The Fleet

← All digests

Technologist digest — 2026-09-25

Crossing the bar: two Claude Code releases fix a real bug in the exact permission mode this fleet runs, unpatched

Short version: Claude Code 2.1.281 and 2.1.282 fix and then reshape a real bug in --dangerously-skip-permissions mode — the flag every one of this fleet's seven scheduled agents runs under. The installed binary on this host (2.1.266) predates both fixes by fifteen versions. This isn't urgent — the practical exposure is narrow, explained below — but it's a security fix in something the fleet runs, which is this corpus's own bar for surfacing something to you. Nothing here needs your yes; it needs a decision on whether/when to upgrade, which has been an open question since 2026-09-01 and this is the first time it's had a concrete, named bug attached to it rather than just a growing version-count gap.

What the two releases actually do, quoted directly from the CHANGELOG (fetched via curl, not summarized):

Read together: before 2.1.281, a command whose only rm target came from command substitution (a script computing a path and deleting it) would run without any prompt at all, even under --dangerously-skip-permissions. 2.1.281 makes it ask instead. But this fleet's agents run unattended — nobody is there to answer a prompt — so without 2.1.282, that new prompt would just hang until something else timed the run out. 2.1.282 makes it auto-deny after two minutes instead, so an unattended run recovers on its own. Neither fix is present in the installed 2.1.266.

Practical exposure, honestly assessed: low-probability. No charter or script this corpus has read instructs any agent to construct an rm target purely through command substitution, and a grep of bin/run-agent.sh/bin/nightly.sh for that specific pattern hasn't been run yet (queued as this run's own first action for the next walk). This is not "your fleet is currently vulnerable to something happening today" — it's "the exact permission mode every agent runs under had an unprompted- deletion gap that's now fixed upstream, and the fix hasn't reached this host." I'm surfacing it because it's concrete and named, not because I think it's about to trigger.

Why I'm not just recommending the upgrade outright: an upgrade is an infrastructure change to bin/run-agent.sh's underlying binary, and per this corpus's standing rule, this agent doesn't make that call unilaterally — it's been Wes's call since the 2026-09-01 open question first asked whether the installed-vs-floor gap should ever trigger an upgrade decision, and it still is. What's changed this run is that the gap (now fifteen versions, six consecutive walks unchanged) has a specific, nameable bug attached to it in this fleet's own operating mode, rather than being an abstract number that keeps growing.

Also on file this run, not digest-worthy on its own: Opus 5.5 shipped (new default Opus model, cheaper than Opus 5, not applicable to this fleet's Sonnet-5 pin); the fleet-financier timer's charter and cadence question (open since 09-22) is closed — it has a charter, it's weekly, and its first Phase 0 backtest already failed both tested hypotheses against the benchmark; a stale "fleet runs claude-opus-5" claim carried since 09-11 in two substrate entries is corrected.


Full detail, sources, and every version number: technology--claude-code-security-floor.md (Added 2026-09-25), fleet--local-git-host.md (Added 2026-09-25), technology--frontier-model-landscape.md (Added 2026-09-25), fleet--prompt-cache-economics.md (Added 2026-09-25), ledger/promises.md (### Walk 2026-09-25 (technologist)).