the Battle UI.
  *
- * POST /api/arena        — create an arena with 2-5 contestants
- * GET  /api/arena/:id    — get all tasks in an arena
- * POST /api/arena/:id/select/:task_id — mark a task as the arena winner
+ * POST /api/battles                         — launch a battle
+ * GET  /api/battles?project_id=             — list battles for a project
+ * GET  /api/battles/:id                     — one battle + contestants + cross-exams
+ * POST /api/battles/:id/stop                — cancel a running battle
+ * POST /api/battles/:id/analyze             — trigger analysis (Phase 5 fills the logic)
+ * POST /api/battles/:id/cross-examine       — start a cross-examination (Phase 5 fills the logic)
+ *
+ * Mirrors the shape of runs.ts (Orchestrator routes). Battle creation delegates to
+ * the battle-runner; cancellation calls cancelBattle then aborts in-flight tasks
+ * via the dispatcher's cancelExternalTask.
  */
 import type { FastifyInstance } from 'fastify';
 import { z } from 'zod';
+import { readFile } from 'node:fs/promises';
+import { join } from 'node:path';
 import type { Sql } from '../db.js';
+import type { Config } from '../config.js';
+import type { BattleRunner } from '../services/arena-runner.js';
+import type { ExternalCancelFn } from './tasks.js';
+import { arenaModelCall } from '../services/arena-model-call.js';
 
-const ContestantSchema = z.object({
-  agent: z.string().max(100).optional(),
-  model: z.string().max(200).optional(),
-  mode_id: z.string().max(200).optional(),
-  thinking_option_id: z.string().max(200).optional(),
+// ─── Validation schemas ───────────────────────────────────────────────────────
+
+const UuidParam = z.string().uuid();
+
+const ContestantInput = z.object({
+  identity: z.string().min(1).max(200),
+  model: z.string().min(1).max(200),
 });
 
-const CreateArenaBody = z.object({
+const CreateBattleBody = z.object({
   project_id: z.string().uuid(),
-  input: z.string().min(1).max(64_000),
-  contestants: z.array(ContestantSchema).min(2).max(5),
+  battle_type: z.enum(['coding', 'qa']),
+  prompt: z.string().min(1).max(64_000),
+  contestants: z
+    .array(ContestantInput)
+    .min(2, 'at least 2 contestants required')
+    .max(6, 'at most 6 contestants allowed'),
 });
 
-interface TaskRow {
-  id: string;
-  agent: string | null;
-  model: string | null;
-  mode_id: string | null;
-  thinking_option_id: string | null;
-  state: string;
-}
+const ListBattlesQuery = z.object({
+  project_id: z.string().uuid(),
+});
 
-export function registerArenaRoutes(app: FastifyInstance, sql: Sql): void {
-  // POST /api/arena — create a new arena
-  app.post('/api/arena', async (req, reply) => {
-    const parsed = CreateArenaBody.safeParse(req.body);
+const CrossExamineBody = z.object({
+  identity: z.string().min(1).max(200),
+  model: z.string().min(1).max(200),
+});
+
+const SetWinnerBody = z.object({
+  winner_contestant_id: z.string().uuid().nullable(),
+});
+
+// ─── Route registration ───────────────────────────────────────────────────────
+
+const GeneratePromptBody = z.object({
+  description: z.string().min(1).max(2_000),
+});
+
+export function registerArenaRoutes(
+  app: FastifyInstance,
+  sql: Sql,
+  battleRunner: BattleRunner,
+  cancelExternal: ExternalCancelFn,
+  config: Config,
+): void {
+
+  // POST /api/battles/generate-prompt — draft a fuller battle prompt from a
+  // short description using the default BooChat model. One-shot, non-streaming.
+  // Must be registered BEFORE /api/battles/:id so the literal 'generate-prompt'
+  // path is not mistaken for a UUID param.
+  app.post('/api/battles/generate-prompt', async (req, reply) => {
+    const parsed = GeneratePromptBody.safeParse(req.body);
     if (!parsed.success) {
       reply.code(400);
       return { error: 'invalid body', details: parsed.error.flatten() };
     }
 
-    const { project_id, input, contestants } = parsed.data;
-    const arenaId = crypto.randomUUID();
+    const { description } = parsed.data;
 
-    const tasks: TaskRow[] = [];
-    for (const contestant of contestants) {
-      const [task] = await sql<TaskRow[]>`
-        INSERT INTO tasks (project_id, input, agent, model, mode_id, thinking_option_id, arena_id)
-        VALUES (
-          ${project_id},
-          ${input},
-          ${contestant.agent ?? null},
-          ${contestant.model ?? null},
-          ${contestant.mode_id ?? null},
-          ${contestant.thinking_option_id ?? null},
-          ${arenaId}
-        )
-        RETURNING id, agent, model, mode_id, thinking_option_id, state
-      `;
-      tasks.push(task!);
+    try {
+      const prompt = await arenaModelCall({
+        config,
+        model: config.DEFAULT_MODEL,
+        system: [
+          'You are a battle-prompt writer for an AI Arena.',
+          'The user gives you a short description of a coding or Q&A challenge.',
+          'Expand it into a clear, self-contained prompt (2–6 sentences) that any AI model can act on.',
+          'Include specific acceptance criteria where helpful.',
+          'Output ONLY the prompt — no preamble, no labels, no meta-commentary.',
+        ].join(' '),
+        user: description,
+        maxTokens: 400,
+        temperature: 0.6,
+      });
+      return { prompt };
+    } catch (err) {
+      app.log.warn(
+        { err: err instanceof Error ? err.message : String(err) },
+        'arena generate-prompt: model call failed',
+      );
+      reply.code(502);
+      return { error: 'model call failed' };
     }
+  });
+
+  // POST /api/battles — launch a battle
+  app.post('/api/battles', async (req, reply) => {
+    const parsed = CreateBattleBody.safeParse(req.body);
+    if (!parsed.success) {
+      reply.code(400);
+      return { error: 'invalid body', details: parsed.error.flatten() };
+    }
+
+    const { project_id, battle_type, prompt, contestants } = parsed.data;
+
+    // Reject duplicate (identity, model) pairs up front — the schema UNIQUE
+    // constraint would catch it too, but an early 422 is friendlier.
+    const seen = new Set<string>();
+    for (const c of contestants) {
+      const key = `${c.identity}::${c.model}`;
+      if (seen.has(key)) {
+        reply.code(422);
+        return {
+          error: 'duplicate_contestant',
+          message: `duplicate contestant: identity="${c.identity}" model="${c.model}"`,
+        };
+      }
+      seen.add(key);
+    }
+
+    // Verify project exists
+    const [proj] = await sql<{ id: string }[]>`SELECT id FROM projects WHERE id = ${project_id}`;
+    if (!proj) {
+      reply.code(404);
+      return { error: 'project not found' };
+    }
+
+    const { battleId } = await battleRunner.startBattle({
+      projectId: project_id,
+      battleType: battle_type,
+      prompt,
+      contestants,
+    });
 
     reply.code(201);
-    return {
-      arena_id: arenaId,
-      tasks: tasks.map((t) => ({
-        id: t.id,
-        agent: t.agent,
-        model: t.model,
-        mode_id: t.mode_id,
-        thinking_option_id: t.thinking_option_id,
-        state: t.state,
-      })),
-    };
+    return { battle_id: battleId };
   });
 
-  // GET /api/arena/:arena_id — list all tasks in an arena
-  app.get<{ Params: { arena_id: string } }>('/api/arena/:arena_id', async (req, reply) => {
-    const { arena_id } = req.params;
-
-    // Validate UUID format
-    const uuidRegex = /^[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}$/i;
-    if (!uuidRegex.test(arena_id)) {
+  // GET /api/battles?project_id= — list battles, most-recent-first
+  app.get('/api/battles', async (req, reply) => {
+    const parsed = ListBattlesQuery.safeParse(req.query);
+    if (!parsed.success) {
       reply.code(400);
-      return { error: 'invalid arena_id format' };
+      return { error: 'invalid query', details: parsed.error.flatten() };
     }
 
-    const tasks = await sql`
-      SELECT id, project_id, state, input, output_summary, agent, model, mode_id, thinking_option_id, execution_path, session_id, started_at, ended_at, created_at, arena_id
-      FROM tasks
-      WHERE arena_id = ${arena_id}
-      ORDER BY created_at
+    const battles = await sql`
+      SELECT id, project_id, battle_type, prompt, status,
+             winner_contestant_id, results_path, error,
+             created_at, updated_at
+      FROM battles
+      WHERE project_id = ${parsed.data.project_id}
+      ORDER BY created_at DESC
+      LIMIT 100
     `;
 
-    if (tasks.length === 0) {
-      reply.code(404);
-      return { error: 'arena not found' };
-    }
-
-    return { arena_id, tasks };
+    return { battles };
   });
 
-  // POST /api/arena/:arena_id/select/:task_id — mark the winner
-  app.post<{ Params: { arena_id: string; task_id: string } }>(
-    '/api/arena/:arena_id/select/:task_id',
-    async (req, reply) => {
-      const { arena_id, task_id } = req.params;
-
-      // Verify the task belongs to this arena
-      const rows = await sql<{ id: string; state: string; arena_id: string | null }[]>`
-        SELECT id, state, arena_id FROM tasks WHERE id = ${task_id}
-      `;
-
-      if (rows.length === 0) {
-        reply.code(404);
-        return { error: 'task not found' };
-      }
-
-      const task = rows[0]!;
-      if (task.arena_id !== arena_id) {
-        reply.code(409);
-        return { error: 'task does not belong to this arena' };
-      }
-
-      // Mark as selected via output_summary prefix (lightweight — no schema change)
-      await sql`
-        UPDATE tasks
-        SET output_summary = COALESCE('[SELECTED] ' || output_summary, '[SELECTED]')
-        WHERE id = ${task_id}
-      `;
-
-      return { selected: true, task_id, arena_id };
+  // GET /api/battles/:id — one battle + its contestants + cross-examinations
+  app.get<{ Params: { id: string } }>('/api/battles/:id', async (req, reply) => {
+    const parsedId = UuidParam.safeParse(req.params.id);
+    if (!parsedId.success) {
+      reply.code(400);
+      return { error: 'invalid id' };
     }
-  );
+    const id = parsedId.data;
+
+    const [battle] = await sql<{
+      id: string;
+      project_id: string;
+      battle_type: string;
+      prompt: string;
+      status: string;
+      winner_contestant_id: string | null;
+      results_path: string | null;
+      error: string | null;
+      created_at: unknown;
+      updated_at: unknown;
+    }[]>`
+      SELECT id, project_id, battle_type, prompt, status,
+             winner_contestant_id, results_path, error,
+             created_at, updated_at
+      FROM battles WHERE id = ${id}
+    `;
+
+    if (!battle) {
+      reply.code(404);
+      return { error: 'battle not found' };
+    }
+
+    const contestants = await sql`
+      SELECT id, battle_id, identity, model, lane, task_id, worktree_id,
+             status, duration_ms, tokens_per_sec, cost_tokens, result_path, error,
+             created_at, updated_at
+      FROM contestants
+      WHERE battle_id = ${id}
+      ORDER BY created_at ASC
+    `;
+
+    const crossExaminations = await sql`
+      SELECT id, battle_id, identity, model, verdict, created_at
+      FROM cross_examinations
+      WHERE battle_id = ${id}
+      ORDER BY created_at ASC
+    `;
+
+    return { battle, contestants, cross_examinations: crossExaminations };
+  });
+
+  // POST /api/battles/:id/stop — cancel a running battle
+  app.post<{ Params: { id: string } }>('/api/battles/:id/stop', async (req, reply) => {
+    const parsedId = UuidParam.safeParse(req.params.id);
+    if (!parsedId.success) {
+      reply.code(400);
+      return { error: 'invalid id' };
+    }
+    const id = parsedId.data;
+
+    const [row] = await sql<{ id: string; status: string }[]>`
+      SELECT id, status FROM battles WHERE id = ${id}
+    `;
+    if (!row) {
+      reply.code(404);
+      return { error: 'battle not found' };
+    }
+    if (row.status !== 'running') {
+      reply.code(409);
+      return { error: `cannot stop battle in status '${row.status}'` };
+    }
+
+    const { cancelled, taskIds } = await battleRunner.cancelBattle(id);
+    if (!cancelled) {
+      reply.code(409);
+      return { error: 'battle is no longer running' };
+    }
+
+    // Abort any in-flight dispatcher tasks (cloud contestants running externally).
+    for (const taskId of taskIds) {
+      cancelExternal(taskId);
+    }
+
+    return { cancelled: true };
+  });
+
+  // GET /api/battles/:id/analysis — read analysis.md from the battle's results_path
+  app.get<{ Params: { id: string } }>('/api/battles/:id/analysis', async (req, reply) => {
+    const parsedId = UuidParam.safeParse(req.params.id);
+    if (!parsedId.success) {
+      reply.code(400);
+      return { error: 'invalid id' };
+    }
+    const id = parsedId.data;
+
+    const [row] = await sql<{ results_path: string | null }[]>`
+      SELECT results_path FROM battles WHERE id = ${id}
+    `;
+    if (!row) {
+      reply.code(404);
+      return { error: 'battle not found' };
+    }
+    if (!row.results_path) {
+      reply.code(404);
+      return { error: 'analysis not ready' };
+    }
+
+    try {
+      const text = await readFile(join(row.results_path, 'analysis.md'), 'utf8');
+      return { text };
+    } catch {
+      reply.code(404);
+      return { error: 'analysis not ready' };
+    }
+  });
+
+  // POST /api/battles/:id/analyze — trigger or re-trigger analysis
+  app.post<{ Params: { id: string } }>('/api/battles/:id/analyze', async (req, reply) => {
+    const parsedId = UuidParam.safeParse(req.params.id);
+    if (!parsedId.success) {
+      reply.code(400);
+      return { error: 'invalid id' };
+    }
+    const id = parsedId.data;
+
+    const [row] = await sql<{ id: string; status: string }[]>`
+      SELECT id, status FROM battles WHERE id = ${id}
+    `;
+    if (!row) {
+      reply.code(404);
+      return { error: 'battle not found' };
+    }
+    if (row.status === 'running') {
+      reply.code(409);
+      return { error: 'battle is still running — wait for all contestants to finish' };
+    }
+
+    const result = await battleRunner.triggerAnalysis(id);
+    if (!result.triggered) {
+      reply.code(404);
+      return { error: 'battle not found' };
+    }
+
+    reply.code(202);
+    return { triggered: true };
+  });
+
+  // PATCH /api/battles/:id/winner — manually set or clear the winner.
+  // Validates the contestant belongs to the battle; publishes battle_updated so
+  // the pane badge reflects the override immediately. Human is authoritative.
+  app.patch<{ Params: { id: string } }>('/api/battles/:id/winner', async (req, reply) => {
+    const parsedId = UuidParam.safeParse(req.params.id);
+    if (!parsedId.success) {
+      reply.code(400);
+      return { error: 'invalid id' };
+    }
+
+    const parsed = SetWinnerBody.safeParse(req.body);
+    if (!parsed.success) {
+      reply.code(400);
+      return { error: 'invalid body', details: parsed.error.flatten() };
+    }
+
+    const result = await battleRunner.setWinner(parsedId.data, parsed.data.winner_contestant_id);
+    if (!result.ok) {
+      if (result.notFound) { reply.code(404); return { error: 'battle not found' }; }
+      if (result.invalidContestant) { reply.code(422); return { error: 'contestant not found in this battle' }; }
+      reply.code(500); return { error: 'unknown error' };
+    }
+    return { ok: true };
+  });
+
+  // GET /api/battles/:id/contestants/:cid/diff — read the diff.patch for a coding contestant.
+  app.get<{ Params: { id: string; cid: string } }>('/api/battles/:id/contestants/:cid/diff', async (req, reply) => {
+    const parsedId = UuidParam.safeParse(req.params.id);
+    const parsedCid = UuidParam.safeParse(req.params.cid);
+    if (!parsedId.success || !parsedCid.success) {
+      reply.code(400);
+      return { error: 'invalid id' };
+    }
+
+    const [contestant] = await sql<{ result_path: string | null }[]>`
+      SELECT result_path FROM contestants
+      WHERE id = ${parsedCid.data} AND battle_id = ${parsedId.data}
+    `;
+    if (!contestant) {
+      reply.code(404);
+      return { error: 'contestant not found' };
+    }
+    if (!contestant.result_path) {
+      reply.code(404);
+      return { error: 'diff not available' };
+    }
+
+    try {
+      const text = await readFile(join(contestant.result_path, 'diff.patch'), 'utf8');
+      return { diff: text };
+    } catch {
+      reply.code(404);
+      return { error: 'diff not available' };
+    }
+  });
+
+  // POST /api/battles/:id/cross-examine — start a cross-examination
+  app.post<{ Params: { id: string } }>('/api/battles/:id/cross-examine', async (req, reply) => {
+    const parsedId = UuidParam.safeParse(req.params.id);
+    if (!parsedId.success) {
+      reply.code(400);
+      return { error: 'invalid id' };
+    }
+    const id = parsedId.data;
+
+    const parsed = CrossExamineBody.safeParse(req.body);
+    if (!parsed.success) {
+      reply.code(400);
+      return { error: 'invalid body', details: parsed.error.flatten() };
+    }
+
+    const [row] = await sql<{ id: string; status: string }[]>`
+      SELECT id, status FROM battles WHERE id = ${id}
+    `;
+    if (!row) {
+      reply.code(404);
+      return { error: 'battle not found' };
+    }
+    if (row.status === 'running') {
+      reply.code(409);
+      return { error: 'battle is still running — cross-examine after all contestants finish' };
+    }
+
+    const { crossExamId } = await battleRunner.startCrossExam(id, {
+      identity: parsed.data.identity,
+      model: parsed.data.model,
+    });
+
+    reply.code(202);
+    return { cross_exam_id: crossExamId };
+  });
 }

apps/coder/src/routes/messages.ts

@@ -4,7 +4,7 @@ import type { Sql } from '../db.js';
 import type { Broker } from '@boocode/server/broker';
 import type { WsFrame } from '@boocode/contracts/ws-frames';
 import { resolveChatId } from './chat-resolve.js';
-import { applyAll } from '../services/pending_changes.js';
+import { asPermissionMode } from '../services/tools/types.js';
 
 const AnswerUserInputBody = z.object({
   tool_call_id: z.string().min(1),
@@ -44,7 +44,13 @@ const SendBody = z.object({
 });
 
 interface InferenceApi {
-  enqueue: (sessionId: string, chatId: string, assistantId: string, user: string) => void;
+  enqueue: (
+    sessionId: string,
+    chatId: string,
+    assistantId: string,
+    user: string,
+    permissionMode?: 'plan' | 'ask' | 'bypass',
+  ) => void;
   cancel: (sessionId: string, chatId: string) => Promise<boolean>;
   hasActive: (chatId: string) => boolean;
 }
@@ -246,36 +252,16 @@ export function registerMessageRoutes(
         RETURNING id
       `;
 
-      inference.enqueue(sessionId, chatId, assistantMsg!.id, 'default');
-
-      // Bypass permission mode (native BooCode): auto-apply staged edits to disk
-      // once the turn settles. `enqueue` registers synchronously, so hasActive is
-      // true immediately; poll until it clears, apply, then re-publish
-      // message_complete so the DiffPanel reflects the now-applied (non-pending)
-      // state. Best-effort — failures stay in the pending queue for manual apply.
-      if (mode_id === 'bypass') {
-        const projectId = sessionRows[0]!.project_id;
-        const assistantId = assistantMsg!.id;
-        void (async () => {
-          try {
-            const [proj] = await sql<{ path: string }[]>`SELECT path FROM projects WHERE id = ${projectId}`;
-            if (!proj?.path) return;
-            for (let i = 0; i < 1200 && inference.hasActive(chatId); i++) {
-              await new Promise((r) => setTimeout(r, 1000));
-            }
-            const applied = await applyAll(sql, sessionId, proj.path);
-            if (applied.length > 0) {
-              broker.publishFrame(sessionId, {
-                type: 'message_complete',
-                message_id: assistantId,
-                chat_id: chatId,
-              } as unknown as WsFrame);
-            }
-          } catch {
-            /* best-effort auto-apply — leave staged changes for manual apply */
-          }
-        })();
-      }
+      // Native BooCode permission gate (plan/ask/bypass) — threaded into the
+      // write-tool context so create/edit/delete and apply_pending honor it.
+      // Plan = read-only, Ask = stage to the queue (agent can't self-apply),
+      // Bypass = apply each write immediately. Other mode ids (e.g. an external
+      // fallback's native mode) leave the gate undefined = legacy behavior.
+      req.log.info(
+        { provider, mode_id, permissionMode: asPermissionMode(mode_id), chatId },
+        'native enqueue — permission gate',
+      );
+      inference.enqueue(sessionId, chatId, assistantMsg!.id, 'default', asPermissionMode(mode_id));
 
       reply.code(202);
       return { user_message_id: userMsg!.id, assistant_message_id: assistantMsg!.id };

apps/coder/src/schema.sql

@@ -54,9 +54,6 @@ DO $$ BEGIN
   END IF;
 END $$;
 
--- v2.0.5: arena support — group tasks into competitive arenas.
-ALTER TABLE tasks ADD COLUMN IF NOT EXISTS arena_id UUID;
-
 -- Human inbox: tasks needing attention
 CREATE OR REPLACE VIEW human_inbox AS
   SELECT * FROM tasks WHERE state IN ('blocked', 'failed');
@@ -81,6 +78,7 @@ ALTER TABLE tasks ADD COLUMN IF NOT EXISTS thinking_option_id TEXT;
 DROP VIEW IF EXISTS human_inbox;
 ALTER TABLE tasks DROP COLUMN IF EXISTS feature_values;
 ALTER TABLE tasks DROP COLUMN IF EXISTS worktree_path;
+ALTER TABLE tasks DROP COLUMN IF EXISTS arena_id;
 CREATE OR REPLACE VIEW human_inbox AS
   SELECT * FROM tasks WHERE state IN ('blocked', 'failed');
 
@@ -157,7 +155,7 @@ CREATE UNIQUE INDEX IF NOT EXISTS worktrees_active_path_uidx ON worktrees(path)
 DROP TABLE IF EXISTS session_worktrees;
 
 -- Dispatch hint: which chat (tab) a task belongs to. The coder message route and
--- skills route set it from the frontend tab; session-less creators (arena, MCP,
+-- skills route set it from the frontend tab; session-less creators (MCP,
 -- new_task, generic /api/tasks) leave it NULL and the dispatcher creates a chat.
 ALTER TABLE tasks ADD COLUMN IF NOT EXISTS chat_id UUID REFERENCES chats(id) ON DELETE SET NULL;
 
@@ -271,7 +269,7 @@ ALTER TABLE agent_sessions ADD CONSTRAINT agent_sessions_backend_chk
   CHECK (backend IN ('opencode_server', 'acp_warm', 'claude_sdk'));
 
 -- LISTEN/NOTIFY fast path: every tasks INSERT (from any call site — routes,
--- new_task tool, arena, MCP server) fires pg_notify('tasks_new') in the same
+-- new_task tool, MCP server) fires pg_notify('tasks_new') in the same
 -- transaction, so the dispatcher reacts immediately instead of waiting for the
 -- fallback poll. Postgres holds the notification until COMMIT, so the listener
 -- always sees the committed row. A trigger covers all insert paths with no
@@ -357,3 +355,71 @@ DO $$ BEGIN
       CHECK (status IN ('pending', 'running', 'completed', 'failed', 'skipped', 'cancelled'));
   END IF;
 END $$;
+
+-- Arena: battles + contestants + cross_examinations.
+-- project_id carries no FK (matches tasks.project_id + flow_runs.project_id convention).
+-- winner_contestant_id FK is deferred (forward reference): added via guarded ALTER below.
+CREATE TABLE IF NOT EXISTS battles (
+  id                   UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+  project_id           UUID NOT NULL,
+  battle_type          TEXT NOT NULL,
+  prompt               TEXT NOT NULL,
+  status               TEXT NOT NULL DEFAULT 'pending',
+  winner_contestant_id UUID,
+  results_path         TEXT,
+  error                TEXT,
+  created_at           TIMESTAMPTZ NOT NULL DEFAULT clock_timestamp(),
+  updated_at           TIMESTAMPTZ NOT NULL DEFAULT clock_timestamp(),
+  CONSTRAINT battles_type_chk   CHECK (battle_type IN ('coding', 'qa')),
+  CONSTRAINT battles_status_chk CHECK (status IN ('pending', 'running', 'completed', 'failed', 'cancelled'))
+);
+
+CREATE TABLE IF NOT EXISTS contestants (
+  id             UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+  battle_id      UUID NOT NULL REFERENCES battles(id) ON DELETE CASCADE,
+  identity       TEXT NOT NULL,
+  model          TEXT NOT NULL,
+  lane           TEXT NOT NULL,
+  task_id        UUID REFERENCES tasks(id) ON DELETE SET NULL,
+  worktree_id    UUID REFERENCES worktrees(id) ON DELETE SET NULL,
+  status         TEXT NOT NULL DEFAULT 'queued',
+  duration_ms    INTEGER,
+  tokens_per_sec DOUBLE PRECISION,
+  cost_tokens    INTEGER,
+  result_path    TEXT,
+  error          TEXT,
+  created_at     TIMESTAMPTZ NOT NULL DEFAULT clock_timestamp(),
+  updated_at     TIMESTAMPTZ NOT NULL DEFAULT clock_timestamp(),
+  CONSTRAINT contestants_lane_chk   CHECK (lane   IN ('local', 'cloud')),
+  CONSTRAINT contestants_status_chk CHECK (status IN ('queued', 'running', 'done', 'error')),
+  UNIQUE (battle_id, identity, model)
+);
+
+CREATE TABLE IF NOT EXISTS cross_examinations (
+  id         UUID PRIMARY KEY DEFAULT gen_random_uuid(),
+  battle_id  UUID NOT NULL REFERENCES battles(id) ON DELETE CASCADE,
+  identity   TEXT NOT NULL,
+  model      TEXT NOT NULL,
+  verdict    TEXT,
+  created_at TIMESTAMPTZ NOT NULL DEFAULT clock_timestamp()
+);
+
+-- Add the winner FK now that contestants exists.
+DO $$ BEGIN
+  IF NOT EXISTS (SELECT 1 FROM pg_constraint WHERE conname = 'battles_winner_contestant_id_fkey') THEN
+    ALTER TABLE battles ADD CONSTRAINT battles_winner_contestant_id_fkey
+      FOREIGN KEY (winner_contestant_id) REFERENCES contestants(id) ON DELETE SET NULL;
+  END IF;
+END $$;
+
+-- battles query (GET /api/battles?project_id=).
+CREATE INDEX IF NOT EXISTS battles_project_created_idx ON battles(project_id, created_at DESC);
+
+-- Lane-scheduler advance scans (contestants WHERE battle_id = ? AND status = ?).
+CREATE INDEX IF NOT EXISTS contestants_battle_status_idx ON contestants(battle_id, status);
+
+-- onTaskTerminal callback: look up the contestant owning a completed task.
+CREATE INDEX IF NOT EXISTS contestants_task_id_idx ON contestants(task_id);
+
+-- Cross-examination listing per battle.
+CREATE INDEX IF NOT EXISTS cross_examinations_battle_idx ON cross_examinations(battle_id);

