Status: ⬜ Not Started

File: lib/env.ts

• Update default POLLINATIONS_BASE_URL → https://gen.pollinations.ai
• Add POLLINATIONS_CHAT_PATH (default: /v1/chat/completions)
• Remove POLLINATIONS_MODEL env var (will be dynamic)
• Add optional POLLINATIONS_PRIMARY_MODEL for override

Status: ⬜ Not Started

File: lib/db.ts

• Add model_cache table for discovered models
• Add model_config table for selected primary/fallback models
• Add upsertModelCache(model) function
• Add getModelCache(type) function
• Add getModelConfig(key) / setModelConfig(key, value) functions

Schema:

CREATE TABLE IF NOT EXISTS model_cache (
  name TEXT PRIMARY KEY,
  type TEXT NOT NULL,           -- 'text' or 'image'
  aliases TEXT,                 -- JSON array
  description TEXT,
  tools INTEGER DEFAULT 0,      -- boolean
  reasoning INTEGER DEFAULT 0,  -- boolean
  context_window INTEGER,
  pricing_json TEXT,
  input_modalities TEXT,        -- JSON array
  output_modalities TEXT,       -- JSON array
  last_updated TEXT NOT NULL
);

CREATE TABLE IF NOT EXISTS model_config (
  key TEXT PRIMARY KEY,
  value TEXT NOT NULL,
  updated_at TEXT NOT NULL
);

Status: ⬜ Not Started

File: lib/llm_pollinations.ts

• Add discoverTextModels() - GET /v1/models + /text/models
• Add discoverImageModels() - GET /image/models
• Add refreshModelCache() - run discovery and cache to SQLite
• Add selectBestModel(options) - pick model based on needs
• Add getModelWithFallbacks() - return primary + fallback list

Model Priority:

1. gemini (cheap, tools=true) - primary
2. openai (baseline) - fallback 1
3. gemini-large (complex tasks) - fallback 2

Status: ⬜ Not Started

File: lib/llm_pollinations.ts

• Update generateChat() with model hints (useTools, complexTask)
• Add auto-retry with fallback models on error
• Add modelUsed and fallbacksAttempted to response
• Update generateJSON() to use new wrapper
• Update generateText() to use new wrapper

Status: ⬜ Not Started

File: lib/llm_pollinations.ts

• Add generateImage(prompt, options) function
• Download image and store in Val Town blob storage
• Add getImageModels() function

File: app/http.ts

• Add POST /api/image/generate endpoint
• Add GET /api/image/models endpoint

Status: ⬜ Not Started

File: jobs/sync/interval.ts

• Add model cache refresh (daily)
• Check last_discovery timestamp before refreshing

Status: ⬜ Not Started

File: lib/agent.ts

• Use complexTask: true for plan generation
• Handle content_blocks in Gemini responses
• Use cheaper model for simple triage suggestions

Status: ⬜ Not Started

File: app/http.ts

• Add GET /api/ai/models - list cached models + config
• Add POST /api/ai/discover - trigger model rediscovery

(none yet)

• Image storage: Using Val Town blob storage (option B)
• Model discovery: Once daily via interval job
• No account balance tracking (provider-agnostic design)
• Plan stored in grounding.md