lm-assist 0.1.65 → 0.1.67

package/.claude-plugin/marketplace.json

@@ -12,7 +12,7 @@
   "plugins": [
     {
       "name": "lm-assist",
-      "version": "0.1.65",
+      "version": "0.1.67",
       "source": "./",
       "description": "Auto-build knowledge from Claude Code sessions and inject it as context. Inspect sessions with deep views into agents, tasks, teams, plans, and tool calls. Access and control all Claude Code terminals from any device via web browser. Includes MCP server (search, detail, feedback), context-injection hook, and 6 slash commands.",
       "category": "productivity",

package/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "lm-assist",
-  "version": "0.1.65",
+  "version": "0.1.67",
   "description": "Knowledge management, session inspector, and web terminal control for Claude Code",
   "author": {
     "name": "langmartai"

package/README.md

@@ -1,59 +1,83 @@
 # lm-assist
 
-Knowledge management, session inspector, and web terminal control for [Claude Code](https://docs.anthropic.com/en/docs/claude-code). Auto-build knowledge from your sessions and inject it as context. Inspect agents, tasks, teams, plans, and tool calls. Access and control all Claude Code terminals from any device via browser.
+## The Observability Platform for Claude Code & Agent SDK
+
+Monitor, debug, and control AI coding agents with full session visibility, real-time execution tracking, and 155+ REST API endpoints.
 
 [![Discord](https://img.shields.io/discord/1475647234669543558?logo=discord&label=Discord&color=5865F2)](https://discord.gg/xb2BNnk4)
 
-### Install
+- **Monitor** — real-time execution tracking, per-model cost & token breakdown, SSE event stream
+- **Debug** — 15 insight views per session: Chat, Thinking, Agents, Plans, Team, DAG, Files, Git & more
+- **Control** — web terminal from any browser, start/abort agents via API, remote access from anywhere
 
-In Claude Code, run:
+<a href="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/architecture-observability.svg"><img src="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/architecture-observability.svg" alt="lm-assist Architecture — Observability Platform for Claude Code & Agent SDK" width="700"></a>
+
+### Install
 
 ```
 /plugin marketplace add langmartai/lm-assist
-
 /plugin install lm-assist@langmartai
 ```
 
-Then **open a new Claude Code session** and run:
+Then **open a new Claude Code session** and run `/assist-setup`.
 
-```
-/assist-setup
-```
+> **Read:** [Inside Claude Code: The Session File Format and How to Inspect It](https://databunny.medium.com/inside-claude-code-the-session-file-format-and-how-to-inspect-it-b9998e66d56b) — technical breakdown of the JSONL session format, message types, subagent trees, and how lm-assist surfaces it all.
+
+---
 
-This automatically registers:
-- **MCP server** — `search`, `detail`, `feedback` tools available in Claude Code
-- **Context hook** — injects relevant knowledge into each prompt
-- **Slash commands** — 6 commands for managing lm-assist
+## Why lm-assist
 
-### How It Works
+Claude Code and the Agent SDK have no built-in dashboard. You get a terminal or logs. When you're running multiple agents, debugging a failed execution, or tracking costs across a fleet of machines — you need full visibility into every session, every subagent, every tool call, every token spent.
 
-<a href="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/architecture-diagram.svg"><img src="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/architecture-diagram.svg" alt="How lm-assist Works — architecture diagram" width="700"></a>
+| Without lm-assist | With lm-assist |
+|-------------------|---------------|
+| Scroll through terminal output | 15 specialized views per session |
+| No cost visibility | Per-model token & cost breakdown |
+| Can't see what agents are doing | Real-time execution dashboard |
+| No way to inspect subagent trees | Full DAG visualization |
+| Terminal-only access | Web UI from any device, anywhere |
+| Agent SDK runs are black boxes | Same session inspection as CLI |
 
-> *Left: Any MCP-compatible IDE (Claude Code, VS Code, Cursor, Codex CLI, Gemini CLI, Antigravity) accesses the knowledge base via MCP. Right: You access the Web UI from localhost, LAN, or langmart.ai for terminal management, session inspection, and more. [Full details →](docs/how-it-works.md)*
+---
 
