@ailib-official/ai-protocol 0.8.4

package/v1/models/zhipu.yaml

@@ -0,0 +1,163 @@
+$schema: "https://raw.githubusercontent.com/ailib-official/ai-protocol/main/schemas/v1.json"
+protocol_version: "1.1"
+
+# 智谱 AI (Zhipu/GLM) Model Instances
+#
+# VERIFIED_AT: 2026-01-31 (based on public Zhipu documentation)
+#
+# Notes:
+# - GLM (General Language Model) is Zhipu's flagship model series.
+# - Founded by Tsinghua University researchers.
+# - Known for strong Chinese language capabilities.
+
+models:
+  # GLM-4 series (latest)
+  glm-4-plus:
+    provider: zhipu
+    model_id: "glm-4-plus"
+    display_name: "GLM-4 Plus"
+    verification:
+      status: unverified
+      verified_at: "2026-01-31"
+      source: "documentation"
+      notes: "Latest flagship model with best performance."
+    context_window: 131072
+    capabilities: [chat, tools, streaming]
+    status: active
+    tags: ["zhipu", "glm", "flagship", "cn"]
+    pricing:
+      input_per_token: 0.00005
+      output_per_token: 0.00005
+
+  glm-4-0520:
+    provider: zhipu
+    model_id: "glm-4-0520"
+    display_name: "GLM-4 0520"
+    verification:
+      status: unverified
+      verified_at: "2026-01-31"
+      source: "documentation"
+    context_window: 131072
+    capabilities: [chat, tools, streaming]
+    status: active
+    tags: ["zhipu", "glm", "cn"]
+    pricing:
+      input_per_token: 0.0001
+      output_per_token: 0.0001
+
+  glm-4-air:
+    provider: zhipu
+    model_id: "glm-4-air"
+    display_name: "GLM-4 Air"
+    verification:
+      status: unverified
+      verified_at: "2026-01-31"
+      source: "documentation"
+      notes: "Cost-effective model with good performance."
+    context_window: 131072
+    capabilities: [chat, tools, streaming]
+    status: active
+    tags: ["zhipu", "glm", "cost_effective", "cn"]
+    pricing:
+      input_per_token: 0.000001
+      output_per_token: 0.000001
+
+  glm-4-airx:
+    provider: zhipu
+    model_id: "glm-4-airx"
+    display_name: "GLM-4 AirX"
+    verification:
+      status: unverified
+      verified_at: "2026-01-31"
+      source: "documentation"
+      notes: "Fast inference version of GLM-4 Air."
+    context_window: 8192
+    capabilities: [chat, tools, streaming]
+    status: active
+    tags: ["zhipu", "glm", "fast", "cn"]
+    pricing:
+      input_per_token: 0.00001
+      output_per_token: 0.00001
+
+  glm-4-flash:
+    provider: zhipu
+    model_id: "glm-4-flash"
+    display_name: "GLM-4 Flash"
+    verification:
+      status: unverified
+      verified_at: "2026-01-31"
+      source: "documentation"
+      notes: "Free tier model with rate limits."
+    context_window: 131072
+    capabilities: [chat, tools, streaming]
+    status: active
+    tags: ["zhipu", "glm", "free_tier", "cn"]
+    pricing:
+      input_per_token: 0
+      output_per_token: 0
+
+  glm-4-long:
+    provider: zhipu
+    model_id: "glm-4-long"
+    display_name: "GLM-4 Long"
+    verification:
+      status: unverified
+      verified_at: "2026-01-31"
+      source: "documentation"
+    context_window: 1048576
+    capabilities: [chat, streaming]
+    status: active
+    tags: ["zhipu", "glm", "ultra_long_context", "cn"]
+    pricing:
+      input_per_token: 0.000001
+      output_per_token: 0.000001
+
+  # GLM-4V Vision model
+  glm-4v-plus:
+    provider: zhipu
+    model_id: "glm-4v-plus"
+    display_name: "GLM-4V Plus"
+    verification:
+      status: unverified
+      verified_at: "2026-01-31"
+      source: "documentation"
+    context_window: 8192
+    capabilities: [chat, vision, streaming]
+    status: active
+    tags: ["zhipu", "glm", "vision", "multimodal", "cn"]
+    pricing:
+      input_per_token: 0.00001
+      output_per_token: 0.00001
+
+  glm-4v:
+    provider: zhipu
+    model_id: "glm-4v"
+    display_name: "GLM-4V"
+    verification:
+      status: unverified
+      verified_at: "2026-01-31"
+      source: "documentation"
+    context_window: 2048
+    capabilities: [chat, vision, streaming]
+    status: active
+    tags: ["zhipu", "glm", "vision", "multimodal", "cn"]
+    pricing:
+      input_per_token: 0.00005
+      output_per_token: 0.00005
+
+  # Embedding model
+  embedding-3:
+    provider: zhipu
+    model_id: "embedding-3"
+    display_name: "Zhipu Embedding 3"
+    verification:
+      status: unverified
+      verified_at: "2026-01-31"
+      source: "documentation"
+    context_window: 8192
+    capabilities: [embedding]
+    status: active
+    tags: ["zhipu", "embedding", "cn"]
+    pricing:
+      input_per_token: 0.0000005
+      output_per_token: 0

