@graphorin/provider 0.6.0 → 0.7.0

package/dist/internal/openai-shaped.js

@@ -82,7 +82,7 @@ function applyOpenAIShapedPreflight(req, capabilities) {
 	};
 }
 async function* streamOpenAIShaped(opts, url, req) {
-	const body = buildBody(opts.model, req, true, opts.capabilities?.structuredOutput ?? true);
+	const body = buildBody(opts.model, req, true, opts.capabilities?.structuredOutput ?? true, conversionOptionsFor(opts));
 	const resp = await callJsonHttp({
 		providerName: opts.providerName,
 		url,
@@ -177,7 +177,7 @@ async function* streamOpenAIShaped(opts, url, req) {
 	};
 }
 async function generateOpenAIShaped(opts, url, req) {
-	const body = buildBody(opts.model, req, false, opts.capabilities?.structuredOutput ?? true);
+	const body = buildBody(opts.model, req, false, opts.capabilities?.structuredOutput ?? true, conversionOptionsFor(opts));
 	const resp = await callJsonHttp({
 		providerName: opts.providerName,
 		url,
@@ -212,13 +212,13 @@ async function generateOpenAIShaped(opts, url, req) {
 		...toolCalls !== void 0 ? { toolCalls } : {}
 	};
 }
-function buildBody(model, req, stream, structuredOutput) {
+function buildBody(model, req, stream, structuredOutput, conversion) {
 	const body = {
 		model,
 		messages: toOpenAIChatMessages(req.systemMessage !== void 0 ? [{
 			role: "system",
 			content: req.systemMessage
-		}, ...req.messages] : req.messages),
+		}, ...req.messages] : req.messages, conversion),
 		stream
 	};
 	if (stream) body.stream_options = { include_usage: true };
@@ -307,6 +307,23 @@ function defaultLogger(level, message, meta) {
 	if (meta !== void 0) fn(`[graphorin/provider] ${message}`, meta);
 	else fn(`[graphorin/provider] ${message}`);
 }
+/**
+* W-095: one dropped-content WARN per adapter INSTANCE (keyed on the
+* factory's options object, following KNOWN_LOOPBACK_OVERRIDES_WARNED),
+* so a chat loop does not repeat it every call.
+*/
+const droppedContentWarned = /* @__PURE__ */ new WeakSet();
+function conversionOptionsFor(opts) {
+	const log = opts.logger ?? defaultLogger;
+	return {
+		multimodal: opts.capabilities?.multimodal ?? false,
+		warn: (message) => {
+			if (droppedContentWarned.has(opts)) return;
+			droppedContentWarned.add(opts);
+			log("warn", message);
+		}
+	};
+}
 
 //#endregion
 export { buildOpenAIShapedProvider };

