Fall back to the schema in the prompt when Anthropic can't compile it

src/lib/ai/providers.test.ts

@@ -200,6 +200,26 @@ describe("AnthropicProvider", () => {
     expect(requests[0].body).toMatchObject({ output_config: { format: { type: "json_schema" } } });
   });
 
+  it("puts the schema in the prompt instead when structured outputs can't compile it", async () => {
+    const requests = stubFetch(({ body }) =>
+      (body as { output_config?: unknown }).output_config
+        ? json({ type: "error", error: { type: "invalid_request_error", message: "The compiled grammar is too large, which would cause performance issues." } }, { status: 400 })
+        : anthropicMessage('{"name":"Jane","years":9}'),
+    );
+    const provider = new AnthropicProvider("https://api.anthropic.com", "sk-ant", "claude-sonnet-5", 60_000);
+    expect(await provider.complete({ ...ask, schema: Person })).toEqual({ name: "Jane", years: 9 });
+    expect(requests).toHaveLength(2);
+    expect(requests[1].body).not.toHaveProperty("output_config");
+    expect((requests[1].body as { system: string }).system).toMatch(/^Be brief\.\n\nReply with only a JSON object[\s\S]*"years"/);
+  });
+
+  it("doesn't retry other refused requests", async () => {
+    const requests = stubFetch(() => json({ type: "error", error: { type: "invalid_request_error", message: "messages: field required" } }, { status: 400 }));
+    const provider = new AnthropicProvider("https://api.anthropic.com", "sk-ant", "claude-sonnet-5", 60_000);
+    expect(await failure(provider.complete({ ...ask, schema: Person }))).toMatchObject({ kind: "provider" });
+    expect(requests).toHaveLength(1);
+  });
+
   it("says when the model declines", async () => {
     stubFetch(() => anthropicMessage("", "refusal"));
     const provider = new AnthropicProvider("https://api.anthropic.com", "sk-ant", "claude-sonnet-5", 60_000);

src/lib/ai/providers/anthropic.ts

@@ -4,7 +4,7 @@ import Anthropic from "@anthropic-ai/sdk";
 import { zodOutputFormat } from "@anthropic-ai/sdk/helpers/zod";
 
 import { AiError } from "../errors";
-import { BaseProvider, type ChatAnswer, type ChatCall, type Effort } from "../provider";
+import { BaseProvider, type ChatAnswer, type ChatCall, type Effort, withJsonSchema } from "../provider";
 import { keyGuardedFetch } from "./guarded-fetch";
 
 /** Offered when the model list can't be fetched, e.g. before a key is saved. */
@@ -21,6 +21,9 @@ const THINKING_TOKENS = 12_000;
 // models, such as Claude Haiku 4.5, refuse the parameter.
 const effortLevels = new Map<string, ReadonlySet<Effort>>();
 
+/** The 400 a schema too complex for structured outputs gets. */
+const GRAMMAR_TOO_LARGE = /compiled grammar is too large/i;
+
 /** Claude through the Messages API with the user's own key. */
 export class AnthropicProvider extends BaseProvider {
   readonly kind = "ANTHROPIC" as const;
@@ -44,17 +47,29 @@ export class AnthropicProvider extends BaseProvider {
   protected async chat(call: ChatCall): Promise<ChatAnswer> {
     if (!this.client) throw AiError.config("Add your Anthropic API key in Settings → AI provider first.");
     const effort = call.effort && (await this.effortLevels(this.client, call.model, call.signal)).has(call.effort) ? call.effort : undefined;
-    const outputConfig = { ...(call.json ? { format: zodOutputFormat(call.json.schema) } : {}), ...(effort ? { effort } : {}) };
-    const message = await this.client.messages.create(
-      {
-        model: call.model,
-        max_tokens: call.maxTokens + THINKING_TOKENS,
-        system: call.system,
-        messages: call.messages,
-        ...(Object.keys(outputConfig).length > 0 ? { output_config: outputConfig } : {}),
-      },
-      { signal: call.signal },
-    );
+    const send = (structured: boolean) => {
+      const outputConfig = { ...(structured && call.json ? { format: zodOutputFormat(call.json.schema) } : {}), ...(effort ? { effort } : {}) };
+      return this.client!.messages.create(
+        {
+          model: call.model,
+          max_tokens: call.maxTokens + THINKING_TOKENS,
+          system: !structured && call.json ? withJsonSchema(call.system, call.json.schema) : call.system,
+          messages: call.messages,
+          ...(Object.keys(outputConfig).length > 0 ? { output_config: outputConfig } : {}),
+        },
+        { signal: call.signal },
+      );
+    };
+    let message: Anthropic.Message;
+    try {
+      message = await send(true);
+    } catch (error) {
+      // Structured outputs refuse schemas with many nullable fields (the
+      // profile import has 32) with a 400; the schema then goes in the prompt
+      // instead, and the answer is still validated against it (#24).
+      if (!call.json || !(error instanceof Anthropic.BadRequestError) || !GRAMMAR_TOO_LARGE.test(error.message)) throw error;
+      message = await send(false);
+    }
     if (message.stop_reason === "refusal") {
       throw new AiError("provider", "The model declined to answer this one. Try rewording it, or try again.", "refusal");
     }