@peakinfer/cli 1.0.133

package/prompts/unified-analyzer.yaml

@@ -0,0 +1,241 @@
+# =============================================================================
+# SYNC NOTE: This file is copied from peakinfer-site (SOURCE OF TRUTH)
+# Source: peakinfer-site/prompts/unified-analyzer.yaml
+#
+# DO NOT MODIFY THIS FILE DIRECTLY IN THE CLI REPO.
+# All changes must be made in peakinfer-site first, then synced here.
+# =============================================================================
+
+id: unified-analyzer
+name: Unified Static Analyzer
+version: "1.7"
+description: |
+  Unified multi-dimensional LLM inference analyzer.
+  Single-call approach: one LLM call per file analyzing all 4 dimensions
+  (cost, latency, throughput, reliability) with actionable fixes.
+
+system: |
+  You are an expert LLM inference performance analyzer. Analyze the provided code file and return a comprehensive performance analysis with actionable fixes.
+
+  Your task:
+  1. Find all LLM inference points including:
+     - OpenAI SDK (client.chat.completions.create, openai.ChatCompletion.create, client.completions.create)
+     - Anthropic SDK (client.messages.create, anthropic.messages)
+     - Google AI / Vertex AI (genai.GenerativeModel, generate_content)
+     - Azure OpenAI (AzureOpenAI)
+     - AWS Bedrock (invoke_model, bedrock.converse, BedrockRuntime)
+     - Mistral, Cohere, Groq, Together AI, Fireworks, Replicate, Perplexity
+     - Vercel AI SDK (@ai-sdk/openai, @ai-sdk/anthropic, streamText, generateText, streamObject)
+     - LangChain (ChatOpenAI, ChatAnthropic, LLMChain, ConversationChain)
+     - LlamaIndex (llama_index.llms)
+     - Self-hosted (vLLM, TGI, TensorRT-LLM, Triton, Ollama, llama.cpp, sglang)
+     - HTTP calls to /v1/chat/completions or similar LLM API endpoints
+     - Claude Agent SDK usage
+  2. For each inference point, analyze across ALL 4 dimensions: cost, latency, throughput, and reliability
+  3. For each issue found, provide the EXACT original code line and a suggested fix
+
+  Return JSON in this exact format:
+  {
+    "inference_points": [
+      {
+        "id": "unique_id",
+        "line": <line_number>,
+        "provider": "<provider_name>",
+        "model": "<model_name_or_null>",
+        "call_type": "direct|wrapper|framework|http",
+        "original_code": "<exact code line from source>",
+        "cost_profile": {
+          "tier": "premium|standard|budget|unknown",
+          "estimated_cost_per_call": <number>,
+          "optimizations": [{"type": "string", "description": "string", "savings_percent": <number>}]
+        },
+        "latency_profile": {
+          "estimated_p95_ms": <number>,
+          "is_blocking": <boolean>,
+          "has_streaming": <boolean>,
+          "optimizations": [{"type": "string", "description": "string", "improvement_percent": <number>}]
+        },
+        "throughput_profile": {
+          "has_rate_limiting": <boolean>,
+          "has_batching": <boolean>,
+          "bottlenecks": [{"type": "string", "description": "string"}],
+          "optimizations": [{"type": "string", "description": "string", "improvement": "string"}]
+        },
+        "reliability_profile": {
+          "has_error_handling": <boolean>,
+          "has_retry": <boolean>,
+          "has_timeout": <boolean>,
+          "has_fallback": <boolean>,
+          "anti_patterns": [{"type": "string", "description": "string"}],
+          "optimizations": [{"type": "string", "description": "string", "priority": "high|medium|low"}]
+        },
+        "issues": [
+          {
+            "type": "model_downgrade|add_streaming|add_error_handling|add_retry|add_timeout|add_fallback",
+            "severity": "critical|warning|info",
+            "headline": "<short description>",
+            "evidence": "<why this is a problem>",
+            "original_code": "<exact line(s) that need to change>",
+            "suggested_fix": "<replacement code with correct indentation>",
+            "ai_agent_prompt": "<detailed instructions for AI agents like Claude to fix this>"
+          }
+        ]
+      }
+    ],
+    "imports": {
+      "llm_providers": ["provider1", "provider2"],
+      "frameworks": ["framework1"]
+    },
+    "insights": [
+      {
+        "severity": "critical|warning|info",
+        "category": "cost|latency|throughput|reliability",
+        "headline": "string",
+        "evidence": "string",
+        "recommendation": "string"
+      }
+    ]
+  }
+
+  ============================================================
+  CRITICAL: CODE FIX REQUIREMENTS (READ CAREFULLY)
+  ============================================================
+
+  ABSOLUTELY FORBIDDEN in suggested_fix - these will break the tool:
+  - "// ... existing code ..."
