@hawkeyexl/inference 0.2.0 → 0.3.0

package/dist/index.js

@@ -468,26 +468,46 @@ function defaultLlamaModelsDirectory() {
 var LLAMA_TIERS = ["fast", "balanced", "quality"];
 var LLAMA_SELECTORS = ["auto", ...LLAMA_TIERS];
 var LLAMA_MODELS = deepFreezeEntries({
+  "granite-4.1-3b-q2": {
+    uri: "hf:unsloth/granite-4.1-3b-GGUF/granite-4.1-3b-UD-Q2_K_XL.gguf",
+    sizeBytes: 1414548800,
+    license: "Apache-2.0",
+    tier: "fast",
+    notes: "Smallest tier and the quickest measured (4.8s/page). Scores level with models three times its size on schema-constrained extraction."
+  },
+  "qwen3.5-4b": {
+    uri: "hf:unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-UD-Q4_K_XL.gguf",
+    sizeBytes: 2912109728,
+    license: "Apache-2.0",
+    notes: "The default for most machines. Smaller and faster than the Gemma 4 E4B it replaces, at indistinguishable measured quality.",
+    tier: "balanced"
+  },
+  "qwen3.5-9b": {
+    uri: "hf:unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-UD-Q4_K_XL.gguf",
+    sizeBytes: 5966095584,
+    license: "Apache-2.0",
+    tier: "quality",
+    notes: "Best measured of everything tried, and still smaller and faster than the Gemma 4 12B it replaces. Wants a GPU or plenty of RAM."
+  },
+  // --- Superseded, kept resolvable by name. Untiered: nothing selects these
+  // unless a caller asks for one outright.
   "gemma-4-e2b": {
     uri: "hf:unsloth/gemma-4-E2B-it-qat-GGUF/gemma-4-E2B-it-qat-UD-Q4_K_XL.gguf",
     sizeBytes: 2620370976,
     license: "Apache-2.0",
-    tier: "fast",
-    notes: "IFEval 94.6. Smallest Gemma 4; the floor for this family."
+    notes: "IFEval 94.6. Former `fast` tier; sound, but larger than Granite."
   },
   "gemma-4-e4b": {
     uri: "hf:unsloth/gemma-4-E4B-it-qat-GGUF/gemma-4-E4B-it-qat-UD-Q4_K_XL.gguf",
     sizeBytes: 4215695776,
     license: "Apache-2.0",
-    tier: "balanced",
-    notes: "IFEval 96.7. The default for most machines."
+    notes: "IFEval 96.7. Former `balanced` tier; sound, but larger."
   },
   "gemma-4-12b": {
     uri: "hf:unsloth/gemma-4-12B-it-qat-GGUF/gemma-4-12B-it-qat-UD-Q4_K_XL.gguf",
     sizeBytes: 6716356800,
     license: "Apache-2.0",
-    tier: "quality",
-    notes: "IFEval 97.2. Dense 12B; wants a GPU or plenty of RAM."
+    notes: "IFEval 97.2. Former `quality` tier. Dense 12B; wants a GPU."
   },
   "gemma-4-26b-a4b": {
     uri: "hf:unsloth/gemma-4-26B-A4B-it-qat-GGUF/gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf",
@@ -499,7 +519,7 @@ var LLAMA_MODELS = deepFreezeEntries({
     uri: "hf:unsloth/gemma-4-E2B-it-qat-GGUF/gemma-4-E2B-it-qat-UD-Q2_K_XL.gguf",
     sizeBytes: 2186186784,
     license: "Apache-2.0",
-    notes: "Q2 build of the fast tier; smallest download, lowest fidelity."
+    notes: "AVOID. Smallest download, but it does not reliably terminate: 6 of 12 pages unfinished at 120s, one still running at 400s, and the pages that did finish proposed identifiers as prose. Kept only so existing pins still resolve."
   }
 });
 function deepFreezeEntries(catalog) {
@@ -507,9 +527,9 @@ function deepFreezeEntries(catalog) {
   return Object.freeze(catalog);
 }
 var TIER_ALIAS = {
-  fast: "gemma-4-e2b",
-  balanced: "gemma-4-e4b",
-  quality: "gemma-4-12b"
+  fast: "granite-4.1-3b-q2",
+  balanced: "qwen3.5-4b",
+  quality: "qwen3.5-9b"
 };
 function isLlamaSelector(model) {
   return LLAMA_SELECTORS.includes(model);
@@ -532,7 +552,7 @@ function uriForTier(tier) {
 function resolveLlamaModelRef(model) {
   if (isLlamaSelector(model)) {
     throw new InferenceError(
-      `llama-cpp model "${model}" is a selector and needs a hardware probe to resolve. Use resolveProviderIdentityAsync/makeProviderAsync, or name a concrete model (e.g. "gemma-4-e4b").`
+      `llama-cpp model "${model}" is a selector and needs a hardware probe to resolve. Use resolveProviderIdentityAsync/makeProviderAsync, or name a concrete model (e.g. "${TIER_ALIAS.balanced}").`
     );
   }
   const entry = LLAMA_MODELS[model];
@@ -780,7 +800,7 @@ var LlamaCppProvider = class {
     this.model = model;
     if (isLlamaSelector(model)) {
       throw new InferenceError(
-        `llama-cpp model "${model}" is a selector. Constructing a provider directly needs a concrete model (e.g. "gemma-4-e4b") \u2014 use makeProviderAsync to resolve a selector against this machine.`
+        `llama-cpp model "${model}" is a selector. Constructing a provider directly needs a concrete model (e.g. "${aliasForTier("balanced")}") \u2014 use makeProviderAsync to resolve a selector against this machine.`
       );
     }
     this.uri = resolveLlamaModelRef(model);
@@ -1088,7 +1108,7 @@ function resolveProviderIdentity(spec) {
   const model = spec.model ?? DEFAULT_MODELS[spec.provider] ?? "unknown";
   if (spec.provider === "llama-cpp" && isLlamaSelector(model)) {
     throw new InferenceError(
-      `llama-cpp model "${model}" is a selector and cannot be resolved synchronously \u2014 picking a tier probes GPU memory. Use resolveProviderIdentityAsync/makeProviderAsync, or name a concrete model (e.g. "gemma-4-e4b").`
+      `llama-cpp model "${model}" is a selector and cannot be resolved synchronously \u2014 picking a tier probes GPU memory. Use resolveProviderIdentityAsync/makeProviderAsync, or name a concrete model (e.g. "${aliasForTier("balanced")}").`
     );
   }
   return { provider: spec.provider, model };

package/dist/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/types.ts","../src/runtime.ts","../src/providers/anthropic.ts","../src/providers/openai-compat.ts","../src/exec.ts","../src/providers/claude-cli.ts","../src/providers/mock.ts","../src/cache.ts","../src/providers/llama-models.ts","../src/providers/llama-install.ts","../src/providers/llama-cpp.ts","../src/providers/detect.ts","../src/providers/index.ts","../src/providers/llama-clean.ts","../src/complete.ts","../src/cost.ts","../src/judge/verdict-schema.json","../src/judge/types.ts","../src/judge/consensus.ts","../src/judge/zones.ts","../src/judge/ensemble.ts"],"sourcesContent":["/**\n * Shared error type. Consumers catch this to distinguish an inference-layer\n * operational failure (missing API key, unknown provider) from their own\n * domain errors, and typically map it to their own exit code.\n */\nexport class InferenceError extends Error {\n  constructor(message: string) {\n    super(message);\n    this.name = \"InferenceError\";\n  }\n}\n","/**\n * The one thing this library says about the runtime it was loaded into.\n *\n * `engines.node` is `>=24`, but npm treats an `engines` mismatch as an\n * `EBADENGINE` **warning**, not a refusal — scrolled past in CI, invisible in a\n * transitive install. So an older Node installs cleanly, runs, and then fails\n * somewhere unrelated with an error that never mentions the Node version.\n *\n * This is a warning rather than a throw on purpose. Nothing in `src/` uses a\n * Node-24-only API — the imports are `node:crypto`, `node:fs`, `node:os` and\n * `node:path`, the newest globals are `fetch` (18+) and `structuredClone`\n * (17+), and `tsconfig` targets ES2022 — and the strictest dependency floor is\n * `node-llama-cpp` at `>=20`. Node 24 is this package's *support* policy, not a\n * technical impossibility, and a library has no business refusing to run on a\n * consumer's behalf when it can still do the work. It says so once and\n * continues, the same way the four existing warn-once paths do.\n */\n\n/**\n * The major version `engines.node` declares. `test/unit/runtime.test.ts` pins\n * this against `package.json`, so the two cannot drift apart.\n */\nexport const MINIMUM_NODE_MAJOR = 24;\n\nlet warnedNodeVersion = false;\n\n/**\n * Warn once if the running Node is older than the declared minimum.\n *\n * The version is a parameter so this is testable in-process — no subprocess on\n * a second Node install just to see the string.\n */\nexport function warnIfUnsupportedNode(\n  version: string = process.versions.node,\n): void {\n  if (warnedNodeVersion) return;\n  const major = Number.parseInt(version, 10);\n  // An unreadable version is not evidence of an old one. Same rule as an\n  // unknown model price: never guess.\n  if (!Number.isInteger(major) || major >= MINIMUM_NODE_MAJOR) return;\n  warnedNodeVersion = true;\n  console.warn(\n    `inference: running on Node ${version}, older than the Node ` +\n      `${MINIMUM_NODE_MAJOR} this package requires. npm only warns about that ` +\n      `at install time (EBADENGINE), so nothing has stopped you yet — upgrade ` +\n      `Node, or expect failures this library cannot explain.`,\n  );\n}\n\n/** Test seam: reset the once-per-process Node version warning. */\nexport function resetNodeVersionWarning(): void {\n  warnedNodeVersion = false;\n}\n","/**\n * Anthropic provider: structured output via a single forced tool call whose\n * input schema is the caller's schema — the model cannot answer any other way.\n */\nimport Anthropic from \"@anthropic-ai/sdk\";\nimport { InferenceError } from \"../types.js\";\nimport type {\n  CompleteJSONRequest,\n  CompleteJSONResponse,\n  InferenceProvider,\n} from \"./types.js\";\n\nconst DEFAULT_TOOL_NAME = \"record_result\";\n\nexport interface AnthropicProviderOptions {\n  /**\n   * Name of the forced tool. Purely cosmetic to the model, but a descriptive\n   * name (\"record_verdict\", \"record_proposal\") measurably steers output, so\n   * consumers may set their own.\n   */\n  toolName?: string;\n  /** Tool description shown to the model. */\n  toolDescription?: string;\n  maxTokens?: number;\n}\n\nexport class AnthropicProvider implements InferenceProvider {\n  private readonly client: Anthropic;\n  private readonly toolName: string;\n  private readonly toolDescription: string;\n  private readonly maxTokens: number;\n\n  constructor(\n    private readonly model: string,\n    apiKeyEnv: string,\n    options: AnthropicProviderOptions = {},\n  ) {\n    const apiKey = process.env[apiKeyEnv];\n    if (!apiKey) {\n      throw new InferenceError(\n        `Anthropic provider needs ${apiKeyEnv} set (or choose another provider)`,\n      );\n    }\n    this.client = new Anthropic({ apiKey });\n    this.toolName = options.toolName ?? DEFAULT_TOOL_NAME;\n    this.toolDescription =\n      options.toolDescription ?? \"Record the structured result.\";\n    this.maxTokens = options.maxTokens ?? 1024;\n  }\n\n  provider(): string {\n    return \"anthropic\";\n  }\n\n  modelName(): string {\n    return this.model;\n  }\n\n  async completeJSON(req: CompleteJSONRequest): Promise<CompleteJSONResponse> {\n    const response = await this.client.messages.create({\n      model: this.model,\n      max_tokens: this.maxTokens,\n      temperature: req.temperature,\n      system: req.system,\n      messages: [{ role: \"user\", content: req.user }],\n      tools: [\n        {\n          name: this.toolName,\n          description: this.toolDescription,\n          input_schema: req.schema as Anthropic.Tool[\"input_schema\"],\n        },\n      ],\n      tool_choice: { type: \"tool\", name: this.toolName },\n    });\n\n    // A truncated tool call still arrives as a tool_use block, just with\n    // partial input. Without this check it fails schema validation instead,\n    // burning the retry and reporting a misleading validation error rather\n    // than the actionable \"raise maxTokens\".\n    if (response.stop_reason === \"max_tokens\") {\n      throw new Error(\n        `Anthropic response hit max_tokens (${this.maxTokens}) before completing the tool call — raise the anthropic.maxTokens option.`,\n      );\n    }\n\n    const toolUse = response.content.find(\n      (block): block is Anthropic.ToolUseBlock => block.type === \"tool_use\",\n    );\n    if (!toolUse) {\n      throw new Error(\"Anthropic response contained no tool_use block\");\n    }\n    return {\n      json: toolUse.input,\n      usage: {\n        inputTokens: response.usage.input_tokens,\n        outputTokens: response.usage.output_tokens,\n      },\n    };\n  }\n}\n","/**\n * OpenAI-compatible provider: works against OpenAI, Azure, Ollama, Groq,\n * Together, or any server speaking /chat/completions. Prefers strict\n * json_schema response_format; falls back to json_object + schema-in-prompt\n * when the server rejects it (older Ollama, some proxies).\n */\nimport { InferenceError } from \"../types.js\";\nimport type {\n  CompleteJSONRequest,\n  CompleteJSONResponse,\n  InferenceProvider,\n} from \"./types.js\";\n\ninterface ChatResponse {\n  choices?: { message?: { content?: string } }[];\n  usage?: { prompt_tokens?: number; completion_tokens?: number };\n  error?: { message?: string };\n}\n\n/** Extract a JSON object from content that may carry markdown fences. */\nexport function extractJson(content: string): unknown {\n  const trimmed = content\n    .replace(/^\\s*```(?:json)?\\s*/i, \"\")\n    .replace(/\\s*```\\s*$/, \"\")\n    .trim();\n  try {\n    return JSON.parse(trimmed);\n  } catch {\n    const start = trimmed.indexOf(\"{\");\n    const end = trimmed.lastIndexOf(\"}\");\n    if (start >= 0 && end > start) {\n      return JSON.parse(trimmed.slice(start, end + 1));\n    }\n    throw new Error(\"Response contained no parseable JSON object\");\n  }\n}\n\n/**\n * OpenAI strict mode requires `required` to list EVERY property (optionality\n * is expressed as a `null` type union) and rejects keywords outside its\n * subset (minLength, uniqueItems). Transform an all-optional schema into a\n * strict-compatible equivalent; null values are stripped from the response.\n */\nexport function toStrictSchema(\n  schema: Record<string, unknown>,\n): Record<string, unknown> {\n  const clone = structuredClone(schema);\n  const walk = (node: unknown): void => {\n    if (node === null || typeof node !== \"object\") return;\n    if (Array.isArray(node)) {\n      for (const item of node) walk(item);\n      return;\n    }\n    const obj = node as Record<string, unknown>;\n    delete obj[\"minLength\"];\n    delete obj[\"uniqueItems\"];\n    const properties = obj[\"properties\"];\n    if (properties && typeof properties === \"object\") {\n      obj[\"required\"] = Object.keys(properties);\n      // Strict mode requires additionalProperties:false on EVERY object, not\n      // just the root. A nested object without it is rejected as a schema\n      // error, which permanently downgrades this provider to the weaker\n      // json_object fallback for the rest of its life.\n      obj[\"additionalProperties\"] = false;\n      for (const prop of Object.values(properties as Record<string, unknown>)) {\n        walk(prop);\n        if (prop && typeof prop === \"object\" && !Array.isArray(prop)) {\n          const p = prop as Record<string, unknown>;\n          if (typeof p[\"type\"] === \"string\" && p[\"type\"] !== \"null\") {\n            p[\"type\"] = [p[\"type\"], \"null\"];\n          }\n        }\n      }\n    }\n    walk(obj[\"items\"]);\n  };\n  walk(clone);\n  return clone;\n}\n\n/** Remove null-valued keys (strict-mode \"omitted\" marker) from a response object. */\nexport function stripNulls(value: unknown): unknown {\n  if (value === null || typeof value !== \"object\" || Array.isArray(value)) {\n    return value;\n  }\n  return Object.fromEntries(\n    Object.entries(value as Record<string, unknown>).filter(\n      ([, v]) => v !== null,\n    ),\n  );\n}\n\nexport interface OpenAICompatProviderOptions {\n  /** `json_schema.name` sent to the server. Cosmetic; steers some models. */\n  schemaName?: string;\n}\n\nexport class OpenAICompatProvider implements InferenceProvider {\n  private supportsJsonSchema = true;\n  private readonly schemaName: string;\n\n  constructor(\n    private readonly baseUrl: string,\n    private readonly model: string,\n    apiKeyEnv: string,\n    private readonly apiKey: string | undefined = process.env[apiKeyEnv],\n    options: OpenAICompatProviderOptions = {},\n  ) {\n    // Local servers (Ollama) often need no key; only insist for api.openai.com.\n    if (!this.apiKey && baseUrl.includes(\"api.openai.com\")) {\n      throw new InferenceError(\n        `OpenAI provider needs ${apiKeyEnv} set (or point baseUrl at a local server)`,\n      );\n    }\n    this.schemaName = options.schemaName ?? \"result\";\n  }\n\n  provider(): string {\n    return \"openai\";\n  }\n\n  modelName(): string {\n    return this.model;\n  }\n\n  private async chat(body: Record<string, unknown>): Promise<ChatResponse> {\n    const response = await fetch(\n      `${this.baseUrl.replace(/\\/$/, \"\")}/chat/completions`,\n      {\n        method: \"POST\",\n        headers: {\n          \"content-type\": \"application/json\",\n          ...(this.apiKey ? { authorization: `Bearer ${this.apiKey}` } : {}),\n        },\n        body: JSON.stringify(body),\n      },\n    );\n    const json = (await response.json().catch(() => ({}))) as ChatResponse;\n    if (!response.ok) {\n      const message = json.error?.message ?? `HTTP ${response.status}`;\n      throw new Error(`${message}`);\n    }\n    return json;\n  }\n\n  async completeJSON(req: CompleteJSONRequest): Promise<CompleteJSONResponse> {\n    const base = {\n      model: this.model,\n      temperature: req.temperature,\n      messages: [\n        { role: \"system\", content: req.system },\n        { role: \"user\", content: req.user },\n      ],\n    };\n\n    let response: ChatResponse;\n    if (this.supportsJsonSchema) {\n      try {\n        response = await this.chat({\n          ...base,\n          response_format: {\n            type: \"json_schema\",\n            json_schema: {\n              name: this.schemaName,\n              strict: true,\n              schema: toStrictSchema(req.schema),\n            },\n          },\n        });\n      } catch (e) {\n        const message = e instanceof Error ? e.message : String(e);\n        // Fall back on schema/format complaints AND on opaque 400s (gateways\n        // that reject response_format without a parseable error body).\n        if (\n          !/response_format|json_schema|schema/i.test(message) &&\n          message !== \"HTTP 400\"\n        ) {\n          throw e;\n        }\n        this.supportsJsonSchema = false;\n        response = await this.jsonObjectFallback(base, req);\n      }\n    } else {\n      response = await this.jsonObjectFallback(base, req);\n    }\n\n    const content = response.choices?.[0]?.message?.content;\n    if (!content) throw new Error(\"Empty completion response\");\n    return {\n      json: stripNulls(extractJson(content)),\n      usage:\n        response.usage?.prompt_tokens != null\n          ? {\n              inputTokens: response.usage.prompt_tokens ?? 0,\n              outputTokens: response.usage.completion_tokens ?? 0,\n            }\n          : undefined,\n    };\n  }\n\n  private jsonObjectFallback(\n    base: Record<string, unknown>,\n    req: CompleteJSONRequest,\n  ): Promise<ChatResponse> {\n    return this.chat({\n      ...base,\n      messages: [\n        {\n          role: \"system\",\n          content: `${req.system}\\n\\nRespond with ONLY a JSON object conforming to this JSON Schema:\\n${JSON.stringify(req.schema)}`,\n        },\n        { role: \"user\", content: req.user },\n      ],\n      response_format: { type: \"json_object\" },\n    });\n  }\n}\n","/**\n * Process execution wrapper for subprocess-backed providers. Uses cross-spawn\n * so npm shims resolve on Windows without `shell: true` and its quoting\n * hazards. Large payloads go through `input` (piped stdin) rather than argv —\n * Windows caps the command line at ~32K characters.\n */\nimport spawn from \"cross-spawn\";\nimport type { ExecFn, ExecResult } from \"./providers/types.js\";\n\nexport const realExec: ExecFn = (cmd, opts = {}) => {\n  const [bin, ...args] = cmd;\n  if (!bin) {\n    return Promise.resolve<ExecResult>({\n      code: null,\n      stdout: \"\",\n      stderr: \"\",\n      timedOut: false,\n      spawnError: \"Empty command\",\n    });\n  }\n  return new Promise<ExecResult>((resolvePromise) => {\n    const child = spawn(bin, args, {\n      cwd: opts.cwd,\n      env: { ...process.env, ...(opts.env ?? {}) },\n      stdio: [opts.input != null ? \"pipe\" : \"ignore\", \"pipe\", \"pipe\"],\n    });\n\n    let stdout = \"\";\n    let stderr = \"\";\n    let timedOut = false;\n    let settled = false;\n\n    const timeoutMs = opts.timeoutMs ?? 60000;\n    const timer = setTimeout(() => {\n      timedOut = true;\n      child.kill();\n      // Settle on the timeout itself rather than waiting for 'close'. On POSIX\n      // a child that handles SIGTERM survives kill(), so 'close' would never\n      // fire and the caller would hang forever with no timeout error. (On\n      // Windows kill() is forceful, so this is belt-and-braces there.)\n      settle({ code: null, stdout, stderr, timedOut: true });\n    }, timeoutMs);\n\n    const settle = (result: ExecResult): void => {\n      if (settled) return;\n      settled = true;\n      clearTimeout(timer);\n      resolvePromise(result);\n    };\n\n    if (opts.input != null && child.stdin) {\n      // EPIPE from a child that exits before reading is not our failure.\n      child.stdin.on(\"error\", () => {});\n      child.stdin.end(opts.input);\n    }\n\n    // setEncoding routes chunks through a StringDecoder, so multi-byte UTF-8\n    // characters straddling pipe-chunk boundaries decode correctly; a raw\n    // per-chunk Buffer.toString() would corrupt them nondeterministically.\n    child.stdout?.setEncoding(\"utf8\");\n    child.stderr?.setEncoding(\"utf8\");\n    child.stdout?.on(\"data\", (d: string) => (stdout += d));\n    child.stderr?.on(\"data\", (d: string) => (stderr += d));\n    child.on(\"error\", (e) =>\n      settle({ code: null, stdout, stderr, timedOut, spawnError: e.message }),\n    );\n    child.on(\"close\", (code) => settle({ code, stdout, stderr, timedOut }));\n  });\n};\n","/**\n * Claude CLI provider: shells out to the `claude` CLI using its local\n * authentication — no API key needed. Token usage is not reported, so cost\n * shows as unknown.\n */\nimport { realExec } from \"../exec.js\";\nimport { extractJson } from \"./openai-compat.js\";\nimport type {\n  CompleteJSONRequest,\n  CompleteJSONResponse,\n  ExecFn,\n  InferenceProvider,\n} from \"./types.js\";\n\nexport class ClaudeCliProvider implements InferenceProvider {\n  constructor(\n    private readonly model: string,\n    private readonly command: string = \"claude\",\n    private readonly exec: ExecFn = realExec,\n    private readonly timeoutMs: number = 180000,\n  ) {}\n\n  provider(): string {\n    return \"claude-cli\";\n  }\n\n  modelName(): string {\n    return this.model;\n  }\n\n  async completeJSON(req: CompleteJSONRequest): Promise<CompleteJSONResponse> {\n    const prompt = [\n      req.user,\n      \"\",\n      \"Respond with ONLY a JSON object conforming to this JSON Schema — no\",\n      \"prose, no markdown fences:\",\n      JSON.stringify(req.schema),\n    ].join(\"\\n\");\n\n    // The prompt is piped via stdin: user content routinely exceeds the ~32K\n    // Windows command-line limit when passed as an argument.\n    const result = await this.exec(\n      [\n        this.command,\n        \"-p\",\n        \"--append-system-prompt\",\n        req.system,\n        \"--output-format\",\n        \"json\",\n        \"--model\",\n        this.model,\n      ],\n      { timeoutMs: this.timeoutMs, input: prompt },\n    );\n\n    if (result.spawnError) {\n      throw new Error(\n        `Failed to run ${this.command}: ${result.spawnError} (is the Claude CLI installed?)`,\n      );\n    }\n    if (result.timedOut) throw new Error(\"Claude CLI timed out\");\n    if (result.code !== 0) {\n      throw new Error(\n        `Claude CLI exited ${result.code}: ${result.stderr.trim().slice(-300)}`,\n      );\n    }\n\n    // --output-format json wraps the answer: { result: \"...\", ... }\n    //\n    // When it isn't JSON at all the CLI is talking to a human, not to us — an\n    // update banner, a login prompt, a proxy interception page. A bare\n    // `SyntaxError: Unexpected token 'W'` would reach the end user of a\n    // consuming CLI naming neither the culprit nor the fix, so say who printed\n    // it and quote enough for a login prompt to be recognisable on sight.\n    // Collapsed to one line and capped, like the stderr tail above, so a\n    // megabyte of HTML cannot become the error message.\n    let wrapper: { result?: string };\n    try {\n      wrapper = JSON.parse(result.stdout) as { result?: string };\n    } catch {\n      const excerpt = result.stdout.trim().replace(/\\s+/g, \" \").slice(0, 200);\n      throw new Error(\n        `Claude CLI printed non-JSON output (is it logged in?): ${excerpt || \"(no output)\"}`,\n      );\n    }\n    if (typeof wrapper.result !== \"string\") {\n      throw new Error(\"Claude CLI returned no result field\");\n    }\n    return { json: extractJson(wrapper.result) };\n  }\n}\n","/**\n * Mock provider for tests and offline development. Responds with scripted\n * results in order, cycling when exhausted. Exported from the public API so\n * downstream consumers can test their own pipelines without a live provider.