Skip to content

Changelog for Four Tools: Claude Code / Codex / OpenCode / WorkBuddy ​

Take two minutes to read this page before upgrading. It summarizes the Claude Code changelog and the version highlights of Codex, OpenCode, and WorkBuddy, listing only major version changes to help you decide "whether to update right now"; for the full set of changes, refer to each official changelog.

Claude Code ​

Entries on this page are current as of: 2.1.329 (2026-11-19)

Versions iterate quickly, so first run claude --version in your terminal to confirm which version you have, then check it against the official changelog. The dates below are taken from the npm release times of each version (UTC), because the official CHANGELOG.md itself does not include dates. Note: npm's stable tag still points to 2.1.311, and latest points to 2.1.329.

2.1.318–2.1.329 Highlights ​

  • 2.1.321 (2026-11-09): Added Claude Fable 5.2 (claude-fable-5-2), which becomes the default Fable model; 1M context, $8/$40 per million tokens, cache reads $0.20.
  • 2.1.323 (2026-11-12): Added the experimental Code Mode feature; when enabled, Claude writes a piece of JavaScript to call tools. It is off by default—see the official docs at code.claude.com/docs/en/code-mode for how to enable it. When a model requires usage credits but the quota is disabled or used up, an options menu pops up: stop, switch to usage credits, or upgrade. For model IDs Claude Code doesn't recognize, the default context changes from 200K to 1M; if you use another vendor's small-context model, set CLAUDE_CODE_DISABLE_1M_CONTEXT=1 to keep 200K.
  • 2.1.325 (2026-11-13): Bash commands running in the sandbox now go through the normal permission confirmation by default. To restore confirmation-free execution, set sandbox.autoAllowBashIfSandboxed: true.
  • 2.1.326 (2026-11-16): On first run, the default theme is now "Auto (match terminal)" instead of Dark; in full-screen mode, subscription users opening /usage now see the plan quota bar first; auto-compaction in plan mode now uses the model currently answering you, which may be more expensive.
  • 2.1.327 (2026-11-17): Added claude mcp add-json <name> -, which reads server JSON from standard input so tokens don't need to be written on the command line (it cannot be used inside a Claude Code session); /status now lists broken agent, skill, and other files; the non-full-screen renderer was renamed to /tui classic, and /tui default still works.
  • 2.1.328 (2026-11-18): On most models, /code-review now only reports bugs and violations of CLAUDE.md; cleanup suggestions about reuse, simplification, and efficiency moved to /simplify; each finding is labeled Important, Pre-existing, or Nit, and when there are no issues it shows "Code review found no issues.". /bug asks Yes or No before sending.
  • 2.1.329 (2026-11-19): /tasks now shows this session's scheduled tasks and /loop; fixed issues such as ~/.claude.json occasionally being reset to default values, repeated logouts on macOS, /code-review <PR 号> sometimes reviewing the current branch, and /usage-credits forcing Pro and Max users to log in again.

To handle before 11/30

2.1.316 states clearly: the fast mode of Opus 4.8 is deprecated as of 2026-10-30 and will be removed in 2026-11-30; after removal, fast mode requests for Opus 4.8 will fail outright rather than falling back to standard speed. The official recommendation is to migrate to Opus 5.5.

  • I have run claude --version, know which version I'm on, and know that /code-review and /simplify now have different responsibilities

Changes in 2.1.317 (recorded on 11/3) ​

Added:

  • /schedule added a reminder: it warns you when cloud scheduled tasks are pushed into connectors you didn't request, and tells you how to remove them.
  • Privacy confirmation: when a web or shell tool call needs to send your account email over the network, it asks you first.
  • Cloud sessions add link cards: when an MCP server asks you to access a link, you can open or reject it directly, so requests are no longer silent.
  • The Agent SDK's hook callbacks add onFailure: "block": a failed callback blocks the action instead of letting it through.
  • Claude Desktop's Code tab now indicates whether your organization's remote hosting settings have loaded; when auto mode has no answer for a subagent tool call, it also gives a hint instead of leaving it blank.
  • For team administrators: the gateway management API adds a cost report GET /v1/organizations/analytics/cost, letting you query historical spend grouped by developer, by day, or by IdP; OTel metrics add fields such as anthropic.plan and anthropic.provider.

Adjusted:

  • Messages sent to a running background agent arrive sooner: the command it is executing moves to the background instead of being stopped.
  • Subagents / teammates using a different model than the main conversation now use the effort level saved for that model, rather than the one for the whole session.

How 2.1.317 affects you ​

If you mainly run tasks in the terminal, this upgrade's two most direct changes are: email sending requires confirmation first, and background agents no longer interrupt commands when you send them messages. If you use cloud sessions, the MCP link cards make "open this link" requests visible and rejectable. For teams automating with the SDK, setting the callback to onFailure: "block" makes failures safer by default. The official changelog is generated from CHANGELOG.md on GitHub; the official page is authoritative for the complete list.

