Release 0.16.0

[github-actions]

Release readiness review (v0.15.3 -> TARGET ea965bb)

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

Diff

v0.15.3...ea965bb

Release call:

🟢 GREEN LIGHT TO SHIP No confirmed blocking regression, breaking API mismatch, data-loss issue, security regression, or packaging/runtime break was found; remaining items are targeted validation and release-note follow-ups.

Scope summary:

• 54 files changed (+2480/-925); key areas touched: runtime model defaults and model settings resolution, MCP tool-name prefixing, local function-tool concurrency, max_turns=None, RunState schema 1.10, Chat Completions streaming tool-call indexes, sandbox tar/symlink safety, docs/translations, and tests.
• Version bump is 0.15.3 -> 0.16.0.
• Commit range includes 12 commits: sandbox symlink hardening, max_turns=None, default model switch, translated docs, hashable sandbox permissions, MCP server-prefixed tool names, tool concurrency config, Chat Completions streaming index fix, realtime docs, and release version bump.

Risk assessment (ordered by impact):

1. Default model changed from gpt-4.1 to gpt-5.4-mini

   • Risk: 🟡 MODERATE. This is an intentional user-facing default behavior change when users do not explicitly set a model, with possible output, latency, pricing, or availability differences.
   • Evidence: src/agents/models/default_models.py now returns gpt-5.4-mini; src/agents/agent.py documentation and implicit default settings logic were updated; pyproject.toml bumps to 0.16.0.
   • Files: src/agents/models/default_models.py, src/agents/agent.py, src/agents/run_internal/turn_preparation.py, tests/models/test_default_models.py, tests/test_run_config.py
   • Action: Confirm release notes call out the default model change, then run uv run pytest tests/models/test_default_models.py tests/test_run_config.py -q; pass criteria: tests pass and the note tells users to set OPENAI_DEFAULT_MODEL or Agent(model=...) to preserve the prior default.

2. RunState schema bumped to 1.10 for nullable max_turns

   • Risk: 🟡 MODERATE. New snapshots that serialize max_turns: null require the newer SDK, while older SDKs intentionally reject newer schema versions.
   • Evidence: CURRENT_SCHEMA_VERSION changed from 1.9 to 1.10, SCHEMA_VERSION_SUMMARIES adds nullable max_turns, and tests cover None round-trip.
   • Files: src/agents/run_state.py, tests/test_run_state.py, tests/test_max_turns.py
   • Action: Run uv run pytest tests/test_run_state.py -q -k 'schema or max_turns_none' tests/test_max_turns.py -q; pass criteria: schema/version tests and non-streamed/streamed max_turns=None tests pass. Include a release-note line that snapshots written by 0.16.0 use schema 1.10.

3. Sandbox hydration now rejects external symlink targets

   • Risk: 🟡 MODERATE. This security hardening can reject workspace archives that previously hydrated successfully if they contain absolute symlink targets or links escaping the archive root.
   • Evidence: sandbox backends now call validate_tar_bytes/validate_tarfile/safe_extract_tarfile with allow_external_symlink_targets=False; tests assert rejection before upload or extraction.
   • Files: src/agents/sandbox/util/tar_utils.py, src/agents/extensions/sandbox/*/sandbox.py, src/agents/sandbox/sandboxes/docker.py, src/agents/sandbox/sandboxes/unix_local.py, tests/sandbox/test_tar_utils.py, tests/sandbox/test_extract.py, tests/extensions/sandbox/test_vercel.py
   • Action: Run uv run pytest tests/sandbox/test_tar_utils.py tests/sandbox/test_extract.py tests/extensions/sandbox/test_vercel.py -q; pass criteria: strict symlink rejection tests pass. Mention this as a security hardening behavior change in release notes.

4. Opt-in MCP server-prefixed tool names add new public naming behavior

   • Risk: 🟢 LOW. The behavior is disabled by default and has targeted tests for collisions, normalization, handoff/local-tool reservation, and streaming/non-streaming runner paths.
   • Evidence: MCPConfig adds include_server_in_tool_names; MCPUtil builds deterministic ASCII-safe names up to 64 chars and preserves original MCP invocation names; docs describe the option.
   • Files: src/agents/agent.py, src/agents/mcp/util.py, docs/mcp.md, tests/mcp/test_mcp_util.py, tests/mcp/test_runner_calls_mcp.py
   • Action: Run uv run pytest tests/mcp/test_mcp_util.py tests/mcp/test_runner_calls_mcp.py -q; pass criteria: prefixed-name and duplicate-name tests pass.

5. Local function-tool concurrency limit is a new public RunConfig option

   • Risk: 🟢 LOW. Default None preserves existing “start all calls” behavior, and the new dataclass is appended after sandbox, preserving positional compatibility.
   • Evidence: ToolExecutionConfig(max_function_tool_concurrency=None) defaults to old behavior; constructor compatibility tests assert sandbox position is preserved and tool_execution is appended.
   • Files: src/agents/run_config.py, src/agents/run.py, src/agents/__init__.py, src/agents/run_internal/tool_execution.py, tests/test_run_step_execution.py, tests/test_source_compat_constructors.py
   • Action: Run uv run pytest tests/test_run_step_execution.py tests/test_source_compat_constructors.py -q -k 'concurrency or tool_execution or run_config'; pass criteria: concurrency ordering/failure tests and positional compatibility tests pass.

6. Chat Completions streaming tool-call output indexes were stabilized

   • Risk: 🟢 LOW. The change targets a regression-prone streaming edge case but includes focused tests for fallback and mixed streamed/fallback tool calls.
   • Evidence: StreamingState.function_call_output_idx now allocates stable output indexes when tool-call deltas appear; new tests assert unique indexes and preserved ordering.
   • Files: src/agents/models/chatcmpl_stream_handler.py, tests/models/test_openai_chatcompletions_stream.py
   • Action: Run uv run pytest tests/models/test_openai_chatcompletions_stream.py -q; pass criteria: streaming function-call index tests pass.

Notes:

• Latest release tag was determined from local tags only with git tag -l 'v*' --sort=-v:refname | head -n1: v0.15.3.
• TARGET was determined with git rev-parse HEAD: ea965bb13f6735085cea49ef7a2b34c2cf7354f6.
• Working tree status was clean.
• Local verification was not run as part of this report; assume CI has run the mandatory release checks unless this job is configured otherwise.