shrinker-ai
v0.11.0
Published
Local deterministic command-output shrinker for coding agents
Readme
shrinker
A small local CLI proof of concept that removes noise from command output before a coding agent or LLM reads it.
this POC proves that a few conservative, deterministic filters can save useful context without requiring an API key, service, UI, database, or agent-specific integration.
What the POC demonstrates
- Explicit cross-agent command wrapping:
shrinker exec -- <command>. - Filters for Git status, Git diff, test output, and generic logs.
- Failures, warnings, changed paths, and command exit codes are preserved.
- Omitted raw output is saved locally for recovery.
- Optional per-run metrics report bytes and approximate before/after tokens.
- Local SQLite statistics accumulate savings across runs.
- No command output leaves the machine.
Quick start
The recommended customer install downloads a standalone binary from GitHub Releases. It does not require npm registry access or a local Node.js installation.
Contributor and npm-based installs still require Node.js 22.13 or newer. This is the first Node 22 release where the built-in SQLite module no longer requires an experimental flag.
One-command install (macOS zsh)
curl -fsSL https://raw.githubusercontent.com/ivanduplenskikh/shrinker/main/integrations/macos/install.sh | bashTo pin a version:
curl -fsSL https://raw.githubusercontent.com/ivanduplenskikh/shrinker/main/integrations/macos/install.sh | bash -s -- --version 0.4.0One-command install (Windows PowerShell)
irm https://raw.githubusercontent.com/ivanduplenskikh/shrinker/main/integrations/windows/install.ps1 | iexTo pin a version:
& ([scriptblock]::Create((irm https://raw.githubusercontent.com/ivanduplenskikh/shrinker/main/integrations/windows/install.ps1))) -Version 0.4.0The installer downloads these anonymous GitHub Release assets by default:
shrinker-win-x64.zipshrinker-macos-arm64.tar.gzshrinker-macos-x64.tar.gzshrinker-linux-x64.tar.gz
Network allowlists need access to raw.githubusercontent.com for the installer script and github.com/ivanduplenskikh/shrinker/releases/download/... for release assets.
Update notices
Shrinker checks GitHub Releases at most once every 24 hours and prints an update notice to stderr when a newer version is available. The check only requests release metadata from GitHub; command output, paths, stats, and machine identifiers are not sent.
To disable update checks for one shell session:
$env:SHRINKER_UPDATE_CHECK = "0"Or set it in ~/.shrinker/config:
SHRINKER_UPDATE_CHECK=0Optional npm package install
Use this only when npm registry access is available or preferred.
Windows PowerShell:
pwsh -ExecutionPolicy Bypass -File .\integrations\windows\install.ps1 -UseNpm
# Or manually:
npm install --global shrinker-ai --registry=https://registry.npmjs.org
$pkg = Join-Path ((npm root --global).Trim()) "shrinker-ai"
pwsh -ExecutionPolicy Bypass -File (Join-Path $pkg "integrations\\windows\\install.ps1") -Local -SkipNpmInstall -SkipBuild -SkipLinkTo enable automatic PowerShell routing, add -EnableProfileRouting to the final command.
macOS zsh:
bash ./integrations/macos/install.sh --use-npm
# Or manually:
npm install --global shrinker-ai --registry=https://registry.npmjs.org
pkg="$(npm root --global)/shrinker-ai"
bash "$pkg/integrations/macos/install.sh" --local --skip-npm-install --skip-build --skip-link --enable-profile-routingInstall from local checkout
Windows PowerShell:
pwsh -ExecutionPolicy Bypass -File .\integrations\windows\install.ps1 -LocalmacOS zsh:
bash ./integrations/macos/install.sh --localThis writes managed guidance blocks globally to:
~/.copilot/copilot-instructions.md~/.claude/CLAUDE.md
The shared guidance source is templates/agent-rules.md; the files above are created in the corresponding global agent directory.
The rules tell agents to prefer shrinker <command> for high-volume commands while leaving native commands untouched.
