- codeclash/agents/mini_anthropic_model.py: mini-swe-agent model class driving Claude through a
  custom Anthropic-compatible base URL (native Anthropic API, so prompt caching is preserved)
  with explicit per-token cost tracking. API key via api_key_env; base URL / model optionally
  from env (base_url_env / model_name_env).
- pyproject.toml: optional 'anthropic' dep as a [llama] extra (locked in uv.lock).
- configs/mini/models/llama_*.yaml: reusable model blocks for the 5 gateway models
  (Opus 4.8, Opus 4.7, Sonnet 5 via AnthropicModel; Gemini 3.5 Flash via litellm OpenAI-compat;
  GPT-5.5 via litellm Responses API).
- configs/ablations/ladder/rungs/{robotrumble,corewar}.yaml: rung lists factored out for reuse.
- configs/ablations/ladder/{robotrumble,corewar}__<model>.yaml: 10 ready-to-run ladder configs,
  one per (game, model) pair — 'uv run codeclash ladder run <config>'.
- configs/ablations/ladder/battlesnake_llama_smoke.yaml: cheap wiring smoke (Opus 4.8).
- .env.example documents the LLAMA_* vars. Only LLAMA_API_KEY is required at runtime; model
  names + endpoints are verbatim in the configs.

…nto john/ladder-run

- Claude gateway models (Opus 4.8/4.7, Sonnet 5) run with extended thinking enabled, which
  requires temperature unset (or 1.0). The model blocks set temperature: 0.2, causing every
  request to fail with HTTP 400. Remove model_kwargs/temperature and raise max_tokens
  4096 -> 16384. Verified against the anthropic-passthrough endpoint with the full agent
  request shape (system + tools + cache_control): now returns 200 / tool_use.
- Register litellm prices for gemini-3-5-flash-fair and gpt-5-5-genai-responses (from reveng's
  model_prices.json) so their cost is tracked instead of ignored. Claude self-computes cost
  from its cost: block (verified against reveng: opus/sonnet numbers match).
- Remove robotrumble_llama.yaml (stale env-indirection config reintroduced by a merge).