kaaroSessions is a live observability surface for your AI coding agents. It normalizes whatever transcripts are still on disk — seven harnesses, one record vocabulary, nothing leaves the machine. It does not outlive a retention sweep. The JSONL was never the tool. Now it isn't even permanent.
This is the hop: raw harness transcripts become one sixteen-kind vocabulary, then a pulse, then History / Live / Sonic. Nothing in the views reads JSONL.
Five stages, left to right. Click one (or [ ]) to see the selected kind at that stage in the inspector: raw transcript, adapter, NormalizedRecord, pulse, or the three views.
Rows are the sixteen kinds. Columns are harnesses. Green ● means that adapter emits the kind; · means the format has no such record. A cyan outline is the kind in the inspector. stream rows feed the live pulse; snapshot rows feed the session bundle.
The selected kind at the selected stage, for one harness. Chips with body-colored labels have a sample hop. H cycles harnesses. Empty hop = no sample for that pair, not a missing kind.
A canned turn. Space or ▸ play a turn walks kinds only: user_turn → tool_use → tool_result → content_block → tokens → context_reset. The pipeline stage stays where you left it. Click a chip to jump. Audio is opt-in (A).