Add Kimi K2 Thinking model and re-enable brain toggle #346
Conversation
|
Note Other AI code review bot(s) detectedCodeRabbit has detected other AI code review bot(s) in this pull request and will avoid duplicating their findings in the review comments. This may lead to a less comprehensive review. WalkthroughAdds a new reasoning-capable model entry ("kimi-k2-thinking"), remaps category reasoning to use it, exports ThinkingBlock/props from markdown, and integrates streaming reasoning items (delta/done) and thinking-toggle rendering into the chat UI. Changes
Sequence Diagram(s)sequenceDiagram
autonumber
participant Browser as Browser (UnifiedChat)
participant Server as Backend (Streaming API)
participant Model as Model/Reasoner
Browser->>Server: Send chat request (assistant message, reasoning enabled)
Server->>Model: Forward request (model id includes reasoning)
Model-->>Server: Provide serverReasoningId, initial reasoning output
Server-->>Browser: Event: reasoning_text.delta (serverReasoningId, text chunk)
Note right of Browser: UnifiedChat creates/updates ReasoningItem\nrenders ThinkingBlock with partial text
Server-->>Browser: Additional reasoning_text.delta events...
Server-->>Browser: Event: reasoning_text.done (serverReasoningId)
Note right of Browser: UnifiedChat finalizes ReasoningItem,\naggregates assistant text and updates UI (toggle/copy)
Estimated code review effort🎯 4 (Complex) | ⏱️ ~45 minutes
Possibly related PRs
Poem
Pre-merge checks and finishing touches❌ Failed checks (1 warning)
✅ Passed checks (2 passed)
✨ Finishing touches
🧪 Generate unit tests (beta)
📜 Recent review detailsConfiguration used: defaults Review profile: CHILL Plan: Pro ⛔ Files ignored due to path filters (1)
📒 Files selected for processing (4)
🚧 Files skipped from review as they are similar to previous changes (2)
🧰 Additional context used📓 Path-based instructions (1)**/*.{ts,tsx}📄 CodeRabbit inference engine (CLAUDE.md)
Files:
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (6)
🔇 Additional comments (11)
Comment |
Deploying maple with
|
| Latest commit: |
e6a20f1
|
| Status: | ✅ Deploy successful! |
| Preview URL: | https://cab51b7e.maple-ca8.pages.dev |
| Branch Preview URL: | https://feature-kimi-k2-thinking-mod.maple-ca8.pages.dev |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 0
🧹 Nitpick comments (1)
frontend/src/components/UnifiedChat.tsx (1)
2805-2837: LGTM! Consistent implementation across both toolbars.The thinking toggle button is correctly duplicated in the bottom toolbar with identical functionality, ensuring consistent behavior across both the centered input (new chat) and bottom toolbar (active chat) views.
While the current duplication is acceptable, consider extracting the thinking toggle into a shared component if more customization is needed in the future. For now, the consistent implementation across both locations ensures a reliable user experience.
📜 Review details
Configuration used: defaults
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (2)
frontend/src/components/ModelSelector.tsx(2 hunks)frontend/src/components/UnifiedChat.tsx(2 hunks)
🧰 Additional context used
📓 Path-based instructions (1)
**/*.{ts,tsx}
📄 CodeRabbit inference engine (CLAUDE.md)
**/*.{ts,tsx}: Use path aliases (@/*maps to./src/*) for imports in TypeScript/React files
Use 2-space indentation, double quotes, and enforce 100-character line limit in TypeScript/React code
Maintain strict TypeScript and avoid usinganytype
Use PascalCase for component names and camelCase for variables and function names
Use functional components with React hooks instead of class components
Use React context for global state management and TanStack Query for server state management
Runjust format,just lint, andjust buildafter making TypeScript/React changes to ensure code quality and compilation
Files:
frontend/src/components/UnifiedChat.tsxfrontend/src/components/ModelSelector.tsx
🧬 Code graph analysis (1)
frontend/src/components/UnifiedChat.tsx (2)
frontend/src/components/ModelSelector.tsx (1)
CATEGORY_MODELS(113-120)frontend/src/components/ui/button.tsx (1)
Button(62-62)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (6)
- GitHub Check: Greptile Review
- GitHub Check: build-macos (universal-apple-darwin)
- GitHub Check: build-ios
- GitHub Check: build-linux
- GitHub Check: build-android
- GitHub Check: Cloudflare Pages
🔇 Additional comments (3)
frontend/src/components/ModelSelector.tsx (2)
71-77: LGTM! Well-configured new model entry.The Kimi K2 Thinking model configuration is complete and follows the established pattern. The 256k token limit is a nice improvement over DeepSeek R1's 130k, providing users with more context capacity.
