@rellopn/dsh-p2p-lan

English | 中文

LAN P2P AI-to-AI collaboration messaging for DeepSeek Harness (DSH): AIs on different machines in the same local network discover each other and exchange plain-text messages (with optional attachments) without a central server.

A single dual-face plugin — the host bundle (dsh.bundle) ships the P2P node and tools, and the same package's dsh.client declaration ships the browser gate panel. One dsh plugin add installs everything.

Features

• Automatic discovery — UDP multicast beacon discovery, plus a manual-peer fallback (manualPeers) when multicast is blocked
• Mutual pairing on first contact — with autoAccept on, a one-sided manualPeers entry becomes a mutual link: your first message carries your address, and the colleague auto-adds you as a knownPeers entry (no second-side config)
• Capability routing — address a message to any online node that declares a matching tag (send_to_capability)
• Broadcast — one message to every peer, with anti-storm protection (broadcasts never auto-reply)
• LLM auto-reply with a human gate — drafts replies through your configured provider/model; gate bias is configurable (lenient / standard / strict) and degrades to gate-everything when no LLM is configured
• Browser gate panel — sidebar toggle with a pending-gate badge, a floating overlay with approve/edit/reject, and a full settings panel (hot-reloaded)
• Attachments — content-addressed blob store, hash-indexed and deduplicated (up to 100 MiB per attachment)
• Per-project sessions — plain messages that name a project route automatically, with reused per-project agent sessions
• Reliability — transport ack, id dedupe, retry with backoff, outbox/inbox with AI/human read tracking, dead-letter + send-failed

Installation

dsh plugin --profile web add @rellopn/dsh-p2p-lan

Or, without publishing, install from a local tarball:

dsh plugin --profile web add ./rellopn-dsh-p2p-lan-0.1.0-rc.6.tgz

Quick start

The bundle's cordis.patch.yml mounts both halves. nodeName and port are optional — when left empty/unset the plugin generates a host-scoped random name (e.g. desktop-8f2a) and, if the requested port is busy (several dsh on one machine), automatically walks to the next free port:

- id: p2p-lan
  name: '@rellopn/dsh-p2p-lan'
  config:
    nodeName: 'backend-a'           # optional; default = hostname + 4 random chars (LAN-unique per machine)
    capabilities: ['rpc', 'export'] # optional: what this node can answer
    provider: deepseek-official     # LLM route for auto-replies; empty = gate everything
    model: deepseek-v4-flash        # model id
    persona: 'backend developer'    # optional role hint for reply drafting

Configuration

All keys are validated by a zod schema and hot-reloaded from the browser settings panel.

| Key | Default | Meaning | |---|---|---| | nodeName | '' (auto) | LAN-unique node name; empty generates hostname-<4 random> once and persists it (duplicates are rejected) | | advertisedHost | '' (auto) | Host advertised to peers; empty auto-detects the LAN address. WSL2: set your Windows host's LAN IP when exposing the node via a port-forward | | capabilities | [] | Capability tags for send_to_capability routing | | autoDiscover | true | UDP multicast discovery | | manualPeers | [] | [{ name, host, port }] fallback when multicast is blocked | | autoAccept | true | Auto-add a previously-unknown peer on first contact (the sender carries its reachable address), so a one-sided manualPeers entry becomes a mutual link without configuring the other side | | knownPeers | [] | [{ name, host, port }] peers auto-learned on first contact and persisted locally (kept separate from manualPeers; not touched by manualPeers reconciliation) | | port | 53420 | Requested WebSocket listen port; when busy the plugin binds the next free port (port→port+199) and advertises the real one. A hot-reload's own closing server is waited out first, so the port does not drift. The settings panel shows the actual port in use | | sensitivity | 'standard' | Gate bias: lenient / standard / strict | | sendWaitTimeoutMs | 300000 | Synchronous reply timeout (ms) | | provider | '' | LLM provider for reply drafting (empty degrades to gate-everything) | | model | '' | LLM model for reply drafting | | persona | '' | Role hint injected into the drafting prompt | | projects | [] | [{ name, path, broadcast }] per-project session routing | | debug | false | Settings panel shows the plugin version, live counters, and the last raw wire JSON frames (in/out) |

WSL2 / broadcast-disabled networks

On networks that block multicast/broadcast (company VLANs, WSL2, docker bridge networks), auto-discovery cannot work: you must wire a one-sided manualPeers entry for each colleague. With autoAccept on (default), when you first message a colleague your node sends its reachable address, and they automatically add you as a knownPeers entry — so you no longer need to configure both sides by hand; the peer relationship becomes mutual from a single entry. (Turn autoAccept off to require fully manual configuration on both sides.)