Codex ​

Entries on this page are current as of: 0.171.0 (2026-11-20)

Codex is OpenAI's coding agent, available as the Codex CLI, an IDE extension, cloud automation, and Codex in the ChatGPT desktop app. Version numbers and release dates follow the official GitHub releases page (dates are UTC).

0.169.0–0.171.0 Highlights ​

  • 0.169.0 (2026-11-13): the terminal workspace supports multiple side conversation tabs; the agent command center can search past tasks by name, prompt, or directory; you can add and edit custom model providers in the terminal; Ctrl+S saves a diagnostics ZIP locally. Behavior changes: in untrusted directories, codex exec disables project config and MCP servers, even if --sandbox workspace-write is added; use --trust (codex exec or codex review) to explicitly save project trust.
  • 2026-11-17: GPT-6.2 Astra and GPT-6.2 Sol are live on ChatGPT Work and Codex. Official guidance: use Astra for the hardest work, and Sol for everyday coding and multi-step tasks; on the same paid plan, 6.2 Sol completes more tasks than 6.1 Sol. The Pro $200 tier can use 6.2 Sol Ultrafast, and the Pro $500 tier can use Ultrafast with both models.
  • 0.170.0 (2026-11-18): the terminal interface supports real-time voice calls (WebRTC).
  • 0.171.0 (2026-11-20): GPT-6.2 Sol and Astra were added to the model catalog, and Sol became the default choice; the terminal workspace is on by default, letting you preview documents, images, and subagent records in the side panel; the input box supports undo and redo; new additions include codex exec --remote <ADDR>, plus /rc and /remote for managing remote control and displaying pairing codes.

How this affects you: after upgrading to 0.171.0 with no config changes, the default is gpt-6.2-sol; for which tasks to switch to Astra, see How to choose a model. If codex exec runs in a new directory in a script and project config doesn't take effect, first check whether --trust is missing.

  • I know that from Codex 0.171.0 onward the default model is gpt-6.2-sol, and that codex exec needs --trust in untrusted directories

Changes in 0.167.0 (recorded on 11/6) ​

Added:

  • codex agents --no-daemon opens the agent command center, and both it and the session selector mark unread results.
  • The terminal session selector can switch between connected backends directly, and drafts are preserved when switching.
  • Local image attachments can be previewed while composing a message (requires terminal support), and local image paths in the transcript are clickable to open.
  • API key and cloud provider sign-in entries are moved into "Sign in another way"; enabling gcp_setup_wizard lets you configure Google Cloud Vertex AI.
  • In /experimental you can optionally use the macOS or Windows system proxy; when an approval prompt is open, you can expand patch details directly in the transcript.

Fixed:

  • After switching model providers, saved sessions remain visible; sessions that can't be restored let you read the transcript and carry the prompts into a new session to continue.
  • After Ctrl+L clears the screen, history messages can still be retrieved with Ctrl+T.
  • Network approval prompts now close after the request finishes; concurrent requests to the same endpoint are prompted one at a time.
  • Automatic context compaction now accounts for the retained output of failed responses, making compaction more accurate.

How 0.167.0 affects you ​

If you use the CLI in the terminal, the most practical changes here are image preview and the Ctrl+L/Ctrl+T clear-screen combo; preserving drafts when switching between multiple backends suits users working with multiple accounts or cloud environments at once. Versions iterate quickly (a 0.169.0-alpha pre-release appeared the day after 0.167.0 was released), so confirm you want the stable version before chasing the new one.

OpenCode ​

Entries on this page are current through: v2.3.2 (2026-11-18)

OpenCode is an open-source AI coding agent, available in terminal, desktop, and web forms. A side note: the official position is that Windows package managers are not supported, so Windows users should use the standalone binary or the desktop app. Dates come from GitHub releases (UTC).

v2.2.6–v2.3.2 Highlights ​

  • v2.2.6 (2026-11-06): MCP servers can enable connection pooling, so projects with the same config share one connection; opencode run -w opens a session in a new git worktree, and --worktree=<name> gives it a name; opencode run --archived keeps sessions started by scripts and CI out of the inbox.
  • v2.2.7 (2026-11-10): if your GitHub Copilot account has the Auto pool, you can select Auto in the model picker; sessions can be marked read or unread. The SDK has a breaking change: @opencode/ai removed ImageResponse.image in favor of reading response.images[0].
  • v2.3.0 (2026-11-13): subagents now run in the background by default, reporting back when done; terminal logs are quieter by default (session.verbosity defaults to low, collapsing a run of tool calls into one line) — to see every step, set Verbosity to medium or high; on desktop, Cmd+Alt+G (Ctrl+Alt+G on Windows and Linux) jumps back to the previous prompt.
  • v1.19.6 (2026-11-17): OpenCode 1 desktop started prompting users to upgrade to 2.0, keeping sessions, projects, and settings, but plugins written for OpenCode 1 must wait for their authors to update them.
  • v2.3.1 (2026-11-18): Rereads CLAUDE.md—directories without AGENTS.md will read CLAUDE.md, and when the global AGENTS.md is missing it falls back to reading ~/.claude/CLAUDE.md; set OPENCODE_DISABLE_CLAUDE_CODE to disable this; timed-out shell commands move to the background and keep running instead of being killed; when auto-compaction is imminent, the summary is written in the background (can be disabled with compaction.background: false); after logging in with ChatGPT you can select GPT 6.2 Sol and Astra (including fast variants); plugins can be checked and updated in Settings › Extensions; the experimental schedules feature was renamed to automations.
  • v2.3.2 (2026-11-18): Fixes an issue where GPT-6.2 models re-billed the system prompt as uncached input on every step.