package/v1/providers/ai21.yaml

@@ -0,0 +1,176 @@
+$schema: "https://raw.githubusercontent.com/ailib-official/ai-protocol/main/schemas/v1.json"
+
+id: ai21
+protocol_version: "1.5"
+
+name: AI21 Labs
+version: "v1"
+status: stable
+category: ai_provider
+official_url: "https://docs.ai21.com"
+support_contact: "https://www.ai21.com/contact"
+
+endpoint:
+  base_url: "https://api.ai21.com/studio/v1"
+  protocol: https
+  timeout_ms: 30000
+
+auth:
+  type: bearer
+  token_env: "AI21_API_KEY"
+
+payload_format: "ai21_style"
+
+# AI21 supports both native and chat endpoints
+api_families: ["chat", "complete"]
+default_api_family: "chat"
+
+endpoints:
+  chat:
+    path: "/chat/completions"
+    method: "POST"
+    adapter: "ai21_chat"
+  # Jamba models use a different endpoint
+  jamba_chat:
+    path: "/jamba-1.5-large/chat/completions"
+    method: "POST"
+    adapter: "openai"
+
+services:
+  list_models:
+    path: "/models"
+    method: "GET"
+
+termination:
+  source_field: "finish_reason"
+  mapping:
+    stop: "end_turn"
+    length: "max_tokens"
+    endoftext: "end_turn"
+  notes:
+    - "AI21 uses finish_reason similar to OpenAI."
+    - "Jamba models follow OpenAI-compatible format."
+
+tooling:
+  source_model: "openai_tool_calls"
+  tool_use:
+    id_path: "id"
+    name_path: "function.name"
+    input_path: "function.arguments"
+    input_format: "json_string"
+  notes:
+    - "Jamba models support function calling in OpenAI format."
+
+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]
+  notes:
+    - "AI21 uses standard rate limiting."
+
+error_classification:
+  by_http_status:
+    "400": "invalid_request"
+    "401": "authentication"
+    "403": "permission_denied"
+    "404": "not_found"
+    "422": "invalid_request"
+    "429": "rate_limited"
+    "500": "server_error"
+    "503": "overloaded"
+
+parameter_mappings:
+  temperature: "temperature"
+  max_tokens: "max_tokens"
+  stream: "stream"
+  top_p: "top_p"
+  stop_sequences: "stop"
+  tools: "tools"
+  tool_choice: "tool_choice"
+  n: "n"
+
+response_format: "ai21_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: "sse"
+  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: "detail"
+      type_path: "type"
+
+capabilities:
+  streaming: true
+  tools: true
+  vision: false
+  agentic: true
+  parallel_tools: false
+  reasoning: false
+
+availability:
+  required: false
+  regions:
+    - global
+  check:
+    method: GET
+    path: "/models"
+    expected_status: [200, 401]
+    timeout_ms: 5000
+
+experimental_features:
+  - "contextual_answers"
+  - "summarization"
+  - "paraphrase"

