feat(zai): support thinking effort for GLM-5.2

[potatogim]

What does this PR do?

GLM-5.2 accepts an effort field (high/max) inside the thinking object, per the Z.AI official docs. Without this PR, Hermes reasoning_effort is silently ignored for Z.AI — the configured value never reaches the API.

This adds a ZAIProfile subclass that overrides build_api_kwargs_extras() to emit thinking.type + optional effort, mirroring the DeepSeek profile pattern.

Related Issue

Supersedes the effort portion of #45483 (closed as duplicate after #45695 merged the model registration). GLM-5.2 model registration and context length are already on main via #45695 — this PR adds only the missing effort support.

Type of Change

• ✨ New feature (non-breaking change that adds functionality)

Changes Made

plugins/model-providers/zai/__init__.py:

• New ZAIProfile(ProviderProfile) class with build_api_kwargs_extras() override
• _model_supports_effort() guard — GLM-5.2 only, with boundary matching (glm-5.2 or glm-5.2-*, not glm-5.20)
• Effort mapping per Z.AI docs:
  • xhigh/max → thinking.effort=max
  • high → thinking.effort=high
  • lower efforts (none/minimal/low/medium) → omit (server default = high)
• When reasoning_config=None, returns empty dicts — preserves default wire format for existing GLM users (5.1, 5, 4.x)

tests/plugins/model_providers/test_zai_profile.py:

• 35 tests across 4 suites: wire shape, model gating, backward compat, transport integration

How to Test

1. Configure Z.AI with GLM-5.2: hermes config set model.default glm-5.2 && hermes config set agent.reasoning_effort xhigh
2. Send a message and verify the API request includes "thinking": {"type": "enabled", "effort": "max"} in extra_body
3. Run tests: pytest tests/plugins/model_providers/test_zai_profile.py -v — 35 passed

Checklist

Code

• I've read the Contributing Guide
• My commit messages follow Conventional Commits
• I searched for existing PRs to make sure this isn't a duplicate
• My PR contains only changes related to this fix/feature (no unrelated commits)
• I've run pytest tests/ -q and all tests pass
• I've added tests for my changes
• I've tested on my platform: Linux 6.18.20-1-lts

Documentation & Housekeeping

• I've updated relevant documentation (docstrings) — or N/A
• N/A — no config keys changed
• N/A — follows existing DeepSeek profile pattern, no architecture change
• N/A — cross-platform safe (provider profile only)

Validation

Effort wire shape (build_api_kwargs_extras):

Config	extra_body
None (no reasoning)	{} (preserved default)
{enabled: True, effort: xhigh}	{thinking: {type: enabled, effort: max}}
{enabled: True, effort: high}	{thinking: {type: enabled, effort: high}}
{enabled: True, effort: medium}	{thinking: {type: enabled}}
{enabled: False}	{thinking: {type: disabled}}

Effort was verified against the live Z.AI API: effort=max produces ~55% more reasoning tokens than baseline.

Screenshots / Logs

N/A

[tonydwb]

Code Review Summary

Verdict: Approved

zai provider now supports thinking effort for GLM-5.2 model. Small, focused change (2 files).

Looks Good

• Well-scoped, clean implementation

---

Reviewed by Hermes Agent

[teknium1]

Thanks for the focused Z.AI investigation. This is now implemented on current main; this is an automated hermes-sweeper review.

• a6079dd3502ee94c47481bc068caeea503045f70 (feat(providers): GLM-5.2 native reasoning_effort controls (#58884)) added GLM-5.2 effort forwarding and is included in v2026.7.7.
• plugins/model-providers/zai/__init__.py:49-106 recognizes GLM-5.2 aliases and maps configured effort to the native top-level reasoning_effort value.
• tests/plugins/model_providers/test_zai_profile.py:81-107 and :237-250 cover high/max mapping and the outgoing transport kwargs.
• The merged implementation also covers OpenCode Go aliases, beyond this PR's direct-Z.AI scope.