Release 0.19.0

[github-actions]

Release readiness review (v0.18.3 -> TARGET origin/main@2fa463571)

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

Diff

v0.18.3...2fa4635

Release call:

🟢 GREEN LIGHT TO SHIP No confirmed runtime, compatibility, security, or packaging blocker; remaining findings are documentation and release-branch synchronization work.

Scope summary:

• 17 commits; 101 files changed (+10,913/-1,388).
• Key areas: Programmatic Tool Calling, nested handoff ownership, RunState schema 1.13, approval validation, Chat Completions conversion, sandbox confinement, E2B dependencies, examples, and documentation.
• Validation: uv lock --check and git diff --check passed; 1,231 focused tests effectively passed; the full documentation build completed; the real gpt-5.6 Programmatic Tool Calling example passed end to end.

Risk assessment (ordered by impact):

1. Programmatic Tool Calling lacks SDK narrative and release documentation

   • Risk: 🟡 MODERATE. The major new public feature is functional but difficult for SDK users to discover and configure correctly.
   • Evidence: ProgrammaticToolCallingTool, ToolCaller, allowed_callers, and output_json_schema are exported, and the live example passed with nine program-issued calls. However, README.md and the English guides contain no narrative PTC section or v0.19.0 release highlight.
   • Files: src/agents/tool.py, src/agents/__init__.py, examples/tools/programmatic_tool_calling.py, docs/tools.md, docs/release.md
   • Action: Add an SDK guide covering Responses-only support, caller permissions, structured output schemas, caller-link preservation, and the example. Run make build-docs; pass when the guide builds and links to the runnable example.

2. Nested-handoff guides describe the previous transcript shape

   • Risk: 🟡 MODERATE. Users enabling the beta may expect one collapsed assistant message, while the implementation now preserves lossless message items in position and creates ordered summary segments around them.
   • Evidence: RunConfig and the new ownership tests implement ordered segments, but docs/handoffs.md and docs/running_agents.md still state that history is collapsed into a single assistant message.
   • Files: src/agents/run_config.py, src/agents/handoffs/history.py, docs/handoffs.md, docs/running_agents.md
   • Action: Update the English descriptions, regenerate translations, and run make build-full-docs; pass when all languages describe ordered summary segments and lossless message preservation.

3. The release branch does not yet contain the reviewed target

   • Risk: 🟡 MODERATE. Publishing the current release branch would omit the latest runnable-documentation fixes, although runtime code is already synchronized.
   • Evidence: origin/release/v0.19.0=b2158f3e9; 2fa463571 is not its ancestor. The target-only change is documentation commit 2fa463571, while the release branch carries the expected 0.19.0 version bump.
   • Files: 29 documentation files, pyproject.toml, uv.lock
   • Action: Sync origin/main into release/v0.19.0. Pass when git merge-base --is-ancestor 2fa463571 origin/release/v0.19.0 succeeds and the remaining release-only diff is limited to the intended version files.

4. No confirmed runtime or non-beta API regression

   • Risk: 🟢 LOW. The reviewed runtime changes behaved correctly across execution, streaming, persistence, resume, provider conversion, and sandbox paths.
   • Evidence: 1,028 core tests passed. All 203 sandbox/E2B cases passed after rerunning six UnixLocal cases outside double confinement. The real gpt-5.6 PTC workflow completed successfully. New FunctionTool fields are keyword-only, top-level exports are additive, and RunState 1.13 retains older readable schemas.
   • Files: src/agents/, tests/
   • Action: Require the normal full CI gate on the synchronized release branch; pass when format, lint, typecheck, tests, and packaging checks are green.

Notes:

• Remote tags and origin/main were freshly fetched; BASE_TAG=v0.18.3, TARGET=2fa463571e76dae8ff267622f1018eaf06ffeb9f.
• The working tree is clean.
• The six initial UnixLocal failures were caused by nested macOS sandbox confinement; their exact node IDs passed outside that wrapper.
• mkdocs build completed successfully with warnings, not errors.