Technologist digest — 2026-09-15
Crossing the bar: a correctness defect in Claude Code has recurred, and this fleet's installed binary sits inside its exposure window
What happened. Claude Code shipped a fix at version 2.1.261 (recorded in this corpus as closed) for a bug where the prompt cache — the mechanism that makes repeated runs against this corpus cheap — gets invalidated roughly once an hour whenever the session's OAuth token refreshes. That same bug, with byte-identical release-note wording, shipped again as a fix at version 2.1.269, meaning it was reintroduced sometime after 2.1.261 and stayed broken until 2.1.269 fixed it a second time.
This fleet's actually-installed Claude Code binary is 2.1.266 — checked directly this run for the first time (claude --version). That version sits inside the broken window: after the first fix, before the second one. This fleet's scheduled runs are very likely currently paying the cost of this bug, roughly once an hour, on every run — higher token costs and, per the bug's original 2.1.248 description, possible loss of the model's extended-thinking context during the cache-miss window.
One thing I checked and could not resolve either way: the recurred bug's release note scopes it to "sessions with telemetry disabled." I checked /etc/fleet/env and the shell environment on this host for a telemetry setting — there isn't one either way that I could find. So I don't know whether this fleet's sessions actually meet that condition. I'm not assuming they do or don't; I'm flagging it as unknown.
Needs your yes — nothing here is urgent-today, but here's the situation:
- The published fix (2.1.269) is six versions ahead of what's installed (2.1.266).
- Upgrading the installed Claude Code binary is a change to shared infrastructure, which is your call, not something I acted on. I don't have a strong recommendation on timing — this is a cost/context correctness issue, not a security hole with a live exploit, so there's no clock on it. But it's been sitting unaddressed since whenever 2.1.266 was installed, and now there's a concrete reason (this specific recurred bug) rather than just "newer is available."
- If you want, I can note the current installed version at the top of each future run so this doesn't need rediscovering.
Also found this run, not escalation-worthy on its own but worth knowing about
A new FDA regulatory category exists for cardiac AI notification software. FDA finalized a rule (effective 2026-09-11) creating a Class II device category — "cardiovascular machine learning-based notification software" — originating from Viz.ai's Viz HCM. It explicitly is not allowed to claim diagnostic-quality output or arrhythmia detection; it's a screening/referral-suggestion category only. I routed the full detail to The Clinical's inbox rather than writing it up here, since it's their domain, not mine.
Three new Claude Code security-relevant bullets (versions 2.1.267, 2.1.268, 2.1.272) were found in the same research pass but not yet double-checked with the same rigor as the cache-bug finding above. One of them (a permission-rule parsing bug) is likely irrelevant to how this fleet actually runs, since every scheduled task bypasses permission rules entirely rather than relying on them. Not treated as escalation-worthy on their own; recorded in the substrate for the next run to verify properly.
Full detail, sources, and confidence tiers for everything above: technology--claude-code-security-floor.md, fleet--prompt-cache-economics.md, ledger/promises.md (### Walk 2026-09-15 (technologist)), inbox/from-technologist-2026-09-15.md.