Bug type
Regression (worked before, now fails)
Summary
Issue Body
Affected Models⚠️ Timeout |⚠️ Timeout |⚠️ Timeout |⚠️ Timeout |
Error Logs
2026-03-14T08:37:51.892+08:00 [agent/embeded] embedded run agent end: runId=7d2b5f3a-14de-42ff-b108-da48205b21e5 isError=true model=glm-5 provider=zai error=LLM request timed out.
2026-03-14T08:40:50.289+08:00 [agent/embeded] embedded run agent end: runId=c2a40dff-91d8-4eff-b1ab-6d87ad641e54 isError=true model=glm-5 provider=zai error=LLM request timed out.
2026-03-14T09:30:46.822+08:00 [agent/embeded] embedded run agent end: runId=d6269439-8945-4a8c-a2e0-ea1fd0209862 isError=true model=glm-5 provider=zai error=LLM request timed out.
2026-03-14T09:32:30.565+08:00 [agent/embedded] embedded run agent end: runId=b10a9f60-9d0d-4d80-aa14-4a086d0cbd65 isError=true model=glm-5 provider=zai error=LLM request timed out.
2026-03-14T09:35:02.842+08:00 [agent/embeded] embedded run agent end: runId=59264d2d-1dfc-4cff-996e-10ce41f25e8b isError=true model=glm-5 provider=zai error=LLM request timed out.
2026-03-14T09:36:52.412+08:00 [agent/embedded] embedded run agent end: runId=2058abfe-8223-4ca3-8fe3-353c88343279 isError=true model=glm-5 provider=zai error=LLM request timed out.
2026-03-14T10:21:26.689+08:00 [agent/embeded] embedded run agent end: runId=3b6f0fdd-51a9-4b06-a916-f536a73863e2 isError=true model=glm-5 provider=zai error=LLM request timed out.
2026-03-14T10:33:39.579+08:00 [agent/embedded] embedded run agent end: runId=b9a96abf-5f66-42ba-8a6f-5468f49a5f46 isError=true model=glm-5 provider=zai error=LLM request timed out.
Pattern Analysis
Key observation: Only models with reasoning: true timeout. Models with reasoning: false work correctly.
Two different API formats affected:
OpenAI-completions (zai/glm-5) → decision=surface_error (no fallback configured)Anthropic-messages (minimax-portal) → decision=fallback_model (fallback configured)
Environment
OpenClaw Version: 2026.3.12 (upgraded from 2026.3.8)OS: macOS 15 (Darwin 25.3.0 arm64)Node: v24.14.0Default Model: minimax-portal/MiniMax-M2.5-highspeed
Root Cause Analysis
From code inspection:
Thinking models emit extra stream events (thinking_start, thinking_delta, thinking_end) that require special handling
2026.3.12 CHANGELOG mentions thinking-related fixes but only for specific providers:
"Ollama/Kimi Cloud: apply the Moonshot Kimi payload compatibility wrapper to Ollama-hosted Kimi models like kimi-k2.5:cloud, so tool routing no longer breaks when thinking is enabled." ([Bug]:fix(ollama): kimi-k2.5:cloud tool routing fails - extend fix to all Kimi variants #41519
zai provider uses createZaiToolStreamWrapper but may not have thinking-specific handling:
if ( provider === "zai" || provider === "z-ai" ) {
if ( merged ?. tool_stream !== false ) {
agent . streamFn = createZaiToolStreamWrapper ( agent . streamFn , true ) ;
}
}
MiniMax portal has no thinking wrapper (unlike Moonshot which has createMoonshotThinkingWrapper)
Neither zai nor minimax-portal have entries in PROVIDER_CAPABILITIES for thinking signature handling
Workaround
{
Suggested Fix#41519 (Moonshot/Kimi thinking fix)#38669 , #39907 , #40552 (Kimi Coding tool_use fix)
Steps to reproduce
Pattern Analysis
关键发现： 只有 reasoning: true 的模型会超时。
Models with reasoning: false work correctly.