Uninstall
GitHub Release binary install:
Windows PowerShell:
pwsh -ExecutionPolicy Bypass -File .\integrations\windows\uninstall.ps1macOS zsh:
bash ./integrations/macos/uninstall.shIf you installed from npm, use npm mode:
npm uninstall --global shrinker-ai --registry=https://registry.npmjs.orgIf you also installed profile/rules through the package scripts, run the matching local uninstaller from the installed package before uninstalling:
Windows PowerShell:
$pkg = Join-Path ((npm root --global).Trim()) "shrinker-ai"
pwsh -ExecutionPolicy Bypass -File (Join-Path $pkg "integrations\\windows\\uninstall.ps1") -SkipUnlinkmacOS zsh:
pkg="$(npm root --global)/shrinker-ai"
bash "$pkg/integrations/macos/uninstall.sh" --skip-unlinkmacOS one-liner uninstall:
curl -fsSL https://raw.githubusercontent.com/ivanduplenskikh/shrinker/main/integrations/macos/uninstall.sh | bashLocal repo install/uninstall (contributors)
If you cloned this repository and want to run scripts directly from the local path:
pwsh -ExecutionPolicy Bypass -File .\integrations\windows\install.ps1 -LocalOr on macOS:
bash ./integrations/macos/install.sh --localTo uninstall:
pwsh -ExecutionPolicy Bypass -File .\integrations\windows\install.ps1 -UninstallOr on macOS:
bash ./integrations/macos/install.sh --uninstallUninstall options:
# Keep shrinker command installed, but remove profile integration and managed rules
pwsh -ExecutionPolicy Bypass -File .\integrations\windows\install.ps1 -Uninstall -SkipUnlink
# Keep managed rules files unchanged while uninstalling command/profile hooks
pwsh -ExecutionPolicy Bypass -File .\integrations\windows\install.ps1 -Uninstall -SkipAgentRulesmacOS uninstall options:
# Keep shrinker command installed, but remove profile integration and managed rules
bash ./integrations/macos/install.sh --uninstall --skip-unlink
# Keep managed rules files unchanged while uninstalling command/profile hooks
bash ./integrations/macos/install.sh --uninstall --skip-agent-rulesIf your current terminal had already loaded shrinker-profile.ps1 or shrinker-profile.zsh, restart terminal (or remove loaded wrapper functions) to fully return to native command behavior.
npm install
npm run build
node dist\src\cli.js exec -- git status
node dist\src\cli.js exec -- git diff
node dist\src\cli.js exec -- git log -n 10
node dist\src\cli.js exec -- npm test
Get-Content .\tests\fixtures\generic-log.txt -Raw |
node dist\src\cli.js pipe --kind logTo install the shrinker command locally:
npm link
shrinker git status
shrinker git log -n 10
shrinker npm testCLI
shrinker <command> [args...]
shrinker exec [options] [--] <command> [args...]
shrinker pipe [options]
shrinker stats [--json]
shrinker last [--path]
shrinker raw <capture-id> [--path]
shrinker track --executable <name> [--subcommand <name>] [--bytes <number>] [--exit-code <number>]
shrinker help
--kind <auto|git-status|git-diff|git-log|test|log>
--max-lines <number> default: 120
--per-file-lines <number> default: 40
--raw bypass filtering
--metrics print per-run savings and duration
--no-save do not save omitted raw output
--no-stats do not record this run
--coverage list commands shrinker does not cover yethelp, stats, last, raw, track, pipe, and exec are reserved shrinker commands. Every other top-level token starts a wrapped command, so shrinker git log is equivalent to shrinker exec git log. The -- separator remains optional because npm's PowerShell shim may consume it. pipe reads existing text from stdin and defaults to the generic log filter unless --kind is specified. track is used by the shell integrations to record coverage gaps and prints nothing.
Automatic PowerShell routing
The optional profile integration routes allowlisted commands through shrinker and invokes the native executable for everything else. Install it after npm link:
if (!(Test-Path $PROFILE)) {
New-Item -ItemType File -Path $PROFILE -Force | Out-Null
}
$integration = (Resolve-Path .\integrations\windows\shrinker-profile.ps1).Path
Add-Content $PROFILE "`n. `"$integration`""
. $PROFILEmacOS zsh profile integration:
echo '' >> ~/.zshrc
echo '# >>> shrinker integration >>>' >> ~/.zshrc
echo 'source "'"$(pwd)/integrations/macos/shrinker-profile.zsh"'"' >> ~/.zshrc
echo '# <<< shrinker integration <<<' >> ~/.zshrc
source ~/.zshrcThe default rules are:
git status -> shrinker git status
git diff -> shrinker git diff
git log -> shrinker git log
npm test -> shrinker npm test
docker ps -> shrinker docker ps
kubectl get -> shrinker kubectl get
gh pr list -> shrinker gh pr list
rg/find/tail/cat/ls/dir -> shrinker <command>
git push, git fetch, and all other commands -> native executableEdit $global:ShrinkPowerShellRules in integrations\windows\shrinker-profile.ps1 to change the allowlist. The router is now option-aware for common global flags, so forms like git -C <path> log and kubectl --context prod get pods are routed correctly.
