← 返回 Skills 市场
luciusrockwing

Windowed Autonomy

作者 luciusrockwing · GitHub ↗ · v1.0.0
cross-platform ✓ 安全检测通过
705
总下载
0
收藏
3
当前安装
1
版本数
在 OpenClaw 中安装
/install autonomy-windowed
功能描述
Time-windowed autonomous task queue. Autonomy works only during specific time windows (e.g., 8 AM - 8 PM UTC), while cron jobs run overnight. Separates conce...
使用说明 (SKILL.md)

Windowed Autonomy

Transform your agent from reactive to autonomous worker during specific time windows.


Concept

The agent works autonomously only during defined time windows. Outside windows, the agent only replies HEARTBEAT_OK. Cron jobs run overnight when autonomy is disabled.

🌙 Overnight (10 PM - 8 AM UTC):  Autonomy OFF
☀️ Daytime (8 AM - 10 PM UTC):   Autonomy ON

Clear separation: Day = work on queue, Night = cron maintenance


How It Works

1. Time Windows

Define active windows in tasks/QUEUE.md:

## ⏰ Autonomy Windows

- 🌙 Overnight: 20:00 - 08:00 UTC → Autonomy OFF
- ☀️ Daytime: 08:00 - 20:00 UTC → Autonomy ON

2. Heartbeat Flow

During Active Window:

Heartbeat → Check urgent → No → Read QUEUE → Pick task → Work 15-30 min → Update QUEUE → Log

Outside Active Window:

Heartbeat → Check urgent → No → Reply "HEARTBEAT_OK" → Sleep

Urgent Override:

Heartbeat → Check urgent → YES → Work immediately (even outside window)

3. Window Transitions

Window Opens (8:00 AM):

  • Autonomous mode enabled
  • Start pulling from queue
  • Log to tasks/MODE_HISTORY.md: "Window opened"

Window Closes (8:00 PM):

  • Autonomous mode disabled
  • Wrap up current task if in progress
  • Log transition: "Window closed"
  • Switch to standby mode (HEARTBEAT_OK only)

Schedule Examples

Time Activity Mode
00:00 GitHub backup (cron) Overnight (no autonomy)
02:00 Heartbeat check Standby (HEARTBEAT_OK)
03:00 Temp cleanup (cron) Overnight (no autonomy)
08:00 Window opens ✅ Autonomy ON
09:00 Work on task Autonomy
12:00 GitHub backup (cron) Autonomy (pause for backup)
14:00 Work on task Autonomy
20:00 Window closes ❌ Autonomy OFF
23:00 Daily ops + memory (cron) Overnights (no autonomy)

Queue Structure

# Task Queue

## ⏰ Autonomy Windows
- 🌙 Overnight: 20:00 - 08:00 UTC → Autonomy OFF
- ☀️ Daytime: 08:00 - 20:00 UTC → Autonomy ON

---

## 🔴 Ready (can be picked up during windows)
- [ ] @priority:high [Task description]
- [ ] @priority:medium [Task description]

## 🟡 In Progress
- [ ] @agent: @priority:high [Task description]
  - Started: 2026-02-16 14:00 UTC

## 🔵 Blocked
- [ ] @priority:medium [Task] (needs: [what's blocking])

## ✅ Done Today
- [x] @agent: @priority:high [Task]
  - Completed: 2026-02-16 14:25 UTC

## 💡 Ideas
- [Idea for future work]

Priority System

Priority affects task selection order during windows:

Priority When to use Selection
@priority:urgent Time-sensitive, deadline \x3C 6h Pick FIRST, even outside window
@priority:high Important, deadline 24h Pick first in window
@priority:medium Normal importance Pick second in window
@priority:low Nice to have Pick last in window

Window Types

Full Day Window (Default)

Active: 8 AM - 8 PM UTC (12 hours)

Use when:

  • youris available during these hours
  • Tasks are not time-sensitive
  • Want predictable work schedule

Heartbeat frequency: Every 2 hours (6 sessions/day)

Extended Window

Active: 6 AM - 10 PM UTC (16 hours)

Use when:

  • Want more work hours
  • yourhas varied schedule

Heartbeat frequency: Every 2 hours (8 sessions/day)

Limited Window

Active: 10 AM - 6 PM UTC (8 hours)

Use when:

  • Want controlled token budget
  • yourhas focused availability

Heartbeat frequency: Every 2-3 hours (3-4 sessions/day)


Token Budget

Recommended: 4-6 sessions/day, ~5-10K tokens each = 20-60K/day

Session strategy:

Window Type Sessions/day Tokens Schedule
Full Day (12h) 6 30-60K Every 2 hours
Extended (16h) 8 40-80K Every 2 hours
Limited (8h) 4 20-40K Every 2-3 hours

When to stop:

  • Window closes (time-based)
  • Tokens remaining \x3C 5K
  • Queue empty
  • youris actively messaging (human priority)

