Tool descriptions now explicitly state when models should use them:
- Memory Store: 'Use when user asks to remember, recall, list memories'
- Task Manager: 'Use when user mentions tasks, todos, things to do'
- Structured Thinking: 'Use for complex questions requiring analysis'
- Decision Matrix: 'Use when comparing options or recommendations'
- Project Planner: 'Use for planning, roadmaps, breaking work into phases'
This helps smaller models understand WHEN to call these tools instead of
just responding with text claiming they don't have memory/capabilities.
When the model calls recall without key or category (e.g., 'what memories
do you have?'), it now returns all memories across all categories instead
of an error. This provides better UX since models often use recall instead
of list for memory queries.
- Add validation for missing key/value in store action
- Return formatted entries in recall (category only) - consistent with list
- Check if category exists before reporting success in forget
- Include value in store response for confirmation
- Include entriesRemoved count when forgetting a category
The list action was only returning keys and counts, not the actual
stored values. Now it returns full entries with key, value, and
stored timestamp for all memories.
Models like ministral output tool calls as plain text (e.g., tool_name[ARGS]{json})
instead of using Ollama's native tool_calls format. This adds a parser that:
- Detects text-based tool call patterns in model output
- Converts them to OllamaToolCall format for execution
- Cleans the raw tool call text from the message
- Shows proper ToolCallDisplay UI with styled output
Supports three formats:
- tool_name[ARGS]{json}
- <tool_call>{"name": "...", "arguments": {...}}</tool_call>
- {"tool_calls": [...]} JSON blobs
- Add curated prompt templates with categories (coding, writing, analysis,
creative, assistant) that users can browse and add to their library
- Add "Browse Templates" tab to the Prompts page with category filtering
and preview functionality
- Add Design Brief Generator tool template for creating structured design
briefs from project requirements
- Add Color Palette Generator tool template for generating harmonious
color schemes from a base color
Prompts included: Code Reviewer, Refactoring Expert, Debug Assistant,
API Designer, SQL Expert, Technical Writer, Marketing Copywriter,
UI/UX Advisor, Security Auditor, Data Analyst, Creative Brainstormer,
Storyteller, Concise Assistant, Patient Teacher, Devil's Advocate,
Meeting Summarizer
Adds the ability to test HTTP endpoint custom tools directly in the
editor, matching the existing test functionality for Python and
JavaScript tools. Closes #6.
- Limit max attachments to 5 files to prevent context overflow
- Fix URL update timing: use SvelteKit's replaceState in onComplete
callback instead of history.replaceState before streaming
- Load attachment content from IndexedDB in conversation history
so follow-up messages have access to file content
- Show error messages in chat when Ollama fails instead of stuck
"Processing..." indicator
- Force file analysis when >3 files attached to reduce context usage
Clear 'Processing...' text only when first token arrives, not before
the LLM request. This keeps the indicator visible during prompt
resolution, RAG retrieval, and LLM initialization.
- Add "Processing X files..." indicator in chat while handling attachments
- Indicator transitions to "Analyzing X files..." for large files needing LLM summarization
- Reuse streaming message for seamless transition to LLM response
- Add FileAnalyzer service for large file summarization with 10s timeout
- Skip analysis for borderline files (within 20% of 8K threshold)
- Read up to 50KB from original file for analysis (not just truncated content)
- Remove base64 blobs from JSON before analysis to reduce prompt size
- Add AttachmentDisplay component for showing file badges on messages
- Persist attachments to IndexedDB with message references
- Add chat state methods: setStreamContent, removeMessage
- Clean up debug logging
- Add postinstall script to copy worker to static/
- Update Dockerfile to copy worker during build
- Update file-processor to try local worker first, fallback to CDN
- Bump version to 0.4.11
Adds two related features for enhanced model customization:
**Model-Specific System Prompts:**
- Assign prompts to models via Settings > Model Prompts
- Capability-based default prompts (vision, tools, thinking, code)
- Auto-select appropriate prompt when switching models in chat
- Per-model prompt mappings stored in IndexedDB
**Custom Ollama Model Creation:**
- Create custom models with embedded system prompts via Models page
- Edit system prompts of existing custom models
- Streaming progress during model creation
- Visual "Custom" badge for models with embedded prompts
- Backend handler for Ollama /api/create endpoint
New files:
- ModelEditorDialog.svelte: Create/edit dialog for custom models
- model-creation.svelte.ts: State management for model operations
- model-prompt-mappings.svelte.ts: Model-to-prompt mapping store
- model-info-service.ts: Fetches and caches model info from Ollama
- modelfile-parser.ts: Parses system prompts from Modelfiles
Custom tools were displayed as enabled in the UI but never sent to
Ollama because getEnabledToolDefinitions() only queried the builtin
tool registry. Now iterates customTools and includes enabled ones.
