feat: implement MAGI 3-agent arbiter system for deadlock resolution

Introduces a third agent role (arbiter) that is summoned on-demand when owner and reviewer reach a deadlock (same verdict 3+ rounds). Architecture: - 3 Discord bots: owner (codex), reviewer (claude), arbiter (claude/codex) - Arbiter is NOT always-on — only invoked when deadlock detected - Arbiter renders binding verdict: PROCEED/REVISE/RESET/ESCALATE - Non-escalate verdicts reset round_trip_count and resume ping-pong - Backward compatible: ARBITER_AGENT_TYPE unset = existing 2-agent mode Changes across 13 source files + 7 test files: - types.ts: PairedRoomRole += 'arbiter', new statuses, ArbiterVerdict type - config.ts: ARBITER_AGENT_TYPE, ARBITER_SERVICE_ID, ARBITER_DEADLOCK_THRESHOLD - db.ts: schema migration (arbiter columns in channel_owner + paired_tasks) - service-routing.ts: arbiter_service_id in lease - room-role-context.ts: arbiter role detection - paired-execution-context.ts: deadlock->arbiter, verdict handling - message-runtime.ts: arbiter turn routing, cursor, sender labeling - message-agent-executor.ts: arbiter mode, failover - agent-runner.ts + environment.ts: arbiter container mode - platform-prompts.ts: arbiter prompt loading New files: - prompts/arbiter-paired-room.md: arbiter system prompt - src/arbiter-context.ts: builds conversation context for arbiter judgment
2026-03-30 22:20:42 +09:00
parent 58ed721d6e
commit a043f2b79a
20 changed files with 587 additions and 53 deletions
--- a/prompts/arbiter-paired-room.md
+++ b/prompts/arbiter-paired-room.md
@@ -0,0 +1,28 @@
+# Arbiter Paired Room Rules
+
+You are the **arbiter** in a MAGI system with three agents: owner (implementer), reviewer (verifier), and you (judge).
+
+You have been summoned because the owner and reviewer reached a deadlock after multiple rounds without progress.
+
+## Your Role
+
+- Read the conversation history between owner and reviewer
+- Understand what each side is arguing
+- Render a binding verdict based on evidence
+
+## Verdict Format
+
+**Start your first line** with one of these four verdicts. This is required.
+
+- **PROCEED** — The owner's approach is correct. The reviewer should approve. Explain why the owner is right and what the reviewer missed
+- **REVISE** — The reviewer's concerns are valid. Tell the owner exactly what to fix. Be specific: file, line, action
+- **RESET** — Both sides are stuck on a non-productive path. Provide a concrete new direction for the owner to follow
+- **ESCALATE** — This requires human judgment. Explain what decision only a human can make
+
+## Rules
+
+- Base your verdict on evidence (code, test output, logs), not on who said what first
+- Your verdict is final for this deadlock cycle — after it, work resumes normally
+- You do NOT implement or review code — you only judge the disagreement
+- Keep your verdict concise — state the decision, the evidence, and the required action
+- If both sides are saying the same thing but not acting on it, call it out and direct the owner to act
--- a/src/agent-runner-environment.test.ts
+++ b/src/agent-runner-environment.test.ts
@@ -44,6 +44,7 @@ vi.mock('./service-routing.js', () => ({
    chat_jid: 'dc:test',
    owner_service_id: 'claude',
    reviewer_service_id: 'codex-main',
+    arbiter_service_id: null,
    activated_at: null,
    reason: null,
    explicit: false,
@@ -206,6 +207,7 @@ describe('prepareGroupEnvironment codex auth handling', () => {
      chat_jid: 'dc:test',
      owner_service_id: 'codex-review',
      reviewer_service_id: 'codex-main',
+      arbiter_service_id: null,
      activated_at: '2026-03-28T00:00:00.000Z',
      reason: 'claude-429',
      explicit: true,
@@ -310,6 +312,7 @@ describe('prepareGroupEnvironment codex auth handling', () => {
      chat_jid: 'dc:test',
      owner_service_id: 'claude',
      reviewer_service_id: 'codex-main',
+      arbiter_service_id: null,
      activated_at: null,
      reason: null,
      explicit: false,
--- a/src/agent-runner-environment.ts
+++ b/src/agent-runner-environment.ts
@@ -23,6 +23,7 @@ import {
  resolveServiceTaskSessionsPath,
 } from './group-folder.js';
 import {
+  readArbiterPrompt,
  readPairedRoomPrompt,
  readPlatformPrompt,
 } from './platform-prompts.js';
@@ -527,8 +528,9 @@ export function prepareContainerSessionEnvironment(args: {
  chatJid: string;
  isMain: boolean;
  memoryBriefing?: string;
+  role?: 'reviewer' | 'arbiter';
 }): void {
-  const { sessionDir, chatJid, isMain, memoryBriefing } = args;
+  const { sessionDir, chatJid, isMain, memoryBriefing, role = 'reviewer' } = args;
  const projectRoot = process.cwd();

  fs.mkdirSync(sessionDir, { recursive: true });
@@ -541,10 +543,12 @@ export function prepareContainerSessionEnvironment(args: {
  ];
  syncDirectoryEntries(skillSources, path.join(sessionDir, 'skills'));

-  // Build CLAUDE.md with reviewer-appropriate prompts
+  // Build CLAUDE.md with role-appropriate prompts (reviewer or arbiter)
  const claudePlatformPrompt = readPlatformPrompt('claude-code', projectRoot);
  const claudePairedRoomPrompt = isPairedRoomJid(chatJid)
-    ? readPairedRoomPrompt('claude-code', projectRoot)
+    ? (role === 'arbiter'
+        ? readArbiterPrompt(projectRoot)
+        : readPairedRoomPrompt('claude-code', projectRoot))
    : undefined;
  const globalDir = path.join(GROUPS_DIR, 'global');
  const globalClaudeMdPath = path.join(globalDir, 'CLAUDE.md');
--- a/src/agent-runner.ts
+++ b/src/agent-runner.ts
@@ -75,7 +75,7 @@ export async function runAgentProcess(
  // ── Reviewer container mode ─────────────────────────────────────
  // Reviewers always run inside a Docker container with read-only source
  // mount for kernel-level write protection. Docker is required.
-  if (envOverrides?.EJCLAW_REVIEWER_RUNTIME === '1') {
+  if (envOverrides?.EJCLAW_REVIEWER_RUNTIME === '1' || envOverrides?.EJCLAW_ARBITER_RUNTIME === '1') {
    const ownerWorkspaceDir =
      envOverrides?.EJCLAW_WORK_DIR || group.workDir || process.cwd();

@@ -83,11 +83,13 @@ export async function runAgentProcess(
    // so the Claude SDK inside the container has platform & paired room prompts.
    const sessionDir = envOverrides?.CLAUDE_CONFIG_DIR;
    if (sessionDir) {
+      const containerRole = envOverrides?.EJCLAW_ARBITER_RUNTIME === '1' ? 'arbiter' as const : 'reviewer' as const;
      prepareContainerSessionEnvironment({
        sessionDir,
        chatJid: input.chatJid,
        isMain: input.isMain,
        memoryBriefing: input.memoryBriefing,
+        role: containerRole,
      });
    }

