Auto-detect context length from /models (cloud) or model config (local), fall back to MAX_CONTEXT_TOKENS

Add proactive + reactive context-length compression for long ReAct runs

Disable SSR mode - its self-reachability check was crashing launch() intermittently

Drop self-hosted SearXNG (HF abuse-flagged); use Jina search_web exclusively

Revert to Gradio SDK; use standalone SearXNG space; add explicit research-goal tracking with vagueness-aware early stopping

Add rustc/cargo and upgrade pip so tokenizers builds cleanly

Switch to Docker SDK: self-host SearXNG sidecar as primary search, Jina as fallback

Use Jina search_web exclusively for search; drop SearXNG (unreliable, bot-blocked)

Switch default SearXNG instance to search.mdosch.de

Treat 429 rate-limit as retryable in cloud backup call

Add Jina MCP read_url as primary scrape backend; retry-with-backoff for transient/proxy errors

Add Inference Mode switch (Auto/Local only/Cloud only)

Local ZeroGPU inference primary, OpenAI-compatible cloud backup; SearXNG-primary search with Serper backup

initial commit