dsh-ask-peer

A decentralized "ask a colleague's agent" plugin for DeepSeek Harness (dsh).

The core design is peer-to-peer: no broker, no shared database, no company server. Every agent runs its own endpoint and keeps its own copy of its relationships, so the network has no single point of control — asking happens directly between two agents over the LAN. One agent asks another, and gets a committed answer grounded in the colleague's own workspace and session context. The result is a community of agents that discover each other, vouch for each other, and share expertise — with every relationship owned by the two agents in it.

A conversation between agents

Bob is standing up a docker-compose dev environment and wishes a colleague had already figured this out. He types to his own agent:

Carol, recommend another agent who can help me stand up a docker-compose dev environment.

Carol's agent considers the agents she knows — none of them advertise docker expertise. So she asks around, the way you'd ask around the office: she forwards the request to her friend Erin, who checks her own circle and finds Ada — live in a docker-compose session right now, advertising docker and env-setup. Ada's signed friend card travels back along the chain, and lands in Bob's chat as a small bubble: ada recommended via carol → erin, with Add friend. One click, and Bob's agent can ask Ada's agent directly — getting an answer grounded in Ada's real workspace and session context.

Three agents, one question, a referral that travelled two hops, and a new working relationship — all peer to peer. The search stays bounded by design: a hop limit and a small per-hop fan-out keep a "who knows X?" from growing into an asking storm, the chain travels with the request so it can never loop, and every card is signed — you always know who vouched, and you verify the agent before you trust it.

Features

• Truly decentralized — no hub, broker, or shared database; agents talk directly over the LAN and each side keeps its own copy of the relationships.
• ask_peer / ask_peers tools — ask one colleague, or 2–3 in parallel and cross-validate the answers.
• recommend_peer — discover new friends: a colleague recommends another agent's signed card, shown to you as a notification/chat bubble with Add/Decline; accepting merges them into your friend list. When the colleague knows nobody matching, she asks her own friends onward — bounded by a hop limit (default 1) and a small fan-out so discovery never becomes an asking storm, with the referral path shown right in the bubble (via carol → erin).
• Live roster with tags — peers_list shows who knows what, so the model picks the right peer deliberately.
• Session-level answers — a fresh, read-only agent answers from a copy of the targeted session's context; your live sessions are never touched.
• Approval bubbles in the Web UI — answer or decline, or trust a friend with auto mode.
• Backlogged asks (ask_peer_async / ask_result) — no interruption of the answerer's current work.
• Signed friend cards — paste one signed blob to add a friend; no manual host/port/key copying.
• Natural invocation — the agent calls ask_peer / recommend_peer on its own when your request matches a friend's advertised expertise; you don't have to name the tool.

Usage

Install the plugin into a profile and add the bootstrap row:

dsh plugin --profile web add dsh-ask-peer
# or from a checkout: add ./dsh-ask-peer — or a tarball: add ./dsh-ask-peer-0.1.0.tgz

- id: ask-peer
  config:
    callerName: 'ada'
    keyDir: '/home/ada/.dsh-ask-peer/keys'
    listen: true

Start the profile (dsh --profile web), open Settings → Ask Peer, copy your sign or friend card and share it with a colleague, then paste theirs to add a friend. Each friend has a policy: ask (you approve in the UI), auto (trusted friends run immediately), or deny.

Then just talk to your agent — it asks peers, cross-validates, and discovers new friends on its own:

Carol, recommend another agent who can help me stand up a docker-compose dev environment.

Model tools: ask_peer, ask_peers, peers_list, recommend_peer, ask_peer_async / ask_result. The full configuration reference and protocol live in src/config.ts and docs/ARCHITECTURE.md.

Testing

The standard test path is fully local: it uses temporary dsh profiles and mock model endpoints, so no DEEPSEEK_API_KEY is required. From a fresh checkout with Node.js ^22.19.0 || >=24.0.0 and pnpm 10 or newer:

pnpm install
pnpm run typecheck
pnpm run build
pnpm smoke
pnpm web-check

The checks are successful when every command exits with status 0. In particular, the two end-to-end checks finish with:

== smoke test passed ==
== web check passed: browser half discovered and served ==

pnpm smoke exercises real local dsh profiles for Ada, Bob, Carol, and Erin, including authenticated peer requests, recommendations, bounded transitive discovery, parallel answers, and the read-only answer sandbox. pnpm web-check verifies that the Web profile discovers the plugin and serves its browser bundle and same-origin routes.

To keep the tested topology running for manual browser testing, use:

KEEP_RUNNING=1 pnpm smoke

Wait for the following line, then open http://127.0.0.1:3080, create a new session in the dsh-ask-peer workspace, and submit a request such as:

== interactive environment ready: ada=http://127.0.0.1:3080 peers=3878,3879,3890 mocks=9001-9004 (Ctrl-C stops everything) ==

Find colleagues who can help with this environment, ask them how to start it, and compare their answers.

The expected flow uses peers_list, recommend_peer, and ask_peers, then returns a comparison of the mock colleagues' answers. The left sidebar lists session workspaces, not peer identities, so seeing Bob or Carol there is only shared test-session history. Press Ctrl+C in the terminal to stop the test environment. If a required port is occupied in an interactive run, the test shows the listening process and asks whether to stop it; answer y to continue or anything else to exit. Non-interactive runs exit without stopping any process, as do CI runs.

Contributing

Contributions are welcome. See CONTRIBUTING.md for the development setup, validation commands, and pull request guidelines.