v1.4.2

TUI

• Fix subagents not being clickable until finished

Desktop

• Removed the forced loading delay while the app connects

v1.4.1

Core

• Fix clangd choosing CMakeLists.txt or Makefile as the project root in C and C++ workspaces. (@nonbanana)
• Add permission prompts for GitLab Duo Workflow tool calls instead of auto-running them. (@vglafirov)
• Hide unsupported variants for Big Pickle models.

TUI

• Show an OpenCode Go subscribe prompt when free usage limits are reached.
• Simplify provider labels in the model and provider pickers.

Desktop

• Fix terminal connections in same-origin desktop and web app setups. (@OpeOginni)
• Fix session review and change lists when diff data arrives in inconsistent shapes.

SDK

• Fix the generated SDK and OpenAPI types for /providers and session shell responses.

Thank you to 3 community contributors:

• @OpeOginni:
  • fix(app): skip url password setting for same-origin server and web app (#19923)
• @vglafirov:
  • feat(llm): integrate GitLab DWS tool approval with permission system (#19955)
• @nonbanana:
  • fix(lsp): remove CMakeLists.txt and Makefile from clangd root markers (#21466)

v1.4.0

Breaking Changes in SDK

Due to some performance issues we had to make a few breaking changes in our data model.

Diff metadata

The diffs contained in the metadata for the edit and patch tool will no longer contain a to and from field with full file contents. they will only contain a patch field with the unified diff

Additionally all snapshot diffs contained in user message and session diff will follow this format

UserMessage.variant

Before the user message had a variant field - it is now nested under model so userMessage.model.variant

Core

• Added OTLP observability export support.
• Fixed failed web fetches leaving stale timeouts behind.
• Improved opencode login transport error messages.
• Retried Alibaba provider rate-limit errors instead of failing immediately.
• Kept model variants scoped to the selected model.
• Added full HTTP proxy support.
• Fixed OpenRouter provider issues.
• Aligned GitHub Copilot Anthropic reasoning levels and removed incorrect Qwen variants.
• Reduced TypeScript LSP memory use by using the native project config. (@derekbar90)

TUI

• Added a keybinding option for "Switch model variant". (@ariane-emory)
• Added PDF drag and drop for attachments. (@gitpush-gitpaid)
• Added opencode run --dangerously-skip-permissions to auto-approve non-denied permission prompts.

Desktop

• Improved subagent sessions with clearer titles, navigation, and progress states.
• Moved auto-accept permissions into Settings.
• Showed full file names on attachment chips.

Thank you to 4 community contributors:

• @derekbar90:
  • fix(lsp): MEMORY LEAK: ensure typescript server uses native project config (#19953)
• @mchenco:
  • docs: update Cloudflare provider setup to reflect /connect prompt flow (#20589)
• @gitpush-gitpaid:
  • feat(opencode): Add PDF attachment Drag and Drop (#16926)
• @ariane-emory:
  • fix(tui): use sentence case for theme mode command palette items (#21192)
  • feat(tui): allow variant_list keybind for the "Switch model variant" command (#21185)

v1.3.17

Core

• Cloudflare Workers AI and AI Gateway now prompt for missing account details and show clear setup errors when required Cloudflare variables are missing (@mchenco)

TUI

• Restored the default kitty keyboard handling on Windows terminals to avoid input issues from the workaround rollback

Thank you to 1 community contributor:

• @mchenco:
  • fix: show clear error when Cloudflare provider env vars are missing (#20399)

v1.3.16

Core

• Support Azure model options on both chat and responses paths (@meruiden)
• Expose session model and mode config options through ACP (@georgeharker)
• Add a separating blank line before read tool file contents for clearer output
• Fix output token totals when reasoning tokens are reported separately

TUI

• Default Ctrl+Z to undo on Windows instead of terminal suspend
• Let you disable TUI mouse capture in config or with OPENCODE_DISABLE_MOUSE (@blackheaven)
• Hide org switching when there is only one org to choose from
• Label Console-managed providers and let you switch orgs from the provider UI

Extensions

• Fix plugin installs from npm aliases and git URLs, including Windows cache path issues

Thank you to 3 community contributors:

