chore: bump version to 10.5.3

Adds a new claude-mem mode tailored for law school study sessions, with
observation types for case holdings, issue patterns, prof frameworks,
doctrine/rule synthesis, argument structures, and cross-case connections.
Includes a chill variant and a CLAUDE.md template for use as a legal
study partner.

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

.claude-plugin/marketplace.json

@@ -10,7 +10,7 @@
   "plugins": [
     {
       "name": "claude-mem",
-      "version": "10.2.4",
+      "version": "10.5.3",
       "source": "./plugin",
       "description": "Persistent memory system for Claude Code - context compression across sessions"
     }

.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-mem",
-  "version": "9.0.6",
+  "version": "10.4.1",
   "description": "Persistent memory system for Claude Code - seamlessly preserve context across sessions",
   "author": {
     "name": "Alex Newman"

.gitignore

@@ -1,6 +1,8 @@
 datasets/
 node_modules/
 dist/
+!installer/dist/
+**/_tree-sitter/
 *.log
 .DS_Store
 .env

.markdownlint.json

@@ -0,0 +1,3 @@
+{
+  "MD013": false
+}

.plan/npx-distribution.md

@@ -0,0 +1,736 @@
+# Plan: NPX Distribution + Universal IDE/CLI Coverage for claude-mem
+
+## Problem
+
+1. **Installation is slow and fragile**: Current install clones the full git repo, runs `npm install`, and builds from source. The npm package already ships pre-built artifacts.
+
+2. **IDE coverage is limited**: claude-mem only supports Claude Code (plugin) and Cursor (hooks installer). The AI coding tools landscape has exploded — Gemini CLI (95k stars), OpenCode (110k stars), Windsurf (~1M users), Codex CLI, Antigravity, Goose, Crush, Copilot CLI, and more all support extensibility.
+
+## Key Insights
+
+- **npm package already has everything**: `plugin/` directory ships pre-built. No git clone or build needed.
+- **Transcript watcher already exists**: `src/services/transcripts/` has a fully built schema-based JSONL tailer. It just needs schemas for more tools.
+- **3 integration tiers exist**: (1) Hook/plugin-based (Claude Code, Gemini CLI, OpenCode, Windsurf, Codex CLI, OpenClaw), (2) MCP-based (Cursor, Copilot CLI, Antigravity, Goose, Crush, Roo Code), (3) Transcript-based (anything with structured log files).
+- **OpenClaw plugin already built**: Full plugin at `openclaw/src/index.ts` (1000+ lines). Needs to be wired into the npx installer.
+- **Gemini CLI is architecturally near-identical to Claude Code**: 11 lifecycle hooks, JSON via stdin/stdout, exit code 0/2 convention, `GEMINI.md` context files, `~/.gemini/settings.json`. This is the easiest high-value integration.
+- **OpenCode has the richest plugin system**: 20+ hook events across 12 categories, JS/TS plugin modules, custom tool creation, MCP support. 110k stars — largest open-source AI CLI.
+- **`npx skills` by Vercel supports 41 agents** — proving the multi-IDE installer UX works. Their agent detection pattern (check if config dir exists) is the right model.
+- **All IDEs share a single worker on port 37777**: One worker serves all integrations. Session source (which IDE) is tracked via the `source` field in hook payloads. No per-IDE worker instances.
+- **This npx CLI fully replaces the old `claude-mem-installer`**: Not a supplement — the complete replacement.
+
+## Solution
+
+`npx claude-mem` becomes a unified CLI: install, configure any IDE, manage the worker, search memory.
+
+```
+npx claude-mem                          # Interactive install + IDE selection
+npx claude-mem install                  # Same as above
+npx claude-mem install --ide windsurf   # Direct IDE setup
+npx claude-mem start / stop / status    # Worker management
+npx claude-mem search <query>           # Search memory from terminal
+npx claude-mem transcript watch         # Start transcript watcher
+```
+
+## Platform Support
+
+**Windows, macOS, and Linux are all first-class targets.** Platform-specific considerations:
+
+- **Config paths**: Use `os.homedir()` and `path.join()` everywhere — never hardcode `/` or `~`
+- **Shebangs**: `#!/usr/bin/env node` for the CLI entry point (cross-platform via Node)
+- **Bun detection**: Check `PATH`, common install locations per platform (`%USERPROFILE%\.bun\bin\bun.exe` on Windows, `~/.bun/bin/bun` on Unix)
+- **File permissions**: `fs.chmod` is a no-op on Windows; don't gate on it
+- **Process management**: Worker start/stop uses signals on Unix, taskkill on Windows — match existing `worker-service.ts` patterns
+- **VS Code paths**: `~/Library/Application Support/Code/` (macOS), `~/.config/Code/` (Linux), `%APPDATA%/Code/` (Windows)
+- **Shell config**: `.bashrc`/`.zshrc` on Unix, PowerShell profile on Windows (for PATH modifications if needed)
+
+---
+
+## Phase 0: Research Findings
+
+### IDE Integration Tiers
+
+**Tier 1 — Native Hook/Plugin Systems** (highest fidelity, real-time capture):
+
+| Tool | Hooks | Config Location | Context Injection | Stars/Users |
+|------|-------|----------------|-------------------|-------------|
+| Claude Code | 5 lifecycle hooks | `~/.claude/settings.json` | CLAUDE.md, plugins | ~25% market |
+| Gemini CLI | 11 lifecycle hooks | `~/.gemini/settings.json` | GEMINI.md | ~95k stars |
+| OpenCode | 20+ event hooks + plugin SDK | `~/.config/opencode/opencode.json` | AGENTS.md + rules dirs | ~110k stars |
+| Windsurf | 11 Cascade hooks | `.windsurf/hooks.json` | `.windsurf/rules/*.md` | ~1M users |
+| Codex CLI | `notify` hook | `~/.codex/config.toml` | `.codex/AGENTS.md`, MCP | Growing (OpenAI) |
+| OpenClaw | 8 event hooks + plugin SDK | `~/.openclaw/openclaw.json` | MEMORY.md sync | ~196k stars |
+
+**Tier 2 — MCP Integration** (tool-based, search + context injection):
+
+| Tool | MCP Support | Config Location | Context Injection |
+|------|------------|----------------|-------------------|
+| Cursor | First-class | `.cursor/mcp.json` | `.cursor/rules/*.mdc` |
+| Copilot CLI | First-class (default MCP) | `~/.copilot/config` | `.github/copilot-instructions.md` |
+| Antigravity | First-class + MCP Store | `~/.gemini/antigravity/mcp_config.json` | `.agent/rules/`, GEMINI.md |
+| Goose | Native MCP (co-developed protocol) | `~/.config/goose/config.yaml` | MCP context |
+| Crush | MCP + Skills | JSON config (charm.land schema) | Skills system |
+| Roo Code | First-class | `.roo/` | `.roo/rules/*.md`, `AGENTS.md` |
+| Warp | MCP + Warp Drive | `WARP.md` + Warp Drive UI | `WARP.md` |
+
+**Tier 3 — Transcript File Watching** (passive, file-based):
+
+| Tool | Transcript Location | Format |
+|------|-------------------|--------|
+| Claude Code | `~/.claude/projects/<proj>/<session>.jsonl` | JSONL |
+| Codex CLI | `~/.codex/sessions/**/*.jsonl` | JSONL |
+| Gemini CLI | `~/.gemini/tmp/<hash>/chats/` | JSON |
+| OpenCode | `.opencode/` (SQLite) | SQLite — needs export |
+
+### What claude-mem Already Has
+
+| Component | Status | Location |
+|-----------|--------|----------|
+| Claude Code plugin | Complete | `plugin/hooks/hooks.json` |
+| Cursor hooks installer | Complete | `src/services/integrations/CursorHooksInstaller.ts` |
+| Platform adapters | Claude Code + Cursor + raw | `src/cli/adapters/` |
+| Transcript watcher | Complete (schema-based JSONL) | `src/services/transcripts/` |
+| Codex transcript schema | Sample exists | `src/services/transcripts/config.ts` |
+| OpenClaw plugin | Complete (1000+ lines) | `openclaw/src/index.ts` |
+| MCP server | Complete | `plugin/scripts/mcp-server.cjs` |
+| Gemini CLI support | Not started | — |
+| OpenCode support | Not started | — |
+| Windsurf support | Not started | — |
+
+### Patterns to Copy
+
+- **Agent detection from `npx skills`** (`vercel-labs/skills/src/agents.ts`): Check if config directory exists
+- **Existing installer logic** (`installer/src/steps/install.ts:29-83`): registerMarketplace, registerPlugin, enablePluginInClaudeSettings — **extract shared logic** from existing installer into reusable modules (DRY with the new CLI)
+- **Bun resolution** (`plugin/scripts/bun-runner.js`): PATH lookup + common locations per platform
+- **CursorHooksInstaller** (`src/services/integrations/CursorHooksInstaller.ts`): Reference implementation for IDE hooks installation
+
+---
+
+## Phase 1: NPX CLI Entry Point
+
+### What to implement
+
+1. **Add `bin` field to `package.json`**:
+   ```json
+   "bin": {
+     "claude-mem": "./dist/cli/index.js"
+   }
+   ```
+
+2. **Create `src/npx-cli/index.ts`** — a Node.js CLI router (NOT Bun) with command categories:
+
+   **Install commands** (pure Node.js, no Bun required):
+   - `npx claude-mem` or `npx claude-mem install` → interactive install (IDE multi-select)
+   - `npx claude-mem install --ide <name>` → direct IDE setup (only for implemented IDEs; unimplemented ones error with "Support for <name> coming soon")
+   - `npx claude-mem update` → update to latest version
+   - `npx claude-mem uninstall` → remove plugin and IDE configs
+   - `npx claude-mem version` → print version
+
+   **Runtime commands** (delegate to Bun via installed plugin):
+   - `npx claude-mem start` → spawns `bun worker-service.cjs start`
+   - `npx claude-mem stop` → spawns `bun worker-service.cjs stop`
+   - `npx claude-mem restart` → spawns `bun worker-service.cjs restart`
+   - `npx claude-mem status` → spawns `bun worker-service.cjs status`
+   - `npx claude-mem search <query>` → hits `GET http://localhost:37777/api/search?q=<query>`
+   - `npx claude-mem transcript watch` → starts transcript watcher
+
+   **Runtime commands must check for installation first**: If plugin directory doesn't exist at `~/.claude/plugins/marketplaces/thedotmack/`, print "claude-mem is not installed. Run: npx claude-mem install" and exit.
+
+3. **The install flow** (fully replaces git clone + build):
+   - Detect the npm package's own location (`import.meta.url` or `__dirname`)
+   - Copy `plugin/` from the npm package to `~/.claude/plugins/marketplaces/thedotmack/`
+   - Copy `plugin/` to `~/.claude/plugins/cache/thedotmack/claude-mem/<version>/`
+   - Register marketplace in `~/.claude/plugins/known_marketplaces.json`
+   - Register plugin in `~/.claude/plugins/installed_plugins.json`
+   - Enable in `~/.claude/settings.json`
+   - Run `npm install` in the marketplace dir (for `@chroma-core/default-embed` — native ONNX binaries, can't be bundled)
+   - Trigger smart-install.js for Bun/uv setup
+   - Run IDE-specific setup for each selected IDE
+
+4. **Interactive IDE selection** (auto-detect + prompt):
+   - Auto-detect installed IDEs by checking config directories
+   - Present multi-select with detected IDEs pre-selected
+   - Detection map:
+     - Claude Code: `~/.claude/` exists
+     - Gemini CLI: `~/.gemini/` exists
+     - OpenCode: `~/.config/opencode/` exists OR `opencode` in PATH
+     - OpenClaw: `~/.openclaw/` exists
+     - Windsurf: `~/.codeium/windsurf/` exists
+     - Codex CLI: `~/.codex/` exists
+     - Cursor: `~/.cursor/` exists
+     - Copilot CLI: `copilot` in PATH (it's a CLI tool, not a config dir)
+     - Antigravity: `~/.gemini/antigravity/` exists
+     - Goose: `~/.config/goose/` exists OR `goose` in PATH
+     - Crush: `crush` in PATH
+     - Roo Code: check for VS Code extension directory containing `roo-code`
+     - Warp: `~/.warp/` exists OR `warp` in PATH
+
+5. **The runtime command routing**:
+   - Locate the installed plugin directory
+   - Find Bun binary (same logic as `bun-runner.js`, platform-aware)
+   - Spawn `bun worker-service.cjs <command>` and pipe stdio through
+   - For `search`: HTTP request to running worker
+
+### Patterns to follow
+
+- `installer/src/steps/install.ts:29-83` for marketplace registration — **extract to shared module**
+- `plugin/scripts/bun-runner.js` for Bun resolution
+- `vercel-labs/skills/src/agents.ts` for IDE auto-detection pattern
+
+### Verification
+
+- `npx claude-mem install` copies plugin to correct directories on macOS, Linux, and Windows
+- Auto-detection finds installed IDEs
+- `npx claude-mem start/stop/status` work after install
+- `npx claude-mem search "test"` returns results
+- `npx claude-mem start` before install prints helpful error message
+- `npx claude-mem update` and `npx claude-mem uninstall` work correctly
+- `npx claude-mem version` prints version
+
+### Anti-patterns
+
+- Do NOT require Bun for install commands — pure Node.js
+- Do NOT clone the git repo
+- Do NOT build from source at install time
+- Do NOT depend on `bun:sqlite` in the CLI entry point
+
+---
+
+## Phase 2: Build Pipeline Integration
+
+### What to implement
+
+1. **Add CLI build step to `scripts/build-hooks.js`**:
+   - Compile `src/npx-cli/index.ts` → `dist/cli/index.js`
+   - Bundle `@clack/prompts` and `picocolors` into the output (self-contained)
+   - Shebang: `#!/usr/bin/env node`
+   - Set executable permissions (no-op on Windows, that's fine)
+
+2. **Move `@clack/prompts` and `picocolors`** to main package.json as dev dependencies (bundled by esbuild into dist/cli/index.js)
+
+3. **Verify `package.json` `files` field**: Currently `["dist", "plugin"]`. `dist/cli/index.js` is already included since it's under `dist/`. No change needed.
+
+4. **Update `prepublishOnly`** to ensure CLI is built before npm publish (already covered — `npm run build` calls `build-hooks.js`)
+
+5. **Pre-build OpenClaw plugin**: Add an esbuild step that compiles `openclaw/src/index.ts` → `openclaw/dist/index.js` so it ships ready-to-use. No `tsc` at install time.
+
+6. **Add `openclaw/dist/` to `package.json` `files` field** (or add `openclaw` if the whole directory should ship)
+
+### Verification
+
+- `npm run build` produces `dist/cli/index.js` with correct shebang
+- `npm run build` produces `openclaw/dist/index.js` pre-built
+- `npm pack` includes both `dist/cli/index.js` and `openclaw/dist/`
+- `node dist/cli/index.js --help` works without Bun
+- Package size is reasonable (check with `npm pack --dry-run`)
+
+---
+
+## Phase 3: Gemini CLI Integration (Tier 1 — Hook-Based)
+
+**Why first among new IDEs**: Near-identical architecture to Claude Code. 11 lifecycle hooks with JSON stdin/stdout, same exit code conventions (0=success, 2=block), `GEMINI.md` context files. 95k GitHub stars. Lowest effort, highest confidence.
+
+### Gemini CLI Hook Events
+
+| Event | Map to claude-mem | Use |
+|-------|-------------------|-----|
+| `SessionStart` | `session-init` | Start tracking session |
+| `BeforeAgent` | `user-prompt` | Capture user prompt |
+| `AfterAgent` | `observation` | Capture full agent response |
+| `BeforeTool` | — | Skip (pre-execution, no result yet) |
+| `AfterTool` | `observation` | Capture tool name + input + response |
+| `BeforeModel` | — | Skip (too low-level, LLM request details) |
+| `AfterModel` | — | Skip (raw LLM response, redundant with AfterAgent) |
+| `BeforeToolSelection` | — | Skip (internal planning step) |
+| `PreCompress` | `summary` | Trigger summary before context compression |
+| `Notification` | — | Skip (system alerts, not session data) |
+| `SessionEnd` | `session-end` | Finalize session |
+
+**Mapped**: 5 of 11 events. **Skipped**: 6 events that are either too low-level (BeforeModel/AfterModel), pre-execution (BeforeTool, BeforeToolSelection), or system-level (Notification).
+
+### Verified Stdin Payload Schemas (from `packages/core/src/hooks/types.ts`)
+
+**Base input (all hooks receive):**
+```typescript
+{ session_id: string, transcript_path: string, cwd: string, hook_event_name: string, timestamp: string }
+```
+
+**Event-specific fields:**
+| Event | Additional Fields |
+|-------|-------------------|
+| `SessionStart` | `source: "startup" \| "resume" \| "clear"` |
+| `SessionEnd` | `reason: "exit" \| "clear" \| "logout" \| "prompt_input_exit" \| "other"` |
+| `BeforeAgent` | `prompt: string` |
+| `AfterAgent` | `prompt: string, prompt_response: string, stop_hook_active: boolean` |
+| `BeforeTool` | `tool_name: string, tool_input: Record<string, unknown>, mcp_context?: McpToolContext, original_request_name?: string` |
+| `AfterTool` | `tool_name: string, tool_input: Record<string, unknown>, tool_response: Record<string, unknown>, mcp_context?: McpToolContext` |
+| `PreCompress` | `trigger: "auto" \| "manual"` |
+| `Notification` | `notification_type: "ToolPermission", message: string, details: Record<string, unknown>` |
+
+**Output (all hooks can return):**
+```typescript
+{ continue?: boolean, stopReason?: string, suppressOutput?: boolean, systemMessage?: string, decision?: "allow" | "deny" | "block" | "approve" | "ask", reason?: string, hookSpecificOutput?: Record<string, unknown> }
+```
+
+**Advisory (non-blocking) hooks:** SessionStart, SessionEnd, PreCompress, Notification — `continue` and `decision` fields are ignored.
+
+**Environment variables provided:** `GEMINI_PROJECT_DIR`, `GEMINI_SESSION_ID`, `GEMINI_CWD`, `CLAUDE_PROJECT_DIR` (compat alias)
+
+### What to implement
+
+1. **Create Gemini CLI platform adapter** at `src/cli/adapters/gemini-cli.ts`:
+   - Normalize Gemini CLI's hook JSON to `NormalizedHookInput`
+   - Base fields always present: `session_id`, `transcript_path`, `cwd`, `hook_event_name`, `timestamp`
+   - Map per event:
+     - `SessionStart`: `source` → session init metadata
+     - `BeforeAgent`: `prompt` → user prompt text
+     - `AfterAgent`: `prompt` + `prompt_response` → full conversation turn
+     - `AfterTool`: `tool_name` + `tool_input` + `tool_response` → observation
+     - `PreCompress`: `trigger` → summary trigger
+     - `SessionEnd`: `reason` → session finalization
+
+2. **Create Gemini CLI hooks installer** at `src/services/integrations/GeminiCliHooksInstaller.ts`:
+   - Write hooks to `~/.gemini/settings.json` under the `hooks` key
+   - Must **merge** with existing settings (read → parse → deep merge → write)
+   - Hook config format (verified against official docs):
+     ```json
+     {
+       "hooks": {
+         "AfterTool": [{
+           "matcher": "*",
+           "hooks": [{ "name": "claude-mem", "type": "command", "command": "<path-to-hook-script>", "timeout": 5000 }]
+         }]
+       }
+     }
+     ```
+   - Note: `matcher` uses regex for tool events, exact string for lifecycle events. `"*"` or `""` matches all.
+   - Hook groups support `sequential: boolean` (default false = parallel execution)
+   - Security: Project-level hooks are fingerprinted — if name/command changes, user is warned
+   - Context injection via `~/.gemini/GEMINI.md` (append claude-mem section with `<claude-mem-context>` tags, same pattern as CLAUDE.md)
+   - Settings hierarchy: project `.gemini/settings.json` > user `~/.gemini/settings.json` > system `/etc/gemini-cli/settings.json`
+
+3. **Register `gemini-cli` in `getPlatformAdapter()`** at `src/cli/adapters/index.ts`
+
+4. **Add Gemini CLI to installer IDE selection**
+
+### Verification
+
+- `npx claude-mem install --ide gemini-cli` merges hooks into `~/.gemini/settings.json`
+- Gemini CLI sessions are captured by the worker
+- `AfterTool` events produce observations with correct `tool_name`, `tool_input`, `tool_response`
+- `GEMINI.md` gets claude-mem context section
+- Existing Gemini CLI settings are preserved (merge, not overwrite)
+- Verify `session_id` from base input is used for session tracking
+
+### Anti-patterns
+
+- Do NOT overwrite `~/.gemini/settings.json` — must deep merge
+- Do NOT map all 11 events — the 6 skipped events would produce noise, not signal
+- Do NOT use `type: "runtime"` — that's for internal extensions only; use `type: "command"`
+- Advisory hooks (SessionStart, SessionEnd, PreCompress, Notification) cannot block — don't set `decision` or `continue` fields on them
+
+---
+
+## Phase 4: OpenCode Integration (Tier 1 — Plugin-Based)
+
+**Why next**: 110k stars, richest plugin ecosystem. OpenCode plugins are JS/TS modules auto-loaded from plugin directories. OpenCode also has a Claude Code compatibility fallback (reads `~/.claude/CLAUDE.md` if no global `AGENTS.md` exists, controllable via `OPENCODE_DISABLE_CLAUDE_CODE_PROMPT=1`).
+
+### Verified Plugin API (from `packages/plugin/src/index.ts`)
+
+**Plugin signature:**
+```typescript
+import { type Plugin, tool } from "@opencode-ai/plugin"
+
+export const ClaudeMemPlugin: Plugin = async (ctx) => {
+  // ctx: { client, project, directory, worktree, serverUrl, $ }
+  return { /* hooks object */ }
+}
+```
+
+**PluginInput type (6 properties, not 4):**
+```typescript
+type PluginInput = {
+  client: ReturnType<typeof createOpencodeClient>  // OpenCode SDK client
+  project: Project                                   // Current project info
+  directory: string                                  // Current working directory
+  worktree: string                                   // Git worktree path
+  serverUrl: URL                                     // Server URL
+  $: BunShell                                        // Bun shell API
+}
+```
+
+**Two hook mechanisms (important distinction):**
+
+1. **Direct interceptor hooks** — keys on the returned `Hooks` object, receive `(input, output)` allowing mutation:
+   - `tool.execute.before`: `(input: { tool, sessionID, callID }, output: { args })`
+   - `tool.execute.after`: `(input: { tool, sessionID, callID, args }, output: { title, output, metadata })`
+   - `shell.env`, `chat.message`, `chat.params`, `chat.headers`, `permission.ask`, `command.execute.before`
+   - Experimental: `experimental.session.compacting`, `experimental.chat.messages.transform`, `experimental.chat.system.transform`
+
+2. **Bus event catch-all** — generic `event` hook, receives `{ event }` where `event.type` is the event name:
+   - `session.created`, `session.compacted`, `session.deleted`, `session.idle`, `session.error`, `session.status`, `session.updated`, `session.diff`
+   - `message.updated`, `message.part.updated`, `message.part.removed`, `message.removed`
+   - `file.edited`, `file.watcher.updated`
+   - `command.executed`, `todo.updated`, `installation.updated`, `server.connected`
+   - `permission.asked`, `permission.replied`
+   - `lsp.client.diagnostics`, `lsp.updated`
+   - `tui.prompt.append`, `tui.command.execute`, `tui.toast.show`
+   - Total: **27 bus events** across **12 categories**
+
+**Custom tool registration (CORRECTED — name is the key, not positional arg):**
+```typescript
+return {
+  tool: {
+    claude_mem_search: tool({
+      description: "Search claude-mem memory database",
+      args: { query: tool.schema.string() },
+      async execute(args, context) {
+        // context: { sessionID, messageID, agent, directory, worktree, abort, metadata, ask }
+        const response = await fetch(`http://localhost:37777/api/search?q=${encodeURIComponent(args.query)}`)
+        return await response.text()
+      },
+    }),
+  },
+}
+```
+
+### What to implement
+
+1. **Create OpenCode plugin** at `src/integrations/opencode-plugin/index.ts`:
+   - Export a `Plugin` function receiving full `PluginInput` context
+   - Use **direct interceptor** `tool.execute.after` for tool observation capture (gives `tool`, `args`, `output`)
+   - Use **bus event catch-all** `event` for session lifecycle:
+
+   | Mechanism | Event | Map to claude-mem |
+   |-----------|-------|-------------------|
+   | interceptor | `tool.execute.after` | `observation` (tool name + args + output) |
+   | bus event | `session.created` | `session-init` |
+   | bus event | `message.updated` | `observation` (assistant messages) |
+   | bus event | `session.compacted` | `summary` |
+   | bus event | `file.edited` | `observation` (file changes) |
+   | bus event | `session.deleted` | `session-end` |
+
+   - Register `claude_mem_search` custom tool using correct `tool({ description, args, execute })` API
+   - Hit `localhost:37777` API endpoints from the plugin
+
+2. **Build the plugin** in the esbuild pipeline → `dist/opencode-plugin/index.js`
+
+3. **Create OpenCode setup in installer** (two options, prefer file-based):
+   - **Option A (file-based):** Copy plugin to `~/.config/opencode/plugins/claude-mem.ts` (auto-loaded at startup)
+   - **Option B (npm-based):** Add to `~/.config/opencode/opencode.json` under `"plugin"` array: `["claude-mem"]`
+   - Config also supports JSONC (`opencode.jsonc`) and legacy `config.json`
+   - Context injection: Append to `~/.config/opencode/AGENTS.md` (or create it) with `<claude-mem-context>` tags
+   - Additional context via `"instructions"` config key (supports file paths, globs, remote URLs)
+
+4. **Add OpenCode to installer IDE selection**
+
+### OpenCode Verification
+
+- `npx claude-mem install --ide opencode` registers the plugin (file or npm)
+- OpenCode loads the plugin on next session
+- `tool.execute.after` interceptor produces observations with `tool`, `args`, `output`
+- Bus events (`session.created`, `session.deleted`) handle session lifecycle
+- `claude_mem_search` custom tool works in OpenCode sessions
+- Context is injected via AGENTS.md
+
+### OpenCode Anti-patterns
+
+- Do NOT try to use OpenCode's `session.diff` for full capture — it's a summary diff, not raw data
+- Do NOT use `tool('name', schema, handler)` — wrong signature. Name is the key in the `tool:{}` map
+- Do NOT assume bus events have the same `(input, output)` mutation pattern — they only receive `{ event }`
+- OpenCode plugins run in Bun — the plugin CAN use Bun APIs (unlike the npx CLI itself)
+- Do NOT hardcode `~/.config/opencode/` — respect `OPENCODE_CONFIG_DIR` env var if set
+
+---
+
+## Phase 5: Windsurf Integration (Tier 1 — Hook-Based)
+
+**Why next**: 11 Cascade hooks, ~1M users. Hook architecture uses JSON stdin with a consistent envelope format.
+
+### Verified Windsurf Hook Events (from docs.windsurf.com/windsurf/cascade/hooks)
+
+**Naming pattern**: `pre_`/`post_` prefix + 5 action categories, plus 2 standalone post-only events.
+
+| Event | Can Block? | Map to claude-mem | Use |
+|-------|-----------|-------------------|-----|
+| `pre_user_prompt` | Yes | `session-init` + `context` | Start session, inject context |
+| `pre_read_code` | Yes | — | Skip (pre-execution, can block file reads) |
+| `post_read_code` | No | — | Skip (too noisy, file reads are frequent) |
+| `pre_write_code` | Yes | — | Skip (pre-execution, can block writes) |
+| `post_write_code` | No | `observation` | Code generation |
+| `pre_run_command` | Yes | — | Skip (pre-execution, can block commands) |
+| `post_run_command` | No | `observation` | Shell command execution |
+| `pre_mcp_tool_use` | Yes | — | Skip (pre-execution, can block MCP calls) |
+| `post_mcp_tool_use` | No | `observation` | MCP tool results |
+| `post_cascade_response` | No | `observation` | Full AI response |
+| `post_setup_worktree` | No | — | Skip (informational) |
+
+**Mapped**: 5 of 11 events (all post-action). **Skipped**: 4 pre-hooks (blocking-capable, pre-execution) + 2 low-value post-hooks.
+
+### Verified Stdin Payload Schema
+
+**Common envelope (all hooks):**
+```json
+{
+  "agent_action_name": "string",
+  "trajectory_id": "string",
+  "execution_id": "string",
+  "timestamp": "ISO 8601 string",
+  "tool_info": { /* event-specific payload */ }
+}
+```
+
+**Event-specific `tool_info` payloads:**
+
+| Event | `tool_info` fields |
+|-------|-------------------|
+| `pre_user_prompt` | `{ user_prompt: string }` |
+| `pre_read_code` / `post_read_code` | `{ file_path: string }` |
+| `pre_write_code` / `post_write_code` | `{ file_path: string, edits: [{ old_string: string, new_string: string }] }` |
+| `pre_run_command` / `post_run_command` | `{ command_line: string, cwd: string }` |
+| `pre_mcp_tool_use` | `{ mcp_server_name: string, mcp_tool_name: string, mcp_tool_arguments: {} }` |
+| `post_mcp_tool_use` | `{ mcp_server_name: string, mcp_tool_name: string, mcp_tool_arguments: {}, mcp_result: string }` |
+| `post_cascade_response` | `{ response: string }` (markdown) |
+| `post_setup_worktree` | `{ worktree_path: string, root_workspace_path: string }` |
+
+**Exit codes:** `0` = success, `2` = block (pre-hooks only; stderr shown to agent), any other = non-blocking warning.
+
+### What to implement
+
+1. **Create Windsurf platform adapter** at `src/cli/adapters/windsurf.ts`:
+   - Normalize Windsurf's hook input format to `NormalizedHookInput`
+   - Common envelope: `agent_action_name`, `trajectory_id`, `execution_id`, `timestamp`, `tool_info`
+   - Map: `trajectory_id` → `sessionId`, `tool_info` fields per event type
+   - For `post_write_code`: `tool_info.file_path` + `tool_info.edits` → file change observation
+   - For `post_run_command`: `tool_info.command_line` + `tool_info.cwd` → command observation
+   - For `post_mcp_tool_use`: `tool_info.mcp_tool_name` + `tool_info.mcp_tool_arguments` + `tool_info.mcp_result` → tool observation
+   - For `post_cascade_response`: `tool_info.response` → full AI response observation
+
+2. **Create Windsurf hooks installer** at `src/services/integrations/WindsurfHooksInstaller.ts`:
+   - Write hooks to `~/.codeium/windsurf/hooks.json` (user-level, for global coverage)
+   - Per-workspace override at `.windsurf/hooks.json` if user chooses workspace-level install
+   - Config format (verified):
+     ```json
+     {
+       "hooks": {
+         "post_write_code": [{
+           "command": "<path-to-hook-script>",
+           "show_output": false,
+           "working_directory": "<optional>"
+         }]
+       }
+     }
+     ```
+   - Note: Tilde expansion (`~`) is NOT supported in `working_directory` — use absolute paths
+   - Merge order: cloud → system → user → workspace (all hooks at all levels execute)
+   - Context injection via `.windsurf/rules/claude-mem-context.md` (workspace-level; Windsurf rules are workspace-scoped)
+   - Rule limits: 6,000 chars per file, 12,000 chars total across all rules
+
+3. **Register `windsurf` in `getPlatformAdapter()`** at `src/cli/adapters/index.ts`
+
+4. **Add Windsurf to installer IDE selection**
+
+### Windsurf Verification
+
+- `npx claude-mem install --ide windsurf` creates hooks config at `~/.codeium/windsurf/hooks.json`
+- Windsurf sessions are captured by the worker via post-action hooks
+- `trajectory_id` is used as session identifier
+- Context is injected via `.windsurf/rules/claude-mem-context.md` (under 6K char limit)
+- Existing hooks.json is preserved (merge, not overwrite)
+
+### Windsurf Anti-patterns
+
+- Do NOT use fabricated event names (`post_search_code`, `post_lint_code`, `on_error`, `pre_tool_execution`) — they don't exist
+- Do NOT assume Windsurf's stdin JSON matches Claude Code's — it uses `tool_info` envelope, not flat fields
+- Do NOT use tilde (`~`) in `working_directory` — not supported, use absolute paths
+- Do NOT exceed 6K chars in the context rule file — Windsurf truncates beyond that
+- Pre-hooks can block actions (exit 2) — only use post-hooks for observation capture
+
+---
+
+## Phase 6: Codex CLI Integration (Tier 1 — Hook + Transcript)
+
+### Dedup strategy
+
+Codex has both a `notify` hook (real-time) and transcript files (complete history). Use **transcript watching only** — it's more complete and avoids the complexity of dual capture paths. The `notify` hook is a simpler mechanism that doesn't provide enough granularity to justify maintaining two integration paths. If transcript watching proves insufficient, add the notify hook later.
+
+### What to implement
+
+1. **Create Codex transcript schema** — the sample in `src/services/transcripts/config.ts` is already production-quality. Verify against current Codex CLI JSONL format and update if needed.
+
+2. **Create Codex setup in installer**:
+   - Write transcript-watch config to `~/.claude-mem/transcript-watch.json`
+   - Set up watch for `~/.codex/sessions/**/*.jsonl` using existing CODEX_SAMPLE_SCHEMA
+   - Context injection via `.codex/AGENTS.md` (Codex reads this natively)
+   - Must merge with existing `config.toml` if it exists (read → parse → merge → write)
+
+3. **Add Codex CLI to installer IDE selection**
+
+### Verification
+
+- `npx claude-mem install --ide codex` creates transcript watch config
+- Codex sessions appear in claude-mem database
+- `AGENTS.md` updated with context after sessions
+- Existing `config.toml` is preserved
+
+---
+
+## Phase 7: OpenClaw Integration (Tier 1 — Plugin-Based)
+
+**Plugin is already fully built** at `openclaw/src/index.ts` (~1000 lines). Has event hooks, SSE observation feed, MEMORY.md sync, slash commands. Only wiring into the installer is needed.
+
+### What to implement
+
+1. **Wire OpenClaw into the npx installer**:
+   - Detect `~/.openclaw/` directory
+   - Copy pre-built plugin from `openclaw/dist/` (built in Phase 2) to OpenClaw plugins location
+   - Register in `~/.openclaw/openclaw.json` under `plugins.claude-mem`
+   - Configure worker port, project name, syncMemoryFile
+   - Optionally prompt for observation feed setup (channel type + target ID)
+
+2. **Add OpenClaw to IDE selection TUI** with hint about messaging channel support
+
+### Verification
+
+- `npx claude-mem install --ide openclaw` registers the plugin
+- OpenClaw gateway loads the plugin on restart
+- Observations are recorded from OpenClaw sessions
+- MEMORY.md syncs to agent workspaces
+
+### Anti-patterns
+
+- Do NOT rebuild the OpenClaw plugin from source at install time — it ships pre-built from Phase 2
+- Do NOT modify the plugin's event handling — it's battle-tested
+
+---
+
+## Phase 8: MCP-Based Integrations (Tier 2)
+
+**These get the MCP server for free** — it already exists at `plugin/scripts/mcp-server.cjs`. The installer just needs to write the right config files per IDE.
+
+MCP-only integrations provide: search tools + context injection. They do NOT capture transcripts or tool usage in real-time.
+
+### What to implement
+
+1. **Copilot CLI MCP setup**:
+   - Write MCP config to `~/.copilot/config` (merge, not overwrite)
+   - Context injection: `.github/copilot-instructions.md`
+   - Detection: `copilot` command in PATH
+
+2. **Antigravity MCP setup**:
+   - Write MCP config to `~/.gemini/antigravity/mcp_config.json` (merge, not overwrite)
+   - Context injection: `~/.gemini/GEMINI.md` (shared with Gemini CLI) and/or `.agent/rules/claude-mem-context.md`
+   - Detection: `~/.gemini/antigravity/` exists
+   - Note: Antigravity has NO hook system — MCP is the only integration path
+
+3. **Goose MCP setup**:
+   - Write MCP config to `~/.config/goose/config.yaml` (YAML merge — use a lightweight YAML parser or write the block manually if config doesn't exist)
+   - Detection: `~/.config/goose/` exists OR `goose` in PATH
+   - Note: Goose co-developed MCP with Anthropic, so MCP support is excellent
+
+4. **Crush MCP setup**:
+   - Write MCP config to Crush's JSON config
+   - Detection: `crush` in PATH
+
+5. **Roo Code MCP setup**:
+   - Write MCP config to `.roo/` or workspace settings
+   - Context injection: `.roo/rules/claude-mem-context.md`
+   - Detection: Check for VS Code extension directory containing `roo-code`
+
+6. **Warp MCP setup**:
+   - Warp uses `WARP.md` in project root for context injection (similar to CLAUDE.md)
+   - MCP servers configured via Warp Drive UI, but also via config files
+   - Detection: `~/.warp/` exists OR `warp` in PATH
+   - Note: Warp is a terminal replacement (~26k stars), not just a CLI tool — multi-agent orchestration with management UI
+
+7. **For each**: Add to installer IDE detection and selection
+
+### Config merging strategy
+
+JSON configs: Read → parse → deep merge → write back. YAML configs (Goose): If file exists, read and append the MCP block. If not, create from template. Avoid pulling in a full YAML parser library — write the MCP block as a string append with proper indentation if the format is predictable.
+
+### Verification
+
+- Each IDE can search claude-mem via MCP tools
+- Context files are written to IDE-specific locations
+- Existing configs are preserved
+
+### Anti-patterns
+
+- MCP-only integrations do NOT capture transcripts — don't claim "full integration"
+- Do NOT overwrite existing config files — always merge
+- Do NOT add a heavy YAML parser dependency for one integration
+
+---
+
+## Phase 9: Remove Old Installer
+
+This is a **full replacement**, not a deprecation.
+
+### What to implement
+
+1. Remove `claude-mem-installer` npm package (unpublish or mark deprecated with message pointing to `npx claude-mem`)
+2. Update `install/public/install.sh` → redirect to `npx claude-mem`
+3. Remove `installer/` directory from the repository (it's replaced by `src/npx-cli/`)
+4. Update docs site to reflect the new install command
+5. Update README.md install instructions
+
+---
+
+## Phase 10: Final Verification
+
+### All platforms (macOS, Linux, Windows)
+
+1. `npm run build` succeeds, produces `dist/cli/index.js` and `openclaw/dist/index.js`
+2. `node dist/cli/index.js install` works clean (no prior install)
+3. Auto-detects installed IDEs correctly per platform
+4. `npx claude-mem start/stop/status/search` all work
+5. `npx claude-mem update` updates correctly
+6. `npx claude-mem uninstall` cleans up all IDE configs
+7. `npx claude-mem version` prints version
+8. `npx claude-mem start` before install shows helpful error
+9. No Bun dependency at install time
+
+### Per-integration verification
+
+| Integration | Type | Captures Sessions | Search via MCP | Context Injection |
+|-------------|------|-------------------|----------------|-------------------|
+| Claude Code | Plugin | Yes (hooks) | Yes | CLAUDE.md |
+| Gemini CLI | Hooks | Yes (AfterTool, AfterAgent) | Yes (via hook) | GEMINI.md |
+| OpenCode | Plugin | Yes (tool.execute.after, message.updated) | Yes (custom tool) | AGENTS.md / rules |
+| Windsurf | Hooks | Yes (post_cascade_response, etc.) | Yes (via hook) | .windsurf/rules/ |
+| Codex CLI | Transcript | Yes (JSONL watcher) | No (passive only) | .codex/AGENTS.md |
+| OpenClaw | Plugin | Yes (event hooks) | Yes (slash commands) | MEMORY.md |
+| Copilot CLI | MCP | No | Yes | copilot-instructions.md |
+| Antigravity | MCP | No | Yes | .agent/rules/ |
+| Goose | MCP | No | Yes | MCP context |
+| Crush | MCP | No | Yes | Skills |
+| Roo Code | MCP | No | Yes | .roo/rules/ |
+| Warp | MCP | No | Yes | WARP.md |
+
+---
+
+## Priority Order & Impact
+
+| Phase | IDE/Tool | Integration Type | Stars/Users | Effort |
+|-------|----------|-----------------|-------------|--------|
+| 1-2 | (infrastructure) | npx CLI + build pipeline | All users | Medium |
+| 3 | Gemini CLI | Hooks (Tier 1) | ~95k stars | Medium (near-identical to Claude Code) |
+| 4 | OpenCode | Plugin (Tier 1) | ~110k stars | Medium (rich plugin SDK) |
+| 5 | Windsurf | Hooks (Tier 1) | ~1M users | Medium |
+| 6 | Codex CLI | Transcript (Tier 3) | Growing (OpenAI) | Low (schema already exists) |
+| 7 | OpenClaw | Plugin (Tier 1) — pre-built | ~196k stars | Low (wire into installer) |
+| 8 | Copilot CLI, Antigravity, Goose, Crush, Warp, Roo Code | MCP (Tier 2) | 20M+ combined | Low per IDE |
+| 9 | (remove old installer) | — | — | Low |
+| 10 | (final verification) | — | — | Low |
+
+## Out of Scope
+
+- **Removing Bun as runtime dependency**: Worker still requires Bun for `bun:sqlite`. Runtime commands delegate to Bun; install commands don't need it.
+- **JetBrains plugin**: Requires Kotlin/Java development — different ecosystem entirely.
+- **Zed extension**: WASM sandbox limits feasibility.
+- **Neovim/Emacs plugins**: Niche audiences, complex plugin ecosystems (Lua/Elisp). Could be added later via MCP (gptel supports it).
+- **Amazon Q / Kiro**: Amazon Q Developer CLI has been sunsetted in favor of Kiro (proprietary, no public extensibility API yet). Revisit when Kiro opens up.
+- **Aider**: Niche audience, writes Markdown transcripts (not JSONL), would require a markdown parser mode in the watcher. Add if demand materializes.
+- **Continue.dev**: Small user base relative to other MCP tools. Can be added as a Tier 2 MCP integration later if requested.
+- **Toad / Qwen Code / Oh-my-pi**: Too early-stage or too niche. Monitor for growth.
+- **OpenClaw plugin development**: The plugin is already complete. Only installer wiring is in scope.

