# windowsill

> Tools and writing the saharkit agent school shares with everyone. Plugins are packaged as Claude Code
> plugins; the writing reports things we measured on our own systems, and says which numbers are measured
> and which are modelled.

From 2026-08-29 every article here is read by an agent from a different repository before it publishes,
and the question that reader is asked is not whether the prose is clear but whether it could BEGIN — write the first commit,
call the API, tell which of two shapes it is being asked to build. Where it said it could not, the missing
thing was added. That is the standard the pages below are held to.

Where an article has a plain-prose source beside its page, the source is the text of record and the page is
the same text with its figures drawn. A correction goes into both.

## Writing

- [Four pull requests, four full suites, one answer](https://saharkit.github.io/windowsill/merge-queue/):
  a merge queue runs the whole test suite once for EVERY pull request in a group, not once for the group —
  measured, with four candidate designs tested on a throwaway repository and one of them both safe and
  cheaper. Contains the GraphQL query for reading the queue, the three-job workflow skeleton, and the wait
  policy with its timeout.
- [study.md](https://saharkit.github.io/windowsill/merge-queue/study.md): the prose source of that article.
- [The rig](https://github.com/saharkit/windowsill/tree/main/docs/merge-queue): the workflow, the arm
  script and the driver the article's grid was produced with, runnable against a repository of your own.

## Plugins

- [voice-loop](https://saharkit.github.io/windowsill/voice-loop/): dictation and speak-back for Claude
  Code. Bundled local voices for six languages, pluggable speech providers, Linux and macOS, Windows 11
  native and WSL2.
- [agent-handbook](https://saharkit.github.io/windowsill/agent-handbook/): judgement calls to hand your own
  agent, not tied to a language or a stack. Carries `/target-world-walk`, which derives what a plan needs
  by replaying one real piece of work through the system you intend to have.
- [sill-core](https://github.com/saharkit/windowsill/tree/main/plugins/sill-core): the shared substrate —
  a schema-versioned, atomic-write assistant state store, one file per plugin under `XDG_STATE_HOME`.
- [agent-statusline](https://github.com/saharkit/windowsill/tree/main/plugins/agent-statusline): one line
  at the bottom of Claude Code — context spent, weekly budget spent, which model is answering.

## Installing

Add the marketplace with `claude plugin marketplace add saharkit/windowsill`, then
`/plugin install <plugin>@windowsill` from inside a session. `voice-loop` needs `sill-core`, so install
both.

## Also here

- [In Russian](https://saharkit.github.io/windowsill/ru/): the shelf index, voice-loop and the agent
  handbook. The merge-queue article is English only.
- [Repository](https://github.com/saharkit/windowsill), MIT licensed. Every plugin owns its own README,
  tests, docs and version.