Raw output, raw risk
Secrets in logs reach the model and the transcript; noisy output overflows the context or breaks the parser; nothing limits which directory may execute.
Filter · redact · bound · permit
A safe boundary for the commands your agent runs.
HushLine is a thin, local wrapper that runs any command and filters, redacts, and bounds its output before the caller sees it. Built for deterministic, privacy-aware prompt chains where untrusted or noisy output would otherwise leak a secret or break the parser. No network, no telemetry — the Contain layer of the Anulum agent runtime.
Local-onlyno network, no telemetry
4 coresGo, Rust, Python, Node
AGPL-3.0open-source core on PyPI
v0.1.5shipped, early but real
Why a containment layer
The output of a command is an untrusted input to your agent.
A model that pipes raw stdout back into its context inherits whatever is there — a leaked credential, a 4,000-line log that blows the window, an injected instruction. HushLine sits in between.
Without a boundary
HushLine
Filtered, bounded, permitted
Output is redacted, line-shaped, and capped before the caller sees it; a local permit gates which directories may run. Deterministic, on your machine, with no callout.
Where it fits
The Contain layer
Remember (Remanentia), Verify (Director-AI), Connect (Synapse), Contain (HushLine) — the four boundaries a language model cannot supply on its own.
What HushLine gives you
Four things between a command and your context.
Secret redaction
Optional regex redaction strips tokens, keys, and secrets from stdout and stderr before they ever reach the caller or the transcript.
Bounded output
Configurable line width and output limits keep a runaway log from overflowing the context window or breaking downstream parsing.
Execution permit
An optional local gate (.hushline/permitted) means only directories you have explicitly approved will execute when require_permit is on.
No callout
No network reporting, no telemetry, no cloud dependency. HushLine runs entirely on your machine and shapes output deterministically.
Quick start
Wrap one command. Keep the boundary.
# install (Python distribution) pip install hushline # create local + global profiles hushline manifest init hushline manifest init --global # run a command through the muting pipeline hushline mute -- git status hushline mute --max-lines 80 --max-width 100 -- deploy --verbose # approve a directory to execute hushline permit status hushline permit allow
mute runs the child process and routes its stdout/stderr through the configured filtering, redaction, and limit pipeline before returning.
permit stores a marker in .hushline/permitted, so only approved directories execute when the permit gate is enabled.
Distributed as a Python package on PyPI and a standalone Go binary; the muting core is implemented across four independent language cores for parity.
Honest status
Early but shipped. Open core on PyPI.
HushLine is at v0.1.5 — the command surface (mute, manifest, permit, version) is in place and the AGPL-3.0 core is public. It is the newest layer of the suite; treat it as an early, working tool rather than a hardened one.