diff --git a/docs/v2/ANVIL-4-REVIEW.md b/docs/v2/ANVIL-4-REVIEW.md
new file mode 100644
index 0000000..705e3a4
--- /dev/null
+++ b/docs/v2/ANVIL-4-REVIEW.md
@@ -0,0 +1,304 @@
+# ANVIL-4-REVIEW: cloud-brain readiness
+
+## Summary
+- Recommendation: **GO** / **GO-WITH-PATCHES** / **HOLD**
+
+## Findings (P0 / P1 / P2 with file + line + issue + suggested fix)
+
+### Path correctness
+...
+
+### Dedup key validity
+...
+
+### Transparency contract (ADR-006)
+...
+
+### chat.history under cloud-brain
+...
+
+### Smoke script soundness
+...
+
+## Single-test-proves-transparency answer
+...
+
+## Out-of-scope notes (optional)
+```
+
+Be terse. Skip nitpicks. Focus on what would actually break under cloud-brain.
+
+# Attached: scripts/cloud-brain-smoke.mjs
+
+#!/usr/bin/env node
+// cloud-brain-smoke.mjs — Phase B validation per
+// `~/.openclaw/wiki/main/_boards/runbooks/platform/benchagi-v2-cloud-brain-pickup.md` §"Validation script".
+//
+// Activates AFTER cloud-brain Phase 1B PRs merge (BenchAGI #872 W1, #874 W4,
+// #878 W2, #988 relay, openclaw#24 W3) AND a developer's agentDeployment is
+// flipped to runtime: 'remote-brain' per the operator-side smoke runbook.
+//
+// What it does:
+//
+//   1. Lists all known agents from the local openclaw gateway.
+//   2. For each agent, queries Firestore (admin REST + gcloud token per
+//      ~/.claude/.../memory/reference_firebase_admin_rest_recipe.md) for
+//      `agentDeployments/{instanceId}_{agentId}` and reads the `runtime` field.
+//   3. If `runtime === 'remote-brain'`, spawns `benchagi --agent <name>
+//      --liveness off "respond: smoke-ok"` with stdout/stderr captured and
+//      a 60s wall-clock timeout.
+//   4. Asserts (per runbook §"Validation script" point 4):
+//        - chat output is non-empty (proves cloud-brain dispatched the LLM turn)
+//        - the run terminated cleanly (proves orchestrator returned)
+//        - no error markers in output
+//        - latency < 60s
+//   5. Emits a JSON summary; exits 0 on all-green, 1 if any agent failed.
+//
+// Required env:
+//   - INSTANCE_ID         — Firestore instance id (e.g. cory's primary instance)
+//   - GCP_PROJECT         — Firebase project id (default: benchagi-8ea90)
+//
+// Optional env:
+//   - SMOKE_AGENT_FILTER  — regex; if set, only test agents matching this
+//   - SMOKE_PROMPT        — override default prompt (default: "respond: smoke-ok")
+//   - SMOKE_TIMEOUT_MS    — override 60s default
+//   - DEBUG_RAW_FRAMES    — if "1", tee raw WS frames to ./smoke-frames-<agent>.jsonl
+//                           (requires bench-cli to honor BENCHAGI_DEBUG_TRACE_FILE,
+//                           which is a V1.1 follow-up — for now this is a no-op)
+//
+// This script is GATED on cloud-brain Phase 1B merging. If the schemas don't
+// support `runtime` field yet, every agent will appear as `runtime: undefined`
+// and the script will report "no remote-brain agents found — gated".
