Release 0.17.0

[github-actions]

Release readiness review (v0.16.1 -> TARGET 38ae4dd)

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

Diff

v0.16.1...38ae4dd

Release call:

🟢 GREEN LIGHT TO SHIP No release-blocking regressions found; the main user-facing behavior changes are covered by a minor version bump and targeted tests.

Scope summary:

• 29 files changed (+759/-110); key areas touched: realtime default/session config, sandbox local artifact source confinement, Responses extra_args handling, examples auto-run skips, version bump to 0.17.0.

Risk assessment (ordered by impact):

1. Realtime default model changes from gpt-realtime-1.5 to gpt-realtime-2

   • Risk: 🟡 MODERATE. Users who rely on implicit realtime defaults will move to a new backend model, which can affect availability, behavior, latency, or cost.
   • Evidence: src/agents/realtime/openai_realtime.py changes DEFAULT_REALTIME_MODEL to gpt-realtime-2; realtime examples and tests were updated to expect gpt-realtime-2; pyproject.toml bumps the package to 0.17.0.
   • Files: src/agents/realtime/openai_realtime.py, src/agents/realtime/config.py, tests/realtime/test_openai_realtime.py, examples/realtime/**
   • Action: Run make tests or, time-constrained, pytest tests/realtime/test_openai_realtime.py tests/realtime/test_session_payload_and_formats.py; pass criteria: default connection/session payload tests continue to assert gpt-realtime-2 and no realtime regressions fail.

2. Sandbox local artifact sources are now constrained to base dir unless granted

   • Risk: 🟡 MODERATE. Existing sandbox manifests that copy local files or lazy skills from paths outside base_dir must now declare SandboxPathGrant, which is an intentional security hardening but a migration-sensitive behavior change.
   • Evidence: commit f47d4869 adds extra_path_grants checks in src/agents/sandbox/entries/artifacts.py and src/agents/sandbox/capabilities/skills.py; examples were updated to add SandboxPathGrant; tests now assert outside-base rejection and granted-source success.
   • Files: src/agents/sandbox/entries/artifacts.py, src/agents/sandbox/capabilities/skills.py, tests/sandbox/test_entries.py, tests/sandbox/capabilities/test_skills_capability.py
   • Action: Run pytest tests/sandbox/test_entries.py tests/sandbox/capabilities/test_skills_capability.py; pass criteria: outside-base paths fail with outside_base_dir, granted paths still materialize, and lazy skill metadata respects grants.

3. Responses extra_args collision behavior changed for omitted kwargs

   • Risk: 🟢 LOW. This fixes a false duplicate-argument rejection when the SDK would otherwise omit the explicit field, with narrow blast radius and direct test coverage.
   • Evidence: src/agents/models/openai_responses.py now ignores duplicate keys whose explicit create kwarg is an OpenAI omitted value; tests/models/test_openai_responses.py adds coverage for context_management supplied through extra_args.
   • Files: src/agents/models/openai_responses.py, tests/models/test_openai_responses.py
   • Action: Run pytest tests/models/test_openai_responses.py -k extra_arg; pass criteria: omitted explicit kwargs allow extra_args, while true duplicate explicit values still raise.

4. Example auto-run behavior now skips more prerequisite-bound examples

   • Risk: 🟢 LOW. CI/example automation should become less flaky, but the changed skip set can reduce automated coverage for examples that require external services.
   • Evidence: examples/run_examples.py adds sandbox/cloud/Codex examples to DEFAULT_AUTO_SKIP; tests/test_run_examples_script.py asserts the new skip set and keeps examples/tools/computer_use.py enabled.
   • Files: examples/run_examples.py, tests/test_run_examples_script.py
   • Action: Run pytest tests/test_run_examples_script.py; pass criteria: prerequisite-bound examples remain skipped and the computer-use example remains enabled.

Notes:

• Latest release tag was determined from local tags only: v0.16.1.
• TARGET is current HEAD: 38ae4dd8b46547cbb53bc1dcd0610238ba09c094.
• Working tree status was clean.
• uv was not available in this CI shell, so I did not run local verification; this is not a release blocker under the skill gate policy.