Preface

The DeepSeek Harness (DSH) plugin ecosystem emphasizes on-demand extension. When developing agents or executing complex tasks, the local environment is good at evidence collection, code modification, and testing, but high-cost reasoning tasks often require calling external models. The dsh-reasoning-bridge plugin solves this split: it delegates high-cost reasoning to web-based AI (defaults to the ChatGPT web version, with switchable targets; currently built in chatgpt and arena), while keeping evidence collection, adoption, modification, and testing entirely local in dsh. The plugin is based on contractual Context Packet / Reasoning Result handoffs, versioned risk authorization gates, an executable browser transport driver, and SHA-256 run receipts.

Core Features

The plugin mainly provides the following capabilities:
1. Separation of reasoning and execution: high-cost reasoning is delegated to web-based AI, while local handles evidence collection, solution adoption, code modification, and test verification.
2. Contractual handoff: structured handoff is performed via Context Packet and Reasoning Result.
3. Versioned risk authorization: a risk statement is displayed before first run, and the authorization file is written after explicit user consent; in non-READY states, the transport layer blocks operations.
4. Executable browser transport driver: supports interacting with web AI through a browser transport driver, and supports multiple targets (chatgpt, arena).
5. SHA-256 run receipt: generates a receipt containing sanitized metadata, artifact hashes, and privacy audit fields.

Installation and Enablement

This plugin depends on the DeepSeek Harness (dsh) environment. The installation command is as follows:

dsh plugin --profile web add dsh-reasoning-bridge

After installation is complete, restart the dsh web service, and reasoning-bridge will appear in the skill directory.

Typical Usage

In a dsh session, initiate the task directly. The plugin automatically handles the handoff between the reasoning request and local execution:

$reasoning-bridge 从第一性原理诊断这个性能回退，验证修复后交付。

$reasoning-bridge 审查这个架构决策，基于本地证据选定方案并完成实现。

Constraints and Boundaries

The plugin sets different constraints at the code level and the process level; developers must distinguish them clearly:

Code Mechanisms (Enforced)

  • Reject unauthorized before sending: when consent is not READY, open/probe/send/fetch/launch fail directly.
  • Accept only supported targets: --site must be one of chatgpt or arena; otherwise an error is reported.
  • Reject suspected credentials before sending: after send reads the request file, it performs credential shape scanning and rejects transmission on a match (--allow-unscanned can be used to explicitly allow it).
  • complete gate rescan: it does not trust self-reported privacy audit booleans and rescans packet/result/evidence content.
  • Reject encoding bypass: non-UTF-8 request files are decoded before scanning.
  • Do not read sensitive APIs: page expressions are prohibited from touching sensitive APIs such as Cookie and Storage.

Workflow Discipline (Non-Code Mechanisms)

  • Adoption gate: web suggestions must be locally re-evidenced by reopening files/symbols before being classified as accepted/rejected/deferred. The plugin cannot block model calls to pwsh/edit/write; final compliance depends on whether the model follows them.
  • Semantic privacy audit: in addition to regex scanning, human judgment is needed for business secrets; the plugin cannot understand business semantics.
  • Credential scanning limitations: it is based only on shape heuristics (e.g., PEM, JWT, API key formats) and cannot intercept business secrets that are split across lines, encoded, or hidden in prose.

Environment Notes

  • Windows fallback launch: dsh’s built-in browser_launch only supports macOS. Windows users must use transport.mjs launch to start a dedicated independent profile, or manually configure the Chrome debug port.
  • Version number discrepancy: the README cover in the materials shows version 0.4.0, while package.json shows 0.5.0; follow the actual environment when installing.

This plugin follows the MIT license and is maintained by the community. For more details, see the Skill Directory or the GitHub repository.