chenxiachan/thoughtdag#dsh

This directory is the DeepSeek Harness plugin of ThoughtDAG, published to npm as dsh-thoughtdag; its version follows the app's, and the ThoughtDAG build it embeds is the one in this tree.

ThoughtDAG for DeepSeek Harness: open DSH sessions (live or on disk) as editable thought graphs on ThoughtDAG's infinite canvas — from inside the harness UI.

This is a DeepSeek Harness plugin (a Cordis plugin distributed as an npm package), built on the same shell pattern as dsh-synapse: the host half mounts the ThoughtDAG SPA under /thoughtdag/ on the EXISTING harness web server (no second process, no second port), and the client half adds a floating "对话 | 思维图" switch at the top of the harness page that shows the canvas in a same-origin full-screen iframe.

The switch floats over the chat, so it is there before any session is open; while the canvas is up, the same switch sits in the canvas's own top bar next to the canvas chip, so it never covers the canvas toolbar. (开关浮 在对话页顶部，空态也能切换；思维图打开时，同一个开关在画布顶栏「画布名」右侧， 不再压住画布工具栏。0.4.18 曾把它放进会话标题栏，0.4.19 撤回。)

What works

• dsh plugin --profile web add dsh-thoughtdag installs the plugin from npm; every GitHub release also carries dsh-thoughtdag-<version>.tgz, and the same command takes that file's URL (dsh plugin forwards to pnpm)

• The harness GUI gains a floating "对话 | 思维图" switch at the top of the page; "思维图" opens ThoughtDAG at /thoughtdag/ (same origin — no CORS, no second server)

• The host serves the SPA plus a read-only session bridge:

  Endpoint	Purpose
  GET /thoughtdag/api/disksessions	list durable sessions on disk (id, title, cwd from the session header; newest first)
  GET /thoughtdag/api/disksessions/<id>/log	one disk session as JSONL — the exact dialect ThoughtDAG's dsh-session adapter parses, so a past harness session imports like a local file
  GET /thoughtdag/api/sessions	list live sessions in this process
  GET /thoughtdag/api/sessions/<id>/log	one live session's events as JSONL
  GET /thoughtdag/api/sessions/<id>/turns	turn boundaries (start/end seq, the person's message id) of a live or on-disk session — the fork points a canvas can name
  POST /thoughtdag/api/sessions/<id>/fork	{ afterTurn } or { atSeq } → a child session inheriting the prefix through that completed turn (sessionController.fork, so the chat UI lists it)
  POST /thoughtdag/api/sessions/<id>/inject	{ text } or { blocks: [{type:'text',text}] } → model-facing context for the next step (agent.inject; no wake; shown in the transcript as injected context from dsh-thoughtdag)
  POST /thoughtdag/api/sessions/<id>/followup	`{ text
  GET /thoughtdag/api/roots · GET /thoughtdag/api/roots/<key>/list|head|read|range	the other agents' session directories on this machine (claude-projects = ~/.claude/projects, codex-sessions = ~/.codex/sessions) with the desktop bridge's file primitives, so Session Atlas inside the harness lists all three sources and mirrors a Claude Code or Codex session as the desktop does; rel never escapes its root
  GET /thoughtdag/api/models	the harness's model catalog in the SPA's list shape (<provider>/<model> ids, the harness default)
  POST /thoughtdag/api/stream · POST /thoughtdag/api/claude	the SPA's own proxy protocol, answered on the harness's providers and credentials (ctx.llm.stream); the embedded SPA is built with VITE_API_BASE=/thoughtdag, so its model picker and every canvas-native generation (summaries, condensing, a canvas that is not a mirrored session) run on the harness's models. These calls do not enter a session log; a mirrored session's turns go through /followup
  harness/agent (a catalog entry)	pick it in the picker and the question goes INTO the harness: a fresh session per call, the canvas's wired context injected first (agent.inject, source dsh-thoughtdag), the question as a user follow-up, the harness's own agent loop with tools; text, reasoning and tool calls stream back as the SPA's frames, the first frame names the session ({ harnessSession }) and, once the question enters the surface, the turn it created ({ harnessTurn: { turn, userMessageId, seq } }), so the canvas can mark its node as that turn's mirror instead of receiving it twice. The session stays in the Chat and the atlas. harness: { cwd } sets the session's working directory (the canvas passes the project it mirrors, else the current chat session's); harness: { session } continues a mirrored session instead of forking a fresh one (a tail follow-up); images (base64) are admitted through the controller's prompt and a vision model is selected for that turn, so a picture the canvas holds is read by the harness's own eyes
  POST /thoughtdag/api/fetch-url	the SPA's link snapshot ({ url } → { title, text, fetchedAt, html? }) through the harness's bounded, credential-free fetcher (ctx.web.fetch)

  Disk logs are zstd concatenated-frame files; the bridge locates frame boundaries structurally and decodes each with node:zlib (the same walk the DSH persistence backend performs), so no external zstd binary or native module is needed.

