feat: add transport layer to provider hierarchy

Иерархия: transport → provider → model → profile. Поле `transport`: direct-api — прямой HTTP к провайдеру aws-bedrock — AWS Bedrock (IAM/role + region) azure-openai — Azure OpenAI (deployment + key) google-vertex — GCP Vertex AI (service-account JSON) local — Ollama / MLX / LM Studio proxy — LiteLLM / OpenRouter subscription — ChatGPT OAuth (Codex) Новые провайдеры: - anthropic-bedrock — Claude через AWS - openai-azure — GPT через Azure - google-vertex — Gemini через Vertex - openrouter — multi-provider proxy Существующие провайдеры получили transport-тег. Используется install/lib-onboarding.sh для интерактивного мастера выбора language → transport → provider → model → credentials. 14 провайдеров суммарно: 5 direct-api + 1 aws + 1 azure + 1 vertex + 3 local + 2 proxy + 1 subscription.
2026-05-17 02:24:19 +08:00 · 2026-05-17 02:24:19 +08:00 · afe0c6f118
commit afe0c6f118
parent 7aaa6a79b0
1 changed files with 79 additions and 6 deletions
--- a/providers.toml
+++ b/providers.toml
@ -1,16 +1,28 @@
 # Реестр провайдеров LLM. SSoT для kei-model-router + keisei-marketplace.
 #
-# Слой 1 из трёх: provider → model → profile (agent-profiles.toml).
-# Провайдер стабилен (endpoint, схема аутентификации, биллинг-аккаунт),
+# Иерархия: transport → provider → model → profile (agent-profiles.toml).
+# Транспорт стабилен (direct-api / aws-bedrock / azure-openai / google-vertex / local),
+# провайдер стабилен (endpoint, схема аутентификации, биллинг-аккаунт),
 # модели за ним меняются — см. models.toml.
 #
 # Все цены/контекст — в models.toml. Здесь только транспорт + лимиты.
 #
-# Конвенция: `id` — kebab-case, без версий. Версии — в моделях.
+# Поле `transport` задаёт способ доступа:
+#   - direct-api      — прямой HTTP к провайдеру (Anthropic, OpenAI, xAI, etc.)
+#   - aws-bedrock     — через AWS Bedrock (IAM/role, регион)
+#   - azure-openai    — через Azure (deployment + Azure key)
+#   - google-vertex   — через GCP Vertex AI (service-account JSON + project/region)
+#   - local           — локальная LLM (Ollama, MLX, LM Studio)
+#   - proxy           — собственный или сторонний прокси (LiteLLM, OpenRouter)
+#   - subscription    — OAuth-привязка через подписку (ChatGPT, Claude.ai)
+#
+# Конвенция: `id` — kebab-case с суффиксом транспорта если не direct-api
+# (anthropic, anthropic-bedrock; openai, openai-azure; google, google-vertex).

 [[provider]]
 id = "anthropic"
-display_name = "Anthropic"
+display_name = "Anthropic (Direct API)"
+transport = "direct-api"
 endpoint = "https://api.anthropic.com/v1/messages"
 auth_scheme = "x-api-key"          # header name; value from env ANTHROPIC_API_KEY
 auth_env = "ANTHROPIC_API_KEY"
@ -22,9 +34,23 @@ rate_limit_rpm = 4000
 billing_currency = "USD"
 notes = "Primary provider for Claude family. RULE 0.20 default for code-implementer-*."

+[[provider]]
+id = "anthropic-bedrock"
+display_name = "Anthropic (AWS Bedrock)"
+transport = "aws-bedrock"
+endpoint = "https://bedrock-runtime.{region}.amazonaws.com"
+auth_scheme = "aws-sigv4"
+auth_env = "AWS_ACCESS_KEY_ID,AWS_SECRET_ACCESS_KEY,AWS_REGION"
+retry_max = 3
+retry_backoff_ms = 500
+rate_limit_rpm = 0                 # per-region quota; see AWS console
+billing_currency = "USD-aws"
+notes = "Claude через AWS Bedrock — для enterprise с AWS-биллингом. Регион в AWS_REGION."
+
 [[provider]]
 id = "openai"
-display_name = "OpenAI"
+display_name = "OpenAI (Direct API)"
+transport = "direct-api"
 endpoint = "https://api.openai.com/v1/chat/completions"
 auth_scheme = "bearer"
 auth_env = "OPENAI_API_KEY"
@ -34,9 +60,23 @@ rate_limit_rpm = 5000
 billing_currency = "USD"
 notes = "GPT family + codex (ChatGPT OAuth, see codex provider for dual-review)."

