The agentic dataset records each turn's original output length, and GuideLLM enforces it with `max_completion_tokens = <recorded count>` **and** `ignore_eos: true` (set unconditionally in `request_handlers.py` when a per-turn `output_tokens_count` exists). `ignore_eos: true` forces the model to emit *exactly* that many tokens, past its own EOS. When the model under test needs more tokens than the recorded count to finish a tool call, the response is **hard-truncated mid-JSON**; the client then feeds that malformed `tool_calls[].function.arguments` back as the next turn's assistant message, and the server's chat preprocessing (`json.loads`) rejects it → a fast, empty-bodied **400** (surfaced as a 503 / `UNUSABLE_BACKEND_RESPONSE`). This is a **client-side request artifact, not a serving/router defect** — it reproduces identically with and without the KV-aware router, and only on the turn after a truncated tool call.
The shipped agentic specs carry a workaround in the backend block:
```j…
Fix GuideLLM default mapping by removing the integer turns column from all three Parquet files. Update the dataset card to explain the sampled workload and concurrent stream testing.
