Codex 更新紀錄

Radar Ultra

217 個版本

0.99.0

CLIGitHub 原文

AI 摘要

  • Direct shell commands can now run while a turn is in progress without interrupting it.
  • You can attach GIF and WebP images, and use /statusline to choose which metadata appears in the TUI footer.
  • The resume picker can sort sessions by creation time or last-updated time, with an indicator showing the selected mode.
  • Windows sign-in startup and several TUI input behaviors have been fixed, including option-label wrapping, Tab in idle steer mode, history cursor placement, and restoring image placeholders in drafts.
  • Admins can set requirements that restrict web search modes and define network constraints through requirements.toml.

由 AI 依官方說明整理,以原文為準

26.210

App官方原文

AI 摘要

  • You can now search for branches in the branch picker.
  • Typing plan in the composer now gives clearer guidance for entering plan mode.
  • Codex now supports parallel approvals.
  • This release also includes performance improvements and bug fixes; no further details are provided.

由 AI 依官方說明整理,以原文為準

26.208

App官方原文

AI 摘要

  • MCP and personality actions are now available in the command palette.
  • Follow-up actions now queue by default.
  • The update includes additional performance improvements and bug fixes; the notes don’t specify details.

由 AI 依官方說明整理,以原文為準

26.206

App官方原文

AI 摘要

  • You can now reveal a referenced file directly in your computer’s file manager.
  • The review pane no longer has an overall limit on diff size, improving its handling of large reviews.
  • The update also includes other performance improvements and bug fixes, without further details in the notes.

由 AI 依官方說明整理,以原文為準

0.98.0

CLIGitHub 原文

AI 摘要

  • Steer mode is stable and on by default. While a task is running, Enter sends input immediately; Tab queues it as a follow-up.
  • Model instructions are now applied correctly when you switch models mid-conversation or resume with a different model.
  • The TypeScript SDK now resumes threads with local images without accidentally starting a new session.
  • Remote conversation trimming now uses matching instructions for estimating and preparing the compacted context, improving trim accuracy and avoiding context overflows.
  • Cloud requirements now reload immediately after login, and the default assistant personality is Pragmatic again.

由 AI 依官方說明整理,以原文為準

0.97.0

CLIGitHub 原文

AI 摘要

  • You can choose “Allow and remember” for MCP/App tool approvals, so repeated calls to the same tool are auto-approved for the rest of that session.
  • Skill file changes are now detected live, without restarting.
  • You can use /debug-config in the TUI to inspect the effective configuration.
  • Logs can be redirected using the new log_dir setting, including through -c overrides.
  • The apps/connectors picker’s description column no longer jitters, and the TUI “working” indicator is more stable during startup and early execution.

由 AI 依官方說明整理,以原文為準

26.205

App官方原文

AI 摘要

  • Codex now supports GPT-5.3-Codex.
  • You can send Codex a message while it’s working to steer what it does.
  • You can attach or drop any file type.
  • The app’s flickering has been fixed.

由 AI 依官方說明整理,以原文為準

0.96.0

CLIGitHub 原文

AI 摘要

  • Thread listings now check the state database first, including archived threads, and use filesystem traversal only when needed. This improves listing correctness and resilience.
  • On non-Windows platforms, unified_exec is now enabled.
  • Websocket clients can receive rate-limit signals through a new event, with ETag and reasoning-metadata handling matching websocket behavior.
  • Clients can start thread compaction through a new API call and track its completion separately; it runs asynchronously.
  • The TUI’s Esc key now exits notes mode instead of interrupting the session when notes are open in the request_user_input overlay.

由 AI 依官方說明整理,以原文為準

0.95.0

CLIGitHub 原文

AI 摘要

  • Git commands that can change or delete data are now less likely to bypass approval checks.
  • You can launch Codex Desktop on macOS with codex app <path>; Codex downloads the DMG automatically if needed.
  • Shell-related tools can run in parallel. Shell executions also receive CODEX_THREAD_ID, which scripts and skills can use to identify the active session.
  • You can load personal skills from ~/.agents/skills (with compatibility for ~/.codex/skills), and app-server APIs and events can list and download public remote skills.
  • /plan accepts prompt arguments and pasted images. The release also fixes interrupted codex exec runs that could hang in websocket or streaming flows.

由 AI 依官方說明整理,以原文為準

26.204

App官方原文

AI 摘要

  • You can now open files and folders in Zed or TextMate.
  • The review panel can now preview PDFs.
  • Performance has been improved.

由 AI 依官方說明整理,以原文為準

26.203

App官方原文

AI 摘要

  • You can now rename a thread by double-clicking it in the thread list.
  • Sync is now called Handoff, and its interface shows clearer source and destination stats.
  • The release also includes unspecified performance improvements and bug fixes.

