Route all agents through Gemini via LiteLLM proxy

[farzanshibu]

Problem

All agents (dispatcher + execution) were hardcoded to Claude models via the Anthropic API. Running Gemini was not supported, and a LiteLLM proxy setup was partially wired but broken in several ways:

• scripts/start-proxy.sh failed to load .env.local (inline comments broke xargs)
• ANTHROPIC_AUTH_TOKEN was used instead of ANTHROPIC_API_KEY — SDK ignored it
• LiteLLM had no model aliases matching the model names Claude Code CLI sends internally (claude-sonnet-4-6, claude-haiku-4-5-20251001, etc.) → 400 errors
• LiteLLM's /v1/messages/count_tokens endpoint sends empty body to Gemini's countTokens API → 500 errors flooding logs
• When the dispatcher model produced no text blocks (tool-only responses), the literal string "(no reply)" was sent to the user's iMessage

Solution (this PR)

See PR #X for implementation. This issue tracks the root causes.

Checklist

• .env.local loading fixed (strip inline comments before xargs)
• ANTHROPIC_API_KEY set correctly for LiteLLM auth
• litellm.config.yaml maps all Claude Code internal model IDs to gemini/gemini-2.5-flash
• Thin proxy on port 4000 intercepts count_tokens → returns mock 200
• LiteLLM runs on port 4001 behind the proxy
• "(no reply)" fallback replaced with a real user-facing fallback message
• System prompt updated to allow self-description questions (e.g. "which model are you")