[Feature] Add custom timout for slow LLM self-hosted #312
Description
🎯 The Goal / Use Case
Allow users to configure HTTP client timeout for LLM requests when using slow models or hardware (when self-hosting with Ollama for example). Currently, PicoClaw appears to have a hardcoded ~2 minute timeout that terminates requests before slower LLM backends (self-hosted Ollama with large models) can complete their inference.
Use case: Running PicoClaw with self-hosted Ollama on AMD integrated GPUs (Radeon Renoir) using Vulkan backend, or any scenario where model inference legitimately takes 3-10 minutes per response.
💡 Proposed Solution
Add a configurable environment variable (e.g., PICOCLAW_HTTP_TIMEOUT or PICOCLAW_REQUEST_TIMEOUT) that allows users to override the default HTTP client timeout. This would enable support for:
- Self-hosted LLMs on consumer hardware (AMD/Intel GPUs via Vulkan)
- Large models (70B+) running on limited resources
- Complex multi-turn reasoning tasks that require extended inference time
Suggested default: Keep 2 minutes for API providers, but allow override for self-hosted scenarios.
🛠 Potential Implementation (Optional)
In the HTTP client initialization (likely in Go code):
func newHTTPClient() *http.Client {
timeout := 2 * time.Minute // default
if envTimeout := os.Getenv("PICOCLAW_HTTP_TIMEOUT"); envTimeout != "" {
if d, err := time.ParseDuration(envTimeout); err == nil {
timeout = d
}
}
return &http.Client{
Timeout: timeout,
}
}Then document in config.json or environment variables section. Or something like that I guess.
💬 Additional Context
Evidence from logs:
# PicoClaw logs - timeout at client side
2026/02/16 13:30:35 [ERROR] agent: LLM call failed {error=context deadline exceeded (Client.Timeout exceeded while awaiting headers)}
# Ollama logs - request processing for ~2 minutes before 500 error
[GIN] 2026/02/16 - 13:28:29 | 500 | 1m55s | 172.27.0.1 | POST "/v1/chat/completions"
[GIN] 2026/02/16 - 13:30:35 | 500 | 1m59s | 172.27.0.1 | POST "/v1/chat/completions"
time=2026-02-16T13:30:35.056Z level=INFO source=runner.go:916 msg="aborting completion request due to client closing the connection"
[GIN] 2026/02/16 - 13:31:00 | 500 | 2m0s | 172.27.0.1 | POST "/v1/chat/completions"
As you can see it looks like the error from timout is from PicoClaw.