$ the-wire · showcase
Lossless sandbox conversions and environment-scoped MCP servers
By RepoJournal · Filed · About OpenAI · Composed from the cited sources · methodology
Codex tightened three of its internal contracts today: permission profiles that survive a round trip, task ownership that follows the backend, and MCP servers whose identities carry their environment instead of just their name.
Preserve legacy sandbox templates in TUI permission conversions openai/codex
Legacy sandbox templates and `PermissionProfile` now convert in both directions without probing workspace metadata, collapsing fixed roots into current-directory access, or deduplicating writable roots, and profiles whose restrictions cannot survive the round trip are rejected outright. The conversions drive sandbox inheritance in `create_thread` and the validation of read-only temporary struct...
Keep same-named MCP servers separate across environments openai/codex
MCP servers found on different executors can share a name, and resolving by name alone exposed only one of them. The new disabled-by-default `multi_environment_mcp` feature exposes environment-scoped servers, including multiple installations of the same plugin, and carries exact `McpServerKey` identities through catalogs, connections, and tool calls.
Preserve Command Center task ownership across backend switches openai/codex
Command Center retains tasks across backend switches, so opening one now switches to the backend that first observed it and rejects conflicting ownership. Restoring a retained draft accounts for history or settings that changed while another backend was displayed, and users must open tasks on other backends before proceeding.
Version answer display instructions and restrict unsupported payloads openai/codex
Free-form `AgentMessageDisplay` actions are replaced by a version `1` contract using `show`, `hide`, and `block` modes with typed annotations, replacements, block reasons, and trusted-contact handling. Invalid or unsupported display instructions are discarded and `moderated` is set to `true` when reading moderation data, with generated JSON schemas, TypeScript types, and the Python SDK updated ...
Reserve Guardian context budget for selected Code Mode exec results openai/codex
Guardian's synchronous elastic budgeting gains `transcript_selected_exec_results`, giving selected Code Mode `exec` outputs their own token reservation and entry cap separate from ordinary tool evidence, selected via `nested_tool_names` and matched against the host-recorded inventory of completed nested calls.