{"repo": "agentcyberrange/cage", "files": {"config/models.example.yml": {"bf246dda34ef8ffe5481aad66bd90b489377af5d": {"t": 1786068431.711207, "n": 10960, "h": "0892644d64a73f0b", "c": "models:\n  # ======================================================================\n  #  CAGE model registry — example / template\n  # ======================================================================\n  #  Copy this file to `config/models.yml`, then fill in your endpoints:\n  #\n  #      cp config/models.example.yml config/models.yml\n  #\n  #  `config/models.yml` is gitignored — keep real keys there, never here.\n  #  Secrets are written as `${ENV_VAR}` refs and expanded at load time, so\n  #  the literal key never lives in any file on disk.\n  #\n  #  Each top-level key below is a model *id* — the name you pass to\n  #  `--model <id>` or list under `agents:` in project.yml. An id maps to ONE\n  #  upstream endpoint; CAGE routes every LLM call through its in-container\n  #  proxy to that endpoint, translating the wire protocol when the agent and\n  #  the endpoint disagree.\n  #\n  # ----------------------------------------------------------------------\n  #  FIELDS\n  # ----------------------------------------------------------------------\n  #  Required\n  #    provider   Wire-protocol family. Decides how CAGE talks to the endpoint:\n  #                 openai | vllm | sglang   -> OpenAI Chat Completions protocol\n  #                 gemini | google          -> Google Generative Language API\n  #                                             (generateContent; for the\n  #                                             gemini_cli agent)\n  #                 anthropic (or any other) -> Anthropic Messages protocol\n  #               `vllm`/`sglang` additionally mark the endpoint as *self-hosted*\n  #               — the only kind eligible for `--wait-for-model` startup polling\n  #               (SaaS providers are assumed always up).\n  #    model      The model string sent upstream, i.e. the provider's own name.\n  #\n  #  Auth — pick the ONE style that fits the endpoint\n  #    api_key    A single key, normally an env ref e.g. ${OPENAI_API_KEY}.\n  #               This is the common case.\n  #    api_keys   A *pool* of keys for the same endpoint — several accounts /\n  #               quotas. When non-empty it OVERRIDES api_key, and CAGE rotates\n  #               across the pool to load-balance concurrent trials. Use this\n  #               INSTEAD of api_key only when you want to spread load over\n  #               multiple accounts. (Internally: pool = api_keys if set, else\n  #               [api_key].)\n  #    auth_source  ONLY for subscription / OAuth agents logged in with an\n  #               account instead of an API key: Claude Code (Max/Pro) or Codex\n  #               (\"Sign in with ChatGPT\"). Points at a host directory holding\n  #               the agent's OAuth credential file (`.credentials.json` for\n  #               Claude Code, `.codex/auth.json` for Codex), which CAGE\n  #               bind-mounts into the agent container. In this mode base_url and\n  #               api_key are left empty. Almost no endpoint needs this — leave\n  #               it out entirely unless you actually run subscription mode.\n  #\n  #  Optional (have sensible defaults — omit unless you need to override)\n  #    base_url            Endpoint URL. Required for normal API / self-hosted\n  #                        endpoints; empty only in subscription (auth_source) mode.\n  #    agent_model_names   Per-agent override of `model`, for when a specific\n  #                        agent CLI expects a different launch name, e.g.\n  #                          agent_model_names: {claude_code: claude-opus-4-7[1m]}\n  #    timeout             Per-request timeout in seconds. Default: 360.