Requesty provider - #84

Summary

Adds Requesty (an OpenAI compatible router) as a tenth provider: run_agent(..., provider="requesty", model="claude-sonnet-5") with REQUESTY_API_KEY from .env or the app's Settings. It follows the Cerebras and Together pattern: chat completions with native tool calling and image input, and the web tool goes through the headless browser agent.

Changes

• AutoCua/llm_provider/requesty/ (service.py, view.py) with three entries: claude-sonnet-5, gemini-3.6-flash, gpt-5.4. Any other vendor/model id is sent as is.
• Registered in llm_manager.py and in the TLS probe hosts. Not added to _CLI_FALLBACK_CANDIDATES, so coder and minions have no second model on this provider.
• Key maps for the browser agent and Telegram, the browser fallback sets for the web tool (Python on all four platforms plus the Android Rust side).
• Settings key row (index.tsx and the matching entry in the built index.js), script.js, ui/service.py provider list and key map.
• README, agent_capabilities.md, model_list.txt, .env.example, requirements_linux.txt comment.

Validation

• py_compile on every touched Python file, ruff check --select E4,E7,E9,F: new files clean, touched files report the same findings as before the change.
• cargo check in AutoCua/android passes.
• Live calls through the linux LLMManager (mode="text", then mode="minion" with a shell call, its tool result and an exit call) on openai/gpt-4o-mini and on all three table entries. All returned the expected text and tool calls.

Disclosure: I work at Requesty. Happy to adjust anything to match project conventions.