Tools left out on the windowsill.

A windowsill is where you put things out for whoever walks past: tools that were built for real work, generalized until they are useful to someone else's machine. Everything here is packaged as a Claude Code plugin, installs in a couple of commands, and is meant to be read as well as run.

What is on the sill

• agent-handbook

  Material you hand to your own agent — the general judgement calls, not tied to a language or a stack. It holds an instrument you reach for, /target-world-walk, which derives what a plan needs by replaying one real piece of work through the system you intend to have; and a standing rule you never invoke, the flagellant rule, which trades apology for fact-plus-fix and exhortations for stopping rules. Nothing to configure, no runtime, no network, no hardware.

• voice-loop

  Dictation and speak-back for Claude Code — hold a key and talk, and it reads the answers worth hearing aloud. Bundled local voices for English, Russian, Ukrainian, German, Spanish and French, with pluggable speech providers you can run local, self-hosted or in the cloud. Linux and macOS; native Windows 11 and WSL2 + WSLg on Windows 11.

• sill-core

  The shared substrate the plugins build on — a plugin-agnostic, schema-versioned, atomic-write assistant state store with remind-once etiquette. One file per plugin under XDG_STATE_HOME, separate from your config.

• agent-statusline

  One line at the bottom of Claude Code: how much of the context window is gone, how much of the weekly budget is spent, and which model is answering. Every figure comes from the payload the harness pipes to the script on stdin, so nothing is host-specific. Needs node and nothing else — no jq, no network, no hardware.

Every plugin owns its own README, its own tests, its own docs and its own version. More will land on the sill over time.

From 2026-08-29 every article here is read, before it publishes, by an agent from a different repository, and the question it is asked is not whether the prose is clear but whether it could begin. Where it said it could not, the missing thing was added. The merge-queue study is the first held to it. There is an agent-facing map of this shelf at llms.txt.

What we measured

• Four pull requests, four full suites, one answer

  A merge queue runs the whole test suite once for every pull request standing in it, and the last of them already contains the other three. We built a rig on a throwaway repository, tested four ways of not running the other three, and one of them is both safe and cheaper. The workflow, the queue API, the wait policy and the full grid are in docs/merge-queue, so you can run it against a repository of your own.

Add the marketplace

In your shell:

claude plugin marketplace add saharkit/windowsill

Then, from inside a running Claude Code session, install what you want:

/plugin install <plugin>@windowsill

voice-loop needs sill-core, so install both.