Neural Ledger · 神经账本

English | 中文

> One-liner: A zero-patch DSH plugin that turns your AI collaboration sessions into a beautiful work ledger — token analytics, smart insights, trend forecasting, and one-click daily / weekly / monthly reports. Live data from your own sessions, no DSH source touched.

Features

• 📊 Living dashboard — Session counts, turns, tool calls, token usage, and AI time in one glance.
• 💡 Smart insights — Auto-generated findings: which session burned the most tokens, what took the longest, how much subagent collaboration matters, and overall conversation efficiency.
• 📈 Token analytics — Daily token bars, token composition donut (uncached input / cache read / output), per-session token & time rankings.
• 🔮 Trend forecasting — Linear-regression prediction of the next 7 days, monthly estimates, and a budget-overrun warning (default 50M tokens/month).
• 🗂 Workspace breakdown — Compare token consumption across projects; tree drill-down workspace → parent agent → subagent.
• 📋 One-click reports — Export daily / weekly / monthly reports as Markdown, with per-turn demand → outcome storylines and subagent task details. Copy to clipboard or download.
• 📤 Export session context — Tree picker (workspace → parent agent → subagent, collapsible, fuzzy search incl. workspace) → generates a distilled context (prompt-like): metadata, goal, per-turn demands/outcomes, tool summary, and a continue-work prompt ready to feed another agent.
• 🌐 i18n — Toggle Chinese / English for the whole UI and exported reports (remembers your choice).
• 🎬 Sample mode — Built-in mock dataset for previewing the whole dashboard without waiting for real data.
• Draggable FAB — The floating action button is draggable, position is remembered, and a hover label follows it live.
• No DSH source modification — Only DOM-level integration; DSH files are never touched.

Quick Start

Install from GitHub

dsh plugin --profile web add github:Elpsycoogroo/dsh-work-report

> pnpm blocks build scripts by default: installing from GitHub runs the project's own build script, and pnpm refuses until you allowlist it. Run the command once — pnpm prints the key to add under allowBuilds in ~/.dsh/profiles/web/pnpm-workspace.yaml. Add it and run again.

Install from npm (once published)

dsh plugin --profile web add dsh-work-report

Manual install in this repo

Clone/symlink the plugin at dsh/plugins/dsh-work-report and build once:

cd dsh/plugins/dsh-work-report
npm install
npm run build

> ⚠️ Don't manually copy only lib/ into the profile's node_modules/ — a copy missing package.json (and cordis.patch.yml) cannot be resolved by the DSH loader. Copy the whole package. For ongoing development use node dev.mjs (watch src/ → auto-build → auto-sync the whole package into the profile).

Usage

1. Open DSH, click the 🧠 floating button (bottom-right by default, draggable anywhere).
2. The Neural Ledger overlay opens with real data from your sessions.
3. Pick a report type (日报 / 周报 / 月报 — auto-switches daily window 1/7/30 days).
4. Hit 📋 Copy or ⬇ Download to get the Markdown report.
5. 📤 Export Context — pick a session from the collapsible tree (workspace → parent → subagent) or fuzzy-search it, then copy/download the distilled context as a prompt for another agent.
6. 🌐 EN — toggle the whole UI (and exports) between Chinese and English.

⚠️ Read before forking/self-hosting (war stories)

1. package.json must export "./package.json": DSH's client-modules reads a plugin's manifest via require.resolve('/package.json'). If exports doesn't expose that subpath, it throws ERR_PACKAGE_PATH_NOT_EXPORTED and the plugin shows in the plugin list but its client.js is never injected into the page.
2. All three names must match: the name in package.json, in the plugin's own cordis.patch.yml, and the name referenced from your profile bundle.
3. You MUST restart dsh after changing the manifest/reinstalling: client-modules caches the "not a client plugin" verdict for the process lifetime.
4. sessionPersistence.readFrom() may be unavailable: the server falls back to projcache (storages/session_projcache.json); token totals for subagents missing from the cache are aggregated from event usage blocks.

Contributor Docs / 给开源作者

Local dev, debugging and integration guides are in DEVELOPING.md / DEVELOPING.zh.md.

