# Auto-summarization

Solo can generate a sidebar-visible, compact summary after recent activity in an agent or terminal. This gives you a quick way to catch up without re-reading the full output.

## What it does

Auto-summarization sends a compact rendered-text snapshot to the summarizer, not the full raw transcript. For PTY-backed agents and terminals, Solo reconstructs recent visible terminal rows plus recent history.

The summarizer is asked for a short summary plus a [state classification](./idle-detection.md) — whether the agent looks idle, waiting for permission, thinking, actively working, or stopped on an error.

## Claude recaps

Claude Code can print its own one-line recap at the end of a turn. When a Claude agent prints a recap line, Solo uses that text directly as the process summary — no summarizer call is needed, and this works even when auto-summarization is not set up.

Solo keeps the most recent complete recap and truncates very long recaps to fit the summary line. Recaps only apply to Claude agents; if you disable recaps in Claude Code's own `/config`, Solo falls back to the summarizer.

## Setup

Open **Settings → Agents → Auto-summarization** and choose a **Summarizer tool**. Auto-summarization is disabled until a tool is selected.

Claude, Codex, Gemini, and Copilot can run summaries natively — select one of them and it works without extra setup. Custom (generic) tools need a **Headless command** configured in the [agent tool editor](./setting-up-tools.md) before they can summarize. Amp, OpenCode, Kimi, and Antigravity don't have a native summarizer in Solo; to use one of those CLIs for summaries, add it as a custom tool with a headless command.

The model field is passed through the selected tool's model flag when Solo uses a native summarizer. Current native defaults are:

- Claude: `sonnet`
- Codex: `gpt-5-codex`
- Gemini: `flash-lite`
- Copilot: `auto`

## Session isolation

Native summarizer runs are one-off headless invocations. Solo explicitly keeps Claude and Codex summary calls out of their normal resume histories:

- Claude summaries run with session persistence disabled, so headless summary calls never show up in Claude Code's resume list.
- Codex summaries run as ephemeral sessions, so they are not saved to Codex's resume list.

Gemini and Copilot summaries also run as one-off headless calls, but Solo does not add a provider-specific history-isolation flag for them. Whether those calls appear in resume history is controlled by the provider CLI.

Your interactive agent sessions are unaffected — the summarizer reads a snapshot of the terminal output and never writes into the conversation you're having with an agent.

## Cadence and triggers

The cadence setting is the minimum time between summary attempts for the same process. The available cadences are **15 sec**, **30 sec**, and **1 min**.

Solo also waits for recent activity before requesting a summary:

- A process becomes eligible for summaries only after you've typed into it.
- A brief quiet window after output stops triggers a summary.
- Long-running continuous output can still produce periodic summaries, based on the selected cadence.
- If there is too little recent output to be worth summarizing, the attempt is skipped.

To save tokens, agent-spawned subagents are skipped for automatic summaries. Their parent group can still show its own summary.

## Testing

Use **Test** in the Auto-summarization settings to run the selected summarizer directly. The health check passes only when the tool replies with `solo okay`.

## Notes

Summaries are a context aid, not a replacement for the full terminal output or chat transcript when you need exact details.

---

Are you a human? Read this doc on the web: https://soloterm.com/docs/agents/auto-summarization