sqlite-utils 4.0rc2 AI-written release by Claude Fable

The release-blocker with real teeth was a transaction-handling flaw in delete_where(). The function issued its DELETE through a bare execute() call with no atomic() wrapper, which meant the connection was left in an open transaction with the write uncommitted ^[1]. Because the delete never committed, that delete and any writes that followed it could be silently rolled back when the connection closed, a quiet path to data loss rather than a loud error ^[1]. The inconsistency was not new: the delete_where() auto-commit behavior had been sitting in the tracker as GitHub Issue #159 since 2020 ^[2]. Related commit-ordering problems showed up in db.query(), which rejected non-row statements only after execute() had already auto-committed the write, and where INSERT...RETURNING only committed once the returned generator was fully exhausted ^[1]. The fix was to make delete_where(), optimize() and rebuild_fts() all route through db.atomic(), bringing them in line with the other write methods ^[1].

The more transferable insight is procedural, not about any single bug. Willison did not just let one model write the code; he had a competing vendor's model check it. OpenAI's GPT-5.5 was tasked with reviewing the changes since the last release candidate to confirm the changelog was up to date, and that audit is where release-blocker problems surfaced ^[1]. Willison describes this as a habit rather than a one-off, saying he has started routinely having Anthropic's best model review OpenAI's work and vice versa because, uncomfortably, it really does work ^[1]. The value here is that each vendor's model tends to catch what the other missed, so the pairing functions as a cheap second reviewer rather than an echo chamber. For a solo maintainer of a widely-depended-on library, that cross-check partly substitutes for the human code review a small project rarely gets.

The headline number is the estimated unsubsidized cost of about $149.25 for a release that moved 37 prompts, 34 commits and +1,321 -190 lines across 30 files ^[1]. Just as telling is the task-selection logic behind the spend. Willison deliberately outsourced the release notes to Claude Fable, arguing that release notes are exactly the kind of writing worth handing to an agent because they need to be boring, predictable and accurate, and he judged the generated notes better than what he would have written himself ^[1]. He kept the final design and release calls for himself. Part of the urgency was timing rather than pure ROI: Claude Fable was only going to remain available on Max subscriptions for a few more days, which pushed him to lean on it now ^[1]. Individual agent runs sometimes took 10 to 15 minutes, which he treated as a feature because it freed him to do other work while the model ground away ^[1].

Not everyone read this as a clean win. Hacker News commenters called them brutal bugs for a release candidate, a pointed reminder that AI-heavy authorship still shipped serious transaction defects into an RC ^[3]. A deeper worry in the same thread is that models never volunteer a clean bill of health: as one commenter put it, there is never a 'Nothing found, good to go', because the model will always give you what you ask for regardless of the truth ^[3]. Another commenter warned that this intensification of work will not be good for workers' health ^[3]. The sharpest disagreement played out on X over AI-written tests: Sentry's David Cramer wildly disagreed that cheap AI-generated tests amount to good engineering, arguing that churn on small changes signals bad test design, while Willison countered that test-code churn no longer carries the cost it once did when a human had to hand-write every test.