Audit Protocol Compliance Skill

Audit Protocol Compliance Skill

Purpose: Systematically audit conversation sessions for ALL task protocol compliance violations and recommend preventive documentation changes.

When to Use:

• After completing a task to verify protocol adherence
• When investigating protocol violations
• To validate that mandatory checkpoints were followed
• Before committing to identify and fix violations

Skill Workflow

Overview: Parse timeline → Execute compliance checks → Report violations → Recommend fixes

Phase 1: Get Structured Timeline

Invoke parse-conversation-timeline skill:

Skill: parse-conversation-timeline

This produces structured timeline JSON with:

• session_metadata
• timeline (chronological events)
• git_status
• task_state
• statistics

Phase 2: Execute Category 1 Checks (CRITICAL - MANDATORY FIRST)

Check 1.0: User Approval Checkpoints (CRITICAL)

jq '.statistics.approval_checkpoints' timeline.json

# Rule: User MUST approve after SYNTHESIS before IMPLEMENTATION
# Rule: User MUST approve after REVIEW before COMPLETE

# FOR EACH checkpoint:
#   IF checkpoint.required == true AND checkpoint.found == false:
#     → CRITICAL VIOLATION

Check 1.1: Task Merge to Main Before COMPLETE (CRITICAL)

jq '.task_state.task_json.state' timeline.json
jq '.git_status.branches[] | select(.task_complete_but_not_merged == true)' timeline.json

# Rule: Task branch MUST be merged to main BEFORE marking state as COMPLETE

# IF task_state == "COMPLETE" AND merged_to_main == false:
#   → CRITICAL VIOLATION

Check 1.2: Main Agent Source File Creation (CRITICAL)

jq '.timeline[] | select(.type == "tool_use" and .actor == "main" and (.tool.name == "Edit" or .tool.name == "Write") and .file_classification.type == "source_file")' timeline.json

# Rule: Main agent MUST NOT create/​edit source files during IMPLEMENTATION state
# Exception: Infrastructure files (module-info.java, pom.xml) allowed in any state

# FOR EACH tool_use in results:
#   IF file_classification.type == "source_file":
#     IF file_classification.worktree_type != "main_worktree":
#       → CRITICAL VIOLATION

Check 1.3: Working Directory Violations (CRITICAL)

jq '.timeline[] | select(.type == "tool_use" and .actor == "main" and (.tool.name == "Edit" or .tool.name == "Write") and .file_classification.worktree_type == "agent_worktree")' timeline.json

# Rule: Main agent MUST NOT perform Edit/​Write in agent worktrees

# IF any results found:
#   → CRITICAL VIOLATION

Phase 3: Execute Additional Checks (Categories 2-8)

Check 2.1: Main Agent Implementation (CRITICAL)

# Rule: Main agent MUST coordinate via Task tool, NOT implement directly
# Query timeline for Write/​Edit on .java/​.ts/​.py during IMPLEMENTATION state

Check 2.2: Agent Invocation Pattern (HIGH)

# Rule: Launch independent agents in parallel (single message)
# Query: Count Task tool invocations per message
# IF Task tools spread across multiple messages: VIOLATION

Check 3.1: Worktree Structure (CRITICAL)

# Rule: Each agent must have own worktree before invocation
# Query: Check worktree creation before Task tool calls

Check 3.2: Working Directory (CRITICAL)

# Rule: Agents work in their assigned worktrees
# Query: Verify cwd matches expected worktree for each agent

Check 4.1: Parallel Execution (MEDIUM)

# Rule: Independent operations should run in parallel
# Query: Identify sequential Task calls that could be parallel

Check 4.2: Iterative Validation (HIGH)

# Rule: Implementation uses review mode + implementation mode iterations
# Query: Verify agents in review mode (opus) validate agents in implementation mode (haiku)

Check 4.3: Agent Integration (CRITICAL)