+
+import { spawn } from "node:child_process";
+import { execFileSync } from "node:child_process";
+import { performance } from "node:perf_hooks";
+import { readFileSync } from "node:fs";
+import { homedir } from "node:os";
+import { join } from "node:path";
+
+const INSTANCE_ID = process.env.INSTANCE_ID;
+const GCP_PROJECT = process.env.GCP_PROJECT ?? "benchagi-8ea90";
+const PROMPT = process.env.SMOKE_PROMPT ?? "respond: smoke-ok";
+const TIMEOUT_MS = parseInt(process.env.SMOKE_TIMEOUT_MS ?? "60000", 10);
+const FILTER = process.env.SMOKE_AGENT_FILTER ? new RegExp(process.env.SMOKE_AGENT_FILTER) : null;
+
+if (!INSTANCE_ID) {
+  console.error("ERROR: INSTANCE_ID env var required");
+  console.error("Usage: INSTANCE_ID=<your-instance> node scripts/cloud-brain-smoke.mjs");
+  process.exit(2);
+}
+
+// --- Firestore admin REST via gcloud user token ---
+
+function gcloudAccessToken() {
+  try {
+    return execFileSync("gcloud", ["auth", "print-access-token"], {
+      encoding: "utf-8",
+      stdio: ["ignore", "pipe", "pipe"],
+    }).trim();
+  } catch (err) {
+    console.error("ERROR: gcloud auth print-access-token failed:", err.message);
+    console.error("Run `gcloud auth login` first.");
+    process.exit(2);
+  }
+}
+
+async function fetchAgentDeployment(instanceId, agentId, token) {
+  const docPath = `instances/${instanceId}/agentDeployments/${instanceId}_${agentId}`;
+  const url = `https://firestore.googleapis.com/v1/projects/${GCP_PROJECT}/databases/(default)/documents/${docPath}`;
+  const resp = await fetch(url, {
+    headers: {
+      Authorization: `Bearer ${token}`,
+      "X-Goog-User-Project": GCP_PROJECT,
+    },
+  });
+  if (resp.status === 404) return null; // No deployment for this agent
+  if (!resp.ok) {
+    throw new Error(`Firestore GET ${docPath} → ${resp.status} ${await resp.text()}`);
+  }
+  const doc = await resp.json();
+  // Firestore REST returns fields wrapped in type tags. Extract `runtime` (string).
+  const runtime = doc?.fields?.runtime?.stringValue ?? null;
+  const tier = doc?.fields?.tier?.stringValue ?? null;
+  return { runtime, tier, raw: doc };
+}
+
+// --- benchagi spawn with stdout capture + timeout ---
+
+function runBenchagi(agentId, prompt, timeoutMs) {
+  return new Promise((resolve) => {
+    const t0 = performance.now();
+    const child = spawn(
+      "node",
+      ["bin/benchagi.mjs", "--agent", agentId, "--liveness", "off", "--no-thinking", prompt],
+      {
+        cwd: process.cwd(),
+        stdio: ["ignore", "pipe", "pipe"],
+        env: { ...process.env, NO_COLOR: "1" },
+      },
+    );
+
+    let stdout = "";
+    let stderr = "";
+    child.stdout.on("data", (d) => { stdout += d.toString(); });
+    child.stderr.on("data", (d) => { stderr += d.toString(); });
+
+    const timer = setTimeout(() => {
+      try { child.kill("SIGINT"); } catch { /* ignore */ }
+      setTimeout(() => { try { child.kill("SIGKILL"); } catch { /* ignore */ } }, 1000);
+    }, timeoutMs);
+
+    child.on("close", (code) => {
+      clearTimeout(timer);
+      const dtMs = performance.now() - t0;
+      resolve({ exitCode: code, stdout, stderr, durationMs: Math.round(dtMs) });
+    });
+  });
+}
+
+// --- Assertions ---
+
+function assertSmokePassed(result, prompt, timeoutMs) {
+  const issues = [];
+
+  if (result.exitCode !== 0) {
+    issues.push(`exit code ${result.exitCode}`);
+  }
+  if (result.durationMs >= timeoutMs) {
+    issues.push(`timed out after ${timeoutMs}ms`);
+  }
+  if (result.durationMs >= 60_000) {
+    issues.push(`latency ${result.durationMs}ms exceeds 60s budget`);
+  }
+  if (!result.stdout || result.stdout.trim().length === 0) {
+    issues.push("empty stdout — no chat output captured");
+  }
+  // Heuristic: look for a few error indicators in the rendered output.
+  if (/error: |chat\.send failed|connection closed|history replay failed/i.test(result.stdout)) {
+    issues.push("error marker in stdout");
+  }
+  // The prompt asks the agent to "respond: smoke-ok" — don't strictly require
+  // this in output (model might paraphrase), but flag if completely absent.
+  if (!/smoke-ok|ok|hello|hi/i.test(result.stdout)) {
+    issues.push("response doesn't contain expected acknowledgement (lenient check failed)");
+  }
+
+  return issues;
+}
+
+// --- Discover known agents from the local gateway ---
+
+function listAgents() {
+  try {
+    const out = execFileSync("node", ["bin/benchagi.mjs", "agents", "list"], {
+      cwd: process.cwd(),
+      encoding: "utf-8",
+      stdio: ["ignore", "pipe", "pipe"],
+    });
+    // Parse lines like "  kestrel-aurelius  pi/aurelius-default"; split on whitespace.