package/v1/providers/anthropic.yaml

@@ -0,0 +1,209 @@
+$schema: "https://raw.githubusercontent.com/ailib-official/ai-protocol/main/schemas/v1.json"
+
+id: anthropic
+protocol_version: "1.5"
+
+name: Anthropic
+version: "v1"
+status: stable
+category: ai_provider
+official_url: "https://docs.anthropic.com"
+support_contact: "https://support.anthropic.com"
+
+# Structured endpoint definition (v1.1+ extension)
+endpoint:
+  base_url: "https://api.anthropic.com/v1"
+  protocol: https
+  timeout_ms: 10000
+
+auth:
+  type: bearer
+  token_env: "ANTHROPIC_API_KEY"
+  extra_headers:
+    - name: "anthropic-version"
+      value: "2023-06-01"
+payload_format: "anthropic_style"
+
+# Rate limits (VERIFIED evidence in research/providers/anthropic.md: rate limits + retry-after)
+rate_limit_headers:
+  retry_after: "retry-after"
+  requests_limit: "anthropic-ratelimit-requests-limit"
+  requests_remaining: "anthropic-ratelimit-requests-remaining"
+  requests_reset: "anthropic-ratelimit-requests-reset"
+  tokens_limit: "anthropic-ratelimit-tokens-limit"
+  tokens_remaining: "anthropic-ratelimit-tokens-remaining"
+  tokens_reset: "anthropic-ratelimit-tokens-reset"
+
+# Retry policy (VERIFIED evidence in research/providers/anthropic.md: SDK retry behavior + docs retry-after)
+retry_policy:
+  strategy: "exponential_backoff"
+  max_retries: 2
+  max_delay_ms: 60000
+  min_delay_ms: 1000
+  jitter: "full"
+  retry_on_http_status: [408, 409, 429, 500, 529]
+  notes:
+    - "On 429, Anthropic returns a retry-after header (seconds) indicating how long to wait."
+    - "SDKs commonly auto-retry certain errors 2 times with a short exponential backoff (429 and >=500 among others)."
+
+# 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"
+    "413": "request_too_large"
+    "429": "rate_limited"
+    "500": "server_error"
+    "529": "overloaded"
+  notes:
+    - "Anthropic uses error.type values (e.g., rate_limit_error, overloaded_error); runtimes SHOULD preserve error.type as raw metadata."
+
+# Anthropic native API families
+api_families: ["messages"]
+default_api_family: "messages"
+endpoints:
+  chat:
+    path: "/messages"
+    method: "POST"
+    adapter: "anthropic"
+
+# V1.5 Service Endpoints
+services:
+  list_models:
+    path: "/models"
+    method: "GET"
+    response_binding: "data"
+  get_usage_report:
+    path: "/organizations/usage_report/messages"
+    method: "POST"
+    headers:
+      "anthropic-beta": "message-batches-2024-09-24"
+
+# Termination reason normalization (see v1/spec.yaml: standard_schema.streaming_events.termination_reasons)
+termination:
+  source_field: "stop_reason"
+  mapping:
+    end_turn: "end_turn"
+    max_tokens: "max_tokens"
+    stop_sequence: "stop_sequence"
+    tool_use: "tool_use"
+    pause_turn: "pause_turn"
+    refusal: "refusal"
+  notes:
+    - "Non-streaming: stop_reason is always non-null."
+    - "Streaming: stop_reason is null in message_start, then non-null otherwise."
+parameter_mappings:
+  temperature: "temperature"
+  max_tokens: "max_tokens"
+  stream: "stream"
+  top_p: "top_p"
+  stop_sequences: "stop_sequences"
+  tools: "tools"
+  tool_choice: "tool_choice"
+  system_message: "system"
+  reasoning_effort: "reasoning_effort"
+response_format: "anthropic_style"
+response_paths:
+  content: "content[0].text"
+  tool_calls: "content[0].tool_calls"
+  usage: "usage"
+  stop_reason: "stop_reason"
+streaming:
+  event_format: "anthropic_sse"
+  decoder:
+    format: "anthropic_sse"
+    strategy: "anthropic_event_stream"
+    delimiter: "\n\n"
+    prefix: "data: "
+    done_signal: "[DONE]"
+  # Official event flow includes message_start and may include ping events.
+  frame_selector: "$.type in ['message_start', 'content_block_start', 'content_block_delta', 'content_block_stop', 'message_delta', 'message_stop', 'ping']"
+  accumulator:
+    stateful_tool_parsing: true
+    key_path: "$.delta.partial_json"
+    flush_on: "$.type == 'content_block_stop'"
+  event_map:
+    # Text content streaming
+    - match: "$.type == 'content_block_delta' && $.delta.type == 'text_delta'"
+      emit: "PartialContentDelta"
+      fields:
+        content: "$.delta.text"
+    # Thinking/reasoning delta (Claude 3.5+ extended thinking)
+    - match: "$.type == 'content_block_delta' && $.delta.type == 'thinking_delta'"
+      emit: "ThinkingDelta"
+      fields:
+        thinking: "$.delta.thinking"
+    # Tool call start (for multi-tool scenarios)
+    - match: "$.type == 'content_block_start' && $.content_block.type == 'tool_use'"
+      emit: "ToolCallStarted"
+      fields:
+        tool_call_id: "$.content_block.id"
+        tool_name: "$.content_block.name"
+        index: "$.index"
+    # Tool call arguments streaming
+    - match: "$.type == 'content_block_delta' && $.delta.type == 'input_json_delta'"
+      emit: "PartialToolCall"
+      fields:
+        arguments: "$.delta.partial_json"
+        index: "$.index"
+    # Tool call end
+    - match: "$.type == 'content_block_stop'"
+      emit: "ToolCallEnded"
+      fields:
+        index: "$.index"
+    # Message delta with usage and stop_reason
+    - match: "$.type == 'message_delta'"
+      emit: "Metadata"
+      fields:
+        stop_reason: "$.delta.stop_reason"
+        usage: "$.usage"
+    # Stream end
+    - match: "$.type == 'message_stop'"
+      emit: "StreamEnd"
+      fields:
+        finish_reason: "end_turn"
+  stop_condition: "$.type == 'message_stop'"
+  extra_metadata_path: "$.usage"
+  content_path: "delta.text"
+  tool_call_path: "delta.tool_calls"
+features:
+  multi_candidate:
+    support_type: "simulated"
+    max_concurrent: 4
+  response_mapping:
+    tool_calls:
+      path: "content"
+      filter: "type == 'tool_use'"
+      fields:
+        id: "id"
+        name: "name"
+        args: "input"
+        id_strategy: "path"
+      array_fan_out: true
+    error:
+      message_path: "error.message"
+      code_path: "error.type"
+      request_id_path: "request_id"
+capabilities:
+  streaming: true
+  tools: true
+  vision: true
+  agentic: true
+  parallel_tools: false
+  reasoning: true
+experimental_features:
+  - "thinking_blocks"
+  - "mcp"
+
+# 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/anyscale.yaml

