lowfat-plugins

Community plugins for lowfat — each one compresses a tool's output to save LLM context.

Quickstart

Install lowfat 0.9.0 or newer first (how), then:

# Clone every plugin into lowfat's scan directory
git clone https://github.com/zdk/lowfat-plugins.git ~/.config/lowfat/plugins

# Use them — installed plugins are active immediately
lowfat terraform plan
lowfat helm status my-release

You'll see noticeably less output than running those commands directly. To wrap commands automatically, add to your shell config:

eval "$(lowfat shell-init zsh)"   # or bash

Available plugins

Savings are at the default full level; real-world numbers are usually higher (see caveats).

Cloud & infrastructure

Plugin	Wraps	Savings
aws-compact	aws	−38%
ansible-compact	ansible-playbook, ansible	−23%
helm-compact	helm	−51%
kubectl-compact	kubectl, k	−33%
terraform-compact	terraform, tf	−47%

Version control

Plugin	Wraps	Savings
git-compact †	git, g	−25%

Language ecosystems

Plugin	Wraps	Savings
cargo-compact	cargo	−51%
go-compact	go	−29%
npm-compact	npm	−62%
pytest-compact	pytest, py.test	−36%

Network & data

Plugin	Wraps	Savings
curl-compact	curl	−82%
psql-compact	psql	−41%

† git-compact is also bundled inside lowfat (covering status / log / diff / show). The community version adds network ops (push, pull, fetch, clone) and edit commands (add, commit, reset, restore, stash, tag).

Shorthands (k, tf, g) work even though they aren't real binaries — lowfat reads the plugin's bin field and routes lowfat k get pods to kubectl.

Trust: overriding bundled plugins

Installed plugins are active right away. trust matters in one case: when a community plugin overlaps a plugin bundled inside lowfat (docker, ls, find, grep, tree, git). Trust decides which wins for the shared commands.

lowfat plugin trust git-compact     # community version wins its overlaps
lowfat plugin untrust git-compact   # fall back to the bundled version

The choice is saved to ~/.config/lowfat/trusted.toml (one community-plugin name per line). Plugins that don't overlap a built-in — terraform-compact, helm-compact, kubectl-compact, … — work whether or not they're listed.

Inspect what lowfat sees:

lowfat plugin list           # all plugins (community + bundled)
lowfat plugin doctor         # same + parse/dependency check + rule count
lowfat plugin info <name>    # path, version, commands for one plugin

Intensity levels

lowfat level ultra   # most compression — verdicts only
lowfat level full    # default — keeps detail, strips noise
lowfat level lite    # mild — most output preserved
lowfat level         # show current level

Level	Best for	Typical savings
lite	Local dev where you want to see almost everything	−0–30%
full (default)	Day-to-day LLM context — strips noise, keeps signal	−20–60%
ultra	Agent loops where you just need "did it work?"	−70–95%

Override per-run with LOWFAT_LEVEL=ultra lowfat terraform plan.

Failure is always safe. Every plugin passes a failed command through raw, with if exit failed: raw or a rule that has exit: failed. So when a command fails you get the full error — compile errors, terraform diagnostics, pytest tracebacks — even at ultra.

Troubleshooting

Nothing happens when I run lowfat <cmd> — the plugin probably isn't in lowfat's scan dir. Check lowfat plugin doctor; if missing, clone or move it into ~/.config/lowfat/plugins/. Also note full passes many subcommands through nearly unchanged — try LOWFAT_LEVEL=ultra to see the filter work.

Two plugin directories — you have plugins in both ~/.lowfat and ~/.config/lowfat. Move everything to ~/.config/lowfat/plugins/ and delete the old ~/.lowfat/.

Wrong plugin runs — a community plugin overlaps a built-in (usually git-compact). Run lowfat plugin info git-compact, then trust/untrust to pick the winner (see Trust).

Preview a filter without running the command — pipe a sample through it; --explain prints per-stage diagnostics to stderr:

cat samples/helm-install-full.txt \
  | lowfat filter helm/helm-compact/filter.lf --sub=install --explain

About the savings

The table percentages come from lowfat plugin bench <name> against each plugin's samples/*.txt at full. They under-represent real savings because:

1. Samples are small (30–100 lines), so head N rarely fires. Unbounded git log or a full helm status see −60–90%.
2. Bench assumes exit 0, skipping the if exit failed: branches where filters do much of their work.
3. ultra is far more aggressive — collapsing a successful run to one status line.

Measure your own:

helm install my-app ./chart > /tmp/helm-real.txt
wc -c /tmp/helm-real.txt
cat /tmp/helm-real.txt | lowfat filter ~/.config/lowfat/plugins/helm/helm-compact/filter.lf --sub=install | wc -c

Writing your own plugin

Scaffold with lowfat plugin new <category>/<name>, or copy an existing folder. Each plugin is three things:

my-tool/my-tool-compact/
├── lowfat.toml     # name, version, commands it wraps
├── filter.lf       # the rules
└── samples/*.txt   # captured output for `lowfat plugin bench`

Rules

A filter.lf file is a list of named rules. lowfat runs the first rule whose fields all hold. Needs lowfat 0.9.0 or newer.

# A failed command keeps its full error output.
rule failed:
    exit: failed
    do: raw

rule build:
    sub: build|check
    do:
        match level:
            ultra:
                keep /(error|warning)/
                head 30
                or "mytool build: ok"
            else:
                drop /^Compiling /
                head 80

rule other:
    do: head 40

The full guide is in docs/DSL.md: fields, ops, levels, macros, JSON output and how to test.

Please keep the failure rule. Start every file with a rule that has exit: failed, so agents get the full error when things break. PRs that filter the failure path will be asked to add it.

Contributing

PRs welcome for:

• New plugins for uncovered tools (helm-diff, golangci-lint, pnpm, yarn, redis-cli, …)
• More samples for existing plugins — bigger/more diverse samples push bench numbers up
• Filter improvements — tighten regexes or head N budgets where full-level savings are weak

Run lowfat plugin bench <name> before and after your change, and include the numbers in the PR.