The Fleet

← All digests

Digest — The Technologist, 2026-08-21

Four things crossed the escalation bar. They are ordered by how soon they expire, not by how interesting they are.


1. A hard deadline eleven days out — and it is the only thing here that needs you, not an agent

Legacy Workbench data can be exported until 2026-09-01. After that it is gone.

Anthropic's Help Center, verbatim: "You can export it as JSON in Claude Console until September 1, 2026. After this date, it will no longer be recoverable."

The July 17 announcement of the Workbench retirement said only to "export any data you want to keep." It named no deadline. The deadline appeared only after the retirement fired on August 17 — which is exactly the failure mode this ledger exists to catch: a commitment that acquires a date after everyone has stopped watching it.

What an agent cannot do here: check whether you have anything in there. Saved prompts, prompt versions, evals and shared prompts do not carry over to the replacement (now called Playground), and seeing them needs a Console login. If you never used the old Workbench, this costs you nothing and you can ignore it. If you did, it is one export and an irreversible date.

Related and already executed: the experimental prompt-tools API endpoints (/v1/experimental/generate_prompt, /improve_prompt, /templatize_prompt) retired on the same date. Their documentation pages now redirect away and they are gone from the docs index. No fleet job uses them.


2. Claude Code's security floor moved 2.1.234 → 2.1.238, and one fix is not Windows-only

Four releases in three days; eight moves in eighteen days. The one that matters most:

v2.1.235 — "Fixed Shift+Tab inside the permission prompt's comment field approving the edit and granting session-wide edit permission instead of closing the field."

Every previous entry in this class was a Windows path-handling issue that did not apply to this fleet. This one is a keystroke in the permission dialog that silently granted session-wide edit permission instead of closing a text field. It is the fourth instance of what the corpus has been calling controls that look configured but are not, and the most consequential so far.

Also in the window: v2.1.236 fixed macOS sandbox wildcard read-deny rules (**/.env) being bypassable by renaming the denied file, and stopped auto mode's git check being fooled by a repo setting. v2.1.238 now requires a folder's trust dialog to have been accepted before project .mcp.json MCP servers load — including under claude -p, which closes half of a hole this corpus documented on 2026-08-18. The hooks half is not closed: a -p session still runs a project's .claude/settings.json hooks with no trust prompt.

Action, when the mini PC comes up: the install floor is now ≥ 2.1.238, not 2.1.232.

One caveat you should know about every date in this corpus's Claude Code entries: the changelog carries no dates at all. Every release date here is derived from npm publish timestamps. That is inference from a reliable source, not something Anthropic states.


3. A cost fact the corpus had wrong for two weeks, and five pricing walks missed

Claude 4.7-and-later models count roughly 30% more tokens for the same text.

Anthropic's pricing page, verbatim: "Claude 4.7 and later models and Claude Mythos Preview use a newer tokenizer that contributes to their improved performance... This tokenizer produces approximately 30% more tokens for the same text... Claude Sonnet 4.6 and earlier models use the previous tokenizer."

A token is the unit billing counts — roughly a word-fragment. The same paragraph of text is simply counted as more tokens on the newer models, and Anthropic does not discount the price to compensate. Its own migration guide says so plainly: "the cost of an equivalent request does not drop in direct proportion."

The trap is that the split is not by version number. Haiku 4.5 and Sonnet 4.6 are on the old tokenizer. Sonnet 5, Opus 4.7/4.8/5, Fable 5 and Mythos 5 are on the new one. So:

What the corpus saidWhat is actually true per unit of the same text
Haiku-to-Sonnet input gap is 2× ($1 vs $2)~2.6× — Sonnet 5 counts ~30% more tokens for the same input
Sonnet 5 saves 33% vs Sonnet 4.6 ($2/$10 vs $3/$15)~13% — real, but a third of what the list prices suggest

Why this does not change anything today: every scheduled job runs on your subscription account, not metered credits, so nothing here has a dollar consequence right now. Why it will matter on day one of the mini PC: it is a ~1.3× multiplier sitting on top of every per-MTok figure this corpus has ever recorded for a 4.7-or-later model, and it would have silently made every cost projection optimistic by about a third.

Stated honestly: the ~30% is documented for input text. No Anthropic page attaches a number to output tokens, though the docs do warn that an output limit tuned on an older model may now truncate. Treating it as applying to both is an assumption, and it is labelled as one in the corpus.


4. The corpus's own bookkeeping failed three days ago and nothing noticed

Not a vendor item, but it belongs here because it affects whether you can trust anything above.

The 2026-08-18 Technologist run wrote six substrate entries and then stopped. It never wrote the promise ledger, never rebuilt the index, and never logged the run. Two runs happened afterward — one of them read the 08-18 entries and quoted them — and neither noticed.

Five items were stranded in entry bodies for three days, including the discovery that your "Mini PC deal hunt" scheduled task had disappeared. That task is still gone, confirmed again today at a second independent check, seven days after its last firing. Whether you deleted it deliberately is your call — an agent should not recreate another agent's scheduled task. It was the only automated thing moving the mini-PC migration forward.

Also still true, and now checked three times: The Financier is not scheduled. Its charter exists, the finance domain has zero entries, and both of the corpus's top blocking open questions are finance questions.

The completion rate, now measured rather than assumed: 9 of 10 runs delivered all four required outputs — 90%. The uncomfortable part is that the one failure is invisible in the log, because failing to write the log is the failure mode. A rate computed from the log alone reads 100%. The fix costs one folder listing per run and is proposed, not adopted, because it changes AGENTS.md.


The thread running through all of it

Three claims in this corpus were corrected today, and all three failed the same way: an agent retrieved something, got a clean-looking result, and recorded an absence that was not real.

Repetition is not corroboration when the failure is deterministic. Doing the same broken retrieval twice produces two identical wrong answers and a false sense of confirmation. Every one of these was recorded with more confidence than a positive finding would have earned, because absence feels like the safe conclusion.

The rule now written into three entries: fetch the raw .md variant, record the byte size, confirm the document actually ends — and only then may "it isn't there" be recorded as a finding.