#2297 Agent history with long `tool_calls` arguments produces `reasoning_content` + empty `content` on final turns

本文

# Agent history with long `tool_calls` arguments produces `reasoning_content` + empty `content` on final turns

## Environment

- Strata **0.1.43.1**, `--serve`
- Model: `Qwen3.8-Flash-Next-UD-IQ4_XS` (served as `qwen3.8-flash-next-unsloth-ud-iq4_xs`)
- Launch flags: `--spec 4 --mtp rt --max-context 262144 --kv int8 --kv-resident 32768 --resident-budget-gib 55 --vision --vram-reserve-mib 700 --spec-min-p 0.5`
- `/metrics` engine block: `expert_slots 15896`, `pcie_frac 0.55`, `spec 6`, `mtp_max 4`, `lookup 3`, `vram_free_mib 1208`, `kv int8`, `kv_resident 32768`
- Client: pi 1.2.0 via `/v1/chat/completions`, `chat_template_kwargs: {"enable_thinking": true}`, assistant history replayed with `reasoning_content`

## Symptom

On the **final turn of an agent loop** — the turn that should produce a user-facing answer — the response comes back with a complete `reasoning_content` block, `finish_reason: "stop"`, several hundred output tokens, and **`content: ""`**.

```
finish_reason = "stop"
reasoning_content = "... 2000+ chars of correct, on-topic analysis ..."
content           = ""
output_tokens     = 340
```

The reasoning is coherent and answers the question — it just never gets emitted as `content`. Sometimes `content` contains a truncated fragment of the reasoning instead of the answer, which looks like the `