DSH Multi-Agent DAG Plugin

English · 中文（简体）

Give your DeepSeek Harness agents the power to run complex multi-agent workflows, declaratively: break one big task into a graph of smaller agent subtasks, run the independent ones in parallel, and combine the results — with deterministic, code-enforced orchestration.

What it is

The DSH main agent is a single long-running turn. Fanning out to many agents by hand — some subtasks depending on others, results flowing back and forth — is error-prone and doesn't scale. DSH already ships imperative fan-out (subagent, workflow), but nothing that declaratively describes a dependency graph and lets the machine run it.

This plugin fills that gap. The model (or any caller) submits a small JSON description of the workflow — nodes, dependencies, success criteria — and the plugin takes care of the rest:

• it validates the graph up front and rejects bad proposals before anything runs,
• it schedules ready nodes in parallel under concurrency limits,
• it retries bounded failures with exponential backoff,
• and it fuses the successful results into one final answer.

You describe what the workflow is. The plugin handles how it runs.

Capabilities at a glance

How it works

DSH main agent
  │   the model calls the dag_run tool (or a plugin calls ctx.dag.start())
  ▼
dsh-dag      the Cordis plugin — tool, service, config, events
  │
  ▼
dag-core     the framework-free engine — model, validation, analysis,
             scheduler, state machine, retry, fusion
  │
  ▼
Parallel agent execution — concurrent child agents via DSH subagents

Every run follows one deterministic pipeline:

1. Compile & validate — the proposal is checked; an invalid graph returns fixable errors and no run is created.
2. Execute by wave — ready nodes run as concurrent child agents; each settlement releases (or blocks) its dependents.
3. Retry & fuse — bounded retries with validation feedback; successful results are fused into one final answer.

Quick start

1. Install the plugin into a DSH profile

dsh plugin --profile <name> add dsh-dag

Installing applies the plugin's bundle patch, which inserts the dsh-dag row into the host composition and provides the dag service on the host plane. Verify it landed:

dsh --profile <name> --dump-config | grep dsh-dag

2. Expose the tool to your agents

The dag_run tool is exposed per agent by composing a preset copy with an isolated dag realm (mirroring how the built-in delegation group isolates workflowEngine):

- id: dag-delegation
  name: cordis:group
  group: true
  isolate:
    dag: true
  config:
    - id: dsh-dag-tool
      name: 'dsh-dag'

3. Use it

Ask your agent for a multi-step task, or call the tool directly with a workflow like the one below.

Using the dag_run tool

Submit a TaskGraphProposal-shaped JSON:

{
  "schemaVersion": "1.0",
  "planId": "research_plan",          // ^[a-z][a-z0-9_-]{0,63}$
  "objective": "Research X and write a report",
  "nodes": [
    {
      "nodeId": "search",
      "title": "Search the literature",
      "prompt": "Find and summarize the top sources on X.",
      "capabilityRequirements": ["web"],
      "outputRequirements": ["A bullet list of sources"],
      "successCriteria": ["at least 3 sources"],
      "executorKind": "runtime",
      "toolLabels": ["web_search"],
      "dependsOn": [],
      "inputSources": []
    },
    {
      "nodeId": "draft",
      "title": "Draft the report",
      "prompt": "Write the report from the search results.",
      "dependsOn": ["search"],
      "inputSources": [{ "sourceNodeId": "search", "purpose": "use search results" }],
      "capabilityRequirements": ["general"],
      "outputRequirements": ["Markdown report"],
      "successCriteria": ["covers all sources"],
      "executorKind": "runtime"
    }
  ]
}

Invalid proposals never create a run — you get back structured, correctable errors instead.

Result envelope

{
  "runId": "dag_a1b2c3d4",
  "status": "completed",              // completed | partial | failed | cancelled
  "value": "…fused final answer…",
  "nodeCount": 2,
  "agentsStarted": 3,
  "failures": [{ "nodeId": "draft", "status": "failed", "error": "…" }]
}

Anything other than completed surfaces as a tool error — never as a silent success.

Programmatic entry

const run = ctx.dag.start({
  proposal,               // TaskGraphProposal
  parent: exec.agent,
  options: { idempotencyKey?, nodeTimeoutMs?, maxTotalNodes?, fusion?, globalLimit?, maxRetries? },
})
const outcome = await run.result
await run.dispose()

Configuration

Observability

Observe-only dag/* events are emitted on the host and (optionally) appended to the calling agent's session:

dag/run-start · dag/node-start · dag/node-end · dag/retry · dag/run-end

Payloads carry scalar facts only — never live handles.

Guarantees and limitations

What you can rely on

• Determinism is code-enforced: graph legality, state transitions, concurrency, retry caps and dependency propagation are all enforced by the engine — the model decides, the code enforces.
• Safe cancellation: run.cancel() / exec.signal aborts in-flight children; every child run is disposed.

Current limitations (V0.1)

• Foreground-only runs — no background start/poll, no journaling/resume.
• No budget/cost ledger; concurrency defaults conservatively to bound token spend.
• The plugin never decides single- vs multi-agent — the model calls dag_run only when it judges the task genuinely needs a DAG.
• Delegated children inherit DSH sub-agent policy (approval never, sandbox scope inherited); nodes that need approval-gated tools fail deterministically.

Uninstall

dsh plugin --profile <name> remove dsh-dag

Then remove the composed preset copy (e.g. the dag-delegation group above) from your agent preset and restart the session. Active runs are cancelled automatically on unload.

For developers

Quick orientation — the full engineering specification and migration plan lives in docs/DSH Multi-Agent DAG Plugin — Engineering Specification & Migration Plan.md.

packages/
├── dag-core/   dsh-dag-core  — framework-free orchestration engine (zero runtime deps)
└── dsh-dag/    dsh-dag   — the Cordis plugin adapter

npm install --ignore-scripts   # workspace install (typescript + vitest)
npm run build                  # tsc → packages/*/lib
npm run typecheck
npm test                       # 110 deterministic tests — no network, no real LLM

dag-core is a faithful TypeScript port of a proven Python orchestration middle layer, kept 1:1 for parity (component mapping in the spec above). The plugin adapter is deliberately thin: everything orchestration-related lives in dag-core, everything DSH-related lives in dsh-dag.

License

MIT