codex

Key Features

• GPT-6.1 Terra model added to Vertex AI and Amazon Bedrock catalogs.
• TUI supports keyboard-based transcript selection (word/line/page motions), integrated search, and custom shortcuts.
• codex queue --resume command allows resuming threads after queuing.
• TUI introduces safety block explanations and safer follow-up prompt selection features.
• Added codex doctor --network-only command.
• Support for incremental skill catalogs (skill_catalog_deltas) and native dollar thread usage display in ENT26 plans.

Architecture and Build Changes

• Guardian review logic (state, context, evidence, etc.) separated into reviewer extension/crate.
• Bazel builds transitioned to source-built V8 (target-matched Rust bindings, Windows MSVC archive support).
• Standalone executor included in canonical CLI and app-server release packages.
• Development/release builds standardized on Node.js 24.22.1 and Rust 1.95.0.

Performance and Stability

• Default thread unload delay reduced to 1 minute.
• Concurrent loading of up to 8 plugins and concurrent image availability checks limited to 8.
• Reduced overhead for remote plugin catalog cache probing.
• Added memory baseline file limits and incomplete snapshot recovery features.
• Fixed app-server race condition and immediate release of completed runtimes.

Bug Fixes

• Preserved approval and permission settings when resuming/forking threads.
• Preserved local image attachments when restoring terminated composer text.
• Fixed macOS installation path errors (non-ASCII, in use).
• Restored child threads when unarchiving parent threads.
• Prevented early compaction by maintaining reasoning-token accounting during resume/fork.
• Extended GitHub release read retries and added jitter.

Other

• Supported negotiated compression in WebSocket remote app-server client.
• Enforced DELETE journaling in NFS environments and allowed DB migrations with shared file permissions.
• Fixed OpenTelemetry header redaction and trace exporter activation during account switching.