Bug type

Regression (worked before, now fails)

Summary

OpenClaw v2026.3.2 sends parallel_tool_calls: true in API requests to all OpenAI-compatible providers, including models that don't support parallel tool calling (e.g. moonshotai/kimi-k2.5 on NVIDIA NIM). This causes a 400 error from the provider and breaks ALL tool execution for the agent.

Steps to reproduce

1. Configure a custom provider using NVIDIA NIM with moonshotai/kimi-k2.5 (or any model that doesn't support parallel tool calls)
2. 2. Send any message that triggers tool use
3. 3. Agent attempts parallel tool calls
4. 4. Provider returns: 400 This model only supports single tool-calls at once!
5. 5. All tools become unavailable - exec, read, write, browser, etc.
6. 6. Agent reports "Tool not found" for any subsequent tool call attempts

Expected behavior

OpenClaw should either:

1. Expose a parallelToolCalls: false config option per model/provider (currently returns Unrecognized key).
2. 2. Auto-detect the 400 response and retry with single tool calls.
3. 3. Not force parallel_tool_calls: true for custom openai-completions providers.

Actual behavior

Provider returns 400 This model only supports single tool-calls at once! and all tool execution breaks.

Error Log:
embedded run agent end: runId=0244ece4 isError=true error=400 This model only supports single tool-calls at once!
lane task error: lane=session:agent:main:main durationMs=600123 error="FailoverError: LLM request timed out."

OpenClaw version

2026.3.2 (85377a2)

Operating system

Linux (Ubuntu)

Install method

npm global

Logs, screenshots, and evidence

Attempted Workarounds (none work):
- "parallelToolCalls": false in model config -> Unrecognized key
- "requestParams": {"parallel_tool_calls": false} in provider -> Unrecognized key

Working Workaround:
Downgrade to 2026.2.25

Environment details:
- Node: 22.22.0
- Provider: NVIDIA NIM (openai-completions API)
- Model: moonshotai/kimi-k2.5

Impact and severity

Affected: All users using providers that don't support parallel tool calls (like Kimi on NVIDIA NIM).
Severity: High (breaks ALL tool execution, making the agent useless).
Frequency: 100% repro on affected models.
Consequence: Agent becomes unable to use any tools after the first error.

Additional information

Last known good version: 2026.2.25
First known bad version: 2026.3.2