noumen 0.5.0 → 0.8.0

package/dist/chunk-HEQQQGK5.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/utils/context.ts","../src/providers/cache.ts"],"sourcesContent":["/**\n * Model context window sizes and effective window calculations.\n */\n\nconst MODEL_CONTEXT_WINDOWS: Record<string, number> = {\n  // Anthropic (evergreen prefixes — also match dated variants via startsWith)\n  \"claude-sonnet-4\": 200_000,\n  \"claude-opus-4\": 200_000,\n  \"claude-haiku-4\": 200_000,\n  \"claude-haiku-3-5\": 200_000,\n  \"claude-3-5-sonnet\": 200_000,\n  \"claude-3-5-haiku\": 200_000,\n  // Bedrock / Vertex model ID patterns (prefix-matched)\n  \"us.anthropic.claude\": 200_000,\n  \"eu.anthropic.claude\": 200_000,\n  \"ap.anthropic.claude\": 200_000,\n  \"anthropic.claude\": 200_000,\n  // OpenAI\n  \"gpt-4.1\": 1_047_576,\n  \"gpt-4.1-mini\": 1_047_576,\n  \"gpt-4.1-nano\": 1_047_576,\n  \"gpt-4o\": 128_000,\n  \"gpt-4o-mini\": 128_000,\n  \"gpt-4-turbo\": 128_000,\n  \"gpt-4\": 8_192,\n  \"o1\": 200_000,\n  \"o1-mini\": 128_000,\n  \"o1-preview\": 128_000,\n  \"o3\": 200_000,\n  \"o3-mini\": 200_000,\n  \"o4-mini\": 200_000,\n  // Google\n  \"gemini-2.5-pro\": 1_048_576,\n  \"gemini-2.5-flash\": 1_048_576,\n  \"gemini-2.0-flash\": 1_048_576,\n  \"gemini-1.5-pro\": 2_097_152,\n  \"gemini-1.5-flash\": 1_048_576,\n};\n\nconst MODEL_MAX_OUTPUT_TOKENS: Record<string, number> = {\n  \"claude-sonnet-4-6\": 128_000,\n  \"claude-opus-4-6\": 128_000,\n  \"claude-sonnet-4\": 64_000,\n  \"claude-opus-4\": 64_000,\n  \"claude-haiku-4\": 64_000,\n  \"claude-haiku-3-5\": 8_192,\n  \"claude-3-5-sonnet\": 8_192,\n  \"claude-3-5-haiku\": 8_192,\n  \"us.anthropic.claude\": 64_000,\n  \"eu.anthropic.claude\": 64_000,\n  \"ap.anthropic.claude\": 64_000,\n  \"anthropic.claude\": 64_000,\n  \"gpt-4.1\": 32_768,\n  \"gpt-4.1-mini\": 32_768,\n  \"gpt-4.1-nano\": 32_768,\n  \"gpt-4o\": 16_384,\n  \"gpt-4o-mini\": 16_384,\n  \"gpt-4-turbo\": 4_096,\n  \"o1\": 100_000,\n  \"o3\": 100_000,\n  \"o3-mini\": 100_000,\n  \"o4-mini\": 100_000,\n  \"gemini-2.5-pro\": 65_536,\n  \"gemini-2.5-flash\": 65_536,\n  \"gemini-2.0-flash\": 8_192,\n};\n\nconst DEFAULT_MAX_OUTPUT_TOKENS = 16_384;\nconst DEFAULT_CONTEXT_WINDOW = 128_000;\nconst AUTOCOMPACT_BUFFER_TOKENS = 13_000;\nconst MAX_OUTPUT_RESERVE = 20_000;\n\nlet customWindows: Record<string, number> = {};\n\n/**\n * Register custom context window sizes for models not in the built-in table.\n */\nexport function registerContextWindows(\n  windows: Record<string, number>,\n): void {\n  customWindows = { ...customWindows, ...windows };\n}\n\n/**\n * Get the context window size for a model. Checks custom overrides first,\n * then built-in table, then prefix-matches, then falls back to default.