CHANGELOG.md

@@ -2,6 +2,278 @@
 
 All notable changes to claude-mem.
 
+## [v10.5.2] - 2026-02-26
+
+## Smart Explore Benchmark Docs & Skill Update
+
+### Documentation
+- Published smart-explore benchmark report to public docs — full A/B comparison with methodology, raw data tables, quality assessment, and decision framework
+- Added benchmark report to docs.json navigation under Best Practices
+
+### Smart Explore Skill
+- Updated token economics with benchmark-accurate data (11-18x savings on exploration, 4-8x on file understanding)
+- Added "map first" core principle as decision heuristic for tool selection
+- Added AST completeness guarantee to smart_unfold documentation (never truncates, unlike Explore agents)
+- Added Explore agent escalation guidance for multi-file synthesis tasks
+- Updated smart_unfold token range from ~1-7k to ~400-2,100 based on measurements
+- Updated Explore agent token range from ~20-40k to ~39-59k based on measurements
+
+## [v10.5.1] - 2026-02-26
+
+### Bug Fix
+
+- Restored hooks.json to pre-smart-explore configuration (re-adds Setup hook, separate worker start command, PostToolUse matcher)
+
+## [v10.5.0] - 2026-02-26
+
+## Smart Explore: AST-Powered Code Navigation
+
+This release introduces **Smart Explore**, a token-optimized structural code search system built on tree-sitter AST parsing. It applies the same progressive disclosure pattern used in human-readable code outlines — but programmatically, for AI agents.
+
+### Why This Matters
+
+The standard exploration cycle (Glob → Grep → Read) forces agents to consume entire files to understand code structure. A typical 800-line file costs ~12,000 tokens to read. Smart Explore replaces this with a 3-layer progressive disclosure workflow that delivers the same understanding at **6-12x lower token cost**.
+
+### 3 New MCP Tools
+
+- **`smart_search`** — Walks directories, parses all code files via tree-sitter, and returns ranked symbols with signatures and line numbers. Replaces the Glob → Grep discovery cycle in a single call (~2-6k tokens).
+- **`smart_outline`** — Returns the complete structural skeleton of a file: all functions, classes, methods, properties, imports (~1-2k tokens vs ~12k for a full Read).
+- **`smart_unfold`** — Expands a single symbol to its full source code including JSDoc, decorators, and implementation (~1-7k tokens).
+
+### Token Economics
+
+| Approach | Tokens | Savings |
+|----------|--------|---------|
+| smart_outline + smart_unfold | ~3,100 | 8x vs Read |
+| smart_search (cross-file) | ~2,000-6,000 | 6-12x vs Explore agent |
+| Read (full file) | ~12,000+ | baseline |
+| Explore agent | ~20,000-40,000 | baseline |
+
+### Language Support
+
+10 languages via tree-sitter grammars: TypeScript, JavaScript, Python, Rust, Go, Java, C, C++, Ruby, PHP.
+
+### Other Changes
+
+- Simplified hooks configuration
+- Removed legacy setup.sh script
+- Security fix: replaced `execSync` with `execFileSync` to prevent command injection in file path handling
+
+## [v10.4.4] - 2026-02-26
+
+## Fix
+
+- **Remove `save_observation` from MCP tool surface** — This tool was exposed as an MCP tool available to Claude, but it's an internal API-only feature. Removing it from the MCP server prevents unintended tool invocation and keeps the tool surface clean.
+
+## [v10.4.3] - 2026-02-25
+
+## Bug Fixes
+
+- **Fix PostToolUse hook crashes and 5-second latency (#1220)**: Added missing `break` statements to all 7 switch cases in `worker-service.ts` preventing fall-through execution, added `.catch()` on `main()` to handle unhandled promise rejections, and removed redundant `start` commands from hook groups that triggered the 5-second `collectStdin()` timeout
+- **Fix CLAUDE_PLUGIN_ROOT fallback for Stop hooks (#1215)**: Added POSIX shell-level `CLAUDE_PLUGIN_ROOT` fallback in `hooks.json` for environments where the variable isn't injected, added script-level self-resolution via `import.meta.url` in `bun-runner.js`, and regression test added in `plugin-distribution.test.ts`
+
+## Maintenance
+
+- Synced all version files (plugin.json was stuck at 10.4.0)
+
+## [v10.4.2] - 2026-02-25
+
+## Bug Fixes
+
+- **Fix PostToolUse hook crashes and 5-second latency (#1220)**: Added missing `break` statements to all 7 switch cases in `worker-service.ts` preventing fall-through execution, added `.catch()` on `main()` to handle unhandled promise rejections, and removed redundant `start` commands from hook groups that triggered the 5-second `collectStdin()` timeout
+- **Fix CLAUDE_PLUGIN_ROOT fallback for Stop hooks (#1215)**: Added POSIX shell-level `CLAUDE_PLUGIN_ROOT` fallback in `hooks.json` for environments where the variable isn't injected, added script-level self-resolution via `import.meta.url` in `bun-runner.js`, and regression test added in `plugin-distribution.test.ts`
+- **Sync plugin.json version**: Fixed `plugin.json` being stuck at 10.4.0 while other version files were at 10.4.1
+
+## [v10.4.1] - 2026-02-24
+
+### Refactor
+- **Skills Conversion**: Converted `/make-plan` and `/do` commands into first-class skills in `plugin/skills/`.
+- **Organization**: Centralized planning and execution instructions alongside `mem-search`.
+- **Compatibility**: Added symlinks for `openclaw/skills/` to ensure seamless integration with OpenClaw.
+
+### Chore
+- **Version Bump**: Aligned all package and plugin manifests to v10.4.1.
+
+## [v10.4.0] - 2026-02-24
+
+## v10.4.0 — Stability & Platform Hardening
+
+Massive reliability release: 30+ root-cause bug fixes across 10 triage phases, plus new features for agent attribution, Chroma control, and broader platform support.
+
+### New Features
+
+- **Session custom titles** — Agents can now set `custom_title` on sessions for attribution (migration 23, new endpoint)
+- **Chroma toggle** — `CLAUDE_MEM_CHROMA_ENABLED` setting allows SQLite-only fallback mode (#707)
+- **Plugin disabled state** — Early exit check in all hook entry points when plugin is disabled (#781)
+- **Context re-injection guard** — `contextInjected` session flag prevents re-injecting context on every UserPromptSubmit turn (#1079)
+
+### Bug Fixes
+
+#### Data Integrity
+- SHA-256 content-hash deduplication on observation INSERT (migration 22 with backfill + index)
+- Project name collision fix: `getCurrentProjectName()` now returns `parent/basename`
+- Empty project string guard with cwd-derived fallback
+- Stuck `isProcessing` reset: pending work older than 5 minutes auto-clears
+
+#### ChromaDB
+- Python version pinning in uvx args for both local and remote mode (#1196, #1206, #1208)
+- Windows backslash-to-forward-slash path conversion for `--data-dir` (#1199)
+- Metadata sanitization: filter null/undefined/empty values in `addDocuments()` (#1183, #1188)
+- Transport error auto-reconnect in `callTool()` (#1162)
+- Stale transport retry with transparent reconnect (#1131)
+
+#### Hook Lifecycle
+- Suppress `process.stderr.write` in `hookCommand()` to prevent diagnostic output showing as error UI (#1181)
+- Route all `console.error()` through logger instead of stderr
+- Verified all 7 handlers return `suppressOutput: true` (#598, #784)
+
+#### Worker Lifecycle
+- PID file mtime guard prevents concurrent restart storms (#1145)
+- `getInstalledPluginVersion()` ENOENT/EBUSY handling (#1042)
+
+#### SQLite Migrations
+- Schema initialization always creates core tables via `CREATE TABLE IF NOT EXISTS`
+- Migrations 5-7 check actual DB state instead of version tracking (fixes version collision between old/new migration systems, #979)
+- Crash-safe temp table rebuilds
+
+#### Platform Support
+- **Windows**: `cmd.exe /c` uvx spawn, PowerShell `$_` elimination with WQL filtering, `windowsHide: true`, FTS5 runtime probe with fallback (#1190, #1192, #1199, #1024, #1062, #1048, #791)
+- **Cursor IDE**: Adapter field fallbacks, tolerant session-init validation (#838, #1049)
+- **Codex CLI**: `session_id` fallbacks, unknown platform tolerance, undefined guard (#744)
+
+#### API & Infrastructure
+- `/api/logs` OOM fix: tail-read replaces full-file `readFileSync` (64KB expanding chunks, 10MB cap, #1203)
+- CORS: explicit methods and allowedHeaders (#1029)
+- MCP type coercion for batch endpoints: string-to-array for `ids` and `memorySessionIds`
+- Defensive observation error handling returns 200 on recoverable errors instead of 500
+- `.git/` directory write guard on all 4 CLAUDE.md/AGENTS.md write sites (#1165)
+
+#### Stale AbortController Fix
+- `lastGeneratorActivity` timestamp tracking with 30s timeout (#1099)
+- Stale generator detection + abort + restart in `ensureGeneratorRunning`
+- `AbortSignal.timeout(30000)` in `deleteSession` prevents indefinite hang
+
+### Installation
+- `resolveRoot()` replaces hardcoded marketplace path using `CLAUDE_PLUGIN_ROOT` env var (#1128, #1166)
+- `installCLI()` path correction and `verifyCriticalModules()` post-install check
+- Build-time distribution verification for skills, hooks, and plugin manifest (#1187)
+
+### Testing
+- 50+ new tests across hook lifecycle, context re-injection, plugin distribution, migration runner, data integrity, stale abort controller, logs tail-read, CORS, MCP type coercion, and smart-install
+- 68 files changed, ~4200 insertions, ~900 deletions
+
+## [v10.3.3] - 2026-02-23
+
+### Bug Fixes
+
+- Fixed session context footer to reference the claude-mem skill instead of MCP search tools for accessing memories
+
+## [v10.3.2] - 2026-02-23
+
+## Bug Fixes
+
+- **Worker startup readiness**: Worker startup hook now waits for full DB/search readiness before proceeding, fixing the race condition where hooks would fire before the worker was initialized on first start (#1210)
+- **MCP tool naming**: Renamed `save_memory` to `save_observation` for consistency with the observation-based data model (#1210)
+- **MCP search instructions**: Updated MCP server tool descriptions to accurately reflect the 3-layer search workflow (#1210)
+- **Installer hosting**: Serve installer JS from install.cmem.ai instead of GitHub raw URLs for reliability
+- **Installer routing**: Added rewrite rule so install.cmem.ai root path correctly serves the install script
+- **Installer build**: Added compiled installer dist so CLI installation works out of the box
+
+## [v10.3.1] - 2026-02-19
+
+## Fix: Prevent Duplicate Worker Daemons and Zombie Processes
+
+Three root causes of chroma-mcp timeouts identified and fixed:
+
+### PID-based daemon guard
+Exit immediately on startup if PID file points to a live process. Prevents the race condition where hooks firing simultaneously could start multiple daemons before either wrote a PID file.
+
+### Port-based daemon guard
+Exit if port 37777 is already bound — runs before WorkerService constructor registers keepalive signal handlers that previously prevented exit on EADDRINUSE.
+
+### Guaranteed process.exit() after HTTP shutdown
+HTTP shutdown (POST /api/admin/shutdown) now calls `process.exit(0)` in a `try/finally` block. Previously, zombie workers stayed alive after shutdown, and background tasks reconnected to chroma-mcp, spawning duplicate subprocesses contending for the same data directory.
+
+## [v10.3.0] - 2026-02-18
+
+## Replace WASM Embeddings with Persistent chroma-mcp MCP Connection
+
+### Highlights
+
+- **New: ChromaMcpManager** — Singleton stdio MCP client communicating with chroma-mcp via `uvx`, replacing the previous ChromaServerManager (`npx chroma run` + `chromadb` npm + ONNX/WASM)
+- **Eliminates native binary issues** — No more segfaults, WASM embedding failures, or cross-platform install headaches
+- **Graceful subprocess lifecycle** — Wired into GracefulShutdown for clean teardown; zombie process prevention with kill-on-failure and stale `onclose` handler guards
+- **Connection backoff** — 10-second reconnect backoff prevents chroma-mcp spawn storms
+- **SQL injection guards** — Added parameterization to ChromaSync ID exclusion queries
+- **Simplified ChromaSync** — Reduced complexity by delegating embedding concerns to chroma-mcp
+
+### Breaking Changes
+
+None — backward compatible. ChromaDB data is preserved; only the connection mechanism changed.
+
+### Files Changed
+
+- `src/services/sync/ChromaMcpManager.ts` (new) — MCP client singleton
+- `src/services/sync/ChromaServerManager.ts` (deleted) — Old WASM/native approach
+- `src/services/sync/ChromaSync.ts` — Simplified to use MCP client
+- `src/services/worker-service.ts` — Updated startup sequence
+- `src/services/infrastructure/GracefulShutdown.ts` — Subprocess cleanup integration
+
+## [v10.2.6] - 2026-02-18
+
+## Bug Fixes
+
+### Zombie Process Prevention (#1168, #1175)
+
+Observer Claude CLI subprocesses were accumulating as zombies — processes that never exited after their session ended, causing massive resource leaks on long-running systems.
+
+**Root cause:** When observer sessions ended (via idle timeout, abort, or error), the spawned Claude CLI subprocesses were not being reliably killed. The existing `ensureProcessExit()` in `SDKAgent` only covered the happy path; sessions terminated through `SessionRoutes` or `worker-service` bypassed process cleanup entirely.
+
+**Fix — dual-layer approach:**
+
+1. **Immediate cleanup:** Added `ensureProcessExit()` calls to the `finally` blocks in both `SessionRoutes.ts` and `worker-service.ts`, ensuring every session exit path kills its subprocess
+2. **Periodic reaping:** Added `reapStaleSessions()` to `SessionManager` — a background interval that scans `~/.claude-mem/observer-sessions/` for stale PID files, verifies the process is still running, and kills any orphans with SIGKILL escalation
+
+This ensures no observer subprocess survives beyond its session lifetime, even in crash scenarios.
+
+## [v10.2.5] - 2026-02-18
+
+### Bug Fixes
+
+- **Self-healing message queue**: Renamed `claimAndDelete` → `claimNextMessage` with atomic self-healing — automatically resets stale processing messages (>60s) back to pending before claiming, eliminating stuck messages from generator crashes without external timers
+- **Removed redundant idle-timeout reset**: The `resetStaleProcessingMessages()` call during idle timeout in worker-service was removed (startup reset kept), since the atomic self-healing in `claimNextMessage` now handles recovery inline
+- **TypeScript diagnostic fix**: Added `QUEUE` to logger `Component` type
+
+### Tests
+
+- 5 new tests for self-healing behavior (stuck recovery, active protection, atomicity, empty queue, session isolation)
+- 1 new integration test for stuck recovery in zombie-prevention suite
+- All existing queue tests updated for renamed method
+
+## [v10.2.4] - 2026-02-18
+
+## Chroma Vector DB Backfill Fix
+
+Fixes the Chroma backfill system to correctly sync all SQLite observations into the vector database on worker startup.
+
+### Bug Fixes
+
+- **Backfill all projects on startup** — `backfillAllProjects()` now runs on worker startup, iterating all projects in SQLite and syncing missing observations to Chroma. Previously `ensureBackfilled()` existed but was never called, leaving Chroma with incomplete data after cache clears.
+
+- **Fixed critical collection routing bug** — Backfill now uses the shared `cm__claude-mem` collection (matching how DatabaseManager and SearchManager operate) instead of creating per-project orphan collections that no search path reads from.
+
+- **Hardened collection name sanitization** — Project names with special characters (e.g., "YC Stuff") are sanitized for Chroma's naming constraints, including stripping trailing non-alphanumeric characters.
+
+- **Eliminated shared mutable state** — `ensureBackfilled()` and `getExistingChromaIds()` now accept project as a parameter instead of mutating instance state, keeping a single Chroma connection while avoiding fragile property mutation across iterations.
+
+- **Chroma readiness guard** — Backfill waits for Chroma server readiness before running, preventing spurious error logs when Chroma fails to start.
+
+### Changed Files
+
+- `src/services/sync/ChromaSync.ts` — Core backfill logic, sanitization, parameter passing
+- `src/services/worker-service.ts` — Startup backfill trigger + readiness guard
+- `src/utils/logger.ts` — Added `CHROMA_SYNC` log component
+
 ## [v10.2.3] - 2026-02-17
 
 ## Fix Chroma ONNX Model Cache Corruption
@@ -917,505 +1189,3 @@ This appears in both terminal (colored) output and markdown format, including em
 
 - Added detailed analysis reports for GitHub issues #511, #514, #517, #520, #527, #531, #532
 
