* feat: add daily token limit with actual usage tracking
- Add DAILY_TOKEN_LIMIT env var for configurable daily token limit
- Track actual tokens from Bedrock API response metadata (not estimates)
- Server sends inputTokens + cachedInputTokens + outputTokens via messageMetadata
- Client increments token count in onFinish callback with actual usage
- Add NaN guards to prevent corrupted localStorage values
- Add token limit toast notification with quota display
- Remove client-side token estimation (was blocking legitimate requests)
- Switch to js-tiktoken for client compatibility (pure JS, no WASM)
* feat: add TPM (tokens per minute) rate limiting
- Add 50k tokens/min client-side rate limit
- Track tokens per minute with automatic minute rollover
- Check TPM limit after daily limits pass
- Show toast when rate limit reached
- NaN guards for localStorage values
* feat: make TPM limit configurable via TPM_LIMIT env var
* chore: restore cache debug logs
* fix: prevent race condition in TPM tracking
checkTPMLimit was resetting TPM count to 0 when checking, which
overwrote the count saved by incrementTPMCount. Now checkTPMLimit
only reads and incrementTPMCount handles all writes.
* chore: improve TPM limit error message clarity
- Add DAILY_REQUEST_LIMIT env var support in config API
- Track request count in localStorage (resets daily)
- Show friendly quota limit toast with self-host/sponsor links
- Apply limit to send, regenerate, and edit message actions
- Fix bug where text after tool calls was merged with initial text
- Group consecutive text/file parts into bubbles while keeping tools in order
- Parts now display as: plan -> tool_result -> additional text
- Remove debug logs from fixToolCallInputs function
Co-authored-by: dayuan.jiang <jiangdy@amazon.co.jp>
- Add client-side cache check in onFormSubmit to bypass API calls for example prompts
- Use findCachedResponse to match input against cached examples
- Directly set messages with cached tool response when example matches
- Hide regenerate button for cached example responses (toolCallId starts with 'cached-')
- Prevents unnecessary API calls when using example buttons
Co-authored-by: dayuan.jiang <jiangdy@amazon.co.jp>
- Add validation to loadDiagram in diagram-context, returns error or null
- display_diagram and edit_diagram tools now check validation result
- Return error to AI agent with state: output-error so it can retry
- Skip validation for trusted sources (localStorage, history, internal templates)
- Add debug logging for tool call inputs to diagnose Bedrock API issues
* Fix: remove hardcoded temperature parameter to support reasoning models
* feat: make temperature configurable via AI_TEMPERATURE env var
- Instead of removing temperature entirely, make it optional via env var
- Set AI_TEMPERATURE=0 for deterministic output (recommended for diagrams)
- Leave unset for models that don't support temperature (e.g., GPT-5.1 reasoning)
* docs: add AI_TEMPERATURE env var documentation
- Update env.example with AI_TEMPERATURE option
- Update README.md configuration section
- Add Temperature Setting section in ai-providers.md
* docs: add TEMPERATURE env var documentation
- Update env.example with TEMPERATURE option
- Update README.md, README_CN.md, README_JA.md configuration sections
- Add Temperature Setting section in ai-providers.md
- Update route.ts to use TEMPERATURE env var
---------
Co-authored-by: dayuan.jiang <jiangdy@amazon.co.jp>
- Add fixToolCallInputs() to fix Bedrock API requirement (JSON object, not string)
- Add experimental_repairToolCall for malformed JSON from model
- Add stepCountIs(5) limit to prevent infinite loops
- Update edit_diagram tool description with JSON escaping warning
Co-authored-by: dayuan.jiang <jiangdy@amazon.co.jp>
- Add Biome as formatter and linter (replaces Prettier)
- Configure Husky + lint-staged for pre-commit hooks
- Add VS Code settings for format on save
- Ignore components/ui/ (shadcn generated code)
- Remove semicolons, use 4-space indent
- Reformat all files to new style
- Add App Context section describing the left/right panel layout
- Add App Features section with icon locations (history, theme, upload, export, clear)
- Dynamically inject model name into system prompt via {{MODEL_NAME}} placeholder
- Expand edit_diagram tool description with usage guidelines
- Remove 15s streaming timeout detection (too slow, added complexity)
- Remove status indicator (issue resolved by switching model)
- Remove streamingError state and related refs
- Simplify onFinish callback (remove 503 detection logging)
- Remove errorHandler function (use default AI SDK errors)
The real fix was switching from global.* to us.* Bedrock model.
