# 08 Compress

Script: `example/08_compress.py`.

`flow.Compressor` runs `run_session_compress(...)` to replace a long transcript with a smaller summary session.

```bash
python example/08_compress.py
```

| Step | Behavior |
| --- | --- |
| Build history | Run an agent turn to create assistant chunks. |
| Count rows | Inspect `len(session.chunk_table.rows)`. |
| Compress | Call `compressor.forward(session)`. |
| Continue | Use the compressed session as the next user-side context. |

Compression disables tools and expects the model to return text only. In
v1.2.2, `finish_reason="length"` raises `RuntimeError` because a truncated
summary must not replace the whole transcript. Increase the provider's output
token budget and retry; OpenRath does not retry this case automatically.

The compressed output contains one **user** chunk. `Session.text()` searches
assistant chunks, so read the summary from `compressed.chunk_table[-1]` rather
than calling `compressed.text()`.

[← Runnable Examples](index.md)