Release 0.14.7

[github-actions]

Release readiness review (v0.14.6 -> TARGET 646bc33)

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

Diff

v0.14.6...646bc33

Release call:

🟢 GREEN LIGHT TO SHIP No blocking regression, packaging break, or unversioned breaking public API change was identified in the release diff.

Scope summary:

• 14 files changed (+506/-28); key areas touched: sandbox archive extraction safety, sandbox local artifact copy safety, tool item convenience helpers, Responses API input normalization, sandbox compaction/memory behavior, focused tests, and version bump to 0.14.7.
• Commit range includes 7 commits: archive validation hardening (fix: tighten tar and zip member validation #3028), tool item helper additions (feat: #2886 add convenience properties (tool_name, call_id) to tool items #3027), symlinked LocalFile rejection (fix: reject symlinked LocalFile sources #2972), Responses API unset-field cleanup (Fix: remove unset fields from calls to Responses API #3026), Phase 2 memory turn-limit increase ([sandbox] Raise Phase 2 memory consolidation turn limit #3038), GPT-5.5 compaction aliases (fix: add GPT-5.5 aliases to sandbox compaction #3039), and the 0.14.7 version bump.

Risk assessment (ordered by impact):

No material risks identified.

1. Sandbox archive and local file validation is stricter

   • Risk: 🟢 LOW. This tightens unsafe archive/member and symlink handling; the only likely user-visible effect is rejection of inputs that previously depended on unsafe Windows-style archive paths, non-directory path conflicts, or symlinked LocalFile sources.
   • Evidence: 2eb8713b adds Windows path rejection, duplicate/non-directory archive path validation, and validate_tarfile(..., allow_symlinks=False) in extraction; ba889de4 routes LocalFile through the safer LocalDir open path and rejects symlinked source ancestors/leaves before checksum/write.
   • Files: src/agents/sandbox/session/archive_extraction.py, src/agents/sandbox/util/tar_utils.py, src/agents/sandbox/entries/artifacts.py, tests/sandbox/test_extract.py, tests/sandbox/test_tar_utils.py, tests/sandbox/test_entries.py
   • Action: Run uv run pytest tests/sandbox/test_extract.py tests/sandbox/test_tar_utils.py tests/sandbox/test_entries.py; pass criteria: all new archive rejection and local symlink rejection tests pass.

2. Tool item helpers and Responses API input cleanup are additive or bug-fix scoped

   • Risk: 🟢 LOW. ToolCallItem.tool_name, ToolCallItem.call_id, and ToolCallOutputItem.call_id are additive properties, while dropping dict status: None before model input reduces invalid API payload risk without changing populated statuses.
   • Evidence: 5df41d30 adds convenience properties with tests for typed raw items and dict raw items; 8025ed0b removes only status keys whose value is None in run_item_to_input_item.
   • Files: src/agents/items.py, src/agents/run_internal/items.py, tests/test_items_helpers.py
   • Action: Run uv run pytest tests/test_items_helpers.py; pass criteria: helper property tests and existing input item normalization tests pass.

3. Sandbox compaction and memory tuning are scoped

   • Risk: 🟢 LOW. GPT-5.5 aliases are mapped to the existing large context-window bucket, and Phase 2 memory consolidation raises only that agent run’s max_turns, which may increase work for unusually long consolidations but does not alter general runner defaults.
   • Evidence: ebdb0f2e adds GPT-5.5 snapshot/pro aliases plus parametrized compaction tests; b3688db7 changes only run_phase_two(... max_turns=500).
   • Files: src/agents/sandbox/capabilities/compaction.py, src/agents/sandbox/memory/phase_two.py, tests/sandbox/test_compaction.py
   • Action: Run uv run pytest tests/sandbox/test_compaction.py; pass criteria: model alias context-window expectations pass. For extra release confidence, run a targeted Phase 2 memory smoke if available in CI.

Notes:

• Latest local release tag used: v0.14.6.
• TARGET used: 646bc33f4ddd5cbaa6ccb801fd629ca7fea98a6e.
• Working tree status was clean.
• Local verification was not run as part of this review; this report assumes normal CI verification covers make format, make lint, make typecheck, and make tests.