• @blackheaven:
  • feat(tui): make the mouse disablable (#6824, #7926) (#13748)
• @georgeharker:
  • fix(core): implement proper configOptions for acp (#21134)
• @meruiden:
  • fix: pass both 'openai' and 'azure' providerOptions keys for @ai-sdk/azure (#20272)

v1.3.15

Core

• Prevent npm installs from failing when Arborist hits the compiled binary's node-gyp path

Thank you to 1 community contributor:

• @Yuxin-Dong:
  • refactor: remove redundant Kimi skill section (#20393)

v1.3.14

Core

• Restored git-backed review modes, including uncommitted and branch diffs.
• Fixed revert chains so restoring an earlier message also restores the right snapshot state. (@natewill)
• Added macOS managed preferences for MDM-enforced config. (@lennyvaknine43)
• Fixed sessions getting stuck after tool calls with OpenAI-compatible providers. (@valenvivaldi)
• Kept compaction summaries in the conversation's language. (@aaron-he-zhu)
• Added Venice AI as a provider. (@dpuyosa)
• Respected model-specific limit.input overrides. (@ykswang)

TUI

• Kept the prompt cursor with dialogs instead of refocusing the background prompt.
• Added a one-time confirmation before sharing a session for the first time.
• Applied scroll settings consistently across TUI scroll views.
• Kept text selections intact during global key handling.
• Fell back to the first available agent if the last-used agent is unavailable.

Desktop

• Added file mentions in review comments.
• Restored prompt focus after closing agent, model, and variant pickers.
• Added keyboard navigation and shortcuts to the question dock.
• Hid model controls in shell mode.
• Stopped the todo dock from auto-scrolling while tasks update.

SDK

• Fixed JS SDK server and TUI launch and shutdown on Windows.
• Fixed object-defined Tool.define() tools from wrapping execute multiple times. (@jpcarranza94)

Extensions

• Added support for theme-only plugin packages.

Thank you to 12 community contributors:

• @github-actions[bot]:
  • Update VOUCHED list
• @JosXa:
  • fix: normalize filepath in FileTime to prevent Windows path mismatch (#20367)
• @valenvivaldi:
  • fix(core): prevent agent loop from stopping after tool calls with OpenAI-compatible providers (#14973)
• @aaron-he-zhu:
  • fix(session): compaction agent responds in same language as conversation (#20581)
• @dpuyosa:
  • feat(opencode): Add Venice AI package as dependency (#20570)
• @noamzbr:
  • feat: add optional messageID to ShellInput (#20657)
  • feat(acp): Add messageID and emit user_message_chunk on prompt/command (#18625)
• @BYK:
  • fix(format): use biome format instead of check to prevent import removal (#20545)
• @lennyvaknine43:
  • feat: add macOS managed preferences support for enterprise MDM deployments (#19178)
• @ykswang:
  • fix(opencode): honor model limit.input overrides (#16306)
• @jpcarranza94:
  • fix: prevent Tool.define() wrapper accumulation on object-defined tools (#16952)
• @kflansburg:
  • feat: Send x-session-affinity and x-parent-session-id headers (#20744)
• @natewill:
  • fix(core): fix restoring earlier messages in a reverted chain (#20780)

v1.3.13

• Bump opentui version

v1.3.12

Core

• Enabled prompt caching and cache token tracking for Google Vertex Anthropic. (@major)
• Fixed Azure provider options being forwarded correctly after the AI SDK v6 migration.

TUI

• Fixed plugin replace slots mounting content more than once.

Thank you to 2 community contributors:

• @major:
  • feat: enable prompt caching and cache token tracking for google-vertex-anthropic (#20266)
• @github-actions[bot]:
  • Update VOUCHED list

v1.3.11

Core

• Add a dedicated system prompt for Kimi models. (@Yuxin-Dong)

TUI

• Improve TUI terminal output passthrough so external command output renders more reliably.

Extensions

• Skip plugins that do not expose a matching server or TUI entrypoint, warn instead of failing, and apply default options from package exports on install.
• Pin explicit plugin versions during install and block package install scripts from running.

Thank you to 1 community contributor:

• @Yuxin-Dong:
  • feat: add a dedicated system prompt for Kimi models (#20259)