How It Works

Architecture

browser (client plugin)
ReportView ── StatCards / Insights / TokenCharts / ForecastCard
├── WorkspaceChart / EfficiencyCharts / ToolRanking
└── SessionTimeline (workspace → parent agent → subagent)
│
└── fetch('/api/work-report?days=7&mock=1')   ← requested by ReportView, shared by all cards
Host (server plugin) [ctx.webServer.register({ kind: 'exact', path: '/api/work-report' })]
ctx.get('sessions')                → attached (in-memory) sessions
ctx.get('sessionPersistence')      → cold (persisted) sessions + events
storages/session_projcache.json    → tokenUsage / sessionStats / contextPressure / subagent labels
→ buildReport(config) → { sessions, token, time, insights, forecast, dailyTokens, workspaceTokens }

Data sources

• Active sessions — ctx.sessions.list() (attached, in-memory).
• Cold sessions — persistence.list() + persistence.readFrom(id, 0) for events and parentSession linkage. readFrom is optional; when absent, metadata comes from projcache.
• Token / stats — projcache projections (tokenUsage.totals, sessionStats, contextPressure) with event-usage aggregation as fallback.
• Subagent labels — projcache subagent.identity.label (e.g. Worker A - 代码开发); parent linkage via meta.parentSession.
• Archived sessions — filtered using workspace.json's global.archivedSessionIds; blank sessions (0 tokens & 0 time) are filtered out too.

Report generation

• Recursive text extraction — demand / outcome text pulled from any message nesting shape, skipping `,Current runtime context.`, and other noise.
• Turn storyline — each turn records user demand, AI outcome, tool calls (✓/✗), and token usage.
• Forecast — linear regression over daily tokens, with fallback base for sparse data; 7-day projection + 30-day estimate vs budget.

Files

dsh-work-report/
├── package.json
├── tsconfig.json
├── tsdown.config.ts
├── cordis.patch.yml
├── README.md            # English docs
├── README.zh.md         # 中文文档
├── DEVELOPING.md        # English contributor guide
├── DEVELOPING.zh.md     # 中文开发者指南
├── CONTRIBUTING.md      # 中文贡献指南
├── CONTRIBUTING.en.md   # English contributing guide
├── GITHUB_SETUP.md      # GitHub repo setup checklist
├── pull_request_template.md
├── pull_request_template.en.md
├── mock-report.json     # built-in sample dataset (🎬 Sample mode)
├── screenshots/         # README screenshots (dashboard previews)
├── .github/             # ISSUE_TEMPLATE (bug_report.yml / feature_request.yml)
└── src/
├── index.ts             # Host entry (re-exports)
├── server/
│   ├── index.ts         # webServer route /api/work-report
│   └── report-data.ts   # data collection, aggregation, insights, forecast
├── client/
│   ├── index.ts         # client entry: draggable FAB + overlay mount
│   ├── i18n.tsx         # zh/en dictionaries + language provider
│   ├── ReportView.tsx   # main dashboard
│   ├── StatCards.tsx    # stat cards
│   ├── Insights.tsx     # smart insight cards
│   ├── TokenCharts.tsx  # daily bars + composition donut
│   ├── ForecastCard.tsx # trend prediction + budget warning
│   ├── WorkspaceChart.tsx # workspace token comparison
│   ├── EfficiencyCharts.tsx # per-session time & token rankings
│   ├── ToolRanking.tsx  # session-type token share
│   ├── SessionTimeline.tsx # 3-level tree session list
│   ├── ContextExporter.tsx # export session context (tree picker + search)
│   ├── markdown.ts      # daily/weekly/monthly report generator
│   └── report-api.ts    # API fetch + formatting utils
└── types/
└── dsh-env.d.ts     # ambient type declarations

Build & Publish to npm

Build locally

cd dsh/plugins/dsh-work-report
npm run build     # tsdown: host ESM (lib/index.js) + browser CJS (lib/client.js)
node dev.mjs      # watch mode: auto-build + auto-sync whole package to profile

> Browser bundle inlines echarts (kept in devDependencies so tsdown bundles it; module-table externals are only react / @deepseek-ai/*).

Publish to npm (once you own the package name)

npm login
exports_subpath=./package.json   # keep exports["./package.json"] — DSH client-modules needs it
npm version patch -m "chore(release): v%s"
npm publish --access public
# verify the tarball contains everything the runtime needs:
npm pack --dry-run | grep -E "package.json|cordis.patch.yml|lib/(index|client)\.js|mock-report"

> The published files are controlled by package.json's files field (lib, src, mock-report.json, cordis.patch.yml, docs). Before the first publish make sure files includes every runtime file — the DSH loader resolves package.json and cordis.patch.yml at runtime, not just lib/.

Console Logs

License

MIT