The rate limit's real cost is the wiped session, and 100 minutes of retries will not cover it

The interesting decision in this design is where it puts the memory. A session transcript is a side effect of a running process, and Lily's account is precise about what dies with that process: `--continue` restores the record of the last exchange, but not the intent behind it, because the only instance that knew "next I'll do X" is the one that got killed ^[6].

So the wrapper does not try to restore a mind. It restores a file. That is why the hardcoded prompt matters more than the retry loop. The first invocation uses whatever task string you passed on the command line; every attempt after that sends the same sentence, "read PROGRESS.md and continue the interrupted work," at line 27 ^[8][9]. Nineteen of the twenty attempts are therefore worth exactly what the file is worth ^[15]. A PROGRESS.md that is three steps stale does not produce a stall; it produces a confident resume at the wrong place, unattended, with a model that has been told not to ask ^[10].

Then there is the arithmetic the post does not finish. Twenty retries at five minutes is 100 minutes of unattended coverage ^[5], against the 300 minutes of the five-hour block the piece opens with ^[1]. That is a third of the worst case ^[14]. Hit a genuine full-length block and the loop burns through all twenty attempts, logs max-retries-exceeded, exits 1 ^[12], and leaves roughly 200 minutes of the block still to run ^[17] with the session cold again and nobody awake.

The published log supports the same reading from the other direction. The rate limit hit around 2 AM, retries went out at 2:05 and 2:10, the third attempt succeeded, and the run continued to 6 AM ^[11]. That is about fifteen minutes of downtime ^[16], not five hours. Either that particular block cleared quickly or the abnormal exit was never a full block at all. Both readings point at the same fix: derive the sleep interval from whatever reset time the tool reports, rather than a five-minute constant repeated twenty times.

One note on provenance. The revenue figures in the post, including 1.2M yen a month within six months of a layoff ^[13], are self-reported and have no bearing on whether the exit-code check works. The mechanism stands or falls on its own.

The honest version of the argument is narrower than "the environment gets past the limit." The environment can own scheduling, permission handling and the notification. It cannot own the sentence that says what you were in the middle of. A human still writes that one, at every step, or the loop resumes into nothing.