$ the-wire · showcase
Codex makes rollout failures loud, SDKs grow agent compaction
By RepoJournal · Filed · About OpenAI · Composed from the cited sources · methodology
Codex now fails thread startup when an enabled rollout recorder cannot initialize and returns web search errors to the model instead of aborting the turn, while the Node and Python SDKs pick up matching agent context compaction and experimental time awareness settings.
Make rollout recording failures and completion explicit openai/codex
Enabled rollout tracing used to fail silently; startup or resume now fails if the recorder cannot initialize, with session resources cleaned up, and recording stops after the first write or serialization failure. A terminal task status no longer implies a complete bundle, so trace consumers should treat interrupted tasks as partial recordings.
Support managed network requirements under `permissions.network` openai/codex
Shared network constraints move into [permissions.network] in requirements.toml alongside [permissions.filesystem], and are enforced through config loading, network approvals, Windows sandbox policy validation, and configRequirements/read, with the source surfaced in /debug-config. [network_proxy] and [experimental_network] still work, but mixing either with [permissions.network] is now rejected.
fix: preserve refusal output and authentication in WebSocket helpers openai/openai-python
Three Responses WebSocket edge cases in the Python SDK: refusal deltas now accumulate through the normal output-item and content-part lifecycle, Azure API-key auth survives a connection replacement including explicit caller header overrides, and local parsed metadata is dropped when replaying typed output items while caller-provided dictionaries are kept. The PR reports 24 focused regression te...
feat(api): add agent context compaction settings openai/openai-node
Beta agents, sessions, and scheduled agent runs accept context compaction settings, with the async setting choosing background or blocking compaction, and effective settings readable back from each configuration. The Python SDK carries the same feature as async_.
Return standalone web search failures to the model openai/codex
Standalone web search errors no longer abort the turn as fatal failures: client setup errors go back to the model, failed requests become unsuccessful tool outputs, and error messages are capped at 1,024 bytes, truncated at a UTF-8 character boundary with a marker. Also landed: an optional markup PNG overlay for image edits that must match the first image's dimensions and cannot be combined wit...