Urgent Tasks Override

Definition: Tasks that cannot wait for window to open

Add to queue with @priority:urgent:

## 🔴 Ready
- [ ] @priority:urgent Emergency: [task description]

Behavior:

  • Autonomy works immediately, even if outside window
  • Override all other considerations
  • Complete urgent task, then check window status

Wrapping Up When Window Closes

If in middle of task when window closes:

  1. Check task progress:

    • If > 80% complete → Finish it
    • If \x3C 80% complete → Save progress, move to In Progress, stop
  2. Save state:

## 🟡 In Progress
- [ ] @agent: @priority:high [Task description]
  - Started: 2026-02-16 19:30 UTC
  - Progress at window close: Completed X section, need to do Y
  - Resume: 2026-02-17 08:00 UTC (next window)
  1. Log to tasks/MODE_HISTORY.md:
## [2026-02-16 20:00 UTC] Window Closed
Status: Task in progress (60% complete)
Action: Saved state, will resume in next window

Mode History Tracking

File: tasks/MODE_HISTORY.md

# Window Mode Transitions

## [2026-02-16 08:00 UTC] Window Opened
Mode: Daytime window
Queue state: 3 tasks ready
Expected sessions: 6

## [2026-02-16 20:00 UTC] Window Closed
Mode: Overnight (standby)
Sessions completed: 5
Tasks completed: 4
Tasks remaining: 1 (in progress, resume tomorrow)

## [2026-02-17 08:00 UTC] Window Opened
Mode: Daytime window
Resumed task in progress

Coordinating With Cron

Daytime (Autonomy ON):

  • Light cron runs fine (Ollama monitor, disk check) - no conflict
  • Heavy cron (backup) - autonomy pauses during execution

Overnight (Autonomy OFF):

  • All cron jobs run freely
  • No autonomy conflicts

Cron schedule:

Time Cron Job Autonomy Mode
Every 5 min Ollama monitor Any (low impact)
Every hour Disk check Any (low impact)
00:00 GitHub backup OFF (overnight)
03:00 Temp cleanup OFF (overnight)
08:00 Window opens ON
12:00 GitHub backup ON (pause during)
14:00 Daily ops + memory (Sun) ON (pause during)
20:00 Window closes OFF (overnight)
23:00 Daily ops + memory OFF (overnight)

GOALS.md Integration

Link queue tasks to RA's long-term goal: MONEY

Add goal references:

- [ ] @priority:high Competitor pricing analysis (GOAL: monetization strategy)
- [ ] @priority:medium Write sales email template (GOAL: improve conversion)

Work toward goals during windows:

  • Research tasks during early window (fresh energy)
  • Writing tasks during mid window (flow state)
  • Review tasks during late window (wind down)

.learnings/ Integration

After completing tasks, add findings to .learnings/:

## [LRN-20260216-001] task-completion
Task: [Task description]

Completed during: Daytime window (14:00-14:25 UTC)
Tokens used: 8K

Key findings: [summarize]

Daily Routine

Window Opens (8:00 AM)

  1. Check queue for tasks in 🔴 Ready
  2. Pick highest priority task
  3. Work 15-30 minutes
  4. Update queue (In Progress or Done)

During Window (9 AM - 8 PM)

  1. Every 2 hours: heartbeat, check queue
  2. Work on next highest priority task
  3. Update queue after completion
  4. Log progress to memory/[today].md

Window Closes (8:00 PM)

  1. Check if task in progress
  2. If yes: Save progress or finish if close
  3. Move to Done Today or keep in In Progress
  4. Log transition tasks/MODE_HISTORY.md
  5. Switch to standby (HEARTBEAT_OK only)

Overnight (8 PM - 8 AM)

  1. Every heartbeat: Check urgent only
  2. If no urgent: Reply HEARTBEAT_OK
  3. Cron jobs run freely
  4. No autonomous work

Task Output Format

## Task: [Title]

**Completed during:** [Window type] [Time range]
**Duration:** X minutes
**Tokens:** YK

### Work Done
[Describe what was accomplished]

### Output
[Attach output file or content]

### Next Steps
[What to do next - add to Ideas if task spawns follow-up]

Evolving Over Time

The "slowly evolve" approach:

This mode allows The agent to gradually take on more work:

Week 1-2:

  • Conservative: 4 sessions/day, limited window
  • Monitor performance, adjust schedule

Week 3-4:

  • Expand: 6 sessions/day, extended window if needed
  • Add more task types to queue

Month 2:

  • Optimize: Tailor window to RA's actual schedule and peak productivity times
  • Adjust based on what yourfinds valuable

Key: Start small, expand based on feedback and results.