\n  #    max_retries         Upstream retry attempts. Default: 2.\n  #    input_cost_per_1m   USD per 1M input tokens — for cost accounting only.\n  #    output_cost_per_1m  USD per 1M output tokens.  (Both unset => not tracked.)\n  #    extra_headers       Extra HTTP headers sent upstream, e.g. {X-Title: cage}.\n  #    max_context_size    The endpoint's REAL context window, in tokens. A model\n  #                        capability (not a per-run knob), so it lives here.\n  #                        Agents whose CLI can be told a window honour it —\n  #                        kimi_code and qwen_code (alias: context_window_size).\n  #                        Omit it and the agent CLI keeps its OWN default; CAGE\n  #                        never invents a number. NOTE: the Claude Code CLI has\n  #                        no knob for a custom window (it infers ~200k from the\n  #                        model id), so claude_code CANNOT honour this — a\n  #                        long-context model still caps ~200k under claude_code.\n  #    reserved_context_size  Tokens of output headroom an agent keeps free below\n  #                        max_context_size before it auto-compacts. kimi_code only.\n  #    context_window_size  Accepted alias for max_context_size (qwen naming).\n  #    rl_reward_sink      RL training integration. Set this to the URL of an\n  #                        external trainer's reward sink to put the model in\n  #                        \"RL mode\": every LLM call it drives carries an\n  #                        `X-Trial-Id` header (so the trainer can group one\n  #                        trajectory's calls), and each finished trial's reward\n  #                        is POSTed here as JSON `{trial_id, reward, success,\n  #                        sample_id, metrics}`. The reward number comes from the\n  #                        benchmark (`Benchmark.reward`, default = its score).\n  #                        Omit ⇒ an ordinary model, behaviour identical. This is\n  #                        the ONLY switch — the whole feature toggles from here.\n  # ======================================================================\n\n  # --- OpenAI-protocol SaaS, single key --------------------------------\n  gpt-5.5:\n    provider: openai\n    model: gpt-5.5\n    base_url: https://api.openai.com/v1\n    api_key: ${OPENAI_API_KEY}\n    timeout: 3600\n    # Cost accounting is optional; set real numbers to track $ per run:\n    # input_cost_per_1m: 1.25\n    # output_cost_per_1m: 10.0\n\n  # --- Multi-account key POOL (use api_keys, not api_key) ---------------\n  #     CAGE rotates trials across these keys to spread quota/load.\n  openai-example:\n    provider: openai\n    model: gpt-5.5\n    base_url: https://api.openai.com/v1\n    api_keys:\n      - ${OPENAI_API_KEY}\n      - ${OPENAI_API_KEY_2}\n\n  # --- Self-hosted vLLM (OpenAI protocol, --wait-for-model eligible) ----\n  vllm-example:\n    provider: vllm\n    model: local-model\n    base_url: http://127.0.0.1:8000/v1\n    api_key: ${VLLM_API_KEY}   # self-hosted servers often accept any/empty key\n    # Declare the model's real context window so kimi_code / qwen_code size their\n    # auto-compaction to it instead of a CLI default. Omit ⇒ the CLI default\n    # stands; CAGE will not invent one. (claude_code can't use this — see the\n    # FIELDS note above: the Claude Code CLI has no custom-window knob.)\n    # max_context_size: 262144\n    # reserved_context_size: 20000      # output headroom kept free (kimi_code)\n\n  # --- RL training policy (verl etc.) — OpenAI endpoint + reward sink ----\n  #     The trainer serves the in-training policy as an OpenAI-compatible\n  #     endpoint; CAGE drives it like any OpenAI model. Adding `rl_reward_sink`\n  #     is the whole switch: trial ids ride on every call and rewards are\n  #     reported back. Drop that one line and this is a plain model again.