Introduction¶
In the use of DeepSeek Harness (DSH), subagents typically inherit the parent session’s agent preset configuration automatically. Although this inheritance mechanism is convenient, when working with models that require specific RL (reinforcement learning) trajectories, subagents may fail to start in an optimal state on their first request without a dedicated bootstrap configuration. DSH’s complete-persona mechanism makes overriding subagent configuration via hooks unreliable. This plugin aims to resolve inconsistent subagent startup states by installing an agent preset.
What It Is¶
dsh-anchored-subagent is a community-maintained DSH plugin developed by GY-Bai. It uses a set of “cues” mechanisms to ensure that the main agent and all subagents start at the beginning of a session with a Minimal condition (bash + str_replace_editor + Minimal persona), and then unlock the full tool catalog afterward. The plugin is designed based on DeepSeek’s RL training trajectories, aiming to place the model into its strongest behavior trajectory on the first request.
Core Features¶
- Startup bootstrap: Installs an agent preset so that the main agent and subagents start with a Minimal condition (
bash+str_replace_editor+ Minimal persona). - Tool unlocking: Unlocks the full tool catalog after the first request.
- Automatic inheritance: Subagents automatically inherit the parent’s agent preset without additional configuration.
- Persona handling: Handles custom subagent personas (stripped at startup and restored after promotion).
Installation and Activation¶
Use the official install command to add the plugin. After installation, restart DSH.
dsh plugin add github:GY-Bai/dsh-anchored-subagent
After restarting DSH, select the dsh-anchored-subagent agent preset in a session (or set it as the default).
The plugin copies the preset to the local directory:
~/.dsh/.agent-presets/dsh-anchored-subagent/
Typical Usage¶
The plugin’s workflow consists of two phases:
- Bootstrap phase: On the first model request (main agent or subagent), only the
bashandstr_replace_editortools are provided, along with the full Minimal persona. At this point, the model begins thinking along the standard trajectory used during RL training (for example,We need...). - Promoted phase: After the first persistent tool call (
tool/call) or assistant message (assistant/message), the plugin automatically restores the full tool catalog to the agent, allowing it to use all tools.
For subagents, if a custom persona is configured, the plugin strips it during the bootstrap phase to ensure the subagent starts with a clean Minimal persona. After the subagent completes its first persistent request, the original persona is restored.
Configuration¶
Edit the configuration file ~/.dsh/.agent-presets/dsh-anchored-subagent/agent.cordis.yml. The key configuration options are as follows:
| Option | Default | Description |
|---|---|---|
promoteOn |
either |
The promotion trigger condition. Options: either, tool-call, or assistant-message. either means either condition can trigger promotion. |
bootstrapMaxTokens |
null |
Optional output limit (token count) for the first request. |
suppressedContextSources |
["agent-instructions","skill-catalog"] |
Controls which injected context sources are automatically stripped during the bootstrap phase. Set to an empty array [] to disable stripping. |
Applicable Scenarios and Notes¶
This plugin is suitable for scenarios where Minimal condition must be enforced when a subagent starts, in order to achieve the best RL trajectory response.
Please note the following:
* DSH must be restarted to apply changes.
* This is a community plugin and has no official endorsement from DeepSeek. It is open source under the MIT license.
* The plugin leverages the “cues” mechanism in DeepSeek’s RL training trajectories.
* Context injection can be controlled via suppressedContextSources.
Summary¶
dsh-anchored-subagent enforces Minimal condition during the startup phase, ensuring that DeepSeek models enter their strongest RL trajectory before full capabilities are unlocked afterward. This is very useful for complex workflows that require stable subagent behavior.