reasoning: false 的模型正常工作。
Two different API formats affected:
影响两种不同的 API 格式：
OpenAI-completions (zai/glm-5) → decision=surface_error (no fallback configured)
OpenAI-completions（zai/glm-5）→ decision=surface_error（未配置降级）
Anthropic-messages (minimax-portal) → decision=fallback_model (fallback configured)
Anthropic-messages（minimax-portal）→ decision=fallback_model（已配置降级）
Expected behavior
Suggested Fix
为 zai 提供商添加 thinking 流事件处理（OpenAI 兼容的 reasoning 模型）
Add minimax-portal to PROVIDER_CAPABILITIES with appropriate thinking handling
将 minimax-portal 添加到 PROVIDER_CAPABILITIES，并配置适当的 thinking 处理
Consider a generic thinking wrapper for providers using Anthropic-compatible APIs without native thinking support
考虑为使用 Anthropic 兼容 API 但无原生 thinking 支持的提供商添加通用的 thinking 包装器
Actual behavior
Pattern Analysis
关键发现： 只有 reasoning: true 的模型会超时。
Models with reasoning: false work correctly.
reasoning: false 的模型正常工作。
Two different API formats affected:
影响两种不同的 API 格式：
OpenAI-completions (zai/glm-5) → decision=surface_error (no fallback configured)
OpenAI-completions（zai/glm-5）→ decision=surface_error（未配置降级）
Anthropic-messages (minimax-portal) → decision=fallback_model (fallback configured)
Anthropic-messages（minimax-portal）→ decision=fallback_model（已配置降级）
OpenClaw version
2026.3.12
Operating system
macos 26.3.1
Install method
No response
Model
zai/glm-5, zai/glm-4.7, minimax-portal/MiniMax-M2.5-highspeed, minimax-portal/MiniMax-M2.5-Lightning
Provider / routing chain
openclaw → zai (Zhipu GLM API) and openclaw → minimax-portal (MiniMax Anthropic API)
Config file / key location
No response
Additional provider/model setup details
No response
Logs, screenshots, and evidence
Impact and severity
No response
Additional information
No response
Bug type
Regression (worked before, now fails)
Summary
Issue Body
Description
After upgrading from OpenClaw 2026.3.8 to 2026.3.12, models with reasoning: true frequently timeout, triggering fallback or surfacing errors. This affects multiple providers.
Affected Models⚠️ Timeout |⚠️ Timeout |⚠️ Timeout |⚠️ Timeout |
| Provider | Model ID | reasoning | API Format | Status |
|----------|------------|--------|
| zai | glm-5 | true | openai-completions |
| zai | glm-4.7 | true | openai-completions |
| minimax-portal | MiniMax-M2.5-highspeed | true | anthropic-messages |
| minimax-portal | MiniMax-M2.5-Lightning | true | anthropic-messages |
| minimax-portal | MiniMax-M2.5 | false | anthropic-messages | ✅ Works |
Error Logs
zai/glm-5 (OpenAI-compatible API)
2026-03-14T08:32:19.504+08:00 [agent/embedded] embedded run agent end: runId=d4c930d0-9d3d-47c4-aa36-d22bf6983399 isError=true model=glm-5 provider=zai error=LLM request timed out.
2026-03-14T08:32:19.527+08:00 [agent/embedded] embedded run failover decision: runId=d4c930d0-9d3d-47c4-aa36-d22bf6983399 stage=assistant decision=surface_error reason=timeout provider=zai/glm-5 profile=sha256:118886372c36
2026-03-14T08:37:51.892+08:00 [agent/embeded] embedded run agent end: runId=7d2b5f3a-14de-42ff-b108-da48205b21e5 isError=true model=glm-5 provider=zai error=LLM request timed out.
