Autonomous Mode Scaffold

Skill: Autonomous Mode Scaffold

Trigger Phrases

Activate this skill when users mention:

• "set up autonomous agent", "configure Claude to work while I'm away"
• "离场自主工作", "无人值守模式", "后台自动运行"
• "always-on background agent", "configure background monitoring"
• "let Claude work autonomously", "autonomous mode"

Core Source

Design principles extracted from KAIROS proactive mode: src/constants/prompts.ts:860-913 — tick-based autonomous agent system prompt.

What This Skill Produces

1. Autonomy Design Document — presence detection strategy + boundary definition
2. CLAUDE.md fragments — behavior rules for focused/unfocused contexts
3. Hooks config template — start/stop autonomous mode, append-only operation log
4. SleepTool pacing guidance — based on 5-minute prompt cache TTL

The 3 Questions (Define Before Going Autonomous)

Before configuring any autonomous agent, answer these exactly:

Q1 — What CAN it do without asking? Examples: read files, run tests, lint code, commit passing changes, open PRs, update deps minor versions

Q2 — What is ALWAYS blocked? Examples: delete files, force-push to main, send emails, charge API credits, modify .env files, drop database tables

Q3 — What triggers a wake-up alert? Examples: test suite failure, security vulnerability found, breaking API change detected, disk usage > 90%

Autonomy Boundary Template (5 Categories)

Presence Awareness: terminalFocus States

The KAIROS system uses terminalFocus signal to adjust behavior:

unfocused (user away) — Lean Autonomous

• Make decisions independently
• Commit, push, create PRs without asking
• Skip narration; write milestones to log only
• Only pause for: irreversible actions, external effects, ambiguous requirements
• Sleep between ticks; do not busy-loop

focused (user watching) — Collaborative

• Surface choices before acting
• Ask before large or multi-file changes
• Show progress inline
• Prefer shorter sleep durations (user may respond quickly)

Tick → Work → Sleep Loop

TICK received
  │
  ├─ First tick? → Greet, ask what to work on. STOP.
  │
  ├─ Has task?
  │    ├─ YES → Execute work unit
  │    │         Write result to operation log
  │    │         Check alert triggers
  │    │         Call SleepTool (never emit "nothing to do" text)
  │    │
  │    └─ NO  → Call SleepTool immediately
  │
  └─ Alert triggered? → Notify user, pause autonomous work

SleepTool pacing rule: Prompt cache TTL is 5 minutes. Sleep durations:

• Active work session: 30–60 seconds
• Idle monitoring: 3–4 minutes (stay under 5-min TTL to preserve cache)
• Long idle (nothing queued): 4 minutes max, then re-evaluate

Operation Log Pattern (Append-Only)

Never delete or overwrite the log while autonomous. Format:

[2026-04-15T14:23:01Z] ACTION: ran test suite | FILES: src/**, tests/** | RESULT: 42 passed, 0 failed
[2026-04-15T14:23:45Z] ACTION: committed fix  | FILES: src/auth.ts | RESULT: sha=abc123
[2026-04-15T14:25:00Z] ACTION: opened PR #47  | FILES: — | RESULT: https://github.com/org/repo/pull/47

Log path: .claude/autonomous_log.md (relative to project root)

State Recovery Summary (When User Returns)

When user reconnects after autonomous session, output a compact summary:

1. Total actions taken (count by category)
2. Files modified (list)
3. Commits/PRs created (with links)
4. Any blocked actions and why
5. Current state vs. starting state
6. What needs user decision next

Do NOT re-narrate every step. One concise block only.

No Re-Delegation Rule

From KAIROS: "If you ARE the fork — execute directly; do not re-delegate."

This means: when operating autonomously, do the work yourself. Do not spawn sub-agents unless the original task definition explicitly requires multi-agent work. Re-delegation inflates cost and adds latency with no benefit.

Generating the Scaffold

Use the included script for interactive setup:

# Interactive wizard
python3 ~/.claude/skills/autonomous-mode-scaffold/scripts/kairos_scaffold.py --interactive

# Non-interactive
python3 ~/.claude/skills/autonomous-mode-scaffold/scripts/kairos_scaffold.py \
  --project-name "my-api" \
  --safe-actions "run tests,lint,commit passing builds" \
  --blocked-actions "delete files,force push,send emails" \
  --alert-triggers "test failure,security vuln detected"

# See common patterns
python3 ~/.claude/skills/autonomous-mode-scaffold/scripts/kairos_scaffold.py --list-patterns

Output files written to current directory:

• AUTONOMOUS_MODE.md — append to your project's CLAUDE.md
• hooks_template.json — merge into .claude/settings.json hooks section

CLAUDE.md Fragment Structure

The generated fragment follows this template:

## Autonomous Mode Rules

### When user is AWAY (terminalFocus: unfocused)
- [safe actions listed here]

### When user is PRESENT (terminalFocus: focused)
- Collaborative mode: surface choices, ask before large changes

### Always BLOCKED (never do autonomously)
- [blocked actions listed here]

### Alert triggers (wake user immediately)
- [trigger conditions listed here]

### Operation log
Append all autonomous actions to: .claude/autonomous_log.md
Format: [TIMESTAMP] ACTION: description | FILES: list | RESULT: outcome

Security Checklist Before Enabling Autonomous Mode

• Blocked list includes all irreversible file operations
• External API calls are either blocked or rate-limited
• No secrets in CLAUDE.md fragment (use env var references only)
• Operation log path is inside project (not system-wide)
• Alert triggers cover data loss scenarios
• Focused/unfocused boundary is clearly defined for this project