API Error: fetch failed

[Sebas0317]

Summary

/init command fails with API Error: fetch failed when using Ollama with qwen2.5-coder:7b, even though the model is correctly configured and working in basic interactions.

---

Steps to Reproduce

1. Start Ollama:

   ollama serve
   

2. Ensure model is installed:

   ollama list
   

3. Launch OpenClaude with:

   $env:CLAUDE_CODE_USE_OPENAI="1"
   $env:OPENAI_BASE_URL="http://localhost:11434/v1"
   $env:OPENAI_MODEL="qwen2.5-coder:7b"
   openclaude
   

4. Run:

   /init
   

---

Expected Behavior

The /init workflow should initialize and execute normally using the configured Ollama model.

---

Actual Behavior

The command fails with the following error:

API Error: fetch failed

---

Environment

• OpenClaude version: v0.1.8
• OS: Windows
• Terminal: PowerShell
• Provider: Ollama (OpenAI-compatible)
• Model: qwen2.5-coder:7b

---

Logs / Screenshots

Provider  Ollama
Model     qwen2.5-coder:7b
Endpoint  http://localhost:11434/v1

● local Ready — type /help to begin

❯ /career-ops
⎿ API Error: fetch failed

---

Additional Context

• The model works correctly when tested directly with:

  ollama run qwen2.5-coder:7b
  

• OpenClaude shows the provider as "Ready"

• Simpler interactions seem to work, but agent-based workflows fail

• Other models tested:

  • qwen:7b → does not support tools
  • gemma4 variants → not compatible with agent workflows

This may be related to tool calling support, API compatibility, or timeout issues when running multi-step agent flows with Ollama.

[kevincodex1]

can you try with latest main branch to see if this is still occuring in development?

[auriti]

I dug through the current code paths and I don’t think fetch failed is enough yet to point to a single root cause here.

A few observations:

• your report says /init, but the screenshot shows /career-ops; that matters because skill / agent workflows can exercise a different path than a simple prompt
• simpler prompts working while agent/skill workflows fail suggests this is probably not a basic Ollama connectivity problem
• it is more likely one of:
  1. a local-provider/tool-calling compatibility issue during multi-step turns
  2. the Ollama endpoint closing/resetting the request on a more complex payload
  3. a Windows-specific transport/runtime issue that only shows up on the heavier workflow path

I also reviewed current open work around this area:

• fix: surface actionable error when fetch fails in _doOpenAIRequest #447 is trying to improve the raw fetch failed error so the client surfaces something more actionable
• Tool calling not executed with Ollama models (outputs raw tool JSON instead of performing actions) #433 tracks a nearby Ollama tool-calling cluster where models return raw tool JSON / don’t execute tools properly

Most useful next step before anyone can safely fix this:

1. retry on the latest main (or a build from latest main)
2. reproduce with the exact same model and command
3. capture the debug output / raw failing request context

If you can, please share:

• whether the failing command is actually /init or /career-ops
• the exact latest-main output
• a debug log captured with a current build (for example via --debug-file)

Without that, this still looks more like a symptom bucket than a single confirmed bug.

[Vasanthdev2004]

Thanks for the extra detail here. Right now API Error: fetch failed is still too broad for us to confidently pin to one root cause, especially since agent/skill workflows can exercise different paths than a simple prompt.

Could you retest on the latest main and, if it still happens, provide:

• the exact command you ran (/init vs another slash command)
• the smallest prompt/workflow that reproduces it
• full startup env/config for the Ollama/OpenAI-compatible path
• whether the same failure happens with a different tool-capable Ollama model

That should help us separate provider compatibility limits from an OpenClaude-side regression.