Fixes #4
- Fix TypeError when check updates returns null updates array
- Display verified capabilities from Ollama runtime in Local Models tab
- Fetch all model capabilities on page mount
- Add data-dev to gitignore
- Add capability verification for installed models using /api/show
- SyncModels now updates real capabilities when fetchDetails=true
- Model browser shows verified/unverified badges for capabilities
- Add info notice that capabilities are sourced from ollama.com
- Fix incorrect capability data (e.g., deepseek-r1 "tools" badge)
Capabilities from ollama.com website may be inaccurate. Once a model
is installed, Vessel fetches actual capabilities from Ollama runtime
and displays a "verified" badge in the model details panel.
- Add size filter (≤3B, 4-13B, 14-70B, >70B) based on model tags
- Add model family filter dropdown with dynamic family list
- Display last updated date on model cards (scraped from ollama.com)
- Add /api/v1/models/remote/families endpoint
- Convert relative time strings ("2 weeks ago") to timestamps during sync
- Add customMaxTokens override to ContextManager
- maxTokens is now derived from custom override or model default
- ChatWindow syncs settings.num_ctx to context manager
- Progress bar now shows custom context length when enabled
- Fetch actual model defaults from Ollama's /api/show endpoint
- Parse YAML-like parameters field (e.g., "temperature 0.7")
- Cache model defaults to avoid repeated API calls
- Initialize sliders with model's actual values when enabling custom params
- Show asterisk indicator when parameter differs from model default
- Reset button now restores to model defaults, not hardcoded values
Tool usage was not showing after page reload because the toolCalls
field was not being included when saving assistant messages to the
database. Now toolCalls are properly persisted and restored.
- Add CodeMirror editor with syntax highlighting for JavaScript and Python
- Add 8 starter templates (4 JS, 4 Python) for common tool patterns
- Add inline documentation panel with language-specific guidance
- Add tool testing UI to run tools with sample inputs before saving
- Add Python tool execution via backend API with 30s timeout
- Add POST /api/v1/tools/execute endpoint for backend tool execution
- Update Dockerfile to include Python 3 for tool execution
- Bump version to 0.4.0
Backend:
- Add /api/v1/version endpoint returning current and latest version
- Fetch latest release from GitHub API with 1-hour cache
- Semver comparison to detect available updates
- Configurable via GITHUB_REPO env var (default: vikingowl/vessel)
Frontend:
- Add VersionState store with 12-hour periodic checking
- Check on app load and periodically for new versions
- Persist dismissed versions in localStorage
- Add UpdateBanner component with teal styling
- Slide-in animation from top, dismissible
The notification appears below TopNav when a new version is available
and remembers dismissals per-version across sessions.
- Remove curl --max-filesize to avoid hard failures on large pages
- Add Truncated/OriginalSize fields to FetchResult for all fetch methods
- Return truncation info in proxy response (truncated, originalSize, returnedSize)
- Add timeout parameter to fetch_url tool (default 30s, max 120s)
- Increase default maxLength from 5KB to 50KB, allow up to 2MB
- Include _hint in response guiding LLM to retry with larger maxLength
Instead of failing when content exceeds limits, the tool now returns
truncated content with guidance for the LLM to request more if needed.
Auto-scroll now stops when the top of the assistant's response
reaches the top of the viewport, allowing users to read from
the beginning while more content streams below.
The SvelteKit hooks proxy was only forwarding pathname without
the query string (event.url.search), causing filters and pagination
to be silently dropped in production builds.
