hasan-aghayev/dsh-task-orchestrator ↗★ 0
dsh-task-orchestrator
基于依赖图的自动规划与角色分发插件 适合需要将复杂任务拆解并分发给多个协同子Agent的复杂工作流。
安裝
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:hasan-aghayev/dsh-task-orchestrator說明文件
閱讀完整 README ↗DSH Task Orchestrator
Automatic plan-first delegation for DeepSeek Harness. The plugin detects complex requests, creates a structured dependency graph, assigns focused roles, runs bounded workers, and sends the evidence to a final reviewer.
Author: Hasan Aghayev
License: MIT
Install
Install the public GitHub package into a DSH profile:
pnpm dsh plugin --profile web add https://github.com/hasan-aghayev/dsh-task-orchestrator.git
The package declares a dsh.bundle manifest in package.json, so the command can discover and apply cordis.patch.yml automatically. Its bundle enables the sandboxed workflow engine required by the orchestrator; it does not enable DSH's separate model-facing workflow tool. Remove it with:
pnpm dsh plugin --profile web remove dsh-task-orchestrator
Restart the profile after installation if it is already running.
What it does
The plugin adds the task_orchestrate tool and an automatic first-step planner. A simple request stays on the normal path. A complex request is scored with a deterministic detector and then sent through these stages:
- A planner returns a strict JSON plan with a summary, risk, roles, dependencies, read-only status, and declared write scopes.
- The scheduler starts a role only after its dependencies complete. Independent read-only roles may run in bounded parallel batches.
- Workers return structured evidence, changed files, tests, blockers, and next steps.
- A final reviewer compares the reports with the current workspace and returns
approved,changes_requested,blocked, orfailed.
Supported roles are researcher, architect, backend, frontend, tester, and documentation. The plugin uses the existing DSH subagent and workflow services and does not modify the agent loop.
Safe defaults
The default mode is hybrid:
- simple requests do not start child agents;
- complex read-only requests can run automatically;
- plans that require writes stop at
plan-onlyuntil a human approves them; - writes and parallel writes are disabled by default;
- planner, workers, and reviewer require a fresh structured-output subagent provider;
- planner and reviewer are included in the total-agent ceiling;
- oversized plans, reports, and parent notices are rejected or truncated at configured limits.
Automatic write execution should be enabled only in a profile that has its own approval and workspace policy:
config:
mode: hybrid
minComplexityScore: 55
subagentProvider: spawn
preferredWorkers: 3
maxWorkers: 6
maxTotalAgents: 8
maxConcurrentAgents: 3
allowWrites: false
allowParallelWrites: false
requireReview: true
The suggest mode always returns a plan first. off disables automatic planning but keeps the explicit tool. auto is available for deployments that intentionally permit automatic execution.
The task_orchestrate tool accepts objective, optional planOnly, optional executeWrites, and an optional maxWorkers cap. A request-side cap can never exceed the configured deployment ceiling.
Development
pnpm install
pnpm check
pnpm test
pnpm build
The package emits JavaScript to lib/ and declaration files to lib/types/. The published package contains only the built runtime, declarations, bundle patch, and paired READMEs.
Limitations
- Workers share the profile workspace. Parallel writes remain disabled by default and the plugin does not create automatic git worktrees.
- Complexity detection uses lexical signals and may miss a short difficult request or classify a long simple request as complex.
- The planner and reviewer are language-model agents. Strict schemas and bounded handoffs reduce malformed output but do not replace human review.
- Write scopes are declared to the scheduler and worker prompts; the surrounding DSH profile remains responsible for filesystem permissions and approval policy.
Repository
Source: