dsh-crew

Why DSH Crew

DSH Crew is a plugin for DeepSeek Harness (DSH) — an open-source agent harness. It makes DSH agents dispatchable from Claude Code, Codex, Antigravity and Grok: the orchestrator keeps its own model, the work runs on a real DSH agent with that harness's tools, sandbox, presets and session history, and the host still shows it as a native subagent with live progress.

What runs the work is a DSH agent, not a bare model call. Tiers (flash / pro) select how much capability that agent gets from the harness's configured roster — DeepSeek V4 Flash and V4 Pro today — so a change of model in DSH needs no change here.

🧵 Native Progress UI

Workers appear as regular subagents in Claude Code / Codex / Antigravity / Grok — dispatch count, running step, tool calls and token usage all show up in the host's own task panel, plus a claude-hud statusline segment: ⚙dsh 1▶pro 2m14s 21.7k/606 ✓3.

🎚️ Tier Policy and Escalation

flash for mechanical work, pro for reasoning, effort from off to max. tier_policy can clamp every dispatch to one tier at the tool layer, and escalate_on_failure retries a failed flash run once on pro — based on evidence, not on guessing difficulty up front.

🏛️ In-Host DSH Sessions

With the bundle installed in a DSH profile, each worker is a first-class DSH session: visible in the Web UI, grouped by working directory, mounted with the Agent preset you choose per tier. Without DSH running, dispatch falls back to a standalone DSH runtime, so CI and headless environments still work.

👁️ Vision and Image Generation

DSH's models are text-only. describe_image now prefers DeepSeek's own VL model (deepseek-v4-flash-vision-exp) whenever a key is available, then falls back to the CLIs you already have — Claude, Codex, Grok, Antigravity — or any OpenAI-compatible API you configure. generate_image borrows the same CLIs' brush. Pasted images stay visible in the conversation and reach the model as text.

🛡️ Dispatch Guardrails

Every dispatch is checked before anything spawns. Worker→worker nesting is capped at origin-chain depth 3 and cycles are refused; a second worker on a workspace another job already holds is refused with the holder's info — never silently queued. Refusals are readable errors: wait or re-scope, don't bypass.

📋 Jobs Board

The DSH Crew panel doubles as a jobs board: every worker job — running or finished — is listed with tier, effort, live progress and tokens, held workspaces show their holders, and a job that vanishes mid-flight (e.g. a hub restart) is surfaced as an orphan ghost instead of disappearing silently.

🔌 Custom Providers

Bring your own endpoint (Base URL + API key + models) or a local command template. Each provider has a connectivity test that checks reachability and auth, then makes one real vision call so you find out now, not mid-task.

📦 One-Click Install

The settings page installs and updates the Claude Code plugin, the Codex role files and the Antigravity / Grok agents, skills and commands for you — marketplace registration, permission allowlist, HUD wiring, absolute paths rendered for this machine — and restores them just as easily. Every settings file is backed up first.

How it works

Claude Code / Codex / Antigravity / Grok (orchestrator, keeps its own model)
  └─ ds-flash / ds-pro  ← native subagent shell (progress shows in the host's task UI)
       └─ MCP: dsh_run_worker(tier, effort, cwd, worker=)
            ├─ worker="agy"/"grok" → that external CLI runs the task (explicit opt-in)
            ├─ hub reachable → session inside DSH (visible in the Web UI, grouped by cwd)
            └─ otherwise     → dsh-jsonrpc-agent runtime (worker.cordis.yml)
                 └─ DeepSeek V4 Flash / Pro (DSH SDK, event stream → progress and token stats)

One run, two views

Dispatch fans out. Below, eighteen workers translate this README in parallel: the host counts them as its own subagents, while the harness runs them as real sessions.

Install

Install into a DSH profile from npm:

dsh plugin --profile web add @zseven-w/dsh-crew@latest
dsh web

Or, for local development straight from the source tree:

dsh plugin --profile web add link:/path/to/dsh-crew
dsh web

The link: protocol symlinks the profile dependency to this repository, so rebuilds are visible immediately.

Configure DeepSeek credentials (standalone only)

In hub mode — the installation above — workers run inside the DSH instance and use the DeepSeek credentials it is already configured with. Nothing else to set up.

Only the standalone fallback needs a key of its own: dispatching from a host with no DSH instance running launches a worker runtime as a separate process. Obtain an API key from platform.deepseek.com and write it to ~/.config/dsh-crew/.env:

DEEPSEEK_API_KEY=sk-...