-> **Read:** [Your Claude Sessions Are Gold: Stop Paying Twice for the Same Knowledge](https://databunny.medium.com/your-claude-sessions-are-gold-stop-paying-twice-for-the-same-knowledge-7632ac6ddb88) — deep dive into session knowledge reuse, CLAUDE.md vs context injection, and token cost savings.
+## Three Pillars
 
-> **Read:** [Inside Claude Code: The Session File Format and How to Inspect It](https://databunny.medium.com/inside-claude-code-the-session-file-format-and-how-to-inspect-it-b9998e66d56b) — technical breakdown of the JSONL session format, message types, subagent trees, and how lm-assist surfaces it all.
+### 1. Monitor
 
----
+Real-time execution dashboard with live session tracking, cost analytics, and multi-machine fleet management.
 
-## Three Core Features
+<a href="https://langmart.ai/images/assist/session-browser.png"><img src="https://langmart.ai/images/assist/session-browser.png" alt="Session Browser — live sessions with cost and token tracking" width="700"></a>
 
-### 1. Access Your Sessions From Anywhere
+> *Browse all sessions with human-readable names, live status, per-model cost breakdown — from any browser, anywhere*
 
-lm-assist runs a web server on your local network. Open any browser on any device — laptop, tablet, phone — and browse all your Claude Code sessions in real time.
+- Session list with human-readable slug names, live status, and running process detection
+- Per-session and per-project cost tracking — total cost in gold at a glance
+- Per-model token breakdown (input, output, cache read, cache creation) in Meta tab
+- SSE event stream for real-time updates
+- Multi-machine fleet dashboard via LangMart Hub
 
-<a href="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshots/session-browsing.gif"><img src="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshots/session-browsing.gif" alt="Session Browsing" width="700"></a>
+**Statusline** — optional status bar showing context %, rate limits (5h/7d usage with time remaining), session cost, RAM, PID, and uptime. Color-coded: green < 50%, yellow 50-80%, red > 80%.
 
-> *Terminal dashboard with 4 live sessions via langmart.ai, then browse into session detail with rich chat history — accessible from any browser, anywhere*
+```
+wt: no worktrees  ctx:42%  $12.34  5h:23% 2h14m left  7d:41%  ram:565M  free:6.1G  pid:12345  up:3h22m
+```
 
-### 2. Deep Insight Views
+**Key endpoints:** `GET /monitor/executions` · `GET /stream` · `GET /sessions` · `GET /projects/sessions`
 
-Every session gets a full breakdown across **15 specialized tabs** — Chat, Thinking, Agents, Skills, Commands, Tasks, Plans, Team, Files, Git, Console, Summary, Meta, JSON, and DB.
+### 2. Debug
 
-<a href="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshots/deep-insight-views.gif"><img src="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshots/deep-insight-views.gif" alt="Deep Insight Views" width="700"></a>
+15 specialized views per session. Trace any decision through conversation flow, extended thinking, subagent hierarchy, tool calls, file changes, and git operations.
 
-> *Click through Chat, Agents, Plans, Files, Thinking, and Git tabs — each surfaces a different dimension of the session*
+<table>
+  <tr>
+    <td><a href="https://langmart.ai/images/assist/agent-tree.png"><img src="https://langmart.ai/images/assist/agent-tree.png" alt="Agent Tree" width="340"></a><br><sub>Subagent hierarchy</sub></td>
+    <td><a href="https://langmart.ai/images/assist/plan-view.png"><img src="https://langmart.ai/images/assist/plan-view.png" alt="Plan View" width="340"></a><br><sub>Plan mode tracking</sub></td>
+  </tr>
+  <tr>
+    <td><a href="https://langmart.ai/images/assist/task-kanban.png"><img src="https://langmart.ai/images/assist/task-kanban.png" alt="Task Kanban" width="340"></a><br><sub>Task kanban board</sub></td>
+    <td><a href="https://langmart.ai/images/assist/team-view.png"><img src="https://langmart.ai/images/assist/team-view.png" alt="Team View" width="340"></a><br><sub>Multi-agent team coordination</sub></td>
+  </tr>
+</table>
 