+    const agents = [];
+    for (const line of out.split("\n")) {
+--
+# ANVIL-4-REVIEW: cloud-brain readiness
+
+## Summary
+- Recommendation: **HOLD**
+- Main blocker: the proposed smoke can pass without exercising cloud-brain. Local `benchagi` still talks to local gateway `chat.send`; W3 adds `/v1/llm_turn` for the relay to call, but I found no local gateway path that routes `chat.send` through cloud-brain based on Firestore `runtime`.
+
+## Findings
+
+### Path correctness
+**P0** — `scripts/cloud-brain-smoke.mjs:97-103` runs `node bin/benchagi.mjs ...`, which sends local gateway `chat.send` via `src/v2/chat-runner.ts:359-364`. The W3 path is a separate HTTP endpoint, `/v1/llm_turn`, registered in `openclaw-w3/src/gateway/server-http.ts:970-984` and handled in `openclaw-w3/src/gateway/llm-turn-http.ts:315-375`.
+
+Issue: selecting agents by Firestore `runtime: remote-brain` does not make local gateway `chat.send` use cloud-brain. Unless one of the unavailable PR diffs adds a local gateway dispatch bridge, this smoke tests local OpenClaw, not cloud-brain.
+
+Suggested fix: either add/verify a gateway `chat.send` remote-brain dispatch path that writes/claims `LlmTurnDirective` and re-emits normal `chat`/`agent` events, or change the smoke to assert cloud artifacts: directive created, claimed, completed, and correlated to the CLI run.
+
+### Dedup key validity
+**OK, conditional.** The CLI key `(prefix, runId, seq, stream, sub)` is valid if forwarded remote-brain events are emitted as normal `agent` frames with stable `runId` and monotonic per-run `seq`. The directive schema carries `runId` per spec lines 611-619, and OpenClaw agent events generate monotonic per-run seq in `openclaw-w3/src/infra/agent-events.ts:200-217`.
+
+Concern is not the key. The gap is proving the gateway actually forwards remote-brain completion as those same event frames.
+
+### Transparency contract (ADR-006)
+**OK in bench-cli code; P0 in integration contract.** `ChatRunner` does not branch on `runtime`; lifecycle handling keys off `chat` final states and `agent.lifecycle` phases only (`src/v2/chat-runner.ts:118-154`). That is the right transparency shape.
+
+But transparency is currently asserted, not proven. The local gateway path must be the component that hides cloud-brain. I did not find that bridge in the readable local code.
+
+### chat.history under cloud-brain
+**P1** — bench-cli calls `chat.history { sessionKey, sinceSeq }` and expects `events`/`frames` for replay (`src/v2/chat-runner.ts:222-245`, `491-505`). The readable OpenClaw `chat.history` handler accepts `sessionKey`, `limit`, `maxChars`, then returns transcript `messages`, not event frames, and does not consume `sinceSeq` (`openclaw-w3/src/gateway/server-methods/chat.ts:1607-1675`).
+
+Suggested fix: gateway needs a real event-frame history contract: `chat.history({ sessionKey, sinceSeq }) -> { events: EventFrame[] }`, including forwarded remote-brain frames. Otherwise reconnect replay is best-effort no-op for both local and cloud-brain runs.
+
+### Smoke script soundness
+**P0** — false-positive risk: non-empty stdout + exit 0 can be produced by local execution. No assertion proves `relayDirectives` had an `llm_turn` for this run, that `/v1/llm_turn` was called, or that a directive completed.
+
+**P1** — Firestore lookup assumes doc id `instances/${instanceId}/agentDeployments/${instanceId}_${agentId}` at `scripts/cloud-brain-smoke.mjs:75-77`. The runbook says this, but the broader spec examples use arbitrary `deploymentId`. If W1 backfill preserves noncanonical deployment ids, the smoke skips real remote-brain agents.
+
+**P2** — Firestore failures are recorded as skips (`scripts/cloud-brain-smoke.mjs:193-200`), so an auth/project outage can produce exit 0 if no agents are tested. For readiness, infra lookup failure should be fatal unless explicitly `--allow-skips`.