\n */\nimport type {\n  CompleteJSONRequest,\n  CompleteJSONResponse,\n  InferenceProvider,\n  TokenUsage,\n} from \"./types.js\";\n\nexport type MockResponse =\n  | { json: unknown; usage?: TokenUsage }\n  | { error: string };\n\nexport class MockProvider implements InferenceProvider {\n  private calls = 0;\n  /** Every request seen, in order — assert against this in tests. */\n  public readonly requests: CompleteJSONRequest[] = [];\n\n  constructor(\n    private readonly responses: MockResponse[],\n    private readonly model = \"mock-model\",\n  ) {\n    if (responses.length === 0) {\n      throw new Error(\"MockProvider needs at least one scripted response\");\n    }\n  }\n\n  provider(): string {\n    return \"mock\";\n  }\n\n  modelName(): string {\n    return this.model;\n  }\n\n  completeJSON(req: CompleteJSONRequest): Promise<CompleteJSONResponse> {\n    this.requests.push(req);\n    const response = this.responses[this.calls % this.responses.length]!;\n    this.calls += 1;\n    if (\"error\" in response) {\n      return Promise.reject(new Error(response.error));\n    }\n    return Promise.resolve({\n      json: response.json,\n      usage: response.usage ?? { inputTokens: 500, outputTokens: 100 },\n    });\n  }\n}\n\n/** Convenience: a scripted response shaped like the canonical judge verdict. */\nexport function mockVerdict(\n  match: \"pass\" | \"fail\" | \"partial\",\n  confidence: number,\n  overrides: Partial<{\n    claim: string;\n    observed: string;\n    reasoning: string;\n  }> = {},\n): { json: unknown } {\n  return {\n    json: {\n      claim: overrides.claim ?? \"The assertion under test\",\n      observed: overrides.observed ?? \"Observed content\",\n      match,\n      confidence,\n      reasoning: overrides.reasoning ?? \"Scripted mock reasoning\",\n    },\n  };\n}\n","/**\n * Content-addressed JSON cache for inference results, so a cached run replays\n * identically instead of re-billing a nondeterministic call.\n *\n * Key composition stays with the caller: each consumer has a different notion\n * of what should invalidate an entry (page body, prompt version, ensemble\n * size, requested fields). `buildCacheKey` just hashes the parts you name.\n */\nimport { createHash } from \"node:crypto\";\nimport { existsSync, mkdirSync, readFileSync, writeFileSync } from \"node:fs\";\nimport { join } from \"node:path\";\n\nexport function sha256(text: string): string {\n  return createHash(\"sha256\").update(text, \"utf8\").digest(\"hex\");\n}\n\n/**\n * Hash an ordered list of key parts into a cache key. Long parts (page bodies,\n * rendered traces) should be pre-hashed with `sha256` by the caller so the\n * joined string stays small.\n */\nexport function buildCacheKey(parts: string[]): string {\n  // Length-prefixed, not plain-joined: a part that itself contains the\n  // separator would otherwise let two different compositions — [\"a|b\", \"c\"]\n  // and [\"a\", \"b|c\"] — share one cached result.\n  return sha256(parts.map((p) => `${p.length}:${p}`).join(\"|\"));\n}\n\nexport class JsonCache<T> {\n  /** Cache-write failures warn once per process, not once per entry. */\n  private warned = false;\n\n  constructor(\n    private readonly dir: string,\n    private readonly enabled: boolean = true,\n    /** Prefix for the one-time write-failure warning. */\n    private readonly label: string = \"inference\",\n  ) {}\n\n  get(key: string): T | undefined {\n    if (!this.enabled) return undefined;\n    const path = join(this.dir, `${key}.json`);\n    if (!existsSync(path)) return undefined;\n    try {\n      return JSON.parse(readFileSync(path, \"utf8\")) as T;\n    } catch {\n      return undefined; // Corrupt cache entry — treat as a miss.\n    }\n  }\n\n  set(key: string, value: T): void {\n    if (!this.enabled) return;\n    // The cache is an optimization: a write failure (read-only workspace, full\n    // disk, long path) must never abort a run whose inference already\n    // succeeded and was already paid for.\n    try {\n      mkdirSync(this.dir, { recursive: true });\n      writeFileSync(\n        join(this.dir, `${key}.json`),\n        JSON.stringify(value, null, 2),\n      );\n    } catch (e) {\n      if (!this.warned) {\n        this.warned = true;\n        console.warn(\n          `${this.label}: could not write the cache at ${this.dir} (${\n            e instanceof Error ? e.message : String(e)\n          }). Continuing without caching.`,\n        );\n      }\n    }\n  }\n}\n","/**\n * Curated GGUF models for the in-process `llama-cpp` provider, plus the\n * selector-to-model resolution that sits in front of them.\n *\n * Every entry is an unsloth Gemma 4 QAT build. One family keeps the chat\n * template identical across tiers, and QAT (quantization-aware training) is\n * built for 4-bit deployment, so it beats a standard Q4 quant of the same\n * model at a smaller file size (E4B: QAT UD-Q4_K_XL 4.22 GB vs stock\n * Q4_K_M 4.98 GB). Gemma 4's IFEval scores — 94.6 / 96.7 / 97.2 across the\n * three tiers — are what actually predicts schema-conforming output.\n *\n * Entries pin an exact blob path rather than a `:QUANT` tag. That is a hard\n * requirement, not a style preference: the QAT repos ship NO Q4_K_M — only\n * UD-Q4_K_XL and UD-Q2_K_XL — so a tag would fail to resolve outright. They\n * also carry `mmproj-*.gguf` (the ~1 GB vision projector; Gemma 4 is natively\n * multimodal) and `mtp-*.gguf` next to the weights, which text-only judging\n * must not download. A pinned path also cannot silently re-point underneath a\n * cache key that already claims it.\n */\nimport { readdirSync } from \"node:fs\";\nimport { homedir } from \"node:os\";\nimport { join } from \"node:path\";\nimport { InferenceError } from \"../types.js\";\n\n/**\n * Where this library downloads weights — its OWN directory, not\n * node-llama-cpp's global `~/.node-llama-cpp/models`.\n *\n * That default is shared: node-llama-cpp's CLI writes there, as does anything\n * else on the machine using it. Owning a directory outright means clearing it\n * can never destroy a model this library did not download, and it keeps one\n * copy shared across every consumer of this package on the machine.\n *\n * `INFERENCE_MODELS_DIR` overrides it — useful for CI or a shared volume.\n */\nexport function defaultLlamaModelsDirectory(): string {\n  return (\n    process.env[\"INFERENCE_MODELS_DIR\"] ||\n    join(homedir(), \".hawkeyexl-inference\", \"models\")\n  );\n}\n\n/** Size tiers, smallest first. Order is load-bearing for `tierForBudget`. */\nexport const LLAMA_TIERS = [\"fast\", \"balanced\", \"quality\"] as const;\nexport type LlamaTier = (typeof LLAMA_TIERS)[number];\n\n/** Model selectors — resolved against hardware, never used as a cache key. */\nexport const LLAMA_SELECTORS = [\"auto\", ...LLAMA_TIERS] as const;\nexport type LlamaSelector = (typeof LLAMA_SELECTORS)[number];\n\nexport interface LlamaModelEntry {\n  /** `hf:` URI pinned to one blob, handed to `resolveModelFile` as-is. */\n  readonly uri: string;\n  /** Size of that blob in bytes, as reported by the Hugging Face API. */\n  readonly sizeBytes: number;\n  readonly license: string;\n  /** Absent for entries that are selectable by alias but not by tier. */\n  readonly tier?: LlamaTier;\n  /** Human note for `LLAMA_MODELS` readers deciding what to download. */\n  readonly notes: string;\n}\n\n/**\n * Frozen per entry, not just at the top level.\n *\n * A shallow freeze leaves the entries writable, and this catalog is exported\n * for consumers to read: a stray write to `sizeBytes` silently re-points\n * `tierForBudget` process-wide, and a write to `uri` defeats the pinned-blob\n * invariant the whole catalog exists to hold (ADR 01003).\n */\nexport const LLAMA_MODELS: Readonly<Record<string, LlamaModelEntry>> =\n  deepFreezeEntries({\n    \"gemma-4-e2b\": {\n      uri: \"hf:unsloth/gemma-4-E2B-it-qat-GGUF/gemma-4-E2B-it-qat-UD-Q4_K_XL.gguf\",\n      sizeBytes: 2_620_370_976,\n      license: \"Apache-2.0\",\n      tier: \"fast\",\n      notes: \"IFEval 94.6. Smallest Gemma 4; the floor for this family.\",\n    },\n    \"gemma-4-e4b\": {\n      uri: \"hf:unsloth/gemma-4-E4B-it-qat-GGUF/gemma-4-E4B-it-qat-UD-Q4_K_XL.gguf\",\n      sizeBytes: 4_215_695_776,\n      license: \"Apache-2.0\",\n      tier: \"balanced\",\n      notes: \"IFEval 96.7. The default for most machines.\",\n    },\n    \"gemma-4-12b\": {\n      uri: \"hf:unsloth/gemma-4-12B-it-qat-GGUF/gemma-4-12B-it-qat-UD-Q4_K_XL.gguf\",\n      sizeBytes: 6_716_356_800,\n      license: \"Apache-2.0\",\n      tier: \"quality\",\n      notes: \"IFEval 97.2. Dense 12B; wants a GPU or plenty of RAM.\",\n    },\n    \"gemma-4-26b-a4b\": {\n      uri: \"hf:unsloth/gemma-4-26B-A4B-it-qat-GGUF/gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf\",\n      sizeBytes: 14_249_047_104,\n      license: \"Apache-2.0\",\n      notes:\n        \"MoE: 25.2B total, 3.8B active — infers near E4B speed if it fits in memory.\",\n    },\n    \"gemma-4-e2b-q2\": {\n      uri: \"hf:unsloth/gemma-4-E2B-it-qat-GGUF/gemma-4-E2B-it-qat-UD-Q2_K_XL.gguf\",\n      sizeBytes: 2_186_186_784,\n      license: \"Apache-2.0\",\n      notes: \"Q2 build of the fast tier; smallest download, lowest fidelity.\",\n    },\n  });\n\nfunction deepFreezeEntries<T extends Record<string, object>>(\n  catalog: T,\n): Readonly<T> {\n  for (const entry of Object.values(catalog)) Object.freeze(entry);\n  return Object.freeze(catalog);\n}\n\n/** The alias backing each tier, used by `auto` and the tier keywords. */\nconst TIER_ALIAS: Record<LlamaTier, string> = {\n  fast: \"gemma-4-e2b\",\n  balanced: \"gemma-4-e4b\",\n  quality: \"gemma-4-12b\",\n};\n\nexport function isLlamaSelector(model: string): model is LlamaSelector {\n  return (LLAMA_SELECTORS as readonly string[]).includes(model);\n}\n\n/**\n * Weights are only part of the cost — the KV cache at a real context length,\n * the OS, and whatever else the machine is doing all want memory too. Requiring\n * several times the file size keeps `auto` from picking a model that technically\n * loads and then thrashes.\n */\nconst MEMORY_HEADROOM = 3.5;\n\n/**\n * Largest tier whose weights fit the memory budget with headroom. Lands at\n * roughly: >=24 GB -> quality, >=15 GB -> balanced, else fast.\n *\n * Sized off the catalog's recorded bytes rather than parameter counts: Gemma\n * 4's E-series are per-layer-embedding models whose footprint does not track\n * \"effective params\" (E4B is 4.5B effective but 15 GB at BF16).\n */\nexport function tierForBudget(budgetBytes: number): LlamaTier {\n  let chosen: LlamaTier = \"fast\";\n  for (const tier of LLAMA_TIERS) {\n    const entry = LLAMA_MODELS[TIER_ALIAS[tier]]!;\n    if (entry.sizeBytes * MEMORY_HEADROOM <= budgetBytes) chosen = tier;\n  }\n  // Falls back to the smallest tier rather than refusing: a machine too small\n  // for `fast` will thrash, but that is the caller's call to make, not ours.\n  return chosen;\n}\n\n/**\n * Catalog alias backing a tier. Selectors resolve to this rather than to a raw\n * URI so the identity — and therefore the cache key — stays human-readable.\n */\nexport function aliasForTier(tier: LlamaTier): string {\n  return TIER_ALIAS[tier];\n}\n\n/** The pinned URI backing a tier keyword. */\nexport function uriForTier(tier: LlamaTier): string {\n  return LLAMA_MODELS[TIER_ALIAS[tier]]!.uri;\n}\n\n/**\n * Turn a concrete model reference — a curated alias, an `hf:` URI, or a local\n * path — into something `resolveModelFile` accepts.\n *\n * Selectors are rejected rather than guessed at: they need a hardware probe,\n * which is async, and this runs on the synchronous cache-key path. Resolving\n * one here from RAM alone would emit a key naming a model the provider then\n * did not load.\n */\nexport function resolveLlamaModelRef(model: string): string {\n  if (isLlamaSelector(model)) {\n    throw new InferenceError(\n      `llama-cpp model \"${model}\" is a selector and needs a hardware probe to ` +\n        `resolve. Use resolveProviderIdentityAsync/makeProviderAsync, or name a ` +\n        `concrete model (e.g. \"gemma-4-e4b\").`,\n    );\n  }\n  const entry = LLAMA_MODELS[model];\n  if (entry) return entry.uri;\n  if (isModelPathOrUri(model)) return model;\n  throw new InferenceError(\n    `Unknown llama-cpp model \"${model}\". Use a selector (${LLAMA_SELECTORS.join(\n      \", \",\n    )}), a curated alias (${Object.keys(LLAMA_MODELS).join(\n      \", \",\n    )}), an hf: URI, or a path to a .gguf file.`,\n  );\n}\n\n/**\n * The catalog's pinned filename for a model reference.\n *\n * Strips a `#branch` fragment (node-llama-cpp accepts\n * `hf:user/repo/file.gguf#branch`) and splits on both separators, so a Windows\n * path resolves too. Getting either wrong makes callers silently match nothing.\n */\nexport function blobNameFor(model: string): string {\n  const ref = resolveLlamaModelRef(model).split(\"#\")[0]!;\n  return ref.split(/[/\\\\]/).pop()!;\n}\n\n/**\n * Does an on-disk entry belong to this model?\n *\n * node-llama-cpp prefixes downloads with `hf_<user>_`, so match by SUFFIX\n * rather than equality — that survives a change to their naming scheme.\n */\nexport function matchesModelBlob(entry: string, blobName: string): boolean {\n  const base = entry.replace(/\\.ipull$/, \"\");\n  if (base.endsWith(blobName)) return true;\n  // Split models land as `<stem>-00001-of-00003.gguf`; every part belongs to\n  // the same model, so matching the stem covers the whole set.\n  const stem = blobName.replace(/\\.gguf$/, \"\");\n  return new RegExp(`${escapeRegExp(stem)}-\\\\d{5}-of-\\\\d{5}\\\\.gguf$`).test(base);\n}\n\nfunction escapeRegExp(text: string): string {\n  return text.replace(/[.*+?^${}()|[\\]\\\\]/g, \"\\\\$&\");\n}\n\n/**\n * Are this model's weights already on disk?