apps/coder/src/services/__tests__/arena-analyzer-helpers.test.ts

@@ -0,0 +1,254 @@
+import { describe, it, expect } from 'vitest';
+import {
+  buildDigestPrompt,
+  buildJudgePrompt,
+  buildCrossExamPrompt,
+  extractWinner,
+  shouldNameWinner,
+  type ContestantDigest,
+  type ContestantDigestInput,
+} from '../arena-analyzer-helpers.js';
+
+// ─── shouldNameWinner ─────────────────────────────────────────────────────────
+
+describe('shouldNameWinner', () => {
+  it('returns false with 0 succeeded contestants', () => {
+    expect(shouldNameWinner(0)).toBe(false);
+  });
+
+  it('returns false with exactly 1 succeeded contestant', () => {
+    expect(shouldNameWinner(1)).toBe(false);
+  });
+
+  it('returns true with exactly 2 succeeded contestants', () => {
+    expect(shouldNameWinner(2)).toBe(true);
+  });
+
+  it('returns true with more than 2 succeeded contestants', () => {
+    expect(shouldNameWinner(3)).toBe(true);
+    expect(shouldNameWinner(6)).toBe(true);
+  });
+});
+
+// ─── extractWinner ────────────────────────────────────────────────────────────
+
+describe('extractWinner', () => {
+  it('extracts identity and model from a WINNER: line', () => {
+    const output = 'Some analysis\n\nWINNER: claude/opus-4-5\n\nMore text.';
+    expect(extractWinner(output)).toEqual({ identity: 'claude', model: 'opus-4-5' });
+  });
+
+  it('is case-insensitive for the WINNER keyword', () => {
+    expect(extractWinner('winner: boocode/qwen3.6-35b')).toEqual({
+      identity: 'boocode',
+      model: 'qwen3.6-35b',
+    });
+    expect(extractWinner('Winner: opencode/some-model')).toEqual({
+      identity: 'opencode',
+      model: 'some-model',
+    });
+  });
+
+  it('returns null when NO_WINNER is declared', () => {
+    expect(extractWinner('WINNER: NO_WINNER')).toBeNull();
+    expect(extractWinner('winner: no_winner')).toBeNull();
+  });
+
+  it('returns null when no WINNER line is present', () => {
+    expect(extractWinner('Just some analysis text with no verdict.')).toBeNull();
+    expect(extractWinner('')).toBeNull();
+  });
+
+  it('returns null when the WINNER line has no slash separator', () => {
+    expect(extractWinner('WINNER: justidentity')).toBeNull();
+  });
+
+  it('returns null when the WINNER line is empty after the colon', () => {
+    expect(extractWinner('WINNER:')).toBeNull();
+    expect(extractWinner('WINNER:   ')).toBeNull();
+  });
+
+  it('handles leading and trailing whitespace around the slash parts', () => {
+    const result = extractWinner('WINNER:  claude / opus-4-5 ');
+    expect(result).toEqual({ identity: 'claude', model: 'opus-4-5' });
+  });
+
+  it('picks the first WINNER line when multiple are present', () => {
+    const output = 'WINNER: claude/opus-4-5\nWINNER: opencode/other-model';
+    expect(extractWinner(output)).toEqual({ identity: 'claude', model: 'opus-4-5' });
+  });
+
+  it('handles model names that contain slashes by splitting at the first slash only', () => {
+    // edge case: model name with a slash — should still split at first slash
+    // identity = 'native', model = 'llama-swap/qwen3.6'
+    const result = extractWinner('WINNER: native/llama-swap/qwen3.6');
+    expect(result).toEqual({ identity: 'native', model: 'llama-swap/qwen3.6' });
+  });
+});
+
+// ─── buildDigestPrompt ────────────────────────────────────────────────────────
+
+describe('buildDigestPrompt', () => {
+  const base: ContestantDigestInput = {
+    identity: 'claude',
+    model: 'opus-4-5',
+    resultMd: '# Output\n\nSome result content.',
+    benchmarkLine: '12000ms',
+  };
+
+  it('returns an object with non-empty system and user strings', () => {
+    const { system, user } = buildDigestPrompt(base);
+    expect(system.length).toBeGreaterThan(0);
+    expect(user.length).toBeGreaterThan(0);
+  });
+
+  it('includes the contestant identity and model in the user prompt', () => {
+    const { user } = buildDigestPrompt(base);
+    expect(user).toContain('claude');
+    expect(user).toContain('opus-4-5');
+  });
+
+  it('includes the benchmark line in the user prompt', () => {
+    const { user } = buildDigestPrompt(base);
+    expect(user).toContain('12000ms');
+  });
+
+  it('includes the result.md content in the user prompt', () => {
+    const { user } = buildDigestPrompt(base);
+    expect(user).toContain('Some result content.');
+  });
+
+  it('includes the diff.patch when provided', () => {
+    const input: ContestantDigestInput = { ...base, diffPatch: '--- a/foo.ts\n+++ b/foo.ts\n+added' };
+    const { user } = buildDigestPrompt(input);
+    expect(user).toContain('added');
+    expect(user).toContain('```diff');
+  });
+
+  it('omits the diff section when diffPatch is undefined', () => {
+    const { user } = buildDigestPrompt(base);
+    expect(user).not.toContain('```diff');
+  });
+
+  it('truncates resultMd longer than 8000 characters', () => {
+    const longResult = 'x'.repeat(10_000);
+    const { user } = buildDigestPrompt({ ...base, resultMd: longResult });
+    // The truncated content must not exceed 8000 chars in the sliced section.
+    // We just check the total user string doesn't balloon unreasonably.
+    expect(user.length).toBeLessThan(15_000);
+  });
+
+  it('truncates diffPatch longer than 5000 characters', () => {
+    const longDiff = '+' + 'x'.repeat(10_000);
+    const { user } = buildDigestPrompt({ ...base, diffPatch: longDiff });
+    expect(user.length).toBeLessThan(16_000);
+  });
+});
+
+// ─── buildJudgePrompt ─────────────────────────────────────────────────────────
+
+describe('buildJudgePrompt', () => {
+  const digests: ContestantDigest[] = [
+    { identity: 'claude', model: 'opus-4-5', digest: 'Good result.', benchmarkLine: '5000ms' },
+    { identity: 'opencode', model: 'qwen3.6', digest: 'Decent result.', benchmarkLine: '8000ms' },
+  ];
+
+  it('includes the original prompt in the user section', () => {
+    const { user } = buildJudgePrompt('Write a sorting algorithm', digests);
+    expect(user).toContain('Write a sorting algorithm');
+  });
+
+  it('includes each contestant heading in the user section', () => {
+    const { user } = buildJudgePrompt('prompt', digests);
+    expect(user).toContain('claude');
+    expect(user).toContain('opus-4-5');
+    expect(user).toContain('opencode');
+    expect(user).toContain('qwen3.6');
+  });
+
+  it('includes each contestant digest text', () => {
+    const { user } = buildJudgePrompt('prompt', digests);
+    expect(user).toContain('Good result.');
+    expect(user).toContain('Decent result.');
+  });
+
+  it('instructs the model to name a WINNER when 2+ digests are provided', () => {
+    const { system } = buildJudgePrompt('prompt', digests);
+    expect(system).toContain('WINNER:');
+  });
+
+  it('instructs the model NOT to name a winner when fewer than 2 digests are provided', () => {
+    const oneDigest = digests.slice(0, 1);
+    const { system } = buildJudgePrompt('prompt', oneDigest);
+    expect(system).toContain('NO_WINNER');
+    expect(system).not.toContain('WINNER: <identity>');
+  });
+
+  it('instructs NO_WINNER when digests list is empty', () => {
+    const { system } = buildJudgePrompt('prompt', []);
+    expect(system).toContain('NO_WINNER');
+  });
+
+  it('truncates originalPrompt longer than 2000 characters', () => {
+    const longPrompt = 'p'.repeat(5_000);
+    const { user } = buildJudgePrompt(longPrompt, digests);
+    // Should not contain more than 2000 chars of the prompt.
+    const promptSection = user.split('# Contestant Digests')[0] ?? '';
+    expect(promptSection.length).toBeLessThan(3_000);
+  });
+});
+
+// ─── buildCrossExamPrompt ─────────────────────────────────────────────────────
+
+describe('buildCrossExamPrompt', () => {
+  const digests: ContestantDigest[] = [
+    { identity: 'claude', model: 'opus-4-5', digest: 'Strong result.', benchmarkLine: '5000ms' },
+    { identity: 'boocode', model: 'qwen3.6-35b', digest: 'Decent result.', benchmarkLine: '12000ms' },
+  ];
+
+  const baseOpts = {
+    originalPrompt: 'Write a sorting algorithm.',
+    digests,
+    analysisContent: '# Arena Analysis\n\nClaude did better.\n\nWINNER: claude/opus-4-5',
+    proposedWinner: 'claude/opus-4-5',
+    examinerIdentity: 'goose',
+    examinerModel: 'gpt-4o',
+  };
+
+  it('includes the examiner identity and model in the system prompt', () => {
+    const { system } = buildCrossExamPrompt(baseOpts);
+    expect(system).toContain('goose');
+    expect(system).toContain('gpt-4o');
+  });
+
+  it('includes the original prompt in the user section', () => {
+    const { user } = buildCrossExamPrompt(baseOpts);
+    expect(user).toContain('Write a sorting algorithm.');
+  });
+
+  it('includes each contestant digest', () => {
+    const { user } = buildCrossExamPrompt(baseOpts);
+    expect(user).toContain('Strong result.');
+    expect(user).toContain('Decent result.');
+  });
+
+  it('includes the proposed analysis content', () => {
+    const { user } = buildCrossExamPrompt(baseOpts);
+    expect(user).toContain('Claude did better.');
+  });
+
+  it('includes the proposed winner when set', () => {
+    const { user } = buildCrossExamPrompt(baseOpts);
+    expect(user).toContain('claude/opus-4-5');
+  });
+
+  it('notes that no winner was proposed when proposedWinner is null', () => {
+    const { user } = buildCrossExamPrompt({ ...baseOpts, proposedWinner: null });
+    expect(user).toContain('No winner was proposed');
+  });
+
+  it('instructs the examiner to provide a VERDICT line', () => {
+    const { system } = buildCrossExamPrompt(baseOpts);
+    expect(system).toContain('VERDICT:');
+  });
+});

apps/coder/src/services/__tests__/arena-decisions.test.ts

@@ -0,0 +1,332 @@
+import { describe, it, expect } from 'vitest';
+import {
+  classifyLane,
+  nextLocalContestant,
+  isBattleComplete,
+  computeBenchmark,
+  sanitizeSlug,
+  buildBattleSlug,
+  buildContestantDir,
+  reconcileContestantResume,
+  reconcileContestants,
+  type ContestantSlot,
+} from '../arena-decisions.js';
+
+// Local models = what the llama-swap server actually serves.
+const LOCAL_MODELS: ReadonlySet<string> = new Set([
+  'qwen3.6-35b-a3b-mxfp4',
+  'qwen2.5-coder-7b',
+]);
+
+// ─── classifyLane ────────────────────────────────────────────────────────────
+
+describe('classifyLane', () => {
+  it('classifies qa battles as local regardless of identity or model', () => {
+    expect(classifyLane('qa', 'boocode', 'qwen3.6-35b-a3b-mxfp4', LOCAL_MODELS)).toBe('local');
+    expect(classifyLane('qa', 'claude', 'claude-opus-4-5', LOCAL_MODELS)).toBe('local');
+    expect(classifyLane('qa', 'Debugger', 'cloud-model', new Set())).toBe('local');
+    expect(classifyLane('qa', 'opencode', 'any-model', LOCAL_MODELS)).toBe('local');
+  });
+
+  it('classifies coding contestants as local when model is in localModels', () => {
+    expect(classifyLane('coding', 'boocode', 'qwen3.6-35b-a3b-mxfp4', LOCAL_MODELS)).toBe('local');
+    expect(classifyLane('coding', 'opencode', 'qwen3.6-35b-a3b-mxfp4', LOCAL_MODELS)).toBe('local');
+    expect(classifyLane('coding', 'qwen', 'qwen2.5-coder-7b', LOCAL_MODELS)).toBe('local');
+  });
+
+  it('classifies coding contestants as cloud when model is not in localModels', () => {
+    expect(classifyLane('coding', 'claude', 'claude-opus-4-5', LOCAL_MODELS)).toBe('cloud');
+    expect(classifyLane('coding', 'opencode', 'claude-opus-4-5', LOCAL_MODELS)).toBe('cloud');
+    expect(classifyLane('coding', 'goose', 'gpt-4o', LOCAL_MODELS)).toBe('cloud');
+    expect(classifyLane('coding', 'qwen', 'unknown-remote-model', LOCAL_MODELS)).toBe('cloud');
+  });
+
+  it('uses the injected localModels set, not a hardcoded list', () => {
+    const custom = new Set(['my-local-model']);
+    expect(classifyLane('coding', 'any-agent', 'my-local-model', custom)).toBe('local');
+    expect(classifyLane('coding', 'boocode', 'other-model', custom)).toBe('cloud');
+  });
+
+  it('defaults to cloud for an empty localModels set', () => {
+    expect(classifyLane('coding', 'boocode', 'qwen3.6-35b-a3b-mxfp4', new Set())).toBe('cloud');
+    expect(classifyLane('coding', 'native', 'any-local-model', new Set())).toBe('cloud');
+  });
+});
+
+// ─── nextLocalContestant ─────────────────────────────────────────────────────
+
+describe('nextLocalContestant', () => {
+  it('returns null for an empty list', () => {
+    expect(nextLocalContestant([])).toBeNull();
+  });
+
+  it('returns null when no local contestants are queued', () => {
+    const slots: ContestantSlot[] = [
+      { id: 'c1', lane: 'local', status: 'running' },
+      { id: 'c2', lane: 'cloud', status: 'queued' },
+    ];
+    expect(nextLocalContestant(slots)).toBeNull();
+  });
+
+  it('returns the first queued local contestant in order', () => {
+    const slots: ContestantSlot[] = [
+      { id: 'c1', lane: 'local', status: 'done' },
+      { id: 'c2', lane: 'local', status: 'queued' },
+      { id: 'c3', lane: 'local', status: 'queued' },
+    ];
+    expect(nextLocalContestant(slots)).toBe('c2');
+  });
+
+  it('skips done/error local contestants and cloud contestants', () => {
+    const slots: ContestantSlot[] = [
+      { id: 'c1', lane: 'cloud', status: 'queued' },
+      { id: 'c2', lane: 'local', status: 'error' },
+      { id: 'c3', lane: 'local', status: 'queued' },
+    ];
+    expect(nextLocalContestant(slots)).toBe('c3');
+  });
+
+  it('returns null when all local contestants are done or error', () => {
+    const slots: ContestantSlot[] = [
+      { id: 'c1', lane: 'local', status: 'done' },
+      { id: 'c2', lane: 'local', status: 'error' },
+    ];
+    expect(nextLocalContestant(slots)).toBeNull();
+  });
+});
+
+// ─── isBattleComplete ────────────────────────────────────────────────────────
+
+describe('isBattleComplete', () => {
+  it('returns false for an empty list', () => {
+    expect(isBattleComplete([])).toBe(false);
+  });
+
+  it('returns true when all contestants are done', () => {
+    expect(isBattleComplete([{ status: 'done' }, { status: 'done' }])).toBe(true);
+  });
+
+  it('returns true when all contestants are error', () => {
+    expect(isBattleComplete([{ status: 'error' }, { status: 'error' }])).toBe(true);
+  });
+
+  it('returns true for a mixed done/error result', () => {
+    expect(isBattleComplete([{ status: 'done' }, { status: 'error' }, { status: 'done' }])).toBe(true);
+  });
+
+  it('returns false while any contestant is still running', () => {
+    expect(isBattleComplete([{ status: 'done' }, { status: 'running' }])).toBe(false);
+  });
+
+  it('returns false while any contestant is still queued', () => {
+    expect(isBattleComplete([{ status: 'done' }, { status: 'queued' }])).toBe(false);
+  });
+});
+
+// ─── computeBenchmark ────────────────────────────────────────────────────────
+
+describe('computeBenchmark', () => {
+  const t0 = new Date('2026-06-06T10:00:00.000Z');
+  const t1 = new Date('2026-06-06T10:00:05.000Z'); // +5 000ms
+
+  it('computes duration in ms for both lanes', () => {
+    const local = computeBenchmark(t0, t1, 100, 'local');
+    expect(local.durationMs).toBe(5000);
+    const cloud = computeBenchmark(t0, t1, null, 'cloud');
+    expect(cloud.durationMs).toBe(5000);
+  });
+
+  it('computes tokens/sec for local lane when costTokens is known', () => {
+    const bench = computeBenchmark(t0, t1, 500, 'local');
+    expect(bench.tokensPerSec).toBeCloseTo(100, 5); // 500 / 5 = 100 tok/s
+  });
+
+  it('omits tokens/sec for cloud lane regardless of costTokens', () => {
+    const bench = computeBenchmark(t0, t1, 500, 'cloud');
+    expect(bench.tokensPerSec).toBeNull();
+  });
+
+  it('omits tokens/sec for local lane when costTokens is null', () => {
+    const bench = computeBenchmark(t0, t1, null, 'local');
+    expect(bench.tokensPerSec).toBeNull();
+  });
+
+  it('returns durationMs = 0 and null tokensPerSec when timestamps are equal', () => {
+    const bench = computeBenchmark(t0, t0, 100, 'local');
+    expect(bench.durationMs).toBe(0);
+    expect(bench.tokensPerSec).toBeNull();
+  });
+
+  it('clamps negative duration to 0 (clock skew)', () => {
+    const bench = computeBenchmark(t1, t0, 50, 'local');
+    expect(bench.durationMs).toBe(0);
+    expect(bench.tokensPerSec).toBeNull();
+  });
+});
+
+// ─── sanitizeSlug ────────────────────────────────────────────────────────────
+
+describe('sanitizeSlug', () => {
+  it('lowercases and preserves alphanumeric + hyphens', () => {
+    expect(sanitizeSlug('claude')).toBe('claude');
+    expect(sanitizeSlug('claude-opus-4-5')).toBe('claude-opus-4-5');
+  });
+
+  it('replaces spaces and special characters with hyphens', () => {
+    expect(sanitizeSlug('Code Reviewer')).toBe('code-reviewer');
+    expect(sanitizeSlug('native/boocode')).toBe('native-boocode');
+    expect(sanitizeSlug('qwen2.5-coder-35b')).toBe('qwen2-5-coder-35b');
+  });
+
+  it('collapses consecutive non-alphanumeric runs to a single hyphen', () => {
+    expect(sanitizeSlug('foo  bar---baz')).toBe('foo-bar-baz');
+  });
+
+  it('strips leading and trailing hyphens', () => {
+    expect(sanitizeSlug('---foo---')).toBe('foo');
+  });
+
+  it('truncates to 64 characters', () => {
+    const long = 'a'.repeat(100);
+    expect(sanitizeSlug(long).length).toBe(64);
+  });
+});
+
+// ─── buildBattleSlug ─────────────────────────────────────────────────────────
+
+describe('buildBattleSlug', () => {
+  it('builds a deterministic dated slug from id, type, and createdAt', () => {
+    const id = 'a1b2c3d4-e5f6-7890-abcd-ef1234567890';
+    const createdAt = new Date('2026-06-06T12:00:00.000Z');
+    const slug = buildBattleSlug(id, 'coding', createdAt);
+    expect(slug).toBe('2026-06-06-coding-a1b2c3d4');
+  });
+
+  it('includes the battle type in the slug', () => {
+    const id = 'aaaaaaaa-0000-0000-0000-000000000000';
+    const createdAt = new Date('2026-01-01T00:00:00.000Z');
+    expect(buildBattleSlug(id, 'qa', createdAt)).toContain('-qa-');
+    expect(buildBattleSlug(id, 'coding', createdAt)).toContain('-coding-');
+  });
+
+  it('uses the first 8 hex chars of the uuid (dashes stripped)', () => {
+    const id = 'deadbeef-0000-0000-0000-000000000000';
+    const slug = buildBattleSlug(id, 'coding', new Date('2026-06-06T00:00:00Z'));
+    expect(slug.endsWith('-deadbeef')).toBe(true);
+  });
+});
+
+// ─── buildContestantDir ──────────────────────────────────────────────────────
+
+describe('buildContestantDir', () => {
+  it('joins sanitized identity and model with a hyphen', () => {
+    expect(buildContestantDir('claude', 'claude-opus-4-5')).toBe('claude-claude-opus-4-5');
+  });
+
+  it('sanitizes both parts independently', () => {
+    expect(buildContestantDir('Code Reviewer', 'qwen2.5-35b')).toBe('code-reviewer-qwen2-5-35b');
+  });
+});
+
+// ─── reconcileContestantResume ───────────────────────────────────────────────
+
+describe('reconcileContestantResume', () => {
+  it('keeps non-running contestants regardless of task state', () => {
+    for (const status of ['queued', 'done', 'error']) {
+      expect(reconcileContestantResume(status, 'tid', 'completed')).toBe('keep');
+      expect(reconcileContestantResume(status, null, null)).toBe('keep');
+    }
+  });
+
+  it('re-dispatches a running contestant with no task_id', () => {
+    expect(reconcileContestantResume('running', null, null)).toBe('re-dispatch');
+  });
+
+  it('re-dispatches a running contestant whose task row is absent', () => {
+    expect(reconcileContestantResume('running', 'tid', null)).toBe('re-dispatch');
+  });
+
+  it('marks done when the task completed before the terminal callback ran', () => {
+    expect(reconcileContestantResume('running', 'tid', 'completed')).toBe('mark-done');
+  });
+
+  it('marks error when the task failed', () => {
+    expect(reconcileContestantResume('running', 'tid', 'failed')).toBe('mark-error');
+  });
+
+  it('marks cancelled when the task was cancelled', () => {
+    expect(reconcileContestantResume('running', 'tid', 'cancelled')).toBe('mark-cancelled');
+  });
+
+  it('keeps a running contestant whose task is pending (dispatcher handles it)', () => {
+    expect(reconcileContestantResume('running', 'tid', 'pending')).toBe('keep');
+  });
+
+  it('re-dispatches when the task is stuck running (process died)', () => {
+    expect(reconcileContestantResume('running', 'tid', 'running')).toBe('re-dispatch');
+  });
+
+  it('re-dispatches when the task is blocked (permission dialog gone on restart)', () => {
+    expect(reconcileContestantResume('running', 'tid', 'blocked')).toBe('re-dispatch');
+  });
+});
+
+// ─── reconcileContestants ────────────────────────────────────────────────────
+
+describe('reconcileContestants', () => {
+  it('returns one decision per contestant', () => {
+    const contestants = [
+      { contestantId: 'c1', taskId: null, status: 'done' },
+      { contestantId: 'c2', taskId: 't1', status: 'running' },
+      { contestantId: 'c3', taskId: 't2', status: 'running' },
+    ];
+    const taskStates = new Map([['t1', 'completed'], ['t2', 'running']]);
+    const decisions = reconcileContestants(contestants, taskStates);
+    expect(decisions).toHaveLength(3);
+    expect(decisions[0]).toEqual({ contestantId: 'c1', action: 'keep' });
+    expect(decisions[1]).toEqual({ contestantId: 'c2', action: 'mark-done' });
+    expect(decisions[2]).toEqual({ contestantId: 'c3', action: 're-dispatch' });
+  });
+
+  it('re-dispatches a running contestant whose taskId is absent from taskStates', () => {
+    const contestants = [{ contestantId: 'c1', taskId: 'orphan', status: 'running' }];
+    const decisions = reconcileContestants(contestants, new Map());
+    expect(decisions[0]?.action).toBe('re-dispatch');
+  });
+
+  it('re-dispatches a running contestant with null taskId', () => {
+    const contestants = [{ contestantId: 'c1', taskId: null, status: 'running' }];
+    const decisions = reconcileContestants(contestants, new Map());
+    expect(decisions[0]?.action).toBe('re-dispatch');
+  });
+
+  it('returns empty array for no contestants', () => {
+    expect(reconcileContestants([], new Map())).toEqual([]);
+  });
+
+  it('keeps a running contestant whose task is pending', () => {
+    const contestants = [{ contestantId: 'c1', taskId: 't1', status: 'running' }];
+    const taskStates = new Map([['t1', 'pending']]);
+    const decisions = reconcileContestants(contestants, taskStates);
+    expect(decisions[0]?.action).toBe('keep');
+  });
+
+  it('handles a mixed battle: done/queued kept, stale running re-dispatched', () => {
+    const contestants = [
+      { contestantId: 'c1', taskId: 't1', status: 'done' },
+      { contestantId: 'c2', taskId: null, status: 'queued' },
+      { contestantId: 'c3', taskId: 't2', status: 'running' },
+      { contestantId: 'c4', taskId: 't3', status: 'running' },
+    ];
+    const taskStates = new Map([
+      ['t1', 'completed'],
+      ['t2', 'running'],  // stuck — process dead
+      ['t3', 'pending'],  // dispatcher will handle
+    ]);
+    const decisions = reconcileContestants(contestants, taskStates);
+    expect(decisions.find((d) => d.contestantId === 'c1')?.action).toBe('keep');
+    expect(decisions.find((d) => d.contestantId === 'c2')?.action).toBe('keep');
+    expect(decisions.find((d) => d.contestantId === 'c3')?.action).toBe('re-dispatch');
+    expect(decisions.find((d) => d.contestantId === 'c4')?.action).toBe('keep');
+  });
+});

apps/coder/src/services/__tests__/fuzzy-match.test.ts

@@ -161,6 +161,52 @@ describe('locateMatch — strategy 4: Levenshtein', () => {
   });
 });
 
+describe('locateMatch — strategy 4: fail-closed on ambiguity (corruption guard)', () => {
+  it('refuses (ambiguous) when two equally-similar anchored blocks both clear the bar', () => {
+    // The repetitive-file case that duplicated blocks: two blocks share the same
+    // first+last anchor lines and their middle lines are EQUALLY similar to the
+    // (drifted) needle. Tier 4 must refuse rather than splice over one of them.
+    const content = [
+      'const x = {',
+      '  total = aa;',
+      '};',
+      'const x = {',
+      '  total = bb;',
+      '};',
+    ].join('\n');
+    const needle = ['const x = {', '  total = ab;', '};'].join('\n');
+    const result = locateMatch(content, needle);
+    expect(result.kind).toBe('ambiguous');
+  });
+
+  it('refuses a below-threshold near-miss that the old 0.66 floor would have spliced', () => {
+    // ~0.7 similar: under the raised 0.85 floor this is now not_found, so the
+    // caller surfaces a correctable error instead of corrupting the file.
+    const content = 'const grandTotalAmount = a + b;\n';
+    const needle = 'const totalValue = a + b;';
+    const result = locateMatch(content, needle);
+    expect(result).toEqual({ kind: 'not_found' });
+  });
+
+  it('still matches a single genuine high-similarity drift uniquely', () => {
+    const content = 'const total = sum + tax;\n';
+    const needle = 'const totals = sum + tax;'; // one-char typo, ~0.96
+    const result = locateMatch(content, needle);
+    expect(result.kind).toBe('fuzzy');
+    const { start, end } = span(result);
+    expect(content.slice(start, end)).toBe('const total = sum + tax;');
+  });
+
+  it('requires an exact first+last line anchor for multi-line needles', () => {
+    // First line drifted too far to anchor → no window is scored → not_found,
+    // even though the middle lines are identical.
+    const content = ['function compute() {', '  return a + b;', '  return done;', '}'].join('\n');
+    const needle = ['totally different opener', '  return a + b;', '}'].join('\n');
+    const result = locateMatch(content, needle);
+    expect(result).toEqual({ kind: 'not_found' });
+  });
+});
+
 describe('locateMatch — edge cases', () => {
   it('returns not_found for an empty needle', () => {
     expect(locateMatch('anything', '')).toEqual({ kind: 'not_found' });

apps/coder/src/services/__tests__/pending_changes_integration.test.ts

@@ -83,6 +83,53 @@ describe.runIf(!!process.env.DATABASE_URL)('pending_changes integration', () =>
     expect(existsSync(resolve(testDir, 'deleteme.txt'))).toBe(false);
   });
 
