DeepSeek Harness (DSH) provides a general-purpose agent runtime. When handling long conversations or complex tasks, loading all history into the context at once is often inefficient and can introduce irrelevant noise. The dsh-point plugin introduces the “Point” concept to discretize history, providing DSH with checkpoint-based session management and inheritance.

Plugin Overview

dsh-point is a DSH plugin maintained by guix4ever. It does not treat the entire session history as a single context; instead, it divides history into immutable Point units. Developers can save the current intermediate state and resume or continue from a specific Point when needed. Models can read raw events from ancestor Points on demand through the API, without reprocessing all history.

Core Features

  1. Point stashing (/point-stash)
    Using the /point-stash [title] command, the plugin saves all participant-visible events since the last Point. The stash boundary waits until the agent is idle, flushes DSH persistence, and advances the cursor after success.

  2. Web workspace view
    The plugin provides a workspace view for DSH Web users. It renders the parent DAG (directed acyclic graph) of Points, displays node details, and offers an explicit Continue from this Point action.

  3. Deterministic session inheritance
    Clicking Continue creates or reuses a deterministic DSH Session that carries the Point’s inheritance chain. This mechanism does not copy historical events into a new context; it references them through the inheritance relationship.

  4. On-demand ancestor data reading (point_read)
    Models can read raw events from ancestor Points through the point_read(point_id) interface. This interface rejects unrelated Point IDs, ensuring that permissions match the inheritance chain.

  5. Crash recovery and cursor synchronization
    The plugin implements crash recovery by coordinating the local cursor with immutable Point metadata, preventing duplicate Points from being created due to unexpected interruptions.

  6. Data integrity preservation
    During processing, the plugin excludes compressed records but fully preserves visible transcript content, content blocks, source metadata, and paired tool calls and results.

Installation and Activation

Before installing this plugin, ensure the following prerequisites are met:
* Node.js 22.19 or later
* DeepSeek Harness version 0.1.0-rc.6
* The point CLI tool is installed and available in the PATH environment variable (it can be installed with python3 -m pip install "git+https://github.com/guix4ever/point.git")

Use the following command to add the plugin to the DSH web configuration file:

dsh plugin --profile web add dsh-point

After installation, DSH must be restarted for the plugin to take effect.

Typical Usage

1. Save current work
Enter the command in the conversation to save current progress:

/point-stash explain the change briefly

2. View and continue a session
Open the Points tab in the DSH conversation view to inspect the current workspace’s DAG structure. Selecting a node to view is a read-only operation. To continue from that node’s state, click the Continue from this Point button.

3. Model reads specific history
After continuing a session, only the selected Point’s ancestor chain is inherited. If the model needs more detailed historical information, it can invoke:

point_read(point_id)

Only Point IDs in the current session’s inheritance chain are accepted.

Notes and Scenarios

  • Failure policy: Projection operations trigger a fail-closed policy when encountering non-contiguous events, open turns, incomplete assistant data blocks, duplicate or missing tool identities, dangling calls, and empty visible Points.
  • Browser limitations: Browser environments cannot independently select filesystem paths, Point binaries, or Shell arguments. Workspace paths and agent presets are derived by the Host based on the addressed DSH Session.
  • Use cases: Suitable for long-chain tasks that require frequent interruption and resumption, and for agent development that requires strict control over context size and historical reference relationships.

Summary

By introducing immutable Point units and an inheritance-based session mechanism, dsh-point optimizes DSH’s resource usage in long-conversation scenarios. It allows developers to more precisely control when and how much historical context is loaded, while the web interface provides intuitive checkpoint management capabilities.

GitHub repository | Skill Hub