$ the-wire · showcase
Typed IDs in Codex traces, bounded guardian retries
By RepoJournal · Filed · About OpenAI · Composed from the cited sources · methodology
Codex spent the day tightening its own plumbing: distinct types for transcript and wire IDs, deferral of Default mode prompts until Plan mode is used, and a hard one-retry ceiling on guardian resampling, while the SDKs clarified the one field most likely to break a Chat Completions call.
Distinguish conversation item IDs from Responses IDs in traces openai/codex
The rollout trace reducer now types transcript node IDs as a Rust newtype, ConversationItemId, and keeps ResponseItemId for receipt and agent delivery lookups; both were plain strings, so callers could mix the two up. The typed IDs are carried through snapshot reconciliation and compaction.
Limit guardian response resampling to one retry openai/codex
Guardian response resampling is capped at one retry: the original prompt is consumed when the retry is scheduled, and the second verdict stays buffered and replays after usage accounting even if it meets the resampling criteria. A retried response that also qualified used to trigger another model request and hold the turn open; usage from the discarded first response still counts toward spending.
Defer built-in Default mode instructions until after Plan mode openai/codex
Bundled and model-catalog Default mode instructions are now omitted until a thread actually uses Plan mode, which fixes threads that carried a collaboration-mode prompt with no Plan mode instructions to supersede. Plan mode instructions emit on entry and Default mode instructions on the switch back, with caller-provided instructions preserved through compaction and resume.
fix(api): clarify when to omit completion choice count openai/openai-python
The API docs now state plainly that newer models reject an explicit n in Chat Completions, n=1 included, and that omitting n in legacy Completions generates one completion per prompt. The same clarification landed in openai-node and openai-python, so a call that passes n out of habit is the likeliest thing to start failing against the newer models.
Keep Cloud threads current in the Agents overview openai/codex
The Agents overview now updates Cloud thread status from stream notifications and polls every 10 seconds while visible with no stream attached, refreshing the visible history while holding the Show more cursor. Separately, the code-mode host and Windows setup helper got signal-specific OpenTelemetry constructors, try_new_http_traces and try_new_http_metrics, so exporter builders for unused sign...
Action items
- → Check Chat Completions and Completions calls that pass n explicitly, including n=1, against newer models openai/openai-python [plan]