Agent loop stops after first LLM call when using @ai-sdk/openai-compatible with LiteLLM (finish_reason: stop instead of tool_calls) #20719

Description

Environment

• OpenCode: 1.3.13
• Adapter: opencode_local via Paperclip
• Provider: @ai-sdk/openai-compatible → LiteLLM v1.82.6 → Ollama (ollama_chat)
• Models tested: qwen3:32b, qwen3.5:122b, glm-4.7-flash, gpt-oss:120b
• Deployment: Docker/Portainer, self-hosted

Describe the bug

When using the opencode_local adapter with any OpenAI-compatible provider (LiteLLM proxy in front of Ollama), the agent loop exits after exactly one LLM call per run — before any tool is called.

The model receives the prompt correctly and produces output tokens, but OpenCode terminates the session with finish_reason: stop instead of continuing the agentic loop.

Root cause

OpenCode's loop exit condition in session/prompt.ts checks finish_reason. The OpenAI spec returns "tool_calls" when a tool was invoked, but @ai-sdk/openai-compatible providers (including LiteLLM) return "stop" even when the response logically should continue. This causes the loop to break after every single LLM call — before any tool is ever executed.

This was previously reported in #14972 but appears to still affect OpenCode 1.3.13 with @ai-sdk/openai-compatible.

Steps to reproduce

1. Set up OpenCode with @ai-sdk/openai-compatible pointing to LiteLLM (https://your-litellm-host/v1)
2. LiteLLM routes to Ollama with custom_llm_provider: ollama_chat
3. Run any prompt that should trigger tool use
4. Observe: OpenCode exits after the first LLM response, no tool is called

Observed behavior

Each run produces output tokens but zero tool calls. LiteLLM logs confirm:

Tools: 10 provided, 0 called
finish_reason: "stop"

The agent requires 10–20+ separate run invocations to complete what should be a 3–5 step task.

Expected behavior

The agent loop should continue after each LLM response until a tool is called, the model signals completion, or max_steps is reached.

Additional context

• Direct curl test to Ollama confirms tool calling works correctly: finish_reason: "tool_calls" is returned when tools are provided
• The bug only manifests through the @ai-sdk/openai-compatible → LiteLLM → Ollama chain
• OpenCode 1.3.13 also sends parallel_tool_calls which ollama_chat does not support — requires drop_params: true in LiteLLM as a workaround
• Tested with multiple models (qwen3:32b, qwen3.5:122b) — all exhibit the same behavior

Suggested fix

In session/prompt.ts, treat finish_reason: "stop" as non-terminal when the response also contains tool calls. The presence of tool calls in the response should take precedence over finish_reason when deciding whether to continue the loop.

Alternatively, expose a config option to treat "stop" as "tool_calls" for providers known to return incorrect finish_reason values.

Plugins

No response

OpenCode version

1.3.13

Steps to reproduce

No response

Screenshot and/or share link

No response

Operating System

No response

Terminal

No response