v0.4.0

Minor Changes

• 9543747: Upgrade @openai/codex-sdk to ^0.142.1 and surface two existing SDK capabilities:

  • Thread resume: capture and expose the Codex thread_id on thread.started
    (metadata.codexAgent.threadId); resume a context's thread after LRU eviction or via a
    client-supplied metadata.codexAgent.threadId (legacy metadata.codexThreadId accepted).
  • Image input: forward A2A image/* file parts to Codex as local_image (inline bytes →
    temp file; local file:// paths scoped to the working directory). Agent card now declares
    image input modes.

  Also exports a typed metadata contract so consumers don't hand-write magic keys:
  threadIdMetadata(), readThreadId(), CodexAgentEventKind, CodexAgentMetadata,
  CODEX_AGENT_METADATA_KEY, LEGACY_THREAD_ID_KEY, and SUPPORTED_IMAGE_MIME_TYPES.
  getWorkingDirectory is deprecated in favor of getThreadOptions().workingDirectory.

v0.3.1

Patch Changes

• ade51ff: update readme

  • Add CI and npm badges
  • Document new configuration options (CORS, maxThreads, shutdownTimeout)
  • Add thread.started and error to event mapping table
  • Switch script examples from npm to pnpm
  • Add license section

v0.3.0

Minor Changes

• 820a927: ### Breaking Changes

  • cancelTask now uses AbortController to truly abort running Codex streams instead of soft polling. Custom getTurnOptions that return a signal will be merged with the internal abort signal.
  • Minimum @openai/codex-sdk version bumped to ^0.116.0.
  • Minimum @a2a-js/sdk version bumped to ^0.3.13.
  • License changed from Apache-2.0 to MIT.

  Features

  • AbortController cancellation — tasks are now cancelled immediately via AbortSignal passed to runStreamed, replacing the previous poll-based approach.
  • Thread cache LRU eviction — new maxThreads option (default 64) with automatic eviction of least-recently-used threads to prevent memory leaks. Added clearThreads() public method.
  • Configurable CORS — new cors option on CodexA2AServer to customize origin, methods, and headers (defaults unchanged).
  • Graceful shutdown — stop() now waits for active connections to drain (configurable via shutdownTimeout, default 5s) and cleans up cached threads.
  • New Codex event support — handles thread.started and ThreadErrorEvent (type: "error") from codex-sdk 0.116.
  • Robust error handling — execute() is wrapped in try-catch; resolveCodex() provides a clear error message when @openai/codex-sdk is missing.
  • Safe port scanning — findAvailablePort uses iteration instead of recursion, with a 100-port scan limit.
  • Changeset + CI — added @changesets/cli for version management and GitHub Actions workflows for CI and automated npm publishing.

  Bug Fixes

  • Fixed potential stack overflow in recursive findAvailablePort when many ports are occupied.
  • Fixed publishFailure not calling eventBus.finished() when runStreamed throws synchronously.