Skip to content
Agents tracked: 501 Releases (7d): 777 Agent status: no reported issuesSince last limit reset: Claude 10d · Codex 8d For AI agents: Markdown Updated Nov 21, 2026

Claude Agent SDK changelog: what's new each month

Every stable Claude Agent SDK release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 7 months covered; the current month updates daily.

November 2026 · so far

11 releases: 0.2.177 → 0.2.187

November saw regular updates to the Claude Agent SDK with incremental improvements and bug fixes. The bundled Claude CLI was updated to maintain compatibility with the latest tools.

Highlights

  • Added on_failure option for HookMatcher to control hook callback error handling.
  • Fixed intermittent test failures on Windows by supporting both Bash and PowerShell in shell-dependent tests.
  • Regular updates to bundled Claude CLI versions for improved tool compatibility.

New

  • on_failure option for HookMatcher to specify behavior when hook callbacks fail.

Improved

  • Test stability on Windows by accepting either Bash or PowerShell shell tools in shell-dependent tests.

Fixed

  • Intermittent test failures on Windows related to shell tool detection.

October 2026

13 releases: 0.2.164 → 0.2.176

October brought improved type safety and reliability to the Claude Agent SDK, with multiple rounds of bundled CLI updates and significant enhancements to how the SDK handles hook events and session message retrieval.

Highlights

  • Hook implementations now receive fully typed output classes and input fields instead of untyped dictionaries.
  • Session message retrieval now correctly includes all parallel tool calls and results.
  • Enhanced security for CI/CD pipelines with network firewalls, permission enforcement, and trusted publishing.
  • System initialization data is now fully typed with nested TypedDict structures for better type information.

New

  • Typed output classes and input fields for hook-specific implementations.
  • SystemInitData TypedDict with nested types for system message payload structure.
  • Slack alerts when a release is blocked or fails.

Improved

  • SandboxSettings.ignoreViolations type corrected to dict[str, list[str]] to match CLI schema.
  • Clarified that include_hook_events does not suppress SessionStart and Setup hook events.
  • CI/CD security hardened with egress-firewall runner, network allow list, and permission enforcement.
  • Release build tools now hash-pinned to prevent tampered downloads.
  • PyPI per-file limit raised to 250 MiB and project size limit raised to 150 GiB.
  • Release uploads switched from long-lived API tokens to trusted publishing with GitHub OIDC.

Fixed

  • Parallel tool calls now correctly included in session messages from get_session_messages and get_session_messages_from_store.
  • Thinking-deltas end-to-end test fixed for CLI 2.1.287+ with explicit thinking display options.

September 2026

13 releases: 0.2.150 → 0.2.163

September brought refinements to system prompt handling and message delivery options, along with stability fixes for multi-turn conversations involving subagents. The bundled Claude CLI was updated throughout the month to track upstream improvements.

Highlights

  • System prompt snapshots can now be recorded in sessions for better reproducibility.
  • New verbatimprompts option lets you send user messages to the CLI without automatic file expansion.
  • Fixed follow-up turns failing when subagents completed just before message processing.
  • Improved docstring accuracy and formatting across the SDK.

New

  • SystemPromptCustom typed dict for custom system prompt definitions.
  • snapshot field on SystemPromptPreset to enable session-level prompt recording.
  • ClaudeAgentOptions.verbatimprompts flag to disable automatic @path file expansion in user messages.

Improved

  • Docstrings aligned with actual code implementation.
  • CI wheel compression and PyPI upload handling for larger distributions.
  • Test infrastructure pinned to specific models to avoid CI failures from upstream default changes.

Fixed

  • Follow-up turns now succeed after background subagents complete their work.
  • stdin closure timing issue during multi-turn queries with hooks or MCP servers.
  • Docstring formatting inconsistencies corrected throughout the codebase.

August 2026

21 releases: 0.2.129 → 0.2.149

August brought significant enhancements to session management, subagent integration, and error handling, alongside important security fixes for skill validation and expanded protocol support.

Highlights

  • Added conversation reset detection and message origin tracking to help applications understand turn context.
  • Implemented session resumption with checkpoint control via resumesessionat and resumedropsturn options.
  • Extended MCP support to version 2.x while maintaining backward compatibility with 1.x.
  • Added forwardsubagenttext option to stream subagent text and thinking blocks for full nested conversation rendering.
  • Fixed skill name validation to prevent injection attacks through the CLI's allowedTools parameter.

