…nd, chat_complete, response)

…rs, search, backup, Vietnamese UI

- BackendRouter: route text/image requests to best provider (chatgpt, opencode, sdwebui...)
- OpenCode free provider (noAuth): bypasses ChatGPT 24KB payload limit for HA
- RateLimitBackoff: exponential backoff 15 levels, per-model locking, error rules
- SearchService: ChatGPT/Gemini/Serper/SearXNG/Brave configurable search backends
- StateBackup: full-system export/import (accounts, providers, config, combos, image tasks)
- Image adapters: 7 providers ported from 9router (SD WebUI, HuggingFace, Cloudflare, Fal.ai, Stability, BFL, Gemini)
- Account service: virtual noAuth connections, health scoring (0.0-1.0)
- API endpoints: /api/v1/backup, /api/v1/restore, /api/v1/backups, /api/v1/health, /api/v1/providers
- UI: Vietnamese dark-mode dashboard, sidebar navigation, providers/combos/search/backup pages
- Config: backends, providers, rate_limit, combo_models, search sections

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

- GitHub Actions: build on main push + version tags, linux/amd64 + linux/arm64
- docker-compose.yml: healthcheck, named volume, clear env vars for Portainer
- Tags: latest (main), sha-XXXX, vX.Y.Z (semver)

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

- BackendRouter: resolve 'auto' → provider default model (nemotron-3-super-free for OpenCode)
- New 9router backup import: extract codex/ChatGPT tokens from 9router backup JSON
- New endpoint: POST /api/v1/import-9router — import tokens from 9router backup file
- Supports .json and .json.gz backup files
- Also extracts combo models and merges into config

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

- 9r/ model prefix proxies chat to 9router (handles OAuth: Claude/Codex/Copilot)
- ha-agent combo: 9router → OpenCode → ChatGPT (best fallback chain)
- 9router auto-connection from backup import
- Config: ninerouter.base_url + api_key section
- No 24KB limit when using 9r/ or oc/ models

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

- cx/ prefix uses Codex OAuth tokens to call api.openai.com directly
- No proxy to 9router needed — tokens from backup work immediately
- Gemini AI Studio provider (free 15 RPM or paid API key)
- 9router backup import marks tokens as type=codex
- add_accounts_with_type() for OAuth tokens
- ha-agent combo: cx/auto → oc/auto → chatgpt/auto

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

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

- Add OAuth token import in Accounts page (type=codex)
- File upload (drag-drop) for 9router backup on Backup page
- POST /api/v1/import-9router-upload for direct JSON upload
- POST /api/accounts/oauth for adding OAuth tokens
- Gemini AI Studio provider fully integrated
- account_service: add_accounts_with_type() for codex tokens

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

9router backup format has accessToken directly on connection objects,
not nested inside a "data" field. Also fix extract_all_oauth_tokens.

Local test: 10 Codex tokens extracted successfully from backup file.

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

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

- Fix import-9router endpoint route name (was incorrectly /api/v1/backup)
- Add oc/, cx/, ha-agent, combo models to /v1/models for HA visibility
- Static models: oc/auto, oc/nemotron-3-super-free, cx/auto, ha-agent

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

- _handle_opencode_chat now uses route.model (resolved) instead of raw model
- import-9router-upload uses Pydantic dict body instead of raw request.body()

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

…alling

- 9r/auto proxies to 9router which handles OAuth for Claude/Codex/Copilot
- ha-agent combo: 9r/auto → cx/auto → oc/auto
- 9router connection configurable via config.json ninerouter section

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

- Calls same endpoint as 9router (not api.openai.com)
- OpenAI chat format → Responses format translation
- Tool calling support via Codex Responses API
- No 9router dependency required
- cx/auto routes to Codex OAuth tokens from backup import

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

- Default model: gpt-5.3-codex (works with free ChatGPT accounts)
- Force stream=true and store=false (Codex API requirements)
- Tested: STATUS 200 from chatgpt.com/backend-api/codex/responses

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

- Add missing 'import json' causing 500 on import-9router-upload
- Accept 'codex' as valid provider (maps to openai_oauth)
- Add codex/ prefix in backend_router

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

- _stream_response now yields parsed dict chunks for FastAPI
- FastAPI auto-serializes dicts to SSE format
- Fixes "not an AssistantContent" error in HA

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

…-agent

- Log upstream error body from Codex endpoint
- ha-agent combo: oc/auto first (working), cx/auto as backup

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>

Codex always returns SSE stream even when we want non-streaming.
Now collects stream chunks into a single response dict.

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

…sing

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>

Codex rejects these params. Like 9router Codex executor transformRequest.

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

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

…erateContent

…or combo models

… error

…fied multimodal models

…d tints

Apply card-3d + card-tint-* styling consistently across all pages:
- Dashboard, Models, Accounts, Providers, Combos, Search, Backup
- Settings (11 components), Video, Video Manager
- Changed border from invisible white to visible thin dark line
- Removed bg-white overrides in favor of pastel tinted backgrounds

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

Changed from near-white rgba(0.95 opacity) to solid Tailwind 100→200
level colors that are clearly distinguishable from page background.

