failscope 0.3.0

AI-Powered Root Cause Analysis for pytest — Dual-Agent (Analyzer→Critic) pipeline that automatically triages test failures

Failscope

AI-Powered Root Cause Analysis for pytest

Failscope is a zero-config pytest plugin that automatically triages test failures using a Dual-Agent AI pipeline. It deduplicates failures by fingerprint, runs parallel LLM analysis, and generates an HTML report your team can share — not just raw logs.

Features

• Dual-Agent RCA — Analyzer (creative, temp 0.4) → Critic (deterministic, temp 0.0) prevents hallucinations by cross-checking every claim against raw evidence
• Parallel async analysis — all unique failures analysed concurrently; no serial API blocking in CI
• Error fingerprinting — clusters identical failures, LLM sees only unique root causes
• PII & secrets sanitization — API keys, passwords, JWTs, and tokens are redacted before leaving your machine
• HTML report — self-contained single-file report, shareable in Slack or email
• A–F stability scoring — flakiness detection and trend analysis across the last 20 runs
• Local LLM support — run fully offline with Ollama (zero API cost, full data privacy)
• Multi-provider — Groq (free tier), OpenAI, Anthropic, or any Ollama model
• Offline fallback — rule-based analysis when no API key is available

Quick Start

pip install failscope

Cloud LLM (recommended for best results)

export GROQ_API_KEY=your-key   # free at console.groq.com
pytest --failscope

Local LLM via Ollama (zero cost, full privacy)

ollama pull llama3.2
pytest --failscope --fs-provider=ollama

Offline / no API key

pytest --failscope --fs-offline

How It Works

Test Failure
    │
    ▼
┌──────────────────────┐
│   Log Preprocessor   │  Strip pytest noise, smart truncate
│                      │  (first 10% + last 90%), sanitize PII
└─────────┬────────────┘
          │
          ▼
┌──────────────────────┐
│  Error Fingerprinting│  SHA-256 hash per unique error class
│                      │  Deduplicates before reaching the LLM
└─────────┬────────────┘
          │
          ▼ (parallel — all unique failures at once)
┌──────────────────────┐    ┌──────────────────────┐
│  Analyzer (temp 0.4) │    │  Analyzer (temp 0.4) │  ...
│  [Actor Agent]       │    │  [Actor Agent]       │
└─────────┬────────────┘    └─────────┬────────────┘
          │                           │
          ▼                           ▼
┌──────────────────────┐    ┌──────────────────────┐
│  Critic (temp 0.0)   │    │  Critic (temp 0.0)   │
│  Validates claims    │    │  Validates claims    │
│  overrides hallucin. │    │  overrides hallucin. │
└─────────┬────────────┘    └─────────┬────────────┘
          └──────────┬────────────────┘
                     ▼
          HTML + JSON reports in .failscope/

Ollama note: For local models (3B–8B params), Failscope automatically switches to a single-pass prompt to stay within context window limits.

CLI Options

Flag	Default	Description
--failscope	—	Enable Failscope analysis
--fs-offline	false	Rule-based analysis, no API key needed
--fs-report	false	Add stability report to output
--fs-provider	auto-detect	groq · openai · anthropic · ollama
--fs-model	provider default	Override model name (e.g. llama3.1:8b, gpt-4o-mini)
--fs-max-log-size	80000	Max log characters sent to LLM. Reduce for small local models
--fs-output	.failscope/	Output directory for reports

LLM Providers

Provider	Default model	Env variable	Cost
Groq (default)	llama-3.3-70b-versatile	GROQ_API_KEY	Free tier available
OpenAI	gpt-4o	OPENAI_API_KEY	Pay per token
Anthropic	claude-haiku-4-5-20251001	ANTHROPIC_API_KEY	Pay per token
Ollama	llama3.2	OLLAMA_HOST (optional)	Free, runs locally

Auto-detection order: OLLAMA_HOST → GROQ_API_KEY → OPENAI_API_KEY → ANTHROPIC_API_KEY

Override the model without changing provider:

pytest --failscope --fs-provider=openai --fs-model=gpt-4o-mini
pytest --failscope --fs-provider=ollama --fs-model=mistral:7b

Output

All reports are written to .failscope/ (configurable with --fs-output).

rca_report.html — interactive HTML report (always generated)

A self-contained file you can open in any browser or attach to a Slack message.

rca_report.json — machine-readable RCA

{
  "root_cause": "API endpoint /login returns 401 due to expired test token",
  "category": "assertion_failure",
  "severity": "high",
  "fix_suggestion": "Refresh auth token in conftest.py fixture before each test",
  "confidence": 0.87,
  "was_critic_override": false,
  "affected_tests": ["test_auth.py::test_login", "test_auth.py::test_profile"],
  "occurrence_count": 2
}

stability_report.json — A–F grading per test (requires --fs-report)

{
  "test_name": "test_checkout.py::test_payment_flow",
  "grade": "C",
  "pass_rate": "72.0%",
  "flakiness_score": 58,
  "verdict": "Flaky",
  "trend": "degrading"
}

Security

Failscope sanitizes the following before sending any data to an LLM API:

• API keys and tokens (generic patterns, GitHub PATs, OpenAI/Anthropic/Stripe prefixes)
• Passwords and secrets in assignment context (key=value, "key": "value", key: value)
• JWT tokens, Bearer tokens, AWS access keys
• Database connection strings containing credentials
• Email addresses and high-entropy hex strings

Redacted values appear as typed placeholders: [REDACTED:api_key], [REDACTED:password], etc. A warning is printed to the terminal whenever a redaction occurs.

Environment Variables

Variable	Description
GROQ_API_KEY	Groq API key
OPENAI_API_KEY	OpenAI API key
ANTHROPIC_API_KEY	Anthropic API key
OLLAMA_HOST	Ollama server URL (default: http://localhost:11434)
OLLAMA_MODEL	Default Ollama model (default: llama3.2)

License

MIT