+  it('re-emitted identical edits dedupe at queue and never duplicate on apply', async () => {
+    // Regression: the 2-3x block-stamping corruption. An anchored insert queued
+    // three times (a local model re-emitting the same tool call) must collapse to
+    // ONE pending row and apply exactly once.
+    await queueCreate(sql, testSessionId, null, 'dup.js', '<script>\nrender();\n', projectRoot)
+      .then((c) => applyOne(sql, c.id, projectRoot));
+
+    const oldStr = '<script>';
+    const newStr = '<script>\nconst recordFormats = ["gif"];';
+    const a = await queueEdit(sql, testSessionId, null, 'dup.js', oldStr, newStr, projectRoot);
+    const b = await queueEdit(sql, testSessionId, null, 'dup.js', oldStr, newStr, projectRoot);
+    const c = await queueEdit(sql, testSessionId, null, 'dup.js', oldStr, newStr, projectRoot);
+    // All three calls return the SAME pending row (deduped).
+    expect(b.id).toBe(a.id);
+    expect(c.id).toBe(a.id);
+
+    await applyOne(sql, a.id, projectRoot);
+    let content = await readFile(resolve(testDir, 'dup.js'), 'utf8');
+    expect((content.match(/const recordFormats/g) || []).length).toBe(1);
+
+    // Even a fresh, separately-queued identical edit re-applied is a no-op, not a stamp.
+    const again = await queueEdit(sql, testSessionId, null, 'dup.js', oldStr, newStr, projectRoot);
+    const res = await applyOne(sql, again.id, projectRoot);
+    expect(res.success).toBe(true);
+    content = await readFile(resolve(testDir, 'dup.js'), 'utf8');
+    expect((content.match(/const recordFormats/g) || []).length).toBe(1);
+  });
+
+  it('preserves CRLF line endings on edit', async () => {
+    await queueCreate(sql, testSessionId, null, 'crlf.txt', 'line one\r\nline two\r\nline three\r\n', projectRoot)
+      .then((c) => applyOne(sql, c.id, projectRoot));
+    const edit = await queueEdit(sql, testSessionId, null, 'crlf.txt', 'line two', 'line TWO', projectRoot);
+    const res = await applyOne(sql, edit.id, projectRoot);
+    expect(res.success).toBe(true);
+    const content = await readFile(resolve(testDir, 'crlf.txt'), 'utf8');
+    expect(content).toBe('line one\r\nline TWO\r\nline three\r\n');
+  });
+
+  it('refuses an edit that matches multiple locations instead of corrupting', async () => {
+    await queueCreate(sql, testSessionId, null, 'ambig.js', 'x=1;\ny=2;\nx=1;\n', projectRoot)
+      .then((ch) => applyOne(sql, ch.id, projectRoot));
+    const edit = await queueEdit(sql, testSessionId, null, 'ambig.js', 'x=1;', 'x=9;', projectRoot);
+    const res = await applyOne(sql, edit.id, projectRoot);
+    expect(res.success).toBe(false);
+    expect(res.error).toMatch(/matches 2 locations/);
+  });
+
   it('rewindOne → verify reverted', async () => {
     // Setup: create and apply a file
     const createChange = await queueCreate(sql, testSessionId, null, 'rewindable.txt', 'initial', projectRoot);

apps/coder/src/services/__tests__/plan-edit.test.ts

@@ -0,0 +1,69 @@
+import { describe, it, expect } from 'vitest';
+import { planEdit } from '../pending_changes.js';
+
+// planEdit is the pure core of applyOne's edit splice. These tests pin the
+// idempotency guards that stop the "block stamped 2-3x" corruption: applying the
+// same queued edit more than once must be a no-op, never a duplicate.
+
+describe('planEdit — normal application', () => {
+  it('applies a unique exact edit', () => {
+    const content = 'a\nfoo\nb\n';
+    const plan = planEdit(content, 'foo', 'bar');
+    expect(plan).toEqual({ kind: 'apply', updated: 'a\nbar\nb\n' });
+  });
+
+  it('reports ambiguous when old_string occurs more than once', () => {
+    const content = 'foo\nx\nfoo\n';
+    const plan = planEdit(content, 'foo', 'bar');
+    expect(plan).toEqual({ kind: 'ambiguous', count: 2 });
+  });
+
+  it('reports not_found when old_string is absent and new is not present', () => {
+    const content = 'alpha\nbeta\n';
+    const plan = planEdit(content, 'gamma that is clearly nowhere', 'delta');
+    expect(plan).toEqual({ kind: 'not_found' });
+  });
+});
+
+describe('planEdit — idempotency (the corruption guard)', () => {
+  it('treats a re-applied anchored insert as already-applied (no duplicate)', () => {
+    // The exact mechanism that tripled `const recordFormats` in settings.html:
+    // an anchored insert (old=anchor, new=anchor+block) where the anchor still
+    // matches uniquely after the first apply.
+    const oldStr = '<script>';
+    const newStr = '<script>\nconst recordFormats = ["gif","mp4"];';
+    const before = '<script>\nfunction render() {}\n</script>\n';
+
+    const first = planEdit(before, oldStr, newStr);
+    expect(first.kind).toBe('apply');
+    const after = first.kind === 'apply' ? first.updated : '';
+    expect((after.match(/const recordFormats/g) || []).length).toBe(1);
+
+    // Re-applying the identical edit to the already-edited content is a no-op.
+    const second = planEdit(after, oldStr, newStr);
+    expect(second).toEqual({ kind: 'noop', reason: 'already-applied' });
+  });
+
+  it('treats an edit whose old_string is gone but new_string is present as already-applied', () => {
+    const content = 'const total = sum + tax;\n';
+    const plan = planEdit(content, 'const subtotal = sum;', 'const total = sum + tax;');
+    expect(plan).toEqual({ kind: 'noop', reason: 'already-applied' });
+  });
+
+  it('treats a no-change splice as a noop', () => {
+    const content = 'a\nfoo\nb\n';
+    const plan = planEdit(content, 'foo', 'foo');
+    expect(plan).toEqual({ kind: 'noop', reason: 'identical' });
+  });
+
+  it('does not duplicate across three repeated applications', () => {
+    const oldStr = 'function f() {';
+    const newStr = 'function f() {\n  const x = 1;';
+    let content = 'function f() {\n  return x;\n}\n';
+    for (let i = 0; i < 3; i++) {
+      const plan = planEdit(content, oldStr, newStr);
+      if (plan.kind === 'apply') content = plan.updated;
+    }
+    expect((content.match(/const x = 1;/g) || []).length).toBe(1);
+  });
+});

apps/coder/src/services/arena-analyzer-helpers.ts

@@ -0,0 +1,191 @@
+/**
+ * Pure, side-effect-free helpers for the Arena analyzer.
+ * No DB, no IO, no network — safe to unit-test directly.
+ *
+ * Covers: digest-prompt assembly, judge-prompt assembly, winner extraction
+ * from the judge output, the <2-survivors no-winner rule, and the
+ * cross-examination prompt.
+ */
+
+// ─── Shared types ─────────────────────────────────────────────────────────────
+
+export interface ContestantDigestInput {
+  identity: string;
+  model: string;
+  resultMd: string;
+  diffPatch?: string;
+  benchmarkLine: string;
+}
+
+export interface ContestantDigest {
+  identity: string;
+  model: string;
+  digest: string;
+  benchmarkLine: string;
+}
+
+// ─── Digest stage ─────────────────────────────────────────────────────────────
+
+/**
+ * Build the system + user prompts for the per-contestant digest call.
+ * The digest is a short structured summary; it keeps each call's context small
+ * so the downstream judge only sees digests (not raw diffs).
+ */
+export function buildDigestPrompt(input: ContestantDigestInput): { system: string; user: string } {
+  const system =
+    'You are an expert technical analyst evaluating the output of an AI coding or Q&A battle. ' +
+    'Produce a concise structured digest (under 300 words, Markdown bullet points) covering: ' +
+    '(1) correctness and quality, (2) completeness, (3) notable strengths, (4) notable weaknesses or issues. ' +
+    'Do not reference the battle or other contestants — focus only on this submission.';
+
+  const parts: string[] = [
+    `# Contestant: ${input.identity} / ${input.model}`,
+    `\nBenchmark: ${input.benchmarkLine}`,
+    '\n## Result\n',
+    input.resultMd.slice(0, 8_000),
+  ];
+
+  if (input.diffPatch) {
+    parts.push('\n## Code Changes (diff)\n```diff');
+    parts.push(input.diffPatch.slice(0, 5_000));
+    parts.push('```');
+  }
+
+  return { system, user: parts.join('\n') };
+}
+
+// ─── Judge stage ──────────────────────────────────────────────────────────────
+
+/**
+ * Build the system + user prompts for the comparative judge call.
+ * Receives contestant digests (NOT raw diffs) to keep context bounded.
+ *
+ * The judge output must contain a line starting with WINNER: or NO_WINNER.
+ * The caller extracts it with extractWinner().
+ */
+export function buildJudgePrompt(
+  originalPrompt: string,
+  digests: ContestantDigest[],
+): { system: string; user: string } {
+  const canName = shouldNameWinner(digests.length);
+
+  const winnerInstruction = canName
+    ? 'After your comparative analysis, name the best submission on its own line in this exact format:\n' +
+      'WINNER: <identity>/<model>\n' +
+      'where <identity> and <model> exactly match the heading above. No other text on that line.'
+    : 'Fewer than 2 contestants succeeded. Do NOT name a winner. Write the following on its own line:\nNO_WINNER';
+
+  const system =
+    'You are an expert judge for an AI battle. You have received digest summaries of each ' +
+    "contestant's work on the same task. Write a comparative analysis, then follow these instructions:\n" +
+    winnerInstruction;
+
+  const parts: string[] = [
+    '# Original Task Prompt\n',
+    originalPrompt.slice(0, 2_000),
+    '\n# Contestant Digests\n',
+  ];
+
+  for (const d of digests) {
+    parts.push(`\n## ${d.identity} / ${d.model}`);
+    parts.push(`Benchmark: ${d.benchmarkLine}`);
+    parts.push(d.digest);
+  }
+
+  parts.push(
+    '\n# Instructions\nCompare the contestants and follow the winner-naming instructions above.',
+  );
+
+  return { system, user: parts.join('\n') };
+}
+
+// ─── No-winner rule ───────────────────────────────────────────────────────────
+
+/**
+ * Returns true when enough contestants succeeded to name a winner.
+ * Rule: at least 2 must have produced a result. With 0 or 1 success the
+ * analysis must NOT name a winner (no meaningful comparison possible).
+ */
+export function shouldNameWinner(succeededCount: number): boolean {
+  return succeededCount >= 2;
+}
+
+// ─── Winner extraction ────────────────────────────────────────────────────────
+
+/**
+ * Parse the judge's text output and extract the declared winner.
+ * Looks for a line matching: WINNER: <identity>/<model>
+ * Returns null when no valid winner line is found, or when the line contains
+ * NO_WINNER.
+ *
+ * The parse is lenient on surrounding whitespace and case for the keyword.
+ */
+export function extractWinner(judgeOutput: string): { identity: string; model: string } | null {
+  for (const line of judgeOutput.split('\n')) {
+    const trimmed = line.trim();
+    if (!trimmed.toUpperCase().startsWith('WINNER:')) continue;
+
+    const rest = trimmed.slice('WINNER:'.length).trim();
+    if (rest.toUpperCase() === 'NO_WINNER' || rest === '') return null;
+
+    const slashIdx = rest.indexOf('/');
+    if (slashIdx === -1) return null;
+
+    const identity = rest.slice(0, slashIdx).trim();
+    const model = rest.slice(slashIdx + 1).trim();
+    if (identity && model) return { identity, model };
+  }
+  return null;
+}
+
+// ─── Cross-examination stage ──────────────────────────────────────────────────
+
+/**
+ * Build the system + user prompts for a cross-examination call.
+ * The cross-examiner sees the original prompt, contestant digests, and the
+ * proposed analysis, and is asked to challenge the result.
+ */
+export function buildCrossExamPrompt(opts: {
+  originalPrompt: string;
+  digests: ContestantDigest[];
+  analysisContent: string;
+  proposedWinner: string | null;
+  examinerIdentity: string;
+  examinerModel: string;
+}): { system: string; user: string } {
+  const system =
+    `You are ${opts.examinerIdentity} (model: ${opts.examinerModel}), acting as an independent ` +
+    'cross-examiner in an AI battle. Your role is to critically challenge the proposed analysis ' +
+    'and winner, then give your own verdict. Be rigorous but fair. ' +
+    'End your response with your verdict on its own line:\n' +
+    'VERDICT: <identity>/<model>  — if you agree or disagree with the proposed winner but can name one\n' +
+    'VERDICT: NO_WINNER  — if no clear winner exists';
+
+  const parts: string[] = [
+    '# Original Task Prompt\n',
+    opts.originalPrompt.slice(0, 2_000),
+    '\n# Contestant Digests\n',
+  ];
+
+  for (const d of opts.digests) {
+    parts.push(`\n## ${d.identity} / ${d.model}`);
+    parts.push(`Benchmark: ${d.benchmarkLine}`);
+    parts.push(d.digest);
+  }
+
+  parts.push('\n# Proposed Analysis\n');
+  parts.push(opts.analysisContent.slice(0, 5_000));
+
+  if (opts.proposedWinner) {
+    parts.push(`\n*(Proposed winner: ${opts.proposedWinner})*`);
+  } else {
+    parts.push('\n*(No winner was proposed — fewer than 2 contestants succeeded.)*');
+  }
+
+  parts.push(
+    '\n# Your Cross-Examination\n' +
+      'Challenge the analysis above, then give your independent verdict (VERDICT: … on its own line).',
+  );
+
+  return { system, user: parts.join('\n') };
+}

apps/coder/src/services/arena-analyzer.ts

@@ -0,0 +1,496 @@
+/**
+ * Arena Analyzer — pluggable seam for battle analysis and cross-examination.
+ *
+ * The Analyzer interface is the plug point: a v2 Han Orchestrator flow can
+ * replace the v1 two-stage digest→judge implementation without a schema change.
+ *
+ * v1 implementation uses DEFAULT_MODEL via direct llama-swap calls (arenaModelCall):
+ *   Digest stage  — one call per succeeded contestant, concurrent; produces a
+ *                   bounded summary of each result (result.md + diff.patch for
+ *                   coding, result.md for Q&A).
+ *   Judge stage   — one call with all digests + the original prompt; writes
+ *                   analysis.md, names a winner (unless < 2 succeeded), and
+ *                   updates battles.winner_contestant_id.
+ *
+ * Cross-examination:
+ *   Local model   — direct arenaModelCall to llama-swap with the chosen model.
+ *   Cloud model   — inserts a tasks row (triggers the dispatcher via pg_notify);
+ *                   polls for completion; reads output_summary as the verdict.
+ *   In both cases the verdict is written to cross_examinations.verdict, appended
+ *   to <resultsPath>/cross-exam.md, and a battle_updated frame is published.
+ *
+ * Never throws — all errors are caught, logged, and swallowed so the caller
+ * (arena-runner's onBattleComplete / onCrossExamStart) is never wedged.
+ */
+
+import { readFile, writeFile, mkdir } from 'node:fs/promises';
+import { join } from 'node:path';
+import type { Sql } from '../db.js';
+import type { Broker } from '@boocode/server/broker';
+import type { WsFrame } from '@boocode/contracts/ws-frames';
+import type { FastifyBaseLogger } from 'fastify';
+import type { Config } from '../config.js';
+import type { BattleType } from '@boocode/contracts/arena';
+import { arenaModelCall } from './arena-model-call.js';
+import {
+  buildDigestPrompt,
+  buildJudgePrompt,
+  buildCrossExamPrompt,
+  extractWinner,
+  shouldNameWinner,
+  type ContestantDigest,
+} from './arena-analyzer-helpers.js';
+
+// ─── Public interface ─────────────────────────────────────────────────────────
+
+/** Pluggable analysis seam — swap to a Han Orchestrator flow in v2. */
+export interface Analyzer {
+  /** Run the two-stage digest→judge analysis for a completed battle. */
+  analyze(battleId: string): Promise<void>;
+  /**
+   * Run a cross-examination for an already-inserted cross_examinations row.
+   * The result is written back to that row and a battle_updated frame is published.
+   */
+  crossExamine(
+    battleId: string,
+    crossExamId: string,
+    opts: { identity: string; model: string },
+  ): Promise<void>;
+}
+
+// ─── Internal DB row types ────────────────────────────────────────────────────
+
+interface BattleRow {
+  id: string;
+  project_id: string;
+  battle_type: BattleType;
+  prompt: string;
+  status: string;
+  results_path: string | null;
+  winner_contestant_id: string | null;
+}
+
+interface ContestantRow {
+  id: string;
+  identity: string;
+  model: string;
+  lane: string;
+  status: string;
+  result_path: string | null;
+  duration_ms: number | null;
+  tokens_per_sec: number | null;
+}
+
+// ─── Factory ──────────────────────────────────────────────────────────────────
+
+interface AnalyzerDeps {
+  sql: Sql;
+  broker: Broker;
+  log: FastifyBaseLogger;
+  config: Pick<Config, 'LLAMA_SWAP_URL' | 'DEFAULT_MODEL'>;
+  /** Model IDs served by local llama-swap — cross-exam routing uses this. */
+  localModels: ReadonlySet<string>;
+}
+
+export function createAnalyzer(deps: AnalyzerDeps): Analyzer {
+  const { sql, broker, log, config, localModels } = deps;
+
+  // ─── analyze ──────────────────────────────────────────────────────────────
+
+  async function analyze(battleId: string): Promise<void> {
+    try {
+      await runAnalysis(battleId);
+    } catch (err) {
+      log.error(
+        { err: errMsg(err), battleId },
+        'arena-analyzer: analysis failed',
+      );
+    }
+  }
+
+  async function runAnalysis(battleId: string): Promise<void> {
+    const battle = await loadBattle(battleId);
+    if (!battle) {
+      log.warn({ battleId }, 'arena-analyzer: battle not found');
+      return;
+    }
+
+    const contestants = await loadContestants(battleId);
+    const succeeded = contestants.filter((c) => c.status === 'done' && c.result_path);
+
+    log.info(
+      { battleId, total: contestants.length, succeeded: succeeded.length },
+      'arena-analyzer: starting analysis',
+    );
+
+    // Digest stage — concurrent, one call per succeeded contestant.
+    const digests = (
+      await Promise.all(succeeded.map((c) => digestContestant(battle, c)))
+    ).filter((d): d is ContestantDigest => d !== null);
+
+    // Failed contestants are noted in the analysis even if they produced no digest.
+    const failedNotes = contestants
+      .filter((c) => c.status === 'error')
+      .map((c) => `- **${c.identity} / ${c.model}**: failed (no result)\n`);
+
+    // Judge stage — single call with all digests.
+    const { analysisText, winner } = await judgeContestants(battle, digests, failedNotes);
+
+    // Write analysis.md to the battle results folder.
+    const resultsPath = battle.results_path;
+    if (resultsPath) {
+      await mkdir(resultsPath, { recursive: true });
+      await writeFile(join(resultsPath, 'analysis.md'), analysisText, 'utf8');
+    }
+
+    // Resolve the winner to a contestant id and update the battle row.
+    let winnerId: string | null = null;
+    if (winner && shouldNameWinner(succeeded.length)) {
+      const winnerContestant = contestants.find(
+        (c) => c.identity === winner.identity && c.model === winner.model,
+      );
+      if (winnerContestant) {
+        winnerId = winnerContestant.id;
+        await sql`
+          UPDATE battles
+          SET winner_contestant_id = ${winnerId}, updated_at = clock_timestamp()
+          WHERE id = ${battleId}
+        `;
+        log.info({ battleId, winnerId, identity: winner.identity, model: winner.model }, 'arena-analyzer: winner set');
+      } else {
+        log.warn({ battleId, winner }, 'arena-analyzer: judge named a winner not found in contestants');
+      }
+    }
+
+    publishUser({
+      type: 'battle_updated',
+      battle_id: battleId,
+      winner_contestant_id: winnerId,
+      analysis_ready: true,
+    });
+
+    log.info({ battleId }, 'arena-analyzer: analysis complete');
+  }
+
+  // ─── crossExamine ─────────────────────────────────────────────────────────
+
+  async function crossExamine(
+    battleId: string,
+    crossExamId: string,
+    opts: { identity: string; model: string },
+  ): Promise<void> {
+    try {
+      await runCrossExam(battleId, crossExamId, opts);
+    } catch (err) {
+      log.error(
+        { err: errMsg(err), battleId, crossExamId },
+        'arena-analyzer: cross-exam failed',
+      );
+    }
+  }
+
+  async function runCrossExam(
+    battleId: string,
+    crossExamId: string,
+    opts: { identity: string; model: string },
+  ): Promise<void> {
+    const battle = await loadBattle(battleId);
+    if (!battle) {
+      log.warn({ battleId }, 'arena-analyzer: battle not found for cross-exam');
+      return;
+    }
+
+    const contestants = await loadContestants(battleId);
+
+    // Re-read the digests (if contestants have results) for context.
+    const succeeded = contestants.filter((c) => c.status === 'done' && c.result_path);
+    const digests = (
+      await Promise.all(succeeded.map((c) => digestContestant(battle, c)))
+    ).filter((d): d is ContestantDigest => d !== null);
+
+    // Read analysis.md for the proposed analysis content.
+    let analysisContent = '';
+    if (battle.results_path) {
+      analysisContent = await readFile(
+        join(battle.results_path, 'analysis.md'), 'utf8',
+      ).catch(() => '');
+    }
+
+    // Resolve proposed winner label.
+    let proposedWinner: string | null = null;
+    if (battle.winner_contestant_id) {
+      const w = contestants.find((c) => c.id === battle.winner_contestant_id);
+      if (w) proposedWinner = `${w.identity}/${w.model}`;
+    }
+
+    const { system, user } = buildCrossExamPrompt({
+      originalPrompt: battle.prompt,
+      digests,
+      analysisContent,
+      proposedWinner,
+      examinerIdentity: opts.identity,
+      examinerModel: opts.model,
+    });
+
+    log.info({ battleId, crossExamId, identity: opts.identity, model: opts.model }, 'arena-analyzer: running cross-exam');
+
+    const verdict = await executeModelCall({
+      battleId,
+      projectId: battle.project_id,
+      identity: opts.identity,
+      model: opts.model,
+      system,
+      user,
+    });
+
+    // Persist verdict and append to cross-exam.md.
+    await sql`
+      UPDATE cross_examinations
+      SET verdict = ${verdict}
+      WHERE id = ${crossExamId}
+    `;
+
+    if (battle.results_path) {
+      const crossExamPath = join(battle.results_path, 'cross-exam.md');
+      const section =
+        `\n---\n\n# Cross-Examination by ${opts.identity} / ${opts.model}\n\n` +
+        `${verdict}\n`;
+      await writeFile(crossExamPath, section, { flag: 'a', encoding: 'utf8' });
+    }
+
+    publishUser({
+      type: 'battle_updated',
+      battle_id: battleId,
+      cross_exam_id: crossExamId,
+    });
+
+    log.info({ battleId, crossExamId }, 'arena-analyzer: cross-exam complete');
+  }
+
+  // ─── Model call routing ───────────────────────────────────────────────────
+
+  /**
+   * Route a one-shot model call to llama-swap (local) or the task dispatcher
+   * (cloud). Cloud dispatch inserts a tasks row and polls for completion.
+   */
+  async function executeModelCall(opts: {
+    battleId: string;
+    projectId: string;
+    identity: string;
+    model: string;
+    system: string;
+    user: string;
+  }): Promise<string> {
+    const isLocal = localModels.has(opts.model) || localModels.has(`llama-swap/${opts.model}`);
+
+    if (isLocal) {
+      return arenaModelCall({
+        config,
+        model: opts.model,
+        system: opts.system,
+        user: opts.user,
+        maxTokens: 2_000,
+        temperature: 0.3,
+      });
+    }
+
+    // Cloud path: dispatch through the task system and poll for completion.
+    return executeCloudModelCall(opts);
+  }
+
+  async function executeCloudModelCall(opts: {
+    projectId: string;
+    identity: string;
+    model: string;
+    system: string;
+    user: string;
+  }): Promise<string> {
+    // The cross-exam prompt is the full input to the external agent. We embed
+    // the system prompt as a preamble in the user message (external agents don't
+    // take a separate system arg through the tasks dispatcher).
+    const input = `${opts.system}\n\n${opts.user}`;
+
+    // For well-known external agents, stamp the agent name so the dispatcher
+    // routes via PTY/ACP. For unknown identities fall back to native inference
+    // (agent = null → DEFAULT_MODEL text generation).
+    const knownAgents = new Set(['claude', 'opencode', 'qwen', 'goose']);
+    const agentName = knownAgents.has(opts.identity) ? opts.identity : null;
+
+    const [task] = await sql<{ id: string }[]>`
+      INSERT INTO tasks (project_id, input, agent, model)
+      VALUES (${opts.projectId}, ${input}, ${agentName}, ${opts.model})
+      RETURNING id
+    `;
+    const taskId = task!.id;
+
+    log.info({ taskId, identity: opts.identity, model: opts.model }, 'arena-analyzer: cloud cross-exam task dispatched');
+
+    // Poll until terminal (up to 5 minutes).
+    const timeoutMs = 5 * 60 * 1_000;
+    const pollMs = 2_000;
+    const deadline = Date.now() + timeoutMs;
+
+    while (Date.now() < deadline) {
+      await sleep(pollMs);
+      const [row] = await sql<{ state: string; output_summary: string | null }[]>`
+        SELECT state, output_summary FROM tasks WHERE id = ${taskId}
+      `;
+      if (!row) break;
+      if (row.state === 'completed') return row.output_summary ?? '';
+      if (row.state === 'failed' || row.state === 'cancelled') {
+        throw new Error(`cross-exam task ${row.state}: ${row.output_summary ?? ''}`);
+      }
+    }
+
+    throw new Error(`cloud cross-exam task timed out after ${timeoutMs / 1000}s`);
+  }
+
+  // ─── Digest helper ────────────────────────────────────────────────────────
+
+  async function digestContestant(
+    battle: BattleRow,
+    c: ContestantRow,
+  ): Promise<ContestantDigest | null> {
+    if (!c.result_path) return null;
+
+    const resultMd = await readFile(join(c.result_path, 'result.md'), 'utf8').catch(() => '');
+
+    let diffPatch: string | undefined;
+    if (battle.battle_type === 'coding') {
+      diffPatch = await readFile(join(c.result_path, 'diff.patch'), 'utf8').catch(
+        () => undefined,
+      );
+    }
+
+    const benchmarkLine = formatBenchmarkLine(c);
+    const { system, user } = buildDigestPrompt({
+      identity: c.identity,
+      model: c.model,
+      resultMd,
+      diffPatch,
+      benchmarkLine,
+    });
+
+    let digest: string;
+    try {
+      digest = await arenaModelCall({
+        config,
+        model: config.DEFAULT_MODEL,
+        system,
+        user,
+        maxTokens: 500,
+        temperature: 0.3,
+      });
+    } catch (err) {
+      log.warn(
+        { err: errMsg(err), identity: c.identity, model: c.model },
+        'arena-analyzer: digest call failed — skipping contestant',
+      );
+      return null;
+    }
+
+    return { identity: c.identity, model: c.model, digest, benchmarkLine };
+  }
+
+  // ─── Judge helper ─────────────────────────────────────────────────────────
+
+  async function judgeContestants(
+    battle: BattleRow,
+    digests: ContestantDigest[],
+    failedNotes: string[],
+  ): Promise<{ analysisText: string; winner: { identity: string; model: string } | null }> {
+    const { system, user } = buildJudgePrompt(battle.prompt, digests);
+
+    let judgeOutput = '';
+    try {
+      judgeOutput = await arenaModelCall({
+        config,
+        model: config.DEFAULT_MODEL,
+        system,
+        user,
+        maxTokens: 2_000,
+        temperature: 0.3,
+      });
+    } catch (err) {
+      log.error({ err: errMsg(err), battleId: battle.id }, 'arena-analyzer: judge call failed');
+      judgeOutput = '*(Judge call failed — no comparison produced.)*';
+    }
+
+    const winner = shouldNameWinner(digests.length) ? extractWinner(judgeOutput) : null;
+
+    const sections: string[] = [
+      `# Arena Analysis`,
+      `\n**Battle type:** ${battle.battle_type}`,
+    ];
+
+    if (failedNotes.length > 0) {
+      sections.push('\n## Failed Contestants\n');
+      sections.push(...failedNotes);
+    }
+
+    if (digests.length > 0) {
+      sections.push('\n## Contestant Digests\n');
+      for (const d of digests) {
+        sections.push(`### ${d.identity} / ${d.model}`);
+        sections.push(`*Benchmark: ${d.benchmarkLine}*\n`);
+        sections.push(d.digest);
+      }
+    }
+
+    sections.push("\n## Judge's Verdict\n");
+    sections.push(judgeOutput);
+
+    if (winner) {
+      sections.push(`\n## Winner\n**${winner.identity} / ${winner.model}**`);
+    } else {
+      const reason =
+        digests.length < 2
+          ? 'fewer than 2 contestants produced results'
+          : 'no clear winner identified';
+      sections.push(`\n## Winner\n*No winner named (${reason}).*`);
+    }
+
+    return { analysisText: sections.join('\n'), winner };
+  }
+
+  // ─── DB helpers ───────────────────────────────────────────────────────────
+
+  async function loadBattle(battleId: string): Promise<BattleRow | null> {
+    const [b] = await sql<BattleRow[]>`
+      SELECT id, project_id, battle_type, prompt, status, results_path, winner_contestant_id
+      FROM battles WHERE id = ${battleId}
+    `;
+    return b ?? null;
+  }
+
+  async function loadContestants(battleId: string): Promise<ContestantRow[]> {
+    return sql<ContestantRow[]>`
+      SELECT id, identity, model, lane, status, result_path, duration_ms, tokens_per_sec
+      FROM contestants WHERE battle_id = ${battleId}
+      ORDER BY created_at ASC
+    `;
+  }
+
+  // ─── Misc helpers ─────────────────────────────────────────────────────────
+
+  function formatBenchmarkLine(c: ContestantRow): string {
+    const parts: string[] = [];
+    if (c.duration_ms !== null) parts.push(`${c.duration_ms}ms`);
+    if (c.tokens_per_sec !== null) parts.push(`${c.tokens_per_sec.toFixed(1)} tok/s`);
+    return parts.length > 0 ? parts.join(', ') : 'no benchmark';
+  }
+
+  function publishUser(frame: Record<string, unknown>): void {
+    broker.publishUserFrame('default', frame as unknown as WsFrame);
+  }
+
+  function sleep(ms: number): Promise<void> {
+    return new Promise((resolve) => setTimeout(resolve, ms));
+  }
+
+  return { analyze, crossExamine };
+}
+
+function errMsg(e: unknown): string {
+  return e instanceof Error ? e.message : String(e);
+}

