Merge pull request #275 from closedloop-ai/fea-1550-sqlite-pglite-migration

FEA-1550: Cut Agent Dashboard over to PGlite

Author: mikeangstadt

CLAUDE.md

@@ -45,18 +45,15 @@ This rule does NOT apply to internal contracts that ship as a single unit with t
 - **[mistake]**: Never fabricate history in changelogs, commit messages, or comments. Do not claim code "replaces" or "fixes" a prior implementation unless that implementation verifiably exists in the codebase or git history. (context: changelog|hallucination|fabrication)
 - **[mistake]**: Before adding a fallback or recovery path, verify the triggering condition can actually occur. Dead fallbacks that read from files never written or variables never set create false confidence in error handling. (context: dead-code|fallback|unreachable)
 
-### Agent Monitor & Sidecar Security
-- **[mistake]**: Treat localhost sidecar routes and iframe messages as privileged surfaces. Mutating routes need origin/trusted-action guards, explicit target origins, and regression coverage. (context: agent-monitor|sidecar|security)
+### Agent Monitor Security
+- **[mistake]**: Treat localhost Agent Monitor hook/listener routes and renderer IPC as privileged surfaces. Mutating routes need explicit local-only binding, trusted-action guards where applicable, and regression coverage. (context: agent-monitor|listener|ipc|security)
 
 ### Process Spawning & Secrets
 - **[mistake]**: Keep large or sensitive data out of spawned argv/env. Use stdin or files for prompts, quote shell args, set approved cwd, and pass minimal child environments. (context: spawn|argv|env|secrets)
 
 ### Boundary Validation
 - **[mistake]**: Runtime-validate gateway, IPC, and persisted payloads before path or file use. TypeScript casts and preload promise types do not protect missing or null fields. (context: validation|ipc|gateway)
 
-### Generated Agent Monitor Runtime
-- **[pattern]**: When generated sidecar overlays, snippets, or patch inputs change, update stamp/materialization inputs and verify generated output so stale assets or bypassed patches cannot ship. (context: agent-monitor|generated|build)
-
 ### State & Lifecycle
 - **[mistake]**: Setting toggles must update persisted state and in-memory side effects together. Avoid one-way restart guards, stale tray state, or stale cloud presence. (context: settings|lifecycle|state)
 

apps/desktop/CLAUDE.md

@@ -138,15 +138,12 @@ The Diagnostics tab shows the current in-memory gateway log plus a bounded previ
 
 ## Agent Monitor
 
-> **Status (FEA-1504):** Agent Monitor has three boot modes. The default user
-> experience is the legacy sidecar-backed dashboard (`agentMonitorEnabled=true`,
-> `agentDashboardDesignSystemEnabled=false`): pnpm-managed upstream packages are
-> materialized into `.generated/agent-monitor`, shipped unpacked, and rendered in
-> the legacy iframe shell. The in-process design-system dashboard is a Labs
-> opt-in only. When `agentDashboardDesignSystemEnabled` is not the literal
-> boolean `true`, the main process must not load `src/main/database/`,
-> `src/main/collectors/`, `AgentHookListener`, `desktop:db:*`, or the `app://`
-> design renderer path.
+> **Status (FEA-1550):** Agent Monitor is an in-process, PGlite-backed desktop
+> feature. The legacy generated runtime tree, embedded web shell, and
+> `agentDashboardDesignSystemEnabled` boot split have been removed. When
+> `agentMonitorEnabled=false`, the main process must not start collectors,
+> `AgentHookListener`, the `desktop:db:*` IPC handlers, cloud session sync, or
+> dashboard-derived cost reads.
 
 The desktop app provides local Claude Code (and opt-in Codex) session/agent
 observability. It powers the **Dashboard** and the agent nav items (Sessions,
@@ -155,23 +152,20 @@ is gated by the persisted `agentMonitorEnabled` desktop setting, which
 **defaults ON**; when disabled, the agent nav items are hidden and only the
 Gateway section remains.
 
