Interaction modes

Context and Problem Statement

Arbor exposes a ui.interaction_mode setting — auto, direction, review, collaborative — that decides how much the operator is asked before the agent acts. The 2026-07-23 jobs review (reviews/2026-07-23-jobs/) proposed the same knob for the kit: one configuration value naming how autonomous a run should be, read by every phase.

The request recurs under several names because the underlying wish is common and reasonable — an autonomy level, a supervision setting, a --interactive / --yolo pair, a confirmation policy, a human-in-the-loop mode. Each arrives as “the kit should let me choose how much it asks me”.

Recorded as French prose in docs/backlog.md §Non-adoptions until this ADR replaced it.

Considered Options

  • A ui.interaction_mode configuration value with four levels, read by every phase.
  • A per-invocation flag with the same four levels.
  • Decline: the entry points already differ by autonomy, and that is the axis.

Decision Outcome

Not pursued by decision (2026-07-23): the kit’s scope variants already are the interaction modes, and they are better ones because each is a different job rather than the same job with a different confirmation policy. /migrate-assess is read-only, /migrate runs the pipeline, /migrate-verify re-runs the final gate; triage-backlog proposes and executes only what the owner confirms, while the lifecycle skills run hands-off (ADR-0005). Adding a fourth orthogonal concept on top of those would mean every phase reading a mode it mostly ignores, and two ways to express “ask me first” that can disagree.

Consequences

Autonomy is chosen by picking an entry point, not by setting a value — so the choice is visible in what the user typed, and a skill can never be half-way between two policies.

The one condition that reopens this: a genuine need the entry points cannot express — a level of supervision that is not “which job am I running” but “how much do I trust this particular run” — demonstrated by a case the existing variants leave with no answer. Reopening is a fresh ADR plus set_status deprecated here, never an edit in place.

Prior requests

  • reviews/2026-07-23-jobs/ — “Modes d’interaction (ui.interaction_mode auto/direction/review/ collaborative)”: a configurable autonomy level, supervision setting, interactive/yolo flag pair or human-in-the-loop confirmation policy, read by every phase (2026-07-23).