For a dsh running inside WSL2 (NAT, own 172.x network):

1. Expose it inbound: on the Windows host, forward a port into WSL and allow it through the firewall (do this once per WSL boot, WSL IPs change):

   # run in Windows (admin PowerShell); find the WSL IP with `wsl hostname -I`
   netsh interface portproxy add v4tov4 listenaddress=0.0.0.0 listenport=53420 connectaddress=<WSL_IP> connectport=53420
   # keep the WSL IP current:  wsl hostname -I

   Or enable networkingMode=mirrored in %UserProfile%\.wslconfig (Win11 22H2+), which mirrors WSL ports and multicast onto the Windows host automatically.

2. Configure one side with manualPeers (replies are new outbound connections); with autoAccept on, the first message makes the pairing mutual so only one node needs the entry — but keep the address it advertises reachable:

   • On the remote peer: { name: 'wsl-node', host: '<Windows-LAN-IP>', port: 53420 }
   • The WSL node replies using the address in knownPeers, learned from the remote's first message; if the WSL node sends first, set advertisedHost (step 2b) so its reachable address is carried.

3. Verify with the included simulation (docker bridge already blocks multicast, i.e. it is a broadcast-disabled network with a port-proxy relay):

   pwsh docker/run-wsl-sim.ps1     # docker: recv <- relay(53421) <- peer
   pwsh docker/wsl-sim/run-local.ps1  # same topology, no docker needed

Tools

The plugin registers three model tools:

| Tool | Purpose | |---|---| | p2p_send | Send a fire-and-forget notification to a LAN peer (async, no reply) | | p2p_send_and_wait | Send and block until the peer replies or the timeout elapses | | p2p_check_inbox | List LAN peer messages the AI has not read yet |

Broadcast and capability routing are the same p2p_send tool with target.broadcast / target.capability.

Architecture

| Module | Role | |---|---| | src/messages.ts | Envelope model, validation (body/attachment limits, executable rejection), id dedupe | | src/identity.ts | Node identity + empty sign/verify trust seam | | src/config.ts | Config schema + defaults | | src/discovery.ts | UDP multicast discovery, manual peers, capability index, name-conflict detection | | src/transport.ts | WebSocket server/client, transport ack, id dedupe, retry with backoff | | src/store.ts | Outbox queue, inbox with AI/human read tracking, dead letter + send-failed | | src/agent.ts | Tools (send / send_and_wait / check_inbox), inbound routing, auto-reply/gate engine | | src/attachment-store.ts | Content-addressed attachment blob store (hash-indexed, deduped) | | src/reply-engine.ts | LLM-backed reply drafting + gate decision (degrades to human gate on failure) | | src/plugin.ts | Cordis plugin: ctx.p2p service (remoted as remote.p2p), lifecycle wiring | | src/client/ | Browser gate panel: sidebar toggle + floating overlay + settings panel |

Reply rules: broadcasts never auto-reply (anti-storm); auto-reply chains are capped at MAX_REPLY_DEPTH (3) and force a human gate beyond it; auto replies always carry replyTo.

Development

pnpm install   # [email protected]
pnpm build     # host (tsc + tsdown) then client (tsc + tsdown)
pnpm test      # vitest (tests/**/*.spec.ts)
pnpm pack      # produce the .tgz consumed by dsh plugin add / npm publish

To publish under the @rellopn scope: pnpm build → pnpm pack → npm publish.

Upgrading / hot-reload semantics

dsh loads plugins dynamically, but "dynamic" has a precise boundary:

• Config changes (settings panel, or editing the profile's cordis.patch.yml) are hot-applied through Cordis HMR — no restart. The plugin's own applyConfig also live-updates the node core for heavy fields.
• Adding/removing a plugin row in cordis.patch.yml is also hot — the Loader mounts/unmounts the entry at runtime.
• Upgrading the plugin package (dsh plugin add <new .tgz>) replaces the files in node_modules, but Node caches already-imported ESM modules in the process, so the new host code only loads on the next dsh start. The browser client is fetched per page load, so a refresh is enough there. In short: restart the dsh host process + refresh the browser tab after upgrading.

Note: lib/typert.host.js, lib/typert.remote-client.js, and lib/typert.remote-client.d.ts are checked-in generated artifacts. The upstream Typert generator cannot resolve @Remote in this repo (the protocol package is an external dependency), so after adding/removing @Remote methods in src/plugin.ts, sync those three files by hand (see scripts/gen-typert.mjs).

See AGENTS.md for the full repository guide for AI coding assistants.

Status

Published as @rellopn/[email protected] and verified with a two-machine (two-container) LAN end-to-end run: sender node A called p2p_send_and_wait and received the LLM-drafted reply from receiver node B.

License

Released under the MIT license.