This removes ~134 lines of unnecessary complexity.
Bedrock streaming responses don't auto-report token usage to OpenTelemetry.
This fix manually sets span attributes (ai.usage.promptTokens, gen_ai.usage.input_tokens)
from the AI SDK onFinish callback to ensure Langfuse captures token counts.
- Add Zod schema validation for log-feedback and log-save endpoints
- Create singleton LangfuseClient to avoid per-request instantiation
- Simplify log-save to only flag trace (no XML content sent)
- Use generic error messages to prevent info leakage
- Update log-feedback API to find existing chat trace by sessionId and attach score to it
- Update log-save API to create span on existing chat trace instead of standalone trace
- Add thumbs up/down feedback buttons on assistant messages
- Add message regeneration and edit functionality
- Add save dialog with format selection (drawio, png, svg)
- Pass sessionId through components for Langfuse linking
- Extract system prompts to dedicated lib/system-prompts.ts module
- Add extended system prompt (~4000 tokens) for models with higher cache minimums (Opus 4.5, Haiku 4.5)
- Clean up debug logs while preserving informational and cache-related logs
- Improve code formatting and organization in chat route
* feat: add trace-level input/output to Langfuse observability
- Add @langfuse/client and @langfuse/tracing dependencies
- Wrap POST handler with observe() for proper tracing
- Use updateActiveTrace() to set trace input, output, sessionId, userId
- Filter Next.js HTTP spans in shouldExportSpan so AI SDK spans become root traces
- Enable recordInputs/recordOutputs in experimental_telemetry
* refactor: extract Langfuse logic to separate lib/langfuse.ts module
When models like DeepSeek (deepseek-chat, deepseek-reasoner) receive image
inputs, they return a cryptic error about 'unknown variant image_url'.
This change detects such errors and shows a clear message asking users
to remove the image or switch to a vision-capable model.
Fixes#42
- Add essential draw.io XML structure rules to system prompt
- Include critical rules about mxCell nesting (all must be direct children of root)
- Add shape/vertex and connector/edge examples with proper structure
- Improve tool description for display_diagram with validation rules
- Update xml_guide.md with better swimlane examples showing flat structure
- Add client-side XML validation to catch nested mxCell errors early
Helps address issues #40 (local Ollama models not working) and #39 (mxCell nesting errors)
- Add lib/cached-responses.ts with pre-generated XML for 4 example prompts
- Modify chat API route to check cache before calling AI
- Cache returns instant response (~0.26s) vs AI generation (~20-25s)
- Add "(cached for instant response)" text to example panel
- Cache only activates for first message with empty diagram
* feat: add Bedrock prompt caching for system and conversation messages
- Add cache point to system message (2558+ tokens cached)
- Add cache point to last assistant message in conversation history
- This caches the entire conversation prefix for subsequent requests
- Reduces latency and costs for multi-turn conversations
* refactor: remove duplicated system prompt
* fix: filter out messages with empty content arrays for Bedrock API
The convertToModelMessages function from AI SDK can produce messages with
empty content arrays when assistant messages have only tool call parts or
when tool results aren't properly converted. Bedrock API rejects these with
400 errors. This fix filters out invalid messages before sending to the API.
* fix: add diagnostic logging for empty message content
Added logging to capture the original UI message structure when empty content
is detected after conversion. This helps debug the root cause while the
filter provides a safety net for Bedrock API compatibility.
* fix: correct anthropic beta header config for fine-grained tool streaming
- Use bedrock.anthropicBeta for Bedrock provider (not additionalModelRequestFields)
- Use top-level headers for direct Anthropic API
- Update @ai-sdk/amazon-bedrock to 3.0.62
- Add headers support to ModelConfig interface
* fix: update @ai-sdk/amazon-bedrock to 3.0.62 for tool streaming support
- Add Examples section to README with 2-column grid layout
- Include demo images for GCP, AWS, Azure, animated connectors, and cat
- Update example panel buttons with clearer labels
- Add animated connector example button
- Add instruction for animated connectors in chat route
- Updated system prompt to allow up to 3 retry attempts with adjusted search patterns
- Simplified error response to provide current diagram XML and reference retry policy
- AI model self-manages retries based on system instructions
- Add formatXML function to format single-line XML with proper indentation
- Format chartXml after fetching to ensure consistency
- Update replaceXMLParts to handle single-line XML with substring fallback
- Improve edit_diagram tool guidance with SEARCH/REPLACE best practices
- Add concrete examples to help AI use minimal, targeted edits