\n */\nexport function getContextWindowForModel(model: string): number {\n  if (customWindows[model] !== undefined) return customWindows[model];\n  if (MODEL_CONTEXT_WINDOWS[model] !== undefined)\n    return MODEL_CONTEXT_WINDOWS[model];\n\n  for (const [prefix, size] of Object.entries(MODEL_CONTEXT_WINDOWS)) {\n    if (model.startsWith(prefix)) return size;\n  }\n  for (const [prefix, size] of Object.entries(customWindows)) {\n    if (model.startsWith(prefix)) return size;\n  }\n\n  return DEFAULT_CONTEXT_WINDOW;\n}\n\n/**\n * Effective context window = total window minus space reserved for the\n * model's output during a compaction/summary request.\n */\nexport function getEffectiveContextWindow(\n  model: string,\n  maxOutputTokens?: number,\n): number {\n  const window = getContextWindowForModel(model);\n  const reserve = Math.min(maxOutputTokens ?? MAX_OUTPUT_RESERVE, MAX_OUTPUT_RESERVE);\n  return window - reserve;\n}\n\n/**\n * Auto-compact threshold: effective window minus a buffer to ensure we\n * compact before we're at the hard limit.\n */\nexport function getAutoCompactThreshold(\n  model: string,\n  maxOutputTokens?: number,\n): number {\n  return getEffectiveContextWindow(model, maxOutputTokens) - AUTOCOMPACT_BUFFER_TOKENS;\n}\n\n/**\n * Get the maximum output tokens a model supports. Used to clamp\n * max_tokens when extended thinking budgets are added.\n */\nexport function getMaxOutputTokensForModel(model: string): number {\n  if (MODEL_MAX_OUTPUT_TOKENS[model] !== undefined)\n    return MODEL_MAX_OUTPUT_TOKENS[model];\n  for (const [prefix, size] of Object.entries(MODEL_MAX_OUTPUT_TOKENS)) {\n    if (model.startsWith(prefix)) return size;\n  }\n  return DEFAULT_MAX_OUTPUT_TOKENS;\n}\n","/**\n * Provider-agnostic prompt caching utilities.\n *\n * Stable tool ordering prevents cache invalidation when the tool set is\n * unchanged. The breakpoint index helper determines which message gets a\n * single cache_control marker per request (matching claude-code's strategy).\n */\n\nimport type { ToolDefinition } from \"./types.js\";\nimport type { ChatMessage } from \"../session/types.js\";\n\nexport type CacheScope = \"global\" | \"org\";\n\nexport interface CacheControlConfig {\n  enabled: boolean;\n  /** TTL for cached content. When set, produces `ttl: '1h'` in cache_control. */\n  ttl?: \"1h\";\n  /** Scope for shared cache across sessions/orgs. */\n  scope?: CacheScope;\n}\n\n/**\n * Sort tool definitions deterministically for prompt cache stability.\n *\n * Strategy (matching claude-code's assembleToolPool): built-in tools form a\n * contiguous prefix sorted by name, followed by MCP/external tools sorted by\n * name. Tools with `mcpInfo` on the original Tool object are treated as MCP;\n * everything else is built-in. Since ToolDefinition doesn't carry mcpInfo,\n * callers can pass an optional set of MCP tool names to partition correctly.