Saktawdi/ha-orchestrator
30Last commit Aug 15, 2026
ha-orchestrator DSH plugin
HA Orchestrator is a DSH plugin that provides high-availability model failover and an orchestrate tool for splitting tasks into subagents. It supports parallel fanout, sequential pipeline, and supervisor review modes, with a settings UI and custom subagent definitions.
How to install the ha-orchestrator DSH plugin
dsh plugin --profile web add dsh-ha-orchestratorCopying does not run this command. Review the repository and version before installing the ha-orchestrator DSH plugin.
ha-orchestrator DSH plugin data source
ha-orchestrator DSH plugin snapshot date: Aug 16, 2026
discovered
What the ha-orchestrator DSH plugin can do
- Model failover: retries failed model calls on backup models with cooldown and retry budget.
- Automatic orchestration: model can call orchestrate tool for fanout, pipeline, or supervisor tasks.
- Custom subagents: define reusable subagents with name, provider, and system prompt in settings.
- AI generate subagent: fill in subagent definition from a one-sentence requirement.
- Slash commands: /ha and /orchestrate commands for inspecting and managing HA state and orchestration runs.
- Persistence: HA runtime state and orchestration run logs are persisted to disk and restored on startup.
Where the ha-orchestrator DSH plugin fits
- Research multiple open-source projects in parallel and compare results.
- Read a large project and summarize its architecture and progress by fanning out per-module reads.
- Run a multi-stage workflow (requirements → design → implementation) with pipeline orchestration.
- Generate a competitive analysis report and have a senior reviewer vet it via supervisor mode.
- Automatically recover from model failures during long-running tasks without manual intervention.
Who the ha-orchestrator DSH plugin is for
- Users who rely on DeepSeek Harness for complex, multi-step AI tasks.
- Developers or researchers who need fault-tolerant model execution and automatic task decomposition.
ha-orchestrator DSH plugin limitations
- Requires DeepSeek Harness with the web profile (no build step, no runtime dependencies).
- Slash commands depend on the DSH commands service; if unavailable, the plugin works but commands are not exposed.
- Auto-triggering of orchestrate may not work with certain persona presets (e.g., `minimal`/`minimal-v3`) that drop plugin system-prompt sections.
- The plugin is in early stage (v0.2.1); features may change or have limited stability.
- Configuration is written to the first writable location; restore behavior depends on filesystem persistence.
ha-orchestrator DSH plugin: from the repository README
Quoted from the Saktawdi/ha-orchestrator README, the upstream source of the ha-orchestrator DSH plugin. Copyright remains with the original authors.
HA Orchestrator is a plugin for [DeepSeek Harness](https://github.com/deepseek-ai/dsh) (dsh): - When a model call fails mid-run, it retries on a backup model and the run continues. - It adds an `orchestrate` tool that the model calls on its own when a task suits it, splitting work across subagents in parallel (`fanout`), in stages (`pipeline`), or with a review pass (`supervisor`). The settings page also lets you define custom subagents (or generate one with AI), and the UI and prompt copy are available in Chinese and English, following your DSH language. [简体中文](README.zh-CN.md) ## What it does ### Failover when a model fails - When a model request errors, it is retried on the next backup model. Backups are tried in order. - The failed model is temporarily skipped and cools down; it comes back on its own when the cooldown expires. - Each failure episode has a retry budget. Once it is spent, the plugin stops retrying instead of looping forever. - If a model error interrupts the run, the plugin restarts the task once so the work is not lost. Backup models, cooldown, failure threshold, and error-code filter are configurable in Settings → "HA 与编排". ### Orchestration, triggered
Read the full READMERepository license: MIT
ha-orchestrator DSH plugin questions
How do I install HA Orchestrator?
Run `dsh plugin --profile web add dsh-ha-orchestrator` in your terminal. The plugin is installed via npm and automatically patched into your DSH profile. No restart is needed; refresh the browser page to see the settings UI.
Does HA Orchestrator work with all DSH persona presets?
It works with most presets, but presets like `minimal` or `minimal-v3` that use `complete: true` drop plugin system-prompt sections. In that case, auto-triggering of orchestration relies only on the tool description. If it doesn't trigger, simply say "use orchestration" in your prompt.
How do I define a custom subagent?
Go to Settings → "HA 与编排" → Custom Subagents. You can add a subagent manually by filling in its name, provider/model, description, and system prompt. Alternatively, click "AI Generate" and describe what you need in one sentence; the current model will generate the full definition.
What happens if a model fails during a long task?
The plugin retries the failed call on the next backup model in order. The failed model is temporarily quarantined with a cooldown. If the retry budget is exhausted, the plugin stops retrying to avoid infinite loops. The task is restarted once to preserve progress.
Can I disable the automatic orchestration feature?
Yes. In Settings → "HA 与编排" → System card, turn off "context injection". The model will then only orchestrate when you explicitly ask, e.g., by saying "use the dsh-ha-orchestrator plugin" or "use orchestration".