116-117: LGTM! Clear model mapping for thinking toggle.The CATEGORY_MODELS update correctly maps reasoning modes to their respective models with clear comments. The existing useEffect (lines 247-254) ensures the thinking state stays synchronized when the model changes.
frontend/src/components/UnifiedChat.tsx (1)
2528-2560: LGTM! Well-implemented thinking toggle with clear visual feedback.The thinking toggle button is properly implemented with:
- Correct conditional rendering for reasoning models only
- Clean state management that updates both
thinkingEnabledand switches between models- Clear visual feedback via Brain icon color (purple when active, muted when inactive)
- Proper accessibility with descriptive aria-labels
Greptile SummarySuccessfully adds Kimi K2 Thinking as a new reasoning model and re-enables the brain toggle feature. The implementation groups reasoning content, tool calls, and message content into a single assistant box for better UX. Key Changes:
Confidence Score: 4/5
Important Files Changed
Sequence DiagramsequenceDiagram
participant User
participant UnifiedChat
participant ModelSelector
participant API
participant MessageList
User->>UnifiedChat: Select "Reasoning" category
UnifiedChat->>ModelSelector: Load reasoning_on model (Kimi K2)
ModelSelector-->>UnifiedChat: Model set to kimi-k2-thinking
User->>UnifiedChat: Click brain toggle button
UnifiedChat->>UnifiedChat: Toggle thinkingEnabled state
alt Brain ON (thinking enabled)
UnifiedChat->>ModelSelector: Set model to reasoning_on
ModelSelector-->>UnifiedChat: Switch to kimi-k2-thinking
UnifiedChat-->>User: Brain icon filled
else Brain OFF (thinking disabled)
UnifiedChat->>ModelSelector: Set model to reasoning_off
ModelSelector-->>UnifiedChat: Switch to deepseek-r1-0528
UnifiedChat-->>User: Brain icon outlined
end
User->>UnifiedChat: Send message
UnifiedChat->>API: Stream request with selected model
alt Kimi K2 (reasoning_on)
API-->>UnifiedChat: response.output_item.added (type: reasoning)
UnifiedChat->>MessageList: Create ReasoningItem
API-->>UnifiedChat: response.reasoning_text.delta
UnifiedChat->>MessageList: Update reasoning content (streaming)
MessageList-->>User: Show ThinkingBlock (collapsible)
API-->>UnifiedChat: response.reasoning_text.done
UnifiedChat->>MessageList: Mark reasoning completed
API-->>UnifiedChat: response.output_text.delta
UnifiedChat->>MessageList: Stream assistant message
MessageList-->>User: Show reasoning + message in Maple box
else DeepSeek R1 (reasoning_off)
API-->>UnifiedChat: response.output_text.delta
UnifiedChat->>MessageList: Stream assistant message directly
MessageList-->>User: Show message in Maple box (no reasoning)
end
|
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 0
🧹 Nitpick comments (1)
frontend/src/components/UnifiedChat.tsx (1)
2593-2625: LGTM! Brain toggle correctly re-enabled with proper model switching.The thinking toggle implementation correctly:
- Shows only when reasoning models are selected (kimi-k2-thinking or deepseek-r1-0528)
- Switches between Kimi K2 (with thinking) and DeepSeek R1 (without thinking) per the PR objectives
- Provides clear visual feedback with icon color changes (purple when enabled, muted when disabled)
- Includes accessible aria-labels that describe the toggle action
The toggle appears in both the centered input view (lines 2593-2625) and bottom input view (lines 2870-2902), which is appropriate for the different layouts.