+
+**P2** — script assumes repo-root cwd for `bin/benchagi.mjs` (`scripts/cloud-brain-smoke.mjs:100-104`). Use `import.meta.url` to resolve repo root.
+
+## Single-test-proves-transparency answer
+The current smoke is not sufficient. A stronger single test:
+
+1. Start `benchagi --agent <remote-brain-agent> "respond: smoke-ok"`.
+2. Capture raw WS frames from the CLI.
+3. Watch Firestore for a new `relayDirectives` doc with `directiveType: "llm_turn"`, matching `instanceId`, `agentId`, and a correlated `runId`.
+4. Assert directive reaches `completed`.
+5. Assert the CLI saw normal `chat` final and/or `agent.lifecycle end` frames with the same run identity and no runtime-specific branch.
+
+That proves both halves: cloud-brain actually ran, and benchagi consumed it transparently.
+tokens used
+164,060
+# ANVIL-4-REVIEW: cloud-brain readiness
+
+## Summary
+- Recommendation: **HOLD**
+- Main blocker: the proposed smoke can pass without exercising cloud-brain. Local `benchagi` still talks to local gateway `chat.send`; W3 adds `/v1/llm_turn` for the relay to call, but I found no local gateway path that routes `chat.send` through cloud-brain based on Firestore `runtime`.
+
+## Findings
+
+### Path correctness
+**P0** — `scripts/cloud-brain-smoke.mjs:97-103` runs `node bin/benchagi.mjs ...`, which sends local gateway `chat.send` via `src/v2/chat-runner.ts:359-364`. The W3 path is a separate HTTP endpoint, `/v1/llm_turn`, registered in `openclaw-w3/src/gateway/server-http.ts:970-984` and handled in `openclaw-w3/src/gateway/llm-turn-http.ts:315-375`.
+
+Issue: selecting agents by Firestore `runtime: remote-brain` does not make local gateway `chat.send` use cloud-brain. Unless one of the unavailable PR diffs adds a local gateway dispatch bridge, this smoke tests local OpenClaw, not cloud-brain.
+
+Suggested fix: either add/verify a gateway `chat.send` remote-brain dispatch path that writes/claims `LlmTurnDirective` and re-emits normal `chat`/`agent` events, or change the smoke to assert cloud artifacts: directive created, claimed, completed, and correlated to the CLI run.
+
+### Dedup key validity
+**OK, conditional.** The CLI key `(prefix, runId, seq, stream, sub)` is valid if forwarded remote-brain events are emitted as normal `agent` frames with stable `runId` and monotonic per-run `seq`. The directive schema carries `runId` per spec lines 611-619, and OpenClaw agent events generate monotonic per-run seq in `openclaw-w3/src/infra/agent-events.ts:200-217`.
+
+Concern is not the key. The gap is proving the gateway actually forwards remote-brain completion as those same event frames.
+
+### Transparency contract (ADR-006)
+**OK in bench-cli code; P0 in integration contract.** `ChatRunner` does not branch on `runtime`; lifecycle handling keys off `chat` final states and `agent.lifecycle` phases only (`src/v2/chat-runner.ts:118-154`). That is the right transparency shape.
+
+But transparency is currently asserted, not proven. The local gateway path must be the component that hides cloud-brain. I did not find that bridge in the readable local code.
+
+### chat.history under cloud-brain
+**P1** — bench-cli calls `chat.history { sessionKey, sinceSeq }` and expects `events`/`frames` for replay (`src/v2/chat-runner.ts:222-245`, `491-505`). The readable OpenClaw `chat.history` handler accepts `sessionKey`, `limit`, `maxChars`, then returns transcript `messages`, not event frames, and does not consume `sinceSeq` (`openclaw-w3/src/gateway/server-methods/chat.ts:1607-1675`).
+
+Suggested fix: gateway needs a real event-frame history contract: `chat.history({ sessionKey, sinceSeq }) -> { events: EventFrame[] }`, including forwarded remote-brain frames. Otherwise reconnect replay is best-effort no-op for both local and cloud-brain runs.
+
+### Smoke script soundness
+**P0** — false-positive risk: non-empty stdout + exit 0 can be produced by local execution. No assertion proves `relayDirectives` had an `llm_turn` for this run, that `/v1/llm_turn` was called, or that a directive completed.
+
+**P1** — Firestore lookup assumes doc id `instances/${instanceId}/agentDeployments/${instanceId}_${agentId}` at `scripts/cloud-brain-smoke.mjs:75-77`. The runbook says this, but the broader spec examples use arbitrary `deploymentId`. If W1 backfill preserves noncanonical deployment ids, the smoke skips real remote-brain agents.
