Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

…"text",...}])

Critical bug: HA messages use list content format, but inject_search_results
only handled string content. Search results were never actually injected.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

- HA addon config.yaml, Dockerfile, run.sh
- Full Vietnamese README in homeassistant-addon/
- Updated root README.md with quick start

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

fix: auto model now respects user config + list models uses intersection across all tokens

- gemini_free/auto now reads model from config.data.providers.gemini_free.model
- backend_router resolves auto by checking user config before hardcoded defaults
- /v1/models fetches models from all active ChatGPT tokens in parallel
- with multi-account, only shows models common to all tokens (intersection)
- falls back to anon if no tokens available

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
@

- Both Add New and Edit mode pickers now 420px wide
- Larger row height (py-2.5) and bigger font (13px)
- Better visual separation with borders between rows
- z-50 elevation for proper layering

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

- Popup width: 640px, max-height: 70vh
- Search input to filter models by name
- Backdrop overlay (click to close)
- ModelSearch state for text filtering

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

- Fixed centered modal 720px, max-height 80vh
- Click outside does NOT close (no backdrop handler)
- Close button (X) + Done button to dismiss
- Selected models show green checkmark + highlight
- Button shows count of selected models
- Edit mode also uses same modal style

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

- Unused providers shown in gray (#9ca3af) with no connection line
- Only providers with actual usage get colored nodes + animated edges
- Provider status determined by usage log, not config
- Falls back to showing all (without edges) if no usage data yet

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Changed prefix matching from exact "geminiapi/" to startsWith("geminiapi")
Now geminiapi1, geminiapi2, geminiapi3, geminiapi4 all get:
- image (image generation)
- vision (image analysis)
- video (video analysis)
capability labels in model list and combo picker.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Pre-configured for ports 8000-8003 with geminiapi1-4 prefixes.
Appears at top of provider presets for easy access.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

_is_model_enabled now checks: if the model's provider has no explicit
filtering configured, treat all its models as enabled. Fixes issue
where enabling model filter for chatgpt would hide all custom provider models.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

…hars

- Now detects relative paths like /media/img_xxx.png?token=yyy
- Prepends base_url for relative image URLs
- Fixed regex to exclude ) ] from URL capture
- Converts relative URLs to absolute before downloading

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

stream_image_outputs_with_pool now checks if model is a combo
and resolves to the first route with an IMAGE_MODELS entry.
Fixes "unsupported image model" error for image combos.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Gemini API Server returns images as markdown links in chat text.
Adapter now extracts URL from ![alt](url) syntax + downloads to base64.
Verified: image access works from within Docker container (200 OK).

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

- build_body now detects image data from edit endpoint
- Edit mode: sends original image as base64 + editing instruction
- Generate mode: existing behavior unchanged
- Adapted to OpenAI vision format (image_url type)

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

ImageComposer now accepts model + imageModels props.
Parent page.tsx needs wiring to fetch image models and pass down.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

- Added imageModel state (default gpt-image-2)
- Fetch image-capable models from /api/v1/models-with-capabilities
- Pass model + imageModels + onModelChange to ImageComposer
- Use selected model instead of hardcoded "gpt-image-2"

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Gemini returns content as array when prompt uses multi-part format.
parse_response now converts array content to string before parsing.
Also handles image_url type in content array.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

format_image_result now strips data:image/...;base64, prefix
before calling base64.b64decode to save image bytes.
Fixes Gemini custom provider image generation.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

- Removed "Hạn mức còn lại" display (frees space)
- Removed "Tải lên" button (frees space)
- Textarea: 16px font-semibold, min-h 120px, no absolute overlay
- Controls bar: flat layout below textarea (not floating on top)
- Removed sm:absolute positioning that caused text compression

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Button opens library image picker for selecting reference images
from previously generated images. Parent page.tsx wiring pending.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

- Button "Thư viện" opens modal with all generated images
- Grid view, click to select and add as reference image
- Downloads image, converts to File, adds via handleReferenceImageChange
- Modal closes on backdrop click or after selection

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

…int)

build_body now accepts 'images' list from _handle_adapter_edit,
encodes raw bytes to base64 for Gemini chat API.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>