\n  rl-policy:\n    provider: openai\n    model: rl-policy\n    base_url: ${RL_ENDPOINT}          # trainer's OpenAI-compatible policy server\n    api_key: dummy                    # trainer usually ignores the key\n    rl_reward_sink: ${RL_REWARD_SINK}  # where per-trial rewards are POSTed\n\n  # --- Anthropic SaaS, single key --------------------------------------\n  claude-opus:\n    provider: anthropic\n    model: claude-opus-4-7\n    base_url: https://api.anthropic.com\n    api_key: ${ANTHROPIC_API_KEY}\n    timeout: 3600\n\n  # --- Subscription / OAuth (auth_source) — Claude Code & Codex ---------\n  #     No base_url / api_key: the agent logs in with a subscription account\n  #     and CAGE bind-mounts the host credentials into the container, keeping\n  #     the on-disk OAuth token fresh via a host-side refresher for the run.\n  claude-opus-sub:\n    provider: anthropic\n    model: claude-opus-4-7\n    agent_model_names:\n      claude_code: claude-opus-4-7[1m]   # decorated launch name (e.g. [1m] = 1M context)\n    auth_source: ${CAGE_CLAUDE_AUTH_SOURCE}   # host dir with .credentials.json\n\n  # --- Codex \"Sign in with ChatGPT\" (OAuth). Run `codex login` on the host\n  #     first so ~/.codex/auth.json holds the ChatGPT OAuth tokens; point\n  #     auth_source at that dir (or the file). No base_url / api_key — the proxy\n  #     forwards the OAuth Bearer to the ChatGPT Codex backend. Use with\n  #     `--agent codex`.\n  codex-chatgpt:\n    provider: openai\n    model: gpt-5-codex\n    auth_source: ${CAGE_CODEX_AUTH_SOURCE}   # host dir with .codex/auth.json (e.g. ~/.codex)\n\n  # --- Anthropic-compatible 3rd party (custom base_url + agent name) ----\n  deepseek-v4-pro:\n    provider: anthropic\n    model: deepseek-v4-pro\n    agent_model_names:\n      claude_code: deepseek-v4-pro[1m]\n    base_url: https://api.deepseek.com/anthropic\n    api_key: ${DEEPSEEK_API_KEY}\n    timeout: 3600\n\n  # --- Self-hosted OpenAI-compatible server -----------------------------\n  local-openai-compatible:\n    provider: openai\n    model: local-model\n    base_url: http://127.0.0.1:8000/v1\n    api_key: ${LOCAL_MODEL_API_KEY}\n    timeout: 3600\n\n  glm-5.1:\n    provider: openai\n    model: GLM-5.1\n    base_url: https://open.bigmodel.cn/api/paas/v4\n    api_key: ${GLM_API_KEY}\n    timeout: 3600\n\n  qwen3.5-397b-w8a8:\n    provider: vllm\n    model: qwen3.5-397b-w8a8\n    base_url: http://127.0.0.1:8000/v1\n    api_key: ${SII_API_KEY}\n    timeout: 3600\n\n  qwen3.6-max-preview:\n    provider: openai\n    model: qwen3.6-max-preview\n    base_url: https://dashscope.aliyuncs.com/compatible-mode/v1\n    api_key: ${QWEN_API_KEY}\n    timeout: 3600\n\n  qwen3.7-max:\n    provider: openai\n    model: qwen3.7-max\n    base_url: https://dashscope.aliyuncs.com/compatible-mode/v1\n    api_key: ${QWEN_API_KEY}\n    timeout: 3600\n\n  kimi-k2.6-cli:\n    provider: openai\n    model: kimi-k2.6\n    base_url: https://api.moonshot.cn/v1\n    api_key: ${KIMI_API_KEY}\n    timeout: 3600\n\n  # Google Gemini, driven by the `gemini_cli` agent. `provider: gemini` routes\n  # the agent's native generateContent traffic through the proxy to Google\n  # (the proxy forwards byte-for-byte and records an OpenAI-shaped projection\n  # so the inspector parses it). `base_url` is the API origin only — gemini-cli\n  # appends the `/v1beta/...` path itself. The key authenticates via the\n  # `x-goog-api-key` header.\n  gemini-2.5-pro:\n    provider: gemini\n    model: gemini-2.5-pro\n    base_url: https://generativelanguage.googleapis.com\n    api_key: ${GEMINI_API_KEY}\n    timeout: 3600\n"}}}}