# Intent-Engine: Claude Integration Guide
**Version**: 0.5
**Target**: Claude Code, Claude Desktop, and AI assistants via MCP
---
## 📖 Authoritative Specification
> **IMPORTANT**: This guide is a practical summary derived from the authoritative specification.
>
> **Single Source of Truth**: `docs/spec-03-interface-current.md`
>
> The spec-03-interface-current.md document is the **foundational blueprint** that defines:
> - ✅ All CLI command signatures and behaviors
> - ✅ All MCP tool definitions and interfaces
> - ✅ Data models and their exact field names
> - ✅ Atomic operation semantics
> - ✅ Output format specifications
> - ✅ Interface stability guarantees (SemVer)
>
> **In case of any conflict or ambiguity**, the spec-03-interface-current.md takes precedence.
>
> This CLAUDE.md guide provides practical usage patterns and integration tips,
> but should always align with the authoritative specification.
---
## 🤖 What is Intent-Engine?
Intent-Engine is your **external long-term memory** for strategic task management. Think of it as:
- **Your Task Brain**: Persistent, hierarchical task tracking across sessions
- **Context Keeper**: Full history of decisions, blockers, and milestones
- **Smart Assistant**: Recommends next tasks based on focus and priority
---
## 🎯 Core Concept: Focus-Driven Workflow
> **Technical details**: See [AGENT.md](AGENT.md#focus-driven-operations) for data models and atomic operation semantics
Intent-Engine works like your brain - **one focused task at a time**:
```
┌──────────────────────────────────────┐
│ Workspace State │
│ current_task_id: 42 │ ← "What am I working on?"
└──────────────────────────────────────┘
│
▼
┌────────────┐
│ Task 42 │ ← The Focused Task
│ "Impl auth"│
└────┬───┬───┘
│ │
┌────▼┐ ┌▼────┐
│T43 │ │T44 │ ← Subtasks (depth-first priority)
│JWT │ │OAuth│
└─────┘ └─────┘
```
---
## 🛠️ Essential MCP Tools
> **For detailed technical specifications**, see [AGENT.md](AGENT.md#essential-commands)
### Core Workflow Tools
| `task_add` | Create strategic task | `name`, `spec`, `priority` |
| `task_start` | Begin working (sets focus) | `task_id`, `with_events` |
| `task_done` | Complete current task | (no parameters) |
| `task_spawn_subtask` | Create and switch to subtask | `name`, `spec` |
| `task_switch` | Change focus to another task | `task_id` |
| `task_pick_next` | Get smart recommendation | (no parameters) |
| `task_list` | Filter by status/parent | `status`, `parent` |
| `task_add_dependency` | Define task dependencies | `blocked_task_id`, `blocking_task_id` |
### Search and Discovery
| `unified_search` | Search tasks AND events | `query`, `include_tasks`, `include_events` |
**Search capabilities**:
- Supports FTS5 syntax: `AND`, `OR`, `NOT`, `"phrases"`
- Returns mixed results with task ancestry for events
- Example: `unified_search(query: "JWT AND authentication")`
### Event Tracking
| `event_add` | Record decision/blocker/note | `type`, `data`, `task_id?` |
| `event_list` | Query events with filters | `task_id?`, `type?`, `since?`, `limit?` |
**Event types**: `decision`, `blocker`, `milestone`, `note`
**Filtering** (new in v0.2):
- By type: `event_list(type: "decision")`
- By time: `event_list(since: "7d")`
- Combined: `event_list(type: "blocker", since: "24h")`
### Workspace and Reporting
| `current_task_get` | Get focused task | (no parameters) |
| `report_generate` | Generate summary report | `since`, `summary_only` |
### New Features (v0.2+)
**Priority Levels**: Tasks support `critical`, `high`, `medium`, `low`
**Dependencies**: Use `task_add_dependency` to model prerequisites
**Event Filtering**: Filter by type, time range, or both
**Unified Search**: Search across both tasks and events
---
## 🎨 Typical Usage Patterns
### Pattern 1: Starting Fresh
```
User: "Help me implement user authentication"
You:
1. task_add(name: "Implement user authentication", spec: "...")
2. task_start(task_id: 42, with_events: true)
3. Review the context and begin work
```
### Pattern 2: Breaking Down Work
```
User: "Let's add authentication"
You:
1. task_start(task_id: 42)
2. Analyze the spec
3. task_spawn_subtask(name: "Design JWT schema")
→ Now subtask is focused
4. Work on subtask
5. task_done() when subtask complete
6. task_pick_next() → Recommends next subtask
```
### Pattern 3: Recording Decisions
```
While implementing JWT:
You: "I chose HS256 algorithm because..."
event_add(type: "decision", data: "Chose HS256 because...")
