Bug type
Regression (worked before, now fails)
Summary
Using volces/deepseek-v3-2-251201 model fails with 400 error: unknown field "prompt_cache_key"
Steps to reproduce
- Configure volces provider with deepseek-v3-2-251201 model
- Send a message to the agent
- API request fails with error
Expected behavior
Agent should respond normally without API errors
Actual behavior
400 json: unknown field "prompt_cache_key" Request id: ...
OpenClaw version
2026.3.13
Operating system
Linux (Ubuntu 24.04)
Install method
No response
Model
volces/deepseek-v3-2-251201
Provider / routing chain
volces (Volcano Engine)
Config file / key location
No response
Additional provider/model setup details
No response
Logs, screenshots, and evidence
Impact and severity
No response
Additional information
Root cause: pi-ai library automatically adds prompt_cache_key and prompt_cache_retention fields to API requests. Volcano Engine DeepSeek API does not support these parameters.
Temporary workaround: Set cacheRetention: "none" in provider config.
Suggested fix: Check if base URL supports prompt cache before adding these fields, or add per-model config to disable prompt cache.
Bug type
Regression (worked before, now fails)
Summary
Using volces/deepseek-v3-2-251201 model fails with 400 error: unknown field "prompt_cache_key"
Steps to reproduce
Expected behavior
Agent should respond normally without API errors
Actual behavior
400 json: unknown field "prompt_cache_key" Request id: ...
OpenClaw version
2026.3.13
Operating system
Linux (Ubuntu 24.04)
Install method
No response
Model
volces/deepseek-v3-2-251201
Provider / routing chain
volces (Volcano Engine)
Config file / key location
No response
Additional provider/model setup details
No response
Logs, screenshots, and evidence
Impact and severity
No response
Additional information
Root cause: pi-ai library automatically adds prompt_cache_key and prompt_cache_retention fields to API requests. Volcano Engine DeepSeek API does not support these parameters.
Temporary workaround: Set cacheRetention: "none" in provider config.
Suggested fix: Check if base URL supports prompt cache before adding these fields, or add per-model config to disable prompt cache.