Release 0.13.1

[github-actions]

Release readiness review (v0.13.0 -> TARGET b6d9b03)

This is a release readiness report done by $final-release-review skill.

Diff

v0.13.0...b6d9b03

Release call:

🟢 GREEN LIGHT TO SHIP No concrete release-blocking regressions or breaking-versioning mismatches were identified in v0.13.0...b6d9b03.

Scope summary:

• 36 files changed (+3986/-128); key areas touched: new any-llm model/provider integration, Realtime response sequencing, MCP streamable HTTP robustness, agent-as-tool output fallback behavior, targeted regression tests, and release/docs/example updates.

Risk assessment (ordered by impact):

1. Realtime response sequencing refactor has residual concurrency sensitivity

• Risk: 🟡 MODERATE. Runtime turn sequencing could still regress in edge websocket timing patterns even though coverage is substantially expanded.
• Evidence: src/agents/realtime/openai_realtime.py introduces a new _ResponseCreateSequencer and deferred task orchestration; tests/realtime/test_openai_realtime.py adds extensive scenario coverage for queued response.create, cancellation, and error correlation.
• Files: src/agents/realtime/openai_realtime.py, tests/realtime/test_openai_realtime.py
• Action: Run uv run pytest -s tests/realtime/test_openai_realtime.py -k "response_create or response_control or close"; pass criteria: all selected sequencing tests pass with no intermittent failures.

2. Any-LLM Responses path depends on provider-private call path

• Risk: 🟡 MODERATE. Future provider SDK internals could drift because transport-header support currently falls back to a private provider entrypoint.
• Evidence: src/agents/extensions/models/any_llm_model.py uses provider._aresponses(...) with a compatibility shim and comments explaining public aresponses() validation limitations in any-llm 1.11.0.
• Files: src/agents/extensions/models/any_llm_model.py, tests/models/test_any_llm_model.py, pyproject.toml
• Action: Run uv run pytest -s tests/models/test_any_llm_model.py on Python 3.11+ with openai-agents[any-llm]; pass criteria: all AnyLLM unit tests pass, including responses/chat selection and streaming normalization paths.

3. MCP initialized-notification tolerance introduces alternate transport path

• Risk: 🟢 LOW. Behavior is opt-in and constrained, but transport customization paths should be validated to avoid accidental masking of non-initialized failures.
• Evidence: src/agents/mcp/server.py adds _InitializedNotificationTolerantStreamableHTTPTransport and ignore_initialized_notification_failure wiring; tests verify initialized failures are tolerated while other failures still raise.
• Files: src/agents/mcp/server.py, tests/mcp/test_streamable_http_client_factory.py
• Action: Run uv run pytest -s tests/mcp/test_streamable_http_client_factory.py; pass criteria: initialized-notification tolerance tests pass and non-initialized failure propagation test remains strict.

Notes:

• Working tree was clean during review.
• Base tag was selected from local tags only (git tag -l 'v*' --sort=-v:refname | head -n1 → v0.13.0) per request.
• Target commit was git rev-parse HEAD → b6d9b03051e383ec98842c9d014bce75ee3865bc.
• This assessment is diff-based; no full local verification stack was executed as part of this report.