docs: revise MCP standalone plan per architecture review

This commit is contained in:
2026-02-28 16:05:44 +01:00
parent 7517bfb4c6
commit 095c212ebc

View File

@@ -5,7 +5,6 @@
Ship a bundled CLI tool (`bds-mcp`) alongside the Electron app that:
- Runs as a self-contained MCP server over **stdio** (for Claude Desktop, VS Code, etc.)
- Spins up **HTTP :4125** alongside stdio for App View callbacks (same shared state)
- Works entirely against the **SQLite database** — no Electron process required
- Signals the running app of changes via a **`db_notifications`** table
- Is installed into agent configs by the app's existing agent-config UI (with add **and remove**)
@@ -19,15 +18,15 @@ Ship a bundled CLI tool (`bds-mcp`) alongside the Electron app that:
│ bds-mcp (bundled CLI binary) │ │ bDS Electron app │
│ │ │ │
│ StdioServerTransport ──────┐ │ │ HTTP MCPServer :4124 │
StreamableHTTP :4125 ──────┤ │ │ ├── same engines │
│ │ ├── same engines │
│ │ │ │ └── NotificationWatcher │
│ PostEngine ├───┼─────┼──► SQLite (WAL mode, shared) │
│ MediaEngine │ │ │ posts / media / scripts … │
│ ScriptEngine │ │ │ db_notifications ← new │
│ TemplateEngine │ │ └────────────────────────────────────┘
│ MCPServer (tools/res) │ │ ▲
│ ProposalStore (in-memory) │ │ setInterval polls
└─────────────────────────────┘ │ refresh caches + IPC to renderer
│ ProposalStore (in-memory) │ │ fs.watch(-wal) + 100 ms debounce
└─────────────────────────────┘ │ invalidate engines + IPC to renderer
agent (Claude Desktop, VS Code…) ─┘
(stdio)
@@ -38,8 +37,7 @@ agent (Claude Desktop, VS Code…) ─┘
| Context | Port | Transport |
|---|---|---|
| bDS app running | :4124 | HTTP (existing, unchanged) |
| CLI stdio mode — agent | — | stdio |
| CLI stdio mode — App View callbacks | :4125 | HTTP |
| `bds-mcp` CLI | — | stdio |
---
@@ -92,6 +90,31 @@ Behaviour:
---
### C0 — Decouple `DatabaseConnection` from Electron
`src/main/database/connection.ts` currently imports `{ app } from 'electron'` for three
things: `app.getPath('userData')`, `app.isPackaged`, and `process.resourcesPath`. The CLI
cannot load this file at all without Electron present.
Refactor `DatabaseConnection` to accept explicit constructor arguments:
```typescript
interface DatabaseConnectionConfig {
dbPath: string; // absolute path to bds.db
migrationsFolder: string; // absolute path to the drizzle/ migrations dir
createDirs?: string[]; // extra directories to mkdir at startup
}
```
The Electron `main.ts` constructs it as today (computing paths via `app.getPath` before
passing them in). The CLI computes the same paths via `platformConfigPath()` and
`__dirname`-relative resolution from the bundle. All Electron imports are removed from
`connection.ts` itself.
This is a prerequisite for every CLI section that follows.
---
### C — `platformConfigPath()` — userData without Electron (`src/cli/platform.ts`)
Pure-Node helper to resolve the same path as Electron's `app.getPath('userData')`:
@@ -123,84 +146,211 @@ Injected as an optional constructor argument on `PostEngine`, `ScriptEngine`,
---
### E — `NotificationWatcher` (new engine, app-side only)
### E — Engine `invalidate()` API
No engine currently exposes a way to signal that its data is stale. Before
`NotificationWatcher` can be implemented, each engine that holds any in-memory state needs
an `invalidate(entityId?: string): void` method that clears the relevant cached entries
(or the whole cache when `entityId` is omitted), forcing a fresh DB read on the next
access. Engines that are already fully stateless (query DB on every call) get a no-op stub
so `NotificationWatcher` can call a uniform API.
Engines to update: `PostEngine`, `MediaEngine`, `ScriptEngine`, `TemplateEngine`.
