sap-ai-core: GPT models cause infinite agent loop — model never stops responding

Bug Report

OpenCode Version: 1.18.23
Provider: sap-ai-core
Model: sap-ai-core/gpt-4.1, sap-ai-core/gpt-5.5 (affects all sap-ai-core/gpt-* models)

Description

When using any GPT/OpenAI model through the SAP AI Core provider, the agent never stops after giving a response. Even for a simple "Hi", the model replies repeatedly in an infinite loop.

Steps to Reproduce

1. Configure OpenCode with sap-ai-core provider and any gpt-* model
2. Send "Hi"
3. Receive a response
4. Observe the agent immediately sends the same conversation again
5. Model responds to "Hi" again
6. This repeats indefinitely — only Ctrl+C stops it

Root Cause

The SAP AI Core Orchestration API (used via @jerome-benoit/sap-ai-provider-v2) does not return a finish_reason: "stop" in the SSE stream in the format OpenCode/AI SDK expects.

OpenCode's agent loop checks finish_reason to decide whether to stop or continue. When finish_reason is missing, empty, or in an unexpected format, OpenCode treats it as "model wants to continue" and re-invokes the loop — sending the full conversation back to the model immediately.

Comparison: OpenCode already handles a similar case for Snowflake Cortex in packages/opencode/src/provider/provider.ts.

Expected Behavior

After the model returns a complete response, the agent loop should stop. The model should respond to "Hi" exactly once.

Suggested Fix

In packages/opencode/src/provider/provider.ts, the sap-ai-core custom loader should add a fetch wrapper that normalizes the SSE stream response — ensuring finish_reason: "stop" is present when the model has finished.