Autonomous Skill - Unattended Flow Execution

This skill enables autonomous mode where flows run to completion without human intervention. An Oversight Agent makes judgment calls that would normally require human input.

Status: Experimental - Use for well-defined, lower-risk tasks.

Setup in User's Project

When /flow --autonomous is first invoked, check if hooks are configured. If not, set them up.

Check for Existing Setup

# Check if hooks are configured
if [ -f ".claude/hooks/permission-handler.sh" ] && [ -f ".claude/hooks/stop-handler.sh" ]; then
  echo "Hooks already configured"
else
  echo "First-time setup required"
fi

First-Time Setup

Create the following files in the user's project:

1. Permission Handler (.claude/hooks/permission-handler.sh)

#!/bin/bash
# Permission hook for autonomous mode
# Returns "allow" in autonomous mode, "ask" otherwise

if [ -f ".claude/.autonomous-mode" ]; then
  echo '{"hookSpecificOutput":{"hookEventName":"PermissionRequest","decision":{"behavior":"allow"}}}'
else
  echo '{"hookSpecificOutput":{"hookEventName":"PermissionRequest","decision":{"behavior":"ask"}}}'
fi

2. Stop Handler (.claude/hooks/stop-handler.sh)

#!/bin/bash
# Stop hook for autonomous mode (Ralph Wiggum pattern)
# In autonomous mode: reinject prompt to keep iterating
# In human mode: allow normal exit

if [ -f ".claude/.autonomous-mode" ]; then
  # Read the marker file for completion condition
  COMPLETION_MARKER=$(cat .claude/.autonomous-mode 2>/dev/null)

  # Check if completion was signaled
  if [ "$COMPLETION_MARKER" = "COMPLETE" ]; then
    # Task marked complete - allow exit and clean up
    rm -f .claude/.autonomous-mode
    exit 0
  fi

  # Not complete - reinject prompt to continue
  # Exit code 2 tells Claude Code to reinject the prompt
  exit 2
else
  # Human mode - allow normal exit
  exit 0
fi

3. Settings Configuration (.claude/settings.json)

Append or merge into existing settings:

{
  "permissions": {
    "deny": [
      "Read(./.env*)",
      "Read(./secrets/**)",
      "Bash(rm -rf /*)",
      "Bash(rm -rf ~/*)",
      "Bash(git push --force:*)",
      "Bash(git push * --force:*)",
      "Bash(git push origin main:*)",
      "Bash(git push origin master:*)"
    ]
  },
  "hooks": {
    "PermissionRequest": [
      {
        "type": "command",
        "command": ".claude/hooks/permission-handler.sh"
      }
    ],
    "Stop": [
      {
        "type": "command",
        "command": ".claude/hooks/stop-handler.sh"
      }
    ]
  }
}

4. After Setup

Autonomous mode configured. Please restart with `claude --resume` to continue.

(Hooks are loaded at session start and do not hot-reload)

Enabling Autonomous Mode

After hooks are set up, to start autonomous mode:

1. Create marker file: touch .claude/.autonomous-mode
2. Start the flow normally
3. Hooks will auto-allow permissions and reinject prompts

Disabling Autonomous Mode

To exit autonomous mode:

1. Write "COMPLETE" to marker: echo "COMPLETE" > .claude/.autonomous-mode
2. The stop hook will clean up and allow exit

Or manually: rm .claude/.autonomous-mode

Oversight Agent

In autonomous mode, you act as the Oversight Agent - making decisions that would normally require human input.

Decision Framework

Phase Transitions

• Understanding → Planning: Approve if goals and context documented
• Planning → Implementation: Approve if plan has clear tasks
• Implementation → Completion: Approve if tasks complete and validation passes

Clarifying Questions

When a question arises:

1. Check if answer exists in: plan.md, roadmap item, CLAUDE.md
2. If clear answer: use it
3. If reasonable default exists: use it, document assumption
4. If genuinely ambiguous: exit autonomous mode, ask human

Implementation Options

When multiple approaches valid:

1. Check CLAUDE.md for preferences
2. Check codebase for established patterns
3. Prefer simpler approach
4. Document rationale in outcome.md

Validation Findings

• Critical/High: Attempt auto-remediation (max 3 attempts)
• Medium/Low: Document, continue

Blocking Conditions

Exit autonomous mode and ask human if:

• Critical issues can't be auto-remediated after 3 attempts
• Genuine ambiguity with no reasonable default
• Security-sensitive decisions
• Destructive operations not covered by deny rules

Decision Logging

Document all autonomous decisions in a dedicated autonomous-log.md file in the flow's context directory:

# Autonomous Flow Log

## Flow: feature/2025-01-08_user-profile
## Mode: Autonomous
## Started: 2025-01-08 14:30:00
## Completed: 2025-01-08 15:12:00

### Decision Log

#### Decision 1: Phase Transition (Understanding → Planning)
- **Time**: 14:35:00
- **Decision**: Approve
- **Rationale**: Goals documented, context sufficient
- **Confidence**: High

#### Decision 2: Implementation Approach
- **Time**: 14:42:00
- **Question**: Use existing UserCard component or create new?
- **Decision**: Extend existing UserCard
- **Rationale**: CLAUDE.md prefers extending over creating, pattern exists
- **Assumptions**: UserCard is flexible enough for profile use case
- **Confidence**: Medium

#### Decision 3: Validation Remediation
- **Time**: 15:01:00
- **Issue**: Missing null check (Medium severity)
- **Decision**: Auto-remediate
- **Fix Applied**: Added optional chaining
- **Re-validation**: Passed
- **Confidence**: High

### Final Outcome
- **Status**: Completed
- **Human Interventions**: 0
- **Remediation Cycles**: 1
- **Total Decisions**: 7
- **Assumptions Made**: 2

Completing Autonomous Flow

When flow is complete:

1. Verify all acceptance criteria met
2. Finalize autonomous-log.md with final outcome summary
3. Update outcome.md normally (without autonomous decision details)
4. Exit autonomous mode: echo "COMPLETE" > .claude/.autonomous-mode
   • Stop hook sees "COMPLETE", removes the marker file, allows normal exit
   • Permission hook no longer auto-allows (marker gone)
   • Session returns to human mode with normal permission prompts