--- a/src/arbiter-context.ts
+++ b/src/arbiter-context.ts
@@ -0,0 +1,29 @@
+import { getRecentChatMessages } from './db.js';
+import { formatMessages } from './router.js';
+
+export function buildArbiterContextPrompt(args: {
+  chatJid: string;
+  taskId: string;
+  roundTripCount: number;
+  timezone: string;
+  recentTurnLimit?: number;
+}): string {
+  const { chatJid, taskId, roundTripCount, timezone, recentTurnLimit = 20 } = args;
+
+  const recentMessages = getRecentChatMessages(chatJid, recentTurnLimit);
+  const conversationContext = formatMessages(recentMessages, timezone);
+
+  return [
+    `<arbiter-context>`,
+    `<task-id>${taskId}</task-id>`,
+    `<round-trips>${roundTripCount}</round-trips>`,
+    `<reason>Deadlock detected: owner and reviewer exchanged ${roundTripCount} rounds without resolution</reason>`,
+    `</arbiter-context>`,
+    ``,
+    `<conversation-history>`,
+    conversationContext,
+    `</conversation-history>`,
+    ``,
+    `Review the conversation above and render your verdict.`,
+  ].join('\n');
+}
--- a/src/config.ts
+++ b/src/config.ts
@@ -110,6 +110,28 @@ export const REVIEWER_SERVICE_ID_FOR_TYPE =
    ? CLAUDE_SERVICE_ID
    : CODEX_REVIEW_SERVICE_ID;

+/** Arbiter agent type. Disabled by default. Set ARBITER_AGENT_TYPE=codex or claude-code to enable. */
+const rawArbiterAgentType = getEnv('ARBITER_AGENT_TYPE');
+export const ARBITER_AGENT_TYPE: AgentType | undefined =
+  rawArbiterAgentType === 'codex' || rawArbiterAgentType === 'claude-code'
+    ? rawArbiterAgentType
+    : undefined;
+
+/** Service ID for the arbiter. Re-uses codex-review bot by default when arbiter is enabled. */
+export const ARBITER_SERVICE_ID = ARBITER_AGENT_TYPE
+  ? (getEnv('ARBITER_SERVICE_ID') || CODEX_REVIEW_SERVICE_ID)
+  : null;
+
+/** Number of consecutive owner↔reviewer round trips before arbiter is auto-requested. */
+export const ARBITER_DEADLOCK_THRESHOLD = parseInt(
+  getEnv('ARBITER_DEADLOCK_THRESHOLD') || '3',
+  10,
+);
+
+export function isArbiterEnabled(): boolean {
+  return ARBITER_AGENT_TYPE !== undefined;
+}
+
 // Max owner↔reviewer round trips per task. 0 = unlimited.
 const rawMaxRoundTrips = getEnv('PAIRED_MAX_ROUND_TRIPS') || '10';
 export const PAIRED_MAX_ROUND_TRIPS =
--- a/src/db.test.ts
+++ b/src/db.test.ts
@@ -559,6 +559,8 @@ describe('paired task state', () => {
      round_trip_count: 0,
      review_requested_at: null,
      status: 'active',
+      arbiter_verdict: null,
+      arbiter_requested_at: null,
      created_at: '2026-03-28T00:00:00.000Z',
      updated_at: '2026-03-28T00:00:00.000Z',
    });
@@ -598,6 +600,8 @@ describe('paired task state', () => {
      round_trip_count: 0,
      review_requested_at: null,
      status: 'active',
+      arbiter_verdict: null,
+      arbiter_requested_at: null,
      created_at: '2026-03-28T00:00:00.000Z',
      updated_at: '2026-03-28T00:00:00.000Z',
    });
--- a/src/db.ts
+++ b/src/db.ts
@@ -58,6 +58,7 @@ export interface ChannelOwnerLeaseRow {
  chat_jid: string;
  owner_service_id: string;
  reviewer_service_id: string | null;
+  arbiter_service_id: string | null;
  activated_at: string | null;
  reason: string | null;
 }
@@ -252,9 +253,11 @@ function createSchema(database: Database.Database): void {
      review_requested_at TEXT,
      round_trip_count INTEGER NOT NULL DEFAULT 0,
      status TEXT NOT NULL DEFAULT 'active',
+      arbiter_verdict TEXT,
+      arbiter_requested_at TEXT,
      created_at TEXT NOT NULL,
      updated_at TEXT NOT NULL,
-      CHECK (status IN ('active', 'review_ready', 'in_review', 'merge_ready', 'completed'))
+      CHECK (status IN ('active', 'review_ready', 'in_review', 'merge_ready', 'completed', 'arbiter_requested', 'in_arbitration'))
    );
    CREATE INDEX IF NOT EXISTS idx_paired_tasks_chat_status
      ON paired_tasks(chat_jid, status, updated_at);
@@ -278,6 +281,7 @@ function createSchema(database: Database.Database): void {
      chat_jid TEXT PRIMARY KEY,
      owner_service_id TEXT NOT NULL,
      reviewer_service_id TEXT,
+      arbiter_service_id TEXT,
      activated_at TEXT,
      reason TEXT
    );
@@ -580,15 +584,74 @@ function createSchema(database: Database.Database): void {
        review_requested_at TEXT,
        round_trip_count INTEGER NOT NULL DEFAULT 0,
        status TEXT NOT NULL DEFAULT 'active',
+        arbiter_verdict TEXT,
+        arbiter_requested_at TEXT,
        created_at TEXT NOT NULL,
        updated_at TEXT NOT NULL,
-        CHECK (status IN ('active', 'review_ready', 'in_review', 'merge_ready', 'completed'))
+        CHECK (status IN ('active', 'review_ready', 'in_review', 'merge_ready', 'completed', 'arbiter_requested', 'in_arbitration'))
      );
      CREATE INDEX IF NOT EXISTS idx_paired_tasks_chat_status
        ON paired_tasks(chat_jid, status, updated_at);
    `);
  }

