Introduction¶
DeepSeek Harness (DSH) adopts an “everything is a plugin” architecture. In long-running tasks, Agents can easily fall into repetition or lose coherence. This plugin provides a mechanism for Agents to “daydream” (lose their train of thought) between work segments, briefly reviewing their state so they can continue the task more naturally.
What It Is¶
This is a plugin package named ai_skills, containing the DeepSeek Harness Cordis plugin and the corresponding Codex Skill. It is maintained by developer Stone623. Its core function is to allow Agents to briefly “forget” their line of thinking during continuous work, quickly recap the current task state, and then continue execution.
Core Features¶
- Brief Daydream: Inserts a “losing train of thought” period into the Agent workflow.
- State Recap: The Agent briefly expresses that it is daydreaming, then quickly reviews completed work, current step, and next step.
- Session Snapshot: Generates a recovery snapshot using the current Session event log, ensuring context is not lost.
- Real Scheduling: Uses a real clock and random numbers for scheduling.
- Boundary Injection: Injects context information at Agent step boundaries.
Installation and Enablement¶
Install the plugin using the official DeepSeek Harness command:
dsh plugin --profile web add github:Stone623/ai_skills
dsh --profile web --dump-config
dsh web
After installation is complete, the plugin takes effect at Agent step boundaries.
Typical Usage¶
Explicitly invoke $daydream-break in task instructions.
Basic Usage:
使用 $daydream-break 帮我重构这个模块。
Configure Behavior:
使用 $daydream-break 完成这个任务,最多发呆 2 次,在 3 到 6 分钟内触发,用中文,轻微一点。
The Agent will trigger a daydream after working for a period, then automatically resume the original task.
Configuration Notes¶
Configuration is passed through natural language, without modifying configuration files. Adjustable parameters include:
| Configuration | Default | Example |
|---|---|---|
| Maximum daydream count | 1 per task | up to 2 daydream breaks |
| Trigger time window | 5-10 minutes | trigger within 3 to 6 minutes |
| Language | Follows user language | daydream in English |
| Style | Brief, restrained | lighter, more dramatic |
Capability Boundaries and Notes¶
The plugin has explicit design and implementation limitations:
- Context Preservation: It does not cause the Agent to actually forget context, nor does it display or replay hidden thinking processes.
- Time Estimation: Time and probability are estimated by the Agent between work stages; they are not OS-level background timers or strict random number generators.
- Runtime Limits: It cannot forcibly insert messages in the middle of a single running long command, nor interrupt an already running model request or tool call.
- Use Cases: Suitable for adding a natural and engaging work rhythm. Not suitable for automation scenarios that rely on precise timing or require forced interruption.
Summary¶
stone623/ai_skills provides DSH Agents with a controllable “breather” mechanism. By injecting state recaps at step boundaries, it reduces the monotony of long-running work without breaking context.
- Project Address: https://github.com/Stone623/ai_skills
- Community Catalog: https://www.skillhub.cn/plugins/Stone623/ai_skills