@ailib-official/ai-protocol 0.8.4

package/v1/providers/openai.yaml

@@ -0,0 +1,224 @@
+$schema: "https://raw.githubusercontent.com/ailib-official/ai-protocol/main/schemas/v1.json"
+
+id: openai
+protocol_version: "1.5"
+
+name: OpenAI
+version: "v1"
+status: stable
+category: ai_provider
+official_url: "https://platform.openai.com/docs"
+support_contact: "https://help.openai.com"
+
+# Structured endpoint definition (v1.1+ extension)
+endpoint:
+  base_url: "https://api.openai.com/v1"
+  protocol: https
+  timeout_ms: 10000
+
+auth:
+  type: bearer
+  token_env: "OPENAI_API_KEY"
+payload_format: "openai_style"
+
+# Explicit API families to avoid mixing incompatible request/response models.
+api_families: ["chat_completions", "completions", "assistants", "realtime"]
+default_api_family: "chat_completions"
+endpoints:
+  chat:
+    path: "/chat/completions"
+    method: "POST"
+    adapter: "openai"
+  completions:
+    path: "/completions"
+    method: "POST"
+    adapter: "openai"
+  embeddings:
+    path: "/embeddings"
+    method: "POST"
+    adapter: "openai"
+  # Realtime (non-HTTP transport; provided as absolute URL for WS clients)
+  realtime_ws: "wss://api.openai.com/v1/realtime"
+
+# V1.5 Service Endpoints
+services:
+  list_models:
+    path: "/models"
+    method: "GET"
+    response_binding: "data"
+  list_files:
+    path: "/files"
+    method: "GET"
+    response_binding: "data"
+  create_batch:
+    path: "/batches"
+    method: "POST"
+
+# Termination reason normalization (see v1/spec.yaml: standard_schema.streaming_events.termination_reasons)
+termination:
+  source_field: "finish_reason"
+  mapping:
+    stop: "end_turn"
+    length: "max_tokens"
+    tool_calls: "tool_use"
+    content_filter: "refusal"
+  notes:
+    - "finish_reason is reported per candidate."
+
+# Tool invocation normalization (see v1/spec.yaml: standard_schema.content_blocks)
+tooling:
+  source_model: "openai_tool_calls"
+  tool_use:
+    id_path: "id"
+    name_path: "function.name"
+    input_path: "function.arguments"
+    input_format: "json_string"
+  notes:
+    - "OpenAI tool_calls.function.arguments is typically a JSON string; runtimes SHOULD parse it into an object when normalizing to tool_use.input."
+
+# Rate limits (VERIFIED evidence in research/providers/openai.md: OpenAI rate limit headers)
+rate_limit_headers:
+  requests_limit: "x-ratelimit-limit-requests"
+  tokens_limit: "x-ratelimit-limit-tokens"
+  requests_remaining: "x-ratelimit-remaining-requests"
+  tokens_remaining: "x-ratelimit-remaining-tokens"
+  requests_reset: "x-ratelimit-reset-requests"
+  tokens_reset: "x-ratelimit-reset-tokens"
+
+# Retry policy (VERIFIED guidance: exponential backoff for 429; 500 retry after brief wait)
+retry_policy:
+  max_retries: 3
+  max_delay_ms: 30000
+  strategy: "exponential_backoff"
+  min_delay_ms: 1000
+  jitter: "full"
+  retry_on_http_status: [429, 500, 502, 503]
+  notes:
+    - "429 may be rate limit reached OR quota exhausted; runtimes SHOULD avoid retrying quota-exhausted cases when detectable (e.g., via error message)."
+    - "For rate limit errors, exponential backoff is recommended."
+    - "For 500 server errors, retry after a brief wait."
+
+# Error classification hints (see v1/spec.yaml: standard_schema.error_handling.error_classes)
+error_classification:
+  by_http_status:
+    "400": "invalid_request"
+    "401": "authentication"
+    "403": "permission_denied"
+    "404": "not_found"
+    "429": "rate_limited"
+    "500": "server_error"
+  notes:
+    - "OpenAI 429 may also indicate quota/billing exhaustion; runtimes SHOULD treat it as quota_exhausted when the message indicates plan/billing/quota."