Verify

node scripts/smoke.mjs

The smoke test dispatches one cheap job through whichever path is available — the hub when a DSH instance is running, standalone otherwise — and prints which one it used. Within about ten seconds you should see smoke test passed — configuration OK. On failure the reason is printed, scoped to the path that was tested.

Then open Settings → DSH Crew and install the host integrations — Claude Code, Codex, Antigravity, Grok — with one click, or drive the same installer from the command line:

node src/install/cli.mjs claude   # Claude Code plugin: marketplace + permissions + HUD segment
node src/install/cli.mjs codex    # Codex agents + prompts
node src/install/cli.mjs agy      # Antigravity MCP config + agents + skills
node src/install/cli.mjs grok     # Grok MCP config + agents + commands
node src/install/cli.mjs all      # all four hosts at once
# uninstall symmetrically (uninstall-claude | uninstall-codex | uninstall-agy | uninstall-grok):
node src/install/cli.mjs uninstall-claude

Background and terminology

• DSH (DeepSeek Harness): DeepSeek's open-source agent harness, a code agent in Web UI form, similar to Claude Code but driving DeepSeek models.
• MCP (Model Context Protocol): Anthropic's AI tool integration protocol, enables LLMs to safely call external tools and data sources.
• Cordis bundle: DSH's plugin format; this project can run standalone as an MCP service or install into DSH Web as hub mode.
• tier: capability tier — which slot of DSH's configured model roster a worker gets. flash is fast and cheap (simple tasks), pro reasons harder (complex problems). Today they map to DeepSeek V4 Flash and V4 Pro; swap models in DSH and nothing changes here.
• worker: the DSH agent doing the work — a full session with its own tools, sandbox and preset, not a bare model call.
• effort: reasoning strength, off = no reasoning, high = high reasoning investment, max = maximum reasoning investment.

Claude Code

Installation

One-click installation (choose one):

• DSH settings page (when hub mode is installed): Settings → DSH Crew → "Install to Claude Code"
• Command line: node src/install/cli.mjs all

Both do the same thing: register local marketplace (parent directory dsh-plugins/ as marketplace root) + claude plugin install + MCP tool permission allowlist + claude-hud worker status segment config (auto-backup settings.json before changes, idempotent). Restart the session after installation for changes to take effect.

Usage

• Directly in conversation, say "dispatch X to ds-flash" or "dispatch X to ds-pro", and subagent executes the task
• Dispatch count and real-time progress shown in Claude Code task UI
• HUD status line segment: ⚙dsh 1▶pro 2m14s 21.7k/606 ✓3 (current tier / elapsed time / token usage / completion count)
  • For local development, statusline/statusline.sh or statusline/worker-segment.sh can be independently integrated
• Long-running tasks: CC has timeout limits on MCP calls (MCP_TOOL_TIMEOUT adjustable), long tasks can have orchestrator use dsh_spawn_worker + dsh_worker_result(wait_seconds) polling
• Local development and debugging: claude --plugin-dir /path/to/dsh-crew to temporarily load

Session commands

These override the global defaults for the current session only, and are enforced at the tool layer rather than by prompting:

Codex

Installation

Recommended to use the installer (auto-renders paths for this machine, copies the /dsh-config, /dsh-status and /dsh-playbook prompts):

node src/install/cli.mjs codex

Or manually copy (requires manual path modification after copying):

cp codex/agents/*.toml ~/.codex/agents/    # global or project-level .codex/agents/

Role files come pre-configured with:

• MCP server mounting configuration
• default_tools_approval_mode = "approve" (required, otherwise tool calls are auto-cancelled in exec mode)
• tool_timeout_sec = 3600

Note: When manually copying, absolute paths in the args field must be updated to match actual installation location; the installer handles this automatically.

Usage

• In interactive TUI, select "spawn ds-pro to ..." to dispatch tasks; Active/Done panels show progress
• codex exec mode can also directly call dsh_run_worker

Session commands

Three prompts are installed for Codex:

Antigravity (agy)

Installation

node src/install/cli.mjs agy

Registers the dsh-crew MCP server in ~/.gemini/config/mcp_config.json and installs the ds-flash / ds-pro agents plus the dsh-config, dsh-status and dsh-playbook skills into ~/.gemini/config/ (all files backed up first). Restart the session after installation.

