was playing around with AgentProbe on a longer cursor session and noticed one observability gap, once the observer is stopped, the emitted events are gone unless the consuming app records them itself
for debugging longer sessions, it would be useful to have an optional way to persist the event timeline and replay it later, like for ex, questions like "why did the agent keep touching this file for 10 minutes?", "what status transitions happened before the error?" or "what changed between two agent states?" are hard to answer from live events alone
one possible direction could be a structured session log format, maybe jsonl/ndjson where each emitted event is recorded with a timestamp, provider, session/source path, event type and payload, that would make it possible to build replay, timeline export, diffs or even simple debugging tools on top of AgentProbe
i was not sure about two design points-
- should this live in core as an optional recorder api or should it be a separate entry point/package like
@agentprobe/recorder so the core observer stays minimal?
- for long-running observers, should AgentProbe define log rotation/retention behaviour or leave that completely to consumers?
i am happy to take a stab at this if it sounds useful.
was playing around with AgentProbe on a longer cursor session and noticed one observability gap, once the observer is stopped, the emitted events are gone unless the consuming app records them itself
for debugging longer sessions, it would be useful to have an optional way to persist the event timeline and replay it later, like for ex, questions like "why did the agent keep touching this file for 10 minutes?", "what status transitions happened before the error?" or "what changed between two agent states?" are hard to answer from live events alone
one possible direction could be a structured session log format, maybe jsonl/ndjson where each emitted event is recorded with a timestamp, provider, session/source path, event type and payload, that would make it possible to build replay, timeline export, diffs or even simple debugging tools on top of AgentProbe
i was not sure about two design points-
@agentprobe/recorderso the core observer stays minimal?i am happy to take a stab at this if it sounds useful.