@@ -0,0 +1,106 @@
+$schema: ../../schemas/v1.json
+protocol_version: "1.5"
+id: "anyscale"
+name: "Anyscale"
+status: "stable"
+category: "model_provider"
+official_url: "https://www.anyscale.com/"
+support_contact: "https://www.anyscale.com/contact"
+
+endpoint:
+  base_url: "https://api.anyscale.com/v1"
+  protocol: "https"
+  timeout_ms: 60000
+
+auth:
+  type: bearer
+  token_env: "ANYSCALE_API_TOKEN"
+  payload_format: "openai_style"
+
+services:
+  list_models:
+    path: "/models"
+    method: "GET"
+    response_binding: "data"
+
+availability:
+  required: false
+  regions: ["global"]
+  check:
+    method: "GET"
+    path: "/models"
+    expected_status: [200]
+
+capabilities:
+  streaming: true
+  tools: true
+  vision: false
+  agentic: false
+  reasoning: false
+  parallel_tools: false
+
+# Anyscale is OpenAI-compatible
+streaming:
+  event_format: "data_lines"
+  decoder:
+    format: "sse"
+    delimiter: "\n\n"
+    prefix: "data: "
+    done_signal: "[DONE]"
+
+response_format: "openai_style"
+response_paths:
+  content: "choices.0.message.content"
+  finish_reason: "choices.0.finish_reason"
+  model: "model"
+  id: "id"
+  created: "created"
+
+# Stream event normalization moved to streaming.event_map (v1/spec.yaml)
+# event_map is defined in the streaming section below
+# Error response mapping (see v1/spec.yaml)
+error_classification:
+  by_http_status:
+    "400": "invalid_request"
+    "401": "authentication"
+    "403": "permission_denied"
+    "404": "not_found"
+    "409": "conflict"
+    "429": "rate_limited"
+    "500": "server_error"
+    "502": "server_error"
+    "503": "server_error"
+    "504": "timeout"
+
+# Retry policy (standard template)
+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]
+
+rate_limit_headers:
+  requests_limit: "x-ratelimit-limit-requests"
+  requests_remaining: "x-ratelimit-remaining-requests"
+  tokens_limit: "x-ratelimit-limit-tokens"
+  tokens_remaining: "x-ratelimit-remaining-tokens"
+
+# Parameter mappings for chat completions
+parameter_mappings:
+  model: "model"
+  messages: "messages"
+  temperature: "temperature"
+  max_tokens: "max_tokens"
+  top_p: "top_p"
+  frequency_penalty: "frequency_penalty"
+  presence_penalty: "presence_penalty"
+  stop: "stop"
+  stream: "stream"
+
+notes:
+  - "Anyscale provides Ray-based model serving infrastructure"
+  - "OpenAI-compatible API with multiple model options"
+  - "Supports various LLMs including Llama, Mistral, and others"
+  - "Enterprise-grade model serving platform"