Only what changed in your stack.

Skills can now ship with their own executable scripts

Skills are no longer just markdown — they can carry scripts the agent runs directly, which makes them behave more like real commands.

Subagents inherit the parent session permission set

Spawned subagents stop interrupting you for permissions the parent session already granted.

Mentioning a directory builds a scoped retrieval index for that path

Pointing at a folder narrows what the agent searches, instead of letting it wander the whole repo.

Claude Opus 5 is the default model, Opus 4.7 removed from fast mode

The default model changed to Opus 5 with 1M context, and fast mode no longer includes Opus 4.7, which may change your model selection if you relied on it.

Subagents can spawn nested subagents up to depth 3 with stream-json forwarding

The default nesting depth for subagents increased from 1 to 3, and stream-json now shows subagents spawned at deeper levels when forwarding is enabled, giving better visibility into agent hierarchies.

Deep research starts only on manual invocation

Claude will no longer start a /deep-research session on its own — you now have to call it manually, which prevents unexpected research runs.

Code review runs as a background subagent

The /code-review command now delegates to a background subagent instead of running inline, so your main conversation isn't blocked and stacked slash commands still target the review.

Auto mode handles dangerous checks without asking

Auto mode no longer pops up permission dialogs for dangerous file deletions, background processes, or suspicious Windows paths — the classifier decides them automatically.

Agent names can no longer contain colons

If you name an agent with a colon in it, the system will now reject it — colons are reserved for plugin namespacing.

The SDK accepts a streaming input channel for voice front-ends

You can wire speech into a coding session without waiting for the speaker to finish a sentence.

MCP tool output truncation no longer leaks memory

When an MCP tool returned more data than the context window could hold, Claude Code kept the full result in memory even after truncation. This prevents long-running agent sessions from gradually consuming all available RAM.

Subagents no longer spawn nested subagents by default

This changes the default agent hierarchy from potentially deep trees to a flat one-level model. If your workflow relied on subagents delegating to further subagents, you'll need to opt in via the new environment variable.

Concurrent subagents capped at 20 by default

Previously, a single prompt could spawn unlimited background subagents, overwhelming your system. The default limit is 20, with an environment variable to raise it if your workflow requires it.

New sandbox.filesystem.disabled setting skips filesystem isolation

A new setting lets you disable the filesystem sandbox while still blocking network egress. Previously you had to choose between full isolation or no isolation; now you can restrict network access without file restrictions.

Fixed quadratic slowdown in long session message normalization

Long sessions used to slow down dramatically as conversations grew because message processing cost increased with the square of turns. Now it scales linearly, so resuming a long session is fast again.

Resumed background agents restore their prompt and tool limits

When a background agent session resumed, it would forget which agent it was and revert to the default. Now it remembers the agent's prompt and tool limits, so background agents behave consistently across pauses.

Cloud sessions re-run interrupted turns after container restart

If a cloud session's container restarted while the agent was processing a message, the message was lost and the session became unresponsive. Now the interrupted turn re-runs automatically when you resume.

Agent tool calls get typed handlers and automatic schema generation

Defining what a voice agent can do is now one typed function instead of a function plus a hand-written schema.