\n */\nexport function sortToolDefinitionsForCache(\n  tools: ToolDefinition[],\n  mcpToolNames?: ReadonlySet<string>,\n): ToolDefinition[] {\n  const byName = (a: ToolDefinition, b: ToolDefinition) =>\n    a.function.name.localeCompare(b.function.name);\n\n  if (!mcpToolNames || mcpToolNames.size === 0) {\n    return [...tools].sort(byName);\n  }\n\n  const builtIn: ToolDefinition[] = [];\n  const mcp: ToolDefinition[] = [];\n\n  for (const t of tools) {\n    if (mcpToolNames.has(t.function.name)) {\n      mcp.push(t);\n    } else {\n      builtIn.push(t);\n    }\n  }\n\n  return [...builtIn.sort(byName), ...mcp.sort(byName)];\n}\n\n/**\n * Determine which message index should receive the cache_control breakpoint.\n *\n * Exactly one message per request is marked. Normally the last message;\n * for forked agents with skipCacheWrite the second-to-last so the fork\n * doesn't write its own tail into the cache.\n */\nexport function getMessageCacheBreakpointIndex(\n  messages: ChatMessage[],\n  skipCacheWrite?: boolean,\n): number {\n  if (messages.length === 0) return -1;\n  return skipCacheWrite && messages.length >= 2\n    ? messages.length - 2\n    : messages.length - 1;\n}\n"],"mappings":";AAIA,IAAM,wBAAgD;AAAA;AAAA,EAEpD,mBAAmB;AAAA,EACnB,iBAAiB;AAAA,EACjB,kBAAkB;AAAA,EAClB,oBAAoB;AAAA,EACpB,qBAAqB;AAAA,EACrB,oBAAoB;AAAA;AAAA,EAEpB,uBAAuB;AAAA,EACvB,uBAAuB;AAAA,EACvB,uBAAuB;AAAA,EACvB,oBAAoB;AAAA;AAAA,EAEpB,WAAW;AAAA,EACX,gBAAgB;AAAA,EAChB,gBAAgB;AAAA,EAChB,UAAU;AAAA,EACV,eAAe;AAAA,EACf,eAAe;AAAA,EACf,SAAS;AAAA,EACT,MAAM;AAAA,EACN,WAAW;AAAA,EACX,cAAc;AAAA,EACd,MAAM;AAAA,EACN,WAAW;AAAA,EACX,WAAW;AAAA;AAAA,EAEX,kBAAkB;AAAA,EAClB,oBAAoB;AAAA,EACpB,oBAAoB;AAAA,EACpB,kBAAkB;AAAA,EAClB,oBAAoB;AACtB;AAEA,IAAM,0BAAkD;AAAA,EACtD,qBAAqB;AAAA,EACrB,mBAAmB;AAAA,EACnB,mBAAmB;AAAA,EACnB,iBAAiB;AAAA,EACjB,kBAAkB;AAAA,EAClB,oBAAoB;AAAA,EACpB,qBAAqB;AAAA,EACrB,oBAAoB;AAAA,EACpB,uBAAuB;AAAA,EACvB,uBAAuB;AAAA,EACvB,uBAAuB;AAAA,EACvB,oBAAoB;AAAA,EACpB,WAAW;AAAA,EACX,gBAAgB;AAAA,EAChB,gBAAgB;AAAA,EAChB,UAAU;AAAA,EACV,eAAe;AAAA,EACf,eAAe;AAAA,EACf,MAAM;AAAA,EACN,MAAM;AAAA,EACN,WAAW;AAAA,EACX,WAAW;AAAA,EACX,kBAAkB;AAAA,EAClB,oBAAoB;AAAA,EACpB,oBAAoB;AACtB;AAEA,IAAM,4BAA4B;AAClC,IAAM,yBAAyB;AAC/B,IAAM,4BAA4B;AAClC,IAAM,qBAAqB;AAE3B,IAAI,gBAAwC,CAAC;AAKtC,SAAS,uBACd,SACM;AACN,kBAAgB,EAAE,GAAG,eAAe,GAAG,QAAQ;AACjD;AAMO,SAAS,yBAAyB,OAAuB;AAC9D,MAAI,cAAc,KAAK,MAAM,OAAW,QAAO,cAAc,KAAK;AAClE,MAAI,sBAAsB,KAAK,MAAM;AACnC,WAAO,sBAAsB,KAAK;AAEpC,aAAW,CAAC,QAAQ,IAAI,KAAK,OAAO,QAAQ,qBAAqB,GAAG;AAClE,QAAI,MAAM,WAAW,MAAM,EAAG,QAAO;AAAA,EACvC;AACA,aAAW,CAAC,QAAQ,IAAI,KAAK,OAAO,QAAQ,aAAa,GAAG;AAC1D,QAAI,MAAM,WAAW,MAAM,EAAG,QAAO;AAAA,EACvC;AAEA,SAAO;AACT;AAMO,SAAS,0BACd,OACA,iBACQ;AACR,QAAM,SAAS,yBAAyB,KAAK;AAC7C,QAAM,UAAU,KAAK,IAAI,mBAAmB,oBAAoB,kBAAkB;AAClF,SAAO,SAAS;AAClB;AAMO,SAAS,wBACd,OACA,iBACQ;AACR,SAAO,0BAA0B,OAAO,eAAe,IAAI;AAC7D;AAMO,SAAS,2BAA2B,OAAuB;AAChE,MAAI,wBAAwB,KAAK,MAAM;AACrC,WAAO,wBAAwB,KAAK;AACtC,aAAW,CAAC,QAAQ,IAAI,KAAK,OAAO,QAAQ,uBAAuB,GAAG;AACpE,QAAI,MAAM,WAAW,MAAM,EAAG,QAAO;AAAA,EACvC;AACA,SAAO;AACT;;;AC3GO,SAAS,4BACd,OACA,cACkB;AAClB,QAAM,SAAS,CAAC,GAAmB,MACjC,EAAE,SAAS,KAAK,cAAc,EAAE,SAAS,IAAI;AAE/C,MAAI,CAAC,gBAAgB,aAAa,SAAS,GAAG;AAC5C,WAAO,CAAC,GAAG,KAAK,EAAE,KAAK,MAAM;AAAA,EAC/B;AAEA,QAAM,UAA4B,CAAC;AACnC,QAAM,MAAwB,CAAC;AAE/B,aAAW,KAAK,OAAO;AACrB,QAAI,aAAa,IAAI,EAAE,SAAS,IAAI,GAAG;AACrC,UAAI,KAAK,CAAC;AAAA,IACZ,OAAO;AACL,cAAQ,KAAK,CAAC;AAAA,IAChB;AAAA,EACF;AAEA,SAAO,CAAC,GAAG,QAAQ,KAAK,MAAM,GAAG,GAAG,IAAI,KAAK,MAAM,CAAC;AACtD;AASO,SAAS,+BACd,UACA,gBACQ;AACR,MAAI,SAAS,WAAW,EAAG,QAAO;AAClC,SAAO,kBAAkB,SAAS,UAAU,IACxC,SAAS,SAAS,IAClB,SAAS,SAAS;AACxB;","names":[]}

package/dist/chunk-L3L3FG5T.js

@@ -1,16 +0,0 @@
-// src/providers/types.ts
-var ChatStreamError = class extends Error {
-  status;
-  retryAfter;
-  constructor(message, opts) {
-    super(message, { cause: opts?.cause });
-    this.name = "ChatStreamError";
-    this.status = opts?.status;
-    this.retryAfter = opts?.retryAfter;
-  }
-};
-
-export {
-  ChatStreamError
-};
-//# sourceMappingURL=chunk-L3L3FG5T.js.map

package/dist/chunk-L3L3FG5T.js.map

@@ -1 +0,0 @@
-{"version":3,"sources":["../src/providers/types.ts"],"sourcesContent":["import type { ChatMessage } from \"../session/types.js\";\nimport type { ThinkingConfig } from \"../thinking/types.js\";\n\nexport interface ToolParameterProperty {\n  type: string;\n  description?: string;\n  enum?: string[];\n  default?: unknown;\n  minimum?: number;\n  maximum?: number;\n}\n\nexport interface ToolDefinition {\n  type: \"function\";\n  function: {\n    name: string;\n    description: string;\n    parameters: {\n      type: \"object\";\n      properties: Record<string, ToolParameterProperty>;\n      required?: string[];\n    };\n  };\n}\n\n// Streaming chunk types (OpenAI-compatible)\n\nexport interface ChatStreamDelta {\n  role?: \"assistant\";\n  content?: string | null;\n  thinking_content?: string | null;\n  thinking_signature?: string | null;\n  /** Opaque data payload for Anthropic redacted_thinking blocks. */\n  redacted_thinking_data?: string | null;\n  tool_calls?: Array<{\n    index: number;\n    id?: string;\n    type?: \"function\";\n    function?: {\n      name?: string;\n      arguments?: string;\n    };\n  }>;\n}\n\nexport interface ChatStreamChoice {\n  index: number;\n  delta: ChatStreamDelta;\n  finish_reason: string | null;\n}\n\nexport interface ChatStreamChunk {\n  id: string;\n  choices: ChatStreamChoice[];\n  model: string;\n  usage?: {\n    prompt_tokens: number;\n    