T — Terse mode

Output-token reduction. Two layers: a system-prompt injection that asks the model to strip filler, and a post-filter that strips filler phrases from tool outputs before they reach the model.

System-prompt injection (layer 1)

When T is enabled for a pane, the MCP session prepends this message to every tool result:

// Injected before every tool result when T is on
CRITICAL: You are in TERSE MODE. ALL responses must minimize
output tokens ruthlessly. No preamble, no summaries, no apologies,
no "I'll help you with", no "Let me know if you need anything else."
Respond only with exactly what was asked. Code blocks pass through
unchanged — only prose is compressed.

The model receives this before processing each response and adjusts its output accordingly.

Algorithmic minimization (layer 2)

A second layer strips filler phrases from tool outputs before forwarding them — removing the noise before it reaches the model, not after. Code blocks are passed through unchanged.

Stripped phrase patterns (case-insensitive): "It looks like", "I'll help you with", "Certainly!", "Of course!", "Let me know if", "I hope this helps", "Feel free to", "Is there anything else", "Happy to help", "I'd be happy to".

Synergy

Terse works synergistically with F/C/S/I: those reduce input tokens; T reduces output tokens. Combined effect on a typical session: 40–65% total token reduction vs no compression.

Env var

SENKANI_TERSE=on|off — default off. Enable it when you want tight, machine-readable responses; disable it when you're pair-programming and want the model's explanations intact.