DSH Planner–Executor

简体中文 · Setup · Routing contract

Let a capable main model plan, design, decide and accept results. Delegate substantial implementation, research and testing to a suitable API model; handle simple tasks directly. Not DeepSeek-only. Existing installs keep their stable technical IDs.

• Practical model setup: create a custom OpenAI/Anthropic-compatible API provider, fetch or enter model IDs, declare per-model reasoning, and save keys through DSH’s official write-only credential service.
• Clear roles: select planner models and one default executor. Explicit child model/reasoning selections keep priority. Existing children are never migrated.
• Useful observation: actual models, running direct children, previous-turn results, native Open/Stop, collapse/hide.
• Cost-aware, not cost-guaranteed: concise worker scopes and acceptance evidence; prefer fresh workers to full-history forks. Delegation can reduce expensive-model execution, but may increase total tokens, latency or cost.

Preview

Screenshots use actual plugin components with public synthetic data, not private sessions or a paid model run. The strategy defaults OFF.

Install

Targets DSH APIs from 0.2.1-rc.2, reviewed against rc.3; declared range >=0.2.1-rc.2 <0.3.0 is not a certification of every build. Use your intended DSH installation’s CLI:

git clone https://github.com/O-OII/dsh-planner-executor.git
cd dsh-planner-executor
dsh plugin --profile desktop add --ignore-scripts .

Replace desktop for another profile, then refresh the existing DSH page. Disable old local activity/routing copies first. Nothing in this repository automatically installs, rewrites a live profile or restarts DSH.

Three steps

1. Open Settings → Plugins → DSH Planner–Executor. In Add custom model API, enter a new Provider ID, protocol, Base URL, optional key and at least one exact model ID. Fetch is an explicit endpoint request; manual model IDs always remain available. Existing/catalog providers are managed in Settings → Models.
2. After Host-confirmed creation, choose the model under Default executor, and add the exact models allowed to act as planners. Configure runtime reasoning only when supported. Save routing separately; provider creation never enables it.
3. Enable Planner–Executor when needed. Only future, route-omitted delegations from matching main agents use the default executor. Explicit selections, configured tool routes and worker descendants remain native.

Setup requires the official llm-pi-ai namespace, schema and LLM directory services; credential/discovery availability is checked separately. Missing services and read-only/overlay refusals are shown honestly. No copied official editor component, private navigation, localStorage switch or custom secret endpoint is used.

More

• Model setup, credential failures and independent saves
• Exact selection precedence, policy and limits
• Panel-only setup, migration and compatibility
• Reproducible tests and screenshot labels
• Contributing · Security · Changelog

MIT License · © 2026 O-OII contributors