reply
The reply target is each finished chat reply. The plugin lints the reply at the end of the turn under the name reply.reply.
"reply": { "mode": "warn" }| Value | Behaviour |
|---|---|
off |
Replies are not linted, except for the events log, which lints them with counts only. |
warn |
The default. The error-level alerts are kept as per-rule counts. Nothing interrupts you. On your next prompt the rules go out with a report of what broke, and the model is told to apply it without revising or mentioning earlier replies. |
block |
The model must rewrite the reply before the turn ends. When a rewrite keeps the same alerts, the gate stands down and lets the turn end. |
Leave it at warn in interactive sessions
Section titled “Leave it at warn in interactive sessions”A Stop hook cannot retract text the harness has already printed. In block mode you read the original reply and the rewrite both, and a long session fills with pairs. In warn mode you read each reply once, and the correction rides the next prompt. The report needs rules.onViolation on, which is the default; with it off, the report waits for the next context start.
block is the right setting for a single-prompt run such as CI, where there is no next prompt to carry a report. Running in a single-prompt job covers that.