Dynamic model fetching, pricing scraping, and context window metadata

[manmeet0409]

Summary

This PR makes the Command Code proxy self-maintaining: it discovers models and pricing dynamically, so changes upstream don't require code updates.

What's new

1. Dynamic model fetching — On startup and every 6 hours, the proxy fetches the live model catalog from https://api.commandcode.ai/provider/v1/models. Replaces the original 12-model hardcoded list with the real, up-to-date catalog.

2. Model probing — Each upstream model is probed with a minimal chat request to verify it actually exists. Models returning 403 ("not recognized") are filtered out. This catches stale entries that upstream doesn't actually serve.

3. Pricing/deal scraping — The public pricing page (https://commandcode.ai/docs/resources/pricing-limits) is scraped every 24 hours for active deals. Each model can carry a pricing field showing the deal multiplier, description, and status (permanent / expiration date).

4. Context window metadata — Each model reports its context_length (131K to 1M tokens) so clients like Hermes can determine compression thresholds.

5. Tomorrow-proof design:

   • Model list refreshes every 6 hours
   • Deals refresh every 24 hours
   • Lenient HTML parser with fallback warnings
   • Fuzzy matching handles name normalization (dots, dashes, prefixes)
   • Graceful fallback to static list when upstream is unreachable
   • Default context window (128K) for unknown models

Files changed

• internal/proxy/contextmap.go (new) — local override map for context windows
• internal/proxy/modelfetch.go (rewritten) — dynamic fetch + probe + pricing integration
• internal/proxy/pricing.go (new) — pricing page scraper + fuzzy model-deal matching
• internal/proxy/proxy.go — HandleModels now reads from cache with fallback
• internal/api/openai.go — OpenAIModel includes context_length and pricing fields
• internal/api/commandcode.go — added reasoning field for effort passthrough
• internal/proxy/model.go — added short aliases (sonnet, opus, haiku, glm-5.2, etc.)
• .gitignore — ignore local-only scripts with API tokens
• bin/command-code-proxy.exe — rebuilt Windows binary

Removed

• Static injection of taste-1 (CLI-only, not on Provider API)
• Static injection of MiniMax-M3-Promo (not served upstream)
• Hardcoded 12-model list (replaced by dynamic fetch)

Total models exposed

33 models matching the full Command Code catalog, with up to 6 active deals visible in /v1/models responses.

Example response

{
  "id": "deepseek/deepseek-v4-pro",
  "object": "model",
  "owned_by": "deepseek",
  "context_length": 1000000,
  "pricing": {
    "multiplier": "4x",
    "description": "4× usage. Every dollar of credit goes 4× further on this model.",
    "status": "permanent"
  }
}

Test plan

• /v1/models returns all models with context_length and pricing
• Probing filters non-existent models
• Pricing scrape finds all 6 current deals
• Fuzzy matching handles name variations (Qwen3.7-Max ↔ qwen-3.7-max)
• Graceful fallback when upstream is down
• No API tokens leaked in commit