bazley13/dsh-multi-model-orchestrator
Makes the main agent split complex work across GLM, Kimi, Qwen or any OpenAI-compatible route, tracking tokens per model.
Turns the primary model into a dispatcher. A system-prompt section coaches it to decompose a complex task, inspect the routes that are actually available with list_subagent_models, and then send each subtask to the model best suited for it through subagent, naming an explicit provider and model. You describe each route's strengths in configuration, so the choice is made against your notes rather than against the model's guesses about vendors. Subagent model selection is bounded by an allowedModels allowlist, so a dispatch cannot wander onto a route you did not intend to pay for. It also tracks per-model token usage, which is the part that makes a mixed setup auditable: you can see which route is consuming the budget. todo_write is used to record the decomposed subtasks so the plan survives the dispatch.
Install
dsh plugin --profile web add dsh-multi-model-orchestratorThe npm line from the README's Install section. It is step one of two: the plugin only orchestrates routes that exist, so the README's quick start has you add your third-party model routes to $DSH_HOME/settings.yaml (default ~/.dsh/settings.yaml) afterwards, with GLM, Kimi and Qwen presets ready to paste from the repo's examples/settings.yaml.
Compatibility
Each route needs only an API key behind apiKeyEnv, which can be an environment variable, $DSH_HOME/.credentials.yaml, or the credentials UI. Third-party routes are registered as non-reasoning models by default, and reasoning flags can be added per model on the web Models page. Two limitations the README states rather than hides: some reasoning models reject a call without an explicit tier — Zhipu's GLM-5.3 line fails with default parameters and only answers when the sub-agent is dispatched with an explicit reasoning_effort such as low — and a manual link: install with a hand-made node_modules junction fails with ERR_MODULE_NOT_FOUND for the @deepseek-ai/* peers, because the plugin has to be a real dependency of the profile.
Details
- Repo: bazley13/dsh-multi-model-orchestrator
- Category: Agent Capabilities
- Stars: 4
- Version: 0.2.2 (npm)
Alternatives
linkbag/dsh-swarm-orchestrator · mario841859784/dsh-expert-orchestrator · cybtachyon/dsh-plugin-degen-investigator