+  // Migration: add arbiter columns to paired_tasks and rebuild CHECK constraint
+  // if it doesn't include arbiter statuses
+  {
+    const ptSqlRow = database
+      .prepare(
+        `SELECT sql FROM sqlite_master WHERE type = 'table' AND name = 'paired_tasks'`,
+      )
+      .get() as { sql?: string } | undefined;
+    const ptSql = ptSqlRow?.sql || '';
+    if (ptSql && !ptSql.includes('arbiter_requested')) {
+      // CHECK constraint cannot be altered in SQLite — rebuild table
+      database.exec(`
+        CREATE TABLE paired_tasks_new (
+          id TEXT PRIMARY KEY,
+          chat_jid TEXT NOT NULL,
+          group_folder TEXT NOT NULL,
+          owner_service_id TEXT NOT NULL,
+          reviewer_service_id TEXT NOT NULL,
+          title TEXT,
+          source_ref TEXT,
+          plan_notes TEXT,
+          review_requested_at TEXT,
+          round_trip_count INTEGER NOT NULL DEFAULT 0,
+          status TEXT NOT NULL DEFAULT 'active',
+          arbiter_verdict TEXT,
+          arbiter_requested_at TEXT,
+          created_at TEXT NOT NULL,
+          updated_at TEXT NOT NULL,
+          CHECK (status IN ('active', 'review_ready', 'in_review', 'merge_ready', 'completed', 'arbiter_requested', 'in_arbitration'))
+        );
+        INSERT INTO paired_tasks_new (
+          id, chat_jid, group_folder, owner_service_id, reviewer_service_id,
+          title, source_ref, plan_notes, review_requested_at,
+          round_trip_count, status, created_at, updated_at
+        )
+        SELECT
+          id, chat_jid, group_folder, owner_service_id, reviewer_service_id,
+          title, source_ref, plan_notes, review_requested_at,
+          round_trip_count, status, created_at, updated_at
+        FROM paired_tasks;
+        DROP TABLE paired_tasks;
+        ALTER TABLE paired_tasks_new RENAME TO paired_tasks;
+        CREATE INDEX IF NOT EXISTS idx_paired_tasks_chat_status
+          ON paired_tasks(chat_jid, status, updated_at);
+      `);
+    }
+  }
+
+  // Migration: add arbiter_service_id column to channel_owner if it doesn't exist
+  try {
+    database.exec(
+      `ALTER TABLE channel_owner ADD COLUMN arbiter_service_id TEXT`,
+    );
+  } catch {
+    /* column already exists */
+  }
+
  // Drop legacy tables that are no longer used
  for (const table of [
    'paired_executions',
@@ -1871,10 +1934,12 @@ export function createPairedTask(task: PairedTask): void {
        review_requested_at,
        round_trip_count,
        status,
+        arbiter_verdict,
+        arbiter_requested_at,
        created_at,
        updated_at
      )
-      VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+      VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
    `,
  ).run(
    task.id,
@@ -1888,6 +1953,8 @@ export function createPairedTask(task: PairedTask): void {
    task.review_requested_at,
    task.round_trip_count,
    task.status,
+    task.arbiter_verdict,
+    task.arbiter_requested_at,
    task.created_at,
    task.updated_at,
  );
@@ -1943,6 +2010,8 @@ export function updatePairedTask(
      | 'review_requested_at'
      | 'round_trip_count'
      | 'status'
+      | 'arbiter_verdict'
+      | 'arbiter_requested_at'
      | 'updated_at'
    >
  >,
@@ -1974,6 +2043,14 @@ export function updatePairedTask(
    fields.push('status = ?');
    values.push(updates.status);
  }
+  if (updates.arbiter_verdict !== undefined) {
+    fields.push('arbiter_verdict = ?');
+    values.push(updates.arbiter_verdict);
+  }
+  if (updates.arbiter_requested_at !== undefined) {
+    fields.push('arbiter_requested_at = ?');
+    values.push(updates.arbiter_requested_at);
+  }
  if (updates.updated_at !== undefined) {
    fields.push('updated_at = ?');
    values.push(updates.updated_at);
@@ -2070,7 +2147,7 @@ export function getChannelOwnerLease(
 ): ChannelOwnerLeaseRow | undefined {
  return db
    .prepare(
-      `SELECT chat_jid, owner_service_id, reviewer_service_id, activated_at, reason
+      `SELECT chat_jid, owner_service_id, reviewer_service_id, arbiter_service_id, activated_at, reason
       FROM channel_owner
       WHERE chat_jid = ?`,
    )
@@ -2080,7 +2157,7 @@ export function getChannelOwnerLease(
 export function getAllChannelOwnerLeases(): ChannelOwnerLeaseRow[] {
  return db
    .prepare(
-      `SELECT chat_jid, owner_service_id, reviewer_service_id, activated_at, reason
+      `SELECT chat_jid, owner_service_id, reviewer_service_id, arbiter_service_id, activated_at, reason
       FROM channel_owner`,
    )
    .all() as ChannelOwnerLeaseRow[];
@@ -2090,6 +2167,7 @@ export function setChannelOwnerLease(input: {
  chat_jid: string;
  owner_service_id: string;
  reviewer_service_id?: string | null;
+  arbiter_service_id?: string | null;
  activated_at?: string | null;
  reason?: string | null;
 }): void {
@@ -2098,13 +2176,15 @@ export function setChannelOwnerLease(input: {
      chat_jid,
      owner_service_id,
      reviewer_service_id,
+      arbiter_service_id,
      activated_at,
      reason
-    ) VALUES (?, ?, ?, ?, ?)`,
+    ) VALUES (?, ?, ?, ?, ?, ?)`,
  ).run(
    input.chat_jid,
    input.owner_service_id,
    input.reviewer_service_id ?? null,
+    input.arbiter_service_id ?? null,
    input.activated_at ?? new Date().toISOString(),
    input.reason ?? null,
  );
--- a/src/message-agent-executor.test.ts
+++ b/src/message-agent-executor.test.ts
@@ -34,6 +34,7 @@ vi.mock('./service-routing.js', () => ({
    chat_jid: 'group@test',
    owner_service_id: 'claude',
    reviewer_service_id: 'codex-main',
+    arbiter_service_id: null,
    activated_at: null,
    reason: null,
    explicit: false,
@@ -278,6 +279,7 @@ describe('runAgentForGroup room memory', () => {
      chat_jid: 'group@test',
      owner_service_id: 'claude',
      reviewer_service_id: 'claude',
+      arbiter_service_id: null,
      activated_at: null,
      reason: null,
      explicit: false,
@@ -307,6 +309,7 @@ describe('runAgentForGroup room memory', () => {
      chat_jid: 'group@test',
      owner_service_id: 'codex-main',
      reviewer_service_id: 'claude',
+      arbiter_service_id: null,
      activated_at: null,
      reason: null,
      explicit: false,
@@ -326,6 +329,8 @@ describe('runAgentForGroup room memory', () => {
        plan_notes: null,
        round_trip_count: 0,
        status: 'active',
+        arbiter_verdict: null,
+        arbiter_requested_at: null,
        created_at: '2026-03-29T00:00:00.000Z',
        updated_at: '2026-03-29T00:00:00.000Z',
      },
@@ -388,6 +393,8 @@ describe('runAgentForGroup room memory', () => {
        round_trip_count: 0,
        review_requested_at: null,
        status: 'active',
+        arbiter_verdict: null,
+        arbiter_requested_at: null,
        created_at: '2026-03-28T00:00:00.000Z',
        updated_at: '2026-03-28T00:00:00.000Z',
      },
@@ -452,6 +459,7 @@ describe('runAgentForGroup room memory', () => {
      chat_jid: 'group@test',
      owner_service_id: 'codex-main',
      reviewer_service_id: 'claude',
+      arbiter_service_id: null,
      activated_at: null,
      reason: null,
      explicit: false,
@@ -471,6 +479,8 @@ describe('runAgentForGroup room memory', () => {
        round_trip_count: 0,
        review_requested_at: null,
        status: 'active',
+        arbiter_verdict: null,
+        arbiter_requested_at: null,
        created_at: '2026-03-28T00:00:00.000Z',
        updated_at: '2026-03-28T00:00:00.000Z',
      },
--- a/src/message-agent-executor.ts
+++ b/src/message-agent-executor.ts
@@ -34,6 +34,7 @@ import {
  shouldRetryFreshSessionOnAgentFailure,
 } from './session-recovery.js';
 import {
+  ARBITER_AGENT_TYPE,
  CODEX_MAIN_SERVICE_ID,
  CODEX_REVIEW_SERVICE_ID,
  SERVICE_SESSION_SCOPE,
@@ -96,29 +97,41 @@ export async function runAgentForGroup(
    : null;
  const effectiveServiceId =
    pairedTask &&
-    (pairedTask.status === 'review_ready' || pairedTask.status === 'in_review')
-      ? currentLease.reviewer_service_id!
-      : currentLease.owner_service_id;
+    (pairedTask.status === 'arbiter_requested' || pairedTask.status === 'in_arbitration')
+      ? currentLease.arbiter_service_id!
+      : pairedTask &&
+          (pairedTask.status === 'review_ready' || pairedTask.status === 'in_review')
+        ? currentLease.reviewer_service_id!
+        : currentLease.owner_service_id;
  const reviewerMode = effectiveServiceId === currentLease.reviewer_service_id;
+  const arbiterMode = effectiveServiceId === currentLease.arbiter_service_id;

