fix(prompt-runner): crawl requests get a real deadline — 4 minutes per round-trip, not the SDK's silent 10

A carnegie-foundation team crawl held one non-streaming connection for
the SDK's full default 10-minute timeout before dying with "Request
timed out" — the operator watched a spinner for 10 minutes to learn
nothing. Crawl model calls now pass an explicit per-request budget
(240s, 1 retry — ~480s worst case, inside the 600s dispatch ceiling) so
a stuck request fails fast with a real error while normal crawls
(observed ~211s total) are untouched. Each pause_turn continuation gets
the same per-request budget.

runPrompt grows optional timeoutMs/maxRetries per-request overrides;
enrichment runs (run.ts, drafter.ts) keep SDK defaults deliberately —
uncapped pack crawls have legitimately run 15+ minutes.

Env dial: CRAWL_REQUEST_TIMEOUT_MS.

Files changed:
- services/prompt-runner/src/anthropic.ts
- services/prompt-runner/src/crawl.ts

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RW28dw3kQAKXr2ZNefCukE

Author: mpstaton

services/prompt-runner/src/anthropic.ts

@@ -56,17 +56,33 @@ function extractText(content: Anthropic.ContentBlock[]): string {
  */
 export async function runPrompt(
   request: Anthropic.MessageCreateParamsNonStreaming,
-  options?: { signal?: AbortSignal },
+  options?: {
+    signal?: AbortSignal;
+    // Per-REQUEST deadline + retry override. Without it the SDK defaults
+    // apply (10-minute timeout, 2 retries) — observed live 2026-07-28: a
+    // carnegie-foundation team crawl held one connection the full 10
+    // minutes before "Request timed out". Callers with their own dispatch
+    // ceilings (crawls: 600s) pass a tighter budget so a stuck request
+    // fails fast and localizes. Each pause_turn continuation gets the same
+    // per-request budget — the cap is per round-trip, not per crawl.
+    timeoutMs?: number;
+    maxRetries?: number;
+  },
 ): Promise<string> {
-  const signal = options?.signal;
+  const { signal, timeoutMs, maxRetries } = options ?? {};
+  const reqOpts = {
+    signal,
+    ...(timeoutMs !== undefined ? { timeout: timeoutMs } : {}),
+    ...(maxRetries !== undefined ? { maxRetries } : {}),
+  };
   let messages: Anthropic.MessageParam[] = request.messages;
-  let response = await getClient().messages.create(request, { signal });
+  let response = await getClient().messages.create(request, reqOpts);
 
   let continuations = 0;
   while (response.stop_reason === 'pause_turn' && continuations < MAX_PAUSE_CONTINUATIONS) {
     continuations += 1;
     messages = [...messages, { role: 'assistant', content: response.content }];
-    response = await getClient().messages.create({ ...request, messages }, { signal });
+    response = await getClient().messages.create({ ...request, messages }, reqOpts);
   }
 
   return extractText(response.content);

services/prompt-runner/src/crawl.ts

@@ -25,6 +25,12 @@ const CRAWL_MAX_TOKENS = Number(process.env.CRAWL_MAX_TOKENS ?? 4096);
 // (SearXNG/Exa, no Anthropic tokens). Uncapped, a crawl of a huge publisher
 // searched open-endedly (see request.ts).
 const CRAWL_MAX_WEB_SEARCHES = Number(process.env.CRAWL_MAX_WEB_SEARCHES ?? 5);
+// Per-request deadline for crawl model calls. The SDK default (10 min ×
+// 2 retries) let a stuck request eat the whole 600s dispatch ceiling and
+// then time out anyway (carnegie-foundation team crawl, 2026-07-28).
+// 240s × (1 retry + 1) ≈ 480s worst case — inside the ceiling, so the
+// operator sees a real error instead of a silent 10-minute hang.
+const CRAWL_REQUEST_TIMEOUT_MS = Number(process.env.CRAWL_REQUEST_TIMEOUT_MS ?? 240_000);
 
 export type CrawlTarget = 'links' | 'streams' | 'team';
 
@@ -220,7 +226,10 @@ async function handleCrawl(nc: NatsConnection, msg: CrawlMsg): Promise<void> {
       tools: ['web_search'],
       webSearchMaxUses: CRAWL_MAX_WEB_SEARCHES,
     });
-    const text = await runPrompt(request);
+    const text = await runPrompt(request, {
+      timeoutMs: CRAWL_REQUEST_TIMEOUT_MS,
+      maxRetries: 1,
+    });
     const parsed = extractJson(text);
 
     if (args.target === 'team') {