feat(providers): declarative reasoning_effort_max for deep-reasoning providers

[leriou]

What

Some providers (Z.AI/GLM-5.2, etc.) expose a deeper-than-default reasoning tier via a single top-level reasoning_effort parameter (e.g. "max"). Hermes' /reasoning xhigh command had no effect on these providers — they don't override build_api_kwargs_extras, so the base method returned ({}, {}) unconditionally.

How

Add a declarative reasoning_effort_max: str | None field to ProviderProfile. When set, the base build_api_kwargs_extras maps effort=xhigh → reasoning_effort=<value> as a top-level kwarg. No subclass needed.

# providers/base.py
reasoning_effort_max: str | None = None

# In build_api_kwargs_extras:
if self.reasoning_effort_max and effort == "xhigh":
    return {}, {"reasoning_effort": self.reasoning_effort_max}
return {}, {}

# plugins/model-providers/zai/__init__.py
zai = ProviderProfile(
    ...,
    reasoning_effort_max="max",
)

Why not override

The existing pattern is Nous/OpenRouter/MiniMax each overriding build_api_kwargs_extras with full custom logic. But providers like Z.AI/GLM only need a single parameter — a full override for one line of actual logic is noise. The declarative field follows the same philosophy as default_aux_model and default_max_tokens: simple things stay declarative.

Impact

• Providers that don't set reasoning_effort_max → zero change (still ({}, {}))
• Providers that override build_api_kwargs_extras entirely → unaffected (their override wins)
• Only providers that set the field AND don't override → new behavior (intended)

[teknium1]

Thanks for the declarative provider-profile proposal. This is now implemented on main through a more provider-specific path.

• Automated hermes-sweeper review verified that ZaiProfile.build_api_kwargs_extras() maps GLM-5.2 xhigh/max/ultra to top-level reasoning_effort="max" in plugins/model-providers/zai/__init__.py:79-106.
• The transport applies profile top-level kwargs to the outgoing request in agent/transports/chat_completions.py:580-592.
• Coverage verifies both the mapping and the final transport kwargs in tests/plugins/model_providers/test_zai_profile.py:100-107 and :237-250.
• This shipped in commit a6079dd3502ee94c47481bc068caeea503045f70 (feat(providers): GLM-5.2 native reasoning_effort controls, feat(providers): GLM-5.2 native reasoning_effort controls (salvage #51108) #58884), included in v2026.7.7.

The linked #51482 concerns a related Ollama Cloud reasoning-effort path; the direct Z.AI behavior requested here is already covered on main.