gateway.yaml

# gateway.yaml
#
# Three ideas:
#   models:    decide WHAT exists, once. Each entry is a model (or an ordered
#              fallback chain of models) with the name you want to use.
#   clients:   decide WHICH model serves which traffic per client.
#   providers: only needed for endpoints the built-in presets don't know.
#
# The name you give a model is the id you use everywhere:
# `claude --model <name>` and Codex's `model = "<name>"`.
#
# Secrets never live in this file: provider keys come from env vars
# (auto-loaded from .env.local / .env), and any value here may reference an
# env var with ${VAR} syntax.

server:
  bind: "127.0.0.1:4000"        # default; non-localhost requires a token
  token: local-dev-token        # accepted as Bearer or x-api-key
  # tokens: [token-a, token-b]  # multiple tokens
  # token: "${GATEWAY_TOKEN}"   # or from the environment
  # Env overrides (no config edit needed): GATEWAY_BIND, GATEWAY_TOKEN

# Providers
#
# Built-in presets imply base URL + key env var — use them directly in model
# strings with zero declaration here:
#
#   openrouter (OPENROUTER_API_KEY)   openai    (OPENAI_API_KEY)
#   anthropic  (ANTHROPIC_API_KEY)    moonshot  (MOONSHOT_API_KEY)
#   fireworks  (FIREWORKS_API_KEY)    together  (TOGETHER_API_KEY)
#   groq       (GROQ_API_KEY)         deepinfra (DEEPINFRA_API_KEY)
#   deepseek   (DEEPSEEK_API_KEY)     mistral   (MISTRAL_API_KEY)
#   xai        (XAI_API_KEY)          cerebras  (CEREBRAS_API_KEY)
#   ollama     (localhost:11434, no key)
#   azure      (AZURE_AI_API_KEY, needs `resource:` below)
#
# Declare a provider only for custom endpoints or to override a preset.
providers:
  azure:
    resource: "${AZURE_AI_RESOURCE}"
  vllm-local:
    base_url: "http://localhost:8000/v1"
    api_key: dummy
  modal-kimi:
    type: openai
    base_url: "https://superagentai--huihui-kimi-k2-7-code-inference-server.modal.run/v1"
    api_key: unused                   # Modal auth is via headers, not bearer
    headers:                          # extra headers sent on every request
      Modal-Key: "${MODAL_KEY}"
      Modal-Secret: "${MODAL_SECRET}"
  # my-anthropic-proxy:
  #   type: anthropic                 # anthropic | openai (default: openai)
  #   base_url: "https://proxy.example.com"
  #   api_key_env: PROXY_API_KEY

# Models
#
# Three forms:
#   name: provider/model-id                single model
#   name: [provider/a, provider/b]         fallback chain (see below)
#   name: { model: ..., <overrides> }      long form, see qwen-local
#                                          (model: also accepts a list —
#                                          a chain with display name/overrides)
#
# Fallback: a list IS the fallback definition. The first entry is primary;
# on connection failure, 429, or 5xx the next entry is tried, in order.
# Fallback only happens before the response starts streaming — a model that
# has begun answering (or calling a tool) is never silently swapped.
# The `kimi` and `glm` models below define opposite fallback chains.
#
# Known model families (kimi, gpt, claude, ...) get capabilities from a
# built-in quirk table — e.g. Kimi automatically drops temperature/top_p
# (fixed by the API), keeps reasoning across tool turns, accepts images.
models:
  abliterated:
    model: modal-kimi/huihui-ai/Huihui-Kimi-K2.7-Code-abliterated-GGUF
    display_name: "Huihui Kimi K2.7 Code Abliterated (Modal)"
    # kimi quirks apply automatically: native thinking, temperature/top_p
    # dropped, 300s timeout
  sol:
    model: [azure/gpt-5.6-sol, openrouter/moonshotai/kimi-k2.7-code]
    display_name: "Azure GPT-5.6 Sol (fallback to OpenRouter Kimi K2.7 Code)"
  fw-kimi-k3:
    model: [azure/FW-Kimi-K3, azure/gpt-5.6-sol]
    display_name: "Azure FW Kimi K3 (fallback to Azure GPT-5.6 Sol)"
    drop_message_params: [reasoning] # Azure accepts reasoning_content only
  kimi-k3:
    model: [openrouter/moonshotai/kimi-k3, openrouter/moonshotai/kimi-k2.7-code]
    display_name: "OpenRouter Kimi K3 Code"
  kimi-27:
    model: openrouter/moonshotai/kimi-k2.7-code
    display_name: "OpenRouter Kimi K2.7 Code"
  kimi:
    model: [azure/kimi-k2.6, openrouter/z-ai/glm-5.2]
    display_name: "Azure Kimi K2.6 (fallback to OpenRouter GLM 5.2)"
  glm:
    model: [openrouter/z-ai/glm-5.2, azure/kimi-k2.6]
    display_name: "OpenRouter GLM 5.2 (fallback to Azure Kimi K2.6)"
  glm-53-flash:
    model: [openrouter/z-ai/glm-5.3-flash, azure/gpt-5.6-sol]
    display_name: "OpenRouter GLM 5.3 Flash (fallback to Azure GPT-5.6 Sol)"
    images: true
  qwen-local:
    model: vllm-local/qwen3-coder
    # model: [vllm-local/qwen3-coder, openrouter/qwen/qwen3-coder]  # with fallback
    display_name: "Qwen3 Coder (local)"
    # images: false                   # vision on/off
    # tools: openai                   # openai | native | none
    # thinking: none                  # native | none (reasoning round-trip)
    # cache_control: none             # native | none
    # drop_params: [temperature]      # strip params the API rejects
    # drop_message_params: [reasoning] # strip fields from chat messages
    # timeout_ms: 120000
    # expose: [claude_code, codex]    # which clients may select it (default: all)

# Clients
#
# Roles route traffic to models by name (or inline provider/model-id string):
#   main:        the conversation model; a list is a fallback chain
#   subagent:    Task-tool subagent requests            (default: unused)
#   background:  cheap lane: titles, summaries          (default: unused)
#   unknown:     where unrecognized model ids go — a role, a model, or
#                "reject" to return 404                 (default: main)
#
# Roles may reference other roles (e.g. `subagent: main`).
clients:
  claude_code:
    main: sol
    subagent: kimi-k3
    background: kimi-k3
  codex:
    main: kimi

# Retry tuning (defaults shown)
#
# WHICH models fall back to which is defined above in models: (a list is a
# fallback chain). This section only tunes WHEN a failure counts as
# retryable and how many attempts are made in total (the first entry is
# retried once before moving down the chain).
# fallback:
#   max_attempts: 3
#   retryable_statuses: [408, 429, 500, 502, 503, 504]

# Telemetry (defaults shown)
#
# telemetry:
#   log_prompts: false                # true logs full request bodies
#   log_tool_calls: true
#   redact_headers: [authorization, x-api-key]