---
### F — `NotificationWatcher` (new engine, app-side only)
`src/main/engine/NotificationWatcher.ts`
- `setInterval` polling every 2 s
- Reads `db_notifications WHERE seenAt IS NULL AND origin = 'cli'`
- For each row: invalidates the relevant engine's cache, calls
`mainWindow.webContents.send('entity:changed', { entity, entityId, action })`
- Stamps `seenAt = Date.now()` on processed rows
- Started in `main.ts` alongside `PreviewServer` and `MCPServer`
---
### F — `MCPServer.startCli()` (extend existing `MCPServer.ts`)
Uses `fs.watch()` on the SQLite WAL file (`bds.db-wal`) for near-instant change
detection instead of polling. When a write lands in WAL mode, the `-wal` file is modified
before a checkpoint; watching it gives sub-100 ms latency.
```typescript
async startCli(appViewsPort = 4125): Promise<void> {
// 1. Stdio transport — primary agent connection
const stdioTransport = new StdioServerTransport();
const stdioMcp = this.createServer({ mode: 'cli', appViewsPort });
await stdioMcp.connect(stdioTransport);
export class NotificationWatcher {
private dbWatcher: FSWatcher | null = null;
private walWatcher: FSWatcher | null = null;
private debounceTimer: ReturnType<typeof setTimeout> | null = null;
// 2. HTTP :4125 — App View callbacks, same ProposalStore
await this.startHttp(appViewsPort, { mode: 'cli' });
constructor(
private readonly dbPath: string, // absolute path to bds.db
private readonly db: DrizzleDB,
private readonly engines: WatchableEngines,
private readonly mainWindow: BrowserWindow,
private readonly debounceMs = 100,
) {}
start(): void {
// Watch the main db file (catches checkpoints)
this.dbWatcher = watch(this.dbPath, { persistent: false },
() => this.schedule());
// Watch the WAL file (catches every write in WAL mode)
// The WAL file may not exist yet if no writes have occurred; we create
// a short-lived retry watcher that tries once per second until the file
// appears, at which point walWatcher is assigned and the retry cancels.
this.attachWalWatcher();
}
private attachWalWatcher(): void {
const walPath = this.dbPath + '-wal';
try {
this.walWatcher = watch(walPath, { persistent: false },
() => this.schedule());
} catch {
// -wal does not exist yet; retry in 1 s
setTimeout(() => this.attachWalWatcher(), 1000);
}
}
private schedule(): void {
if (this.debounceTimer) clearTimeout(this.debounceTimer);
this.debounceTimer = setTimeout(() => this.process(), this.debounceMs);
}
private async process(): Promise<void> {
const rows = await this.db
.select()
.from(dbNotifications)
.where(and(
isNull(dbNotifications.seenAt),
eq(dbNotifications.origin, 'cli'),
));
for (const row of rows) {
this.engines[row.entity]?.invalidate(row.entityId);
this.mainWindow.webContents.send(
'entity:changed',
{ entity: row.entity, entityId: row.entityId, action: row.action },
);
await this.db
.update(dbNotifications)
.set({ seenAt: Date.now() })
.where(eq(dbNotifications.id, row.id));
}
}
stop(): void {
if (this.debounceTimer) clearTimeout(this.debounceTimer);
this.dbWatcher?.close();
this.walWatcher?.close();
}
}
```
Both `McpServer` instances are created via the same `createServer()` factory sharing the
same `ProposalStore` instance. App View HTML receives the `:4125` base URL injected as a
`<meta name="bds-mcp-base-url">` tag at render time so `callServerTool()` points to the
right endpoint regardless of mode.
`WatchableEngines` is a plain record keyed by entity name (`post`, `media`, `script`,
`template`) mapping to objects that expose `invalidate()`. Started in `main.ts` alongside
`PreviewServer` and `MCPServer`; stopped in the `app.on('before-quit')` handler.
---
### G — `src/cli/bds-mcp.ts` — CLI entrypoint (no Electron imports)
### G — `MCPServer.startCli()` (extend existing `MCPServer.ts`)
The App View's `app.callServerTool()` routes through the MCP host (Claude Desktop, VS
Code, etc.) over the **same** connection the server is on — it does not make a direct HTTP
call to a secondary endpoint. The CLI therefore needs exactly one server and one transport:
```typescript
async startCli(): Promise<void> {
const server = this.createMcpServer();
const transport = new StdioServerTransport();
await server.connect(transport);
// Keep alive until the host closes stdin or the process receives a signal
await new Promise<void>((resolve) => {
process.stdin.on('close', resolve);
process.once('SIGTERM', resolve);
process.once('SIGINT', resolve);
});
await server.close();
}
```
No secondary HTTP port. No `bds-mcp-base-url` meta tag. The `ProposalStore` is owned by
the single `MCPServer` instance as before.
