Maya: Anthropic's September fourteenth engineering report puts a striking number on its own workflow: continuous integration jobs rose twenty-five fold over six months, while the test suite grew tenfold and headcount barely moved. That traces back to Claude authoring eighty percent of their code and helping review pull requests.

James: What's useful there, Maya, is where the bottleneck moved. Faster coding and review shifted the constraint into CI, after code was written. And their early fixes kept expiring: a bigger machine bought seventy days, package-level sharding bought twenty-nine, and restarting the service eventually bought less than one day.

Maya: The root problem was one process holding all the running history. When the listener fell twenty minutes behind, tens of thousands of updates could go missing from the selector. Anthropic ran a long-running Claude Tag session to watch the backlog and resume the same investigation whenever lag passed fifty thousand jobs.

James: The durable fix was architectural. They moved state out of the listener into an in-memory journal, with a separate consumer rolling events into per-test history every few seconds. One engineer finished it in three weeks, Claude handled much of the tuning like journal sizing and worker counts, and the backlog stayed flat after cutover.

Maya: That thinking shows up in the smaller releases too. Claude Code two point one point two six seven added a maxEffortLevel setting, so administrators can cap model effort at the top level or per model under modelSettings. That ceiling applies across providers, including Amazon Bedrock, Google Vertex AI, and Foundry, and users can still pick anything below the maximum.

James: There's also a nice option for tuning prompts. With system-prompt-snapshot off, Claude Code renders the system prompt fresh on every request instead of reusing the one recorded with the conversation, which Anthropic frames as a way to iterate on prompt text. And reloading a session larger than five megabytes no longer drops parallel tool calls and their hook output.

Maya: The managed-work fixes stood out to me. Cowork scheduled cloud tasks no longer fail at startup when an organization requires sandboxing, and if managed HTTP-hook URLs or channel plugins can't be read, Claude Code now admits none of them rather than all of them. That fail-closed behavior is the safer default for a team relying on managed settings.

James: The next release leaned into gateway teams. Two point one point two six eight lets administrators put pricing in gateway dot yaml, and signed-in clients receive those rates through managed settings. Those values feed both slash cost and telemetry, so the numbers inside a client can match the gateway's own spend meter.

Maya: It also added two network-safety signals. Claude Code warns at startup when access control allow cidrs is empty, and gives a one-time warning when the first request arrives from a public address. A new gatewayInternalNetworks setting lets an administrator permit login from the organization's own public block, though Anthropic is clear these don't replace the organization's network policy.

James: For automation, plugin administration got easier. The install, uninstall, update, enable, and disable commands now accept dash dash json, and plugin list json rows include errorDetails and noteDetails, so scripts can consume structured results without parsing terminal text. Self-hosted runner operators can also start with remove-session-state, which is off by default and only clears per-session directories, not the repository.

Maya: Two point one point two seven three added something remote-control users will appreciate: you can fork a session that began through remote-control from the Claude app, and the fork runs as a background session on your own computer. That gives you a second branch without moving the new work to a separate cloud host.

James: The release also added five optional gateway hint headers, which operators opt into with a single environment variable, exposing request context a gateway can use operationally without changing model output. And there's better failure visibility for MCP: when a server disconnects and automatic reconnection gives up, Claude Code now posts a notification pointing you to slash mcp.

Maya: The permission fixes here matter for safety. Bash commands the checker couldn't fully analyze no longer skip the prompt under blockReadsOutsideWorkingDirectories, and a subshell can no longer hide a dangerous rm while bypass mode is active. Managed MCP and connector restrictions delivered through MDM also stay effective when server-managed settings are present.

James: And two point one point two seven one extended fast mode to remote sessions, whether they run in Anthropic's cloud or on self-hosted runners. The host's fast-mode setting, or slash fast typed in the session, applies when the organization permits it.

Maya: The network change is the practical one for reviewers. Bash, PowerShell, and Monitor commands can now request an allowed-domains list for that individual command in auto mode. The requested hosts are reviewed with the command and opened only for it, while other hosts stay refused. That's much narrower than opening broad access for a whole session.

James: There's a thoughtful fallback fix too. A fast request that hits overload or a usage-credit limit now falls back to standard speed in the covered retry path instead of failing the turn. And the fullscreen config panel gained mouse support, so the wheel scrolls and clicking a value changes it.

Maya: Custom and plugin subagents can also use omitClaudeMd to run without the CLAUDE dot md files, though managed policy still loads, so it never bypasses organization policy.

Maya: That's our walk through the week's releases here at Claude Code Review. Keep shipping carefully.