A translation batch whose reply has the wrong cue count or numbering is now split in half and both halves are sent again, up to four times (down to 1/16 of the batch), before the translation fails. A mismatched reply is still never accepted. This covers both manual translations and the reconcile loop, since both go through arr_subs::translate::translate.
Worst case for a backend that mangles every reply is 31 requests per batch instead of 1.
Why
qwen-14b on the local llama server returned 145 cues for a 139-cue batch, failing a Coyote vs. Acme translation after 7.5 minutes. On a 59-cue batch it kept count.
Verification
just ci passes (830 tests). New tests cover a batch that only answers correctly at 2 cues (8 → 4 → 2, cues stay in order) and the split limit.
<!-- claude-projects-attribution: {"github_login":"naps62"} -->
_Requested by **naps**_
A translation batch whose reply has the wrong cue count or numbering is now split in half and both halves are sent again, up to four times (down to 1/16 of the batch), before the translation fails. A mismatched reply is still never accepted. This covers both manual translations and the reconcile loop, since both go through `arr_subs::translate::translate`.
Worst case for a backend that mangles every reply is 31 requests per batch instead of 1.
<details><summary>Why</summary>
qwen-14b on the local llama server returned 145 cues for a 139-cue batch, failing a Coyote vs. Acme translation after 7.5 minutes. On a 59-cue batch it kept count.
</details>
<details><summary>Verification</summary>
`just ci` passes (830 tests). New tests cover a batch that only answers correctly at 2 cues (8 → 4 → 2, cues stay in order) and the split limit.
</details>
🤖 Generated with [Claude Code](https://claude.com/claude-code)
qwen-14b returned 145 cues for a 139-cue batch and failed a whole movie
after 7.5 minutes. A mismatched reply is still never accepted, but the
batch is now halved and resent, up to four times, before giving up.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
Requested by naps
A translation batch whose reply has the wrong cue count or numbering is now split in half and both halves are sent again, up to four times (down to 1/16 of the batch), before the translation fails. A mismatched reply is still never accepted. This covers both manual translations and the reconcile loop, since both go through
arr_subs::translate::translate.Worst case for a backend that mangles every reply is 31 requests per batch instead of 1.
Why
qwen-14b on the local llama server returned 145 cues for a 139-cue batch, failing a Coyote vs. Acme translation after 7.5 minutes. On a 59-cue batch it kept count.
Verification
just cipasses (830 tests). New tests cover a batch that only answers correctly at 2 cues (8 → 4 → 2, cues stay in order) and the split limit.🤖 Generated with Claude Code