pi-squeeze

Shrinks bash output before it reaches the agent's context: native compactors for git, search, listings, tests, builds (cargo, gradle, Maven, javac, Xcode, tsc, docker), Android (adb, lint, detekt), linters, package installs and logs, plus rtk when install

Packages

Package details

extension

Install pi-squeeze from npm and Pi will load the resources declared by the package manifest.

npm repohome report
$ pi install npm:pi-squeeze
Package
pi-squeeze
Version
0.4.1
Published
Nov 11, 2026
Downloads
1,111/mo · 308/wk
Author
greenspand
License
MIT
Types
extension
Size
163.9 KB
Dependencies
0 dependencies · 1 peer
Pi manifest JSON
{
  "extensions": [
    "./extensions/squeeze.ts"
  ]
}

Security note

Pi packages can execute code and influence agent behavior. Review the source before installing third-party packages.

README

pi-squeeze

npm CI License: MIT Audit Fuzz

Shrinks bash output before it reaches the agent's context: native compactors for git, search, listings, tests, builds (cargo, gradle, Maven, javac, Xcode, tsc, docker), Android (adb, lint, detekt), linters, package installs and logs, plus rtk when installed. Full output stays one temp file away.

A pi extension.

Install

pi install npm:pi-squeeze

What it does

After a bash command, pi-squeeze replaces its output with a shorter form with the same facts. It only touches output it fully understands, and otherwise passes it through unchanged:

Command Becomes
rg -n / grep -rn / ack totals line, then matches grouped by file (path written once, indent trimmed, long lines and busy files cut)
rg pattern dir / grep -r (no -n) the same, but only when every path: prefix is a file that exists in the session's directory (a colon in the text is never taken for a path)
fd, find paths grouped by directory
lsd -l, eza -l, ls -l type, size, short date and name (owner, group, permissions and the total line dropped)
git diff one header line per file (path (new, +12 −0)), index/mode/---/+++ lines dropped, every changed line kept, context trimmed to 1 line, long files cut
cargo test totals line, then only warnings, failures, panics and result lines (build progress and passing tests dropped); also when the run failed
rg -l / rg --files / grep -l / git ls-files / git diff --name-only paths grouped by directory
git status ## main...origin/main [ahead 2], then staged: / unstaged: / untracked: with M/A/D/R codes; the hints go
git log one line per commit (short hash, date, author, subject); the body stays, capped at 8 lines; one author is said once
git show that commit line, then the compacted diff
git blame one header per run of lines from the same commit
dust size, tree and share; the bar chart goes
cargo build / check / clippy / doc errors keep their whole block; warnings become one line each, grouped by lint (clippy::needless_return (2)); Compiling/Finished lines go
cargo audit one line per advisory (id, crate, version, kind, title, fix)
cargo tree, tree, eza --tree cut below 3 levels, … [+15 lines deeper] under each cut
tsc errors grouped by file (a.ts (8), 1:7 TS2322 …)
gradle, ./gradlew task lines counted (11 tasks (5 executed, 4 no-source, 2 up-to-date)); failures, compiler errors, test failures and the BUILD line stay; daemon, configuration-cache, JVM and --scan chatter goes; reported files lose their shared directory
bun test, go test passing tests counted, not listed; failures and totals stay
pytest session header, progress dots and PASSED lines go; failures, warnings and the short summary stay; full-width ==== banners shrink
jest, vitest passing tests and files go, skipped ones are counted; failures with their diff and frame, and the totals, stay
xcodebuild, swift build / test build steps (CompileSwift, Ld, CodeSign and their command lines) and progress counted, not listed; warnings grouped by message; errors keep frame and notes; passing XCTest / Swift Testing tests counted; the ** BUILD SUCCEEDED ** line stays
eslint, biome, ruff, flake8, mypy, pylint, ktlint, detekt, swiftlint, shellcheck, flutter analyze findings grouped by rule and message, each location once (error eqeqeq: … (5) then m.js:2:46 m1.js:2:47 …), the shared directory said once; frames and fix suggestions go
npm / pnpm / yarn / bun install, pip install, uv sync, pod install, bundle install, flutter pub get download and progress lines go; per-package lines become one line (Collecting (9): requests, rich, …); deprecation warnings become one list of names; errors and the summary stay
docker build [internal] steps, layer downloads, CACHED / DONE lines go; each step keeps its header and what it printed
docker logs, docker compose logs, kubectl logs, journalctl, adb logcat (not -f / --follow) 3+ lines that differ only in timestamps and numbers keep the first and last; framework stack frames fold; logcat drops the date and thread id; past the limit the start goes, not the end
gradle on Android also: Kotlin / KSP warnings, detekt / ktlint findings and Android Lint issues grouped by message with their locations (lint's source lines, explanations and baseline advice go); repeated warnings and --warning-mode all deprecations said once with a count, without the upgrade-guide links; R8 Missing class and Duplicate class lines become one list of classes; "Searched in the following locations" becomes a count; A failure occurred while executing … wrappers go and the failure summary does not repeat what was printed above it; framework frames fold
gradle tasks, gradle projects, dependencyInsight, dependencies tasks: names per group, descriptions go; projects: every path once, grouped by the first segment (:api (261): …); insight: the attribute tables go; dependencies: cut like a tree
mvn, ./mvnw, mvnd Downloading / Downloaded lines and their progress counted; [INFO] lines go except the result, the test totals and modules that failed; a passing test class takes its output (Spring banners, logs) with it; errors said once, the [Help 1] footer goes; dependency:tree is cut like a tree
javac errors keep their source line, caret and symbol: / location:; warnings grouped by kind (warning rawtypes: found raw type: List (2)); Recompile with -Xlint notes go
a JVM stack trace from anything (java -jar, a test JVM, a log) framework frames (java., kotlin., android., org.junit., org.springframework.…) fold to … N framework frames; the exception, each Caused by:, your own frames and ... N more stay; also when the command failed
adb exec-out uiautomator dump /dev/tty one line per node that says something, indented by nesting, with where to tap: TextView "Network & internet" #title @375,608; empty wrappers and off-screen nodes go
adb shell getprop, settings list … keys that share a prefix written once, every value kept: aaudio: hw_burst_min_usec=2000 mmap_policy=2
adb shell ps -A, top -n 1 kernel threads and idle processes counted, not listed; column padding goes
adb shell dumpsys … lines deeper than an entry's own go with a count; 3+ lines of the same shape keep their ends
adb shell pm list packages [-f] grouped by the first two name segments; with -f, name then path, hashed install directories elided
adb logcat (threadtime, time, brief) as logs above; with adb -s <serial> too
… | tail -50, … | head -n 20 the command before the pipe picks the compactor
anything else colour codes dropped, progress bars (\r) keep their last frame, 4+ identical lines become line [×N], 3+ blank lines become one. Never for cat, head, jq, curl… (content commands)
git, cargo, docker… rewritten through rtk when it is installed (switch: /squeeze rtk off)
rtk rg …, rtk git diff… run plain again when a native compactor handles the command (typed by the model or rewritten by another extension), so the native compactor gets the output