 <details>
 <summary><strong>All 15 tabs at a glance</strong></summary>
@@ -64,72 +88,83 @@ Every session gets a full breakdown across **15 specialized tabs** — Chat, Thi
 | **Thinking** | Claude's extended thinking / chain-of-thought |
 | **Agents** | Subagent tree — Explore, Plan, Bash, and custom agents |
 | **Skills** | Skill invocation timeline with chain flow, span attribution, and deep trace |
-| **Commands** | Slash command invocations (e.g., /trade-analyze) with args and timing |
-| **Tasks** | Todo lists created during the session |
+| **Commands** | Slash command invocations with args and timing |
+| **Tasks** | Task lists with dependency tracking |
 | **Plans** | Plan mode entries with approval status |
-| **Team** | Team/swarm coordination (Opus 4.6 multi-agent) |
+| **Team** | Team/swarm coordination (multi-agent teams) |
 | **Files** | All files read, written, or edited during the session |
 | **Git** | Commits, pushes, and diffs from the session |
 | **Console** | Terminal output and process management |
 | **Summary** | AI-generated session summary |
-| **Meta** | Session metadata — timing, model, token usage |
+| **Meta** | Session metadata — slug, timing, model, token usage |
 | **JSON** | Raw session JSONL data |
 | **DB** | Internal cache and index data |
 
 </details>
 
-### 3. Auto-Built Knowledge Base
+- Session DAG visualization (message graph + cross-session subagent/team graph)
+- Fork tracking and branch visualization
+- Tool call traces with full inputs and results
+
+**Key endpoints:** `GET /sessions/:id` · `GET /sessions/:id/dag` · `GET /sessions/:id/subagents` · `GET /sessions/:id/conversation`
 
-lm-assist automatically generates knowledge from your Claude Code sessions, then injects it back into future prompts — giving Claude Code memory of what you've worked on before.
+### 3. Control
 
-<a href="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshots/knowledge-base.gif"><img src="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshots/knowledge-base.gif" alt="Knowledge Base" width="700"></a>
+Full runtime management API. Start, stop, and monitor agent executions from any device. Web terminal access to running Claude Code sessions.
 
-> *Browse 667 knowledge entries, expand details with code snippets and file references, inspect context injection logs showing exactly what knowledge gets injected into each prompt*
+<a href="https://langmart.ai/images/assist/session-terminal.png"><img src="https://langmart.ai/images/assist/session-terminal.png" alt="Web Terminal — control Claude Code from any browser" width="700"></a>
 
-**How it works:**
-1. **Generate** — Analyzes your sessions and extracts reusable knowledge (patterns, decisions, debugging insights)
-2. **Search** — Indexed with BM25 + vector similarity for fast retrieval
-3. **Inject** — On every prompt, the context-injection hook finds relevant knowledge and injects it as context
-4. **MCP tools** — Claude Code can also actively search and retrieve knowledge using `search`, `detail`, and `feedback` tools
+> *Live terminal access to running Claude Code sessions — from any browser, anywhere*
 
-<a href="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshots/context-injection-cli.png"><img src="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshots/preview/context-injection-cli.png" alt="Context Injection in Claude Code CLI" width="700"></a>
+- Start and abort agent executions via REST API
+- Web terminal (ttyd) from any browser — control Claude Code remotely
+- SDK runner for programmatic headless execution
+- Session cache warm/clear for performance tuning
+- Remote access via LangMart Hub — no VPN or port forwarding needed
 
-> *Claude Code CLI — MCP tools search and inject knowledge before Claude responds*
+**Key endpoints:** `POST /monitor/abort/:executionId` · `POST /ttyd/start` · `POST /agent/execute` · `POST /hub/connect`
 
 ### Settings
 
 <a href="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshots/settings.png"><img src="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshots/preview/settings.png" alt="Settings" width="700"></a>
 
-> *Settings — cloud sign-in, LAN access, connection status, and more*
+> *Settings — connection status, Claude Code config, knowledge kill switch, and more*
 
 ### Mobile & Tablet Support
 
-The web UI is fully responsive. Access everything from your phone or tablet — control terminals, browse sessions, review knowledge, and manage tasks on the go.
+The web UI is fully responsive. Monitor sessions, debug agents, and control terminals from your phone or tablet.
 
