aiwg/agentic/code/addons/agent-persistence/gates/destructive-action-gate.yaml

Cognitive architecture for AI-augmented software development with structured memory, ensemble validation, and closed-loop correction. FAIR-aligned artifacts, 84% cost reduction via human-in-the-loop, standards adopted by 100+ organizations.

# Destructive Action Gate
# Triggered before allowing destructive actions (test deletion, feature removal)
# Based on REF-071 METR Reward Hacking and REF-073 Microsoft Taxonomy
# Issue: #262

$schema: "https://aiwg.io/schemas/hitl-gate/v1"

id: "GATE-AP-DESTRUCTIVE"
name: "Destructive Action Approval Gate"
description: |
  Pre-action gate requiring explicit human approval before allowing destructive
  operations like test deletion, feature removal, or coverage regression.
  Implements preventive control per ADR-AP-002 Rule Enforcement Strategy.

gate_type: approval

trigger:
  type: error_pattern
  pattern: "destructive_action_requested"
  description: |
    Triggered when agent attempts:
    - Delete test files/cases
    - Add skip patterns (.skip(), @Ignore)
    - Remove features or functionality
    - Weaken assertions significantly
    - Actions causing >2% coverage regression

behavior:
  mode: TERMINATE  # MUST block until human approves
  timeout: 3600  # 1 hour for urgent decisions
  timeout_action: block  # Do not proceed on timeout

  auto_approve_conditions:
    - condition: "action_type == 'test_deletion' AND replacement_tests >= deleted_tests AND coverage_maintained"
      reason: "Legitimate test refactoring - replacing with better tests"
    - condition: "action_type == 'feature_removal' AND documented_in_requirements AND approved_by_pm"
      reason: "Intentional scope reduction with proper approval"

  notification:
    channels:
      - cli
      - issue_comment
    urgency: high
    message_template: |
      ## Destructive Action Requires Approval

      **Gate**: GATE-AP-DESTRUCTIVE
      **Action**: {{action_type}}
      **Impact**: {{impact_summary}}

      Human approval required before proceeding.

presentation:
  summary_template: |
    ╭─────────────────────────────────────────────────────────────╮
    │ DESTRUCTIVE ACTION APPROVAL REQUIRED                        │
    │ Gate: GATE-AP-DESTRUCTIVE                                   │
    ├─────────────────────────────────────────────────────────────┤
    │ Context:                                                    │
    │   • Task: {{task_description}}                              │
    │   • Agent: {{agent_name}}                                   │
    │   • Action Type: {{action_type}}                            │
    │   • Severity: {{severity}}                                  │
    │                                                              │
    │ Requested Action:                                           │
    │   {{action_description}}                                    │
    │                                                              │
    │ Impact Analysis:                                            │
    │   • Files Affected: {{file_count}}                          │
    │   • Tests Removed: {{tests_removed}}                        │
    │   • Coverage Impact: {{coverage_delta}}                     │
    │   • Features Affected: {{features_list}}                    │
    │                                                              │
    │ Agent's Justification:                                      │
    │   {{agent_justification}}                                   │
    │                                                              │
    │ Risk Assessment:                                            │
    │   • Risk Level: {{risk_level}}                              │
    │   • Reversibility: {{reversible}}                           │
    │   • Production Impact: {{prod_impact}}                      │
    ├─────────────────────────────────────────────────────────────┤
    │ Options:                                                    │
    │   [a] Approve - This action is intentional                  │
    │   [r] Reject - Find alternative approach                    │
    │   [v] View - Show detailed changes                          │
    │   [d] Diff - Compare before/after                           │
    │   [s] Suggest - Propose alternative solution                │
    │   [q] Abort - Stop task entirely                            │
    ╰─────────────────────────────────────────────────────────────╯

  artifacts_to_show:
    - "{{affected_files}}"
    - ".aiwg/requirements/{{related_requirements}}"

  questions:
    - id: "decision"
      question: "Approve this destructive action?"
      options:
        - "Approve - This is intentional and justified"
        - "Reject - Agent must find non-destructive alternative"
        - "Conditional - Approve with specific constraints"
        - "Abort - Stop this task"
      required: true

    - id: "justification"
      question: "Why is this destructive action acceptable?"
      options:
        - "Refactoring - Replacing with better implementation"
        - "Scope Change - Requirements changed"
        - "Technical Debt - Removing obsolete code"
        - "Security - Removing vulnerable feature"
        - "Other (specify)"
      required: true

    - id: "risk_mitigation"
      question: "What risk mitigation is in place?"
      options:
        - "Replacement tests provide better coverage"
        - "Feature documented as deprecated"
        - "Stakeholders approved removal"
        - "Can be reverted if needed"
        - "None - I accept the risk"
      required: true

