bug: infinite loop when provider returns finishReason "unknown"

[okuyam2y]

Description

When using OpenAI-compatible providers (e.g. GLM-5 Turbo via gateway), the model returns finishReason: "unknown" instead of "stop" or "end_turn". The prompt loop in prompt.ts treats "unknown" as non-terminal (same as "tool-calls"), causing an infinite loop when no tool calls are present.

Steps to reproduce

1. Configure an OpenAI-compatible provider that returns finishReason: "unknown" (e.g. ZhipuAI GLM-5 Turbo)
2. Send any message
3. The session loops indefinitely — each step has finishReason: "unknown" and hasToolCalls: false

Expected behavior

The prompt loop should treat "unknown" as a terminal finish reason (like "stop") when no tool calls are detected, and exit the loop.

Relevant code

packages/opencode/src/session/prompt.ts L324 and L720 — "unknown" is excluded from terminal reasons alongside "tool-calls":

!["tool-calls", "unknown"].includes(processor.message.finish)

[github-actions]

---

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

• z.ai/zhipuai finish_reason: "model_context_window_exceeded" is treated as unknown, causing loop instead of stopping #18813: Very similar issue — ZhipuAI provider returns an unrecognized finish_reason (treated as unknown), causing the prompt loop to not terminate
• GLM-5 model gets stuck in infinite loop, repeating same actions #18415: GLM-5 model (same provider: ZhipuAI GLM-5) gets stuck in an infinite loop

If your issue is distinct from these, please clarify the difference. Otherwise, consider adding a reaction or comment to the existing issue instead.

[okuyam2y]

This is the same root cause as #18813. Fix PR: #19342

[heimoshuiyu]

same root cause as #18813 but not #18415 . In #18813 model stop with unknow/something_their_defined_stop_reason. in #18415 model just response the repeated message. That is their glm-5 infra bug. https://zhipu-ai.feishu.cn/share/base/form/shrcnfiAPrQ3o9CewRdCtXXJwBd

[okuyam2y]

Thanks for the clarification. You're right — #18415 is a different issue (GLM-5 repeating the same actions due to a model-side infra bug), while #19339 and #18813 share the same root cause (non-standard finish reasons being treated as unknown and keeping the prompt loop running). PR #19342 addresses #18813 and #19339 but would not fix #18415.