Introduction

DeepSeek Harness (DSH) trajectory evaluation often requires the model to maintain strict behavioral constraints during the early part of a session. If the tool catalog is unlocked immediately after the first request, the model may easily retain old habits learned during the onboarding phase—for example, using long sleep + pgrep polling in foreground jobs instead of invoking the background job interface. The dsh-preset-anchored-standard plugin solves this issue by controlling when the tool set is exposed.

Plugin Positioning

This is an agent preset for DeepSeek Harness, designed to anchor the first request in the session. It uses the bash and str_replace_editor tool pair from the official Minimal preset and maintains this constraint during the early part of the conversation; subsequently, after detecting the first persistent promotion signal (a tool call or an assistant message), it unlocks the full Standard catalog.

Core Capabilities

  1. Anchor the First Request: Ensures the first round of interaction uses only the bash and str_replace_editor tools.
  2. Persistent Promotion: After the first promotion condition is met, exposes the full Standard catalog starting from the second request.
  3. Behavioral Guidance Notification: Injects a one-time user-role notification during the promotion step, informing the model of the difference between background jobs and foreground polling to prevent old habits from persisting.

Installation and Activation

Bundled installation is recommended; the plugin is synchronized to ~/.dsh/.agent-presets/ on startup.

dsh plugin --profile web add "github:ruby1304/dsh-preset-anchored-standard"

After installation, enable it by setting it as the default preset in ~/.dsh/settings.yaml:

agent-presets:
  default: anchored-standard

Configuration Items

All configuration is located in the tool-bootstrap line of presets/anchored-standard/agent.cordis.yml. The main configuration items are as follows:

Key Default Meaning
bootstrapTools [bash, str_replace_editor] Tool catalog exposed for the first request.
promoteOn either Promotion signal condition: tool-call, assistant-message, or either.
suppressedContextSources [agent-instructions, skill-catalog] Context sources automatically injected during bootstrap; set to an empty list to disable filtering.
bootstrapMaxTokens unset Output length limit for the first request, optional.
promotionNotice built-in text One-time notification content injected during the promotion step; can be customized or set to false to disable.

Applicable Scenarios and Notes

This plugin is suitable for scenarios that require strictly restricting early-stage behavior in trajectory evaluations while still needing full capabilities later.

Dependencies and Notes:
* It depends only on the @deepseek-ai/dsh-* plugins bundled with DSH and the local tool-bootstrap.mjs, with no third-party dependencies.
* Before installation, please inspect the source code and license to ensure compliance with your environment requirements.

By using behavioral anchoring rather than mere structural changes, the plugin helps the model quickly switch to background-job mode after acquiring full capabilities.