dsh-a2a
Agent2Agent (A2A) v1.0.1 dual-end plugin for DeepSeek Harness — English · 中文
@hanphone/dsh-a2a publishes DeepSeek Harness agents over A2A and brings
remote A2A agents into DSH as model tools. Both directions are multi-instance
and are managed entirely from the Settings UI: instances live in the a2a
storage domain, not in a config file.
What you get
- Inbound servers. Each instance has its own endpoint
(
/a2a/<id>), AgentCard, preset binding, auth token and enabled flag. Its
skills are the model-invocable entries of that preset's skill directory, so
there is no skill form to fill in.
- Outbound connections. Each connection has its own remote card URL, token,
timeout and optional preset. Every enabled remote skill becomes a model tool
named
a2a__<connection>__<skill>.
- A2A 1.0.1 JSON-RPC surface.
SendMessage, SendStreamingMessage,
GetTask, ListTasks, CancelTask, SubscribeToTask and
GetExtendedAgentCard, including SSE catch-up and the official task-state
and error vocabularies.
- Durable tasks. Tasks and context→session bindings survive restarts. Live
sessions and SSE subscriptions do not.
- Two executors.
session runs one turn on a per-context DSH session;
subagent delegates to ctx.subagents when the agent loop and that service
are mounted.
- Governed and observable. Inbound tasks pass through the
a2a/inbound-task waterfall, and the dashboard shows peers, sessions, tasks,
counters and an audit ring.
- Credentials, not secrets in records. The UI writes tokens through the
harness credentials service; records keep only the generated environment
variable name.
- Version and update check. The panel shows the installed version, the Node
runtime and each card's version, and can check npm/GitHub for a newer release.
Install
dsh plugin --profile web add @hanphone/dsh-a2a
Any profile name works. For a local tarball:
pnpm build
npm pack
dsh plugin --profile <name> add <path-to>/hanphone-dsh-a2a-<version>.tgz
Restart the web app after installing; the browser bundle is scanned at host
startup.
Quick start
- Restart DSH and open Settings → A2A 连接.
- Create an inbound server: pick a name, a preset and optionally a bearer
token. The endpoint and card are live immediately.
- Add an outbound connection: paste a remote AgentCard URL, click Import,
review the discovered skills, then Connect.
The getting-started guide walks through both
directions with curl examples. The exact payloads are in the
HTTP API reference.
Documentation
| Document | Answers |
|---|
| Getting started | How do I install it, publish an agent and call a remote one? |
| Concepts | What are instances, presets, skills, tasks, contexts and executors? |
| Operations | How do the panel and /a2a command work, and how do I configure storage and access? |
| HTTP API | What does the dashboard API expose, and what is the wire protocol? |
| Troubleshooting | Why is a reply empty, a connection failed or a tool missing? |
| Architecture | How are the host half, managers, stores and browser half put together? |
| Harness compatibility | How do I move the plugin to another DSH release? |
| Testing | Which test layer catches which breakage? |
The manual describes 0.5.1 against DeepSeek Harness 0.2.1-rc.3; ranges and
the verification record are in
harness compatibility.
Development
The plugin typechecks against a sibling harness checkout through TypeScript
project references, so a built ../deepseek-harness is required.
pnpm typecheck # host + client + compile-time mirror check
pnpm test # unit + composition + real-persistence tests
pnpm build # lib/index.js (host) + lib/client.js (browser)
pnpm test:e2e # two-instance A2A smoke against the built artifact
pnpm test:all # all of the above
Testing layers and their limits are described in Testing.
License
MIT