-- **Legacy sidecar (default):** `src/main/agent-monitor-sidecar.ts` launches the
-  generated Claude-Code-Agent-Monitor runtime tree. `build:agent-monitor`
-  materializes the tree from pnpm-managed upstream packages; package/stage logic
-  must keep `.generated/agent-monitor` available for default users.
-- **Design-system runtime (Labs opt-in):** `src/main/agent-dashboard-design-system-runtime.ts`
+- **Runtime:** `src/main/agent-dashboard-design-system-runtime.ts`
   is the only module allowed to import `src/main/database/`,
   `src/main/collectors/`, `AgentHookListener`, or register `desktop:db:*`. It is
-  reached only through `await import()` after boot mode resolves to
-  `design-system`.
-- **Disabled mode:** `agentMonitorEnabled=false` starts no sidecar, no
-  design-system runtime, no dashboard-derived sync source, and no
-  dashboard-derived cost source.
+  reached only through `await import()` after the Agent Monitor setting is
+  enabled.
+- **Disabled mode:** `agentMonitorEnabled=false` starts no design-system
+  runtime, no dashboard-derived sync source, and no dashboard-derived cost
+  source.
 - **Hook listener:** in design-system mode, `src/main/agent-monitor-listener.ts`
   binds `127.0.0.1:4820` in the main process and accepts the hook payload
-  (`POST /api/hooks/event`, `GET /api/health`). Each event is gated by the
-  FEA-1407 sandbox check, harness-stamped from `__provider`, and applied by the
-  lifecycle state machine.
+  (`POST /api/hooks/event`, `GET /api/health`). Each event is
+  harness-stamped from `__provider` and applied by the lifecycle state machine.
+  Local import is ungated — all hook events are written to the local DB
+  regardless of the sandbox directory (FEA-1550).
 - **Collection layer (`src/main/collectors/`):** design-system mode uses
   `CollectorManager` for best-effort boot bulk import and live file watchers for
   all five agent CLIs, writing through the first-party `importSession` into the
@@ -182,12 +176,12 @@ Gateway section remains.
   `~/.claude/settings.json` at install time, so 4820 means hooks need zero
   per-hook env. 4820 is outside `PORT_PROBE_ORDER`, so it never collides with
   the gateway. (FEA-1500 tracks migrating this transport later.)
-- **Durable DB:** `app.getPath("userData")/agent-dashboard.sqlite` (schema in
-  `src/main/database/schema.ts`), Node's built-in `node:sqlite`. Persisted
-  collector caches live under `<userData>/agent-monitor/`.
-- **UI:** a first-party React app in the main window (`src/renderer/`) — NO
-  iframe. The left sidebar drives the **Dashboard** + agent nav items; live
-  updates arrive via the `desktop:db:changed` IPC push after each write.
+- **Durable DB:** `app.getPath("userData")/agent-dashboard.pgdata` (PGlite,
+  schema in `src/main/database/pglite.ts`). Persisted collector caches live
+  under `<userData>/agent-dashboard-ingest/`.
+- **UI:** a first-party React app in the main window (`src/renderer/`). The
+  left sidebar drives the **Dashboard** + agent nav items; live updates arrive
+  via the `desktop:db:changed` IPC push after each write.
 - **Hooks are explicit opt-in (consent-bearing).** The user enables/disables
   tracking via the toggle → `src/main/agent-monitor-hooks.ts` writes/removes the
   hook entries in `~/.claude/settings.json` (and, opt-in, `~/.codex/hooks.json`).
@@ -206,10 +200,12 @@ Gateway section remains.
   tree.
 - **Security model (by design):** the collectors + listener read the agent-CLI
   home dirs (`~/.claude`, `~/.codex`, …) **directly**, outside the gateway
-  `isPathAllowed` sandbox, but every captured session is dropped unless its
-  `cwd` is inside the FEA-1407 sandbox base directory (fail-closed). The listener
-  is bound to `127.0.0.1` only; no cloud egress from collectors. Hooks only
-  mutate global Claude/Codex config on explicit user opt-in and are reversible.
+  `isPathAllowed` sandbox. Local import and hook capture are ungated — all
+  sessions are written to the local PGlite DB regardless of working directory
+  (FEA-1550). Cloud sync sends all local sessions without sandbox filtering;
+  payloads are sanitized before upload. The listener is bound to `127.0.0.1`
+  only; no cloud egress from collectors. Hooks only mutate global Claude/Codex
+  config on explicit user opt-in and are reversible.
 - **Multi-harness support (5 agent tools):** ingests sessions from **Claude
   Code** (hooks live + file historical), **OpenAI Codex** (rollout JSONL under
   `~/.codex/sessions/`), **Cursor** (agent transcripts under `~/.cursor/projects/`),

apps/desktop/electron-builder.yml

@@ -14,20 +14,6 @@ extraResources:
     to: trayIconTemplate.png
   - from: resources/trayIconTemplate@2x.png
     to: trayIconTemplate@2x.png
-  # Generated Claude-Code-Agent-Monitor runtime tree, shipped unpacked
-  # (outside the asar) so the spawned Node server, the built client, and the
-  # hook scripts resolve as real files. Built by scripts/build-agent-monitor.mjs
-  # before packaging (chained into `build`). `client/dist/**/*` (NOT
-  # `client/**/*`) — the server resolves ../client/dist relative to server/, so
-  # the server/ <-> client/dist/ layout must be preserved.
-  - from: .generated/agent-monitor
-    to: agent-monitor
-    filter:
-      - server/**/*
-      - client/dist/**/*
-      - scripts/**/*
-      - package.json
-      - LICENSE
   # First-party agent-monitor hook handlers (FEA-1503), shipped unpacked outside
   # the asar so they resolve as real files. agent-monitor-hooks.ts copies them
   # into userData at install time; the installed hook command runs them via the

