Editing Openai/694057b6-101c-8007-9a65-c40578c7252d (section)

==== I’d spec it like this: ====
* Maintain last_line and repeat_count.
* As streamed text arrives, split into completed lines (buffer partial line until newline).
* Normalize a line before comparison: - strip trailing whitespace - optionally collapse runs of spaces/tabs - ignore empty/very-short lines (e.g. < 8 chars)
* If line == last_line: repeat_count += 1, else reset to 1 and update last_line.
* If repeat_count >= N and line length >= min_len, trigger: - cancel request (free slot) - mark terminal: FAILED(reason="repeated_line_loop") - attach loop_line (truncated) and repeat_count to debug/status

Conservative defaults that won’t trip on normal output:
* N = 8–12
* min_len = 32–64
* ignore lines that are just punctuation/whitespace
* optional: don’t start checking until min_generated_chars (e.g. 512) so you don’t kill harmless prompt echoes

This plus max_tokens is a very solid “early kill the worst derangements” combo.