--output-schema silently ignored on the deepseek provider — prose instead of JSON

[justrach]

Caught by graff-evals (run-20260816-191053, schema-output task, deepseek-v4-flash): the model answered correctly but in markdown prose; the check expects the schema'd JSON. On xai/openai the flag is enforced (two-phase on strict wires). On deepseek it is dropped without a warning.

Expected: either wire it to DeepSeek's response_format json_schema support, or fail fast with 'structured outputs unsupported on provider deepseek' so callers know. Silent degradation is the worst of the three.

Repro: cd graff-evals && ./run.py --harness graff --task schema-output --model deepseek-v4-flash

[justrach]

Fixed in 3f297dd: response_format json_schema rejection now degrades to json_object with the schema moved into the prompt (same learned quirk ladder as cap_new, reset per provider switch/resume). Live-verified on deepseek-v4-flash: graff -p ... --output-schema @schema.json emits {"distinct_skus": 4, "top_sku": "SKU-B"}, and the graff-evals schema-output task now passes (run-20260816-193329).

[justrach]

Follow-up shipped: studied how dsh does structured outputs on their own API (which rejects json_schema) — they never touch response_format at all. Their subagent driver registers a structured_output tool whose parameters ARE the schema, adds a MUST-call prompt instruction, validates arguments client-side (validateJsonSchemaValue), and bounces violations back as tool errors so the model self-repairs; capture concludes the turn, and a run that ends without capturing is an error, never silent prose.

graff's degrade path now does the same (unforced — deepseek thinking mode rejects tool_choice). A/B on deepseek-v4-flash, 6 reps each: json_object mode 6/6 with 167,664 in-tokens (one 96.5k flail rep); tool mode 6/6 with 113,754 (−32%, worst rep 29.9k). Hard nested/enum/array schema: 3/3 valid. json_object mode remains only for agentic turns where real tools occupy the tools array.

[justrach]

Wire-by-wire audit + kimi verification (release/v0.0.263, 9e2da0a):

• kimi native (k3, openai wire): needs NO degrade — Moonshot accepts response_format: json_schema server-side. Verified live on the flat-rate seat: simple schema clean in 3 calls, hard nested/enum schema VALID. $0.
• anthropic wire (anthropic / kimi-anthropic transport / minimax): was a SILENT NO-OP — buildBody never wrote the schema anywhere and never used the schema-aware prompt; prose out, no error to even degrade on. Fixed: the wire now always carries the schema as the structured_output tool (input_schema shape, unforced), same capture/validation path as deepseek. Byte-level tests pin it; live probe blocked by a stale anthropic key in the store.
• responses wire (codex / xai / openai-responses): native text.format json_schema, unaffected.
• remaining blind spot: an openai-compatible server that silently ACCEPTS json_schema and ignores it (some local/proxy stacks) is undetectable from the client — nothing to reject, nothing to learn from.