jiruidai/dsh-meta-orchestrator
30Last commit Aug 14, 2026
dsh-meta-orchestrator DSH plugin
dsh-meta-orchestrator adds a plan-first protocol and five pattern playbooks to your DSH agent. The agent analyzes the request, selects a pattern (chain, parallel, router, supervisor, evaluation loop), writes a structured plan via orchestrate, and records it durably. The plugin never executes the plan—it only records, leaving execution to DSH's native primitives.
How to install the dsh-meta-orchestrator DSH plugin
dsh plugin --profile web add dsh-meta-orchestrator # from npmCopying does not run this command. Review the repository and version before installing the dsh-meta-orchestrator DSH plugin.
dsh-meta-orchestrator DSH plugin data source
dsh-meta-orchestrator DSH plugin snapshot date: Aug 16, 2026
discovered
What the dsh-meta-orchestrator DSH plugin can do
- Adds a ~330-word instruction block to the system prompt that teaches the plan-first protocol
- Provides five canonical workflow patterns: prompt-chaining, parallel-workers, router, supervisor, evaluation-loop
- Exposes three tools: orchestrate (record plan), adapt_workflow (revise plan), complete_workflow (close out)
- Plans persist across restarts in DSH's storage domain as an append-only chain
- No runtime overhead—no extra model calls, background processes, or custom session event types
Where the dsh-meta-orchestrator DSH plugin fits
- Audit a codebase, rank findings, then write a report—the agent plans a chain and executes step by step
- Research multiple topics in parallel, then synthesize a summary—uses parallel-workers pattern
- Classify support tickets and route to different resolution paths—uses router pattern
- Delegate subtasks to a supervisor agent and review results—uses supervisor pattern
- Draft, score, and iteratively improve an article—uses evaluation-loop pattern
Who the dsh-meta-orchestrator DSH plugin is for
- DSH users who run multi-step, multi-agent tasks and want explicit, versioned, and durable workflow plans
- Developers building complex agentic applications on DeepSeek Harness who need structured planning without hard-coding pipelines
dsh-meta-orchestrator DSH plugin limitations
- Requires a profile with storage stack (standard web profile; headless profile does not ship it)
- Only verified against DSH npm packages `^0.1.0-rc.5` and `0.1.0-rc.6`; prerelease semver range may break on future rc lines
- The plugin records plans only—execution must be done by DSH's own tools (subagent, todos, workflow scripts, plan mode)
- Tested on Windows 11 and Ubuntu only; other platforms may have unknown issues
dsh-meta-orchestrator DSH plugin: from the repository README
Quoted from the jiruidai/dsh-meta-orchestrator README, the upstream source of the dsh-meta-orchestrator DSH plugin. Copyright remains with the original authors.
A [DeepSeek Harness](https://github.com/deepseek-ai/deepseek-harness) plugin that teaches the agent to **synthesize a task-specific workflow at runtime** — instead of hard-coding pipelines or fixed agent topologies. For anyone running multi-step work in DSH who wants the plan the agent follows to be explicit, versioned, and durable rather than implicit. </div> --- ## 🧭 How it works Normally an agent just starts working on your request. With this plugin, it **plans first, then works the plan** — and the workflow is dynamic: written by the model itself, per request: 1. **Analyze** — the agent reads the request; if it is genuinely ambiguous, it asks before doing anything. 2. **Pick a pattern** — it chooses one of five proven ways to organize the work: ⛓️ steps in order (`prompt-chaining`), ⚡ fan out independent parts (`parallel-workers`), 🔀 classify, then dispatch (`router`), 🎯 delegate and review (`supervisor`), 🔁 draft → score → improve (`evaluation-loop`). Each pattern's detailed playbook is loaded only when picked. 3. **Write the plan down** — it calls `orchestrate` with the stages, the roles it will delegate, and verifiable success criteria. The plugin validates the stru
Read the full READMERepository license: MIT
dsh-meta-orchestrator DSH plugin questions
Does this plugin execute the workflow for me?
No, dsh-meta-orchestrator only records the plan. It provides three tools (orchestrate, adapt_workflow, complete_workflow) to save and version the workflow. The actual execution uses DSH's native primitives like subagent, todos, workflow scripts, and plan mode. The plugin is designed to be lightweight and non-invasive.
What profiles does this plugin work with?
It requires a profile that includes the storage stack. The standard web profile (from the `@deepseek-ai/dsh-web-app` bundle) ships this. The headless profile does not. You can check your profile's capabilities by running `dsh profile info` or referring to the compatibility section in the README.
Can I disable the plugin temporarily without uninstalling?
Yes. You can add a `disabled: true` entry in the profile's `cordis.patch.yml` file under the plugin's id. This takes effect on the next request, no restart or reinstall needed. Deleting those two lines re-enables it. Recorded workflows are untouched.
How do I know which pattern the agent chose?
The agent will typically announce the pattern in a sentence like "this is a chain: audit → rank → write." It then loads the pattern's playbook via a `skill` call, and finally calls `orchestrate` to record the plan. You can also inspect the stored workflow file at `$DSH_HOME/storages/meta_orchestrator.json` to see the full spec snapshot.
Will this plugin slow down my agent?
No. The plugin adds only a short instruction block (~330 words) to the system prompt and five pattern skills that are loaded only when the model selects a pattern. It does not make extra model calls, run background processes, or introduce any runtime overhead. The planning step is part of the agent's normal reasoning and does not add significant latency.