---
### H — `src/cli/bds-mcp.ts` — CLI entrypoint (no Electron imports)
```
1. platformConfigPath() → userData dir
2. getDatabase(path.join(userData, 'bds.db'))
3. SELECT active project (isActive = 1); exit with message if none
4. instantiate engines with CliNotifier DB writer
5. new MCPServer({ postEngine, mediaEngine, … })
6. mcpServer.startCli()
2. migrationsFolder = path.join(__dirname, '../drizzle') // bundled alongside .cjs
3. new DatabaseConnection({ dbPath, migrationsFolder })
4. await db.initializeLocal()
5. SELECT active project (isActive = 1); exit with message if none
6. instantiate engines with CliNotifier DB writer
7. new MCPServer({ postEngine, mediaEngine, … })
8. await mcpServer.startCli()
```
No `electron` imports. Fully standalone Node process.
**Graceful shutdown** — register handlers *before* `startCli()` so cleanup runs on both
normal exit and forced signals:
```typescript
async function shutdown(db: DatabaseConnection, store: ProposalStore): Promise<void> {
store.destroy();
await db.close(); // flushes any in-flight @libsql writes
process.exit(0);
}
process.once('SIGTERM', () => shutdown(db, mcpServer.proposalStore));
process.once('SIGINT', () => shutdown(db, mcpServer.proposalStore));
process.once('beforeExit', () => shutdown(db, mcpServer.proposalStore));
```
(`startCli()` awaits `process.stdin` close / signal and then falls through to the same
cleanup, so shutdown runs exactly once with the `once` guards.)
---
### H — esbuild/Vite build target for CLI bundle
### I — esbuild/Vite build target for CLI bundle
New build target in `vite.config.ts` (or `package.json` scripts):
```
src/cli/bds-mcp.ts → dist/cli/bds-mcp.cjs
target: node, format: cjs, bundle: true
externals: native modules only (better-sqlite3)
```
The project uses `@libsql/client` with platform-specific native binaries (e.g.
`@libsql/darwin-arm64`). These **cannot** be bundled by esbuild. Externalize all
`@libsql/*` packages and ensure the platform binary directories are included alongside
the `.cjs` file in `extraResources`:
```
externals: ['@libsql/client', '@libsql/darwin-arm64', '@libsql/linux-x64-gnu', …]
```
Also bundle the `drizzle/` migrations folder next to `bds-mcp.cjs` so the path resolution
in section H (`path.join(__dirname, '../drizzle')`) works at runtime inside the app bundle.
Runs as part of `npm run build` before the Electron build.
---
### I — electron-builder `extraResources`
### J — electron-builder `extraResources`
```json
"build": {
"extraResources": [
{ "from": "dist/cli/bds-mcp.cjs", "to": "bds-mcp.cjs" }
{ "from": "dist/cli/bds-mcp.cjs", "to": "bds-mcp.cjs" },
{ "from": "drizzle", "to": "drizzle" }
]
}
```
Placed at `Contents/Resources/bds-mcp.cjs` in the macOS app bundle.
`bds-mcp.cjs` is placed at `Contents/Resources/bds-mcp.cjs`; the `drizzle/` migrations
folder is placed at `Contents/Resources/drizzle/` so the CLI migration resolver
(`path.join(__dirname, '../drizzle')`) resolves correctly from `Contents/Resources/`.
---
### J — `MCPAgentConfigEngine` — add Claude Desktop + stdio support + remove
### K — `MCPAgentConfigEngine` — add Claude Desktop + stdio support + remove
#### New agent type
@@ -259,7 +409,7 @@ buttons, so users never have to hunt down config files manually.
---
### K — `ScriptEngine` / `TemplateEngine` — draft lifecycle methods
### L — `ScriptEngine` / `TemplateEngine` — draft lifecycle methods
New methods on each engine:
@@ -277,7 +427,7 @@ deleteDraftTemplate(id: string): Promise<boolean>
---
### L — `ProposalStore` — map proposals to DB row IDs
### M — `ProposalStore` — map proposals to DB row IDs
For `draftPost` proposals the store already maps `proposalId → postId` (the DB row). Extend
this consistently:
@@ -290,35 +440,45 @@ this consistently:
`accept_proposal` dispatches to `publishScript` / `publishTemplate` / `publishPost` etc.
