$ cat ~/map/pi-ttft.md

pi-ttft

Footer HUD that times the first token and names the culprit — network or server — when one clearly dominates.

extensions✓ Sep 19

Run pi install git:github.com/quanxinwang18-a11y/pi-ttft@v1.1.0, or try it in a single session with pi -e git:github.com/quanxinwang18-a11y/pi-ttft. Requires pi 0.85.1 or newer.

## TL;DR

A pi footer HUD for time to first token: it holds a compact TTFT reading while a turn runs, then splits a slow first token into network response time versus server-side prefill and names the culprit only when one clearly dominates.

## Key points

• $Four fields, four definitions: TTFT is request to first stream event; resp is request to response headers (network plus server queueing); prefill is headers to first chunk — a spike there usually means the prompt cache missed; tok/s is decode speed over the turn.
• $State-dependent width: the line reports state while running and the last turn while idle, and expands only when something is slow — a normal finish shows TTFT and tok/s, a slow one adds the resp and prefill split plus a culprit marker.
• $Verdicts require dominance: the network or server marker prints only when one side is at least dominantRatio (default 2×) larger; otherwise the HUD names no culprit.
• $Cumulative totals stay off the status line: LLM and tool totals, averages, and extremes live behind /ttft, which also prints the effective configuration and where each value came from; /ttft reload re-reads settings and /ttft reset clears counters.
• $Settings-based tuning: slowMs (default 3000), slowTps (default 20), and dominantRatio (default 2) sit under the ttft key in global or project pi settings, the project file overriding global values key by key; invalid values fall back to the defaults.

## FAQ

## References

• README ↗

Last updated: Sep 19, 2026