hush

agent
Guvenlik Denetimi
Gecti
Health Gecti
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Community trust — 21 GitHub stars
Code Gecti
  • Code scan — Scanned 12 files during light audit, no dangerous patterns found
Permissions Gecti
  • Permissions — No dangerous permissions requested

Bu listing icin henuz AI raporu yok.

SUMMARY

Token-lean sessions at the harness level — a terse output style, output-shrinking hooks, and log compression cut both input and output tokens.

README.md
hush

hush

Shuts Claude up so your session stops costing money to read.

Every session in the benchmark suite drawn as a waveform, one spike per run, amplitude is words of play-by-play before the answer. The no-plugin lane spikes across most of the suite and peaks at 78 words, silent in 11 of 32 sessions. The hush lane is close to a flat line — silent in 23 of 32, with the remaining spikes on the rename task

This is what a session sounds like.

License: MIT Claude Code


What is this?

You've seen it: "Let me start by looking at the codebase." "Now I'll check the config." Four hundred lines of build output you didn't ask for, followed — eventually — by the one sentence you actually needed. Every word of that is billed. All of it.

hush doesn't ask Claude to "be more concise" and hope for the best. It trims the actual bulk — logs, command output, narration — at the source, as it happens, before any of it hits your bill.

It's built for real engineering sessions — the kind that read logs, run builds, and dig through output — because that's where the noise actually lives.

Why you'd want it

  • Cheaper sessions. It shrinks the two biggest sources of bulk — noisy output and narration — so long sessions cost less.
  • Easier to read. The answer sits at the top of one final message, not buried in a play-by-play.
  • Nothing important is lost. Failing command output, code, diffs, and security warnings are kept whole.
  • Zero setup. Install it and it's on. Tune it later only if you feel like it.

How it works

Four small habits, picked up the moment it's installed:

Moment What happens
Progress narration Swapped for one clean summary at the end, not a running commentary
Command output & log files Trimmed as it comes in — a short tail from a clean run, the whole thing from a failing one
Mid-turn rambling Caught by a running word count and cut off the moment it starts
Really large output (a huge log, a giant lockfile) Moved to a local file behind a short summary, so it's not re-sent in full every turn

That's the whole list. No workflow to learn, no dial to find first — it's just how Claude behaves now.

Install

Inside Claude Code, run:

/plugin marketplace add V-Songbird/foundry
/plugin install hush@foundry

Takes effect at your next session. There's nothing to invoke — hush just works in the background.

Benchmarks

We put hush up against plain Claude Code and a plugin that just tells Claude to talk less, on real engineering work — full agent sessions that explore, edit, and run code, not a single canned reply. Same jobs, phrased the way a developer actually types them, real cost read straight from the API.

Average bill across the benchmark suite: no plugin $0.226, a 'be brief' plugin $0.214, hush $0.169. hush takes $0.057 off the bill; asking Claude to be brief takes off $0.012

Being brief isn't enough. Asking Claude to talk less takes about a cent off the bill. hush takes off closer to six. Turns out asking politely and actually doing the work are two different things.

One average session itemised: what Claude read — files, logs, command output — $0.216; what Claude wrote back, the reply, $0.010; the session $0.226. A plugin that only shortens the reply is working on the $0.010

Almost the whole bill is what Claude reads, not what it writes back. A plugin that only shortens the reply is working on one cent of a twenty-three cent session. hush trims the noisy output and bulky logs before they hit your bill.

A multi-turn debugging session — triage an outage, dig a version out of a huge lockfile, write the handoff: no plugin $0.47, a 'be brief' plugin $0.42, hush $0.29. hush takes $0.18 off the bill

It shows most in longer sessions. Drag a huge file into a multi-turn conversation and that bulk gets re-sent every turn. hush keeps a tidy summary in the chat and the full copy one click away — that outage session came in at $0.29 against $0.47, where being brief barely moved it.

