dsh-project-workbench

Project Workbench is a DSH sidebar plugin that groups the conversations of one requirement together and manages their progress and memory, so you no longer have to repeat the background in every chat.

What it solves

• One requirement often needs several conversations for analysis, implementation, and verification — and each new chat starts from scratch.
• The sidebar lists chats by time, making it hard to see which chats belong to one requirement and how far it has progressed.

How to use it

Open the workbench: click the "Project Workbench" button at the bottom of the sidebar to expand or collapse it.

Create a group: pick a project, click "New group", give it a name (e.g. "Sign-in redesign"), and put all conversations of that requirement into it.

Add chats: use the "+" dropdown inside a group —

• New chat: opens a fresh conversation in the group right away, like the native New Chat;
• the other entries are existing conversations that belong to no group yet — click one to attach it.

Manage chats: each chat's "⋯" menu lets you rename, fork (continue from the last step on a new line), archive, or remove it from the group (handy for moving it to another group). While a chat is running, its row shows the same live indicator as the native sidebar.

Take notes: every group keeps a "summary" and "notes". The summary refreshes automatically after a conversation that actually advanced the requirement (used a tool or produced a substantive result, and was not a status inquiry), and is carried into the first conversation of each group session so the session knows the background; it is never re-injected afterwards. Status questions, chit-chat, and unrelated Q&A never change it. Notes are maintained by you and are never carried into conversations.

Wrap up: when the requirement is done, archive the group and its chats go with it.

Screenshots

Projects and requirement groups

Auto-updated group memory

Collapsed state

Install

The plugin requires a profile that includes @deepseek-ai/dsh-web-app.

pnpm dsh plugin --profile web add ./plugins/dsh-project-workbench
pnpm dsh --profile web

Data

• All data stays in the DSH data directory on the current device — nothing is uploaded or synced across devices.
• Labels follow the current DSH interface language automatically.

Technical overview

The plugin has a browser half and a host half, integrated through public Cordis slots and host services without taking over the native sidebar, conversations, or composer.

Integration surface

• The browser half (client.js, a window.__ModuleLoader__ bundle) registers two slots:
  • sidebar.footer.action (order 20): the "Project Workbench" entry button in the sidebar footer, styled identically to the Settings and plugin-panel buttons;
  • shell.overlay: the workbench panel itself (a full-height 320px column), positioned beside the native sidebar by measuring the column siblings preceding the overlay, with matching space reserved in the center conversation column.
• The host half (index.js) registers local routes through webServer, listens to session/event to refresh the group summary after every conversation, and attaches the summary injection on agent/pre-step.

Host routes

• GET/PUT /project-workbench/state: reads and writes the workbench state, persisted to project-workbench/state.json under the active profile.
• POST /project-workbench/create-session: truly mints a new session via ctx.apiProxy.sessions.create (the browser first scans for a reusable blank session and only calls this when none exists).
• GET /project-workbench/session-summary?sessionId=: returns the session's first user message (firstUserMessageSummary, a pure function, truncated at 240 chars) for group-memory seeding.

Data model

State is shaped as projects[workspaceId].groups[]; each requirement group carries id, name, status (in progress / review / completed), archivedAt, sessionIds, and memory { summary, notes }. Browser storage is only an offline cache and migration source for older data.

Auto-updated group memory and injection

The host listens to session/event and, only after a turn that actually advanced the requirement (the request was not a status inquiry, and the turn used a tool or produced a substantive reply) finishes (turn/end), updates the group summary ON TOP of its current content and persists it to state.json. The digest keeps a structured 需求：… / 进展：… shape (the whole group always carries exactly one requirement line and one progress line, compact and easy for the model to read): the progress line is refreshed in place to the key sentence of the latest progress reply (auto-condensed to a single, complete sentence; no detail lists). Status questions, chit-chat, and unrelated Q&A never change it. agent/pre-step then adds the current summary as a source-attributed context message ([project-workbench-memory group=… revision=…]) only to the first conversation (first turn) of each group session; later turns are not re-injected. Notes stay local and never enter any model request.

Group-memory generation: current state and the LLM-mode toggle

By default the digest is generated deterministically, with rules (no model call): the host folds session events into turns (collectTurns), keeps only turns that actually advanced the requirement (turnCountsAsProgress), picks a key sentence per session for 需求/进展 (condenseText, skipping "好的/我先…" openers and detail lists), and supplementGroupSummary maintains the single-line 需求：… / 进展：… digest in place. The whole digest stays within 240 characters; when over budget, whole sentences are dropped — never a mid-sentence cut, never an ellipsis.

Every requirement group's memory pane has an Auto / Manual toggle on the summary tab: in manual mode the host never rewrites the summary after conversations (and does not auto-seed it) — you maintain it fully, yet it is still injected into the first conversation of each group session.

A toggle for LLM summary mode is reserved in the plugin config (rule-based by default, for future improvement):

• summaryMode: "rule" | "llm" (default "rule"): with "llm", after a progress turn the configured model generates the 需求：… / 进展：… digest from a dedicated prompt (llmSummaryPrompt);
• summaryModel: "provider/model" (e.g. "deepseek-official/deepseek-chat"): the model used in LLM mode; when unset, the session's own active model is resolved automatically (via the host apiProxy session models — the config system's own conversation model); on any failure or unavailability it falls back to the rule digest.

Future improvements: ① let the model really rewrite/condense instead of rule extraction; ② async queue + retry for the LLM call; ③ configurable budget and prompt; ④ a "pin this digest" option so manual edits are not overwritten.

Conversation behavior

• "New chat" follows native New Chat semantics: it reuses the project's unused blank session when one exists (blank chats show no timestamp and do not accumulate), and only mints a fresh one through the host gateway otherwise.
• "Add existing chat" lists only conversations claimed by no requirement group and not archived.
• Group conversation rows support rename, fork (ctx.sessions.fork), archive (ctx.workspaces.archiveSession), and remove-from-group; the running indicator matches the native sidebar.

For the design and product boundary, see docs/design.md.