[berkayturk]

Problem

With the Anthropic provider (claude-sonnet-5-5), Profile → Import from a resume always fails with "The AI provider ran into a problem. Try again in a moment."

The API refuses the extracted_profile schema with a 400:

Schemas contains too many parameters with union types (32 parameters with type arrays or anyOf). This causes exponential compilation cost. Reduce the number of nullable or union-typed parameters (limit: 16 parameters with unions).

Making the schema's text fields non-nullable (32 → 13 unions) isn't enough. The same request is then refused with:

The compiled grammar is too large, which would cause performance issues. Simplify your tool schemas or reduce the number of strict tools.

I reproduced both errors with a direct /v1/messages call that used the task's schema and a three-line resume.

Fix

When a structured request fails with one of these "schema too complex" errors, AnthropicProvider sends it again with:

• the schema in the system prompt, through withJsonSchema, which the OpenAI-compatible and Ollama providers already use;
• no output_config.format.

The provider remembers this per server, model and schema name, so later calls skip the refused request. BaseProvider.complete still validates the answer against the schema and retries once if it doesn't match. Other errors are unchanged.

Tests

• New test in src/lib/ai/providers.test.ts: a 400 "grammar is too large" leads to one retry with the schema in the prompt, and later calls go straight to the prompt.
• Before the fix, this test failed with the same error users see. With the fix, providers.test.ts, structured-output.test.ts and extract-profile.test.ts pass (68 tests).
• tsc --noEmit and eslint pass on the changed files.
• Checked in a running instance: the resume import now works with claude-sonnet-5-5.

🤖 Generated with Claude Code

[philorien]

Hit the same thing today and landed on the same fix independently. Confirming it isn't specific to sonnet-5-5: with the full extracted_profile schema I got the "compiled grammar is too large" 400 on claude-sonnet-5, claude-opus-5-5, claude-haiku-5-5 and claude-haiku-4-5, while small schemas went through fine. Falling back to the schema in the prompt fixed resume import on my instance.

[AienTech]

Thanks, @berkayturk! I took this into Dossier with you as co-author, and it's out in 1.10.0. On the way in, the refusal is matched by the wordings seen and documented ("grammar is too large", "parameters with union types", "too complex"), so other 400s still fail as before. The fallback is remembered per model and schema, so other schemas on the same model keep structured output. Thanks also to @philorien for confirming it across models. Closing in favour of the release.