auto-compact-percent

Pi extension that shows a live context-usage bar in the status line and auto-compacts early — either at a configurable percentage of the context window, or when the real input token count (including the system prompt) reaches a fixed threshold.

Unlike pi's native compaction (which fires near the model limit, based on reserved response tokens), this lets you compact early at any percentage you choose, or at a specific token count, with a visible countdown bar.

What it does

• Context bar: after every turn, renders a bar in the pi status line using the real total input from the last assistant response: [█████░░░░░ 14%] 27k/200k
  • Both the bar % and the 27k/200k suffix reflect everything the model actually has in context: system prompt + tools + conversation + cached tokens (usage.input + cacheRead + cacheWrite, same as pi's own "Input" accounting).
  • Before the first response it shows a chars/4 estimate of just the system prompt, so the bar is never a misleading 0/1M.
  • After compaction the total is reset (it's stale) and the bar falls back to the system-prompt floor until the next response reports fresh usage.
• Trigger marker: once usage passes the threshold, appends trigger to the bar so you can see it's about to compact.
• Auto-compaction: when either trigger fires (see below) and the debounce window has passed, calls ctx.compact() with instructions focused on the most recent work. Shows a compacting bar and notifies on start/complete.
• Two triggers: compacts when the real total reaches threshold fraction of the context window, OR when the real total token count reaches maxTokens (whichever comes first). The threshold is now measured on the REAL total (system prompt + cache included), so at 40% you're compacting when the model's context is actually 40% full — not 40% of conversation.
• Debounce: skips compaction if fewer than debounceTurns turns have passed since the last one (or since session start), so a burst of activity doesn't trigger compaction repeatedly.
• Eligibility check: skips compaction when context usage is below minTokens, so a short session with nothing old to summarize isn't compacted wastefully.

Commands

| Command | Purpose | | ---------------------------------------- | ------------------------------------------------------------------------------------ | | /auto-compact | Toggle auto-compaction for the current session (in-memory only; resets on /reload) | | /auto-compact on / /auto-compact off | Enable or disable explicitly |

The toggle is session-scoped and never writes to your settings files.

Install

As a pi package (from git)

pi install git:github.com/christopherarter/pi-auto-compact-percent

npm publishing is planned; until then install from git above.

Manual (local copy)

Pi auto-discovers extensions in ~/.pi/agent/extensions/:

mkdir -p ~/.pi/agent/extensions/auto-compact-percent
cp index.ts README.md ~/.pi/agent/extensions/auto-compact-percent/

Then reload pi (/reload) or restart. Loads from ~/.pi/agent/extensions/auto-compact-percent/index.ts.

Configuration

Configure via autoCompactPct in either:

• ~/.pi/agent/settings.json (global, all projects)
• .pi/settings.json in the current project (project override, wins)

| Key | Default | Description | | ------------------------ | -------------------------------------------------------------------- | ------------------------------------------------------------------------------------------- | | threshold | 0.4 | Fraction (0–1) of context window (REAL total, incl. system prompt) that triggers compaction | | maxTokens | 0 | Compact when real input tokens (incl. system prompt + cache) reach this. 0 disables | | minContextWindow | 32000 | Don't auto-compact if the model's context window is smaller than this | | minTokens | 8000 | Don't auto-compact if context usage is below this (nothing old to summarize) | | debounceTurns | 5 | Minimum turns between compaction attempts | | barWidth | 20 | Width of the rendered progress bar in characters | | compactionInstructions | "Focus on the most recent work, decisions, and in-progress tasks." | Custom instructions passed to ctx.compact() |

Example:

{
  "autoCompactPct": {
    "threshold": 0.5,
    "maxTokens": 200000,
    "minContextWindow": 64000,
    "minTokens": 10000,
    "debounceTurns": 3,
    "barWidth": 30,
    "compactionInstructions": "Preserve unfinished work and exact details required to resume safely."
  }
}

Invalid values (threshold out of range, negative numbers, tiny bar widths) fall back to defaults.

Development

Single-file extension + node:test suite. To package for npm:

npm test        # run the 26-test suite (node --test + module mocks)
npm pack        # verify tarball contents
npm publish     # requires "pi-package" keyword (already set)

Dependencies: only the pi core package (@earendil-works/pi-coding-agent), listed as a peer dependency — pi bundles it at runtime, so no install needed.

Behavior notes

• The bar is always rendered (even below threshold) while enabled; compaction only fires when a trigger is reached.
• The bar's % and the 27k/200k suffix are both the real total from the last assistant response (usage.input + cacheRead + cacheWrite, same as pi's own "Input" accounting) — the accurate measure of window pressure, system prompt included. At session start, before the first response, both show a chars/4 estimate of the system prompt so it's not 0. If maxTokens is set, it compares against this real total.
• Compaction respects pi's own compaction.enabled setting — if compaction is disabled globally this extension's ctx.compact() may be a no-op.
• On compaction failure the bar clears and compaction stays disarmed until usage drops back below the threshold, so a broken compaction can't cause a retry loop.
• The /auto-compact toggle is in-memory: it survives turn_end calls but resets to enabled on /reload or restart. It never modifies settings files.
• No background resources are started from the factory; all work happens on session_start and turn_end events.
• UI calls (notify) are guarded with ctx.hasUI so it degrades gracefully in print/JSON modes.

Similar packages

• @thunstack/auto-compact — percentage-based auto-compaction with a /auto-compact toggle and TUI config panel, but no live context bar.
• pi-powerline-footer — shows context as text (12k/200k (6.2%) or 83%) but doesn't auto-compact and has no progress bar.

This package combines the live visual bar with early percentage-based compaction, plus a real-token (maxTokens) trigger for a tight, window- independent leash.