slim

A pluggable pipe filter that strips verbose CLI output (huge YAML/JSON dumps, lockfiles, long diffs) before it reaches the LLM, so an agent spends its context budget on signal, not noise. Lossless cleanups always-on; large dumps clamped to head+tail with a clear elision marker.

Agents rarely need a 10k-line kubectl get -o yaml, a full package-lock.json, or a 2,000-line diff to make the next decision. slim sits in the pipe, applies always-safe cleanups, and clamps genuinely huge dumps to head + tail with a clear elision marker the agent can act on (or re-run without slim for full fidelity). Pure Python standard library.

The executable is {baseDir}/bin/slim.

When to use this

Pipe a command through slim when its output is likely to be large and mostly noise — cluster dumps, dependency trees, lockfiles, long diffs, verbose install logs — and you only need the signal. Do NOT use it when you need exact, complete output (then run the command directly).

How to use it

text

# pipe mode — pass the command so slim can pick the right clamp
kubectl get pods -o yaml | {baseDir}/bin/slim --cmd "kubectl get pods -o yaml"

# exec wrapper
{baseDir}/bin/slim --report -- git log -p -8

# savings summary printed to stderr; filtered output to stdout
some-build | {baseDir}/bin/slim --report

--report prints a one-line savings summary to stderr; filtered output always goes to stdout, so slim stays composable in a pipe.

What it does

Always-on (lossless of signal): strips ANSI colour codes, trims trailing whitespace, collapses blank-line runs, collapses 3+ identical consecutive lines to ... (repeated Nx).

Opt-in clamp (lossy by design, per-command): unrecognised commands clamp only when > 250 lines (60 head + 30 tail); kubectl -o yaml/json clamps to 25 head + 10 tail; pip install logs drop download/progress noise.

Notes for the agent

• Reference the tool as {baseDir}/bin/slim — never hardcode a path.
• The clamp is lossy: if a clamped result hides what you needed, re-run the command without slim.
• Add a plugin by registering a Callable[[str], str] in slim/plugins.py.
• On Claude Code specifically, a PostToolUse hook cannot rewrite tool output — so the saving has to happen at the tool layer, i.e. piping through slim.
• Built by Workloft (https://workloft.ai/labs). Inspired by lowfat.