github openai/openai-agents-js @openai/agents-core@0.19.0

Minor Changes

  • 39d5e9e: fix: bound agent tool streaming callbacks to 1024 pending events by default, with onStreamMaxPendingEvents to adjust the limit or restore unlimited buffering with null.
  • f0f3066: fix: reject started checkpoints without completed initial input validation, including older snapshots without completion evidence; restart with the original input and context.
  • aeef9e1: fix: bind resumed MCP calls to their original recipients and require a fresh run for historical pending function calls without recipient provenance.
  • 39decd7: fix: bind conditional tool approvals to isolated normalized execution input in core and Realtime, re-evaluate current policies on durable approval resumes, reject uncopyable normalized values before conditional approval, and preserve invalid-input handling (#1914, #1915).
  • 3515c29: fix: redact default function-tool and Realtime approval parse failures; require explicit invocation policy for sensitive error traces and clarify when custom errorFunction feedback applies.

Patch Changes

  • c55d64d: fix: add an optional rollback item budget for Responses session compaction.
  • c345409: feat: add optional page limits for automatic MCP tool listing
  • a0b1c6f: fix: Preserve current permanent function approval decisions when restoring nested runs from saved state.
  • 9c94138: fix: honor the invoking Runner's sensitive-data tracing policy in Codex spans, including mixed CommonJS and ESM applications.
  • c3e02d2: fix: Refresh cached sandbox application tools and preserve public agent identity in tool filters.
  • f80c91c: fix: Preserve guarded final outputs, including empty strings, in default Agent.asTool output extraction.
  • 3949bb1: fix: Apply static MCP tool filters when discovering tools without full agent and run context.
  • 186f6ec: fix: Trim local workspace logical path separators in linear time.
  • fdaf0a6: fix: handle MCP tools from one server whose names normalize to the same function tool name
  • 817161d: build: Generate package metadata with native Node.js TypeScript support.
  • 20e5946: chore: Run PTY signal regression tests with native Node.js instead of the tsx CLI.
  • ec8bb33: fix: Preserve post-construction agent storage settings through model requests and cloning.
  • 9623a22: fix: preserve multiline skill frontmatter descriptions and keep nested fields out of skill index metadata.
  • 58844c7: fix: Withhold unvetted final assistant output from session history when output guardrails fail to complete.
  • d0e5d43: fix: Keep parent transport overrides out of agent tool runs with an explicit Model while preserving common settings and explicit child overrides.
  • 2127475: fix: preserve stored history excluded from the latest successful model exchange during automatic compaction.
  • b9a4fb8: fix: respect sensitive-data trace settings for escaping tool callbacks.
  • 7f135af: fix: update compatible runtime and development dependencies.
  • ee65014: fix: Reject read-only host binds combined with privileged in-container Docker storage before provider or filesystem effects.
  • 00eef0c: fix: Reject selected handoff input filters with server-managed conversationId or previousResponseId before handoff side effects; use client-managed history or a Session without continuation options instead.
  • 218a2b8: fix: snapshot configured tool candidates during asynchronous enablement so in-place array changes do not duplicate or skip tools.
  • 8ac97df: fix: Reject computer calls without aborting the run when onSafetyCheck explicitly returns false.
  • 96aa754: fix: respect sensitive tracing settings for computer safety-check callback errors and enclosing run-span error names.
  • 7efc536: fix: update compatible workspace dependencies while preserving supported provider and Expo versions.
  • 8a26e30: fix: Serialize hosted MCP approval tool references without live configuration.
  • 4ef9008: fix: Preserve cached MCP tools for unrelated servers with colon-suffixed names when invalidating another server's cache (#1976).
  • f507590: fix: accept deferred hosted MCP listings before tool-search results (#1978)
  • d187747: fix: Apply current MCP tool filters on every lookup while reusing cached tool discovery.
  • fc2d2a4: feat: allow configuring phase-two memory consolidation turns.
  • 0bdea85: fix: Compute sandbox parent paths in linear time while preserving existing path semantics.
  • a9b543b: fix: preserve prompt formatting when rendering large multiline skill descriptions.
  • 06bbcc0: chore: migrate tracing test mocks to Vitest 5.

Don't miss a new openai-agents-js release

NewReleases is sending notifications on new releases.