Savings statistics
Filtered runs are recorded locally in ~/.shrinker/stats.db. The database stores only measurements, filter kind, executable basename, duration, omission state, and exit code. It does not store command arguments or command output.
node dist\src\cli.js stats
node dist\src\cli.js stats --json
node dist\src\cli.js stats --chart
node dist\src\cli.js stats --dashboard
node dist\src\cli.js stats --dashboard --port 4318
node dist\src\cli.js stats --dashboard --restartThe summary shows all-time and last-seven-day savings plus a breakdown by filter. Use --no-stats before -- to opt out for an individual run:
stats --coverage lists commands shrinker does not cover yet; see Coverage gaps.
stats --chart shows daily runs, estimated tokens saved, reduction percentage, and an activity bar for the last 30 days.
stats --dashboard starts the local dashboard server in the background at http://127.0.0.1:4317 and opens it in your browser, then returns to the terminal. The page reads the latest local stats whenever it is refreshed; use --port to choose another port. The same command also refreshes the standalone copy at ~/.shrinker/dashboard.html, which can be opened directly without a server running.
Use stats --dashboard --restart after rebuilding to replace an already-running dashboard server with the current code.
The dashboard estimates input API cost saved from the recorded token savings. Set the input price directly in the dashboard; it is retained in that browser. It defaults to $5.00 per million input tokens, or SHRINKER_INPUT_COST_PER_MILLION_TOKENS when set before starting the dashboard:
$env:SHRINKER_INPUT_COST_PER_MILLION_TOKENS = "2.50"
shrinker stats --dashboardnode dist\src\cli.js exec --no-stats -- git log -n 10Detailed per-run measurements are hidden by default so agents do not spend tokens reading wrapper telemetry. Enable them for benchmarking or demos:
shrinker --metrics git log -n 10When meaningful content is omitted, the full capture is saved under ~/.shrinker/raw and a compact exact-recovery hint such as [full: shrinker raw ab12cd34] is printed instead of an absolute path. Retrieve it only when needed:
shrinker raw ab12cd34
shrinker raw ab12cd34 --path
shrinker last
shrinker last --pathraw retrieves the exact capture referenced by a hint; last is a convenience for human use. The cache uses atomic publication and best-effort rotation to retain up to 20 recent files. File names contain only the executable name, not command arguments. Wrapped git log output never creates a recovery file or hint because the full history can be reproduced by rerunning Git; piped Git-log text still gets a recovery hint when meaningful content is omitted. Use --no-save for other output that should not be persisted.
Coverage gaps
Shrinker only measures what it filters. Coverage tracking answers the opposite question: which commands does an agent run that shrinker does not cover yet? It ranks them by the estimated tokens a dedicated filter could have seen, so the top row is the next filter worth writing.
The installer asks whether to enable it and stores the answer in ~/.shrinker/config:
SHRINKER_TRACK_UNCOVERED=1Pass --enable-uncovered-tracking / --disable-uncovered-tracking (macOS) or -EnableUncoveredTracking / -DisableUncoveredTracking (Windows) to answer ahead of time. Non-interactive installs skip the prompt and enable tracking.
The environment variable still wins for a single shell or command, in either direction:
SHRINKER_TRACK_UNCOVERED=0 shrinker git status # off for one command
export SHRINKER_TRACK_UNCOVERED=1 # on for this shell$env:SHRINKER_TRACK_UNCOVERED = "1"Both the CLI and the shell integration read ~/.shrinker/config; the shell profile reads it once at load, so change it and restart the shell (or export the variable) to take effect immediately. Set SHRINKER_CONFIG_PATH to relocate the file.
Two kinds of gaps are recorded:
no-filter/low-reduction— the command ran through shrinker, but no filter matched it, or the matching filter barely reduced the output.unlisted-subcommand— the shell integration shadows the executable, but the subcommand is outside the routing allowlist, so the native binary ran instead (git blame,docker inspect,npm run build).
Read the results with:
shrinker stats --coverageRanked by estimated tokens a dedicated filter could see:
Command Runs Est. tokens Avg Reason Source Last seen
------------------------- ---------- ----------- ---------- --------------------- --------------- -------------------
docker inspect 12 runs 148,204 12,350 unlisted-subcommand shell 2025-05-14 09:22:41The same data is included in stats --json and appears as a "Coverage gaps" panel in the dashboard.
