dsh-acp-interactive
用于 DeepSeek Harness (dsh) 的交互式 ACP (Agent Client Protocol) v1 插件：作为 dsh bundle 插件在 Zed 的 Agent Panel 中运行 dsh agents。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:dangpangch/dsh-acp#32e3ae87b61e42b32feed3b53a705eb039c85d79Acp Interactive 以 dsh-acp-interactive 发布，当前版本为 0.2.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
dsh-acp-interactive is, at heart, a dsh plugin: it supplies the ACP capabilities missing from dsh's built-in ACP, and serves the DeepSeek Harness to the Zed editor as a custom ACP agent-server extension (an interactive ACP v1 server).
Run DeepSeek Harness agents inside Zed's Agent Panel over the Agent Client Protocol (v1): create/close threads, stream text + reasoning, live tool cards, plan updates, slash commands and installed agent skills, session history, permissions, model/thought-level/preset selects, elicitation forms — while every tool runs inside the dsh sandbox with dsh's own model route.
dsh CLI (tested on 0.1.2-rc.1) — install globally first:
npm install -g @deepseek-ai/dsh
dsh --version # → 0.1.2-rc.1
pnpm (the dsh plugin command delegates to pnpm)
Zed with the Agent Panel (ACP v1)
A DeepSeek API key: DEEPSEEK_API_KEY env var, or configured once in dsh
Web (Models settings → writes ~/.dsh/.credentials.yaml)
The plugin is installed as a profile bundle. Pick a profile name (the
examples use acp); the boot command is then dsh --profile acp.
# HTTPS (public repository; use a credentialed URL for a private one)
dsh plugin --profile acp add https://github.com/dangpangch/dsh-acp.git
The repository ships its prebuilt bundle (lib/), so the install is a plain
fetch — no build scripts, no extra allowlist. Repeat add (or remove +
add) after pulling new commits to upgrade the installed copy.
dsh plugin --profile acp add /path/to/dsh-acp
This installs a pnpm link to the local checkout. After changing the source code, rebuild and the running profile picks it up on next boot:
pnpm build # tsdown -> lib/
Add a Custom Agent to Zed's settings.json
(~/.config/zed/settings.json on Linux; Cmd+, → "Open Zed Settings" from
the agent panel otherwise):
{
"agent_servers": {
"DeepSeek Harness (acp)": {
"type": "custom",
"command": "dsh",
"args": ["--profile", "acp"]
}
}
}
Notes:
command: "dsh" assumes dsh is on PATH (npm global install). If a
GUI-launched Zed cannot find it, start Zed from a terminal that has dsh
on PATH, or set command to the absolute path of the dsh binary.DeepSeek Harness (acp). The thread gear menu offers Model / Thought Level /
Write permission selects.DEEPSEEK_API_KEY is optional in agent_servers[].env — without it dsh uses
the credentials already stored by dsh Web.stdout must contain only JSON-RPC; EOF must exit 0:
printf '%s\n%s\n' \
'{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":1,"clientCapabilities":{}}}' \
'{"jsonrpc":"2.0","id":2,"method":"session/new","params":{"cwd":"/tmp","mcpServers":[]}}' |
dsh --profile acp
Expect two result frames (initialize → protocolVersion 1, session/new →
sessionId), then exit 0.
session/new · list · load · resume · close · delete with durable
history (session-query/persistence); load replays committed content per
ACP semantics.agent_message_chunk, streamed reasoning, tool cards
(execute cards carry the concrete command line in their title), plan/todo
updates, usage_update, available_commands_update slash catalog = dsh
command plane + user-invocable skills (~/.agents/skills,
<project>/.agents/skills, .dsh/skills). Skills follow the pi-acp naming
convention: announced as skill:<name> (/skill:find-skills in the /
popup), commands keep plain names; picking one loads the skill body through
dsh's tool-skill pre-step.clientCapabilities.terminal, each
bash/pwsh result is echoed through a client terminal
(terminal/create → embed {type:'terminal'} → terminal/release) so the
card shows the output like a native Zed terminal card (auto-expanded per
expand_terminal_card). dsh still runs the command under its own
sandbox/approval — the client terminal only displays the captured text.
Without the capability (or on any failure) the card falls back to plain
text content.session/request_permission (allow-once /
reject-once).authenticate via DEEPSEEK_API_KEY or dsh Web credentials;
AUTH_REQUIRED with a sign-in method when missing.ask_user_question → ACP form (when the client declares
elicitation.form).Honestly not implemented (never advertised): session fork, delegated
terminal/fs execution (tools stay in the dsh sandbox — the client terminal
above only displays already-captured output, it never runs the agent's
command), additionalDirectories,
audio/embeddedContext, MCP mounting (non-empty mcpServers is rejected with an
explanation), fine-grained diff cards, Windows.
pnpm install
pnpm typecheck # tsc --noEmit
pnpm build # tsdown -> lib/
pnpm test # vitest (89 tests incl. spawned frame-purity + history probes)
node scripts/history-probe.mjs # session history end-to-end (isolated DSH_HOME)
Layout: src/bridge/index.ts (plugin entry), catalog.ts (slash catalog),
replay.ts (history → ACP frames), tool-cards.ts (card titles/kinds),
{codec,updates,content,config-options,session-store}.ts (wire builders /
decision tables), src/dev-bin.ts (isolated dev/test boot),
cordis.patch.yml (bundle patch).
docs/design.zh.md