 <table>
   <tr>
     <td align="center"><a href="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshot/session-terminal%20(mobile).png"><img src="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshot/session-terminal%20(mobile).png" alt="Terminal on mobile" width="180"></a><br><sub>Live Terminal</sub></td>
     <td align="center"><a href="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshot/session-detail-chat%20(mobile).png"><img src="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshot/session-detail-chat%20(mobile).png" alt="Session detail on mobile" width="180"></a><br><sub>Session Detail</sub></td>
     <td align="center"><a href="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshot/task-kanban%20(mobile).png"><img src="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshot/task-kanban%20(mobile).png" alt="Task kanban on mobile" width="180"></a><br><sub>Task Kanban</sub></td>
-    <td align="center"><a href="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshot/knowledge-list%20(mobile).png"><img src="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshot/knowledge-list%20(mobile).png" alt="Knowledge list on mobile" width="180"></a><br><sub>Knowledge Base</sub></td>
-  </tr>
-  <tr>
-    <td align="center"><a href="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshot/knowledge-detail-full%20(mobile).png"><img src="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshot/knowledge-detail-full%20(mobile).png" alt="Knowledge detail on mobile" width="180"></a><br><sub>Knowledge Detail</sub></td>
-    <td align="center"><a href="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshot/context-hook-logs%20(mobile).png"><img src="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshot/context-hook-logs%20(mobile).png" alt="Context hook logs on mobile" width="180"></a><br><sub>Context Injection Logs</sub></td>
     <td align="center"><a href="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshot/settings-connection%20(mobile).png"><img src="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshot/settings-connection%20(mobile).png" alt="Settings on mobile" width="180"></a><br><sub>Settings</sub></td>
-    <td align="center"><a href="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshot/settings-claude-code%20(mobile).png"><img src="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshot/settings-claude-code%20(mobile).png" alt="Claude Code settings on mobile" width="180"></a><br><sub>Claude Code Config</sub></td>
   </tr>
 </table>
 
-### Bring Your Knowledge to Every IDE
+---
+
+## Data Sources
+
+lm-assist reads the same JSONL session files regardless of how they were created:
+
+| Source | What It Produces | lm-assist Coverage |
+|--------|-----------------|-------------------|
+| **Claude Code CLI** | Interactive sessions with subagents, teams, worktrees | Full parsing — all message types, tool calls, metadata |
+| **Agent SDK** (Python/TypeScript) | Programmatic agent executions | Same JSONL format — full session inspection |
+| **Headless mode** (`claude -p`) | Background/CI runs | Detected via process status store |
+| **Running processes** | Live PID, tmux, terminal state | Real-time monitoring with zero polling overhead |
+
+---
 
-lm-assist builds knowledge from your Claude Code sessions — but that knowledge isn't locked to Claude Code. Any MCP-compatible IDE can connect to the lm-assist MCP server and access the same knowledge base: search entries, view details, and provide feedback.
+## Knowledge (Optional)
 
-One-click activation from the Settings page:
+lm-assist includes an optional knowledge subsystem that auto-extracts reusable knowledge from your sessions and injects it into future prompts via MCP tools and a context hook. Enable or disable it at runtime from Settings — the kill switch unloads the embedder and vector store to save ~100MB of memory.
 
-<a href="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshot/settings-ide-mcp%20(mobile).png"><img src="https://raw.githubusercontent.com/langmartai/lm-assist/main/docs/screenshot/settings-ide-mcp%20(mobile).png" alt="IDE MCP Integration" width="280"></a>
+> **Read:** [Your Claude Sessions Are Gold: Stop Paying Twice for the Same Knowledge](https://databunny.medium.com/your-claude-sessions-are-gold-stop-paying-twice-for-the-same-knowledge-7632ac6ddb88)
 
-**Supported IDEs:**
+When enabled, any MCP-compatible IDE can access the knowledge base:
 
 | IDE | MCP Config |
 |-----|-----------|
@@ -139,9 +174,6 @@ One-click activation from the Settings page:
 | **Windsurf** | `~/.windsurf/mcp.json` |
 | **Codex CLI** (OpenAI) | `~/.codex/config.toml` |
 | **Gemini CLI** (Google) | `~/.gemini/settings.json` |
-| **Google Antigravity** | `~/.gemini/antigravity/settings.json` |
-
-All IDEs get access to the same 3 MCP tools (`search`, `detail`, `feedback`) and the same knowledge base. Generate knowledge once in Claude Code, use it everywhere.
 