💡 Optional: Consider extracting duplicate toggle logic to reduce code duplication
The toggle button logic is duplicated between the centered and bottom input views. While acceptable given the different rendering contexts, you could optionally extract it to a memoized component or helper function:
const ThinkingToggleButton = memo(({ className }: { className?: string }) => { const localState = useLocalState(); if ( localState.model !== CATEGORY_MODELS.reasoning_on && localState.model !== CATEGORY_MODELS.reasoning_off ) { return null; } return ( <Button type="button" variant="ghost" size="sm" className={`h-8 w-8 p-0 ${className || ''}`} onClick={() => { const newThinkingEnabled = !localState.thinkingEnabled; localState.setThinkingEnabled(newThinkingEnabled); localState.setModel( newThinkingEnabled ? CATEGORY_MODELS.reasoning_on : CATEGORY_MODELS.reasoning_off ); }} aria-label={ localState.thinkingEnabled ? "Disable thinking mode" : "Enable thinking mode" } > <Brain className={`h-4 w-4 ${ localState.thinkingEnabled ? "text-purple-500" : "text-muted-foreground" }`} /> </Button> ); });This is a low-priority refactor and can be deferred if preferred.
Also applies to: 2870-2902
📜 Review details
Configuration used: defaults
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (1)
frontend/src/components/UnifiedChat.tsx(5 hunks)
🧰 Additional context used
📓 Path-based instructions (1)
**/*.{ts,tsx}
📄 CodeRabbit inference engine (CLAUDE.md)
**/*.{ts,tsx}: Use path aliases (@/*maps to./src/*) for imports in TypeScript/React files
Use 2-space indentation, double quotes, and enforce 100-character line limit in TypeScript/React code
Maintain strict TypeScript and avoid usinganytype
Use PascalCase for component names and camelCase for variables and function names
Use functional components with React hooks instead of class components
Use React context for global state management and TanStack Query for server state management
Runjust format,just lint, andjust buildafter making TypeScript/React changes to ensure code quality and compilation
Files:
frontend/src/components/UnifiedChat.tsx
🧬 Code graph analysis (1)
frontend/src/components/UnifiedChat.tsx (1)
frontend/src/components/ModelSelector.tsx (1)
CATEGORY_MODELS(113-120)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (5)
- GitHub Check: build-ios
- GitHub Check: Greptile Review
- GitHub Check: build-android
- GitHub Check: build-linux
- GitHub Check: build-macos (universal-apple-darwin)
🔇 Additional comments (2)
frontend/src/components/UnifiedChat.tsx (2)
1655-1655: LGTM! Reasoning streaming integration is well-implemented.The reasoning accumulation logic correctly handles the streaming flow:
- Incremental deltas are accumulated in
accumulatedReasoning- During streaming, reasoning is displayed with an opening
<think>tag (no closing tag to indicate in-progress state)- On
reasoning_done, the reasoning is finalized with a closing</think>tag and appended to any existing content- Message deduplication is properly handled via
mergeMessagesByIdThis implementation aligns with the PR objective to add Kimi K2 Thinking model support with extended reasoning visualization.
Also applies to: 1806-1864
1878-1882: LGTM! Content display properly integrates reasoning.The display concatenation logic correctly prepends accumulated reasoning to the content when both exist. The
\n\nspacing between the closing</think>tag and content provides visual separation in the markdown rendering, which enhances readability.
There was a problem hiding this comment.
Actionable comments posted: 0
🧹 Nitpick comments (1)
frontend/src/components/UnifiedChat.tsx (1)
2650-2682: Consider extracting the thinking toggle button to a reusable component.This toggle button logic is duplicated in the compact/bottom input section (lines 2927-2959). While acceptable for now given the "Chill" review mode, extracting to a shared component would improve maintainability.
