JujuLeaf Worker watches normalized Overleaf comment events through the JujuLeaf Bridge, dispatches selected @worker tasks to Codex or Kimi Code, and reports progress and a final summary back to the original comment thread.

What it does

• Comment-driven work. Start an Agent task without leaving the Overleaf review thread where the request originated.
• Explicit Agent policy. Choose Codex, Kimi Code, or an ordered list when the worker starts; only configured Agents can receive tasks.
• Visible progress. Keep collaborators informed with concise stage updates and a structured final result instead of hidden terminal output.
• Durable coordination. Deduplicate comment events, serialize Agent turns, recover interrupted tasks, and resume per-thread Agent sessions.
• JujuLeaf-native safety. Leave Overleaf authentication, project context, Skill installation, synchronization, and conflict checks to JujuLeaf.

Requirements

• Node.js 22.13 or newer
• JujuLeaf with Bridge protocol 1 support (v0.1.2 or newer) installed on PATH
• An authenticated JujuLeaf profile and an Overleaf project ID or JujuLeaf clone
• Codex and/or Kimi Code CLI installed and authenticated
• The JujuLeaf Skill installed for each enabled agent with jujuleaf skill install

JujuLeaf Worker does not install JujuLeaf, log in to Overleaf, or manage Agent Skills. Those remain JujuLeaf responsibilities.

Quick start

Run the worker from any directory and select the Overleaf project explicitly:

npx @jujuleaf/worker --project-id PROJECT_ID

No clone is required. The current directory is only the Agent workspace. JujuLeaf uses its default profile unless --profile is supplied.

The fully explicit form is:

mkdir -p agent-workspace
npx @jujuleaf/worker \
  --project-id PROJECT_ID \
  --profile overleaf \
  --workspace ./agent-workspace \
  --agent codex

--profile and --workspace are optional. For backward compatibility, you can also omit --project-id inside a JujuLeaf clone; JujuLeaf then discovers the project from that workspace.

The data flow stays deliberately small:

Overleaf comment → JujuLeaf Bridge → JujuLeaf Worker → Codex / Kimi Code
       ↑                                      │
       └──────── progress and result ─────────┘

The first snapshot processes recent unhandled mentions from open threads. Subsequent snapshots and events are deduplicated by Overleaf message ID.

Choose the Agent

Codex is the only enabled Agent by default, so plain @worker comments go to Codex. Choose Kimi Code instead when starting the worker:

npx @jujuleaf/worker --agent kimi

To enable per-comment routing, configure more than one Agent. The first Agent handles plain @worker comments:

npx @jujuleaf/worker --agent codex,kimi

Supported comment forms for a multi-Agent worker:

@worker Explain this derivation.
@worker ask Explain this derivation.
@worker codex suggest Rewrite this paragraph more concisely.
@worker kimi edit Fix this typo directly.
@worker-codex compile Diagnose the current LaTeX error.
@worker-kimi suggest Improve the academic prose.

@worker uses the first configured Agent. @worker-codex, @worker-kimi, or the first word after the mention selects one of the other enabled Agents. Selectors for Agents not enabled by --agent are ignored. Messages without an explicit action default to suggest.

The mention is configurable. For example, --mention @paperbot enables @paperbot, @paperbot-codex, and @paperbot-kimi.

Run diagnostics without starting the worker:

npx @jujuleaf/worker doctor --project-id PROJECT_ID --agent codex,kimi

Options

--project-id <id>            Overleaf project (otherwise discover from clone)
--profile <name>             JujuLeaf login profile (default: JujuLeaf default)
--workspace <path>           Existing Agent directory (default: current directory)
--mention <name>             Mention prefix (default: @worker)
--agent <names>              Agent(s), comma-separated; first handles @worker
                             (default: codex)
--protocol <number>          JujuLeaf Bridge protocol (default: 1)
--reconcile-interval <secs>  Authoritative snapshot interval (default: 60)
--bootstrap <mode>           recent, ignore, or all (default: recent)
--lookback-minutes <mins>    Recent bootstrap window (default: 30)
--status-interval <secs>     Minimum progress edit interval (default: 15)
--state <path>               SQLite state database path
--jujuleaf <path>            JujuLeaf executable (default: jujuleaf)
--codex <path>               Codex executable (default: codex)
--kimi <path>                Kimi Code executable (default: kimi)
-C, --cwd <path>             Alias for --workspace

Safety model

• JujuLeaf owns authentication, Overleaf protocol handling, synchronization, and Skill installation.
• JujuLeaf Bridge snapshots are authoritative; live events are wake-up hints.
• One worker handles one project, selected by --project-id or discovered from a JujuLeaf clone, and runs one Agent turn at a time.
• The worker never parses private Overleaf payloads.
• Progress comments contain factual stages only, never hidden reasoning or raw command output.
• Codex runs with workspace-write isolation and automatic approval review. Kimi Code runs in non-interactive prompt mode, whose tool calls are auto-approved by Kimi Code. --workspace is therefore the local Agent boundary; it is not implicitly cloned or synchronized.
• JujuLeaf still performs version, hash, and conflict checks.

The default comment action is suggest. Small, exact changes are submitted as JujuLeaf tracked suggestions so collaborators can accept or reject them. Direct remote edits require an explicit edit action. Larger multi-file review batches should run from a dedicated JujuLeaf clone.

Development

npm install
npm test

Maintainers can follow the release guide to bootstrap npm Trusted Publishing and publish later versions from Git tags.

Licensed under the MIT License.