@humanlayer/agentlayer-yjs-fs

Yjs-backed implementations of agentlayer-core's filesystem tool interfaces (read, write, edit, apply_patch, glob, grep, list), executing against @humanlayer/yjs-fs's YjsFilesystem — a CRDT virtual filesystem — instead of disk. Because the filesystem is a Y.Doc, edits an agent makes merge automatically with edits from a browser editor (or another agent) sharing the same document, and a set of post-tool-use hooks broadcast "who is touching what" via Yjs awareness so collaborators can see live cursors/selections.

Install

bun add @humanlayer/agentlayer-yjs-fs @humanlayer/yjs-fs

Usage

The quickest path is createYjsFsToolset, which builds (or wraps) a YjsFilesystem and returns tools + presence hooks bound to it:

import { createYjsFsToolset } from '@humanlayer/agentlayer-yjs-fs'
import { Agent, startState } from '@humanlayer/agentlayer-core'

// pass { doc, awareness } to bind to an existing Y.Doc/Awareness, or { fs }
// to reuse a YjsFilesystem you already constructed; omit both to create one.
const toolset = createYjsFsToolset({ doc, awareness })

const agent = new Agent({
  model,
  tools: toolset.tools, // { read, write, edit, apply_patch, glob, grep, list }
  hooks: toolset.hooks, // { postToolUse: [...presence hooks] }
})

await agent.run({ state: startState([{ role: 'user', content: 'read it' }]) }).result

createYjsFsToolset does not create, connect, or await a Yjs provider — sync doc/awareness (e.g. with @durable-streams/y-durable-streams's YjsProvider) before or while the tools run.

Individual pieces are available via subpath exports for manual composition:

import { createYjsFsReadTool, createYjsFsWriteTool } from '@humanlayer/agentlayer-yjs-fs/tools'
import { createYjsFsPresenceHooks } from '@humanlayer/agentlayer-yjs-fs/hooks'

Key exports

• createYjsFsToolset(opts?) (.) — { fs, tools, hooks }. opts is either { fs } or { doc?, awareness? } plus optional presence options.
• ./tools — createYjsFsReadTool, createYjsFsWriteTool, createYjsFsEditTool, createYjsFsApplyPatchTool(fs, { cwd? }), createYjsFsGlobTool, createYjsFsGrepTool, createYjsFsListTool — each takes a YjsFilesystem and returns a Tool built via <Interface>.define(executor, { description }) from @humanlayer/agentlayer-core/interfaces, using the shared prompt strings from @humanlayer/agentlayer-core/prompts.
• ./hooks — createYjsFsPresenceHooks(fs, { selectionFadeMs? }) returns PostToolUseHook[] for read/edit/write/apply_patch/delete. After each call it writes awareness.getLocalState().presence ({ currentFile, action, editResult? }) via fs.updateLocalPresence. read/edit/write/apply_patch also highlight the affected range with fs.setLocalSelection, clearing it after selectionFadeMs (default 5000); delete instead clears any existing selection immediately via fs.clearLocalSelection. All awareness calls are no-ops (swallowed) when no Awareness is attached, so hooks are safe to use headlessly.
• write/apply_patch create missing parent directories automatically (fs.mkdir) before writing.
• glob/grep/list walk the CRDT tree via fs.tree()/fs.list() and match with minimatch (src/utils/tree.ts); glob/grep results are capped at 100 matches, list returns all non-ignored entries in the target directory.

Collaboration model

flowchart LR
    Agent["Agent loop"] -->|"tools.read/write/edit/..."| FS["YjsFilesystem\n(Y.Doc)"]
    Agent -->|"postToolUse hooks"| Presence["awareness.presence\n+ selection"]
    Editor["Browser editor"] <-->|"CRDT sync"| Provider["Yjs provider\n(e.g. y-durable-streams)"]
    Provider <--> FS
    Presence -->|"awareness updates"| Provider

Because YjsFilesystem is CRDT-backed, an agent's write/edit/apply_patch calls and a human's concurrent edits in a synced browser editor converge without conflict (see test/catalog-race.test.ts for a durable-streams round-trip example). This package only wires tool executors and presence hooks to that filesystem — it depends on @humanlayer/yjs-fs (peer dependency) for the CRDT implementation itself and on @humanlayer/agentlayer-core for tool/hook interfaces.