my-plugin

<div align="center"> <img src="./docs/reagent-mark.svg" alt="ReAgent logo" width="108" /> <h1>ReAgent</h1> Standalone local-first runtime and CLI for long-running research workflows. Use one always-on runtime for research, memory, channels, jobs, and durable delivery artifacts. <a href="./README.md">English</a>  |  <a href="./README.zh-CN.md">简体中文</a> <a href="./docs/index.md">Docs</a>  |  <a href="./ROADMAP.md">Roadmap</a>  |  <a href="./OPERATIONS.md">Operations</a>  |  <a href="./CONTRIBUTING.md">Contributing</a> </div>

At A Glance

reagent is the primary product surface.
The only official npm install target is @sinlair/reagent.
The public docs hub starts at docs/index.md.
One always-on runtime acts as the control plane for research, memory, channels, jobs, and delivery.
The CLI is the main control surface. The web console is a companion inspection surface.
Research work is durable: briefs, tasks, reports, reviews, presentations, workstreams, and module assets stay visible after one run.
Runtime configuration is managed locally: LLM routes, MCP servers, workspace skills, and inbound command policy all stay inspectable.
Hermes Agent is an explicit runtime-design reference, especially for seams such as hooks, tool orchestration, skills, jobs, and delegation.

What ReAgent Is

ReAgent is not a chat wrapper with a few extra research commands.

It is a standalone runtime and CLI for people who want:

structured research briefs instead of ad-hoc prompts
durable memory instead of chat-only context
visible long-running task state instead of hidden background work
reusable outputs instead of one-off answers
inspectable runtime seams instead of opaque orchestration

ReAgent is currently strongest for:

research task orchestration
file-backed workspace memory
evidence-backed reports
reusable delivery artifacts
local runtime and service control

Core Runtime Capabilities

The current runtime includes:

entry-aware toolsets for direct, UI, WeChat, and OpenClaw surfaces
a canonical agent runtime surface across root CLI (reagent agent ...) and HTTP (/api/agent/*)
a tool registry and tool execution pipeline
runtime hooks, tool policies, and built-in audit trails
structured session digests instead of relying only on recent chat turns
progressive workspace skill disclosure with optional reference files
explicit host-session linkage for OpenClaw-backed conversations
unified job runtime observability for scheduled discovery and memory compaction
research handoff workstreams for search, reading, and synthesis
bounded research-only delegation with durable delegation records and workstream artifacts

Operator-facing examples:

reagent runtime jobs
reagent agent runtime
reagent agent sessions
reagent agent session wechat:wx-user-1
reagent agent host sessions
reagent research discovery scheduler runtime
reagent memory scheduler runtime
reagent research handoff <taskId>
reagent research workstream <taskId> search

Current Product Areas

ReAgent currently spans five practical surfaces:

research: queue or run research tasks, manage direction profiles, inspect discovery plans and scheduler runs, record feedback, traverse the research memory graph, and reopen paper, repo, module, presentation, and direction-report artifacts.
memory: keep file-backed notes searchable and recallable, write durable entries, and run manual or scheduled compaction with visible policy and run history.
channels: receive UI or WeChat messages through mock, native, or openclaw providers, push outbound replies, and inspect lifecycle state, OpenClaw events, and cached sessions.
agent runtime: inspect canonical sessions, tune per-session role, skills, model route, fallback routes, reasoning effort, hook history, host linkage, and bounded delegations instead of treating every conversation as one shared opaque agent.
workspace control: manage LLM providers, MCP servers, workspace skills, and inbound command authorization from the CLI or JSON-backed config files.

Examples:

reagent research directions
reagent research graph report --view asset
reagent memory recall "recent research choices"
reagent agent session wx-user-1
reagent agent profile wechat:wx-user-1
reagent agent hooks wechat:wx-user-1
reagent agent delegate wx-user-1 search --task <taskId>
reagent models routes
reagent mcp list
reagent skills list
reagent commands policy

Canonical Agent Surface

ReAgent now exposes one canonical agent runtime surface:

root CLI: reagent agent ...
HTTP API: /api/agent/*
web console: Agents and Sessions panels

That surface covers:

runtime overview
canonical session list and detail
profile read/write (role, skills, model, fallbacks, reasoning)
runtime history and hook/audit inspection
OpenClaw host session list/history under agent host
bounded research-only delegation under agent delegate / agent delegates

Compatibility paths remain available:

reagent channels agent ...
top-level reagent sessions, reagent history, reagent watch
/api/channels/wechat/agent*

The intent is to keep existing operator flows working while moving new documentation and new tooling toward the canonical agent surface.