+  - "// existing logic"
+  - "// handle error"
+  - "// retry logic here"
+  - "// add your code here"
+  - Any placeholder or ellipsis comment
+
+  REQUIRED in suggested_fix:
+  - The COMPLETE, COMPILABLE replacement code
+  - Copy ALL the original code lines and ADD the fix around them
+  - Real error class names (Anthropic.RateLimitError, not generic Error)
+  - Actual implementation, not descriptions
+
+  For original_code:
+  - Include the ENTIRE function body that needs to change
+  - This will be shown in GitHub's "Suggested change" UI
+
+  ============================================================
+  EXAMPLE: Adding error handling to a function
+  ============================================================
+
+  If the source code is:
+  ```
+  export async function chat(prompt: string): Promise<string> {
+    const response = await client.messages.create({
+      model: 'claude-sonnet-4-20250514',
+      max_tokens: 2000,
+      messages: [{ role: 'user', content: prompt }],
+    });
+    return response.content[0].type === 'text' ? response.content[0].text : '';
+  }
+  ```
+
+  Then the issue should be:
+  {
+    "type": "add_error_handling",
+    "severity": "critical",
+    "headline": "No error handling",
+    "evidence": "LLM calls can fail unexpectedly.",
+    "original_code": "export async function chat(prompt: string): Promise<string> {\n  const response = await client.messages.create({\n    model: 'claude-sonnet-4-20250514',\n    max_tokens: 2000,\n    messages: [{ role: 'user', content: prompt }],\n  });\n  return response.content[0].type === 'text' ? response.content[0].text : '';\n}",
+    "suggested_fix": "export async function chat(prompt: string): Promise<string> {\n  try {\n    const response = await client.messages.create({\n      model: 'claude-sonnet-4-20250514',\n      max_tokens: 2000,\n      messages: [{ role: 'user', content: prompt }],\n    });\n    return response.content[0].type === 'text' ? response.content[0].text : '';\n  } catch (error) {\n    if (error instanceof Anthropic.RateLimitError) {\n      await new Promise(r => setTimeout(r, 1000));\n      return chat(prompt);\n    }\n    throw error;\n  }\n}",
+    "ai_agent_prompt": "Wrap the Anthropic API call in try-catch with retry on rate limit."
+  }
+
+  Notice: suggested_fix contains the COMPLETE function with REAL code, not placeholders.
+
+  Be thorough but concise. Focus on actionable insights across all 4 performance dimensions.
+
+  IMPORTANT: You MUST detect ALL of these LLM inference patterns:
+  - OpenAI SDK: client.chat.completions.create, openai.ChatCompletion.create, client.completions.create, client.embeddings.create
+  - Anthropic SDK: client.messages.create, anthropic.messages, create_message
+  - Google AI: genai.GenerativeModel, generate_content, vertexai.generative_models
+  - Azure OpenAI: AzureOpenAI client calls
+  - AWS Bedrock: invoke_model, bedrock.converse, BedrockRuntime
+  - Mistral: MistralClient, mistral.chat
+  - Cohere: cohere.chat, cohere.generate, CohereClient
+  - Groq: groq.chat, Groq()
+  - Together AI: together.chat, Together()
+  - Fireworks: fireworks.chat, Fireworks()
+  - Replicate: replicate.run, replicate.predictions.create
+  - Perplexity: perplexity.chat
+  - Vercel AI SDK: streamText(), generateText(), streamObject(), anthropic(), openai(), google() from @ai-sdk/*
+  - LangChain: ChatOpenAI, ChatAnthropic, ChatGoogleGenerativeAI, LLMChain, ConversationChain, .invoke(), .ainvoke()
+  - LlamaIndex: llama_index.llms, OpenAILike, query_engine
+  - Self-hosted: vLLM, TGI (InferenceClient), TensorRT-LLM, Triton, Ollama, llama.cpp, sglang
+  - HTTP calls: fetch/axios/requests to /v1/chat/completions, /v1/completions, or similar LLM endpoints
+  - Claude Agent SDK: query() function usage
+
+user_template: |
+  Analyze this {{language}} file for LLM inference points and their performance characteristics:
+
+  File: {{file_path}}
+
+  ```{{language}}
+  {{content}}
+  ```
+
+  Find ALL LLM inference calls in this file. Look for:
+  - OpenAI (client.chat.completions.create, etc.)
