intent-guardian（意向守护者）

Intent Guardian

Never lose your train of thought again.

Intent Guardian continuously monitors your desktop activity, builds a real-time understanding of what you're working on, and detects when you've been derailed by an interruption and forgotten your original task. When that happens, it nudges you back on track.

The Problem

Knowledge workers are interrupted every 3 minutes on average. After an interruption, it takes ~23 minutes to return to the original task -- and often, you never return at all. Not because the task isn't important, but because you simply *forgot*.

Intent Guardian solves this by maintaining a task stack -- a living model of what you're doing, what you were doing, and what got interrupted -- so you don't have to.

When to Use

• You are a multi-tasker who frequently context-switches between apps and tasks
• You often finish replying to a message and think "wait, what was I doing?"
• You want an AI companion that understands your work rhythm, not just your commands
• You want to build long-term awareness of your own focus patterns

Core Capabilities

1. Real-Time Activity Sensing

Polls the active window title and application name to build a continuous activity stream.

macOS:

bash command:"scripts/sense_activity.sh"

With ActivityWatch (richer data):

bash command:"scripts/sense_activitywatch.sh"

With Screen Capture (optional, requires vision model):

bash command:"scripts/sense_screen.sh"

2. Task Stack Maintenance

The agent maintains a task stack in memory/skills/intent-guardian/task_stack.json:

{
  "stack": [
    {
      "id": "task_001",
      "intent": "Writing product requirements doc, section 3",
      "app": "Google Docs",
      "window_title": "Product Requirements v2 - Google Docs",
      "started_at": "2026-02-26T14:28:00",
      "status": "suspended",
      "suspended_at": "2026-02-26T14:31:00",
      "suspended_by": "Slack notification from Li Si",
      "completion_estimate": 0.6
    },
    {
      "id": "task_002",
      "intent": "Replying to Li Si about the API bug",
      "app": "Slack",
      "window_title": "Slack - #engineering",
      "started_at": "2026-02-26T14:31:00",
      "status": "completed",
      "completed_at": "2026-02-26T14:35:00"
    },
    {
      "id": "task_003",
      "intent": "Looking up React useEffect cleanup pattern",
      "app": "Chrome",
      "window_title": "Stack Overflow - React useEffect cleanup",
      "started_at": "2026-02-26T14:35:00",
      "status": "active"
    }
  ],
  "current_focus": "task_003",
  "forgotten_candidates": ["task_001"]
}

3. Interruption & Forgetting Detection

The agent analyzes the task stack to detect potential forgetting:

Signals that suggest forgetting:

• A suspended task has not been resumed for longer than interruption_threshold_minutes
• The user has moved through 2+ unrelated contexts since the suspension
• The user shows "wandering" behavior: rapid app-switching without sustained focus
• The suspended task had low completion estimate (was in the middle of something)

Signals that suggest deliberate abandonment (do NOT remind):

• The user explicitly closed the application associated with the task
• The task was at a natural completion point (e.g., finished sending an email)
• The user started a clearly intentional new deep-work session

4. Gentle Reminders

When forgetting is detected, the agent sends a non-intrusive reminder:

Simple reminder:

"You were working on the product requirements doc (section 3) about 25 minutes ago, before the Slack interruption. Want to go back to it?"

Smart reminder (with context carried forward):

"You left the product requirements doc to look up the useEffect cleanup pattern for Li Si's bug. The answer is: return a cleanup function in useEffect. Ready to bring this back to the doc?"

The second form is more powerful -- it doesn't just remind you what you forgot, it completes the reason you left, so you can return to your main task with the answer in hand.

5. Long-Term Pattern Learning

Over time, the agent builds a focus profile in memory/skills/intent-guardian/focus_profile.json:

{
  "user_id": "default",
  "updated_at": "2026-02-26",
  "patterns": {
    "avg_focus_duration_minutes": 12,
    "interruption_sources": {
      "Slack": { "count": 45, "forget_rate": 0.78 },
      "Mail": { "count": 22, "forget_rate": 0.18 },
      "Chrome": { "count": 31, "forget_rate": 0.42 }
    },
    "peak_focus_hours": ["09:00-11:00", "14:00-16:00"],
    "high_risk_transitions": [
      { "from": "VSCode", "to": "Slack", "forget_rate": 0.82 },
      { "from": "Google Docs", "to": "Chrome", "forget_rate": 0.55 }
    ],
    "reminder_effectiveness": {
      "accepted": 34,
      "dismissed": 8,
      "ignored": 5
    }
  }
}

This profile is used to personalize detection thresholds and reminder timing.