  // Override agent type based on role config (OWNER_AGENT_TYPE / REVIEWER_AGENT_TYPE).
  // When the reviewer uses a different agent type than the group default,
  // swap in the configured type for the duration of this execution.
  const reviewerAgentTypeOverride =
    reviewerMode && REVIEWER_AGENT_TYPE !== (group.agentType || 'claude-code');
-  const effectiveAgentType = reviewerAgentTypeOverride
-    ? REVIEWER_AGENT_TYPE
-    : group.agentType || 'claude-code';
-  const effectiveGroup = reviewerAgentTypeOverride
-    ? { ...group, agentType: REVIEWER_AGENT_TYPE }
-    : group;
+  const arbiterAgentTypeOverride =
+    arbiterMode && ARBITER_AGENT_TYPE != null && ARBITER_AGENT_TYPE !== (group.agentType || 'claude-code');
+  const effectiveAgentType = arbiterAgentTypeOverride
+    ? ARBITER_AGENT_TYPE!
+    : reviewerAgentTypeOverride
+      ? REVIEWER_AGENT_TYPE
+      : group.agentType || 'claude-code';
+  const effectiveGroup = arbiterAgentTypeOverride
+    ? { ...group, agentType: ARBITER_AGENT_TYPE! }
+    : reviewerAgentTypeOverride
+      ? { ...group, agentType: REVIEWER_AGENT_TYPE }
+      : group;
  const isClaudeCodeAgent = effectiveAgentType === 'claude-code';

-  // When the reviewer uses a different agent type than the group default,
-  // use a separate session key so owner and reviewer sessions don't collide.
-  const sessionFolder = reviewerAgentTypeOverride
-    ? `${group.folder}:reviewer`
-    : group.folder;
+  // When the reviewer/arbiter uses a different agent type than the group default,
+  // use a separate session key so owner, reviewer, and arbiter sessions don't collide.
+  const sessionFolder = arbiterAgentTypeOverride
+    ? `${group.folder}:arbiter`
+    : reviewerAgentTypeOverride
+      ? `${group.folder}:reviewer`
+      : group.folder;
  const sessionId = sessions[sessionFolder];
  const memoryBriefing = sessionId
    ? undefined
@@ -194,6 +207,26 @@ export async function runAgentForGroup(
      return false;
    }

+    if (arbiterMode) {
+      // Arbiter failed (e.g. Claude 401/429) — re-trigger arbitration with codex
+      createServiceHandoff({
+        chat_jid: chatJid,
+        group_folder: group.folder,
+        source_service_id: SERVICE_SESSION_SCOPE,
+        target_service_id: CODEX_REVIEW_SERVICE_ID,
+        target_agent_type: 'codex',
+        prompt,
+        start_seq: startSeq ?? null,
+        end_seq: endSeq ?? null,
+        reason: `arbiter-claude-${reason}`,
+      });
+      logger.warn(
+        { chatJid, group: group.name, runId, reason },
+        'Claude arbiter unavailable, handed off arbiter turn to codex',
+      );
+      return true;
+    }
+
    if (reviewerMode) {
      // Reviewer failed (e.g. Claude 401/429) — re-trigger review with codex
      // instead of swapping owner/reviewer roles.
@@ -421,7 +454,9 @@ export async function runAgentForGroup(
        runId,
        provider: effectiveAgentType,
        reviewerMode,
+        arbiterMode,
        reviewerAgentTypeOverride,
+        arbiterAgentTypeOverride,
      },
      `Using provider: ${effectiveAgentType}`,
    );
--- a/src/message-runtime.test.ts
+++ b/src/message-runtime.test.ts
@@ -19,6 +19,7 @@ vi.mock('./config.js', () => ({
  CODEX_MAIN_SERVICE_ID: 'codex-main',
  CODEX_REVIEW_SERVICE_ID: 'codex-review',
  REVIEWER_AGENT_TYPE: 'claude-code',
+  ARBITER_AGENT_TYPE: undefined,
  normalizeServiceId: vi.fn((serviceId: string) => serviceId),
  isClaudeService: vi.fn(() => true),
  isReviewService: vi.fn(() => false),
@@ -556,6 +557,8 @@ describe('createMessageRuntime', () => {
      review_requested_at: '2026-03-30T00:00:00.000Z',
      round_trip_count: 0,
      status: 'review_ready',
+      arbiter_verdict: null,
+      arbiter_requested_at: null,
      created_at: '2026-03-30T00:00:00.000Z',
      updated_at: '2026-03-30T00:00:00.000Z',
    });
--- a/src/message-runtime.ts
+++ b/src/message-runtime.ts
@@ -21,6 +21,7 @@ import {
  type WorkItem,
 } from './db.js';
 import {
+  ARBITER_AGENT_TYPE,
  CLAUDE_SERVICE_ID,
  CODEX_MAIN_SERVICE_ID,
  CODEX_REVIEW_SERVICE_ID,
@@ -46,6 +47,7 @@ import {
  hasAllowedTrigger,
  shouldSkipBotOnlyCollaboration,
 } from './message-runtime-rules.js';
+import { buildArbiterContextPrompt } from './arbiter-context.js';
 import { runAgentForGroup } from './message-agent-executor.js';
 import { MessageTurnController } from './message-turn-controller.js';
 import {
@@ -152,7 +154,9 @@ export function createMessageRuntime(deps: MessageRuntimeDeps): {
          ? 'owner'
          : serviceId === lease.reviewer_service_id
            ? 'reviewer'
-            : msg.sender_name;
+            : serviceId === lease.arbiter_service_id
+              ? 'arbiter'
+              : msg.sender_name;
      return { ...msg, sender_name: role };
    });
  };
@@ -435,14 +439,19 @@ export function createMessageRuntime(deps: MessageRuntimeDeps): {
      return;
    }

