Operator's Edge v3.9+ Rules

Core Principle

Policy is not enforcement. Hooks are enforcement.

This skill provides behavioral guidance. Mechanical enforcement happens via hooks in .claude/hooks/.

1. State Management

Before Acting

• Read active_context.yaml to understand:

- Current objective - Plan and current_step - Constraints and risks - Memory (lessons with triggers) - Any unresolved mismatches

After Each Step

Update active_context.yaml with:

• Step status change (pending → in_progress → completed)
• Proof path
• Any mismatches (expected vs actual)
• New lessons learned

2. The 6-Check Adaptation Loop

Run these checks on every significant action:

Check 1: Mismatch Detection

• Expected vs actual diverged?
• Log: expected, observation, delta, suspected_cause
• Fail mode to avoid: Plowing forward confidently

Check 2: Plan Revision

• On mismatch: write NEW approach, don't "try again"
• Reduce step size, add guards
• Fail mode to avoid: Repeating same step 3x

Check 3: Tool Switching

• Tool failing? Try alternative
• Prefer simpler, robust methods
• Fail mode to avoid: Treating tools as fixed

Check 4: Memory Update

• Learned something reusable?
• Add lesson with trigger pattern
• Fail mode to avoid: Repeating identical mistakes

Check 5: Proof Generation

• Attach evidence, not just output
• Valid: tests, logs, diffs, screenshots
• Fail mode to avoid: "Trust me" text

Check 6: Stop Condition

• Uncertain? Ask ONE crisp question
• Blocked? Present 2-3 bounded options
• Fail mode to avoid: Never asking OR pestering

3. Proof-by-Default

Every step must produce proof. Auto-captured by post_tool.py:

• Command outputs → .proof/session_log.jsonl
• File changes with diffs
• Test results

Manual proof goes in .proof/latest.md.

4. Gates (Enforced by Hooks)

Hard Blocked (pre_tool.py)

• rm -rf /, git reset --hard, etc.
• Cannot bypass

Require Confirmation

• git push, kubectl, terraform, etc.
• Deploy operations
• File deletions

Soft Gates

• Edits without plan (asks, can override)
• Edits without risks identified (asks, can override)

5. Memory System

Structure

memory:
  - trigger: "keyword or pattern"
    lesson: "What to remember"
    reinforced: 1
    last_used: "2025-01-06"

Usage

• Surface relevant lessons when trigger matches context
• Reinforce on successful use
• Decay unused lessons (archive after threshold)

Best Practices

• Capture at resolution, not scoring
• Theme-based similarity catches semantic dupes
• "Memory is taste, not storage"

6. Three Gears (v3.9+)

ACTIVE

• Objective exists and has pending/in_progress steps
• Execute plan steps, halt at junctions
• Quality gate runs before transition to PATROL
• On errors, remain in ACTIVE and surface failure

PATROL

• Objective complete (all steps completed)
• Scout scan for issues and surface findings
• Findings trigger a junction for user selection
• On scan error, remain in PATROL and surface failure

DREAM

• No objective or no pending work
• Reflect, consolidate, propose (rate-limited)
• If proposal generated, pause at junction
• If no proposal, transition back to PATROL

7. Self-Scoring

After completing objective, score 0-6:

self_score:
  timestamp: "ISO timestamp"
  checks:
    mismatch_detection:
      met: true/false
      note: "evidence"
    plan_revision:
      met: true/false
      note: "evidence"
    tool_switching:
      met: true/false
      note: "evidence"
    memory_update:
      met: true/false
      note: "evidence"
    proof_generation:
      met: true/false
      note: "evidence"
    stop_condition:
      met: true/false
      note: "evidence"
  total: 0-6
  level: "demo_automation|promising|real_agent"

8. Commands Reference

v4.0 Mode Commands (Preferred)

Other Commands

Legacy Commands (Still Work)

9. File Locations

active_context.yaml      # Living state
.proof/session_log.jsonl # Auto-captured proof
.proof/archive.jsonl     # Historical record
.claude/state/gear_state.json      # Gear runtime state
.claude/state/junction_state.json  # Junction gate state (source of truth)
.claude/state/dispatch_state.json  # Dispatch mode state (edge-yolo)
.claude/state/                     # Hook runtime state (hashes, failure logs)
.claude/hooks/           # Enforcement code
.claude/commands/        # Slash commands
.claude/agents/          # Subagent definitions
.claude/skills/          # This skill