Fall back to JSON object mode when model rejects json_schema

[DavertMik]

Problem

Models that don't support strict JSON Schema structured outputs failed every structured-output call (test planning, supervisor verdicts, etc.) with:

This model does not support response format `json_schema`.

The AI SDK's default generateObject strategy always asks the provider for response_format: json_schema; providers like Groq (and many OpenRouter models, especially reasoning models) reject it for unsupported models, so the whole step failed and retried identically.

Fix

Provider.generateObject now degrades gracefully:

1. Detect the capability error (isJsonSchemaUnsupportedError) and remember the model for the session (so later calls skip the failed first attempt).
2. Retry in JSON object mode — AI SDK output: 'no-schema' (which makes the provider send response_format: { type: "json_object" }), the JSON Schema injected into the prompt, and the reply validated against the original Zod schema. This matches the provider's documented "JSON Object Mode".
3. The object-mode attempt treats the deterministic json_schema error as non-retryable, so it falls back immediately instead of burning retries.

Also in this PR

• [Planner] forwards agentName: 'planner' to generateObject. Previously the planner passed no options, so mergeProviderOptions dropped its per-agent config — meaning ai.agents.planner.providerOptions (e.g. reasoningEffort) was silently ignored. This matters because a reasoning model with full reasoning can spend its entire output budget thinking and never emit the plan JSON; the fix lets users disable/tune reasoning for the planner.
• Suppress the AI SDK allowSystemInMessages warning on provider calls (the codebase intentionally carries the system prompt as the first message).

Which models this helps

The fallback is provider-agnostic — it triggers on the error string and works for any provider whose unsupported models still accept json_object. From the live OpenRouter model list (340 models): 264 already support json_schema (never needed this), and 23 support json_object but not json_schema — those now work. Examples: qwen/qwen3-235b-a22b, nousresearch/hermes-4-405b/70b, z-ai/glm-4.5, google/gemma-4-31b-it, ai21/jamba-large-1.7, microsoft/wizardlm-2-8x22b, nvidia/llama-3.3-nemotron-super-49b-v1.5. (53 models support neither and remain unsupported.)

Reasoning models in this set also need their reasoning budgeted (via provider options like Groq reasoningEffort / OpenRouter reasoning) so they leave output room for the JSON — that's config, enabled here by the planner agentName fix.

Verification

• Reproduced against Groq qwen/qwen3.6-27b through the real Provider: without the fix → Failed to validate JSON; with it → 4 scenarios returned.
• bun test tests/integration/ → 62/62 pass
• bun run format / bun run lint:fix clean

🤖 Generated with Claude Code