2026-03-14T08:37:51.915+08:00 [agent/embedded] embedded run failover decision: runId=7d2b5f3a-14de-42ff-b108-da48205b21e5 stage=assistant decision=surface_error reason=timeout provider=zai/glm-5 profile=sha256:118886372c36
2026-03-14T08:40:50.289+08:00 [agent/embeded] embedded run agent end: runId=c2a40dff-91d8-4eff-b1ab-6d87ad641e54 isError=true model=glm-5 provider=zai error=LLM request timed out.
2026-03-14T08:40:50.312+08:00 [agent/embedded] embedded run failover decision: runId=c2a40dff-91d8-4eff-b1ab-6d87ad641e54 stage=assistant decision=surface_error reason=timeout provider=zai/glm-5 profile=sha256:118886372c36
2026-03-14T09:30:46.822+08:00 [agent/embeded] embedded run agent end: runId=d6269439-8945-4a8c-a2e0-ea1fd0209862 isError=true model=glm-5 provider=zai error=LLM request timed out.
2026-03-14T09:30:46.841+08:00 [agent/embedded] embedded run failover decision: runId=d6269439-8945-4a8c-a2e0-ea1fd0209862 stage=assistant decision=surface_error reason=timeout provider=zai/glm-5 profile=sha256:118886372c36
2026-03-14T09:32:30.565+08:00 [agent/embedded] embedded run agent end: runId=b10a9f60-9d0d-4d80-aa14-4a086d0cbd65 isError=true model=glm-5 provider=zai error=LLM request timed out.
2026-03-14T09:32:30.593+08:00 [agent/embedded] embedded run failover decision: runId=b10a9f60-9d0d-4d80-aa14-4a086d0cbd65 stage=assistant decision=surface_error reason=timeout provider=zai/glm-5 profile=sha256:118886372c36
2026-03-14T09:35:02.842+08:00 [agent/embeded] embedded run agent end: runId=59264d2d-1dfc-4cff-996e-10ce41f25e8b isError=true model=glm-5 provider=zai error=LLM request timed out.
2026-03-14T09:35:02.858+08:00 [agent/embedded] embedded run failover decision: runId=59264d2d-1dfc-4cff-996e-10ce41f25e8b stage=assistant decision=surface_error reason=timeout provider=zai/glm-5 profile=sha256:118886372c36
2026-03-14T09:36:52.412+08:00 [agent/embedded] embedded run agent end: runId=2058abfe-8223-4ca3-8fe3-353c88343279 isError=true model=glm-5 provider=zai error=LLM request timed out.
2026-03-14T09:36:52.434+08:00 [agent/embeded] embedded run failover decision: runId=2058abfe-8223-4ca3-8fe3-353c88343279 stage=assistant decision=surface_error reason=timeout provider=zai/glm-5 profile=sha256:118886372c36
2026-03-14T10:21:26.689+08:00 [agent/embeded] embedded run agent end: runId=3b6f0fdd-51a9-4b06-a916-f536a73863e2 isError=true model=glm-5 provider=zai error=LLM request timed out.
2026-03-14T10:21:26.717+08:00 [agent/embedded] embedded run failover decision: runId=3b6f0fdd-51a9-4b06-a916-f536a73863e2 stage=assistant decision=surface_error reason=timeout provider=zai/glm-5 profile=sha256:118886372c36
2026-03-14T10:33:39.579+08:00 [agent/embedded] embedded run agent end: runId=b9a96abf-5f66-42ba-8a6f-5468f49a5f46 isError=true model=glm-5 provider=zai error=LLM request timed out.