apps/coder/src/services/arena-decisions.ts

@@ -0,0 +1,186 @@
+/**
+ * Pure scheduling and classification decisions for the Arena battle-runner.
+ * No database, no IO. Mirrors the pattern of flow-runner-decisions.ts.
+ *
+ * Vocabulary:
+ *   local lane  — llama-swap-backed contestants, run strictly one at a time
+ *   cloud lane  — cloud-backed contestants, run all in parallel
+ *
+ * A contestant's status lifecycle:
+ *   queued → running → done | error
+ */
+import type { BattleType, ContestantLane } from '@boocode/contracts/arena';
+
+// ─── Lane classification ──────────────────────────────────────────────────────
+
+/**
+ * Classify a contestant into a lane.
+ *
+ * Q&A contestants always run on the native (llama-swap) backend → local.
+ * Coding contestants: their MODEL is checked against the localModels set
+ * (all model IDs served by the local llama-swap server). This means an
+ * opencode or qwen contestant pointed at a local model counts as local,
+ * which correctly captures GPU-contention and fair benchmarking (ADR 0001).
+ *
+ * @param battleType  'coding' | 'qa'
+ * @param identity    backend name (coding) or persona name (qa) — not used for lane logic
+ * @param model       the contestant's model id
+ * @param localModels set of model IDs served by the local llama-swap server
+ */
+export function classifyLane(
+  battleType: BattleType,
+  _identity: string,
+  model: string,
+  localModels: ReadonlySet<string>,
+): ContestantLane {
+  if (battleType === 'qa') return 'local';
+  return localModels.has(model) ? 'local' : 'cloud';
+}
+
+// ─── Local-lane queue ─────────────────────────────────────────────────────────
+
+export interface ContestantSlot {
+  id: string;
+  lane: ContestantLane;
+  status: string;
+}
+
+/**
+ * The next queued local contestant to dispatch — the first 'queued' contestant
+ * in the local lane, in creation order (caller must supply rows in created_at ASC).
+ * Returns null when the local queue is empty or all local slots are non-queued.
+ */
+export function nextLocalContestant(contestants: readonly ContestantSlot[]): string | null {
+  for (const c of contestants) {
+    if (c.lane === 'local' && c.status === 'queued') return c.id;
+  }
+  return null;
+}
+
+// ─── Battle completion ────────────────────────────────────────────────────────
+
+/**
+ * True when every contestant has reached a terminal state (done | error).
+ * Returns false for an empty list — a battle with no contestants never completes.
+ */
+export function isBattleComplete(contestants: readonly { status: string }[]): boolean {
+  if (contestants.length === 0) return false;
+  return contestants.every((c) => c.status === 'done' || c.status === 'error');
+}
+
+// ─── Benchmark ────────────────────────────────────────────────────────────────
+
+export interface Benchmark {
+  durationMs: number;
+  tokensPerSec: number | null;
+}
+
+/**
+ * Compute the benchmark for a contestant.
+ * Wall-clock duration is captured for every contestant; tokens/sec is only
+ * meaningful for local (llama-swap) contestants where the model has sole
+ * access to the GPU and the measurement is fair.
+ */
+export function computeBenchmark(
+  startedAt: Date,
+  endedAt: Date,
+  costTokens: number | null,
+  lane: ContestantLane,
+): Benchmark {
+  const durationMs = Math.max(0, endedAt.getTime() - startedAt.getTime());
+  const tokensPerSec =
+    lane === 'local' && costTokens !== null && durationMs > 0
+      ? (costTokens / durationMs) * 1000
+      : null;
+  return { durationMs, tokensPerSec };
+}
+
+// ─── Slug / path helpers ──────────────────────────────────────────────────────
+
+/**
+ * Sanitize a string for use as a directory name component.
+ * Lowercases, replaces non-alphanumeric runs with '-', trims leading/trailing
+ * dashes, and caps at 64 characters.
+ */
+export function sanitizeSlug(s: string): string {
+  return s
+    .toLowerCase()
+    .replace(/[^a-z0-9]+/g, '-')
+    .replace(/^-+|-+$/g, '')
+    .slice(0, 64);
+}
+
+/**
+ * Build the dated battle slug used as the Arena results folder name.
+ * Format: YYYY-MM-DD-<battleType>-<first-8-hex-of-uuid>
+ * Deterministic: callers can rebuild it from (id, type, created_at) on resume.
+ */
+export function buildBattleSlug(battleId: string, battleType: BattleType, createdAt: Date): string {
+  const date = createdAt.toISOString().slice(0, 10);
+  const shortId = battleId.replace(/-/g, '').slice(0, 8);
+  return `${date}-${battleType}-${shortId}`;
+}
+
+/**
+ * Build the per-contestant results directory name within a battle folder.
+ * Format: <sanitized-identity>-<sanitized-model>
+ */
+export function buildContestantDir(identity: string, model: string): string {
+  return `${sanitizeSlug(identity)}-${sanitizeSlug(model)}`;
+}
+
+// ─── Resume reconciliation ────────────────────────────────────────────────────
+
+export type ContestantResumeAction =
+  | 'keep'
+  | 're-dispatch'
+  | 'mark-done'
+  | 'mark-error'
+  | 'mark-cancelled';
+
+export interface ContestantResumeDecision {
+  contestantId: string;
+  action: ContestantResumeAction;
+}
+
+/**
+ * Decide what to do with ONE contestant during startup resume.
+ * Mirrors reconcileResumeStep from flow-runner-decisions.ts.
+ *
+ * @param status    contestants.status
+ * @param taskId    contestants.task_id (null when not yet dispatched)
+ * @param taskState tasks.state for taskId, or null if the task row is absent
+ */
+export function reconcileContestantResume(
+  status: string,
+  taskId: string | null,
+  taskState: string | null,
+): ContestantResumeAction {
+  if (status !== 'running') return 'keep';
+  if (!taskId || taskState === null) return 're-dispatch';
+  switch (taskState) {
+    case 'completed': return 'mark-done';
+    case 'failed':    return 'mark-error';
+    case 'cancelled': return 'mark-cancelled';
+    case 'pending':   return 'keep'; // dispatcher startup poll will run it normally
+    default:          return 're-dispatch'; // 'running'/'blocked' — process is dead
+  }
+}
+
+/**
+ * Reconcile every contestant of an in-flight battle for startup resume.
+ * Returns one decision per contestant. Pure — no IO.
+ */
+export function reconcileContestants(
+  contestants: ReadonlyArray<{ contestantId: string; taskId: string | null; status: string }>,
+  taskStates: ReadonlyMap<string, string>,
+): ContestantResumeDecision[] {
+  return contestants.map((c) => ({
+    contestantId: c.contestantId,
+    action: reconcileContestantResume(
+      c.status,
+      c.taskId,
+      c.taskId ? (taskStates.get(c.taskId) ?? null) : null,
+    ),
+  }));
+}

apps/coder/src/services/arena-model-call.ts

@@ -0,0 +1,70 @@
+/**
+ * One-shot model completion for the Arena analyzer.
+ *
+ * Calls the local llama-swap server directly for a single non-streaming
+ * completion. Used for the digest and judge stages (always DEFAULT_MODEL)
+ * and for local-model cross-examinations (any local model).
+ *
+ * Mirrors apps/server/src/services/task-model.ts but targets the coder's
+ * config shape and uses a longer timeout appropriate for analysis calls.
+ */
+
+import type { Config } from '../config.js';
+
+const TIMEOUT_MS = 120_000;
+
+export async function arenaModelCall(opts: {
+  config: Pick<Config, 'LLAMA_SWAP_URL'>;
+  model: string;
+  system: string;
+  user: string;
+  maxTokens?: number;
+  temperature?: number;
+}): Promise<string> {
+  const { config, model, system, user } = opts;
+  const maxTokens = opts.maxTokens ?? 2_000;
+  const temperature = opts.temperature ?? 0.3;
+
+  const res = await fetch(`${config.LLAMA_SWAP_URL}/v1/chat/completions`, {
+    method: 'POST',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify({
+      model,
+      messages: [
+        { role: 'system', content: system },
+        { role: 'user', content: user },
+      ],
+      max_tokens: maxTokens,
+      temperature,
+      stream: false,
+      chat_template_kwargs: { enable_thinking: false },
+    }),
+    signal: AbortSignal.timeout(TIMEOUT_MS),
+  });
+
+  if (!res.ok) {
+    const text = await res.text().catch(() => '');
+    throw new Error(`llama-swap responded ${res.status}: ${text.slice(0, 200)}`);
+  }
+
+  const data = (await res.json()) as {
+    choices?: Array<{
+      message?: { content?: string; reasoning_content?: string };
+    }>;
+  };
+
+  const choice = data.choices?.[0]?.message;
+  if (!choice) return '';
+
+  const content = (choice.content ?? '').trim();
+  if (content.length > 0) return content;
+
+  // For thinking-mode models the answer sometimes only lands in reasoning_content.
+  const reasoning = (choice.reasoning_content ?? '').trim();
+  if (reasoning.length > 0) {
+    const lines = reasoning.split('\n').filter((l) => l.trim().length > 0);
+    return lines[lines.length - 1] ?? '';
+  }
+
+  return '';
+}

apps/coder/src/services/arena-runner.ts

@@ -0,0 +1,895 @@
+/**
+ * Arena battle-runner — DB-backed execution engine for Arena battles.
+ *
+ * Mirrors flow-runner.ts but implements the Arena's two-lane scheduler instead
+ * of the Orchestrator's wave scheduler. Persists to battles/contestants tables
+ * (not flow_runs/flow_steps). Each contestant is dispatched as a real tasks row
+ * via an injected DispatchContestantFn (Phase 4 wires this to the dispatcher).
+ * Advances on the dispatcher's onTaskTerminal hook.
+ *
+ * Scheduling:
+ *   - Cloud lane: all contestants start immediately, in parallel.
+ *   - Local lane: contestants run strictly one at a time (serial queue). Only
+ *     the first local contestant runs at start; the next is dispatched when the
+ *     current one terminates. Both lanes run concurrently with each other.
+ *
+ * Results:
+ *   Written to <projectRoot>/Arena/<battleSlug>/<identity>-<model>/
+ *   Coding: result.md + diff.patch (from the contestant's worktree).
+ *   Q&A:    result.md with the text answer.
+ *
+ * Analyzer seam:
+ *   onBattleComplete is called when all contestants are terminal. Phase 5 wires
+ *   this to the two-stage digest→judge analyzer. A failed contestant does NOT
+ *   abort the battle — others continue and the analyzer judges survivors.
+ */
+import type { Sql } from '../db.js';
+import type { Broker } from '@boocode/server/broker';
+import type { WsFrame } from '@boocode/contracts/ws-frames';
+import type { FastifyBaseLogger } from 'fastify';
+import type { BattleType, ContestantLane } from '@boocode/contracts/arena';
+import { mkdir, writeFile } from 'node:fs/promises';
+import { join } from 'node:path';
+import { diffWorktree } from './worktrees.js';
+import {
+  buildBattleSlug,
+  buildContestantDir,
+  classifyLane,
+  computeBenchmark,
+  isBattleComplete,
+  nextLocalContestant,
+  reconcileContestants,
+  type ContestantResumeAction,
+  type ContestantSlot,
+} from './arena-decisions.js';
+
+// ─── Public types ─────────────────────────────────────────────────────────────
+
+export interface ContestantSpec {
+  /** Backend name (coding) or persona name (qa). */
+  identity: string;
+  model: string;
+}
+
+export interface BattleStartOpts {
+  projectId: string;
+  battleType: BattleType;
+  prompt: string;
+  /** 2–6 contestants. Duplicate (identity, model) pairs are rejected by the schema UNIQUE constraint. */
+  contestants: ContestantSpec[];
+}
+
+/**
+ * Injected dispatch function — Phase 4 wires this to the real task inserter.
+ * Must INSERT a tasks row and return its id. The arena-runner sets the
+ * contestant's task_id and status after this call.
+ * `sessionId` is returned when already known (Q&A pre-creates the session);
+ * null for coding contestants whose session is created lazily by the dispatcher.
+ */
+export type DispatchContestantFn = (opts: {
+  projectId: string;
+  contestantId: string;
+  prompt: string;
+  identity: string;
+  model: string;
+  battleType: BattleType;
+}) => Promise<{ taskId: string; sessionId: string | null }>;
+
+/**
+ * Called once when every contestant in a battle has reached a terminal state.
+ * Phase 5 wires this to the two-stage digest→judge analyzer.
+ * Must never throw — the caller swallows errors.
+ */
+export type OnBattleComplete = (battleId: string) => void;
+
+/**
+ * Called after a cross_examinations row has been inserted, with its id.
+ * Phase 5 wires this to the analyzer's cross-examination runner.
+ * Must never throw — the caller swallows errors.
+ */
+export type OnCrossExamStart = (opts: {
+  battleId: string;
+  crossExamId: string;
+  identity: string;
+  model: string;
+}) => void;
+
+export interface BattleRunner {
+  /** Start a battle: persist it + its contestants, classify lanes, dispatch initial wave. */
+  startBattle(opts: BattleStartOpts): Promise<{ battleId: string }>;
+  /**
+   * Wire to createDispatcher({ onTaskTerminal }). Fires when ANY task settles;
+   * the runner ignores tasks it doesn't own. Never throws.
+   */
+  handleTaskTerminal(taskId: string, state: string): void;
+  /**
+   * Re-advance any battles still marked 'running' after a coder restart.
+   * Mirrors flow-runner's initResume (D-9). Never throws.
+   */
+  initResume(): Promise<void>;
+  /**
+   * Cancel a running battle. Marks it and all non-terminal contestants cancelled,
+   * publishes frames, and returns the task_ids of in-flight contestants so the
+   * route can abort them via the dispatcher's cancelExternalTask.
+   */
+  cancelBattle(battleId: string): Promise<{ cancelled: boolean; taskIds: string[] }>;
+  /**
+   * Trigger analysis for a completed (or manually re-analyzed) battle.
+   * Phase 5 wires this to the two-stage digest→judge analyzer. For now, calls
+   * the injected onBattleComplete seam directly.
+   */
+  triggerAnalysis(battleId: string): Promise<{ triggered: boolean }>;
+  /**
+   * Start a cross-examination on a battle. Inserts a cross_examinations row and
+   * invokes the analyzer seam. Phase 5 fills the actual verdict logic.
+   */
+  startCrossExam(
+    battleId: string,
+    opts: { identity: string; model: string },
+  ): Promise<{ crossExamId: string }>;
+  /**
+   * Manually set (or clear) the winner. Validates the contestant belongs to the
+   * battle, updates battles.winner_contestant_id, and publishes a battle_updated
+   * frame so the pane reflects the override immediately.
+   */
+  setWinner(battleId: string, winnerId: string | null): Promise<{
+    ok: boolean;
+    notFound?: boolean;
+    invalidContestant?: boolean;
+  }>;
+}
+
+// ─── Internal row shapes ──────────────────────────────────────────────────────
+
+interface ContestantRow {
+  id: string;
+  battle_id: string;
+  identity: string;
+  model: string;
+  lane: ContestantLane;
+  task_id: string | null;
+  worktree_id: string | null;
+  status: string;
+}
+
+interface BattleRow {
+  id: string;
+  project_id: string;
+  battle_type: BattleType;
+  prompt: string;
+  status: string;
+  results_path: string | null;
+  created_at: Date;
+}
+
+// ─── Deps / factory ───────────────────────────────────────────────────────────
+
+interface Deps {
+  sql: Sql;
+  broker: Broker;
+  log: FastifyBaseLogger;
+  dispatch: DispatchContestantFn;
+  onBattleComplete: OnBattleComplete;
+  /**
+   * Called after a cross_examinations row is inserted. Phase 5 wires this to
+   * the analyzer's cross-examination runner. Optional: absent → no cross-exam
+   * logic runs (stub behaviour for tests).
+   */
+  onCrossExamStart?: OnCrossExamStart;
+  /**
+   * Model IDs served by the local llama-swap server. Used for lane classification:
+   * a contestant whose model is in this set runs in the local lane (serial, GPU-fair).
+   * Q&A contestants are always local regardless of this set.
+   * Defaults to an empty set → all coding contestants go to the cloud lane.
+   */
+  localModels?: ReadonlySet<string>;
+}
+
+const DEFAULT_LOCAL_MODELS: ReadonlySet<string> = new Set();
+
+export function createBattleRunner(deps: Deps): BattleRunner {
+  const { sql, broker, log, dispatch, onBattleComplete, onCrossExamStart } = deps;
+  const localModels = deps.localModels ?? DEFAULT_LOCAL_MODELS;
+
+  // Serialize local-lane advance per battle so two near-simultaneous terminal
+  // callbacks don't double-dispatch the next local contestant.
+  const advanceChain = new Map<string, Promise<void>>();
+
+  // Delta bridge: per-contestant broker unsubscribe functions.
+  // 'terminated' sentinel prevents a late-arriving setupDeltaBridge from
+  // registering a subscription that would never be cleaned up.
+  const deltaUnsubs = new Map<string, (() => void) | 'terminated'>();
+
+  function publishUser(frame: Record<string, unknown>): void {
+    broker.publishUserFrame('default', frame as unknown as WsFrame);
+  }
+
+  /**
+   * Subscribe to the contestant's inference session and forward delta frames
+   * to the user channel as contestant_updated{delta}. Polls for session_id
+   * when not immediately known (coding contestants whose session is created
+   * lazily by the dispatcher). Unsubscribes on termination or max retries.
+   */
+  async function setupDeltaBridge(
+    battleId: string,
+    contestantId: string,
+    taskId: string,
+    knownSessionId: string | null,
+  ): Promise<void> {
+    let sessionId = knownSessionId;
+    if (!sessionId) {
+      // Coding contestant: session_id is written by the dispatcher just before
+      // inference starts. Poll until it appears or the contestant terminates.
+      for (let i = 0; i < 50; i++) {
+        if (deltaUnsubs.get(contestantId) === 'terminated') return;
+        const [row] = await sql<{ session_id: string | null }[]>`
+          SELECT session_id FROM tasks WHERE id = ${taskId}
+        `.catch(() => []);
+        if (row?.session_id) { sessionId = row.session_id; break; }
+        await new Promise((r) => setTimeout(r, 200));
+      }
+    }
+    if (!sessionId) return;
+    if (deltaUnsubs.get(contestantId) === 'terminated') return;
+
+    const unsub = broker.subscribe(sessionId, (frame) => {
+      if (frame.type === 'delta') {
+        const deltaContent = (frame as unknown as { content?: unknown }).content;
+        if (typeof deltaContent === 'string') {
+          publishUser({
+            type: 'contestant_updated',
+            battle_id: battleId,
+            contestant_id: contestantId,
+            delta: deltaContent,
+          });
+        }
+      }
+    });
+
+    const existing = deltaUnsubs.get(contestantId);
+    if (existing === 'terminated') {
+      unsub();
+    } else {
+      deltaUnsubs.set(contestantId, unsub);
+    }
+  }
+
+  function teardownDeltaBridge(contestantId: string): void {
+    const entry = deltaUnsubs.get(contestantId);
+    if (typeof entry === 'function') {
+      entry();
+      deltaUnsubs.delete(contestantId);
+    } else {
+      deltaUnsubs.set(contestantId, 'terminated');
+    }
+  }
+
+  // ─── startBattle ────────────────────────────────────────────────────────────
+
+  async function startBattle(opts: BattleStartOpts): Promise<{ battleId: string }> {
+    if (opts.contestants.length < 2 || opts.contestants.length > 6) {
+      throw new Error(`battle requires 2–6 contestants; got ${opts.contestants.length}`);
+    }
+
+    const [proj] = await sql<{ path: string }[]>`SELECT path FROM projects WHERE id = ${opts.projectId}`;
+    if (!proj) throw new Error(`project not found: ${opts.projectId}`);
+
+    // Insert the battle row as 'running'; update results_path once we have the id.
+    const [battle] = await sql<{ id: string; created_at: Date }[]>`
+      INSERT INTO battles (project_id, battle_type, prompt, status)
+      VALUES (${opts.projectId}, ${opts.battleType}, ${opts.prompt}, 'running')
+      RETURNING id, created_at
+    `;
+    const battleId = battle!.id;
+    const battleSlug = buildBattleSlug(battleId, opts.battleType, battle!.created_at);
+    const resultsPath = join(proj.path, 'Arena', battleSlug);
+
+    await sql`
+      UPDATE battles SET results_path = ${resultsPath}, updated_at = clock_timestamp()
+      WHERE id = ${battleId}
+    `;
+
+    // Insert all contestant rows with lane classification.
+    const contestantRows: Array<{ id: string; identity: string; model: string; lane: ContestantLane }> = [];
+    for (const spec of opts.contestants) {
+      const lane = classifyLane(opts.battleType, spec.identity, spec.model, localModels);
+      const [row] = await sql<{ id: string }[]>`
+        INSERT INTO contestants (battle_id, identity, model, lane, status)
+        VALUES (${battleId}, ${spec.identity}, ${spec.model}, ${lane}, 'queued')
+        RETURNING id
+      `;
+      contestantRows.push({ id: row!.id, identity: spec.identity, model: spec.model, lane });
+    }
+
+    // Write initial manifest so the results folder is always populated.
+    await writeManifest(
+      battleId, resultsPath, opts.battleType, opts.prompt, battle!.created_at,
+      contestantRows.map((c) => ({ identity: c.identity, model: c.model, lane: c.lane })),
+      null,
+    ).catch((err) => {
+      log.warn({ err: errMsg(err), battleId }, 'arena-runner: initial manifest write failed');
+    });
+
+    publishUser({
+      type: 'battle_started',
+      battle_id: battleId,
+      battle_type: opts.battleType,
+      prompt: opts.prompt,
+      contestants: contestantRows.map((c) => ({
+        id: c.id,
+        identity: c.identity,
+        model: c.model,
+        lane: c.lane,
+      })),
+    });
+
+    // Dispatch: cloud lane starts all contestants in parallel; local lane starts
+    // only the first queued contestant (serial queue).
+    let localStarted = false;
+    for (const c of contestantRows) {
+      if (c.lane === 'cloud') {
+        await dispatchContestant(battleId, opts.projectId, opts.battleType, opts.prompt, c);
+      } else if (!localStarted) {
+        await dispatchContestant(battleId, opts.projectId, opts.battleType, opts.prompt, c);
+        localStarted = true;
+        // remaining local contestants stay 'queued' until this one finishes
+      }
+    }
+
+    return { battleId };
+  }
+
+  async function dispatchContestant(
+    battleId: string,
+    projectId: string,
+    battleType: BattleType,
+    prompt: string,
+    c: { id: string; identity: string; model: string; lane: ContestantLane },
+  ): Promise<void> {
+    const { taskId, sessionId } = await dispatch({
+      projectId,
+      contestantId: c.id,
+      prompt,
+      identity: c.identity,
+      model: c.model,
+      battleType,
+    });
+    await sql`
+      UPDATE contestants
+      SET task_id = ${taskId}, status = 'running', updated_at = clock_timestamp()
+      WHERE id = ${c.id}
+    `;
+    publishContestantFrame(battleId, c.id, { status: 'running' });
+    // Start the delta bridge in the background; unsubscribe when the contestant
+    // terminates (teardownDeltaBridge called in handleTaskTerminal).
+    void setupDeltaBridge(battleId, c.id, taskId, sessionId ?? null);
+  }
+
+  // ─── local-lane advance (serialized per battle) ───────────────────────────
+
+  function advanceLocalLane(battleId: string): Promise<void> {
+    const prev = advanceChain.get(battleId) ?? Promise.resolve();
+    const next = prev
+      .catch(() => {})
+      .then(() =>
+        advanceLocalLaneInner(battleId).catch((err) => {
+          log.error({ err: errMsg(err), battleId }, 'arena-runner: advanceLocalLane failed');
+        }),
+      );
+    advanceChain.set(battleId, next);
+    void next.finally(() => {
+      if (advanceChain.get(battleId) === next) advanceChain.delete(battleId);
+    });
+    return next;
+  }
+
+  async function advanceLocalLaneInner(battleId: string): Promise<void> {
+    const battle = await loadBattle(battleId);
+    if (!battle || battle.status !== 'running') return;
+
+    const contestants = await loadContestants(battleId);
+    const slots: ContestantSlot[] = contestants.map((c) => ({
+      id: c.id,
+      lane: c.lane,
+      status: c.status,
+    }));
+
+    // Nothing to do if the local lane is still busy.
+    const localRunning = slots.some((c) => c.lane === 'local' && c.status === 'running');
+    if (localRunning) return;
+
+    const nextId = nextLocalContestant(slots);
+    if (!nextId) return; // local queue is exhausted
+
+    const next = contestants.find((c) => c.id === nextId)!;
+    await dispatchContestant(battleId, battle.project_id, battle.battle_type, battle.prompt, {
+      id: next.id,
+      identity: next.identity,
+      model: next.model,
+      lane: next.lane,
+    });
+  }
+
+  // ─── handleTaskTerminal ───────────────────────────────────────────────────
+
+  function handleTaskTerminal(taskId: string, state: string): void {
+    void (async () => {
+      // Look up which contestant owns this task (contestants_task_id_idx).
+      const [row] = await sql<ContestantRow[]>`
+        SELECT id, battle_id, identity, model, lane, task_id, worktree_id, status
+        FROM contestants WHERE task_id = ${taskId}
+      `;
+      if (!row) return; // not an arena task — ignore
+      if (row.status !== 'running') return; // already settled (idempotent)
+
+      const battle = await loadBattle(row.battle_id);
+
+      // Pull the task row for benchmark + output.
+      const [task] = await sql<{
+        chat_id: string | null;
+        started_at: Date | null;
+        ended_at: Date | null;
+        cost_tokens: number | null;
+      }[]>`SELECT chat_id, started_at, ended_at, cost_tokens FROM tasks WHERE id = ${taskId}`;
+
+      const endedAt = task?.ended_at ?? new Date();
+
+      if (state === 'completed') {
+        const startedAt = task?.started_at ?? endedAt;
+        const bench = computeBenchmark(startedAt, endedAt, task?.cost_tokens ?? null, row.lane);
+
+        const output = task?.chat_id ? await readChatOutput(task.chat_id) : '';
+
+        const resultPath = battle
+          ? await writeContestantResults(battle, row, output, bench).catch((err) => {
+              log.warn({ err: errMsg(err), contestantId: row.id }, 'arena-runner: result write failed');
+              return null;
+            })
+          : null;
+
+        await sql`
+          UPDATE contestants
+          SET status      = 'done',
+              duration_ms = ${Math.round(bench.durationMs)},
+              tokens_per_sec = ${bench.tokensPerSec},
+              cost_tokens = ${task?.cost_tokens ?? null},
+              result_path = ${resultPath},
+              updated_at  = clock_timestamp()
+          WHERE id = ${row.id} AND status = 'running'
+        `;
+        teardownDeltaBridge(row.id);
+
+        // Check if this was the last contestant.
+        const allContestants = await loadContestants(row.battle_id);
+        const battleDone = isBattleComplete(allContestants);
+
+        publishContestantFrame(row.battle_id, row.id, {
+          status: 'done',
+          duration_ms: Math.round(bench.durationMs),
+          ...(bench.tokensPerSec !== null ? { tokens_per_sec: bench.tokensPerSec } : {}),
+          ...(battleDone ? { battle_status: 'completed' } : {}),
+        });
+
+        if (battleDone) {
+          await completeBattle(row.battle_id);
+        } else if (row.lane === 'local') {
+          void advanceLocalLane(row.battle_id);
+        }
+      } else {
+        // failed or cancelled — the contest continues; this contestant is error.
+        const errorMsg = state === 'cancelled' ? 'cancelled' : `task ${state}`;
+        await sql`
+          UPDATE contestants
+          SET status = 'error', error = ${errorMsg}, updated_at = clock_timestamp()
+          WHERE id = ${row.id} AND status = 'running'
+        `;
+        teardownDeltaBridge(row.id);
+
+        const allContestants = await loadContestants(row.battle_id);
+        const battleDone = isBattleComplete(allContestants);
+
+        publishContestantFrame(row.battle_id, row.id, {
+          status: 'error',
+          error: errorMsg,
+          ...(battleDone ? { battle_status: 'completed' } : {}),
+        });
+
+        if (battleDone) {
+          await completeBattle(row.battle_id);
+        } else if (row.lane === 'local') {
+          void advanceLocalLane(row.battle_id);
+        }
+      }
+    })().catch((err) => {
+      log.error({ err: errMsg(err), taskId }, 'arena-runner: handleTaskTerminal failed');
+    });
+  }
+
+  // ─── battle finalization ──────────────────────────────────────────────────
+
+  async function completeBattle(battleId: string): Promise<void> {
+    const updated = await sql`
+      UPDATE battles SET status = 'completed', updated_at = clock_timestamp()
+      WHERE id = ${battleId} AND status = 'running'
+    `;
+    if (updated.count === 0) return; // already terminal (race guard)
+    log.info({ battleId }, 'arena-runner: battle completed');
+
+    // Update manifest with finished_at timestamp.
+    const completedBattle = await loadBattle(battleId);
+    if (completedBattle?.results_path) {
+      const contestants = await loadContestants(battleId);
+      await writeManifest(
+        battleId,
+        completedBattle.results_path,
+        completedBattle.battle_type,
+        completedBattle.prompt,
+        completedBattle.created_at,
+        contestants.map((c) => ({ identity: c.identity, model: c.model, lane: c.lane })),
+        new Date(),
+      ).catch((err) => {
+        log.warn({ err: errMsg(err), battleId }, 'arena-runner: manifest update failed');
+      });
+    }
+
+    onBattleComplete(battleId);
+  }
+
+  // ─── manifest writer ─────────────────────────────────────────────────────
+
+  async function writeManifest(
+    battleId: string,
+    resultsPath: string,
+    battleType: BattleType,
+    prompt: string,
+    createdAt: Date,
+    contestants: Array<{ identity: string; model: string; lane: ContestantLane }>,
+    finishedAt: Date | null,
+  ): Promise<void> {
+    await mkdir(resultsPath, { recursive: true });
+    const manifest = {
+      id: battleId,
+      battle_type: battleType,
+      prompt,
+      contestants,
+      created_at: createdAt.toISOString(),
+      finished_at: finishedAt?.toISOString() ?? null,
+    };
+    await writeFile(join(resultsPath, 'manifest.json'), JSON.stringify(manifest, null, 2), 'utf8');
+  }
+
+  // ─── results writer ───────────────────────────────────────────────────────
+
+  async function writeContestantResults(
+    battle: BattleRow,
+    contestant: { identity: string; model: string; lane: ContestantLane; worktree_id: string | null },
+    output: string,
+    bench: { durationMs: number; tokensPerSec: number | null },
+  ): Promise<string> {
+    const resultsPath = await getOrBuildResultsPath(battle);
+    if (!resultsPath) throw new Error('cannot resolve results path for battle ' + battle.id);
+
+    const contestantDir = buildContestantDir(contestant.identity, contestant.model);
+    const dir = join(resultsPath, contestantDir);
+    await mkdir(dir, { recursive: true });
+
+    const benchLines = [
+      `duration: ${bench.durationMs}ms`,
+      bench.tokensPerSec != null ? `tokens/sec: ${bench.tokensPerSec.toFixed(1)}` : null,
+    ]
+      .filter(Boolean)
+      .join('\n');
+
+    const resultMd =
+      `# ${contestant.identity} / ${contestant.model}\n\n` +
+      `## Benchmark\n\n${benchLines}\n\n` +
+      `## Output\n\n${output}\n`;
+    await writeFile(join(dir, 'result.md'), resultMd, 'utf8');
+
+    if (battle.battle_type === 'coding' && contestant.worktree_id) {
+      const [wt] = await sql<{ path: string; base_commit: string | null }[]>`
+        SELECT path, base_commit FROM worktrees WHERE id = ${contestant.worktree_id}
+      `;
+      if (wt) {
+        const [proj] = await sql<{ path: string }[]>`
+          SELECT path FROM projects WHERE id = ${battle.project_id}
+        `;
+        if (proj) {
+          const diff = await diffWorktree(wt.path, proj.path, {
+            baseRef: wt.base_commit ?? undefined,
+          }).catch(() => '');
+          await writeFile(join(dir, 'diff.patch'), diff, 'utf8');
+        }
+      }
+    }
+
+    return dir;
+  }
+
+  /** Resolve or rebuild results_path for a battle (handles crash-before-UPDATE). */
+  async function getOrBuildResultsPath(battle: BattleRow): Promise<string | null> {
+    if (battle.results_path) return battle.results_path;
+    const [proj] = await sql<{ path: string }[]>`SELECT path FROM projects WHERE id = ${battle.project_id}`;
+    if (!proj) return null;
+    const slug = buildBattleSlug(battle.id, battle.battle_type, battle.created_at);
+    const resultsPath = join(proj.path, 'Arena', slug);
+    await sql`
+      UPDATE battles SET results_path = ${resultsPath}, updated_at = clock_timestamp()
+      WHERE id = ${battle.id}
+    `;
+    return resultsPath;
+  }
+
+  // ─── helpers ──────────────────────────────────────────────────────────────
+
+  async function readChatOutput(chatId: string): Promise<string> {
+    const [m] = await sql<{ content: string | null }[]>`
+      SELECT content FROM messages
+      WHERE chat_id = ${chatId} AND role = 'assistant'
+      ORDER BY created_at DESC LIMIT 1
+    `;
+    return m?.content ?? '';
+  }
+
+  async function loadBattle(battleId: string): Promise<BattleRow | null> {
+    const [b] = await sql<BattleRow[]>`
+      SELECT id, project_id, battle_type, prompt, status, results_path, created_at
+      FROM battles WHERE id = ${battleId}
+    `;
+    return b ?? null;
+  }
+
+  async function loadContestants(battleId: string): Promise<ContestantRow[]> {
+    return sql<ContestantRow[]>`
+      SELECT id, battle_id, identity, model, lane, task_id, worktree_id, status
+      FROM contestants WHERE battle_id = ${battleId}
+      ORDER BY created_at ASC
+    `;
+  }
+
+  function publishContestantFrame(
+    battleId: string,
+    contestantId: string,
+    extra: Record<string, unknown>,
+  ): void {
+    publishUser({
+      type: 'contestant_updated',
+      battle_id: battleId,
+      contestant_id: contestantId,
+      ...extra,
+    });
+  }
+
+  // ─── initResume ───────────────────────────────────────────────────────────
+
+  async function initResume(): Promise<void> {
+    const battles = await sql<BattleRow[]>`
+      SELECT id, project_id, battle_type, prompt, status, results_path, created_at
+      FROM battles WHERE status = 'running'
+    `;
+    if (battles.length === 0) return;
+    log.info({ count: battles.length }, 'arena-runner: resuming in-flight battles on startup');
+    for (const battle of battles) {
+      await resumeBattle(battle).catch((err) => {
+        log.error({ err: errMsg(err), battleId: battle.id }, 'arena-runner: initResume failed for battle');
+      });
+    }
+  }
+
+  async function resumeBattle(battle: BattleRow): Promise<void> {
+    const contestants = await loadContestants(battle.id);
+
+    const taskIds = contestants.map((c) => c.task_id).filter((id): id is string => id !== null);
+    const taskStates = new Map<string, string>();
+    if (taskIds.length > 0) {
+      const tasks = await sql<{ id: string; state: string }[]>`
+        SELECT id, state FROM tasks WHERE id = ANY(${taskIds})
+      `;
+      for (const t of tasks) taskStates.set(t.id, t.state);
+    }
+
+    const decisions = reconcileContestants(
+      contestants.map((c) => ({ contestantId: c.id, taskId: c.task_id, status: c.status })),
+      taskStates,
+    );
+
+    for (const decision of decisions) {
+      if (decision.action === 'keep') continue;
+      const contestant = contestants.find((c) => c.id === decision.contestantId)!;
+      await applyResumeDecision(battle, contestant, decision.action);
+    }
+
+    // Re-check completion after applying decisions.
+    const updated = await loadContestants(battle.id);
+    if (isBattleComplete(updated)) {
+      await completeBattle(battle.id);
+    } else {
+      // Advance local lane in case a slot opened up.
+      void advanceLocalLane(battle.id);
+    }
+
+    log.info({ battleId: battle.id }, 'arena-runner: battle resumed');
+  }
+
+  async function applyResumeDecision(
+    battle: BattleRow,
+    contestant: ContestantRow,
+    action: ContestantResumeAction,
+  ): Promise<void> {
+    switch (action) {
+      case 'keep': break;
+
+      case 'mark-done': {
+        const taskRow = contestant.task_id
+          ? (await sql<{ started_at: Date | null; ended_at: Date | null; cost_tokens: number | null; chat_id: string | null }[]>`
+              SELECT started_at, ended_at, cost_tokens, chat_id FROM tasks WHERE id = ${contestant.task_id}`)[0]
+          : null;
+        const endedAt = taskRow?.ended_at ?? new Date();
+        const startedAt = taskRow?.started_at ?? endedAt;
+        const bench = computeBenchmark(startedAt, endedAt, taskRow?.cost_tokens ?? null, contestant.lane);
+        const output = taskRow?.chat_id ? await readChatOutput(taskRow.chat_id) : '';
+        const resultPath = battle
+          ? await writeContestantResults(battle, contestant, output, bench).catch((err) => {
+              log.warn({ err: errMsg(err), contestantId: contestant.id }, 'arena-runner: resume result write failed');
+              return null;
+            })
+          : null;
+        await sql`
+          UPDATE contestants
+          SET status = 'done',
+              duration_ms = ${Math.round(bench.durationMs)},
+              tokens_per_sec = ${bench.tokensPerSec},
+              result_path = ${resultPath},
+              updated_at = clock_timestamp()
+          WHERE id = ${contestant.id}
+        `;
+        break;
+      }
+
+      case 'mark-error':
+        await sql`
+          UPDATE contestants
+          SET status = 'error', error = 'task failed before callback',
+              updated_at = clock_timestamp()
+          WHERE id = ${contestant.id}
+        `;
+        break;
+
+      case 'mark-cancelled':
+        await sql`
+          UPDATE contestants
+          SET status = 'error', error = 'cancelled before callback',
+              updated_at = clock_timestamp()
+          WHERE id = ${contestant.id}
+        `;
+        break;
+
+      case 're-dispatch': {
+        const { taskId } = await dispatch({
+          projectId: battle.project_id,
+          contestantId: contestant.id,
+          prompt: battle.prompt,
+          identity: contestant.identity,
+          model: contestant.model,
+          battleType: battle.battle_type,
+        });
+        await sql`
+          UPDATE contestants
+          SET task_id = ${taskId}, updated_at = clock_timestamp()
+          WHERE id = ${contestant.id}
+        `;
+        log.info(
+          { battleId: battle.id, contestantId: contestant.id, taskId },
+          'arena-runner: contestant re-dispatched on resume',
+        );
+        break;
+      }
+    }
+  }
+
+  // ─── cancelBattle ─────────────────────────────────────────────────────────
+
+  async function cancelBattle(battleId: string): Promise<{ cancelled: boolean; taskIds: string[] }> {
+    const updated = await sql`
+      UPDATE battles SET status = 'cancelled', updated_at = clock_timestamp()
+      WHERE id = ${battleId} AND status = 'running'
+    `;
+    if (updated.count === 0) return { cancelled: false, taskIds: [] };
+
+    // Mark all non-terminal contestants cancelled and collect in-flight task_ids.
+    const contestants = await sql<{ id: string; task_id: string | null; status: string }[]>`
+      SELECT id, task_id, status FROM contestants
+      WHERE battle_id = ${battleId} AND status NOT IN ('done', 'error')
+    `;
+
+    if (contestants.length > 0) {
+      await sql`
+        UPDATE contestants
+        SET status = 'error', error = 'battle cancelled', updated_at = clock_timestamp()
+        WHERE battle_id = ${battleId} AND status NOT IN ('done', 'error')
+      `;
+      for (const c of contestants) {
+        publishContestantFrame(battleId, c.id, {
+          status: 'error',
+          error: 'battle cancelled',
+          battle_status: 'cancelled',
+        });
+      }
+    }
+
+    const taskIds = contestants
+      .filter(
+        (c): c is typeof c & { task_id: string } =>
+          c.task_id !== null && c.status === 'running',
+      )
+      .map((c) => c.task_id);
+
+    log.info({ battleId }, 'arena-runner: battle cancelled by request');
+    return { cancelled: true, taskIds };
+  }
+
+  // ─── triggerAnalysis (Phase 5 seam) ──────────────────────────────────────
+
+  async function triggerAnalysis(battleId: string): Promise<{ triggered: boolean }> {
+    const battle = await loadBattle(battleId);
+    if (!battle) return { triggered: false };
+    log.info({ battleId }, 'arena-runner: triggerAnalysis requested');
+    // Calls the injected onBattleComplete seam — Phase 5 replaces this with the
+    // real two-stage digest→judge analyzer (see ADR 0002 + plan Phase 5).
+    onBattleComplete(battleId);
+    return { triggered: true };
+  }
+
+  // ─── startCrossExam (Phase 5 seam) ───────────────────────────────────────
+
+  async function startCrossExam(
+    battleId: string,
+    opts: { identity: string; model: string },
+  ): Promise<{ crossExamId: string }> {
+    const [row] = await sql<{ id: string }[]>`
+      INSERT INTO cross_examinations (battle_id, identity, model)
+      VALUES (${battleId}, ${opts.identity}, ${opts.model})
+      RETURNING id
+    `;
+    const crossExamId = row!.id;
+    log.info({ battleId, crossExamId, ...opts }, 'arena-runner: cross-exam inserted, triggering analyzer');
+    if (onCrossExamStart) {
+      try {
+        onCrossExamStart({ battleId, crossExamId, identity: opts.identity, model: opts.model });
+      } catch (err) {
+        log.error({ err: err instanceof Error ? err.message : String(err), battleId, crossExamId }, 'arena-runner: onCrossExamStart threw');
+      }
+    }
+    return { crossExamId };
+  }
+
+  // ─── setWinner (user override) ────────────────────────────────────────────
+
+  async function setWinner(
+    battleId: string,
+    winnerId: string | null,
+  ): Promise<{ ok: boolean; notFound?: boolean; invalidContestant?: boolean }> {
+    const [row] = await sql<{ id: string }[]>`SELECT id FROM battles WHERE id = ${battleId}`;
+    if (!row) return { ok: false, notFound: true };
+
+    if (winnerId !== null) {
+      const [c] = await sql<{ id: string }[]>`
+        SELECT id FROM contestants WHERE id = ${winnerId} AND battle_id = ${battleId}
+      `;
+      if (!c) return { ok: false, invalidContestant: true };
+    }
+
+    await sql`
+      UPDATE battles SET winner_contestant_id = ${winnerId}, updated_at = clock_timestamp()
+      WHERE id = ${battleId}
+    `;
+    publishUser({ type: 'battle_updated', battle_id: battleId, winner_contestant_id: winnerId });
+    return { ok: true };
+  }
+
+  return { startBattle, handleTaskTerminal, initResume, cancelBattle, triggerAnalysis, startCrossExam, setWinner };
+}
+
+function errMsg(e: unknown): string {
+  return e instanceof Error ? e.message : String(e);
+}

