* Release V3.67.0
Bump version from 3.66.0 to 3.67.0 in package.json and
package-lock.json. Add changelog entry for v3.67.0 covering new
features (subagent skills, AgentConfigLoader, Responses API, websocket
preconnect, CLI /q command), bug fixes (reasoning delta crash, OpenAI
tool ID, auth checks, Gemini 3.1 Pro), and other changes. Update
WhatsNewItems fallback banners to reflect current promotions.
* Fixing stuff
* refactor: consolidate subagent request usage tracking into state object
Replace scattered per-request token tracking variables with a structured
`SubagentUsageState` interface containing `currentRequest` and
`lastRequest` states. This improves code organization by grouping related
token metrics (input, output, cache write/read, total tokens, cost) into
a cohesive `SubagentRequestUsageState` object, reducing variable sprawl
and making the usage lifecycle (current → last) more explicit.
* feat(subagent): add support for skills and optional modelId in agent config
- Update `AgentBaseConfigSchema` and `AgentConfigFrontmatterSchema` to include an optional `skills` field and make `modelId` optional.
- Implement `parseSkills` and `normalizeSkillName` in `AgentConfigLoader` to handle skill parsing from YAML frontmatter.
- Update `SubagentBuilder` to provide access to configured skills.
- Modify `SubagentRunner` to filter available skills based on the agent's configuration, falling back to all available skills if none are specified.
- Update host retrieval to use `HostRegistryInfo` instead of `HostProvider`.
This allows subagents to be restricted to specific skills and provides more flexibility in model configuration.
* update unit test
* feat(cli): fetch featured models from backend with local fallback
- Add async getFeaturedModelsForCline() to fetch models via controller
- Load featured models dynamically in AuthView with useEffect
- Update FeaturedModelPicker to accept featuredModels as optional prop
- Refactor helper functions to accept models parameter for flexibility
- Keep local hardcoded models as fallback when backend fetch fails
* Fixing stuff
* Fixing stuff
* Fixing stuff
* refactor: centralize tool handler registration and filter by allowed tools
- Create centralized toolHandlersMap for all tool handler instantiation
- Add registerToolHandlers method to register only allowed tools from config
- Filter subagent tools to only include allowed tools from allowedTools config
- Remove scattered tool handler registration logic in favor of single source of truth
- Improve maintainability by consolidating tool handler creation in one place
This refactoring ensures subagents only have access to explicitly allowed tools
and makes the tool registration process more maintainable and consistent.
It also improves separation of concerns by having the coordinator manage all tool handler registration, while the executor focuses on orchestration. The allowedTools parameter enables runtime filtering of available tools for different contexts.
Also update PromptRegistry to load synchronous and simplify variant lookup
- Convert async load() to synchronous, called in constructor as both loadVariants and loadComponents are not async functions
- Remove health check logic and loaded state tracking
- Extract getVariant() method with proper generic fallback
- Add getComponents() accessor and simplify component loading
- Convert variant/component loaders from async to synchronous
- Remove unnecessary await calls throughout the codebase
- Add PromptRegistry tests for variant resolution and components
* fix test
* feat: add AgentConfigLoader for file-based agent configs
Add AgentConfigLoader singleton to manage agent configurations loaded from
YAML files in the agents directory. Supports hot-reloading via file watcher,
validates config schema with Zod, and integrates with extension lifecycle
(StateManager initialization and tearDown disposal).
* add tests
* add missing export
* update tests
* feat(tools): implement dynamic tool registration for subagents
Updates the tool system to support dynamically registered subagents as individual tools.
- Modifies `ClineToolSet` to generate specific tool definitions for configured subagents via `AgentConfigLoader`.
- Updates `parseAssistantMessageV2` to use `getToolUseNames()` instead of a static list, enabling the parser to recognize dynamic tool tags.
- Replaces the generic `USE_SUBAGENTS` tool with specific subagent instances when available in the system prompt context.
* update config path and refine tool descriptions
- Relocate the subagent configuration directory from `~/.cline/data/agents` to `~/Documents/Cline/Agents` to improve user accessibility.
- Update subagent tool descriptions and parameter instructions in the system prompt to be more descriptive and helpful for the model.
* revert unrelated changes
* revert unrelated changes
* update unit test
* fix: await AgentConfigLoader initialization before StateManager completes
Ensure agent configs are fully loaded during StateManager initialization
by awaiting the `ready()` promise. Previously, `AgentConfigLoader` was
instantiated without waiting for the initial load to complete, causing
potential race conditions where configs might not be available when
needed.
- Add `initialLoadPromise` field to track the async initial load
- Expose a `ready()` method to allow callers to await initialization
- Await `AgentConfigLoader.getInstance().ready()` in StateManager
* set previousRequestTotalTokens
* Made messages api changes
* Made changes
* Added changes
* Made maxTokens point to the right thing
* Removed deprecated max_tokens field
* Reverted the change
* Making an additional change to not cause any issues with chat completions
* reverting changes so we can make them in the backend
* Added changeset
* Fixed changes based on AI comments
Warm up the OpenAI WebSocket connection early in WebSocket mode to avoid handshake latency on the first response.create call. This introduces a responsesWsReadyPromise to track the connection state and prevent duplicate connection attempts while the initial connection is in flight.