package/dist/internal/openai-shaped.js.map

@@ -1 +1 @@
-{"version":3,"file":"openai-shaped.js","names":["DEFAULT_CAPABILITIES: ProviderCapabilities","capabilities: ProviderCapabilities","finishReason: FinishReason","usage: Usage | undefined","parsed: OpenAIChunk","json: OpenAIChunk","body: Record<string, unknown>","headers: Record<string, string>"],"sources":["../../src/internal/openai-shaped.ts"],"sourcesContent":["/**\n * Shared implementation backing the two OpenAI-compatible adapters\n * (`llamaCppServerAdapter` and `openAICompatibleAdapter`). The wire\n * shape is identical for the `/v1/chat/completions` endpoint, so we\n * factor the streaming + one-shot logic here and the adapters become\n * thin shells declaring trust class, defaults, and capability flags.\n *\n * @internal\n */\n\nimport type {\n  FinishReason,\n  Provider,\n  ProviderCapabilities,\n  ProviderEvent,\n  ProviderRequest,\n  ProviderResponse,\n  Sensitivity,\n  Usage,\n} from '@graphorin/core';\n\nimport { LocalProviderInsecureTransportError, ProviderStreamParseError } from '../errors/errors.js';\nimport { applyReasoningPolicy } from '../reasoning/apply-policy.js';\nimport { resolveReasoningRetention } from '../reasoning/retention.js';\nimport { foldToolExamples } from '../tool-examples.js';\nimport {\n  classifyLocalProvider,\n  type LocalProviderClassification,\n} from '../trust/classify-local-provider.js';\nimport { callJsonHttp, makeStreamStartEvent, toOpenAIChatMessages } from './http.js';\nimport { parseEventStream } from './sse.js';\nimport { stripTrailingSlashes } from './url-utils.js';\n\n/**\n * Options shared by every OpenAI-compatible adapter.\n *\n * @internal\n */\nexport interface OpenAIShapedOptions {\n  readonly providerName: string;\n  readonly model: string;\n  readonly baseUrl: string;\n  readonly chatPath?: string;\n  readonly apiKey?: string;\n  readonly headers?: Readonly<Record<string, string>>;\n  readonly fetchImpl?: typeof fetch;\n  readonly allowInsecureTransport?: boolean;\n  readonly capabilities?: Partial<ProviderCapabilities>;\n  /**\n   * Time-to-response budget per request (PS-24). Default\n   * `DEFAULT_REQUEST_TIMEOUT_MS` (120s); `0` disables.\n   */\n  readonly timeoutMs?: number;\n  /**\n   * Overrides the default `acceptsSensitivity` value derived from the\n   * trust classifier. Loud users can opt out of the conservative\n   * default, but the framework still emits one WARN per process so\n   * the override is visible.\n   */\n  readonly acceptsSensitivity?: ReadonlyArray<Sensitivity>;\n  /**\n   * Optional log sink. When unset, the adapter falls back to\n   * `console.warn`. Tests pass a fixture-driven sink to silence the\n   * console.\n   */\n  readonly logger?: (level: 'warn' | 'info', message: string, meta?: object) => void;\n}\n\nconst DEFAULT_CAPABILITIES: ProviderCapabilities = {\n  streaming: true,\n  toolCalling: true,\n  parallelToolCalls: false,\n  multimodal: false,\n  structuredOutput: true,\n  reasoning: false,\n  contextWindow: 8_192,\n  maxOutput: 4_096,\n  reasoningContract: 'optional',\n};\n\nconst KNOWN_LOOPBACK_OVERRIDES_WARNED = new Set<string>();\n\n/**\n * Build a Graphorin `Provider` backed by an OpenAI-compatible HTTP\n * server. Returns the wrapped object plus the resolved trust\n * classification so the caller can attach it to span attributes if\n * desired.\n *\n * @internal\n */\nexport function buildOpenAIShapedProvider(opts: OpenAIShapedOptions): {\n  readonly provider: Provider;\n  readonly classification: LocalProviderClassification;\n} {\n  const classification = classifyLocalProvider(opts.baseUrl);\n  if (classification.trust === 'public-cleartext' && opts.allowInsecureTransport !== true) {\n    throw new LocalProviderInsecureTransportError(opts.baseUrl);\n  }\n  const log = opts.logger ?? defaultLogger;\n  if (classification.trust === 'public-cleartext') {\n    log('warn', `[${opts.providerName}] allowInsecureTransport=true accepted for ${opts.baseUrl}`, {\n      baseUrl: opts.baseUrl,\n    });\n  } else if (classification.trust === 'public-tls') {\n    log('warn', `[${opts.providerName}] public-TLS endpoint; treating as cloud-tier`, {\n      baseUrl: opts.baseUrl,\n    });\n  } else if (classification.trust === 'private') {\n    log(\n      'warn',\n      `[${opts.providerName}] private-network endpoint detected (${classification.reason})`,\n      {\n        baseUrl: opts.baseUrl,\n        acceptsSensitivity: classification.acceptsSensitivity,\n      },\n    );\n  }\n\n  const acceptsSensitivity = opts.acceptsSensitivity ?? classification.acceptsSensitivity;\n  if (\n    opts.acceptsSensitivity !== undefined &&\n    !sameSensitivity(opts.acceptsSensitivity, classification.acceptsSensitivity) &&\n    !KNOWN_LOOPBACK_OVERRIDES_WARNED.has(opts.baseUrl)\n  ) {\n    KNOWN_LOOPBACK_OVERRIDES_WARNED.add(opts.baseUrl);\n    log(\n      'info',\n      `[${opts.providerName}] sensitivity override accepted; default would be [${classification.acceptsSensitivity.join(', ')}]`,\n      {\n        baseUrl: opts.baseUrl,\n      },\n    );\n  }\n\n  const capabilities: ProviderCapabilities = {\n    ...DEFAULT_CAPABILITIES,\n    ...opts.capabilities,\n  };\n  const chatPath = opts.chatPath ?? '/v1/chat/completions';\n  const url = `${stripTrailingSlashes(opts.baseUrl)}${chatPath}`;\n\n  const provider: Provider = {\n    name: opts.providerName,\n    modelId: opts.model,\n    capabilities,\n    acceptsSensitivity,\n    stream(req) {\n      return streamOpenAIShaped(opts, url, applyOpenAIShapedPreflight(req, capabilities));\n    },\n    async generate(req) {\n      return generateOpenAIShaped(opts, url, applyOpenAIShapedPreflight(req, capabilities));\n    },\n  };\n\n  return { provider, classification };\n}\n\nfunction applyOpenAIShapedPreflight(\n  req: ProviderRequest,\n  capabilities: ProviderCapabilities,\n): ProviderRequest {\n  const retention = resolveReasoningRetention({\n    ...(req.reasoningRetention !== undefined ? { requested: req.reasoningRetention } : {}),\n    ...(capabilities.reasoningContract !== undefined\n      ? { contract: capabilities.reasoningContract }\n      : {}),\n  });\n  if (retention === 'pass-through-all') return req;\n  const filtered = applyReasoningPolicy({ messages: req.messages, retention });\n  if (filtered === req.messages) return req;\n  return { ...req, messages: filtered };\n}\n\nasync function* streamOpenAIShaped(\n  opts: OpenAIShapedOptions,\n  url: string,\n  req: ProviderRequest,\n): AsyncIterable<ProviderEvent> {\n  const body = buildBody(opts.model, req, true, opts.capabilities?.structuredOutput ?? true);\n  const resp = await callJsonHttp({\n    providerName: opts.providerName,\n    url,\n    headers: buildHeaders(opts),\n    body,\n    ...(req.signal !== undefined ? { signal: req.signal } : {}),\n    ...(opts.fetchImpl !== undefined ? { fetchImpl: opts.fetchImpl } : {}),\n    ...(opts.timeoutMs !== undefined ? { timeoutMs: opts.timeoutMs } : {}),\n  });\n\n  yield makeStreamStartEvent({ providerName: opts.providerName, modelId: opts.model });\n\n  let finishReason: FinishReason = 'stop';\n  let usage: Usage | undefined;\n  const toolCallBuffer = new Map<\n    number,\n    { id: string; name: string; args: string; emittedStart: boolean }\n  >();\n\n  for await (const payload of parseEventStream(\n    resp.body,\n    req.signal !== undefined ? { signal: req.signal } : {},\n  )) {\n    if (req.signal?.aborted) {\n      // PS-12: report the abort honestly instead of the default 'stop'.\n      finishReason = 'aborted';\n      break;\n    }\n    let parsed: OpenAIChunk;\n    try {\n      parsed = JSON.parse(payload) as OpenAIChunk;\n    } catch (cause) {\n      throw new ProviderStreamParseError(\n        opts.providerName,\n        `failed to parse SSE chunk as JSON: ${(cause as Error).message}`,\n        cause,\n      );\n    }\n    const choice = parsed.choices?.[0];\n    if (choice === undefined) {\n      if (parsed.usage !== undefined) usage = mapOpenAIUsage(parsed.usage);\n      continue;\n    }\n    const delta = choice.delta ?? {};\n    if (typeof delta.content === 'string' && delta.content.length > 0) {\n      yield { type: 'text-delta', delta: delta.content };\n    }\n    if (typeof (delta as { reasoning_content?: unknown }).reasoning_content === 'string') {\n      const r = (delta as { reasoning_content: string }).reasoning_content;\n      if (r.length > 0) yield { type: 'reasoning-delta', delta: r };\n    }\n    if (Array.isArray(delta.tool_calls)) {\n      for (const tc of delta.tool_calls) {\n        const idx = tc.index ?? 0;\n        const slot = toolCallBuffer.get(idx) ?? {\n          id: '',\n          name: '',\n          args: '',\n          emittedStart: false,\n        };\n        if (typeof tc.id === 'string' && tc.id.length > 0) slot.id = tc.id;\n        if (tc.function?.name !== undefined) slot.name = tc.function.name;\n        if (typeof tc.function?.arguments === 'string') slot.args += tc.function.arguments;\n        toolCallBuffer.set(idx, slot);\n        if (!slot.emittedStart && slot.id.length > 0 && slot.name.length > 0) {\n          yield {\n            type: 'tool-call-start',\n            toolCallId: slot.id,\n            toolName: slot.name,\n          };\n          slot.emittedStart = true;\n        }\n        if (typeof tc.function?.arguments === 'string' && tc.function.arguments.length > 0) {\n          yield {\n            type: 'tool-call-input-delta',\n            toolCallId: slot.id,\n            argsDelta: tc.function.arguments,\n          };\n        }\n      }\n    }\n    if (typeof choice.finish_reason === 'string') {\n      finishReason = mapFinishReason(choice.finish_reason);\n      for (const slot of toolCallBuffer.values()) {\n        const finalArgs = parseFinalArgs(slot.args);\n        yield {\n          type: 'tool-call-end',\n          toolCallId: slot.id,\n          finalArgs,\n        };\n      }\n    }\n    if (parsed.usage !== undefined) usage = mapOpenAIUsage(parsed.usage);\n  }\n\n  yield {\n    type: 'finish',\n    finishReason,\n    usage: usage ?? { promptTokens: 0, completionTokens: 0, totalTokens: 0 },\n  };\n}\n\nasync function generateOpenAIShaped(\n  opts: OpenAIShapedOptions,\n  url: string,\n  req: ProviderRequest,\n): Promise<ProviderResponse> {\n  const body = buildBody(opts.model, req, false, opts.capabilities?.structuredOutput ?? true);\n  const resp = await callJsonHttp({\n    providerName: opts.providerName,\n    url,\n    headers: buildHeaders(opts),\n    body,\n    ...(req.signal !== undefined ? { signal: req.signal } : {}),\n    ...(opts.fetchImpl !== undefined ? { fetchImpl: opts.fetchImpl } : {}),\n    ...(opts.timeoutMs !== undefined ? { timeoutMs: opts.timeoutMs } : {}),\n  });\n  let json: OpenAIChunk;\n  try {\n    json = (await resp.json()) as OpenAIChunk;\n  } catch (cause) {\n    throw new ProviderStreamParseError(\n      opts.providerName,\n      'response body was not valid JSON',\n      cause,\n    );\n  }\n  const choice = json.choices?.[0];\n  const message = choice?.message ?? {};\n  const text = typeof message.content === 'string' ? message.content : undefined;\n  const toolCalls = Array.isArray(message.tool_calls)\n    ? message.tool_calls.map((tc) => ({\n        toolCallId: tc.id ?? '',\n        toolName: tc.function?.name ?? '',\n        args: parseFinalArgs(\n          typeof tc.function?.arguments === 'string' ? tc.function.arguments : '',\n        ),\n      }))\n    : undefined;\n  const usage = mapOpenAIUsage(json.usage) ?? {\n    promptTokens: 0,\n    completionTokens: 0,\n    totalTokens: 0,\n  };\n  return {\n    usage,\n    finishReason: mapFinishReason(choice?.finish_reason),\n    ...(text !== undefined ? { text } : {}),\n    ...(toolCalls !== undefined ? { toolCalls } : {}),\n  };\n}\n\nfunction buildBody(\n  model: string,\n  req: ProviderRequest,\n  stream: boolean,\n  structuredOutput: boolean,\n): Record<string, unknown> {\n  const messages =\n    req.systemMessage !== undefined\n      ? [{ role: 'system' as const, content: req.systemMessage }, ...req.messages]\n      : req.messages;\n  const body: Record<string, unknown> = {\n    model,\n    messages: toOpenAIChatMessages(messages),\n    stream,\n  };\n  if (stream) {\n    // core-provider-09: vanilla OpenAI-shaped servers (vLLM, Together,\n    // OpenAI proper) only emit the final usage chunk when asked;\n    // without it streamed calls finish with {0,0,0} usage and cost\n    // tracking silently zeroes. llama.cpp accepts the flag too.\n    // `providerOptions` is merged last, so callers can override it for\n    // servers that reject the field.\n    body.stream_options = { include_usage: true };\n  }\n  if (req.temperature !== undefined) body.temperature = req.temperature;\n  if (req.maxTokens !== undefined) body.max_tokens = req.maxTokens;\n  if (req.tools !== undefined && req.tools.length > 0) {\n    // C2: fold worked examples in the adapter itself (idempotent when an\n    // upstream createProvider fold already ran).\n    body.tools = foldToolExamples(req.tools).map((t) => ({\n      type: 'function',\n      function: {\n        name: t.name,\n        ...(t.description !== undefined ? { description: t.description } : {}),\n        parameters: t.inputSchema,\n      },\n    }));\n  }\n  if (req.toolChoice !== undefined) {\n    body.tool_choice = mapToolChoice(req.toolChoice);\n  }\n  // PS-24: structured output finally reaches the wire - gated on the\n  // declared capability so a structuredOutput:false override keeps the\n  // request clean for servers that reject response_format.\n  if (structuredOutput && req.outputType?.kind === 'structured') {\n    body.response_format =\n      req.outputType.jsonSchema !== undefined\n        ? {\n            type: 'json_schema',\n            json_schema: { name: 'output', schema: req.outputType.jsonSchema, strict: true },\n          }\n        : { type: 'json_object' };\n  }\n  if (req.providerOptions !== undefined) {\n    Object.assign(body, req.providerOptions);\n  }\n  return body;\n}\n\nfunction buildHeaders(opts: OpenAIShapedOptions): Record<string, string> {\n  const headers: Record<string, string> = {\n    'content-type': 'application/json',\n    accept: 'application/json',\n    ...opts.headers,\n  };\n  if (opts.apiKey !== undefined && opts.apiKey.length > 0) {\n    headers.authorization = `Bearer ${opts.apiKey}`;\n  }\n  return headers;\n}\n\nfunction mapToolChoice(value: unknown): unknown {\n  if (value === 'auto' || value === 'none' || value === 'required') return value;\n  if (typeof value === 'object' && value !== null && 'tool' in value) {\n    return { type: 'function', function: { name: (value as { tool: string }).tool } };\n  }\n  return 'auto';\n}\n\nfunction mapFinishReason(value: string | null | undefined): FinishReason {\n  switch (value) {\n    case 'stop':\n    case 'length':\n    case 'content-filter':\n      return value;\n    // PS-12: the OpenAI wire format spells it with an underscore; the dashed\n    // form alone never matched, so content-filtered completions reported 'stop'.\n    case 'content_filter':\n      return 'content-filter';\n    case 'tool_calls':\n    case 'function_call':\n      return 'tool-calls';\n    case null:\n    case undefined:\n      return 'stop';\n    default:\n      return 'stop';\n  }\n}\n\nfunction mapOpenAIUsage(input: unknown): Usage | undefined {\n  if (input === undefined || input === null || typeof input !== 'object') return undefined;\n  const u = input as {\n    prompt_tokens?: number;\n    completion_tokens?: number;\n    total_tokens?: number;\n    reasoning_tokens?: number;\n    // OpenAI reports cache reads as a SUBSET of prompt_tokens\n    // (core-provider-02); there is no cache-write leg on this wire.\n    prompt_tokens_details?: { cached_tokens?: number };\n  };\n  const promptTokens = u.prompt_tokens ?? 0;\n  const completionTokens = u.completion_tokens ?? 0;\n  const totalTokens = u.total_tokens ?? promptTokens + completionTokens;\n  const cachedTokens = u.prompt_tokens_details?.cached_tokens;\n  const usage: Usage = {\n    promptTokens,\n    completionTokens,\n    totalTokens,\n    ...(u.reasoning_tokens !== undefined ? { reasoningTokens: u.reasoning_tokens } : {}),\n    ...(typeof cachedTokens === 'number' && Number.isFinite(cachedTokens) && cachedTokens > 0\n      ? { cachedReadTokens: cachedTokens }\n      : {}),\n  };\n  return usage;\n}\n\nfunction parseFinalArgs(raw: string): unknown {\n  if (raw.length === 0) return {};\n  try {\n    return JSON.parse(raw);\n  } catch {\n    return raw;\n  }\n}\n\nfunction sameSensitivity(a: ReadonlyArray<Sensitivity>, b: ReadonlyArray<Sensitivity>): boolean {\n  if (a.length !== b.length) return false;\n  for (let i = 0; i < a.length; i++) {\n    if (a[i] !== b[i]) return false;\n  }\n  return true;\n}\n\nfunction defaultLogger(level: 'warn' | 'info', message: string, meta?: object): void {\n  const fn = level === 'warn' ? console.warn : console.info;\n  if (meta !