+
+**P2** — Firestore failures are recorded as skips (`scripts/cloud-brain-smoke.mjs:193-200`), so an auth/project outage can produce exit 0 if no agents are tested. For readiness, infra lookup failure should be fatal unless explicitly `--allow-skips`.
+
+**P2** — script assumes repo-root cwd for `bin/benchagi.mjs` (`scripts/cloud-brain-smoke.mjs:100-104`). Use `import.meta.url` to resolve repo root.
+
+## Single-test-proves-transparency answer
+The current smoke is not sufficient. A stronger single test:
+
+1. Start `benchagi --agent <remote-brain-agent> "respond: smoke-ok"`.
+2. Capture raw WS frames from the CLI.
+3. Watch Firestore for a new `relayDirectives` doc with `directiveType: "llm_turn"`, matching `instanceId`, `agentId`, and a correlated `runId`.
+4. Assert directive reaches `completed`.
+5. Assert the CLI saw normal `chat` final and/or `agent.lifecycle end` frames with the same run identity and no runtime-specific branch.
+
+That proves both halves: cloud-brain actually ran, and benchagi consumed it transparently.
+=== EXIT 0 ===
+DONE: Wed May  6 13:10:35 MDT 2026
diff --git a/scripts/cloud-brain-smoke.mjs b/scripts/cloud-brain-smoke.mjs
new file mode 100755
index 0000000..cff7a38
--- /dev/null
+++ b/scripts/cloud-brain-smoke.mjs
@@ -0,0 +1,251 @@
+#!/usr/bin/env node
+// cloud-brain-smoke.mjs — Phase B validation per
+// `~/.openclaw/wiki/main/_boards/runbooks/platform/benchagi-v2-cloud-brain-pickup.md` §"Validation script".
+//
+// Activates AFTER cloud-brain Phase 1B PRs merge (BenchAGI #872 W1, #874 W4,
+// #878 W2, #988 relay, openclaw#24 W3) AND a developer's agentDeployment is
+// flipped to runtime: 'remote-brain' per the operator-side smoke runbook.
+//
+// What it does:
+//
+//   1. Lists all known agents from the local openclaw gateway.
+//   2. For each agent, queries Firestore (admin REST + gcloud token per
+//      ~/.claude/.../memory/reference_firebase_admin_rest_recipe.md) for
+//      `agentDeployments/{instanceId}_{agentId}` and reads the `runtime` field.
+//   3. If `runtime === 'remote-brain'`, spawns `benchagi --agent <name>
+//      --liveness off "respond: smoke-ok"` with stdout/stderr captured and
+//      a 60s wall-clock timeout.
+//   4. Asserts (per runbook §"Validation script" point 4):
+//        - chat output is non-empty (proves cloud-brain dispatched the LLM turn)
+//        - the run terminated cleanly (proves orchestrator returned)
+//        - no error markers in output
+//        - latency < 60s
+//   5. Emits a JSON summary; exits 0 on all-green, 1 if any agent failed.
+//
+// Required env:
+//   - INSTANCE_ID         — Firestore instance id (e.g. cory's primary instance)
+//   - GCP_PROJECT         — Firebase project id (default: benchagi-8ea90)
+//
+// Optional env:
+//   - SMOKE_AGENT_FILTER  — regex; if set, only test agents matching this
+//   - SMOKE_PROMPT        — override default prompt (default: "respond: smoke-ok")
+//   - SMOKE_TIMEOUT_MS    — override 60s default
+//   - DEBUG_RAW_FRAMES    — if "1", tee raw WS frames to ./smoke-frames-<agent>.jsonl
+//                           (requires bench-cli to honor BENCHAGI_DEBUG_TRACE_FILE,
+//                           which is a V1.1 follow-up — for now this is a no-op)
+//
+// This script is GATED on cloud-brain Phase 1B merging. If the schemas don't
+// support `runtime` field yet, every agent will appear as `runtime: undefined`
+// and the script will report "no remote-brain agents found — gated".