* fix: restrict OpenAI tool ID transformation to native provider
Update `convertToOpenAiMessages` and `transformToolCallId` to only apply tool ID transformations when the provider is explicitly set to `openai-native`. This prevents unintended ID modifications for other providers (like OpenRouter or local LLMs) that use the OpenAI format but may have different tool ID requirements or already provide compatible IDs.
* update tests
* transformToolCallIdForNativeApi
* fix: openai native provider token usage mapping
- Add `store` parameter support to OpenAI native provider options to allow persisting completions.
- Fix incorrect mapping of `cached_tokens` and `reasoning_tokens` in usage statistics.
- Include `thoughtsTokenCount` in the final usage report to track reasoning model performance and costs.
* feat: add websocket support for OpenAI Responses API
This commit introduces WebSocket support for the OpenAI Native provider's Responses API, providing an alternative to the standard HTTP streaming.
- Implement `createResponseStreamWebsocket` in `OpenAiNativeHandler` with a fallback to HTTP on failure.
- Refactor `OpenAiNativeHandler` to modularize tool mapping and parameter construction for the Responses API.
- Update `OcaHandler` to explicitly disable `previousResponseId` when using the Responses API and add validation for model information.
- Integrate `undici` WebSocket for better compatibility in the extension environment.
* disablePreviousResponseId
* feat: add timestamp to conversation messages for response chaining
Add `ts` field to `ClineStorageMessage` to track when messages were
created. Use this timestamp to enforce a 23-hour validity window when
chaining OpenAI responses via `previousResponseId`, since the API only
retains responses for 24 hours. Also fix non-null assertion operators
in tests to use optional chaining for safer access.
* add OpenAI Responses Websocket Mode ApiFormat support
- Add `OPENAI_RESPONSES_WEBSOCKET_MODE` to the `ApiFormat` enum in proto definitions.
- Update `OpenAiNativeHandler` to use the new API format for determining when to use websocket mode, replacing previous environment-based logic.
- Refactor tool mapping for OpenAI Responses to support strict mode and correctly handle null parameters.
- Ensure the `store` option is disabled when `previous_response_id` is present in websocket mode.
- Bump version to 2.4.1 and update package dependencies.
* use abortController
* add support for websocket mode to openai-codex
* set behind feature flag
- added a method to StateManager, setSessionOverride, which overrides
state settings while the statemanager lives in memory
Co-authored-by: Max Paulus 🥪 <max@cline.bot>
* fix: inline focus-chain slider within its feature row
Moves the focus-chain reminder interval `SettingsSlider` from a
standalone element rendered after all experimental feature rows to
being rendered directly beneath the focus-chain `FeatureRow`. The
slider now renders conditionally when `feature.id === "focus-chain"`
and the feature is enabled, improving UI cohesion and making the
relationship between the toggle and its configuration more explicit.
Additionally:
- Relocates focus-chain from `experimentalFeatures` to `agentFeatures`
- Removes the `isExperimental` prop and "Experimental:" label badge
from `FeatureRow` and related feature toggle definitions
- Simplifies `SettingsSlider` markup by removing the wrapper card
styling, making it suitable for inline embedding
- Removes unused line from common.ts
* nestedKey
Handle plain Cmd/Ctrl+A directly in `ChatTextArea` keydown to force
textarea-wide selection via `setSelectionRange`, while preventing default
and propagation. This avoids intermittent failures caused by global shortcut
listener races, and keeps Cmd/Ctrl+Shift+A behavior unchanged.fix(chat): make Cmd/Ctrl+A select-all deterministic
Handle plain Cmd/Ctrl+A directly in `ChatTextArea` keydown to force
textarea-wide selection via `setSelectionRange`, while preventing default
and propagation. This avoids intermittent failures caused by global shortcut
listener races, and keeps Cmd/Ctrl+Shift+A behavior unchanged.
* fix: flaky Cancel behavior by preventing duplicate cancel actions
This PR fixes chat cancel behavior where users sometimes had to click Cancel multiple times, and repeated clicks could accidentally transition into Resume/restart behavior.
* move to finally
* refactor: replace non-null assertions with safe null checks in PatchParser
Replace all forbidden non-null assertions (`!`) in PatchParser.ts with
safe alternatives using optional chaining (`?.`) and nullish coalescing
(`?? ""`/`?? 0`). Also refactor the Levenshtein distance matrix from a
2D array to a flat array to eliminate index-based non-null assertions,
improving type safety and code robustness.
No feature behavior changes.
* simplify Levenshtein matrix indexing
Initialize the distance matrix with zeroes and add `at`/`set` helpers for flat-array access in `levenshteinDistance`.