+[[provider]]
+id = "openai-azure"
+display_name = "OpenAI (Azure)"
+transport = "azure-openai"
+endpoint = "https://{azure_resource}.openai.azure.com/openai/deployments/{deployment}"
+auth_scheme = "api-key"
+auth_env = "AZURE_OPENAI_API_KEY,AZURE_OPENAI_ENDPOINT,AZURE_OPENAI_DEPLOYMENT"
+retry_max = 3
+retry_backoff_ms = 500
+rate_limit_rpm = 0                 # per-deployment quota; see Azure portal
+billing_currency = "USD-azure"
+notes = "OpenAI через Azure — для enterprise с Azure-биллингом. Нужны resource+deployment."
+
 [[provider]]
 id = "codex"
 display_name = "OpenAI Codex (ChatGPT OAuth)"
+transport = "subscription"
 endpoint = "local:codex-cli"       # invoked via `codex` CLI, not HTTP
 auth_scheme = "oauth-subscription"
 auth_env = "_"                      # no env — CLI handles auth
@ -49,6 +89,7 @@ notes = "RULE 0.23 DUAL REVIEW. Separate billing from API providers."
 [[provider]]
 id = "xai"
 display_name = "xAI"
+transport = "direct-api"
 endpoint = "https://api.x.ai/v1/chat/completions"
 auth_scheme = "bearer"
 auth_env = "XAI_API_KEY"
@ -61,6 +102,7 @@ notes = "Grok family. OpenAI-compatible API."
 [[provider]]
 id = "deepseek"
 display_name = "DeepSeek"
+transport = "direct-api"
 endpoint = "https://api.deepseek.com/v1/chat/completions"
 auth_scheme = "bearer"
 auth_env = "DEEPSEEK_API_KEY"
@ -72,7 +114,8 @@ notes = "Cheapest tier for batch reasoning. R1 reasoner family."

 [[provider]]
 id = "google"
-display_name = "Google Gemini"
+display_name = "Google Gemini (Direct API)"
+transport = "direct-api"
 endpoint = "https://generativelanguage.googleapis.com/v1beta/models"
 auth_scheme = "query-key"          # ?key=<...> in URL
 auth_env = "GEMINI_API_KEY"
@ -82,9 +125,23 @@ rate_limit_rpm = 1000
 billing_currency = "USD"
 notes = "1M-context window for long-doc analysis."

+[[provider]]
+id = "google-vertex"
+display_name = "Google Gemini (Vertex AI)"
+transport = "google-vertex"
+endpoint = "https://{region}-aiplatform.googleapis.com/v1/projects/{project}/locations/{region}/publishers/google/models"
+auth_scheme = "gcp-service-account"
+auth_env = "GOOGLE_APPLICATION_CREDENTIALS,GCP_PROJECT_ID,GCP_REGION"
+retry_max = 3
+retry_backoff_ms = 500
+rate_limit_rpm = 0                 # per-project quota; see GCP console
+billing_currency = "USD-gcp"
+notes = "Gemini через Vertex AI — для enterprise с GCP-биллингом. Нужен service-account JSON."
+
 [[provider]]
 id = "ollama-local"
 display_name = "Ollama (local)"
+transport = "local"
 endpoint = "http://127.0.0.1:11434/api/chat"
 auth_scheme = "none"
 auth_env = "_"
@ -97,6 +154,7 @@ notes = "Self-hosted. Models loaded via `ollama pull`."
 [[provider]]
 id = "mlx-local"
 display_name = "MLX (Apple silicon local)"
+transport = "local"
 endpoint = "http://127.0.0.1:8080/v1/chat/completions"
 auth_scheme = "none"
 auth_env = "_"
@ -109,6 +167,7 @@ notes = "Native Apple-silicon inference. mlx_lm.server."
 [[provider]]
 id = "lmstudio-local"
 display_name = "LM Studio (local)"
+transport = "local"
 endpoint = "http://127.0.0.1:1234/v1/chat/completions"
 auth_scheme = "none"
 auth_env = "_"
@ -121,6 +180,7 @@ notes = "Desktop GUI runner. OpenAI-compatible."
 [[provider]]
 id = "litellm-proxy"
 display_name = "LiteLLM proxy (keisei.app)"
+transport = "proxy"
 endpoint = "https://api.keisei.app/llm/v1/chat/completions"
 auth_scheme = "bearer"
 auth_env = "KEI_LITELLM_KEY"
@ -129,3 +189,16 @@ retry_backoff_ms = 500
 rate_limit_rpm = 1000
 billing_currency = "USD-keisei"
 notes = "Per-user-keyed proxy. Routes to anthropic/openai/xai/deepseek under one bearer."
+
+[[provider]]
+id = "openrouter"
+display_name = "OpenRouter (multi-provider proxy)"
+transport = "proxy"
+endpoint = "https://openrouter.ai/api/v1/chat/completions"
+auth_scheme = "bearer"
+auth_env = "OPENROUTER_API_KEY"
+retry_max = 3
+retry_backoff_ms = 500
+rate_limit_rpm = 1000
+billing_currency = "USD"
+notes = "Меняет провайдеров по pricing. Подходит для эксплоративных задач."