Introduction¶
When developing or debugging agents in DeepSeek Harness (DSH), long conversation trajectories can be difficult to parse quickly. @nydsg/dsh-mindmap renders the conversation as a horizontal hierarchical tree: the leftmost node is a single starting master title (whose content is the round 1 question), and each subsequent round expands outward to the right, level by level. Card surfaces display only the question. The plugin automatically continues branches by comparing the similarity of each round’s question with earlier questions (signal-word threshold: 0.50), and also allows manual changes to the parent node. All analysis is performed locally in the browser, with no network requests, no model calls, and no build steps.
Installation and Activation¶
After installing the plugin, restart Harness. A third tab, Mind Map, appears at the top of the session, at the same level as Chat and Trace.
dsh plugin --profile web add @nydsg/dsh-mindmap
Core Features¶
Horizontal Hierarchical Tree and Master Title¶
The mind map view uses a layout where columns represent depth and rows represent placement. The leftmost column is fixed as the starting master title, whose content is taken from the round 1 question. It is not a conversational round, has no module, cannot be selected, and does not generate its own connector line. All branch starting points are attached under the master title, forming a single-entry tree structure.
Automatic Branching and Following the Previous Round¶
The plugin determines branch direction by comparing surface-level similarity of questions. The rules are as follows:
1. Surface hit: If the current question shares enough signal words with a previous question (threshold: 0.50), it is attached under that round (backtrack).
2. Follow the previous round: If the question does not match any earlier round on the surface, the current round becomes a child of the previous round. This is a structural decision based on conversation order, not a guess.
3. New branch: If it neither matches an earlier question nor follows the previous round, it becomes a new starting branch.
Hierarchy Limit and Wrap¶
The default maximum hierarchy level is 5. If a round would be placed at level 6 under the normal rules, it does not go deeper; instead, it is placed alongside the previous round (a wrap creates it as a sibling). This prevents endless horizontal expansion in long sessions. In the toolbar, you can set the maximum hierarchy level to 0 to remove the hierarchy limit.
Manual Control and Persistence¶
After selecting a card, you can use the right-side panel to manually attach the round under any earlier question or convert it into a new branch. Manually specified connections are saved locally per session, remain after reload, and are not overwritten by automatic recalculation.
Model Judgment (Optional)¶
Starting from version 1.4.0, “Model Judgment” is supported. It uses the model configured in DSH and the system prompt to determine whether the current round should be “pushed down,” “wrapped,” or “backtracked.” The judgment is invoked through the local HTTP bridge (/plugin-mindmap/*), is disabled by default, and takes effect when you click “Run Model Judgment.” Manual specification always overrides the model’s conclusion.
Export and Offline Use¶
The plugin supports one-click export of the entire mind map as a Markdown outline. All analysis is performed locally in the browser, with zero network requests, zero model calls, and zero build steps. The offline default behavior is to push down according to conversation order.
Judgment Logic¶
The plugin uses signal words (appearing no more than twice in the session or above their own IDF median) rather than cosine similarity to compute scores. Genuine follow-up questions (for example, “Then what’s next?”) almost never repeat nouns from the previous round’s reply, so similarity alone cannot determine “follow the previous round.” The plugin defines “follow the previous round” as structural continuation and defines “backtrack” as a return to an older topic with clear surface-level evidence.
Use Cases¶
Suitable for developers who need to organize the logic of long conversations, debug an agent’s planning capability, or visualize session structure.
Resources¶
- Repository: https://github.com/nydsg/dsh-mindmap