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

Each card within the same page now has a different tint color:
- Models: providers mapped to emerald/amber/violet/indigo
- Providers: each provider gets a unique tint
- Combos: cycle through 6 tints for combo cards
- Settings: 11 components each assigned a distinct tint
- Search: indigo/violet/sky for 3 config cards
- Backup: amber import + cycling row tints
- Video: indigo/violet, Manager: sky/rose
- Accounts: emerald tint

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

- Filter bar moved above stat cards, centered layout
- Stat cards now use card-3d for consistent thin dark borders
- New tree view mode groups accounts by provider type (pro/codex/free/...)
- Each group shows expandable header with status summary (active/limited/error counts)
- Toggle between List view and Tree view

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

- New .card-main class: white gradient + warm shadow for largest cards in 2-3 card layouts
- Refined .card-tint-* gradients: aurora-colored light cast effect ("bóng")
- Softer shadows: multi-layer subtle depth replacing harsh shadows
- Thinner border: rgba(0,0,0,0.08) for cleaner look
- Applied largest-card-white rule: accounts table, search combo, combos add-new, settings config
- Dashboard status bar uses card-main for clean footer

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

- Removed standalone filter section above stat cards
- Search + Type + Status filters now sit right below action buttons row
- Fixed Status filter: labels now use i18n translation instead of empty label prop
- More compact layout: filters h-9, smaller inputs

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

Backend:
- _check_gemini_status now scans ALL custom providers, not just "geminiapi"
- Returns per-instance status with port, models, clients, error details
- New /api/v1/provider-tree endpoint: aggregates accounts + providers + custom APIs
  into hierarchical tree structure (ChatGPT types, Provider APIs, Custom instances)

Frontend:
- Dashboard: stat cards split into "Gemini API" + "Custom APIs" count
- Dashboard: status bar dynamically renders all custom provider instances
- Accounts: removed list view, tree view is now the only mode
- Accounts: 3-level tree — Provider category → Sub-type/API → Account detail
- Level 1: ChatGPT (grouped by type: free/pro/codex), Providers, Custom APIs
- Level 2: account types or individual API instances with status indicators
- Level 3: click account to expand quota + request stats panel
- Each custom provider instance shows live health check status

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

When backend returns old format without "instances" array, the expression
instances?.filter(...).length throws TypeError because ?.filter() returns
undefined on undefined, and .length on undefined crashes. Added missing ?.

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

- Reduced card padding p-4→p-3, icon circle size-9→size-7
- Smaller fonts: label 11px→10px, value text-2xl→text-xl
- Grid: 2 cols mobile, 3 tablet, 5 desktop (was 6 but too cramped)
- Tighter gap: gap-3→gap-2

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

- mergedTree combines groupedAccounts (from filteredAccounts) + providerTree
- ChatGPT groups now built from filteredAccounts directly (supports search/filter)
- Providers/Custom APIs branches still come from /api/v1/provider-tree
- Fixed property mismatch: items vs accounts, added key/label/count fields
- Accounts now visible even if provider-tree API fails

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

…PI fails

- fetchProviderTree now has fallback: if /api/v1/provider-tree fails,
  builds tree from /api/v1/health (gemini instances) + /api/v1/custom-providers
- Providers branch: shows Gemini API status from health check
- Custom APIs branch: shows all custom provider instances with status
- Ensures all branches visible even without updated backend

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

Fallback now handles both old and new health API formats:
- New: health.gemini.instances array
- Old: health.gemini.geminiapi string + geminiapi_port/clients/entries
- Always merges custom-providers list from config
- Uses seenIds to avoid duplicates between formats
- Shows Gemini even with "no_key" status

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

- Grid: lg:grid-cols-6 for 6 cards on one row
- Reduced padding p-3→p-2.5, icon size-7→size-6, icon 3.5→3
- Value font text-xl→text-lg, gap 2→1.5
- Removed shadow-sm from icon circle

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

Row 1: "Tài khoản" title + count badge
Row 2: Import, Custom API, Export Token buttons
Row 3: Refresh, Refresh All, Search, Type filter, Status filter
Removed old subtitle, cleaned up duplicate title in tree section

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

…hboard)

- fetchProviderTree now directly uses /api/v1/health + /api/v1/custom-providers
- Same data sources as dashboard status bar (health.gemini.instances)
- No dependency on /api/v1/provider-tree backend endpoint
- Handles both new format (instances array) and old format (geminiapi field)
- Always shows Providers + Custom APIs branches when data available

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

Simplified buildProviderTree to fetch from:
- /api/v1/custom-providers (all custom provider instances)
- /api/v1/providers (built-in providers like opencode, gemini_free, etc.)
Added console.error for debugging issues

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