dsh-subagent-model: Enforce Subagent Model and Reasoning Effort
This plugin is designed to resolve the issue in DeepSeek Harness where subagents inherit the parent route, leading to unpredictable model usage. It provides a minimal forced-configuration approach that directly specifies the model and reasoning effort when invoking subagents, ensuring controlled execution. Core features include enforcing model selection, validating reasoning effort values, joining the current turn rather than waiting for the next one, and displaying the route in logs. Users can configure the allowlist and effort level via the UI or YAML. After installation, replace the delegation line in the preset and set `maxDepth` and `toolFilter` to prevent deep delegation.
Read Moredsh-subagent-model: Independent Model and Preset Configuration for DSH Subagents
The DeepSeek Harness plugin is designed to address the resource overhead caused by a main agent invoking subagents that inherit the main model. It adds a "Subagent Model" menu, allowing users to uniformly configure independent providers, models, and reasoning effort for all delegated subagents. The plugin provides official UI integration, a globally applied delegation policy, and preset tier management, while supporting foreground waiting and context inheritance. Installation requires mounting an npm package and configuring dependencies; configuration is stored in a local JSON file and is designed for the DSH Web environment.
Read Moredsh-subagent-model: DSH Plugin for Per-Call Model Routing and Context Inheritance
dsh-subagent-model is a DSH plugin maintained by Momojie-S and based on @deepseek-ai/dsh-tool-subagent, providing the subagent_model tool to enhance subagent delegation capabilities. Its core features include dynamically specifying provider, model, and max_tokens for each call, and flexibly controlling context via the fresh_context parameter: a clean context is enabled when the parameter is omitted or set to true, while passing false inherits the current conversation based on the latest state of the parent session at the time of the call. The plugin supports background execution, distinguishes between continuable (sustainable interaction) and one-shot (one-time task) strategies, and prioritizes routing based on call parameters over the plugin.
Read More