Introduction¶
The plugin-based architecture of DeepSeek Harness (DSH) allows capabilities to be extended on demand. TaskWatch is an early preview plugin for DSH, designed to address intent drift during agent execution. As agent autonomy increases, small modifications can evolve into refactoring that departs from the user’s original goals. TaskWatch prevents this by separating the user request, model reasoning, execution evidence, and supervisory feedback.
Plugin Focus¶
- One-line positioning: User-intent tracking and agent supervision for DeepSeek Harness (DSH) · User intent oversight and controlled correction
- Maintainer: cjx12036
- Core problem solved: Keep requests, explanations, evidence, and feedback separate, preventing inferences from silently becoming new instructions.
Core Features¶
TaskWatch maintains three linked views of a task:
- Intent: the original user message, explicit requirements, and independently versioned model reasoning.
- Evidence: task-bound public conversation and tool evidence, including visible gaps, timeouts, truncations, and stale results.
- Supervision: evaluations and recommendations comparing evidence against task intent. These do not replace user utterances or expand authorization.
The sidebar can display the current goal, requirements, trade-offs, usage, and reports. Its private supervision chat only updates TaskWatch’s own understanding; it does not send messages to the main agent or rewrite the main task input.
Coach mode: available only with explicit consent. After enabling project Coach, the consent dialog for new tasks clearly discloses that the main agent may be automatically steered away from execution. Accepting the dialog grants supervisory consent and task Coach consent. Tasks previously bound must separately perform the “Consent Coach” action after project Coach has been enabled.
Installation and Configuration¶
TaskWatch is installed from GitHub into the DSH configuration file. Use the standard web profile to install:
dsh plugin --profile web add github:cjx12036/TaskWatch
After installation, restart the web profile using the normal DSH startup command, then open the TaskWatch sidebar. If you need to diagnose the installation, you can confirm whether the package exists in the configuration:
dsh --profile web --dump-config
Usage Steps¶
- Open TaskWatch and select Settings.
- Select and save the supervision provider and model, then explicitly allow supervision model calls. Saving these settings only does not enable the project or task.
- Enable supervision for the current project.
- When DSH asks for a task decision, select Enable supervision for the specific task. If project Coach is already enabled, read its explicit Coach disclosure before accepting; that single disclosure decision also grants Coach consent to new tasks. Skipping a task does not affect later tasks.
- View the reports in the sidebar. Pause or resume observation of the task as needed.
- If Coach is needed, explicitly enable it for the project, and grant Coach consent for each applicable task according to the description above.
Notes¶
- Early preview: The plugin was smoke-tested for installation and the sidebar UI flow on DSH
0.1.5-rc.3, but real-time runtime generation and Coach behavior have not been verified. - Environment requirements: DSH web client and Node.js
>=22.23.2are required. - Supervision scope: only tasks explicitly bound through the consent flow are supervised.
- Local storage: The default local storage path is
~/.dsh/taskwatch/ledger.sqlite. The path can be changed by settingDSH_HOME(for example,DSH_HOME="$HOME/.dsh-private"). - Remote access: Remote access to TaskWatch endpoints is not supported.
Conclusion¶
TaskWatch provides a mechanism that lets users see more clearly whether an agent’s reasoning process and execution results align with the original intent. For developers who need to keep agent behavior controllable in complex workflows, the plugin offers an intermediate zone between “observation” and “active intervention.”