-## [v8.5.7] - 2026-01-04
-
-## Modular Architecture Refactor
-
-This release refactors the monolithic service architecture into focused, single-responsibility modules with comprehensive test coverage.
-
-### Architecture Improvements
-
-- **SQLite Repositories** (`src/services/sqlite/`) - Modular repositories for sessions, observations, prompts, summaries, and timeline
-- **Worker Agents** (`src/services/worker/agents/`) - Extracted response processing, error handling, and session cleanup
-- **Search Strategies** (`src/services/worker/search/`) - Modular search with Chroma, SQLite, and Hybrid strategies plus orchestrator
-- **Context Generation** (`src/services/context/`) - Separated context building, token calculation, formatters, and renderers
-- **Infrastructure** (`src/services/infrastructure/`) - Graceful shutdown, health monitoring, and process management
-- **Server** (`src/services/server/`) - Express server setup, middleware, and error handling
-
-### Test Coverage
-
-- **595 tests** across 36 test files
-- **1,120 expect() assertions**
-- Coverage for SQLite repos, worker agents, search, context, infrastructure, and server modules
-
-### Session ID Refactor
-
-- Aligned tests with NULL-based memory session initialization pattern
-- Updated `SESSION_ID_ARCHITECTURE.md` documentation
-
-### Other Improvements
-
-- Added missing logger imports to 34 files for better observability
-- Updated esbuild and MCP SDK to latest versions
-- Removed `bun.lock` from version control
-
-**Full Changelog**: https://github.com/thedotmack/claude-mem/compare/v8.5.6...v8.5.7
-
-## [v8.5.6] - 2026-01-04
-
-## Major Architectural Refactoring
-
-Decomposes monolithic services into modular, maintainable components:
-
-### Worker Service
-Extracted infrastructure (GracefulShutdown, HealthMonitor, ProcessManager), server layer (ErrorHandler, Middleware, Server), and integrations (CursorHooksInstaller)
-
-### Context Generator
-Split into ContextBuilder, ContextConfigLoader, ObservationCompiler, TokenCalculator, formatters (Color/Markdown), and section renderers (Header/Footer/Summary/Timeline)
-
-### Search System
-Extracted SearchOrchestrator, ResultFormatter, TimelineBuilder, and strategy pattern (Chroma/SQLite/Hybrid search strategies) with dedicated filters (Date/Project/Type)
-
-### Agent System
-Extracted shared logic into ResponseProcessor, ObservationBroadcaster, FallbackErrorHandler, and SessionCleanupHelper
-
-### SQLite Layer
-Decomposed SessionStore into domain modules (observations, prompts, sessions, summaries, timeline) with proper type exports
-
-## Bug Fixes
-- Fixed duplicate observation storage bug (observations stored multiple times when messages were batched)
-- Added duplicate observation cleanup script for production database remediation
-- Fixed FOREIGN KEY constraint and missing `failed_at_epoch` column issues
-
-## Coming Next
-Comprehensive test suite in a new PR, targeting **v8.6.0**
-
-🤖 Generated with [Claude Code](https://claude.com/claude-code)
-
-## [v8.5.5] - 2026-01-03
-
-## Improved Error Handling and Logging
-
-This patch release enhances error handling and logging across all worker services for better debugging and reliability.
-
-### Changes
-- **Enhanced Error Logging**: Improved error context across SessionStore, SearchManager, SDKAgent, GeminiAgent, and OpenRouterAgent
-- **SearchManager**: Restored error handling for Chroma calls with improved logging
-- **SessionStore**: Enhanced error logging throughout database operations
-- **Bug Fix**: Fixed critical bug where `memory_session_id` could incorrectly equal `content_session_id`
-- **Hooks**: Streamlined error handling and loading states for better maintainability
-
-### Investigation Reports
-- Added detailed analysis documents for generator failures and observation duplication regressions
-
-**Full Changelog**: https://github.com/thedotmack/claude-mem/compare/v8.5.4...v8.5.5
-
-## [v8.5.4] - 2026-01-02
-
-## Bug Fixes
-
-### Chroma Connection Error Handling
-Fixed a critical bug in ChromaSync where connection-related errors were misinterpreted as missing collections. The `ensureCollection()` method previously caught ALL errors and assumed they meant the collection doesn't exist, which caused connection errors to trigger unnecessary collection creation attempts. Now connection-related errors like "Not connected" are properly distinguished and re-thrown immediately, preventing false error handling paths and inappropriate fallback behavior.
-
-### Removed Dead last_user_message Code
-Cleaned up dead code related to `last_user_message` handling in the summary flow. This field was being extracted from transcripts but never used anywhere - in Claude Code transcripts, "user" type messages are mostly tool_results rather than actual user input, and the user's original request is already stored in the user_prompts table. Removing this unused field eliminates confusing warnings like "Missing last_user_message when queueing summary". Changes span summary-hook, SessionRoutes, SessionManager, interface definitions, and all agent implementations.
-
-## Improvements
-
-### Enhanced Error Handling Across Services
-Comprehensive improvement to error handling across 8 core services:
-- **BranchManager** - Now logs recovery checkout failures
-- **PaginationHelper** - Logs when file paths are plain strings instead of valid JSON
-- **SDKAgent** - Enhanced logging for Claude executable detection failures
-- **SearchManager** - Logs plain string handling for files read and edited
-- **paths.ts** - Improved logging for git root detection failures
-- **timeline-formatting** - Enhanced JSON parsing errors with input previews
-- **transcript-parser** - Logs summary of parse errors after processing
-- **ChromaSync** - Logs full error context before attempting collection creation
-
-### Error Handling Documentation & Tooling
-- Created `error-handling-baseline.txt` establishing baseline error handling practices
-- Documented error handling anti-pattern rules in CLAUDE.md
-- Added `detect-error-handling-antipatterns.ts` script to identify empty catch blocks, improper logging practices, and oversized try-catch blocks
-
-## New Features
-
-### Console Filter Bar with Log Parsing
-Implemented interactive log filtering in the viewer UI:
-- **Structured Log Parsing** - Extracts timestamp, level, component, correlation ID, and message content using regex pattern matching
-- **Level Filtering** - Toggle visibility for DEBUG, INFO, WARN, ERROR log levels
-- **Component Filtering** - Filter by 9 component types: HOOK, WORKER, SDK, PARSER, DB, SYSTEM, HTTP, SESSION, CHROMA
-- **Color-Coded Rendering** - Visual distinction with component-specific icons and log level colors
-- **Special Message Detection** - Recognizes markers like → (dataIn), ← (dataOut), ✓ (success), ✗ (failure), ⏱ (timing), [HAPPY-PATH]
-- **Smart Auto-Scroll** - Maintains scroll position when reviewing older logs
-- **Responsive Design** - Filter bar adapts to smaller screens
-
-## [v8.5.3] - 2026-01-02
-
-# 🛡️ Error Handling Hardening & Developer Tools
-
-Version 8.5.3 introduces comprehensive error handling improvements that prevent silent failures and reduce debugging time from hours to minutes. This release also adds new developer tools for queue management and log monitoring.
-
----
-
-## 🔴 Critical Error Handling Improvements
-
-### The Problem
-A single overly-broad try-catch block caused a **10-hour debugging session** by silently swallowing errors. This pattern was pervasive throughout the codebase, creating invisible failure modes.
-
-### The Solution
-
-**Automated Anti-Pattern Detection** (`scripts/detect-error-handling-antipatterns.ts`)
-- Detects 7 categories of error handling anti-patterns
-- Enforces zero-tolerance policy for empty catch blocks
-- Identifies large try-catch blocks (>10 lines) that mask specific errors
-- Flags missing error logging that causes silent failures
-- Supports approved overrides with justification comments
-- Exit code 1 if critical issues detected (enforceable in CI)
-
-**New Error Handling Standards** (Added to `CLAUDE.md`)
-- **5-Question Pre-Flight Checklist**: Required before writing any try-catch
-  1. What SPECIFIC error am I catching?
-  2. Show documentation proving this error can occur
-  3. Why can't this error be prevented?
-  4. What will the catch block DO?
-  5. Why shouldn't this error propagate?
-- **Forbidden Patterns**: Empty catch, catch without logging, large try blocks, promise catch without handlers
-- **Allowed Patterns**: Specific errors, logged failures, minimal scope, explicit recovery
-- **Meta-Rule**: Uncertainty triggers research, NOT try-catch
-
-### Fixes Applied
-
-**Wave 1: Empty Catch Blocks** (5 files)
-- `import-xml-observations.ts` - Log skipped invalid JSON
-- `bun-path.ts` - Log when bun not in PATH
-- `cursor-utils.ts` - Log failed registry reads & corrupt MCP config
-- `worker-utils.ts` - Log failed health checks
-
-**Wave 2: Promise Catches on Critical Paths** (8 locations)
-- `worker-service.ts` - Background initialization failures
-- `SDKAgent.ts` - Session processor errors (2 locations)
-- `GeminiAgent.ts` - Finalization failures (2 locations)
-- `OpenRouterAgent.ts` - Finalization failures (2 locations)
-- `SessionManager.ts` - Generator promise failures
-
-**Wave 3: Comprehensive Audit** (29 catch blocks)
-- Added logging to 16 catch blocks (UI, servers, worker, routes, services)
-- Documented 13 intentional exceptions with justification comments
-- All patterns now follow error handling guidelines with appropriate log levels
-
-### Approved Override System
-
-For justified exceptions (performance-critical paths, expected failures), use:
-```typescript
-// [APPROVED OVERRIDE]: Brief technical justification
-try {
-  // code
-} catch {
-  // allowed exception
-}
-```
-
-**Progress**: 163 anti-patterns → 26 approved overrides (84% reduction in silent failures)
-
----
-
-## 🗂️ Queue Management Features
-
-**New Commands**
-- `npm run queue:clear` - Interactive removal of failed messages
-- `npm run queue:clear -- --all` - Clear all messages (pending, processing, failed)
-- `npm run queue:clear -- --force` - Non-interactive mode
-
-**HTTP API Endpoints**
-- `DELETE /api/pending-queue/failed` - Remove failed messages
-- `DELETE /api/pending-queue/all` - Complete queue reset
-
-Failed messages exceed max retry count and remain for debugging. These commands provide clean queue maintenance.
-
----
-
-## 🪵 Developer Console (Chrome DevTools Style)
-
-**UI Improvements**
-- Bottom drawer console (slides up from bottom-left corner)
-- Draggable resize handle for height adjustment
-- Auto-refresh toggle (2s interval)
-- Clear logs button with confirmation
-- Monospace font (SF Mono/Monaco/Consolas)
-- Minimum height: 150px, adjustable to window height - 100px
-
-**API Endpoints**
-- `GET /api/logs` - Fetch last 1000 lines of current day's log
-- `DELETE /api/logs` - Clear current log file
-
-Logs viewer accessible via floating console button in UI.
-
----
-
-## 📚 Architecture Documentation
-
-**Session ID Architecture** (`docs/SESSION_ID_ARCHITECTURE.md`)
-- Comprehensive documentation of 1:1 session mapping guarantees
-- 19 validation tests proving UNIQUE constraints and resume consistency
-- Documents single-transition vulnerability (application-level enforcement)
-- Complete reference for session lifecycle management
-
----
-
-## 📊 Impact Summary
-
-- **Debugging Time**: 10 hours → minutes (proper error visibility)
-- **Test Coverage**: +19 critical architecture validation tests
-- **Silent Failures**: 84% reduction (163 → 26 approved exceptions)
-- **Protection**: Automated detection prevents regression
-- **Developer UX**: Console logs, queue management, comprehensive docs
-
----
-
-## 🔧 Technical Details
-
-**Files Changed**: 25+ files across error handling, queue management, UI, and documentation
-
-**Critical Path Protection**
-These files now have strict error propagation (no catch-and-continue):
-- `SDKAgent.ts`
-- `GeminiAgent.ts`
-- `OpenRouterAgent.ts`
-- `SessionStore.ts`
-- `worker-service.ts`
-
-**Build Verification**: All changes tested, build successful
-
----
-
-**Full Changelog**: https://github.com/thedotmack/claude-mem/compare/v8.5.2...v8.5.3
-
-## [v8.5.2] - 2025-12-31
-
-## Bug Fixes
-
-### Fixed SDK Agent Memory Leak (#499)
-
-Fixed a critical memory leak where Claude SDK child processes were never terminated after sessions completed. Over extended usage, this caused hundreds of orphaned processes consuming 40GB+ of RAM.
-
-**Root Cause:**
-- When the SDK agent generator completed naturally (no more messages to process), the `AbortController` was never aborted
-- Child processes spawned by the Agent SDK remained running indefinitely
-- Sessions stayed in memory (by design for future events) but underlying processes were never cleaned up
-
-**Fix:**
-- Added proper cleanup to SessionRoutes finally block
-- Now calls `abortController.abort()` when generator completes with no pending work
-- Creates new `AbortController` when crash recovery restarts generators
-- Ensures cleanup happens even if recovery logic fails
-
-**Impact:**
-- Prevents orphaned `claude` processes from accumulating
-- Eliminates multi-gigabyte memory leaks during normal usage
-- Maintains crash recovery functionality with proper resource cleanup
-
-Thanks to @yonnock for the detailed bug report and investigation in #499!
-
-## [v8.5.1] - 2025-12-30
-
-## Bug Fix
-
-**Fixed**: Migration 17 column rename failing for databases in intermediate states (#481)
-
-### Problem
-Migration 17 renamed session ID columns but used a single check to determine if ALL tables were migrated. This caused errors for databases in partial migration states:
-- `no such column: sdk_session_id` (when columns already renamed)
-- `table observations has no column named memory_session_id` (when not renamed)
-
-### Solution
-- Rewrote migration 17 to check **each table individually** before renaming
-- Added `safeRenameColumn()` helper that handles all edge cases gracefully
-- Handles all database states: fresh, old, and partially migrated
-
-### Who was affected
-- Users upgrading from pre-v8.2.6 versions
-- Users whose migration was interrupted (crash, restart, etc.)
-- Users who restored database from backup
-
----
-
-🤖 Generated with [Claude Code](https://claude.com/claude-code)
-
-## [v8.5.0] - 2025-12-30
-
-# Cursor Support Now Available 🎉
-
-This is a major release introducing **full Cursor IDE support**. Claude-mem now works with Cursor, bringing persistent AI memory to Cursor users with or without a Claude Code subscription.
-
-## Highlights
-
-**Give Cursor persistent memory.** Every Cursor session starts fresh - your AI doesn't remember what it worked on yesterday. Claude-mem changes that. Your agent builds cumulative knowledge about your codebase, decisions, and patterns over time.
-
-### Works Without Claude Code
-
-You can now use claude-mem with Cursor using free AI providers:
-- **Gemini** (recommended): 1,500 free requests/day, no credit card required
-- **OpenRouter**: Access to 100+ models including free options
-- **Claude SDK**: For Claude Code subscribers
-
-### Cross-Platform Support
-
-Full support for all major platforms:
-- **macOS**: Bash scripts with `jq` and `curl`
-- **Linux**: Same toolchain as macOS
-- **Windows**: Native PowerShell scripts, no WSL required
-
-## New Features
-
-### Interactive Setup Wizard (`bun run cursor:setup`)
-A guided installer that:
-- Detects your environment (Claude Code present or not)
-- Helps you choose and configure an AI provider
-- Installs Cursor hooks automatically
-- Starts the worker service
-- Verifies everything is working
-
-### Cursor Lifecycle Hooks
-Complete hook integration with Cursor's native hook system:
-- `session-init.sh/.ps1` - Session start with context injection
-- `user-message.sh/.ps1` - User prompt capture
-- `save-observation.sh/.ps1` - Tool usage logging
-- `save-file-edit.sh/.ps1` - File edit tracking
-- `session-summary.sh/.ps1` - Session end summary
-- `context-inject.sh/.ps1` - Load relevant history
-
-### Context Injection via `.cursor/rules`
-Relevant past context is automatically injected into Cursor sessions via the `.cursor/rules/claude-mem-context.mdc` file, giving your AI immediate awareness of prior work.
-
-### Project Registry
-Multi-project support with automatic project detection:
-- Projects registered in `~/.claude-mem/cursor-projects.json`
-- Context automatically scoped to current project
-- Works across multiple workspaces simultaneously
-
-### MCP Search Tools
-Full MCP server integration for Cursor:
-- `search` - Find observations by query, date, type
-- `timeline` - Get context around specific observations
-- `get_observations` - Fetch full details for filtered IDs
-
-## New Commands
-
-| Command | Description |
-|---------|-------------|
-| `bun run cursor:setup` | Interactive setup wizard |
-| `bun run cursor:install` | Install Cursor hooks |
-| `bun run cursor:uninstall` | Remove Cursor hooks |
-| `bun run cursor:status` | Check hook installation status |
-
-## Documentation
-
-Full documentation available at [docs.claude-mem.ai/cursor](https://docs.claude-mem.ai/cursor):
-- Cursor Integration Overview
-- Gemini Setup Guide (free tier)
-- OpenRouter Setup Guide
-- Troubleshooting
-
-## Getting Started
-
-### For Cursor-Only Users (No Claude Code)
-
-```bash
-git clone https://github.com/thedotmack/claude-mem.git
-cd claude-mem && bun install && bun run build
-bun run cursor:setup
-```
-
-### For Claude Code Users
-
-```bash
-/plugin marketplace add thedotmack/claude-mem
-/plugin install claude-mem
-claude-mem cursor install
-```
-
-**Full Changelog**: https://github.com/thedotmack/claude-mem/compare/v8.2.10...v8.5.0
-
-## [v8.2.10] - 2025-12-30
-
-## Bug Fixes
-
-- **Auto-restart worker on version mismatch** (#484): When the plugin updates but the worker was already running on the old version, the worker now automatically restarts instead of failing with 400 errors.
-
-### Changes
-- `/api/version` endpoint now returns the built-in version (compiled at build time) instead of reading from disk
-- `worker-service start` command checks for version mismatch and auto-restarts if needed
-- Downgraded hook version mismatch warning to debug logging (now handled by auto-restart)
-
-Thanks @yungweng for the detailed bug report!
-
-## [v8.2.9] - 2025-12-29
-
-## Bug Fixes
-
-- **Worker Service**: Remove file-based locking and improve Windows stability
-  - Replaced file-based locking with health-check-first approach for cleaner mutual exclusion
-  - Removed AbortSignal.timeout() calls to reduce Bun libuv assertion errors on Windows
-  - Added 500ms shutdown delays on Windows to prevent zombie ports
-  - Reduced hook timeout values for improved responsiveness
-  - Increased worker readiness polling duration from 5s to 15s
-
-## Internal Changes
-
-- Updated worker CLI scripts to reference worker-service.cjs directly
-- Simplified hook command configurations
-
-## [v8.2.8] - 2025-12-29
-
-## Bug Fixes
-
-- Fixed orphaned chroma-mcp processes during shutdown (#489)
-  - Added graceful shutdown handling with signal handlers registered early in WorkerService lifecycle
-  - Ensures ChromaSync subprocess cleanup even when interrupted during initialization
-  - Removes PID file during shutdown to prevent stale process tracking
-
-## Technical Details
-
-This patch release addresses a race condition where SIGTERM/SIGINT signals arriving during ChromaSync initialization could leave orphaned chroma-mcp processes. The fix moves signal handler registration from the start() method to the constructor, ensuring cleanup handlers exist throughout the entire initialization lifecycle.
-
-**Full Changelog**: https://github.com/thedotmack/claude-mem/compare/v8.2.7...v8.2.8
-
-## [v8.2.7] - 2025-12-29
-
-## What's Changed
-
-### Token Optimizations
-- Simplified MCP server tool definitions for reduced token usage
-- Removed outdated troubleshooting and mem-search skill documentation
-- Enhanced search parameter descriptions for better clarity
-- Streamlined MCP workflows for improved efficiency
-
-This release significantly reduces the token footprint of the plugin's MCP tools and documentation.
-
-**Full Changelog**: https://github.com/thedotmack/claude-mem/compare/v8.2.6...v8.2.7
-
-## [v8.2.6] - 2025-12-29
-
-## What's Changed
-
-### Bug Fixes & Improvements
-- Session ID semantic renaming for clarity (content_session_id, memory_session_id)
-- Queue system simplification with unified processing logic
-- Memory session ID capture for agent resume functionality
-- Comprehensive test suite for session ID refactoring
-
-**Full Changelog**: https://github.com/thedotmack/claude-mem/compare/v8.2.5...v8.2.6
-
-## [v8.2.5] - 2025-12-28
-
-## Bug Fixes
-
-- **Logger**: Enhanced Error object handling in debug mode to prevent empty JSON serialization
-- **ChromaSync**: Refactored DatabaseManager to initialize ChromaSync lazily, removing background backfill on startup
-- **SessionManager**: Simplified message handling and removed linger timeout that was blocking completion
-
-## Technical Details
-
-This patch release addresses several issues discovered after the session continuity fix:
-
-1. Logger now properly serializes Error objects with stack traces in debug mode
-2. ChromaSync initialization is now lazy to prevent silent failures during startup
-3. Session linger timeout removed to eliminate artificial 5-second delays on session completion
-
-Full changelog: https://github.com/thedotmack/claude-mem/compare/v8.2.4...v8.2.5
-
-## [v8.2.4] - 2025-12-28
-
-Patch release v8.2.4
-

CLAUDE.md

@@ -14,6 +14,10 @@ Claude-mem is a Claude Code plugin providing persistent memory across sessions.
 
 **Search Skill** (`plugin/skills/mem-search/SKILL.md`) - HTTP API for searching past work, auto-invoked when users ask about history
 
+**Planning Skill** (`plugin/skills/make-plan/SKILL.md`) - Orchestrator instructions for creating phased implementation plans with documentation discovery
+
+**Execution Skill** (`plugin/skills/do/SKILL.md`) - Orchestrator instructions for executing phased plans using subagents
+
 **Chroma** (`src/services/sync/ChromaSync.ts`) - Vector embeddings for semantic search
 
 **Viewer UI** (`src/ui/viewer/`) - React interface at http://localhost:37777, built to `plugin/ui/viewer.html`

README.md

@@ -198,7 +198,7 @@ See [Architecture Overview](https://docs.claude-mem.ai/architecture/overview) fo
 
 ## MCP Search Tools
 
-Claude-Mem provides intelligent memory search through **5 MCP tools** following a token-efficient **3-layer workflow pattern**:
+Claude-Mem provides intelligent memory search through **4 MCP tools** following a token-efficient **3-layer workflow pattern**:
 
 **The 3-Layer Workflow:**
 
@@ -211,7 +211,6 @@ Claude-Mem provides intelligent memory search through **5 MCP tools** following
 - Start with `search` to get an index of results
 - Use `timeline` to see what was happening around specific observations
 - Use `get_observations` to fetch full details for relevant IDs
-- Use `save_memory` to manually store important information
 - **~10x token savings** by filtering before fetching details
 
 **Available MCP Tools:**
@@ -219,8 +218,6 @@ Claude-Mem provides intelligent memory search through **5 MCP tools** following
 1. **`search`** - Search memory index with full-text queries, filters by type/date/project
 2. **`timeline`** - Get chronological context around a specific observation or query
 3. **`get_observations`** - Fetch full observation details by IDs (always batch multiple IDs)
-4. **`save_memory`** - Manually save a memory/observation for semantic search
-5. **`__IMPORTANT`** - Workflow documentation (always visible to Claude)
 
 **Example Usage:**
 
@@ -232,9 +229,6 @@ search(query="authentication bug", type="bugfix", limit=10)
 