This removes repeated index math and nullish fallbacks, making the algorithm easier to read while keeping bounds-safe access and identical behavior.refactor(patch-parser): simplify Levenshtein matrix indexing
Initialize the distance matrix with zeroes and add `at`/`set` helpers for flat-array access in `levenshteinDistance`.
This removes repeated index math and nullish fallbacks, making the algorithm easier to read while keeping bounds-safe access and identical behavior.
* feat: add welcome banner support from backend
* make DB banner format conform with existing banners
* add support for welcome banner actions
* remove debugging helper that bypass dismissal, dismissal should work again
* undo changes to make welcome banner always appear during debugging
* remove console logs for debugging
* clean up bannerservice
* clean up welcomesection.tsx
* add new tests for ide type filtering
* add welcome banner own feature flag and conditionally display between hard coded welcome banner and DB backed ones
* turn on welcome banner flag locally by default
* close welcome banners when clicking on actions
* apply bot review suggestion, fix memory leak
* address feedback: use p without span
* split welcome banners into a seperate component to keep whatsnewmodal clean
* get action through api schema instead of extractin it from rules_json
* use only bannerWaitTimeoutRef, remove waitingForBannersRef
* resolve new merge conflict
* linter
* cerebra
- Remove `auto_condense_threshold` from `Settings` and `UpdateSettingsRequest` in `state.proto`.
- Remove `autoCondenseThreshold` from `ApiProviderInfo` interface.
- Update `generate-state-proto.mjs` to remove double field handling and improve integer parsing.
- Add error handling to `ContextManager` when parsing previous request JSON to prevent crashes on malformed data.
* fix(models): keep Sonnet 4.5 as default
* chore(changeset): add release note for Sonnet 4.5 default
* fix(models): remove Sonnet 4.6 from curated model lists
* fix(models): restore Sonnet 4.6 in web recommended list
* feat(cerebras): remove deprecated llama-3.3-70b and qwen-3-32b models
These models have been deprecated from the Cerebras inference platform.
- Remove llama-3.3-70b and qwen-3-32b from cerebrasModels in api.ts
- Update supported models documentation in cerebras.mdx
- Add changeset for the deprecation
* fix: remove stale llama-3.3-70b and qwen-3-32b references from rate limits
Remove dead switch cases in getRateLimits() that referenced deprecated models
no longer present in cerebrasModels.
* feat(cli): add /skills slash command for managing skills
- Add /skills to CLI_ONLY_COMMANDS in slashCommands.ts
- Create SkillsPanelContent component with:
- Display global and workspace skills with toggle indicators
- Enter to use skill (inserts @path into input)
- Space to toggle skill enabled/disabled
- Selectable marketplace link to skills.sh
- Keyboard navigation with arrow keys and vim keys
- Wire up panel in ChatView.tsx
- Add comprehensive tests for keyboard interactions
* refactor(cli): use static skill controller imports
* fix(cli): add React import to skills panel test
* fix(cli): suppress required React import lint in skills test
* fix(cli): harden /skills panel interactions
Revert optimistic skill toggle state when persistence fails, and surface a fallback URL when opening the marketplace fails. Also tighten and extend tests to verify exact marketplace URL handling and rollback behavior.
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix: disable click-to-set auto-condense threshold and hardcode default
Clicking anywhere on the context window progress bar silently set
autoCondenseThreshold to a value based on click position (e.g. 0.05),
persisting in globalState. This caused compaction to fire at ~10K tokens
instead of the intended ~150K, resulting in ~20 context resets per task.
- Comment out click and keyboard handlers on progress bar (keep components
for future release with proper UX)
- Hardcode threshold to 0.75 default, ignoring corrupted stored values
Co-authored-by: Cursor <cursoragent@cursor.com>
* test: add shouldCompactContextWindow unit tests
Cover threshold math including the accidental low-threshold bug case,
undefined/zero fallbacks, cache token inclusion, and maxAllowedSize cap.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix: hardcode autoCondenseThreshold in all remaining callsites
Address Greptile review: SubagentRunner.ts, task/index.ts display
logic, and controller/index.ts webview state all still read the
corrupted value from globalState. Hardcode 0.75 everywhere.
Co-authored-by: Cursor <cursoragent@cursor.com>
* style: remove unnecessary union type on hardcoded threshold
Drop `number | undefined` annotation from the hardcoded 0.75 literal
in SubagentRunner.ts per Greptile review feedback.
Co-authored-by: Cursor <cursoragent@cursor.com>
* refactor: use SETTINGS_DEFAULTS constant, remove commented-out code, clarify test
- Replace hardcoded 0.75 with SETTINGS_DEFAULTS.autoCondenseThreshold
across all 4 callsites for a single source of truth
- Delete commented-out click/keyboard handlers in ContextWindow.tsx,
replace with TODO referencing PR #9348
- Make bug-case test self-documenting by deriving token values from
the threshold calculation
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Ara