2026-03-14T10:33:39.603+08:00 [agent/embeded] embedded run failover decision: runId=b9a96abf-5f66-42ba-8a6f-5468f49a5f46 stage=assistant decision=surface_error reason=timeout provider=zai/glm-5 profile=sha256:118886372c36
minimax-portal/MiniMax-M2.5-highspeed (Anthropic-compatible API)
2026-03-14T10:44:09.986+08:00 [agent/embeded] embedded run failover decision: runId=slug-gen-177345623467 stage=assistant decision=fallback_model reason=timeout provider=minimax-portal/MiniMax-M2.5-highspeed profile=sha256:9e08bd6be9c1
``
Pattern Analysis
Key observation: Only models with
reasoning: truetimeout. Models withreasoning: falsework correctly.Two different API formats affected:
decision=surface_error(no fallback configured)decision=fallback_model(fallback configured)Environment
minimax-portal/MiniMax-M2.5-highspeedRoot Cause Analysis
From code inspection:
Thinking models emit extra stream events (
thinking_start,thinking_delta,thinking_end) that require special handling2026.3.12 CHANGELOG mentions thinking-related fixes but only for specific providers:
zai provider uses
createZaiToolStreamWrapperbut may not have thinking-specific handling:MiniMax portal has no thinking wrapper (unlike Moonshot which has createMoonshotThinkingWrapper)
Neither zai nor minimax-portal have entries in PROVIDER_CAPABILITIES for thinking signature handling
Workaround
Set reasoning: false in model config:
{
"id": "MiniMax-M2.5-highspeed",
"reasoning": false
}
Or use models without reasoning enabled.
Suggested Fix
Add thinking stream event handling for zai provider (OpenAI-compatible reasoning models)
Add minimax-portal to PROVIDER_CAPABILITIES with appropriate thinking handling
Consider a generic thinking wrapper for providers using Anthropic-compatible APIs without native thinking support
Related
#41519 (Moonshot/Kimi thinking fix)
#38669, #39907, #40552 (Kimi Coding tool_use fix)
src/agents/provider-capabilities.ts
src/agents/pi-embedded-runner/stream-wrappers.ts
Steps to reproduce
Pattern Analysis
模式分析
Key observation: Only models with reasoning: true timeout.
关键发现： 只有 reasoning: true 的模型会超时。
Models with reasoning: false work correctly.
reasoning: false 的模型正常工作。
Two different API formats affected:
影响两种不同的 API 格式：
OpenAI-completions (zai/glm-5) → decision=surface_error (no fallback configured)
OpenAI-completions（zai/glm-5）→ decision=surface_error（未配置降级）
Anthropic-messages (minimax-portal) → decision=fallback_model (fallback configured)
Anthropic-messages（minimax-portal）→ decision=fallback_model（已配置降级）
Expected behavior
Suggested Fix
建议修复
Add thinking stream event handling for zai provider (OpenAI-compatible reasoning models)
为 zai 提供商添加 thinking 流事件处理（OpenAI 兼容的 reasoning 模型）
Add minimax-portal to PROVIDER_CAPABILITIES with appropriate thinking handling
将 minimax-portal 添加到 PROVIDER_CAPABILITIES，并配置适当的 thinking 处理
Consider a generic thinking wrapper for providers using Anthropic-compatible APIs without native thinking support
考虑为使用 Anthropic 兼容 API 但无原生 thinking 支持的提供商添加通用的 thinking 包装器
Actual behavior
Pattern Analysis
模式分析
Key observation: Only models with reasoning: true timeout.
关键发现： 只有 reasoning: true 的模型会超时。
Models with reasoning: false work correctly.
reasoning: false 的模型正常工作。
Two different API formats affected:
影响两种不同的 API 格式：
OpenAI-completions (zai/glm-5) → decision=surface_error (no fallback configured)
OpenAI-completions（zai/glm-5）→ decision=surface_error（未配置降级）
Anthropic-messages (minimax-portal) → decision=fallback_model (fallback configured)
Anthropic-messages（minimax-portal）→ decision=fallback_model（已配置降级）
OpenClaw version
2026.3.12
Operating system
macos 26.3.1
Install method
No response
Model
zai/glm-5, zai/glm-4.7, minimax-portal/MiniMax-M2.5-highspeed, minimax-portal/MiniMax-M2.5-Lightning
Provider / routing chain
openclaw → zai (Zhipu GLM API) and openclaw → minimax-portal (MiniMax Anthropic API)
Config file / key location
No response
Additional provider/model setup details
No response
Logs, screenshots, and evidence
Impact and severity
No response
Additional information
No response