== undefined) {\n    fn(`[graphorin/provider] ${message}`, meta);\n  } else {\n    fn(`[graphorin/provider] ${message}`);\n  }\n}\n\ninterface OpenAIChunk {\n  readonly id?: string;\n  readonly choices?: ReadonlyArray<{\n    readonly delta?: {\n      readonly content?: string;\n      readonly tool_calls?: ReadonlyArray<{\n        readonly index?: number;\n        readonly id?: string;\n        readonly type?: string;\n        readonly function?: { readonly name?: string; readonly arguments?: string };\n      }>;\n      readonly reasoning_content?: string;\n    };\n    readonly message?: {\n      readonly content?: string | null;\n      readonly tool_calls?: ReadonlyArray<{\n        readonly id?: string;\n        readonly type?: string;\n        readonly function?: { readonly name?: string; readonly arguments?: string };\n      }>;\n    };\n    readonly finish_reason?: string | null;\n  }>;\n  readonly usage?: {\n    readonly prompt_tokens?: number;\n    readonly completion_tokens?: number;\n    readonly total_tokens?: number;\n    readonly reasoning_tokens?: number;\n    readonly prompt_tokens_details?: { readonly cached_tokens?: number };\n  };\n}\n"],"mappings":";;;;;;;;;;AAoEA,MAAMA,uBAA6C;CACjD,WAAW;CACX,aAAa;CACb,mBAAmB;CACnB,YAAY;CACZ,kBAAkB;CAClB,WAAW;CACX,eAAe;CACf,WAAW;CACX,mBAAmB;CACpB;AAED,MAAM,kDAAkC,IAAI,KAAa;;;;;;;;;AAUzD,SAAgB,0BAA0B,MAGxC;CACA,MAAM,iBAAiB,sBAAsB,KAAK,QAAQ;AAC1D,KAAI,eAAe,UAAU,sBAAsB,KAAK,2BAA2B,KACjF,OAAM,IAAI,oCAAoC,KAAK,QAAQ;CAE7D,MAAM,MAAM,KAAK,UAAU;AAC3B,KAAI,eAAe,UAAU,mBAC3B,KAAI,QAAQ,IAAI,KAAK,aAAa,6CAA6C,KAAK,WAAW,EAC7F,SAAS,KAAK,SACf,CAAC;UACO,eAAe,UAAU,aAClC,KAAI,QAAQ,IAAI,KAAK,aAAa,gDAAgD,EAChF,SAAS,KAAK,SACf,CAAC;UACO,eAAe,UAAU,UAClC,KACE,QACA,IAAI,KAAK,aAAa,uCAAuC,eAAe,OAAO,IACnF;EACE,SAAS,KAAK;EACd,oBAAoB,eAAe;EACpC,CACF;CAGH,MAAM,qBAAqB,KAAK,sBAAsB,eAAe;AACrE,KACE,KAAK,uBAAuB,UAC5B,CAAC,gBAAgB,KAAK,oBAAoB,eAAe,mBAAmB,IAC5E,CAAC,gCAAgC,IAAI,KAAK,QAAQ,EAClD;AACA,kCAAgC,IAAI,KAAK,QAAQ;AACjD,MACE,QACA,IAAI,KAAK,aAAa,qDAAqD,eAAe,mBAAmB,KAAK,KAAK,CAAC,IACxH,EACE,SAAS,KAAK,SACf,CACF;;CAGH,MAAMC,eAAqC;EACzC,GAAG;EACH,GAAG,KAAK;EACT;CACD,MAAM,WAAW,KAAK,YAAY;CAClC,MAAM,MAAM,GAAG,qBAAqB,KAAK,QAAQ,GAAG;AAepD,QAAO;EAAE,UAbkB;GACzB,MAAM,KAAK;GACX,SAAS,KAAK;GACd;GACA;GACA,OAAO,KAAK;AACV,WAAO,mBAAmB,MAAM,KAAK,2BAA2B,KAAK,aAAa,CAAC;;GAErF,MAAM,SAAS,KAAK;AAClB,WAAO,qBAAqB,MAAM,KAAK,2BAA2B,KAAK,aAAa,CAAC;;GAExF;EAEkB;EAAgB;;AAGrC,SAAS,2BACP,KACA,cACiB;CACjB,MAAM,YAAY,0BAA0B;EAC1C,GAAI,IAAI,uBAAuB,SAAY,EAAE,WAAW,IAAI,oBAAoB,GAAG,EAAE;EACrF,GAAI,aAAa,sBAAsB,SACnC,EAAE,UAAU,aAAa,mBAAmB,GAC5C,EAAE;EACP,CAAC;AACF,KAAI,cAAc,mBAAoB,QAAO;CAC7C,MAAM,WAAW,qBAAqB;EAAE,UAAU,IAAI;EAAU;EAAW,CAAC;AAC5E,KAAI,aAAa,IAAI,SAAU,QAAO;AACtC,QAAO;EAAE,GAAG;EAAK,UAAU;EAAU;;AAGvC,gBAAgB,mBACd,MACA,KACA,KAC8B;CAC9B,MAAM,OAAO,UAAU,KAAK,OAAO,KAAK,MAAM,KAAK,cAAc,oBAAoB,KAAK;CAC1F,MAAM,OAAO,MAAM,aAAa;EAC9B,cAAc,KAAK;EACnB;EACA,SAAS,aAAa,KAAK;EAC3B;EACA,GAAI,IAAI,WAAW,SAAY,EAAE,QAAQ,IAAI,QAAQ,GAAG,EAAE;EAC1D,GAAI,KAAK,cAAc,SAAY,EAAE,WAAW,KAAK,WAAW,GAAG,EAAE;EACrE,GAAI,KAAK,cAAc,SAAY,EAAE,WAAW,KAAK,WAAW,GAAG,EAAE;EACtE,CAAC;AAEF,OAAM,qBAAqB;EAAE,cAAc,KAAK;EAAc,SAAS,KAAK;EAAO,CAAC;CAEpF,IAAIC,eAA6B;CACjC,IAAIC;CACJ,MAAM,iCAAiB,IAAI,KAGxB;AAEH,YAAW,MAAM,WAAW,iBAC1B,KAAK,MACL,IAAI,WAAW,SAAY,EAAE,QAAQ,IAAI,QAAQ,GAAG,EAAE,CACvD,EAAE;AACD,MAAI,IAAI,QAAQ,SAAS;AAEvB,kBAAe;AACf;;EAEF,IAAIC;AACJ,MAAI;AACF,YAAS,KAAK,MAAM,QAAQ;WACrB,OAAO;AACd,SAAM,IAAI,yBACR,KAAK,cACL,sCAAuC,MAAgB,WACvD,MACD;;EAEH,MAAM,SAAS,OAAO,UAAU;AAChC,MAAI,WAAW,QAAW;AACxB,OAAI,OAAO,UAAU,OAAW,SAAQ,eAAe,OAAO,MAAM;AACpE;;EAEF,MAAM,QAAQ,OAAO,SAAS,EAAE;AAChC,MAAI,OAAO,MAAM,YAAY,YAAY,MAAM,QAAQ,SAAS,EAC9D,OAAM;GAAE,MAAM;GAAc,OAAO,MAAM;GAAS;AAEpD,MAAI,OAAQ,MAA0C,sBAAsB,UAAU;GACpF,MAAM,IAAK,MAAwC;AACnD,OAAI,EAAE,SAAS,EAAG,OAAM;IAAE,MAAM;IAAmB,OAAO;IAAG;;AAE/D,MAAI,MAAM,QAAQ,MAAM,WAAW,CACjC,MAAK,MAAM,MAAM,MAAM,YAAY;GACjC,MAAM,MAAM,GAAG,SAAS;GACxB,MAAM,OAAO,eAAe,IAAI,IAAI,IAAI;IACtC,IAAI;IACJ,MAAM;IACN,MAAM;IACN,cAAc;IACf;AACD,OAAI,OAAO,GAAG,OAAO,YAAY,GAAG,GAAG,SAAS,EAAG,MAAK,KAAK,GAAG;AAChE,OAAI,GAAG,UAAU,SAAS,OAAW,MAAK,OAAO,GAAG,SAAS;AAC7D,OAAI,OAAO,GAAG,UAAU,cAAc,SAAU,MAAK,QAAQ,GAAG,SAAS;AACzE,kBAAe,IAAI,KAAK,KAAK;AAC7B,OAAI,CAAC,KAAK,gBAAgB,KAAK,GAAG,SAAS,KAAK,KAAK,KAAK,SAAS,GAAG;AACpE,UAAM;KACJ,MAAM;KACN,YAAY,KAAK;KACjB,UAAU,KAAK;KAChB;AACD,SAAK,eAAe;;AAEtB,OAAI,OAAO,GAAG,UAAU,cAAc,YAAY,GAAG,SAAS,UAAU,SAAS,EAC/E,OAAM;IACJ,MAAM;IACN,YAAY,KAAK;IACjB,WAAW,GAAG,SAAS;IACxB;;AAIP,MAAI,OAAO,OAAO,kBAAkB,UAAU;AAC5C,kBAAe,gBAAgB,OAAO,cAAc;AACpD,QAAK,MAAM,QAAQ,eAAe,QAAQ,EAAE;IAC1C,MAAM,YAAY,eAAe,KAAK,KAAK;AAC3C,UAAM;KACJ,MAAM;KACN,YAAY,KAAK;KACjB;KACD;;;AAGL,MAAI,OAAO,UAAU,OAAW,SAAQ,eAAe,OAAO,MAAM;;AAGtE,OAAM;EACJ,MAAM;EACN;EACA,OAAO,SAAS;GAAE,cAAc;GAAG,kBAAkB;GAAG,aAAa;GAAG;EACzE;;AAGH,eAAe,qBACb,MACA,KACA,KAC2B;CAC3B,MAAM,OAAO,UAAU,KAAK,OAAO,KAAK,OAAO,KAAK,cAAc,oBAAoB,KAAK;CAC3F,MAAM,OAAO,MAAM,aAAa;EAC9B,cAAc,KAAK;EACnB;EACA,SAAS,aAAa,KAAK;EAC3B;EACA,GAAI,IAAI,WAAW,SAAY,EAAE,QAAQ,IAAI,QAAQ,GAAG,EAAE;EAC1D,GAAI,KAAK,cAAc,SAAY,EAAE,WAAW,KAAK,WAAW,GAAG,EAAE;EACrE,GAAI,KAAK,cAAc,SAAY,EAAE,WAAW,KAAK,WAAW,GAAG,EAAE;EACtE,CAAC;CACF,IAAIC;AACJ,KAAI;AACF,SAAQ,MAAM,KAAK,MAAM;UAClB,OAAO;AACd,QAAM,IAAI,yBACR,KAAK,cACL,oCACA,MACD;;CAEH,MAAM,SAAS,KAAK,UAAU;CAC9B,MAAM,UAAU,QAAQ,WAAW,EAAE;CACrC,MAAM,OAAO,OAAO,QAAQ,YAAY,WAAW,QAAQ,UAAU;CACrE,MAAM,YAAY,MAAM,QAAQ,QAAQ,WAAW,GAC/C,QAAQ,WAAW,KAAK,QAAQ;EAC9B,YAAY,GAAG,MAAM;EACrB,UAAU,GAAG,UAAU,QAAQ;EAC/B,MAAM,eACJ,OAAO,GAAG,UAAU,cAAc,WAAW,GAAG,SAAS,YAAY,GACtE;EACF,EAAE,GACH;AAMJ,QAAO;EACL,OANY,eAAe,KAAK,MAAM,IAAI;GAC1C,cAAc;GACd,kBAAkB;GAClB,aAAa;GACd;EAGC,cAAc,gBAAgB,QAAQ,cAAc;EACpD,GAAI,SAAS,SAAY,EAAE,MAAM,GAAG,EAAE;EACtC,GAAI,cAAc,SAAY,EAAE,WAAW,GAAG,EAAE;EACjD;;AAGH,SAAS,UACP,OACA,KACA,QACA,kBACyB;CAKzB,MAAMC,OAAgC;EACpC;EACA,UAAU,qBALV,IAAI,kBAAkB,SAClB,CAAC;GAAE,MAAM;GAAmB,SAAS,IAAI;GAAe,EAAE,GAAG,IAAI,SAAS,GAC1E,IAAI,SAGgC;EACxC;EACD;AACD,KAAI,OAOF,MAAK,iBAAiB,EAAE,eAAe,MAAM;AAE/C,KAAI,IAAI,gBAAgB,OAAW,MAAK,cAAc,IAAI;AAC1D,KAAI,IAAI,cAAc,OAAW,MAAK,aAAa,IAAI;AACvD,KAAI,IAAI,UAAU,UAAa,IAAI,MAAM,SAAS,EAGhD,MAAK,QAAQ,iBAAiB,IAAI,MAAM,CAAC,KAAK,OAAO;EACnD,MAAM;EACN,UAAU;GACR,MAAM,EAAE;GACR,GAAI,EAAE,gBAAgB,SAAY,EAAE,aAAa,EAAE,aAAa,GAAG,EAAE;GACrE,YAAY,EAAE;GACf;EACF,EAAE;AAEL,KAAI,IAAI,eAAe,OACrB,MAAK,cAAc,cAAc,IAAI,WAAW;AAKlD,KAAI,oBAAoB,IAAI,YAAY,SAAS,aAC/C,MAAK,kBACH,IAAI,WAAW,eAAe,SAC1B;EACE,MAAM;EACN,aAAa;GAAE,MAAM;GAAU,QAAQ,IAAI,WAAW;GAAY,QAAQ;GAAM;EACjF,GACD,EAAE,MAAM,eAAe;AAE/B,KAAI,IAAI,oBAAoB,OAC1B,QAAO,OAAO,MAAM,IAAI,gBAAgB;AAE1C,QAAO;;AAGT,SAAS,aAAa,MAAmD;CACvE,MAAMC,UAAkC;EACtC,gBAAgB;EAChB,QAAQ;EACR,GAAG,KAAK;EACT;AACD,KAAI,KAAK,WAAW,UAAa,KAAK,OAAO,SAAS,EACpD,SAAQ,gBAAgB,UAAU,KAAK;AAEzC,QAAO;;AAGT,SAAS,cAAc,OAAyB;AAC9C,KAAI,UAAU,UAAU,UAAU,UAAU,UAAU,WAAY,QAAO;AACzE,KAAI,OAAO,UAAU,YAAY,UAAU,QAAQ,UAAU,MAC3D,QAAO;EAAE,MAAM;EAAY,UAAU,EAAE,MAAO,MAA2B,MAAM;EAAE;AAEnF,QAAO;;AAGT,SAAS,gBAAgB,OAAgD;AACvE,SAAQ,OAAR;EACE,KAAK;EACL,KAAK;EACL,KAAK,iBACH,QAAO;EAGT,KAAK,iBACH,QAAO;EACT,KAAK;EACL,KAAK,gBACH,QAAO;EACT,KAAK;EACL,KAAK,OACH,QAAO;EACT,QACE,QAAO;;;AAIb,SAAS,eAAe,OAAmC;AACzD,KAAI,UAAU,UAAa,UAAU,QAAQ,OAAO,UAAU,SAAU,QAAO;CAC/E,MAAM,IAAI;CASV,MAAM,eAAe,EAAE,iBAAiB;CACxC,MAAM,mBAAmB,EAAE,qBAAqB;CAChD,MAAM,cAAc,EAAE,gBAAgB,eAAe;CACrD,MAAM,eAAe,EAAE,uBAAuB;AAU9C,QATqB;EACnB;EACA;EACA;EACA,GAAI,EAAE,qBAAqB,SAAY,EAAE,iBAAiB,EAAE,kBAAkB,GAAG,EAAE;EACnF,GAAI,OAAO,iBAAiB,YAAY,OAAO,SAAS,aAAa,IAAI,eAAe,IACpF,EAAE,kBAAkB,cAAc,GAClC,EAAE;EACP;;AAIH,SAAS,eAAe,KAAsB;AAC5C,KAAI,IAAI,WAAW,EAAG,QAAO,EAAE;AAC/B,KAAI;AACF,SAAO,KAAK,MAAM,IAAI;SAChB;AACN,SAAO;;;AAIX,SAAS,gBAAgB,GAA+B,GAAwC;AAC9F,KAAI,EAAE,WAAW,EAAE,OAAQ,QAAO;AAClC,MAAK,IAAI,IAAI,GAAG,IAAI,EAAE,QAAQ,IAC5B,KAAI,EAAE,OAAO,EAAE,GAAI,QAAO;AAE5B,QAAO;;AAGT,SAAS,cAAc,OAAwB,SAAiB,MAAqB;CACnF,MAAM,KAAK,UAAU,SAAS,QAAQ,OAAO,QAAQ;AACrD,KAAI,SAAS,OACX,IAAG,wBAAwB,WAAW,KAAK;KAE3C,IAAG,wBAAwB,UAAU"}
+{"version":3,"file":"openai-shaped.js","names":["DEFAULT_CAPABILITIES: ProviderCapabilities","capabilities: ProviderCapabilities","finishReason: FinishReason","usage: Usage | undefined","parsed: OpenAIChunk","json: OpenAIChunk","body: Record<string, unknown>","headers: Record<string, string>"],"sources":["../../src/internal/openai-shaped.ts"],"sourcesContent":["/**\n * Shared implementation backing the two OpenAI-compatible adapters\n * (`llamaCppServerAdapter` and `openAICompatibleAdapter`). The wire\n * shape is identical for the `/v1/chat/completions` endpoint, so we\n * factor the streaming + one-shot logic here and the adapters become\n * thin shells declaring trust class, defaults, and capability flags.\n *\n * @internal\n */\n\nimport type {\n  FinishReason,\n  Provider,\n  ProviderCapabilities,\n  ProviderEvent,\n  ProviderRequest,\n  ProviderResponse,\n  Sensitivity,\n  Usage,\n} from '@graphorin/core';\n\nimport { LocalProviderInsecureTransportError, ProviderStreamParseError } from '../errors/errors.js';\nimport { applyReasoningPolicy } from '../reasoning/apply-policy.js';\nimport { resolveReasoningRetention } from '../reasoning/retention.js';\nimport { foldToolExamples } from '../tool-examples.js';\nimport {\n  classifyLocalProvider,\n  type LocalProviderClassification,\n} from '../trust/classify-local-provider.js';\nimport {\n  type ChatMessageConversionOptions,\n  callJsonHttp,\n  makeStreamStartEvent,\n  toOpenAIChatMessages,\n} from './http.js';\nimport { parseEventStream } from './sse.js';\nimport { stripTrailingSlashes } from './url-utils.js';\n\n/**\n * Options shared by every OpenAI-compatible adapter.\n *\n * @internal\n */\nexport interface OpenAIShapedOptions {\n  readonly providerName: string;\n  readonly model: string;\n  readonly baseUrl: string;\n  readonly chatPath?: string;\n  readonly apiKey?: string;\n  readonly headers?: Readonly<Record<string, string>>;\n  readonly fetchImpl?: typeof fetch;\n  readonly allowInsecureTransport?: boolean;\n  readonly capabilities?: Partial<ProviderCapabilities>;\n  /**\n   * Time-to-response budget per request (PS-24). Default\n   * `DEFAULT_REQUEST_TIMEOUT_MS` (120s); `0` disables.\n   */\n  readonly timeoutMs?: number;\n  /**\n   * Overrides the default `acceptsSensitivity` value derived from the\n   * trust classifier. Loud users can opt out of the conservative\n   * default, but the framework still emits one WARN per process so\n   * the override is visible.\n   */\n  readonly acceptsSensitivity?: ReadonlyArray<Sensitivity>;\n  /**\n   * Optional log sink. When unset, the adapter falls back to\n   * `console.warn`. Tests pass a fixture-driven sink to silence the\n   * console.\n   */\n  readonly logger?: (level: 'warn' | 'info', message: string, meta?: object) => void;\n}\n\nconst DEFAULT_CAPABILITIES: ProviderCapabilities = {\n  streaming: true,\n  toolCalling: true,\n  parallelToolCalls: false,\n  multimodal: false,\n  structuredOutput: true,\n  reasoning: false,\n  contextWindow: 8_192,\n  maxOutput: 4_096,\n  reasoningContract: 'optional',\n};\n\nconst KNOWN_LOOPBACK_OVERRIDES_WARNED = new Set<string>();\n\n/**\n * Build a Graphorin `Provider` backed by an OpenAI-compatible HTTP\n * server. Returns the wrapped object plus the resolved trust\n * classification so the caller can attach it to span attributes if\n * desired.\n *\n * @internal\n */\nexport function buildOpenAIShapedProvider(opts: OpenAIShapedOptions): {\n  readonly provider: Provider;\n  readonly classification: LocalProviderClassification;\n} {\n  const classification = classifyLocalProvider(opts.baseUrl);\n  if (classification.trust === 'public-cleartext' && opts.allowInsecureTransport !== true) {\n    throw new LocalProviderInsecureTransportError(opts.baseUrl);\n  }\n  const log = opts.logger ?? defaultLogger;\n  if (classification.trust === 'public-cleartext') {\n    log('warn', `[${opts.providerName}] allowInsecureTransport=true accepted for ${opts.baseUrl}`, {\n      baseUrl: opts.baseUrl,\n    });\n  } else if (classification.trust === 'public-tls') {\n    log('warn', `[${opts.providerName}] public-TLS endpoint; treating as cloud-tier`, {\n      baseUrl: opts.baseUrl,\n    });\n  } else if (classification.trust === 'private') {\n    log(\n      'warn',\n      `[${opts.providerName}] private-network endpoint detected (${classification.reason})`,\n      {\n        baseUrl: opts.baseUrl,\n        acceptsSensitivity: classification.acceptsSensitivity,\n      },\n    );\n  }\n\n  const acceptsSensitivity = opts.acceptsSensitivity ?? classification.acceptsSensitivity;\n  if (\n    opts.acceptsSensitivity !== undefined &&\n    !sameSensitivity(opts.acceptsSensitivity, classification.acceptsSensitivity) &&\n    !KNOWN_LOOPBACK_OVERRIDES_WARNED.has(opts.baseUrl)\n  ) {\n    KNOWN_LOOPBACK_OVERRIDES_WARNED.add(opts.baseUrl);\n    log(\n      'info',\n      `[${opts.providerName}] sensitivity override accepted; default would be [${classification.acceptsSensitivity.join(', ')}]`,\n      {\n        baseUrl: opts.baseUrl,\n      },\n    );\n  }\n\n  const capabilities: ProviderCapabilities = {\n    ...DEFAULT_CAPABILITIES,\n    ...opts.capabilities,\n  };\n  const chatPath = opts.chatPath ?? '/v1/chat/completions';\n  const url = `${stripTrailingSlashes(opts.baseUrl)}${chatPath}`;\n\n  const provider: Provider = {\n    name: opts.providerName,\n    modelId: opts.model,\n    capabilities,\n    acceptsSensitivity,\n    stream(req) {\n      return streamOpenAIShaped(opts, url, applyOpenAIShapedPreflight(req, capabilities));\n    },\n    async generate(req) {\n      return generateOpenAIShaped(opts, url, applyOpenAIShapedPreflight(req, capabilities));\n    },\n  };\n\n  return { provider, classification };\n}\n\nfunction applyOpenAIShapedPreflight(\n  req: ProviderRequest,\n  capabilities: ProviderCapabilities,\n): ProviderRequest {\n  const retention = resolveReasoningRetention({\n    ...(req.reasoningRetention !== undefined ? { requested: req.reasoningRetention } : {}),\n    ...(capabilities.reasoningContract !== undefined\n      ? { contract: capabilities.reasoningContract }\n      : {}),\n  });\n  if (retention === 'pass-through-all') return req;\n  const filtered = applyReasoningPolicy({ messages: req.messages, retention });\n  if (filtered === req.messages) return req;\n  return { ...req, messages: filtered };\n}\n\nasync function* streamOpenAIShaped(\n  opts: OpenAIShapedOptions,\n  url: string,\n  req: ProviderRequest,\n): AsyncIterable<ProviderEvent> {\n  const body = buildBody(\n    opts.model,\n    req,\n    true,\n    opts.capabilities?.structuredOutput ?? true,\n    conversionOptionsFor(opts),\n  );\n  const resp = await callJsonHttp({\n    providerName: opts.providerName,\n    url,\n    headers: buildHeaders(opts),\n    body,\n    ...(req.signal !== undefined ? { signal: req.signal } : {}),\n    ...(opts.fetchImpl !== undefined ? { fetchImpl: opts.fetchImpl } : {}),\n    ...(opts.timeoutMs !== undefined ? { timeoutMs: opts.timeoutMs } : {}),\n  });\n\n  yield makeStreamStartEvent({ providerName: opts.providerName, modelId: opts.model });\n\n  let finishReason: FinishReason = 'stop';\n  let usage: Usage | undefined;\n  const toolCallBuffer = new Map<\n    number,\n    { id: string; name: string; args: string; emittedStart: boolean }\n  >();\n\n  for await (const payload of parseEventStream(\n    resp.body,\n    req.signal !== undefined ? { signal: req.signal } : {},\n  )) {\n    if (req.signal?.aborted) {\n      // PS-12: report the abort honestly instead of the default 'stop'.\n      finishReason = 'aborted';\n      break;\n    }\n    let parsed: OpenAIChunk;\n    try {\n      parsed = JSON.parse(payload) as OpenAIChunk;\n    } catch (cause) {\n      throw new ProviderStreamParseError(\n        opts.providerName,\n        `failed to parse SSE chunk as JSON: ${(cause as Error).message}`,\n        cause,\n      );\n    }\n    const choice = parsed.choices?.[0];\n    if (choice === undefined) {\n      if (parsed.usage !== undefined) usage = mapOpenAIUsage(parsed.usage);\n      continue;\n    }\n    const delta = choice.delta ?? {};\n    if (typeof delta.content === 'string' && delta.content.length > 0) {\n      yield { type: 'text-delta', delta: delta.content };\n    }\n    if (typeof (delta as { reasoning_content?: unknown }).reasoning_content === 'string') {\n      const r = (delta as { reasoning_content: string }).reasoning_content;\n      if (r.length > 0) yield { type: 'reasoning-delta', delta: r };\n    }\n    if (Array.isArray(delta.tool_calls)) {\n      for (const tc of delta.tool_calls) {\n        const idx = tc.index ?? 0;\n        const slot = toolCallBuffer.get(idx) ?? {\n          id: '',\n          name: '',\n          args: '',\n          emittedStart: false,\n        };\n        if (typeof tc.id === 'string' && tc.id.length > 0) slot.id = tc.id;\n        if (tc.function?.name !