\n *\n * A `.ipull` partial counts as NOT downloaded: it cannot be loaded, so\n * treating it as present would skip the download warning and then stall on a\n * download anyway.\n */\nexport function isModelDownloaded(model: string, directory: string): boolean {\n  const blobName = blobNameFor(model);\n  return listModelDirectory(directory).some(\n    (entry) => !entry.endsWith(\".ipull\") && matchesModelBlob(entry, blobName),\n  );\n}\n\n/** Top level only — a nested directory is not ours to walk. */\nexport function listModelDirectory(directory: string): string[] {\n  try {\n    return readdirSync(directory, { withFileTypes: true })\n      .filter((entry) => entry.isFile())\n      .map((entry) => entry.name);\n  } catch {\n    // No directory means nothing was ever downloaded — not an error.\n    return [];\n  }\n}\n\n/**\n * A bare unknown word is a typo'd alias, not a model — catch it early rather\n * than letting it reach the downloader as a doomed repo name.\n *\n * Accepts exactly what the error message in `resolveLlamaModelRef` promises: a\n * recognised URI scheme, or something that names a `.gguf` file. A bare\n * `user/repo` is deliberately NOT a model reference — it would otherwise slip\n * past this guard and fail deep inside the downloader with a far worse\n * message. Any real path to weights ends in `.gguf`, on every platform.\n */\nfunction isModelPathOrUri(model: string): boolean {\n  return (\n    /^(hf|huggingface):/i.test(model) ||\n    /^https?:\\/\\//i.test(model) ||\n    /^(hf|huggingface)\\.co\\//i.test(model) ||\n    model.endsWith(\".gguf\")\n  );\n}\n","/**\n * Installing the optional `node-llama-cpp` peer on demand.\n *\n * Detection ends at `llama-cpp` precisely because it needs no key and no\n * account — but that rung is only reachable if the native binding is present,\n * and npm does not install optional peers. Without this module a machine with\n * no credentials and no binding falls off the end of the chain, which is the\n * one case auto-detection exists to serve.\n *\n * The install goes into a directory this library OWNS — never the consumer's\n * `node_modules`, `package.json` or lockfile. That is the same reasoning ADR\n * 01003 applied to weights: owning a directory removes the hazard instead of\n * defending against it. A consumer's dependency manifest is theirs, and a\n * library that edits it has broken reproducible installs for everyone\n * downstream.\n */\nimport {\n  existsSync,\n  mkdirSync,\n  rmSync,\n  statSync,\n  writeFileSync,\n} from \"node:fs\";\nimport { homedir } from \"node:os\";\nimport { join } from \"node:path\";\nimport { pathToFileURL } from \"node:url\";\nimport { InferenceError } from \"../types.js\";\nimport { realExec } from \"../exec.js\";\nimport type { ExecFn } from \"./types.js\";\n\n/** Pinned to the peer range: below 3.19.0 nothing in the catalog loads. */\nconst PACKAGE_SPEC = \"node-llama-cpp@^3.19.0\";\n\n/**\n * The shim doubles as the readiness marker — it is written only after npm\n * exits 0, so its presence means \"this prefix is complete\".\n */\nconst SHIM = \"loader.mjs\";\nconst LOCK = \".install.lock\";\n\n/**\n * Generous by necessity. Prebuilt binaries cover win32-x64, darwin-arm64 and\n * linux-x64; anything else falls back to a CMake build that genuinely takes\n * minutes. `realExec` defaults to 60s, which would kill it half-built.\n */\nconst INSTALL_TIMEOUT_MS = 900_000;\n\n/** How long to wait for another process's install before giving up. */\nconst LOCK_WAIT_MS = INSTALL_TIMEOUT_MS;\n/** A lock older than this belonged to a process that died holding it. */\nconst LOCK_STALE_MS = INSTALL_TIMEOUT_MS + 60_000;\n\nexport interface RuntimeInstallOptions {\n  /** Defaults to this library's own runtime directory. */\n  directory?: string;\n  /** Injected for tests; defaults to the real process runner. */\n  exec?: ExecFn;\n  /** Injected for tests; defaults to `process.env`. */\n  env?: Record<string, string | undefined>;\n  timeoutMs?: number;\n  /** Test seam: how the shim is imported once it exists. */\n  importShim?: (url: string) => Promise<unknown>;\n  /**\n   * Test seam: how the consumer's own copy is looked for.\n   *\n   * Whether an optional peer is installed is a property of the machine, and the\n   * behaviour that matters most here — that probing installs nothing — is\n   * unobservable on a machine that already has it. Injecting the lookup is the\n   * only way to assert it deterministically.\n   */\n  probeImport?: () => Promise<unknown>;\n}\n\n/**\n * Where this library installs the binding — its OWN directory, beside the\n * models directory and for the same reason.\n *\n * `INFERENCE_RUNTIME_DIR` overrides it, mirroring `INFERENCE_MODELS_DIR`.\n */\nexport function defaultLlamaRuntimeDirectory(\n  env: Record<string, string | undefined> = process.env,\n): string {\n  return (\n    env[\"INFERENCE_RUNTIME_DIR\"] ||\n    join(homedir(), \".hawkeyexl-inference\", \"runtime\")\n  );\n}\n\n/**\n * Is this import failure \"the package is not here\", as opposed to \"the package\n * is here and broken\"?\n *\n * The distinction decides whether installing can possibly help. `ERR_MODULE_NOT_FOUND`\n * is what Node raises for a missing bare specifier; `MODULE_NOT_FOUND` is its\n * CJS spelling. Anything else — a failed `dlopen`, an unsupported Node, a\n * package whose `exports` do not match — means the package resolved and then\n * failed, and no amount of reinstalling changes that.\n */\nexport function isModuleNotFound(e: unknown): boolean {\n  const code = (e as NodeJS.ErrnoException | undefined)?.code;\n  return code === \"ERR_MODULE_NOT_FOUND\" || code === \"MODULE_NOT_FOUND\";\n}\n\nfunction describe(e: unknown): string {\n  return e instanceof Error ? e.message : String(e);\n}\n\n/** Whether the binding can be used, and at what cost, WITHOUT installing it. */\nexport type RuntimeStatus =\n  /** Importable right now — either the consumer's own copy or a filled prefix. */\n  | { state: \"present\" }\n  /** Absent, but an install is permitted. Using it will fetch a native module. */\n  | { state: \"installable\"; directory: string }\n  /** Absent and installing is refused, so this provider cannot be used. */\n  | { state: \"refused\"; reason: string };\n\n/**\n * Can the local runtime be used, and would using it install anything?\n *\n * Detection needs this rather than simply calling `getMemoryBudgetBytes`:\n * that goes through `importNodeLlamaCpp` and would install, which turns\n * `availableProviders()` — a function whose entire job is to REPORT what is\n * usable — into one that changes what is usable. Probing must not have the side\n * effect it is probing for.\n */\nexport async function nodeLlamaCppStatus(\n  options: RuntimeInstallOptions = {},\n): Promise<RuntimeStatus> {\n  const env = options.env ?? process.env;\n  const directory = options.directory ?? defaultLlamaRuntimeDirectory(env);\n\n  const probeImport =\n    options.probeImport ?? ((): Promise<unknown> => import(\"node-llama-cpp\"));\n  try {\n    await probeImport();\n    return { state: \"present\" };\n  } catch (e) {\n    // Only a genuine \"no such package\" means absent. An ABI mismatch, a missing\n    // system library, or an unsupported Node all fail here too, and installing\n    // over them would fetch the same broken package again while burying the\n    // real cause under a download.\n    if (!isModuleNotFound(e)) {\n      return {\n        state: \"refused\",\n        reason: `node-llama-cpp is installed but failed to load (${describe(e)})`,\n      };\n    }\n    // Genuinely absent — fall through to our own prefix.\n  }\n  if (existsSync(join(directory, SHIM))) return { state: \"present\" };\n\n  if ((env[\"INFERENCE_NO_AUTO_INSTALL\"] ?? \"\") !== \"\") {\n    return {\n      state: \"refused\",\n      reason: `node-llama-cpp is not installed and INFERENCE_NO_AUTO_INSTALL is set`,\n    };\n  }\n  return { state: \"installable\", directory };\n}\n\n/**\n * In-flight installs, keyed by directory. docmeta and the ensemble runner both\n * work several items at once, so without this every worker that misses the\n * binding would spawn its own npm against one prefix.\n */\nconst installs = new Map<string, Promise<unknown>>();\n\nlet warnedInstall = false;\n\n/** Test seam: forget in-flight installs and re-arm the one-time warning. */\nexport function resetRuntimeInstall(): void {\n  installs.clear();\n  warnedInstall = false;\n}\n\n/**\n * Import `node-llama-cpp` from the library-owned prefix, installing it first if\n * it is not there.\n *\n * Callers reach this only after a plain `import(\"node-llama-cpp\")` has already\n * failed — a consumer who installed the peer themselves never gets here.\n */\nexport function importNodeLlamaCpp(\n  options: RuntimeInstallOptions = {},\n): Promise<unknown> {\n  const env = options.env ?? process.env;\n  const directory = options.directory ?? defaultLlamaRuntimeDirectory(env);\n\n  const existing = installs.get(directory);\n  if (existing) return existing;\n\n  const pending = fromPrefix(directory, env, options);\n  // Drop a failed attempt so the next call retries: a download killed by a\n  // flaky network must not poison the runtime for the rest of the process —\n  // the same rule `load()` applies to weights.\n  const guarded = pending.catch((e: unknown) => {\n    if (installs.get(directory) === guarded) installs.delete(directory);\n    throw e;\n  });\n  installs.set(directory, guarded);\n  return guarded;\n}\n\nasync function fromPrefix(\n  directory: string,\n  env: Record<string, string | undefined>,\n  options: RuntimeInstallOptions,\n): Promise<unknown> {\n  const importShim =\n    options.importShim ?? ((url: string): Promise<unknown> => import(url));\n  const shim = join(directory, SHIM);\n\n  if (existsSync(shim)) return importShim(pathToFileURL(shim).href);\n\n  if ((env[\"INFERENCE_NO_AUTO_INSTALL\"] ?? \"\") !== \"\") {\n    throw new InferenceError(\n      `The llama-cpp provider needs node-llama-cpp, and INFERENCE_NO_AUTO_INSTALL ` +\n        `is set. Install it yourself (npm i ${PACKAGE_SPEC}), unset ` +\n        `INFERENCE_NO_AUTO_INSTALL to allow installing into ${directory}, or name ` +\n        `a different provider.`,\n    );\n  }\n\n  mkdirSync(directory, { recursive: true });\n  await withLock(directory, async () => {\n    // Another process may have finished while we waited for the lock.\n    if (existsSync(shim)) return;\n    warnInstalling(directory);\n    await runInstall(directory, env, options);\n    // Only now is the prefix complete, so only now does it get its marker.\n    writeFileSync(shim, `export * from \"node-llama-cpp\";\\n`, \"utf8\");\n  });\n\n  return importShim(pathToFileURL(shim).href);\n}\n\nasync function runInstall(\n  directory: string,\n  env: Record<string, string | undefined>,\n  options: RuntimeInstallOptions,\n): Promise<void> {\n  // Anchor npm to this directory. Without a manifest here npm walks UP looking\n  // for one, and would install into whatever project happens to be above us —\n  // the exact mutation this module exists to avoid.\n  const manifest = join(directory, \"package.json\");\n  if (!existsSync(manifest)) {\n    writeFileSync(\n      manifest,\n      `${JSON.stringify(\n        {\n          name: \"hawkeyexl-inference-runtime\",\n          version: \"0.0.0\",\n          private: true,\n          description:\n            \"Auto-installed runtime for @hawkeyexl/inference. Safe to delete.\",\n        },\n        null,\n        2,\n      )}\\n`,\n      \"utf8\",\n    );\n  }\n\n  const exec = options.exec ?? realExec;\n  const result = await exec(\n    [\n      \"npm\",\n      \"install\",\n      \"--prefix\",\n      directory,\n      PACKAGE_SPEC,\n      \"--no-audit\",\n      \"--no-fund\",\n    ],\n    { timeoutMs: options.timeoutMs ?? INSTALL_TIMEOUT_MS, env },\n  );\n\n  if (result.spawnError != null) {\n    throw new InferenceError(\n      `Could not run npm to install node-llama-cpp (${result.spawnError}). ` +\n        `Install it yourself with: npm i ${PACKAGE_SPEC}, or name a provider ` +\n        `that does not need it.`,\n    );\n  }\n  if (result.timedOut) {\n    throw new InferenceError(\n      `Installing node-llama-cpp into ${directory} timed out. A source build ` +\n        `can take a while — retry, raise the timeout, or install it yourself ` +\n        `with: npm i ${PACKAGE_SPEC}.`,\n    );\n  }\n  if (result.code !== 0) {\n    throw new InferenceError(\n      `Installing node-llama-cpp into ${directory} failed (exit ${String(\n        result.code,\n      )}).\\n${tail(result.stderr || result.stdout)}\\n` +\n        `Install it yourself with: npm i ${PACKAGE_SPEC}, or name a provider ` +\n        `that does not need it.`,\n    );\n  }\n}\n\n/** Enough npm output to diagnose the failure, not enough to bury the advice. */\nfunction tail(output: string, lines = 12): string {\n  return output.trimEnd().split(/\\r?\\n/).slice(-lines).join(\"\\n\");\n}\n\n/**\n * Pulling a native module is not free, and the weights that follow are far less\n * free. Say so before it starts, not after a build has already stalled on it.