Use ThoughtDAG inside DeepSeek Harness

The plugin embeds the ThoughtDAG canvas in the Harness web UI. Compose the next turn's context on the graph and let the harness execute it. Session Atlas also lets you explore Claude Code, Codex, and Harness conversations.

Install

Prerequisites: a configured DeepSeek Harness installation, 0.1.2-rc.1 or later, with Node.js 22.19+ (22.x) or 24+.

dsh plugin --profile web add dsh-thoughtdag
dsh web

Open the address printed by dsh web. Above the chat, switch to the thought-graph view (思维图); select the chat view (对话) to return.

The plugin includes the canvas. No separate ThoughtDAG desktop app, CLI, or MCP installation is required. Models and tools use your Harness configuration.

Execution reference

Choose how to run a question in the canvas model picker:

Selection	Execution	Where the result is stored
A regular model provided by Harness	Calls the model directly; does not run the full agent tool loop	ThoughtDAG canvas
DeepSeek Harness · Agent	Runs a real harness turn, with tools as configured in Harness	Canvas and Harness session log
An Agent follow-up at the tail of a mirrored Harness session	Continues the corresponding Harness session	A new turn in that session, with the result reflected on the canvas

Selecting a regular model is not the same as running the Harness Agent. Choose DeepSeek Harness · Agent when you want tool execution and a recorded harness turn.

Example: explore an answer

This is a recording of the actual interface:

1. Switch from chat to the thought graph in Harness and open an existing node.
2. Double-click the first question-and-answer node to read its full answer in the right panel.
3. Type a follow-up in the panel's lower input and send it. The answer grows into a second, connected node.
4. Select a passage in the second answer and choose Explore.
5. The new question and answer form a third node. Check its incoming wires to decide what the next question receives.

Both regular models and Agent mode can create canvas nodes. Whether a turn enters the Harness log depends on the execution mode above, not on whether a node appears.

Context and sessions

What reaches the next turn

Upstream nodes, enabled materials, and notes wired into the question become its context. Disconnecting an edge excludes that branch from downstream model context along that path. The nodes remain active on the canvas and can grow or reconnect.

PDF import, text extraction, and material wiring work as in the standalone app. See Material nodes and reader and Control context.

Append a turn; do not rewrite history

Editing or deleting a mirror node changes the canvas, not an existing Harness session log. Sending through Agent mode executes and records a new turn in Harness.

Open Session Atlas from the canvas menu to browse other conversations. The plugin accesses supported sessions on the machine running Harness, not arbitrary files on a browser client.

Query history

The published dsh-thoughtdag@0.4.4 provides the canvas and session integration above. For standalone history queries, see Why layer: CLI and MCP.

Native query tools: awaiting a new npm release

The repository's main branch includes the following native tools and /why command. They are not included in the published dsh-thoughtdag@0.4.4 package and will not appear after installing that version.

Tool / command	Parameters	Purpose
why_check	path	Check whether an object has history
why_file	path; optional include_read, limit	Retrieve related turns and observed changes
why_find	phrase; optional in, limit	Find exact words in questions, answers, or materials
why_recall	session, turn	Read one complete turn
/why <path|url|arxiv:id>	File path, URL, or paper identifier	Query directly without sending another model message

Native tools share the CLI's ~/.thoughtdag index. Relative paths resolve against the current Harness session's working directory. why_find is not semantic search, and candidate explanations are not verified causes.

The main-branch implementation enables native tools and a check-history-before-editing prompt by default. Plugin configuration whyPrompt: false disables the prompt; whyTools: false disables the tools, /why, and their prompt. See the Why manual for output markers.

Troubleshooting

Symptom	Check first
No thought-graph switch	Install with --profile web, restart dsh web, and refresh; check the Harness version and startup logs
A model is missing	Check its model configuration and credentials in Harness
An answer appears on the canvas but not in the Harness log	Check whether you selected a regular model; use Agent mode for tool execution and a recorded turn
Atlas cannot find another tool's sessions	Check that the logs are on the Harness host and use a supported format
/why is missing after installation	Version 0.4.4 does not include native query tools; see the release note above

A local index does not mean retrieved content can never leave the device. When an agent uses results in a remote model request, the matching history may be sent with that request. See Privacy and storage.