sqlite-utils 4.0rc2: Written by Claude Fable, $149.25, Fixes Data Loss Bug

Simon Willison shipped sqlite-utils 4.0rc2, a Python library for SQLite, written mostly by Claude Fable. The total cost: about $149.25 in API fees over 37 prompts and 34 commits, changing +1,321 -190 lines across 30 files.
What Changed
The release candidate fixes a critical data loss bug in Table.delete_where() — previously, it never committed, leaving the connection in-transaction and poisoning all subsequent writes. Here's the reproduction from Fable's analysis:
db = sqlite_utils.Database("dw.db")
db["t"].insert_all([{"id": i} for i in range(3)], pk="id")
db["t"].delete_where("id = ?", [0])
# conn.in_transaction is now True
db["t"].insert({"id": 50})
db["u"].insert({"a": 1})
db.close()
# Reopen: rows are [0, 1, 2] — the delete, row 50, AND table u are all gone.
Willison calls this a “really bad bug” — one he’s glad Fable caught before the stable 4.0 release.
New Transaction Model
Every write method (insert(), upsert(), update(), delete(), delete_where(), transform(), create_table(), enable_fts(), etc.) now runs inside its own transaction and commits before returning. No manual commit() needed.
Two exceptions:
- Use
db.atomic()to group multiple operations atomically - If you open a transaction with
db.begin(), the library will never commit it — you control the commit
Python 3.12+ autocompat Support
Fable also ensured compatibility with Python 3.12's new sqlite3.connect(..., autocommit=True|False) mode. The library's automatic per-method transactions are designed for default connection mode, and Fable patched the test suite to pass under the new autocommit settings.
Review Process
Willison started the work on his iPhone via Claude Code for web, then switched to a laptop for final review via GitHub PRs. He says reviewing documentation edits first is an excellent way to understand what changed. A final review by GPT-5.5 caught additional edge cases — the full PR transcript is shared.
Who It's For
Python developers who use sqlite-utils for lightweight database ops and want a production-ready 4.0 release with proper transaction safety.
📖 Read the full source: HN AI Agents
👀 See Also

PocketTeam: A Claude Code Pipeline with Hook-Based Safety and Learning Agents
PocketTeam is a Claude Code pipeline that implements 9 safety layers at the tool-call level to block dangerous operations like writes to .env or rm -rf commands. The system includes an Observer agent that analyzes completed tasks and writes structured learnings to improve future agent performance.

PicoClaw Fails to Build F1 AI Agent, Burns $20 in API Credits
A developer attempted to build an F1 information bot using PicoClaw on a Raspberry Pi Zero 2W, but the tool defaulted to version 11, generated hallucinated Python code, and consumed $20 in DeepSeek API credits without producing a working solution.

Claude Code HUD: Terminal Dashboard for Monitoring AI Coding Sessions
claude-code-hud is a terminal dashboard that provides real-time monitoring for Claude Code sessions, showing context window usage, API rate limits, and file changes without requiring an IDE. Run it with npx claude-code-hud.

Claude Proposal: Scope Memory & Hermetic Instance Isolation
A formal proposal to Anthropic from a power user: global/local memory scoping and hermetic instance isolation for Claude, enabling deterministic, auditable sessions.