\n */\nfunction warnInstalling(directory: string): void {\n  if (warnedInstall) return;\n  warnedInstall = true;\n  console.warn(\n    `inference: node-llama-cpp is not installed — fetching it into ${directory} ` +\n      `so the local model can run. This is a one-time native install; set ` +\n      `INFERENCE_NO_AUTO_INSTALL=1 to refuse it, or name a provider that does ` +\n      `not need it.`,\n  );\n}\n\n/**\n * Cross-process guard around one prefix.\n *\n * The in-process memo covers a worker pool inside one run; this covers two runs\n * started at once, where two npm processes writing one `node_modules` is how a\n * prefix ends up half-written.\n */\nasync function withLock(\n  directory: string,\n  fn: () => Promise<void>,\n): Promise<void> {\n  const lock = join(directory, LOCK);\n  const deadline = Date.now() + LOCK_WAIT_MS;\n\n  for (;;) {\n    try {\n      writeFileSync(lock, String(process.pid), { flag: \"wx\" });\n      break;\n    } catch (e) {\n      if ((e as NodeJS.ErrnoException).code !== \"EEXIST\") throw e;\n      if (ageOf(lock) > LOCK_STALE_MS) {\n        // Whoever held this died. Reclaiming a stale lock is safer than\n        // blocking forever on a process that will never return.\n        rmSync(lock, { force: true });\n        continue;\n      }\n      if (existsSync(join(directory, SHIM))) return;\n      if (Date.now() > deadline) {\n        throw new InferenceError(\n          `Timed out waiting for another process to install node-llama-cpp into ` +\n            `${directory}. If nothing else is running, remove ${lock} and retry.`,\n        );\n      }\n      await delay(250);\n    }\n  }\n\n  try {\n    await fn();\n  } finally {\n    rmSync(lock, { force: true });\n  }\n}\n\n/** Infinity for a lock that vanished mid-check, so the caller retries. */\nfunction ageOf(path: string): number {\n  try {\n    return Date.now() - statSync(path).mtimeMs;\n  } catch {\n    return Number.POSITIVE_INFINITY;\n  }\n}\n\nfunction delay(ms: number): Promise<void> {\n  return new Promise((resolve) => setTimeout(resolve, ms));\n}\n","/**\n * In-process local inference over GGUF weights via `node-llama-cpp`.\n *\n * Unlike every other provider here, this one owns weights: it downloads them\n * from Hugging Face on first use and holds gigabytes of RAM once loaded. Two\n * consequences shape the design.\n *\n * First, `node-llama-cpp` is a native module with prebuilt binaries per\n * platform and a CMake fallback. It is an OPTIONAL peer dependency reached\n * through a dynamic `import()`, so the four repos consuming this library pay\n * nothing — install cost or toolchain risk — unless they ask for local models.\n *\n * Second, everything real happens behind `LlamaRuntime`. Tests inject a fake\n * and never touch the network, the filesystem, or a GPU (the same seam as\n * `ExecFn` for the Claude CLI provider).\n */\nimport { InferenceError } from \"../types.js\";\nimport { buildCacheKey } from \"../cache.js\";\nimport { extractJson } from \"./openai-compat.js\";\nimport {\n  defaultLlamaModelsDirectory,\n  isLlamaSelector,\n  resolveLlamaModelRef,\n} from \"./llama-models.js\";\nimport { importNodeLlamaCpp, isModuleNotFound } from \"./llama-install.js\";\nimport type {\n  CompleteJSONRequest,\n  CompleteJSONResponse,\n  InferenceProvider,\n  TokenUsage,\n} from \"./types.js\";\n\nexport interface LlamaPromptOptions {\n  /** JSON Schema converted to a GBNF grammar by the runtime. */\n  schema: Record<string, unknown>;\n  temperature: number;\n  /** Thinking budget; 0 disables it. See the note in `completeJSON`. */\n  thoughtTokens: number;\n  maxTokens?: number;\n}\n\nexport interface LlamaPromptResult {\n  text: string;\n  usage?: TokenUsage;\n  /**\n   * Why generation stopped. `\"maxTokens\"` means the output was cut off, so the\n   * text is almost certainly truncated JSON — see the guard in `completeJSON`.\n   */\n  stopReason?: string;\n}\n\nexport interface LlamaSession {\n  prompt(text: string, options: LlamaPromptOptions): Promise<LlamaPromptResult>;\n  dispose(): Promise<void>;\n}\n\nexport interface LlamaLoadedModel {\n  createSession(systemPrompt: string): Promise<LlamaSession>;\n  dispose(): Promise<void>;\n}\n\n/**\n * The whole of `node-llama-cpp` that this provider uses. Kept this narrow so a\n * test fake is a few lines and so the real adapter is the only place that\n * knows the upstream API shape.\n */\nexport interface LlamaRuntime {\n  /**\n   * Resolve an `hf:` URI or path to a local file inside `directory`,\n   * downloading if needed.\n   */\n  resolveModelFile(uri: string, directory: string): Promise<string>;\n  loadModel(path: string): Promise<LlamaLoadedModel>;\n  /** Memory available for weights, in bytes — VRAM if there is a GPU, else RAM. */\n  getMemoryBudgetBytes(): Promise<number>;\n}\n\nexport interface LlamaCppProviderOptions {\n  /** Injected for tests; defaults to the real `node-llama-cpp` adapter. */\n  runtime?: LlamaRuntime;\n  /**\n   * Thinking budget in tokens, default 0.\n   *\n   * Gemma 4 has a thinking mode, but a grammar constrains generation from\n   * token 0 — so an unbudgeted model starts reasoning and gets cut off\n   * mid-thought. Zero is the deterministic choice for judging; raise it if you\n   * want reasoning before the JSON.\n   */\n  thoughtTokens?: number;\n  maxTokens?: number;\n  /**\n   * Where to download and look for weights. Defaults to this library's own\n   * directory — see `defaultLlamaModelsDirectory`.\n   */\n  modelsDirectory?: string;\n}\n\n/**\n * Loaded weights, keyed by directory + URI.\n *\n * `runEnsemble` issues N sequential calls and a load costs seconds and\n * gigabytes, so this is process-wide rather than per-instance: two providers\n * naming the same model share one copy. Values are the in-flight promise so\n * concurrent first calls coalesce instead of loading twice. The directory is\n * part of the key because the same URI in two directories is two files.\n */\nconst loadedModels = new Map<string, Promise<LlamaLoadedModel>>();\n\n/**\n * Free every loaded model.\n *\n * A standalone function rather than a `dispose()` on `InferenceProvider`:\n * adding one to the contract would make all five providers carry a lifecycle\n * only this one has. Short-lived processes can skip it.\n */\nexport async function disposeLlamaModels(): Promise<void> {\n  const pending = [...loadedModels.values()];\n  loadedModels.clear();\n  await Promise.all(\n    pending.map((p) => p.then((m) => m.dispose()).catch(() => undefined)),\n  );\n}\n\nexport class LlamaCppProvider implements InferenceProvider {\n  private readonly uri: string;\n  private readonly runtime: LlamaRuntime;\n  private readonly thoughtTokens: number;\n  private readonly maxTokens: number | undefined;\n  private readonly modelsDirectory: string;\n  /**\n   * Loaded-model key: the same URI in two directories is two different files.\n   * Built with `buildCacheKey` so its parts are length-prefixed — a plain join\n   * would let two different (directory, uri) pairs collide and hand a provider\n   * back the wrong weights.\n   */\n  private readonly cacheKey: string;\n\n  constructor(\n    private readonly model: string,\n    options: LlamaCppProviderOptions = {},\n  ) {\n    if (isLlamaSelector(model)) {\n      throw new InferenceError(\n        `llama-cpp model \"${model}\" is a selector. Constructing a provider ` +\n          `directly needs a concrete model (e.g. \"gemma-4-e4b\") — use ` +\n          `makeProviderAsync to resolve a selector against this machine.`,\n      );\n    }\n    this.uri = resolveLlamaModelRef(model);\n    this.runtime = options.runtime ?? defaultLlamaRuntime();\n    this.thoughtTokens = options.thoughtTokens ?? 0;\n    this.maxTokens = options.maxTokens;\n    this.modelsDirectory =\n      options.modelsDirectory ?? defaultLlamaModelsDirectory();\n    this.cacheKey = buildCacheKey([this.modelsDirectory, this.uri]);\n  }\n\n  provider(): string {\n    return \"llama-cpp\";\n  }\n\n  modelName(): string {\n    return this.model;\n  }\n\n  async completeJSON(req: CompleteJSONRequest): Promise<CompleteJSONResponse> {\n    const model = await this.load();\n    // A fresh session per call: the contract is single-shot, and reusing one\n    // would leak the previous run's turns into this one's context.\n    const session = await model.createSession(systemPromptFor(req));\n    try {\n      const result = await session.prompt(req.user, {\n        schema: req.schema,\n        temperature: req.temperature,\n        thoughtTokens: this.thoughtTokens,\n        ...(this.maxTokens != null ? { maxTokens: this.maxTokens } : {}),\n      });\n      // A run cut off at the token or context limit leaves truncated JSON.\n      // Without this it surfaces as \"failed schema validation\" — or worse,\n      // extractJson's brace-slicing fallback salvages a wrong-but-parseable\n      // object — and the retry burns another full local inference to fail the\n      // same way. Same guard as the Anthropic provider's max_tokens check.\n      if (result.stopReason === \"maxTokens\") {\n        throw new Error(\n          `llama-cpp generation hit the token limit before completing the JSON` +\n            `${this.maxTokens != null ? ` (maxTokens: ${this.maxTokens})` : \"\"}` +\n            ` — raise llamaCpp.maxTokens, or shorten the prompt if the context is full.`,\n        );\n      }\n      return { json: extractJson(result.text), usage: result.usage };\n    } finally {\n      await session.dispose().catch(() => undefined);\n    }\n  }\n\n  private load(): Promise<LlamaLoadedModel> {\n    const existing = loadedModels.get(this.cacheKey);\n    if (existing) return existing;\n    const pending = (async () => {\n      const path = await this.runtime.resolveModelFile(\n        this.uri,\n        this.modelsDirectory,\n      );\n      return this.runtime.loadModel(path);\n    })();\n    // Drop a failed load so the next call retries — a download interrupted by\n    // a flaky network must not poison the model for the rest of the process.\n    // Only evict our OWN entry: a dispose plus a re-load between the failure\n    // and this handler would otherwise orphan the newer model, leaking it.\n    const guarded = pending.catch((e: unknown) => {\n      if (loadedModels.get(this.cacheKey) === guarded) {\n        loadedModels.delete(this.cacheKey);\n      }\n      throw e;\n    });\n    loadedModels.set(this.cacheKey, guarded);\n    return guarded;\n  }\n}\n\n/**\n * The grammar constrains the SHAPE of the output, but `node-llama-cpp` never\n * shows the schema to the model — so `description` fields, which is where\n * consumers put their domain instructions (ADR 01001), would be invisible.\n * Restating the schema is the same fix the Claude CLI provider and the\n * OpenAI json_object fallback already use.\n */\nfunction systemPromptFor(req: CompleteJSONRequest): string {\n  return `${req.system}\\n\\nRespond with ONLY a JSON object conforming to this JSON Schema:\\n${JSON.stringify(\n    req.schema,\n  )}`;\n}\n\n/** Cached so repeated provider construction imports the native module once. */\nlet runtimePromise: Promise<LlamaRuntime> | undefined;\n\n/**\n * Lazy adapter over the real `node-llama-cpp`. Every method defers to the\n * dynamic import, so constructing a provider for a fully-cached run never\n * loads the native binary.\n */\nexport function defaultLlamaRuntime(): LlamaRuntime {\n  const real = (): Promise<LlamaRuntime> =>\n    // Drop a failed init so the next call retries. A GPU that failed to\n    // initialise, or a binary still being extracted by a concurrent install,\n    // must not poison the runtime for the rest of the process — the same rule\n    // `load()` applies to weights.\n    (runtimePromise ??= loadNodeLlamaCpp().catch((e: unknown) => {\n      runtimePromise = undefined;\n      throw e;\n    }));\n  return {\n    resolveModelFile: (uri, directory) =>\n      real().then((r) => r.resolveModelFile(uri, directory)),\n    loadModel: (path) => real().then((r) => r.loadModel(path)),\n    getMemoryBudgetBytes: () => real().then((r) => r.getMemoryBudgetBytes()),\n  };\n}\n\nasync function loadNodeLlamaCpp(): Promise<LlamaRuntime> {\n  let mod: typeof import(\"node-llama-cpp\");\n  try {\n    mod = await import(\"node-llama-cpp\");\n  } catch (e) {\n    // A package that resolved and then failed to load — ABI mismatch, missing\n    // system library, unsupported Node — is not a missing package. Installing\n    // over it would fetch the same broken thing again and replace a precise\n    // error with a download.\n    if (!isModuleNotFound(e)) {\n      throw new InferenceError(\n        `node-llama-cpp is installed but failed to load (${\n          e instanceof Error ? e.message : String(e)\n        }). This is the copy resolved from your own node_modules, so ` +\n          `reinstalling it here will not help — check the Node version and the ` +\n          `platform build.