🔎 Example extraction:
// Could be defined near other button components function ThinkingToggleButton() { const localState = useLocalState(); if (localState.model !== CATEGORY_MODELS.reasoning_on && localState.model !== CATEGORY_MODELS.reasoning_off) { return null; } return ( <Button type="button" variant="ghost" size="sm" className="h-8 w-8 p-0" onClick={() => { const newThinkingEnabled = !localState.thinkingEnabled; localState.setThinkingEnabled(newThinkingEnabled); localState.setModel( newThinkingEnabled ? CATEGORY_MODELS.reasoning_on : CATEGORY_MODELS.reasoning_off ); }} aria-label={ localState.thinkingEnabled ? "Disable thinking mode" : "Enable thinking mode" } > <Brain className={`h-4 w-4 ${ localState.thinkingEnabled ? "text-purple-500" : "text-muted-foreground" }`} /> </Button> ); }
📜 Review details
Configuration used: defaults
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (2)
frontend/src/components/UnifiedChat.tsx(9 hunks)frontend/src/components/markdown.tsx(1 hunks)
🧰 Additional context used
📓 Path-based instructions (1)
**/*.{ts,tsx}
📄 CodeRabbit inference engine (CLAUDE.md)
**/*.{ts,tsx}: Use path aliases (@/*maps to./src/*) for imports in TypeScript/React files
Use 2-space indentation, double quotes, and enforce 100-character line limit in TypeScript/React code
Maintain strict TypeScript and avoid usinganytype
Use PascalCase for component names and camelCase for variables and function names
Use functional components with React hooks instead of class components
Use React context for global state management and TanStack Query for server state management
Runjust format,just lint, andjust buildafter making TypeScript/React changes to ensure code quality and compilation
Files:
frontend/src/components/markdown.tsxfrontend/src/components/UnifiedChat.tsx
🧬 Code graph analysis (1)
frontend/src/components/UnifiedChat.tsx (2)
frontend/src/components/markdown.tsx (1)
ThinkingBlock(31-119)frontend/src/components/ModelSelector.tsx (1)
CATEGORY_MODELS(113-120)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (6)
- GitHub Check: Greptile Review
- GitHub Check: build-linux
- GitHub Check: build-macos (universal-apple-darwin)
- GitHub Check: build-ios
- GitHub Check: build-android
- GitHub Check: Cloudflare Pages
🔇 Additional comments (10)
frontend/src/components/markdown.tsx (1)
25-31: LGTM!Clean export of
ThinkingBlockPropsinterface andThinkingBlockcomponent to enable reuse inUnifiedChat.tsxfor rendering reasoning items.frontend/src/components/UnifiedChat.tsx (9)
53-53: LGTM!Correct import using path alias as per coding guidelines.
106-123: LGTM!The
ReasoningItemtype correctly models reasoning content with appropriate status values that align with theThinkingBlockcomponent'sisThinkingprop logic. The types are kept internal (non-exported) which is appropriate for this use case.
473-494: LGTM!Correct integration of
ThinkingBlockfor reasoning items. The status-basedisThinkingdetection properly reflects the streaming state, and the text extraction handles the content array structure appropriately.
1686-1690: LGTM!The new tracking variables follow the established pattern for
serverAssistantIdandaccumulatedContent, maintaining consistency in the streaming response processing logic.
1737-1754: LGTM!The reasoning item creation handler follows the established pattern for other output item types, correctly initializing the item with
status: "in_progress"and tracking the ID for subsequent delta/done events.
1859-1884: LGTM!The reasoning delta handler correctly accumulates text and updates the reasoning item. The fallback from
event.item_idtoserverReasoningIdprovides resilience against missing event properties.
1885-1908: LGTM!The done handler correctly finalizes the reasoning item, using the event's
textproperty as the authoritative final content when available, and properly transitioning status to"completed".
1937-1970: LGTM!Good refactoring to make the
response.output_item.donehandler generic, supporting any item type (reasoning, message, etc.) while maintaining backward compatibility with theserverAssistantIdfallback.
2927-2959: LGTM!The thinking toggle button in the compact input correctly mirrors the centered input behavior. The duplication concern was noted above.
There was a problem hiding this comment.
Actionable comments posted: 0
🧹 Nitpick comments (1)
frontend/src/components/UnifiedChat.tsx (1)
3048-3080: Consider extracting the thinking toggle to a shared component.This toggle button is duplicated from lines 2771-2803. While both work correctly, extracting to a shared component (e.g.,
ThinkingToggleButton) would reduce duplication and prevent future drift if the logic needs updating.🔎 Example extraction:
// Could be defined near the top of the file or in a separate file function ThinkingToggleButton() { const localState = useLocalState(); if (localState.model !== CATEGORY_MODELS.reasoning_on && localState.model !== CATEGORY_MODELS.reasoning_off) { return null; } return ( <Button type="button" variant="ghost" size="sm" className="h-8 w-8 p-0" onClick={() => { const newThinkingEnabled = !localState.thinkingEnabled; localState.setThinkingEnabled(newThinkingEnabled); localState.setModel( newThinkingEnabled ? CATEGORY_MODELS.reasoning_on : CATEGORY_MODELS.reasoning_off ); }} aria-label={ localState.thinkingEnabled ? "Disable thinking mode" : "Enable thinking mode" } > <Brain className={`h-4 w-4 ${ localState.thinkingEnabled ? "text-purple-500" : "text-muted-foreground" }`} /> </Button> ); }Then replace both instances with
<ThinkingToggleButton />.
