Release readiness review (v0.14.4 -> TARGET fc9e78c)

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

Diff

v0.14.4...fc9e78c

Release call:

🟢 GREEN LIGHT TO SHIP Patch release with narrowly scoped runtime fixes, additive sandbox configuration, matching regression tests, and no concrete blocking evidence in v0.14.4...fc9e78c2.

Scope summary:

• 24 files changed (+1126/-634); key areas touched: src/agents/run_internal/ streamed/runtime resume fixes, src/agents/extensions/sandbox/modal/ additive idle_timeout support, focused test additions, sandbox docs/examples, and version bump to 0.14.5.

Risk assessment (ordered by impact):

No material risks identified.

1. Interrupted server-conversation resume handling changed in a stateful path

• Risk: 🟢 LOW. The change affects a subtle resume/HITL path, but the behavior is constrained to preserving unsent tool outputs from the interrupted turn and has direct regression coverage.
• Evidence: Commit 16e04092 adds get_unsent_tool_call_ids_for_interrupted_state() and threads unsent_tool_call_ids through OpenAIServerConversationTracker.hydrate_from_state() so interrupted local tool outputs are not incorrectly marked as already sent; tests/test_server_conversation_tracker.py adds test_hydrate_from_state_preserves_unsent_outputs_from_interrupted_turn.
• Files: src/agents/run_internal/agent_runner_helpers.py, src/agents/run_internal/oai_conversation.py, src/agents/run.py, src/agents/run_internal/run_loop.py, tests/test_server_conversation_tracker.py
• Action: Run uv run pytest tests/test_server_conversation_tracker.py -q -k interrupted_turn; pass criteria: the new resume-path test passes and prepared input still includes the expected function-call outputs on resume.

2. Streamed terminal-output backfill adjusts final response assembly

• Risk: 🟢 LOW. This changes only the case where a streaming backend emits response.output_item.done events but leaves response.completed.output empty, and the fix is covered with targeted ordering/preservation tests.
• Evidence: Commit 81c57c5e stores ResponseOutputItemDoneEvent items in streamed_response_output and backfills terminal_response.output before step resolution when the completed response is empty; new file tests/test_streamed_terminal_output_backfill.py covers empty terminal output, preserving a non-empty terminal output, and multi-tool ordering.
• Files: src/agents/run_internal/run_loop.py, tests/test_streamed_terminal_output_backfill.py
• Action: Run uv run pytest tests/test_streamed_terminal_output_backfill.py -q; pass criteria: all new streaming regression tests pass and streamed tool calls still resolve into final output.

3. Modal sandbox idle_timeout is a user-facing additive option

• Risk: 🟢 LOW. The new field is optional, threaded end-to-end, appended without breaking constructor field order, and deserialization keeps backward compatibility by defaulting missing values to None.
• Evidence: Commit 5be06a16 adds idle_timeout to ModalSandboxClientOptions and ModalSandboxSessionState, forwards it into sandbox creation and snapshot restore, updates compatibility guards for stable field ordering, and adds focused tests for create/restore behavior.
• Files: src/agents/extensions/sandbox/modal/sandbox.py, tests/extensions/test_sandbox_modal.py, tests/sandbox/test_client_options.py, tests/sandbox/test_compatibility_guards.py
• Action: Run uv run pytest tests/extensions/test_sandbox_modal.py -q -k idle_timeout && uv run pytest tests/sandbox/test_client_options.py tests/sandbox/test_compatibility_guards.py -q -k idle_timeout; pass criteria: idle_timeout is forwarded during create/restore and legacy serialized state without the field still restores cleanly.

Notes:

• Base tag was selected from local tags only, per request: v0.14.4.
• Target was HEAD: fc9e78c2f318d1ecdb9625c6296e73c9fee55d78.
• Working tree was clean during review (git status --short returned no changes).
• This report is diff-based; no local full verification stack was rerun as part of the review.