dsh-workflow-isolate

简体中文 | Architecture | Security model | Compatibility

dsh-workflow-isolate is a drop-in WorkflowEngine for DeepSeek Harness that runs model-written orchestration scripts in QuickJS/WASM. It preserves the DSH workflow hooks and lifecycle events while adding an independent JavaScript runtime, bounded guest memory and stack, interrupt fuel, a host wall deadline, forced worker termination, and child-agent budgets.

DeepSeek Harness's official worker-thread engine deliberately uses node:vm as an API-shaping mechanism. Its documentation states that it is not a security boundary and calls for a different engine when scripts are not trusted. This project explores that engine seam without changing the model-facing workflow tool.

[!IMPORTANT] QuickJS/WASM is a stronger language boundary than node:vm, not a claim of perfect isolation. The host-side subagent provider remains trusted and can use its configured model, tools, network, and credentials. Runtime vulnerabilities, side channels, supply-chain compromise, and host denial of service remain relevant. Read the security model before deploying it across trust boundaries.

Why this exists

Model-written workflow code sits in an unusual middle ground: it needs enough JavaScript to coordinate many agents, but it should not inherit Node.js authority merely because the harness is written in Node. dsh-workflow-isolate narrows that gap:

• A fresh QuickJS runtime and realm are created for every workflow run.
• The guest receives only args, agent, parallel, pipeline, phase, and log; no process, require, Node module loader, filesystem, network, or timers are injected.
• Only a plain-JSON projection crosses the guest boundary. Functions and symbols cannot cross; cycles, sparse arrays, non-finite numbers, and exotic prototypes are rejected.
• QuickJS heap, stack, and interrupt-fuel limits constrain script computation. A host wall deadline and worker termination cover code that does not cooperate with cancellation.
• Concurrency, total-agent, and per-combinator item caps bound orchestration fan-out.
• Child agents stay on the host and are reached only through a narrow RPC bridge to the configured DSH subagent provider.
• The engine preserves WorkflowRun, never-rejecting results, bounded disposal, and paired workflow/* lifecycle events expected by DSH consumers.

Architecture

flowchart LR
    T["DSH workflow tool"] --> E["IsolatedWorkflowEngine"]
    E --> H["Host run controller"]
    H --> W["Node worker thread"]
    W --> Q["Fresh QuickJS/WASM realm"]
    Q -->|"JSON child request"| H
    H -->|"trusted host RPC"| S["ctx.subagents"]
    S -->|"JSON result projection"| H
    H --> Q
    H --> O["workflow/* observers"]

The Node worker is a lifecycle and termination container. The QuickJS runtime inside it is the language boundary: guest constructors and prototypes belong to QuickJS, not V8, and no Node object is intentionally placed in the realm. See Architecture for the run sequence and Threat model for boundary assumptions.

Compatibility status

The current release targets @deepseek-ai/dsh-workflow@0.1.0-rc.7 and the associated DSH 0.1.0-rc.7 workflow packages. DSH's workflow API is still a release candidate, so compatibility is intentionally version-bounded.

The following surfaces are retained:

• ctx.workflowEngine service provider
• WorkflowStartRequest, WorkflowRun, and WorkflowResult
• agent, parallel, pipeline, phase, log, and args
• structured child output through DSH's supported object JSON Schema subset
• workflow/start, workflow/phase, workflow/log, workflow/agent-start, workflow/agent-end, and workflow/end
• per-run subagent-provider and total-agent policy overrides

QuickJS is not V8. Workflow bodies must use portable JavaScript and cannot depend on Node APIs, V8-only behavior, dynamic module loading, or ambient timers. Error text and stack formatting may also differ. The detailed behavior matrix is in Compatibility.

Benchmark

pnpm benchmark measures cold and warm fresh-runtime overhead and prints median/p95 JSON. The Node baseline is included only as a scale reference; it is not a security-equivalent engine. See the benchmark methodology.

