A 7,000-character quality check read a one-line quota message as a short article

The quality check had no way to tell a truncated body from an exhausted quota. `article_quality_ok` asks whether the body clears MIN_CHARS of 7,000 and whether it carries at least two affiliate links ^[3]. A one-line limit notice fails both tests. The reason handed back to the loop was "quality NG reason: not enough characters" ^[4]. That is accurate and unusable. At the byte count a truncated generation and an exhausted plan look identical, and only one of them gets better on the next pass ^[15]. The post's author wrote that a retry loop needs more than the bare fact of failure ^[12].

The loop already held the evidence. `RAW` comes back from `generate_article_raw` as a string before `write_article_file` touches the disk. The fix reads it there: a case-insensitive `grep -qiE` for limit wording, and a `break` out of the pass loop on a match ^[5]. When `QUOTA_HIT` is set, the script deletes the half-written output, removes the images it downloaded for that product, posts a Discord notification and exits 1 ^[7].

Three calls a run across four failing runs is 12 `claude -p` calls that produced no article ^[13]. The patched path would have spent 4.

What the match tests is the provider's prose. The pattern contains the literal strings "usage limit reached" and "rate limit", plus a reset-time form that matches things like "resets at 3pm" ^[6]. Reword the notice upstream and the grep stops firing, silently, and the loop goes back to three passes. The sturdier signal is already in the report: the limit message arrives as one line with no `TITLE:` line ^[2].

That distinction matters more for the siblings than for note2. `article-daily-stock.sh` generates the article series this post belongs to ^[10]. It writes technical posts about quota handling, and their bodies contain the phrase "usage limit reached" for the same reason this one does. A case-insensitive substring match over a whole generated body would abort on its own subject matter. So before the pattern is copied across, it needs scoping to the first line, or replacing with the shape test.

`run_pass()` in `series-daily-stock.sh` is the closest structural match to what note2 tripped over. Three attempts, each a `timeout 600` on `claude -p`, a WARN on empty output, sleeps of 30 then 60 seconds between attempts ^[11]. Inside a limit window that is up to 1,890 seconds, about 31 minutes, of a scheduled job waiting on a body the provider will not return ^[14]. The grep across all three siblings found no detection pattern in any of them ^[9].

One grep per pass is the cheapest operation in a script that shells out to a language model three times. The post does not report patching the other three.