Architecture

 Your Desktop
      |
      v
 [Sensing Layer]  ---- scripts/sense_activity.sh (window title, every N sec)
      |                 scripts/sense_activitywatch.sh (ActivityWatch API)
      |                 scripts/sense_screen.sh (optional screenshot + vision)
      v
 [Activity Log]   ---- memory/skills/intent-guardian/activity_log.jsonl
      |
      v
 [Understanding]  ---- LLM analyzes activity stream on each Heartbeat:
      |                 - Segments activities into logical tasks
      |                 - Infers intent for each task
      |                 - Detects interruptions and context switches
      v
 [Task Stack]     ---- memory/skills/intent-guardian/task_stack.json
      |
      v
 [Detection]      ---- Compares stack state against forgetting heuristics
      |                 Consults focus_profile.json for personalized thresholds
      v
 [Reminder]       ---- Sends reminder via configured notification channel
      |                 Logs user response for feedback loop
      v
 [Learning]       ---- Updates focus_profile.json with new patterns

Heartbeat Integration

Add to your HEARTBEAT.md:

## Intent Guardian Check
Every heartbeat, run the intent guardian cycle:
1. Read latest activity from memory/skills/intent-guardian/activity_log.jsonl
2. Update task_stack.json with new activities
3. Check for forgotten tasks (suspended > threshold, not resumed)
4. If forgotten task detected, send reminder
5. Log any reminder responses to focus_profile.json

Cron Integration

Add a daily focus report:

openclaw cron add --name "intent-guardian-daily" --schedule "0 18 * * *" \
  --prompt "Generate my daily focus report using intent-guardian data"

Setup

Minimal Setup (macOS, no dependencies)

1. Install the skill:

npx playbooks add skill openclaw/skills --skill intent-guardian

1. The skill uses native macOS commands (osascript) to get the active window.

No additional software required.

1. Add the Heartbeat integration above to your HEARTBEAT.md.

Enhanced Setup (with ActivityWatch)

1. Install ActivityWatch for richer window-level tracking.
2. Set activitywatch_enabled: true in skill config.
3. ActivityWatch provides detailed per-window timelines including URLs and document titles.

Full Setup (with Screen Capture)

1. Set screen_capture_enabled: true and configure vision_model.
2. Requires a vision-capable model (Claude, Gemini, Qwen3-VL).
3. Screenshots are captured locally, analyzed locally or via API, and never stored permanently.

Privacy & Security

• All data stays local. Activity logs, task stacks, and focus profiles are stored as local files in your OpenClaw memory directory.
• No raw screenshots are persisted. If screen capture is enabled, frames are analyzed and immediately discarded. Only the semantic summary is kept.
• Excluded apps are respected. Configure excluded_apps to prevent tracking specific applications.
• Working hours enforced. Sensing only runs during configured working_hours_start to working_hours_end.
• You own the data. All files are human-readable JSON/JSONL. Review, edit, or delete at any time.

Configuration Reference

Example Interactions

User: "What was I doing before I got sidetracked?" Agent reads task_stack.json, responds:

"Before the Slack interruption at 14:31, you were writing section 3 of the product requirements doc in Google Docs. You were about 60% through. Want to jump back?"

User: "Show me my task stack" Agent responds with current stack visualization:

`` Task Stack (3 items): [ACTIVE] Looking up React useEffect cleanup (Chrome) - 17 min [COMPLETED] Replying to Li Si about API bug (Slack) - 4 min [SUSPENDED] Writing product requirements doc, section 3 (Google Docs) - 25 min ago !! ``

User: "How's my focus today?" Agent reads activity log and focus profile:

"Today you've had 6 deep work sessions averaging 18 minutes each. You were interrupted 11 times, mostly by Slack (7 times). You forgot to return to your original task 4 times -- I caught 3 of those. Your peak focus was 09:30-10:45."

Proactive (via Heartbeat):

"Heads up -- you've been away from the quarterly report (Excel) for 40 minutes since that phone call. It's due tomorrow. Want to wrap up the current task and go back to it?"

Integration with Other Skills

Intent Guardian works well alongside these existing skills:

Roadmap

• [ ] Cross-device sync (remind on phone about tasks left on laptop)
• [ ] Team mode (with consent: "Zhang San is waiting for your review, 2 days now")
• [ ] Proactive task delegation ("I can draft that email reply while you focus on the doc")
• [ ] Calendar-aware urgency ("The quarterly report meeting is in 2 hours, and you're only 60% done")
• [ ] Intent hierarchy (L1-L4 goal tracking, not just immediate tasks)

License

MIT