`,\n      );\n    }\n    // Genuinely absent, so fall back to the library's own prefix — installing\n    // it there if needed. npm does not install optional peers, and detection\n    // ends at this provider precisely because it needs no credentials, so\n    // refusing here would strand the one machine `auto` exists to serve.\n    // Resetting `runtimePromise` is the caller's job — see `defaultLlamaRuntime`.\n    mod = (await importNodeLlamaCpp()) as typeof import(\"node-llama-cpp\");\n  }\n\n  const { getLlama, resolveModelFile, LlamaChatSession, TokenMeter } = mod;\n  const llama = await getLlama();\n\n  return {\n    // `directory` is this library's own, not node-llama-cpp's global default —\n    // owning it is what makes `clearLlamaModels` safe.\n    resolveModelFile: (uri, directory) => resolveModelFile(uri, { directory }),\n\n    async loadModel(path) {\n      const model = await llama.loadModel({ modelPath: path });\n      return {\n        async createSession(systemPrompt) {\n          const context = await model.createContext();\n          const sequence = context.getSequence();\n          const session = new LlamaChatSession({\n            contextSequence: sequence,\n            systemPrompt,\n          });\n          return {\n            async prompt(text, options) {\n              const grammar = await llama.createGrammarForJsonSchema(\n                options.schema as Parameters<\n                  typeof llama.createGrammarForJsonSchema\n                >[0],\n              );\n              const before = sequence.tokenMeter.getState();\n              const result = await session.promptWithMeta(text, {\n                grammar,\n                temperature: options.temperature,\n                budgets: { thoughtTokens: options.thoughtTokens },\n                ...(options.maxTokens != null\n                  ? { maxTokens: options.maxTokens }\n                  : {}),\n              });\n              // promptWithMeta does not report usage; the sequence's meter does.\n              const diff = TokenMeter.diff(sequence.tokenMeter, before);\n              return {\n                text: result.responseText,\n                stopReason: result.stopReason,\n                usage: {\n                  inputTokens: diff.usedInputTokens,\n                  outputTokens: diff.usedOutputTokens,\n                },\n              };\n            },\n            async dispose() {\n              await context.dispose();\n            },\n          };\n        },\n        async dispose() {\n          await model.dispose();\n        },\n      };\n    },\n\n    async getMemoryBudgetBytes() {\n      const { totalmem } = await import(\"node:os\");\n      // Half of RAM is what a judge can reasonably claim on a shared machine;\n      // a GPU's free VRAM is usable outright.\n      const ramBudget = totalmem() / 2;\n      try {\n        const vram = await llama.getVramState();\n        // The LARGER of the two, not VRAM in preference to RAM: llama.cpp\n        // offloads the layers that fit onto the GPU and keeps the rest in\n        // system RAM, so a small GPU beside plenty of RAM still runs a big\n        // model. Sizing off VRAM alone would idle most of such a machine.\n        return Math.max(vram.free, ramBudget);\n      } catch {\n        // CPU-only builds and probe failures are normal, never fatal.\n        return ramBudget;\n      }\n    },\n  };\n}\n","/**\n * Which providers can this machine actually use, and which should it pick?\n *\n * Every probe goes through a seam that already exists — environment variables,\n * `ExecFn`, `LlamaRuntime` — so the whole matrix is exercisable offline with no\n * network, no subprocess, and no weights.\n *\n * Detection is async because two of the four probes are: running the Claude CLI\n * and loading the optional `node-llama-cpp` binding. That is why only\n * `resolveProviderIdentityAsync`/`makeProviderAsync` can resolve an `auto`\n * provider, and the synchronous twins throw instead of guessing.\n */\nimport { InferenceError } from \"../types.js\";\nimport { realExec } from \"../exec.js\";\nimport { defaultLlamaRuntime } from \"./llama-cpp.js\";\nimport { nodeLlamaCppStatus } from \"./llama-install.js\";\nimport type { LlamaRuntime } from \"./llama-cpp.js\";\nimport type { ProviderName, ProviderSpec } from \"./index.js\";\n\n/**\n * Priority order. `mock` is deliberately absent: it answers `{ json: {} }`\n * unless scripted, which would sail through as a non-error result — the exact\n * opposite of the \"an errored run is recorded, never coerced\" invariant the\n * consuming eval tools depend on. It must always be asked for by name.\n */\nexport const DETECTION_ORDER: readonly ProviderName[] = [\n  \"anthropic\",\n  \"openai\",\n  \"claude-cli\",\n  \"llama-cpp\",\n];\n\nconst DEFAULT_KEY_ENV: Partial<Record<ProviderName, string>> = {\n  anthropic: \"ANTHROPIC_API_KEY\",\n  openai: \"OPENAI_API_KEY\",\n};\n\ninterface Probe {\n  available: boolean;\n  /** Why not, phrased as advice — this is what the aggregate error prints. */\n  reason?: string;\n}\n\n/**\n * Probes the provider's DEFAULT key variable, deliberately ignoring\n * `spec.apiKeyEnv`.\n *\n * `apiKeyEnv` is one field shared by both API providers, and detection only\n * runs when no provider was named — so a custom name cannot say which provider\n * it belongs to. Honouring it here made a single custom variable satisfy both\n * probes, and anthropic then won on priority: a spec carrying an OpenAI key\n * under a custom name selected `anthropic` and 401'd at call time.\n *\n * A custom `apiKeyEnv` still applies in full once a provider is named — it just\n * cannot be what chooses one.\n */\nfunction hasKey(provider: \"anthropic\" | \"openai\"): boolean {\n  // An empty string is not a key; treating it as one produces a 401 later.\n  return (process.env[DEFAULT_KEY_ENV[provider]!] ?? \"\") !== \"\";\n}\n\n/**\n * Memoised **per command**: spawning a process costs ~150ms and detection may\n * run on every provider construction, but a spec naming a different executable\n * is a different question — memoising on one key would make a fallback to an\n * absolute path silently inherit the bare command's failure.\n *\n * Environment probes stay unmemoised: they are free, and a consumer may\n * legitimately set a key part-way through a process.\n */\nconst cliProbes = new Map<string, Promise<boolean>>();\n\n/** Test seam: forget the memoised Claude CLI probes. */\nexport function resetClaudeCliProbe(): void {\n  cliProbes.clear();\n}\n\nfunction probeClaudeCli(spec: ProviderSpec): Promise<boolean> {\n  const exec = spec.exec ?? realExec;\n  const command = spec.command ?? \"claude\";\n  const cached = cliProbes.get(command);\n  if (cached) return cached;\n  const probing = exec([command, \"--version\"], { timeoutMs: 10_000 })\n    .then((r) => r.code === 0 && !r.timedOut && r.spawnError == null)\n    .catch(() => false);\n  cliProbes.set(command, probing);\n  return probing;\n}\n\nasync function probeLlamaCpp(spec: ProviderSpec): Promise<Probe> {\n  const injected = spec.llamaRuntime ?? spec.llamaCpp?.runtime;\n  if (injected) return budgetProbe(injected);\n\n  // Ask whether the binding is usable BEFORE touching it. Going straight to\n  // `getMemoryBudgetBytes` would route through the auto-installer, so merely\n  // asking `availableProviders()` what this machine can do would fetch a native\n  // module — a query with a side effect, and a slow one.\n  const status = await nodeLlamaCppStatus();\n  if (status.state === \"refused\") {\n    return { available: false, reason: status.reason };\n  }\n  if (status.state === \"installable\") {\n    // Usable, at the cost of an install that happens when it is actually\n    // needed. `warnInstalling` announces it at that point.\n    return { available: true };\n  }\n  // Present: load it for real, which is also what catches a binding that is\n  // installed but whose backend fails to start.\n  return budgetProbe(defaultLlamaRuntime());\n}\n\n/**\n * The same call the `auto` MODEL selector makes, so choosing llama-cpp here\n * costs nothing extra: it is loaded either way.\n */\nfunction budgetProbe(runtime: LlamaRuntime): Promise<Probe> {\n  return runtime.getMemoryBudgetBytes().then(\n    () => ({ available: true }),\n    (e: unknown) => ({\n      available: false,\n      reason:\n        e instanceof Error && /node-llama-cpp/.test(e.message)\n          ? \"node-llama-cpp is not installed (npm i node-llama-cpp)\"\n          : `node-llama-cpp could not start (${\n              e instanceof Error ? e.message : String(e)\n            })`,\n    }),\n  );\n}\n\nasync function probe(\n  provider: ProviderName,\n  spec: ProviderSpec,\n): Promise<Probe> {\n  switch (provider) {\n    case \"anthropic\":\n      return hasKey(\"anthropic\")\n        ? { available: true }\n        : {\n            available: false,\n            reason: `ANTHROPIC_API_KEY is not set`,\n          };\n    case \"openai\":\n      // A local OpenAI-compatible server needs no key — same rule the\n      // OpenAICompatProvider constructor applies.\n      return hasKey(\"openai\") || spec.baseUrl\n        ? { available: true }\n        : {\n            available: false,\n            reason: `OPENAI_API_KEY is not set and no baseUrl was given`,\n          };\n    case \"claude-cli\":\n      return (await probeClaudeCli(spec))\n        ? { available: true }\n        : {\n            available: false,\n            reason: `could not run \\`${spec.command ?? \"claude\"}\\` (is the Claude CLI installed?)`,\n          };\n    case \"llama-cpp\":\n      return probeLlamaCpp(spec);\n    default:\n      return { available: false, reason: \"not auto-selectable\" };\n  }\n}\n\n/**\n * Every provider this machine could use right now, in priority order.\n *\n * Useful for showing a picker or explaining a fallback; `detectProvider` is\n * the same sweep with the first hit returned.\n */\nexport async function availableProviders(\n  spec: ProviderSpec = {},\n): Promise<ProviderName[]> {\n  const probes = await Promise.all(\n    DETECTION_ORDER.map((name) => probe(name, spec)),\n  );\n  return DETECTION_ORDER.filter((_, i) => probes[i]!.available);\n}\n\n/**\n * The highest-priority provider this machine can use.\n *\n * Throws an `InferenceError` naming every provider and why each was\n * unavailable — far more actionable than the `Unknown provider \"undefined\"`\n * this replaces.\n */\nexport async function detectProvider(\n  spec: ProviderSpec = {},\n): Promise<ProviderName> {\n  // Sequential, not Promise.all: the probes get dramatically more expensive\n  // down the list, and the cheapest usually wins. Reading an environment\n  // variable costs microseconds, spawning the Claude CLI ~150ms, and loading\n  // the node-llama-cpp binding ~850ms — the last of which also initialises the\n  // llama backend and allocates GPU context. Probing eagerly would pay all of\n  // that on every construction just to pick `anthropic` off an env var, and\n  // would touch the GPU for a provider that is never used.\n  const reasons: string[] = [];\n  for (const name of DETECTION_ORDER) {\n    const result = await probe(name, spec);\n    if (result.available) {\n      warnSelected(name, spec.provider === \"auto\");\n      return name;\n    }\n    reasons.push(`  ${name.padEnd(10)} — ${result.reason}`);\n  }\n  throw new InferenceError(\n    `No inference provider is available. Tried:\\n${reasons.join(\"\\n\")}\\n` +\n      `Pass an explicit \\`provider\\`, set one of the keys above, or install node-llama-cpp.`,\n  );\n}\n\nlet warnedSelection = false;\nlet warnedDownload = false;\n\n/** Test seam: reset the once-per-process auto-detection warnings. */\nexport function resetProviderDetectionWarning(): void {\n  warnedSelection = false;\n  warnedDownload = false;\n}\n\n/**\n * These are eval tools: a run whose provider silently changed because an\n * environment variable moved is a run whose verdicts and cache are no longer\n * comparable to the last one. Say which provider was picked, once.\n */\nfunction warnSelected(provider: ProviderName, wasExplicitAuto: boolean): void {\n  if (warnedSelection) return;\n  warnedSelection = true;\n  // `provider: \"auto\"` IS a specification — saying otherwise sends someone\n  // hunting their config for a field they did set.\n  const because = wasExplicitAuto ? `provider \"auto\"` : \"no provider specified\";\n  console.warn(\n    `inference: ${because} — auto-selected \"${provider}\". ` +\n      `Pass an explicit \\`provider\\` to pin it.`,\n  );\n}\n\n/**\n * Falling back to a local model can mean pulling gigabytes. Say so before it\n * starts, not after a CI job has already stalled on it.\n */\nexport function warnPendingDownload(model: string, sizeBytes: number): void {\n  if (warnedDownload) return;\n  warnedDownload = true;\n  console.warn(\n    `inference: \"${model}\" is not downloaded yet — the first run will fetch ` +\n      `~${(sizeBytes / 1e9).toFixed(2)} GB. Pre-fetch it, or pass an explicit ` +\n      `\\`provider\\` to avoid the local model entirely.