Claude Code 2.1.299 and 2.1.300
Anthropic published v2.1.299 and v2.1.300 on October 14, 2026. The biggest change is about cost: "Added idle compaction in the terminal: a 200K+ token conversation left idle for about 54 minutes is compacted before its one-hour prompt cache expires, so your next message doesn't reprocess it all". The threshold can be changed with CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS, and "idleCompaction": false turns it off.
The fix to read first is a sandbox gap: "Fixed a sandboxed shell command being able to write into a --plugin-dir folder, whose code Claude Code reloads and runs outside the sandbox". In other words, a sandboxed command could plant code that later ran without the sandbox. Two other fixes concern credentials. One stops an agent teammate started without a description from keeping part of a credential from its prompt in the task description shown for it. The other: "Fixed a login refused by forceLoginOrgUUID leaving the refused organization's token stored, or live after /login".
For observability, 2.1.300 adds cache_creation_5m_tokens and cache_creation_1h_tokens to the OpenTelemetry api_request event, and GenAI semantic-convention attributes such as gen_ai.provider.name and gen_ai.tool.name to trace spans. Agent hooks now respect their configured model, and an unavailable model "falls back to the caller's instead of allowing the action". A hook that used to fail open now fails safe.
OpenAI Agents SDK 0.24.0
The Python SDK's first minor release since 0.23.1 shipped on October 15. It will "expose MCP annotations on converted tools", so hints from an MCP server, such as read-only or destructive, now reach the SDK's tool layer. It will also "defer automatic compaction for pending approvals", so context is not rewritten while a human is still deciding. It also adds "run-wide tool exception formatting".
ADK Java 1.12.0
Released October 14, it adds an experimental Workflow type and JoinNode, and makes every agent a workflow node ("make BaseAgent extend BaseNode"). That moves the Java SDK toward ADK Python's graph workflows. One fix deserves a check: ADK Java will now "create and use Gemini context caches when ContextCacheConfig is set", which suggests Java teams who set that config before were not getting caching. Another change will "allow rejecting config-supplied stdio MCP servers", narrowing one path to command execution.
Pydantic AI 2.56.0
Released October 13, it adds SmolSandbox, "a workspace that runs harness capabilities in a local smolvm microVM", and Postgres-backed task durability. Its compatibility note: "Run BrowserUse on the host model by default and make browser-use telemetry opt-in".
What remains uncertain. None of the notes says whether the sandbox or credential fixes were reported as vulnerabilities, or which earlier versions were affected.
Analysis: Two ideas run through this week's releases. Compaction is becoming a cost and safety control: Claude Code times it to the cache, and the Agents SDK holds it during approvals. And a sandbox is only as strong as the paths out of it. The plugin-folder fix is a prompt to audit every directory a sandboxed agent can write to that something unsandboxed later reads, in any harness.
Claude Code 2.1.299/2.1.300 (October 14) compact idle 200K+ sessions before the one-hour prompt cache expires, fix a sandboxed command writing into a plugin folder run outside the sandbox, and add OpenTelemetry GenAI attributes; OpenAI Agents SDK 0.24 defers compaction during approvals, and ADK Java 1.12 adds workflow graphs and working context caching.