# Show HN: Compactly – Compact Claude Code sessions between tasks

> Source: <https://github.com/swapnildahiphale/compactly>
> Published: 2026-10-26 12:58:47+00:00

**Compact smartly.**

A Claude Code plugin that compacts your session at task boundaries, with a handoff summary for
the next task, and shows what each turn costs and when the prompt cache expires. The name is
the idea: *compact smartly*, with the "smart" squeezed out.

## compactly.mp4

<sub>Illustrative session · costs vary by model and context.</sub>

Claude Code re-reads the whole conversation on every request, so a large context carried into an unrelated task is paid for again on every turn. Compacting in the middle of a task drops what it still needs. compactly compacts in between: right after the work is done.

- **Compacts at natural breaks** , with a summary written as a handoff for the next task.
- **Shows what each turn costs** and when the prompt cache expires.
- **Learns from your 👍 / 👎** on each automatic compaction.

Requires Claude Code 2.1.295 or later. compactly is a plugin and installs through Claude Code's plugin system. Inside Claude Code:

```
/plugin marketplace add swapnildahiphale/compactly
/plugin install compactly@compactly
```

Or from a shell:

```
claude plugin install compactly --marketplace swapnildahiphale/compactly
```

Start a new session afterwards. See [installation](https://github.com/swapnildahiphale/compactly/blob/main/docs/installation.md) for a local clone,
scopes, updating and uninstalling.

- `/cost-panel` opens the cost & cache panel, or press the footer line beside the prompt.
- `/compactly` compacts now.
- By default (`auto` mode), Claude Sonnet reads each new prompt once the context is large. If
the prompt starts after a natural break, the plugin compacts first and then sends the prompt.
A band above the prompt asks**👍 Yes** or**👎 No, I was mid-task** , and the judge learns
from your answers.
- In `suggest` mode, the plugin only tells you when a git commit, push or PR is a good moment to
compact.

Open it with `/cost-panel`, or press the footer line beside the prompt.

- **Header** : the model, the cache TTL, a**⚙ Settings** button and a**Compact** button.
- **Stat tiles** with sparklines: session cost, cache hit, re-read per turn, rebuild if cold.
- **Cache clock** : drains as the cache ages, with a compact window before it goes cold.
- **Context** : tokens used against the model's window.
- **Hints** : ranked actions: compact now, snooze 10 min, not this session.
- **Cost per turn** : cache read, cache write, output and uncached input.
- **Footer line** : a one-line summary such as`● warm 59m · $0.012 if cold` , visible even
when the panel is closed.

**⚙ Settings** chooses which sections the panel draws and which fields the footer line
carries, by preset (`minimal`, `cost`, `standard`) or one at a time. See
[settings](https://github.com/swapnildahiphale/compactly/blob/main/docs/panel.md#settings).

**Saved by compacting** keeps score of what your compactions saved after their own cost, and
an all-time total earns levels from Cache Rookie to Compaction Legend.

Details: [the panel](https://github.com/swapnildahiphale/compactly/blob/main/docs/panel.md).

In `auto` mode, a judge model decides when you send your next prompt:

- If no prompt follows a commit, push or PR, it compacts 5 minutes before the cache expires,
once the context is at least `minTokens` and your work isn't paused.
- If your recent prompts say work is paused, git signals don't compact automatically.
- In `suggest` mode it only tells you when a commit, push or PR is a good moment.

Details: [how it works](https://github.com/swapnildahiphale/compactly/blob/main/docs/how-it-works.md), [decision flows](https://github.com/swapnildahiphale/compactly/blob/main/diagrams/architecture.md).

| Option | When it runs | Summary | 
|---|---|---|
| `/compact` | when you type it | general history of the conversation | 
| autocompact (built in) | when the context nears the window limit | general | 
| idle compaction (built in) | after about 55 min idle, context ≥ 200K | general | 
| `/clear` , new session | when you run it | none; only CLAUDE.md and memory remain | 
| **`/compactly`** (or automatic) | when you type it, or at a natural break before the next prompt | handoff for the next task | 

The trade-off: the judge is sometimes wrong, so some automatic compactions land mid-task.
Rate them and it becomes more careful, or set `mode` to `suggest` to confirm each one.
Details: [how it works](https://github.com/swapnildahiphale/compactly/blob/main/docs/how-it-works.md#how-it-compares).

The defaults work as they are. The settings people change most:

| Setting | Default | Meaning | 
|---|---|---|
| `mode` | `auto` | `suggest` : notify only;`auto` : compact at a natural break;`off` : do nothing | 
| `judge` | `sonnet` | `sonnet` ,`haiku` (cheaper, compacts less often), or`off` | 
| `minTokens` | `150000` | act only when the context is at least this large | 

```
claude plugin install compactly --marketplace swapnildahiphale/compactly \
  --config mode=suggest --config minTokens=100000
```

Details: [configuration](https://github.com/swapnildahiphale/compactly/blob/main/docs/configuration.md), [installation](https://github.com/swapnildahiphale/compactly/blob/main/docs/installation.md).

The plugin makes no network calls of its own. The judge runs through Claude Code with the model
you pick. Ratings and savings totals are kept in the plugin's own store under `~/.claude`, on your
machine.

Issues and pull requests are welcome. To work on the plugin from a clone, see
[Develop](https://github.com/swapnildahiphale/compactly/blob/main/plugin/compactly/README.md#develop).