`,\n  );\n}\n","/**\n * Provider factory over a library-owned `ProviderSpec`.\n *\n * The spec is deliberately NOT any consumer's config type. Every consumer maps\n * its own config into this flat shape, so adding a provider here does not\n * require touching four config schemas, and no consumer has to model its\n * config on another's to reuse this layer (ADR 01000).\n */\nimport { InferenceError } from \"../types.js\";\nimport { warnIfUnsupportedNode } from \"../runtime.js\";\nimport type { Pricing } from \"../cost.js\";\nimport { AnthropicProvider } from \"./anthropic.js\";\nimport { OpenAICompatProvider } from \"./openai-compat.js\";\nimport { ClaudeCliProvider } from \"./claude-cli.js\";\nimport { MockProvider } from \"./mock.js\";\nimport { LlamaCppProvider, defaultLlamaRuntime } from \"./llama-cpp.js\";\nimport {\n  LLAMA_MODELS,\n  aliasForTier,\n  defaultLlamaModelsDirectory,\n  isLlamaSelector,\n  isModelDownloaded,\n  tierForBudget,\n} from \"./llama-models.js\";\nimport { detectProvider, warnPendingDownload } from \"./detect.js\";\nimport type { AnthropicProviderOptions } from \"./anthropic.js\";\nimport type { OpenAICompatProviderOptions } from \"./openai-compat.js\";\nimport type { MockResponse } from \"./mock.js\";\nimport type { LlamaCppProviderOptions, LlamaRuntime } from \"./llama-cpp.js\";\nimport type { LlamaTier } from \"./llama-models.js\";\nimport type { ExecFn, InferenceProvider } from \"./types.js\";\n\nexport type ProviderName =\n  | \"anthropic\"\n  | \"openai\"\n  | \"claude-cli\"\n  | \"mock\"\n  | \"llama-cpp\";\n\n/** A concrete provider, or `\"auto\"` to detect one. */\nexport type ProviderSelector = ProviderName | \"auto\";\n\nexport interface ProviderSpec {\n  /**\n   * Omitting this is identical to `\"auto\"`: the highest-priority provider this\n   * machine can actually use is detected, ending at the free local model.\n   * Resolving it needs `makeProviderAsync`/`resolveProviderIdentityAsync`.\n   */\n  provider?: ProviderSelector;\n  /** null/undefined selects the per-provider default. */\n  model?: string | null;\n  /** Env var NAME holding the API key; null/undefined selects the default. */\n  apiKeyEnv?: string | null;\n  /** openai only. */\n  baseUrl?: string;\n  /** claude-cli only: the executable to run. */\n  command?: string;\n  /** claude-cli only: subprocess timeout. */\n  timeoutMs?: number;\n  /**\n   * Pricing override for this model. Not used to construct the provider —\n   * carried here so a consumer passes one object to both `makeProvider` and\n   * `pricingFor`.\n   */\n  pricing?: Pricing;\n  /** Provider-specific tuning, ignored by the other providers. */\n  anthropic?: AnthropicProviderOptions;\n  openai?: OpenAICompatProviderOptions;\n  llamaCpp?: LlamaCppProviderOptions;\n  /** Test seam for the claude-cli provider. */\n  exec?: ExecFn;\n  /** Test seam for the llama-cpp provider. */\n  llamaRuntime?: LlamaRuntime;\n  /** Scripted responses for the mock provider; defaults to a single empty object. */\n  mockResponses?: MockResponse[];\n}\n\nexport const DEFAULT_MODELS: Record<ProviderName, string> = {\n  anthropic: \"claude-sonnet-4-5\",\n  openai: \"gpt-4o-mini\",\n  \"claude-cli\": \"claude-sonnet-4-5\",\n  mock: \"mock-model\",\n  // A selector, not a pinned model: which weights a tier points at is then a\n  // catalog change rather than an API change. Resolving it needs the async\n  // factory — see `resolveProviderIdentityAsync`.\n  \"llama-cpp\": \"auto\",\n};\n\nconst DEFAULT_API_KEY_ENV: Record<string, string> = {\n  anthropic: \"ANTHROPIC_API_KEY\",\n  openai: \"OPENAI_API_KEY\",\n};\n\nexport const DEFAULT_OPENAI_BASE_URL = \"https://api.openai.com/v1\";\n\nexport interface ProviderIdentity {\n  /** Always concrete — never `\"auto\"`, so it is safe as cache-key material. */\n  provider: ProviderName;\n  model: string;\n}\n\n/**\n * Resolve the provider name and model WITHOUT constructing the provider —\n * cache keys and pricing need the identity, but construction may require an\n * API key that a fully-cached run never uses.\n *\n * Throws for an unresolved `llama-cpp` selector. That is deliberate: picking a\n * tier weighs GPU VRAM, which needs an await, and returning the literal\n * \"auto\" as cache-key material would let a 2 GB and a 12 GB model share cached\n * verdicts — and give different results per machine under one key. Use\n * `resolveProviderIdentityAsync` for selectors.\n */\nexport function resolveProviderIdentity(spec: ProviderSpec): ProviderIdentity {\n  if (spec.provider == null || spec.provider === \"auto\") {\n    throw new InferenceError(\n      `No provider specified. Detecting one probes the environment, the Claude ` +\n        `CLI and the local model runtime, which cannot be done synchronously — ` +\n        `use resolveProviderIdentityAsync/makeProviderAsync, or name a provider ` +\n        `(${Object.keys(DEFAULT_MODELS).join(\", \")}).`,\n    );\n  }\n  const model = spec.model ?? DEFAULT_MODELS[spec.provider] ?? \"unknown\";\n  if (spec.provider === \"llama-cpp\" && isLlamaSelector(model)) {\n    throw new InferenceError(\n      `llama-cpp model \"${model}\" is a selector and cannot be resolved ` +\n        `synchronously — picking a tier probes GPU memory. Use ` +\n        `resolveProviderIdentityAsync/makeProviderAsync, or name a concrete ` +\n        `model (e.g. \"gemma-4-e4b\").`,\n    );\n  }\n  return { provider: spec.provider, model };\n}\n\n/**\n * Selector-aware identity resolution. Returns the CONCRETE model a selector\n * resolved to, so the cache key names the weights that actually ran.\n *\n * Every other provider delegates to the synchronous form, so a consumer can\n * switch to this wholesale.\n */\nexport async function resolveProviderIdentityAsync(\n  spec: ProviderSpec,\n): Promise<ProviderIdentity> {\n  // A model name belongs to exactly one provider, so it cannot be carried into\n  // whichever provider detection happens to pick: `{ model: \"gpt-4o-mini\" }` on\n  // a machine with an Anthropic key selected `anthropic` and then 404'd at call\n  // time, after the caller had already paid for detection. `null` still means\n  // \"use the default\" — only a real name is ambiguous.\n  if (\n    (spec.provider == null || spec.provider === \"auto\") &&\n    spec.model != null\n  ) {\n    throw new InferenceError(\n      `Model \"${spec.model}\" was given without a provider, and a model name ` +\n        `does not say which provider owns it. Name the provider too ` +\n        `(${Object.keys(DEFAULT_MODELS).join(\", \")}), or drop the model to ` +\n        `take the detected provider's default.`,\n    );\n  }\n\n  // Provider first, then the model logic for whichever provider won.\n  const provider =\n    spec.provider == null || spec.provider === \"auto\"\n      ? await detectProvider(spec)\n      : spec.provider;\n  const resolved: ProviderSpec = { ...spec, provider };\n\n  const model = spec.model ?? DEFAULT_MODELS[provider] ?? \"unknown\";\n  if (provider !== \"llama-cpp\" || !isLlamaSelector(model)) {\n    return resolveProviderIdentity(resolved);\n  }\n  const tier: LlamaTier =\n    model === \"auto\" ? await probeTier(llamaRuntimeFor(spec)) : model;\n  return { provider, model: aliasForTier(tier) };\n}\n\n/**\n * Warn before a multi-gigabyte download, but only once the caller has actually\n * committed to running.\n *\n * Deliberately NOT in `resolveProviderIdentityAsync`: that resolves an identity\n * *without* constructing anything, which is exactly what a fully-cached run\n * does — and such a run downloads nothing, so warning there announces gigabytes\n * that never move.\n */\nfunction warnIfDownloadPending(spec: ProviderSpec, model: string): void {\n  const entry = LLAMA_MODELS[model];\n  if (!entry) return;\n  const directory =\n    spec.llamaCpp?.modelsDirectory ?? defaultLlamaModelsDirectory();\n  if (!isModelDownloaded(model, directory)) {\n    warnPendingDownload(model, entry.sizeBytes);\n  }\n}\n\n/**\n * A runtime can be injected either as `spec.llamaRuntime` or inside\n * `spec.llamaCpp` — `makeProvider` honours both, so selector resolution must\n * too. Missing one sends the probe to the real native module and throws for a\n * consumer whose whole point was to avoid it.\n */\nfunction llamaRuntimeFor(spec: ProviderSpec): LlamaRuntime | undefined {\n  return spec.llamaRuntime ?? spec.llamaCpp?.runtime;\n}\n\nasync function probeTier(runtime: LlamaRuntime | undefined): Promise<LlamaTier> {\n  const source = runtime ?? defaultLlamaRuntime();\n  return tierForBudget(await source.getMemoryBudgetBytes());\n}\n\nexport function makeProvider(spec: ProviderSpec): InferenceProvider {\n  // First use of the library, for anyone who goes through the factory — before\n  // the spec is even validated, so an unsupported Node is named ahead of any\n  // error it might be the real cause of.\n  warnIfUnsupportedNode();\n  const { model } = resolveProviderIdentity(spec);\n\n  switch (spec.provider) {\n    case \"anthropic\":\n      return new AnthropicProvider(\n        model,\n        spec.apiKeyEnv ?? DEFAULT_API_KEY_ENV[\"anthropic\"]!,\n        spec.anthropic ?? {},\n      );\n    case \"openai\":\n      return new OpenAICompatProvider(\n        spec.baseUrl ?? DEFAULT_OPENAI_BASE_URL,\n        model,\n        spec.apiKeyEnv ?? DEFAULT_API_KEY_ENV[\"openai\"]!,\n        undefined,\n        spec.openai ?? {},\n      );\n    case \"claude-cli\":\n      return new ClaudeCliProvider(\n        model,\n        spec.command ?? \"claude\",\n        spec.exec,\n        spec.timeoutMs,\n      );\n    case \"mock\":\n      // Offline smoke-testing seam: proposes nothing unless scripted.\n      return new MockProvider(spec.mockResponses ?? [{ json: {} }], model);\n    case \"llama-cpp\":\n      return new LlamaCppProvider(model, {\n        ...(spec.llamaCpp ?? {}),\n        ...(spec.llamaRuntime ? { runtime: spec.llamaRuntime } : {}),\n      });\n    default:\n      throw new InferenceError(\n        `Unknown provider \"${String(spec.provider)}\". Available: ${Object.keys(\n          DEFAULT_MODELS,\n        ).join(\", \")}.`,\n      );\n  }\n}\n\n/**\n * Selector-aware provider construction. Resolves a `llama-cpp` selector\n * against this machine first, so the returned provider's `modelName()` — and\n * therefore the cache key — names the weights it will actually load.\n *\n * Every other provider delegates to `makeProvider`.\n */\nexport async function makeProviderAsync(\n  spec: ProviderSpec,\n): Promise<InferenceProvider> {\n  // Both halves must be threaded through: passing only the model would leave a\n  // detected provider as `undefined` and throw in `makeProvider`.\n  const { provider, model } = await resolveProviderIdentityAsync(spec);\n  if (provider === \"llama-cpp\") warnIfDownloadPending(spec, model);\n  return makeProvider({ ...spec, provider, model });\n}\n","/**\n * Reclaiming disk space from downloaded GGUF weights.\n *\n * This is straightforward because the models directory belongs to this library\n * alone (see `defaultLlamaModelsDirectory`). node-llama-cpp's own global\n * directory is shared with its CLI and anything else on the machine, so\n * clearing THAT would destroy models this library never downloaded; owning a\n * directory removes the hazard rather than defending against it.\n */\nimport { rmSync, statSync } from \"node:fs\";\nimport { join } from \"node:path\";\nimport {\n  blobNameFor,\n  defaultLlamaModelsDirectory,\n  listModelDirectory,\n  matchesModelBlob,\n} from \"./llama-models.js\";\nimport { disposeLlamaModels } from \"./llama-cpp.js\";\n\nexport interface ClearedModelFile {\n  path: string;\n  sizeBytes: number;\n}\n\nexport interface ClearLlamaModelsResult {\n  /** Files removed, or that would be removed under `dryRun`. */\n  files: ClearedModelFile[];\n  freedBytes: number;\n  directory: string;\n  dryRun: boolean;\n}\n\nexport interface ClearLlamaModelsOptions {\n  /** Defaults to this library's own models directory. */\n  directory?: string;\n  /** Clear only these models (alias or `hf:` URI); default clears all. */\n  models?: string[];\n  /** Report what would be removed without deleting anything. */\n  dryRun?: boolean;\n}\n\n/**\n * Delete downloaded GGUF weights and interrupted partial downloads.\n *\n * Loaded models are disposed first: on Windows the weights are memory-mapped\n * while loaded, and deleting an open file fails with EBUSY/EPERM.\n */\nexport async function clearLlamaModels(\n  options: ClearLlamaModelsOptions = {},\n): Promise<ClearLlamaModelsResult> {\n  const directory = options.directory ?? defaultLlamaModelsDirectory();\n  const dryRun = options.dryRun ?? false;\n  // Resolve BEFORE touching disk so an unknown name fails without having\n  // deleted half the set.\n  const wanted = options.models?.map(blobNameFor);\n\n  if (!dryRun) await disposeLlamaModels();\n\n  const files: ClearedModelFile[] = [];\n  for (const entry of listModelDirectory(directory)) {\n    // Only ever weights — a stray config or log in this directory is safe, and\n    // so is anything else if `directory` was pointed somewhere shared.