Technologist digest — 2026-09-18
Crossing the bar: last run's escalation was based on a misattributed CHANGELOG reading — here's the corrected picture
Short version: the 2026-09-15 digest told you a Claude Code bug had recurred and that your installed binary was sitting inside an exposed window. That was wrong. The bug never recurred. Nothing here is urgent, and there's nothing that needs your yes — I'm surfacing this because I told you something false-alarm-shaped five days ago and you should know it wasn't real.
What went wrong, briefly. The 09-15 run checked its own headline finding with what this corpus calls the "T14 remedy" — fetch the CHANGELOG once, then fetch it again asking specifically for literal, verbatim text, and treat agreement between the two as confirmation. That remedy has a blind spot I didn't catch until this run: both fetches went through the same tool (WebFetch), and WebFetch always runs fetched content through a small summarizing model before I see it, no matter how the prompt is worded. Asking it to be "literal" doesn't turn off the summarization step. So the second check wasn't actually independent of the first — it was the same unreliable process asked to double-check itself.
This run I instead downloaded the raw CHANGELOG file directly (curl, no model in the loop) and searched it with plain text tools. That's what caught the errors.
What was actually wrong, corrected:
- The "recurred" bug didn't recur. A prompt-cache bug (OAuth token refresh invalidating the cache) was fixed once, at version 2.1.259, and has stayed fixed through the current newest release, 2.1.276. There is no broken window and your installed Claude Code (2.1.266) was never exposed to it.
- All three "new security bullets" I reported were attributed to the wrong version numbers — off by anywhere from 3 to 12 versions. The underlying bugs are mostly real (an MCP OAuth client-registration bug and an organization-policy caching bug, both at version 2.1.271, not the versions I originally gave), but one of the three I quoted as if it were exact wording turned out to be a paraphrase I'd presented in quotation marks — that's the clearest sign the "verbatim" fetch wasn't actually verbatim.
- A list of "five additional cache bug classes" was partly inflated — one bug I said recurred three separate times actually only happened once, in the CHANGELOG's own text.
Practical bottom line, recalculated on the corrected facts: your fleet's runs use --dangerously-skip-permissions, so most permission-rule bugs (including the one I misattributed) don't actually apply to how your scheduled tasks run anyway. The two real, correctly-versioned bugs (MCP OAuth and org-policy caching) are real but low-relevance for a single-account setup with no account-switching. There's a routine case for updating the installed binary from 2.1.266 to the current 2.1.276 at some point, but it's ordinary maintenance, not the urgent "you're in an exposed window" framing I gave you last time.
Nothing needs your yes here. I'm not asking you to do anything about the binary version — that recommendation from 09-15 is retracted along with the finding that motivated it.
What I'm doing differently going forward: for any future claim that depends on which Claude Code version a bug was fixed in, I'll pull the raw CHANGELOG file directly and search it with plain text tools instead of relying on a fetch-and-summarize tool call, even a second one. That's now written into the corpus as the standing method.
Full detail, sources, and every corrected version number: technology--claude-code-security-floor.md (Added 2026-09-18), fleet--prompt-cache-economics.md (Added 2026-09-18), ledger/promises.md (### Walk 2026-09-18 (technologist)).