== undefined) slot.name = tc.function.name;\n        if (typeof tc.function?.arguments === 'string') slot.args += tc.function.arguments;\n        toolCallBuffer.set(idx, slot);\n        if (!slot.emittedStart && slot.id.length > 0 && slot.name.length > 0) {\n          yield {\n            type: 'tool-call-start',\n            toolCallId: slot.id,\n            toolName: slot.name,\n          };\n          slot.emittedStart = true;\n        }\n        if (typeof tc.function?.arguments === 'string' && tc.function.arguments.length > 0) {\n          yield {\n            type: 'tool-call-input-delta',\n            toolCallId: slot.id,\n            argsDelta: tc.function.arguments,\n          };\n        }\n      }\n    }\n    if (typeof choice.finish_reason === 'string') {\n      finishReason = mapFinishReason(choice.finish_reason);\n      for (const slot of toolCallBuffer.values()) {\n        const finalArgs = parseFinalArgs(slot.args);\n        yield {\n          type: 'tool-call-end',\n          toolCallId: slot.id,\n          finalArgs,\n        };\n      }\n    }\n    if (parsed.usage !== undefined) usage = mapOpenAIUsage(parsed.usage);\n  }\n\n  yield {\n    type: 'finish',\n    finishReason,\n    usage: usage ?? { promptTokens: 0, completionTokens: 0, totalTokens: 0 },\n  };\n}\n\nasync function generateOpenAIShaped(\n  opts: OpenAIShapedOptions,\n  url: string,\n  req: ProviderRequest,\n): Promise<ProviderResponse> {\n  const body = buildBody(\n    opts.model,\n    req,\n    false,\n    opts.capabilities?.structuredOutput ?? true,\n    conversionOptionsFor(opts),\n  );\n  const resp = await callJsonHttp({\n    providerName: opts.providerName,\n    url,\n    headers: buildHeaders(opts),\n    body,\n    ...(req.signal !== undefined ? { signal: req.signal } : {}),\n    ...(opts.fetchImpl !== undefined ? { fetchImpl: opts.fetchImpl } : {}),\n    ...(opts.timeoutMs !== undefined ? { timeoutMs: opts.timeoutMs } : {}),\n  });\n  let json: OpenAIChunk;\n  try {\n    json = (await resp.json()) as OpenAIChunk;\n  } catch (cause) {\n    throw new ProviderStreamParseError(\n      opts.providerName,\n      'response body was not valid JSON',\n      cause,\n    );\n  }\n  const choice = json.choices?.[0];\n  const message = choice?.message ?? {};\n  const text = typeof message.content === 'string' ? message.content : undefined;\n  const toolCalls = Array.isArray(message.tool_calls)\n    ? message.tool_calls.map((tc) => ({\n        toolCallId: tc.id ?? '',\n        toolName: tc.function?.name ?? '',\n        args: parseFinalArgs(\n          typeof tc.function?.arguments === 'string' ? tc.function.arguments : '',\n        ),\n      }))\n    : undefined;\n  const usage = mapOpenAIUsage(json.usage) ?? {\n    promptTokens: 0,\n    completionTokens: 0,\n    totalTokens: 0,\n  };\n  return {\n    usage,\n    finishReason: mapFinishReason(choice?.finish_reason),\n    ...(text !== undefined ? { text } : {}),\n    ...(toolCalls !== undefined ? { toolCalls } : {}),\n  };\n}\n\nfunction buildBody(\n  model: string,\n  req: ProviderRequest,\n  stream: boolean,\n  structuredOutput: boolean,\n  conversion: ChatMessageConversionOptions,\n): Record<string, unknown> {\n  const messages =\n    req.systemMessage !== undefined\n      ? [{ role: 'system' as const, content: req.systemMessage }, ...req.messages]\n      : req.messages;\n  const body: Record<string, unknown> = {\n    model,\n    messages: toOpenAIChatMessages(messages, conversion),\n    stream,\n  };\n  if (stream) {\n    // core-provider-09: vanilla OpenAI-shaped servers (vLLM, Together,\n    // OpenAI proper) only emit the final usage chunk when asked;\n    // without it streamed calls finish with {0,0,0} usage and cost\n    // tracking silently zeroes. llama.cpp accepts the flag too.\n    // `providerOptions` is merged last, so callers can override it for\n    // servers that reject the field.\n    body.stream_options = { include_usage: true };\n  }\n  if (req.temperature !== undefined) body.temperature = req.temperature;\n  if (req.maxTokens !== undefined) body.max_tokens = req.maxTokens;\n  if (req.tools !== undefined && req.tools.length > 0) {\n    // C2: fold worked examples in the adapter itself (idempotent when an\n    // upstream createProvider fold already ran).\n    body.tools = foldToolExamples(req.tools).map((t) => ({\n      type: 'function',\n      function: {\n        name: t.name,\n        ...(t.description !== undefined ? { description: t.description } : {}),\n        parameters: t.inputSchema,\n      },\n    }));\n  }\n  if (req.toolChoice !== undefined) {\n    body.tool_choice = mapToolChoice(req.toolChoice);\n  }\n  // PS-24: structured output finally reaches the wire - gated on the\n  // declared capability so a structuredOutput:false override keeps the\n  // request clean for servers that reject response_format.\n  if (structuredOutput && req.outputType?.kind === 'structured') {\n    body.response_format =\n      req.outputType.jsonSchema !== undefined\n        ? {\n            type: 'json_schema',\n            json_schema: { name: 'output', schema: req.outputType.jsonSchema, strict: true },\n          }\n        : { type: 'json_object' };\n  }\n  if (req.providerOptions !== undefined) {\n    Object.assign(body, req.providerOptions);\n  }\n  return body;\n}\n\nfunction buildHeaders(opts: OpenAIShapedOptions): Record<string, string> {\n  const headers: Record<string, string> = {\n    'content-type': 'application/json',\n    accept: 'application/json',\n    ...opts.headers,\n  };\n  if (opts.apiKey !== undefined && opts.apiKey.length > 0) {\n    headers.authorization = `Bearer ${opts.apiKey}`;\n  }\n  return headers;\n}\n\nfunction mapToolChoice(value: unknown): unknown {\n  if (value === 'auto' || value === 'none' || value === 'required') return value;\n  if (typeof value === 'object' && value !== null && 'tool' in value) {\n    return { type: 'function', function: { name: (value as { tool: string }).tool } };\n  }\n  return 'auto';\n}\n\nfunction mapFinishReason(value: string | null | undefined): FinishReason {\n  switch (value) {\n    case 'stop':\n    case 'length':\n    case 'content-filter':\n      return value;\n    // PS-12: the OpenAI wire format spells it with an underscore; the dashed\n    // form alone never matched, so content-filtered completions reported 'stop'.\n    case 'content_filter':\n      return 'content-filter';\n    case 'tool_calls':\n    case 'function_call':\n      return 'tool-calls';\n    case null:\n    case undefined:\n      return 'stop';\n    default:\n      return 'stop';\n  }\n}\n\nfunction mapOpenAIUsage(input: unknown): Usage | undefined {\n  if (input === undefined || input === null || typeof input !== 'object') return undefined;\n  const u = input as {\n    prompt_tokens?: number;\n    completion_tokens?: number;\n    total_tokens?: number;\n    reasoning_tokens?: number;\n    // OpenAI reports cache reads as a SUBSET of prompt_tokens\n    // (core-provider-02); there is no cache-write leg on this wire.\n    prompt_tokens_details?: { cached_tokens?: number };\n  };\n  const promptTokens = u.prompt_tokens ?? 0;\n  const completionTokens = u.completion_tokens ?? 0;\n  const totalTokens = u.total_tokens ?? promptTokens + completionTokens;\n  const cachedTokens = u.prompt_tokens_details?.cached_tokens;\n  const usage: Usage = {\n    promptTokens,\n    completionTokens,\n    totalTokens,\n    ...(u.reasoning_tokens !== undefined ? { reasoningTokens: u.reasoning_tokens } : {}),\n    ...(typeof cachedTokens === 'number' && Number.isFinite(cachedTokens) && cachedTokens > 0\n      ? { cachedReadTokens: cachedTokens }\n      : {}),\n  };\n  return usage;\n}\n\nfunction parseFinalArgs(raw: string): unknown {\n  if (raw.length === 0) return {};\n  try {\n    return JSON.parse(raw);\n  } catch {\n    return raw;\n  }\n}\n\nfunction sameSensitivity(a: ReadonlyArray<Sensitivity>, b: ReadonlyArray<Sensitivity>): boolean {\n  if (a.length !== b.length) return false;\n  for (let i = 0; i < a.length; i++) {\n    if (a[i] !== b[i]) return false;\n  }\n  return true;\n}\n\nfunction defaultLogger(level: 'warn' | 'info', message: string, meta?: object): void {\n  const fn = level === 'warn' ? console.warn : console.info;\n  if (meta !== undefined) {\n    fn(`[graphorin/provider] ${message}`, meta);\n  } else {\n    fn(`[graphorin/provider] ${message}`);\n  }\n}\n\n/**\n * W-095: one dropped-content WARN per adapter INSTANCE (keyed on the\n * factory's options object, following KNOWN_LOOPBACK_OVERRIDES_WARNED),\n * so a chat loop does not repeat it every call.\n */\nconst droppedContentWarned = new WeakSet<object>();\n\nfunction conversionOptionsFor(opts: {\n  readonly capabilities?: { readonly multimodal?: boolean };\n  readonly logger?: (level: 'warn' | 'info', message: string, meta?: object) => void;\n}): ChatMessageConversionOptions {\n  const log = opts.logger ?? defaultLogger;\n  return {\n    multimodal: opts.capabilities?.multimodal ?? false,\n    warn: (message) => {\n      if (droppedContentWarned.has(opts)) return;\n      droppedContentWarned.add(opts);\n      log('warn', message);\n    },\n  };\n}\n\ninterface OpenAIChunk {\n  readonly id?: string;\n  readonly choices?: ReadonlyArray<{\n    readonly delta?: {\n      readonly content?: string;\n      readonly tool_calls?: ReadonlyArray<{\n        readonly index?: number;\n        readonly id?: string;\n        readonly type?: string;\n        readonly function?: { readonly name?: string; readonly arguments?: string };\n      }>;\n      readonly reasoning_content?: string;\n    };\n    readonly message?: {\n      readonly content?: string | null;\n      readonly tool_calls?: ReadonlyArray<{\n        readonly id?: string;\n        readonly type?: string;\n        readonly function?: { readonly name?: string; readonly arguments?: string };\n      }>;\n    };\n    readonly finish_reason?: string | null;\n  }>;\n  readonly usage?: {\n    readonly prompt_tokens?: number;\n    readonly completion_tokens?: number;\n    readonly total_tokens?: number;\n    readonly reasoning_tokens?: number;\n    readonly prompt_tokens_details?: { readonly cached_tokens?: number };\n  };\n}\n"],"mappings":";;;;;;;;;;AAyEA,MAAMA,uBAA6C;CACjD,WAAW;CACX,aAAa;CACb,mBAAmB;CACnB,YAAY;CACZ,kBAAkB;CAClB,WAAW;CACX,eAAe;CACf,WAAW;CACX,mBAAmB;CACpB;AAED,MAAM,kDAAkC,IAAI,KAAa;;;;;;;;;AAUzD,SAAgB,0BAA0B,MAGxC;CACA,MAAM,iBAAiB,sBAAsB,KAAK,QAAQ;AAC1D,KAAI,eAAe,UAAU,sBAAsB,KAAK,2BAA2B,KACjF,OAAM,IAAI,oCAAoC,KAAK,QAAQ;CAE7D,MAAM,MAAM,KAAK,UAAU;AAC3B,KAAI,eAAe,UAAU,mBAC3B,KAAI,QAAQ,IAAI,KAAK,aAAa,6CAA6C,KAAK,WAAW,EAC7F,SAAS,KAAK,SACf,CAAC;UACO,eAAe,UAAU,aAClC,KAAI,QAAQ,IAAI,KAAK,aAAa,gDAAgD,EAChF,SAAS,KAAK,SACf,CAAC;UACO,eAAe,UAAU,UAClC,KACE,QACA,IAAI,KAAK,aAAa,uCAAuC,eAAe,OAAO,IACnF;EACE,SAAS,KAAK;EACd,oBAAoB,eAAe;EACpC,CACF;CAGH,MAAM,qBAAqB,KAAK,sBAAsB,eAAe;AACrE,KACE,KAAK,uBAAuB,UAC5B,CAAC,gBAAgB,KAAK,oBAAoB,eAAe,mBAAmB,IAC5E,CAAC,gCAAgC,IAAI,KAAK,QAAQ,EAClD;AACA,kCAAgC,IAAI,KAAK,QAAQ;AACjD,MACE,QACA,IAAI,KAAK,aAAa,qDAAqD,eAAe,mBAAmB,KAAK,KAAK,CAAC,IACxH,EACE,SAAS,KAAK,SACf,CACF;;CAGH,MAAMC,eAAqC;EACzC,GAAG;EACH,GAAG,KAAK;EACT;CACD,MAAM,WAAW,KAAK,YAAY;CAClC,MAAM,MAAM,GAAG,qBAAqB,KAAK,QAAQ,GAAG;AAepD,QAAO;EAAE,UAbkB;GACzB,MAAM,KAAK;GACX,SAAS,KAAK;GACd;GACA;GACA,OAAO,KAAK;AACV,WAAO,mBAAmB,MAAM,KAAK,2BAA2B,KAAK,aAAa,CAAC;;GAErF,MAAM,SAAS,KAAK;AAClB,WAAO,qBAAqB,MAAM,KAAK,2BAA2B,KAAK,aAAa,CAAC;;GAExF;EAEkB;EAAgB;;AAGrC,SAAS,2BACP,KACA,cACiB;CACjB,MAAM,YAAY,0BAA0B;EAC1C,GAAI,IAAI,uBAAuB,SAAY,EAAE,WAAW,IAAI,oBAAoB,GAAG,EAAE;EACrF,GAAI,aAAa,sBAAsB,SACnC,EAAE,UAAU,aAAa,mBAAmB,GAC5C,EAAE;EACP,CAAC;AACF,KAAI,cAAc,mBAAoB,QAAO;CAC7C,MAAM,WAAW,qBAAqB;EAAE,UAAU,IAAI;EAAU;EAAW,CAAC;AAC5E,KAAI,aAAa,IAAI,SAAU,QAAO;AACtC,QAAO;EAAE,GAAG;EAAK,UAAU;EAAU;;AAGvC,gBAAgB,mBACd,MACA,KACA,KAC8B;CAC9B,MAAM,OAAO,UACX,KAAK,OACL,KACA,MACA,KAAK,cAAc,oBAAoB,MACvC,qBAAqB,KAAK,CAC3B;CACD,MAAM,OAAO,MAAM,aAAa;EAC9B,cAAc,KAAK;EACnB;EACA,SAAS,aAAa,KAAK;EAC3B;EACA,GAAI,IAAI,WAAW,SAAY,EAAE,QAAQ,IAAI,QAAQ,GAAG,EAAE;EAC1D,GAAI,KAAK,cAAc,SAAY,EAAE,WAAW,KAAK,WAAW,GAAG,EAAE;EACrE,GAAI,KAAK,cAAc,SAAY,EAAE,WAAW,KAAK,WAAW,GAAG,EAAE;EACtE,CAAC;AAEF,OAAM,qBAAqB;EAAE,cAAc,KAAK;EAAc,SAAS,KAAK;EAAO,CAAC;CAEpF,IAAIC,eAA6B;CACjC,IAAIC;CACJ,MAAM,iCAAiB,IAAI,KAGxB;AAEH,YAAW,MAAM,WAAW,iBAC1B,KAAK,MACL,IAAI,WAAW,SAAY,EAAE,QAAQ,IAAI,QAAQ,GAAG,EAAE,CACvD,EAAE;AACD,MAAI,IAAI,QAAQ,SAAS;AAEvB,kBAAe;AACf;;EAEF,IAAIC;AACJ,MAAI;AACF,YAAS,KAAK,MAAM,QAAQ;WACrB,OAAO;AACd,SAAM,IAAI,yBACR,KAAK,cACL,sCAAuC,MAAgB,WACvD,MACD;;EAEH,MAAM,SAAS,OAAO,UAAU;AAChC,MAAI,WAAW,QAAW;AACxB,OAAI,OAAO,UAAU,OAAW,SAAQ,eAAe,OAAO,MAAM;AACpE;;EAEF,MAAM,QAAQ,OAAO,SAAS,EAAE;AAChC,MAAI,OAAO,MAAM,YAAY,YAAY,MAAM,QAAQ,SAAS,EAC9D,OAAM;GAAE,MAAM;GAAc,OAAO,MAAM;GAAS;AAEpD,MAAI,OAAQ,MAA0C,sBAAsB,UAAU;GACpF,MAAM,IAAK,MAAwC;AACnD,OAAI,EAAE,SAAS,EAAG,OAAM;IAAE,MAAM;IAAmB,OAAO;IAAG;;AAE/D,MAAI,MAAM,QAAQ,MAAM,WAAW,CACjC,MAAK,MAAM,MAAM,MAAM,YAAY;GACjC,MAAM,MAAM,GAAG,SAAS;GACxB,MAAM,OAAO,eAAe,IAAI,IAAI,IAAI;IACtC,IAAI;IACJ,MAAM;IACN,MAAM;IACN,cAAc;IACf;AACD,OAAI,OAAO,GAAG,OAAO,YAAY,GAAG,GAAG,SAAS,EAAG,MAAK,KAAK,GAAG;AAChE,OAAI,GAAG,UAAU,SAAS,OAAW,MAAK,OAAO,GAAG,SAAS;AAC7D,OAAI,OAAO,GAAG,UAAU,cAAc,SAAU,MAAK,QAAQ,GAAG,SAAS;AACzE,kBAAe,IAAI,KAAK,KAAK;AAC7B,OAAI,CAAC,KAAK,gBAAgB,KAAK,GAAG,SAAS,KAAK,KAAK,KAAK,SAAS,GAAG;AACpE,UAAM;KACJ,MAAM;KACN,YAAY,KAAK;KACjB,UAAU,KAAK;KAChB;AACD,SAAK,eAAe;;AAEtB,OAAI,OAAO,GAAG,UAAU,cAAc,YAAY,GAAG,SAAS,UAAU,SAAS,EAC/E,OAAM;IACJ,MAAM;IACN,YAAY,KAAK;IACjB,WAAW,GAAG,SAAS;IACxB;;AAIP,MAAI,OAAO,OAAO,kBAAkB,UAAU;AAC5C,kBAAe,gBAAgB,OAAO,cAAc;AACpD,QAAK,MAAM,QAAQ,eAAe,QAAQ,EAAE;IAC1C,MAAM,YAAY,eAAe,KAAK,KAAK;AAC3C,UAAM;KACJ,MAAM;KACN,YAAY,KAAK;KACjB;KACD;;;AAGL,MAAI,OAAO,UAAU,OAAW,SAAQ,eAAe,OAAO,MAAM;;AAGtE,OAAM;EACJ,MAAM;EACN;EACA,OAAO,SAAS;GAAE,cAAc;GAAG,kBAAkB;GAAG,aAAa;GAAG;EACzE;;AAGH,eAAe,qBACb,MACA,KACA,KAC2B;CAC3B,MAAM,OAAO,UACX,KAAK,OACL,KACA,OACA,KAAK,cAAc,oBAAoB,MACvC,qBAAqB,KAAK,CAC3B;CACD,MAAM,OAAO,MAAM,aAAa;EAC9B,cAAc,KAAK;EACnB;EACA,SAAS,aAAa,KAAK;EAC3B;EACA,GAAI,IAAI,WAAW,SAAY,EAAE,QAAQ,IAAI,QAAQ,GAAG,EAAE;EAC1D,GAAI,KAAK,cAAc,SAAY,EAAE,WAAW,KAAK,WAAW,GAAG,EAAE;EACrE,GAAI,KAAK,cAAc,SAAY,EAAE,WAAW,KAAK,WAAW,GAAG,EAAE;EACtE,CAAC;CACF,IAAIC;AACJ,KAAI;AACF,SAAQ,MAAM,KAAK,MAAM;UAClB,OAAO;AACd,QAAM,IAAI,yBACR,KAAK,cACL,oCACA,MACD;;CAEH,MAAM,SAAS,KAAK,UAAU;CAC9B,MAAM,UAAU,QAAQ,WAAW,EAAE;CACrC,MAAM,OAAO,OAAO,QAAQ,YAAY,WAAW,QAAQ,UAAU;CACrE,MAAM,YAAY,MAAM,QAAQ,QAAQ,WAAW,GAC/C,QAAQ,WAAW,KAAK,QAAQ;EAC9B,YAAY,GAAG,MAAM;EACrB,UAAU,GAAG,UAAU,QAAQ;EAC/B,MAAM,eACJ,OAAO,GAAG,UAAU,cAAc,WAAW,GAAG,SAAS,YAAY,GACtE;EACF,EAAE,GACH;AAMJ,QAAO;EACL,OANY,eAAe,KAAK,MAAM,IAAI;GAC1C,cAAc;GACd,kBAAkB;GAClB,aAAa;GACd;EAGC,cAAc,gBAAgB,QAAQ,cAAc;EACpD,GAAI,SAAS,SAAY,EAAE,MAAM,GAAG,EAAE;EACtC,GAAI,cAAc,SAAY,EAAE,WAAW,GAAG,EAAE;EACjD;;AAGH,SAAS,UACP,OACA,KACA,QACA,kBACA,YACyB;CAKzB,MAAMC,OAAgC;EACpC;EACA,UAAU,qBALV,IAAI,kBAAkB,SAClB,CAAC;GAAE,MAAM;GAAmB,SAAS,IAAI;GAAe,EAAE,GAAG,IAAI,SAAS,GAC1E,IAAI,UAGiC,WAAW;EACpD;EACD;AACD,KAAI,OAOF,MAAK,iBAAiB,EAAE,eAAe,MAAM;AAE/C,KAAI,IAAI,gBAAgB,OAAW,MAAK,cAAc,IAAI;AAC1D,KAAI,IAAI,cAAc,OAAW,MAAK,aAAa,IAAI;AACvD,KAAI,IAAI,UAAU,UAAa,IAAI,MAAM,SAAS,EAGhD,MAAK,QAAQ,iBAAiB,IAAI,MAAM,CAAC,KAAK,OAAO;EACnD,MAAM;EACN,UAAU;GACR,MAAM,EAAE;GACR,GAAI,EAAE,gBAAgB,SAAY,EAAE,aAAa,EAAE,aAAa,GAAG,EAAE;GACrE,YAAY,EAAE;GACf;EACF,EAAE;AAEL,KAAI,IAAI,eAAe,OACrB,MAAK,cAAc,cAAc,IAAI,WAAW;AAKlD,KAAI,oBAAoB,IAAI,YAAY,SAAS,aAC/C,MAAK,kBACH,IAAI,WAAW,eAAe,SAC1B;EACE,MAAM;EACN,aAAa;GAAE,MAAM;GAAU,QAAQ,IAAI,WAAW;GAAY,QAAQ;GAAM;EACjF,GACD,EAAE,MAAM,eAAe;AAE/B,KAAI,IAAI,oBAAoB,OAC1B,QAAO,OAAO,MAAM,IAAI,gBAAgB;AAE1C,QAAO;;AAGT,SAAS,aAAa,MAAmD;CACvE,MAAMC,UAAkC;EACtC,gBAAgB;EAChB,QAAQ;EACR,GAAG,KAAK;EACT;AACD,KAAI,KAAK,WAAW,UAAa,KAAK,OAAO,SAAS,EACpD,SAAQ,gBAAgB,UAAU,KAAK;AAEzC,QAAO;;AAGT,SAAS,cAAc,OAAyB;AAC9C,KAAI,UAAU,UAAU,UAAU,UAAU,UAAU,WAAY,QAAO;AACzE,KAAI,OAAO,UAAU,YAAY,UAAU,QAAQ,UAAU,MAC3D,QAAO;EAAE,MAAM;EAAY,UAAU,EAAE,MAAO,MAA2B,MAAM;EAAE;AAEnF,QAAO;;AAGT,SAAS,gBAAgB,OAAgD;AACvE,SAAQ,OAAR;EACE,KAAK;EACL,KAAK;EACL,KAAK,iBACH,QAAO;EAGT,KAAK,iBACH,QAAO;EACT,KAAK;EACL,KAAK,gBACH,QAAO;EACT,KAAK;EACL,KAAK,OACH,QAAO;EACT,QACE,QAAO;;;AAIb,SAAS,eAAe,OAAmC;AACzD,KAAI,UAAU,UAAa,UAAU,QAAQ,OAAO,UAAU,SAAU,QAAO;CAC/E,MAAM,IAAI;CASV,MAAM,eAAe,EAAE,iBAAiB;CACxC,MAAM,mBAAmB,EAAE,qBAAqB;CAChD,MAAM,cAAc,EAAE,gBAAgB,eAAe;CACrD,MAAM,eAAe,EAAE,uBAAuB;AAU9C,QATqB;EACnB;EACA;EACA;EACA,GAAI,EAAE,qBAAqB,SAAY,EAAE,iBAAiB,EAAE,kBAAkB,GAAG,EAAE;EACnF,GAAI,OAAO,iBAAiB,YAAY,OAAO,SAAS,aAAa,IAAI,eAAe,IACpF,EAAE,kBAAkB,cAAc,GAClC,EAAE;EACP;;AAIH,SAAS,eAAe,KAAsB;AAC5C,KAAI,IAAI,WAAW,EAAG,QAAO,EAAE;AAC/B,KAAI;AACF,SAAO,KAAK,MAAM,IAAI;SAChB;AACN,SAAO;;;AAIX,SAAS,gBAAgB,GAA+B,GAAwC;AAC9F,KAAI,EAAE,WAAW,EAAE,OAAQ,QAAO;AAClC,MAAK,IAAI,IAAI,GAAG,IAAI,EAAE,QAAQ,IAC5B,KAAI,EAAE,OAAO,EAAE,GAAI,QAAO;AAE5B,QAAO;;AAGT,SAAS,cAAc,OAAwB,SAAiB,MAAqB;CACnF,MAAM,KAAK,UAAU,SAAS,QAAQ,OAAO,QAAQ;AACrD,KAAI,SAAS,OACX,IAAG,wBAAwB,WAAW,KAAK;KAE3C,IAAG,wBAAwB,UAAU;;;;;;;AASzC,MAAM,uCAAuB,IAAI,SAAiB;AAElD,SAAS,qBAAqB,MAGG;CAC/B,MAAM,MAAM,KAAK,UAAU;AAC3B,QAAO;EACL,YAAY,KAAK,cAAc,cAAc;EAC7C,OAAO,YAAY;AACjB,OAAI,qBAAqB,IAAI,KAAK,CAAE;AACpC,wBAAqB,IAAI,KAAK;AAC9B,OAAI,QAAQ,QAAQ;;EAEvB"}

