feat(zai): wire reasoning_effort for GLM-5.x thinking models

[kivo360]

Summary

GLM-5.x is a reasoning (thinking) family: every response carries a reasoning_content field alongside content, and Z.AI's API accepts a top-level reasoning_effort parameter that controls thinking depth (minimal disables it entirely; low/medium/high scale it).

The problem: the agent's reasoning_effort config was a no-op for Z.AI. The generic transport gate (_supports_reasoning_extra_body) only forwards reasoning params for OpenRouter / Nous / GitHub / LM Studio, so the effort level never reached the wire — the model kept its server default regardless of config.

The fix: override build_api_kwargs_extras on ZaiProfile to emit reasoning_effort as a top-level API param for GLM-5.x models. This mirrors the established pattern from DeepSeekProfile (commit cd9470f41) and KimiProfile — reasoning controls routed through the provider profile, not the legacy transport fallback.

Behavior

• GLM-5.x + effort set → reasoning_effort sent as top-level param
• xhigh → clamped to high (Hermes-internal level with no Z.AI equivalent; empirically yields fewer reasoning tokens than high)
• enabled: false → emits minimal (turns Z.AI thinking off entirely)
• No effort / no config → omits the param so Z.AI applies its server default
• GLM-4.x (non-reasoning) → untouched, never sends reasoning_effort

Empirical verification

Tested against api.z.ai/api/coding/paas/v4/chat/completions — reasoning token counts by effort:

effort sent	reasoning tokens
(none — current behavior)	237
minimal	0 (reasoning off)
low	184
medium	171
high	235
xhigh	173 (no distinct level — clamped to high)

Test plan

• TestZaiProfile — 7 new tests covering full VALID_REASONING_EFFORTS range, disabled state, no-config, and non-reasoning model guard
• Full tests/providers/ suite passes (118 passed; 1 pre-existing failure in test_bundled_plugins_discovered unrelated to this change — ai-gateway missing __init__.py on main)
• ruff check passes
• End-to-end verified on live Z.AI gateway: agent turns with tools + multi-turn reasoning work cleanly

[teknium1]

Thanks for the provider-profile work. This is now implemented on main by a later, more precise GLM-5.2 implementation. This is an automated hermes-sweeper review.

• a6079dd3502ee94c47481bc068caeea503045f70 (feat(providers): GLM-5.2 native reasoning_effort controls, PR feat(providers): GLM-5.2 native reasoning_effort controls (salvage #51108) #58884) added ZaiProfile reasoning-effort handling.
• plugins/model-providers/zai/__init__.py:49-106 recognizes GLM-5.2 aliases and sends top-level reasoning_effort with the supported high/max mapping.
• agent/transports/chat_completions.py:579-592 invokes the profile and merges its top-level kwargs into the live request.
• tests/plugins/model_providers/test_zai_profile.py:78-163 covers effort mapping, disabled/unset behavior, aliases, and exclusion of non-5.2 models.

The implementation shipped in v2026.7.7.