apps/coder/src/services/dispatcher.ts

@@ -4,7 +4,7 @@ import type { Broker } from '@boocode/server/broker';
 import type { WsFrame } from '@boocode/contracts/ws-frames';
 import type { Config } from '../config.js';
 import { createWorktree, diffWorktree, cleanupWorktree, ensureSessionWorktree } from './worktrees.js';
-import { applyAll } from './pending_changes.js';
+import { asPermissionMode } from './tools/types.js';
 import { createCheckpoint } from './checkpoints.js';
 import { makeDcpStreamStripper } from './dcp-strip.js';
 import { dispatchViaAcp } from './acp-dispatch.js';
@@ -32,7 +32,13 @@ import {
 import { shouldFailOnMissingAgent } from './flow-runner-decisions.js';
 
 interface InferenceRunner {
-  enqueue: (sessionId: string, chatId: string, assistantId: string, user: string) => void;
+  enqueue: (
+    sessionId: string,
+    chatId: string,
+    assistantId: string,
+    user: string,
+    permissionMode?: 'plan' | 'ask' | 'bypass',
+  ) => void;
   cancel: (sessionId: string, chatId: string) => Promise<boolean>;
   hasActive: (chatId: string) => boolean;
 }
@@ -310,6 +316,9 @@ export function createDispatcher(deps: Deps): {
     const taskId = task.id;
     log.info({ taskId }, 'dispatcher: starting task (path A — native)');
 
+    // Declared before try so the catch block can write it back on the task row.
+    let chatId: string | null = null;
+
     try {
       // Mark running
       await sql`
@@ -318,26 +327,29 @@ export function createDispatcher(deps: Deps): {
         WHERE id = ${taskId}
       `;
 
-      // Create session + chat for this task
+      // Session setup: reuse a pre-created session (e.g. Q&A arena contestants
+      // whose persona is stamped on the session via agent_id) or create a fresh one.
       const model = task.model ?? config.DEFAULT_MODEL;
-      const sessionName = 'Task: ' + task.input.slice(0, 40);
-
-      const [session] = await sql<{ id: string }[]>`
-        INSERT INTO sessions (project_id, name, model, status)
-        VALUES (${task.project_id}, ${sessionName}, ${model}, 'open')
-        RETURNING id
-      `;
-      const sessionId = session!.id;
+      let sessionId: string;
+      if (task.session_id) {
+        sessionId = task.session_id;
+      } else {
+        const sessionName = 'Task: ' + task.input.slice(0, 40);
+        const [session] = await sql<{ id: string }[]>`
+          INSERT INTO sessions (project_id, name, model, status)
+          VALUES (${task.project_id}, ${sessionName}, ${model}, 'open')
+          RETURNING id
+        `;
+        sessionId = session!.id;
+        await sql`UPDATE tasks SET session_id = ${sessionId} WHERE id = ${taskId}`;
+      }
 
       const [chat] = await sql<{ id: string }[]>`
         INSERT INTO chats (session_id, name, status)
         VALUES (${sessionId}, 'Task execution', 'open')
         RETURNING id
       `;
-      const chatId = chat!.id;
-
-      // Link task to session
-      await sql`UPDATE tasks SET session_id = ${sessionId} WHERE id = ${taskId}`;
+      chatId = chat!.id;
 
       // Create user message + streaming assistant
       await sql<{ id: string }[]>`
@@ -352,8 +364,9 @@ export function createDispatcher(deps: Deps): {
       `;
       const assistantId = assistantMsg!.id;
 
-      // Enqueue inference
-      inference.enqueue(sessionId, chatId, assistantId, 'default');
+      // Enqueue inference — pass the native permission gate (plan/ask/bypass)
+      // through to the write-tool context. Non-unified mode ids → undefined.
+      inference.enqueue(sessionId, chatId, assistantId, 'default', asPermissionMode(task.mode_id));
 
       // Wait for inference to complete (poll message status)
       const finalStatus = await waitForCompletion(assistantId);
@@ -382,26 +395,10 @@ export function createDispatcher(deps: Deps): {
         const summary = (msg?.content ?? '').slice(0, 500);
         await sql`
           UPDATE tasks
-          SET state = 'completed', ended_at = clock_timestamp(), output_summary = ${summary}, cost_tokens = ${costTokens}
+          SET state = 'completed', ended_at = clock_timestamp(), output_summary = ${summary}, cost_tokens = ${costTokens}, chat_id = ${chatId}
           WHERE id = ${taskId}
         `;
         log.info({ taskId, costTokens }, 'dispatcher: task completed (native)');
-        // Bypass permission mode: auto-apply the staged edits to disk after the
-        // turn. Ask/Plan leave them in the pending-changes queue for review.
-        if (task.mode_id === 'bypass') {
-          try {
-            const [proj] = await sql<{ path: string }[]>`SELECT path FROM projects WHERE id = ${task.project_id}`;
-            if (proj?.path) {
-              const applied = await applyAll(sql, sessionId, proj.path);
-              log.info({ taskId, applied: applied.length }, 'dispatcher: native bypass auto-applied pending changes');
-            }
-          } catch (applyErr) {
-            log.warn(
-              { taskId, err: applyErr instanceof Error ? applyErr.message : String(applyErr) },
-              'dispatcher: native bypass auto-apply failed',
-            );
-          }
-        }
       } else {
         const [msg] = await sql<{ content: string | null }[]>`
           SELECT content FROM messages WHERE id = ${assistantId}
@@ -409,7 +406,7 @@ export function createDispatcher(deps: Deps): {
         const summary = (msg?.content ?? 'Inference failed').slice(0, 500);
         await sql`
           UPDATE tasks
-          SET state = 'failed', ended_at = clock_timestamp(), output_summary = ${summary}, cost_tokens = ${costTokens}
+          SET state = 'failed', ended_at = clock_timestamp(), output_summary = ${summary}, cost_tokens = ${costTokens}, chat_id = ${chatId}
           WHERE id = ${taskId}
         `;
         log.warn({ taskId, finalStatus }, 'dispatcher: task failed (native)');
@@ -419,7 +416,7 @@ export function createDispatcher(deps: Deps): {
       log.error({ taskId, err: errMsg }, 'dispatcher: task error (native)');
       await sql`
         UPDATE tasks
-        SET state = 'failed', ended_at = clock_timestamp(), output_summary = ${errMsg.slice(0, 500)}
+        SET state = 'failed', ended_at = clock_timestamp(), output_summary = ${errMsg.slice(0, 500)}, chat_id = ${chatId}
         WHERE id = ${taskId}
       `.catch(() => {});
     }

apps/coder/src/services/fuzzy-match.ts

@@ -21,7 +21,16 @@
 //                         punctuation to ASCII on both sides; the match is
 //                         mapped back to original offsets.
 //   4. levenshtein      — best line-window by normalized edit-distance
-//                         similarity; accepted only at >= SIMILARITY_THRESHOLD.
+//                         similarity; accepted only at >= SIMILARITY_THRESHOLD,
+//                         anchored on an exact first+last line for multi-line
+//                         needles, and REFUSED (ambiguous) when a second window
+//                         scores within AMBIGUITY_EPSILON of the best. Like the
+//                         exact/whitespace tiers, this tier fails CLOSED — it
+//                         never splices over a merely-plausible guess, because a
+//                         wrong-window splice corrupts the file (it leaves the
+//                         real target intact and duplicates it). This mirrors
+//                         opencode/cline/qwen, whose fuzzy tiers all keep the
+//                         unique-match requirement rather than picking a winner.
 //
 // Pure and dependency-free (Levenshtein is the standard iterative two-row DP),
 // reimplemented from the general technique — no vendored source.
@@ -31,8 +40,31 @@ export type MatchResult =
   | { kind: 'ambiguous'; count: number }
   | { kind: 'not_found' };
 
-/** Levenshtein similarity floor for the final fuzzy fallback (strategy 4). */
-export const SIMILARITY_THRESHOLD = 0.66;
+/**
+ * Levenshtein similarity floor for the final fuzzy fallback (strategy 4).
+ * 0.66 was far too low — at two-thirds similarity a structurally-wrong window
+ * (e.g. one of three near-identical form blocks) clears the bar and gets spliced
+ * over, leaving the real target intact and duplicated. Competent agents anchor
+ * far tighter (opencode's BlockAnchor needs an exact anchor; cline needs exact
+ * first+last lines). 0.85 keeps genuine quantized-model drift (a typo, an indent
+ * shift) while refusing a different block.
+ */
+export const SIMILARITY_THRESHOLD = 0.85;
+
+/**
+ * If a second candidate window scores within this of the best, the match is
+ * ambiguous and tier 4 refuses rather than guessing — the same fail-closed
+ * stance the exact and whitespace tiers take on multiple hits. Repetitive files
+ * (the duplicate-block corruption case) produce near-tied windows; this is what
+ * turns that into a clean "add more context" error instead of a wrong splice.
+ */
+export const AMBIGUITY_EPSILON = 0.05;
+
+/** Multi-line needles at or above this length must anchor on an exact (after
+ *  trim + unicode-fold) first AND last line before similarity is even scored —
+ *  the cline/opencode block-anchor rule. Below it, threshold + uniqueness alone
+ *  guard the match. */
+const ANCHOR_MIN_LINES = 3;
 
 export function locateMatch(content: string, needle: string): MatchResult {
   // Empty needle has no meaningful match.
@@ -252,20 +284,39 @@ function locateByLevenshtein(content: string, needle: string): MatchResult | nul
 
   const needleJoined = needleLines.map((l) => l.trim()).join('\n');
 
-  let best = -1;
-  let bestSpan: { start: number; end: number } | null = null;
+  // Block-anchor gate for multi-line needles: the first and last lines must match
+  // exactly (after trim + unicode-fold) or the window is not even scored. This
+  // stops a high interior-similarity from dragging a structurally-wrong window
+  // over the threshold — the failure that duplicates blocks in repetitive files.
+  const anchored = n >= ANCHOR_MIN_LINES;
+  const needleFirst = canonicalize(needleLines[0]!.trim());
+  const needleLast = canonicalize(needleLines[n - 1]!.trim());
+
+  const scored: Array<{ score: number; start: number; end: number }> = [];
   for (let i = 0; i + n <= contentLines.length; i++) {
     const window = contentLines.slice(i, i + n);
-    const windowJoined = window.map((l) => l.text.trim()).join('\n');
-    const score = similarity(windowJoined, needleJoined);
-    if (score > best) {
-      best = score;
-      bestSpan = { start: window[0]!.start, end: window[n - 1]!.end };
+    if (anchored) {
+      const winFirst = canonicalize(window[0]!.text.trim());
+      const winLast = canonicalize(window[n - 1]!.text.trim());
+      if (winFirst !== needleFirst || winLast !== needleLast) continue;
     }
+    const windowJoined = window.map((l) => l.text.trim()).join('\n');
+    scored.push({
+      score: similarity(windowJoined, needleJoined),
+      start: window[0]!.start,
+      end: window[n - 1]!.end,
+    });
   }
 
-  if (bestSpan && best >= SIMILARITY_THRESHOLD) {
-    return { kind: 'fuzzy', start: bestSpan.start, end: bestSpan.end };
-  }
-  return null;
+  if (scored.length === 0) return null;
+  scored.sort((a, b) => b.score - a.score);
+  const best = scored[0]!;
+  if (best.score < SIMILARITY_THRESHOLD) return null;
+
+  // Uniqueness guard: refuse when a second window is within epsilon of the best.
+  // Fail closed (ambiguous) rather than silently splicing one of several lookalikes.
+  const tied = scored.filter((s) => s.score >= best.score - AMBIGUITY_EPSILON);
+  if (tied.length > 1) return { kind: 'ambiguous', count: tied.length };
+
+  return { kind: 'fuzzy', start: best.start, end: best.end };
 }

apps/coder/src/services/pending_changes.ts

@@ -1,9 +1,120 @@
-import { readFile, writeFile, unlink, mkdir } from 'node:fs/promises';
-import { dirname } from 'node:path';
+import { readFile, writeFile, unlink, mkdir, rename, realpath } from 'node:fs/promises';
+import { dirname, join, basename } from 'node:path';
+import { randomBytes } from 'node:crypto';
 import type { Sql } from '../db.js';
 import { resolveWritePath } from './write_guard.js';
 import { locateMatch } from './fuzzy-match.js';
 
+/**
+ * Write a file atomically: stage to a sibling temp file, then rename over the
+ * target. rename(2) on the same filesystem is atomic, so a crash mid-write can
+ * never leave a half-written (truncated/corrupt) source file — readers see
+ * either the old content or the complete new content. The temp lives in the same
+ * directory to guarantee a same-filesystem rename.
+ *
+ * Symlinks: a plain writeFile FOLLOWS a symlink and writes through to its target;
+ * a bare rename would REPLACE the link with a regular file. We realpath an
+ * existing target first so the rename lands on the real file and the link
+ * survives — preserving the prior follow-through behavior. A missing target
+ * (create, or a broken link) just writes the literal path.
+ */
+async function writeFileAtomic(filePath: string, content: string): Promise<void> {
+  let target = filePath;
+  try {
+    target = await realpath(filePath);
+  } catch {
+    // ENOENT (new file) or broken link — write the literal path.
+  }
+  const tmp = join(dirname(target), `.${basename(target)}.tmp.${process.pid}.${randomBytes(6).toString('hex')}`);
+  await writeFile(tmp, content, 'utf8');
+  try {
+    await rename(tmp, target);
+  } catch (err) {
+    await unlink(tmp).catch(() => {});
+    throw err;
+  }
+}
+
+/** Detect a file's dominant line ending so an edit can preserve it. */
+function detectEol(text: string): '\r\n' | '\n' {
+  return text.includes('\r\n') ? '\r\n' : '\n';
+}
+
+/**
+ * Serialize the read-modify-write of a single file so two concurrent applies
+ * (e.g. two chat tabs sharing one worktree, or a Bypass write racing an
+ * apply_pending) can't lose an update. In-process keying is sufficient —
+ * BooCoder is a single Fastify process. One Map entry per distinct path.
+ */
+const fileLocks = new Map<string, Promise<void>>();
+async function withFileLock<T>(filePath: string, fn: () => Promise<T>): Promise<T> {
+  const prev = fileLocks.get(filePath) ?? Promise.resolve();
+  let release!: () => void;
+  const current = new Promise<void>((r) => { release = r; });
+  fileLocks.set(filePath, prev.then(() => current));
+  await prev.catch(() => {});
+  try {
+    return await fn();
+  } finally {
+    release();
+  }
+}
+
+// --- Edit-apply planning (pure, unit-tested) ---------------------------------
+
+/**
+ * Decision for applying one queued edit to a file's current content. Pulled out
+ * of `applyOne` so the splice — the part that actually corrupted files — is pure
+ * and testable without a DB or filesystem. Mirrors how opencode/cline/qwen keep
+ * their matchers fail-closed and idempotent.
+ */
+export type EditPlan =
+  | { kind: 'apply'; updated: string }
+  | { kind: 'noop'; reason: 'identical' | 'already-applied' }
+  | { kind: 'ambiguous'; count: number }
+  | { kind: 'not_found' };
+
+/**
+ * Decide how (or whether) to apply an `old → new` edit to `content`.
+ *
+ * Idempotency is the whole point here: a queued edit can legitimately be
+ * re-applied (a local model re-emits the same tool call; a turn is retried; the
+ * same change sits in the queue twice). A naive splice stamps the new text again
+ * each time — the 2–3× block duplication. Two guards make re-application a no-op:
+ *
+ *  - already-applied (anchored insert): when `new` is `old` + an appended block
+ *    (`old="anchor"`, `new="anchor\n<block>"`), `old` still matches uniquely after
+ *    the first apply, so a second apply would duplicate `<block>`. If the full
+ *    `new` text is already present at the match site, the edit is already applied.
+ *  - already-applied (old gone): if `old` can't be located but `new` is already
+ *    in the file, the change landed on a prior pass — treat as a no-op, not an error.
+ *  - identical: the splice would not change the file.
+ *
+ * Anything ambiguous or genuinely absent fails CLOSED so the caller surfaces a
+ * correctable error instead of writing a guess.
+ */
+export function planEdit(content: string, oldStr: string, newStr: string): EditPlan {
+  const match = locateMatch(content, oldStr);
+
+  if (match.kind === 'ambiguous') return { kind: 'ambiguous', count: match.count };
+
+  if (match.kind === 'not_found') {
+    if (newStr.length > 0 && content.includes(newStr)) {
+      return { kind: 'noop', reason: 'already-applied' };
+    }
+    return { kind: 'not_found' };
+  }
+
+  const updated = content.slice(0, match.start) + newStr + content.slice(match.end);
+  // No-change splice first (covers old === new), then the anchored re-stamp guard:
+  // the full replacement already sits at the match site (re-emitted anchored insert).
+  if (updated === content) return { kind: 'noop', reason: 'identical' };
+  if (content.slice(match.start, match.start + newStr.length) === newStr) {
+    return { kind: 'noop', reason: 'already-applied' };
+  }
+  return { kind: 'apply', updated };
+}
+
 // --- Types -------------------------------------------------------------------
 
 export interface PendingChange {
@@ -47,6 +158,13 @@ export async function queueEdit(
   const resolved = resolveWritePath(projectRoot, filePath);
   const diff = JSON.stringify({ old: oldString, new: newString });
 
+  // Idempotent queue: collapse an identical edit that is still pending. Local
+  // quantized models re-emit the same edit_file call within a turn, and a retried
+  // turn re-queues — each duplicate row would apply and stamp another copy. One
+  // pending row per (session, file, operation, diff) is enough.
+  const existing = await findPendingDuplicate(sql, sessionId, resolved, 'edit', diff);
+  if (existing) return existing;
+
   const [row] = await sql<PendingChange[]>`
     INSERT INTO pending_changes (session_id, task_id, file_path, operation, diff, agent)
     VALUES (${sessionId}, ${taskId}, ${resolved}, 'edit', ${diff}, ${agent})
@@ -55,6 +173,28 @@ export async function queueEdit(
   return row!;
 }
 
+/** Return an identical still-pending change for this (session, file, op, diff),
+ *  or undefined. Used to keep the queue idempotent against re-emitted edits. */
+async function findPendingDuplicate(
+  sql: Sql,
+  sessionId: string,
+  resolvedPath: string,
+  operation: 'create' | 'edit' | 'delete',
+  diff: string,
+): Promise<PendingChange | undefined> {
+  const [row] = await sql<PendingChange[]>`
+    SELECT * FROM pending_changes
+    WHERE session_id = ${sessionId}
+      AND file_path = ${resolvedPath}
+      AND operation = ${operation}
+      AND diff = ${diff}
+      AND status = 'pending'
+    ORDER BY created_at ASC
+    LIMIT 1
+  `;
+  return row;
+}
+
 export async function queueCreate(
   sql: Sql,
   sessionId: string,
@@ -68,6 +208,9 @@ export async function queueCreate(
 ): Promise<PendingChange> {
   const resolved = resolveWritePath(projectRoot, filePath);
 