 // Step 3: Fetch full details
 get_observations(ids=[123, 456])
-
-// Save important information manually
-save_memory(text="API requires auth header X-API-Key", title="API Auth")
 ```
 
 See [Search Tools Guide](https://docs.claude-mem.ai/usage/search-tools) for detailed examples.

docs/public/docs.json

@@ -62,7 +62,8 @@
         "icon": "lightbulb",
         "pages": [
           "context-engineering",
-          "progressive-disclosure"
+          "progressive-disclosure",
+          "smart-explore-benchmark"
         ]
       },
       {

docs/public/smart-explore-benchmark.mdx

@@ -0,0 +1,196 @@
+---
+title: "Smart Explore Benchmark"
+description: "Token efficiency comparison between AST-based and traditional code exploration"
+---
+
+# Smart Explore Benchmark
+
+Smart Explore uses tree-sitter AST parsing to provide structural code navigation through three MCP tools: `smart_search`, `smart_outline`, and `smart_unfold`. This report documents a rigorous A/B comparison against the standard Explore agent (which uses Glob, Grep, and Read tools) to quantify the token savings and quality trade-offs.
+
+## Executive Summary
+
+| Metric | Smart Explore | Explore Agent | Advantage |
+|--------|:---:|:---:|---|
+| Discovery (cross-file search) | ~14,200 tokens | ~252,500 tokens | **17.8x cheaper** |
+| Targeted reads (specific symbols) | ~5,650 tokens | ~109,400 tokens | **19.4x cheaper** |
+| End-to-end (search + read) | ~4,200 tokens | ~45,000 tokens | **10-12x cheaper** |
+| Completeness | 5/5 full source returned | 4/5 (truncated longest method) | Smart Explore more reliable |
+| Speed | Under 2s per call | 5-66s per call | **10-30x faster** |
+
+## Methodology
+
+### Test Environment
+
+- **Codebase**: claude-mem (`src/` directory, 194 TypeScript files, 1,206 parsed symbols)
+- **Model**: Claude Opus 4.6 for both approaches
+- **Measurement**: Token counts from tool response metadata (`total_tokens` for Explore agents, self-reported `~N tokens for folded view` for Smart Explore)
+
+### Controls
+
+The Explore agents were explicitly instructed: *"Do NOT use smart_search, smart_outline, or smart_unfold tools. Only use Glob, Grep, and Read tools."* This was verified necessary after an initial round where agents opportunistically used the Smart Explore tools, invalidating the comparison.
+
+### Queries
+
+Five queries were selected to represent common exploration tasks:
+
+1. **"session processing"** -- Cross-cutting feature spanning multiple services
+2. **"shutdown"** -- Infrastructure concern touching 6+ files
+3. **"hook registration"** -- Architecture question about plugin system
+4. **"sqlite database"** -- Technology-specific search across the data layer
+5. **"worker-service.ts outline"** -- Single large file (1,225 lines) structural understanding
+
+## Round 1: Discovery
+
+*"What exists and where is it?"* -- Finding relevant files and symbols across the codebase.
+
+### Results
+
+| Query | Smart Explore | Explore Agent | Ratio | Explore Tool Calls |
+|-------|:---:|:---:|:---:|:---:|
+| session processing | ~4,391 t | 51,659 t | **11.8x** | 15 |
+| shutdown | ~3,852 t | 51,523 t | **13.4x** | 18 |
+| hook registration | ~1,930 t | 51,688 t | **26.8x** | 37 |
+| sqlite database | ~2,543 t | 58,633 t | **23.1x** | 16 |
+| worker-service outline | ~1,500 t | 38,973 t | **26.0x** | 15 |
+| **Total** | **~14,216 t** | **252,476 t** | **17.8x** | **101** |
+
+### What Each Returned
+
+**Smart Explore** (1 tool call each): 10 ranked symbols with signatures, line numbers, and JSDoc summaries, plus folded structural views of all matching files showing every function/class/interface with bodies collapsed.
+
+**Explore Agent** (15-37 tool calls each): Synthesized narrative reports with architecture diagrams, design pattern analysis, data flow explanations, complete interface dumps, and file structure maps. Significantly more explanatory prose.
+
+### Analysis
+
+The token gap is widest for narrowly-scoped queries ("hook registration" at 26.8x) because the Explore agent reads multiple full files to find relatively few relevant symbols. For broad queries ("session processing" at 11.8x), more of the file content is relevant, narrowing the ratio.
+
+Smart Explore's consistent 1-tool-call pattern means its cost is predictable. The Explore agent's cost varies with how many files it reads and how much it synthesizes -- ranging from 15 to 37 tool calls for comparable scope.
+
+## Round 2: Targeted Reads
+
+*"Show me this specific function."* -- Reading the implementation of a known symbol after discovery.
+
+Based on the Round 1 results, five specific symbols were selected as natural drill-down targets:
+
+| Target Symbol | File | Lines |
+|---------------|------|:---:|
+| `SessionManager.initializeSession` | services/worker/SessionManager.ts | 135 |
+| `performGracefulShutdown` | services/infrastructure/GracefulShutdown.ts | 48 |
+| `hookCommand` | cli/hook-command.ts | 45 |
+| `DatabaseManager.initialize` | services/sqlite/Database.ts | 27 |
+| `WorkerService.startSessionProcessor` | services/worker-service.ts | 158 |
+
+### Results
+
+| Symbol | Smart Unfold | Explore Agent | Ratio | Completeness |
+|--------|:---:|:---:|:---:|---|
+| initializeSession (135 lines) | ~1,800 t | 27,816 t | **15.5x** | Both returned full source |
+| performGracefulShutdown (48 lines) | ~700 t | 19,621 t | **28.0x** | Both returned full source |
+| hookCommand (45 lines) | ~650 t | 18,680 t | **28.7x** | Both returned full source |
+| DatabaseManager.initialize (27 lines) | ~400 t | 22,334 t | **55.8x** | Both returned full source |
+| startSessionProcessor (158 lines) | ~2,100 t | 20,906 t | **10.0x** | Smart Unfold: complete. Explore: **truncated** |
+| **Total** | **~5,650 t** | **109,357 t** | **19.4x** | |
+
+### Analysis
+
+**The ratio scales inversely with symbol size.** The smallest function (`initialize`, 27 lines) shows the biggest gap at 55.8x because the Explore agent still reads the entire 235-line file to extract 27 lines. The largest method (`startSessionProcessor`, 158 lines) narrows to 10x since more of the file is "useful."
+
+**Smart Unfold returned more complete code.** For the longest method (158 lines), the Explore agent truncated the error handling section with "... error handling continues ...", while `smart_unfold` returned the complete implementation. This is because smart_unfold extracts by AST node boundaries, guaranteeing completeness regardless of symbol size.
+
+**Explore agents add zero unique information for targeted reads.** When you already know the file path and symbol name, the agent's overhead is pure waste -- it reads the file, locates the function, and echoes it back. The only addition is a brief explanatory paragraph.
+
+## Combined Workflow
+
+The realistic workflow is discovery followed by targeted reading. Here is the end-to-end cost comparison for understanding a single function:
+
+### Smart Explore: search + unfold
+
+```
+smart_search("shutdown", path="./src")     ~3,852 tokens
+smart_unfold("GracefulShutdown.ts", "performGracefulShutdown")  ~700 tokens
+────────────────────────────────────────────────────────────────
+Total: ~4,552 tokens (2 tool calls, under 3 seconds)
+```
+
+### Explore Agent: single query
+
+```
+"Find and explain the shutdown logic"      ~51,523 tokens
+────────────────────────────────────────────────────────────────
+Total: ~51,523 tokens (18 tool calls, ~43 seconds)
+```
+
+**End-to-end ratio: 11.3x** -- and the Smart Explore workflow gives you the actual source code, while the Explore agent gives you a prose summary that may paraphrase or truncate.
+
+## Quality Assessment
+
+Neither approach is universally better. They optimize for different outcomes.
+
+### Smart Explore Strengths
+
+- **Predictable cost**: 1 tool call per operation, consistent token ranges
+- **Complete source code**: AST-based extraction guarantees full symbol bodies
+- **Structural context**: Folded views show every symbol in matching files
+- **Speed**: Sub-second responses enable rapid iteration
+- **Composability**: Search, outline, and unfold chain naturally
+
+### Explore Agent Strengths
+
+- **Synthesized understanding**: Produces architecture narratives, data flow diagrams, and design pattern analysis
+- **Cross-cutting explanation**: Connects concepts across files that individual symbol reads cannot
+- **Onboarding quality**: Output reads like documentation, not raw code
+- **Error handling insight**: Identifies edge cases and design decisions that require reading multiple related functions
+- **No prior knowledge needed**: Can answer open-ended questions without knowing file paths or symbol names
+
+### Quality by Task Type
+
+| Task | Better Tool | Why |
+|------|-------------|-----|
+| "Where is X defined?" | Smart Explore | One call, exact answer |
+| "What functions are in this file?" | Smart Explore | Outline returns complete structural map |
+| "Show me this function" | Smart Explore | Unfold returns exact source, never truncates |
+| "How does feature X work end-to-end?" | Explore Agent | Reads multiple files and synthesizes narrative |
+| "What design patterns are used here?" | Explore Agent | Requires reading and interpreting, not just extracting |
+| "Help me understand this codebase" | Explore Agent | Produces onboarding-quality documentation |
+
+## When to Use Which
+
+**Use Smart Explore when:**
+- You know what you are looking for (function name, concept, file)
+- You need source code, not explanation
+- You are iterating quickly (read, modify, read again)
+- Token budget matters (large codebases, long sessions)
+- You need file structure at a glance
+
+**Use the Explore Agent when:**
+- You need synthesized cross-cutting understanding
+- The question is open-ended ("how does this system work?")
+- You are writing documentation or architecture reviews
+- You need to understand *why*, not just *what*
+- You are onboarding to an unfamiliar codebase
+
+**Use both when:**
+- Start with Smart Explore for discovery and navigation
+- Escalate to Explore Agent only for deep analysis that requires multi-file synthesis
+- This hybrid approach captures most of the token savings while preserving access to deep understanding when needed
+
+## Token Economics Reference
+
+| Operation | Tokens | Use Case |
+|-----------|:---:|----------|
+| `smart_search` | 2,000-6,000 | Cross-file symbol discovery |
+| `smart_outline` | 1,000-2,000 | Single file structural map |
+| `smart_unfold` | 400-2,100 | Single symbol full source |
+| `smart_search` + `smart_unfold` | 3,000-8,000 | End-to-end: find and read |
+| Explore Agent (targeted) | 18,000-28,000 | Single function with explanation |
+| Explore Agent (cross-cutting) | 39,000-59,000 | Architecture-level understanding |
+| Read (full file) | 8,000-15,000+ | Complete file contents |
+
+### Savings by Workflow
+
+| Workflow | Smart Explore | Traditional | Savings |
+|----------|:---:|:---:|:---:|
+| Understand one file | outline + unfold (~3,100 t) | Read full file (~12,000 t) | **4x** |
+| Find a function across codebase | search (~3,500 t) | Explore agent (~50,000 t) | **14x** |
+| Find and read a specific function | search + unfold (~4,500 t) | Explore agent (~50,000 t) | **11x** |
+| Navigate a 1,200-line file | outline (~1,500 t) | Read full file (~12,000 t) | **8x** |

install/public/install.sh

@@ -5,7 +5,7 @@ set -euo pipefail
 # Usage: curl -fsSL https://install.cmem.ai | bash
 #   or:  curl -fsSL https://install.cmem.ai | bash -s -- --provider=gemini --api-key=YOUR_KEY
 
-INSTALLER_URL="https://raw.githubusercontent.com/thedotmack/claude-mem/main/installer/dist/index.js"
+INSTALLER_URL="https://install.cmem.ai/installer.js"
 
 # Colors
 RED='\033[0;31m'

install/vercel.json

@@ -1,5 +1,8 @@
 {
   "$schema": "https://openapi.vercel.sh/vercel.json",
+  "rewrites": [
+    { "source": "/", "destination": "/install.sh" }
+  ],
   "headers": [
     {
       "source": "/(.*)\\.sh",

openclaw/openclaw.plugin.json

@@ -3,10 +3,10 @@
   "name": "Claude-Mem (Persistent Memory)",
   "description": "Official OpenClaw plugin for Claude-Mem. Records observations from embedded runner sessions and streams them to messaging channels.",
   "kind": "memory",
-  "version": "1.0.0",
+  "version": "10.4.1",
   "author": "thedotmack",
   "homepage": "https://claude-mem.com",
-  "skills": ["skills/make-plan", "skills/do-plan"],
+  "skills": ["skills/make-plan", "skills/do"],
   "configSchema": {
     "type": "object",
     "additionalProperties": false,

openclaw/skills/do/SKILL.md

@@ -0,0 +1 @@
+../../../plugin/skills/do/SKILL.md

openclaw/skills/make-plan/SKILL.md

@@ -1,63 +0,0 @@
----
-name: make-plan
-description: Create a detailed, phased implementation plan with documentation discovery. Use when asked to plan a feature, task, or multi-step implementation — especially before executing with do-plan.
----
-
-# Make Plan
-
-You are an ORCHESTRATOR. Create an LLM-friendly plan in phases that can be executed consecutively in new chat contexts.
-
-## Delegation Model
-
-Use subagents for *fact gathering and extraction* (docs, examples, signatures, grep results). Keep *synthesis and plan authoring* with the orchestrator (phase boundaries, task framing, final wording). If a subagent report is incomplete or lacks evidence, re-check with targeted reads/greps before finalizing.
-
-### Subagent Reporting Contract (MANDATORY)
-
-Each subagent response must include:
-1. Sources consulted (files/URLs) and what was read
-2. Concrete findings (exact API names/signatures; exact file paths/locations)
-3. Copy-ready snippet locations (example files/sections to copy)
-4. "Confidence" note + known gaps (what might still be missing)
-
-Reject and redeploy the subagent if it reports conclusions without sources.
-
-## Plan Structure
-
-### Phase 0: Documentation Discovery (ALWAYS FIRST)
-
-Before planning implementation, deploy "Documentation Discovery" subagents to:
-1. Search for and read relevant documentation, examples, and existing patterns
-2. Identify the actual APIs, methods, and signatures available (not assumed)
-3. Create a brief "Allowed APIs" list citing specific documentation sources
-4. Note any anti-patterns to avoid (methods that DON'T exist, deprecated parameters)
-
-The orchestrator consolidates findings into a single Phase 0 output.
-
-### Each Implementation Phase Must Include
-
-1. **What to implement** — Frame tasks to COPY from docs, not transform existing code
-   - Good: "Copy the V2 session pattern from docs/examples.ts:45-60"
-   - Bad: "Migrate the existing code to V2"
-2. **Documentation references** — Cite specific files/lines for patterns to follow
-3. **Verification checklist** — How to prove this phase worked (tests, grep checks)
-4. **Anti-pattern guards** — What NOT to do (invented APIs, undocumented params)
-
-### Final Phase: Verification
-
-1. Verify all implementations match documentation
-2. Check for anti-patterns (grep for known bad patterns)
-3. Run tests to confirm functionality
-
-## Key Principles
-
-- Documentation Availability ≠ Usage: Explicitly require reading docs
-- Task Framing Matters: Direct agents to docs, not just outcomes
-- Verify > Assume: Require proof, not assumptions about APIs
-- Session Boundaries: Each phase should be self-contained with its own doc references
-
-## Anti-Patterns to Prevent
-
-- Inventing API methods that "should" exist
-- Adding parameters not in documentation
-- Skipping verification steps
-- Assuming structure without checking examples

openclaw/skills/make-plan/SKILL.md

@@ -0,0 +1 @@
+../../../plugin/skills/make-plan/SKILL.md

openclaw/src/index.ts

@@ -642,6 +642,9 @@ export default function claudeMemPlugin(api: OpenClawPluginApi): void {
     const toolName = event.toolName;
     if (!toolName) return;
 
+    // Skip memory_ tools to prevent recursive observation loops
+    if (toolName.startsWith("memory_")) return;
+
     const contentSessionId = getContentSessionId(ctx.sessionKey);
 
     // Extract result text from all content blocks
@@ -654,6 +657,12 @@ export default function claudeMemPlugin(api: OpenClawPluginApi): void {
         .join("\n");
     }
 
+    // Truncate long responses to prevent oversized payloads
+    const MAX_TOOL_RESPONSE_LENGTH = 1000;
+    if (toolResponseText.length > MAX_TOOL_RESPONSE_LENGTH) {
+      toolResponseText = toolResponseText.slice(0, MAX_TOOL_RESPONSE_LENGTH);
+    }
+
     // Fire-and-forget: send observation + sync MEMORY.md in parallel
     workerPostFireAndForget(workerPort, "/api/sessions/observations", {
       contentSessionId,

package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-mem",
-  "version": "10.2.4",
+  "version": "10.5.3",
   "description": "Memory compression system for Claude Code - persist context across sessions",
   "keywords": [
     "claude",
@@ -98,9 +98,7 @@
   "dependencies": {
     "@anthropic-ai/claude-agent-sdk": "^0.1.76",
     "@modelcontextprotocol/sdk": "^1.25.1",
-    "@chroma-core/default-embed": "^0.1.9",
     "ansi-to-html": "^0.7.2",
-    "chromadb": "^3.2.2",
     "dompurify": "^3.3.1",
     "express": "^4.18.2",
     "glob": "^11.0.3",
@@ -119,6 +117,16 @@
     "@types/react-dom": "^18.3.0",
     "esbuild": "^0.27.2",
     "np": "^11.0.2",
+    "tree-sitter-c": "^0.24.1",
+    "tree-sitter-cli": "^0.26.5",
+    "tree-sitter-cpp": "^0.23.4",
+    "tree-sitter-go": "^0.25.0",
+    "tree-sitter-java": "^0.23.5",
+    "tree-sitter-javascript": "^0.25.0",
+    "tree-sitter-python": "^0.25.0",
+    "tree-sitter-ruby": "^0.23.1",
+    "tree-sitter-rust": "^0.24.0",
+    "tree-sitter-typescript": "^0.23.2",
     "tsx": "^4.20.6",
     "typescript": "^5.3.0"
   }

plugin/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-mem",
-  "version": "10.2.4",
+  "version": "10.5.3",
   "description": "Persistent memory system for Claude Code - seamlessly preserve context across sessions",
   "author": {
     "name": "Alex Newman"

plugin/commands/do.md

@@ -1,43 +0,0 @@
----
-description: "Execute a plan using subagents for implementation"
-argument-hint: "[task or plan reference]"
----
-
-You are an ORCHESTRATOR.
-
-Primary instruction: deploy subagents to execute *all* work for #$ARGUMENTS.
-Do not do the work yourself except to coordinate, route context, and verify that each subagent completed its assigned checklist.
-
-Deploy subagents to execute each phase of #$ARGUMENTS independently and consecutively. For every checklist item below, explicitly deploy (or reuse) a subagent responsible for that item and record its outcome before proceeding.
-
-## Execution Protocol (Orchestrator-Driven)
-
-Orchestrator rules:
-- Each phase uses fresh subagents where noted (or when context is large/unclear).
-- The orchestrator assigns one clear objective per subagent and requires evidence (commands run, outputs, files changed).
-- Do not advance to the next step until the assigned subagent reports completion and the orchestrator confirms it matches the plan.
-
-### During Each Phase:
-Deploy an "Implementation" subagent to:
-1. Execute the implementation as specified
-2. COPY patterns from documentation, don't invent
-3. Cite documentation sources in code comments when using unfamiliar APIs
-4. If an API seems missing, STOP and verify - don't assume it exists
-
-### After Each Phase:
-Deploy subagents for each post-phase responsibility:
-1. **Run verification checklist** - Deploy a "Verification" subagent to prove the phase worked
-2. **Anti-pattern check** - Deploy an "Anti-pattern" subagent to grep for known bad patterns from the plan
-3. **Code quality review** - Deploy a "Code Quality" subagent to review changes
-4. **Commit only if verified** - Deploy a "Commit" subagent *only after* verification passes; otherwise, do not commit
-
-### Between Phases:
-Deploy a "Branch/Sync" subagent to:
-- Push to working branch after each verified phase
-- Prepare the next phase handoff so the next phase's subagents start fresh but have plan context
-
-## Failure Modes to Prevent
-- Don't invent APIs that "should" exist - verify against docs
-- Don't add undocumented parameters - copy exact signatures
-- Don't skip verification - deploy a verification subagent and run the checklist
-- Don't commit before verification passes (or without explicit orchestrator approval)

plugin/commands/make-plan.md

@@ -1,66 +0,0 @@
----
-description: "Create an implementation plan with documentation discovery"
-argument-hint: "[feature or task description]"
----
-
-You are an ORCHESTRATOR.
-
-Create an LLM-friendly plan in phases that can be executed consecutively in new chat contexts.
-
-Delegation model (because subagents can under-report):
-- Use subagents for *fact gathering and extraction* (docs, examples, signatures, grep results).
-- Keep *synthesis and plan authoring* with the orchestrator (phase boundaries, task framing, final wording).
-- If a subagent report is incomplete or lacks evidence, the orchestrator must re-check with targeted reads/greps before finalizing the plan.
-
-Subagent reporting contract (MANDATORY):
-- Each subagent response must include:
-   1) Sources consulted (files/URLs) and what was read
-   2) Concrete findings (exact API names/signatures; exact file paths/locations)
-   3) Copy-ready snippet locations (example files/sections to copy)
-   4) "Confidence" note + known gaps (what might still be missing)
-- Reject and redeploy the subagent if it reports conclusions without sources.
-
-## Plan Structure Requirements
-
-### Phase 0: Documentation Discovery (ALWAYS FIRST)
-Before planning implementation, you MUST:
-Deploy one or more "Documentation Discovery" subagents to:
-1. Search for and read relevant documentation, examples, and existing patterns
-2. Identify the actual APIs, methods, and signatures available (not assumed)
-3. Create a brief "Allowed APIs" list citing specific documentation sources
-4. Note any anti-patterns to avoid (methods that DON'T exist, deprecated parameters)
-
-Then the orchestrator consolidates their findings into a single Phase 0 output.
-
-### Each Implementation Phase Must Include:
-1. **What to implement** - Frame tasks to COPY from docs, not transform existing code
-   - Good: "Copy the V2 session pattern from docs/examples.ts:45-60"
-   - Bad: "Migrate the existing code to V2"
-2. **Documentation references** - Cite specific files/lines for patterns to follow
-3. **Verification checklist** - How to prove this phase worked (tests, grep checks)
-4. **Anti-pattern guards** - What NOT to do (invented APIs, undocumented params)
-
-Subagent-friendly split:
-- Subagents can propose candidate doc references and verification commands.
-- The orchestrator must write the final phase text, ensuring tasks are copy-based, scoped, and independently executable.
-
-### Final Phase: Verification
-1. Verify all implementations match documentation
-2. Check for anti-patterns (grep for known bad patterns)
-3. Run tests to confirm functionality
-
-Delegation guidance:
-- Deploy a "Verification" subagent to draft the checklist and commands.
-- The orchestrator must review the checklist for completeness and ensure it maps to earlier phase outputs.
-
-## Key Principles
-- Documentation Availability ≠ Usage: Explicitly require reading docs
-- Task Framing Matters: Direct agents to docs, not just outcomes
-- Verify > Assume: Require proof, not assumptions about APIs
-- Session Boundaries: Each phase should be self-contained with its own doc references
-
-## Anti-Patterns to Prevent
-- Inventing API methods that "should" exist
-- Adding parameters not in documentation
-- Skipping verification steps
-- Assuming structure without checking examples

plugin/hooks/hooks.json

@@ -7,8 +7,8 @@
         "hooks": [
           {
             "type": "command",
-            "command": "${CLAUDE_PLUGIN_ROOT}/scripts/setup.sh",
-            "timeout": 120
+            "command": "_R=\"${CLAUDE_PLUGIN_ROOT}\"; [ -z \"$_R\" ] && _R=\"$HOME/.claude/plugins/marketplaces/thedotmack/plugin\"; \"$_R/scripts/setup.sh\"",
+            "timeout": 300
           }
         ]
       }
@@ -19,17 +19,22 @@
         "hooks": [
           {
             "type": "command",
-            "command": "node \"${CLAUDE_PLUGIN_ROOT}/scripts/smart-install.js\"",
+            "command": "_R=\"${CLAUDE_PLUGIN_ROOT}\"; [ -z \"$_R\" ] && _R=\"$HOME/.claude/plugins/marketplaces/thedotmack/plugin\"; node \"$_R/scripts/smart-install.js\"",
             "timeout": 300
-          },
+          }
+        ]
+      },
+      {
+        "matcher": "startup|clear|compact",
+        "hooks": [
           {
             "type": "command",
-            "command": "node \"${CLAUDE_PLUGIN_ROOT}/scripts/bun-runner.js\" \"${CLAUDE_PLUGIN_ROOT}/scripts/worker-service.cjs\" start",
+            "command": "_R=\"${CLAUDE_PLUGIN_ROOT}\"; [ -z \"$_R\" ] && _R=\"$HOME/.claude/plugins/marketplaces/thedotmack/plugin\"; node \"$_R/scripts/bun-runner.js\" \"$_R/scripts/worker-service.cjs\" start",
             "timeout": 60
           },
           {
             "type": "command",
-            "command": "node \"${CLAUDE_PLUGIN_ROOT}/scripts/bun-runner.js\" \"${CLAUDE_PLUGIN_ROOT}/scripts/worker-service.cjs\" hook claude-code context",
+            "command": "_R=\"${CLAUDE_PLUGIN_ROOT}\"; [ -z \"$_R\" ] && _R=\"$HOME/.claude/plugins/marketplaces/thedotmack/plugin\"; node \"$_R/scripts/bun-runner.js\" \"$_R/scripts/worker-service.cjs\" hook claude-code context",
             "timeout": 60
           }
         ]
@@ -40,12 +45,7 @@
         "hooks": [
           {
             "type": "command",
-            "command": "node \"${CLAUDE_PLUGIN_ROOT}/scripts/bun-runner.js\" \"${CLAUDE_PLUGIN_ROOT}/scripts/worker-service.cjs\" start",
-            "timeout": 60
-          },
-          {
-            "type": "command",
-            "command": "node \"${CLAUDE_PLUGIN_ROOT}/scripts/bun-runner.js\" \"${CLAUDE_PLUGIN_ROOT}/scripts/worker-service.cjs\" hook claude-code session-init",
+            "command": "_R=\"${CLAUDE_PLUGIN_ROOT}\"; [ -z \"$_R\" ] && _R=\"$HOME/.claude/plugins/marketplaces/thedotmack/plugin\"; node \"$_R/scripts/bun-runner.js\" \"$_R/scripts/worker-service.cjs\" hook claude-code session-init",
             "timeout": 60
           }
         ]
@@ -57,12 +57,7 @@
         "hooks": [
           {
             "type": "command",
-            "command": "node \"${CLAUDE_PLUGIN_ROOT}/scripts/bun-runner.js\" \"${CLAUDE_PLUGIN_ROOT}/scripts/worker-service.cjs\" start",
-            "timeout": 60
-          },
-          {
-            "type": "command",
-            "command": "node \"${CLAUDE_PLUGIN_ROOT}/scripts/bun-runner.js\" \"${CLAUDE_PLUGIN_ROOT}/scripts/worker-service.cjs\" hook claude-code observation",
+            "command": "_R=\"${CLAUDE_PLUGIN_ROOT}\"; [ -z \"$_R\" ] && _R=\"$HOME/.claude/plugins/marketplaces/thedotmack/plugin\"; node \"$_R/scripts/bun-runner.js\" \"$_R/scripts/worker-service.cjs\" hook claude-code observation",
             "timeout": 120
           }
         ]
@@ -73,17 +68,12 @@
         "hooks": [
           {
             "type": "command",
-            "command": "node \"${CLAUDE_PLUGIN_ROOT}/scripts/bun-runner.js\" \"${CLAUDE_PLUGIN_ROOT}/scripts/worker-service.cjs\" start",
-            "timeout": 60
-          },
-          {
-            "type": "command",
-            "command": "node \"${CLAUDE_PLUGIN_ROOT}/scripts/bun-runner.js\" \"${CLAUDE_PLUGIN_ROOT}/scripts/worker-service.cjs\" hook claude-code summarize",
+            "command": "_R=\"${CLAUDE_PLUGIN_ROOT}\"; [ -z \"$_R\" ] && _R=\"$HOME/.claude/plugins/marketplaces/thedotmack/plugin\"; node \"$_R/scripts/bun-runner.js\" \"$_R/scripts/worker-service.cjs\" hook claude-code summarize",
             "timeout": 120
           },
           {
             "type": "command",
-            "command": "node \"${CLAUDE_PLUGIN_ROOT}/scripts/bun-runner.js\" \"${CLAUDE_PLUGIN_ROOT}/scripts/worker-service.cjs\" hook claude-code session-complete",
+            "command": "_R=\"${CLAUDE_PLUGIN_ROOT}\"; [ -z \"$_R\" ] && _R=\"$HOME/.claude/plugins/marketplaces/thedotmack/plugin\"; node \"$_R/scripts/bun-runner.js\" \"$_R/scripts/worker-service.cjs\" hook claude-code session-complete",
             "timeout": 30
           }
         ]

plugin/modes/law-study--chill.json

@@ -0,0 +1,7 @@
+{
+  "name": "Law Study (Chill)",
+  "prompts": {
+    "recording_focus": "WHAT TO RECORD (HIGH SIGNAL ONLY)\n----------------------------------\nOnly record what would be painful to reconstruct later:\n- Issue-spotting triggers: specific fact patterns that signal a testable issue\n- Professor's explicit emphasis, frameworks, or exam tips\n- Counterintuitive holdings or gotchas that contradict intuition\n- Cross-case connections that reframe how a doctrine works\n- A synthesized rule only if it distills something non-obvious from multiple sources\n\nSkip anything that could be looked up in a casebook in under 60 seconds.\n\nUse verbs like: held, established, revealed, distinguished, flagged",
+    "skip_guidance": "WHEN TO SKIP (LIBERAL — WHEN IN DOUBT, SKIP)\n---------------------------------------------\nSkip freely:\n- All case briefs, even condensed ones, unless the holding is counterintuitive\n- Any rule or doctrine stated plainly in the casebook without nuance\n- Definitions of standard legal terms\n- Procedural history\n- Any fact pattern or case that wasn't specifically emphasized by the professor\n- Anything you could find again in under 60 seconds\n- **No output necessary if skipping.**"
+  }
+}

plugin/modes/law-study-CLAUDE.md

@@ -0,0 +1,85 @@
+# Legal Study Assistant
+
+You are a rigorous legal study partner for a law student. Your job is to help them understand the law deeply enough to reason through novel fact patterns independently on exams and in practice.
+
+---
+
+## Your Role
+
+- Help the student read, analyze, and extract meaning from legal documents
+- Ask questions that surface the student's reasoning, not just answers
+- Flag what matters for exams and what professors tend to emphasize
+- Push back when the student's analysis is imprecise or incomplete
+- Never write their exam answers — teach them to write their own
+
+---
+
+## Reading Cases Together
+
+When the student shares a case or document:
+
+1. Read it fully before saying anything. No skimming.
+2. Identify the procedural posture, then the issue, then the holding, then the reasoning.
+3. Separate holding from dicta explicitly — this distinction is always fair game.
+4. Surface ambiguity when the court was evasive. That ambiguity is often the exam question.
+5. Ask: "Which facts were outcome-determinative? What if those facts changed?"
+
+**Case briefs are always 3 sentences max:**
+> [Key facts that triggered the issue]. The court held [holding + extracted rule]. [Why this rule exists or how it fits the doctrine — only if non-obvious.]
+
+---
+
+## Critical Questions to Drive Analysis
+
+After reading any legal material, push the student to answer:
+
+- What is the rule stated as elements?
+- What did the dissent argue and why does it matter?
+- How does this fit — or conflict with — earlier cases?
+- What fact pattern on an exam triggers this rule?
+- What does the professor emphasize about this? Their framing is the exam framing.
+- Is the law settled or contested here?
+
+---
+
+## Issue Spotting
+
+When working through a fact pattern:
+
+1. Read the entire hypo before naming any issues.
+2. List every potential claim and defense — err toward inclusion.
+3. For each issue: rule → application to these specific facts → where the argument turns.
+4. Treat "irrelevant" facts as planted triggers. Nothing in an exam hypo is accidental.
+5. Calibrate to the professor's emphasis — they wrote the exam.
+
+---
+
+## Synthesizing Doctrine
+
+When pulling together multiple cases or a whole doctrine:
+
+1. Find the common principle across all the cases.
+2. Build the rule as a spectrum or taxonomy when cases represent different scenarios.
+3. State the limiting principle — where does this rule stop and why.
+4. Majority rule first, then minority positions with their rationale.
+5. Identify the live tension — what the courts haven't resolved yet.
+
+---
+
+## Tone and Pace
+
+- Be direct. Law school trains precision — model it.
+- When the student is vague, say so and ask them to be specific.
+- Celebrate when they spot something sharp. Legal reasoning is hard.
+- Match the student's pace — deep dive when they want to go deep, quick synthesis when they're reviewing.
+
+---
+
+## Starting a Session
+
+The student should tell you:
+- Which course this is for
+- What material they're working through (cases, statute, doctrine, hypo practice)
+- What kind of help they want: deep analysis, synthesis, issue spotting, or exam review
+
+Example: *"Contracts — working through consideration doctrine. Here are four cases. Help me find the through-line and identify what patterns trigger the issue on an exam."*

plugin/modes/law-study.json

@@ -0,0 +1,120 @@
+{
+  "name": "Law Study",
+  "description": "Legal study and exam preparation for law students",
+  "version": "1.0.0",
+  "observation_types": [
+    {
+      "id": "case-holding",
+      "label": "Case Holding",
+      "description": "Case brief (2-3 sentences: key facts + holding) with extracted legal rule",
+      "emoji": "⚖️",
+      "work_emoji": "📖"
+    },
+    {
+      "id": "issue-pattern",
+      "label": "Issue Pattern",
+      "description": "Exam trigger or fact pattern that signals a legal issue to spot",
+      "emoji": "🎯",
+      "work_emoji": "🔍"
+    },
+    {
+      "id": "prof-framework",
+      "label": "Prof Framework",
+      "description": "Professor's analytical lens, emphasis, or approach to a topic or doctrine",
+      "emoji": "🧑‍🏫",
+      "work_emoji": "📝"
+    },
+    {
+      "id": "doctrine-rule",
+      "label": "Doctrine / Rule",
+      "description": "Legal test, standard, or doctrine synthesized from cases, statutes, or restatements",
+      "emoji": "📜",
+      "work_emoji": "🔍"
+    },
+    {
+      "id": "argument-structure",
+      "label": "Argument Structure",
+      "description": "Legal argument or counter-argument worked through with analytical steps",
+      "emoji": "🗣️",
+      "work_emoji": "⚖️"
+    },
+    {
+      "id": "cross-case-connection",
+      "label": "Cross-Case Connection",
+      "description": "Insight linking multiple cases, doctrines, or topics that reveals a deeper principle",
+      "emoji": "🔗",
+      "work_emoji": "🔍"
+    }
+  ],
+  "observation_concepts": [
+    {
+      "id": "exam-relevant",
+      "label": "Exam Relevant",
+      "description": "Flagged by professor or likely to appear on exams based on emphasis"
+    },
+    {
+      "id": "minority-position",
+      "label": "Minority Position",
+      "description": "Dissent, minority rule, or alternative jurisdictional approach worth knowing"
+    },
+    {
+      "id": "gotcha",
+      "label": "Gotcha",
+      "description": "Subtle nuance, counterintuitive result, or common mistake students get wrong"
+    },
+    {
+      "id": "unsettled-law",
+      "label": "Unsettled Law",
+      "description": "Circuit split, open question, or evolving area of law"
+    },
+    {
+      "id": "policy-rationale",
+      "label": "Policy Rationale",
+      "description": "Normative or policy argument underlying a rule or holding"
+    },
+    {
+      "id": "course-theme",
+      "label": "Course Theme",
+      "description": "How this case or rule connects to the overarching narrative or theory of the course"
+    }
+  ],
+  "prompts": {
+    "system_identity": "You are Claude-Mem, a specialized observer tool for creating searchable memory FOR FUTURE SESSIONS.\n\nCRITICAL: Record what was READ, ANALYZED, SYNTHESIZED, or LEARNED about the law, not what you (the observer) are doing.\n\nYou do not have access to tools. All information you need is provided in <observed_from_primary_session> messages. Create observations from what you observe - no investigation needed.",
+    "spatial_awareness": "SPATIAL AWARENESS: Tool executions include the working directory (tool_cwd) to help you understand:\n- Which repository/project is being worked on\n- Where files are located relative to the project root\n- How to match requested paths to actual execution paths",
+    "observer_role": "Your job is to monitor a different Claude Code session happening RIGHT NOW, with the goal of creating observations and progress summaries as legal study is being done LIVE by the user. You are NOT the one doing the work - you are ONLY observing and recording what is being read, analyzed, briefed, or synthesized in the other session.",
