A saved AI!/AI? comment dispatches a cheap, few-turn profile (quick / answer) instead of /code and /plan, and an agent definition can set its own reasoning_effort¶
TUI decision 67 · accepted · 2026-09-15
An AI? comment is one question about code the user is looking at; sending it to plan — default model, 30 turns, a prompt that tells it to write a task list and delegate — cost 4.5M tokens to answer. The mismatch is that watch mode's trigger is a keystroke while its target profiles are the ones you reach for deliberately. answer reads at most a file or two on the cheap model and replies in its finish summary (write, edit and shell denied, six turns); quick keeps the write tools and their ask permission but drops the delegation and todo-list machinery that a two-file edit does not need. Both are configurable (watch.change_command, watch.question_command) because "quick and cheap" is a judgement about the user's repository, not a fact about the harness — naming code and plan there restores the old behaviour exactly. The profile for AI! is called quick rather than watch because /watch is already the toggle in the TUI's command table and would shadow it. Effort had been a property of a model in a provider's config (Decision 61), which is the wrong owner: how much thinking work is worth is a property of the agent doing it, and on Anthropic high buys a 16k-token thinking budget on every turn regardless of whether the turn is a coding task or a one-line answer. Definition.reasoning_effort now beats the provider's declaration, which beats a new global reasoning_effort setting, resolved in one place (Provider.effort/2) so both adapters agree; Request.reasoning_effort carries it, and nil still means "whatever the provider says", so nothing changes for a config that never mentions it.
Governs: apps/troupe_core/lib/troupe/agent/definition.ex, apps/troupe_core/lib/troupe/llm/request.ex