+parameter_mappings:
+  temperature: "temperature"
+  max_tokens: "max_tokens"
+  stream: "stream"
+  top_p: "top_p"
+  stop_sequences: "stop"
+  frequency_penalty: "frequency_penalty"
+  presence_penalty: "presence_penalty"
+  logprobs: "logprobs"
+  top_logprobs: "top_logprobs"
+  seed: "seed"
+  tools: "tools"
+  tool_choice: "tool_choice"
+  reasoning_effort: "reasoning_effort"
+response_format: "openai_style"
+response_paths:
+  content: "choices[0].message.content"
+  tool_calls: "choices[0].message.tool_calls"
+  usage: "usage"
+  finish_reason: "choices[0].finish_reason"
+streaming:
+  event_format: "data_lines"
+  decoder:
+    format: "sse"
+    delimiter: "\n\n"
+    prefix: "data: "
+    done_signal: "[DONE]"
+  # Allow both normal Chat Completions frames and stream/realtime error frames.
+  frame_selector: "exists($.choices) || exists($.error) || $.type == 'error'"
+  candidate:
+    candidate_id_path: "$.choices[*].index"
+    fan_out: true
+  event_map:
+    # Stream/realtime error event (first-class event)
+    - match: "exists($.error) || $.type == 'error'"
+      emit: "StreamError"
+      fields:
+        type: "$.type"
+        event_id: "$.event_id"
+        error: "$.error"
+    # Text content delta (supports multi-candidate via choices array)
+    - match: "exists($.choices[*].delta.content)"
+      emit: "PartialContentDelta"
+      fields:
+        content: "$.choices[*].delta.content"
+    # Tool call start (first chunk with tool_calls)
+    - match: "exists($.choices[*].delta.tool_calls[*].function.name)"
+      emit: "ToolCallStarted"
+      fields:
+        tool_call_id: "$.choices[*].delta.tool_calls[*].id"
+        tool_name: "$.choices[*].delta.tool_calls[*].function.name"
+        index: "$.choices[*].delta.tool_calls[*].index"
+    # Tool call arguments streaming
+    - match: "exists($.choices[*].delta.tool_calls[*].function.arguments)"
+      emit: "PartialToolCall"
+      fields:
+        arguments: "$.choices[*].delta.tool_calls[*].function.arguments"
+        index: "$.choices[*].delta.tool_calls[*].index"
+    # Usage metadata (OpenAI returns usage in final chunk with stream_options)
+    - match: "exists($.usage)"
+      emit: "Metadata"
+      fields:
+        usage: "$.usage"
+        prompt_tokens: "$.usage.prompt_tokens"
+        completion_tokens: "$.usage.completion_tokens"
+    # Finish with reason
+    - match: "exists($.choices[*].finish_reason)"
+      emit: "FinalCandidate"
+      fields:
+        finish_reason: "$.choices[*].finish_reason"
+        candidate_index: "$.choices[*].index"
+  stop_condition: "$.choices[0].finish_reason != null"
+  extra_metadata_path: "$.usage"
+  content_path: "choices[0].delta.content"
+  tool_call_path: "choices[0].delta.tool_calls"
+features:
+  multi_candidate:
+    support_type: "native"
+    param_name: "n"
+  response_mapping:
+    tool_calls:
+      path: "choices[0].message.tool_calls"
+      fields:
+        id: "id"
+        name: "function.name"
+        args: "function.arguments"
+    error:
+      message_path: "error.message"
+      code_path: "error.code"
+      type_path: "error.type"
+      param_path: "error.param"
+capabilities:
+  streaming: true
+  tools: true
+  vision: true
+  agentic: true
+  parallel_tools: true
+  reasoning: true
+experimental_features:
+  - "strict_tools"
+  - "parallel_tool_calls"
+  - "responses_api"
+
+# Availability and health checking (v1.1+ extension)
+availability:
+  required: false
+  regions:
+    - global
+  check:
+    method: GET
+    path: "/models"
+    expected_status: [200, 401]
+    timeout_ms: 3000

package/v1/providers/openrouter.yaml

@@ -0,0 +1,181 @@
+$schema: "https://raw.githubusercontent.com/ailib-official/ai-protocol/main/schemas/v1.json"
+
+id: openrouter
+protocol_version: "1.5"
+
+name: OpenRouter
+version: "v1"
+status: stable
+category: third_party_aggregator
+official_url: "https://openrouter.ai/docs"
+support_contact: "https://openrouter.ai/contact"
+
+endpoint:
+  base_url: "https://openrouter.ai/api/v1"
+  protocol: https
+  timeout_ms: 120000
+
+auth:
+  type: bearer
+  token_env: "OPENROUTER_API_KEY"
+
+payload_format: "openai_style"
+
+# OpenAI-compatible API (aggregates multiple providers)
+api_families: ["chat_completions"]
+default_api_family: "chat_completions"
+
+endpoints:
+  chat:
+    path: "/chat/completions"
+    method: "POST"
+    adapter: "openai"
+
+services:
+  list_models:
+    path: "/models"
+    method: "GET"
+    response_binding: "data"
+  get_generation:
+    path: "/generation"
+    method: "GET"
+
+termination:
+  source_field: "finish_reason"
+  mapping:
+    stop: "end_turn"
+    length: "max_tokens"
+    tool_calls: "tool_use"
+    content_filter: "refusal"
+  notes:
+    - "OpenRouter normalizes responses to OpenAI format."