+    "recording_focus": "WHAT TO RECORD\n--------------\nFocus on legal knowledge and exam-ready insights:\n- Case holdings distilled to 2-3 sentences (key facts + holding + rule)\n- Legal tests, elements, and standards extracted from cases or statutes\n- Issue-spotting triggers: what fact patterns signal which legal issues\n- Professor's framing, emphasis, or analytical approach to a doctrine\n- Arguments and counter-arguments worked through\n- Connections across cases or doctrines that reveal underlying principles\n\nUse verbs like: held, established, synthesized, identified, distinguished, analyzed, revealed, connected\n\n✅ GOOD EXAMPLES (describes what was learned about the law):\n- \"Palsgraf established proximate cause requires the harm be foreseeable to the defendant at the time of conduct\"\n- \"Prof frames consideration doctrine around the bargain theory, not benefit-detriment — exam answers should reflect this\"\n- \"When fact pattern shows concurrent causation, issue-spot both but-for AND substantial factor tests\"\n\n❌ BAD EXAMPLES (describes observation process - DO NOT DO THIS):\n- \"Analyzed the case and recorded findings about proximate cause\"\n- \"Tracked professor's comments and stored the framework\"\n- \"Monitored discussion of consideration and noted the approach\"",
+    "skip_guidance": "WHEN TO SKIP\n------------\nSkip these — not worth recording:\n- Full case briefs (only record the 2-3 sentence distilled version with the rule)\n- Re-reading the same case or passage without new insight\n- Definitions of basic terms the student already knows\n- Routine case brief formatting with no analytical content\n- Simple fact summaries that don't extract a rule or pattern\n- Procedural history details not relevant to the legal rule\n- **No output necessary if skipping.**",
+    "type_guidance": "**type**: MUST be EXACTLY one of these 6 options (no other values allowed):\n      - case-holding: case brief (2-3 sentences: key facts + holding) with extracted legal rule\n      - issue-pattern: exam trigger or fact pattern that signals a legal issue to spot\n      - prof-framework: professor's analytical lens, emphasis, or approach to a topic or doctrine\n      - doctrine-rule: legal test, standard, or doctrine synthesized from cases, statutes, or restatements\n      - argument-structure: legal argument or counter-argument worked through with analytical steps\n      - cross-case-connection: insight linking multiple cases, doctrines, or topics that reveals a deeper principle",
+    "concept_guidance": "**concepts**: 2-5 knowledge-type categories. MUST use ONLY these exact keywords:\n      - exam-relevant: flagged by professor or likely to appear on exams\n      - minority-position: dissent, minority rule, or alternative jurisdictional approach\n      - gotcha: subtle nuance, counterintuitive result, or common mistake\n      - unsettled-law: circuit split, open question, or evolving area\n      - policy-rationale: normative or policy argument underlying a rule\n      - course-theme: connects to the overarching narrative or theory of the course\n\n    IMPORTANT: Do NOT include the observation type (case-holding/issue-pattern/etc.) as a concept.\n    Types and concepts are separate dimensions.",
+    "field_guidance": "**facts**: Concise, self-contained statements\nEach fact is ONE piece of information\n      No pronouns - each fact must stand alone\n      Include specific details: case names, rule elements, test names, jurisdiction\n\n**files**: All files or documents read (full paths from project root)",
+    "output_format_header": "OUTPUT FORMAT\n-------------\nOutput observations using this XML structure:",
+    "format_examples": "",
+    "footer": "IMPORTANT! DO NOT do any work right now other than generating this OBSERVATIONS from tool use messages - and remember that you are a memory agent designed to summarize a DIFFERENT claude code session, not this one.\n\nNever reference yourself or your own actions. Do not output anything other than the observation content formatted in the XML structure above. All other output is ignored by the system, and the system has been designed to be smart about token usage. Please spend your tokens wisely on useful observations.\n\nRemember that we record these observations as a way of helping us stay on track with our progress, and to help us keep important decisions and changes at the forefront of our minds! :) Thank you so much for your help!",
+
+    "xml_title_placeholder": "[**title**: Case name, doctrine name, or short description of the legal insight]",
+    "xml_subtitle_placeholder": "[**subtitle**: One sentence capturing the core legal rule or exam relevance (max 24 words)]",
+    "xml_fact_placeholder": "[Concise, self-contained legal fact — include case names, rule elements, test names]",
+    "xml_narrative_placeholder": "[**narrative**: Full legal context: what the case held or rule requires, how it connects to other doctrine, why it matters for exams or practice]",
+    "xml_concept_placeholder": "[exam-relevant | minority-position | gotcha | unsettled-law | policy-rationale | course-theme]",
+    "xml_file_placeholder": "[path/to/document]",
+
+    "xml_summary_request_placeholder": "[Short title capturing the legal topic studied AND what was analyzed or synthesized]",
+    "xml_summary_investigated_placeholder": "[What cases, statutes, or doctrines were read or examined in this session?]",
+    "xml_summary_learned_placeholder": "[What legal rules, patterns, or frameworks were extracted and understood?]",
+    "xml_summary_completed_placeholder": "[What study work was completed? Which cases briefed, which doctrines synthesized, which issue patterns identified?]",
+    "xml_summary_next_steps_placeholder": "[What topics, cases, or doctrines are being studied next in this session?]",
+    "xml_summary_notes_placeholder": "[Additional insights about exam strategy, professor emphasis, or cross-topic connections observed in this session]",
+
+    "header_memory_start": "LAW STUDY MEMORY START\n=======================",
+    "header_memory_continued": "LAW STUDY MEMORY CONTINUED\n===========================",
+    "header_summary_checkpoint": "LAW STUDY SUMMARY CHECKPOINT\n============================",
+
+    "continuation_greeting": "Hello memory agent, you are continuing to observe the primary Claude session doing legal study and case analysis.",
+    "continuation_instruction": "IMPORTANT: Continue generating observations from tool use messages using the XML structure below.",
+
+    "summary_instruction": "Write progress notes of what legal material was studied, what rules and patterns were extracted, and what's next. This is a checkpoint to capture study progress so far. The session is ongoing - more cases or doctrines may be analyzed after this summary. Write \"next_steps\" as the current study trajectory (what topics or cases are actively being worked through), not as post-session plans. Always write at least a minimal summary explaining current progress, even if study is still early, so that users see a summary output tied to each study block.",
+    "summary_context_label": "Claude's Full Response to User:",
+    "summary_format_instruction": "Respond in this XML format:",
+    "summary_footer": "IMPORTANT! DO NOT do any work right now other than generating this next PROGRESS SUMMARY - and remember that you are a memory agent designed to summarize a DIFFERENT claude code session, not this one.\n\nNever reference yourself or your own actions. Do not output anything other than the summary content formatted in the XML structure above. All other output is ignored by the system, and the system has been designed to be smart about token usage. Please spend your tokens wisely on useful summary content.\n\nThank you, this summary will be very useful for keeping track of legal study progress!"
+  }
+}

plugin/package.json

@@ -1,11 +1,20 @@
 {
   "name": "claude-mem-plugin",
-  "version": "10.2.4",
+  "version": "10.5.3",
   "private": true,
   "description": "Runtime dependencies for claude-mem bundled hooks",
   "type": "module",
   "dependencies": {
-    "@chroma-core/default-embed": "^0.1.9"
+    "tree-sitter-cli": "^0.26.5",
+    "tree-sitter-c": "^0.24.1",
+    "tree-sitter-cpp": "^0.23.4",
+    "tree-sitter-go": "^0.25.0",
+    "tree-sitter-java": "^0.23.5",
+    "tree-sitter-javascript": "^0.25.0",
+    "tree-sitter-python": "^0.25.0",
+    "tree-sitter-ruby": "^0.23.1",
+    "tree-sitter-rust": "^0.24.0",
+    "tree-sitter-typescript": "^0.23.2"
   },
   "engines": {
     "node": ">=18.0.0",

plugin/scripts/CLAUDE.md

@@ -1,3 +1,5 @@
+Never read built source files in this directory. These are compiled outputs — read the source files in `src/` instead.
+
 <claude-mem-context>
 # Recent Activity
 

plugin/scripts/bun-runner.js

@@ -12,12 +12,37 @@
  * Fixes #818: Worker fails to start on fresh install
  */
 import { spawnSync, spawn } from 'child_process';
-import { existsSync } from 'fs';
-import { join } from 'path';
+import { existsSync, readFileSync } from 'fs';
+import { join, dirname, resolve } from 'path';
 import { homedir } from 'os';
+import { fileURLToPath } from 'url';
 
 const IS_WINDOWS = process.platform === 'win32';
 
+// Self-resolve plugin root when CLAUDE_PLUGIN_ROOT is not set by Claude Code.
+// Upstream bug: anthropics/claude-code#24529 — Stop hooks (and on Linux, all hooks)
+// don't receive CLAUDE_PLUGIN_ROOT, causing script paths to resolve to /scripts/...
+// which doesn't exist. This fallback derives the plugin root from bun-runner.js's
+// own filesystem location (this file lives in <plugin-root>/scripts/).
+const __bun_runner_dirname = dirname(fileURLToPath(import.meta.url));
+const RESOLVED_PLUGIN_ROOT = process.env.CLAUDE_PLUGIN_ROOT || resolve(__bun_runner_dirname, '..');
+
+/**
+ * Fix script path arguments that were broken by empty CLAUDE_PLUGIN_ROOT.
+ * When CLAUDE_PLUGIN_ROOT is empty, "${CLAUDE_PLUGIN_ROOT}/scripts/foo.cjs"
+ * expands to "/scripts/foo.cjs" which doesn't exist. Detect this and rewrite
+ * the path using our self-resolved plugin root.
+ */
+function fixBrokenScriptPath(argPath) {
+  if (argPath.startsWith('/scripts/') && !existsSync(argPath)) {
+    const fixedPath = join(RESOLVED_PLUGIN_ROOT, argPath);
+    if (existsSync(fixedPath)) {
+      return fixedPath;
+    }
+  }
+  return argPath;
+}
+
 /**
  * Find Bun executable - checks PATH first, then common install locations
  */
@@ -54,6 +79,24 @@ function findBun() {
   return null;
 }
 
+// Early exit if plugin is disabled in Claude Code settings (#781).
+// Sync read + JSON parse — fastest possible check before spawning Bun.
+function isPluginDisabledInClaudeSettings() {
+  try {
+    const configDir = process.env.CLAUDE_CONFIG_DIR || join(homedir(), '.claude');
+    const settingsPath = join(configDir, 'settings.json');
+    if (!existsSync(settingsPath)) return false;
+    const settings = JSON.parse(readFileSync(settingsPath, 'utf-8'));
+    return settings?.enabledPlugins?.['claude-mem@thedotmack'] === false;
+  } catch {
+    return false;
+  }
+}
+
+if (isPluginDisabledInClaudeSettings()) {
+  process.exit(0);
+}
+
 // Get args: node bun-runner.js <script> [args...]
 const args = process.argv.slice(2);
 
@@ -62,6 +105,9 @@ if (args.length === 0) {
   process.exit(1);
 }
 
+// Fix broken script paths caused by empty CLAUDE_PLUGIN_ROOT (#1215)
+args[0] = fixBrokenScriptPath(args[0]);
+
 const bunPath = findBun();
 
 if (!bunPath) {

plugin/scripts/setup.sh

@@ -1,228 +0,0 @@
-#!/usr/bin/env bash
-#
-# claude-mem Setup Hook
-# Ensures dependencies are installed before plugin runs
-#
-
-set -euo pipefail
-
-# Use CLAUDE_PLUGIN_ROOT if available, otherwise detect from script location
-if [[ -z "${CLAUDE_PLUGIN_ROOT:-}" ]]; then
-  SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-  ROOT="$(dirname "$SCRIPT_DIR")"
-else
-  ROOT="$CLAUDE_PLUGIN_ROOT"
-fi
-
-MARKER="$ROOT/.install-version"
-PKG_JSON="$ROOT/package.json"
-
-# Colors (when terminal supports it)
-if [[ -t 2 ]]; then
-  RED='\033[0;31m'
-  GREEN='\033[0;32m'
-  YELLOW='\033[0;33m'
-  BLUE='\033[0;34m'
-  NC='\033[0m' # No Color
-else
-  RED='' GREEN='' YELLOW='' BLUE='' NC=''
-fi
-
-log_info()  { echo -e "${BLUE}ℹ${NC} $*" >&2; }
-log_ok()    { echo -e "${GREEN}✓${NC} $*" >&2; }
-log_warn()  { echo -e "${YELLOW}⚠${NC} $*" >&2; }
-log_error() { echo -e "${RED}✗${NC} $*" >&2; }
-
-#
-# Detect Bun - check PATH and common locations
-#
-find_bun() {
-  # Try PATH first
-  if command -v bun &>/dev/null; then
-    echo "bun"
-    return 0
-  fi
-  
-  # Check common install locations
-  local paths=(
-    "$HOME/.bun/bin/bun"
-    "/usr/local/bin/bun"
-    "/opt/homebrew/bin/bun"
-  )
-  
-  for p in "${paths[@]}"; do
-    if [[ -x "$p" ]]; then
-      echo "$p"
-      return 0
-    fi
-  done
-  
-  return 1
-}
-
-#
-# Detect uv - check PATH and common locations
-#
-find_uv() {
-  # Try PATH first
-  if command -v uv &>/dev/null; then
-    echo "uv"
-    return 0
-  fi
-  
-  # Check common install locations
-  local paths=(
-    "$HOME/.local/bin/uv"
-    "$HOME/.cargo/bin/uv"
-    "/usr/local/bin/uv"
-    "/opt/homebrew/bin/uv"
-  )
-  
-  for p in "${paths[@]}"; do
-    if [[ -x "$p" ]]; then
-      echo "$p"
-      return 0
-    fi
-  done
-  
-  return 1
-}
-
-#
-# Get package.json version
-#
-get_pkg_version() {
-  if [[ -f "$PKG_JSON" ]]; then
-    # Simple grep-based extraction (no jq dependency)
-    grep -o '"version"[[:space:]]*:[[:space:]]*"[^"]*"' "$PKG_JSON" | head -1 | sed 's/.*"\([^"]*\)"$/\1/'
-  fi
-}
-
-#
-# Get marker version (if exists)
-#
-get_marker_version() {
-  if [[ -f "$MARKER" ]]; then
-    grep -o '"version"[[:space:]]*:[[:space:]]*"[^"]*"' "$MARKER" | head -1 | sed 's/.*"\([^"]*\)"$/\1/'
-  fi
-}
-
-#
-# Get marker's recorded bun version
-#
-get_marker_bun() {
-  if [[ -f "$MARKER" ]]; then
-    grep -o '"bun"[[:space:]]*:[[:space:]]*"[^"]*"' "$MARKER" | head -1 | sed 's/.*"\([^"]*\)"$/\1/'
-  fi
-}
-
-#
-# Check if install is needed
-#
-needs_install() {
-  # No node_modules? Definitely need install
-  if [[ ! -d "$ROOT/node_modules" ]]; then
-    return 0
-  fi
-  
-  # No marker? Need install
-  if [[ ! -f "$MARKER" ]]; then
-    return 0
-  fi
-  
-  local pkg_ver marker_ver bun_ver marker_bun
-  pkg_ver=$(get_pkg_version)
-  marker_ver=$(get_marker_version)
-  
-  # Version mismatch? Need install
-  if [[ "$pkg_ver" != "$marker_ver" ]]; then
-    return 0
-  fi
-  
-  # Bun version changed? Need install
-  if BUN_PATH=$(find_bun); then
-    bun_ver=$("$BUN_PATH" --version 2>/dev/null || echo "")
-    marker_bun=$(get_marker_bun)
-    if [[ -n "$bun_ver" && "$bun_ver" != "$marker_bun" ]]; then
-      return 0
-    fi
-  fi
-  
-  # All good, no install needed
-  return 1
-}
-
-#
-# Write version marker after successful install
-#
-write_marker() {
-  local bun_ver uv_ver pkg_ver
-  pkg_ver=$(get_pkg_version)
-  bun_ver=$("$BUN_PATH" --version 2>/dev/null || echo "unknown")
-  
-  if UV_PATH=$(find_uv); then
-    uv_ver=$("$UV_PATH" --version 2>/dev/null | head -1 || echo "unknown")
-  else
-    uv_ver="not-installed"
-  fi
-  
-  cat > "$MARKER" <<EOF
-{
-  "version": "$pkg_ver",
-  "bun": "$bun_ver",
-  "uv": "$uv_ver",
-  "installedAt": "$(date -u +%Y-%m-%dT%H:%M:%SZ)"
-}
-EOF
-}
-
-#
-# Main
-#
-
-# 1. Check for Bun
-BUN_PATH=$(find_bun) || true
-if [[ -z "$BUN_PATH" ]]; then
-  log_error "Bun runtime not found!"
-  echo "" >&2
-  echo "claude-mem requires Bun to run. Please install it:" >&2
-  echo "" >&2
-  echo "  curl -fsSL https://bun.sh/install | bash" >&2
-  echo "" >&2
-  echo "Or on macOS with Homebrew:" >&2
-  echo "" >&2
-  echo "  brew install oven-sh/bun/bun" >&2
-  echo "" >&2
-  echo "Then restart your terminal and try again." >&2
-  exit 1
-fi
-
-BUN_VERSION=$("$BUN_PATH" --version 2>/dev/null || echo "unknown")
-log_ok "Bun $BUN_VERSION found at $BUN_PATH"
-
-# 2. Check for uv (optional - for Python/Chroma support)
-UV_PATH=$(find_uv) || true
-if [[ -z "$UV_PATH" ]]; then
-  log_warn "uv not found (optional - needed for Python/Chroma vector search)"
-  echo "  To install: curl -LsSf https://astral.sh/uv/install.sh | sh" >&2
-else
-  UV_VERSION=$("$UV_PATH" --version 2>/dev/null | head -1 || echo "unknown")
-  log_ok "uv $UV_VERSION found"
-fi
-
-# 3. Install dependencies if needed
-if needs_install; then
-  log_info "Installing dependencies with Bun..."
-  
-  if ! "$BUN_PATH" install --cwd "$ROOT"; then
-    log_error "Failed to install dependencies"
-    exit 1
-  fi
-  
-  write_marker
-  log_ok "Dependencies installed ($(get_pkg_version))"
-else
-  log_ok "Dependencies up to date ($(get_marker_version))"
-fi
-
-exit 0

plugin/scripts/smart-install.js

@@ -4,16 +4,72 @@
  *
  * Ensures Bun runtime and uv (Python package manager) are installed
  * (auto-installs if missing) and handles dependency installation when needed.
+ *
+ * Resolves the install directory from CLAUDE_PLUGIN_ROOT (set by Claude Code
+ * for both cache and marketplace installs), falling back to script location
+ * and legacy paths.
  */
 import { existsSync, readFileSync, writeFileSync } from 'fs';
 import { execSync, spawnSync } from 'child_process';
-import { join } from 'path';
+import { join, dirname } from 'path';
 import { homedir } from 'os';
+import { fileURLToPath } from 'url';
 
-const ROOT = join(homedir(), '.claude', 'plugins', 'marketplaces', 'thedotmack');
-const MARKER = join(ROOT, '.install-version');
+// Early exit if plugin is disabled in Claude Code settings (#781)
+function isPluginDisabledInClaudeSettings() {
+  try {
+    const configDir = process.env.CLAUDE_CONFIG_DIR || join(homedir(), '.claude');
+    const settingsPath = join(configDir, 'settings.json');
+    if (!existsSync(settingsPath)) return false;
+    const settings = JSON.parse(readFileSync(settingsPath, 'utf-8'));
+    return settings?.enabledPlugins?.['claude-mem@thedotmack'] === false;
+  } catch {
+    return false;
+  }
+}
+
+if (isPluginDisabledInClaudeSettings()) {
+  process.exit(0);
+}
 const IS_WINDOWS = process.platform === 'win32';
 
+/**
+ * Resolve the plugin root directory where dependencies should be installed.
+ *
+ * Priority:
+ * 1. CLAUDE_PLUGIN_ROOT env var (set by Claude Code for hooks — works for
+ *    both cache-based and marketplace installs)
+ * 2. Script location (dirname of this file, up one level from scripts/)
+ * 3. XDG path (~/.config/claude/plugins/marketplaces/thedotmack)
+ * 4. Legacy path (~/.claude/plugins/marketplaces/thedotmack)
+ */
+function resolveRoot() {
+  // CLAUDE_PLUGIN_ROOT is the authoritative location set by Claude Code
+  if (process.env.CLAUDE_PLUGIN_ROOT) {
+    const root = process.env.CLAUDE_PLUGIN_ROOT;
+    if (existsSync(join(root, 'package.json'))) return root;
+  }
+
+  // Derive from script location (this file is in <root>/scripts/)
+  try {
+    const scriptDir = dirname(fileURLToPath(import.meta.url));
+    const candidate = dirname(scriptDir);
+    if (existsSync(join(candidate, 'package.json'))) return candidate;
+  } catch {
+    // import.meta.url not available
+  }
+
+  // Probe XDG path, then legacy
+  const marketplaceRel = join('plugins', 'marketplaces', 'thedotmack');
+  const xdg = join(homedir(), '.config', 'claude', marketplaceRel);
+  if (existsSync(join(xdg, 'package.json'))) return xdg;
+
+  return join(homedir(), '.claude', marketplaceRel);
+}
+
+const ROOT = resolveRoot();
+const MARKER = join(ROOT, '.install-version');
+
 /**
  * Check if Bun is installed and accessible
  */
@@ -287,7 +343,7 @@ function installUv() {
  * Add shell alias for claude-mem command
  */
 function installCLI() {
-  const WORKER_CLI = join(ROOT, 'plugin', 'scripts', 'worker-service.cjs');
+  const WORKER_CLI = join(ROOT, 'scripts', 'worker-service.cjs');
   const bunPath = getBunPath() || 'bun';
   const aliasLine = `alias claude-mem='${bunPath} "${WORKER_CLI}"'`;
   const markerPath = join(ROOT, '.cli-installed');
@@ -405,6 +461,31 @@ function installDeps() {
   }));
 }
 
+/**
+ * Verify that critical runtime modules are resolvable from the install directory.
+ * Returns true if all critical modules exist, false otherwise.
+ */
+function verifyCriticalModules() {
+  const pkg = JSON.parse(readFileSync(join(ROOT, 'package.json'), 'utf-8'));
+  const dependencies = Object.keys(pkg.dependencies || {});
+
+  const missing = [];
+  for (const dep of dependencies) {
+    // Check that the module directory exists in node_modules
+    const modulePath = join(ROOT, 'node_modules', ...dep.split('/'));
+    if (!existsSync(modulePath)) {
+      missing.push(dep);
+    }
+  }
+
+  if (missing.length > 0) {
+    console.error(`❌ Post-install check failed: missing modules: ${missing.join(', ')}`);
+    return false;
+  }
+
+  return true;
+}
+
 // Main execution
 try {
   // Step 1: Ensure Bun is installed and meets minimum version (REQUIRED)
@@ -456,6 +537,21 @@ try {
     const newVersion = pkg.version;
 
     installDeps();
+
+    // Verify critical modules are resolvable
+    if (!verifyCriticalModules()) {
+      console.error('⚠️  Retrying install with npm...');
+      try {
+        execSync('npm install --production', { cwd: ROOT, stdio: 'inherit', shell: IS_WINDOWS });
+      } catch {
+        // npm also failed
+      }
+      if (!verifyCriticalModules()) {
+        console.error('❌ Dependencies could not be installed. Plugin may not work correctly.');
+        process.exit(1);
+      }
+    }
+
     console.error('✅ Dependencies installed');
 
     // Auto-restart worker to pick up new code

plugin/skills/do/SKILL.md

@@ -1,5 +1,5 @@
 ---
-name: do-plan
+name: do
 description: Execute a phased implementation plan using subagents. Use when asked to execute, run, or carry out a plan — especially one created by make-plan.
 ---
 

plugin/skills/make-plan/SKILL.md

@@ -0,0 +1,63 @@
+---
+name: make-plan
+description: Create a detailed, phased implementation plan with documentation discovery. Use when asked to plan a feature, task, or multi-step implementation — especially before executing with do.
+---
+
+# Make Plan
+
+You are an ORCHESTRATOR. Create an LLM-friendly plan in phases that can be executed consecutively in new chat contexts.
+
+## Delegation Model
+
+Use subagents for *fact gathering and extraction* (docs, examples, signatures, grep results). Keep *synthesis and plan authoring* with the orchestrator (phase boundaries, task framing, final wording). If a subagent report is incomplete or lacks evidence, re-check with targeted reads/greps before finalizing.
+
+### Subagent Reporting Contract (MANDATORY)
+
+Each subagent response must include:
+1. Sources consulted (files/URLs) and what was read
+2. Concrete findings (exact API names/signatures; exact file paths/locations)
+3. Copy-ready snippet locations (example files/sections to copy)
+4. "Confidence" note + known gaps (what might still be missing)
+
+Reject and redeploy the subagent if it reports conclusions without sources.
+
+## Plan Structure
+
+### Phase 0: Documentation Discovery (ALWAYS FIRST)
+
+Before planning implementation, deploy "Documentation Discovery" subagents to:
+1. Search for and read relevant documentation, examples, and existing patterns
+2. Identify the actual APIs, methods, and signatures available (not assumed)
+3. Create a brief "Allowed APIs" list citing specific documentation sources
+4. Note any anti-patterns to avoid (methods that DON'T exist, deprecated parameters)
+
+The orchestrator consolidates findings into a single Phase 0 output.
+
+### Each Implementation Phase Must Include
+
+1. **What to implement** — Frame tasks to COPY from docs, not transform existing code
+   - Good: "Copy the V2 session pattern from docs/examples.ts:45-60"
+   - Bad: "Migrate the existing code to V2"
+2. **Documentation references** — Cite specific files/lines for patterns to follow
+3. **Verification checklist** — How to prove this phase worked (tests, grep checks)
+4. **Anti-pattern guards** — What NOT to do (invented APIs, undocumented params)
+
+### Final Phase: Verification
+
+1. Verify all implementations match documentation
+2. Check for anti-patterns (grep for known bad patterns)
+3. Run tests to confirm functionality
+
+## Key Principles
+
+- Documentation Availability ≠ Usage: Explicitly require reading docs
+- Task Framing Matters: Direct agents to docs, not just outcomes
+- Verify > Assume: Require proof, not assumptions about APIs
+- Session Boundaries: Each phase should be self-contained with its own doc references
+
+## Anti-Patterns to Prevent
+
+- Inventing API methods that "should" exist
+- Adding parameters not in documentation
+- Skipping verification steps
+- Assuming structure without checking examples

plugin/skills/mem-search/SKILL.md

@@ -93,20 +93,6 @@ get_observations(ids=[11131, 10942])
 
 **Returns:** Complete observation objects with title, subtitle, narrative, facts, concepts, files (~500-1000 tokens each)
 
-## Saving Memories
-
-Use the `save_memory` MCP tool to store manual observations:
-
-```
-save_memory(text="Important discovery about the auth system", title="Auth Architecture", project="my-project")
-```
-
-**Parameters:**
-
-- `text` (string, required) - Content to remember
-- `title` (string, optional) - Short title, auto-generated if omitted
-- `project` (string, optional) - Project name, defaults to "claude-mem"
-
 ## Examples
 
 **Find recent bug fixes:**

plugin/skills/smart-explore/SKILL.md

@@ -0,0 +1,145 @@
+---
+name: smart-explore
+description: Token-optimized structural code search using tree-sitter AST parsing. Use instead of reading full files when you need to understand code structure, find functions, or explore a codebase efficiently.
+---
+
+# Smart Explore
+
+Structural code exploration using AST parsing. **This skill overrides your default exploration behavior.** While this skill is active, use smart_search/smart_outline/smart_unfold as your primary tools instead of Read, Grep, and Glob.
+
+**Core principle:** Index first, fetch on demand. Give yourself a map of the code before loading implementation details. The question before every file read should be: "do I need to see all of this, or can I get a structural overview first?" The answer is almost always: get the map.
+
+## Your Next Tool Call
+
+This skill only loads instructions. You must call the MCP tools yourself. Your next action should be one of:
+
+```
+smart_search(query="<topic>", path="./src")    -- discover files + symbols across a directory
+smart_outline(file_path="<file>")              -- structural skeleton of one file
+smart_unfold(file_path="<file>", symbol_name="<name>")  -- full source of one symbol
+```
+
+Do NOT run Grep, Glob, Read, or find to discover files first. `smart_search` walks directories, parses all code files, and returns ranked symbols in one call. It replaces the Glob → Grep → Read discovery cycle.
+
+## 3-Layer Workflow
+
+### Step 1: Search -- Discover Files and Symbols
+
+```
+smart_search(query="shutdown", path="./src", max_results=15)
+```
+
+**Returns:** Ranked symbols with signatures, line numbers, match reasons, plus folded file views (~2-6k tokens)
+
+```
+-- Matching Symbols --
+  function performGracefulShutdown (services/infrastructure/GracefulShutdown.ts:56)
+  function httpShutdown (services/infrastructure/HealthMonitor.ts:92)
+  method WorkerService.shutdown (services/worker-service.ts:846)
+
+-- Folded File Views --
+  services/infrastructure/GracefulShutdown.ts (7 symbols)
+  services/worker-service.ts (12 symbols)
+```
+
+This is your discovery tool. It finds relevant files AND shows their structure. No Glob/find pre-scan needed.
+
+**Parameters:**
+
+- `query` (string, required) -- What to search for (function name, concept, class name)
+- `path` (string) -- Root directory to search (defaults to cwd)
+- `max_results` (number) -- Max matching symbols, default 20, max 50
+- `file_pattern` (string, optional) -- Filter to specific files/paths
+
+### Step 2: Outline -- Get File Structure
+
+```
+smart_outline(file_path="services/worker-service.ts")
+```
+
+**Returns:** Complete structural skeleton -- all functions, classes, methods, properties, imports (~1-2k tokens per file)
+
+**Skip this step** when Step 1's folded file views already provide enough structure. Most useful for files not covered by the search results.
+
+**Parameters:**
+
+- `file_path` (string, required) -- Path to the file
+
+### Step 3: Unfold -- See Implementation
+
+Review symbols from Steps 1-2. Pick the ones you need. Unfold only those:
+
+```
+smart_unfold(file_path="services/worker-service.ts", symbol_name="shutdown")
+```
+
+**Returns:** Full source code of the specified symbol including JSDoc, decorators, and complete implementation (~400-2,100 tokens depending on symbol size). AST node boundaries guarantee completeness regardless of symbol size — unlike Read + agent summarization, which may truncate long methods.
+
+**Parameters:**
+
+- `file_path` (string, required) -- Path to the file (as returned by search/outline)
+- `symbol_name` (string, required) -- Name of the function/class/method to expand
+
+## When to Use Standard Tools Instead
+
+Use these only when smart_* tools are the wrong fit:
+
+- **Grep:** Exact string/regex search ("find all TODO comments", "where is `ensureWorkerStarted` defined?")
+- **Read:** Small files under ~100 lines, non-code files (JSON, markdown, config)
+- **Glob:** File path patterns ("find all test files")
+- **Explore agent:** When you need synthesized understanding across 6+ files, architecture narratives, or answers to open-ended questions like "how does this entire system work end-to-end?" Smart-explore is a scalpel — it answers "where is this?" and "show me that." It doesn't synthesize cross-file data flows, design decisions, or edge cases across an entire feature.
+
+For code files over ~100 lines, prefer smart_outline + smart_unfold over Read.
+
+## Workflow Examples
+
+**Discover how a feature works (cross-cutting):**
+
+```
+1. smart_search(query="shutdown", path="./src")
+   -> 14 symbols across 7 files, full picture in one call
+2. smart_unfold(file_path="services/infrastructure/GracefulShutdown.ts", symbol_name="performGracefulShutdown")
+   -> See the core implementation
+```
+
+**Navigate a large file:**
+
+```
+1. smart_outline(file_path="services/worker-service.ts")
+   -> 1,466 tokens: 12 functions, WorkerService class with 24 members
+2. smart_unfold(file_path="services/worker-service.ts", symbol_name="startSessionProcessor")
+   -> 1,610 tokens: the specific method you need
+Total: ~3,076 tokens vs ~12,000 to Read the full file
+```
+
+**Write documentation about code (hybrid workflow):**
+
+```
+1. smart_search(query="feature name", path="./src")    -- discover all relevant files and symbols
+2. smart_outline on key files                           -- understand structure
+3. smart_unfold on important functions                  -- get implementation details
+4. Read on small config/markdown/plan files             -- get non-code context
+```
+
+Use smart_* tools for code exploration, Read for non-code files. Mix freely.
+
+**Exploration then precision:**
+
+```
+1. smart_search(query="session", path="./src", max_results=10)
+   -> 10 ranked symbols: SessionMetadata, SessionQueueProcessor, SessionSummary...
+2. Pick the relevant one, unfold it
+```
+
+## Token Economics
+
+| Approach | Tokens | Use Case |
+|----------|--------|----------|
+| smart_outline | ~1,000-2,000 | "What's in this file?" |
+| smart_unfold | ~400-2,100 | "Show me this function" |
+| smart_search | ~2,000-6,000 | "Find all X across the codebase" |
+| search + unfold | ~3,000-8,000 | End-to-end: find and read (the primary workflow) |
+| Read (full file) | ~12,000+ | When you truly need everything |
+| Explore agent | ~39,000-59,000 | Cross-file synthesis with narrative |
+
+**4-8x savings** on file understanding (outline + unfold vs Read). **11-18x savings** on codebase exploration vs Explore agent. The narrower the query, the wider the gap — a 27-line function costs 55x less to read via unfold than via an Explore agent, because the agent still reads the entire file.

scripts/CLAUDE.md

@@ -0,0 +1 @@
+Never read built source files in this directory. These are compiled outputs — read the source files in `src/` instead.

scripts/build-hooks.js

@@ -59,8 +59,16 @@ async function buildHooks() {
       description: 'Runtime dependencies for claude-mem bundled hooks',
       type: 'module',
       dependencies: {
-        // Chroma embedding function with native ONNX binaries (can't be bundled)
-        '@chroma-core/default-embed': '^0.1.9'
+        'tree-sitter-cli': '^0.26.5',
+        'tree-sitter-c': '^0.24.1',
+        'tree-sitter-cpp': '^0.23.4',
+        'tree-sitter-go': '^0.25.0',
+        'tree-sitter-java': '^0.23.5',
+        'tree-sitter-javascript': '^0.25.0',
+        'tree-sitter-python': '^0.25.0',
+        'tree-sitter-ruby': '^0.23.1',
+        'tree-sitter-rust': '^0.24.0',
+        'tree-sitter-typescript': '^0.23.2',
       },
       engines: {
         node: '>=18.0.0',
@@ -128,7 +136,19 @@ async function buildHooks() {
       outfile: `${hooksDir}/${MCP_SERVER.name}.cjs`,
       minify: true,
       logLevel: 'error',
-      external: ['bun:sqlite'],
+      external: [
+        'bun:sqlite',
+        'tree-sitter-cli',
+        'tree-sitter-javascript',
+        'tree-sitter-typescript',
+        'tree-sitter-python',
+        'tree-sitter-go',
+        'tree-sitter-rust',
+        'tree-sitter-ruby',
+        'tree-sitter-java',
+        'tree-sitter-c',
+        'tree-sitter-cpp',
+      ],
       define: {
         '__DEFAULT_PACKAGE_VERSION__': `"${version}"`
       },
@@ -162,6 +182,21 @@ async function buildHooks() {
     const contextGenStats = fs.statSync(`${hooksDir}/${CONTEXT_GENERATOR.name}.cjs`);
     console.log(`✓ context-generator built (${(contextGenStats.size / 1024).toFixed(2)} KB)`);
 
+    // Verify critical distribution files exist (skills are source files, not build outputs)
+    console.log('\n📋 Verifying distribution files...');
+    const requiredDistributionFiles = [
+      'plugin/skills/mem-search/SKILL.md',
+      'plugin/skills/smart-explore/SKILL.md',
+      'plugin/hooks/hooks.json',
+      'plugin/.claude-plugin/plugin.json',
+    ];
+    for (const filePath of requiredDistributionFiles) {
+      if (!fs.existsSync(filePath)) {
+        throw new Error(`Missing required distribution file: ${filePath}`);
+      }
+    }
+    console.log('✓ All required distribution files present');
+
     console.log('\n✅ Worker service, MCP server, and context generator built successfully!');
     console.log(`   Output: ${hooksDir}/`);
     console.log(`   - Worker: worker-service.cjs`);

scripts/regenerate-claude-md.ts

@@ -279,6 +279,11 @@ function formatObservationsForClaudeMd(observations: ObservationRow[], folderPat
  * which only writes to existing folders.
  */
 function writeClaudeMdToFolderForRegenerate(folderPath: string, newContent: string): void {
+  const resolvedPath = path.resolve(folderPath);
+
+  // Never write inside .git directories — corrupts refs (#1165)
+  if (resolvedPath.includes('/.git/') || resolvedPath.includes('\\.git\\') || resolvedPath.endsWith('/.git') || resolvedPath.endsWith('\\.git')) return;
+
   const claudeMdPath = path.join(folderPath, 'CLAUDE.md');
   const tempFile = `${claudeMdPath}.tmp`;
 

scripts/smart-install.js

@@ -7,34 +7,40 @@
  */
 import { existsSync, readFileSync, writeFileSync } from 'fs';
 import { execSync, spawnSync } from 'child_process';
-import { join } from 'path';
+import { join, dirname } from 'path';
 import { homedir } from 'os';
+import { fileURLToPath } from 'url';
 
 const IS_WINDOWS = process.platform === 'win32';
 
 /**
- * Resolve the marketplace root directory.
+ * Resolve the plugin root directory where dependencies should be installed.
  *
- * Claude Code may store plugins under either `~/.claude/plugins/` (legacy) or
- * `~/.config/claude/plugins/` (XDG-compliant, e.g. Nix-managed installs).
- * When `CLAUDE_PLUGIN_ROOT` is set we derive the base from it; otherwise we
- * probe both candidate paths and fall back to the legacy location.
+ * Priority:
+ * 1. CLAUDE_PLUGIN_ROOT env var (set by Claude Code for hooks — works for
+ *    both cache-based and marketplace installs)
+ * 2. Script location (dirname of this file, up one level from scripts/)
+ * 3. XDG path (~/.config/claude/plugins/marketplaces/thedotmack)
+ * 4. Legacy path (~/.claude/plugins/marketplaces/thedotmack)
  */
 function resolveRoot() {
-  const marketplaceRel = join('plugins', 'marketplaces', 'thedotmack');
-
-  // Derive from CLAUDE_PLUGIN_ROOT (e.g. .../plugins/cache/thedotmack/claude-mem/<ver>)
+  // CLAUDE_PLUGIN_ROOT is the authoritative location set by Claude Code
   if (process.env.CLAUDE_PLUGIN_ROOT) {
-    let dir = process.env.CLAUDE_PLUGIN_ROOT;
-    const cacheIndex = dir.indexOf(join('plugins', 'cache'));
-    if (cacheIndex !== -1) {
-      const base = dir.substring(0, cacheIndex);
-      const candidate = join(base, marketplaceRel);
-      if (existsSync(join(candidate, 'package.json'))) return candidate;
-    }
+    const root = process.env.CLAUDE_PLUGIN_ROOT;