Output with a marker of its own (a tsc error list, a gradle or Maven build, javac errors, a JVM stack trace, a uiautomator dump, a bun test run, rustc diagnostics) is recognised whatever ran it: just check, npm run build, make test, … 2>&1. A failed command is compacted too, but only for test and build output, where the errors are the point. git -C dir …, git --no-pager …, --color=never and cargo +nightly … are read as the plain command.

Safety rules:

  • Pipes, redirects, chains and --json/-c/-l/-x style flags are never touched: the output is no longer the tool's own format.
  • A line that does not parse makes the whole output pass through.
  • The result must be clearly smaller (under 90%) and the output at least 400 characters.
  • Every cut leaves a marker (… [+12 more in this file]), and the full original goes to a temp file whose path is in the output.
  • read, edit and write are never changed.

Commands

Command
/squeeze or /squeeze status the settings, this session's savings, and the all-time savings by compactor and command plus the biggest outputs no compactor took
/squeeze [on|off] switch everything
/squeeze [search|paths|listing|diff|tests|git|build|lint|install|logs|tree|generic] [on|off] switch one compactor
/squeeze rtk [on|off] rewrite git, cargo, docker… through rtk, when installed
/squeeze min <chars> smallest output worth compacting (default 400)

Settings are remembered in <pi agent dir>/squeeze.json (PI_CODING_AGENT_DIR, else ~/.pi/agent); it also holds lineChars, perGroup and total, the limits of a compacted result, maxRatio (a result is used only when at most this share of the original, 0.5 to 1, default 0.9), and logSavings, logMissed and missedChars (below). PI_SQUEEZE=off disables it at start. The status line shows the tokens saved.

Savings log

Each compacted output adds one line to <pi agent dir>/squeeze-savings.jsonl, so the savings can be tracked over time; the model never sees it:

{"ts":"2026-01-02T09:15:00.000Z","session":"…","cwd":"/work/app","tokens":1840,"chars":7360,"kind":"build","cmd":"cargo build"}

cmd is the program and, for tools with subcommands, the subcommand: never arguments, paths or patterns.

tokenburn and pi-tokenburn read it as the "pi squeeze" saving. When rtk ran the command, rtk logs its own part; this log holds only what pi-squeeze cut after it. "logSavings": false in squeeze.json turns it off.

Outputs of at least missedChars characters (default 2000) from commands no compactor handles go to <pi agent dir>/squeeze-missed.jsonl as {"ts","session","cwd","cmd","chars"}: the commands worth a compactor next. Content commands (cat, jq, head…) are never logged. "logMissed": false turns it off.

Development

bun install
just check   # typecheck + tests (95% coverage gate) + pack check + nu script tests
just --list  # everything else: fuzz, audit, release, remotes

Layout:

Path What
extensions/squeeze.ts the extension: pure functions plus a thin default export
tests/ bun test: core (pure logic), extension (hooks), fuzz (randomised properties)
scripts/ Nushell only, each with a test in scripts/tests/
.github/, .gitea/ CI, release, nightlies on both hosts

Releases run in CI: Actions → Release → Run workflow with a bump (patch, minor, major, or current for the version already in package.json), or just release-ci patch. The workflow runs the quality gate, bumps package.json, regenerates CHANGELOG.md with git-cliff, commits, tags, pushes and publishes to npm (secret NPM_TOKEN). Locally, just release patch does the same up to the tag push.

License

MIT