Summary

Implements issue #25 - FailureAnalyzer class that systematically records and analyzes PR failures to help prevent repeated mistakes. This addresses the development focus on agent reliability and error recovery.

Key capabilities:

• Records failures to logs/failures.jsonl with structured metadata (issue ID, category, root cause, evidence)
• Categorizes failures into standardized types (missing_tests, ci_failures, scope_creep, etc.)
• Provides similar failure matching so Engineers get warned about past mistakes before starting work
• Generates failure pattern analysis for Auditor health reports
• Configurable backoff policy to avoid reworking issues that repeatedly fail

Evidence

• Linked issue: Agent failure analysis and adaptive recovery system #25 (Failure pattern tracking and analysis)
• Repro: Currently when PRs fail, there's no systematic tracking of why they failed or mechanism to warn Engineers about similar past failures

Dependencies

• Lockfile changes: NO
• Dependency changes: NONE

Changes

New Files

• src/barbossa/utils/failure_analyzer.py (663 lines) - Core FailureAnalyzer class with:

  • record_failure() - Record failures with structured metadata
  • get_similar_failures() - Find matching past failures by keywords/labels
  • get_failure_warnings() - Formatted warnings for Engineer prompts
  • should_skip_issue() - Backoff policy check for repeatedly failing issues
  • analyze_failure_patterns() - Pattern analysis for Auditor reports
  • rotate_failures() - 90-day retention cleanup

• tests/test_failure_analyzer.py (542 lines) - Comprehensive test suite with 23 tests

Modified Files

• src/barbossa/agents/tech_lead.py - Records failures when closing PRs
• src/barbossa/agents/engineer.py - Queries failure history before starting issues
• src/barbossa/agents/auditor.py - Includes failure patterns in health reports
• prompts/engineer.txt - Added {{failure_warnings_section}} placeholder
• config/repositories.schema.json - Added failure_analyzer settings schema

Testing

• All 23 new tests pass: PYTHONPATH=src python -m unittest tests.test_failure_analyzer -v
• Agent imports verified: All three agents (Tech Lead, Engineer, Auditor) import successfully
• Existing tests unaffected (pre-existing import issues in older tests are separate)

Configuration Example

{
  "settings": {
    "failure_analyzer": {
      "enabled": true,
      "retention_days": 90,
      "backoff_policy": {
        "skip_runs_after_failures": 1,
        "consecutive_failures_threshold": 2
      }
    }
  }
}

🤖 Generated with Claude Code