 ---
 
@@ -153,12 +185,17 @@ All IDEs get access to the same 3 MCP tools (`search`, `detail`, `feedback`) and
 curl -fsSL https://raw.githubusercontent.com/langmartai/lm-assist/main/install.sh | bash
 ```
 
-This clones the repo, builds, adds the marketplace, and installs the plugin. Then **open a new Claude Code session** and run:
+Then **open a new Claude Code session** and run `/assist-setup`.
 
-```
-/assist-setup
+### Install via npm
+
+```bash
+npm install -g lm-assist
+lm-assist start
 ```
 
+Then in Claude Code, run `/assist-setup`.
+
 ### Install from source
 
 ```bash
@@ -168,105 +205,185 @@ npm install && npm run build
 ./core.sh start
 ```
 
-Then in Claude Code, run `/plugin install .` to register the plugin. **Open a new Claude Code session** and run `/assist-setup`.
+Then in Claude Code, run `/plugin install .` and `/assist-setup`.
 
-### Install via npm
+### What gets installed
+
+| Component | Auto-installed | Purpose |
+|-----------|---------------|---------|
+| Core API + Web UI | Yes (via npm/source) | 155+ endpoint REST API + Next.js dashboard |
+| MCP server | Yes (via plugin) | `search`, `detail`, `feedback` tools |
+| Context hook | Yes (via plugin) | Knowledge injection (optional) |
+| Slash commands | Yes (via plugin) | 6 `/assist-*` commands |
+| Skills | Yes (via plugin) | `observe` (auto-triggered observability) + `route` (cross-project routing) |
+| Slash commands | Yes (via plugin) | 9 commands: `/sessions`, `/summary`, `/run` + 6 `/assist-*` |
+| Statusline | Optional | Git branch, context %, rate limits, process stats |
+
+## Services
+
+| Service | Port | Description |
+|---------|------|-------------|
+| Core API | 3100 | REST API — sessions, monitor, agents, tasks, knowledge |
+| Web UI | 3848 | Next.js dashboard — 15 insight tabs, terminal, settings |
 
 ```bash
-npm install -g lm-assist
-lm-assist start
+lm-assist start       # Start both services
+lm-assist stop        # Stop all services
+lm-assist status      # Health check + process info
+lm-assist upgrade     # Upgrade to latest version
 ```
 
-Then in Claude Code, run `/assist-setup`.
+## Skills & Commands
 
-### What gets installed
+lm-assist ships with **2 auto-triggered skills** and **9 slash commands** that work inside Claude Code.
 
-| Component | Auto-installed by plugin | Purpose |
-|-----------|-------------------------|---------|
-| MCP server | Yes | `search`, `detail`, `feedback` tools in Claude Code |
-| Context hook (Node.js) | Yes | Injects relevant knowledge into each prompt |
-| Slash commands | Yes | 6 `/assist-*` commands |
-| Statusline | No (optional) | Git branch, context %, process stats in status bar |
+### Skills (auto-triggered)
 
-The context hook uses Node.js for cross-platform support (Windows, macOS, Linux). The statusline is optional — install via `/assist-setup --statusline` or the web UI settings page.
+Skills activate automatically when Claude detects relevant intent — no slash command needed.
 
-## Slash Commands
+| Skill | Triggers on | What it does |
+|-------|------------|--------------|
+| **observe** | "what's running?", "session costs", "show subagents", "run this on project X" | Full observability — monitor sessions, debug agents, control executions, manage summaries |
+| **route** | Prompt mentions another project's features or codebase | Cross-project routing — evaluates whether to stay, resume, fork, queue, or start new |
 
-Use these from within any Claude Code session:
+### Commands
 
 | Command | Description |
 |---------|-------------|
+| `/sessions` | Session list with costs, turns, running status |
+| `/summary` | Summarize current session, generate display name, record learning |
+| `/run <prompt>` | Execute an agent with pre-flight checks |
 | `/assist` | Open the web UI in your browser |
 | `/assist-status` | Show status of all components |
 | `/assist-setup` | Start services and verify integrations |
-| `/assist-search <query>` | Search the knowledge base |
+| `/assist-search` | Search the knowledge base |
 | `/assist-logs` | View context-inject hook logs |
 | `/assist-mcp-logs` | View MCP tool call logs |
 