package/dist/middleware/with-fallback.js

@@ -1,5 +1,5 @@
-import { defineProviderMiddleware } from "./compose.js";
 import { isAbortError } from "../internal/abort.js";
+import { defineProviderMiddleware } from "./compose.js";
 
 //#region src/middleware/with-fallback.ts
 /**

package/dist/middleware/with-rate-limit.d.ts

@@ -1,15 +1,8 @@
 import * as _graphorin_core4 from "@graphorin/core";
+import { ProviderRequest } from "@graphorin/core";
 
 //#region src/middleware/with-rate-limit.d.ts
 
-/**
- * `withRateLimit` - token-bucket rate limiter keyed per `provider × model`.
- * Configurable mode: `'throw'` (default) raises
- * {@link RateLimitExceededError} on overflow; `'queue'` waits for the
- * next available slot.
- *
- * @packageDocumentation
- */
 /**
  * Options for {@link withRateLimit}.
  *
@@ -20,6 +13,26 @@ interface WithRateLimitOptions {
   readonly requestsPerMinute: number;
   /** Burst size - defaults to `requestsPerMinute / 4` (rounded up to >= 1). */
   readonly burst?: number;
+  /**
+   * Optional token budget per minute (W-145). When set, each request
+   * additionally reserves its estimated token weight from a second
+   * bucket whose capacity is the full minute budget; a request whose
+   * weight exceeds the remaining budget waits (queue mode) or throws
+   * with the TPM-aware `retryAfterMs` (throw mode) even when the RPM
+   * bucket has room. Unset: behaviour is byte-identical to the
+   * RPM-only limiter.
+   */
+  readonly tokensPerMinute?: number;
+  /**
+   * Estimator for a request's token weight (only consulted when
+   * `tokensPerMinute` is set). The default is the deliberate cheap
+   * heuristic `ceil(textChars / 4) + (maxTokens ?? 0)` - synchronous
+   * and allocation-free, because this runs in the request hot path and
+   * must not add latency or network. Wire the counter from
+   * `@graphorin/provider/counters` (`createDefaultCounter`) here when
+   * you need provider-accurate weights.
+   */
+  readonly estimateTokens?: (req: ProviderRequest) => number;
   /** What to do on overflow. Default `'throw'`. */
   readonly mode?: 'throw' | 'queue';
   /** Test hook overriding `Date.now`. */

package/dist/middleware/with-rate-limit.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"with-rate-limit.d.ts","names":[],"sources":["../../src/middleware/with-rate-limit.ts"],"sourcesContent":[],"mappings":";;;;;;;AAmBA;AAgCA;;;;;;;;;UAhCiB,oBAAA;;;;;;;;;;6CAU4B,gBAAgB;;;;;cAsBhD,sBAAa,yBAAA,gBAAA,CAAA"}
+{"version":3,"file":"with-rate-limit.d.ts","names":[],"sources":["../../src/middleware/with-rate-limit.ts"],"sourcesContent":[],"mappings":";;;;;;;;;;UA0BiB,oBAAA;;;;;;;;;;;;;;;;;;;;;;;;kCAwBiB;;;;;;6CAMW,gBAAgB;;;;;cA2BhD,sBAAa,yBAAA,gBAAA,CAAA"}