-    // Reviewer failover handoffs (reason starts with "reviewer-") should
-    // run via the reviewer channel so they execute in reviewer mode.
+    // Reviewer/arbiter failover handoffs should run via the appropriate
+    // channel so they execute in the correct role mode.
    const isReviewerHandoff = handoff.reason?.startsWith('reviewer-');
+    const isArbiterHandoff = handoff.reason?.startsWith('arbiter-');
    let handoffChannel = channel;
    if (isReviewerHandoff) {
      const revChName =
        REVIEWER_AGENT_TYPE === 'claude-code' ? 'discord' : 'discord-review';
      handoffChannel = findChannelByName(deps.channels, revChName) || channel;
+    } else if (isArbiterHandoff) {
+      const arbChName =
+        ARBITER_AGENT_TYPE === 'claude-code' ? 'discord' : 'discord-review';
+      handoffChannel = findChannelByName(deps.channels, arbChName) || channel;
    }

    const runId = `handoff-${handoff.id}`;
@@ -517,6 +526,18 @@ export function createMessageRuntime(deps: MessageRuntimeDeps): {
      reviewerChannelName,
    );
    const reviewerChannel = foundReviewerChannel || channel;
+
+    // Arbiter channel: route arbiter output through the appropriate bot.
+    const arbiterChannelName =
+      isPairedRoomJid(chatJid) && ARBITER_AGENT_TYPE === 'claude-code'
+        ? 'discord'
+        : 'discord-review';
+    const foundArbiterChannel = findChannelByName(
+      deps.channels,
+      arbiterChannelName,
+    );
+    const arbiterChannel = foundArbiterChannel || channel;
+
    if (isPairedRoomJid(chatJid)) {
      logger.info(
        {
@@ -524,9 +545,12 @@ export function createMessageRuntime(deps: MessageRuntimeDeps): {
          reviewerChannelName,
          foundChannel: foundReviewerChannel?.name ?? null,
          usingChannel: reviewerChannel.name,
+          arbiterChannelName,
+          foundArbiterChannel: foundArbiterChannel?.name ?? null,
+          usingArbiterChannel: arbiterChannel.name,
          availableChannels: deps.channels.map((c) => c.name),
        },
-        'Paired room reviewer channel resolution',
+        'Paired room reviewer/arbiter channel resolution',
      );
    }

@@ -545,7 +569,15 @@ export function createMessageRuntime(deps: MessageRuntimeDeps): {
        pendingTask &&
        (pendingTask.status === 'review_ready' ||
          pendingTask.status === 'in_review');
-      const deliveryChannel = isReviewerWorkItem ? reviewerChannel : channel;
+      const isArbiterWorkItem =
+        pendingTask &&
+        (pendingTask.status === 'arbiter_requested' ||
+          pendingTask.status === 'in_arbitration');
+      const deliveryChannel = isArbiterWorkItem
+        ? arbiterChannel
+        : isReviewerWorkItem
+          ? reviewerChannel
+          : channel;
      const delivered = await deliverOpenWorkItem(
        deliveryChannel,
        openWorkItem,
@@ -619,6 +651,42 @@ export function createMessageRuntime(deps: MessageRuntimeDeps): {
          return deliverySucceeded;
        }

+        // arbiter_requested / in_arbitration: run arbiter turn
+        if (
+          pendingReviewTask &&
+          (pendingReviewTask.status === 'arbiter_requested' ||
+            pendingReviewTask.status === 'in_arbitration')
+        ) {
+          const lastRaw = rawMissedMessages[rawMissedMessages.length - 1];
+          const cursor = lastRaw?.seq ?? lastRaw?.timestamp;
+          if (cursor != null) {
+            advanceLastAgentCursor(
+              deps.getLastAgentTimestamps(),
+              deps.saveState,
+              chatJid,
+              cursor,
+            );
+          }
+
+          const arbiterPrompt = buildArbiterContextPrompt({
+            chatJid,
+            taskId: pendingReviewTask.id,
+            roundTripCount: pendingReviewTask.round_trip_count,
+            timezone: deps.timezone,
+          });
+
+          const { deliverySucceeded } = await executeTurn({
+            group,
+            prompt: arbiterPrompt,
+            chatJid,
+            runId,
+            channel: arbiterChannel,
+            startSeq: null,
+            endSeq: null,
+          });
+          return deliverySucceeded;
+        }
+
        // merge_ready: reviewer approved, owner gets final turn to finalize
        if (pendingReviewTask && pendingReviewTask.status === 'merge_ready') {
          const lastRaw = rawMissedMessages[rawMissedMessages.length - 1];
@@ -756,13 +824,34 @@ export function createMessageRuntime(deps: MessageRuntimeDeps): {
        pendingTaskForChannel &&
        (pendingTaskForChannel.status === 'review_ready' ||
          pendingTaskForChannel.status === 'in_review');
-      const turnChannel = useReviewerChannel ? reviewerChannel : channel;
-      const cursorKey = pairedCursorKey(chatJid, !!useReviewerChannel);
+      const useArbiterChannel =
+        pendingTaskForChannel &&
+        (pendingTaskForChannel.status === 'arbiter_requested' ||
+          pendingTaskForChannel.status === 'in_arbitration');
+      const turnChannel = useArbiterChannel
+        ? arbiterChannel
+        : useReviewerChannel
+          ? reviewerChannel
+          : channel;
+      const cursorKey = useArbiterChannel
+        ? `${chatJid}:arbiter`
+        : pairedCursorKey(chatJid, !!useReviewerChannel);

-      const prompt = formatMessages(
-        labelPairedSenders(chatJid, missedMessages),
-        deps.timezone,
-      );
+      // Arbiter turns use a dedicated context prompt; regular turns use formatted messages.
+      let prompt: string;
+      if (useArbiterChannel && pendingTaskForChannel) {
+        prompt = buildArbiterContextPrompt({
+          chatJid,
+          taskId: pendingTaskForChannel.id,
+          roundTripCount: pendingTaskForChannel.round_trip_count,
+          timezone: deps.timezone,
+        });
+      } else {
+        prompt = formatMessages(
+          labelPairedSenders(chatJid, missedMessages),
+          deps.timezone,
+        );
+      }
      const startSeq = missedMessages[0].seq ?? null;
      const endSeq = missedMessages[missedMessages.length - 1].seq ?? null;
      if (endSeq !== null) {
@@ -953,7 +1042,13 @@ export function createMessageRuntime(deps: MessageRuntimeDeps): {
              !!loopPendingTask &&
              (loopPendingTask.status === 'review_ready' ||
                loopPendingTask.status === 'in_review');
-            const loopCursorKey = pairedCursorKey(chatJid, loopIsReviewerTurn);
+            const loopIsArbiterTurn =
+              !!loopPendingTask &&
+              (loopPendingTask.status === 'arbiter_requested' ||
+                loopPendingTask.status === 'in_arbitration');
+            const loopCursorKey = loopIsArbiterTurn
+              ? `${chatJid}:arbiter`
+              : pairedCursorKey(chatJid, loopIsReviewerTurn);

