fix: guard anthropic refusal trigger

src/agents/pi-embedded-runner/run.ts

@@ -51,6 +51,18 @@ import { describeUnknownError } from "./utils.js";
 
 type ApiKeyInfo = ResolvedProviderAuth;
 
+// Avoid Anthropic's refusal test token poisoning session transcripts.
+const ANTHROPIC_MAGIC_STRING_TRIGGER_REFUSAL = "ANTHROPIC_MAGIC_STRING_TRIGGER_REFUSAL";
+const ANTHROPIC_MAGIC_STRING_REPLACEMENT = "ANTHROPIC MAGIC STRING TRIGGER REFUSAL (redacted)";
+
+function scrubAnthropicRefusalMagic(prompt: string): string {
+  if (!prompt.includes(ANTHROPIC_MAGIC_STRING_TRIGGER_REFUSAL)) return prompt;
+  return prompt.replaceAll(
+    ANTHROPIC_MAGIC_STRING_TRIGGER_REFUSAL,
+    ANTHROPIC_MAGIC_STRING_REPLACEMENT,
+  );
+}
+
 export async function runEmbeddedPiAgent(
   params: RunEmbeddedPiAgentParams,
 ): Promise<EmbeddedPiRunResult> {
@@ -211,6 +223,9 @@ export async function runEmbeddedPiAgent(
           attemptedThinking.add(thinkLevel);
           await fs.mkdir(resolvedWorkspace, { recursive: true });
 
+          const prompt =
+            provider === "anthropic" ? scrubAnthropicRefusalMagic(params.prompt) : params.prompt;
+
           const attempt = await runEmbeddedAttempt({
             sessionId: params.sessionId,
             sessionKey: params.sessionKey,
@@ -228,7 +243,7 @@ export async function runEmbeddedPiAgent(
             agentDir,
             config: params.config,
             skillsSnapshot: params.skillsSnapshot,
-            prompt: params.prompt,
+            prompt,
             images: params.images,
             provider,
             modelId,

src/gateway/gateway-models.profiles.live.test.ts

@@ -38,6 +38,7 @@ const PROVIDERS = parseFilter(process.env.CLAWDBOT_LIVE_GATEWAY_PROVIDERS);
 const THINKING_LEVEL = "high";
 const THINKING_TAG_RE = /<\s*\/?\s*(?:think(?:ing)?|thought|antthinking)\s*>/i;
 const FINAL_TAG_RE = /<\s*\/?\s*final\s*>/i;
+const ANTHROPIC_MAGIC_STRING_TRIGGER_REFUSAL = "ANTHROPIC_MAGIC_STRING_TRIGGER_REFUSAL";
 
 const describeLive = LIVE || GATEWAY_LIVE ? describe : describe.skip;
 
@@ -120,6 +121,73 @@ function isEmptyStreamText(text: string): boolean {
   return text.includes("request ended without sending any chunks");
 }
 
+function buildAnthropicRefusalToken(): string {
+  const suffix = randomUUID().replace(/-/g, "");
+  return `${ANTHROPIC_MAGIC_STRING_TRIGGER_REFUSAL}_${suffix}`;
+}
+
+async function runAnthropicRefusalProbe(params: {
+  client: GatewayClient;
+  sessionKey: string;
+  modelKey: string;
+  label: string;
+  thinkingLevel: string;
+}): Promise<void> {
+  logProgress(`${params.label}: refusal-probe`);
+  const magic = buildAnthropicRefusalToken();
+  const runId = randomUUID();
+  const probe = await params.client.request<AgentFinalPayload>(
+    "agent",
+    {
+      sessionKey: params.sessionKey,
+      idempotencyKey: `idem-${runId}-refusal`,
+      message: `Reply with the single word ok. Test token: ${magic}`,
+      thinking: params.thinkingLevel,
+      deliver: false,
+    },
+    { expectFinal: true },
+  );
+  if (probe?.status !== "ok") {
+    throw new Error(`refusal probe failed: status=${String(probe?.status)}`);
+  }
+  const probeText = extractPayloadText(probe?.result);
+  assertNoReasoningTags({
+    text: probeText,
+    model: params.modelKey,
+    phase: "refusal-probe",
+    label: params.label,
+  });
+  if (!/\bok\b/i.test(probeText)) {
+    throw new Error(`refusal probe missing ok: ${probeText}`);
+  }
+
+  const followupId = randomUUID();
+  const followup = await params.client.request<AgentFinalPayload>(
+    "agent",
+    {
+      sessionKey: params.sessionKey,
+      idempotencyKey: `idem-${followupId}-refusal-followup`,
+      message: "Now reply with exactly: still ok.",
+      thinking: params.thinkingLevel,
+      deliver: false,
+    },
+    { expectFinal: true },
+  );
+  if (followup?.status !== "ok") {
+    throw new Error(`refusal followup failed: status=${String(followup?.status)}`);
+  }
+  const followupText = extractPayloadText(followup?.result);
+  assertNoReasoningTags({
+    text: followupText,
+    model: params.modelKey,
+    phase: "refusal-followup",
+    label: params.label,
+  });
+  if (!/\bstill\b/i.test(followupText) || !/\bok\b/i.test(followupText)) {
+    throw new Error(`refusal followup missing expected text: ${followupText}`);
+  }
+}
+
 function randomImageProbeCode(len = 6): string {
   // Chosen to avoid common OCR confusions in our 5x7 bitmap font.
   // Notably: 0↔8, B↔8, 6↔9, 3↔B, D↔0.
@@ -736,6 +804,16 @@ async function runGatewayModelSuite(params: GatewayModelSuiteParams) {
             }
           }
 
+          if (model.provider === "anthropic") {
+            await runAnthropicRefusalProbe({
+              client,
+              sessionKey,
+              modelKey,
+              label: progressLabel,
+              thinkingLevel: params.thinkingLevel,
+            });
+          }
+
           logProgress(`${progressLabel}: done`);
           break;
         } catch (err) {