fix(openai_legacy): improve handling of empty streaming responses from OpenAI-compatible backends

[jhinpan]

Problem

When using kimi-cli with an OpenAI-compatible backend (e.g. SGLang serving Kimi-K2.5), the server occasionally returns streaming responses that complete without any content or tool calls. After exhausting 5 retries (tenacity), the error surfaces to the user as:

LLM provider error: The API returned an empty response.

The current turn is lost and the user must re-enter their prompt.

Root Cause

In packages/kosong/src/kosong/_generate.py, after the stream finishes, if the assembled message has neither content nor tool_calls, an APIEmptyResponseError is raised. The retry logic in kimisoul.py (_kosong_step_with_retry) correctly treats this as retryable, but with only 5 attempts and short backoff (0.3s initial, 5s max), all retries can exhaust quickly when the backend is transiently returning empty responses.

The shell UI (src/kimi_cli/ui/shell/__init__.py) has no specific handler for APIEmptyResponseError — it falls through to the generic ChatProviderError case with a terse red error message, which can be confusing for users since the server is reachable.

Proposed Fix

1. Better diagnostics in _generate.py: Add a logger.warning() before raising APIEmptyResponseError that captures the stream ID and the number of chunks received. Include these details in the error message for easier troubleshooting.

2. Dedicated UI handler in shell/__init__.py: Add an APIEmptyResponseError-specific branch in the error handler that displays a yellow (warning-level) message like: "Server returned an empty response after retries. This is typically a transient issue — please try again." This makes the error less alarming than a generic red "LLM provider error" and signals that it's retryable.

Environment

• Backend: SGLang v0.4.x serving Kimi-K2.5 via OpenAI-compatible API
• Provider: openai_legacy with stream=True
• Hardware: AMD MI355X
• kimi-cli version: 1.12.0

Reference

Fix implemented in fork: https://github.com/jhinpan/kimi-cli-fork/tree/fix/handle-empty-response-from-openai-compatible-backends