fix(zai): comprehensive Z.AI/GLM provider support

[nibzard]

Summary

Fixes multiple issues with Z.AI (api.z.ai) GLM models that caused timeouts, empty responses, and missing reasoning support.

Root Causes Identified

1. tool_stream not sent — Z.AI has a server-side 30s idle timeout. When tool_call arguments are generated in a single batch (no tool_stream), the connection goes silent and gets killed, causing ReadTimeout errors.
2. thinking parameter missing — Z.AI uses thinking: {type: "enabled/disabled"} (not OpenRouter's reasoning key) for extended thinking. Without this, GLM-5+ models put all content into reasoning_content with empty content field.
3. No Coding Plan endpoint support — The zai provider only knew about api.z.ai/api/paas/v4 (standard API), not /api/coding/paas/v4 (Coding Plan).

Changes

File	Change
plugins/model-providers/zai/__init__.py	Replace bare ProviderProfile with ZaiProfile subclass. Register 4 provider variants: zai (Global), zai-cn (China), zai-coding-global, zai-coding-cn — each with correct base URLs, env vars, and X-Title headers for Coding Plan endpoints.
agent/transports/chat_completions.py	In both legacy and profile paths: detect Z.AI endpoints (by provider name or base_url), inject tool_stream=true when tools present, inject thinking parameter. This ensures provider: custom configs pointing at api.z.ai also work.
run_agent.py	Add _is_zai detection in _build_api_kwargs, pass is_zai flag to legacy path. Add "z-ai/" to _supports_reasoning_extra_body prefixes. Add new provider names to _anthropic_preserve_dots allowlist.
tests/agent/transports/test_chat_completions.py	15 new tests covering all Z.AI transport behavior (legacy + profile paths, tool_stream, thinking, URL detection).

Provider Variants

Hermes ID	Endpoint	Env Var
zai	api.z.ai/api/paas/v4	ZAI_API_KEY
zai-cn	open.bigmodel.cn/api/paas/v4	GLM_API_KEY
zai-coding-global	api.z.ai/api/coding/paas/v4	ZAI_CODING_API_KEY
zai-coding-cn	open.bigmodel.cn/api/coding/paas/v4	GLM_CODING_API_KEY

Testing

• 122 tests pass (81 transport + 41 provider profiles)
• 15 new Z.AI-specific tests covering:
  • Legacy path: tool_stream with/without tools, thinking enabled/disabled, URL-based detection (z.ai, bigmodel.cn), all 4 provider names
  • Profile path: tool_stream, thinking enabled/disabled, all profile variants
  • Negative: non-Z.AI providers don't get tool_stream

Related PRs

Closes #12758 (tool_stream fix), #14619 (tool_stream in transport), #16592 (thinking injection), #11494 (thinking support)
Relates-to #13911 (4-plan provider split), #16479 (preserved thinking), #18863

Config Migration

Users currently on provider: custom + base_url: https://api.z.ai/api/coding/paas/v4 should switch to:

model:
  provider: zai-coding-global
  api_key: <your-key>

Or set ZAI_CODING_API_KEY env var and use provider: auto.

[DeamonDev888]

👋 Superseded — please see the new PR

@nibzard thanks for the comprehensive 4-variant proposal. #62080 is the new rebased PR that closes PR #61575.

Your 4-variant provider split (zai / zai-cn / zai-coding-global / zai-coding-cn) is a valid approach. #62080 takes a different one: instead of adding 4 provider entries to the registry, it extends ZAI_ENDPOINTS with 6 candidates (added anthropic-global and anthropic-cn to the existing 4) and lets detect_zai_endpoint() pick the right wire format automatically.

Both approaches are valid. If maintainers prefer the 4-variant split approach from this PR (#24915) over the unified probe approach in #62080, that's a maintainer design decision — both achieve the same routing correctness. Either way, end users get the same result.

[DeamonDev888]

Related: our PR #62467 takes a different approach to the same root cause. Instead of splitting Z.AI into 4 provider variants (zai, zai-cn, zai-coding-global, zai-coding-cn), we use a single unified resolver that detects the correct endpoint per-key at runtime. Both approaches are valid; ours avoids adding provider-registry bloat. Would welcome your feedback on #62467.