            const rawPendingMessages = getMessagesSinceSeq(
              chatJid,
--- a/src/paired-execution-context.test.ts
+++ b/src/paired-execution-context.test.ts
@@ -110,6 +110,8 @@ function buildPairedTask(overrides: Partial<PairedTask> = {}): PairedTask {
    review_requested_at: null,
    round_trip_count: 0,
    status: 'active',
+    arbiter_verdict: null,
+    arbiter_requested_at: null,
    created_at: '2026-03-28T00:00:00.000Z',
    updated_at: '2026-03-28T00:00:00.000Z',
    ...overrides,
@@ -198,6 +200,8 @@ describe('paired execution context', () => {
      round_trip_count: 0,
      review_requested_at: '2026-03-28T00:00:00.000Z',
      status: 'review_ready',
+      arbiter_verdict: null,
+      arbiter_requested_at: null,
      created_at: '2026-03-28T00:00:00.000Z',
      updated_at: '2026-03-28T00:00:00.000Z',
    });
@@ -230,6 +234,8 @@ describe('paired execution context', () => {
      round_trip_count: 0,
      review_requested_at: '2026-03-28T00:00:00.000Z',
      status: 'review_ready',
+      arbiter_verdict: null,
+      arbiter_requested_at: null,
      created_at: '2026-03-28T00:00:00.000Z',
      updated_at: '2026-03-28T00:00:00.000Z',
    });
@@ -265,6 +271,8 @@ describe('paired execution context', () => {
      round_trip_count: 0,
      review_requested_at: null,
      status: 'active',
+      arbiter_verdict: null,
+      arbiter_requested_at: null,
      created_at: '2026-03-28T00:00:00.000Z',
      updated_at: '2026-03-28T00:00:00.000Z',
    });
@@ -306,6 +314,8 @@ describe('paired execution context', () => {
      round_trip_count: 0,
      review_requested_at: '2026-03-28T00:00:00.000Z',
      status: 'in_review',
+      arbiter_verdict: null,
+      arbiter_requested_at: null,
      created_at: '2026-03-28T00:00:00.000Z',
      updated_at: '2026-03-28T00:00:00.000Z',
    });
--- a/src/paired-execution-context.ts
+++ b/src/paired-execution-context.ts
@@ -3,7 +3,12 @@ import crypto from 'crypto';
 import fs from 'fs';
 import path from 'path';

-import { DATA_DIR, PAIRED_MAX_ROUND_TRIPS } from './config.js';
+import {
+  ARBITER_DEADLOCK_THRESHOLD,
+  DATA_DIR,
+  PAIRED_MAX_ROUND_TRIPS,
+  isArbiterEnabled,
+} from './config.js';
 import {
  createPairedTask,
  getLatestPairedTaskForChat,
@@ -21,6 +26,7 @@ import {
  provisionOwnerWorkspaceForPairedTask,
 } from './paired-workspace-manager.js';
 import type {
+  PairedRoomRole,
  PairedTask,
  PairedWorkspace,
  RegisteredGroup,
@@ -57,6 +63,24 @@ function classifyReviewerVerdict(
  return 'continue';
 }

+// ---------------------------------------------------------------------------
+// Arbiter verdict detection
+// ---------------------------------------------------------------------------
+
+type ArbiterVerdictResult = 'proceed' | 'revise' | 'reset' | 'escalate' | 'unknown';
+
+function classifyArbiterVerdict(summary: string | null | undefined): ArbiterVerdictResult {
+  if (!summary) return 'unknown';
+  const cleaned = summary.replace(/<internal>[\s\S]*?<\/internal>/g, '').trim();
+  if (!cleaned) return 'unknown';
+  const firstLine = cleaned.split('\n')[0].trim();
+  if (/^\*{0,2}PROCEED\*{0,2}\b/i.test(firstLine)) return 'proceed';
+  if (/^\*{0,2}REVISE\*{0,2}\b/i.test(firstLine)) return 'revise';
+  if (/^\*{0,2}RESET\*{0,2}\b/i.test(firstLine)) return 'reset';
+  if (/^\*{0,2}ESCALATE\*{0,2}\b/i.test(firstLine)) return 'escalate';
+  return 'unknown';
+}
+
 // ---------------------------------------------------------------------------
 // Helpers
 // ---------------------------------------------------------------------------
@@ -156,6 +180,8 @@ function ensureActiveTask(
    review_requested_at: null,
    round_trip_count: 0,
    status: 'active',
+    arbiter_verdict: null,
+    arbiter_requested_at: null,
    created_at: now,
    updated_at: now,
  };
@@ -239,7 +265,7 @@ export function preparePairedExecutionContext(args: {
    // Use a stable per-channel worktree (not per-task) so the Claude SDK
    // session persists across tasks. Different channels still get isolation.
    workspace = provisionOwnerWorkspaceForPairedTask(latestTask.id);
-  } else {
+  } else if (roomRoleContext.role === 'reviewer') {
    const reviewerWorkspace = prepareReviewerWorkspaceForExecution(latestTask);
    workspace = reviewerWorkspace.workspace;
    blockMessage = reviewerWorkspace.blockMessage;
@@ -250,6 +276,18 @@ export function preparePairedExecutionContext(args: {
        updated_at: now,
      });
    }
+  } else if (roomRoleContext.role === 'arbiter') {
+    // Arbiter uses same read-only workspace as reviewer
+    const reviewerWorkspace = prepareReviewerWorkspaceForExecution(latestTask);
+    workspace = reviewerWorkspace.workspace;
+    blockMessage = reviewerWorkspace.blockMessage;
+    const refreshedTask = getPairedTaskById(latestTask.id) ?? latestTask;
+    if (workspace && refreshedTask.status === 'arbiter_requested') {
+      updatePairedTask(latestTask.id, {
+        status: 'in_arbitration',
+        updated_at: now,
+      });
+    }
  }

  const envOverrides: Record<string, string> = {
@@ -272,6 +310,15 @@ export function preparePairedExecutionContext(args: {
    );
    fs.mkdirSync(reviewerSessionDir, { recursive: true });
    envOverrides.CLAUDE_CONFIG_DIR = reviewerSessionDir;
+  } else if (roomRoleContext.role === 'arbiter') {
+    envOverrides.EJCLAW_ARBITER_RUNTIME = '1';
+    const arbiterSessionDir = path.join(
+      DATA_DIR,
+      'sessions',
+      `${group.folder}-arbiter`,
+    );
+    fs.mkdirSync(arbiterSessionDir, { recursive: true });
+    envOverrides.CLAUDE_CONFIG_DIR = arbiterSessionDir;
  }

  return {
@@ -288,7 +335,7 @@ export function preparePairedExecutionContext(args: {

