Model-written wording

Replies, approval questions, failure explanations and recovery direction go through the model. Deterministic code validates, normalizes, enforces schemas, orchestrates retries and records diagnostics, and it may hand the model safe facts (failure stage, error code, attempted actions), but it does not compose sentences for people. When a failure needs a judgment, the runtime asks for structured output first and feeds that decision into the reply.

A real task failure always produces a reply. The failure notice is tried in order: written by the task's model, possibly rewritten by a review step, written by the local or device model, and last a raw error with secrets redacted, compacted to 600 characters (bluecollar's agentcontract/failure_notice.go). Replies are suppressed only for duplicates, cancelled output and messages from the bot itself. The exception is the exact /stop and /stop-all acknowledgement.