How this affects you: For people using both Claude Code and OpenCode, starting from v2.3.1 projects with only CLAUDE.md can also be read by OpenCode, but when both files exist it only recognizes AGENTS.md; see AGENTS.md and CLAUDE.md for configuration. If you use GPT-6.2, upgrading to v2.3.2 can save you wasted money.

  • I know that starting from OpenCode v2.3.0, when per-step details are not visible, I should adjust Verbosity

Changes in v2.2.5 (recorded on 10/30) ​

Added:

  • Worktree and Shell mode shortcuts: on macOS press ⌘⌥O (on Windows/Linux press Ctrl+Alt+O) to select a worktree, and ⌘⌥B / Ctrl+Alt+B to pick its base branch; ⌘⇧X / Ctrl+Shift+X toggles Shell mode in the composer.
  • Send directly from the quote popover: hold ⌘ or Ctrl to turn the checkbox into a send action, or press ⌘Enter / Ctrl+Enter to send only that quote and comment, leaving the rest of the composer untouched.
  • The compaction buffer supports percentages: write "compaction": { "buffer": "20%" } in opencode.json to reserve headroom as a proportion of the model's context window, so one configuration fits both large and small windows; the token-count syntax still works.

Fixed:

  • MCP servers whose stdio messages exceed 10 MiB are no longer disconnected, and large messages are read faster.
  • When read fails due to a missing folder, the error names the missing folder and the nearest existing directory; when edit can't find oldString, it shows the closest line in the file or notes that the change is already in the file.
  • Tool calls named bash are now executed as shell instead of erroring, with permissions unchanged; plugin tools actually named bash take precedence.

How v2.2.5 affects you ​

If you often work in parallel across multiple worktrees, these shortcuts save a lot of menu-clicking time; users whose context tends to overflow can switch the compaction buffer to a percentage and skip recalculating token counts when changing models. The fix for MCP large-message disconnections is a real stability improvement for heavy MCP users.

WorkBuddy ​

Entries on this page are up to date as of: 5.9.3 (2026-10-27)

WorkBuddy is an all-scenario AI workstation launched by Tencent Cloud Code Assistant. Billing uses credits: consumption depends on both model pricing and task complexity; credits are issued monthly, valid only for the current month with no carryover; WeChat Pay is supported.

Changes in This Release ​

5.9.3 (2026-10-27) is mainly about fixes:

  • When migrating data from another device, you are no longer limited to a full import: you can select by category, and the import limit has been raised to 100GB.
  • Fixed issues on Windows where upgrade checks failed due to installation directory validation, all tool executions errored after upgrade because sandbox initialization failed, and environment variables were incomplete in the new process after a silent upgrade.
  • Fixed issues where the model of a historical session was incorrectly rolled back after restart, and where the session area occasionally crashed when opening a long session task.

5.9.2(2026-10-26):

  • The final speech recognition transcript is now filled directly into the input box without additional polishing; the voice shortcut has restored support for key combinations.
  • Fixed issues such as exceptions when saving or saving local files as, messages being sent by mistake when pressing Enter after a cloud session stopped generating, and the paid label being indistinguishable in dark mode.

How This Affects You ​

If you have run into "all tools erroring after upgrade" or "long sessions failing to open" on Windows, 5.9.3 is aimed exactly at these issues, and updating is recommended. Users switching computers can now migrate data by category instead of importing the whole package. Voice input users take note: the final transcript is no longer polished—whatever is recognized is what gets filled in.

Next steps after updating ​

  • Claude Code: The native installation updates automatically in the background; if you can't wait, run claude update to upgrade manually, then use claude --version to verify the version and check the official changelog to decide whether to follow up.
  • Codex: If installed via npm, you can rerun npm install -g @openai/codex to update and use codex --version to verify the version; refer to the official releases page for version updates.
  • OpenCode / WorkBuddy: Follow their respective official CHANGELOGs and client update prompts; WorkBuddy can also be downloaded as the latest installer from the official website workbuddy.cn.

Related reading: Quick reference for the four tools: install commands and key operations · Install Claude Code: ready in 10 minutes

gaodong.aiAboutPrivacyContact

Independent tutorials, not affiliated with any product mentioned. Product names and trademarks belong to their owners. Check official pages for prices and policies.