AGENT SESSION INSIGHTS / VENN DESIGN
{{TITLE}}
Follow the complete visible conversation, measured model activity, and evidence-backed directions.
The recorded session
Drag to zoom · Scroll below to pan · Click to seekUsage details · API-equivalent estimates, not a bill
Input includes cached reads/writes. Prompt costs cover whole recorded turns, including later calls, independent of zoom. Visible-range usage follows time and harness filters. Session totals include undated calls; time-window subtotals do not.
Models · entire session
| Model | Scope | Calls | Tokens in | Tokens out | Estimated USD |
|---|
Spend & tokens by turn
One row per genuine human input, including logistics. Whole recorded turns, independent of zoom/filters. Recorded turn IDs take precedence; other ownership is inferred from response links or preceding input in the same conversation. Unmatched calls stay unattributed. Expand a turn's attribution for the method.
| Turn / attribution | Models | Calls | Tokens in | Tokens out | Estimated USD |
|---|
Pricing evidence & assumptions
USD per million tokens; long-context rules are applied per request to inclusive input.
| Model / verified date | Fresh in | Cache read | Cache write | Out | Long-context rule / write TTL |
|---|
Session text
First recorded input
Source and attribution
Prompt usage covers the complete recorded turn, not just the zoomed area or calls reached by the playhead. Text is preserved with best-effort credential redaction. Submitted material may contain pasted, AI-drafted plans. A timestamp marks submission, not continuous human effort.
Agent response
Full session log
Each user prompt includes its complete recorded turn's model, estimated cost and tokens. Agent rows do not repeat the charge.
Outside the clock: {{UNDATED}} undated messages
These messages have no reliable timestamps. Source order is preserved; dates are not inferred.