+  const existing = await findPendingDuplicate(sql, sessionId, resolved, 'create', content);
+  if (existing) return existing;
+
   const [row] = await sql<PendingChange[]>`
     INSERT INTO pending_changes (session_id, task_id, file_path, operation, diff, agent)
     VALUES (${sessionId}, ${taskId}, ${resolved}, 'create', ${content}, ${agent})
@@ -87,6 +230,9 @@ export async function queueDelete(
 ): Promise<PendingChange> {
   const resolved = resolveWritePath(projectRoot, filePath);
 
+  const existing = await findPendingDuplicate(sql, sessionId, resolved, 'delete', '');
+  if (existing) return existing;
+
   const [row] = await sql<PendingChange[]>`
     INSERT INTO pending_changes (session_id, task_id, file_path, operation, diff, agent)
     VALUES (${sessionId}, ${taskId}, ${resolved}, 'delete', '', ${agent})
@@ -110,48 +256,60 @@ export async function applyOne(
   }
 
   try {
-    // Re-validate path in case projectRoot has shifted
-    resolveWritePath(projectRoot, change.file_path);
+    return await withFileLock(change.file_path, async () => {
+      // Re-validate path in case projectRoot has shifted
+      resolveWritePath(projectRoot, change.file_path);
 
-    switch (change.operation) {
-      case 'create': {
-        await mkdir(dirname(change.file_path), { recursive: true });
-        await writeFile(change.file_path, change.diff, 'utf8');
-        break;
-      }
-      case 'edit': {
-        const { old: oldStr, new: newStr } = JSON.parse(change.diff) as { old: string; new: string };
-        const content = await readFile(change.file_path, 'utf8');
-        const match = locateMatch(content, oldStr);
-        if (match.kind === 'ambiguous') {
-          throw new Error(
-            `old_string matches ${match.count} locations — add surrounding context to disambiguate`,
-          );
+      switch (change.operation) {
+        case 'create': {
+          await mkdir(dirname(change.file_path), { recursive: true });
+          await writeFileAtomic(change.file_path, change.diff);
+          break;
         }
-        if (match.kind === 'not_found') {
-          throw new Error(
-            'old_string not found in file (even fuzzily) — file may have changed since the edit was queued',
-          );
+        case 'edit': {
+          const { old: oldStr, new: newStr } = JSON.parse(change.diff) as { old: string; new: string };
+          const raw = await readFile(change.file_path, 'utf8');
+          // Normalize to LF for matching, then write back in the file's native EOL
+          // so an LF-emitting model doesn't leave a CRLF file with mixed endings.
+          const eol = detectEol(raw);
+          const toLf = (t: string) => t.replaceAll('\r\n', '\n');
+          const plan = planEdit(toLf(raw), toLf(oldStr), toLf(newStr));
+          if (plan.kind === 'ambiguous') {
+            throw new Error(
+              `old_string matches ${plan.count} locations — add surrounding context to disambiguate`,
+            );
+          }
+          if (plan.kind === 'not_found') {
+            throw new Error(
+              'old_string not found in file (even fuzzily) — file may have changed since the edit was queued',
+            );
+          }
+          if (plan.kind === 'apply') {
+            const out = eol === '\r\n' ? plan.updated.replaceAll('\n', '\r\n') : plan.updated;
+            await writeFileAtomic(change.file_path, out);
+          } else {
+            // noop: the edit is already applied (re-emitted / retried) or a no-change.
+            // Mark it applied without rewriting so it can't stamp a duplicate.
+            console.log(`[pending] edit ${change.file_path} is a no-op (${plan.reason}) — not rewriting`);
+          }
+          break;
         }
-        const updated = content.slice(0, match.start) + newStr + content.slice(match.end);
-        await writeFile(change.file_path, updated, 'utf8');
-        break;
-      }
-      case 'delete': {
-        // Stash current content in diff for potential rewind
-        try {
-          const existing = await readFile(change.file_path, 'utf8');
-          await sql`UPDATE pending_changes SET diff = ${existing} WHERE id = ${changeId}`;
-        } catch {
-          // File may already be gone — proceed with status update
+        case 'delete': {
+          // Stash current content in diff for potential rewind
+          try {
+            const existing = await readFile(change.file_path, 'utf8');
+            await sql`UPDATE pending_changes SET diff = ${existing} WHERE id = ${changeId}`;
+          } catch {
+            // File may already be gone — proceed with status update
+          }
+          await unlink(change.file_path);
+          break;
         }
-        await unlink(change.file_path);
-        break;
       }
-    }
 
-    await sql`UPDATE pending_changes SET status = 'applied' WHERE id = ${changeId}`;
-    return { id: change.id, file_path: change.file_path, operation: change.operation, success: true };
+      await sql`UPDATE pending_changes SET status = 'applied' WHERE id = ${changeId}`;
+      return { id: change.id, file_path: change.file_path, operation: change.operation, success: true };
+    });
   } catch (err) {
     const message = err instanceof Error ? err.message : String(err);
     return { id: change.id, file_path: change.file_path, operation: change.operation, success: false, error: message };
@@ -220,13 +378,13 @@ export async function rewindOne(
           );
         }
         const reverted = content.slice(0, match.start) + oldStr + content.slice(match.end);
-        await writeFile(change.file_path, reverted, 'utf8');
+        await writeFileAtomic(change.file_path, reverted);
         break;
       }
       case 'delete': {
         // Reverse a delete: recreate the file (diff holds the original content stashed at apply time)
         await mkdir(dirname(change.file_path), { recursive: true });
-        await writeFile(change.file_path, change.diff, 'utf8');
+        await writeFileAtomic(change.file_path, change.diff);
         break;
       }
     }

apps/coder/src/services/tools/__tests__/inference_context.test.ts

@@ -0,0 +1,38 @@
+import { describe, it, expect } from 'vitest';
+import { runWithInferenceContext, getInferenceContext } from '../inference_context.js';
+import type { Sql } from '../../../db.js';
+
+const fakeSql = {} as unknown as Sql;
+
+describe('inference context (AsyncLocalStorage isolation)', () => {
+  it('throws when read outside a run', () => {
+    expect(() => getInferenceContext()).toThrow(/outside inference context/);
+  });
+
+  it('keeps each run its own context across overlapping awaits', async () => {
+    // The race the global `let current` had: run B starts (and would overwrite a
+    // shared global) while run A is awaiting. After A resumes it must still read
+    // its OWN sessionId, not B's.
+    const run = (id: string, delay: number) =>
+      runWithInferenceContext({ sql: fakeSql, sessionId: id, taskId: null }, async () => {
+        await new Promise((r) => setTimeout(r, delay));
+        return getInferenceContext().sessionId;
+      });
+
+    const [a, b] = await Promise.all([run('A', 20), run('B', 5)]);
+    expect(a).toBe('A');
+    expect(b).toBe('B');
+  });
+
+  it('carries permissionMode and taskId per run', async () => {
+    const result = await runWithInferenceContext(
+      { sql: fakeSql, sessionId: 's1', taskId: 't1', permissionMode: 'bypass' },
+      async () => {
+        await Promise.resolve();
+        const ctx = getInferenceContext();
+        return { taskId: ctx.taskId, mode: ctx.permissionMode };
+      },
+    );
+    expect(result).toEqual({ taskId: 't1', mode: 'bypass' });
+  });
+});

apps/coder/src/services/tools/apply_pending.ts

@@ -26,6 +26,15 @@ export const applyPendingTool: ToolDef<ApplyPendingInputT> = {
     },
   },
   async execute(_input: ApplyPendingInputT, projectRoot: string, context: ToolContext): Promise<unknown> {
+    // Under Ask (and Plan) the human approves via the Pending Changes panel — the
+    // agent must not auto-apply. Bypass and legacy (undefined) may apply.
+    if (context.permissionMode === 'ask' || context.permissionMode === 'plan') {
+      return {
+        status: 'denied',
+        message:
+          'Permission mode is Ask — staged changes must be approved by the user in the Pending Changes panel, not applied by the agent.',
+      };
+    }
     const results = await applyAll(context.sql, context.sessionId, projectRoot);
     const succeeded = results.filter((r) => r.success).length;
     const failed = results.filter((r) => !r.success).length;

apps/coder/src/services/tools/create_file.ts

@@ -1,6 +1,7 @@
 import { z } from 'zod';
 import type { ToolDef, ToolContext } from './types.js';
 import { queueCreate } from '../pending_changes.js';
+import { denyReadOnly, finalizeWrite } from './write-gate.js';
 
 const CreateFileInput = z.object({
   file_path: z.string().min(1),
@@ -32,6 +33,7 @@ export const createFileTool: ToolDef<CreateFileInputT> = {
     },
   },
   async execute(input: CreateFileInputT, projectRoot: string, context: ToolContext): Promise<unknown> {
+    if (context.permissionMode === 'plan') return denyReadOnly('create_file');
     const change = await queueCreate(
       context.sql,
       context.sessionId,
@@ -40,12 +42,11 @@ export const createFileTool: ToolDef<CreateFileInputT> = {
       input.content,
       projectRoot,
     );
-    return {
-      status: 'queued',
-      change_id: change.id,
-      file_path: change.file_path,
-      operation: 'create',
-      message: `File creation queued: ${change.file_path}. Use apply_pending to write changes to disk.`,
-    };
+    return finalizeWrite(
+      context,
+      projectRoot,
+      change,
+      `File creation queued: ${change.file_path}. Use apply_pending to write changes to disk.`,
+    );
   },
 };

apps/coder/src/services/tools/delete_file.ts

@@ -1,6 +1,7 @@
 import { z } from 'zod';
 import type { ToolDef, ToolContext } from './types.js';
 import { queueDelete } from '../pending_changes.js';
+import { denyReadOnly, finalizeWrite } from './write-gate.js';
 
 const DeleteFileInput = z.object({
   file_path: z.string().min(1),
@@ -30,6 +31,7 @@ export const deleteFileTool: ToolDef<DeleteFileInputT> = {
     },
   },
   async execute(input: DeleteFileInputT, projectRoot: string, context: ToolContext): Promise<unknown> {
+    if (context.permissionMode === 'plan') return denyReadOnly('delete_file');
     const change = await queueDelete(
       context.sql,
       context.sessionId,
@@ -37,12 +39,11 @@ export const deleteFileTool: ToolDef<DeleteFileInputT> = {
       input.file_path,
       projectRoot,
     );
-    return {
-      status: 'queued',
-      change_id: change.id,
-      file_path: change.file_path,
-      operation: 'delete',
-      message: `File deletion queued: ${change.file_path}. Use apply_pending to write changes to disk.`,
-    };
+    return finalizeWrite(
+      context,
+      projectRoot,
+      change,
+      `File deletion queued: ${change.file_path}. Use apply_pending to write changes to disk.`,
+    );
   },
 };

apps/coder/src/services/tools/edit_file.ts

@@ -1,6 +1,7 @@
 import { z } from 'zod';
 import type { ToolDef, ToolContext } from './types.js';
 import { queueEdit } from '../pending_changes.js';
+import { denyReadOnly, finalizeWrite } from './write-gate.js';
 
 const EditFileInput = z.object({
   file_path: z.string().min(1),
@@ -34,6 +35,7 @@ export const editFileTool: ToolDef<EditFileInputT> = {
     },
   },
   async execute(input: EditFileInputT, projectRoot: string, context: ToolContext): Promise<unknown> {
+    if (context.permissionMode === 'plan') return denyReadOnly('edit_file');
     const change = await queueEdit(
       context.sql,
       context.sessionId,
@@ -43,12 +45,11 @@ export const editFileTool: ToolDef<EditFileInputT> = {
       input.new_string,
       projectRoot,
     );
-    return {
-      status: 'queued',
-      change_id: change.id,
-      file_path: change.file_path,
-      operation: 'edit',
-      message: `Edit queued for ${change.file_path}. Use apply_pending to write changes to disk.`,
-    };
+    return finalizeWrite(
+      context,
+      projectRoot,
+      change,
+      `Edit queued for ${change.file_path}. Use apply_pending to write changes to disk.`,
+    );
   },
 };

apps/coder/src/services/tools/inference_context.ts

@@ -1,36 +1,49 @@
+import { AsyncLocalStorage } from 'node:async_hooks';
 import type { Sql } from '../../db.js';
+import type { PermissionMode } from './types.js';
 
 /**
- * Module-level inference context for write tools.
+ * Per-run inference context for write tools.
  *
- * Set via `setInferenceContext()` before each inference run starts.
- * Write tools read it via `getInferenceContext()` during execute.
- * Same pattern as BooChat's `loadConfig()` singleton — tools need
- * ambient state that can't be threaded through the tool-phase execute
- * signature (which is `execute(input, projectRoot, extraRoots?)`).
+ * Write tools need ambient state (sql, sessionId, the permission gate) that the
+ * BooChat tool-phase `execute(input, projectRoot, extraRoots?)` signature can't
+ * carry. This used to be a single module-level `let current` — but the inference
+ * runner's `enqueue()` is fire-and-forget, so two overlapping runs (a user
+ * message racing a dispatcher-polled native task; two chat tabs streaming) would
+ * clobber each other's context, and `cancel()` cleared it for ALL in-flight runs.
+ *
+ * AsyncLocalStorage gives each run its own context: `enqueue()` starts its async
+ * loop synchronously inside `runWithInferenceContext`, so the store propagates
+ * through every awaited tool execution in that run — and only that run.
  */
 
 export interface InferenceContext {
   sql: Sql;
   sessionId: string;
   taskId: string | null;
+  /** Native-BooCode permission gate, set per run from the request/task mode. */
+  permissionMode?: PermissionMode;
 }
 
-let current: InferenceContext | null = null;
+const storage = new AsyncLocalStorage<InferenceContext>();
 
-export function setInferenceContext(ctx: InferenceContext): void {
-  current = ctx;
-}
-
-export function clearInferenceContext(): void {
-  current = null;
+/**
+ * Bind `ctx` for the duration of the (possibly detached) async chain `fn` starts.
+ * The inference runner kicks off its loop synchronously within this call, so all
+ * downstream `await`s — including write-tool `execute` via the adapter — read the
+ * same store. Concurrent runs each get their own; nothing is shared or cleared
+ * out from under an in-flight run.
+ */
+export function runWithInferenceContext<T>(ctx: InferenceContext, fn: () => T): T {
+  return storage.run(ctx, fn);
 }
 
 export function getInferenceContext(): InferenceContext {
-  if (!current) {
+  const ctx = storage.getStore();
+  if (!ctx) {
     throw new Error(
-      'Write tool called outside inference context — setInferenceContext() was not called before this run',
+      'Write tool called outside inference context — runWithInferenceContext() did not wrap this run',
     );
   }
-  return current;
+  return ctx;
 }

apps/coder/src/services/tools/types.ts

@@ -1,6 +1,22 @@
 import type { z } from 'zod';
 import type { Sql } from '../../db.js';
 
+/**
+ * Unified permission ladder for native BooCode inference. Gates the write tools:
+ *   plan   — read-only: create/edit/delete are denied (no staging).
+ *   ask    — stage to the pending-changes queue; `apply_pending` is denied so the
+ *            agent cannot self-apply (the human approves via the Diff panel).
+ *   bypass — apply each write immediately (no queue, no approval).
+ * Undefined preserves the historical behavior (stage + `apply_pending` allowed).
+ */
+export type PermissionMode = 'plan' | 'ask' | 'bypass';
+
+/** Narrow a raw task/request mode id to a unified PermissionMode, else undefined
+ *  (e.g. an external agent's native mode id, or null). */
+export function asPermissionMode(id: string | null | undefined): PermissionMode | undefined {
+  return id === 'plan' || id === 'ask' || id === 'bypass' ? id : undefined;
+}
+
 export interface ToolJsonSchema {
   type: 'function';
   function: {
@@ -21,6 +37,8 @@ export interface ToolContext {
   sql: Sql;
   sessionId: string;
   taskId: string | null;
+  /** Native-BooCode permission gate for write tools (undefined = legacy behavior). */
+  permissionMode?: PermissionMode;
 }
 
 export interface ToolDef<TInput> {

apps/coder/src/services/tools/write-gate.ts

@@ -0,0 +1,53 @@
+/**
+ * Permission-gate helpers for native BooCode write tools. The gate comes from
+ * the per-run inference context (`ToolContext.permissionMode`):
+ *   plan   — deny the write (read-only); nothing is staged.
+ *   bypass — apply the staged change immediately (no queue, no approval).
+ *   ask / undefined — leave it in the pending-changes queue for review.
+ */
+import type { ToolContext } from './types.js';
+import { applyOne } from '../pending_changes.js';
+
+/** Result returned when a write is denied under Plan (read-only) mode. */
+export function denyReadOnly(operation: string): unknown {
+  return {
+    status: 'denied',
+    operation,
+    message: `Read-only (Plan) permission mode — ${operation} is not permitted. Switch to Ask or Bypass to make changes.`,
+  };
+}
+
+/** Finalize a just-staged change per the permission gate: apply now under Bypass,
+ *  otherwise return it as queued for the human to approve. */
+export async function finalizeWrite(
+  context: ToolContext,
+  projectRoot: string,
+  change: { id: string; file_path: string; operation: string },
+  queuedHint: string,
+): Promise<unknown> {
+  if (context.permissionMode === 'bypass') {
+    const res = await applyOne(context.sql, change.id, projectRoot);
+    console.log(
+      `[write-gate] bypass apply ${change.operation} ${change.file_path} -> ${res.success ? 'applied' : 'FAILED: ' + (res.error ?? '?')}`,
+    );
+    return {
+      status: res.success ? 'applied' : 'failed',
+      change_id: change.id,
+      file_path: change.file_path,
+      operation: change.operation,
+      message: res.success
+        ? `${change.operation} applied to ${change.file_path}.`
+        : `Apply failed for ${change.file_path}: ${res.error ?? 'unknown error'}. Left in the pending queue.`,
+    };
+  }
+  console.log(
+    `[write-gate] ${context.permissionMode ?? 'legacy'} queued ${change.operation} ${change.file_path}`,
+  );
+  return {
+    status: 'queued',
+    change_id: change.id,
+    file_path: change.file_path,
+    operation: change.operation,
+    message: queuedHint,
+  };
+}

apps/server/src/schema.sql

@@ -372,13 +372,12 @@ ALTER TABLE messages ADD COLUMN IF NOT EXISTS tail_start_id UUID REFERENCES mess
 ALTER TABLE chats ADD COLUMN IF NOT EXISTS needs_compaction BOOLEAN NOT NULL DEFAULT FALSE;
 CREATE INDEX IF NOT EXISTS idx_messages_chat_compacted ON messages (chat_id, compacted_at);
 
--- tasks table (provider dispatch, arena)
+-- tasks table (provider dispatch)
 CREATE TABLE IF NOT EXISTS tasks (
   id                UUID PRIMARY KEY DEFAULT gen_random_uuid(),
   project_id        UUID NOT NULL REFERENCES projects(id) ON DELETE CASCADE,
   session_id        UUID REFERENCES sessions(id) ON DELETE CASCADE,
   parent_task_id    UUID REFERENCES tasks(id),
-  arena_id          UUID,
   state             TEXT NOT NULL DEFAULT 'pending'
                     CHECK (state IN ('pending','running','completed','failed','blocked','cancelled')),
   input             TEXT NOT NULL,
@@ -405,3 +404,6 @@ DO $$ BEGIN
       FOREIGN KEY (session_id) REFERENCES sessions(id) ON DELETE CASCADE;
   END IF;
 END $$;
+
+-- Remove the v2.0.5 arena_id column (replaced by the new Arena feature).
+ALTER TABLE tasks DROP COLUMN IF EXISTS arena_id;

apps/server/src/services/inference/types.ts

@@ -44,7 +44,11 @@ export interface InferenceFrame {
     | 'chat_renamed'
     | 'error'
     | 'flow_run_started'
-    | 'flow_run_step_updated';
+    | 'flow_run_step_updated'
+    // arena frames
+    | 'battle_started'
+    | 'contestant_updated'
+    | 'battle_updated';
   message_id?: string;
   message_ids?: string[];
   chat_id?: string;
@@ -84,6 +88,19 @@ export interface InferenceFrame {
   status?: string;
   run_status?: 'running' | 'completed' | 'failed' | 'cancelled';
   report?: string;
+  // arena frames
+  battle_id?: string;
+  battle_type?: 'coding' | 'qa';
+  prompt?: string;
+  contestants?: Array<{ id: string; identity: string; model: string; lane: 'local' | 'cloud' }>;
+  contestant_id?: string;
+  battle_status?: 'pending' | 'running' | 'completed' | 'failed' | 'cancelled';
+  duration_ms?: number;
+  tokens_per_sec?: number;
+  winner_contestant_id?: string | null;
+  analysis_ready?: boolean;
+  cross_exam_id?: string;
+  delta?: string;
 }
 
 export type FramePublisher = (sessionId: string, frame: InferenceFrame) => void;

apps/web/src/App.tsx

@@ -16,6 +16,7 @@ import { RightRailDrawerProvider, useRightRailDrawer } from '@/hooks/useRightRai
 import { useViewport } from '@/hooks/useViewport';
 import { ThemeFx } from '@/components/fx/ThemeFx';
 import { FlowLauncherDialog } from '@/components/FlowLauncherDialog';
+import { ArenaLauncherDialog } from '@/components/ArenaLauncherDialog';
 
 function SessionRightRail() {
   const { id } = useParams<{ id: string }>();
@@ -102,6 +103,7 @@ function AppShell() {
         </Routes>
         <Toaster position="bottom-right" />
         <FlowLauncherDialog />
+        <ArenaLauncherDialog />
       </div>
     </>
   );

apps/web/src/api/client.ts

@@ -27,6 +27,9 @@ import type {
   WorkspaceState,
   FlowRunRow,
   FlowStepRow,
+  BattleShape,
+  ContestantShape,
+  CrossExaminationShape,
 } from './types';
 
 // v2.6 Phase 1-UX §9b: chat-scoped agent-session rows. Returned by
@@ -518,6 +521,63 @@ export const api = {
       request<AgentsResponse>(`/api/projects/${projectId}/agents`),
   },
 
+  // Arena battle API — proxied to boocoder at /api/coder/battles/*.
+  battles: {
+    create: (body: {
+      project_id: string;
+      battle_type: 'coding' | 'qa';
+      prompt: string;
+      contestants: Array<{ identity: string; model: string }>;
+    }) =>
+      request<{ battle_id: string }>('/api/coder/battles', {
+        method: 'POST',
+        body: JSON.stringify(body),
+      }),
+    list: (projectId: string) =>
+      request<{ battles: BattleShape[] }>(
+        `/api/coder/battles?project_id=${encodeURIComponent(projectId)}`,
+      ),
+    get: (battleId: string) =>
+      request<{
+        battle: BattleShape;
+        contestants: ContestantShape[];
+        cross_examinations: CrossExaminationShape[];
+      }>(`/api/coder/battles/${encodeURIComponent(battleId)}`),
+    stop: (battleId: string) =>
+      request<{ cancelled: boolean }>(
+        `/api/coder/battles/${encodeURIComponent(battleId)}/stop`,
+        { method: 'POST' },
+      ),
+    analyze: (battleId: string) =>
+      request<{ triggered: boolean }>(
+        `/api/coder/battles/${encodeURIComponent(battleId)}/analyze`,
+        { method: 'POST' },
+      ),
+    crossExamine: (battleId: string, body: { identity: string; model: string }) =>
+      request<{ cross_exam_id: string }>(
+        `/api/coder/battles/${encodeURIComponent(battleId)}/cross-examine`,
+        { method: 'POST', body: JSON.stringify(body) },
+      ),
+    getAnalysis: (battleId: string) =>
+      request<{ text: string }>(
+        `/api/coder/battles/${encodeURIComponent(battleId)}/analysis`,
+      ),
+    generatePrompt: (description: string) =>
+      request<{ prompt: string }>('/api/coder/battles/generate-prompt', {
+        method: 'POST',
+        body: JSON.stringify({ description }),
+      }),
+    setWinner: (battleId: string, body: { winner_contestant_id: string | null }) =>
+      request<{ ok: boolean }>(
+        `/api/coder/battles/${encodeURIComponent(battleId)}/winner`,
+        { method: 'PATCH', body: JSON.stringify(body) },
+      ),
+    getDiff: (battleId: string, contestantId: string) =>
+      request<{ diff: string }>(
+        `/api/coder/battles/${encodeURIComponent(battleId)}/contestants/${encodeURIComponent(contestantId)}/diff`,
+      ),
+  },
+
   skills: {
     list: () => request<{ skills: Skill[] }>('/api/skills'),
   },

apps/web/src/api/types.ts

@@ -391,7 +391,8 @@ export type WorkspacePaneKind =
   | 'settings'
   | 'markdown_artifact'
   | 'html_artifact'
-  | 'orchestrator';
+  | 'orchestrator'
+  | 'arena';
 
 // Mixed tabs: a pane can hold tabs of different kinds (a BooChat tab next to a
 // BooCode tab next to a Terminal tab). Each tab carries its own kind; the active
@@ -424,6 +425,10 @@ export interface OrchestratorState {
   band: 'small' | 'medium' | 'large';
 }
 
+// Arena pane state — single-sourced in @boocode/contracts; edit the package, not here.
+import type { ArenaState, BattleShape, ContestantShape, CrossExaminationShape, BattleType, BattleStatus, ContestantStatus, ContestantLane } from '@boocode/contracts/arena';
+export type { ArenaState, BattleShape, ContestantShape, CrossExaminationShape, BattleType, BattleStatus, ContestantStatus, ContestantLane };
+
 // Orchestrator run API types (returned by GET /api/coder/runs/:id).
 export interface FlowRunRow {
   id: string;
@@ -475,6 +480,8 @@ export interface WorkspacePane {
   html_artifact_state?: HtmlArtifactState;
   // orchestrator pane: populated only when kind === 'orchestrator'.
   orchestrator_state?: OrchestratorState;
+  // arena pane: populated only when kind === 'arena'.
+  arena_state?: ArenaState;
 }
 
