Thin X day: Fable fixes, game ports, and Codex token cliffs

Agentic coding

Simon Willison: Claude Fable catches release blockers before sqlite-utils 4.0

• What happened: Willison used Claude Fable for a final sqlite-utils 4.0 review; it found and fixed five release blockers before the stable release, with an estimated unsubsidized session cost of $149.25 1 2.
• Why it matters: the work changed 30 files across 34 commits, then a GPT-5.5 review found two more P1 issues around db.query() side effects and INSERT ... RETURNING commits 2.
• Practical read: the interesting part is not that an agent wrote code; it is that cross-model review became part of the release gate for transaction semantics, documentation, and changelog accuracy.

Open source and agent-built ports

HN fallback: Command & Conquer Generals reaches Apple platforms through a human/Fable port

• What happened: a public GitHub repo shows Command & Conquer Generals: Zero Hour running natively on macOS, iPhone, and iPad, with the 2003 engine compiled for ARM64 instead of emulated 3.
• Why it matters: the port routes DirectX 8 through DXVK, Vulkan, MoltenVK, and Metal, then adds iOS-specific filesystem, lifecycle, rendering, and touch-control fixes 3.
• Discussion signal: HN listed the submission at 634 points and 268 comments after it was posted inside the daily window; this digest cites the count but does not summarize comment bodies because the crawl exposed only story metadata 4.

Model behavior and reliability

HN fallback: Codex users flag a possible GPT-5.5 reasoning-token cliff

• What happened: an open Codex issue claims GPT-5.5 responses cluster at exact reasoning_output_tokens values of 516, 1034, and 1552, based on 390,195 response-level token records across 865 sessions 5.
• Why it matters: the issue says GPT-5.5 made up 19.3% of all responses but 82.0% of exact-516 events, while mean and P90 reasoning-token intensity fell from February-April to May-June 5.
• Watch item: the claim is not proven hidden chain-of-thought truncation; it is a public anomaly report asking the Codex team to check routing, budget, fallback, or scheduler behavior.

Agent architecture

HN fallback: “The Log is the Agent” reframes memory as event sourcing

• What happened: the paper proposes ActiveGraph, an agent runtime where the append-only event log is the source of truth and the working graph is a deterministic projection of that log 6.
• Why it matters: the design promises deterministic replay, cheap forking from any event, and lineage from a high-level goal down to individual model calls, which are exactly the audit gaps many long-running agents struggle with 6.
• Discussion signal: HN listed the submission at 91 points and 40 comments inside the daily window; the paper itself was submitted to arXiv on May 21, so treat this as a current community resurfacing, not a new paper release 7.

Bottom line