fix(llm-parse): ask Anthropic models for the extraction tool instead of forcing it

Claude Sonnet 5.5, Opus 5.5 and Fable 5.1 reject tool_choice type "tool" and
"any" with a 400, so every booking import through the Anthropic provider failed
on them. The request now sends tool_choice auto with parallel tool use off,
which every Anthropic model and Messages-compatible gateway accepts, and the
system prompt tells the model to answer through the emit tool exactly once. A
model that still answers in text is reported as before.

Closes #2902

server/src/nest/llm-parse/clients/anthropic.client.ts

@@ -10,10 +10,13 @@ const DEFAULT_MAX_TOKENS = 8192;
 const ANTHROPIC_VERSION = '2023-06-01';
 
 /**
- * Anthropic Messages API client. Structured output via forced tool-use: a single
+ * Anthropic Messages API client. Structured output via tool use: a single
  * `emit_<rootKey>` tool (`emit_reservations` by default) whose `input_schema` is
- * the caller's schema, with `tool_choice` forcing it — the documented, reliable
- * way to get structured JSON.
+ * the caller's schema, which the system prompt asks the model to answer with.
+ * The tool is not forced through `tool_choice`: the current models (Claude
+ * Sonnet 5.5, Opus 5.5, Fable 5.1) refuse `{ type: 'tool' }` and `{ type: 'any' }`
+ * with a 400, while `auto` is accepted by every model and by Messages-compatible
+ * gateways (#2902). A model that answers in text instead is reported as such below.
  * PDFs go as native base64 `document` blocks (Anthropic reads scanned PDFs), photos
  * as `image` blocks.
  * Raw fetch (no SDK) to match the codebase's HTTP style.
@@ -48,7 +51,7 @@ export class AnthropicClient implements LlmExtractionClient {
     const body = {
       model: input.model,
       max_tokens: maxTokens,
-      system: input.prompt,
+      system: `${input.prompt}\n\n${toolInstruction(toolName)}`,
       tools: [
         {
           name: toolName,
@@ -56,7 +59,8 @@ export class AnthropicClient implements LlmExtractionClient {
           input_schema: input.jsonSchema,
         },
       ],
-      tool_choice: { type: 'tool', name: toolName },
+      // At most one call: the whole list goes into it.
+      tool_choice: { type: 'auto', disable_parallel_tool_use: true },
       messages: [{ role: 'user', content }],
     };
 
@@ -97,7 +101,7 @@ export class AnthropicClient implements LlmExtractionClient {
       throw new Error('Anthropic declined to process this document');
     }
     // A run that hits the cap stops mid-tool-call, so what arrives is a fragment
-    // of the list or nothing at all — and a forced tool that was not called left
+    // of the list or nothing at all — and a tool that was not called left
     // no list either. Neither is "this document holds no booking", but both came
     // back as [] and reached the person as an empty preview with nothing in the
     // log: the #2375 symptom, on the provider the dropdown offers first.
@@ -114,3 +118,8 @@ export class AnthropicClient implements LlmExtractionClient {
 }
 
 const USER_TEXT = 'Extract every travel reservation from the following document as schema.org JSON-LD.';
+
+/** What stands in for the forced tool call: the answer goes through the tool, once, and nowhere else. */
+function toolInstruction(toolName: string): string {
+  return `Answer only by calling the ${toolName} tool, exactly once, with everything you extracted. Call it with an empty list when the document holds nothing to extract. Do not answer in text.`;
+}

server/tests/unit/nest/llm-parse/clients.test.ts

@@ -576,7 +576,7 @@ describe('AnthropicClient', () => {
     expect(fetchFn.mock.calls[0][0]).toBe('http://gateway.lan:4000/v1/messages');
     expect(fetchFn.mock.calls[1][0]).toBe('https://api.anthropic.com/v1/messages');
     const body = JSON.parse((fetchFn.mock.calls[0][1] as RequestInit).body as string);
-    expect(body.tool_choice).toMatchObject({ type: 'tool', name: 'emit_reservations' });
+    expect(body.tool_choice).toEqual({ type: 'auto', disable_parallel_tool_use: true });
   });
 
   it('tells a prompt too long for a gateway apart from other refusals', async () => {
@@ -632,7 +632,7 @@ describe('AnthropicClient', () => {
     expect(blockOf(1).type).toBe('document');
   });
 