Install from a checkout

Prerequisites: Node.js 22.19.x or 24.x, pnpm 11, a DSH 0.1.0-rc.7 installation, and a working spawn subagent provider.

git clone https://github.com/Linxiushen/dsh-workflow-isolate.git
cd dsh-workflow-isolate
corepack enable
pnpm install --frozen-lockfile
pnpm check
pnpm pack

Install the generated tarball into the DSH profile you use. A tarball includes built output and does not require a git dependency build allowance:

dsh plugin --profile web add ./dsh-workflow-isolate-0.1.0.tgz
dsh --profile web --dump-config

The bundled cordis.patch.yml disables the stock workflow-worker-thread row and inserts this engine. DSH permits one ctx.workflowEngine provider per context, so the two engines must not be mounted together.

For local source iteration, DSH also accepts a linked checkout after it has been built:

dsh plugin --profile web add .

Configuration

The bundle ships with a conservative deployment profile:

- id: workflow-worker-thread
  disabled: true

- insert:
    - id: workflow-isolate
      name: dsh-workflow-isolate
      config:
        provider: spawn
        memoryLimitBytes: 67108864
        maxInterruptTicks: 250000
        maxAgentRequestBytes: 1048576
        maxWallTimeMs: 600000
        maxConcurrentAgents: 0
        maxTotalAgents: 1000
        maxItemsPerCall: 4096
        disposeGraceMs: 3000

maxConcurrentAgents: 0 asks the engine to derive a bounded value from available CPU parallelism. A profile's own cordis.patch.yml is applied after bundle layers and can replace this row's config. Cordis row configs replace rather than deep-merge, so restate every value you need when overriding it.

Resource limits are deployment policy, not script options. Lower them for exposed or multi-user deployments, and remember that child-agent cost is primarily controlled by maxConcurrentAgents and maxTotalAgents, not QuickJS memory.

Millisecond timer settings are capped at Node.js's maximum single-delay value (2,147,483,647) so oversized values cannot be clamped to an immediate timeout.

All engine defaults are listed below. The bundle patch relies on static defaults for values it does not restate.

Workflow example

The model-facing tool still receives meta, args, and a plain JavaScript function body. A body can fan research questions out through structured child calls and then synthesize the surviving results:

phase("Research");

const findings = await pipeline(args.questions, async (question) =>
  agent("Investigate this question and cite concrete evidence: " + question, {
    label: question,
    schema: {
      type: "object",
      properties: {
        answer: { type: "string" },
        evidence: { type: "array", items: { type: "string" } },
      },
      required: ["answer", "evidence"],
      additionalProperties: false,
    },
  }),
);

phase("Synthesis");
const usable = findings.filter(Boolean);
const summary = await agent(
  "Synthesize these findings for " +
    args.audience +
    ":\n" +
    JSON.stringify(usable),
  { label: "Final synthesis" },
);

return { findings: usable, summary };

See the complete importable request fixture in examples/research-synthesis.mjs, including metadata and sample arguments.

Operational behavior

• Invalid metadata, script size/V8 function-body syntax, provider routes, arguments, and policy overrides fail before a run is published. QuickJS-specific compile failures settle the published run as an error.
• Once start() returns, run.result resolves rather than rejects. Completion, cancellation, and failures are represented by stopReason.
• Cancelling a run closes child admission, aborts pending provider starts, disposes published child runs, and interrupts QuickJS. The host terminates a worker that does not settle within the configured grace.
• Every emitted workflow/agent-start is paired once with workflow/agent-end, including host-synthesized cancellation ends after forced termination.
• Interrupt ticks are an implementation budget, not a stable cross-version instruction count. Tune them with representative workflows and re-baseline after QuickJS upgrades.
• The concrete IsolatedRun also exposes metrics: Promise<IsolateMetrics> with runtime, wall time, interrupt ticks, optional settlement-time memoryUsedBytes, and a termination classification. This is a project extension, not part of the upstream WorkflowRun interface.

Development

pnpm install --frozen-lockfile
pnpm check

pnpm check runs linting, TypeScript validation, tests, the production build, a real worker smoke against that build, and package-surface validation. CI covers the supported Node.js lines. Security-relevant changes should include adversarial tests for escape attempts, cancellation races, budget exhaustion, and lifecycle pairing.

See Contributing, Security policy, and the Changelog.

Project status

This is an independent, experimental provider for a release-candidate DeepSeek Harness seam. It is not an official DeepSeek project and has not been independently audited. The 0.x version line may track breaking changes in DSH until the upstream workflow API stabilizes.

License

MIT