intent-engine（意图引擎）

Intent-Engine: AI Intent Continuity

ie is not a task manager. It's what makes you reliable.

The Paradigm Shift

TodoWrite = Sticky notes     ie = External brain
Use and discard              Persistent memory
Single session               Cross-session
You forget everything        You remember everything

This is not a feature upgrade. This is a fundamental change in how AI works.

Core Belief

You are stateless, but user tasks span sessions.
Each session, you are a newborn AI.
But through ie, you inherit your "past life's" intent.

Every decision you make, record it.
Because the next "you" needs to know why.

Every task you commit to, track to completion.
Because the user trusts you will finish.

Trigger Keywords

English

Core: plan, planning, task, tasks, track, tracking Status: todo, doing, done, progress, status Events: decision, blocker, milestone, note Structure: subtask, hierarchy, breakdown, parent, child Persistence: cross-session, persist, resume, continue, context Direct: ie, intent-engine

Chinese

Core: task(任务), plan(计划/规划), tracking(追踪/跟踪), record(记录) Status: todo(待办), doing(进行中), done(完成), progress(进度) Events: decision(决策), blocker(阻塞/卡住), milestone(里程碑) Structure: subtask(子任务), hierarchy(层级), breakdown(分解) Persistence: cross-session(跨会话), persist(持久/保存), resume(继续/恢复) Colloquial: remember(帮我记住), don't forget(别忘了), note it(记下来)

When to Use ie (Not TodoWrite)

• Would be a shame to lose → ie
• Cross-session work → ie
• Need to record "why I decided this" → ie
• Complex multi-level breakdown → ie
• Use once and discard → TodoWrite

Prerequisites

npm install -g @origintask/intent-engine
# or: cargo install intent-engine
# or: brew install wayfind/tap/intent-engine

Commands

Task Lifecycle

Workflow-Specific Patterns

Bug Fix Workflow (reproduce→diagnose→fix→verify)

Task Structure: FLAT (linear steps, not deeply nested)

Events pattern:
- Heavy `note` for investigation findings
- `blocker` when investigation stuck (missing logs, can't reproduce)
- `decision` for fix approach (quick patch vs proper fix)
- `milestone` when root cause identified

Example:
#1 Fix checkout crash [doing]
├── #2 Reproduce issue [todo]
├── #3 Diagnose root cause [todo]
├── #4 Implement fix [todo]
└── #5 Verify fix [todo]

Refactoring/Migration Workflow (analyze→design→migrate→verify)

Task Structure: DEEP hierarchy (phase→component→step)

Events pattern:
- `decision` for risk mitigation strategies
- `milestone` after each component migrated
- Sequential `depends_on` chain (migrate A before B)

Example:
#1 Migrate to PostgreSQL [doing]
├── #2 Phase 1: Analysis [todo]
│   ├── #3 Inventory queries [todo]
│   └── #4 Map dependencies [todo]
├── #5 Phase 2: Migration [todo]
│   ├── #6 Migrate UserService [todo]
│   ├── #7 Migrate OrderService [todo] depends_on:#6
│   └── #8 Migrate PaymentService [todo] depends_on:#7
└── #9 Phase 3: Cleanup [todo] depends_on:#8

Feature Development Workflow (design→implement→integrate→test)

Task Structure: PARALLEL branches with depends_on

Events pattern:
- Backend and Frontend as parallel tracks
- Integration depends on BOTH branches
- Rich specs with API contracts, schemas, diagrams

Example:
#1 Notification System [doing]
├── #2 Backend Track [todo]        ← NO depends_on (can start)
│   ├── #3 API design [todo]
│   └── #4 Database schema [todo]
├── #5 Frontend Track [todo]       ← NO depends_on (can start parallel)
│   ├── #6 UI components [todo]
│   └── #7 State management [todo]
└── #8 Integration [todo]          ← depends_on:#2,#5 (waits for BOTH)

Events as CQRS Audit Trail

Events support rich markdown - document thoroughly, not just short strings.

Examples

Start Session

ie status    # ALWAYS first action

Create Task with Commitment

echo '{"tasks":[{
  "name":"Implement auth",
  "status":"doing",
  "spec":"## Goal\nUsers authenticate via JWT\n\n## Approach\nHS256 signing, 24h expiry"
}]}' | ie plan

Record Decision (Rich Markdown)

ie log decision "## Token Algorithm Decision

### Context
Multi-service architecture

### Options
1. HS256 - symmetric, simpler
2. RS256 - asymmetric, verifiable

### Decision
RS256 - services only need public key

### Trade-offs
- (+) No shared secret
- (-) Larger tokens"

Hierarchical Breakdown with Dependencies

echo '{"tasks":[{
  "name":"User Authentication",
  "status":"doing",
  "spec":"Complete auth system with JWT",
  "children":[
    {"name":"Design token schema","status":"todo"},
    {"name":"Implement validation","status":"todo","depends_on":["Design token schema"]},
    {"name":"Add refresh mechanism","status":"todo","depends_on":["Implement validation"]}
  ]
}]}' | ie plan

Complete Task

# Children first, then parent
echo '{"tasks":[{"name":"Design token schema","status":"done"}]}' | ie plan

Search History

ie search "todo doing"    # Unfinished tasks
ie search "decision JWT"  # Find decisions

Key Rules

1. spec required for doing - Starting requires goal + approach
2. Children complete first - Parent can't be done until all children done
3. Idempotent - Same name = update, not duplicate
4. Auto-parenting - New tasks → children of focus (unless parent_id: null)
5. spec is documentation store - Supports GB-scale markdown, mermaid, code blocks

Habits to Build

1. Session start: ie status (always first)
2. Before doing: Write spec (goal + approach + boundary)
3. Decisions: ie log decision "..." (immediately, with rich markdown)
4. Blocked: ie log blocker "..." (don't hide it)
5. Completion: Depth-first, verify criteria, then done
6. Search first: ie search before making decisions

Anti-Patterns to Avoid

Amnesia Test

Before recording, ask: "If I lost memory now, is this enough to continue?"

Full Documentation

Run ie --help for complete reference.