package/dist/middleware/with-rate-limit.js

@@ -9,8 +9,12 @@ const withRateLimit = defineProviderMiddleware({
 	kind: "withRateLimit",
 	factory: (opts) => {
 		if (!(opts.requestsPerMinute > 0)) throw new RangeError("withRateLimit: requestsPerMinute must be > 0.");
+		if (opts.tokensPerMinute !== void 0 && !(opts.tokensPerMinute > 0)) throw new RangeError("withRateLimit: tokensPerMinute must be > 0 when set.");
 		const burst = Math.max(opts.burst ?? Math.ceil(opts.requestsPerMinute / 4), 1);
 		const refillPerMs = opts.requestsPerMinute / 6e4;
+		const tpm = opts.tokensPerMinute;
+		const tpmRefillPerMs = tpm === void 0 ? 0 : tpm / 6e4;
+		const estimate = opts.estimateTokens ?? estimateTokensHeuristic;
 		const mode = opts.mode ?? "throw";
 		const now = opts.nowImpl ?? (() => Date.now());
 		const sleep = opts.sleepImpl ?? defaultSleep;
@@ -23,6 +27,8 @@ const withRateLimit = defineProviderMiddleware({
 					bucket = {
 						tokens: burst,
 						lastRefill: now(),
+						tpmTokens: tpm ?? 0,
+						tpmLastRefill: now(),
 						queue: [],
 						draining: false
 					};
@@ -34,6 +40,34 @@ const withRateLimit = defineProviderMiddleware({
 				const ts = now();
 				bucket.tokens = Math.min(burst, bucket.tokens + (ts - bucket.lastRefill) * refillPerMs);
 				bucket.lastRefill = ts;
+				if (tpm !== void 0) {
+					bucket.tpmTokens = Math.min(tpm, bucket.tpmTokens + (ts - bucket.tpmLastRefill) * tpmRefillPerMs);
+					bucket.tpmLastRefill = ts;
+				}
+			};
+			/**
+			* A request's TPM weight, clamped to the bucket capacity - a
+			* single request estimated above the whole minute budget must
+			* degrade to "wait for a full bucket", not deadlock the queue
+			* forever behind an unsatisfiable reservation.
+			*/
+			const weightFor = (req) => {
+				if (tpm === void 0) return 0;
+				const raw = estimate(req);
+				return Math.min(Math.max(0, Number.isFinite(raw) ? raw : tpm), tpm);
+			};
+			/** Both dimensions available for this weight? */
+			const available = (bucket, weight) => bucket.tokens >= 1 && (tpm === void 0 || bucket.tpmTokens >= weight);
+			/** Wait until BOTH dimensions can serve `weight`, in ms. */
+			const waitMsFor = (bucket, weight) => {
+				const reqWait = bucket.tokens >= 1 ? 0 : (1 - bucket.tokens) / refillPerMs;
+				const tpmWait = tpm === void 0 || bucket.tpmTokens >= weight ? 0 : (weight - bucket.tpmTokens) / tpmRefillPerMs;
+				return Math.ceil(Math.max(reqWait, tpmWait));
+			};
+			/** Deduct BOTH dimensions atomically (caller checked `available`). */
+			const take = (bucket, weight) => {
+				bucket.tokens -= 1;
+				if (tpm !== void 0) bucket.tpmTokens -= weight;
 			};
 			const drain = (bucket) => {
 				if (bucket.draining) return;
@@ -47,13 +81,13 @@ const withRateLimit = defineProviderMiddleware({
 							continue;
 						}
 						refill(bucket);
-						if (bucket.tokens >= 1) {
-							bucket.tokens -= 1;
+						if (available(bucket, head.weight)) {
+							take(bucket, head.weight);
 							bucket.queue.shift();
 							head.resolve();
 							continue;
 						}
-						const waitMs = Math.ceil((1 - bucket.tokens) / refillPerMs);
+						const waitMs = waitMsFor(bucket, head.weight);
 						try {
 							await sleep(waitMs);
 						} catch {}
@@ -61,21 +95,22 @@ const withRateLimit = defineProviderMiddleware({
 					bucket.draining = false;
 				})();
 			};
-			const acquire = async (signal) => {
+			const acquire = async (signal, weight = 0) => {
 				if (signal?.aborted === true) throw signal.reason ?? /* @__PURE__ */ new Error("aborted");
 				const bucket = bucketFor();
 				refill(bucket);
-				if (bucket.tokens >= 1 && bucket.queue.length === 0) {
-					bucket.tokens -= 1;
+				if (available(bucket, weight) && bucket.queue.length === 0) {
+					take(bucket, weight);
 					return;
 				}
-				if (mode === "throw") throw new RateLimitExceededError(Math.ceil((1 - bucket.tokens) / refillPerMs));
+				if (mode === "throw") throw new RateLimitExceededError(waitMsFor(bucket, weight));
 				await new Promise((resolve, reject) => {
 					const waiter = {
 						resolve,
 						reject,
 						...signal ? { signal } : {},
-						cancelled: false
+						cancelled: false,
+						weight
 					};
 					if (signal !== void 0) {
 						const onAbort = () => {
@@ -99,10 +134,10 @@ const withRateLimit = defineProviderMiddleware({
 				capabilities: next.capabilities,
 				...next.acceptsSensitivity !== void 0 ? { acceptsSensitivity: next.acceptsSensitivity } : {},
 				stream(req) {
-					return rateLimitedStream(next, req, acquire);
+					return rateLimitedStream(next, req, acquire, weightFor(req));
 				},
 				async generate(req) {
-					await acquire(req.signal);
+					await acquire(req.signal, weightFor(req));
 					return next.generate(req);
 				},
 				...next.countTokens ? { countTokens: next.countTokens.bind(next) } : {}
@@ -110,8 +145,26 @@ const withRateLimit = defineProviderMiddleware({
 		};
 	}
 });
-async function* rateLimitedStream(next, req, acquire) {
-	await acquire(req.signal);
+/**
+* Default TPM weight heuristic: `ceil(textChars / 4) + (maxTokens ?? 0)`.
+* Counts string contents and the `text` of content parts plus the
+* system message; images / other parts are not weighed (heuristic by
+* design - see {@link WithRateLimitOptions.estimateTokens}).
+*/
+function estimateTokensHeuristic(req) {
+	let chars = req.systemMessage?.length ?? 0;
+	for (const message of req.messages) {
+		const content = message.content;
+		if (typeof content === "string") chars += content.length;
+		else if (Array.isArray(content)) for (const part of content) {
+			const text = part.text;
+			if (typeof text === "string") chars += text.length;
+		}
+	}
+	return Math.ceil(chars / 4) + (req.maxTokens ?? 0);
+}
+async function* rateLimitedStream(next, req, acquire, weight) {
+	await acquire(req.signal, weight);
 	for await (const event of next.stream(req)) yield event;
 }
 function defaultSleep(ms, signal) {