New

  • ConversationResetMessage dataclass surfaces when conversations are cleared or reset.
  • Message origin field on UserMessage and ResultMessage indicates why a turn was initiated.
  • resumesessionat and resumedropsturn options for truncating and validating session resumption.
  • forwardsubagenttext option to forward subagent text and thinking blocks in the message stream.
  • ResultError exception with structured error payload for terminal CLI errors.
  • canusetool callback support for query() and string prompts with stdin kept open.
  • Recovery of parenttooluseid when reading subagent transcripts via getsubagentmessages().
  • MCP 2.x support for in-process SDK MCP servers using in-memory transport.

Improved

  • Skill name validation now prevents characters that could cause injection through CLI parameters.
  • Session resumption now seeds settings.json and coworksettings.json into temporary config directories.
  • Error messages for failed resume operations are now more descriptive for pending control requests.
  • Bundled Claude CLI updated throughout the month to version 2.1.252 with cumulative improvements.

Fixed

  • Skill names containing parentheses, commas, control characters, wildcards, or leading slashes are now validated.
  • Skill name validation prevents --allowedTools injection via unchecked characters in ClaudeAgentOptions.skills.
  • Session resumption now properly surfaces CLI rejection errors to pending control requests like initialize().

July 2026

18 releases: 0.2.111 → 0.2.128

July brought significant security hardening and reliability improvements to the Claude Agent SDK, with numerous fixes for subprocess management, input validation, and Windows compatibility. The month also introduced better visibility into query termination reasons and improved type safety.

Highlights

  • Fixed critical security vulnerabilities in argument handling to prevent command injection on Windows and across platforms.
  • Resolved zombie process issues in subprocess cleanup that could leave orphaned Claude processes.
  • Added terminal reason tracking to ResultMessage so you can see why queries ended.
  • Improved type safety with typed ModelUsage on ResultMessage.
  • Fixed stdin closure timing that was prematurely ending queries with background tasks.

New

  • ResultMessage.terminalreason field indicates why the query loop ended (completed, maxturns, abortedstreaming, etc.).
  • ModelUsage type definition for better type safety on usage tracking.

Improved

  • Subprocess cleanup now shields from asyncio cancellation to ensure proper SIGTERM/SIGKILL teardown.
  • NDJSON parser now correctly handles whitespace on large lines exceeding 64 KiB buffer.
  • CLI argument handling uses --flag=value form for dash-prefixed values to prevent flag misinterpretation.
  • Windows support now blocks spawning .bat/.cmd scripts to prevent cmd.exe metacharacter injection.
  • Slack notification workflow now properly escapes untrusted fields using jq.
  • Build scripts hardened against command injection via CLAUDECLIVERSION validation.
  • CI now fails properly on curl errors during CLI installation.

Fixed

  • TypeError when message content is a plain string instead of expected list.
  • canusetool callback shadowed by allowedtools or bypassPermissions now produces runtime warning.
  • Resume and session ID values containing cmd.exe metacharacters or newlines now raise ValueError on Windows.
  • Premature stdin closure when background tasks are still in flight.
  • E2E stderr test flakiness.
  • CLI workspace trust issues preventing project-scoped permission grants.

June 2026

23 releases: 0.2.88 → 0.2.110

June brought improved async compatibility and better lifecycle event handling to the Claude Agent SDK. The month focused on strengthening the foundation for different async backends and exposing more detailed task state information.

Highlights

  • Trio compatibility for session stores, fixing TypeError crashes when using Trio backend
  • Typed task lifecycle events (TaskUpdatedMessage) with taskid, patch, status, and sessionid fields
  • Comprehensive async backend testing via anyio's pytest plugin for both asyncio and trio
  • Workload identity federation for CI with automatic short-lived token refresh

New

  • TaskUpdatedMessage typed events exposing terminal system/taskupdated events with structured fields

Improved

  • Session store code paths (TranscriptMirrorBatcher, sessionresume, sessions) ported from asyncio to anyio for broader async runtime support
  • Test suite now validates async backend compatibility by running every test under both asyncio and trio runners

Fixed

  • Trio compatibility crash (TypeError) in session store handling
  • MCP dependency constraint to prevent incompatible breaking changes from version 2.0.0+

May 2026

3 releases: 0.2.85 → 0.2.87

May brought security improvements and maintenance updates to the Claude Agent SDK, with a focus on strengthening authentication practices.

Highlights

  • Switched authentication to use short-lived tokens via Workload Identity Federation instead of long-lived API keys
  • Updated bundled Claude CLI to version 2.1.150

Improved

  • CI workflows now use Workload Identity Federation for more secure Claude authentication

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Claude Agent SDK: pricing, features and alternatives · All changelogs