docs: Claude Code v2.1.267 - maxEffortLevel cap, Linux sandbox, tool name limit doubled, new Anthropic repo in discovery

[claude-yolo]

Night shift report

WHY THIS MATTERS: v2.1.267 lands a meaningful set of enterprise-facing controls (a global effort cap, Linux sandbox parity with macOS) alongside a quiet but important API change (tool name limit doubled to 128 chars), two security fixes that close real holes, and a new Anthropic public repo that appeared in discovery.json — it has "incident transcript" in the name, which may or may not warrant archiving but a human should see it.

HIGHLIGHTS:

• maxEffortLevel setting (new feature): Admins can now cap the effort level globally — top-level or per-model under modelSettings — covering Bedrock, Vertex, and Foundry. Users can still pick lower, but not higher. This is the org-wide cost/speed governance knob enterprises have been missing.
• Linux sandbox support (Government security doc): The OS-level sandbox for Code sessions (previously macOS-only) now applies to Linux too, requiring bubblewrap + socat. Container environments get a sandbox.enableWeakerNestedSandbox escape hatch and a sandbox.enabled: false fallback. The doc was previously silent on Linux — this is a real gap closed.
• Tool name max length: 64 → 128 chars (define-tools.md): The API tool name regex changed from {1,64} to {1,128}. Breaks no existing tools but lets builders use longer, more descriptive names without truncating.
• Security fix — managed settings fail-safe: allowedHttpHookUrls, httpHookAllowedEnvVars, and allowedChannelPlugins used to admit everything when unreadable; they now admit nothing. Unreadable policy → deny, not allow. Important for managed deployments where misconfigured policy files previously opened the gates.
• Security fix — marketplace path traversal: A marketplace entry path containing a backslash could bypass the containment check for fetched marketplaces on macOS and Linux. Fixed.
• discovery.json: new repo anthropics/mythos-5-incident-transcript: Anthropic published a new public GitHub repo. The name suggests it may be a transparency artifact (an AI incident log?). Not yet fetched — whether to add it to sources.json is a human call, but the archive should have an opinion.
• --system-prompt-snapshot off (new flag): Forces the system prompt to render fresh on every request instead of reusing the recorded one. Aimed at prompt iteration workflows — previously you'd have to start a new session to see prompt changes reflected.
• API docs cleanup: BetaManagedAgentsStreamSessionEvents and BetaManagedAgentsStreamSessionThreadEvents verbose type-union definitions removed across all language SDK reference pages (CLI, C#, Go, Java, PHP, Python, Ruby). Looks like schema simplification on the platform side — the union definition is gone, not just the docs for it.
• Government — bootstrap address aliases: Added explicit guidance that DNS aliases, redirects, or reverse proxies under an agency's own domain won't work as the bootstrapUrl value. This clarifies a support headache (sign-in silently fails with an alias).
• Government — connector tool policy rewrite: Per-tool control for connectors now documented in detail. Tools listed in a connector's Tool Policy prompt on each use; unlisted tools stay under user control; M365 write actions (send mail, calendar edits, Teams posts) are permanently set to ask regardless of policy.
• Large CHANGELOG batch: 52 entries total including prompt-cache stability fixes (subagents, resumed sessions, MCP reconnects), VSCode fixes (CRLF diffs, RTL text, Remote-SSH), Claude Tag fixes (usage credits error messaging, thread reply routing), Cowork cloud startup fix for sandboxed orgs, and claude remote-control credential renewal (was dropping sessions after ~30 days).

---

Created by night-shift claude-yolo
Day-shift claude-yolo will review and merge this in the morning