Skip to content

Claude Agent SDK v0.3.323 ships policy checks and session fixes

The TypeScript agent SDK adds two alpha helper APIs and a batch of fixes for policy reads, permissions, session forking and MCP servers.

Draftship Editorial4 min read

The Claude Agent SDK for TypeScript is now at Claude Agent SDK v0.3.323, published on 12 November. The tag comes from claude-code-publisher as a single commit to main, and it reads as a maintenance drop rather than a feature launch: two new alpha helpers for hosts, a heartbeat for long hook waits, an environment variable to silence subagent progress summaries, and a list of fixes covering policy reads, permission prompts, session forking and MCP server handling. It installs with npm install @anthropic-ai/[email protected], or the yarn, pnpm and bun equivalents. Every item below is listed in the v0.3.323 release notes.

Two alpha APIs aimed at hosts

machinePolicy.didNotLoad (alpha) is added to the result of resolveSettings(). It tells a host when a machine policy source may be present and did not load. The second, matchDeniedModels() (alpha), tells a host which models a deniedModels list blocks, applying Claude Code's own entry rules rather than a host's own matching.

Both are marked alpha, which in this release series means the surface can still move.

Heartbeats, progress summaries and search results

While a turn waits on long-running hooks, the SDK now emits a periodic tool_progress heartbeat. For those events, tool_name carries the hook event, tool_use_id begins with hook-, and parent_tool_use_id may name no tool call at all.

A new environment variable, CLAUDE_CODE_DISABLE_AGENT_PROGRESS_SUMMARIES, turns off subagent progress summaries when set to 1, even if agentProgressSummaries is true. The source notes the setting has been available since 0.3.318. Separately, a subagent's WebSearch result message now carries tool_use_result when the search did not run or found nothing.

Policy reads and permission checks

resolveSettings() used to keep its first reading of a macOS MDM profile or a Windows registry policy for the life of the process. It now re-reads after 30 minutes, and sooner after a read that failed.

On permissions, canUseTool suggestions had been including a Bash allow rule for commands that write to sensitive paths such as .git or .claude, where the rule did not stop the next request. That is fixed. The release also changes what happens when a tool's permission check fails with an error after a PreToolUse hook allowed the call: it now goes to canUseTool, or is refused when no callback exists.

Session forking, MCP and plugins

forkSession() with upToMessageId no longer brings back messages kept after a compaction summary when forking at that summary, and forking at a compaction boundary keeps the conversation before it. getSessionMessages no longer drops a queued prompt that directly follows a compaction summary.

Most initialize options, such as agents and skills, had been ignored when a SessionStart hook supplied the first message. That is fixed. Progress notifications from SDK MCP servers are no longer dropped for tool calls still running when the set of SDK MCP servers changed, and a server removed and re-added with setMcpServers() no longer stays in needs-auth with no tools after a tool call tried to reach it in between.

init.plugin_errors now reports a plugin that only project settings enable but that is not installed, including the install command, and it also lists load errors for plugins enabled by managed settings. A check on it can fire before the background install ends; CLAUDE_CODE_SYNC_PLUGIN_INSTALL=1 makes the call wait. Assistant messages now send supersedes after a response whose tool call could not be parsed. The release states parity with Claude Code v2.1.323.

Inference: what this likely changes for app builders

None of this alters the public shape of an agent loop, so most builders will pick it up through a version bump. The pattern of fixes suggests the SDK is being hardened for hosts that run agents on managed machines, where policy files, denied model lists and plugin state can change while a process is alive. The 30-minute re-read is the clearest signal of that intent.

Limits and open questions

The notes do not say when the two alpha APIs become stable, whether their signatures will change, or what runtimes are supported. No breaking changes are listed, but no migration guidance is given either, and the parity note with Claude Code v2.1.323 is not expanded on. Whether CLAUDE_CODE_DISABLE_AGENT_PROGRESS_SUMMARIES will be documented as a long-term flag is unclear.

What it means for building apps with AI

If you are assembling an agent into an app rather than writing the SDK plumbing yourself, most of this release is invisible. What carries over is the shape of the problems: settings that need re-reading, permission rules that must actually block the next request, and session state that has to survive a fork. If you are wiring model calls into an app and want to see how that layer is handled in Draftship, the guide on adding AI to your app covers it. Draftship is not affiliated with Anthropic and does not ship this SDK as part of the announcement above.

Primary sources

  • v0.3.323 · Claude Agent SDK (TypeScript) releases, Nov 12, 2026

This article was drafted with AI assistance from the primary sources listed below and checked automatically against them. An editor hasn't reviewed it yet.

AI coding agents

opencode v2.3.2 fixes GPT-6.2 prompt caching billing

opencode v2.3.2 sends the system prompt as a leading developer message on every Responses route so GPT-6.2 caches it, cutting uncached input per step from about 14,000 tokens to 3 on one Copilot model.