由 AI 依官方說明整理,以原文為準

0.94.0

CLIGitHub 原文

AI 摘要

  • Plan mode is now on by default, with updated guidance in its prompt.
  • Skills can now be loaded from .agents/skills; instructions for relative paths are clearer, and nested folders can be marked as skills.
  • Personality settings are now stable: the default is friendly, the setting is named personality, and existing settings migrate forward.
  • Console output now includes runtime metrics to help with diagnostics.
  • Unarchiving a thread refreshes its timestamp and sidebar ordering; conversation rules are capped and duplicate prefix rules are removed; override turn context no longer adds extra items.

由 AI 依官方說明整理,以原文為準

0.93.0

CLIGitHub 原文

AI 摘要

  • Smart approvals are enabled by default, and MCP tool calls now show explicit approval prompts.
  • Plan mode now streams proposed plans in a dedicated TUI view; a feature-gated /plan shortcut lets you switch modes quickly.
  • Use /apps to browse connectors in the TUI and $ to insert apps into prompts.
  • The app server can accept ChatGPT authentication tokens from a host app and request refreshed tokens when needed.
  • Several fixes improve everyday use: file search follows session folder changes and handles multiple roots; input history restores image attachments and rich text; and connections fall back to HTTP if WebSocket proxies fail.

由 AI 依官方說明整理,以原文為準

0.92.0

CLIGitHub 原文

AI 摘要

  • Collaboration flows use less CPU, and new safeguards limit how deeply agents can collaborate. They also add an explorer role and improve collaboration event handling.
  • The web search tool now shows activity while it works and handles all action types. Cached web search is now the default client behavior.
  • You can filter the app’s thread list to browse large collections more easily, and restore archived threads to active sessions with a new RPC.
  • MCP servers can define OAuth scopes in config.toml, so you need to pass --scopes on login less often.
  • Fixed a TUI freeze under heavy streaming, and corrected codex resume --last --json prompt parsing.

由 AI 依官方說明整理,以原文為準

0.91.0

CLIGitHub 原文

AI 摘要

  • The maximum number of sub-agents is now 6, tightening limits on how many can be used at once.
  • This change is intended to tighten resource usage and guardrails for agent fan-out.
  • No other changes are described in these notes.

由 AI 依官方說明整理,以原文為準

0.90.0

CLIGitHub 原文

AI 摘要

  • A network sandbox proxy now helps control outbound network access through policy enforcement.
  • Collaboration mode is available as a TUI beta, with a clearer plan-to-execute handoff and simpler Coding or Plan mode selection.
  • Connectors support has begun through the app server and MCP integration, with related configuration and documentation updates.
  • Resume is more reliable: Codex handles out-of-order events and prompts for the working directory when it differs.
  • Several usability fixes make errors and defaults more dependable: configuration errors are clearer, filtered model presets are respected, and --yolo no longer fails outside a Git repository.

由 AI 依官方說明整理,以原文為準

0.89.0

CLIGitHub 原文

AI 摘要

  • You can now use /permissions for a shorter approval workflow; /approvals remains available for compatibility.
  • A new /skill screen lets you turn individual skills on or off.
  • Slash-command suggestions now favor exact and prefix matches over fuzzy matches.
  • The app server can read individual threads and filter archived threads from thread lists. Its clients also support layered configuration files, and config/read can calculate effective settings for a specified working directory.
  • Tilde paths such as ~//... can no longer expand outside your home directory.

由 AI 依官方說明整理,以原文為準

0.88.0

CLIGitHub 原文

AI 摘要

  • Headless environments now have device-code authentication as a standalone fallback.
  • Configuration is loaded only from trusted folders, and symlinked config.toml files are now resolved correctly.
  • Interrupted turns no longer repeat.
  • Fixed Azure endpoint input errors, a memory leak in the core runtime, and an image-paste regression in the WSL interface.
  • Updated MCP documentation links and collaboration-mode developer instructions, and corrected a naming typo.

由 AI 依官方說明整理,以原文為準

0.87.0

CLIGitHub 原文

AI 摘要

  • Piped commands that don’t use a terminal no longer hang while waiting for input.
  • User shell commands now use your shell settings and aliases.
  • The TUI now shows approval requests from spawned or unsubscribed threads.
  • Token estimates during compaction are more accurate, improving budget estimates in long sessions.
  • User-message annotations can now survive history rebuilds, and MCP tool results include a thread ID and defined output schema for compatibility.

由 AI 依官方說明整理,以原文為準

0.86.0

CLIGitHub 原文

AI 摘要

  • Skill metadata can now be set in SKILL.toml, including names, descriptions, icons, brand color, and a default prompt, and is surfaced in the app server and TUI.
  • Clients can explicitly disable web search and send an eligibility header, supporting alignment with server-side rollout controls.
  • Accepting an MCP elicitation now sends an empty JSON payload instead of null, for servers that expect content.
  • Several interaction and process issues were fixed: prompt placeholders are no longer italic, empty paste events no longer read clipboard images, and Unified exec cleans up background processes to prevent late End events.
  • Internal changes refresh the orchestrator prompt to improve routing behavior and reduce noisy needs_follow_up error logs.

由 AI 依官方說明整理,以原文為準