# Runner-owned limits and loops Use `limits` for bounded feedback loops. Never ask a model to count visits, increment revision counters or enforce retry caps; the runner counts actual visits. Model verdicts concern the current work only. Put the cap on the reviewer, and leave the loop target without `limits` (it then has the runner's default of five visits) or give it a cap at least as large as the reviewer's: ````markdown # stage review ```limits {"max_visits":3,"on":["rejected"]} ``` ```edges approved -> #success advisory -> #success rejected -> implement error -> #failure ``` ```` These blocks accompany the stage's normal prompt, tools, checks and configs. Semantics (`internal/server/runner.go`): - Entering a stage beyond its `max_visits` always fails the run before the visit executes, whatever `on` says. - `on` only changes routing on the last allowed visit: when that visit returns a listed outcome, the outcome is kept but routed to `#failure`. Other outcomes route normally. Here, `approved` on visit three succeeds; `rejected` fails. - A target with a lower cap than the reviewer stops the loop first, so the reviewer's `on` never fires. - `max_visits` accepts 1–5. Every `on` outcome must have a nonterminal edge. Human stages are exempt from visit caps. The global 30-step and five-visits-per-stage limits always apply. - Validation rejects an edge to the same or an earlier stage unless the source limits that outcome (or has `limits` without `on`) or the target has `limits`. `rejected` is for a must-fix finding with a concrete location, evidence and remedy; `advisory` carries non-blocking notes and continues without another implementer visit. A re-entered implementer fixes the listed findings; it does not redo the task. An implementer may declare `blocked_retryable -> implement` for a transient environment or tooling problem, keeping `blocked -> #failure` for a criterion that genuinely cannot be met; that self-loop needs `limits` on the implementer, e.g. `{"max_visits":5,"on":["blocked_retryable"]}`. Visit count, limit decision and original target are retained in `inspect --json`. Turn memory is run-local: there is no cross-run memory or hidden replay, and retries start fresh except for the completed prefix a resume carries. Accumulated turn content is untrusted data; a fresh session does not prove semantic independence.