completion_tokens: number;\n    total_tokens: number;\n    cache_read_tokens?: number;\n    cache_creation_tokens?: number;\n    thinking_tokens?: number;\n  };\n}\n\nexport interface ChatCompletionUsage {\n  prompt_tokens: number;\n  completion_tokens: number;\n  total_tokens: number;\n  cache_read_tokens?: number;\n  cache_creation_tokens?: number;\n  thinking_tokens?: number;\n}\n\n/**\n * Structured output format. When provided, the model is constrained to\n * produce a response matching the given JSON schema.\n *\n * - `json_schema`: the model must produce JSON conforming to the given schema.\n * - `json_object`: the model must produce valid JSON (no specific schema).\n */\nexport type OutputFormat = JsonSchemaOutputFormat | JsonObjectOutputFormat;\n\nexport interface JsonSchemaOutputFormat {\n  type: \"json_schema\";\n  /** JSON Schema object describing the expected output shape. */\n  schema: Record<string, unknown>;\n  /** Optional name for the schema (required by some providers). */\n  name?: string;\n  /** When true, the provider enforces strict schema adherence. */\n  strict?: boolean;\n}\n\nexport interface JsonObjectOutputFormat {\n  type: \"json_object\";\n}\n\nexport interface ChatParams {\n  model: string;\n  messages: ChatMessage[];\n  tools?: ToolDefinition[];\n  max_tokens?: number;\n  system?: string;\n  temperature?: number;\n  thinking?: ThinkingConfig;\n  /** Constrain the model to produce structured output matching this schema. */\n  outputFormat?: OutputFormat;\n  /**\n   * When true, the provider should place the cache breakpoint on the\n   * second-to-last message instead of the last. Used by subagent forks\n   * to avoid writing fork-only tails into the shared prompt cache.\n   */\n  skipCacheWrite?: boolean;\n  /** Abort signal — providers should forward this to cancel in-flight HTTP requests. */\n  signal?: AbortSignal;\n}\n\nexport interface AIProvider {\n  chat(params: ChatParams): AsyncIterable<ChatStreamChunk>;\n}\n\n/**\n * Extended error type that providers can throw to convey retry-relevant metadata.\n * Consumers (like the retry engine) can inspect these fields without knowing\n * provider-specific SDK error types.\n */\nexport class ChatStreamError extends Error {\n  status?: number;\n  retryAfter?: string;\n\n  constructor(\n    message: string,\n    opts?: { status?: number; retryAfter?: string; cause?: unknown },\n  ) {\n    super(message, { cause: opts?.cause });\n    this.name = \"ChatStreamError\";\n    this.status = opts?.status;\n    this.retryAfter = opts?.retryAfter;\n  }\n}\n"],"mappings":";AA8HO,IAAM,kBAAN,cAA8B,MAAM;AAAA,EACzC;AAAA,EACA;AAAA,EAEA,YACE,SACA,MACA;AACA,UAAM,SAAS,EAAE,OAAO,MAAM,MAAM,CAAC;AACrC,SAAK,OAAO;AACZ,SAAK,SAAS,MAAM;AACpB,SAAK,aAAa,MAAM;AAAA,EAC1B;AACF;","names":[]}