+  - Anthropic (client.messages.create, etc.)
+  - Google AI, Azure OpenAI, AWS Bedrock
+  - Mistral, Cohere, Groq, Together, Fireworks, Replicate, Perplexity
+  - Vercel AI SDK (streamText, generateText, @ai-sdk/*)
+  - LangChain, LlamaIndex
+  - Self-hosted (vLLM, TGI, Ollama, llama.cpp)
+  - HTTP calls to LLM endpoints
+
+  Return the JSON analysis for each inference point found.
+
+input_schema:
+  file_path: string
+  content: string
+  language: string
+
+output_format:
+  inference_points:
+    - id: string
+      line: number
+      provider: string
+      model: string|null
+      call_type: "direct|wrapper|framework|http"
+      original_code: string
+      cost_profile: object
+      latency_profile: object
+      throughput_profile: object
+      reliability_profile: object
+      issues:
+        - type: string
+          severity: "critical|warning|info"
+          headline: string
+          evidence: string
+          original_code: string
+          suggested_fix: string|null
+          ai_agent_prompt: string
+  imports:
+    llm_providers: string[]
+    frameworks: string[]
+  insights:
+    - severity: string
+      category: string
+      headline: string
+      evidence: string
+      recommendation: string
+
+constraints:
+  - Return valid JSON only
+  - original_code must be EXACT match from source (include full function if fixing function-level issue)
+  - suggested_fix must be COMPLETE, WORKING code - NO placeholders like "// existing code" or "// handle error"
+  - suggested_fix must be one-click usable - user clicks "Commit suggestion" and code works immediately
+  - ai_agent_prompt must be actionable instruction with specific file path and line number
+  - All line numbers must match actual source
+  - For error handling fixes, include REAL error types (Anthropic.RateLimitError, etc.) not generic "handle error"
+  - For model downgrade fixes, show the ACTUAL model string replacement
+
+defaults:
+  max_inference_points_per_file: 20
+  min_confidence: 0.5

package/schemas/inference-map.v0.1.json

@@ -0,0 +1,215 @@
+{
+  "$schema": "http://json-schema.org/draft-07/schema#",
+  "$id": "https://peakinfer.dev/schemas/inference-map.v0.1.json",
+  "title": "InferenceMap",
+  "description": "PeakInfer InferenceMap v0.1 - Schema for LLM inference point analysis results",
+  "type": "object",
+  "required": ["version", "root", "generatedAt", "summary", "callsites"],
+  "properties": {
+    "version": {
+      "type": "string",
+      "description": "Schema version",
+      "const": "0.1"
+    },
+    "root": {
+      "type": "string",
+      "description": "Root directory path that was analyzed"
+    },
+    "generatedAt": {
+      "type": "string",
+      "format": "date-time",
+      "description": "ISO 8601 timestamp when the analysis was generated"
+    },
+    "metadata": {
+      "type": "object",
+      "description": "Optional metadata about the analysis run",
+      "properties": {
+        "absolutePath": {
+          "type": "string",
+          "description": "Full absolute path that was analyzed"
+        },
+        "promptId": {
+          "type": "string",
+          "description": "ID of the analysis prompt used"
+        },
+        "promptVersion": {
+          "type": "string",
+          "description": "Version of the analysis prompt"
+        },
+        "templatesVersion": {
+          "type": "string",
+          "description": "Version of peakinfer-templates used"
+        },
+        "llmProvider": {
+          "type": "string",
+          "description": "LLM provider used for analysis (e.g., 'anthropic')"
+        },
+        "llmModel": {
+          "type": "string",
+          "description": "LLM model used for analysis"
+        }
+      }
+    },
+    "summary": {
+      "type": "object",
+      "description": "Summary statistics for the analyzed codebase",
+      "required": ["totalCallsites", "providers", "models", "patterns"],
+      "properties": {
+        "totalCallsites": {
+          "type": "integer",
+          "minimum": 0,
+          "description": "Total number of inference points found"
+        },
+        "providers": {
+          "type": "array",
+          "items": { "type": "string" },
+          "description": "List of unique LLM providers detected"
+        },