• The bridge is Host-header fenced (localhost/127.0.0.1 + trustedHosts)

Verified (2026-09)

• installed into the local web profile via dsh plugin --profile web add <dir>
• a fresh web-profile boot serves /thoughtdag/ (SPA) and the bridge API; the index injection table lists dsh-thoughtdag/client.js
• headless-browser walk: the "对话 | 思维图" switch floats at the top of the page, the full-screen same-origin iframe opens, and ThoughtDAG boots inside it
• fetching the bridge from INSIDE the thoughtdag origin works with no extra auth: GET /thoughtdag/api/disksessions lists every on-disk session (new sessions created by the running instance appear automatically) and GET .../log returns the full event stream
• real session logs decode correctly (multi-frame zstd → 15k+ JSONL lines, identical to the zstd CLI)
• the live bridge (/api/sessions) reflects the running instance's active session as soon as the GUI opens one (verified: opening the UI surfaces the auto-created session with its live seq), complementing the disk archive

Canvas-side import

When a turn launched from the canvas needs the person's approval (the harness's sandbox asking to escalate, a policy hook answering ask), the plugin answers first: it registers ahead of the harness's own panel on the approval/request waterfall for the sessions it is running, streams the question to the canvas as an approval frame, and the node shows it — tool, command, the harness's reason, allow once or reject. The decision returns through POST /thoughtdag/api/approvals/:id and stays on the node as part of the turn's record. If the canvas stream is gone, the request is handed down the chain to the harness's panel unchanged.

Inside the iframe the canvas runs with this bridge as its session source (the SPA is built with VITE_DSH_BRIDGE=/thoughtdag/api): Session Atlas lists the harness's sessions beside Claude Code's, Codex's and Pi's (served by this host from their directories), opens one as a graph, and follows it live by polling the session's seq. The client half tells the canvas which session the chat shows (td:current-session) and switches the chat when the canvas names one (td:select-session).

The why layer inside the harness

The plugin bundles the thoughtdag CLI's library (lib/why.mjs, built at pack time) and answers the same four questions over the same ~/.thoughtdag index — no CLI or MCP install needed:

Surface	What
Native tools	why_check(path), why_file(path, include_read?, limit?), why_find(phrase, in?, limit?), why_recall(session, turn) — the agent calls them like any harness tool; relative paths resolve against the session's working directory
/why <path | url | arxiv:id>	a person asks from the chat; the answer is the CLI's why output, no model message
System-prompt section	four lines telling the model to why_check before editing a file with history and how to read the other three

Both are on by default; whyTools: false / whyPrompt: false in the plugin config turn them off.

Not yet (roadmap)

• Canvas why panel (a node's file → its history through /thoughtdag/api)
• Canvas UI for the write bridge: name a fork point on a node, hand a compiled context to inject, continue with followup; the client already switches the chat to a session the canvas names (td:select-session)
• replace: shadow a surface range from the canvas (DSH's compaction primitive); needs a live check of how the chat renders a replaced range
• Theme and language sync with the harness UI
• Open the chat's current session on the canvas automatically when the view switches

Development

# 1. build the embedded ThoughtDAG SPA with a /thoughtdag/ subpath base (from the repo root)
npm run dsh:build            # writes dsh/dist-app, which git ignores

# 2. install into a profile from this directory (pnpm link)
dsh plugin --profile web add /abs/path/to/thoughtdag/dsh

# 3. restart the profile's web app; the switch floats at the top of the page

The script drives tsc, vite and esbuild through Node directly, so the same code path builds on Windows, macOS and Linux. Always run it as npm run dsh:build instead of passing the flags through Git Bash by hand: MSYS path conversion rewrites --base=/thoughtdag/ into a Git-install path, which bakes a broken base into dist-app — every asset request then falls through to the SPA fallback HTML and the canvas mounts as a blank white page.

While developing, the profile keeps a pnpm link to this directory, so edits to lib/ apply on the next restart without reinstalling.

License: MIT