When to Use This Skill

Use this skill when:

  • You want predictable work schedule
  • You want autonomy to work during specific hours
  • Tasks are not time-sensitive (can wait for window)
  • You want clear temporal separation between autonomy and maintenance
  • You're okay with "slowly evolving" - start small, expand over time

When NOT to Use This Skill

Do not use this skill when:

  • You want 24/7 autonomous work → Use autonomy-type-based for type-based filtering
  • Tasks are highly time-sensitive → Use urgency override or different approach
  • You want task-based separation rather than time-based → Use autonomy-type-based
  • youris awake during non-standard hours → Adjust windows to match schedule

Quick Reference

Default window: 8 AM - 8 PM UTC (adjustable) Session frequency: Every 2-3 hours during window Priority order: urgent (override) → highmediumlow Queue location: tasks/QUEUE.md


See templates/QUEUE.md for full template structure

安全使用建议
This skill is coherent and appears to do what it says: manage a time-windowed autonomous task queue by reading and writing local task/memory files. Before installing: (1) Confirm you are comfortable with the agent having read/write access to the workspace paths named (memory/, tasks/, GOALS.md, .learnings) and that those files don't contain secrets you don't want persisted; (2) Verify how your agent runtime exposes token-usage and 'human active' signals — the skill assumes those signals exist but doesn't define how to get them; (3) Decide whether you want the 'urgent' override (allows work outside windows) and whether the agent should be allowed to edit GOALS.md or .learnings automatically; (4) Test in a conservative, limited window and review MODE_HISTORY and logs to ensure behavior matches expectations. If you want tighter control, restrict file-permission scope or add an approval step for urgent tasks and edits to goal files.
功能分析
Type: OpenClaw Skill Name: autonomy-windowed Version: 1.0.0 The skill bundle provides detailed instructions for an AI agent to manage its tasks, schedule, and logging based on time windows. All file operations are directed to specific, internal markdown files within its workspace (e.g., `tasks/QUEUE.md`, `memory/YYYY-MM-DD.md`, `tasks/MODE_HISTORY.md`, `GOALS.md`, `.learnings/`). There are no instructions for data exfiltration, unauthorized execution, persistence, or other malicious activities. The instructions, while extensive, are focused on the agent's self-management and operational workflow, which is a legitimate function for an autonomous agent.
能力评估
Purpose & Capability
Name/description match the actual instructions: the skill coordinates autonomous work during configured time windows and uses local queue/checkpoint files (tasks/QUEUE.md, memory/*, tasks/MODE_HISTORY.md). It does not request unrelated binaries, env vars, or external services.
Instruction Scope
Instructions direct the agent to read and write many workspace files (QUEUE.md, MODE_HISTORY.md, memory/episodic/[today].md, GOALS.md, .learnings). That is consistent with a task queue. Two points to note: (1) the SKILL assumes the agent/runtime can determine 'tokens remaining' and whether the human (RA) is actively messaging — the skill does not specify how to obtain those signals, which is a gap in the doc; (2) the agent is given discretionary authority (urgent override, choosing tasks, updating GOALS.md/.learnings), which is functional but grants broad action within the workspace.
Install Mechanism
Instruction-only skill with no install spec and no downloaded code. Nothing is written to disk by an installer — the only writes are the runtime file operations the instructions describe.
Credentials
The skill requests no environment variables or credentials. The only privileges it needs are read/write access to project workspace files (task and memory files), which is proportionate to its stated purpose.
Persistence & Privilege
always: false and the skill does not request persistent platform-level privileges or modifications to other skills. It does instruct writing to its own workspace files (mode history, memory files), which is expected for this kind of skill.
如何使用
  1. 确保已安装 OpenClaw(本地或 Docker 部署)
  2. 在对话框中输入安装命令:/install autonomy-windowed
  3. 安装完成后,直接呼叫该 Skill 的名称或使用 /autonomy-windowed 触发
  4. 根据 Skill 的参数说明提供必要输入,即可获得结构化输出
版本历史
v1.0.0
Initial release - time-windowed autonomous task queue system. Works only during defined windows (e.g., 8 AM - 8 PM UTC). Day = autonomy, night = cron maintenance. Predictable schedule with controlled token budget.
元数据
Slug autonomy-windowed
版本 1.0.0
许可证
累计安装 3
当前安装数 3
历史版本数 1
常见问题

Windowed Autonomy 是什么?

Time-windowed autonomous task queue. Autonomy works only during specific time windows (e.g., 8 AM - 8 PM UTC), while cron jobs run overnight. Separates conce... 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件,目前累计下载 705 次。

如何安装 Windowed Autonomy?

在 OpenClaw 或 Claude Code 对话框中运行命令「/install autonomy-windowed」即可一键安装,无需额外配置。

Windowed Autonomy 是免费的吗?

是的,Windowed Autonomy 完全免费(开源免费),可自由下载、安装和使用。

Windowed Autonomy 支持哪些平台?

Windowed Autonomy 跨平台运行,可在任意部署了 OpenClaw / Claude Code 的环境中使用(cross-platform)。

谁开发了 Windowed Autonomy?

由 luciusrockwing(@luciusrockwing)开发并维护,当前版本 v1.0.0。

💬 留言讨论