 // Reopen LIFO stack entry. Shape unchanged from the prior module-level stack;
@@ -592,4 +599,31 @@ export type WsFrame =
       status: 'pending' | 'running' | 'completed' | 'failed' | 'skipped' | 'cancelled';
       run_status?: 'running' | 'completed' | 'failed' | 'cancelled';
       report?: string;
+    }
+  // arena frames: battle lifecycle + per-contestant streaming
+  | {
+      type: 'battle_started';
+      battle_id: string;
+      battle_type: 'coding' | 'qa';
+      prompt: string;
+      contestants: Array<{ id: string; identity: string; model: string; lane: 'local' | 'cloud' }>;
+    }
+  | {
+      type: 'contestant_updated';
+      battle_id: string;
+      contestant_id: string;
+      status?: 'queued' | 'running' | 'done' | 'error';
+      duration_ms?: number;
+      tokens_per_sec?: number;
+      battle_status?: 'pending' | 'running' | 'completed' | 'failed' | 'cancelled';
+      delta?: string;
+      error?: string;
+    }
+  | {
+      type: 'battle_updated';
+      battle_id: string;
+      status?: 'pending' | 'running' | 'completed' | 'failed' | 'cancelled';
+      winner_contestant_id?: string | null;
+      analysis_ready?: boolean;
+      cross_exam_id?: string;
     };

apps/web/src/components/ArenaLauncherDialog.tsx

@@ -0,0 +1,410 @@
+// ArenaLauncherDialog — mirrors FlowLauncherDialog.
+// Opens via sessionEvents 'open_arena_launcher'.
+// Flow: pick Battle Type → write/generate prompt → add 2–6 contestants → Start.
+
+import { useCallback, useEffect, useRef, useState } from 'react';
+import { Loader2, Minus, Plus, Swords, TriangleAlert, X } from 'lucide-react';
+import { toast } from 'sonner';
+import {
+  Dialog,
+  DialogContent,
+  DialogFooter,
+  DialogHeader,
+  DialogTitle,
+} from '@/components/ui/dialog';
+import { Button } from '@/components/ui/button';
+import { Label } from '@/components/ui/label';
+import { api } from '@/api/client';
+import type { Agent, ProviderSnapshotEntry } from '@/api/types';
+import { sessionEvents } from '@/hooks/sessionEvents';
+import { useProviderSnapshot } from '@/hooks/useProviderSnapshot';
+import { cn } from '@/lib/utils';
+
+// ─── types ────────────────────────────────────────────────────────────────────
+
+type BattleType = 'coding' | 'qa';
+
+interface Contestant {
+  key: string; // local unique key for React
+  identity: string;
+  model: string;
+}
+
+// ─── helpers ─────────────────────────────────────────────────────────────────
+
+function newContestant(): Contestant {
+  return { key: crypto.randomUUID(), identity: '', model: '' };
+}
+
+function isDuplicate(contestants: Contestant[], c: Contestant): boolean {
+  const dups = contestants.filter(
+    (x) => x.key !== c.key && x.identity === c.identity && x.model === c.model && x.identity !== '',
+  );
+  return dups.length > 0;
+}
+
+function hasDuplicatePair(contestants: Contestant[]): boolean {
+  return contestants.some((c) => isDuplicate(contestants, c));
+}
+
+function localCount(battleType: BattleType, contestants: Contestant[], snapshot: ProviderSnapshotEntry[] | null): number {
+  if (battleType === 'qa') return contestants.filter((c) => c.identity !== '').length;
+  const boocode = snapshot?.find((e) => e.name === 'boocode');
+  const localModelIds = new Set(boocode?.models.map((m) => m.id) ?? []);
+  return contestants.filter((c) => {
+    // Match bare IDs (boocode/native) and llama-swap/-prefixed IDs used by
+    // opencode and other external agents pointing at the local llama-swap server.
+    return localModelIds.has(c.model) || localModelIds.has(c.model.replace(/^llama-swap\//, ''));
+  }).length;
+}
+
+// ─── ContestantRow ────────────────────────────────────────────────────────────
+
+function ContestantRow({
+  contestant,
+  battleType,
+  snapshot,
+  agents,
+  allContestants,
+  onUpdate,
+  onRemove,
+  removable,
+}: {
+  contestant: Contestant;
+  battleType: BattleType;
+  snapshot: ProviderSnapshotEntry[] | null;
+  agents: Agent[];
+  allContestants: Contestant[];
+  onUpdate: (patch: Partial<Contestant>) => void;
+  onRemove: () => void;
+  removable: boolean;
+}) {
+  const dup = isDuplicate(allContestants, contestant);
+
+  // Identity options for Coding: installed provider names.
+  // Identity options for Q&A: agents by id.
+  const identityOptions =
+    battleType === 'coding'
+      ? (snapshot ?? [])
+          .filter((e) => e.installed && e.enabled)
+          .map((e) => ({ value: e.name, label: e.label }))
+      : agents.map((a) => ({ value: a.id, label: a.name }));
+
+  // Model options: for Coding use the selected provider's models; for Q&A use boocode models.
+  const modelOptions: { value: string; label: string }[] = (() => {
+    if (battleType === 'coding') {
+      const provider = (snapshot ?? []).find((e) => e.name === contestant.identity);
+      return (provider?.models ?? []).map((m) => ({ value: m.id, label: m.label }));
+    }
+    // Q&A: native backend only — use boocode models
+    const boocode = (snapshot ?? []).find((e) => e.name === 'boocode');
+    return (boocode?.models ?? []).map((m) => ({ value: m.id, label: m.label }));
+  })();
+
+  function handleIdentityChange(value: string) {
+    // Reset model when identity changes so stale model doesn't persist.
+    onUpdate({ identity: value, model: '' });
+  }
+
+  function handleModelChange(value: string) {
+    onUpdate({ model: value });
+  }
+
+  return (
+    <div className={cn('flex items-center gap-2', dup && 'opacity-60')}>
+      <select
+        value={contestant.identity}
+        onChange={(e) => handleIdentityChange(e.target.value)}
+        className="flex-1 min-w-0 text-xs border border-border rounded bg-background px-2 py-1.5 text-foreground focus:outline-none focus:ring-1 focus:ring-ring"
+        aria-label={battleType === 'coding' ? 'Backend' : 'Persona'}
+      >
+        <option value="">{battleType === 'coding' ? 'Backend…' : 'Persona…'}</option>
+        {identityOptions.map((o) => (
+          <option key={o.value} value={o.value}>{o.label}</option>
+        ))}
+      </select>
+      <select
+        value={contestant.model}
+        onChange={(e) => handleModelChange(e.target.value)}
+        disabled={!contestant.identity}
+        className="flex-1 min-w-0 text-xs border border-border rounded bg-background px-2 py-1.5 text-foreground focus:outline-none focus:ring-1 focus:ring-ring disabled:opacity-50"
+        aria-label="Model"
+      >
+        <option value="">Model…</option>
+        {modelOptions.map((o) => (
+          <option key={o.value} value={o.value}>{o.label}</option>
+        ))}
+      </select>
+      {dup && (
+        <span title="Duplicate contestant" className="shrink-0 text-destructive">
+          <TriangleAlert size={12} />
+        </span>
+      )}
+      {removable && (
+        <button
+          type="button"
+          onClick={onRemove}
+          className="shrink-0 inline-flex items-center justify-center p-1 rounded text-muted-foreground hover:bg-muted hover:text-foreground"
+          aria-label="Remove contestant"
+        >
+          <Minus size={12} />
+        </button>
+      )}
+    </div>
+  );
+}
+
+// ─── ArenaLauncherDialog ──────────────────────────────────────────────────────
+
+export function ArenaLauncherDialog() {
+  const [open, setOpen] = useState(false);
+  const [projectId, setProjectId] = useState('');
+  const [placement, setPlacement] = useState<'new' | 'split'>('new');
+  const [battleType, setBattleType] = useState<BattleType>('coding');
+  const [prompt, setPrompt] = useState('');
+  const [contestants, setContestants] = useState<Contestant[]>(() => [
+    newContestant(),
+    newContestant(),
+  ]);
+  const [generating, setGenerating] = useState(false);
+  const [starting, setStarting] = useState(false);
+  const [agents, setAgents] = useState<Agent[]>([]);
+  const promptRef = useRef<HTMLTextAreaElement>(null);
+
+  const snapshot = useProviderSnapshot();
+
+  useEffect(() => {
+    return sessionEvents.subscribe((ev) => {
+      if (ev.type !== 'open_arena_launcher') return;
+      setProjectId(ev.project_id);
+      setPlacement(ev.placement ?? 'new');
+      setBattleType('coding');
+      setPrompt('');
+      setContestants([newContestant(), newContestant()]);
+      setGenerating(false);
+      setStarting(false);
+      setOpen(true);
+    });
+  }, []);
+
+  // Load agents list when dialog opens (for Q&A mode).
+  useEffect(() => {
+    if (!open || !projectId) return;
+    api.agents.list(projectId)
+      .then((r) => setAgents(r.agents))
+      .catch(() => {});
+  }, [open, projectId]);
+
+  const handleGeneratePrompt = useCallback(async () => {
+    const description = prompt.trim();
+    if (!description || generating) return;
+    setGenerating(true);
+    try {
+      const { prompt: generated } = await api.battles.generatePrompt(description);
+      setPrompt(generated);
+      promptRef.current?.focus();
+    } catch (err) {
+      toast.error(err instanceof Error ? err.message : 'Generate failed');
+    } finally {
+      setGenerating(false);
+    }
+  }, [prompt, generating]);
+
+  function updateContestant(key: string, patch: Partial<Contestant>) {
+    setContestants((prev) => prev.map((c) => (c.key === key ? { ...c, ...patch } : c)));
+  }
+
+  function removeContestant(key: string) {
+    setContestants((prev) => prev.filter((c) => c.key !== key));
+  }
+
+  function addContestant() {
+    if (contestants.length >= 6) return;
+    setContestants((prev) => [...prev, newContestant()]);
+  }
+
+  const canStart =
+    !starting &&
+    prompt.trim().length > 0 &&
+    contestants.length >= 2 &&
+    contestants.every((c) => c.identity !== '' && c.model !== '') &&
+    !hasDuplicatePair(contestants);
+
+  const localLaneCount = localCount(battleType, contestants, snapshot);
+  const showLocalWarning = localLaneCount >= 3;
+
+  async function handleStart() {
+    if (!canStart) return;
+    setStarting(true);
+    try {
+      const { battle_id } = await api.battles.create({
+        project_id: projectId,
+        battle_type: battleType,
+        prompt: prompt.trim(),
+        contestants: contestants.map((c) => ({ identity: c.identity, model: c.model })),
+      });
+      sessionEvents.emit({
+        type: 'open_arena_pane',
+        state: { battle_id, battle_type: battleType, prompt: prompt.trim() },
+        placement,
+      });
+      setOpen(false);
+    } catch (err) {
+      toast.error(err instanceof Error ? err.message : 'Failed to start battle');
+    } finally {
+      setStarting(false);
+    }
+  }
+
+  return (
+    <Dialog open={open} onOpenChange={setOpen}>
+      <DialogContent
+        className="flex flex-col gap-0 p-0 max-h-[85vh] sm:max-w-lg overflow-hidden"
+        showCloseButton={false}
+      >
+        <DialogHeader className="gap-1.5 px-4 pt-4 pb-3 border-b shrink-0">
+          <div className="flex items-center gap-2">
+            <Swords size={14} className="text-muted-foreground shrink-0" />
+            <DialogTitle className="text-sm font-medium">New Arena Battle</DialogTitle>
+          </div>
+          <p className="text-xs text-muted-foreground">
+            Run the same prompt against multiple AI competitors and pick the best result.
+          </p>
+        </DialogHeader>
+
+        <div className="flex flex-col gap-4 overflow-y-auto overscroll-contain px-4 py-3">
+          {/* Battle type */}
+          <div className="flex flex-col gap-1.5">
+            <Label className="text-xs text-muted-foreground">Battle type</Label>
+            <div className="flex gap-1.5">
+              {(['coding', 'qa'] as const).map((t) => (
+                <button
+                  key={t}
+                  type="button"
+                  onClick={() => { setBattleType(t); setContestants([newContestant(), newContestant()]); }}
+                  aria-pressed={battleType === t}
+                  className={cn(
+                    'flex-1 rounded-lg border py-1.5 text-xs transition-colors capitalize',
+                    battleType === t
+                      ? 'border-primary bg-primary/10 text-primary font-medium'
+                      : 'border-border text-muted-foreground hover:bg-muted hover:text-foreground',
+                  )}
+                >
+                  {t === 'coding' ? 'Coding' : 'Q&A'}
+                </button>
+              ))}
+            </div>
+            <p className="text-xs text-muted-foreground">
+              {battleType === 'coding'
+                ? 'Each contestant works in its own isolated worktree. Results include a diff.'
+                : 'Contestants answer the prompt as text. No code changes.'}
+            </p>
+          </div>
+
+          {/* Prompt */}
+          <div className="flex flex-col gap-1.5">
+            <div className="flex items-center justify-between">
+              <Label htmlFor="arena-prompt" className="text-xs text-muted-foreground">
+                Prompt
+              </Label>
+              <button
+                type="button"
+                onClick={() => void handleGeneratePrompt()}
+                disabled={generating || prompt.trim().length === 0}
+                className="text-xs text-primary hover:text-primary/80 disabled:opacity-40 disabled:cursor-default flex items-center gap-1"
+                title="Expand your description into a fuller battle prompt"
+              >
+                {generating && <Loader2 size={10} className="animate-spin" />}
+                Generate prompt
+              </button>
+            </div>
+            <textarea
+              id="arena-prompt"
+              ref={promptRef}
+              value={prompt}
+              onChange={(e) => setPrompt(e.target.value)}
+              placeholder={
+                battleType === 'coding'
+                  ? 'Describe a coding task, or enter a short description and click Generate prompt…'
+                  : 'Ask a question or describe a topic, or enter a short description and click Generate prompt…'
+              }
+              rows={4}
+              className="w-full text-sm border border-border rounded bg-background px-3 py-2 text-foreground placeholder:text-muted-foreground focus:outline-none focus:ring-1 focus:ring-ring resize-none"
+            />
+          </div>
+
+          {/* Contestants */}
+          <div className="flex flex-col gap-2">
+            <div className="flex items-center justify-between">
+              <Label className="text-xs text-muted-foreground">
+                Contestants ({contestants.length}/6)
+              </Label>
+              <span className="text-xs text-muted-foreground">
+                {battleType === 'coding' ? 'Backend + Model' : 'Persona + Model'}
+              </span>
+            </div>
+
+            <div className="flex flex-col gap-1.5">
+              {contestants.map((c) => (
+                <ContestantRow
+                  key={c.key}
+                  contestant={c}
+                  battleType={battleType}
+                  snapshot={snapshot}
+                  agents={agents}
+                  allContestants={contestants}
+                  onUpdate={(patch) => updateContestant(c.key, patch)}
+                  onRemove={() => removeContestant(c.key)}
+                  removable={contestants.length > 2}
+                />
+              ))}
+            </div>
+
+            {contestants.length < 6 && (
+              <button
+                type="button"
+                onClick={addContestant}
+                className="flex items-center gap-1.5 text-xs text-muted-foreground hover:text-foreground py-1"
+              >
+                <Plus size={12} /> Add contestant
+              </button>
+            )}
+
+            {hasDuplicatePair(contestants) && (
+              <div className="flex items-center gap-1.5 text-xs text-destructive">
+                <TriangleAlert size={12} />
+                Duplicate contestants (same identity + model) are not allowed.
+              </div>
+            )}
+
+            {showLocalWarning && (
+              <div className="flex items-center gap-1.5 text-xs text-amber-600 dark:text-amber-400">
+                <TriangleAlert size={12} />
+                {localLaneCount} local contestants will run serially (one GPU load at a time). This battle will take a while.
+              </div>
+            )}
+          </div>
+        </div>
+
+        <DialogFooter className="px-4 py-3 border-t shrink-0 flex items-center justify-between">
+          <button
+            type="button"
+            onClick={() => setOpen(false)}
+            className="flex items-center gap-1.5 text-xs text-muted-foreground hover:text-foreground"
+          >
+            <X size={12} /> Cancel
+          </button>
+          <Button
+            type="button"
+            size="sm"
+            onClick={() => void handleStart()}
+            disabled={!canStart}
+          >
+            {starting ? <Loader2 className="animate-spin" /> : <Swords size={14} />}
+            Start battle
+          </Button>
+        </DialogFooter>
+      </DialogContent>
+    </Dialog>
+  );
+}

apps/web/src/components/ChatTabBar.tsx

@@ -37,6 +37,7 @@ interface Props {
   onNewTab: (kind: WorkspaceTabKind) => void;
   onSplitPane: (kind: 'chat' | 'terminal' | 'coder') => void;
   onNewOrchestrator?: () => void;
+  onNewArena?: () => void;
   onReopenPane?: () => void;
   onShowHistory: () => void;
   onRename: (chatId: string, name: string) => Promise<void>;
@@ -69,6 +70,7 @@ export function ChatTabBar({
   onNewTab,
   onSplitPane,
   onNewOrchestrator,
+  onNewArena,
   onReopenPane,
   onShowHistory,
   onRename,
@@ -230,6 +232,7 @@ export function ChatTabBar({
           onNewTab={onNewTab}
           onSplitPane={onSplitPane}
           onNewOrchestrator={onNewOrchestrator}
+          onNewArena={onNewArena}
           onReopenPane={onReopenPane}
           onShowHistory={onShowHistory}
           onRemovePane={onRemovePane}

apps/web/src/components/PaneHeaderActions.tsx

@@ -1,4 +1,4 @@
-import { Code, Columns2, History, MessageSquare, Plus, RotateCcw, Terminal, Workflow, X } from 'lucide-react';
+import { Code, Columns2, History, MessageSquare, Plus, RotateCcw, Swords, Terminal, Workflow, X } from 'lucide-react';
 import {
   DropdownMenu,
   DropdownMenuContent,
@@ -19,6 +19,8 @@ interface Props {
   // When provided, shows a "New Orchestrator" item that opens the flow launcher.
   // Orchestrators are always split (run-bound; can't live as a tab in another pane).
   onNewOrchestrator?: () => void;
+  // When provided, shows a "New Arena" item that opens the arena launcher.
+  onNewArena?: () => void;
   onReopenPane?: () => void;
   onShowHistory: () => void;
   onRemovePane?: () => void;
@@ -35,6 +37,7 @@ export function PaneHeaderActions({
   onNewTab,
   onSplitPane,
   onNewOrchestrator,
+  onNewArena,
   onReopenPane,
   onShowHistory,
   onRemovePane,
@@ -71,6 +74,11 @@ export function PaneHeaderActions({
               <Workflow size={14} /> New Orchestrator
             </DropdownMenuItem>
           )}
+          {onNewArena && (
+            <DropdownMenuItem onSelect={onNewArena}>
+              <Swords size={14} /> New Arena
+            </DropdownMenuItem>
+          )}
         </DropdownMenuContent>
       </DropdownMenu>
 
@@ -101,6 +109,11 @@ export function PaneHeaderActions({
               <Workflow size={14} /> New Orchestrator
             </DropdownMenuItem>
           )}
+          {onNewArena && (
+            <DropdownMenuItem onSelect={onNewArena}>
+              <Swords size={14} /> New Arena
+            </DropdownMenuItem>
+          )}
         </DropdownMenuContent>
       </DropdownMenu>
 

apps/web/src/components/Workspace.tsx

@@ -13,6 +13,7 @@ import { CoderPane } from '@/components/panes/CoderPane';
 import { MarkdownArtifactPane } from '@/components/MarkdownArtifactPane';
 import { HtmlArtifactPane } from '@/components/HtmlArtifactPane';
 import { OrchestratorPane } from '@/components/panes/OrchestratorPane';
+import { ArenaPane } from '@/components/panes/ArenaPane';
 import { ChatTabBar, type TabDescriptor } from '@/components/ChatTabBar';
 import { SessionLandingPage } from '@/components/SessionLandingPage';
 import { cn } from '@/lib/utils';
@@ -134,6 +135,14 @@ export function Workspace({
     });
   }
 
+  function handleNewArena() {
+    sessionEvents.emit({
+      type: 'open_arena_launcher',
+      project_id: projectId,
+      placement: 'split',
+    });
+  }
+
   // v1.10 booterm + mixed tabs: per-terminal-TAB label, keyed by the terminal
   // tab id (which keys its tmux session). Numbered across the workspace.
   const terminalLabels = useMemo(() => {
@@ -180,6 +189,7 @@ export function Workspace({
           const isTerminal = pane.kind === 'terminal';
           const isCoder = pane.kind === 'coder';
           const isOrchestrator = pane.kind === 'orchestrator';
+          const isArena = pane.kind === 'arena';
           const isArtifact = pane.kind === 'markdown_artifact' || pane.kind === 'html_artifact';
           // v1.9: when maximized, hide every pane except the settings one.
           // display:none keeps the React tree mounted so streams / drafts
@@ -192,8 +202,8 @@ export function Workspace({
             }
             return null;
           }
-          // Terminal + coder + orchestrator panes own their tab strip (no chats, no ChatTabBar).
-          const isChromeless = isSettings || isTerminal || isCoder || isArtifact || isOrchestrator;
+          // Terminal + coder + orchestrator + arena panes own their tab strip (no chats, no ChatTabBar).
+          const isChromeless = isSettings || isTerminal || isCoder || isArtifact || isOrchestrator || isArena;
           return (
           <div
             key={pane.id}
@@ -218,7 +228,7 @@ export function Workspace({
                   (chat / coder / terminal / empty-landing). The "+" adds a tab
                   of any kind; Split adds a pane. Settings/artifact panes own
                   their own headers. Hidden on mobile (mobile uses pane panes). */}
-              {!isMobile && !isSettings && !isArtifact && !isOrchestrator && (
+              {!isMobile && !isSettings && !isArtifact && !isOrchestrator && !isArena && (
                 <ChatTabBar
                   pane={pane}
                   tabs={paneTabs(pane)}
@@ -231,6 +241,7 @@ export function Workspace({
                   onNewTab={(kind) => void createTab(idx, kind)}
                   onSplitPane={(kind) => onAddPane(kind)}
                   onNewOrchestrator={handleNewOrchestrator}
+                  onNewArena={handleNewArena}
                   onReopenPane={hasClosedPanes ? reopenPane : undefined}
                   onShowHistory={() => openSessionHistory(idx)}
                   onRename={renameChat}
@@ -277,6 +288,12 @@ export function Workspace({
                   state={pane.orchestrator_state}
                   onClose={() => removePane(idx)}
                 />
+              ) : pane.kind === 'arena' && pane.arena_state ? (
+                <ArenaPane
+                  state={pane.arena_state}
+                  projectId={projectId}
+                  onClose={() => removePane(idx)}
+                />
               ) : pane.kind === 'markdown_artifact' && pane.markdown_artifact_state ? (
                 <MarkdownArtifactPane
                   chatId={pane.markdown_artifact_state.chat_id}