+    - "Aggregates 100+ models from various providers."
+
+tooling:
+  source_model: "openai_tool_calls"
+  tool_use:
+    id_path: "id"
+    name_path: "function.name"
+    input_path: "function.arguments"
+    input_format: "json_string"
+  notes:
+    - "Tool support depends on underlying model."
+
+retry_policy:
+  strategy: "exponential_backoff"
+  max_retries: 3
+  min_delay_ms: 1000
+  max_delay_ms: 60000
+  jitter: "full"
+  retry_on_http_status: [429, 500, 502, 503]
+  notes:
+    - "OpenRouter may have longer latencies due to routing."
+
+error_classification:
+  by_http_status:
+    "400": "invalid_request"
+    "401": "authentication"
+    "402": "quota_exhausted"
+    "403": "permission_denied"
+    "404": "not_found"
+    "429": "rate_limited"
+    "500": "server_error"
+    "502": "server_error"
+    "503": "overloaded"
+
+parameter_mappings:
+  temperature: "temperature"
+  max_tokens: "max_tokens"
+  stream: "stream"
+  top_p: "top_p"
+  frequency_penalty: "frequency_penalty"
+  presence_penalty: "presence_penalty"
+  stop_sequences: "stop"
+  tools: "tools"
+  tool_choice: "tool_choice"
+  seed: "seed"
+  top_k: "top_k"
+
+response_format: "openai_style"
+
+response_paths:
+  content: "choices[0].message.content"
+  tool_calls: "choices[0].message.tool_calls"
+  usage: "usage"
+  finish_reason: "choices[0].finish_reason"
+
+streaming:
+  event_format: "data_lines"
+  decoder:
+    format: "sse"
+    delimiter: "\n\n"
+    prefix: "data: "
+    done_signal: "[DONE]"
+  content_path: "choices[0].delta.content"
+  tool_call_path: "choices[0].delta.tool_calls"
+  usage_path: "usage"
+  frame_selector: "exists($.choices) || exists($.error)"
+  event_map:
+    - match: "exists($.choices[*].delta.content)"
+      emit: "PartialContentDelta"
+      fields:
+        content: "$.choices[*].delta.content"
+    - match: "exists($.choices[*].delta.tool_calls[*].function.name)"
+      emit: "ToolCallStarted"
+      fields:
+        tool_call_id: "$.choices[*].delta.tool_calls[*].id"
+        tool_name: "$.choices[*].delta.tool_calls[*].function.name"
+        index: "$.choices[*].delta.tool_calls[*].index"
+    - match: "exists($.choices[*].delta.tool_calls[*].function.arguments)"
+      emit: "PartialToolCall"
+      fields:
+        arguments: "$.choices[*].delta.tool_calls[*].function.arguments"
+        index: "$.choices[*].delta.tool_calls[*].index"
+    - match: "exists($.usage)"
+      emit: "Metadata"
+      fields:
+        usage: "$.usage"
+    - match: "exists($.choices[*].finish_reason)"
+      emit: "FinalCandidate"
+      fields:
+        finish_reason: "$.choices[*].finish_reason"
+  stop_condition: "$.choices[0].finish_reason != null"
+
+features:
+  multi_candidate:
+    support_type: "native"
+    param_name: "n"
+  response_mapping:
+    tool_calls:
+      path: "choices[0].message.tool_calls"
+      fields:
+        id: "id"
+        name: "function.name"
+        args: "function.arguments"
+    error:
+      message_path: "error.message"
+      code_path: "error.code"
+      type_path: "error.type"
+
+capabilities:
+  streaming: true
+  tools: true
+  vision: true
+  agentic: true
+  parallel_tools: true
+  reasoning: true
+
+availability:
+  required: false
+  regions:
+    - global
+  check:
+    method: GET
+    path: "/models"
+    expected_status: [200, 401]
+    timeout_ms: 5000
+
+experimental_features:
+  - "model_routing"
+  - "fallback_models"
+  - "cost_tracking"

package/v1/providers/perplexity.yaml

@@ -0,0 +1,144 @@
+$schema: "https://raw.githubusercontent.com/ailib-official/ai-protocol/main/schemas/v1.json"
+
+id: perplexity
+protocol_version: "1.5"
+
+name: Perplexity
+version: "v1"
+status: stable
+category: ai_provider
+official_url: "https://docs.perplexity.ai"
+support_contact: "https://www.perplexity.ai/contact"
+
+endpoint:
+  base_url: "https://api.perplexity.ai"
+  protocol: https
+  timeout_ms: 60000
+
+auth:
+  type: bearer
+  token_env: "PPLX_API_KEY"
+
+payload_format: "openai_style"
+
+# OpenAI-compatible API
+api_families: ["chat_completions"]
+default_api_family: "chat_completions"
+
+endpoints:
+  chat:
+    path: "/chat/completions"
+    method: "POST"
+    adapter: "openai"
+
+services:
+  list_models:
+    path: "/models"
+    method: "GET"
+    response_binding: "data"
+
+termination:
+  source_field: "finish_reason"
+  mapping:
+    stop: "end_turn"
+    length: "max_tokens"
+  notes:
+    - "Perplexity is known for search-augmented AI responses."
+    - "Returns citations in responses when using online models."
+
+tooling:
+  source_model: "unknown"
+  notes:
+    - "Perplexity focuses on search/RAG, limited tool support."
+
+retry_policy:
+  strategy: "exponential_backoff"
+  max_retries: 3
+  min_delay_ms: 1000
+  max_delay_ms: 30000
+  jitter: "full"
+  retry_on_http_status: [429, 500, 502, 503]
+
+error_classification:
+  by_http_status:
+    "400": "invalid_request"
+    "401": "authentication"
+    "403": "permission_denied"
+    "404": "not_found"
+    "429": "rate_limited"
+    "500": "server_error"
+    "503": "overloaded"
+
+parameter_mappings:
+  temperature: "temperature"
+  max_tokens: "max_tokens"
+  stream: "stream"
+  top_p: "top_p"
+  frequency_penalty: "frequency_penalty"
+  presence_penalty: "presence_penalty"
+
+response_format: "openai_style"
+
+response_paths:
+  content: "choices[0].message.content"
+  usage: "usage"
+  finish_reason: "choices[0].finish_reason"
+  citations: "citations"
+
+streaming:
+  event_format: "data_lines"
+  decoder:
+    format: "sse"
+    delimiter: "\n\n"
+    prefix: "data: "
+    done_signal: "[DONE]"
+  content_path: "choices[0].delta.content"
+  usage_path: "usage"
+  frame_selector: "exists($.choices) || exists($.error)"
+  event_map:
+    - match: "exists($.choices[*].delta.content)"
+      emit: "PartialContentDelta"
+      fields:
+        content: "$.choices[*].delta.content"
+    - match: "exists($.usage)"
+      emit: "Metadata"
+      fields:
+        usage: "$.usage"
+    - match: "exists($.choices[*].finish_reason)"
+      emit: "FinalCandidate"
+      fields:
+        finish_reason: "$.choices[*].finish_reason"
+  stop_condition: "$.choices[0].finish_reason != null"
+
+features:
+  multi_candidate:
+    support_type: "simulated"
+    max_concurrent: 1
+  response_mapping:
+    error:
+      message_path: "error.message"
+      code_path: "error.code"
+      type_path: "error.type"
+
+capabilities:
+  streaming: true
+  tools: false
+  vision: false
+  agentic: false
+  parallel_tools: false
+  reasoning: false
+
+availability:
+  required: false
+  regions:
+    - global
+  check:
+    method: GET
+    path: "/models"
+    expected_status: [200, 401]
+    timeout_ms: 5000
+
+experimental_features:
+  - "online_search"
+  - "citations"
+  - "related_questions"