+    if (existsSync(join(root, 'package.json'))) return root;
   }
 
-  // Probe XDG path first, then legacy
+  // Derive from script location (this file is in <root>/scripts/)
+  try {
+    const scriptDir = dirname(fileURLToPath(import.meta.url));
+    const candidate = dirname(scriptDir);
+    if (existsSync(join(candidate, 'package.json'))) return candidate;
+  } catch {
+    // import.meta.url not available
+  }
+
+  // Probe XDG path, then legacy
+  const marketplaceRel = join('plugins', 'marketplaces', 'thedotmack');
   const xdg = join(homedir(), '.config', 'claude', marketplaceRel);
   if (existsSync(join(xdg, 'package.json'))) return xdg;
 
@@ -275,12 +281,42 @@ function installDeps() {
   }));
 }
 
+/**
+ * Verify that critical runtime modules are resolvable from the install directory.
+ * Returns true if all critical modules exist, false otherwise.
+ */
+function verifyCriticalModules() {
+  const pkg = JSON.parse(readFileSync(join(ROOT, 'package.json'), 'utf-8'));
+  const dependencies = Object.keys(pkg.dependencies || {});
+
+  const missing = [];
+  for (const dep of dependencies) {
+    const modulePath = join(ROOT, 'node_modules', ...dep.split('/'));
+    if (!existsSync(modulePath)) {
+      missing.push(dep);
+    }
+  }
+
+  if (missing.length > 0) {
+    console.error(`❌ Post-install check failed: missing modules: ${missing.join(', ')}`);
+    return false;
+  }
+
+  return true;
+}
+
 // Main execution
 try {
   if (!isBunInstalled()) installBun();
   if (!isUvInstalled()) installUv();
   if (needsInstall()) {
     installDeps();
+
+    if (!verifyCriticalModules()) {
+      console.error('❌ Dependencies could not be installed. Plugin may not work correctly.');
+      process.exit(1);
+    }
+
     console.error('✅ Dependencies installed');
   }
 } catch (e) {

scripts/sync-marketplace.cjs

@@ -76,7 +76,7 @@ try {
   const gitignoreExcludes = getGitignoreExcludes(rootDir);
 
   execSync(
-    `rsync -av --delete --exclude=.git --exclude=/.mcp.json --exclude=bun.lock --exclude=package-lock.json ${gitignoreExcludes} ./ ~/.claude/plugins/marketplaces/thedotmack/`,
+    `rsync -av --delete --exclude=.git --exclude=bun.lock --exclude=package-lock.json ${gitignoreExcludes} ./ ~/.claude/plugins/marketplaces/thedotmack/`,
     { stdio: 'inherit' }
   );
 

scripts/wipe-chroma.cjs

@@ -0,0 +1,19 @@
+#!/usr/bin/env node
+/**
+ * Wipes the Chroma data directory so backfillAllProjects rebuilds it on next worker start.
+ * Chroma is always rebuildable from SQLite — this is safe.
+ */
+const fs = require('fs');
+const path = require('path');
+const os = require('os');
+
+const chromaDir = path.join(os.homedir(), '.claude-mem', 'chroma');
+
+if (fs.existsSync(chromaDir)) {
+  const before = fs.readdirSync(chromaDir);
+  console.log(`Wiping ${chromaDir} (${before.length} items)...`);
+  fs.rmSync(chromaDir, { recursive: true, force: true });
+  console.log('Done. Chroma will rebuild from SQLite on next worker restart.');
+} else {
+  console.log('Chroma directory does not exist, nothing to wipe.');
+}

src/cli/adapters/claude-code.ts

@@ -6,7 +6,7 @@ export const claudeCodeAdapter: PlatformAdapter = {
   normalizeInput(raw) {
     const r = (raw ?? {}) as any;
     return {
-      sessionId: r.session_id,
+      sessionId: r.session_id ?? r.id ?? r.sessionId,
       cwd: r.cwd ?? process.cwd(),
       prompt: r.prompt,
       toolName: r.tool_name,

src/cli/adapters/cursor.ts

@@ -3,15 +3,20 @@ import type { PlatformAdapter, NormalizedHookInput, HookResult } from '../types.
 // Maps Cursor stdin format - field names differ from Claude Code
 // Cursor uses: conversation_id, workspace_roots[], result_json, command/output
 // Handle undefined input gracefully for hooks that don't receive stdin
+//
+// Cursor payload variations (#838, #1049):
+//   Session ID: conversation_id, generation_id, or id
+//   Prompt: prompt, query, input, or message (varies by Cursor version/hook type)
+//   CWD: workspace_roots[0] or cwd
 export const cursorAdapter: PlatformAdapter = {
   normalizeInput(raw) {
     const r = (raw ?? {}) as any;
     // Cursor-specific: shell commands come as command/output instead of tool_name/input/response
     const isShellCommand = !!r.command && !r.tool_name;
     return {
-      sessionId: r.conversation_id || r.generation_id,  // conversation_id preferred
-      cwd: r.workspace_roots?.[0] ?? process.cwd(),     // First workspace root
-      prompt: r.prompt,
+      sessionId: r.conversation_id || r.generation_id || r.id,
+      cwd: r.workspace_roots?.[0] ?? r.cwd ?? process.cwd(),
+      prompt: r.prompt ?? r.query ?? r.input ?? r.message,
       toolName: isShellCommand ? 'Bash' : r.tool_name,
       toolInput: isShellCommand ? { command: r.command } : r.tool_input,
       toolResponse: isShellCommand ? { output: r.output } : r.result_json,  // result_json not tool_response

src/cli/adapters/index.ts

@@ -8,7 +8,8 @@ export function getPlatformAdapter(platform: string): PlatformAdapter {
     case 'claude-code': return claudeCodeAdapter;
     case 'cursor': return cursorAdapter;
     case 'raw': return rawAdapter;
-    default: throw new Error(`Unknown platform: ${platform}`);
+    // Codex CLI and other compatible platforms use the raw adapter (accepts both camelCase and snake_case fields)
+    default: return rawAdapter;
   }
 }
 

src/cli/claude-md-commands.ts

@@ -264,6 +264,11 @@ function formatObservationsForClaudeMd(observations: ObservationRow[], folderPat
  * Only writes to folders that exist — never creates directories.
  */
 function writeClaudeMdToFolder(folderPath: string, newContent: string): void {
+  const resolvedPath = path.resolve(folderPath);
+
+  // Never write inside .git directories — corrupts refs (#1165)
+  if (resolvedPath.includes('/.git/') || resolvedPath.includes('\\.git\\') || resolvedPath.endsWith('/.git') || resolvedPath.endsWith('\\.git')) return;
+
   const claudeMdPath = path.join(folderPath, 'CLAUDE.md');
   const tempFile = `${claudeMdPath}.tmp`;
 

src/cli/handlers/context.ts

@@ -10,6 +10,8 @@ import { ensureWorkerRunning, getWorkerPort } from '../../shared/worker-utils.js
 import { getProjectContext } from '../../utils/project-name.js';
 import { HOOK_EXIT_CODES } from '../../shared/hook-constants.js';
 import { logger } from '../../utils/logger.js';
+import { SettingsDefaultsManager } from '../../shared/SettingsDefaultsManager.js';
+import { USER_SETTINGS_PATH } from '../../shared/paths.js';
 
 export const contextHandler: EventHandler = {
   async execute(input: NormalizedHookInput): Promise<HookResult> {
@@ -30,6 +32,10 @@ export const contextHandler: EventHandler = {
     const context = getProjectContext(cwd);
     const port = getWorkerPort();
 
+    // Check if terminal output should be shown (load settings early)
+    const settings = SettingsDefaultsManager.loadFromFile(USER_SETTINGS_PATH);
+    const showTerminalOutput = settings.CLAUDE_MEM_CONTEXT_SHOW_TERMINAL_OUTPUT === 'true';
+
     // Pass all projects (parent + worktree if applicable) for unified timeline
     const projectsParam = context.allProjects.join(',');
     const url = `http://127.0.0.1:${port}/api/context/inject?projects=${encodeURIComponent(projectsParam)}`;
@@ -37,11 +43,11 @@ export const contextHandler: EventHandler = {
     // Note: Removed AbortSignal.timeout due to Windows Bun cleanup issue (libuv assertion)
     // Worker service has its own timeouts, so client-side timeout is redundant
     try {
-      // Fetch both markdown (for Claude context) and colored (for user display) truly in parallel
+      // Fetch markdown (for Claude context) and optionally colored (for user display)
       const colorUrl = `${url}&colors=true`;
       const [response, colorResponse] = await Promise.all([
         fetch(url),
-        fetch(colorUrl).catch(() => null)
+        showTerminalOutput ? fetch(colorUrl).catch(() => null) : Promise.resolve(null)
       ]);
 
       if (!response.ok) {
@@ -60,7 +66,8 @@ export const contextHandler: EventHandler = {
 
       const additionalContext = contextResult.trim();
       const coloredTimeline = colorResult.trim();
-      const systemMessage = coloredTimeline
+
+      const systemMessage = showTerminalOutput && coloredTimeline
         ? `${coloredTimeline}\n\nView Observations Live @ http://localhost:${port}`
         : undefined;
 

src/cli/handlers/index.ts

@@ -6,6 +6,7 @@
 
 import type { EventHandler } from '../types.js';
 import { HOOK_EXIT_CODES } from '../../shared/hook-constants.js';
+import { logger } from '../../utils/logger.js';
 import { contextHandler } from './context.js';
 import { sessionInitHandler } from './session-init.js';
 import { observationHandler } from './observation.js';
@@ -46,7 +47,7 @@ const handlers: Record<EventType, EventHandler> = {
 export function getEventHandler(eventType: string): EventHandler {
   const handler = handlers[eventType as EventType];
   if (!handler) {
-    console.error(`[claude-mem] Unknown event type: ${eventType}, returning no-op`);
+    logger.warn('HOOK', `Unknown event type: ${eventType}, returning no-op`);
     return {
       async execute() {
         return { continue: true, suppressOutput: true, exitCode: HOOK_EXIT_CODES.SUCCESS };

src/cli/handlers/session-init.ts

@@ -24,6 +24,12 @@ export const sessionInitHandler: EventHandler = {
 
     const { sessionId, cwd, prompt: rawPrompt } = input;
 
+    // Guard: Codex CLI and other platforms may not provide a session_id (#744)
+    if (!sessionId) {
+      logger.warn('HOOK', 'session-init: No sessionId provided, skipping (Codex CLI or unknown platform)');
+      return { continue: true, suppressOutput: true, exitCode: HOOK_EXIT_CODES.SUCCESS };
+    }
+
     // Check if project is excluded from tracking
     const settings = SettingsDefaultsManager.loadFromFile(USER_SETTINGS_PATH);
     if (cwd && isProjectExcluded(cwd, settings.CLAUDE_MEM_EXCLUDED_PROJECTS)) {
@@ -63,11 +69,12 @@ export const sessionInitHandler: EventHandler = {
       promptNumber: number;
       skipped?: boolean;
       reason?: string;
+      contextInjected?: boolean;
     };
     const sessionDbId = initResult.sessionDbId;
     const promptNumber = initResult.promptNumber;
 
-    logger.debug('HOOK', 'session-init: Received from /api/sessions/init', { sessionDbId, promptNumber, skipped: initResult.skipped });
+    logger.debug('HOOK', 'session-init: Received from /api/sessions/init', { sessionDbId, promptNumber, skipped: initResult.skipped, contextInjected: initResult.contextInjected });
 
     // Debug-level alignment log for detailed tracing
     logger.debug('HOOK', `[ALIGNMENT] Hook Entry | contentSessionId=${sessionId} | prompt#=${promptNumber} | sessionDbId=${sessionDbId}`);
@@ -80,6 +87,16 @@ export const sessionInitHandler: EventHandler = {
       return { continue: true, suppressOutput: true };
     }
 
+    // Skip SDK agent re-initialization if context was already injected for this session (#1079)
+    // The prompt was already saved to the database by /api/sessions/init above —
+    // no need to re-start the SDK agent on every turn
+    if (initResult.contextInjected) {
+      logger.info('HOOK', `INIT_COMPLETE | sessionDbId=${sessionDbId} | promptNumber=${promptNumber} | skipped_agent_init=true | reason=context_already_injected`, {
+        sessionId: sessionDbId
+      });
+      return { continue: true, suppressOutput: true };
+    }
+
     // Only initialize SDK agent for Claude Code (not Cursor)
     // Cursor doesn't use the SDK agent - it only needs session/observation storage
     if (input.platform !== 'cursor' && sessionDbId) {

src/cli/hook-command.ts

@@ -2,6 +2,7 @@ import { readJsonFromStdin } from './stdin-reader.js';
 import { getPlatformAdapter } from './adapters/index.js';
 import { getEventHandler } from './handlers/index.js';
 import { HOOK_EXIT_CODES } from '../shared/hook-constants.js';
+import { logger } from '../utils/logger.js';
 
 export interface HookCommandOptions {
   /** If true, don't call process.exit() - let caller handle process lifecycle */
@@ -65,6 +66,12 @@ export function isWorkerUnavailableError(error: unknown): boolean {
 }
 
 export async function hookCommand(platform: string, event: string, options: HookCommandOptions = {}): Promise<number> {
+  // Suppress stderr in hook context — Claude Code shows stderr as error UI (#1181)
+  // Exit 1: stderr shown to user. Exit 2: stderr fed to Claude for processing.
+  // All diagnostics go to log file via logger; stderr must stay clean.
+  const originalStderrWrite = process.stderr.write.bind(process.stderr);
+  process.stderr.write = (() => true) as typeof process.stderr.write;
+
   try {
     const adapter = getPlatformAdapter(platform);
     const handler = getEventHandler(event);
@@ -84,18 +91,22 @@ export async function hookCommand(platform: string, event: string, options: Hook
   } catch (error) {
     if (isWorkerUnavailableError(error)) {
       // Worker unavailable — degrade gracefully, don't block the user
-      console.error(`[claude-mem] Worker unavailable, skipping hook: ${error instanceof Error ? error.message : error}`);
+      // Log to file instead of stderr (#1181)
+      logger.warn('HOOK', `Worker unavailable, skipping hook: ${error instanceof Error ? error.message : error}`);
       if (!options.skipExit) {
         process.exit(HOOK_EXIT_CODES.SUCCESS);  // = 0 (graceful)
       }
       return HOOK_EXIT_CODES.SUCCESS;
     }
 
-    // Handler/client bug — show as blocking error so developers see it
-    console.error(`Hook error: ${error}`);
+    // Handler/client bug — log to file instead of stderr (#1181)
+    logger.error('HOOK', `Hook error: ${error instanceof Error ? error.message : error}`, {}, error instanceof Error ? error : undefined);
     if (!options.skipExit) {
       process.exit(HOOK_EXIT_CODES.BLOCKING_ERROR);  // = 2
     }
     return HOOK_EXIT_CODES.BLOCKING_ERROR;
+  } finally {
+    // Restore stderr for non-hook code paths (e.g., when skipExit is true and process continues as worker)
+    process.stderr.write = originalStderrWrite;
   }
 }

src/servers/mcp-server.ts

@@ -28,6 +28,10 @@ import {
   ListToolsRequestSchema,
 } from '@modelcontextprotocol/sdk/types.js';
 import { getWorkerPort, getWorkerHost } from '../shared/worker-utils.js';
+import { searchCodebase, formatSearchResults } from '../services/smart-file-read/search.js';
+import { parseFile, formatFoldedView, unfoldSymbol } from '../services/smart-file-read/parser.js';
+import { readFile } from 'node:fs/promises';
+import { resolve } from 'node:path';
 
 /**
  * Worker HTTP API configuration
@@ -235,28 +239,115 @@ NEVER fetch full details without filtering first. 10x token savings.`,
     }
   },
   {
-    name: 'save_memory',
-    description: 'Save a manual memory/observation for semantic search. Use this to remember important information.',
+    name: 'smart_search',
+    description: 'Search codebase for symbols, functions, classes using tree-sitter AST parsing. Returns folded structural views with token counts. Use path parameter to scope the search.',
     inputSchema: {
       type: 'object',
       properties: {
-        text: {
+        query: {
           type: 'string',
-          description: 'Content to remember (required)'
+          description: 'Search term — matches against symbol names, file names, and file content'
         },
-        title: {
+        path: {
           type: 'string',
-          description: 'Short title (auto-generated from text if omitted)'
+          description: 'Root directory to search (default: current working directory)'
         },
-        project: {
+        max_results: {
+          type: 'number',
+          description: 'Maximum results to return (default: 20)'
+        },
+        file_pattern: {
           type: 'string',
-          description: 'Project name (uses "claude-mem" if omitted)'
+          description: 'Substring filter for file paths (e.g. ".ts", "src/services")'
         }
       },
-      required: ['text']
+      required: ['query']
     },
     handler: async (args: any) => {
-      return await callWorkerAPIPost('/api/memory/save', args);
+      const rootDir = resolve(args.path || process.cwd());
+      const result = await searchCodebase(rootDir, args.query, {
+        maxResults: args.max_results || 20,
+        filePattern: args.file_pattern
+      });
+      const formatted = formatSearchResults(result, args.query);
+      return {
+        content: [{ type: 'text' as const, text: formatted }]
+      };
+    }
+  },
+  {
+    name: 'smart_unfold',
+    description: 'Expand a specific symbol (function, class, method) from a file. Returns the full source code of just that symbol. Use after smart_search or smart_outline to read specific code.',
+    inputSchema: {
+      type: 'object',
+      properties: {
+        file_path: {
+          type: 'string',
+          description: 'Path to the source file'
+        },
+        symbol_name: {
+          type: 'string',
+          description: 'Name of the symbol to unfold (function, class, method, etc.)'
+        }
+      },
+      required: ['file_path', 'symbol_name']
+    },
+    handler: async (args: any) => {
+      const filePath = resolve(args.file_path);
+      const content = await readFile(filePath, 'utf-8');
+      const unfolded = unfoldSymbol(content, filePath, args.symbol_name);
+      if (unfolded) {
+        return {
+          content: [{ type: 'text' as const, text: unfolded }]
+        };
+      }
+      // Symbol not found — show available symbols
+      const parsed = parseFile(content, filePath);
+      if (parsed.symbols.length > 0) {
+        const available = parsed.symbols.map(s => `  - ${s.name} (${s.kind})`).join('\n');
+        return {
+          content: [{
+            type: 'text' as const,
+            text: `Symbol "${args.symbol_name}" not found in ${args.file_path}.\n\nAvailable symbols:\n${available}`
+          }]
+        };
+      }
+      return {
+        content: [{
+          type: 'text' as const,
+          text: `Could not parse ${args.file_path}. File may be unsupported or empty.`
+        }]
+      };
+    }
+  },
+  {
+    name: 'smart_outline',
+    description: 'Get structural outline of a file — shows all symbols (functions, classes, methods, types) with signatures but bodies folded. Much cheaper than reading the full file.',
+    inputSchema: {
+      type: 'object',
+      properties: {
+        file_path: {
+          type: 'string',
+          description: 'Path to the source file'
+        }
+      },
+      required: ['file_path']
+    },
+    handler: async (args: any) => {
+      const filePath = resolve(args.file_path);
+      const content = await readFile(filePath, 'utf-8');
+      const parsed = parseFile(content, filePath);
+      if (parsed.symbols.length > 0) {
+        return {
+          content: [{ type: 'text' as const, text: formatFoldedView(parsed) }]
+        };
+      }
+      return {
+        content: [{
+          type: 'text' as const,
+          text: `Could not parse ${args.file_path}. File may use an unsupported language or be empty.`
+        }]
+      };
     }
   }
 ];

src/services/context/formatters/ColorFormatter.ts

@@ -74,8 +74,8 @@ export function renderColorContextIndex(): string[] {
     `${colors.dim}Context Index: This semantic index (titles, types, files, tokens) is usually sufficient to understand past work.${colors.reset}`,
     '',
     `${colors.dim}When you need implementation details, rationale, or debugging context:${colors.reset}`,
-    `${colors.dim}  - Use MCP tools (search, get_observations) to fetch full observations on-demand${colors.reset}`,
-    `${colors.dim}  - Critical types ( bugfix, decision) often need detailed fetching${colors.reset}`,
+    `${colors.dim}  - Fetch by ID: get_observations([IDs]) for observations visible in this index${colors.reset}`,
+    `${colors.dim}  - Search history: Use the mem-search skill for past decisions, bugs, and deeper research${colors.reset}`,
     `${colors.dim}  - Trust this index over re-reading code for past decisions and learnings${colors.reset}`,
     ''
   ];
@@ -226,7 +226,7 @@ export function renderColorFooter(totalDiscoveryTokens: number, totalReadTokens:
   const workTokensK = Math.round(totalDiscoveryTokens / 1000);
   return [
     '',
-    `${colors.dim}Access ${workTokensK}k tokens of past research & decisions for just ${totalReadTokens.toLocaleString()}t. Use MCP search tools to access memories by ID.${colors.reset}`
+    `${colors.dim}Access ${workTokensK}k tokens of past research & decisions for just ${totalReadTokens.toLocaleString()}t. Use the claude-mem skill to access memories by ID.${colors.reset}`
   ];
 }
 

src/services/context/formatters/MarkdownFormatter.ts

@@ -72,8 +72,8 @@ export function renderMarkdownContextIndex(): string[] {
     `**Context Index:** This semantic index (titles, types, files, tokens) is usually sufficient to understand past work.`,
     '',
     `When you need implementation details, rationale, or debugging context:`,
-    `- Use MCP tools (search, get_observations) to fetch full observations on-demand`,
-    `- Critical types ( bugfix, decision) often need detailed fetching`,
+    `- Fetch by ID: get_observations([IDs]) for observations visible in this index`,
+    `- Search history: Use the mem-search skill for past decisions, bugs, and deeper research`,
     `- Trust this index over re-reading code for past decisions and learnings`,
     ''
   ];
@@ -229,7 +229,7 @@ export function renderMarkdownFooter(totalDiscoveryTokens: number, totalReadToke
   const workTokensK = Math.round(totalDiscoveryTokens / 1000);
   return [
     '',
-    `Access ${workTokensK}k tokens of past research & decisions for just ${totalReadTokens.toLocaleString()}t. Use MCP search tools to access memories by ID.`
+    `Access ${workTokensK}k tokens of past research & decisions for just ${totalReadTokens.toLocaleString()}t. Use the claude-mem skill to access memories by ID.`
   ];
 }
 

src/services/infrastructure/GracefulShutdown.ts

@@ -30,9 +30,9 @@ export interface CloseableDatabase {
 }
 
 /**
- * Stoppable service interface for Chroma server
+ * Stoppable service interface for ChromaMcpManager
  */
-export interface StoppableServer {
+export interface StoppableService {
   stop(): Promise<void>;
 }
 
@@ -44,7 +44,7 @@ export interface GracefulShutdownConfig {
   sessionManager: ShutdownableService;
   mcpClient?: CloseableClient;
   dbManager?: CloseableDatabase;
-  chromaServer?: StoppableServer;
+  chromaMcpManager?: StoppableService;
 }
 
 /**
@@ -79,11 +79,11 @@ export async function performGracefulShutdown(config: GracefulShutdownConfig): P
     logger.info('SYSTEM', 'MCP client closed');
   }
 
-  // STEP 5: Stop Chroma server (local mode only)
-  if (config.chromaServer) {
-    logger.info('SHUTDOWN', 'Stopping Chroma server...');
-    await config.chromaServer.stop();
-    logger.info('SHUTDOWN', 'Chroma server stopped');
+  // STEP 5: Stop Chroma MCP connection
+  if (config.chromaMcpManager) {
+    logger.info('SHUTDOWN', 'Stopping Chroma MCP connection...');
+    await config.chromaMcpManager.stop();
+    logger.info('SHUTDOWN', 'Chroma MCP connection stopped');
   }
 
   // STEP 6: Close database connection (includes ChromaSync cleanup)

src/services/infrastructure/HealthMonitor.ts

@@ -29,31 +29,49 @@ export async function isPortInUse(port: number): Promise<boolean> {
 }
 
 /**
- * Wait for the worker HTTP server to become responsive (liveness check)
- * Uses /api/health instead of /api/readiness because:
- * - /api/health returns 200 as soon as HTTP server is listening
- * - /api/readiness waits for full initialization (MCP connection can take 5+ minutes)
- * See: https://github.com/thedotmack/claude-mem/issues/811
- * @param port Worker port to check
- * @param timeoutMs Maximum time to wait in milliseconds
- * @returns true if worker became responsive, false if timeout
+ * Poll a localhost endpoint until it returns 200 OK or timeout.
+ * Shared implementation for liveness and readiness checks.
  */
-export async function waitForHealth(port: number, timeoutMs: number = 30000): Promise<boolean> {
+async function pollEndpointUntilOk(
+  port: number,
+  endpointPath: string,
+  timeoutMs: number,
+  retryLogMessage: string
+): Promise<boolean> {
   const start = Date.now();
   while (Date.now() - start < timeoutMs) {
     try {
       // Note: Removed AbortSignal.timeout to avoid Windows Bun cleanup issue (libuv assertion)
-      const response = await fetch(`http://127.0.0.1:${port}/api/health`);
+      const response = await fetch(`http://127.0.0.1:${port}${endpointPath}`);
       if (response.ok) return true;
     } catch (error) {
       // [ANTI-PATTERN IGNORED]: Retry loop - expected failures during startup, will retry
-      logger.debug('SYSTEM', 'Service not ready yet, will retry', { port }, error as Error);
+      logger.debug('SYSTEM', retryLogMessage, { port }, error as Error);
     }
     await new Promise(r => setTimeout(r, 500));
   }
   return false;
 }
 
+/**
+ * Wait for the worker HTTP server to become responsive (liveness check).
+ * Uses /api/health which returns 200 as soon as the HTTP server is listening.
+ * For full initialization (DB + search), use waitForReadiness() instead.
+ */
+export function waitForHealth(port: number, timeoutMs: number = 30000): Promise<boolean> {
+  return pollEndpointUntilOk(port, '/api/health', timeoutMs, 'Service not ready yet, will retry');
+}
+
+/**
+ * Wait for the worker to be fully initialized (DB + search ready).
+ * Uses /api/readiness which returns 200 only after core initialization completes.
+ * Now that initializationCompleteFlag is set after DB/search init (not MCP),
+ * this typically completes in a few seconds.
+ */
+export function waitForReadiness(port: number, timeoutMs: number = 30000): Promise<boolean> {
+  return pollEndpointUntilOk(port, '/api/readiness', timeoutMs, 'Worker not ready yet, will retry');
+}
+
 /**
  * Wait for a port to become free (no longer responding to health checks)
  * Used after shutdown to confirm the port is available for restart
@@ -97,12 +115,22 @@ export async function httpShutdown(port: number): Promise<boolean> {
 
 /**
  * Get the plugin version from the installed marketplace package.json
- * This is the "expected" version that should be running
+ * This is the "expected" version that should be running.
+ * Returns 'unknown' on ENOENT/EBUSY (shutdown race condition, fix #1042).
  */
 export function getInstalledPluginVersion(): string {
-  const packageJsonPath = path.join(MARKETPLACE_ROOT, 'package.json');
-  const packageJson = JSON.parse(readFileSync(packageJsonPath, 'utf-8'));
-  return packageJson.version;
+  try {
+    const packageJsonPath = path.join(MARKETPLACE_ROOT, 'package.json');
+    const packageJson = JSON.parse(readFileSync(packageJsonPath, 'utf-8'));
+    return packageJson.version;
+  } catch (error: unknown) {
+    const code = (error as NodeJS.ErrnoException).code;
+    if (code === 'ENOENT' || code === 'EBUSY') {
+      logger.debug('SYSTEM', 'Could not read plugin version (shutdown race)', { code });
+      return 'unknown';
+    }
+    throw error;
+  }
 }
 
 /**
@@ -137,8 +165,8 @@ export async function checkVersionMatch(port: number): Promise<VersionCheckResul
   const pluginVersion = getInstalledPluginVersion();
   const workerVersion = await getRunningWorkerVersion(port);
 
-  // If we can't get worker version, assume it matches (graceful degradation)
-  if (!workerVersion) {
+  // If either version is unknown/null, assume match (graceful degradation, fix #1042)
+  if (!workerVersion || pluginVersion === 'unknown') {
     return { matches: true, pluginVersion, workerVersion };
   }
 

src/services/infrastructure/ProcessManager.ts

@@ -10,7 +10,7 @@
 
 import path from 'path';
 import { homedir } from 'os';
-import { existsSync, writeFileSync, readFileSync, unlinkSync, mkdirSync } from 'fs';
+import { existsSync, writeFileSync, readFileSync, unlinkSync, mkdirSync, rmSync, statSync, utimesSync } from 'fs';
 import { exec, execSync, spawn } from 'child_process';
 import { promisify } from 'util';
 import { logger } from '../../utils/logger.js';
@@ -54,7 +54,8 @@ function lookupBinaryInPath(binaryName: string, platform: NodeJS.Platform): stri
   try {
     const output = execSync(command, {
       stdio: ['ignore', 'pipe', 'ignore'],
-      encoding: 'utf-8'
+      encoding: 'utf-8',
+      windowsHide: true
     });
 
     const firstMatch = output
@@ -191,10 +192,10 @@ export async function getChildProcesses(parentPid: number): Promise<number[]> {
   }
 
   try {
-    // PowerShell Get-Process instead of WMIC (deprecated in Windows 11)
-    const cmd = `powershell -NoProfile -NonInteractive -Command "Get-Process | Where-Object { $_.ParentProcessId -eq ${parentPid} } | Select-Object -ExpandProperty Id"`;
-    const { stdout } = await execAsync(cmd, { timeout: HOOK_TIMEOUTS.POWERSHELL_COMMAND });
-    // PowerShell outputs just numbers (one per line), simpler than WMIC's "ProcessId=1234" format
+    // Use WQL -Filter to avoid $_ pipeline syntax that breaks in Git Bash (#1062, #1024).
+    // Get-CimInstance with server-side filtering is also more efficient than piping through Where-Object.
+    const cmd = `powershell -NoProfile -NonInteractive -Command "Get-CimInstance Win32_Process -Filter 'ParentProcessId=${parentPid}' | Select-Object -ExpandProperty ProcessId"`;
+    const { stdout } = await execAsync(cmd, { timeout: HOOK_TIMEOUTS.POWERSHELL_COMMAND, windowsHide: true });
     return stdout
       .split('\n')
       .map(line => line.trim())
@@ -223,7 +224,7 @@ export async function forceKillProcess(pid: number): Promise<void> {
   try {
     if (process.platform === 'win32') {
       // /T kills entire process tree, /F forces termination
-      await execAsync(`taskkill /PID ${pid} /T /F`, { timeout: HOOK_TIMEOUTS.POWERSHELL_COMMAND });
+      await execAsync(`taskkill /PID ${pid} /T /F`, { timeout: HOOK_TIMEOUTS.POWERSHELL_COMMAND, windowsHide: true });
     } else {
       process.kill(pid, 'SIGKILL');
     }
@@ -315,13 +316,14 @@ export async function cleanupOrphanedProcesses(): Promise<void> {
 
   try {
     if (isWindows) {
-      // Windows: Use PowerShell Get-CimInstance with JSON output for age filtering
-      const patternConditions = ORPHAN_PROCESS_PATTERNS
-        .map(p => `$_.CommandLine -like '*${p}*'`)
-        .join(' -or ');
+      // Windows: Use WQL -Filter for server-side filtering (no $_ pipeline syntax).
+      // Avoids Git Bash $_ interpretation (#1062) and PowerShell syntax errors (#1024).
+      const wqlPatternConditions = ORPHAN_PROCESS_PATTERNS
+        .map(p => `CommandLine LIKE '%${p}%'`)
+        .join(' OR ');
 
-      const cmd = `powershell -NoProfile -NonInteractive -Command "Get-CimInstance Win32_Process | Where-Object { (${patternConditions}) -and $_.ProcessId -ne ${currentPid} } | Select-Object ProcessId, CreationDate | ConvertTo-Json"`;
-      const { stdout } = await execAsync(cmd, { timeout: HOOK_TIMEOUTS.POWERSHELL_COMMAND });
+      const cmd = `powershell -NoProfile -NonInteractive -Command "Get-CimInstance Win32_Process -Filter '(${wqlPatternConditions}) AND ProcessId != ${currentPid}' | Select-Object ProcessId, CreationDate | ConvertTo-Json"`;
+      const { stdout } = await execAsync(cmd, { timeout: HOOK_TIMEOUTS.POWERSHELL_COMMAND, windowsHide: true });
 
       if (!stdout.trim() || stdout.trim() === 'null') {
         logger.debug('SYSTEM', 'No orphaned claude-mem processes found (Windows)');
@@ -406,7 +408,7 @@ export async function cleanupOrphanedProcesses(): Promise<void> {
         continue;
       }
       try {
-        execSync(`taskkill /PID ${pid} /T /F`, { timeout: HOOK_TIMEOUTS.POWERSHELL_COMMAND, stdio: 'ignore' });
+        execSync(`taskkill /PID ${pid} /T /F`, { timeout: HOOK_TIMEOUTS.POWERSHELL_COMMAND, stdio: 'ignore', windowsHide: true });
       } catch (error) {
         // [ANTI-PATTERN IGNORED]: Cleanup loop - process may have exited, continue to next PID
         logger.debug('SYSTEM', 'Failed to kill process, may have already exited', { pid }, error as Error);
@@ -426,6 +428,184 @@ export async function cleanupOrphanedProcesses(): Promise<void> {
   logger.info('SYSTEM', 'Orphaned processes cleaned up', { count: pidsToKill.length });
 }
 
+// Patterns that should be killed immediately at startup (no age gate)
+// These are child processes that should not outlive their parent worker
+const AGGRESSIVE_CLEANUP_PATTERNS = ['worker-service.cjs', 'chroma-mcp'];
+
+// Patterns that keep the age-gated threshold (may be legitimately running)
+const AGE_GATED_CLEANUP_PATTERNS = ['mcp-server.cjs'];
+
+/**
+ * Aggressive startup cleanup for orphaned claude-mem processes.
+ *
+ * Unlike cleanupOrphanedProcesses() which age-gates everything at 30 minutes,
+ * this function kills worker-service.cjs and chroma-mcp processes immediately
+ * (they should not outlive their parent worker). Only mcp-server.cjs keeps
+ * the age threshold since it may be legitimately running.
+ *
+ * Called once at daemon startup.
+ */
+export async function aggressiveStartupCleanup(): Promise<void> {
+  const isWindows = process.platform === 'win32';
+  const currentPid = process.pid;
+  const pidsToKill: number[] = [];
+  const allPatterns = [...AGGRESSIVE_CLEANUP_PATTERNS, ...AGE_GATED_CLEANUP_PATTERNS];
+
+  try {
+    if (isWindows) {
+      // Use WQL -Filter for server-side filtering (no $_ pipeline syntax).
+      // Avoids Git Bash $_ interpretation (#1062) and PowerShell syntax errors (#1024).
+      const wqlPatternConditions = allPatterns
+        .map(p => `CommandLine LIKE '%${p}%'`)
+        .join(' OR ');
+
+      const cmd = `powershell -NoProfile -NonInteractive -Command "Get-CimInstance Win32_Process -Filter '(${wqlPatternConditions}) AND ProcessId != ${currentPid}' | Select-Object ProcessId, CommandLine, CreationDate | ConvertTo-Json"`;
+      const { stdout } = await execAsync(cmd, { timeout: HOOK_TIMEOUTS.POWERSHELL_COMMAND, windowsHide: true });
+
+      if (!stdout.trim() || stdout.trim() === 'null') {
+        logger.debug('SYSTEM', 'No orphaned claude-mem processes found (Windows)');
+        return;
+      }
+
+      const processes = JSON.parse(stdout);
+      const processList = Array.isArray(processes) ? processes : [processes];
+      const now = Date.now();
+
+      for (const proc of processList) {
+        const pid = proc.ProcessId;
+        if (!Number.isInteger(pid) || pid <= 0 || pid === currentPid) continue;
+
+        const commandLine = proc.CommandLine || '';
+        const isAggressive = AGGRESSIVE_CLEANUP_PATTERNS.some(p => commandLine.includes(p));
+
+        if (isAggressive) {
+          // Kill immediately — no age check
+          pidsToKill.push(pid);
+          logger.debug('SYSTEM', 'Found orphaned process (aggressive)', { pid, commandLine: commandLine.substring(0, 80) });
+        } else {
+          // Age-gated: only kill if older than threshold
+          const creationMatch = proc.CreationDate?.match(/\/Date\((\d+)\)\//);
+          if (creationMatch) {
+            const creationTime = parseInt(creationMatch[1], 10);
+            const ageMinutes = (now - creationTime) / (1000 * 60);
+            if (ageMinutes >= ORPHAN_MAX_AGE_MINUTES) {
+              pidsToKill.push(pid);
+              logger.debug('SYSTEM', 'Found orphaned process (age-gated)', { pid, ageMinutes: Math.round(ageMinutes) });
+            }
+          }
+        }
+      }
+    } else {
+      // Unix: Use ps with elapsed time
+      const patternRegex = allPatterns.join('|');
+      const { stdout } = await execAsync(
+        `ps -eo pid,etime,command | grep -E "${patternRegex}" | grep -v grep || true`
+      );
+
+      if (!stdout.trim()) {
+        logger.debug('SYSTEM', 'No orphaned claude-mem processes found (Unix)');
+        return;
+      }
+
+      const lines = stdout.trim().split('\n');
+      for (const line of lines) {
+        const match = line.trim().match(/^(\d+)\s+(\S+)\s+(.*)$/);
+        if (!match) continue;
+
+        const pid = parseInt(match[1], 10);
+        const etime = match[2];
+        const command = match[3];
+
+        if (!Number.isInteger(pid) || pid <= 0 || pid === currentPid) continue;
+
+        const isAggressive = AGGRESSIVE_CLEANUP_PATTERNS.some(p => command.includes(p));
+
+        if (isAggressive) {
+          // Kill immediately — no age check
+          pidsToKill.push(pid);
+          logger.debug('SYSTEM', 'Found orphaned process (aggressive)', { pid, command: command.substring(0, 80) });
+        } else {
+          // Age-gated: only kill if older than threshold
+          const ageMinutes = parseElapsedTime(etime);
+          if (ageMinutes >= ORPHAN_MAX_AGE_MINUTES) {
+            pidsToKill.push(pid);
+            logger.debug('SYSTEM', 'Found orphaned process (age-gated)', { pid, ageMinutes, command: command.substring(0, 80) });
