pisama 0.1.0

Multi-agent failure detection for production AI systems

Pisama

Multi-agent failure detection for production AI systems. Runs locally, no API key needed.

Install

pip install pisama

Quick Start

import pisama

result = pisama.analyze("trace.json")
for issue in result.issues:
    print(f"[{issue.type}] {issue.summary} (severity={issue.severity})")

CLI

# Analyze a trace file
pisama analyze trace.json

# List available detectors
pisama detectors

Detectors

Pisama ships with 18 failure detectors that run locally:

Detector	Catches
loop	Exact, structural, and semantic loops
corruption	State corruption and invalid transitions
persona_drift	Persona drift and role confusion
coordination	Agent handoff and communication failures
hallucination	Factual inaccuracy in agent output
injection	Prompt injection attempts
overflow	Context window exhaustion
derailment	Task focus deviation
context	Context neglect in responses
communication	Inter-agent communication breakdown
specification	Output vs specification mismatch
decomposition	Task breakdown failures
workflow	Workflow execution issues
withholding	Information withholding
completion	Premature or delayed task completion
cost	Token and cost budget overruns
convergence	Metric plateau, regression, and divergence
repetition	Repetitive output patterns

Docs

Full documentation: docs.pisama.ai

License

MIT