Release readiness review (v0.18.3 -> TARGET da82ee7)

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

Diff

v0.18.3...da82ee7

Release call:

🟢 GREEN LIGHT TO SHIP No confirmed regression, incompatible versioning, security issue, or release-critical packaging failure remains.

Scope summary:

• 291 files changed (+29,472/-2,289); key areas include Programmatic Tool Calling, RunState and handoff persistence, model retries, sandbox providers and mounts, configuration coercion, sensitive logging, callable tools, Realtime/voice paths, documentation, and packaged integration coverage.

Risk assessment (ordered by impact):

1. Programmatic Tool Calling and durable resume contract

   • Risk: 🟢 LOW. This is a substantial opt-in Responses feature, but its durable-state boundary is explicitly versioned and older state remains readable.
   • Evidence: ProgrammaticToolCallingTool, ToolCaller, allowed_callers, caller validation, replay filtering, and raw program/program_output preservation were added together. RunState now writes schema 1.13, retains prior supported versions, and rejects PTC data falsely labeled with an older schema.
   • Files: src/agents/tool.py, src/agents/run_state.py, src/agents/run_internal/tool_caller.py, src/agents/extensions/handoff_filters.py, tests/test_programmatic_tool_calling.py, tests/test_run_state.py
   • Action: Ensure generated 0.19.0 release notes state that PTC is Responses-only and opt-in, and describe the schema 1.13 resume boundary. Pass when those exact constraints appear in the published notes.

2. Sandbox/provider and distribution expansion

   • Risk: 🟢 LOW. Vercel cloud mounts and updated E2B pins expand optional provider behavior without changing the core dependency set.
   • Evidence: Vercel mount strategy has dedicated implementation and extensive provider tests; E2B remains an optional extra pinned to e2b==2.31.0 and e2b-code-interpreter==2.8.1. The exact origin/release/v0.19.0 tree produced installable wheel and sdist artifacts, preserved Python >=3.10, and excluded root automation/integration directories.
   • Files: src/agents/extensions/sandbox/vercel/, src/agents/extensions/sandbox/e2b/sandbox.py, pyproject.toml, .github/scripts/run_integration_tests.py, integration_tests/packaging/
   • Action: Preserve the provider names, optional-extra requirements, and E2B pin changes in generated release notes. Pass when users can identify which installations are affected.

3. Configuration and callable-tool API expansion

   • Risk: 🟢 LOW. Public configuration now accepts equivalent dictionaries, and plain callable instances are supported without changing existing positional constructor meaning.
   • Evidence: RunConfig, ModelSettings, session/sandbox settings, and agent registration retain typed-object inputs while coercing dictionaries through their owning models. New FunctionTool fields are keyword-only. Plain sync/async callable objects use the existing schema/invocation pipeline; ambiguous partials, wrappers, descriptors, and unresolved generic annotations fail during construction with an explicit wrapper-function alternative. The exact-target focused suite exercised these paths successfully.
   • Files: src/agents/_config_coercion.py, src/agents/run_config.py, src/agents/model_settings.py, src/agents/tool.py, src/agents/function_schema.py, src/agents/decorators.py
   • Action: Mention dictionary configuration support, the tool decorator alias, and plain callable-object support in the release notes. Pass when the wording does not imply support for arbitrary callable wrapper graphs.

4. Sensitive diagnostic handling

   • Risk: 🟢 LOW. Default-redacted logging now covers model, tool, Realtime, MCP, RunState, tracing, sandbox, and cleanup error paths without suppressing operational messages.
   • Evidence: Shared logging helpers avoid inspecting or attaching exception payloads while redaction is active; Realtime validation and RunState deserialization remove event, agent, and item identifiers from redacted records. Diagnostic payloads remain available only when the corresponding model/tool logging policies are explicitly disabled.
   • Files: src/agents/logger.py, src/agents/realtime/openai_realtime.py, src/agents/run_state.py, src/agents/mcp/_logging.py, tests/test_error_logging_redaction.py
   • Action: Retain the exact OPENAI_AGENTS_DONT_LOG_MODEL_DATA=0 and OPENAI_AGENTS_DONT_LOG_TOOL_DATA=0 opt-in wording in release documentation. Pass when secure defaults and diagnostic opt-in are unambiguous.

5. Responses WebSocket overload retry

   • Risk: 🟢 LOW. Pre-response overload errors can participate in opt-in runner-managed retries, while partial responses and replay-unsafe stateful requests remain protected.
   • Evidence: Only ResponsesWebSocketError with event type error and code server_is_overloaded receives positive provider advice. Errors marked after response activity retain replay_safety="unsafe". The official API guidance recommends retrying overload failures with backoff, matching this bounded implementation.
   • Files: src/agents/models/openai_responses.py, src/agents/run_internal/model_retry.py, tests/models/test_openai_responses.py
   • Action: State that overload retries require configured runner retry settings and do not authorize unsafe replay. Pass when the release notes do not imply unconditional automatic retries.

Notes:

• BASE was resolved from freshly fetched remote tags as v0.18.3; TARGET was freshly fetched origin/main at da82ee786ad5968f52001f271a9144bf000aa039.
• Refreshed origin/release/v0.19.0 at 330f8c537daf65f99b1221ec3d7eac5368effd16 contains TARGET and differs only by the 0.19.0 version updates in pyproject.toml and uv.lock.
• Earlier review context was used only to focus the incremental audit; refs, ancestry, full diff scope, exact-target checks, and exact release-tree packaging were refreshed in this review.
• No repository files were changed.