[teknium1]

Thanks for identifying the Z.AI tool-streaming gap. Current main already implements the thinking portion through ZaiProfile.build_api_kwargs_extras() (plugins/model-providers/zai/__init__.py:88-108) and covers omitted preferences, unsupported GLM models, and GLM-5.2 effort mapping in tests/plugins/model_providers/test_zai_profile.py:38-203.

Problems

• The proposed ZaiProfile.build_extra_body() always enables thinking when no preference is provided and does not gate by model version. That conflicts with current behavior that intentionally omits the field without a preference and avoids unsupported pre-4.5 GLM models (plugins/model-providers/zai/__init__.py:94-101).
• The proposed build_api_kwargs_extras() returns empty values, which would remove the current GLM-5.2 native reasoning_effort mapping (plugins/model-providers/zai/__init__.py:103-106).

Suggested changes

• Salvage only tool_stream into the current profile-path extra-body assembly at agent/transports/chat_completions.py:577-610; that path currently receives tools at :537-541 but has no Z.AI tool-stream insertion.
• Add profile-path and custom-base-URL regression tests while retaining the existing Z.AI reasoning contract.

Automated hermes-sweeper review.

[GottZ]

This was generated by AI during triage.

Summary

Four PRs address overlapping Z.AI failures: #24915 covers tool-call streaming, thinking behavior, and endpoint variants; #46615/#46666 target missing picker models and the Coding Plan URL; #46724 targets regional model discovery and omitted free Flash models. Current main already provides Z.AI thinking handling, resolved-endpoint fetching, and live-plus-curated catalog merging, leaving tool_stream and any independently verified missing model IDs as the substantive gaps.

Related pull requests

• #24915 related — (+324/-9) — salvage and narrow: the diff uniquely addresses the reported 30-second tool-call timeout by adding tool_stream when Z.AI tools are present, including custom Z.AI URLs. Keep the contributor keep_open review in force, but remove the provider split and thinking replacement because they conflict with current ZaiProfile model gating, omitted-preference behavior, and GLM-5.2 reasoning_effort mapping.
• #46615 [closed] related — (+8711/-1221) — superseded: this closed PR remains relevant as the original model-picker proposal, but its Z.AI change is buried among thousands of unrelated stale-branch changes. The contributor explicitly rejected that scope, and the clean continuation is #46666.
• #46666 related — (+47/-2) — mostly obsolete, with limited salvage: its current override accepts and forwards base_url, but current main already performs live-plus-curated merging and resolved-endpoint discovery, while changing the profile base_url does not change the runtime default. Despite the keep_open review on #46666, the diff now shows only candidate missing IDs such as glm-4.5-air/glm-4.6 remain worth consolidating after verification.
• #46724 related — (+27/-1) — salvage verified free-model IDs only: resolved China-endpoint fetching has already landed on main, while this diff's monkey-patched fetch_models wrapper neither accepts nor forwards base_url and would regress authenticated discovery. Despite the keep_open review on #46724, its five verified free Flash IDs can be folded into a proper ZaiProfile.fetch_models implementation; the wrapper and removal of glm-5.2 should not merge.

Duplicates

#46615 and #46666 are successive versions of substantially the same curated-model/base-URL change; #46724 overlaps their catalog and endpoint work but contributes a narrower free-model allowlist. #24915 overlaps endpoint/provider handling but uniquely contains the tool_stream fix.

Suggested consolidation

Merge #24915 only after narrowing it to the contributor-requested tool_stream fix on top of the current ZaiProfile, preserving existing thinking and GLM-5.2 reasoning behavior; optionally fold in only independently verified missing IDs from #46666/#46724 through the existing catalog path. Then #46615 can remain closed as superseded, and #46666 plus #46724 can be closed as consolidated duplicates despite their keep_open reviews because current main already contains their endpoint and primary picker fixes, while their remaining useful model IDs are separable from the obsolete implementations.

Cross-PR triage: Reviewed 4 pull requests and 0 issues in this complex. Each diff was read against this issue; Assessment working set: 631 kB of PR diffs, 8 kB of issue/PR text, 8 kB of discussion (9 comments), 1 verify verdict. verdicts reflect diff content, not PR titles. Part of an automated triage batch.