package/dist/middleware/with-rate-limit.js.map

@@ -1 +1 @@
-{"version":3,"file":"with-rate-limit.js","names":["waiter: Waiter"],"sources":["../../src/middleware/with-rate-limit.ts"],"sourcesContent":["/**\n * `withRateLimit` - token-bucket rate limiter keyed per `provider × model`.\n * Configurable mode: `'throw'` (default) raises\n * {@link RateLimitExceededError} on overflow; `'queue'` waits for the\n * next available slot.\n *\n * @packageDocumentation\n */\n\nimport type { Provider } from '@graphorin/core';\n\nimport { RateLimitExceededError } from '../errors/errors.js';\nimport { defineProviderMiddleware } from './compose.js';\n\n/**\n * Options for {@link withRateLimit}.\n *\n * @stable\n */\nexport interface WithRateLimitOptions {\n  /** Allowed requests per minute. */\n  readonly requestsPerMinute: number;\n  /** Burst size - defaults to `requestsPerMinute / 4` (rounded up to >= 1). */\n  readonly burst?: number;\n  /** What to do on overflow. Default `'throw'`. */\n  readonly mode?: 'throw' | 'queue';\n  /** Test hook overriding `Date.now`. */\n  readonly nowImpl?: () => number;\n  /** Test hook overriding `setTimeout`-based wait. */\n  readonly sleepImpl?: (ms: number, signal?: AbortSignal) => Promise<void>;\n}\n\ninterface Waiter {\n  resolve: () => void;\n  reject: (err: unknown) => void;\n  signal?: AbortSignal;\n  cancelled: boolean;\n}\n\ninterface BucketState {\n  tokens: number;\n  lastRefill: number;\n  /** FIFO waiters in `'queue'` mode (PS-18). */\n  queue: Waiter[];\n  /** Whether the single drain loop for this bucket is running. */\n  draining: boolean;\n}\n\n/**\n * @stable\n */\nexport const withRateLimit = defineProviderMiddleware<WithRateLimitOptions>({\n  kind: 'withRateLimit',\n  factory: (opts: WithRateLimitOptions) => {\n    if (!(opts.requestsPerMinute > 0)) {\n      throw new RangeError('withRateLimit: requestsPerMinute must be > 0.');\n    }\n    const burst = Math.max(opts.burst ?? Math.ceil(opts.requestsPerMinute / 4), 1);\n    const refillPerMs = opts.requestsPerMinute / 60_000;\n    const mode = opts.mode ?? 'throw';\n    const now = opts.nowImpl ?? (() => Date.now());\n    const sleep = opts.sleepImpl ?? defaultSleep;\n    const buckets = new Map<string, BucketState>();\n    return (next: Provider): Provider => {\n      const key = `${next.name}::${next.modelId}`;\n      const bucketFor = (): BucketState => {\n        let bucket = buckets.get(key);\n        if (bucket === undefined) {\n          bucket = { tokens: burst, lastRefill: now(), queue: [], draining: false };\n          buckets.set(key, bucket);\n        }\n        return bucket;\n      };\n      const refill = (bucket: BucketState): void => {\n        const ts = now();\n        bucket.tokens = Math.min(burst, bucket.tokens + (ts - bucket.lastRefill) * refillPerMs);\n        bucket.lastRefill = ts;\n      };\n      // PS-18: a single drain loop per bucket grants queued waiters ONE token at\n      // a time, sleeping until the next token actually refills. Previously every\n      // concurrent waiter computed the same wait, slept once, and then all\n      // passed - letting N requests through on ~1 token (a burst). FIFO order is\n      // preserved and grants track the real refill schedule.\n      const drain = (bucket: BucketState): void => {\n        if (bucket.draining) return;\n        bucket.draining = true;\n        void (async () => {\n          while (bucket.queue.length > 0) {\n            const head = bucket.queue[0];\n            if (head === undefined) break;\n            if (head.cancelled) {\n              bucket.queue.shift();\n              continue;\n            }\n            refill(bucket);\n            if (bucket.tokens >= 1) {\n              bucket.tokens -= 1;\n              bucket.queue.shift();\n              head.resolve();\n              continue;\n            }\n            const waitMs = Math.ceil((1 - bucket.tokens) / refillPerMs);\n            try {\n              await sleep(waitMs);\n            } catch {\n              // The shared drain sleep carries no per-waiter signal; individual\n              // aborts are handled on the waiter via its listener.\n            }\n          }\n          bucket.draining = false;\n        })();\n      };\n      const acquire = async (signal?: AbortSignal): Promise<void> => {\n        if (signal?.aborted === true) throw signal.reason ?? new Error('aborted');\n        const bucket = bucketFor();\n        refill(bucket);\n        // Fast path: a token is free AND nobody is queued ahead (preserve FIFO).\n        if (bucket.tokens >= 1 && bucket.queue.length === 0) {\n          bucket.tokens -= 1;\n          return;\n        }\n        if (mode === 'throw') {\n          throw new RateLimitExceededError(Math.ceil((1 - bucket.tokens) / refillPerMs));\n        }\n        await new Promise<void>((resolve, reject) => {\n          const waiter: Waiter = {\n            resolve,\n            reject,\n            ...(signal ? { signal } : {}),\n            cancelled: false,\n          };\n          if (signal !== undefined) {\n            const onAbort = (): void => {\n              waiter.cancelled = true;\n              reject(signal.reason ?? new Error('aborted'));\n            };\n            signal.addEventListener('abort', onAbort, { once: true });\n            const settle = resolve;\n            waiter.resolve = (): void => {\n              signal.removeEventListener('abort', onAbort);\n              settle();\n            };\n          }\n          bucket.queue.push(waiter);\n          drain(bucket);\n        });\n      };\n      return {\n        name: next.name,\n        modelId: next.modelId,\n        capabilities: next.capabilities,\n        ...(next.acceptsSensitivity !== undefined\n          ? { acceptsSensitivity: next.acceptsSensitivity }\n          : {}),\n        stream(req) {\n          return rateLimitedStream(next, req, acquire);\n        },\n        async generate(req) {\n          await acquire(req.signal);\n          return next.generate(req);\n        },\n        ...(next.countTokens ? { countTokens: next.countTokens.bind(next) } : {}),\n      };\n    };\n  },\n});\n\nasync function* rateLimitedStream(\n  next: Provider,\n  req: import('@graphorin/core').ProviderRequest,\n  acquire: (signal?: AbortSignal) => Promise<void>,\n): AsyncIterable<import('@graphorin/core').ProviderEvent> {\n  await acquire(req.signal);\n  for await (const event of next.stream(req)) {\n    yield event;\n  }\n}\n\nfunction defaultSleep(ms: number, signal?: AbortSignal): Promise<void> {\n  return new Promise<void>((resolve, reject) => {\n    if (signal?.aborted) {\n      reject(signal.reason ?? new Error('aborted'));\n      return;\n    }\n    const timer = setTimeout(() => {\n      cleanup();\n      resolve();\n    }, ms);\n    const onAbort = () => {\n      cleanup();\n      reject(signal?.reason ?? new Error('aborted'));\n    };\n    function cleanup(): void {\n      clearTimeout(timer);\n      signal?.removeEventListener('abort', onAbort);\n    }\n    signal?.addEventListener('abort', onAbort, { once: true });\n  });\n}\n"],"mappings":";;;;;;;AAmDA,MAAa,gBAAgB,yBAA+C;CAC1E,MAAM;CACN,UAAU,SAA+B;AACvC,MAAI,EAAE,KAAK,oBAAoB,GAC7B,OAAM,IAAI,WAAW,gDAAgD;EAEvE,MAAM,QAAQ,KAAK,IAAI,KAAK,SAAS,KAAK,KAAK,KAAK,oBAAoB,EAAE,EAAE,EAAE;EAC9E,MAAM,cAAc,KAAK,oBAAoB;EAC7C,MAAM,OAAO,KAAK,QAAQ;EAC1B,MAAM,MAAM,KAAK,kBAAkB,KAAK,KAAK;EAC7C,MAAM,QAAQ,KAAK,aAAa;EAChC,MAAM,0BAAU,IAAI,KAA0B;AAC9C,UAAQ,SAA6B;GACnC,MAAM,MAAM,GAAG,KAAK,KAAK,IAAI,KAAK;GAClC,MAAM,kBAA+B;IACnC,IAAI,SAAS,QAAQ,IAAI,IAAI;AAC7B,QAAI,WAAW,QAAW;AACxB,cAAS;MAAE,QAAQ;MAAO,YAAY,KAAK;MAAE,OAAO,EAAE;MAAE,UAAU;MAAO;AACzE,aAAQ,IAAI,KAAK,OAAO;;AAE1B,WAAO;;GAET,MAAM,UAAU,WAA8B;IAC5C,MAAM,KAAK,KAAK;AAChB,WAAO,SAAS,KAAK,IAAI,OAAO,OAAO,UAAU,KAAK,OAAO,cAAc,YAAY;AACvF,WAAO,aAAa;;GAOtB,MAAM,SAAS,WAA8B;AAC3C,QAAI,OAAO,SAAU;AACrB,WAAO,WAAW;AAClB,KAAM,YAAY;AAChB,YAAO,OAAO,MAAM,SAAS,GAAG;MAC9B,MAAM,OAAO,OAAO,MAAM;AAC1B,UAAI,SAAS,OAAW;AACxB,UAAI,KAAK,WAAW;AAClB,cAAO,MAAM,OAAO;AACpB;;AAEF,aAAO,OAAO;AACd,UAAI,OAAO,UAAU,GAAG;AACtB,cAAO,UAAU;AACjB,cAAO,MAAM,OAAO;AACpB,YAAK,SAAS;AACd;;MAEF,MAAM,SAAS,KAAK,MAAM,IAAI,OAAO,UAAU,YAAY;AAC3D,UAAI;AACF,aAAM,MAAM,OAAO;cACb;;AAKV,YAAO,WAAW;QAChB;;GAEN,MAAM,UAAU,OAAO,WAAwC;AAC7D,QAAI,QAAQ,YAAY,KAAM,OAAM,OAAO,0BAAU,IAAI,MAAM,UAAU;IACzE,MAAM,SAAS,WAAW;AAC1B,WAAO,OAAO;AAEd,QAAI,OAAO,UAAU,KAAK,OAAO,MAAM,WAAW,GAAG;AACnD,YAAO,UAAU;AACjB;;AAEF,QAAI,SAAS,QACX,OAAM,IAAI,uBAAuB,KAAK,MAAM,IAAI,OAAO,UAAU,YAAY,CAAC;AAEhF,UAAM,IAAI,SAAe,SAAS,WAAW;KAC3C,MAAMA,SAAiB;MACrB;MACA;MACA,GAAI,SAAS,EAAE,QAAQ,GAAG,EAAE;MAC5B,WAAW;MACZ;AACD,SAAI,WAAW,QAAW;MACxB,MAAM,gBAAsB;AAC1B,cAAO,YAAY;AACnB,cAAO,OAAO,0BAAU,IAAI,MAAM,UAAU,CAAC;;AAE/C,aAAO,iBAAiB,SAAS,SAAS,EAAE,MAAM,MAAM,CAAC;MACzD,MAAM,SAAS;AACf,aAAO,gBAAsB;AAC3B,cAAO,oBAAoB,SAAS,QAAQ;AAC5C,eAAQ;;;AAGZ,YAAO,MAAM,KAAK,OAAO;AACzB,WAAM,OAAO;MACb;;AAEJ,UAAO;IACL,MAAM,KAAK;IACX,SAAS,KAAK;IACd,cAAc,KAAK;IACnB,GAAI,KAAK,uBAAuB,SAC5B,EAAE,oBAAoB,KAAK,oBAAoB,GAC/C,EAAE;IACN,OAAO,KAAK;AACV,YAAO,kBAAkB,MAAM,KAAK,QAAQ;;IAE9C,MAAM,SAAS,KAAK;AAClB,WAAM,QAAQ,IAAI,OAAO;AACzB,YAAO,KAAK,SAAS,IAAI;;IAE3B,GAAI,KAAK,cAAc,EAAE,aAAa,KAAK,YAAY,KAAK,KAAK,EAAE,GAAG,EAAE;IACzE;;;CAGN,CAAC;AAEF,gBAAgB,kBACd,MACA,KACA,SACwD;AACxD,OAAM,QAAQ,IAAI,OAAO;AACzB,YAAW,MAAM,SAAS,KAAK,OAAO,IAAI,CACxC,OAAM;;AAIV,SAAS,aAAa,IAAY,QAAqC;AACrE,QAAO,IAAI,SAAe,SAAS,WAAW;AAC5C,MAAI,QAAQ,SAAS;AACnB,UAAO,OAAO,0BAAU,IAAI,MAAM,UAAU,CAAC;AAC7C;;EAEF,MAAM,QAAQ,iBAAiB;AAC7B,YAAS;AACT,YAAS;KACR,GAAG;EACN,MAAM,gBAAgB;AACpB,YAAS;AACT,UAAO,QAAQ,0BAAU,IAAI,MAAM,UAAU,CAAC;;EAEhD,SAAS,UAAgB;AACvB,gBAAa,MAAM;AACnB,WAAQ,oBAAoB,SAAS,QAAQ;;AAE/C,UAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,MAAM,CAAC;GAC1D"}
+{"version":3,"file":"with-rate-limit.js","names":["waiter: Waiter","content: unknown"],"sources":["../../src/middleware/with-rate-limit.ts"],"sourcesContent":["/**\n * `withRateLimit` - token-bucket rate limiter keyed per `provider × model`.\n * Configurable mode: `'throw'` (default) raises\n * {@link RateLimitExceededError} on overflow; `'queue'` waits for the\n * next available slot.\n *\n * W-145: an optional second dimension, `tokensPerMinute`, budgets\n * MODEL tokens alongside requests. For agentic workloads the binding\n * provider limit is TPM, not RPM - a 150k-token compacted transcript\n * used to occupy the same single slot as a 200-token reranker call,\n * so bursts of long-context steps sailed through the RPM gate and\n * surfaced as provider 429s that only `withRetry` absorbed.\n *\n * @packageDocumentation\n */\n\nimport type { Provider, ProviderRequest } from '@graphorin/core';\n\nimport { RateLimitExceededError } from '../errors/errors.js';\nimport { defineProviderMiddleware } from './compose.js';\n\n/**\n * Options for {@link withRateLimit}.\n *\n * @stable\n */\nexport interface WithRateLimitOptions {\n  /** Allowed requests per minute. */\n  readonly requestsPerMinute: number;\n  /** Burst size - defaults to `requestsPerMinute / 4` (rounded up to >= 1). */\n  readonly burst?: number;\n  /**\n   * Optional token budget per minute (W-145). When set, each request\n   * additionally reserves its estimated token weight from a second\n   * bucket whose capacity is the full minute budget; a request whose\n   * weight exceeds the remaining budget waits (queue mode) or throws\n   * with the TPM-aware `retryAfterMs` (throw mode) even when the RPM\n   * bucket has room. Unset: behaviour is byte-identical to the\n   * RPM-only limiter.\n   */\n  readonly tokensPerMinute?: number;\n  /**\n   * Estimator for a request's token weight (only consulted when\n   * `tokensPerMinute` is set). The default is the deliberate cheap\n   * heuristic `ceil(textChars / 4) + (maxTokens ?? 0)` - synchronous\n   * and allocation-free, because this runs in the request hot path and\n   * must not add latency or network. Wire the counter from\n   * `@graphorin/provider/counters` (`createDefaultCounter`) here when\n   * you need provider-accurate weights.\n   */\n  readonly estimateTokens?: (req: ProviderRequest) => number;\n  /** What to do on overflow. Default `'throw'`. */\n  readonly mode?: 'throw' | 'queue';\n  /** Test hook overriding `Date.now`. */\n  readonly nowImpl?: () => number;\n  /** Test hook overriding `setTimeout`-based wait. */\n  readonly sleepImpl?: (ms: number, signal?: AbortSignal) => Promise<void>;\n}\n\ninterface Waiter {\n  resolve: () => void;\n  reject: (err: unknown) => void;\n  signal?: AbortSignal;\n  cancelled: boolean;\n  /** TPM weight this waiter must reserve (0 when TPM is unset). */\n  weight: number;\n}\n\ninterface BucketState {\n  tokens: number;\n  lastRefill: number;\n  /** W-145: remaining TPM budget (only meaningful when TPM is set). */\n  tpmTokens: number;\n  tpmLastRefill: number;\n  /** FIFO waiters in `'queue'` mode (PS-18). */\n  queue: Waiter[];\n  /** Whether the single drain loop for this bucket is running. */\n  draining: boolean;\n}\n\n/**\n * @stable\n */\nexport const withRateLimit = defineProviderMiddleware<WithRateLimitOptions>({\n  kind: 'withRateLimit',\n  factory: (opts: WithRateLimitOptions) => {\n    if (!(opts.requestsPerMinute > 0)) {\n      throw new RangeError('withRateLimit: requestsPerMinute must be > 0.');\n    }\n    if (opts.tokensPerMinute !== undefined && !(opts.tokensPerMinute > 0)) {\n      throw new RangeError('withRateLimit: tokensPerMinute must be > 0 when set.');\n    }\n    const burst = Math.max(opts.burst ?? Math.ceil(opts.requestsPerMinute / 4), 1);\n    const refillPerMs = opts.requestsPerMinute / 60_000;\n    const tpm = opts.tokensPerMinute;\n    const tpmRefillPerMs = tpm === undefined ? 0 : tpm / 60_000;\n    const estimate = opts.estimateTokens ?? estimateTokensHeuristic;\n    const mode = opts.mode ?? 'throw';\n    const now = opts.nowImpl ?? (() => Date.now());\n    const sleep = opts.sleepImpl ?? defaultSleep;\n    const buckets = new Map<string, BucketState>();\n    return (next: Provider): Provider => {\n      const key = `${next.name}::${next.modelId}`;\n      const bucketFor = (): BucketState => {\n        let bucket = buckets.get(key);\n        if (bucket === undefined) {\n          bucket = {\n            tokens: burst,\n            lastRefill: now(),\n            tpmTokens: tpm ?? 0,\n            tpmLastRefill: now(),\n            queue: [],\n            draining: false,\n          };\n          buckets.set(key, bucket);\n        }\n        return bucket;\n      };\n      const refill = (bucket: BucketState): void => {\n        const ts = now();\n        bucket.tokens = Math.min(burst, bucket.tokens + (ts - bucket.lastRefill) * refillPerMs);\n        bucket.lastRefill = ts;\n        if (tpm !== undefined) {\n          // TPM bucket capacity = the full minute budget, so any\n          // (clamped) weight is eventually satisfiable.\n          bucket.tpmTokens = Math.min(\n            tpm,\n            bucket.tpmTokens + (ts - bucket.tpmLastRefill) * tpmRefillPerMs,\n          );\n          bucket.tpmLastRefill = ts;\n        }\n      };\n      /**\n       * A request's TPM weight, clamped to the bucket capacity - a\n       * single request estimated above the whole minute budget must\n       * degrade to \"wait for a full bucket\", not deadlock the queue\n       * forever behind an unsatisfiable reservation.\n       */\n      const weightFor = (req: ProviderRequest): number => {\n        if (tpm === undefined) return 0;\n        const raw = estimate(req);\n        return Math.min(Math.max(0, Number.isFinite(raw) ? raw : tpm), tpm);\n      };\n      /** Both dimensions available for this weight? */\n      const available = (bucket: BucketState, weight: number): boolean =>\n        bucket.tokens >= 1 && (tpm === undefined || bucket.tpmTokens >= weight);\n      /** Wait until BOTH dimensions can serve `weight`, in ms. */\n      const waitMsFor = (bucket: BucketState, weight: number): number => {\n        const reqWait = bucket.tokens >= 1 ? 0 : (1 - bucket.tokens) / refillPerMs;\n        const tpmWait =\n          tpm === undefined || bucket.tpmTokens >= weight\n            ? 0\n            : (weight - bucket.tpmTokens) / tpmRefillPerMs;\n        return Math.ceil(Math.max(reqWait, tpmWait));\n      };\n      /** Deduct BOTH dimensions atomically (caller checked `available`). */\n      const take = (bucket: BucketState, weight: number): void => {\n        bucket.tokens -= 1;\n        if (tpm !== undefined) bucket.tpmTokens -= weight;\n      };\n      // PS-18: a single drain loop per bucket grants queued waiters ONE slot at\n      // a time, sleeping until the next token actually refills. Previously every\n      // concurrent waiter computed the same wait, slept once, and then all\n      // passed - letting N requests through on ~1 token (a burst). FIFO order is\n      // preserved and grants track the real refill schedule. W-145: a grant now\n      // requires BOTH the request token and the head's TPM weight.\n      const drain = (bucket: BucketState): void => {\n        if (bucket.draining) return;\n        bucket.draining = true;\n        void (async () => {\n          while (bucket.queue.length > 0) {\n            const head = bucket.queue[0];\n            if (head === undefined) break;\n            if (head.cancelled) {\n              bucket.queue.shift();\n              continue;\n            }\n            refill(bucket);\n            if (available(bucket, head.weight)) {\n              take(bucket, head.weight);\n              bucket.queue.shift();\n              head.resolve();\n              continue;\n            }\n            const waitMs = waitMsFor(bucket, head.weight);\n            try {\n              await sleep(waitMs);\n            } catch {\n              // The shared drain sleep carries no per-waiter signal; individual\n              // aborts are handled on the waiter via its listener.\n            }\n          }\n          bucket.draining = false;\n        })();\n      };\n      const acquire = async (signal?: AbortSignal, weight = 0): Promise<void> => {\n        if (signal?.aborted === true) throw signal.reason ?? new Error('aborted');\n        const bucket = bucketFor();\n        refill(bucket);\n        // Fast path: both dimensions free AND nobody queued ahead (FIFO).\n        if (available(bucket, weight) && bucket.queue.length === 0) {\n          take(bucket, weight);\n          return;\n        }\n        if (mode === 'throw') {\n          throw new RateLimitExceededError(waitMsFor(bucket, weight));\n        }\n        await new Promise<void>((resolve, reject) => {\n          const waiter: Waiter = {\n            resolve,\n            reject,\n            ...(signal ? { signal } : {}),\n            cancelled: false,\n            weight,\n          };\n          if (signal !== undefined) {\n            const onAbort = (): void => {\n              waiter.cancelled = true;\n              reject(signal.reason ?? new Error('aborted'));\n            };\n            signal.addEventListener('abort', onAbort, { once: true });\n            const settle = resolve;\n            waiter.resolve = (): void => {\n              signal.removeEventListener('abort', onAbort);\n              settle();\n            };\n          }\n          bucket.queue.push(waiter);\n          drain(bucket);\n        });\n      };\n      return {\n        name: next.name,\n        modelId: next.modelId,\n        capabilities: next.capabilities,\n        ...(next.acceptsSensitivity !== undefined\n          ? { acceptsSensitivity: next.acceptsSensitivity }\n          : {}),\n        stream(req) {\n          return rateLimitedStream(next, req, acquire, weightFor(req));\n        },\n        async generate(req) {\n          await acquire(req.signal, weightFor(req));\n          return next.generate(req);\n        },\n        ...(next.countTokens ? { countTokens: next.countTokens.bind(next) } : {}),\n      };\n    };\n  },\n});\n\n/**\n * Default TPM weight heuristic: `ceil(textChars / 4) + (maxTokens ?? 0)`.\n * Counts string contents and the `text` of content parts plus the\n * system message; images / other parts are not weighed (heuristic by\n * design - see {@link WithRateLimitOptions.estimateTokens}).\n */\nfunction estimateTokensHeuristic(req: ProviderRequest): number {\n  let chars = req.systemMessage?.length ?? 0;\n  for (const message of req.messages) {\n    const content: unknown = message.content;\n    if (typeof content === 'string') {\n      chars += content.length;\n    } else if (Array.isArray(content)) {\n      for (const part of content) {\n        const text = (part as { readonly text?: unknown }).text;\n        if (typeof text === 'string') chars += text.length;\n      }\n    }\n  }\n  return Math.ceil(chars / 4) + (req.maxTokens ?? 0);\n}\n\nasync function* rateLimitedStream(\n  next: Provider,\n  req: import('@graphorin/core').ProviderRequest,\n  acquire: (signal?: AbortSignal, weight?: number) => Promise<void>,\n  weight: number,\n): AsyncIterable<import('@graphorin/core').ProviderEvent> {\n  await acquire(req.signal, weight);\n  for await (const event of next.stream(req)) {\n    yield event;\n  }\n}\n\nfunction defaultSleep(ms: number, signal?: AbortSignal): Promise<void> {\n  return new Promise<void>((resolve, reject) => {\n    if (signal?.aborted) {\n      reject(signal.reason ?? new Error('aborted'));\n      return;\n    }\n    const timer = setTimeout(() => {\n      cleanup();\n      resolve();\n    }, ms);\n    const onAbort = () => {\n      cleanup();\n      reject(signal?.reason ?? new Error('aborted'));\n    };\n    function cleanup(): void {\n      clearTimeout(timer);\n      signal?.removeEventListener('abort', onAbort);\n    }\n    signal?.addEventListener('abort', onAbort, { once: true });\n  });\n}\n"],"mappings":";;;;;;;AAmFA,MAAa,gBAAgB,yBAA+C;CAC1E,MAAM;CACN,UAAU,SAA+B;AACvC,MAAI,EAAE,KAAK,oBAAoB,GAC7B,OAAM,IAAI,WAAW,gDAAgD;AAEvE,MAAI,KAAK,oBAAoB,UAAa,EAAE,KAAK,kBAAkB,GACjE,OAAM,IAAI,WAAW,uDAAuD;EAE9E,MAAM,QAAQ,KAAK,IAAI,KAAK,SAAS,KAAK,KAAK,KAAK,oBAAoB,EAAE,EAAE,EAAE;EAC9E,MAAM,cAAc,KAAK,oBAAoB;EAC7C,MAAM,MAAM,KAAK;EACjB,MAAM,iBAAiB,QAAQ,SAAY,IAAI,MAAM;EACrD,MAAM,WAAW,KAAK,kBAAkB;EACxC,MAAM,OAAO,KAAK,QAAQ;EAC1B,MAAM,MAAM,KAAK,kBAAkB,KAAK,KAAK;EAC7C,MAAM,QAAQ,KAAK,aAAa;EAChC,MAAM,0BAAU,IAAI,KAA0B;AAC9C,UAAQ,SAA6B;GACnC,MAAM,MAAM,GAAG,KAAK,KAAK,IAAI,KAAK;GAClC,MAAM,kBAA+B;IACnC,IAAI,SAAS,QAAQ,IAAI,IAAI;AAC7B,QAAI,WAAW,QAAW;AACxB,cAAS;MACP,QAAQ;MACR,YAAY,KAAK;MACjB,WAAW,OAAO;MAClB,eAAe,KAAK;MACpB,OAAO,EAAE;MACT,UAAU;MACX;AACD,aAAQ,IAAI,KAAK,OAAO;;AAE1B,WAAO;;GAET,MAAM,UAAU,WAA8B;IAC5C,MAAM,KAAK,KAAK;AAChB,WAAO,SAAS,KAAK,IAAI,OAAO,OAAO,UAAU,KAAK,OAAO,cAAc,YAAY;AACvF,WAAO,aAAa;AACpB,QAAI,QAAQ,QAAW;AAGrB,YAAO,YAAY,KAAK,IACtB,KACA,OAAO,aAAa,KAAK,OAAO,iBAAiB,eAClD;AACD,YAAO,gBAAgB;;;;;;;;;GAS3B,MAAM,aAAa,QAAiC;AAClD,QAAI,QAAQ,OAAW,QAAO;IAC9B,MAAM,MAAM,SAAS,IAAI;AACzB,WAAO,KAAK,IAAI,KAAK,IAAI,GAAG,OAAO,SAAS,IAAI,GAAG,MAAM,IAAI,EAAE,IAAI;;;GAGrE,MAAM,aAAa,QAAqB,WACtC,OAAO,UAAU,MAAM,QAAQ,UAAa,OAAO,aAAa;;GAElE,MAAM,aAAa,QAAqB,WAA2B;IACjE,MAAM,UAAU,OAAO,UAAU,IAAI,KAAK,IAAI,OAAO,UAAU;IAC/D,MAAM,UACJ,QAAQ,UAAa,OAAO,aAAa,SACrC,KACC,SAAS,OAAO,aAAa;AACpC,WAAO,KAAK,KAAK,KAAK,IAAI,SAAS,QAAQ,CAAC;;;GAG9C,MAAM,QAAQ,QAAqB,WAAyB;AAC1D,WAAO,UAAU;AACjB,QAAI,QAAQ,OAAW,QAAO,aAAa;;GAQ7C,MAAM,SAAS,WAA8B;AAC3C,QAAI,OAAO,SAAU;AACrB,WAAO,WAAW;AAClB,KAAM,YAAY;AAChB,YAAO,OAAO,MAAM,SAAS,GAAG;MAC9B,MAAM,OAAO,OAAO,MAAM;AAC1B,UAAI,SAAS,OAAW;AACxB,UAAI,KAAK,WAAW;AAClB,cAAO,MAAM,OAAO;AACpB;;AAEF,aAAO,OAAO;AACd,UAAI,UAAU,QAAQ,KAAK,OAAO,EAAE;AAClC,YAAK,QAAQ,KAAK,OAAO;AACzB,cAAO,MAAM,OAAO;AACpB,YAAK,SAAS;AACd;;MAEF,MAAM,SAAS,UAAU,QAAQ,KAAK,OAAO;AAC7C,UAAI;AACF,aAAM,MAAM,OAAO;cACb;;AAKV,YAAO,WAAW;QAChB;;GAEN,MAAM,UAAU,OAAO,QAAsB,SAAS,MAAqB;AACzE,QAAI,QAAQ,YAAY,KAAM,OAAM,OAAO,0BAAU,IAAI,MAAM,UAAU;IACzE,MAAM,SAAS,WAAW;AAC1B,WAAO,OAAO;AAEd,QAAI,UAAU,QAAQ,OAAO,IAAI,OAAO,MAAM,WAAW,GAAG;AAC1D,UAAK,QAAQ,OAAO;AACpB;;AAEF,QAAI,SAAS,QACX,OAAM,IAAI,uBAAuB,UAAU,QAAQ,OAAO,CAAC;AAE7D,UAAM,IAAI,SAAe,SAAS,WAAW;KAC3C,MAAMA,SAAiB;MACrB;MACA;MACA,GAAI,SAAS,EAAE,QAAQ,GAAG,EAAE;MAC5B,WAAW;MACX;MACD;AACD,SAAI,WAAW,QAAW;MACxB,MAAM,gBAAsB;AAC1B,cAAO,YAAY;AACnB,cAAO,OAAO,0BAAU,IAAI,MAAM,UAAU,CAAC;;AAE/C,aAAO,iBAAiB,SAAS,SAAS,EAAE,MAAM,MAAM,CAAC;MACzD,MAAM,SAAS;AACf,aAAO,gBAAsB;AAC3B,cAAO,oBAAoB,SAAS,QAAQ;AAC5C,eAAQ;;;AAGZ,YAAO,MAAM,KAAK,OAAO;AACzB,WAAM,OAAO;MACb;;AAEJ,UAAO;IACL,MAAM,KAAK;IACX,SAAS,KAAK;IACd,cAAc,KAAK;IACnB,GAAI,KAAK,uBAAuB,SAC5B,EAAE,oBAAoB,KAAK,oBAAoB,GAC/C,EAAE;IACN,OAAO,KAAK;AACV,YAAO,kBAAkB,MAAM,KAAK,SAAS,UAAU,IAAI,CAAC;;IAE9D,MAAM,SAAS,KAAK;AAClB,WAAM,QAAQ,IAAI,QAAQ,UAAU,IAAI,CAAC;AACzC,YAAO,KAAK,SAAS,IAAI;;IAE3B,GAAI,KAAK,cAAc,EAAE,aAAa,KAAK,YAAY,KAAK,KAAK,EAAE,GAAG,EAAE;IACzE;;;CAGN,CAAC;;;;;;;AAQF,SAAS,wBAAwB,KAA8B;CAC7D,IAAI,QAAQ,IAAI,eAAe,UAAU;AACzC,MAAK,MAAM,WAAW,IAAI,UAAU;EAClC,MAAMC,UAAmB,QAAQ;AACjC,MAAI,OAAO,YAAY,SACrB,UAAS,QAAQ;WACR,MAAM,QAAQ,QAAQ,CAC/B,MAAK,MAAM,QAAQ,SAAS;GAC1B,MAAM,OAAQ,KAAqC;AACnD,OAAI,OAAO,SAAS,SAAU,UAAS,KAAK;;;AAIlD,QAAO,KAAK,KAAK,QAAQ,EAAE,IAAI,IAAI,aAAa;;AAGlD,gBAAgB,kBACd,MACA,KACA,SACA,QACwD;AACxD,OAAM,QAAQ,IAAI,QAAQ,OAAO;AACjC,YAAW,MAAM,SAAS,KAAK,OAAO,IAAI,CACxC,OAAM;;AAIV,SAAS,aAAa,IAAY,QAAqC;AACrE,QAAO,IAAI,SAAe,SAAS,WAAW;AAC5C,MAAI,QAAQ,SAAS;AACnB,UAAO,OAAO,0BAAU,IAAI,MAAM,UAAU,CAAC;AAC7C;;EAEF,MAAM,QAAQ,iBAAiB;AAC7B,YAAS;AACT,YAAS;KACR,GAAG;EACN,MAAM,gBAAgB;AACpB,YAAS;AACT,UAAO,QAAQ,0BAAU,IAAI,MAAM,UAAU,CAAC;;EAEhD,SAAS,UAAgB;AACvB,gBAAa,MAAM;AACnB,WAAQ,oBAAoB,SAAS,QAAQ;;AAE/C,UAAQ,iBAAiB,SAAS,SAAS,EAAE,MAAM,MAAM,CAAC;GAC1D"}

