$ the-wire · showcase
openai-python's required call_id change lands, then gets reverted
By RepoJournal · Filed · About OpenAI · Composed from the cited sources · methodology
A breaking constructor change for local-shell output models shipped and was undone within the same window, while codex tightened proxy and TUI plumbing and openai-node cut 7.21.0.
Revert "fix(api)!: declare required local shell output call IDs (#3923)" (#3929) openai/openai-python
Mark Stuart reverted the required local shell output call IDs change, explaining that "the generated SDK changes merged before their source-of-truth changes." The tracked tree now matches the reverted commit's parent exactly, so the breaking constructor requirement from the original patch is not in the released line.
fix(api)!: declare required local shell output call IDs (#3923) openai/openai-python
The change now undone had made historical local-shell output models expose a required call_id separate from the output item id, forcing callers to supply it when constructing those items. It also dropped a phantom request_id requirement from MCP approval objects; historical call IDs gained no length restriction.
release: 3.17.0 (#3904) openai/openai-python
3.17.0 shipped with external storage configuration management and safety case retrieval, and it carries the call_id change that was reverted minutes later. If you pulled this version and construct local-shell output models, wait for the next release rather than pinning here.
Support caller-provided MITM CAs in the network proxy openai/codex
Codex's network proxy now accepts caller-provided MITM CAs through a new network.mitm_ca setting with certificate_file and private_key_file. Validation is strict: absolute paths, regular files, size limits, cert/key matching, Unix private-key permissions, with symlinks and embedded private keys rejected.
Preserve assistant answers in bounded TUI task responses (#47100) openai/codex
Eric Traut raised the TUI response budget from 999 bytes to 8,000 and removed the separate 2,000-character cap on assistant messages in read_thread and wait_threads. Assistant answers can no longer be silently truncated alongside supporting history when a thread list is compacted to fit.
fix(api): preserve model choices and improve request handling openai/openai-node
openai-node's request path now resolves resource request options before building requests while preserving parsing helpers, upload metadata, skill filenames, and admin-key auth, and it sanitizes logged values and normalizes invalid retry counts. gpt-5.1-mini and existing audio model choices are preserved, which the notes say does not imply model availability.
chore(deps): bump httpx2 from 2.7.0 to 2.12.0 in the python-security group across 1 directory openai/openai-python
The published HTTPX2 minimum rose to 2.12.0 so SDK installs pick up incremental response decompression and stream cleanup on decoding errors, with the matching HTTPcore2 constraint updated. Regression tests cover bounded, lossless decoding of a large gzip response and immediate stream closure on malformed gzip or deflate data.
Action items
- → Avoid openai-python 3.17.0 if you construct local shell output models; wait for the next release, which drops the req... openai/openai-python [plan]