 export function completePairedExecutionContext(args: {
  taskId: string;
-  role: 'owner' | 'reviewer';
+  role: PairedRoomRole;
  status: 'succeeded' | 'failed';
  summary?: string | null;
 }): void {
@@ -487,13 +534,25 @@ export function completePairedExecutionContext(args: {
      case 'continue':
      default:
        // If both sides keep echoing DONE_WITH_CONCERNS without progress,
-        // stop the loop after 3 round trips to prevent infinite ping-pong.
-        if (task.round_trip_count >= 3) {
-          updatePairedTask(taskId, { status: 'completed', updated_at: now });
-          logger.info(
-            { taskId, verdict, roundTrips: task.round_trip_count },
-            'Stopped ping-pong after repeated DONE_WITH_CONCERNS — escalating to user',
-          );
+        // request arbiter intervention (or escalate to user if arbiter not configured).
+        if (task.round_trip_count >= ARBITER_DEADLOCK_THRESHOLD) {
+          if (isArbiterEnabled()) {
+            updatePairedTask(taskId, {
+              status: 'arbiter_requested',
+              arbiter_requested_at: now,
+              updated_at: now,
+            });
+            logger.info(
+              { taskId, verdict, roundTrips: task.round_trip_count },
+              'Deadlock detected — requesting arbiter intervention',
+            );
+          } else {
+            updatePairedTask(taskId, { status: 'completed', updated_at: now });
+            logger.info(
+              { taskId, verdict, roundTrips: task.round_trip_count },
+              'Stopped ping-pong — escalating to user (arbiter not configured)',
+            );
+          }
          break;
        }
        // Owner needs to address feedback — ping-pong continues
@@ -505,4 +564,52 @@ export function completePairedExecutionContext(args: {
        break;
    }
  }
+
+  // Arbiter finished → classify verdict and route accordingly
+  if (role === 'arbiter') {
+    const now = new Date().toISOString();
+    const arbiterVerdict = classifyArbiterVerdict(args.summary);
+
+    logger.info(
+      { taskId, arbiterVerdict, summary: args.summary?.slice(0, 200) },
+      'Arbiter verdict rendered',
+    );
+
+    switch (arbiterVerdict) {
+      case 'proceed':
+      case 'revise':
+      case 'reset':
+        // Non-escalate: reset counter, continue ping-pong
+        updatePairedTask(taskId, {
+          status: 'active',
+          round_trip_count: 0,
+          arbiter_verdict: arbiterVerdict,
+          updated_at: now,
+        });
+        logger.info(
+          { taskId, arbiterVerdict },
+          'Arbiter resolved deadlock — resuming ping-pong',
+        );
+        break;
+      case 'escalate':
+        updatePairedTask(taskId, {
+          status: 'completed',
+          arbiter_verdict: 'escalate',
+          updated_at: now,
+        });
+        logger.info(
+          { taskId },
+          'Arbiter escalated to user — task completed',
+        );
+        break;
+      default:
+        // Unknown verdict — treat as escalate
+        updatePairedTask(taskId, {
+          status: 'completed',
+          arbiter_verdict: 'unknown',
+          updated_at: now,
+        });
+        break;
+    }
+  }
 }
--- a/src/paired-workspace-manager.test.ts
+++ b/src/paired-workspace-manager.test.ts
@@ -78,6 +78,8 @@ describe('paired workspace manager', () => {
      round_trip_count: 0,
      review_requested_at: null,
      status: 'active',
+      arbiter_verdict: null,
+      arbiter_requested_at: null,
      created_at: now,
      updated_at: now,
    });
@@ -167,6 +169,8 @@ describe('paired workspace manager', () => {
      round_trip_count: 0,
      review_requested_at: null,
      status: 'active',
+      arbiter_verdict: null,
+      arbiter_requested_at: null,
      created_at: now,
      updated_at: now,
    });
@@ -243,6 +247,8 @@ describe('paired workspace manager', () => {
      round_trip_count: 0,
      review_requested_at: null,
      status: 'active',
+      arbiter_verdict: null,
+      arbiter_requested_at: null,
      created_at: now,
      updated_at: now,
    });
@@ -318,6 +324,8 @@ describe('paired workspace manager', () => {
      round_trip_count: 0,
      review_requested_at: null,
      status: 'active',
+      arbiter_verdict: null,
+      arbiter_requested_at: null,
      created_at: now,
      updated_at: now,
    });
@@ -459,6 +467,8 @@ describe('paired workspace manager', () => {
      round_trip_count: 0,
      review_requested_at: null,
      status: 'active',
+      arbiter_verdict: null,
+      arbiter_requested_at: null,
      created_at: now,
      updated_at: now,
    });
@@ -521,6 +531,8 @@ describe('paired workspace manager', () => {
      round_trip_count: 0,
      review_requested_at: '2026-03-28T00:01:00.000Z',
      status: 'review_ready',
+      arbiter_verdict: null,
+      arbiter_requested_at: null,
      created_at: now,
      updated_at: '2026-03-28T00:01:00.000Z',
    });
@@ -587,6 +599,8 @@ describe('paired workspace manager', () => {
      round_trip_count: 0,
      review_requested_at: null,
      status: 'active',
+      arbiter_verdict: null,
+      arbiter_requested_at: null,
      created_at: now,
      updated_at: now,
    });
--- a/src/platform-prompts.ts
+++ b/src/platform-prompts.ts
@@ -15,6 +15,8 @@ const PAIRED_ROOM_PROMPT_FILES: Record<AgentType, string> = {
  codex: 'claude-paired-room.md',
 };

+const ARBITER_PROMPT_FILE = 'arbiter-paired-room.md';
+
 export function getPlatformPromptsDir(projectRoot = process.cwd()): string {
  return path.join(projectRoot, 'prompts');
 }
@@ -60,3 +62,22 @@ export function readPairedRoomPrompt(
  const prompt = fs.readFileSync(promptPath, 'utf-8').trim();
  return prompt || undefined;
 }
+
+export function getArbiterPromptPath(
+  projectRoot = process.cwd(),
+): string {
+  return path.join(
+    getPlatformPromptsDir(projectRoot),
+    ARBITER_PROMPT_FILE,
+  );
+}
+
+export function readArbiterPrompt(
+  projectRoot = process.cwd(),
+): string | undefined {
+  const promptPath = getArbiterPromptPath(projectRoot);
+  if (!fs.existsSync(promptPath)) return undefined;
+
+  const prompt = fs.readFileSync(promptPath, 'utf-8').trim();
+  return prompt || undefined;
+}
--- a/src/room-role-context.test.ts
+++ b/src/room-role-context.test.ts
@@ -10,6 +10,7 @@ describe('buildRoomRoleContext', () => {
          chat_jid: 'group@test',
          owner_service_id: 'claude',
          reviewer_service_id: 'codex-main',
+          arbiter_service_id: null,
          activated_at: null,
          reason: null,
          explicit: false,
@@ -22,6 +23,7 @@ describe('buildRoomRoleContext', () => {
      ownerServiceId: 'claude',
      reviewerServiceId: 'codex-main',
      failoverOwner: false,
+      arbiterServiceId: undefined,
    });
  });

@@ -32,6 +34,7 @@ describe('buildRoomRoleContext', () => {
          chat_jid: 'group@test',
          owner_service_id: 'codex-review',
          reviewer_service_id: 'codex-main',
+          arbiter_service_id: null,
          activated_at: '2026-03-28T10:00:00.000Z',
          reason: 'claude-429',
          explicit: true,
@@ -44,6 +47,31 @@ describe('buildRoomRoleContext', () => {
      ownerServiceId: 'codex-review',
      reviewerServiceId: 'codex-main',
      failoverOwner: true,
+      arbiterServiceId: undefined,
+    });
+  });
+
+  it('returns arbiter context when service matches arbiter_service_id', () => {
+    expect(
+      buildRoomRoleContext(
+        {
+          chat_jid: 'group@test',
+          owner_service_id: 'codex-main',
+          reviewer_service_id: 'claude',
+          arbiter_service_id: 'codex-review',
+          activated_at: null,
+          reason: null,
+          explicit: false,
+        },
+        'codex-review',
+      ),
+    ).toEqual({
+      serviceId: 'codex-review',
+      role: 'arbiter',
+      ownerServiceId: 'codex-main',
+      reviewerServiceId: 'claude',
+      failoverOwner: false,
+      arbiterServiceId: 'codex-review',
    });
  });

