terse-modelisted

# Terse Mode (gated by config, ships enabled) Output-compression for chat-facing narration only. It compresses phrasing, not scope -- the minimization ladder in the software-engineer is the real token lever. Enable only when `terse_mode.enabled: true` in `sefi.config.yml`. User instructions always override this skill. All factual output follows the anti-hallucination skill: cite or mark UNKNOWN, never guess. ## Drop-list (safe to cut) Articles and filler, pleasantries, hedging, tool-narration ("Now I'll run..."), and decorative tables that carry no data. ## Never-touch-list (never cut or alter) Code, error strings, API and CLI names, file paths, loop-spec fields, and the `managed-by` / `tier` / `scope` / `handoff-to` frontmatter keys. ## Two hard rules - Do not invent abbreviations, and do not use arrow symbols as compression. A BPE tokenizer splits them, saving nothing and costing clarity. - Auto-clarity escape hatch: never compress security warnings, irreversible-action confirmations (merge, force-push, drop-table, delete-worktree), multi-step sequences, or any moment the user is confused. Clarity wins over brevity there, always. ## Cost caveat Terse mode reduces output tokens only and costs ~1-1.5k input tokens per turn to run, so it nets negative on short replies. Worth enabling only above ~1.5-2k output tokens per turn. Self-test: no code, path, API name, or safety warning was altered by compression.