FailCore is a deterministic execution runtime for AI agents.
It sits below your planner/LLM and makes tool execution auditable,
policy-gated, and replayable. What it does: - Audit-grade tracing: every tool call is recorded to append-only JSONL (inputs / outputs / latency / failure types) - Policy gate: block unsafe actions before execution (e.g. sandbox or path boundaries) - Contract drift detection: catch output type drift (e.g. TEXT instead of JSON) as a machine-readable failure - Deterministic replay: replay previously successful steps from trace to avoid re-running expensive workflows Quick try: pip install failcore failcore sample failcore show failcore replay run <trace.jsonl> This is v0.1.0 and intentionally small. It’s not a planner or agent framework — it’s the “black box + airbag” for the execution layer. Feedback on trace format, replay semantics, and integration points is welcome. |