- search_places: detect HTTP 429 and mark retryable=False to stop the
retry loop immediately instead of spiraling until MAX_ITERATIONS
- get_weather: extract collection coordinates (lat/lng from first
location with coords) and retry when LLM omits required params;
uses sync_to_async for the DB query in the async view
- AITravelChat: deduplicate context-only tools (get_trip_details,
get_weather) in the render pipeline to prevent duplicate place cards
from appearing when the retry loop causes multiple get_trip_details calls
- Tests: 5 new tests covering 429 non-retryable path and weather
coord fallback; all 39 chat tests pass
- Add _COMMAND_VERBS guard to _is_likely_location_reply() so messages
starting with imperative verbs (find, search, show, get, ...) are not
mistakenly treated as user location replies. This prevented 'Find
good places' from being used as a retry location, which was causing
the clarification path to never fire and the tool loop to exhaust
MAX_ALL_FAILURE_ROUNDS instead.
- Extract city from comma-delimited fallback address strings when
city/country FKs are absent, e.g. 'Little Turnstile 6, London'
→ 'London', so context-based location retry works for manually-
entered itinerary stops without geocoded FK data.
- Add attempted_location_retry flag: if retry was attempted but all
retry attempts failed, convert result to an execution failure rather
than emitting a clarification prompt (user already provided context
via their itinerary).
- Fix test assertion ordering in test_collection_context_retry_extracts_
city_from_fallback_address: streaming_content must be consumed before
checking mock call counts since StreamingHttpResponse is lazy.
Replaced `node:22-alpine` + `npm install -g bun@1.2.22` with
`oven/bun:1.2.22-alpine` as the builder stage base image.
The npm-based bun install was failing on linux/arm64 with:
"Failed to find package @oven/bun-linux-aarch64"
The official oven/bun Docker image supports both linux/amd64 and
linux/arm64 natively, eliminating the need to install bun via npm.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
- Fetch models dynamically from OpenCode Zen API (36 models vs 5 hardcoded)
- Add function calling support check before using tools
- Add retry logic (num_retries=2) for transient failures
- Improve logging for debugging API calls and errors
- Update system prompt for multi-stop itinerary context
- Clean up unused imports in frontend components
- Remove deleted views.py (moved to views/__init__.py)
- Connection error: wrap sync get_llm_api_key() in sync_to_async in
stream_chat_completion() to fix SynchronousOnlyOperation raised when
the async SSE generator calls a synchronous Django ORM function
- Models not loading: add opencode_zen handler to models endpoint
returning its default model; fix frontend to show 'Default' instead
of 'Loading...' indefinitely when no model list is returned
- Location in header: remove destination subtitle from Travel Assistant
header — collection-wide chat has no single meaningful location
Fix 1: Provider/Model Selection (Critical - unblocks LLM)
- Add /api/chat/providers/{id}/models/ endpoint to fetch available models
- Auto-select first configured provider instead of hardcoded 'openai'
- Add model dropdown populated from provider API
- Filter provider list to only show configured providers
- Show helpful error when no providers configured
Fix 2: Auto-Learn Preferences (Replaces manual input)
- Create auto_profile.py utility to infer preferences from user data
- Learn interests from Activity sport types and Location categories
- Learn trip style from Lodging types (hostel=budget, resort=luxury, etc.)
- Learn geographic preferences from VisitedRegion/VisitedCity
- Call auto-learn on every chat start (send_message)
- System prompt now indicates preferences are auto-inferred
Fix 3: Remove Manual Preference UI
- Remove travel_preferences section from Settings
- Remove preference form fields and save logic
- Remove preference fetch from server-side load
- Keep UserRecommendationPreferenceProfile type for backend use
The LLM should now work correctly:
- Users with any configured provider will have it auto-selected
- Model list is fetched dynamically from provider API
- Preferences are learned from actual travel history
Add two new sections to AGENTS.md and CLAUDE.md:
- .memory Files: consult knowledge.md, decisions.md, plans/, research/ at task start
- Instruction File Sync: keep AGENTS.md, CLAUDE.md, .cursorrules, and Copilot CLI instructions in sync when any is updated
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
The optimize function always started nearest-neighbor from the first
array element, ignoring where the traveler actually is after preceding
anchors (flights, lodging). Now passes the preceding anchor's exit
coordinates (destination for transportation) so the algorithm picks
the spatially nearest item as the starting point.
Prevent API key and sensitive info leakage through exception messages:
- Replace str(exc) with generic error messages in all catch-all handlers
- Add server-side exception logging via logger.exception()
- Add ALLOWED_KWARGS per-tool allowlist to filter untrusted LLM kwargs
- Bound tool execution loop to MAX_TOOL_ITERATIONS=10
- Fix tool_call delta merge to use tool_call index
Implement a full chat-based travel agent using LiteLLM for multi-provider
LLM support (OpenAI, Anthropic, Gemini, Ollama, Groq, Mistral, etc.).
Backend:
- New 'chat' Django app with ChatConversation and ChatMessage models
- Streaming SSE endpoint via StreamingHttpResponse
- 5 agent tools: search_places, list_trips, get_trip_details,
add_to_itinerary, get_weather
- LiteLLM client wrapper with per-user API key retrieval
- System prompt with user preference context injection
Frontend:
- New /chat route with full-page chat UI (DaisyUI + Tailwind)
- Collapsible conversation sidebar with CRUD
- SSE streaming response display with tool call visualization
- Provider selector dropdown
- SSE proxy fix to stream text/event-stream without buffering
- Navbar link and i18n keys
Temperature display always showed 'unavailable' because formatDayTemperature()
read dayTemperatures from closure, which Svelte 4's compiler doesn't track in
template expressions. Same issue affected getLocationConnector() reading
connectorMetricsMap from closure.
Fix: pass both state variables as explicit function parameters so they appear
in template expressions and trigger re-renders on async updates.
Also improve optimize button diagnostics with console logging and better
toast feedback when items lack coordinates.
- fix optimizeDayOrder() dual-update: directly set days[dayIndex].items + days before
saveReorderedItems() so it reads the correct post-optimization order synchronously
(Svelte 4 batches reactive statements; days wasn't updated before save read it)
- also patch collection.itinerary order values so reactive rebuild uses new order
- replace single google_maps <option> with 8 AI LLM provider options
(anthropic, openai, gemini, ollama, groq, mistral, github_models, openrouter)
- add getApiKeyProviderLabel() helper for saved key display with google_maps fallback
- add i18n keys for all new provider labels in en.json and de.json
Unify collections and shared items under a single Collections tab while keeping Archive separate, and fix card layering so menus render correctly. Restore invite discoverability by adding navbar access to /invites and add missing i18n keys to prevent raw key labels in collections/invites UI.
