ACP adapter for DeepSeek Harness

Use DeepSeek Harness from Agent Client Protocol clients such as Zed.

@openma/deepseek-harness-acp is a dsh profile plugin (a harness bundle) and a standalone stdio ACP server. In both shapes it maps the harness session-event log onto the full ACP update vocabulary and reuses your existing dsh setup — including the API key you saved in the dsh Web UI. No credentials in your editor config.

Install (recommended: as a dsh plugin)

1. Have DeepSeek Harness (you probably already do):

npm install -g @deepseek-ai/dsh
dsh web        # first run: save your DeepSeek API key in Settings → Models

2. Add the ACP bundle to a profile:

dsh plugin --profile acp add -w @openma/deepseek-harness-acp

This creates $DSH_HOME/profiles/acp, installs the package, and registers the bundle (its dsh.bundle patch mounts the bridge over @deepseek-ai/dsh-base — the same product baseline as dsh web, with the module-reload watcher off).

3. Point Zed at it (settings.json):

{
  "agent_servers": {
    "DeepSeek Harness": {
      "command": "dsh",
      "args": ["--profile", "acp"]
    }
  }
}

That's it — no env, no keys in the editor. Credentials come from the harness's own credential store ($DSH_HOME/.credentials.yaml, the file the Web UI writes, hot-reloaded), with the process environment (DEEPSEEK_API_KEY / DEEPSEEK_BASE_URL) as a fallback layer.

Because the profile rides dsh-base, the agent in your editor is the full product: sandboxed bash and filesystem tools, todo plans, skills, subagents, workflows, web search, plan mode, LLM session titles, compaction — and it shares $DSH_HOME/sessions, so conversations started in the dsh Web UI can be listed and loaded from the editor.

MCP servers configured in your client (Zed context servers, for example) are attached per session: each session/new server entry mounts a @deepseek-ai/dsh-mcp-client instance (stdio and streamable HTTP transports), whose tools join the agent as mcp__<server>__<tool>. A server that fails to start never takes the session down.

Override provider/model per profile in $DSH_HOME/profiles/acp/cordis.patch.yml (id-targeted patch on the acp-bridge row), or via DSH_MODEL / DSH_PERMISSION_MODE in the launch environment.

Install (alternative: standalone server)

The package is also a self-contained ACP server that attaches to a DeepSeek Harness installation — the way codex-acp runs the Codex you point it at:

npm install -g @openma/deepseek-harness-acp
dsh-acp --help

{
  "agent_servers": {
    "DeepSeek Harness": {
      "command": "dsh-acp",
      "args": []               // optional: ["--dsh-path", "/path/to/dsh"]
    }
  }
}

It finds the harness via --dsh-path/DSH_PATH, its own tree, ./node_modules, dsh on PATH, or npm root -g — and composes a fixed coding-agent tree from it (spine, sandboxed bash/fs, todo, compaction). Use this mode when you want an ACP server without creating a profile; use the profile mode when you want your full dsh composition in the editor.

Features

• Streamed assistant text and reasoning (agent_message_chunk / agent_thought_chunk), with assembled-message fallback when an adapter emits no deltas.
• Tool calls with ACP kinds, human titles, file locations, raw input/output — and real file diffs sourced from the fs tool's hunk metadata.
• todo_write snapshots as ACP plans; token accounting as usage_update (context pressure) plus per-turn PromptResponse.usage.
• Real cancellation: session/cancel interrupts the live turn through the harness agent, not by killing a process.
• Permission requests: sandboxed wider-access retries surface as session/request_permission with allow-once / always-allow / reject options.
• Session modes mapped to the harness sandbox policy: read-only, workspace-write, danger-full-access — switchable per session at runtime.
• Model switching through session config options (session/set_config_option), preserving full conversation history via durable resume.
• session/load with complete history replay from JSONL persistence, session/list from the same store, session titles as session_info_update.
• Slash commands (/status) advertised through available_commands_update.
• Env-var auth method advertisement (DEEPSEEK_API_KEY), honored by authenticate.

Configuration

Flags win over environment variables, which win over defaults.

Flag	Env	Default	Purpose
--dsh-path	DSH_PATH	auto-detect	DeepSeek Harness installation
--provider	DSH_PROVIDER	deepseek-official	Provider route for created agents
--model	DSH_MODEL	deepseek-v4-flash	Default model
--models	DSH_ACP_MODELS	deepseek-v4-flash,deepseek-v4-pro	Selectable models for the session Model option
--max-tokens	DSH_MAX_TOKENS	provider default	Per-request output-token cap
--permission-mode	DSH_PERMISSION_MODE	workspace-write	Initial sandbox mode (read-only / workspace-write / danger-full-access)
--session-root	DSH_SESSION_ROOT	~/.dsh-acp/sessions	JSONL session store
--persona	DSH_SYSTEM_PROMPT	built-in coding persona	System-prompt persona ({{model}}, {{cwd}} interpolate)
--reasoning-effort	DSH_REASONING_EFFORT	high	off / high / max
--no-thinking	—	thinking on	Disable model thinking output
--bash-timeout	DSH_BASH_TIMEOUT_MS	60000	Foreground bash timeout (ms)
—	DEEPSEEK_API_KEY	—	API credential (advertised as the ACP auth method)
—	DEEPSEEK_BASE_URL	DeepSeek endpoint	OpenAI-compatible endpoint override
—	DSH_ACP_DEBUG	off	Verbose stderr diagnostics

Permissions and sandboxing

Sessions start in workspace-write: bash and file mutations are confined to the session's cwd (plus shared temp roots), and a model retry requesting wider access raises an ACP permission request. Choosing Always allow (this session) flips the harness approval policy to never for that session. danger-full-access disables both the sandbox and the prompts — use it only in disposable checkouts or containers.

Architecture

ACP client (Zed, …)
   │  ACP JSON-RPC over stdio
   ▼
dsh-acp
   ├─ src/harness.ts          host discovery + module loading (DSH_PATH / auto-detect)
   ├─ src/app.ts              composition built from the host's packages
   ├─ src/bridge/             the ACP bridge (cordis plugin)
   │    ├─ index.ts           sessions, prompts, cancel, modes, options, permissions
   │    ├─ translate.ts       session-event → ACP update projection (pure)
   │    ├─ history.ts         stored-log replay for session/load (pure)
   │    └─ prompt.ts          ACP prompt blocks → harness content blocks (pure)
   ▼
your @deepseek-ai/dsh installation   (agent spine, llm-deepseek, persistence,
                                      sandbox, bash, fs, approvals, todo,
                                      token meter, compaction)

The bridge consumes the harness session/event firehose (the same append-only log that persistence stores), so live streaming, history replay, and session/list all agree by construction. All harness modules — including cordis itself — load from the host tree, so plugin and service identity is never split across copies.

Development

npm install         # dev deps include the harness packages (types + tests)
npm run typecheck   # tsc --noEmit
npm test            # vitest: unit + e2e smoke (boots the real composition; no model calls)
npm run build       # esbuild → dist/index.js
npm run pack:local  # build + npm pack

The e2e suite exercises initialize, session/new, modes, model switching, /status, session/list, and cross-process session/load — all without a model credential. To also test against a standalone host install:

npm install --prefix /tmp/dsh-host @deepseek-ai/dsh
DSH_ACP_TEST_HOST=/tmp/dsh-host npm test

License

Apache-2.0.