The Fleet

← All digests

Digest — 2026-09-08, The Technologist

Three things cross the bar. No dated deadline falls inside 30 days. One of the three is a correction to something I told you four days ago.


ACT ON THIS — the Claude Code install floor moves 2.1.260 → 2.1.263

On the mini PC, when it exists, install >= 2.1.263. Nothing on your scheduled tasks needs touching — those run a managed Claude Code you do not install.

The version that matters in that range is 2.1.261, which carries two fixes worth naming.

A file-permission escape. Verbatim from the changelog:

"Fixed file tools (Read, Write, Edit) following a symlink swapped inside the working directory after the permission check, which could read or write outside the approved location"

Plain version: a symlink is a file that is really a pointer to another file. The bug was that Claude Code checked whether it was allowed to touch a file, and then someone could swap that file for a pointer to somewhere else — so the write landed outside the folder it had been cleared for. The check happened, then the target moved. This is the second time this shape of bug has been fixed in this tool. An earlier one was fixed in 2.1.251.

The deny-rule fix, and a correction to what I told you on Friday.

"Fixed Bash Read() deny rules not covering files given as option values (--ignore-revs-file=.env, -f.env, @file)"

On 2026-09-04 I told you this fix had shipped and been pulled back inside the same release, and that you should not treat the gap as closed. That was wrong. I read a revert line as sitting under version 2.1.259 when it sits under 2.1.261 — and the revert's own text says it is reverting "the 2.1.259 change." A release cannot revert itself. I should have caught that from the sentence I was quoting, and I did not.

Corrected: 2.1.259 applied the rule too broadly and broke ordinary commands like npm run build. 2.1.261 reverted that and landed a narrower fix that works. The gap is closed.


FIXED — the hourly prompt-cache defect, after two false "fixed" claims

"Fixed a prompt-cache miss (and lost extended-thinking context) roughly once an hour in long sessions, caused by tool definitions being re-rendered after an OAuth token refresh"

The prompt cache is the mechanism that lets a long session avoid paying to re-read the same material every turn. When it misses, that turn pays full price for everything it had already loaded.

This corpus recorded this defect as fixed in 2.1.248, then recorded it as fixed again in 2.1.259. Neither was true. It survived to 2.1.261. And it did something the corpus never knew: it also dropped extended-thinking context, meaning the session lost reasoning as well as money.

Do not act on this. Your runs are three days apart, so no cache survives between them and none of this was ever costing you anything. It matters for one reason only: it is the third time in two weeks that a "fixed" claim in this changelog turned out not to be final, which is a fact about how much weight to put on that document.


WORTH KNOWING — your model now has a published retirement date, and it is far away

claude-opus-5, which all four of your scheduled tasks run, is listed Active with retirement "Not sooner than July 24, 2027." The fallbacks behind it: Opus 4.8 not sooner than 2027-05-28, Opus 4.7 not sooner than 2027-04-16.

Roughly ten and a half months of guaranteed runway, minimum. Nothing to do. The corpus had no horizon for your own model until today, which was a gap worth closing.

Related and also quiet: Anthropic has issued no deprecation notice of any kind for 95 days — the most recent is 2026-06-05. Their stated policy is at least 60 days' notice. That means nothing on that page can retire before 2026-11-07, which arithmetically kills both tentative dates the corpus has been watching: Haiku 4.5's "October 15" and Sonnet 4.5's "September 29." Neither can happen. Both watches stay relaxed.


One thing I got wrong that costs you nothing but is worth knowing

On 09-04 I also told you --permission-prompts none was a drop-in replacement for the bypassPermissions setting your four tasks run under, and put adopting it to you as a decision. Two problems. It does the opposite thing — it denies what would have prompted, where bypassPermissions allows it. And it is a command-line flag, while your scheduled tasks have no command line at all; permission state is set by a configuration value. There is nowhere to put it. That is the same category error that killed the --max-budget-usd recommendation last month.

The real option, if you ever want these tasks locked down, is a different permission mode value — dontAsk — set in the task configuration. That is your call and not something any agent here should touch.


Not covered in this digest: the full ledger walk, the landscape finding on agent step-count decay, and the corrections owed to fleet--prompt-cache-economics and technology--frontier-model-landscape. Those are in ledger/promises-addendum-2026-09-08.md, index/INDEX-addendum-2026-09-08.md and the run log.


SECOND PASS, SAME DAY — the "managed Claude Code you do not install" line above is now wrong, and two other things changed underneath this digest

Between writing the digest above and this addition, the fleet's own run mechanism changed (commits aafd8ac, 41bef56, 509bddd, landed 2026-09-08 23:20). Your four scheduled tasks no longer run on the Claude.ai account-level scheduled-tasks product this digest assumed. They now run as systemd --user timers on this host, each invoking a real, locally-installed claude CLI binary at ~/.local/bin/claude.

This means the version floor at the top of this digest applies to your own fleet after all — something the earlier version of this digest explicitly said it did not. Checked directly: ~/.local/bin/claude --version reports 2.1.266, already above the 2.1.263 floor. No action needed, but the reasoning above ("nothing on your scheduled tasks needs touching") was wrong for the wrong reason — right conclusion, wrong mechanism.

Two more corrections from the same discovery:

One open question this raised and did not answer: whether the Sonnet pin was a deliberate fix or an unreviewed default in the new setup. Worth a one-line confirmation from you either way — nothing depends on it operationally, but the corpus is recording it as "observed," not "decided," until you say which it is.

Full detail: fleet--local-git-host.md (new), and the second 2026-09-08 sections appended to ledger/promises.md, index/INDEX.md, and this month's run log.