Default agent is brittle against local OpenAI-compatible tool-call quirks

[krystophny]

Summary

When the default OpenCode agent talks to local OpenAI-compatible backends, two brittle cases can break or stall normal tool use:

• the bash tool can fail hard if the model omits description but provides a valid command
• some local backends return finish_reason: "tool_calls" together with tool_calls: [], which should not be treated as a real tool-call turn

Reproduction

Use the default build agent against a local OpenAI-compatible backend such as llama.cpp or LM Studio.

Observed failures:

• The bash tool was called with invalid arguments ... expected string, received undefined on description
• session loop/hang behavior when empty tool_calls: [] is returned

Expected

• harmless missing bash descriptions should be normalized or defaulted
• empty tool_calls arrays should be treated like a normal stop completion

Verification

This was reproduced against a local llama.cpp backend with the default agent/tool stack.

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• bug: infinite loop when provider returns finishReason "unknown" #19339: Infinite loop when provider returns finishReason "unknown" — very similar to the empty tool_calls: [] causing a session loop/hang
• Session loop doesn't stop when model returns finish_reason=stop #11153: Session loop doesn't stop when model returns finish_reason=stop — also covers the session continuing when no tool calls are present

The first part of this issue (bash tool failing on missing description) may be distinct, but please review those issues to see if your second symptom is already tracked.

[ByteBand1t]

This also fixes #20719 — our Paperclip + LiteLLM + Ollama setup where agents stop after first LLM call. Please prioritize merging

[daniel-farina]

I tested this and helped get gemma working with open code
https://gist.github.com/daniel-farina/87dc1c394b94e45bb700d27e9ea03193