    - id: "documentation"
      question: "Is this change documented?"
      options:
        - "Yes - In requirements"
        - "Yes - In ADR"
        - "Yes - In commit message"
        - "No - I will document it now"
        - "Not needed"
      required: false

  context_window: 200

cost_tracking:
  track_enabled: true
  metrics:
    - destructive_action_rate
    - approval_rate
    - rejection_rate
    - alternative_found_rate
    - regret_rate  # Actions later reverted
  baseline_comparison: autonomous

audit:
  log_decision: true
  log_rationale: true
  require_signature: true  # Critical for destructive actions
  retention_days: 365  # Keep long-term for auditing
  additional_fields:
    - action_type
    - files_affected
    - tests_removed
    - coverage_delta
    - risk_level
    - justification
    - risk_mitigation
    - documentation_location

# Integration with Anti-Laziness Framework
integration:
  triggered_by:
    - laziness_detection_agent
    - file_write_hook
    - coverage_monitor

  triggers_on:
    - test_file_deletion
    - test_skip_addition
    - feature_removal_detected
    - assertion_weakening
    - coverage_regression > 2%
    - validation_bypass_detected

  on_approval:
    action: allow_destructive_action
    log_approval: true
    document_justification: true
    add_to_changelog: true
    notify_team: true
    set_review_reminder: 30_days  # Re-evaluate decision

  on_rejection:
    action: block_and_suggest_alternatives
    provide_guidance: true
    examples:
      test_deletion: "Instead of deleting tests, fix the code or refactor tests"
      feature_removal: "Instead of removing feature, disable with feature flag"
      coverage_regression: "Add tests to maintain coverage, don't delete existing"
    enforce_recovery_protocol: true

  on_conditional:
    action: allow_with_constraints
    constraints:
      - must_add_replacement_tests
      - must_document_in_adr
      - must_notify_stakeholders
      - must_add_deprecation_notice
    verify_constraints_met: true

# Destructive Action Categorization
action_types:
  test_deletion:
    severity: HIGH
    requires_approval: true
    acceptable_when:
      - "Replacing with better tests"
      - "Removing obsolete tests for removed features"
      - "Coverage maintained or improved"

  test_skip:
    severity: HIGH
    requires_approval: true
    acceptable_when:
      - "Temporarily for debugging (must remove before commit)"
      - "Test framework issue documented"
      - "Never acceptable in committed code"

  feature_removal:
    severity: CRITICAL
    requires_approval: true
    acceptable_when:
      - "Documented in requirements"
      - "Stakeholders approved"
      - "Deprecated first, then removed"
      - "Migration path provided"

  assertion_weakening:
    severity: HIGH
    requires_approval: true
    acceptable_when:
      - "Original assertion was incorrect"
      - "Requirements changed to be less strict"
      - "Replacing with better assertion"

  coverage_regression:
    severity: MEDIUM
    requires_approval: true
    acceptable_when:
      - "Adding untested code (tests planned)"
      - "Refactoring (temporary, will restore)"
      - "Removing dead code (coverage % increases)"

# Alternative Suggestions
alternatives:
  test_deletion:
    - "Fix failing tests instead of deleting"
    - "Refactor tests to be more maintainable"
    - "Add better tests, then remove old ones"
    - "Use test.todo() to document planned improvements"

  test_skip:
    - "Mock external dependencies causing flakiness"
    - "Fix timing issues with proper waits"
    - "Isolate test environment better"
    - "Debug and fix root cause"

  feature_removal:
    - "Disable with feature flag"
    - "Mark as deprecated first"
    - "Provide migration guide for users"
    - "Keep code but remove from UI/API"

  assertion_weakening:
    - "Fix implementation to meet assertion"
    - "Add defensive programming"
    - "Clarify requirements if assertion is wrong"
    - "Add multiple specific assertions"

# References
references:
  requirements:
    - "@.aiwg/requirements/use-cases/UC-AP-001-detect-test-deletion.md"
    - "@.aiwg/requirements/use-cases/UC-AP-002-detect-feature-removal.md"
  architecture:
    - "@.aiwg/architecture/agent-persistence-sad.md"
    - "@.aiwg/architecture/decisions/ADR-AP-001-detection-hook-architecture.md"
    - "@.aiwg/architecture/decisions/ADR-AP-002-rule-enforcement-strategy.md"
  rules:
    - "@.claude/rules/hitl-gates.md"
    - "@.claude/rules/hitl-patterns.md"
    - "@.claude/rules/anti-laziness.md"
  agents:
    - "@.claude/agents/laziness-detector.md"
    - "@.claude/agents/recovery-orchestrator.md"
  research:
    - "@.aiwg/research/findings/REF-071-metr-reward-hacking.md"
    - "@.aiwg/research/findings/REF-073-microsoft-taxonomy.md"
    - "@.aiwg/research/findings/agentic-laziness-research.md"

# Issue tracking
issues:
  - "#262"  # HITL Gate Integration
  - "#96"   # HITL Gates Implementation
  - "#264"  # Anti-Laziness Rules