📜 Review details
Configuration used: defaults
Review profile: CHILL
Plan: Pro
📒 Files selected for processing (1)
frontend/src/components/UnifiedChat.tsx(13 hunks)
🧰 Additional context used
📓 Path-based instructions (1)
**/*.{ts,tsx}
📄 CodeRabbit inference engine (CLAUDE.md)
**/*.{ts,tsx}: Use path aliases (@/*maps to./src/*) for imports in TypeScript/React files
Use 2-space indentation, double quotes, and enforce 100-character line limit in TypeScript/React code
Maintain strict TypeScript and avoid usinganytype
Use PascalCase for component names and camelCase for variables and function names
Use functional components with React hooks instead of class components
Use React context for global state management and TanStack Query for server state management
Runjust format,just lint, andjust buildafter making TypeScript/React changes to ensure code quality and compilation
Files:
frontend/src/components/UnifiedChat.tsx
🧬 Code graph analysis (1)
frontend/src/components/UnifiedChat.tsx (2)
frontend/src/components/markdown.tsx (2)
ThinkingBlock(31-119)Markdown(676-705)frontend/src/components/ModelSelector.tsx (1)
CATEGORY_MODELS(113-120)
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (5)
- GitHub Check: Greptile Review
- GitHub Check: build-ios
- GitHub Check: build-android
- GitHub Check: build-macos (universal-apple-darwin)
- GitHub Check: build-linux
🔇 Additional comments (11)
frontend/src/components/UnifiedChat.tsx (11)
53-53: LGTM!The
ThinkingBlockimport is correctly added using the path alias convention.
106-124: LGTM!The new
ReasoningItemtype is well-structured with proper discriminant (type: "reasoning") for the union type. The content array structure mirrors existing patterns and allows for future extensibility.
364-368: LGTM!The
MessageGrouptype cleanly separates user and assistant turns, with the assistant turn properly containing an array of related items (reasoning, tool calls, messages).
400-439: LGTM!The grouping logic correctly separates user messages from assistant-related items (reasoning, tool calls, messages). The use of
useMemooptimizes re-computation, and the flush-then-push pattern ensures proper ordering.
484-498: LGTM!The reasoning item rendering correctly extracts text content and determines the
isThinkingstate based on both"in_progress"and"streaming"statuses. TheThinkingBlockcomponent will display the animated "Thinking for X seconds" indicator while streaming and "Thought for X seconds" when complete.
570-583: Reasoning text is excluded from copy.The copy button only captures assistant message content, not the reasoning/thinking text. This seems intentional for typical use cases. If users should be able to copy reasoning content too, consider including it or providing a separate option.
729-741: LGTM!The loading indicator logic correctly accounts for streaming reasoning items, preventing duplicate loading indicators when
ThinkingBlockis already displaying its own "Thinking..." animation.
1858-1875: LGTM!The reasoning item is correctly created when
response.output_item.addedfires withtype: "reasoning". The initial state with empty content and"in_progress"status matches the pattern used for assistant messages.
1980-2029: LGTM!The delta and done handlers for reasoning text follow the established pattern for streaming content:
- Delta accumulates text and sets status to
"streaming"- Done finalizes with the complete text (from event or accumulated) and sets status to
"completed"- Both handlers correctly fall back to
serverReasoningIdwhenitem_idis not in the event
2058-2077: LGTM!The
response.output_item.donehandler is correctly generalized to handle completion for any item type (reasoning, message, etc.), updating the status to"completed"and thelastSeenItemIdfor polling.
2771-2803: LGTM!The thinking toggle correctly:
- Shows only when a reasoning model is selected
- Toggles between Kimi K2 (with thinking) and DeepSeek R1 (without thinking)
- Updates both
thinkingEnabledstate and the active model- Uses appropriate visual feedback (purple when enabled)
There was a problem hiding this comment.
Actionable comments posted: 0
🧹 Nitpick comments (1)
frontend/src/components/Marketing.tsx (1)
436-436: Consider verifying mobile layout on small screens.The grid layout change from 5-column to 3-column desktop creates a balanced 2×3 grid with 6 models. However, the mobile layout now displays 2 columns instead of 1, which may be tight on smaller devices (320px-375px width) given the badge images and labels.