-## MCP Server
+### Use Cases with Examples
 
-The MCP server (`lm-assist`) provides 3 tools that any MCP-compatible IDE can use directly (Claude Code, VS Code, Cursor, Codex CLI, Gemini CLI, Antigravity):
+**"What sessions are running and how much have they cost?"**
+```
+> /sessions
+Sessions (3 running, 415 total)
+───────────────────────────────────────────────────────────────────────
+Status  Name                         Project            Model      Cost  Turns
+───────────────────────────────────────────────────────────────────────
+[RUN]   observability-platform-build  lm-assist          opus    $307.66  1462
+[RUN]   trade-delta-analysis          lm-unified-trade   opus    $153.20   822
+[RUN]   anti-kelly-system             lm-unified-trade   opus     $35.49   433
+        two-track-dashboard           lm-unified-trade   opus    $107.43   516
+        skill-validation-test         lm-assist          opus      $0.63    11
+───────────────────────────────────────────────────────────────────────
+Total cost: $604.41
+```
 
-| Tool | Description |
-|------|-------------|
-| `search` | Unified search across knowledge and file history |
-| `detail` | Progressive disclosure — expand any item by ID (e.g., `K001`, `arch:component`) |
-| `feedback` | Flag context as outdated, wrong, irrelevant, or useful |
+**"Summarize what this session has been doing"**
+```
+> /summary
+Session Summary
+═══════════════
+Name:    observability-platform-build
+Project: lm-assist
+Turns:   1462 | Cost: $307.66
+Status:  in progress
+
+What this session is about:
+  Building the lm-assist observability platform — session summaries,
+  cross-project routing, prompt queue, auto-learning.
+
+What was accomplished:
+  • Session slug/customTitle parsing from Claude Code JSONL
+  • Knowledge system kill switch (saves ~100MB)
+  • Observability skill + routing skill + 3 new commands
+  • Persistent session summaries with delta tracking
+  • Universal prompt queue with source/target context
+  • Auto-learning signals that evolve routing over time
+```
 
-When installed as a plugin, the MCP server is registered automatically. For non-plugin installs:
+**"Fix the delta analysis bug in the trading pipeline"**
 
-```bash
-curl -X POST http://localhost:3100/claude-code/mcp/install
+The route skill auto-detects this belongs to another project:
 ```
+This task appears to belong to lm-unified-trade, not the current project.
 
-## Services
+Relevant session found: trade-delta-analysis
+Summary: Implemented delta analysis mode for /trade-analyze command...
+Status: idle (not running)
 
-lm-assist runs two services:
+Recommendation: RESUME
+Reason: This session built all existing delta modes and knows the spec,
+flag table, and pre-launch copy patterns.
 
-| Service | Port | Description |
-|---------|------|-------------|
-| Core API | 3100 | REST API — sessions, knowledge, tasks |
-| Web UI | 3848 | Next.js dashboard — accessible from any device on your network |
+To resume:  claude --resume 36a30ba6-4bd1-4587-8bb3-c1d21a2ab616
+```
 
-```bash
-./core.sh start        # Start both services
-./core.sh stop         # Stop all services
-./core.sh restart      # Restart (auto-rebuilds if TypeScript changed)
-./core.sh status       # Health check
-./core.sh logs core    # View API logs
-./core.sh logs web     # View web logs
+**"Run a security review on this project"**
+```
+> /run review all API endpoints for security vulnerabilities
+Checking running executions... OK: none running.
+
+Execution started: agent-1774468578687-1p2cnh
+  Session: nifty-wishing-boole
+  Status: running
+  Monitor with: /sessions
+  Web UI: http://localhost:3848
+```
+
+**"What's the cost breakdown by project?"**
+
+The observe skill activates:
+```
+Project costs:
+  lm-assist              $308.29  (80 sessions)
+  lm-unified-trade       $892.15  (440 sessions)
+  LangMartDesign          $45.20  (12 sessions)
+  tier-agent              $23.50  (15 sessions)
+```
+
+### Session Intelligence
+
+lm-assist learns from every interaction to route prompts smarter over time:
+
+- **Session summaries** — LLM-generated descriptions stored with delta tracking. Updated incrementally as sessions grow.
+- **Project summaries** — Comprehensive reference generated by agents that explore each project's CLAUDE.md, scripts, configs, and codebase.
+- **Prompt queue** — When a session is busy, new work is queued with full context (original intent, formatted prompt, routing reason, context hint).
+- **Auto-learning** — Keywords, commands, and routing patterns accumulate with frequency counts. After a few interactions, routing skips deep scans entirely.
+
+```
+Learning context for lm-unified-trade:
+  Frequently mentioned: delta analysis(3x), anti-kelly(2x), regime analysis(1x)
+  Areas worked in: analysis pipeline(1x)
+  Routing patterns: delta analysis → trade-delta-analysis session
 ```
 
