fix(aux): structured-output 400s no longer kill fallback candidates or cost a doomed first request on DeepSeek (#83390, #105191, salvage #92908)

[teknium1]

Auxiliary tasks that attach response_format (session titles via agent/title_generator.py, plugin structured calls via agent/plugin_llm.py) no longer fail when a fallback provider rejects response_format, and providers known to lack json_schema no longer pay a guaranteed-fail first request.

• _call_fallback_candidate_sync/_async get the primary path's rung: one retry without response_format on a structured-output 400 (before: any non-auth error re-raised and killed the task after the primary provider had already failed — Ollama timeout → DeepSeek fallback → 400, per the Auxiliary title_generation fails on DeepSeek: HTTP 400 "This response_format type is unavailable now" #83390 thread).
• New sibling agent/auxiliary_structured_output.py: a provider profile may declare unsupported_response_formats (DeepSeek: ("json_schema",), its API implements only json_object), and the recovery ladder remembers every (route host:port, model, type) triple a capability rejection produced; _build_call_kwargs — shared by primary and fallback requests, sync and async — omits the field up front. Dropping (not downgrading to json_object) is the same end state the retry already produced; json_object needs a JSON-mentioning prompt and some relays return empty content under it.
• ProviderProfile.unsupported_response_formats documented in the model-provider plugin guide; tests/agent/conftest.py resets the process-level memo per test.

Live repro: live against a stand-in replaying DeepSeek's documented 400 ({"error": {"message": "This response_format type is unavailable now", "type": "invalid_request_error"}}, https://api-docs.deepseek.com/guides/json_mode), driving the real call_llm / async_call_llm / generate_title with temp HERMES_HOMEs.
before: fallback_chain → DeepSeek: BadRequestError … This response_format type is unavailable now re-raised (sync + async), generate_title → None; pinned DeepSeek route: 6 wire requests for 3 calls (3×400 + 3 retries).
after: fallback → fallback candidate fallback_chain[0](custom) rejected the structured-output format field; retrying once without it → title returned (sync + async); pinned route: the field is omitted up front (1 request per call; profile flag on provider: deepseek, memo after one rejection elsewhere). Control: a json_schema-capable stand-in and a real OpenAI gpt-5-mini call still receive response_format.type=json_schema (3/3, 1/1). New tests: 3 red on base by source swap, green on the head; 383 focused tests pass.

Root cause: the structured-output rejection rung existed only on the primary ladder, and nothing consulted provider capability before attaching the field.

Fixes #83390
Fixes #105191
Fixes #84976
Fixes #88830
Fixes #102849
Fixes #113064
Supersedes #92908 (@Legion-is-life, co-authored — same fallback rung, slimmer shape), #113111 (@fangliquanflq), #110633 (@29206394, its by-shape detection already landed on main), #107963 (@dacheah), #88953 (@gerryqi), #85115 (@Johannnnn506); #87870 (@f-trycua) is already covered on main by the Anthropic output_config translation (8f2d61e, f709bd8).

Dropped hunks

• fix(aux): retry fallback candidates once without response_format on structured-output rejection #92908: threading the stripped extra_body into the auth-refresh retry — not needed; a 401 on the same credential right after a 400 is not a real sequence, so the retry after the strip simply raises on any further error.
• fix(auxiliary): avoid failed DeepSeek requests before structured tasks #113111: json_schema → json_object downgrade + supports_structured_output model override — replaced by dropping the field (see thread: json_object empty-content and "prompt must contain json" preconditions).

Review follow-up (241411a):

• MAJOR memo keyed by (host, type) blinded every model on an aggregator host after one model's json_schema 400, and the by-shape rule (response_format + json_schema in the body) also fed the memo from schema-VALIDATION 400s of json_schema-capable providers — fixed: memo is now (endpoint, model, type), every remember_structured_output_rejection caller (both fallback candidate paths, the primary ladder rung) hands over the rejecting error and only capability messages (unavailable / not supported / unknown or unsupported parameter / extra inputs not permitted, never "Invalid schema …") are memoised; the by-shape match still drives the one-shot retry. Reproduced the reviewer's probe on ab4d3a8 (openrouter openai/gpt-5-mini lost the field after another model's rejection; validation 400 memoised) → both False on base, green on the head; test_rejection_memo_is_per_model_and_ignores_schema_validation_errors red on base by source swap.
• COMPOSITION _fallback_structured_output_retry_kwargs + its block in _call_fallback_candidate_sync/_async — fixed on the rebase (3e0481e): fix(aux): title generation recovers from stacked parameter rejections on primary and fallback routes (#78273, #72351, #109774, salvage #72515) #113958 landed first, its rung table now wraps every fallback candidate and carries the remember column, so the fallback-only copy was deleted here instead of shipping two seams for one behaviour.
• MINOR caller list named the kanban decomposer / goal judge — fixed: only agent/title_generator.py and agent/plugin_llm.py attach response_format on aux calls (body first line + test module docstring).
• MINOR profile gate fired only for provider == deepseek — fixed (4 LOC): a custom route whose base_url host is api.deepseek.com resolves to the DeepSeek profile via the existing base_url→provider inference; asserted in test_known_unsupported_route_skips_response_format_before_first_request.

Infographic

[github-actions]

૮ >ﻌ< ა ci review

_{ran on 3e0481e — refactor(aux): drop the fallback-only structured-output rung}

debug info

CI timings

CI timings · View report · View job

Wall time 6m12s vs 5m48s (+6.9%). 7 job(s) slower, 5 faster, 1 unchanged.

• Docs Site / docs-site-checks: -93.0s
• OS-specific tests / macOS-only tests: +20.0s
• Python tests / Run tests: -13.0s
• Python lints / ruff enforcement (blocking): -4.0s
• OS-specific tests / Windows-only tests: +4.0s