@@ -54,6 +82,7 @@ describe('buildRoomRoleContext', () => {
          chat_jid: 'solo@test',
          owner_service_id: 'codex-main',
          reviewer_service_id: null,
+          arbiter_service_id: null,
          activated_at: null,
          reason: null,
          explicit: false,
--- a/src/room-role-context.ts
+++ b/src/room-role-context.ts
@@ -20,12 +20,20 @@ export function buildRoomRoleContext(
  }

  const ownerServiceId = normalizeServiceId(lease.owner_service_id);
+  const arbiterServiceId = lease.arbiter_service_id
+    ? normalizeServiceId(lease.arbiter_service_id)
+    : undefined;
+
+  // Check arbiter role first: if this service matches the arbiter_service_id,
+  // it takes the arbiter role (even if it also matches owner or reviewer)
  const role =
-    ownerServiceId === normalizedServiceId
-      ? 'owner'
-      : reviewerServiceId === normalizedServiceId
-        ? 'reviewer'
-        : null;
+    arbiterServiceId && arbiterServiceId === normalizedServiceId
+      ? 'arbiter'
+      : ownerServiceId === normalizedServiceId
+        ? 'owner'
+        : reviewerServiceId === normalizedServiceId
+          ? 'reviewer'
+          : null;

  if (!role) {
    return undefined;
@@ -39,5 +47,6 @@ export function buildRoomRoleContext(
    failoverOwner:
      ownerServiceId === CODEX_REVIEW_SERVICE_ID &&
      reviewerServiceId === CODEX_MAIN_SERVICE_ID,
+    arbiterServiceId,
  };
 }
--- a/src/service-routing.ts
+++ b/src/service-routing.ts
@@ -1,4 +1,5 @@
 import {
+  ARBITER_SERVICE_ID,
  CLAUDE_SERVICE_ID,
  CODEX_MAIN_SERVICE_ID,
  CODEX_REVIEW_SERVICE_ID,
@@ -6,6 +7,7 @@ import {
  REVIEWER_SERVICE_ID_FOR_TYPE,
  SERVICE_AGENT_TYPE,
  SERVICE_ID,
+  isArbiterEnabled,
  normalizeServiceId,
 } from './config.js';
 import {
@@ -20,6 +22,7 @@ export interface EffectiveChannelLease {
  chat_jid: string;
  owner_service_id: string;
  reviewer_service_id: string | null;
+  arbiter_service_id: string | null;
  activated_at: string | null;
  reason: string | null;
  explicit: boolean;
@@ -40,6 +43,9 @@ function normalizeLeaseRow(
    reviewer_service_id: row.reviewer_service_id
      ? normalizeServiceId(row.reviewer_service_id)
      : null,
+    arbiter_service_id: row.arbiter_service_id
+      ? normalizeServiceId(row.arbiter_service_id)
+      : null,
    activated_at: row.activated_at,
    reason: row.reason,
    explicit,
@@ -59,6 +65,7 @@ function getDefaultLease(chatJid: string): EffectiveChannelLease {
      chat_jid: chatJid,
      owner_service_id: ownerServiceId,
      reviewer_service_id: REVIEWER_SERVICE_ID_FOR_TYPE,
+      arbiter_service_id: isArbiterEnabled() ? ARBITER_SERVICE_ID : null,
      activated_at: null,
      reason: null,
      explicit: false,
@@ -70,6 +77,7 @@ function getDefaultLease(chatJid: string): EffectiveChannelLease {
      chat_jid: chatJid,
      owner_service_id: CODEX_MAIN_SERVICE_ID,
      reviewer_service_id: null,
+      arbiter_service_id: null,
      activated_at: null,
      reason: null,
      explicit: false,
@@ -81,6 +89,7 @@ function getDefaultLease(chatJid: string): EffectiveChannelLease {
      chat_jid: chatJid,
      owner_service_id: CLAUDE_SERVICE_ID,
      reviewer_service_id: null,
+      arbiter_service_id: null,
      activated_at: null,
      reason: null,
      explicit: false,
@@ -94,6 +103,7 @@ function getDefaultLease(chatJid: string): EffectiveChannelLease {
        ? CODEX_MAIN_SERVICE_ID
        : CLAUDE_SERVICE_ID,
    reviewer_service_id: null,
+    arbiter_service_id: null,
    activated_at: null,
    reason: null,
    explicit: false,
@@ -143,6 +153,17 @@ export function isReviewerServiceForChat(
  );
 }

+export function isArbiterServiceForChat(
+  chatJid: string,
+  serviceId: string = SERVICE_ID,
+): boolean {
+  const lease = getEffectiveChannelLease(chatJid);
+  return (
+    lease.arbiter_service_id !== null &&
+    normalizeServiceId(serviceId) === lease.arbiter_service_id
+  );
+}
+
 export function shouldServiceProcessChat(
  _chatJid: string,
  _serviceId: string = SERVICE_ID,
@@ -156,6 +177,7 @@ export function activateCodexFailover(chatJid: string, reason: string): void {
    chat_jid: chatJid,
    owner_service_id: CODEX_REVIEW_SERVICE_ID,
    reviewer_service_id: CODEX_MAIN_SERVICE_ID,
+    arbiter_service_id: null,
    activated_at: now,
    reason,
  };
--- a/src/types.ts
+++ b/src/types.ts
@@ -23,14 +23,18 @@ export type VisiblePhase = 'silent' | 'progress' | 'final';

 export type AgentVisibility = 'public' | 'silent';

-export type PairedRoomRole = 'owner' | 'reviewer';
+export type PairedRoomRole = 'owner' | 'reviewer' | 'arbiter';

 export type PairedTaskStatus =
  | 'active'
  | 'review_ready'
  | 'in_review'
  | 'merge_ready'
-  | 'completed';
+  | 'completed'
+  | 'arbiter_requested'
+  | 'in_arbitration';
+
+export type ArbiterVerdict = 'proceed' | 'revise' | 'reset' | 'escalate';

 export type PairedWorkspaceRole = 'owner' | 'reviewer';

@@ -42,6 +46,7 @@ export interface RoomRoleContext {
  ownerServiceId: string;
  reviewerServiceId: string;
  failoverOwner: boolean;
+  arbiterServiceId?: string;
 }

 export interface PairedProject {
@@ -64,6 +69,8 @@ export interface PairedTask {
  review_requested_at: string | null;
  round_trip_count: number;
  status: PairedTaskStatus;
+  arbiter_verdict: string | null;
+  arbiter_requested_at: string | null;
  created_at: string;
  updated_at: string;
 }