+## Key API Endpoints
+
+| Category | Endpoints | Highlights |
+|----------|-----------|------------|
+| **Sessions** | 27 | List, detail, delta fetch, batch-check, conversation, subagents, forks, DAG |
+| **Monitor** | 6 | Running executions, summary, abort, SSE stream |
+| **Projects** | 12 | List projects, sessions per project, git info, worktree detection, costs |
+| **Terminal** | 13 | ttyd start/stop/status, WebSocket proxy, tmux attach |
+| **Tasks** | 22 | Task lists, aggregated tasks, ready tasks, dependency tracking |
+| **Summaries** | 10 | Session summaries, project summaries, needs-update check |
+| **Queue** | 10 | Prompt queue with source/target, priority, dispatch/complete lifecycle |
+| **Learning** | 4 | Record signals, query by project, learning context generation |
+| **Search** | 4 | Session content search, recent sessions, vector search |
+| **Skills** | 9 | Skills analytics, chains, traces, per-session breakdown |
+| **Knowledge** | 21 | List, search, generate, review (optional — can be disabled) |
+
+All endpoints support `ifModifiedSince` for efficient polling. Session data supports three indexing dimensions: `lineIndex` (JSONL position), `turnIndex` (conversation turn), and `userPromptIndex` (user message count).
+
 ## Configuration
 
-No API key is needed — lm-assist works entirely with your local Claude Code session data. Optionally copy `.env.example` to `.env` to customize ports:
+No API key needed — lm-assist works entirely with your local Claude Code session data. Optionally copy `.env.example` to `.env`:
 
 ```bash
 API_PORT=3100                    # Core API port (default: 3100)
 WEB_PORT=3848                    # Web UI port (default: 3848)
 ```
 
-## Project Structure
-
-```
-lm-assist/
-├── core/                    ← Backend API (TypeScript)
-│   ├── src/
-│   │   ├── mcp-server/      ← MCP server (search, detail, feedback tools)
-│   │   ├── routes/core/     ← REST API routes (155 endpoints)
-│   │   ├── knowledge/       ← Knowledge generation pipeline
-│   │   └── vector/          ← Embeddings + vector store
-│   └── hooks/               ← Claude Code hook scripts (Node.js, cross-platform)
-├── web/                     ← Web UI (Next.js, React 19)
-├── commands/                ← Slash command definitions
-├── hooks/                   ← Plugin hook registration
-├── docs/screenshots/        ← Product screenshots
-├── .claude-plugin/          ← Claude Code plugin metadata
-├── .mcp.json                ← MCP server auto-registration
-├── core.sh                  ← Service manager
-└── bin/lm-assist.js         ← CLI entry point
-```
-
 ## Platform Support
 
 | Platform | Support | Notes |
@@ -274,9 +391,14 @@ lm-assist/
 | Linux | Full | All features including web terminal |
 | macOS | Full | All features including web terminal |
 | Windows | Partial | Everything except console/terminal access (ttyd not available) |
-| Mobile / Tablet | Web UI | Browse sessions, tasks, knowledge from any device on your network |
+| Mobile / Tablet | Web UI | Monitor, debug, and control from any device on your network |
+
+## Who It's For
 
-The web UI is fully responsive — optimized for phone, tablet, and desktop viewports.
+- **Solo developers** using Claude Code — see what's happening across all your sessions
+- **Teams building with Agent SDK** — observability for your agent pipelines
+- **DevOps managing agent fleets** — multi-machine dashboard, cost tracking, process management
+- **AI product builders** — debug agent behavior with 15 insight views
 
 ## Changelog