+
+import { spawn } from "node:child_process";
+import { execFileSync } from "node:child_process";
+import { performance } from "node:perf_hooks";
+import { readFileSync } from "node:fs";
+import { homedir } from "node:os";
+import { join } from "node:path";
+
+const INSTANCE_ID = process.env.INSTANCE_ID;
+const GCP_PROJECT = process.env.GCP_PROJECT ?? "benchagi-8ea90";
+const PROMPT = process.env.SMOKE_PROMPT ?? "respond: smoke-ok";
+const TIMEOUT_MS = parseInt(process.env.SMOKE_TIMEOUT_MS ?? "60000", 10);
+const FILTER = process.env.SMOKE_AGENT_FILTER ? new RegExp(process.env.SMOKE_AGENT_FILTER) : null;
+
+if (!INSTANCE_ID) {
+  console.error("ERROR: INSTANCE_ID env var required");
+  console.error("Usage: INSTANCE_ID=<your-instance> node scripts/cloud-brain-smoke.mjs");
+  process.exit(2);
+}
+
+// --- Firestore admin REST via gcloud user token ---
+
+function gcloudAccessToken() {
+  try {
+    return execFileSync("gcloud", ["auth", "print-access-token"], {
+      encoding: "utf-8",
+      stdio: ["ignore", "pipe", "pipe"],
+    }).trim();
+  } catch (err) {
+    console.error("ERROR: gcloud auth print-access-token failed:", err.message);
+    console.error("Run `gcloud auth login` first.");
+    process.exit(2);
+  }
+}
+
+async function fetchAgentDeployment(instanceId, agentId, token) {
+  const docPath = `instances/${instanceId}/agentDeployments/${instanceId}_${agentId}`;
+  const url = `https://firestore.googleapis.com/v1/projects/${GCP_PROJECT}/databases/(default)/documents/${docPath}`;
+  const resp = await fetch(url, {
+    headers: {
+      Authorization: `Bearer ${token}`,
+      "X-Goog-User-Project": GCP_PROJECT,
+    },
+  });
+  if (resp.status === 404) return null; // No deployment for this agent
+  if (!resp.ok) {
+    throw new Error(`Firestore GET ${docPath} → ${resp.status} ${await resp.text()}`);
+  }
+  const doc = await resp.json();
+  // Firestore REST returns fields wrapped in type tags. Extract `runtime` (string).
+  const runtime = doc?.fields?.runtime?.stringValue ?? null;
+  const tier = doc?.fields?.tier?.stringValue ?? null;
+  return { runtime, tier, raw: doc };
+}
+
+// --- benchagi spawn with stdout capture + timeout ---
+
+function runBenchagi(agentId, prompt, timeoutMs) {
+  return new Promise((resolve) => {
+    const t0 = performance.now();
+    const child = spawn(
+      "node",
+      ["bin/benchagi.mjs", "--agent", agentId, "--liveness", "off", "--no-thinking", prompt],
+      {
+        cwd: process.cwd(),
+        stdio: ["ignore", "pipe", "pipe"],
+        env: { ...process.env, NO_COLOR: "1" },
+      },
+    );
+
+    let stdout = "";
+    let stderr = "";
+    child.stdout.on("data", (d) => { stdout += d.toString(); });
+    child.stderr.on("data", (d) => { stderr += d.toString(); });
+
+    const timer = setTimeout(() => {
+      try { child.kill("SIGINT"); } catch { /* ignore */ }
+      setTimeout(() => { try { child.kill("SIGKILL"); } catch { /* ignore */ } }, 1000);
+    }, timeoutMs);
+
+    child.on("close", (code) => {
+      clearTimeout(timer);
+      const dtMs = performance.now() - t0;
+      resolve({ exitCode: code, stdout, stderr, durationMs: Math.round(dtMs) });
+    });
+  });
+}
+
+// --- Assertions ---
+
+function assertSmokePassed(result, prompt, timeoutMs) {
+  const issues = [];
+
+  if (result.exitCode !== 0) {
+    issues.push(`exit code ${result.exitCode}`);
+  }
+  if (result.durationMs >= timeoutMs) {
+    issues.push(`timed out after ${timeoutMs}ms`);
+  }
+  if (result.durationMs >= 60_000) {
+    issues.push(`latency ${result.durationMs}ms exceeds 60s budget`);
+  }
+  if (!result.stdout || result.stdout.trim().length === 0) {
+    issues.push("empty stdout — no chat output captured");
+  }
+  // Heuristic: look for a few error indicators in the rendered output.