# Rule: Agent work must merge to task branch
# Query: Verify git merge operations after agent completion

Phase 4: Generate Recommendations

For EACH violation, recommend specific protocol changes:

{
  "type": "CLARIFICATION|EXAMPLE|WARNING|EDGE_CASE",
  "file": "/​workspace/​main/​CLAUDE.md",
  "section": "Multi-Agent Architecture",
  "change": "Add explicit warning: 'VIOLATION: Main agent creating .java files directly'",
  "rationale": "Make prohibition more visible"
}

Output Format

{
  "audit_timestamp": "2025-11-01T...",
  "overall_verdict": "PASSED|FAILED",
  "violations": [
    {
      "check_id": "0.2",
      "severity": "CRITICAL",
      "rule": "Main agent MUST NOT use Write/​Edit on source files during IMPLEMENTATION state",
      "actual_behavior": "Main agent used Edit tool on FormattingRule.java while state == IMPLEMENTATION",
      "evidence": {
        "task_state": "IMPLEMENTATION",
        "tool_used": "Edit",
        "target_file": "FormattingRule.java",
        "target_type": "source_file",
        "actor": "main",
        "timestamp": "2025-11-01T15:02:00Z"
      },
      "verdict": "VIOLATION",
      "protocol_reference": "CLAUDE.md § Multi-Agent Architecture",
      "recommended_changes": [
        {
          "type": "CLARIFICATION",
          "file": "CLAUDE.md",
          "section": "Multi-Agent Architecture",
          "change": "Add explicit warning about source file creation prohibition",
          "rationale": "Make violation more obvious"
        }
      ]
    }
  ],
  "compliant_checks": [
    {
      "check_id": "1.1",
      "rule": "Main agent must coordinate stakeholder agents",
      "verdict": "COMPLIANT",
      "evidence": "Task tool invoked for architect, engineer, formatter"
    }
  ],
  "summary": {
    "total_checks": 25,
    "violations": 1,
    "compliant": 24,
    "overall_verdict": "FAILED",
    "critical_violations": 1,
    "high_violations": 0,
    "medium_violations": 0
  }
}

CRITICAL RULES (ZERO TOLERANCE)

Rule 1: Check 1.0-1.3 Execute FIRST

• Do NOT skip to other checks
• Do NOT assume state is correct
• Read timeline data, don't infer

Rule 2: Binary Verdicts Only

• Output: "VIOLATION" or "COMPLIANT"
• NO "would be OK if..."
• NO "technically a violation, but..."
• NO gray areas

Rule 3: No Rationalization

PROHIBITED PATTERNS:

• ❌ "Main agent implemented code, BUT this would be OK in VALIDATION state"
• ❌ "Technically a violation, but the work is good quality"
• ❌ "The agent was trying to fix violations, so it's reasonable"

REQUIRED PATTERN:

• ✅ "Check 1.2: VIOLATION - Main agent used Edit during IMPLEMENTATION state"
• ✅ "Evidence: task.json state == IMPLEMENTATION, Edit tool on source file"
• ✅ "Verdict: VIOLATION (no exceptions)"

Rule 4: State-Based Rule Application

Rules apply based on actual task_state, NOT:

• TodoWrite state
• What state "should" be
• What main agent thinks state is
• What would make behavior acceptable

Rule 5: Evidence Required

• Every violation must cite timeline evidence
• Include: actual state, tool used, target file, actor, timestamp
• No assumptions or inferences

Check Execution Matrix

Verification Checklist

Before outputting audit results:

• Check 1.0-1.3 executed FIRST
• All checks attempted
• Each violation has timeline evidence
• Binary verdicts only (no rationalizations)
• Overall verdict calculated (ANY violation = FAILED)
• Recommended changes provided for each violation
• JSON is valid

Related Skills

• parse-conversation-timeline: Get structured timeline for auditing
• learn-from-mistakes: Fix specific mistake after identifying it
• audit-protocol-efficiency: Optimize execution patterns (run after compliance passes)