Usage

• Pick ds-flash or ds-pro as the agent to dispatch tasks
• dsh_worker_config reads or overrides the session defaults

Session skills

Caveats

• agy runs workers with full approval (--dangerously-skip-permissions + accept-edits): agy 1.1.16 has no workspace-scoped permission mode, so a headless worker must auto-approve tool requests.

Uninstall: node src/install/cli.mjs uninstall-agy

Grok

Installation

node src/install/cli.mjs grok

Writes the [mcp_servers.dsh-crew] section into ~/.grok/config.toml and installs the ds-flash / ds-pro agents plus the /dsh-config, /dsh-status and /dsh-playbook commands into ~/.grok/ (all files backed up first).

Usage

• Pick ds-flash or ds-pro as the agent to dispatch tasks

Session commands

Caveats

• By security design, grok does not start repo-level MCP servers in untrusted project directories (grok mcp doctor reports "folder untrusted"); a global install is unaffected — change directory or pass --trust.
• grok workers run with bypassPermissions (always-approve, as the grok docs recommend for headless automation); deny rules and hooks still apply.

Uninstall: node src/install/cli.mjs uninstall-grok

MCP tools

Progress is simultaneously mirrored to ~/.config/dsh-crew/status.d/ (one shard file per writer, can be read by statusline / external monitoring).

Dispatch guardrails

Every dispatch is checked before anything spawns — refusals are readable errors, never silent queues:

• Origin chain: each dispatch appends a hop to the worker→worker origin chain. Nesting deeper than the cap (origin_depth_limit, default 3) is refused, and so is any cycle (the same backend + cwd appearing twice) — the guard that stops recursive worker self-amplification.
• cwd advisory lock: one running worker per workspace. A second dispatch into a held workspace is refused with the holder's job id, backend and start time — wait for it to settle, cancel it with dsh_worker_cancel, or pass allow_concurrent_cwd: true (read-only tasks only).

Dispatch playbook

How to dispatch well — flash vs pro, self-contained briefs, safe parallelism, verifying results, and the guardrails above — is bundled as a per-host playbook: /dsh-crew:playbook (Claude Code skill), /dsh-playbook (Codex prompt, Antigravity skill, Grok command).

Explicit CLI backends

worker="agy" / worker="grok" pins a dispatch to that external CLI (backend × model × effort) instead of the DSH tier logic. It is explicit opt-in — there is no default, so set it only when the user asks for that CLI. Caveats: grok refuses to start repo-local MCP servers in untrusted folders, and agy runs workers with full approval (no workspace-scoped permission mode).

Multimodal: vision and image generation

DeepSeek is a text-only model and does not support image input or generation. This plugin sources these capabilities externally through MCP tools:

Native vision first: when the vision provider is a built-in CLI (or explicitly native), describe_image first tries DeepSeek's own VL model deepseek-v4-flash-vision-exp (direct API call; key from DEEPSEEK_API_KEY or ~/.config/dsh-crew/.env). Any failure degrades gracefully to the CLI provider chain below, which is kept as the fallback. Image generation is untouched — the native model only ever looks at pictures.

Session image pasting: In DSH, switch model to DeepSeek (vision) ◉ to directly paste images. Images remain in session and display normally; the plugin appends transcribed text after them and strips images before sending—you see the image, the model reads the text. The transcription follows the same native-first ladder: DeepSeek's VL model when a key is available, then your configured CLI provider.

Configuration

In DSH settings page → DSH Crew → Multimodal (or directly edit ~/.config/dsh-crew/config.json):

Vision provider (image viewing):

• native / deepseek-native (DeepSeek's own VL model — tried first automatically for every built-in provider whenever a key is available)
• claude-code (default, uses haiku, inexpensive)
• codex (uses GPT, can specify specific model)
• grok (uses Grok)
• agy (Antigravity)
• custom (OpenAI-compatible API or local command)
• off (disabled)

Image generation provider (image generation):

• codex ($imagegen, gpt-image-2)
• agy (Nano Banana)
• grok (Imagine)
• custom (OpenAI-compatible API or local command)
• off (disabled)

Custom provider

Two integration methods:

API: Any OpenAI-compatible endpoint

• Fill Base URL, API Key, model list
• Vision uses /chat/completions with inline base64 images
• Image generation uses /images/generations
• Must specify "image generation model" to have generation capability, otherwise provider only appears in vision selection

CLI: Local command template, placeholders substituted with safe references

• Vision: {image} {question} {model} → stdout as answer
• Image generation: {prompt} {output} {size} → command must write file to {output}
• Fill at least one command; whichever is filled determines capability

Connectivity test: Each custom provider has a test button

• API: Check endpoint reachability, auth, send real vision request to verify
• CLI: Check executable file, run real command to verify
• Image generation: Validate config only, no actual image output

Borrowed subscription CLIs (claude / codex / grok / agy) require you to be logged in locally; the plugin won't bypass their permissions for you.