Product Shape

The main product loop is:

1. reagent onboard 2. reagent home 3. reagent service ... 4. reagent runtime ... 5. reagent research ... 6. reagent memory ... 7. optional web console inspection

The design principle is simple:

one standalone runtime
one CLI
many durable artifacts

Architecture

At the code level, ReAgent is a Fastify runtime that serves the web console, exposes HTTP APIs, and starts background schedulers for long-running work.

Structured research data is persisted through Prisma, with SQLite as the default local database. Workspace memory, channel traces, digests, and research artifacts remain file-backed so operators can inspect them directly.

The current provider stack combines fallback or OpenAI LLM routes, Crossref and arXiv discovery, PDF parsing for paper reading, and WeChat delivery through mock, native, or OpenClaw bridge modes.

Hermes Agent Context

Hermes Agent is not just an inspiration link in this repo. It is a concrete reference point for how ReAgent thinks about long-running agent runtime design.

The working position in this repository is:

Hermes Agent is closer to a general-purpose agent runtime.
ReAgent is closer to a research workspace / research operating system.
ReAgent should borrow Hermes-style runtime seams such as tool registry, runtime hooks, progressive skill disclosure, context compression, job runtime, and constrained delegation.
ReAgent should not flatten itself into "another Hermes" and lose its research-first artifact and memory model.

Those choices are reflected directly in the runtime, CLI, and repository structure rather than a separate design memo.

Get Started

ReAgent is designed to be installed and driven as a product, not as a repo-first demo.

Official install path:

npm install -g @sinlair/reagent

First-run path:

reagent onboard
reagent onboard --apply
reagent home
reagent service run

Default Web URL:

http://127.0.0.1:3000/

The starter profile uses fallback + mock. It is intended for first-run evaluation and walkthroughs, not as the final research-quality route.

Product Surfaces

ReAgent is meant to be used through a few stable product surfaces:

reagent home

the product overview and next-step entrypoint

reagent research ...

briefs, tasks, reports, workstreams, and reusable artifacts

reagent memory ...

durable notes and recall

reagent agent ...

canonical runtime sessions, hooks, host linkage, and delegation

reagent openclaw ...

host-facing OpenClaw inspection

reagent workspace ...

backup snapshots, restore preview or apply, and support bundles

Daily Use

Typical product commands:

reagent home
reagent runtime status
reagent agent runtime
reagent openclaw status
reagent research recent
reagent research tasks
reagent memory recall "recent research choices"
reagent workspace snapshot

Typical research flow:

reagent research directions
reagent research enqueue "multimodal web agents" --question "Which open-source baselines are strongest?"
reagent research tasks
reagent research report <taskId>
reagent research handoff <taskId>

Web Console

The Web console is the companion inspection surface.

Use it for:

first-run orientation
discovery review
report and artifact reading
recent artifact reopen
scheduler and digest configuration
runtime and session inspection

The CLI remains the primary control plane.

Recovery And Safety

ReAgent now treats recovery as a product surface, not just an internal maintenance concern.

Use:

reagent workspace snapshot
reagent workspace restore preview <snapshotPath>
reagent workspace restore apply <snapshotPath> --yes
reagent workspace support-bundle

These flows are designed to keep workspace state inspectable and recoverable without exposing raw secrets by default.

OpenClaw Host Surface

Use the dedicated OpenClaw entrypoint when you want bridge-facing host state:

reagent openclaw
reagent openclaw status
reagent openclaw sessions
reagent openclaw history <sessionKey>

Use reagent agent ... when you want canonical runtime state. Use reagent openclaw ... when you want host-facing bridge state.

Docs

Start here:

Product Positioning

ReAgent is a research-first workspace.

It borrows Hermes-style runtime seams such as:

hooks
audit trails
progressive skill disclosure
bounded delegation
explicit host/runtime inspection

It does not present itself as a finished general-purpose agent runtime. Its core product value is still durable research work, durable memory, and reusable artifacts.

Documentation

Chinese README: README.zh-CN.md
Changelog: CHANGELOG.md
Roadmap: ROADMAP.md
Security: SECURITY.md

Inspiration

ReAgent is informed by projects such as:

License

Released under the MIT License.

Summary

Install to Claude Code