apps/desktop/package.json

@@ -1,22 +1,20 @@
 {
   "name": "desktop",
-  "version": "0.15.117",
+  "version": "0.16.0",
   "description": "ClosedLoop Desktop",
   "author": "ClosedLoop AI <support@closedloop.ai>",
   "private": true,
   "type": "module",
   "main": "dist/main/index.js",
   "scripts": {
     "dev": "pnpm build && ELECTRON_BIN=$(bash scripts/patch-electron-plist.sh | tail -1) && \"$ELECTRON_BIN\" .",
-    "start": "pnpm build:agent-monitor && ELECTRON_BIN=$(bash scripts/patch-electron-plist.sh | tail -1) && \"$ELECTRON_BIN\" .",
+    "start": "ELECTRON_BIN=$(bash scripts/patch-electron-plist.sh | tail -1) && \"$ELECTRON_BIN\" .",
     "clean:dist": "node -e \"require('fs').rmSync('dist',{recursive:true,force:true})\"",
     "clean:package": "node -e \"require('fs').rmSync('dist-dmg',{recursive:true,force:true})\"",
     "prebuild": "node -e \"const{execSync:e}=require('child_process'),{writeFileSync:w}=require('fs');const h=e('git rev-parse HEAD').toString().trim();w('src/shared/build-info.ts','// AUTO-GENERATED — do not edit\\nexport const BUILD_COMMIT_HASH = \\\"'+h+'\\\";\\n');\"",
-    "build": "pnpm clean:dist && pnpm prebuild && tsc -p tsconfig.json && pnpm build:renderer && pnpm build:agent-monitor",
+    "build": "pnpm clean:dist && pnpm prebuild && tsc -p tsconfig.json && pnpm build:renderer",
     "build:renderer": "vite build --config vite.renderer.config.ts",
-    "build:agent-monitor": "node scripts/build-agent-monitor.mjs",
     "dashboard:reset": "node scripts/reset-dashboard-db.mjs",
-    "dashboard:reset-packs": "node scripts/reset-dashboard-db.mjs --packs-only",
     "stage:package": "node scripts/stage-packaging-app.mjs",
     "typecheck": "tsc -p tsconfig.json --noEmit && pnpm typecheck:renderer",
     "typecheck:renderer": "tsc -p tsconfig.renderer.json --noEmit",
@@ -25,28 +23,15 @@
     "test:boot:design-system-off": "node --import tsx scripts/assert-design-system-boot-off.mjs",
     "measure:agent-dashboard-storage": "node scripts/measure-agent-dashboard-storage.mjs",
     "verify:electron-binary": "node scripts/ensure-electron-binary.mjs",
-    "test": "tsx --test --test-concurrency=1 test/*.test.ts && node --test \"scripts/agent-monitor-packs/__tests__/*.test.js\" \"scripts/agent-monitor-pull-requests/__tests__/*.test.js\"",
-    "pretest:contract": "pnpm build:agent-monitor",
-    "test:contract": "node --test \"test-e2e/agent-monitor/specs/api-contract/*.test.mjs\"",
-    "pretest:e2e": "pnpm build:agent-monitor",
-    "test:e2e": "playwright test --config test-e2e/agent-monitor/playwright.config.ts",
-    "pretest:audit": "pnpm build:agent-monitor && pnpm audit:scan && pnpm audit:classify",
-    "audit:scan": "node test-e2e/agent-monitor/inventory/scan-tiles.mjs",
-    "audit:classify": "node test-e2e/agent-monitor/inventory/coverage-classifier.mjs",
-    "audit:coverage": "bash scripts/check-audit-coverage.sh",
-    "test:audit": "node --test \"test-e2e/agent-monitor/specs/audit/*.test.mjs\"",
-    "pretest:audit:ui": "pnpm build:agent-monitor",
-    "test:audit:ui": "playwright test --config test-e2e/agent-monitor/playwright.audit.config.ts",
-    "preaudit:report": "pnpm build:agent-monitor",
-    "audit:report": "node test-e2e/agent-monitor/inventory/run-report.mjs",
+    "test": "tsx --test --test-concurrency=1 test/*.test.ts",
     "package": "pnpm clean:package && pnpm build && pnpm stage:package && node scripts/run-electron-builder.mjs",
     "release": "pnpm clean:package && pnpm build && pnpm stage:package && node scripts/run-electron-builder.mjs --publish always"
   },
   "dependencies": {
     "@closedloop-ai/design-system": "0.1.1-dev.26892521643.1",
     "@closedloop-ai/loops-api": ">=0.3.1",
+    "@electric-sql/pglite": "^0.4.6",
     "@pydantic/genai-prices": "0.0.62",
-    "agent-dashboard": "github:hoangsonww/Claude-Code-Agent-Monitor#840c518d7fa69231de049e41b893938228b67e40",
     "busboy": "^1.6.0",
     "electron-log": "^5.4.3",
     "electron-store": "^8.2.0",
@@ -72,7 +57,6 @@
     "@typescript-eslint/eslint-plugin": "^8.57.1",
     "@typescript-eslint/parser": "^8.57.1",
     "@vitejs/plugin-react": "^5.1.3",
-    "agent-dashboard-client": "github:hoangsonww/Claude-Code-Agent-Monitor#840c518d7fa69231de049e41b893938228b67e40&path:/client",
     "autoprefixer": "10.4.20",
     "electron": "^35.0.2",
     "electron-builder": "^26.8.1",