+          }
+        }
+      }
+    }
+  } catch (error) {
+    logger.error('SYSTEM', 'Failed to enumerate orphaned processes during aggressive cleanup', {}, error as Error);
+    return;
+  }
+
+  if (pidsToKill.length === 0) {
+    return;
+  }
+
+  logger.info('SYSTEM', 'Aggressive startup cleanup: killing orphaned processes', {
+    platform: isWindows ? 'Windows' : 'Unix',
+    count: pidsToKill.length,
+    pids: pidsToKill
+  });
+
+  if (isWindows) {
+    for (const pid of pidsToKill) {
+      if (!Number.isInteger(pid) || pid <= 0) continue;
+      try {
+        execSync(`taskkill /PID ${pid} /T /F`, { timeout: HOOK_TIMEOUTS.POWERSHELL_COMMAND, stdio: 'ignore', windowsHide: true });
+      } catch (error) {
+        logger.debug('SYSTEM', 'Failed to kill process, may have already exited', { pid }, error as Error);
+      }
+    }
+  } else {
+    for (const pid of pidsToKill) {
+      try {
+        process.kill(pid, 'SIGKILL');
+      } catch (error) {
+        logger.debug('SYSTEM', 'Process already exited', { pid }, error as Error);
+      }
+    }
+  }
+
+  logger.info('SYSTEM', 'Aggressive startup cleanup complete', { count: pidsToKill.length });
+}
+
+const CHROMA_MIGRATION_MARKER_FILENAME = '.chroma-cleaned-v10.3';
+
+/**
+ * One-time chroma data wipe for users upgrading from versions with duplicate
+ * worker bugs that could corrupt chroma data. Since chroma is always rebuildable
+ * from SQLite (via backfillAllProjects), this is safe.
+ *
+ * Checks for a marker file. If absent, wipes ~/.claude-mem/chroma/ and writes
+ * the marker. If present, skips. Idempotent.
+ *
+ * @param dataDirectory - Override for DATA_DIR (used in tests)
+ */
+export function runOneTimeChromaMigration(dataDirectory?: string): void {
+  const effectiveDataDir = dataDirectory ?? DATA_DIR;
+  const markerPath = path.join(effectiveDataDir, CHROMA_MIGRATION_MARKER_FILENAME);
+  const chromaDir = path.join(effectiveDataDir, 'chroma');
+
+  if (existsSync(markerPath)) {
+    logger.debug('SYSTEM', 'Chroma migration marker exists, skipping wipe');
+    return;
+  }
+
+  logger.warn('SYSTEM', 'Running one-time chroma data wipe (upgrade from pre-v10.3)', { chromaDir });
+
+  if (existsSync(chromaDir)) {
+    rmSync(chromaDir, { recursive: true, force: true });
+    logger.info('SYSTEM', 'Chroma data directory removed', { chromaDir });
+  }
+
+  // Write marker file to prevent future wipes
+  mkdirSync(effectiveDataDir, { recursive: true });
+  writeFileSync(markerPath, new Date().toISOString());
+  logger.info('SYSTEM', 'Chroma migration marker written', { markerPath });
+}
+
 /**
  * Spawn a detached daemon process
  * Returns the child PID or undefined if spawn failed
@@ -523,10 +703,10 @@ export function spawnDaemon(
  *
  * EPERM is treated as "alive" because it means the process exists but
  * belongs to a different user/session (common in multi-user setups).
- * PID 0 (Windows WMIC sentinel for unknown PID) is treated as alive.
+ * PID 0 (Windows sentinel for unknown PID) is treated as alive.
  */
 export function isProcessAlive(pid: number): boolean {
-  // PID 0 is the Windows WMIC sentinel value — process was spawned but PID unknown
+  // PID 0 is the Windows sentinel value — process was spawned but PID unknown
   if (pid === 0) return true;
 
   // Invalid PIDs are not alive
@@ -544,6 +724,39 @@ export function isProcessAlive(pid: number): boolean {
   }
 }
 
+/**
+ * Check if the PID file was written recently (within thresholdMs).
+ *
+ * Used to coordinate restarts across concurrent sessions: if the PID file
+ * was recently written, another session likely just restarted the worker.
+ * Callers should poll /api/health instead of attempting their own restart.
+ *
+ * @param thresholdMs - Maximum age in ms to consider "recent" (default: 15000)
+ * @returns true if the PID file exists and was modified within thresholdMs
+ */
+export function isPidFileRecent(thresholdMs: number = 15000): boolean {
+  try {
+    const stats = statSync(PID_FILE);
+    return (Date.now() - stats.mtimeMs) < thresholdMs;
+  } catch {
+    return false;
+  }
+}
+
+/**
+ * Touch the PID file to update its mtime without changing contents.
+ * Used after a restart to signal other sessions that a restart just completed.
+ */
+export function touchPidFile(): void {
+  try {
+    if (!existsSync(PID_FILE)) return;
+    const now = new Date();
+    utimesSync(PID_FILE, now, now);
+  } catch {
+    // Best-effort — failure to touch doesn't affect correctness
+  }
+}
+
 /**
  * Read the PID file and remove it if the recorded process is dead (stale).
  *

src/services/queue/SessionQueueProcessor.ts

@@ -20,8 +20,9 @@ export class SessionQueueProcessor {
 
   /**
    * Create an async iterator that yields messages as they become available.
-   * Uses atomic claim-and-delete to prevent duplicates.
-   * The queue is a pure buffer: claim it, delete it, process in memory.
+   * Uses atomic claim-confirm to prevent duplicates.
+   * Messages are claimed (marked processing) and stay in DB until confirmProcessed().
+   * Self-heals stale processing messages before each claim.
    * Waits for 'message' event when queue is empty.
    *
    * CRITICAL: Calls onIdleTimeout callback after 3 minutes of inactivity.
@@ -34,14 +35,14 @@ export class SessionQueueProcessor {
 
     while (!signal.aborted) {
       try {
-        // Atomically claim AND DELETE next message from DB
-        // Message is now in memory only - no "processing" state tracking needed
-        const persistentMessage = this.store.claimAndDelete(sessionDbId);
+        // Atomically claim next pending message (marks as 'processing')
+        // Self-heals any stale processing messages before claiming
+        const persistentMessage = this.store.claimNextMessage(sessionDbId);
 
         if (persistentMessage) {
           // Reset activity time when we successfully yield a message
           lastActivityTime = Date.now();
-          // Yield the message for processing (it's already deleted from queue)
+          // Yield the message for processing (it's marked as 'processing' in DB)
           yield this.toPendingMessageWithId(persistentMessage);
         } else {
           // Queue empty - wait for wake-up event or timeout

src/services/server/Server.ts

@@ -248,8 +248,14 @@ export class Server {
         process.send!({ type: 'restart' });
       } else {
         // Unix or standalone Windows - handle restart ourselves
+        // The spawner (ensureWorkerStarted/restart command) handles spawning the new daemon.
+        // This process just needs to shut down and exit.
         setTimeout(async () => {
-          await this.options.onRestart();
+          try {
+            await this.options.onRestart();
+          } finally {
+            process.exit(0);
+          }
         }, 100);
       }
     });
@@ -268,7 +274,14 @@ export class Server {
       } else {
         // Unix or standalone Windows - handle shutdown ourselves
         setTimeout(async () => {
-          await this.options.onShutdown();
+          try {
+            await this.options.onShutdown();
+          } finally {
+            // CRITICAL: Exit the process after shutdown completes (or fails).
+            // Without this, the daemon stays alive as a zombie — background tasks
+            // (backfill, reconnects) keep running and respawn chroma-mcp subprocesses.
+            process.exit(0);
+          }
         }, 100);
       }
     });

src/services/smart-file-read/parser.ts

@@ -0,0 +1,666 @@
+/**
+ * Code structure parser — shells out to tree-sitter CLI for AST-based extraction.
+ *
+ * No native bindings. No WASM. Just the CLI binary + query patterns.
+ *
+ * Supported: JS, TS, Python, Go, Rust, Ruby, Java, C, C++
+ *
+ * by Copter Labs
+ */
+
+import { execFileSync } from "node:child_process";
+import { writeFileSync, mkdtempSync, rmSync, existsSync } from "node:fs";
+import { join, dirname } from "node:path";
+import { tmpdir } from "node:os";
+import { createRequire } from "node:module";
+
+// CJS-safe require for resolving external packages at runtime.
+// In ESM: import.meta.url works. In CJS bundle (esbuild): __filename works.
+// typeof check avoids ReferenceError in ESM where __filename doesn't exist.
+const _require = typeof __filename !== 'undefined'
+  ? createRequire(__filename)
+  : createRequire(import.meta.url);
+
+// --- Types ---
+
+export interface CodeSymbol {
+  name: string;
+  kind: "function" | "class" | "method" | "interface" | "type" | "const" | "variable" | "export" | "struct" | "enum" | "trait" | "impl" | "property" | "getter" | "setter";
+  signature: string;
+  jsdoc?: string;
+  lineStart: number;
+  lineEnd: number;
+  parent?: string;
+  exported: boolean;
+  children?: CodeSymbol[];
+}
+
+export interface FoldedFile {
+  filePath: string;
+  language: string;
+  symbols: CodeSymbol[];
+  imports: string[];
+  totalLines: number;
+  foldedTokenEstimate: number;
+}
+
+// --- Language detection ---
+
+const LANG_MAP: Record<string, string> = {
+  ".js": "javascript",
+  ".mjs": "javascript",
+  ".cjs": "javascript",
+  ".jsx": "tsx",
+  ".ts": "typescript",
+  ".tsx": "tsx",
+  ".py": "python",
+  ".pyw": "python",
+  ".go": "go",
+  ".rs": "rust",
+  ".rb": "ruby",
+  ".java": "java",
+  ".c": "c",
+  ".h": "c",
+  ".cpp": "cpp",
+  ".cc": "cpp",
+  ".cxx": "cpp",
+  ".hpp": "cpp",
+  ".hh": "cpp",
+};
+
+export function detectLanguage(filePath: string): string {
+  const ext = filePath.slice(filePath.lastIndexOf("."));
+  return LANG_MAP[ext] || "unknown";
+}
+
+// --- Grammar path resolution ---
+
+const GRAMMAR_PACKAGES: Record<string, string> = {
+  javascript: "tree-sitter-javascript",
+  typescript: "tree-sitter-typescript/typescript",
+  tsx: "tree-sitter-typescript/tsx",
+  python: "tree-sitter-python",
+  go: "tree-sitter-go",
+  rust: "tree-sitter-rust",
+  ruby: "tree-sitter-ruby",
+  java: "tree-sitter-java",
+  c: "tree-sitter-c",
+  cpp: "tree-sitter-cpp",
+};
+
+function resolveGrammarPath(language: string): string | null {
+  const pkg = GRAMMAR_PACKAGES[language];
+  if (!pkg) return null;
+  try {
+    const packageJsonPath = _require.resolve(pkg + "/package.json");
+    return dirname(packageJsonPath);
+  } catch {
+    return null;
+  }
+}
+
+// --- Query patterns (declarative symbol extraction) ---
+
+const QUERIES: Record<string, string> = {
+  jsts: `
+(function_declaration name: (identifier) @name) @func
+(lexical_declaration (variable_declarator name: (identifier) @name value: [(arrow_function) (function_expression)])) @const_func
+(class_declaration name: (type_identifier) @name) @cls
+(method_definition name: (property_identifier) @name) @method
+(interface_declaration name: (type_identifier) @name) @iface
+(type_alias_declaration name: (type_identifier) @name) @tdef
+(enum_declaration name: (identifier) @name) @enm
+(import_statement) @imp
+(export_statement) @exp
+`,
+
+  python: `
+(function_definition name: (identifier) @name) @func
+(class_definition name: (identifier) @name) @cls
+(import_statement) @imp
+(import_from_statement) @imp
+`,
+
+  go: `
+(function_declaration name: (identifier) @name) @func
+(method_declaration name: (field_identifier) @name) @method
+(type_declaration (type_spec name: (type_identifier) @name)) @tdef
+(import_declaration) @imp
+`,
+
+  rust: `
+(function_item name: (identifier) @name) @func
+(struct_item name: (type_identifier) @name) @struct_def
+(enum_item name: (type_identifier) @name) @enm
+(trait_item name: (type_identifier) @name) @trait_def
+(impl_item type: (type_identifier) @name) @impl_def
+(use_declaration) @imp
+`,
+
+  ruby: `
+(method name: (identifier) @name) @func
+(class name: (constant) @name) @cls
+(module name: (constant) @name) @cls
+(call method: (identifier) @name) @imp
+`,
+
+  java: `
+(method_declaration name: (identifier) @name) @method
+(class_declaration name: (identifier) @name) @cls
+(interface_declaration name: (identifier) @name) @iface
+(enum_declaration name: (identifier) @name) @enm
+(import_declaration) @imp
+`,
+
+  generic: `
+(function_declaration name: (identifier) @name) @func
+(function_definition name: (identifier) @name) @func
+(class_declaration name: (identifier) @name) @cls
+(class_definition name: (identifier) @name) @cls
+(import_statement) @imp
+(import_declaration) @imp
+`,
+};
+
+function getQueryKey(language: string): string {
+  switch (language) {
+    case "javascript":
+    case "typescript":
+    case "tsx":
+      return "jsts";
+    case "python": return "python";
+    case "go": return "go";
+    case "rust": return "rust";
+    case "ruby": return "ruby";
+    case "java": return "java";
+    default: return "generic";
+  }
+}
+
+// --- Temp file management ---
+
+let queryTmpDir: string | null = null;
+const queryFileCache = new Map<string, string>();
+
+function getQueryFile(queryKey: string): string {
+  if (queryFileCache.has(queryKey)) return queryFileCache.get(queryKey)!;
+
+  if (!queryTmpDir) {
+    queryTmpDir = mkdtempSync(join(tmpdir(), "smart-read-queries-"));
+  }
+
+  const filePath = join(queryTmpDir, `${queryKey}.scm`);
+  writeFileSync(filePath, QUERIES[queryKey]);
+  queryFileCache.set(queryKey, filePath);
+  return filePath;
+}
+
+// --- CLI execution ---
+
+let cachedBinPath: string | null = null;
+
+function getTreeSitterBin(): string {
+  if (cachedBinPath) return cachedBinPath;
+
+  // Try direct binary from tree-sitter-cli package
+  try {
+    const pkgPath = _require.resolve("tree-sitter-cli/package.json");
+    const binPath = join(dirname(pkgPath), "tree-sitter");
+    if (existsSync(binPath)) {
+      cachedBinPath = binPath;
+      return binPath;
+    }
+  } catch { /* fall through */ }
+
+  // Fallback: assume it's on PATH
+  cachedBinPath = "tree-sitter";
+  return cachedBinPath;
+}
+
+interface RawCapture {
+  tag: string;
+  startRow: number;
+  startCol: number;
+  endRow: number;
+  endCol: number;
+  text?: string;
+}
+
+interface RawMatch {
+  pattern: number;
+  captures: RawCapture[];
+}
+
+function runQuery(queryFile: string, sourceFile: string, grammarPath: string): RawMatch[] {
+  const result = runBatchQuery(queryFile, [sourceFile], grammarPath);
+  return result.get(sourceFile) || [];
+}
+
+function runBatchQuery(queryFile: string, sourceFiles: string[], grammarPath: string): Map<string, RawMatch[]> {
+  if (sourceFiles.length === 0) return new Map();
+
+  const bin = getTreeSitterBin();
+  const execArgs = ["query", "-p", grammarPath, queryFile, ...sourceFiles];
+
+  let output: string;
+  try {
+    output = execFileSync(bin, execArgs, { encoding: "utf-8", timeout: 30000, stdio: ["pipe", "pipe", "pipe"] });
+  } catch {
+    return new Map();
+  }
+
+  return parseMultiFileQueryOutput(output);
+}
+
+function parseMultiFileQueryOutput(output: string): Map<string, RawMatch[]> {
+  const fileMatches = new Map<string, RawMatch[]>();
+  let currentFile: string | null = null;
+  let currentMatch: RawMatch | null = null;
+
+  for (const line of output.split("\n")) {
+    // File header: a line that doesn't start with whitespace and isn't empty
+    if (line.length > 0 && !line.startsWith(" ") && !line.startsWith("\t")) {
+      currentFile = line.trim();
+      if (!fileMatches.has(currentFile)) {
+        fileMatches.set(currentFile, []);
+      }
+      currentMatch = null;
+      continue;
+    }
+
+    if (!currentFile) continue;
+
+    const patternMatch = line.match(/^\s+pattern:\s+(\d+)/);
+    if (patternMatch) {
+      currentMatch = { pattern: parseInt(patternMatch[1]), captures: [] };
+      fileMatches.get(currentFile)!.push(currentMatch);
+      continue;
+    }
+
+    const captureMatch = line.match(
+      /^\s+capture:\s+(?:\d+\s*-\s*)?(\w+),\s*start:\s*\((\d+),\s*(\d+)\),\s*end:\s*\((\d+),\s*(\d+)\)(?:,\s*text:\s*`([^`]*)`)?/
+    );
+    if (captureMatch && currentMatch) {
+      currentMatch.captures.push({
+        tag: captureMatch[1],
+        startRow: parseInt(captureMatch[2]),
+        startCol: parseInt(captureMatch[3]),
+        endRow: parseInt(captureMatch[4]),
+        endCol: parseInt(captureMatch[5]),
+        text: captureMatch[6],
+      });
+    }
+  }
+
+  return fileMatches;
+}
+
+// --- Symbol building ---
+
+const KIND_MAP: Record<string, CodeSymbol["kind"]> = {
+  func: "function",
+  const_func: "function",
+  cls: "class",
+  method: "method",
+  iface: "interface",
+  tdef: "type",
+  enm: "enum",
+  struct_def: "struct",
+  trait_def: "trait",
+  impl_def: "impl",
+};
+
+const CONTAINER_KINDS = new Set(["class", "struct", "impl", "trait"]);
+
+function extractSignatureFromLines(lines: string[], startRow: number, endRow: number, maxLen: number = 200): string {
+  const firstLine = lines[startRow] || "";
+  let sig = firstLine;
+
+  if (!sig.trimEnd().endsWith("{") && !sig.trimEnd().endsWith(":")) {
+    const chunk = lines.slice(startRow, Math.min(startRow + 10, endRow + 1)).join("\n");
+    const braceIdx = chunk.indexOf("{");
+    if (braceIdx !== -1 && braceIdx < 500) {
+      sig = chunk.slice(0, braceIdx).replace(/\n/g, " ").replace(/\s+/g, " ").trim();
+    }
+  }
+
+  sig = sig.replace(/\s*[{:]\s*$/, "").trim();
+  if (sig.length > maxLen) sig = sig.slice(0, maxLen - 3) + "...";
+  return sig;
+}
+
+function findCommentAbove(lines: string[], startRow: number): string | undefined {
+  const commentLines: string[] = [];
+  let foundComment = false;
+
+  for (let i = startRow - 1; i >= 0; i--) {
+    const trimmed = lines[i].trim();
+    if (trimmed === "") {
+      if (foundComment) break;
+      continue;
+    }
+    if (trimmed.startsWith("/**") || trimmed.startsWith("*") || trimmed.startsWith("*/") ||
+        trimmed.startsWith("//") || trimmed.startsWith("///") || trimmed.startsWith("//!") ||
+        trimmed.startsWith("#") || trimmed.startsWith("@")) {
+      commentLines.unshift(lines[i]);
+      foundComment = true;
+    } else {
+      break;
+    }
+  }
+
+  return commentLines.length > 0 ? commentLines.join("\n").trim() : undefined;
+}
+
+function findPythonDocstringFromLines(lines: string[], startRow: number, endRow: number): string | undefined {
+  for (let i = startRow + 1; i <= Math.min(startRow + 3, endRow); i++) {
+    const trimmed = lines[i]?.trim();
+    if (!trimmed) continue;
+    if (trimmed.startsWith('"""') || trimmed.startsWith("'''")) return trimmed;
+    break;
+  }
+  return undefined;
+}
+
+function isExported(
+  name: string, startRow: number, endRow: number,
+  exportRanges: Array<{ startRow: number; endRow: number }>,
+  lines: string[], language: string
+): boolean {
+  switch (language) {
+    case "javascript":
+    case "typescript":
+    case "tsx":
+      return exportRanges.some(r => startRow >= r.startRow && endRow <= r.endRow);
+    case "python":
+      return !name.startsWith("_");
+    case "go":
+      return name.length > 0 && name[0] === name[0].toUpperCase() && name[0] !== name[0].toLowerCase();
+    case "rust":
+      return lines[startRow]?.trimStart().startsWith("pub") ?? false;
+    default:
+      return true;
+  }
+}
+
+function buildSymbols(matches: RawMatch[], lines: string[], language: string): { symbols: CodeSymbol[]; imports: string[] } {
+  const symbols: CodeSymbol[] = [];
+  const imports: string[] = [];
+  const exportRanges: Array<{ startRow: number; endRow: number }> = [];
+  const containers: Array<{ sym: CodeSymbol; startRow: number; endRow: number }> = [];
+
+  // Collect exports and imports
+  for (const match of matches) {
+    for (const cap of match.captures) {
+      if (cap.tag === "exp") {
+        exportRanges.push({ startRow: cap.startRow, endRow: cap.endRow });
+      }
+      if (cap.tag === "imp") {
+        imports.push(cap.text || lines[cap.startRow]?.trim() || "");
+      }
+    }
+  }
+
+  // Build symbols
+  for (const match of matches) {
+    const kindCapture = match.captures.find(c => KIND_MAP[c.tag]);
+    const nameCapture = match.captures.find(c => c.tag === "name");
+    if (!kindCapture) continue;
+
+    const name = nameCapture?.text || "anonymous";
+    const startRow = kindCapture.startRow;
+    const endRow = kindCapture.endRow;
+    const kind = KIND_MAP[kindCapture.tag];
+
+    const comment = findCommentAbove(lines, startRow);
+    const docstring = language === "python" ? findPythonDocstringFromLines(lines, startRow, endRow) : undefined;
+
+    const sym: CodeSymbol = {
+      name,
+      kind,
+      signature: extractSignatureFromLines(lines, startRow, endRow),
+      jsdoc: comment || docstring,
+      lineStart: startRow,
+      lineEnd: endRow,
+      exported: isExported(name, startRow, endRow, exportRanges, lines, language),
+    };
+
+    if (CONTAINER_KINDS.has(kind)) {
+      sym.children = [];
+      containers.push({ sym, startRow, endRow });
+    }
+
+    symbols.push(sym);
+  }
+
+  // Nest methods inside containers
+  const nested = new Set<CodeSymbol>();
+  for (const container of containers) {
+    for (const sym of symbols) {
+      if (sym === container.sym) continue;
+      if (sym.lineStart > container.startRow && sym.lineEnd <= container.endRow) {
+        if (sym.kind === "function") sym.kind = "method";
+        container.sym.children!.push(sym);
+        nested.add(sym);
+      }
+    }
+  }
+
+  return { symbols: symbols.filter(s => !nested.has(s)), imports };
+}
+
+// --- Main parse functions ---
+
+export function parseFile(content: string, filePath: string): FoldedFile {
+  const language = detectLanguage(filePath);
+  const lines = content.split("\n");
+
+  const grammarPath = resolveGrammarPath(language);
+  if (!grammarPath) {
+    return {
+      filePath, language, symbols: [], imports: [],
+      totalLines: lines.length, foldedTokenEstimate: 50,
+    };
+  }
+
+  const queryKey = getQueryKey(language);
+  const queryFile = getQueryFile(queryKey);
+
+  // Write content to temp file with correct extension for language detection
+  const ext = filePath.slice(filePath.lastIndexOf(".")) || ".txt";
+  const tmpDir = mkdtempSync(join(tmpdir(), "smart-src-"));
+  const tmpFile = join(tmpDir, `source${ext}`);
+  writeFileSync(tmpFile, content);
+
+  try {
+    const matches = runQuery(queryFile, tmpFile, grammarPath);
+    const result = buildSymbols(matches, lines, language);
+
+    const folded = formatFoldedView({
+      filePath, language,
+      symbols: result.symbols, imports: result.imports,
+      totalLines: lines.length, foldedTokenEstimate: 0,
+    });
+
+    return {
+      filePath, language,
+      symbols: result.symbols, imports: result.imports,
+      totalLines: lines.length,
+      foldedTokenEstimate: Math.ceil(folded.length / 4),
+    };
+  } finally {
+    rmSync(tmpDir, { recursive: true, force: true });
+  }
+}
+
+/**
+ * Batch parse multiple on-disk files. Groups by language for one CLI call per language.
+ * Much faster than calling parseFile() per file (one process spawn per language vs per file).
+ */
+export function parseFilesBatch(
+  files: Array<{ absolutePath: string; relativePath: string; content: string }>
+): Map<string, FoldedFile> {
+  const results = new Map<string, FoldedFile>();
+
+  // Group files by language (and thus by query + grammar)
+  const languageGroups = new Map<string, typeof files>();
+  for (const file of files) {
+    const language = detectLanguage(file.relativePath);
+    if (!languageGroups.has(language)) languageGroups.set(language, []);
+    languageGroups.get(language)!.push(file);
+  }
+
+  for (const [language, groupFiles] of languageGroups) {
+    const grammarPath = resolveGrammarPath(language);
+    if (!grammarPath) {
+      // No grammar — return empty results for these files
+      for (const file of groupFiles) {
+        const lines = file.content.split("\n");
+        results.set(file.relativePath, {
+          filePath: file.relativePath, language, symbols: [], imports: [],
+          totalLines: lines.length, foldedTokenEstimate: 50,
+        });
+      }
+      continue;
+    }
+
+    const queryKey = getQueryKey(language);
+    const queryFile = getQueryFile(queryKey);
+
+    // Run one batch query for all files of this language
+    const absolutePaths = groupFiles.map(f => f.absolutePath);
+    const batchResults = runBatchQuery(queryFile, absolutePaths, grammarPath);
+
+    // Build FoldedFile for each file using the batch results
+    for (const file of groupFiles) {
+      const lines = file.content.split("\n");
+      const matches = batchResults.get(file.absolutePath) || [];
+      const symbolResult = buildSymbols(matches, lines, language);
+
+      const folded = formatFoldedView({
+        filePath: file.relativePath, language,
+        symbols: symbolResult.symbols, imports: symbolResult.imports,
+        totalLines: lines.length, foldedTokenEstimate: 0,
+      });
+
+      results.set(file.relativePath, {
+        filePath: file.relativePath, language,
+        symbols: symbolResult.symbols, imports: symbolResult.imports,
+        totalLines: lines.length,
+        foldedTokenEstimate: Math.ceil(folded.length / 4),
+      });
+    }
+  }
+
+  return results;
+}
+
+// --- Formatting ---
+
+export function formatFoldedView(file: FoldedFile): string {
+  const parts: string[] = [];
+
+  parts.push(`📁 ${file.filePath} (${file.language}, ${file.totalLines} lines)`);
+  parts.push("");
+
+  if (file.imports.length > 0) {
+    parts.push(`  📦 Imports: ${file.imports.length} statements`);
+    for (const imp of file.imports.slice(0, 10)) {
+      parts.push(`    ${imp}`);
+    }
+    if (file.imports.length > 10) {
+      parts.push(`    ... +${file.imports.length - 10} more`);
+    }
+    parts.push("");
+  }
+
+  for (const sym of file.symbols) {
+    parts.push(formatSymbol(sym, "  "));
+  }
+
+  return parts.join("\n");
+}
+
+function formatSymbol(sym: CodeSymbol, indent: string): string {
+  const parts: string[] = [];
+
+  const icon = getSymbolIcon(sym.kind);
+  const exportTag = sym.exported ? " [exported]" : "";
+  const lineRange = sym.lineStart === sym.lineEnd
+    ? `L${sym.lineStart + 1}`
+    : `L${sym.lineStart + 1}-${sym.lineEnd + 1}`;
+
+  parts.push(`${indent}${icon} ${sym.name}${exportTag} (${lineRange})`);
+  parts.push(`${indent}  ${sym.signature}`);
+
+  if (sym.jsdoc) {
+    const jsdocLines = sym.jsdoc.split("\n");
+    const firstLine = jsdocLines.find(l => {
+      const t = l.replace(/^[\s*/]+/, "").replace(/^['"`]{3}/, "").trim();
+      return t.length > 0 && !t.startsWith("/**");
+    });
+    if (firstLine) {
+      const cleaned = firstLine.replace(/^[\s*/]+/, "").replace(/^['"`]{3}/, "").replace(/['"`]{3}$/, "").trim();
+      if (cleaned) {
+        parts.push(`${indent}  💬 ${cleaned}`);
+      }
+    }
+  }
+
+  if (sym.children && sym.children.length > 0) {
+    for (const child of sym.children) {
+      parts.push(formatSymbol(child, indent + "  "));
+    }
+  }
+
+  return parts.join("\n");
+}
+
+function getSymbolIcon(kind: CodeSymbol["kind"]): string {
+  const icons: Record<string, string> = {
+    function: "ƒ", method: "ƒ", class: "◆", interface: "◇",
+    type: "◇", const: "●", variable: "○", export: "→",
+    struct: "◆", enum: "▣", trait: "◇", impl: "◈",
+    property: "○", getter: "⇢", setter: "⇠",
+  };
+  return icons[kind] || "·";
+}
+
+// --- Unfold ---
+
+export function unfoldSymbol(content: string, filePath: string, symbolName: string): string | null {
+  const file = parseFile(content, filePath);
+
+  const findSymbol = (symbols: CodeSymbol[]): CodeSymbol | null => {
+    for (const sym of symbols) {
+      if (sym.name === symbolName) return sym;
+      if (sym.children) {
+        const found = findSymbol(sym.children);
+        if (found) return found;
+      }
+    }
+    return null;
+  };
+
+  const symbol = findSymbol(file.symbols);
+  if (!symbol) return null;
+
+  const lines = content.split("\n");
+
+  // Include preceding comments/decorators
+  let start = symbol.lineStart;
+  for (let i = symbol.lineStart - 1; i >= 0; i--) {
+    const trimmed = lines[i].trim();
+    if (trimmed === "" || trimmed.startsWith("*") || trimmed.startsWith("/**") ||
+        trimmed.startsWith("///") || trimmed.startsWith("//") ||
+        trimmed.startsWith("#") || trimmed.startsWith("@") ||
+        trimmed === "*/") {
+      start = i;
+    } else {
+      break;
+    }
+  }
+
+  const extracted = lines.slice(start, symbol.lineEnd + 1).join("\n");
+  return `// 📍 ${filePath} L${start + 1}-${symbol.lineEnd + 1}\n${extracted}`;
+}

src/services/smart-file-read/search.ts

@@ -0,0 +1,316 @@
+/**
+ * Search module — finds code files and symbols matching a query.
+ *
+ * Two search modes:
+ * 1. Grep-style: find files/lines containing the query string
+ * 2. Structural: parse files and match against symbol names/signatures
+ *
+ * Both return folded views, not raw content.
+ *
+ * Uses batch parsing (one CLI call per language) for fast multi-file search.
+ */
+
+import { readFile, readdir, stat } from "node:fs/promises";
+import { join, relative } from "node:path";
+import { parseFilesBatch, formatFoldedView, type FoldedFile } from "./parser.js";
+
+const CODE_EXTENSIONS = new Set([
+  ".js", ".jsx", ".ts", ".tsx", ".mjs", ".cjs",
+  ".py", ".pyw",
+  ".go",
+  ".rs",
+  ".rb",
+  ".java",
+  ".cs",
+  ".cpp", ".c", ".h", ".hpp",
+  ".swift",
+  ".kt",
+  ".php",
+  ".vue", ".svelte",
+]);
+
+const IGNORE_DIRS = new Set([
+  "node_modules", ".git", "dist", "build", ".next", "__pycache__",
+  ".venv", "venv", "env", ".env", "target", "vendor",
+  ".cache", ".turbo", "coverage", ".nyc_output",
+  ".claude", ".smart-file-read",
+]);
+
+const MAX_FILE_SIZE = 512 * 1024; // 512KB — skip huge files
+
+export interface SearchResult {
+  foldedFiles: FoldedFile[];
+  matchingSymbols: SymbolMatch[];
+  totalFilesScanned: number;
+  totalSymbolsFound: number;
+  tokenEstimate: number;
+}
+
+export interface SymbolMatch {
+  filePath: string;
+  symbolName: string;
+  kind: string;
+  signature: string;
+  jsdoc?: string;
+  lineStart: number;
+  lineEnd: number;
+  matchReason: string; // why this matched
+}
+
+/**
+ * Walk a directory recursively, yielding file paths.
+ */
+async function* walkDir(dir: string, rootDir: string, maxDepth: number = 20): AsyncGenerator<string> {
+  if (maxDepth <= 0) return;
+
+  let entries;
+  try {
+    entries = await readdir(dir, { withFileTypes: true });
+  } catch {
+    return; // permission denied, etc.
+  }
+
+  for (const entry of entries) {
+    if (entry.name.startsWith(".") && entry.name !== ".") continue;
+    if (IGNORE_DIRS.has(entry.name)) continue;
+
+    const fullPath = join(dir, entry.name);
+
+    if (entry.isDirectory()) {
+      yield* walkDir(fullPath, rootDir, maxDepth - 1);
+    } else if (entry.isFile()) {
+      const ext = entry.name.slice(entry.name.lastIndexOf("."));
+      if (CODE_EXTENSIONS.has(ext)) {
+        yield fullPath;
+      }
+    }
+  }
+}
+
+/**
+ * Read a file safely, skipping if too large or binary.
+ */
+async function safeReadFile(filePath: string): Promise<string | null> {
+  try {
+    const stats = await stat(filePath);
+    if (stats.size > MAX_FILE_SIZE) return null;
+    if (stats.size === 0) return null;
+
+    const content = await readFile(filePath, "utf-8");
+
+    // Quick binary check — if first 1000 chars have null bytes, skip
+    if (content.slice(0, 1000).includes("\0")) return null;
+
+    return content;
+  } catch {
+    return null;
+  }
+}
+
+/**
+ * Search a codebase for symbols matching a query.
+ *
+ * Phase 1: Collect files and read content
+ * Phase 2: Batch parse all files (one CLI call per language)
+ * Phase 3: Match query against parsed symbols
+ */
+export async function searchCodebase(
+  rootDir: string,
+  query: string,
+  options: {
+    maxResults?: number;
+    includeImports?: boolean;
+    filePattern?: string;
+  } = {}
+): Promise<SearchResult> {
+  const maxResults = options.maxResults || 20;
+  const queryLower = query.toLowerCase();
+  const queryParts = queryLower.split(/[\s_\-./]+/).filter(p => p.length > 0);
+
+  // Phase 1: Collect files
+  const filesToParse: Array<{ absolutePath: string; relativePath: string; content: string }> = [];
+
+  for await (const filePath of walkDir(rootDir, rootDir)) {
+    if (options.filePattern) {
+      const relPath = relative(rootDir, filePath);
+      if (!relPath.toLowerCase().includes(options.filePattern.toLowerCase())) continue;
+    }
+
+    const content = await safeReadFile(filePath);
+    if (!content) continue;
+
+    filesToParse.push({
+      absolutePath: filePath,
+      relativePath: relative(rootDir, filePath),
+      content,
+    });
+  }
+
+  // Phase 2: Batch parse (one CLI call per language)
+  const parsedFiles = parseFilesBatch(filesToParse);
+
+  // Phase 3: Match query against symbols
+  const foldedFiles: FoldedFile[] = [];
+  const matchingSymbols: SymbolMatch[] = [];
+  let totalSymbolsFound = 0;
+
+  for (const [relPath, parsed] of parsedFiles) {
+    totalSymbolsFound += countSymbols(parsed);
+
+    const pathMatch = matchScore(relPath.toLowerCase(), queryParts);
+    let fileHasMatch = pathMatch > 0;
+    const fileSymbolMatches: SymbolMatch[] = [];
+
+    const checkSymbols = (symbols: typeof parsed.symbols, parent?: string) => {
+      for (const sym of symbols) {
+        let score = 0;
+        let reason = "";
+
+        const nameScore = matchScore(sym.name.toLowerCase(), queryParts);
+        if (nameScore > 0) {
+          score += nameScore * 3;
+          reason = "name match";
+        }
+
+        if (sym.signature.toLowerCase().includes(queryLower)) {
+          score += 2;
+          reason = reason ? `${reason} + signature` : "signature match";
+        }
+
+        if (sym.jsdoc && sym.jsdoc.toLowerCase().includes(queryLower)) {
+          score += 1;
+          reason = reason ? `${reason} + jsdoc` : "jsdoc match";
+        }
+
+        if (score > 0) {
+          fileHasMatch = true;
+          fileSymbolMatches.push({
+            filePath: relPath,
+            symbolName: parent ? `${parent}.${sym.name}` : sym.name,
+            kind: sym.kind,
+            signature: sym.signature,
+            jsdoc: sym.jsdoc,
+            lineStart: sym.lineStart,
+            lineEnd: sym.lineEnd,
+            matchReason: reason,
+          });
+        }
+
+        if (sym.children) {
+          checkSymbols(sym.children, sym.name);
+        }
+      }
+    };
+
+    checkSymbols(parsed.symbols);
+
+    if (fileHasMatch) {
+      foldedFiles.push(parsed);
+      matchingSymbols.push(...fileSymbolMatches);
+    }
+  }
+
+  // Sort by relevance and trim
+  matchingSymbols.sort((a, b) => {
+    const aScore = matchScore(a.symbolName.toLowerCase(), queryParts);
+    const bScore = matchScore(b.symbolName.toLowerCase(), queryParts);
+    return bScore - aScore;
+  });
+
+  const trimmedSymbols = matchingSymbols.slice(0, maxResults);
+  const relevantFiles = new Set(trimmedSymbols.map(s => s.filePath));
+  const trimmedFiles = foldedFiles.filter(f => relevantFiles.has(f.filePath)).slice(0, maxResults);
+
+  const tokenEstimate = trimmedFiles.reduce((sum, f) => sum + f.foldedTokenEstimate, 0);
+
+  return {
+    foldedFiles: trimmedFiles,
+    matchingSymbols: trimmedSymbols,
+    totalFilesScanned: filesToParse.length,
+    totalSymbolsFound,
+    tokenEstimate,
+  };
+}
+
+/**
+ * Score how well query parts match a string.
+ * Returns 0 for no match, higher for better matches.
+ */
+function matchScore(text: string, queryParts: string[]): number {
+  let score = 0;
+  for (const part of queryParts) {
+    if (text === part) {
+      score += 10; // exact match
+    } else if (text.includes(part)) {
+      score += 5; // substring match
+    } else {
+      // Fuzzy: check if all chars appear in order
+      let ti = 0;
+      let matched = 0;
+      for (const ch of part) {
+        const idx = text.indexOf(ch, ti);
+        if (idx !== -1) {
+          matched++;
+          ti = idx + 1;
+        }
+      }
+      if (matched === part.length) {
+        score += 1; // loose fuzzy match
+      }
+    }
+  }
+  return score;
+}
+
+function countSymbols(file: FoldedFile): number {
+  let count = file.symbols.length;
+  for (const sym of file.symbols) {
+    if (sym.children) count += sym.children.length;
+  }
+  return count;
+}
+
+/**
+ * Format search results for LLM consumption.
+ */
+export function formatSearchResults(result: SearchResult, query: string): string {
+  const parts: string[] = [];
+
+  parts.push(`🔍 Smart Search: "${query}"`);
+  parts.push(`   Scanned ${result.totalFilesScanned} files, found ${result.totalSymbolsFound} symbols`);
+  parts.push(`   ${result.matchingSymbols.length} matches across ${result.foldedFiles.length} files (~${result.tokenEstimate} tokens for folded view)`);
+  parts.push("");
+
+  if (result.matchingSymbols.length === 0) {
+    parts.push("   No matching symbols found.");
+    return parts.join("\n");
+  }
+