apps/web/src/components/panes/ArenaPane.tsx

@@ -0,0 +1,664 @@
+// ArenaPane — live view for an Arena battle.
+// Mirrors OrchestratorPane: header with status/winner, contestant roster
+// (collapsed rows, expand-one), analysis panel, cross-examination control.
+//
+// Subscribes to the coder user channel (via useCoderUserEvents → sessionEvents)
+// for battle_started / contestant_updated / battle_updated frames.
+
+import { useCallback, useEffect, useRef, useState } from 'react';
+import { ChevronDown, ChevronRight, Loader2, MoreHorizontal, RotateCcw, Swords, Trophy, X } from 'lucide-react';
+import { toast } from 'sonner';
+import { api } from '@/api/client';
+import type { ArenaState, BattleShape, ContestantShape, CrossExaminationShape, ProviderSnapshotEntry } from '@/api/types';
+import { sessionEvents } from '@/hooks/sessionEvents';
+import { useProviderSnapshot } from '@/hooks/useProviderSnapshot';
+import {
+  DropdownMenu,
+  DropdownMenuContent,
+  DropdownMenuItem,
+  DropdownMenuTrigger,
+} from '@/components/ui/dropdown-menu';
+import { cn } from '@/lib/utils';
+
+// ─── Status dot (mirrors FlowStepStatusDot) ───────────────────────────────────
+
+function ContestantStatusDot({ status }: { status: ContestantShape['status'] }) {
+  if (status === 'running') {
+    return (
+      <span
+        aria-label="running"
+        className="inline-block w-3 h-3 rounded-full border-2 border-emerald-500 border-t-transparent animate-spin shrink-0"
+      />
+    );
+  }
+  const cls =
+    status === 'done'
+      ? 'bg-emerald-500'
+      : status === 'error'
+        ? 'bg-destructive'
+        : 'bg-muted-foreground/40'; // queued
+  return <span aria-label={status} className={cn('inline-block w-1.5 h-1.5 rounded-full shrink-0', cls)} />;
+}
+
+// ─── Lane badge ───────────────────────────────────────────────────────────────
+
+function LaneBadge({ lane }: { lane: ContestantShape['lane'] }) {
+  return (
+    <span
+      className={cn(
+        'text-[10px] px-1 py-0.5 rounded shrink-0',
+        lane === 'local'
+          ? 'bg-sky-500/10 text-sky-600 dark:text-sky-400'
+          : 'bg-violet-500/10 text-violet-600 dark:text-violet-400',
+      )}
+    >
+      {lane}
+    </span>
+  );
+}
+
+// ─── Duration formatter ───────────────────────────────────────────────────────
+
+function formatDuration(ms: number | null): string {
+  if (ms == null) return '';
+  const s = Math.round(ms / 1000);
+  if (s < 60) return `${s}s`;
+  return `${Math.floor(s / 60)}m${String(s % 60).padStart(2, '0')}s`;
+}
+
+// ─── Live ticker for running contestants ─────────────────────────────────────
+
+function LiveDuration({ startedAt }: { startedAt: number }) {
+  const [elapsed, setElapsed] = useState(() => Date.now() - startedAt);
+  useEffect(() => {
+    const id = setInterval(() => setElapsed(Date.now() - startedAt), 1000);
+    return () => clearInterval(id);
+  }, [startedAt]);
+  return <span>{formatDuration(elapsed)}</span>;
+}
+
+// ─── DiffView ─────────────────────────────────────────────────────────────────
+
+function DiffView({ diff }: { diff: string }) {
+  const lines = diff.split('\n');
+  return (
+    <div className="border-t border-border/50">
+      <div className="px-3 pt-2 pb-1 text-[10px] font-medium uppercase tracking-wide text-muted-foreground">
+        Diff
+      </div>
+      <pre className="px-3 pb-3 text-xs font-mono whitespace-pre leading-relaxed overflow-x-auto">
+        {lines.map((line, i) => {
+          const cls =
+            line.startsWith('+') && !line.startsWith('+++')
+              ? 'text-emerald-600 dark:text-emerald-400'
+              : line.startsWith('-') && !line.startsWith('---')
+                ? 'text-destructive'
+                : line.startsWith('@@')
+                  ? 'text-violet-500 dark:text-violet-400'
+                  : 'text-muted-foreground';
+          return (
+            <span key={i} className={cn('block', cls)}>
+              {line || ' '}
+            </span>
+          );
+        })}
+      </pre>
+    </div>
+  );
+}
+
+// ─── ContestantRow ────────────────────────────────────────────────────────────
+
+interface ContestantRowState {
+  data: ContestantShape;
+  output: string;
+  startedAt: number | null;
+}
+
+function ContestantRow({
+  row,
+  isExpanded,
+  onToggle,
+  isWinner,
+  battleId,
+  battleType,
+}: {
+  row: ContestantRowState;
+  isExpanded: boolean;
+  onToggle: () => void;
+  isWinner: boolean;
+  battleId: string;
+  battleType: 'coding' | 'qa';
+}) {
+  const { data, output, startedAt } = row;
+  const label = `${data.identity} / ${data.model}`;
+
+  // Lazy-fetch diff for coding contestants once they are done and expanded.
+  const [diff, setDiff] = useState<string | null>(null);
+  useEffect(() => {
+    if (!isExpanded || battleType !== 'coding' || data.status !== 'done') return;
+    if (diff !== null) return;
+    api.battles.getDiff(battleId, data.id)
+      .then(({ diff: d }) => setDiff(d))
+      .catch(() => setDiff(''));
+  }, [isExpanded, battleType, data.status, data.id, battleId, diff]);
+
+  async function handleSetWinner(contestantId: string | null) {
+    try {
+      await api.battles.setWinner(battleId, { winner_contestant_id: contestantId });
+    } catch {
+      // WS frame updates the badge; a failed call just leaves it unchanged
+    }
+  }
+
+  return (
+    <div>
+      <button
+        type="button"
+        onClick={onToggle}
+        className="w-full flex items-center gap-2 px-3 py-2.5 text-left hover:bg-muted/30 transition-colors"
+      >
+        <ContestantStatusDot status={data.status} />
+        <span className="text-sm flex-1 truncate min-w-0">{label}</span>
+        {isWinner && (
+          <Trophy size={11} className="shrink-0 text-emerald-500" aria-label="winner" />
+        )}
+        <LaneBadge lane={data.lane} />
+        {data.status === 'running' && startedAt != null ? (
+          <span className="text-xs text-muted-foreground shrink-0 tabular-nums">
+            <LiveDuration startedAt={startedAt} />
+          </span>
+        ) : data.duration_ms != null ? (
+          <span className="text-xs text-muted-foreground shrink-0 tabular-nums">
+            {formatDuration(data.duration_ms)}
+          </span>
+        ) : null}
+        {data.tokens_per_sec != null && (
+          <span className="text-xs text-muted-foreground shrink-0 hidden sm:block tabular-nums">
+            {data.tokens_per_sec.toFixed(1)} tok/s
+          </span>
+        )}
+        {data.status === 'error' && (
+          <span className="text-xs text-destructive shrink-0 hidden sm:block truncate max-w-[100px]" title={data.error ?? ''}>
+            {data.error ?? 'error'}
+          </span>
+        )}
+        {isExpanded ? (
+          <ChevronDown size={12} className="shrink-0 text-muted-foreground" />
+        ) : (
+          <ChevronRight size={12} className="shrink-0 text-muted-foreground" />
+        )}
+        {/* Row menu: winner override. Stop propagation so the row toggle isn't triggered. */}
+        <span onClick={(e) => e.stopPropagation()}>
+          <DropdownMenu>
+            <DropdownMenuTrigger asChild>
+              <button
+                type="button"
+                className="shrink-0 p-0.5 rounded text-muted-foreground hover:text-foreground hover:bg-muted"
+                aria-label="Contestant options"
+              >
+                <MoreHorizontal size={12} />
+              </button>
+            </DropdownMenuTrigger>
+            <DropdownMenuContent align="end">
+              {!isWinner && (
+                <DropdownMenuItem onSelect={() => void handleSetWinner(data.id)}>
+                  <Trophy size={12} /> Set as winner
+                </DropdownMenuItem>
+              )}
+              {isWinner && (
+                <DropdownMenuItem onSelect={() => void handleSetWinner(null)}>
+                  Clear winner
+                </DropdownMenuItem>
+              )}
+            </DropdownMenuContent>
+          </DropdownMenu>
+        </span>
+      </button>
+
+      {isExpanded && (
+        <div className="border-t border-border/50 bg-muted/10 max-h-[55vh] overflow-y-auto">
+          {output.length === 0 ? (
+            <div className="flex items-center justify-center py-6 text-sm text-muted-foreground">
+              {data.status === 'queued'
+                ? 'Waiting to start…'
+                : data.status === 'error'
+                  ? data.error ?? 'Error'
+                  : 'Connecting…'}
+            </div>
+          ) : (
+            <pre className="p-3 text-xs font-mono whitespace-pre-wrap leading-relaxed break-all text-foreground">
+              {output}
+            </pre>
+          )}
+          {battleType === 'coding' && data.status === 'done' && diff && (
+            <DiffView diff={diff} />
+          )}
+        </div>
+      )}
+    </div>
+  );
+}
+
+// ─── CrossExaminationPanel ────────────────────────────────────────────────────
+
+function CrossExaminationPanel({
+  battleId,
+  crossExams,
+  snapshot,
+}: {
+  battleId: string;
+  crossExams: CrossExaminationShape[];
+  snapshot: ProviderSnapshotEntry[] | null;
+}) {
+  const [identity, setIdentity] = useState('');
+  const [model, setModel] = useState('');
+  const [running, setRunning] = useState(false);
+
+  const identityOptions = (snapshot ?? [])
+    .filter((e) => e.installed && e.enabled)
+    .map((e) => ({ value: e.name, label: e.label }));
+
+  const modelOptions = (() => {
+    const provider = (snapshot ?? []).find((e) => e.name === identity);
+    return (provider?.models ?? []).map((m) => ({ value: m.id, label: m.label }));
+  })();
+
+  async function handleRun() {
+    if (!identity || !model || running) return;
+    setRunning(true);
+    try {
+      await api.battles.crossExamine(battleId, { identity, model });
+      // The verdict arrives via battle_updated frame; ArenaPane will refetch.
+    } catch (err) {
+      toast.error(err instanceof Error ? err.message : 'Cross-examination failed');
+    } finally {
+      setRunning(false);
+    }
+  }
+
+  return (
+    <div className="border-t border-border p-4 flex flex-col gap-3">
+      <div className="text-xs font-medium text-muted-foreground uppercase tracking-wide">
+        Cross-examination
+      </div>
+      <p className="text-xs text-muted-foreground">
+        Challenge the results with any model. The verdict is advisory and never changes the recorded winner.
+      </p>
+      <div className="flex gap-2 items-center flex-wrap">
+        <select
+          value={identity}
+          onChange={(e) => { setIdentity(e.target.value); setModel(''); }}
+          className="flex-1 min-w-[120px] text-xs border border-border rounded bg-background px-2 py-1.5 text-foreground focus:outline-none focus:ring-1 focus:ring-ring"
+          aria-label="Backend"
+        >
+          <option value="">Backend…</option>
+          {identityOptions.map((o) => (
+            <option key={o.value} value={o.value}>{o.label}</option>
+          ))}
+        </select>
+        <select
+          value={model}
+          onChange={(e) => setModel(e.target.value)}
+          disabled={!identity}
+          className="flex-1 min-w-[120px] text-xs border border-border rounded bg-background px-2 py-1.5 text-foreground focus:outline-none focus:ring-1 focus:ring-ring disabled:opacity-50"
+          aria-label="Model"
+        >
+          <option value="">Model…</option>
+          {modelOptions.map((o) => (
+            <option key={o.value} value={o.value}>{o.label}</option>
+          ))}
+        </select>
+        <button
+          type="button"
+          onClick={() => void handleRun()}
+          disabled={!identity || !model || running}
+          className="inline-flex items-center gap-1 text-xs px-2 py-1.5 rounded border border-border text-foreground hover:bg-muted disabled:opacity-50"
+        >
+          {running && <Loader2 size={10} className="animate-spin" />}
+          Run
+        </button>
+      </div>
+
+      {crossExams.length > 0 && (
+        <div className="flex flex-col gap-3 mt-1">
+          {crossExams.map((xe) => (
+            <div key={xe.id} className="rounded border border-border/50 bg-muted/20 p-3">
+              <div className="text-xs font-medium text-muted-foreground mb-1.5">
+                {xe.identity} / {xe.model}
+              </div>
+              {xe.verdict ? (
+                <div className="text-sm whitespace-pre-wrap leading-relaxed">{xe.verdict}</div>
+              ) : (
+                <div className="text-xs text-muted-foreground flex items-center gap-1.5">
+                  <Loader2 size={10} className="animate-spin" /> Running…
+                </div>
+              )}
+            </div>
+          ))}
+        </div>
+      )}
+    </div>
+  );
+}
+
+// ─── ArenaPane ────────────────────────────────────────────────────────────────
+
+interface Props {
+  state: ArenaState;
+  projectId: string; // available for future use (e.g. file browser affordance)
+  onClose: () => void;
+}
+
+export function ArenaPane({ state, onClose }: Props) {
+  const [battle, setBattle] = useState<BattleShape | null>(null);
+  const [contestantRows, setContestantRows] = useState<ContestantRowState[]>([]);
+  const [crossExams, setCrossExams] = useState<CrossExaminationShape[]>([]);
+  const [analysis, setAnalysis] = useState<string | null>(null);
+  const [expandedId, setExpandedId] = useState<string | null>(null);
+  const [stopping, setStopping] = useState(false);
+  const [reanalyzing, setReanalyzing] = useState(false);
+  const startTimesRef = useRef<Map<string, number>>(new Map());
+
+  const snapshot = useProviderSnapshot();
+
+  // Fetch current battle state on mount / battle_id change.
+  useEffect(() => {
+    setBattle(null);
+    setContestantRows([]);
+    setCrossExams([]);
+    setAnalysis(null);
+    setExpandedId(null);
+
+    api.battles.get(state.battle_id)
+      .then(({ battle: b, contestants, cross_examinations }) => {
+        setBattle(b);
+        setContestantRows(
+          contestants.map((c) => ({
+            data: c,
+            output: '',
+            startedAt: c.status === 'running' ? Date.now() : null,
+          })),
+        );
+        setCrossExams(cross_examinations);
+        // Fetch analysis text if battle is already completed.
+        if (b.status === 'completed') {
+          api.battles.getAnalysis(state.battle_id)
+            .then(({ text }) => setAnalysis(text))
+            .catch(() => {});
+        }
+        // Auto-expand first running contestant.
+        const firstRunning = contestants.find((c) => c.status === 'running');
+        if (firstRunning) setExpandedId(firstRunning.id);
+      })
+      .catch(() => {});
+  }, [state.battle_id]);
+
+  // Subscribe to live battle/contestant frames.
+  useEffect(() => {
+    return sessionEvents.subscribe((ev) => {
+      if (ev.type === 'battle_started' && ev.battle_id === state.battle_id) {
+        setContestantRows((prev) => {
+          if (prev.length > 0) return prev;
+          return ev.contestants.map((c) => ({
+            data: {
+              id: c.id,
+              battle_id: ev.battle_id,
+              identity: c.identity,
+              model: c.model,
+              lane: c.lane,
+              task_id: null,
+              worktree_id: null,
+              status: 'queued' as const,
+              duration_ms: null,
+              tokens_per_sec: null,
+              cost_tokens: null,
+              result_path: null,
+              error: null,
+              created_at: new Date().toISOString(),
+              updated_at: new Date().toISOString(),
+            },
+            output: '',
+            startedAt: null,
+          }));
+        });
+      } else if (ev.type === 'contestant_updated' && ev.battle_id === state.battle_id) {
+        setContestantRows((prev) =>
+          prev.map((row) => {
+            if (row.data.id !== ev.contestant_id) return row;
+            const updatedData: ContestantShape = {
+              ...row.data,
+              ...(ev.status != null ? { status: ev.status } : {}),
+              ...(ev.duration_ms != null ? { duration_ms: ev.duration_ms } : {}),
+              ...(ev.tokens_per_sec != null ? { tokens_per_sec: ev.tokens_per_sec } : {}),
+              ...(ev.error != null ? { error: ev.error } : {}),
+            };
+            const newStartedAt =
+              ev.status === 'running' && row.startedAt == null
+                ? Date.now()
+                : ev.status === 'done' || ev.status === 'error'
+                  ? null
+                  : row.startedAt;
+            if (ev.status === 'running') {
+              startTimesRef.current.set(ev.contestant_id, newStartedAt ?? Date.now());
+              setExpandedId(ev.contestant_id);
+            }
+            return {
+              data: updatedData,
+              output: ev.delta ? row.output + ev.delta : row.output,
+              startedAt: newStartedAt,
+            };
+          }),
+        );
+        if (ev.battle_status) {
+          setBattle((prev) => prev ? { ...prev, status: ev.battle_status! } : prev);
+        }
+      } else if (ev.type === 'battle_updated' && ev.battle_id === state.battle_id) {
+        setBattle((prev) => {
+          if (!prev) return prev;
+          return {
+            ...prev,
+            ...(ev.status != null ? { status: ev.status } : {}),
+            ...(ev.winner_contestant_id !== undefined ? { winner_contestant_id: ev.winner_contestant_id } : {}),
+          };
+        });
+        if (ev.analysis_ready) {
+          api.battles.getAnalysis(state.battle_id)
+            .then(({ text }) => setAnalysis(text))
+            .catch(() => setAnalysis('Analysis ready — failed to load text.'));
+        }
+        if (ev.cross_exam_id) {
+          // Refetch cross-exams to get the latest verdict.
+          api.battles.get(state.battle_id)
+            .then(({ cross_examinations }) => setCrossExams(cross_examinations))
+            .catch(() => {});
+        }
+      }
+    });
+  }, [state.battle_id]);
+
+  const toggleExpand = useCallback((id: string) => {
+    setExpandedId((prev) => (prev === id ? null : id));
+  }, []);
+
+  async function handleStop() {
+    if (stopping) return;
+    setStopping(true);
+    try {
+      await api.battles.stop(state.battle_id);
+    } catch {
+      // non-fatal
+    } finally {
+      setStopping(false);
+    }
+  }
+
+  async function handleReanalyze() {
+    if (reanalyzing) return;
+    setReanalyzing(true);
+    try {
+      await api.battles.analyze(state.battle_id);
+      toast.success('Re-analysis triggered');
+    } catch (err) {
+      toast.error(err instanceof Error ? err.message : 'Re-analysis failed');
+    } finally {
+      setReanalyzing(false);
+    }
+  }
+
+  function handleOpenResults() {
+    if (!battle?.results_path) return;
+    sessionEvents.emit({ type: 'open_file_in_browser', path: battle.results_path });
+  }
+
+  function handleCopyAnalysis() {
+    if (!analysis) return;
+    navigator.clipboard.writeText(analysis).catch(() => toast.error('Clipboard write failed'));
+  }
+
+  const battleStatus = battle?.status ?? 'running';
+  const isRunning = battleStatus === 'running' || battleStatus === 'pending';
+  const isCompleted = battleStatus === 'completed';
+  const winnerId = battle?.winner_contestant_id;
+  const winnerRow = winnerId ? contestantRows.find((r) => r.data.id === winnerId) : null;
+  const winnerLabel = winnerRow ? `${winnerRow.data.identity} / ${winnerRow.data.model}` : null;
+
+  return (
+    <div className="flex flex-col h-full min-h-0 overflow-hidden">
+      {/* Header */}
+      <div className="flex items-center gap-2 border-b border-border bg-muted/20 px-3 py-2 shrink-0">
+        <Swords size={13} className="text-muted-foreground shrink-0" />
+        <span className="text-sm font-medium truncate min-w-0 flex-1" title={state.prompt}>
+          {state.prompt.length > 60 ? state.prompt.slice(0, 60) + '…' : state.prompt}
+        </span>
+        <span className="text-xs text-muted-foreground shrink-0 capitalize">{state.battle_type}</span>
+
+        {winnerLabel && (
+          <span
+            className="text-xs px-1.5 py-0.5 rounded bg-emerald-500/10 text-emerald-600 dark:text-emerald-400 shrink-0 hidden sm:block truncate max-w-[130px]"
+            title={`Winner: ${winnerLabel}`}
+          >
+            ✓ {winnerLabel}
+          </span>
+        )}
+
+        <div className="ml-auto flex items-center gap-1 shrink-0">
+          {isRunning ? (
+            <button
+              type="button"
+              onClick={() => void handleStop()}
+              disabled={stopping}
+              className="inline-flex items-center gap-1 text-xs px-1.5 py-0.5 rounded border border-border text-muted-foreground hover:text-foreground hover:bg-muted disabled:opacity-50"
+              title="Stop battle"
+            >
+              Stop
+            </button>
+          ) : (
+            <span
+              className={cn(
+                'text-xs px-1.5 py-0.5 rounded',
+                isCompleted
+                  ? 'text-emerald-600 bg-emerald-500/10'
+                  : battleStatus === 'failed' || battleStatus === 'cancelled'
+                    ? 'text-destructive bg-destructive/10'
+                    : 'text-muted-foreground bg-muted/40',
+              )}
+            >
+              {battleStatus}
+            </span>
+          )}
+
+          {isCompleted && (
+            <DropdownMenu>
+              <DropdownMenuTrigger asChild>
+                <button
+                  type="button"
+                  className="inline-flex items-center justify-center p-1 rounded text-muted-foreground hover:bg-muted hover:text-foreground"
+                  aria-label="Battle options"
+                >
+                  <MoreHorizontal size={14} />
+                </button>
+              </DropdownMenuTrigger>
+              <DropdownMenuContent align="end">
+                <DropdownMenuItem onSelect={() => void handleReanalyze()} disabled={reanalyzing}>
+                  <RotateCcw size={14} /> Re-analyze
+                </DropdownMenuItem>
+                {battle?.results_path && (
+                  <DropdownMenuItem onSelect={handleOpenResults}>
+                    Open results folder
+                  </DropdownMenuItem>
+                )}
+                {analysis && (
+                  <DropdownMenuItem onSelect={handleCopyAnalysis}>
+                    Copy analysis
+                  </DropdownMenuItem>
+                )}
+              </DropdownMenuContent>
+            </DropdownMenu>
+          )}
+
+          <button
+            type="button"
+            onClick={onClose}
+            className="inline-flex items-center justify-center p-1 rounded text-muted-foreground hover:bg-muted hover:text-foreground"
+            aria-label="Close pane"
+            title="Close pane"
+          >
+            <X size={12} />
+          </button>
+        </div>
+      </div>
+
+      {/* Body */}
+      <div className="flex-1 min-h-0 overflow-y-auto">
+        {/* Analysis panel */}
+        {analysis && (
+          <div className="border-b border-border p-4">
+            <div className="text-xs font-medium text-muted-foreground uppercase tracking-wide mb-2 pb-1 border-b border-border/50">
+              Analysis
+            </div>
+            <div className="text-sm text-foreground whitespace-pre-wrap leading-relaxed">
+              {analysis}
+            </div>
+            {winnerLabel && (
+              <div className="mt-2 text-sm font-medium text-emerald-600 dark:text-emerald-400">
+                Winner: {winnerLabel}
+              </div>
+            )}
+          </div>
+        )}
+
+        {/* Empty state */}
+        {contestantRows.length === 0 && !analysis && (
+          <div className="flex items-center justify-center h-24 text-sm text-muted-foreground">
+            Starting battle…
+          </div>
+        )}
+
+        {/* Contestant roster */}
+        <div className="divide-y divide-border">
+          {contestantRows.map((row) => (
+            <ContestantRow
+              key={row.data.id}
+              row={row}
+              isExpanded={expandedId === row.data.id}
+              onToggle={() => toggleExpand(row.data.id)}
+              isWinner={winnerId === row.data.id}
+              battleId={state.battle_id}
+              battleType={battle?.battle_type ?? state.battle_type}
+            />
+          ))}
+        </div>
+
+        {/* Cross-examination panel — available after battle finishes */}
+        {!isRunning && (
+          <CrossExaminationPanel
+            battleId={state.battle_id}
+            crossExams={crossExams}
+            snapshot={snapshot}
+          />
+        )}
+      </div>
+    </div>
+  );
+}

apps/web/src/hooks/sessionEvents.ts

@@ -3,7 +3,11 @@
 // also refresh the sidebar's session list).
 
 import type {
+  ArenaState,
+  BattleShape,
   Chat,
+  ContestantShape,
+  CrossExaminationShape,
   ErrorReason,
   HtmlArtifactState,
   MarkdownArtifactState,
@@ -231,6 +235,53 @@ export interface FlowRunStepUpdatedEvent {
   report?: string;
 }
 
+// Arena: emitted by "New Arena" menu items to request the launcher dialog.
+export interface OpenArenaLauncherEvent {
+  type: 'open_arena_launcher';
+  project_id: string;
+  placement?: 'new' | 'split';
+}
+
+// Arena: emitted after a battle is created to open/focus the arena pane.
+export interface OpenArenaPaneEvent {
+  type: 'open_arena_pane';
+  state: ArenaState;
+  placement?: 'new' | 'split';
+}
+
+// Arena: battle lifecycle frames forwarded from the coder user channel.
+export interface BattleStartedEvent {
+  type: 'battle_started';
+  battle_id: string;
+  battle_type: 'coding' | 'qa';
+  prompt: string;
+  contestants: Array<{ id: string; identity: string; model: string; lane: 'local' | 'cloud' }>;
+}
+
+export interface ContestantUpdatedEvent {
+  type: 'contestant_updated';
+  battle_id: string;
+  contestant_id: string;
+  status?: 'queued' | 'running' | 'done' | 'error';
+  duration_ms?: number;
+  tokens_per_sec?: number;
+  battle_status?: 'pending' | 'running' | 'completed' | 'failed' | 'cancelled';
+  delta?: string;
+  error?: string;
+}
+
+export interface BattleUpdatedEvent {
+  type: 'battle_updated';
+  battle_id: string;
+  status?: 'pending' | 'running' | 'completed' | 'failed' | 'cancelled';
+  winner_contestant_id?: string | null;
+  analysis_ready?: boolean;
+  cross_exam_id?: string;
+}
+
+// Re-export arena API shapes for consumers that need the full battle data.
+export type { BattleShape, ContestantShape, CrossExaminationShape };
+
 export type SessionEvent =
   | SessionRenamedEvent
   | ProjectCreatedEvent
@@ -262,7 +313,12 @@ export type SessionEvent =
   | OpenOrchestratorPaneEvent
   | FlowRunStartedEvent
   | FlowRunStepUpdatedEvent
-  | OpenFlowLauncherEvent;
+  | OpenFlowLauncherEvent
+  | OpenArenaLauncherEvent
+  | OpenArenaPaneEvent
+  | BattleStartedEvent
+  | ContestantUpdatedEvent
+  | BattleUpdatedEvent;
 type Listener = (event: SessionEvent) => void;
 
 const listeners = new Set<Listener>();

apps/web/src/hooks/useCoderUserEvents.ts

@@ -8,7 +8,13 @@
 import { useEffect } from 'react';
 import { WsFrameSchema } from '@boocode/contracts/ws-frames';
 import { sessionEvents } from './sessionEvents';
-import type { FlowRunStartedEvent, FlowRunStepUpdatedEvent } from './sessionEvents';
+import type {
+  BattleStartedEvent,
+  BattleUpdatedEvent,
+  ContestantUpdatedEvent,
+  FlowRunStartedEvent,
+  FlowRunStepUpdatedEvent,
+} from './sessionEvents';
 
 const RECONNECT_INITIAL_MS = 1000;
 const RECONNECT_MAX_MS = 30_000;
@@ -49,6 +55,12 @@ export function useCoderUserEvents(): void {
           sessionEvents.emit(frame as unknown as FlowRunStartedEvent);
         } else if (frame.type === 'flow_run_step_updated') {
           sessionEvents.emit(frame as unknown as FlowRunStepUpdatedEvent);
+        } else if (frame.type === 'battle_started') {
+          sessionEvents.emit(frame as unknown as BattleStartedEvent);
+        } else if (frame.type === 'contestant_updated') {
+          sessionEvents.emit(frame as unknown as ContestantUpdatedEvent);
+        } else if (frame.type === 'battle_updated') {
+          sessionEvents.emit(frame as unknown as BattleUpdatedEvent);
         }
       };
 

apps/web/src/hooks/useSessionStream.ts

@@ -204,6 +204,13 @@ function applyFrame(state: State, frame: WsFrame): State {
       // No-op here to keep TS exhaustiveness satisfied.
       return state;
     }
+    case 'battle_started':
+    case 'contestant_updated':
+    case 'battle_updated': {
+      // Arena frames consumed by ArenaPane's own subscription.
+      // No-op here to keep TS exhaustiveness satisfied.
+      return state;
+    }
   }
 }
 

apps/web/src/hooks/useSidebar.ts

@@ -195,6 +195,13 @@ function applyEvent(prev: SidebarResponse, event: import('./sessionEvents').Sess
     case 'flow_run_step_updated':
       // Consumed by useWorkspacePanes / OrchestratorPane / FlowLauncherDialog; sidebar has no stake.
       return prev;
+    case 'open_arena_launcher':
+    case 'open_arena_pane':
+    case 'battle_started':
+    case 'contestant_updated':
+    case 'battle_updated':
+      // Consumed by useWorkspacePanes / ArenaPane / ArenaLauncherDialog; sidebar has no stake.
+      return prev;
     case 'project_archived': {
       const next = prev.projects.filter((p) => p.id !== event.project_id);
       if (next.length === prev.projects.length) return prev;

apps/web/src/hooks/useWorkspacePanes.ts

@@ -3,6 +3,7 @@ import type { DragEvent } from 'react';
 import { toast } from 'sonner';
 import { api } from '@/api/client';
 import type {
+  ArenaState,
   ClosedPaneEntry,
   HtmlArtifactState,
   MarkdownArtifactState,
@@ -187,6 +188,16 @@ function orchestratorPane(state: OrchestratorState): WorkspacePane {
   };
 }
 
+function arenaPane(state: ArenaState): WorkspacePane {
+  return {
+    id: generateId(),
+    kind: 'arena',
+    chatIds: [],
+    activeChatIdx: -1,
+    arena_state: state,
+  };
+}
+
 // v1.9: settings panes are ephemeral. Filter them out before persisting so a
 // page reload always returns to a clean workspace; the user re-opens via the
 // sidebar Settings button when needed.
@@ -290,6 +301,8 @@ export interface UseWorkspacePanesResult {
   createTab: (paneIdx: number, kind: WorkspaceTabKind) => Promise<void>;
   /** Open an orchestrator run pane (or focus an existing one for the same run_id). */
   addOrchestratorPane: (state: OrchestratorState) => string | null;
+  /** Open an arena battle pane (or focus an existing one for the same battle_id). */
+  addArenaPane: (state: ArenaState) => string | null;
   /** Back-compat alias for createTab(paneIdx, 'coder'). */
   createCoderTab: (paneIdx: number) => Promise<void>;
   // Open-on-first-click, close-on-second-click. Singleton — settings panes
@@ -877,6 +890,38 @@ export function useWorkspacePanes(sessionId: string): UseWorkspacePanesResult {
     });
   }, [addOrchestratorPane]);
 
+  const addArenaPane = useCallback((state: ArenaState): string | null => {
+    let openedId: string | null = null;
+    setPanes((prev) => {
+      const existingIdx = prev.findIndex(
+        (p) => p.kind === 'arena' && p.arena_state?.battle_id === state.battle_id,
+      );
+      if (existingIdx >= 0) {
+        setActivePaneIdx(existingIdx);
+        openedId = prev[existingIdx]!.id;
+        return prev;
+      }
+      if (nonSettingsCount(prev) >= MAX_PANES) {
+        toast.error(`Maximum ${MAX_PANES} panes`);
+        return prev;
+      }
+      const newPane = arenaPane(state);
+      openedId = newPane.id;
+      const next = [...prev, newPane];
+      setActivePaneIdx(next.length - 1);
+      return next;
+    });
+    return openedId;
+  }, []);
+
+  // Arena pane: open via sessionEvents (fired by the launcher).
+  useEffect(() => {
+    return sessionEvents.subscribe((ev) => {
+      if (ev.type !== 'open_arena_pane') return;
+      addArenaPane(ev.state);
+    });
+  }, [addArenaPane]);
+
   // Returns the new settings pane id when one is OPENED (so mobile callers can
   // push ?pane= atomically — see addPaneAndSwitch), or null when it was closed.
   // Id generated outside the updater so a strict-mode double-invoke agrees.
@@ -1121,6 +1166,7 @@ export function useWorkspacePanes(sessionId: string): UseWorkspacePanesResult {
     addSplitPane,
     createTab,
     addOrchestratorPane,
+    addArenaPane,
     createCoderTab,
     toggleSettingsPane,
     removePane,

docs/adr/0001-arena-two-lane-scheduling.md

@@ -0,0 +1,19 @@
+# Arena schedules contestants in a local lane (serial) and a cloud lane (parallel)
+
+A Battle runs the same prompt against 2–6 Contestants. The local llama-swap
+server can only hold one model in memory at a time, so llama-swap-backed
+Contestants are placed in a **local lane** and run strictly one at a time, while
+cloud-backed Contestants (Claude Code, OpenCode-on-cloud) run all in parallel in
+a **cloud lane**; the two lanes run concurrently. We chose this over running
+everything serially (too slow for cloud) or everything in parallel (impossible
+for local, and it would corrupt the speed Benchmark) because the single-model
+constraint is physical and the serial local lane also gives each local model an
+uncontended, fair tokens/sec measurement.
+
+## Consequences
+
+- A Battle's wall-clock is roughly `max(slowest cloud contestant, sum of local
+  contestants)`. Deep local lanes (especially all-local Q&A battles) are slow by
+  design; the launcher warns when the local lane is deep.
+- The speed Benchmark (tokens/sec) is only meaningful for local-lane Contestants,
+  which is acceptable since external CLI agents don't report token usage anyway.

docs/adr/0002-arena-dedicated-tables-not-flow-runner.md

@@ -0,0 +1,22 @@
+# Arena gets dedicated battles/contestants tables and replaces the old API-only arena
+
+The Arena feature reuses the dispatcher, the `onTaskTerminal` advance hook, the
+streaming→WS-frame pipeline, and the pane pattern from the Orchestrator, but
+persists to its **own `battles` + `contestants` tables** rather than the
+Orchestrator's `flow_runs`/`flow_steps`. A Battle is not shaped like a flow — it
+has two scheduling lanes, per-contestant benchmarks, on-disk results folders, a
+two-stage analysis, and cross-examinations — so modelling it as flow steps would
+fight the schema. Each Contestant links to a real `tasks` row via `task_id`,
+inheriting all worktree/streaming/dispatch machinery. This also **replaces the
+earlier v2.0.5 API-only arena** (`POST /api/arena`, `tasks.arena_id`,
+select-winner): that feature had no UI and no users, and the new Arena is a
+strict superset, so the old routes and the `tasks.arena_id` column are removed
+rather than left as a second, competing "arena" concept.
+
+## Consequences
+
+- Analysis and cross-examination run through a small pluggable **Analyzer** seam
+  (v1 = default-model two-stage judge). A v2 that drives a Han Orchestrator flow
+  as the analyzer slots in behind that seam without a schema change.
+- The `arena` pane kind, `ArenaState`, and `battle_*` WS frames are added
+  alongside (not folded into) the Orchestrator's, mirroring its patterns.

packages/contracts/package.json

@@ -28,6 +28,10 @@
     "./worktree-risk": {
       "types": "./dist/worktree-risk.d.ts",
       "default": "./dist/worktree-risk.js"
+    },
+    "./arena": {
+      "types": "./dist/arena.d.ts",
+      "default": "./dist/arena.js"
     }
   },
   "scripts": {

packages/contracts/src/arena.ts

@@ -0,0 +1,55 @@
+/** Arena types — single source of truth for cross-app Arena wire contracts. */
+
+export type BattleType = 'coding' | 'qa';
+export type BattleStatus = 'pending' | 'running' | 'completed' | 'failed' | 'cancelled';
+export type ContestantStatus = 'queued' | 'running' | 'done' | 'error';
+export type ContestantLane = 'local' | 'cloud';
+
+// Pane state — carried on the WorkspacePane row, mirrors OrchestratorState.
+export interface ArenaState {
+  battle_id: string;
+  battle_type: BattleType;
+  prompt: string;
+}
+
+export interface BattleShape {
+  id: string;
+  project_id: string;
+  battle_type: BattleType;
+  prompt: string;
+  status: BattleStatus;
+  winner_contestant_id: string | null;
+  results_path: string | null;
+  error: string | null;
+  created_at: string;
+  updated_at: string;
+}
+
+export interface ContestantShape {
+  id: string;
+  battle_id: string;
+  /** Backend name (coding) or persona name (qa). Unique per (battle, model) pair. */
+  identity: string;
+  model: string;
+  lane: ContestantLane;
+  task_id: string | null;
+  worktree_id: string | null;
+  status: ContestantStatus;
+  duration_ms: number | null;
+  tokens_per_sec: number | null;
+  cost_tokens: number | null;
+  result_path: string | null;
+  error: string | null;
+  created_at: string;
+  updated_at: string;
+}
+
+export interface CrossExaminationShape {
+  id: string;
+  battle_id: string;
+  /** Backend + model performing the cross-examination. */
+  identity: string;
+  model: string;
+  verdict: string | null;
+  created_at: string;
+}

packages/contracts/src/ws-frames.ts

@@ -358,6 +358,53 @@ export const FlowRunStepUpdatedFrame = z.object({
   report: z.string().optional(),
 });
 
+// ---- arena frames ----------------------------------------------------------
+
+const ContestantManifestEntry = z.object({
+  id: Uuid,
+  identity: z.string().min(1),
+  model: z.string().min(1),
+  lane: z.enum(['local', 'cloud']),
+});
+
+// Published once when a battle starts. Carries the contestant roster so the
+// ArenaPane can build its grid immediately.
+export const BattleStartedFrame = z.object({
+  type: z.literal('battle_started'),
+  battle_id: Uuid,
+  battle_type: z.enum(['coding', 'qa']),
+  prompt: z.string(),
+  contestants: z.array(ContestantManifestEntry),
+});
+
+// Published on every contestant status change or streaming update.
+// `delta` carries the latest chunk of streaming output while status='running'.
+// `battle_status` is present only on the final transition that closes the battle.
+export const ContestantUpdatedFrame = z.object({
+  type: z.literal('contestant_updated'),
+  battle_id: Uuid,
+  contestant_id: Uuid,
+  status: z.enum(['queued', 'running', 'done', 'error']).optional(),
+  duration_ms: z.number().int().nonnegative().optional(),
+  tokens_per_sec: z.number().nonnegative().optional(),
+  battle_status: z.enum(['pending', 'running', 'completed', 'failed', 'cancelled']).optional(),
+  delta: z.string().optional(),
+  error: z.string().optional(),
+});
+
+// Published when battle-level state changes that don't ride on a contestant
+// update: analysis finished, winner set, cross-exam verdict ready. The pane
+// uses this to update its analysis panel and winner badge without a refetch.
+// Fields are all optional — publishers include only what changed.
+export const BattleUpdatedFrame = z.object({
+  type: z.literal('battle_updated'),
+  battle_id: Uuid,
+  status: z.enum(['pending', 'running', 'completed', 'failed', 'cancelled']).optional(),
+  winner_contestant_id: Uuid.nullable().optional(),
+  analysis_ready: z.boolean().optional(),
+  cross_exam_id: Uuid.optional(),
+});
+
 // ---- discriminated union ---------------------------------------------------
 
 export const WsFrameSchema = z.discriminatedUnion('type', [
@@ -381,6 +428,10 @@ export const WsFrameSchema = z.discriminatedUnion('type', [
   // orchestrator
   FlowRunStartedFrame,
   FlowRunStepUpdatedFrame,
+  // arena
+  BattleStartedFrame,
+  ContestantUpdatedFrame,
+  BattleUpdatedFrame,
   // per-user
   ChatStatusFrame,
   SessionUpdatedFrame,
@@ -425,6 +476,9 @@ export const KNOWN_FRAME_TYPES: readonly WsFrame['type'][] = [
   'agent_status_updated',
   'flow_run_started',
   'flow_run_step_updated',
+  'battle_started',
+  'contestant_updated',
+  'battle_updated',
   'chat_status',
   'session_updated',
   'session_renamed',