-  it('forces the emit_reservations tool and reads its input', async () => {
+  it('asks for the emit_reservations tool without forcing it and reads its input (#2902)', async () => {
     const fetchFn = mockFetch(() =>
       jsonResponse({
         stop_reason: 'tool_use',
@@ -644,8 +644,12 @@ describe('AnthropicClient', () => {
     const out = await new AnthropicClient().extract(baseInput);
     expect(out).toEqual([{ '@type': 'LodgingReservation' }]);
     const body = JSON.parse((fetchFn.mock.calls[0][1] as RequestInit).body as string);
-    expect(body.tool_choice).toEqual({ type: 'tool', name: 'emit_reservations' });
+    // Claude Sonnet 5.5, Opus 5.5 and Fable 5.1 refuse a forced tool_choice with a 400;
+    // auto with one call at most is accepted by every model, the prompt names the tool.
+    expect(body.tool_choice).toEqual({ type: 'auto', disable_parallel_tool_use: true });
     expect(body.tools[0].name).toBe('emit_reservations');
+    expect(body.system.startsWith(baseInput.prompt)).toBe(true);
+    expect(body.system).toContain('Answer only by calling the emit_reservations tool, exactly once');
   });
 
   /*
@@ -702,7 +706,7 @@ describe('AnthropicClient', () => {
   });
 
   /*
-   * A forced tool that was never called, and a run that stopped at the cap with a
+   * A tool that was never called, and a run that stopped at the cap with a
    * fragment of one, both came back as [] — so a long voucher reached the person
    * as "no reservations found", with an empty warnings array and nothing in the
    * log. That is the #2375 symptom on the provider the dropdown offers first, so
@@ -734,7 +738,7 @@ describe('AnthropicClient', () => {
     }
   });
 
-  it('throws when the forced tool was never called (#2375)', async () => {
+  it('throws when the tool was never called (#2375)', async () => {
     mockFetch(() =>
       jsonResponse({
         stop_reason: 'end_turn',

wiki/AI-Booking-Import.md

@@ -84,7 +84,7 @@ The **AI Parsing** tab starts with **Reads bookings with**, which shows the conn
 The form for an **Own connection** has these fields:
 
 - **Provider**: **Local (Ollama)**, **OpenAI** or **Anthropic**.
-- **Base URL**: empty means the provider's own address: `http://localhost:11434/v1` for a local Ollama server, `https://api.openai.com/v1` for OpenAI and `https://api.anthropic.com` for Anthropic. Any OpenAI-compatible endpoint goes here. For Anthropic it can name a gateway that speaks the Anthropic Messages API; give it without `/v1`, since TREK adds `/v1/messages`, and the gateway has to accept PDF document blocks and forced tool calls. Picking another provider empties the field, so an address meant for one never receives another provider's key.
+- **Base URL**: empty means the provider's own address: `http://localhost:11434/v1` for a local Ollama server, `https://api.openai.com/v1` for OpenAI and `https://api.anthropic.com` for Anthropic. Any OpenAI-compatible endpoint goes here. For Anthropic it can name a gateway that speaks the Anthropic Messages API; give it without `/v1`, since TREK adds `/v1/messages`, and the gateway has to accept PDF document blocks and tool calls. Picking another provider empties the field, so an address meant for one never receives another provider's key.
 - **API key**: optional for a local server (the field says *Often not needed*), required for the cloud providers. It is stored **encrypted** and shown masked (`••••••••`) once saved; leaving it unchanged keeps the stored key.
 - **Model**: the model id, for example `qwen3.5:4b`, `gpt-4o` or `claude-opus-4-8`.