The same three-turn job — fix a currency bug, ask why the conversion order matters, write it up — with every reply drawn at actual size. A typical run with no plugin fills 497 words; a typical hush run fills 287, a visibly shorter column. Both fixed the bug and passed the same check

And you read about half as much. Across the suite the replies come to 242 words against 493. The answer lands in one message at the end — outcome first, one fact per line, instead of arriving in pieces while Claude works.

And it mostly says nothing until it's done. That's the waveform at the top of this page — every session in the suite, one spike per run. Plain Claude Code is already quiet about a third of the time, but when it isn't, it can run to 78 words of commentary first. hush is silent in 23 sessions out of 32. It isn't a gag order: Claude still speaks up to flag something you'd want to stop, or when it's blocked and needs an answer from you.

The full picture

Every job, every setup — the wins and the ties and losses. Cheapest per row in bold.

What Claude did no plugin "be brief" hush
Triage a production outage log $0.29 $0.29 $0.15
Multi-turn incident + write the handoff $0.47 $0.42 $0.29
Track a pool leak through two logs $0.25 $0.23 $0.19
Fix a failing test suite $0.20 $0.13 $0.15
Summarize a repo $0.12 $0.12 $0.13
Find the error in a noisy build $0.18 $0.23 $0.20
Answer a code question (no tools) $0.07 $0.07 $0.08
Average $0.23 $0.21 $0.17

Every job passed its correctness check in every setup — compression never bought a cheaper-but-wrong answer.

[!NOTE]
hush wins where there's noise to cut — logs, long sessions, debugging — and roughly ties on short or low-output jobs, where a session's fixed overhead dwarfs anything a plugin can trim. On a couple it costs a hair more; that's the honest shape, and it's why the average is what to read.

How we tested: the same jobs, three setups, several runs each in fresh throwaway workspaces, on Sonnet — a full multi-turn agent session every time, never a single generated reply — costs from the API, not estimates. Numbers move a few percent between runs. Reproduce it yourself — see benchmarks/.

Compress a memory file

/hush:hush-compress <path> shrinks a CLAUDE.md or notes file into a tighter form, so every future session that loads it costs a little less.

[!IMPORTANT]
It never touches your original — it writes a copy alongside it (CLAUDE.mdCLAUDE.hush.md) for you to review and swap in yourself.

Under the hood

Every check above runs locally as Claude works — read the plugin's files if you want the exact mechanics. Pairs naturally with razor: razor cuts the code and the cost, hush cuts the noise. Run both and neither notices the other — measured together, they add no overhead of their own.

Settings

Most people never touch these, but a few environment variables tune the caps or turn parts off:

Variable What it does
HUSH_DISABLE=1 Turns the hooks off
HUSH_CAP_PASS=60 Lines kept from successful command output
HUSH_CAP_FAIL=250 Lines kept from failing output
HUSH_NARRATION_BUDGET=120 Words of narration before a gentle nudge
HUSH_WRAP=1 Captures real exit codes from shell commands in permission-checked sessions — needs blanket Bash/PowerShell permission rules (no command pattern); sessions that bypass permissions get this automatically
HUSH_NOTE=off Skips the one-time per-session note that tells the model hush's compression markers are trusted tooling
HUSH_SUBAGENT=off Stops extending the terse-report style to spawned subagents
HUSH_ADAPTIVE=off Keeps compression caps fixed instead of tightening them in very long sessions
HUSH_SIDECAR=off Keeps very large outputs inline instead of moving them to a local file behind a digest
HUSH_SIDECAR_MIN=15000 Size (characters) at which an output moves to a sidecar file
HUSH_SIDECAR_SHELL_MAX=28000 Above this size, a command's output stays inline (past here the terminal already keeps its own copy)
HUSH_COMPACT=off Stops shaping the compaction summarizer's format instructions
HUSH_TEMPLATE=off Stops collapsing runs of same-shaped log lines (e.g. repeated worker/job entries) into one example line + a count
HUSH_DEBUG=1 Appends a per-decision JSON line to a temp-dir manifest file for every tool output hush looks at — for debugging and benchmark tooling, not everyday use

License

MIT — see LICENSE.

Yorumlar (0)

Sonuc bulunamadi