failproofai-sdk

Install the failproofai-sdk package with pip, call configure() once at startup, and wrap each run in a failproofai_sdk.agent() scope. Use tool_call() around tool executions and event.model_request/model_response around LLM calls; events are written to local .jsonl files for a collector to ship.

Sessions are defined as runs that emitted agent_start, so a run emitting only tool or model events produces zero sessions. Emit agent_start at the top of every run and agent_end at every exit, then confirm the events appear in the local spool directory.

Yes, adapters for all four frameworks ship inside the SDK wheel. Import your framework first, then call failproofai_sdk.instrument() to auto-detect and record sessions, tools, models, and errors with no call-site changes.

Contextvars propagate into asyncio tasks automatically but not into new threads, so worker threads start with no bound identity. Wrap the callable with failproofai_sdk.propagate(fn) before submitting it to a thread pool or executor.

The agenteye name on PyPI resolves to a stranded old CLI build, not the SDK, and installing it can replace a pre-rename SDK installation. Always install failproofai-sdk for instrumentation and install the CLI separately as fp-cloud-cli via pipx.

A run counts as failed only when outcome is one of failed, error, timeout, or rejected. Passing outcome="failure" is silently treated as a non-failure, so use outcome="failed" on the agent_end event.