Per-session analytics
Every session in Confabulous has an analytics summary surfaced alongside the transcript. The summary panel renders one card per dimension; together they answer “what did this session cost, how long did it take, and what did it do?”
See it live on the demo instance.
What it shows
Section titled “What it shows”- Cost — estimated USD cost, computed from token usage and current model pricing.
- Tokens — input, output, and cache token totals, broken down by model.
- Duration — wall-clock duration, split into assistant response time and user active time.
- Tools — every tool the agent invoked, with call counts.
- Conversation shape — message count by role (user / assistant), plus a timeline visualization of when each message arrived.
How it’s computed
Section titled “How it’s computed”Analytics are derived from the transcript at sync time. No additional LLM calls are made — the numbers come straight from the structured session data the CLI uploads.
Model pricing tables live in the backend (backend/internal/analytics/pricing.go) and are kept in sync with the frontend display table (src/utils/tokenStats.ts) via a parity test. When a new model ships, both tables update together.
Related
Section titled “Related”- Organization analytics rolls the same numbers up across every user in your instance.
- Smart recap layers an LLM-generated narrative summary on top of the structural analytics.