+  // Show matching symbols first (compact)
+  parts.push("── Matching Symbols ──");
+  parts.push("");
+  for (const match of result.matchingSymbols) {
+    parts.push(`  ${match.kind} ${match.symbolName} (${match.filePath}:${match.lineStart + 1})`);
+    parts.push(`    ${match.signature}`);
+    if (match.jsdoc) {
+      const firstLine = match.jsdoc.split("\n").find(l => l.replace(/^[\s*/]+/, "").trim().length > 0);
+      if (firstLine) {
+        parts.push(`    💬 ${firstLine.replace(/^[\s*/]+/, "").trim()}`);
+      }
+    }
+    parts.push("");
+  }
+
+  // Show folded file views
+  parts.push("── Folded File Views ──");
+  parts.push("");
+  for (const file of result.foldedFiles) {
+    parts.push(formatFoldedView(file));
+    parts.push("");
+  }
+
+  parts.push("── Actions ──");
+  parts.push('  To see full implementation: use smart_unfold with file path and symbol name');
+
+  return parts.join("\n");
+}

src/services/sqlite/PendingMessageStore.ts

@@ -2,6 +2,9 @@ import { Database } from './sqlite-compat.js';
 import type { PendingMessage } from '../worker-types.js';
 import { logger } from '../../utils/logger.js';
 
+/** Messages processing longer than this are considered stale and reset to pending by self-healing */
+const STALE_PROCESSING_THRESHOLD_MS = 60_000;
+
 /**
  * Persistent pending message record from database
  */
@@ -26,12 +29,17 @@ export interface PersistentPendingMessage {
 /**
  * PendingMessageStore - Persistent work queue for SDK messages
  *
- * Messages are persisted before processing using a claim-and-delete pattern.
+ * Messages are persisted before processing using a claim-confirm pattern.
  * This simplifies the lifecycle and eliminates duplicate processing bugs.
  *
  * Lifecycle:
  * 1. enqueue() - Message persisted with status 'pending'
- * 2. claimAndDelete() - Atomically claims and deletes message (process in memory)
+ * 2. claimNextMessage() - Atomically claims next pending message (marks as 'processing')
+ * 3. confirmProcessed() - Deletes message after successful processing
+ *
+ * Self-healing:
+ * - claimNextMessage() resets stale 'processing' messages (>60s) back to 'pending' before claiming
+ * - This eliminates stuck messages from generator crashes without external timers
  *
  * Recovery:
  * - getSessionsWithPendingMessages() - Find sessions that need recovery on startup
@@ -78,13 +86,29 @@ export class PendingMessageStore {
 
   /**
    * Atomically claim the next pending message by marking it as 'processing'.
-   * CRITICAL FIX: Does NOT delete - message stays in DB until confirmProcessed() is called.
-   * This prevents message loss if the generator crashes mid-processing.
+   * Self-healing: resets any stale 'processing' messages (>60s) back to 'pending' first.
+   * Message stays in DB until confirmProcessed() is called.
    * Uses a transaction to prevent race conditions.
    */
-  claimAndDelete(sessionDbId: number): PersistentPendingMessage | null {
-    const now = Date.now();
+  claimNextMessage(sessionDbId: number): PersistentPendingMessage | null {
     const claimTx = this.db.transaction((sessionId: number) => {
+      // Capture time inside transaction so it's fresh if WAL contention causes retry
+      const now = Date.now();
+      // Self-healing: reset stale 'processing' messages back to 'pending'
+      // This recovers from generator crashes without external timers
+      // Note: strict < means messages must be OLDER than threshold to be reset
+      const staleCutoff = now - STALE_PROCESSING_THRESHOLD_MS;
+      const resetStmt = this.db.prepare(`
+        UPDATE pending_messages
+        SET status = 'pending', started_processing_at_epoch = NULL
+        WHERE session_db_id = ? AND status = 'processing'
+          AND started_processing_at_epoch < ?
+      `);
+      const resetResult = resetStmt.run(sessionId, staleCutoff);
+      if (resetResult.changes > 0) {
+        logger.info('QUEUE', `SELF_HEAL | sessionDbId=${sessionId} | recovered ${resetResult.changes} stale processing message(s)`);
+      }
+
       const peekStmt = this.db.prepare(`
         SELECT * FROM pending_messages
         WHERE session_db_id = ? AND status = 'pending'
@@ -374,9 +398,22 @@ export class PendingMessageStore {
   }
 
   /**
-   * Check if any session has pending work
+   * Check if any session has pending work.
+   * Excludes 'processing' messages stuck for >5 minutes (resets them to 'pending' as a side effect).
    */
   hasAnyPendingWork(): boolean {
+    // Reset stuck 'processing' messages older than 5 minutes before checking
+    const stuckCutoff = Date.now() - (5 * 60 * 1000);
+    const resetStmt = this.db.prepare(`
+      UPDATE pending_messages
+      SET status = 'pending', started_processing_at_epoch = NULL
+      WHERE status = 'processing' AND started_processing_at_epoch < ?
+    `);
+    const resetResult = resetStmt.run(stuckCutoff);
+    if (resetResult.changes > 0) {
+      logger.info('QUEUE', `STUCK_RESET | hasAnyPendingWork reset ${resetResult.changes} stuck processing message(s) older than 5 minutes`);
+    }
+
     const stmt = this.db.prepare(`
       SELECT COUNT(*) as count FROM pending_messages
       WHERE status IN ('pending', 'processing')

src/services/sqlite/SessionSearch.ts

@@ -46,6 +46,10 @@ export class SessionSearch {
    * - Tables maintained but search paths removed
    * - Triggers still fire to keep tables synchronized
    *
+   * FTS5 may be unavailable on some platforms (e.g., Bun on Windows #791).
+   * When unavailable, we skip FTS table creation — search falls back to
+   * ChromaDB (vector) and LIKE queries (structured filters) which are unaffected.
+   *
    * TODO: Remove FTS5 infrastructure in future major version (v7.0.0)
    */
   private ensureFTSTables(): void {
@@ -58,91 +62,117 @@ export class SessionSearch {
       return;
     }
 
+    // Runtime check: verify FTS5 is available before attempting to create tables.
+    // bun:sqlite on Windows may not include the FTS5 extension (#791).
+    if (!this.isFts5Available()) {
+      logger.warn('DB', 'FTS5 not available on this platform — skipping FTS table creation (search uses ChromaDB)');
+      return;
+    }
+
     logger.info('DB', 'Creating FTS5 tables');
 
-    // Create observations_fts virtual table
-    this.db.run(`
-      CREATE VIRTUAL TABLE IF NOT EXISTS observations_fts USING fts5(
-        title,
-        subtitle,
-        narrative,
-        text,
-        facts,
-        concepts,
-        content='observations',
-        content_rowid='id'
-      );
-    `);
+    try {
+      // Create observations_fts virtual table
+      this.db.run(`
+        CREATE VIRTUAL TABLE IF NOT EXISTS observations_fts USING fts5(
+          title,
+          subtitle,
+          narrative,
+          text,
+          facts,
+          concepts,
+          content='observations',
+          content_rowid='id'
+        );
+      `);
 
-    // Populate with existing data
-    this.db.run(`
-      INSERT INTO observations_fts(rowid, title, subtitle, narrative, text, facts, concepts)
-      SELECT id, title, subtitle, narrative, text, facts, concepts
-      FROM observations;
-    `);
-
-    // Create triggers for observations
-    this.db.run(`
-      CREATE TRIGGER IF NOT EXISTS observations_ai AFTER INSERT ON observations BEGIN
+      // Populate with existing data
+      this.db.run(`
         INSERT INTO observations_fts(rowid, title, subtitle, narrative, text, facts, concepts)
-        VALUES (new.id, new.title, new.subtitle, new.narrative, new.text, new.facts, new.concepts);
-      END;
+        SELECT id, title, subtitle, narrative, text, facts, concepts
+        FROM observations;
+      `);
 
-      CREATE TRIGGER IF NOT EXISTS observations_ad AFTER DELETE ON observations BEGIN
-        INSERT INTO observations_fts(observations_fts, rowid, title, subtitle, narrative, text, facts, concepts)
-        VALUES('delete', old.id, old.title, old.subtitle, old.narrative, old.text, old.facts, old.concepts);
-      END;
+      // Create triggers for observations
+      this.db.run(`
+        CREATE TRIGGER IF NOT EXISTS observations_ai AFTER INSERT ON observations BEGIN
+          INSERT INTO observations_fts(rowid, title, subtitle, narrative, text, facts, concepts)
+          VALUES (new.id, new.title, new.subtitle, new.narrative, new.text, new.facts, new.concepts);
+        END;
 
-      CREATE TRIGGER IF NOT EXISTS observations_au AFTER UPDATE ON observations BEGIN
-        INSERT INTO observations_fts(observations_fts, rowid, title, subtitle, narrative, text, facts, concepts)
-        VALUES('delete', old.id, old.title, old.subtitle, old.narrative, old.text, old.facts, old.concepts);
-        INSERT INTO observations_fts(rowid, title, subtitle, narrative, text, facts, concepts)
-        VALUES (new.id, new.title, new.subtitle, new.narrative, new.text, new.facts, new.concepts);
-      END;
-    `);
+        CREATE TRIGGER IF NOT EXISTS observations_ad AFTER DELETE ON observations BEGIN
+          INSERT INTO observations_fts(observations_fts, rowid, title, subtitle, narrative, text, facts, concepts)
+          VALUES('delete', old.id, old.title, old.subtitle, old.narrative, old.text, old.facts, old.concepts);
+        END;
 
-    // Create session_summaries_fts virtual table
-    this.db.run(`
-      CREATE VIRTUAL TABLE IF NOT EXISTS session_summaries_fts USING fts5(
-        request,
-        investigated,
-        learned,
-        completed,
-        next_steps,
-        notes,
-        content='session_summaries',
-        content_rowid='id'
-      );
-    `);
+        CREATE TRIGGER IF NOT EXISTS observations_au AFTER UPDATE ON observations BEGIN
+          INSERT INTO observations_fts(observations_fts, rowid, title, subtitle, narrative, text, facts, concepts)
+          VALUES('delete', old.id, old.title, old.subtitle, old.narrative, old.text, old.facts, old.concepts);
+          INSERT INTO observations_fts(rowid, title, subtitle, narrative, text, facts, concepts)
+          VALUES (new.id, new.title, new.subtitle, new.narrative, new.text, new.facts, new.concepts);
+        END;
+      `);
 
-    // Populate with existing data
-    this.db.run(`
-      INSERT INTO session_summaries_fts(rowid, request, investigated, learned, completed, next_steps, notes)
-      SELECT id, request, investigated, learned, completed, next_steps, notes
-      FROM session_summaries;
-    `);
+      // Create session_summaries_fts virtual table
+      this.db.run(`
+        CREATE VIRTUAL TABLE IF NOT EXISTS session_summaries_fts USING fts5(
+          request,
+          investigated,
+          learned,
+          completed,
+          next_steps,
+          notes,
+          content='session_summaries',
+          content_rowid='id'
+        );
+      `);
 
-    // Create triggers for session_summaries
-    this.db.run(`
-      CREATE TRIGGER IF NOT EXISTS session_summaries_ai AFTER INSERT ON session_summaries BEGIN
+      // Populate with existing data
+      this.db.run(`
         INSERT INTO session_summaries_fts(rowid, request, investigated, learned, completed, next_steps, notes)
-        VALUES (new.id, new.request, new.investigated, new.learned, new.completed, new.next_steps, new.notes);
-      END;
+        SELECT id, request, investigated, learned, completed, next_steps, notes
+        FROM session_summaries;
+      `);
 
-      CREATE TRIGGER IF NOT EXISTS session_summaries_ad AFTER DELETE ON session_summaries BEGIN
-        INSERT INTO session_summaries_fts(session_summaries_fts, rowid, request, investigated, learned, completed, next_steps, notes)
-        VALUES('delete', old.id, old.request, old.investigated, old.learned, old.completed, old.next_steps, old.notes);
-      END;
+      // Create triggers for session_summaries
+      this.db.run(`
+        CREATE TRIGGER IF NOT EXISTS session_summaries_ai AFTER INSERT ON session_summaries BEGIN
+          INSERT INTO session_summaries_fts(rowid, request, investigated, learned, completed, next_steps, notes)
+          VALUES (new.id, new.request, new.investigated, new.learned, new.completed, new.next_steps, new.notes);
+        END;
 
-      CREATE TRIGGER IF NOT EXISTS session_summaries_au AFTER UPDATE ON session_summaries BEGIN
-        INSERT INTO session_summaries_fts(session_summaries_fts, rowid, request, investigated, learned, completed, next_steps, notes)
-        VALUES('delete', old.id, old.request, old.investigated, old.learned, old.completed, old.next_steps, old.notes);
-        INSERT INTO session_summaries_fts(rowid, request, investigated, learned, completed, next_steps, notes)
-        VALUES (new.id, new.request, new.investigated, new.learned, new.completed, new.next_steps, new.notes);
-      END;
-    `);
+        CREATE TRIGGER IF NOT EXISTS session_summaries_ad AFTER DELETE ON session_summaries BEGIN
+          INSERT INTO session_summaries_fts(session_summaries_fts, rowid, request, investigated, learned, completed, next_steps, notes)
+          VALUES('delete', old.id, old.request, old.investigated, old.learned, old.completed, old.next_steps, old.notes);
+        END;
 
-    logger.info('DB', 'FTS5 tables created successfully');
+        CREATE TRIGGER IF NOT EXISTS session_summaries_au AFTER UPDATE ON session_summaries BEGIN
+          INSERT INTO session_summaries_fts(session_summaries_fts, rowid, request, investigated, learned, completed, next_steps, notes)
+          VALUES('delete', old.id, old.request, old.investigated, old.learned, old.completed, old.next_steps, old.notes);
+          INSERT INTO session_summaries_fts(rowid, request, investigated, learned, completed, next_steps, notes)
+          VALUES (new.id, new.request, new.investigated, new.learned, new.completed, new.next_steps, new.notes);
+        END;
+      `);
+
+      logger.info('DB', 'FTS5 tables created successfully');
+    } catch (error) {
+      // FTS5 creation failed at runtime despite probe succeeding — degrade gracefully
+      logger.warn('DB', 'FTS5 table creation failed — search will use ChromaDB and LIKE queries', {}, error as Error);
+    }
+  }
+
+  /**
+   * Probe whether the FTS5 extension is available in the current SQLite build.
+   * Creates and immediately drops a temporary FTS5 table.
+   */
+  private isFts5Available(): boolean {
+    try {
+      this.db.run('CREATE VIRTUAL TABLE _fts5_probe USING fts5(test_column)');
+      this.db.run('DROP TABLE _fts5_probe');
+      return true;
+    } catch {
+      return false;
+    }
   }
 
 

src/services/sqlite/SessionStore.ts

@@ -13,6 +13,7 @@ import {
   LatestPromptResult
 } from '../../types/database.js';
 import type { PendingMessageStore } from './PendingMessageStore.js';
+import { computeObservationContentHash, findDuplicateObservation } from './observations/store.js';
 
 /**
  * Session data store for SDK sessions, observations, and summaries
@@ -48,11 +49,17 @@ export class SessionStore {
     this.repairSessionIdColumnRename();
     this.addFailedAtEpochColumn();
     this.addOnUpdateCascadeToForeignKeys();
+    this.addObservationContentHashColumn();
+    this.addSessionCustomTitleColumn();
   }
 
   /**
-   * Initialize database schema using migrations (migration004)
-   * This runs the core SDK tables migration if no tables exist
+   * Initialize database schema (migration004)
+   *
+   * ALWAYS creates core tables using CREATE TABLE IF NOT EXISTS — safe to run
+   * regardless of schema_versions state.  This fixes issue #979 where the old
+   * DatabaseManager migration system (versions 1-7) shared the schema_versions
+   * table, causing maxApplied > 0 and skipping core table creation entirely.
    */
   private initializeSchema(): void {
     // Create schema_versions table if it doesn't exist
@@ -64,90 +71,77 @@ export class SessionStore {
       )
     `);
 
-    // Get applied migrations
-    const appliedVersions = this.db.prepare('SELECT version FROM schema_versions ORDER BY version').all() as SchemaVersion[];
-    const maxApplied = appliedVersions.length > 0 ? Math.max(...appliedVersions.map(v => v.version)) : 0;
+    // Always create core tables — IF NOT EXISTS makes this idempotent
+    this.db.run(`
+      CREATE TABLE IF NOT EXISTS sdk_sessions (
+        id INTEGER PRIMARY KEY AUTOINCREMENT,
+        content_session_id TEXT UNIQUE NOT NULL,
+        memory_session_id TEXT UNIQUE,
+        project TEXT NOT NULL,
+        user_prompt TEXT,
+        started_at TEXT NOT NULL,
+        started_at_epoch INTEGER NOT NULL,
+        completed_at TEXT,
+        completed_at_epoch INTEGER,
+        status TEXT CHECK(status IN ('active', 'completed', 'failed')) NOT NULL DEFAULT 'active'
+      );
 
-    // Only run migration004 if no migrations have been applied
-    // This creates the sdk_sessions, observations, and session_summaries tables
-    if (maxApplied === 0) {
-      logger.info('DB', 'Initializing fresh database with migration004');
+      CREATE INDEX IF NOT EXISTS idx_sdk_sessions_claude_id ON sdk_sessions(content_session_id);
+      CREATE INDEX IF NOT EXISTS idx_sdk_sessions_sdk_id ON sdk_sessions(memory_session_id);
+      CREATE INDEX IF NOT EXISTS idx_sdk_sessions_project ON sdk_sessions(project);
+      CREATE INDEX IF NOT EXISTS idx_sdk_sessions_status ON sdk_sessions(status);
+      CREATE INDEX IF NOT EXISTS idx_sdk_sessions_started ON sdk_sessions(started_at_epoch DESC);
 
-      // Migration004: SDK agent architecture tables
-      this.db.run(`
-        CREATE TABLE IF NOT EXISTS sdk_sessions (
-          id INTEGER PRIMARY KEY AUTOINCREMENT,
-          content_session_id TEXT UNIQUE NOT NULL,
-          memory_session_id TEXT UNIQUE,
-          project TEXT NOT NULL,
-          user_prompt TEXT,
-          started_at TEXT NOT NULL,
-          started_at_epoch INTEGER NOT NULL,
-          completed_at TEXT,
-          completed_at_epoch INTEGER,
-          status TEXT CHECK(status IN ('active', 'completed', 'failed')) NOT NULL DEFAULT 'active'
-        );
+      CREATE TABLE IF NOT EXISTS observations (
+        id INTEGER PRIMARY KEY AUTOINCREMENT,
+        memory_session_id TEXT NOT NULL,
+        project TEXT NOT NULL,
+        text TEXT NOT NULL,
+        type TEXT NOT NULL,
+        created_at TEXT NOT NULL,
+        created_at_epoch INTEGER NOT NULL,
+        FOREIGN KEY(memory_session_id) REFERENCES sdk_sessions(memory_session_id) ON DELETE CASCADE ON UPDATE CASCADE
+      );
 
-        CREATE INDEX IF NOT EXISTS idx_sdk_sessions_claude_id ON sdk_sessions(content_session_id);
-        CREATE INDEX IF NOT EXISTS idx_sdk_sessions_sdk_id ON sdk_sessions(memory_session_id);
-        CREATE INDEX IF NOT EXISTS idx_sdk_sessions_project ON sdk_sessions(project);
-        CREATE INDEX IF NOT EXISTS idx_sdk_sessions_status ON sdk_sessions(status);
-        CREATE INDEX IF NOT EXISTS idx_sdk_sessions_started ON sdk_sessions(started_at_epoch DESC);
+      CREATE INDEX IF NOT EXISTS idx_observations_sdk_session ON observations(memory_session_id);
+      CREATE INDEX IF NOT EXISTS idx_observations_project ON observations(project);
+      CREATE INDEX IF NOT EXISTS idx_observations_type ON observations(type);
+      CREATE INDEX IF NOT EXISTS idx_observations_created ON observations(created_at_epoch DESC);
 
-        CREATE TABLE IF NOT EXISTS observations (
-          id INTEGER PRIMARY KEY AUTOINCREMENT,
-          memory_session_id TEXT NOT NULL,
-          project TEXT NOT NULL,
-          text TEXT NOT NULL,
-          type TEXT NOT NULL,
-          created_at TEXT NOT NULL,
-          created_at_epoch INTEGER NOT NULL,
-          FOREIGN KEY(memory_session_id) REFERENCES sdk_sessions(memory_session_id) ON DELETE CASCADE ON UPDATE CASCADE
-        );
+      CREATE TABLE IF NOT EXISTS session_summaries (
+        id INTEGER PRIMARY KEY AUTOINCREMENT,
+        memory_session_id TEXT UNIQUE NOT NULL,
+        project TEXT NOT NULL,
+        request TEXT,
+        investigated TEXT,
+        learned TEXT,
+        completed TEXT,
+        next_steps TEXT,
+        files_read TEXT,
+        files_edited TEXT,
+        notes TEXT,
+        created_at TEXT NOT NULL,
+        created_at_epoch INTEGER NOT NULL,
+        FOREIGN KEY(memory_session_id) REFERENCES sdk_sessions(memory_session_id) ON DELETE CASCADE ON UPDATE CASCADE
+      );
 
-        CREATE INDEX IF NOT EXISTS idx_observations_sdk_session ON observations(memory_session_id);
-        CREATE INDEX IF NOT EXISTS idx_observations_project ON observations(project);
-        CREATE INDEX IF NOT EXISTS idx_observations_type ON observations(type);
-        CREATE INDEX IF NOT EXISTS idx_observations_created ON observations(created_at_epoch DESC);
+      CREATE INDEX IF NOT EXISTS idx_session_summaries_sdk_session ON session_summaries(memory_session_id);
+      CREATE INDEX IF NOT EXISTS idx_session_summaries_project ON session_summaries(project);
+      CREATE INDEX IF NOT EXISTS idx_session_summaries_created ON session_summaries(created_at_epoch DESC);
+    `);
 
-        CREATE TABLE IF NOT EXISTS session_summaries (
-          id INTEGER PRIMARY KEY AUTOINCREMENT,
-          memory_session_id TEXT UNIQUE NOT NULL,
-          project TEXT NOT NULL,
-          request TEXT,
-          investigated TEXT,
-          learned TEXT,
-          completed TEXT,
-          next_steps TEXT,
-          files_read TEXT,
-          files_edited TEXT,
-          notes TEXT,
-          created_at TEXT NOT NULL,
-          created_at_epoch INTEGER NOT NULL,
-          FOREIGN KEY(memory_session_id) REFERENCES sdk_sessions(memory_session_id) ON DELETE CASCADE ON UPDATE CASCADE
-        );
-
-        CREATE INDEX IF NOT EXISTS idx_session_summaries_sdk_session ON session_summaries(memory_session_id);
-        CREATE INDEX IF NOT EXISTS idx_session_summaries_project ON session_summaries(project);
-        CREATE INDEX IF NOT EXISTS idx_session_summaries_created ON session_summaries(created_at_epoch DESC);
-      `);
-
-      // Record migration004 as applied
-      this.db.prepare('INSERT INTO schema_versions (version, applied_at) VALUES (?, ?)').run(4, new Date().toISOString());
-
-      logger.info('DB', 'Migration004 applied successfully');
-    }
+    // Record migration004 as applied (OR IGNORE handles re-runs safely)
+    this.db.prepare('INSERT OR IGNORE INTO schema_versions (version, applied_at) VALUES (?, ?)').run(4, new Date().toISOString());
   }
 
   /**
    * Ensure worker_port column exists (migration 5)
+   *
+   * NOTE: Version 5 conflicts with old DatabaseManager migration005 (which drops orphaned tables).
+   * We check actual column state rather than relying solely on version tracking.
    */
   private ensureWorkerPortColumn(): void {
-    // Check if migration already applied
-    const applied = this.db.prepare('SELECT version FROM schema_versions WHERE version = ?').get(5) as SchemaVersion | undefined;
-    if (applied) return;
-
-    // Check if column exists
+    // Check actual column existence — don't rely on version tracking alone (issue #979)
     const tableInfo = this.db.query('PRAGMA table_info(sdk_sessions)').all() as TableColumnInfo[];
     const hasWorkerPort = tableInfo.some(col => col.name === 'worker_port');
 
@@ -162,12 +156,12 @@ export class SessionStore {
 
   /**
    * Ensure prompt tracking columns exist (migration 6)
+   *
+   * NOTE: Version 6 conflicts with old DatabaseManager migration006 (which creates FTS5 tables).
+   * We check actual column state rather than relying solely on version tracking.
    */
   private ensurePromptTrackingColumns(): void {
-    // Check if migration already applied
-    const applied = this.db.prepare('SELECT version FROM schema_versions WHERE version = ?').get(6) as SchemaVersion | undefined;
-    if (applied) return;
-
+    // Check actual column existence — don't rely on version tracking alone (issue #979)
     // Check sdk_sessions for prompt_counter
     const sessionsInfo = this.db.query('PRAGMA table_info(sdk_sessions)').all() as TableColumnInfo[];
     const hasPromptCounter = sessionsInfo.some(col => col.name === 'prompt_counter');
@@ -201,13 +195,12 @@ export class SessionStore {
 
   /**
    * Remove UNIQUE constraint from session_summaries.memory_session_id (migration 7)
+   *
+   * NOTE: Version 7 conflicts with old DatabaseManager migration007 (which adds discovery_tokens).
+   * We check actual constraint state rather than relying solely on version tracking.
    */
   private removeSessionSummariesUniqueConstraint(): void {
-    // Check if migration already applied
-    const applied = this.db.prepare('SELECT version FROM schema_versions WHERE version = ?').get(7) as SchemaVersion | undefined;
-    if (applied) return;
-
-    // Check if UNIQUE constraint exists
+    // Check actual constraint state — don't rely on version tracking alone (issue #979)
     const summariesIndexes = this.db.query('PRAGMA index_list(session_summaries)').all() as IndexInfo[];
     const hasUniqueConstraint = summariesIndexes.some(idx => idx.unique === 1);
 
@@ -222,6 +215,9 @@ export class SessionStore {
     // Begin transaction
     this.db.run('BEGIN TRANSACTION');
 
+    // Clean up leftover temp table from a previously-crashed run
+    this.db.run('DROP TABLE IF EXISTS session_summaries_new');
+
     // Create new table without UNIQUE constraint
     this.db.run(`
       CREATE TABLE session_summaries_new (
@@ -335,6 +331,9 @@ export class SessionStore {
     // Begin transaction
     this.db.run('BEGIN TRANSACTION');
 
+    // Clean up leftover temp table from a previously-crashed run
+    this.db.run('DROP TABLE IF EXISTS observations_new');
+
     // Create new table with text as nullable
     this.db.run(`
       CREATE TABLE observations_new (
@@ -428,34 +427,39 @@ export class SessionStore {
       CREATE INDEX idx_user_prompts_lookup ON user_prompts(content_session_id, prompt_number);
     `);
 
-    // Create FTS5 virtual table
-    this.db.run(`
-      CREATE VIRTUAL TABLE user_prompts_fts USING fts5(
-        prompt_text,
-        content='user_prompts',
-        content_rowid='id'
-      );
-    `);
+    // Create FTS5 virtual table — skip if FTS5 is unavailable (e.g., Bun on Windows #791).
+    // The user_prompts table itself is still created; only FTS indexing is skipped.
+    try {
+      this.db.run(`
+        CREATE VIRTUAL TABLE user_prompts_fts USING fts5(
+          prompt_text,
+          content='user_prompts',
+          content_rowid='id'
+        );
+      `);
 
-    // Create triggers to sync FTS5
-    this.db.run(`
-      CREATE TRIGGER user_prompts_ai AFTER INSERT ON user_prompts BEGIN
-        INSERT INTO user_prompts_fts(rowid, prompt_text)
-        VALUES (new.id, new.prompt_text);
-      END;
+      // Create triggers to sync FTS5
+      this.db.run(`
+        CREATE TRIGGER user_prompts_ai AFTER INSERT ON user_prompts BEGIN
+          INSERT INTO user_prompts_fts(rowid, prompt_text)
+          VALUES (new.id, new.prompt_text);
+        END;
 
-      CREATE TRIGGER user_prompts_ad AFTER DELETE ON user_prompts BEGIN
-        INSERT INTO user_prompts_fts(user_prompts_fts, rowid, prompt_text)
-        VALUES('delete', old.id, old.prompt_text);
-      END;
+        CREATE TRIGGER user_prompts_ad AFTER DELETE ON user_prompts BEGIN
+          INSERT INTO user_prompts_fts(user_prompts_fts, rowid, prompt_text)
+          VALUES('delete', old.id, old.prompt_text);
+        END;
 
-      CREATE TRIGGER user_prompts_au AFTER UPDATE ON user_prompts BEGIN
-        INSERT INTO user_prompts_fts(user_prompts_fts, rowid, prompt_text)
-        VALUES('delete', old.id, old.prompt_text);
-        INSERT INTO user_prompts_fts(rowid, prompt_text)
-        VALUES (new.id, new.prompt_text);
-      END;
-    `);
+        CREATE TRIGGER user_prompts_au AFTER UPDATE ON user_prompts BEGIN
+          INSERT INTO user_prompts_fts(user_prompts_fts, rowid, prompt_text)
+          VALUES('delete', old.id, old.prompt_text);
+          INSERT INTO user_prompts_fts(rowid, prompt_text)
+          VALUES (new.id, new.prompt_text);
+        END;
+      `);
+    } catch (ftsError) {
+      logger.warn('DB', 'FTS5 not available — user_prompts_fts skipped (search uses ChromaDB)', {}, ftsError as Error);
+    }
 
     // Commit transaction
     this.db.run('COMMIT');
@@ -463,7 +467,7 @@ export class SessionStore {
     // Record migration
     this.db.prepare('INSERT OR IGNORE INTO schema_versions (version, applied_at) VALUES (?, ?)').run(10, new Date().toISOString());
 
-    logger.debug('DB', 'Successfully created user_prompts table with FTS5 support');
+    logger.debug('DB', 'Successfully created user_prompts table');
   }
 
   /**
@@ -675,6 +679,9 @@ export class SessionStore {
       this.db.run('DROP TRIGGER IF EXISTS observations_ad');
       this.db.run('DROP TRIGGER IF EXISTS observations_au');
 
+      // Clean up leftover temp table from a previously-crashed run
+      this.db.run('DROP TABLE IF EXISTS observations_new');
+
       this.db.run(`
         CREATE TABLE observations_new (
           id INTEGER PRIMARY KEY AUTOINCREMENT,
@@ -744,6 +751,9 @@ export class SessionStore {
       // 2. Recreate session_summaries table
       // ==========================================
 
+      // Clean up leftover temp table from a previously-crashed run
+      this.db.run('DROP TABLE IF EXISTS session_summaries_new');
+
       this.db.run(`
         CREATE TABLE session_summaries_new (
           id INTEGER PRIMARY KEY AUTOINCREMENT,
@@ -825,6 +835,44 @@ export class SessionStore {
     }
   }
 
+  /**
+   * Add content_hash column to observations for deduplication (migration 22)
+   */
+  private addObservationContentHashColumn(): void {
+    const applied = this.db.prepare('SELECT version FROM schema_versions WHERE version = ?').get(22) as SchemaVersion | undefined;
+    if (applied) return;
+
+    const tableInfo = this.db.query('PRAGMA table_info(observations)').all() as TableColumnInfo[];
+    const hasColumn = tableInfo.some(col => col.name === 'content_hash');
+
+    if (!hasColumn) {
+      this.db.run('ALTER TABLE observations ADD COLUMN content_hash TEXT');
+      this.db.run("UPDATE observations SET content_hash = substr(hex(randomblob(8)), 1, 16) WHERE content_hash IS NULL");
+      this.db.run('CREATE INDEX IF NOT EXISTS idx_observations_content_hash ON observations(content_hash, created_at_epoch)');
+      logger.debug('DB', 'Added content_hash column to observations table with backfill and index');
+    }
+
+    this.db.prepare('INSERT OR IGNORE INTO schema_versions (version, applied_at) VALUES (?, ?)').run(22, new Date().toISOString());
+  }
+
+  /**
+   * Add custom_title column to sdk_sessions for agent attribution (migration 23)
+   */
+  private addSessionCustomTitleColumn(): void {
+    const applied = this.db.prepare('SELECT version FROM schema_versions WHERE version = ?').get(23) as SchemaVersion | undefined;
+    if (applied) return;
+
+    const tableInfo = this.db.query('PRAGMA table_info(sdk_sessions)').all() as TableColumnInfo[];
+    const hasColumn = tableInfo.some(col => col.name === 'custom_title');
+
+    if (!hasColumn) {
+      this.db.run('ALTER TABLE sdk_sessions ADD COLUMN custom_title TEXT');
+      logger.debug('DB', 'Added custom_title column to sdk_sessions table');
+    }
+
+    this.db.prepare('INSERT OR IGNORE INTO schema_versions (version, applied_at) VALUES (?, ?)').run(23, new Date().toISOString());
+  }
+
   /**
    * Update the memory session ID for a session
    * Called by SDKAgent when it captures the session ID from the first SDK message
@@ -1290,9 +1338,10 @@ export class SessionStore {
     memory_session_id: string | null;
     project: string;
     user_prompt: string;
+    custom_title: string | null;
   } | null {
     const stmt = this.db.prepare(`
-      SELECT id, content_session_id, memory_session_id, project, user_prompt
+      SELECT id, content_session_id, memory_session_id, project, user_prompt, custom_title
       FROM sdk_sessions
       WHERE id = ?
       LIMIT 1
@@ -1311,6 +1360,7 @@ export class SessionStore {
     memory_session_id: string;
     project: string;
     user_prompt: string;
+    custom_title: string | null;
     started_at: string;
     started_at_epoch: number;
     completed_at: string | null;
@@ -1321,7 +1371,7 @@ export class SessionStore {
 
     const placeholders = memorySessionIds.map(() => '?').join(',');
     const stmt = this.db.prepare(`
-      SELECT id, content_session_id, memory_session_id, project, user_prompt,
+      SELECT id, content_session_id, memory_session_id, project, user_prompt, custom_title,
              started_at, started_at_epoch, completed_at, completed_at_epoch, status
       FROM sdk_sessions
       WHERE memory_session_id IN (${placeholders})
@@ -1366,7 +1416,7 @@ export class SessionStore {
    * Pure get-or-create: never modifies memory_session_id.
    * Multi-terminal isolation is handled by ON UPDATE CASCADE at the schema level.
    */
-  createSDKSession(contentSessionId: string, project: string, userPrompt: string): number {
+  createSDKSession(contentSessionId: string, project: string, userPrompt: string, customTitle?: string): number {
     const now = new Date();
     const nowEpoch = now.getTime();
 
@@ -1383,6 +1433,13 @@ export class SessionStore {
           WHERE content_session_id = ? AND (project IS NULL OR project = '')
         `).run(project, contentSessionId);
       }
+      // Backfill custom_title if provided and not yet set
+      if (customTitle) {
+        this.db.prepare(`
+          UPDATE sdk_sessions SET custom_title = ?
+          WHERE content_session_id = ? AND custom_title IS NULL
+        `).run(customTitle, contentSessionId);
+      }
       return existing.id;
     }
 
@@ -1392,9 +1449,9 @@ export class SessionStore {
     // must NEVER equal contentSessionId - that would inject memory messages into the user's transcript!
     this.db.prepare(`
       INSERT INTO sdk_sessions
-      (content_session_id, memory_session_id, project, user_prompt, started_at, started_at_epoch, status)
-      VALUES (?, NULL, ?, ?, ?, ?, 'active')
-    `).run(contentSessionId, project, userPrompt, now.toISOString(), nowEpoch);
+      (content_session_id, memory_session_id, project, user_prompt, custom_title, started_at, started_at_epoch, status)
+      VALUES (?, NULL, ?, ?, ?, ?, ?, 'active')
+    `).run(contentSessionId, project, userPrompt, customTitle || null, now.toISOString(), nowEpoch);
 
     // Return new ID
     const row = this.db.prepare('SELECT id FROM sdk_sessions WHERE content_session_id = ?')
@@ -1441,6 +1498,7 @@ export class SessionStore {
   /**
    * Store an observation (from SDK parsing)
    * Assumes session already exists (created by hook)
+   * Performs content-hash deduplication: skips INSERT if an identical observation exists within 30s
    */
   storeObservation(
     memorySessionId: string,
@@ -1463,11 +1521,18 @@ export class SessionStore {
     const timestampEpoch = overrideTimestampEpoch ?? Date.now();
     const timestampIso = new Date(timestampEpoch).toISOString();
 
+    // Content-hash deduplication
+    const contentHash = computeObservationContentHash(memorySessionId, observation.title, observation.narrative);
+    const existing = findDuplicateObservation(this.db, contentHash, timestampEpoch);
+    if (existing) {
+      return { id: existing.id, createdAtEpoch: existing.created_at_epoch };
+    }
+
     const stmt = this.db.prepare(`
       INSERT INTO observations
       (memory_session_id, project, type, title, subtitle, facts, narrative, concepts,
-       files_read, files_modified, prompt_number, discovery_tokens, created_at, created_at_epoch)
-      VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+       files_read, files_modified, prompt_number, discovery_tokens, content_hash, created_at, created_at_epoch)
+      VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
     `);
 
     const result = stmt.run(
@@ -1483,6 +1548,7 @@ export class SessionStore {
       JSON.stringify(observation.files_modified),
       promptNumber || null,
       discoveryTokens,
+      contentHash,
       timestampIso,
       timestampEpoch
     );

src/services/sqlite/migrations.ts

@@ -372,6 +372,16 @@ export const migration005: Migration = {
 export const migration006: Migration = {
   version: 6,
   up: (db: Database) => {
+    // FTS5 may be unavailable on some platforms (e.g., Bun on Windows #791).
+    // Probe before creating tables — search falls back to ChromaDB when unavailable.
+    try {
+      db.run('CREATE VIRTUAL TABLE _fts5_probe USING fts5(test_column)');
+      db.run('DROP TABLE _fts5_probe');
+    } catch {
+      console.log('⚠️  FTS5 not available on this platform — skipping FTS migration (search uses ChromaDB)');
+      return;
+    }
+
     // FTS5 virtual table for observations
     // Note: This assumes the hierarchical fields (title, subtitle, etc.) already exist
     // from the inline migrations in SessionStore constructor