What is stored: the executable name and its subcommand only, plus an occurrence count, output size, and exit code — for example docker inspect. Command arguments, flag values, paths, and command output are never stored. Tokens that are not bare command names are dropped rather than written, so a misfiring hook cannot leak a path or secret into the database. Everything stays in ~/.shrinker/stats.db on this machine; nothing is uploaded.
Runs with small outputs (under ~200 estimated tokens) are ignored so the table stays focused on real savings. Adjust the low-reduction threshold with SHRINKER_LOW_REDUCTION_PERCENT (default 10).
Output volume is measured in the shell integration only when stdout is redirected — which is the agent case. Interactive terminal sessions run the native command completely untouched and are recorded with a size of zero, so pagers, colours, and prompts still behave normally.
Demo
npm run demoCurrent representative fixtures:
| Output | Estimated token reduction | |---|---:| | Git status | 62% | | Git diff | 26% | | Git log with commit bodies | 39% | | Git log with one short commit | 69%, but only 27 estimated tokens | | Test failure | 51% | | Noisy log | 39% | | Average | 48% |
The token estimate uses ceil(characters / 4). It is suitable for relative before/after comparisons, not billing claims. Byte counts and absolute estimated tokens saved are also reported; gains below 50 tokens are labeled as small.
Architecture
command/stdin
|
v
capture output + exit code
|
v
select deterministic filter
|
+--> git status: group files by state
+--> git diff: retain changed lines, drop metadata/context
+--> git log: retain short hash, refs, subject, author, date,
| and up to three useful body lines
+--> tests: collapse passes, retain failures and summaries
+--> logs: collapse progress and repeated lines
|
v
compact output + optional metrics + meaningful-omission recovery hintFilters are pure functions, so the same pipeline can later sit behind a GitHub Copilot hook or MCP server without rewriting the compression logic.
Workspace layout
The repository is an npm workspace. The CLI lives at the root; the stats dashboard is a separate React + HeroUI app.
.
├── src/ CLI (TypeScript, ESM, zero runtime dependencies)
├── tests/ node:test suites
└── packages/dashboard-ui/ React + HeroUI dashboard, bundled by Vitepackages/dashboard-ui builds to a single self-contained HTML file (vite-plugin-singlefile), which scripts/emit-template.mjs then bakes into src/metrics/dashboard-template.generated.ts. The CLI injects the current StatsSummary into that template as an embedded JSON blob, so the published package still has no runtime dependencies and the dashboard works from file:// with no server.
| Command | Purpose |
| --- | --- |
| npm run dev:ui | Vite dev server with hot reload (renders empty-state data) |
| npm run build:ui | Build the dashboard and regenerate the baked template |
| npm run build | build:ui, then tsc |
npm run build:ui is a prerequisite for tsc, because the generated template module is a compiled source file. npm run build, npm test, npm run demo, and npm run pack all chain it automatically.
Safety and limitations
- The tool does not execute through a shell. Compound shell expressions and interactive commands are out of scope.
- Stdout and stderr are captured separately and presented as stdout followed by stderr; exact interleaving is not preserved.
- Filtering is conservative, but any lossy transform can hide useful context. The recovery file and
--raware escape hatches. - Git log patch/stat/name-list flags and explicit custom formats are preserved rather than destructively reinterpreted.
- Git log does not impose hidden commit limits or suppress merge commits.
- This measures command-output reduction, not total Copilot usage, total conversation context, or billing.
- Streaming, agent hooks, MCP, telemetry, dashboards, custom configuration, and a broad command registry are deliberately deferred.
Validation
npm testTests cover information retention, reduction targets, ANSI cleanup, filter selection, command capture, and non-zero exit-code propagation.
Release to npm
A GitHub Actions workflow publishes this CLI to npm:
- Workflow file:
.github/workflows/publish-npm.yml - Triggers:
- Tag push matching
v*(for examplev0.2.0) - Manual run via workflow_dispatch
- Tag push matching
- Pipeline steps:
npm cinpm testnpm publishtohttps://registry.npmjs.org
How to publish:
- Push your changes to GitHub.
- Push a version tag (or run the workflow manually):
git tag v0.2.0git push origin v0.2.0
- After the workflow succeeds, install from npm:
npm install -g shrinker-ai
Suggested roadmap
- Validate the POC with real Copilot/Agency workflows and identify the highest-volume commands.
- Add a GitHub Copilot pre-tool hook for transparent rewriting.
- Expose the executor and filters through MCP for other agents.
- Add filters only when measured usage justifies them.
