feat(models): add Z.AI GLM-5.2 to native zai provider + 1M context + thinking effort

[potatogim]

Summary

GLM-5.2 now appears in the Z.AI native model picker and resolves to its correct 1M context length on every surface, with thinking + effort parameter support for reasoning depth control.

Why this needed a manual edit: The Z.AI provider lists are hardcoded in _PROVIDER_MODELS and model_metadata.py — the live /models API returns 7 GLM models but not glm-5.2, so it falls through to the static curated list and must be added by hand.

Why effort matters: GLM-5.2 accepts an effort field (high/max) inside the thinking object. Verified against the live Z.AI API: effort=max produces significantly more reasoning tokens than baseline. Without this PR, Hermes reasoning_effort: xhigh is silently ignored for Z.AI.

Changes

Model registration:

• agent/model_metadata.py: add DEFAULT_CONTEXT_LENGTHS key "glm-5.2": 1_000_000. Longest-key-first substring matching makes it win over the generic "glm": 202752 catch-all.
• hermes_cli/models.py: add to _PROVIDER_MODELS["zai"] curated list (native Z.AI picker only).

Thinking effort implementation:

• plugins/model-providers/zai/__init__.py: ZAIProfile class overrides build_api_kwargs_extras() to emit thinking.type + optional effort. Mirrors the DeepSeek profile pattern (ProviderProfile subclass + override).
  • Effort mapping per Z.AI official docs: xhigh/max → max, high → high, lower efforts omitted (server default = high)
  • Effort is sent for GLM-5.2 only — guarded by _model_supports_effort() with boundary matching (glm-5.2 or glm-5.2-*, not glm-5.20)
  • When reasoning_config=None, returns empty dicts — preserves default wire format for existing GLM users (5.1, 5, 4.x)
  • The thinking parameter itself works across all GLM models on both endpoints (verified); effort is a GLM-5.2 feature.
  • glm-5.2 added to fallback_models for picker display until the live /models API includes it.

Tests:

• tests/plugins/model_providers/test_zai_profile.py — 35 tests: wire shape, model gating, backward compat, transport integration
• tests/agent/test_model_metadata.py — GLM-5.2 context length resolution test

Model facts (from Z.AI docs)

• ID glm-5.2, 1M context window, max output 131,072 tokens, reasoning enabled
• API endpoints: https://api.z.ai/api/paas/v4 (standard) and https://api.z.ai/api/coding/paas/v4 (Coding Plan)
• GLM-5.2 is currently available on the Coding Plan endpoint; standard endpoint support is expected soon
• The thinking and effort parameters are accepted on both endpoints (verified)
• Effort mapping: low/medium/high → high, xhigh/max → max
• https://docs.z.ai/devpack/latest-model

Context length

The context length is set to 1_000_000, matching the Z.AI official docs which explicitly state 1,000,000 tokens. A power-of-two approximation (1,048,576) has been used in some references, but the documented spec value is preferred.

Validation

Input	Resolves to
glm-5.2 / zai/glm-5.2	1,000,000 ctx
glm-5.1 / glm-5	202,752 ctx (unchanged)
glm-4.6 / glm-4.5	202,752 ctx (unchanged)

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}}

Full test suite: 5899 passed, 2 skipped. The 19 pre-existing failures (LSP, coding-context, vision-routing) are unrelated to this change and reproduce on main. GLM-5.2 specific tests: 35 passed (effort profile) + 101 passed (model metadata, includes 1 new GLM-5.2 test).

Related

Also addresses #45519.

• fix(model_metadata): add GLM-5.2 1M context entry to DEFAULT_CONTEXT_LENGTHS #45538 — context length entry only.
• feat(zai): add GLM-5.2 with verified 1M context window #45695 — context length + setup/auth surfaces, with empirical needle-in-haystack verification at 789K tokens.
• feat: add GLM-5.2 to model catalogs #45739 — context length + OpenRouter/Nous catalogs + auth probing.
• fix(models): support glm-5.2, kimi-k2.7-code + merge live/curated mod… #45751 — context length + kimi-k2.7-code + live/curated model merge fix.
• feat(models): add GLM-5.2 with 1M context window #45806 — context length + OpenRouter/curated model lists.

This PR adds thinking/effort support on top, which none of the above cover.

[liuhao1024]

Review: Bundled model addition

The PR title says "add Z.AI GLM-5.2 model" but the diff also adds anthropic/claude-fable-5 to both the curated and free-tier sections of model-catalog.json — that's an unrelated Anthropic model with no description, not mentioned in the title or body.

Concern: If claude-fable-5 is not yet publicly available or its API ID is speculative, adding it to the catalog will show it in the model picker but selecting it will fail at runtime. This is a silent user-facing break.

Suggestion: Either remove the claude-fable-5 entries from this PR and submit them separately with a proper description and pricing info, or update the PR title/body to document both additions.

The GLM-5.2 additions across model_metadata.py, auth.py, models.py, setup.py, and zai/__init__.py look correct — context length, probe models, curated lists, and fallback models are all consistently updated.

[tonydwb]

Code Review Summary

Verdict: Approved

Overview

Adds Z.AI GLM-5.2 model (1M context window) to the providers list, following the standard provider/model registration pattern.

Assessment

• Model registration follows established patterns with appropriate context window (1M).
• One prior informal review from Copilot is noted; this constitutes the formal review.
• No security, performance, or documentation concerns.

---

Reviewed by Hermes Agent

[potatogim]

Thanks @liuhao1024 for catching that! You were right — the claude-fable-5 entries were accidentally reintroduced during a rebase conflict resolution. The upstream had already removed them in #45492, and our theirs pick picked up the stale version.

Fixed in the latest commit — both openrouter and nous catalog entries for claude-fable-5 have been removed. The diff is now GLM-5.2 only. 🙏

[teknium1]

Thanks for jumping on GLM-5.2. We merged the broader GLM-5.2 support via #45695, preserving @am423's authorship in git history:

#45695

That landed the 1M context metadata plus the Z.AI model picker/setup/auth probing/fallback surfaces, so this PR is now covered. Closing as duplicate, with appreciation for helping validate the same issue.