[Bug] K2.7-Code on self-hosted OpenAI-compatible endpoint: model intermittently ends turn after thinking (finish=stop, no text/tool call) in multi-agent runs — APIEmptyResponseError

Summary

Running Kimi-K2.7-Code on a self-hosted OpenAI-compatible endpoint (vLLM 0.19.1, --tool-call-parser kimi_k2 --reasoning-parser kimi_k2), multi-agent workloads (AgentSwarm, 12 sub-agents) intermittently fail with APIEmptyResponseError. Extracting finishReason from every failing retry shows the dominant failure is the model emitting thinking and then a normal stop token — no text, no tool call is ever generated:

failing retries by finishReason, per run (kimi-code 0.39.1):
  7x completed   (vLLM 0.19.1)
  9x completed   (vLLM 0.19.1 + DCP)
  9x completed   (vLLM 0.28.0)
  2x tool_calls  (separate vLLM streaming-parser bug, filed as vllm-project/vllm#54701)

The finishReason=completed cases are engine-independent (reproduced identically on vLLM 0.19.1 and 0.28.0), so this is not the server's streaming parser — the model ends its turn after the reasoning block. Recorded wire-log content parts for a failing step look like:

{"part":{"type":"think","think":" One agent came back empty, so I'll quickly read that file myself and include it in the merged summary."}}
→ (no text part, no tool.call part; turn retried with APIEmptyResponseError)

i.e. the reasoning concludes with a clear intent to act, then the model stops. Retries hit the same behavior often enough that sub-agents exhaust attempts (typical damage: 1–4 failed sub-agents per 12-agent swarm; ~5–30% of tool-call-bearing requests under concurrent load; never observed single-stream).

What we tested (both negative)

1. System-prompt nudge ("after thinking, always emit the tool call or text — never end on thinking alone", injected via workspace AGENTS.md, confirmed present in sub-agent contexts): no improvement (9/12 and 8/12 completions vs the same baseline band).
2. KIMI_MODEL_THINKING_KEEP=off (verified to bind — thinkingKeep disappears from the CLI's request records): one 12/12 clean run, then 10/12 with 8× the same finishReason=completed failures — within run-to-run noise, not a fix.

Questions / ask

• Is think-only-then-stop a known K2.7-Code behavior in multi-turn agentic contexts where prior-turn reasoning is replayed (thinkingKeep: all default) to an OpenAI-compatible endpoint?
• Is the CLI's reasoning replay format for type = "openai" providers (assistant reasoning field) what the model's chat template expects, or should replayed thinking be rendered differently for self-hosted vLLM?
• Any recommended serving-side or CLI-side setting to prevent the model ending a turn inside/at the end of reasoning?

Environment

• kimi-code CLI 0.39.1, provider type = "openai", model entry: max_context_size 131072, max_output_size 32768, capabilities ["thinking","always_thinking","tool_use"], thinking effort high
• Server: vLLM 0.19.1 (also 0.28.0), moonshotai/Kimi-K2.7-Code official INT4, TP=8, --enable-auto-tool-choice --tool-call-parser kimi_k2 --reasoning-parser kimi_k2, temp 1.0 / top_p 0.95
• Full wire logs (CLI session records) and SSE captures available on request.

Start by reproducing the failure with the listed kimi-code CLI, OpenAI-compatible provider, vLLM versions, and multi-agent workload. Inspect the CLI session records and available wire logs, focusing on reasoning replay, thinkingKeep, finishReason=completed, and empty responses. Done means isolating whether the CLI replay format, model behavior, or serving configuration causes the think-only stop and documenting a narrowly scoped fix or confirmed limitation.