fde-console — Super Workbench

English | 中文文档

A local-first workbench plugin for DSH that turns your local AI coding agents (ZCode, Claude Code, Codex, OpenCode, DSH, WorkBuddy…) into an observable, accountable agent team.

The core loop: See → Understand → Attribute → Accept.

• See — live agent activity across six local session sources (per-project feed, running-now section, 28-day heatmap)
• Understand — full transcript reader per session, plus the dispatch prompt: the original human instruction behind each session
• Attribute — commits and sessions auto-attached to tasks via a branch/markdown convention (docs/ATTRIBUTION.md), with a repo guard against cross-repo false matches
• Accept — evidence lives on the task: timeline, session evidence with one-click confirm, daily/weekly reports with Markdown export

Local-first by design: no cloud, no third-party runtime dependencies (Node built-ins only), all data in one directory you own. Agent sessions are read-only observed — the workbench never writes into your agents.

Features

• Task board — kanban with timeline, acceptor sign-off, deep links, priorities, due dates, daily diary per task
• Agent observation — six-source adapters, session lists, transcript reader, running-now detection, per-project tabs
• Work search — Ctrl+K across task titles, session titles, commit messages and diaries
• Dispatch prompts — see the original instruction each agent session was started with
• Session evidence — suggested attachments scored per task, one-click or batch confirm
• Reports — auto-generated daily digest and weekly review, exportable as Markdown
• Heatmap & scorecards — 28-day activity grid per agent; SVG scorecard endpoints
• Integrations — MCP server (stdio) so your agents can query the board; webhook out; IM notify (Feishu/DingTalk bots); Gitea/GitHub/Gitee activity feeds
• Governance — credential never touches disk (git config runtime read), redaction on every output, SSRF allowlist for all outbound, single-instance lock with warning, backup/restore

Requirements

Node.js	≥ 22.13 (uses built-in node:sqlite)
DSH	0.2.1-rc.2 (@deepseek-ai/dsh)
pnpm	on PATH (plugin install forwards to it)

Quick start

# 1. install / update the plugin into a profile (the profile is created on first use)
dsh plugin --profile fde-dev add <this repo path>

# 2. a fresh profile ships @deepseek-ai/dsh-base only — the web surface bundle is separate.
#    Without it the plugin loads but never activates:
#      "fde-console: pending (waiting for service: webServer)"  (port never listens)
#    Pin it to your dsh version; an unpinned add resolves an incompatible old build.
dsh plugin --profile fde-dev add @deepseek-ai/dsh-web-app@0.2.1-rc.2   # ← use `dsh --version`

# 3. start (no `dsh --profile fde-dev web` — web is a reserved alias profile)
dsh --profile fde-dev --port 38082

Open the printed URL (auth token included) → Super Workbench appears in the sidebar. Your local agents are detected automatically — zero config to see your first session.

Data lives in <DSH_HOME|~/.dsh>/fde-console/ (data/ + config). A fresh data directory stores JSON; run node scripts/migrate-storage.js (with every instance stopped) to move the owned units to SQLite + WAL. JSON remains the export/interop format.

Step-by-step verification (test gate → plugin load → API round-trip → UI): docs/VERIFY.md.

Demo dataset

Want a populated workbench before connecting anything? Seed a fully synthetic dataset (members, tasks, minutes, inbox candidates, team events — placeholder names only, zero network):

node scripts/seed-demo.js                          # seeds the default data dir (~/.dsh/fde-console/data)
node scripts/seed-demo.js --data-dir <dir> --force # custom dir / overwrite

Then start the instance and explore; delete the seeded JSON files to reset. The seed goes through the JSON storage channel, so it targets fresh (un-migrated) data dirs — the script refuses a migrated (SQLite) target and asks you to use a new directory.

The six views

View	What it shows
Overview	stats, agent heatmap, digital-worker reports (daily/weekly), version matrix
Task Radar	meeting transcripts → AI task decomposition → inbox gate
Task Board	kanban: timeline, acceptor, deep links, diary, project filters
Agents	live activity feed across six sources + session reader
Members	desktop agents + human members: status, output, diary nudge
Team Activity	multi-source commit/PR feed with task attribution
Settings	sources, engine, security allowlist, one-click env check

Attribution protocol

AI-made commits can self-attach to tasks with three conventions: branch task/<id>, marker [#<id>] in the message, or binding the task to a project directory. See docs/ATTRIBUTION.md — includes a ready-to-paste snippet for your own AGENTS.md and a commit-msg hook.

Writing your own session-source adapter (Cursor, Trae, Aider…): see docs/ADAPTERS.md.

MCP server

scripts/mcp-server.mjs exposes read-only tools (list tasks / search work / agent status) over stdio, so any MCP client can query the workbench:

{ "mcpServers": { "fde": { "command": "node", "args": ["<repo>/scripts/mcp-server.mjs"] } } }

Testing

npm test   # node --test test/*.test.js — 400+ cases, zero network

Documentation

Primary documentation is in Chinese under docs/: SPEC.md (functional contract), SPEC-PLUGIN.md (DSH plugin conformance), SPEC-STORAGE.md (SQLite storage design), ATTRIBUTION.md, ADAPTERS.md, DISTRIBUTION.md.

License

MIT