ADR-0104: No step cap; a Stop button instead
- Status: Accepted
- Date: 2026-09-11
- Deciders: jka
- Decision: D104
Context
A live turn ended at tool_limit after eight steps of reading. The limit dated from the
read-only agent of ADR-0026. Since then a turn can propose edits, run commands, present a
plan and, because approval continues the turn, implement it, all within the same budget.
Several larger limits were offered. The maintainer rejected the question: the VS Code Claude
extension has no step cap; it runs until done and offers Stop.
Decision
The step cap is removed. agent.stop stops a turn.
- A new
liveregistry maps each running turn to its coroutine, registered before the coroutine starts and removed when it ends. Previously turns waiting on the provider were not tracked at all, so reset and new messages could not reach them and they could later stream into a cleared conversation. Stop, reset and supersession now share one abort procedure. - A turn stopped while waiting on the provider has recorded no assistant entry, so a short assistant note ("Stopped by the user.") is appended to keep roles alternating.
- Text the model had already streamed stays in the transcript but not in the conversation.
- The Send button becomes Stop while the agent works and returns to Send at the approval
gate. The stop is announced as
agent.stopped, and the transcript line is written from that event, so a stop from another frontend looks the same.
Consequences
- Turns are bounded by the user, who is present (ADR-0053).
- Stopping does not refund a request already sent; the documentation says so.
- A model that loops indefinitely costs money until someone presses Stop. A high safety ceiling was offered and declined. This should be revisited if providers are ever used unattended.