zai provider: thinking/reasoning parameter never sent to API — affects all endpoints

Summary

The zai provider plugin declares thinkingFormat: "zai" but no code path consumes it. The thinking/reasoning parameter is never sent to the 智谱 API, regardless of which endpoint is used. Two separate code paths both block supportsReasoningEffort: false.

Affected Versions

OpenClaw 2026.4.23 (a979721)

Root Cause

The blocking happens in provider-model-compat-D5P6lFk0.js line 36:

supportsReasoningEffort: !isZai && knownProviderFamily !== "mistral" && endpointClass !== "xai-native" && !usesExplicitProxyLikeEndpoint,

Path 1: Native endpoint (api.z.ai → endpointClass: "zai-native")

• isZai = endpointClass === "zai-native" → true
• supportsReasoningEffort = false (blocked by !isZai)
• thinkingFormat is set to "zai" (line 39), but no code in the stream builder (provider-stream-rd4D2qfi.js) consumes thinkingFormat === "zai" — only "openrouter" is handled
• Result: no thinking parameter sent

Path 2: Custom baseUrl (open.bigmodel.cn → endpointClass: "custom")

This applies to Coding Plan (https://open.bigmodel.cn/api/coding/paas/v4) and CN endpoint (https://open.bigmodel.cn/api/paas/v4).

• isZai = false (neither "zai-native" nor default route with zai provider)
• usesConfiguredBaseUrl = true (endpointClass "custom" ≠ "default")
• usesKnownNativeOpenAIEndpoint = false (not openai-public/openai-codex/azure-openai)
• usesExplicitProxyLikeEndpoint = true (configured && not known native)
• supportsReasoningEffort = false (blocked by !usesExplicitProxyLikeEndpoint)
• thinkingFormat falls to else branch → "openai" (not "zai"), but even "openai" isn't consumed either
• Result: no thinking parameter sent

Why the stream builder doesn't help

In provider-stream-rd4D2qfi.js, the stream builder only has two thinking paths:

// Path A: OpenRouter
if (compat.thinkingFormat === "openrouter") {
  params.reasoning = { effort: ... };
}

// Path B: Standard OpenAI reasoning_effort
if (compat.supportsReasoningEffort && model.reasoning) {
  params.reasoning_effort = ...;
}

Zai matches neither path — thinkingFormat === "zai" has no handler, and supportsReasoningEffort is always false for zai.

Impact

• All GLM models (glm-5.1, glm-5, glm-5-turbo, glm-4.7, etc.) run without any thinking/reasoning parameter
• The binary thinking profile (off/low) registered by the zai extension (resolveThinkingProfile) only controls the UI toggle, not the actual API request
• Users who set /thinking on see the UI change but no think/thinking/reasoning_effort parameter reaches the API
• This is likely a regression — previously the full 5-level spectrum (off/minimal/low/medium/high) was available, and the thinking parameter was sent

Suggested Fix

1. Add a thinkingFormat === "zai" handler in the stream builder that sends the appropriate 智谱 API thinking parameter (likely thinking or think in the request body)
2. For the endpointClass: "custom" path, consider checking knownProviderFamily === "zai" in addition to endpointClass === "zai-native" so Coding Plan / CN endpoints aren't treated as generic proxies
3. Consider adding open.bigmodel.cn to the zai manifest's providerEndpoints so the endpoint resolves as zai-native instead of custom

Key Files

Environment

• OS: Ubuntu 24.04 (x64), Linux 6.17.0-22-generic
• Node: v24.14.0
• Config: models.providers.zai.baseUrl: "https://open.bigmodel.cn/api/coding/paas/v4", all models api: "openai-completions", all reasoning: true