+        "models": {
+          "type": "array",
+          "items": { "type": "string" },
+          "description": "List of unique models detected"
+        },
+        "patterns": {
+          "type": "object",
+          "description": "Count of detected patterns (streaming, batching, etc.)",
+          "additionalProperties": { "type": "integer" }
+        }
+      }
+    },
+    "callsites": {
+      "type": "array",
+      "description": "List of inference points (callsites) found in the codebase",
+      "items": {
+        "$ref": "#/definitions/Callsite"
+      }
+    }
+  },
+  "definitions": {
+    "Provider": {
+      "type": "string",
+      "enum": [
+        "openai", "anthropic", "google", "cohere", "mistral",
+        "bedrock", "azure_openai", "together", "fireworks",
+        "groq", "replicate", "perplexity",
+        "vllm", "sglang", "tgi", "ollama", "llamacpp",
+        "unknown"
+      ],
+      "description": "LLM provider identifier"
+    },
+    "Patterns": {
+      "type": "object",
+      "description": "Detected code patterns at an inference point",
+      "properties": {
+        "streaming": {
+          "type": "boolean",
+          "description": "Whether streaming is enabled"
+        },
+        "batching": {
+          "type": "boolean",
+          "description": "Whether batching is used"
+        },
+        "retries": {
+          "type": "boolean",
+          "description": "Whether retry logic is present"
+        },
+        "caching": {
+          "type": "boolean",
+          "description": "Whether caching is implemented"
+        },
+        "fallback": {
+          "type": "boolean",
+          "description": "Whether fallback logic exists"
+        }
+      }
+    },
+    "Callsite": {
+      "type": "object",
+      "description": "An LLM inference point in the codebase",
+      "required": ["id", "file", "line", "patterns", "confidence"],
+      "properties": {
+        "id": {
+          "type": "string",
+          "description": "Unique identifier for this inference point"
+        },
+        "file": {
+          "type": "string",
+          "description": "Relative file path from root"
+        },
+        "line": {
+          "type": "integer",
+          "minimum": 1,
+          "description": "Line number in the file"
+        },
+        "provider": {
+          "oneOf": [
+            { "$ref": "#/definitions/Provider" },
+            { "type": "null" }
+          ],
+          "description": "Detected LLM provider"
+        },
+        "model": {
+          "type": ["string", "null"],
+          "description": "Detected model name"
+        },
+        "framework": {
+          "type": ["string", "null"],
+          "description": "Framework used (e.g., 'langchain', 'llamaindex')"
+        },
+        "runtime": {
+          "type": ["string", "null"],
+          "description": "Runtime environment (e.g., 'vllm', 'tgi')"
+        },
+        "patterns": {
+          "$ref": "#/definitions/Patterns"
+        },
+        "confidence": {
+          "type": "number",
+          "minimum": 0,
+          "maximum": 1,
+          "description": "Confidence score of the detection (0-1)"
+        }
+      }
+    }
+  },
+  "examples": [
+    {
+      "version": "0.1",
+      "root": "./src",
+      "generatedAt": "2025-12-24T12:00:00.000Z",
+      "metadata": {
+        "absolutePath": "/Users/dev/myproject/src",
+        "promptId": "unified-analyzer",
+        "promptVersion": "1.6",
+        "llmProvider": "anthropic",
+        "llmModel": "claude-sonnet-4-20250514"
+      },
+      "summary": {
+        "totalCallsites": 3,
+        "providers": ["anthropic", "openai"],
+        "models": ["claude-sonnet-4-20250514", "gpt-4o"],
+        "patterns": {
+          "streaming": 2,
+          "retries": 1,
+          "caching": 0
+        }
+      },
+      "callsites": [
+        {
+          "id": "cs-001",
+          "file": "services/chat.ts",
+          "line": 42,
+          "provider": "anthropic",
+          "model": "claude-sonnet-4-20250514",
+          "framework": null,
+          "runtime": null,
+          "patterns": {
+            "streaming": true,
+            "retries": true
+          },
+          "confidence": 0.95
+        }
+      ]
+    }
+  ]
+}