+  if (/error: |chat\.send failed|connection closed|history replay failed/i.test(result.stdout)) {
+    issues.push("error marker in stdout");
+  }
+  // The prompt asks the agent to "respond: smoke-ok" — don't strictly require
+  // this in output (model might paraphrase), but flag if completely absent.
+  if (!/smoke-ok|ok|hello|hi/i.test(result.stdout)) {
+    issues.push("response doesn't contain expected acknowledgement (lenient check failed)");
+  }
+
+  return issues;
+}
+
+// --- Discover known agents from the local gateway ---
+
+function listAgents() {
+  try {
+    const out = execFileSync("node", ["bin/benchagi.mjs", "agents", "list"], {
+      cwd: process.cwd(),
+      encoding: "utf-8",
+      stdio: ["ignore", "pipe", "pipe"],
+    });
+    // Parse lines like "  kestrel-aurelius  pi/aurelius-default"; split on whitespace.
+    const agents = [];
+    for (const line of out.split("\n")) {
+      const m = /^\s+(\S+)\s+(\S+)/.exec(line);
+      if (m) agents.push({ id: m[1], model: m[2] });
+    }
+    return agents;
+  } catch (err) {
+    console.error("ERROR: `benchagi agents list` failed:", err.message);
+    process.exit(2);
+  }
+}
+
+// --- Main ---
+
+async function main() {
+  console.log(`[smoke] instance=${INSTANCE_ID} project=${GCP_PROJECT} timeout=${TIMEOUT_MS}ms`);
+
+  const token = gcloudAccessToken();
+  const agents = listAgents();
+  console.log(`[smoke] discovered ${agents.length} agent(s) from local gateway`);
+
+  const results = [];
+  for (const agent of agents) {
+    if (FILTER && !FILTER.test(agent.id)) continue;
+
+    let deployment;
+    try {
+      deployment = await fetchAgentDeployment(INSTANCE_ID, agent.id, token);
+    } catch (err) {
+      console.log(`[smoke] ${agent.id}: SKIP — Firestore lookup failed: ${err.message}`);
+      results.push({ agent: agent.id, skipped: true, reason: "firestore-lookup-failed", error: err.message });
+      continue;
+    }
+
+    if (!deployment) {
+      console.log(`[smoke] ${agent.id}: SKIP — no agentDeployment doc`);
+      results.push({ agent: agent.id, skipped: true, reason: "no-deployment-doc" });
+      continue;
+    }
+
+    if (deployment.runtime !== "remote-brain") {
+      console.log(`[smoke] ${agent.id}: SKIP — runtime=${deployment.runtime ?? "<unset>"}, not remote-brain`);
+      results.push({ agent: agent.id, skipped: true, reason: "not-remote-brain", runtime: deployment.runtime });
+      continue;
+    }
+
+    console.log(`[smoke] ${agent.id}: RUN — runtime=remote-brain, tier=${deployment.tier ?? "<unset>"}`);
+    const run = await runBenchagi(agent.id, PROMPT, TIMEOUT_MS);
+    const issues = assertSmokePassed(run, PROMPT, TIMEOUT_MS);
+
+    if (issues.length === 0) {
+      console.log(`[smoke] ${agent.id}: PASS (${run.durationMs}ms)`);
+      results.push({ agent: agent.id, pass: true, durationMs: run.durationMs });
+    } else {
+      console.log(`[smoke] ${agent.id}: FAIL — ${issues.join(", ")}`);
+      console.log(`[smoke] --- stdout ---\n${run.stdout}\n[smoke] --- stderr ---\n${run.stderr}`);
+      results.push({ agent: agent.id, pass: false, durationMs: run.durationMs, issues, stdout: run.stdout, stderr: run.stderr });
+    }
+  }
+
+  const tested = results.filter((r) => !r.skipped);
+  const failed = tested.filter((r) => !r.pass);
+
+  console.log("\n[smoke] --- summary ---");
+  console.log(JSON.stringify({
+    totalAgents: results.length,
+    tested: tested.length,
+    passed: tested.length - failed.length,
+    failed: failed.length,
+    results,
+  }, null, 2));
+
+  if (tested.length === 0) {
+    console.log("[smoke] no remote-brain agents found — Phase 1B may not be merged + flipped yet");
+    process.exit(0);
+  }
+
+  process.exit(failed.length === 0 ? 0 : 1);
+}
+
+main().catch((err) => {
+  console.error("[smoke] fatal:", err);
+  process.exit(2);
+});