$ the-wire · showcase
Live Responses gains delegation tools, Codex tightens agent message delivery
By RepoJournal · Filed · About OpenAI · Composed from the cited sources · methodology
Both SDKs shipped Live Responses delegation tools and explicit tool choices while preserving the existing function and MCP contracts, and Codex reworked how injected agent messages are queued against active sampling.
feat(api)!: expand Live Responses delegation tools and choices openai/openai-node
openai-node adds delegation tool definitions and explicit tool choices for Live Responses; existing function and MCP choice names and required fields remain available, so most callers are unaffected. The breaking surface is narrow but real: code that exhaustively switches over delegation tool or choice types may need to handle additional variants.
feat(api): expand Live Responses delegation tools openai/openai-python
openai-python mirrors the change in session creation and updates, preserving the Python function and MCP choice types and keeping MCP tool-choice `name` required and non-null. The MCP tool input still exposes only `type: "mcp"`, so server configuration fields remain untyped; the PR reports 20 focused compatibility and offline transport tests passing.
Respect delivery boundaries for injected agent messages (#58988) openai/codex
Injected `agent_message` items now participate in mailbox preemption checks when the matching turn still accepts mail and `DeferMailboxPreemption` is disabled, and are treated as queue-only mail. The intent, per the commit, is that injected messages "reach active sampling before stale tool calls execute" while messages queued around a final answer wait for the next task without requesting anoth...
Add host cycling and saved defaults for new TUI conversations openai/codex
With multi-backend support enabled, the TUI shows the current host before the first submission and binds `leader b` to cycle connections, exposed as `tui.keymap.global.change_host`. Cycling preserves the conversation name and composer draft, including cursor position, attachments, pending pastes, and whether `!` was literal or a shell command.
Preserve Guardian review framing under elastic context limits openai/codex
Elastic entry caps could previously truncate review framing and authorization warnings along with supporting evidence; protected framing, host notices, and exact planned actions are now preserved, and protection metadata carries through root conversation, trusted answer, and `node_repl` evidence composition. The remaining Codex work is test and plumbing: connection events are now covered agains...
Action items