Hub mode

This package is also a valid DSH bundle (dsh.bundle + cordis.patch.yml). After installing into DSH Web profile with dsh plugin add dsh-crew:

• Worker sessions become first-class citizens: run as first-class sessions in DSH host (agents.create + per-session model/effort waterfall + default preset), appear in Web UI session list, can be opened anytime to view complete execution
• Organize by working directory: manage worker sessions by cwd in Web UI
• Loopback API:
  • POST/GET /_dsh/dsh-crew/jobs: spawn tasks, list, long-poll results, cancel
  • GET /_dsh/dsh-crew/ping: health check (MCP shim uses this to detect if hub is running)
  • POST /_dsh/dsh-crew/install: one-click install of the host integrations — Claude Code / Codex / Antigravity / Grok (backend of src/install/)
• Auto-detection: the hosts' MCP shim auto-detects hub (DSH_CREW_HUB env var, default http://127.0.0.1:3080)
  • DSH Web running → jobs enter hub mode (mode: "hub")
  • Not running → fall back to standalone runtime

Solution selection and limitations

Regular subscribers → shell subagent approach (recommended)

• Current state: Claude Code subagent shell uses haiku as intermediary; each dispatch adds hundreds to thousands of tokens
• Trade-off: Use small amount of Anthropic token in exchange for native task UI, real-time progress display, no extra configuration
• Recommendation: If you already subscribe to Claude Pro or use Claude Code, use this approach—convenient and transparent

Pay-as-you-go / CI environments → direct router approach

• Current state: Claude Code subagent frontmatter doesn't support direct third-party model connection; this repo's router experiment in scratchpad requires API-key credentials for Claude Code, but subscription OAuth is blocked upstream by Anthropic with 403
• Recommendation:
  • If using API-key credentials (not OAuth) and want to save Anthropic tokens, can run local router for direct DeepSeek connection
  • CI environments typically also use API keys; this approach is more economical (all DeepSeek tokens)
  • Requires self-testing of router integration (not officially supported)

Running DSH Web → hub mode auto-enabled

• Current state: If dsh plugin add dsh-crew installed into DSH Web profile, jobs run as first-class sessions in host, appear in Web UI session list
• Recommendation: During local development iteration, recommend enabling hub mode; worker progress can be fully observed in Web UI; for cross-machine collaboration or environments without Web UI, use the dispatching-host shell approach

Known items

• Codex role can theoretically try model_provider pointing directly to DeepSeek (unverified); this bridge doesn't depend on it
• Image generation output is flat bitmap; layer editing requires OpenPencil
• Runtime dependencies: Only @modelcontextprotocol/sdk and zod; @deepseek-ai/* are host runtime, provided by the DSH host (a plain npm install never pulls them)
• Codex must configure: default_tools_approval_mode = "approve", otherwise tool calls are auto-cancelled

Develop

pnpm install
node_modules/.bin/tsdown src/client/index.tsx --format cjs --platform browser \
  --target es2022 --tsconfig tsconfig.client.json --out-dir .client-build --clean
node scripts/build-client.mjs   # wraps the bundle for the DSH module loader
node scripts/smoke.mjs          # dispatches one real flash task end to end

Runtime dependencies are only @modelcontextprotocol/sdk and zod; every @deepseek-ai/* package is host runtime provided by the DSH host (documented in the package's dshHostRuntime field, not in peerDependencies, so a plain npm install never pulls them), which keeps the plugin inside the host's single module realm.

Ecosystem

• DSH Android — a live Android emulator or USB device inside the conversation, driven entirely through adb
• DSH iOS — a live iOS Simulator and a USB-connected iPhone, inside the conversation
• DSH Noema — long-term memory for DSH
• DSH OpenPencil — inspect and edit .op design documents inside a conversation

License

MIT