TPS tracker
The TPS tracker is a bundled extension that shows live tokens-per-second (tok/s) for the main model and subagent runs in the status area. Live status stays approximate; final main-model TPS is computed from provider output usage with reliability gates against stalls, short streams, and tool calls.
Bundled with Selesai
Loaded automatically with Selesai; no per-extension npm install is required.
Setup and prerequisites
- Bundled with Selesai and loaded automatically; no installation or configuration is required.
- The status line appears in interactive TUI sessions; headless runs still record measurements but have no status surface to draw on.
What it sets up
- Shows live tok/s for the main model while it streams, using provider output usage when available and an estimate otherwise.
- Tracks subagent runs separately: live per-agent tok/s during execution, a summary after each run, and aggregated subagent totals.
- Applies pi-tps-style reliability gates: stalls, very short streams, sparse updates, and implausible rates are filtered out of the final measurement.
- Marks tool-call-bound measurements so they do not count as clean main-model TPS.
What you can configure
The TPS tracker is automatic: it listens to message and tool events, shows live tok/s in the status area, and needs no configuration.
This capability does not expose user-editable settings. Its behavior is controlled by the commands and runtime rules documented below.
Source evidence
src/extensions/tps.tsTPS tracker source
There are no user settings: thresholds are fixed constants in the extension source.
What you can do
- Watch main-model throughput live while Selesai answers.
- See per-subagent throughput while delegated agents run, including parallel and chained modes.
- Read the final reliable TPS summary at the end of a generation or subagent run.
Commands, tools, and shortcuts
No slash commands or shortcuts are registered. The extension only writes the tps status slot.
Limits and safety
- Live numbers are estimates by design; the final main-model figure is the only reliability-gated value.
- Rates are capped at a plausible maximum and dropped when stream quality does not meet the gates.
- The extension never reads prompt or message content — only timing and usage counters.