`discard_proposal` dispatches to `deleteDraftScript` / `deleteDraftTemplate` / `deletePost`.
**Known failure mode for in-memory metadata proposals:** if the CLI process is killed after
a `propose_media_metadata` or `propose_post_metadata` call but before accept/discard, the
`proposalId` no longer exists in any `ProposalStore`. The App View buttons will call
`accept_proposal` and receive `{ success: false, message: "Proposal … not found" }`
this is already user-visible in the review UI. No silent corruption occurs; the post or
media item is simply unchanged. This is accepted as the trade-off for keeping metadata
proposals lightweight (no DB rows, no filesystem side effects).
---
## Unresolved Questions
None — all design decisions resolved:
## Design Decisions
| Question | Decision |
|---|---|
| Active project in CLI | Use `isActive = 1` from DB; fail fast if none |
| Proposal durability | Scripts/templates use DB draft rows; metadata proposals stay in-memory |
| App View URL in CLI mode | Injected as `<meta name="bds-mcp-base-url">` at render time |
| Proposal durability | Scripts/templates use DB draft rows; metadata proposals stay in-memory (failure mode documented in M) |
| App View tool calls in CLI mode | `app.callServerTool()` routes through the MCP host, not a direct HTTP call — no secondary port needed |
| Claude Desktop support | stdio via bundled `bds-mcp.cjs` + `ELECTRON_RUN_AS_NODE=1` |
| Remove from config | New `removeFromConfig()` method + Remove buttons in UI |
| DB change detection | `fs.watch()` on `-wal` file + 100 ms debounce; falls back to watching main db file on checkpoint |
| DB coupling to Electron | `DatabaseConnection` accepts explicit paths; Electron-specific path resolution stays in `main.ts` |
---
## Implementation Order (TDD per AGENTS.md)
1. **Migrations** (A + B) — schema changes first, all other work builds on them
2. **`platformConfigPath()`** (C) — pure function, easy to test
3. **`CliNotifier` + DB writer** (D) — unit tests with mock DB
4. **`ScriptEngine`/`TemplateEngine` draft lifecycle** (K) — tests first
5. **`ProposalStore` DB mapping** (L) — extend existing tests
6. **`NotificationWatcher`** (E) — test with mock DB rows
7. **`MCPAgentConfigEngine``claude-desktop` + `removeFromConfig`** (J) — extend existing tests
8. **`MCPServer.startCli()`** (F) — test with in-process transports
9. **`src/cli/bds-mcp.ts`** (G) — integration test: spawn, send `initialize`, assert response
10. **Build target + `extraResources`** (H + I) — verify `npm run build` produces bundle
11. **UI: Remove buttons** — renderer component changes, follows J
12. **`NotificationWatcher` wired in `main.ts`** (E, app side)
2. **Decouple `DatabaseConnection` from Electron** (C0) — prerequisite for all CLI code; update existing DB tests to pass explicit paths
3. **`platformConfigPath()`** (C) — pure function, easy to test
4. **`CliNotifier` + DB writer** (D) — unit tests with mock DB
5. **Engine `invalidate()` API** (E) — add and test on each engine before NotificationWatcher
6. **`ScriptEngine`/`TemplateEngine` draft lifecycle** (L) — tests first
7. **`ProposalStore` DB mapping** (M) — extend existing tests
8. **`NotificationWatcher`** (F) — test with mock DB rows + mock `fs.watch` emitter
9. **`MCPAgentConfigEngine``claude-desktop` + `removeFromConfig`** (K) — extend existing tests
10. **`MCPServer.startCli()`** (G) — test with in-process `StdioServerTransport`
11. **`src/cli/bds-mcp.ts`** (H) — integration test: spawn process, send `initialize`, assert response, assert clean shutdown on SIGTERM
12. **Build target + `extraResources`** (I + J) — verify `npm run build` produces bundle and `drizzle/` is co-located
13. **UI: Remove buttons** — renderer component changes, follows K
14. **`NotificationWatcher` wired in `main.ts`** (F, app side) + stopped in `before-quit`
Each step: write failing test → implement → green → next.