Consider testing the layout on small mobile devices (iPhone SE, small Android phones) to ensure the badge images and labels render comfortably in the 2-column layout.
📜 Review details
Configuration used: defaults
Review profile: CHILL
Plan: Pro
⛔ Files ignored due to path filters (1)
frontend/public/badge-kimi-logo.pngis excluded by!**/*.png
📒 Files selected for processing (1)
frontend/src/components/Marketing.tsx(2 hunks)
🧰 Additional context used
📓 Path-based instructions (1)
**/*.{ts,tsx}
📄 CodeRabbit inference engine (CLAUDE.md)
**/*.{ts,tsx}: Use path aliases (@/*maps to./src/*) for imports in TypeScript/React files
Use 2-space indentation, double quotes, and enforce 100-character line limit in TypeScript/React code
Maintain strict TypeScript and avoid usinganytype
Use PascalCase for component names and camelCase for variables and function names
Use functional components with React hooks instead of class components
Use React context for global state management and TanStack Query for server state management
Runjust format,just lint, andjust buildafter making TypeScript/React changes to ensure code quality and compilation
Files:
frontend/src/components/Marketing.tsx
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (5)
- GitHub Check: build-ios
- GitHub Check: build-android
- GitHub Check: build-macos (universal-apple-darwin)
- GitHub Check: build-linux
- GitHub Check: Cloudflare Pages
🔇 Additional comments (1)
frontend/src/components/Marketing.tsx (1)
30-34: LGTM! Image asset verified.The Kimi K2 model entry is correctly structured and consistent with other entries in the AI_MODELS array. The badge image file exists at
./frontend/public/badge-kimi-logo.png. Code follows all formatting guidelines with proper 2-space indentation and double quotes.
- Add kimi-k2-thinking to MODEL_CONFIG (256k context, Pro tier) - Set Kimi K2 as reasoning_on model (with thinking) - Keep DeepSeek R1 as reasoning_off model (without thinking) - Re-enable brain toggle button in chat UI Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Handle response.reasoning_text.delta and response.reasoning_text.done events for models like Kimi K2 that stream reasoning separately from content. Reasoning is wrapped in <think> tags to reuse existing ThinkingBlock rendering. Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
- Export ThinkingBlock from markdown.tsx for reuse - Add ReasoningItem type to Message union for type: reasoning items - Add rendering for reasoning items in MessageList using ThinkingBlock - Handle response.output_item.added for reasoning items during streaming - Update reasoning_text.delta/done handlers to update reasoning items - Remove inline <think> tag wrapping from output_text.delta Reasoning items are now proper conversation items that can be fetched via the items API and are handled with the same dedup/paging logic as messages and tool calls. Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Update handler to use item.id from the event to mark the correct item as completed, rather than only handling serverAssistantId. This properly supports the separate reasoning and message items in the streaming flow. Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
Refactor MessageList to group consecutive assistant items (reasoning, tool calls, message content) into a single 'Maple' box instead of rendering them separately. This provides a cleaner UI similar to how DeepSeek's <think> tags rendered inline. Changes: - Add MessageGroup type for user/assistant turn grouping - Group messages using useMemo before rendering - Render all assistant items within one Maple wrapper - Fix loading indicator to not show duplicate box during tool use - Export ThinkingBlock from markdown.tsx for reuse Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
- Add Kimi logo (badge-kimi-logo.png) - Add Kimi K2 to AI_MODELS array - Change grid to 3 columns for balanced 2-row layout Closes #345 Co-authored-by: factory-droid[bot] <138933559+factory-droid[bot]@users.noreply.github.com>
AnthonyRonning
force-pushed
the
feature/kimi-k2-thinking-model
branch
from
1a62605 to
e6a20f1
Compare
Summary
Adds Kimi K2 Thinking as a new reasoning model and re-enables the brain toggle feature.
Changes
kimi-k2-thinkingto MODEL_CONFIG (256k context, Pro tier, Reasoning badge)reasoning_onmodel (with thinking enabled)reasoning_offmodel (without thinking)How it works
When users select the Reasoning category, they can use the brain toggle to switch between:
Related
Summary by CodeRabbit
New Features
UI Improvements
✏️ Tip: You can customize this high-level summary in your review settings.