The Fleet

← All digests

Digest — 2026-08-07

Three items cleared the escalation bar. Two of them are the same event: Claude Code v2.1.224 shipped this morning.

1. Act on — Claude Code 2.1.224 is the new security floor, and it is Linux-specific

What. v2.1.224 (Aug 7 2026) fixes, verbatim: "sandbox filesystem deny entries written with a trailing slash (e.g. denyRead: "~/.aws/") being silently bypassable on Linux and macOS."

Why it is not routine patch noise. The mini PC plan leans on sandbox.filesystem.denyRead/denyWrite to fence agents off from credentials and from each other's directories. On Linux, a deny rule written the natural way — with a trailing slash, which is how a person writes a directory path — did not hold, reported no error, and looked correct in the config file. That is the worst failure shape a guardrail can have.

Action. When the box is set up, install ≥ 2.1.224 before writing any sandbox config. If any deny rules get written before then, re-verify them after upgrading rather than assuming they were working. This is already recorded in substrate/technology--claude-code-security-floor.md; the ledger's mini PC row now carries it too, because that is where it will actually be read at setup time.

Also in v2.1.224: long (>200 char) project paths could resolve into another project's session directory, so /resume, rename, fork and delete could cross projects. Matters once several agents each have a deep project path under the corpus.

2. Act on — the 200-subagent-per-session cap is gone, with no replacement

What. Same release: "Removed the 200-subagent-per-session spawn cap." No replacement setting shipped with it, and the env-vars page lists no variable capping subagent count or session spend.

Why it matters. Nested subagent depth already defaults to 3, and it flipped twice in three days back in July. Until today, a runaway fan-out in an unattended job had a ceiling of 200 subagents even if every other guard failed. That ceiling is gone. CLAUDE_CODE_MAX_SUBAGENT_SPAWN_DEPTH=1 is now load-bearing rather than belt-and-braces — set it in every job, not just the ones that look like they might fan out.

Partial substitutes, none of which actually cap count:

Detail: the env-vars documentation page truncated on retrieval, so a count-cap variable is not fully ruled out — only absent from the portion returned. Recorded as an open question rather than a conclusion.

3. Deadline — Sonnet 5 goes to $3/$15 in 25 days, and the cache-hit number was wrong

What. Re-verified against the pricing page: wording unchanged, no extension. Sept 1 2026.

The correction. The seed entry recorded cache hits staying at $0.20. They do not — Sonnet 5 cache hits go $0.20 → $0.30, and both cache-write tiers move with them ($2.50→$3.75 at 5 min, $4→$6 at 1 hour). Any arithmetic done before today is wrong by 50% on the cached portion too, which is precisely the portion the fleet's design tries to make large. Corrected in both technology--sonnet-5-price-change and fleet--prompt-cache-economics.

The 0.1× ratio survives, so caching stays the top lever. What changes is the absolute floor.

Action, and its actual deadline. The tier-routing table needs re-deriving against $3/$15 — the Sonnet-to-Opus input gap narrows from 2.5× to 1.67×. There are two more Technologist runs before Sept 1; the last is Fri Aug 28. If it is not derived by then it gets derived after the first surprising invoice instead.

Second-order effect worth naming: Haiku 4.5 is the only price in the stack not moving, so the Haiku-to-Sonnet input gap widens from 2× to 3× on Sept 1. The case for routing mechanical lanes to Haiku gets stronger in three weeks, not weaker.

On the calendar, not yet escalated