Files

teknium1 df1074b4e5 fix(aux): structured-output rejection no longer kills fallback candidates or costs a doomed first request

Two open atoms of #83390 (DeepSeek "This response_format type is unavailable now"):

* `_call_fallback_candidate_sync/_async` only special-cased auth errors, so when the primary
  aux provider failed (timeout, rate limit, payment) and the fallback landed on a provider that
  rejects `json_schema`, the 400 re-raised and the whole task died — the primary-path rung from
  #89589 never applied there. Both fallback paths now retry once without `response_format`.
* Every structured aux call (titles, kanban decomposer, goal judge, plugin structured calls)
  paid a guaranteed-fail request on providers that lack `json_schema` before the retry. A
  provider profile can now declare `unsupported_response_formats` (DeepSeek: json_schema, per
  https://api-docs.deepseek.com/guides/json_mode) and the recovery ladder remembers any route
  that rejected a type once (host:port scoped), so `_build_call_kwargs` — shared by the primary
  and fallback paths — omits the field before the first request. Dropping rather than
  downgrading to json_object matches the end state the retry already produced; json_object
  needs a JSON-mentioning prompt and some relays return empty content under it.

New logic lives in agent/auxiliary_structured_output.py; the facade only gains the fallback rung
next to the predicate it uses. tests/agent/conftest.py resets the process-level memo per test.

Fixes #83390, #105191. Closes duplicates #84976, #88830, #102849, #113064.
Co-authored-by: Legion-is-life <Legion-is-life@users.noreply.github.com>

2026-09-17 09:11:38 -07:00

__init__.py

fix: key the routed vision-tool veto on is_routing_aggregator

2026-09-15 04:48:20 -07:00

base.py

fix(aux): structured-output rejection no longer kills fallback candidates or costs a doomed first request

2026-09-17 09:11:38 -07:00

README.md

…

README.md

providers/

Registry and ABC for every inference provider Hermes knows about.

Each provider is declared once as a ProviderProfile. Every other layer — auth resolution, transport kwargs, model listing, runtime routing — reads from these profiles instead of maintaining its own parallel data.

Layout

providers/
├── base.py         ProviderProfile dataclass + OMIT_TEMPERATURE sentinel
├── __init__.py     Registry: register_provider(), get_provider_profile(), list_providers()
└── README.md       This file

The profiles themselves live as plugins under plugins/model-providers/<name>/ (bundled in this repo) and $HERMES_HOME/plugins/model-providers/<name>/ (per-user overrides). The registry in providers/__init__.py lazily discovers them the first time any consumer calls get_provider_profile() or list_providers(). See plugins/model-providers/README.md for the plugin contract and examples.

How it wires in

The registry is populated on first access. After that, every downstream layer reads from it:

hermes_cli/auth.py extends PROVIDER_REGISTRY with every api-key profile it sees (skipping copilot, kimi-coding, kimi-coding-cn, zai, openrouter, custom — those need bespoke token resolution).
hermes_cli/models.py extends CANONICAL_PROVIDERS and calls profile.fetch_models() inside provider_model_ids().
hermes_cli/doctor.py adds a /models health check for each auth_type="api_key" profile.
hermes_cli/config.py injects every env_var into OPTIONAL_ENV_VARS so the setup wizard knows about it.
hermes_cli/runtime_provider.py reads profile.api_mode as a fallback when URL detection finds nothing.
agent/model_metadata.py maps hostname → provider via profile.get_hostname().
agent/auxiliary_client.py reads profile.default_aux_model first before falling back to the legacy hardcoded dict.
agent/transports/chat_completions.py::_build_kwargs_from_profile() invokes profile.prepare_messages(), profile.build_extra_body(), and profile.build_api_kwargs_extras() on every call.
run_agent.py passes provider_profile=<ProviderProfile> so the transport takes the profile path instead of the legacy flag path.

Adding a provider

See plugins/model-providers/README.md — drop a new directory there (or under $HERMES_HOME/plugins/model-providers/ for a private plugin).

Hooks you can override on `ProviderProfile`

Hook	Purpose
`get_hostname()`	URL-based detection — default derives from `base_url`.
`prepare_messages(msgs)`	Provider-specific message preprocessing (Qwen normalises to list-of-parts, injects `cache_control`).
`build_extra_body(**ctx)`	Provider-specific `extra_body` (OpenRouter provider prefs, Gemini `thinking_config`).
`build_api_kwargs_extras(**ctx)`	`(extra_body_additions, top_level_kwargs)` — Kimi puts reasoning_effort top-level, Qwen splits `enable_thinking`/`thinking_budget`.
`supported_reasoning_efforts(model)`	Declared per-model reasoning-effort vocabulary for gateways that 400 on unknown levels (Ramp Router reads its live catalog). `None` = defer to transport defaults, `()` = model takes no reasoning params, tuple = clamp target. Must be cache-only — called on the request hot path.
`fetch_models(*, api_key)`	Live catalog fetch — default hits `{models_url or base_url}/models` with Bearer auth. Override for no-REST providers (Bedrock), OAuth catalogs (Anthropic), or public catalogs (OpenRouter).

Configuration fields

Full reference in providers/base.py dataclass definition.

README.md

providers/

Layout

How it wires in

Adding a provider

Hooks you can override on ProviderProfile

Configuration fields

Hooks you can override on `ProviderProfile`