RepoJournal

$ cat openai/week/2026-10-05.log

OpenAI

OpenAI

the week in review · Oct 5 – Oct 11, 2026

Codex hands model defaults to the server

By RepoJournal · composed from the cited sources · human-reviewed weekly · methodology

♥

Both SDKs shipped standalone Decisions and agent tool handling, while Codex picked up Bazel release artifacts and a deny-read ACL recovery fix.

302 commits 323 PRs merged 30 releases 7 briefings covered

all openai reviews →

Use server model defaults for connected TUI fresh starts openai/codex

by copyberry

Fresh TUI threads connected to an app server could bootstrap with stale client model settings before the server configuration was read, and an empty model catalog could block startup even when managed new-thread defaults supplied a model. The server is now the source of truth on fresh starts, so if you drive Codex through an app server, your client-side model config no longer wins at session st...

Honor server reasoning summary defaults in new TUI threads openai/codex

by copyberry

The same correction applied to reasoning summaries: new embedded TUI threads forced summaries off when none was explicitly configured, overriding the model's default, and client settings could override the destination server's configuration. Expect summary behavior in embedded threads to follow the server unless you configure it deliberately, and be aware this changes what you currently see.

feat(api): add standalone Decisions support openai/openai-python

by dpiet-oai

Both the synchronous Python client and the Node client gained a standalone decisions.create that accepts text or user messages with inline images, plus typed predicate, choice, score and refusal answers. The Python client preserves boolean and string choice values when reading responses, so you can branch on the literal value rather than coercing it.

feat(agents): run local tools during streamed session creation openai/openai-python

by apcha-oai

Local function tools now run on the first turn through sessions.create(stream=True), so an application can create a session with environment={"type": "none"}, handle its initial tool calls, and collect the answer in a single stream. Previously automatic handlers only fired on a following sessions.stream() turn, which forced callers to dispatch the initial calls themselves.

feat(agents): run local tools during session creation streams openai/openai-node

by apcha-oai

The Node SDK mirrors the Python change and adds an onToolError callback: when argument validation, application execution, or output serialization fails, the handler submits a sanitized failure to the model while your code gets the details locally. The model-visible result is unchanged, so this is observability rather than behavior.

feat(tools): support tool search (defer_loading=true) in typed function helpers openai/openai-node

by apcha-oai

Typed function helpers now configure deferred discovery without mutating the tool definition or losing argument parsing: the example moved from setting tool.defer_loading = true after construction to passing it into the helper. If you were flipping defer_loading on the returned object, migrate to the helper argument.

Recover malformed Windows deny-read ACL state safely openai/codex

by copyberry

Malformed deny_read_acl_state.json previously broke deny-read ACL reconciliation on Windows. Recovery now restores bookkeeping without dropping unknown existing restrictions or rewriting a linked file's contents, so a corrupted state file no longer leaves you with restrictions changed out from under you.

Build Bazel release artifacts alongside Cargo artifacts openai/codex

by copyberry

Cargo and Bazel build matrices now cover Linux, macOS and Windows and run through the same signing, packaging and verification path, with Bazel binaries, packages and symbols published under a -bazel suffix alongside the existing Cargo artifacts. Alongside these: a prompt prefix compatibility test gates changes to the default model-visible prefix, URL destinations in wrapped hook details render...

$ ls openai/week/ # the briefings behind this review

Keep up with OpenAI in about 3 minutes: what actually shipped — the commits, pull requests, releases, and security advisories that matter.

or

fair warning: these emails are deeply technical. diffs, version numbers, CVEs, benchmark deltas. if that's not your idea of a good read, this isn't your newsletter.

Folds into your digest — weekly by default, monthly if you prefer. Unsubscribe in one click.

all openai reviews →