@anthropic-ai/claude-agent-sdk
0.3.330 · 20 Nov 2026SEE LICENSE IN README.md20.7M / weekRepositorynpmAtom feed
Latest releases
All 351 releases0.3.330
20 Nov 2026Summary· from the note as AI read it
Read before upgrading: 1 change to make.
What's changed
Added a
system/informationalwarning, taggedsettings_not_in_effect, whensettingsor a user, project or local settings file fails validation and is not in effectFixed Claude still waiting for a background agent that
stopTask()stopped just as it finishedFixed variables that a
SessionStarthook writes toCLAUDE_ENV_FILEnot reaching the Bash tool with thecontinueoptionFixed Google Cloud's Agent Platform sessions ignoring
fastModeinsettingswhen the model comes fromANTHROPIC_MODELor a settings file; such sessions now bill at the fast mode priceFixed the initialize response's
fast_mode_statedescribing the default model instead of the model the session runsFixed a lost or swapped tool result from an SDK MCP server when a call overlapped a reconnect; while an earlier request is unanswered, a new one can carry an
idof 1,048,576 or moreChanged the TypeScript SDK's
query()to fail with an error whensettingsis a JSON string that doesn't parseUpdated to parity with Claude Code v2.1.330
Update
npm install @anthropic-ai/claude-agent-sdk@0.3.330 # or yarn add @anthropic-ai/claude-agent-sdk@0.3.330 # or pnpm add @anthropic-ai/claude-agent-sdk@0.3.330 # or bun add @anthropic-ai/claude-agent-sdk@0.3.330
0.3.329
19 Nov 2026Added a
system/informationalwarning, taggedmanaged_hooks_not_loaded, when hooks in managed settings fail to loadAdded
permissions.workingDirectoriesExcludeto theSettingstype: paths, read only from--settingsand settings the host sends, that the working-directory read and write blocks treat as outsideFixed sessions bound with
SpareProcess.claim()ignoringautoCompactWindowfrom the folder's settings or the claim'ssettings
What's changed
Added a
system/informationalwarning, taggedmanaged_hooks_not_loaded, when hooks in managed settings fail to loadAdded
permissions.workingDirectoriesExcludeto theSettingstype: paths, read only from--settingsand settings the host sends, that the working-directory read and write blocks treat as outsideFixed sessions bound with
SpareProcess.claim()ignoringautoCompactWindowfrom the folder's settings or the claim'ssettingsFixed retained memory when an in-process MCP server does not answer its start-up handshake
Fixed memory growth in long sessions that use
--replay-user-messages: the SDK no longer keeps echoes of its own tool results and permission answersFixed an MCP server that a plugin started at launch staying connected after the plugin was disabled or uninstalled and
reload_pluginsran, in headless and SDK sessionsFixed
rewind_conversationremoving history under a message sent right after a model change: the request is now refused with "commands queued", and callers can send it againImproved memory use of idle SDK sessions: after 30 seconds idle, large tool result data that no model request reads is released, except while Remote Control is on
Changed the error when the macOS Keychain can't save a refreshed login:
erroris nowinvalid_request, notauthentication_failed, andapi_error_statusisnull, not 401Updated to parity with Claude Code v2.1.329
Update
npm install @anthropic-ai/claude-agent-sdk@0.3.329 # or yarn add @anthropic-ai/claude-agent-sdk@0.3.329 # or pnpm add @anthropic-ai/claude-agent-sdk@0.3.329 # or bun add @anthropic-ai/claude-agent-sdk@0.3.329
0.3.328
18 Nov 2026Added
onFailuretomcp_toolhooks in theSettingstype: with"block", a hook whose call fails blocks the action it guardsAdded
CLAUDE_CODE_DISABLE_DEFERRED_TOOL_RESUME: set it in theenvoption to stopresumeandcontinuefrom running a tool call that a hook deferred; settings can't set itAdded support for the policy plugin's
allowManagedModsOnlyoption in host-suppliedmanagedSettings; an admin's managed settings still win
What's changed
Added
onFailuretomcp_toolhooks in theSettingstype: with"block", a hook whose call fails blocks the action it guardsAdded
CLAUDE_CODE_DISABLE_DEFERRED_TOOL_RESUME: set it in theenvoption to stopresumeandcontinuefrom running a tool call that a hook deferred; settings can't set itAdded support for the policy plugin's
allowManagedModsOnlyoption in host-suppliedmanagedSettings; an admin's managed settings still winFixed the
agentoption: an inline definition inagentsnow wins over a user, project, plugin or built-in agent of the same name for model, prompt and hooksFixed the
agentoption: an agent from an agents directory now keeps its system prompt whensystemPromptis not setFixed the model saying background agents were still running after an SDK interrupt had stopped them
Changed subagents to wait for their own subagents by default when
CLAUDE_CODE_FORK_SUBAGENT=0is setChanged a shell command rewritten by a
PermissionRequesthook: the rewrite's own permission check now decides the sandbox, so a sandbox-excluded rewrite runs outside it;canUseToolis unchangedChanged
resolveSettings()on Windows to re-read registry policy every 6 hours while none is set (was 30 minutes)Updated to parity with Claude Code v2.1.328
Update
npm install @anthropic-ai/claude-agent-sdk@0.3.328 # or yarn add @anthropic-ai/claude-agent-sdk@0.3.328 # or pnpm add @anthropic-ai/claude-agent-sdk@0.3.328 # or bun add @anthropic-ai/claude-agent-sdk@0.3.328
Dependency tracking
Know when your own dependencies move
@anthropic-ai/claude-agent-sdk published 310 releases in the last 12 months. inlinr folds the routine ones into a line and keeps the few that need you.
Advisories reach you
An email when a critical or high advisory affects a project, before the next install.
Upgrades checked first
Peer ranges, Node, ESM-only and types changes read before you bump.
Only what matters
Majors and breaking notes stay; patches and sibling bumps fold into one line.
Reading releases, release notes and advisories…
Registry facts last checked 21 Nov 2026. Releases and advisories last read 21 Nov 2026.
A whole project to check? Paste its package.json in the dependency check: every deprecated, outdated or vulnerable dependency, with what to upgrade to. Free, no account.