Back to home

apheli0os

deepseek-harness-orchestrate

Declarative task-DAG orchestration plugin for DeepSeek Harness

Stars
0
Language
TypeScript
Created
Aug 15, 2026
Updated
Aug 15, 2026

Introduction

dsh-tool-orchestrate

English | 中文

Declarative task-DAG orchestration for DeepSeek Harness. This optional Cordis plugin registers a model-facing orchestrate_tasks tool, validates a bounded directed acyclic task graph before anything starts, and executes it through the existing ctx.workflowEngine.

This is an independent community plugin and is not an official DeepSeek package.

The package-owned worker script executes deterministic topological layers. Model-authored ids, prompts, dependency ids, schemas, provider names, model names, and upstream results are passed only as JSON data; none is interpolated into JavaScript source.

Install

dsh plugin --profile web add dsh-tool-orchestrate

Replace web with the profile you want to extend. The package declares a DSH bundle manifest, so installation adds its tool-orchestrate Cordis row automatically. This project distributes prebuilt releases through npm; installing directly from the GitHub repository is not supported.

The plugin expects a DeepSeek Harness deployment that already provides:

  • ctx.tools
  • ctx.systemPrompt
  • ctx.workflowEngine, normally supplied by @deepseek-ai/dsh-workflow-worker-thread
  • a subagent provider selected by the workflow engine

Manual Cordis configuration

Normally no manual configuration is needed after dsh plugin add. For a custom composition, add:

- id: tool-orchestrate
  name: 'dsh-tool-orchestrate'

The standard DSH base bundle already loads workflow-worker-thread. A custom composition must provide ctx.workflowEngine; if it disables the workflow engine, re-enable it explicitly in a later patch layer.

Model input

orchestrate_tasks accepts:

  • meta: normalized name and description;
  • tasks: a non-empty array of tasks with lowercase kebab-case id, title, and normalized prompt;
  • optional dependsOn, provider, model, and object-rooted outputSchema per task.

Each completed task returns its child's text or structured value. A failed child becomes { status: 'failed', error: 'subagent_failed' }; a task blocked by failed or skipped direct dependencies becomes skipped with deterministic blockedBy ids. Results keep the original task-array order, with aggregate counts and a top-level completed, partial, or failed status.

API key ownership

This plugin does not own or store an API key. Child tasks use the host deployment's existing LLM route and credential source: for example DEEPSEEK_API_KEY in the process environment, a host credential provider, or another configured route. Subagents inherit the same credential configuration as the parent composition unless a task selects a different provider/model route.

Do not place API keys in prompts, cordis.yml, source code, or GitHub Actions logs. Prefer environment variables or your deployment's credential mechanism.

Config

KeyDefaultMeaning
toolNameorchestrate_tasksModel-facing tool name.
maxTasks64Maximum tasks per call.
maxDependencies16Maximum direct dependencies per task.
maxPromptChars32768Maximum normalized characters per task prompt.
maxResultChars50000Rendered canonical-JSON ceiling.

All numeric limits must be positive safe integers; toolName must be a non-empty normalized string.

Development

pnpm install
pnpm run typecheck
pnpm run check

Publishing is performed from a clean checkout through the manually triggered Publish to npm GitHub Actions workflow. Maintainers must configure its npm environment and NPM_TOKEN secret first.