package/dist/middleware/with-retry.js

@@ -1,5 +1,5 @@
-import { defineProviderMiddleware } from "./compose.js";
 import { isAbortError } from "../internal/abort.js";
+import { defineProviderMiddleware } from "./compose.js";
 
 //#region src/middleware/with-retry.ts
 /**

package/dist/package.js

@@ -0,0 +1,6 @@
+//#region package.json
+var version = "0.7.0";
+
+//#endregion
+export { version };
+//# sourceMappingURL=package.js.map

package/dist/package.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"package.js","names":[],"sources":["../package.json"],"sourcesContent":["{\n  \"name\": \"@graphorin/provider\",\n  \"version\": \"0.7.0\",\n  \"description\": \"Vendor-neutral LLM provider layer for the Graphorin framework. Ships the canonical Provider interface, the Vercel AI SDK adapter (default cloud path), three local-LLM adapters (direct Ollama HTTP, llama-server HTTP, and a generic OpenAI-compatible adapter for LMStudio / LocalAI / vLLM / Together-style endpoints), the shared LocalProviderTrust classifier (loopback / private / public-tls / public-cleartext), the canonical-order middleware composer (withTracing / withRetry / withRateLimit / withCostLimit / withCostTracking / withFallback / withRedaction), the pluggable TokenCounter dispatcher (js-tiktoken default plus per-vendor native counters and a heuristic fallback), the per-provider model-tier auto-classifier (fast / balanced / smart), and the production startup hook that fails fast when withRedaction is missing.\",\n  \"license\": \"MIT\",\n  \"author\": \"Oleksiy Stepurenko\",\n  \"homepage\": \"https://github.com/o-stepper/graphorin/tree/main/packages/provider\",\n  \"repository\": {\n    \"type\": \"git\",\n    \"url\": \"git+https://github.com/o-stepper/graphorin.git\",\n    \"directory\": \"packages/provider\"\n  },\n  \"bugs\": {\n    \"url\": \"https://github.com/o-stepper/graphorin/issues\"\n  },\n  \"keywords\": [\n    \"graphorin\",\n    \"ai\",\n    \"agents\",\n    \"framework\",\n    \"provider\",\n    \"llm\",\n    \"vercel-ai-sdk\",\n    \"ollama\",\n    \"llama-cpp\",\n    \"openai-compatible\",\n    \"middleware\",\n    \"retry\",\n    \"fallback\",\n    \"rate-limit\",\n    \"cost-limit\",\n    \"redaction\",\n    \"token-counter\",\n    \"tiktoken\",\n    \"local-first\"\n  ],\n  \"type\": \"module\",\n  \"sideEffects\": false,\n  \"engines\": {\n    \"node\": \">=22.12.0\"\n  },\n  \"main\": \"./dist/index.js\",\n  \"module\": \"./dist/index.js\",\n  \"types\": \"./dist/index.d.ts\",\n  \"exports\": {\n    \".\": {\n      \"types\": \"./dist/index.d.ts\",\n      \"default\": \"./dist/index.js\"\n    },\n    \"./adapters/vercel\": {\n      \"types\": \"./dist/adapters/vercel.d.ts\",\n      \"default\": \"./dist/adapters/vercel.js\"\n    },\n    \"./adapters/ollama\": {\n      \"types\": \"./dist/adapters/ollama.d.ts\",\n      \"default\": \"./dist/adapters/ollama.js\"\n    },\n    \"./adapters/llamacpp-server\": {\n      \"types\": \"./dist/adapters/llamacpp-server.d.ts\",\n      \"default\": \"./dist/adapters/llamacpp-server.js\"\n    },\n    \"./adapters/openai-compatible\": {\n      \"types\": \"./dist/adapters/openai-compatible.d.ts\",\n      \"default\": \"./dist/adapters/openai-compatible.js\"\n    },\n    \"./middleware\": {\n      \"types\": \"./dist/middleware/index.d.ts\",\n      \"default\": \"./dist/middleware/index.js\"\n    },\n    \"./counters\": {\n      \"types\": \"./dist/counters/index.d.ts\",\n      \"default\": \"./dist/counters/index.js\"\n    },\n    \"./trust\": {\n      \"types\": \"./dist/trust/index.d.ts\",\n      \"default\": \"./dist/trust/index.js\"\n    },\n    \"./model-tier\": {\n      \"types\": \"./dist/model-tier/index.d.ts\",\n      \"default\": \"./dist/model-tier/index.js\"\n    },\n    \"./reasoning\": {\n      \"types\": \"./dist/reasoning/index.d.ts\",\n      \"default\": \"./dist/reasoning/index.js\"\n    },\n    \"./errors\": {\n      \"types\": \"./dist/errors/index.d.ts\",\n      \"default\": \"./dist/errors/index.js\"\n    },\n    \"./package.json\": \"./package.json\"\n  },\n  \"files\": [\n    \"dist\",\n    \"src\",\n    \"README.md\",\n    \"CHANGELOG.md\",\n    \"LICENSE\"\n  ],\n  \"scripts\": {\n    \"build\": \"tsdown\",\n    \"typecheck\": \"tsc --noEmit && tsc -p tsconfig.tests.json\",\n    \"test\": \"vitest run\",\n    \"lint\": \"biome check .\",\n    \"clean\": \"rimraf dist .turbo *.tsbuildinfo\"\n  },\n  \"dependencies\": {\n    \"@graphorin/core\": \"workspace:*\",\n    \"@graphorin/observability\": \"workspace:*\"\n  },\n  \"peerDependencies\": {\n    \"ai\": \"^7.0.0-beta.76\",\n    \"js-tiktoken\": \"^1.0.0\"\n  },\n  \"peerDependenciesMeta\": {\n    \"ai\": {\n      \"optional\": true\n    },\n    \"js-tiktoken\": {\n      \"optional\": true\n    }\n  },\n  \"publishConfig\": {\n    \"access\": \"public\",\n    \"provenance\": true\n  },\n  \"devDependencies\": {\n    \"@graphorin/pricing\": \"workspace:*\",\n    \"ai\": \"^7.0.15\",\n    \"fast-check\": \"^3.23.0\",\n    \"js-tiktoken\": \"^1.0.21\"\n  }\n}\n"],"mappings":";cAEa"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@graphorin/provider",
-  "version": "0.6.0",
+  "version": "0.7.0",
   "description": "Vendor-neutral LLM provider layer for the Graphorin framework. Ships the canonical Provider interface, the Vercel AI SDK adapter (default cloud path), three local-LLM adapters (direct Ollama HTTP, llama-server HTTP, and a generic OpenAI-compatible adapter for LMStudio / LocalAI / vLLM / Together-style endpoints), the shared LocalProviderTrust classifier (loopback / private / public-tls / public-cleartext), the canonical-order middleware composer (withTracing / withRetry / withRateLimit / withCostLimit / withCostTracking / withFallback / withRedaction), the pluggable TokenCounter dispatcher (js-tiktoken default plus per-vendor native counters and a heuristic fallback), the per-provider model-tier auto-classifier (fast / balanced / smart), and the production startup hook that fails fast when withRedaction is missing.",
   "license": "MIT",
   "author": "Oleksiy Stepurenko",
@@ -35,8 +35,9 @@
     "local-first"
   ],
   "type": "module",
+  "sideEffects": false,
   "engines": {
-    "node": ">=22.0.0"
+    "node": ">=22.12.0"
   },
   "main": "./dist/index.js",
   "module": "./dist/index.js",
@@ -44,59 +45,60 @@
   "exports": {
     ".": {
       "types": "./dist/index.d.ts",
-      "import": "./dist/index.js"
+      "default": "./dist/index.js"
     },
     "./adapters/vercel": {
       "types": "./dist/adapters/vercel.d.ts",
-      "import": "./dist/adapters/vercel.js"
+      "default": "./dist/adapters/vercel.js"
     },
     "./adapters/ollama": {
       "types": "./dist/adapters/ollama.d.ts",
-      "import": "./dist/adapters/ollama.js"
+      "default": "./dist/adapters/ollama.js"
     },
     "./adapters/llamacpp-server": {
       "types": "./dist/adapters/llamacpp-server.d.ts",
-      "import": "./dist/adapters/llamacpp-server.js"
+      "default": "./dist/adapters/llamacpp-server.js"
     },
     "./adapters/openai-compatible": {
       "types": "./dist/adapters/openai-compatible.d.ts",
-      "import": "./dist/adapters/openai-compatible.js"
+      "default": "./dist/adapters/openai-compatible.js"
     },
     "./middleware": {
       "types": "./dist/middleware/index.d.ts",
-      "import": "./dist/middleware/index.js"
+      "default": "./dist/middleware/index.js"
     },
     "./counters": {
       "types": "./dist/counters/index.d.ts",
-      "import": "./dist/counters/index.js"
+      "default": "./dist/counters/index.js"
     },
     "./trust": {
       "types": "./dist/trust/index.d.ts",
-      "import": "./dist/trust/index.js"
+      "default": "./dist/trust/index.js"
     },
     "./model-tier": {
       "types": "./dist/model-tier/index.d.ts",
-      "import": "./dist/model-tier/index.js"
+      "default": "./dist/model-tier/index.js"
     },
     "./reasoning": {
       "types": "./dist/reasoning/index.d.ts",
-      "import": "./dist/reasoning/index.js"
+      "default": "./dist/reasoning/index.js"
     },
     "./errors": {
       "types": "./dist/errors/index.d.ts",
-      "import": "./dist/errors/index.js"
+      "default": "./dist/errors/index.js"
     },
     "./package.json": "./package.json"
   },
   "files": [
     "dist",
+    "src",
     "README.md",
     "CHANGELOG.md",
     "LICENSE"
   ],
   "dependencies": {
-    "@graphorin/core": "0.6.0",
-    "@graphorin/observability": "0.6.0"
+    "@graphorin/core": "0.7.0",
+    "@graphorin/observability": "0.7.0"
   },
   "peerDependencies": {
     "ai": "^7.0.0-beta.76",
@@ -118,7 +120,7 @@
     "ai": "^7.0.15",
     "fast-check": "^3.23.0",
     "js-tiktoken": "^1.0.21",
-    "@graphorin/pricing": "0.6.0"
+    "@graphorin/pricing": "0.7.0"
   },
   "scripts": {
     "build": "tsdown",

package/src/adapters/index.ts

@@ -0,0 +1,14 @@
+/**
+ * Adapter barrel - re-exports the four bundled adapters.
+ *
+ * @packageDocumentation
+ */
+
+export type { LlamaCppServerAdapterOptions } from './llamacpp-server.js';
+export { llamaCppServerAdapter } from './llamacpp-server.js';
+export type { OllamaAdapterOptions } from './ollama.js';
+export { ollamaAdapter } from './ollama.js';
+export type { OpenAICompatibleAdapterOptions } from './openai-compatible.js';
+export { openAICompatibleAdapter } from './openai-compatible.js';
+export type { VercelAdapterOptions } from './vercel.js';
+export { vercelAdapter } from './vercel.js';