Preface

The design philosophy of DeepSeek Harness (DSH) is “everything is a plugin.” For developers who need to build research workflows, standardizing literature retrieval, scoring, review, and state management is a common requirement. dsh-research-autoresearch is a DSH tool-based implementation of this protocol, aiming to provide agents with a closed-loop research processing pipeline.

Plugin Overview

This plugin is maintained by yuan-source-666 and is released as a single-bundle AutoResearch all-in-one suite. It is licensed under the MIT license. Its core value lies in transforming the AutoResearch framework protocol into a pluggable DSH toolset.

Core Features

The plugin includes the following six main tools:

  1. arxiv_search: Literature recall. Supports arXiv search and deduplication.
  2. lit_score: LQS five-dimension scoring and bucketing.
  3. research_state: Research state/progress dashboard. Includes 7 operations.
  4. stall_check: Stagnation detection. Provides a traffic-light mechanism with pivot/escalate/cap.
  5. peer_review: Five-persona median review.
  6. research_ui_switch: Visualization master control console. Supports runtime toggling.

Design Philosophy

The design of this plugin is inspired by Deli Chen’s AutoResearch framework protocol, particularly its empirical research on review mechanisms and stagnation detection:
* Review mechanism: Uses MEDIAN instead of the mean to prevent a single persona from inflating scores.
* Honest score reduction: Retains the scoreDropped field to record proactive score reductions after external citation verification.
* Stagnation detection: Follows the “pivot structure, not tactics” principle and sets dual caps of 15 rounds / 30 minutes.

Installation and Enablement

Installation methods are as follows:

Method 1 (git installation):

dsh plugin --profile web add github:<your-github-user>/dsh-research-autoresearch

Method 2 (prebuilt tarball):
Build and pack it first:

npm run build && npm pack

Then install:

dsh plugin --profile web add ./dsh-research-autoresearch-0.3.1.tgz

After installation, you can add a configuration in the profile’s cordis.patch.yml to disable the plugin, or control it through the Research Visualization Master Control Console card in the “Plugin Configuration” section of the Web settings page.

Visualization Controls

The plugin provides three levels of visual control to avoid features becoming invisible:
1. Suite level: Disable the entire bundle via config.enabled=false.
2. Deployment level: Inject progress cards declaratively via *Cards (disabled by default).
3. Session level: Instantly toggle via the command research_ui_switch {feature, action: on|off|reset} or by setting defaultCards.

Version Updates and Fixes

The plugin fixed common deficiencies in research workflows in versions 0.3.0 and 0.4.0:

  • v0.4.0 Orchestration layer completion: Fixed the “installed but not effective” problem caused by missing registration. Added the autoresearch runtime skill to ensure the model can call tools correctly. Fixed a bug where 6 tools were called 0 times.
  • v0.3.0 Literature-driven upgrade:
    • Citation window calibration: For new papers under 400 days old, when the citation count is 0, it imputes using the mean of the remaining dimensions and marks citationEmergingImputed.
    • Review uncertainty band: Added judgeDisagreement (range) and scoreBand (MAD robust interval). Panels with disagreement ≥4 can no longer directly accept.
    • Self-contained build: All source code is bundled to ensure standalone building.

Test Results

The plugin has undergone multiple rounds of testing, with the following results:
* Host semi-smoke: 11 assertions passed.
* Browser semi-replay: 9/9 passed.
* Installed-state stability: 16/16 passed.
* Cold-start stability: 3/3 passed (HTTP 200, 3-7s).
* Real network round trips: arXiv ≈1.0–1.3s; S2 ≈2.9–16.5s.
* Type checking: 0 diagnostics.
* End-to-end simulation: 133/133 passed.

Dependencies

The plugin depends on the following DSH core libraries:
* @deepseek-ai/cordis
* @deepseek-ai/dsh-settings
* @deepseek-ai/dsh-tools
* @deepseek-ai/schemastery

The generated artifact is located in lib/client.js; manual modification is not recommended.

Summary

This plugin provides DSH with a standardized research processing pipeline, covering the full chain from literature recall, scoring, and review to state management. For developers who need to build research-oriented agents, this is a ready-to-deploy toolset.

GitHub Repository
Community Directory