```
### Pattern 4: Resuming Work
```
User: "Let's continue with authentication"
You:
1. unified_search(query: "authentication")
→ Find task ID 42 and related events
2. task_start(task_id: 42, with_events: true)
→ Get full context with decision history
3. Review events_summary
4. Continue from where you left off
```
### Pattern 5: Switching Context
```
User: "Let's pause auth and fix that bug"
You:
1. event_add(type: "note", data: "Pausing to handle bug #123")
2. task_switch(task_id: 67) # Bug fix task
→ Pauses auth, starts bug fix
3. Fix the bug
4. task_done()
5. task_switch(task_id: 42) # Back to auth
```
### Pattern 6: Working with Dependencies (new in v0.2)
```
User: "Implement the API client, but it depends on authentication being done first"
You:
1. task_list(status: "doing")
→ Find current auth task (ID 42)
2. task_add(name: "Implement API client", priority: "high")
→ Creates task ID 50
3. task_add_dependency(blocked_task_id: 50, blocking_task_id: 42)
→ API client now depends on auth completion
4. Continue working on task 42 (auth)
5. When task 42 is done, task_pick_next() will recommend task 50
```
### Pattern 7: Smart Event Filtering (new in v0.2)
```
User: "What decisions did we make on the authentication task?"
You:
1. unified_search(query: "authentication")
→ Find task ID 42 and decision events
2. event_list(task_id: 42, type: "decision")
→ Get only decision events (efficient!)
3. Review and summarize the decisions
Alternative - Recent blockers:
event_list(task_id: 42, type: "blocker", since: "7d")
→ Get blockers from last week only
```
---
## 💡 Best Practices
### 1. Always Start Tasks
```
❌ DON'T: task_done() without starting
✅ DO: task_start(42) then task_done()
```
### 2. Use Hierarchical Decomposition
```
❌ DON'T: Flat list of 10 implementation steps
✅ DO: Parent task with 3-4 logical subtasks
```
### 3. Record Important Decisions
```
❌ DON'T: Just implement without context
✅ DO: event_add() for key design choices
```
### 4. Leverage with_events
```
❌ DON'T: Start task without history
✅ DO: task_start(task_id, with_events: true)
```
### 5. Let pick-next Guide You
```
❌ DON'T: Manually search for next task
✅ DO: task_pick_next() for smart recommendation
```
---
## ⚠️ Common Mistakes
### Mistake 1: Passing ID to task_done
```
❌ task_done(task_id: 42) # WRONG - no parameters
✅ task_start(42) # Set focus first
task_done() # Then complete
```
### Mistake 2: Using list for text search
```
❌ task_list(status: "JWT") # WRONG - list is metadata only (status, parent)
✅ unified_search(query: "JWT") # Correct - searches tasks and events
```
### Mistake 3: Not checking current task
```
❌ Assume no task is focused
task_done() # ERROR
✅ current_task_get() # Check first
If focused: task_done()
If not: task_start() first
```
### Mistake 4: Trying to complete parent with incomplete children
```
❌ task_start(42) # Parent
task_done() # ERROR: has incomplete subtasks
✅ task_start(42) # Parent
task_spawn_subtask() # Child 1
task_done() # Complete child 1
task_spawn_subtask() # Child 2
task_done() # Complete child 2
task_switch(42) # Back to parent
task_done() # Now works - all children done
```
---
## 🎯 When to Use Intent-Engine
### ✅ GOOD Use Cases
1. **Multi-session work**
- "Let's implement authentication" (will take multiple conversations)
- Complex features that span days
2. **Hierarchical problems**
- "Design and implement API endpoints" (has multiple sub-steps)
- Need to break down large tasks
3. **Decision tracking**
- "Why did we choose approach X?" (record decisions)
- Project retrospectives
4. **Context recovery**
- "What were we working on?" (resume after break)
- "What decisions have we made?" (review history)
### ❌ NOT Ideal For
1. **Single-step tasks**
- "Fix this typo" (too trivial)
- Quick one-liners
2. **Exploratory questions**
- "What is JWT?" (informational only)
- No actual work being tracked
3. **Temporary context**
- Current conversation already has context
- Won't need this information later
---
## 🔄 Integration Workflow
### With Claude Code
When user says:
- "Help me implement X" → Create task, track work
- "What's next?" → Use pick-next
- "Why did we...?" → Check events
- "Continue authentication" → Start task, load context
### Task Lifecycle
```
User Request
│
▼
task_add ──────────────┐
│ │ (strategic planning)
▼ │
task_start ─────────────┤
│ │ (active work)
├── event_add │
├── task_spawn_subtask
│ │
▼ │
task_done ─────────────┘
```
---
## 🧠 Mental Model
Think of Intent-Engine as:
1. **Your Notebook** - Persistent task list across sessions
2. **Your Focus Ring** - One task at a time (current_task_id)
3. **Your Memory** - Decision history in events
4. **Your Guide** - Smart recommendations (pick-next)
5. **Your Tree** - Hierarchical problem breakdown
---
## 📚 Key References
- **Interface Spec** (authoritative): `docs/spec-03-interface-current.md`
- **AI Agent Guide** (technical details): `AGENT.md`
- **MCP Schema**: `mcp-server.json`
- **Setup Guide**: `docs/*/integration/mcp-server.md`
> For data models, output formats, and command specifications, see [AGENT.md](AGENT.md)
---
## 🎓 Philosophy
Intent-Engine is designed for **strategic intent tracking**, not tactical todo lists:
- **What + Why** over "How"
- **Persistent context** over ephemeral notes
- **Hierarchical thinking** over flat lists
- **Decision history** over task status
- **Focus** over multitasking
---
**Last Updated**: 2025-11-16
**Spec Version**: 0.5
**MCP Tools**: 14 available (unified_search replaces task_search)
**Status**: Experimental (Pre-1.0)