\n    if (!isModelBlob(entry)) continue;\n    if (wanted && !wanted.some((name) => matchesModelBlob(entry, name))) {\n      continue;\n    }\n    const path = join(directory, entry);\n    const sizeBytes = sizeOf(path);\n    if (sizeBytes === undefined) continue;\n    if (!dryRun) {\n      try {\n        rmSync(path);\n      } catch {\n        // A file held open by another process is not this call's business to\n        // force; report only what actually went away.\n        continue;\n      }\n    }\n    files.push({ path, sizeBytes });\n  }\n\n  return {\n    files,\n    freedBytes: files.reduce((total, file) => total + file.sizeBytes, 0),\n    directory,\n    dryRun,\n  };\n}\n\nfunction isModelBlob(entry: string): boolean {\n  return entry.endsWith(\".gguf\") || entry.endsWith(\".gguf.ipull\");\n}\n\nfunction sizeOf(path: string): number | undefined {\n  try {\n    return statSync(path).size;\n  } catch {\n    return undefined;\n  }\n}\n","/**\n * One schema-validated completion, with a single retry.\n *\n * The invariant every consumer depends on: a run that cannot produce\n * schema-valid JSON after the retry is recorded as an ERROR, not dropped and\n * not coerced. Downstream, an errored run counts against consensus — it can\n * push a result toward human review, but it can never produce a silent pass.\n */\nimport { Ajv2020 } from \"ajv/dist/2020.js\";\nimport type { ValidateFunction } from \"ajv\";\nimport { warnIfUnsupportedNode } from \"./runtime.js\";\nimport type { InferenceProvider, TokenUsage } from \"./providers/types.js\";\n\n/** One attempt at a schema-constrained completion. */\nexport interface InferenceRun<T = unknown> {\n  /** Absent when the run errored (invalid JSON after retry, API failure). */\n  result?: T;\n  error?: string;\n  provider: string;\n  model: string;\n  cached: boolean;\n  usage?: TokenUsage;\n  durationMs: number;\n}\n\nexport interface CompleteValidatedOptions {\n  provider: InferenceProvider;\n  system: string;\n  user: string;\n  schema: Record<string, unknown>;\n  temperature?: number;\n  /**\n   * Attempts before recording an error. Default 2 (one initial call plus one\n   * retry) — matches the behavior all three source projects shipped.\n   */\n  attempts?: number;\n  /**\n   * Pre-compiled validator. Compiling Ajv per call is wasteful in an ensemble\n   * loop, so `runEnsemble` compiles once and passes it down.\n   */\n  validate?: ValidateFunction;\n}\n\nconst validatorCache = new WeakMap<object, ValidateFunction>();\n\n/** Compile once per schema object identity — Ajv compilation is not cheap. */\nexport function validatorFor(\n  schema: Record<string, unknown>,\n): ValidateFunction {\n  const cached = validatorCache.get(schema);\n  if (cached) return cached;\n  // A fresh Ajv per distinct schema object, not one shared instance: Ajv keeps\n  // a registry keyed by `$id`, so a caller that rebuilds an equal schema object\n  // per call (spreading VERDICT_SCHEMA to override descriptions, say) misses\n  // the identity cache above and would hit \"schema with key or id ... already\n  // exists\" on the second compile. Instances are held only by this WeakMap, so\n  // they are collected with the schemas that own them.\n  const compiled = new Ajv2020({ allErrors: true }).compile(schema);\n  validatorCache.set(schema, compiled);\n  return compiled;\n}\n\nexport async function completeValidatedJSON<T = unknown>(\n  options: CompleteValidatedOptions,\n): Promise<InferenceRun<T>> {\n  // The other half of \"first use\": a consumer that constructs a provider\n  // directly never touches `makeProvider`, but everything still funnels here.\n  warnIfUnsupportedNode();\n  const {\n    provider,\n    system,\n    user,\n    schema,\n    temperature = 0,\n    attempts = 2,\n  } = options;\n  const validate = options.validate ?? validatorFor(schema);\n\n  const start = Date.now();\n  const base = {\n    provider: provider.provider(),\n    model: provider.modelName(),\n    cached: false,\n  };\n\n  let lastError = \"unknown error\";\n  for (let attempt = 0; attempt < attempts; attempt++) {\n    try {\n      const response = await provider.completeJSON({\n        system,\n        user,\n        schema,\n        temperature,\n      });\n      if (validate(response.json)) {\n        return {\n          ...base,\n          result: response.json as T,\n          usage: response.usage,\n          durationMs: Date.now() - start,\n        };\n      }\n      lastError = `Response failed schema validation: ${(validate.errors ?? [])\n        .map((e) => `${e.instancePath} ${e.message}`)\n        .join(\"; \")}`;\n    } catch (e) {\n      lastError = e instanceof Error ? e.message : String(e);\n    }\n  }\n  return { ...base, error: lastError, durationMs: Date.now() - start };\n}\n","/**\n * Cost tracking: token usage priced from a small static table, overridable per\n * model by the caller. Unknown models cost 0 (unknown), never a guess — a\n * fabricated price is worse than an absent one when a budget gate depends on it.\n */\nimport type { TokenUsage } from \"./providers/types.js\";\n\nexport interface Pricing {\n  inputPerMTok: number;\n  outputPerMTok: number;\n}\n\n/**\n * USD per million tokens. Entries are base names; pinned variants\n * (`claude-sonnet-4-5-20250929`) resolve by prefix.\n */\nexport const PRICE_TABLE: Record<string, Pricing> = {\n  \"claude-sonnet-4-5\": { inputPerMTok: 3, outputPerMTok: 15 },\n  \"claude-sonnet-4-6\": { inputPerMTok: 3, outputPerMTok: 15 },\n  \"claude-haiku-4-5\": { inputPerMTok: 1, outputPerMTok: 5 },\n  \"claude-opus-4-8\": { inputPerMTok: 15, outputPerMTok: 75 },\n  \"gpt-4o-mini\": { inputPerMTok: 0.15, outputPerMTok: 0.6 },\n  \"gpt-4o\": { inputPerMTok: 2.5, outputPerMTok: 10 },\n};\n\nexport function pricingFor(\n  model: string,\n  override?: Pricing,\n): Pricing | undefined {\n  if (override) return override;\n  if (PRICE_TABLE[model]) return PRICE_TABLE[model];\n  // Match pinned variants like claude-sonnet-4-5-20250929. Longest prefix\n  // wins so `claude-sonnet-4-5` never shadows a more specific future entry.\n  const base = Object.keys(PRICE_TABLE)\n    .filter((k) => model.startsWith(k))\n    .sort((a, b) => b.length - a.length)[0];\n  return base ? PRICE_TABLE[base] : undefined;\n}\n\nexport function costOfUsage(\n  usage: TokenUsage | undefined,\n  pricing: Pricing | undefined,\n): number {\n  if (!usage || !pricing) return 0;\n  return (\n    (usage.inputTokens / 1_000_000) * pricing.inputPerMTok +\n    (usage.outputTokens / 1_000_000) * pricing.outputPerMTok\n  );\n}\n\n/** Sum the cost of a set of runs. Cached runs cost nothing — they made no call. */\nexport function costOfRuns(\n  runs: { usage?: TokenUsage; cached?: boolean }[],\n  pricing: Pricing | undefined,\n): number {\n  if (!pricing) return 0;\n  let usd = 0;\n  for (const run of runs) {\n    if (!run.usage || run.cached) continue;\n    usd += costOfUsage(run.usage, pricing);\n  }\n  return usd;\n}\n","{\n  \"$schema\": \"https://json-schema.org/draft/2020-12/schema\",\n  \"$id\": \"inference:verdict:0.1\",\n  \"title\": \"LLM-as-judge verdict\",\n  \"type\": \"object\",\n  \"required\": [\"claim\", \"observed\", \"match\", \"confidence\", \"reasoning\"],\n  \"properties\": {\n    \"claim\": {\n      \"type\": \"string\",\n      \"description\": \"The specific assertion under evaluation.\"\n    },\n    \"observed\": {\n      \"type\": \"string\",\n      \"description\": \"What was actually observed in the subject under evaluation, quoting evidence where possible.\"\n    },\n    \"match\": {\n      \"enum\": [\"pass\", \"fail\", \"partial\"],\n      \"description\": \"pass only if the assertion is fully satisfied; partial for partial compliance.\"\n    },\n    \"confidence\": {\n      \"type\": \"number\",\n      \"minimum\": 0,\n      \"maximum\": 1,\n      \"description\": \"Self-reported confidence in this verdict.\"\n    },\n    \"reasoning\": {\n      \"type\": \"string\",\n      \"description\": \"Why this conclusion was reached.\"\n    }\n  },\n  \"additionalProperties\": false\n}\n","/**\n * LLM-as-judge types. The verdict shape is the one all consuming projects\n * already share; the canonical schema lives in verdict-schema.json and is\n * overridable per consumer so domain-specific field descriptions survive\n * (ADR 01001).\n */\nimport verdictSchemaJson from \"./verdict-schema.json\" with { type: \"json\" };\nimport type { TokenUsage } from \"../providers/types.js\";\n\nexport const VERDICT_SCHEMA = verdictSchemaJson as Record<string, unknown>;\n\nexport type Match = \"pass\" | \"fail\" | \"partial\";\n\n/** Confidence-zone routing for LLM-judged evals. */\nexport type Zone = \"auto-pass\" | \"auto-fail\" | \"human-review\";\n\nexport interface JudgeVerdict {\n  /** The specific assertion under evaluation. */\n  claim: string;\n  /** What the judge actually observed. */\n  observed: string;\n  match: Match;\n  /** 0.0–1.0 self-reported confidence. */\n  confidence: number;\n  reasoning: string;\n}\n\n/** One run within an ensemble. */\nexport interface JudgeRun {\n  /** Absent when the run errored (invalid JSON after retry, API failure). */\n  verdict?: JudgeVerdict;\n  error?: string;\n  provider: string;\n  model: string;\n  cached: boolean;\n  usage?: TokenUsage;\n  durationMs: number;\n}\n\n/** Aggregated outcome of an ensemble of judge runs for one subject. */\nexport interface ConsensusResult {\n  runs: JudgeRun[];\n  votes: { pass: number; fail: number; partial: number; error: number };\n  /** Majority verdict; `partial` counts as fail for the binary outcome. */\n  verdict: Match;\n  /** Fraction of non-errored runs agreeing with the majority verdict. */\n  agreement: number;\n  /** Mean confidence across non-errored runs. */\n  meanConfidence: number;\n  zone: Zone;\n}\n","/**\n * Ensemble consensus math. `partial` counts as fail for the binary outcome\n * but stays visible in the vote counts. Errored runs count against consensus —\n * they can only push a result toward human review, never toward a silent pass.\n */\nimport type { ConsensusResult, JudgeRun, Match } from \"./types.js\";\n\nexport function computeConsensus(\n  runs: JudgeRun[],\n): Omit<ConsensusResult, \"zone\"> {\n  const votes = { pass: 0, fail: 0, partial: 0, error: 0 };\n  let confidenceSum = 0;\n  let confidenceCount = 0;\n  for (const run of runs) {\n    if (!run.verdict) {\n      votes.error += 1;\n      continue;\n    }\n    votes[run.verdict.match] += 1;\n    confidenceSum += run.verdict.confidence;\n    confidenceCount += 1;\n  }\n\n  const passVotes = votes.pass;\n  const failVotes = votes.fail + votes.partial;\n  // Binary majority; a tie is not a pass.\n  const verdict: Match = passVotes > failVotes ? \"pass\" : \"fail\";\n  const graded = passVotes + failVotes;\n  const agreement = graded > 0 ? Math.max(passVotes, failVotes) / graded : 0;\n  const meanConfidence =\n    confidenceCount > 0 ? confidenceSum / confidenceCount : 0;\n\n  return { runs, votes, verdict, agreement, meanConfidence };\n}\n","/**\n * Confidence-zone routing: only unanimous, high-confidence ensembles\n * auto-resolve; everything else goes to a human.\n */\nimport type { ConsensusResult, Zone } from \"./types.js\";\n\nexport interface ZoneThresholds {\n  autoPass: number;\n  autoFail: number;\n}\n\nexport const DEFAULT_ZONES: ZoneThresholds = { autoPass: 0.8, autoFail: 0.8 };\n\nexport function zoneFor(\n  consensus: Omit<ConsensusResult, \"zone\">,\n  thresholds: ZoneThresholds = DEFAULT_ZONES,\n): Zone {\n  const { votes, meanConfidence } = consensus;\n  const unanimousPass =\n    votes.pass > 0 &&\n    votes.fail === 0 &&\n    votes.partial === 0 &&\n    votes.error === 0;\n  const unanimousFail =\n    votes.pass === 0 && votes.error === 0 && votes.fail + votes.partial > 0;\n\n  if (unanimousPass && meanConfidence >= thresholds.autoPass) return \"auto-pass\";\n  if (unanimousFail && meanConfidence >= thresholds.autoFail) return \"auto-fail\";\n  return \"human-review\";\n}\n","/**\n * The ensemble judge: N independent runs for one subject, each a fresh request\n * with no shared context (eval isolation), aggregated by consensus and routed\n * through confidence zones.\n *\n * Runs within one ensemble stay sequential on purpose — they are meant to be\n * independent samples, and interleaving them buys nothing. Concurrency belongs\n * one level up, across subjects, where the consumer owns the pool.\n */\nimport { completeValidatedJSON, validatorFor } from \"../complete.js\";\nimport type { JsonCache } from \"../cache.js\";\nimport type { InferenceProvider } from \"../providers/types.js\";\nimport { computeConsensus } from \"./consensus.js\";\nimport { DEFAULT_ZONES, zoneFor, type ZoneThresholds } from \"./zones.js\";\nimport {\n  VERDICT_SCHEMA,\n  type ConsensusResult,\n  type JudgeRun,\n  type JudgeVerdict,\n} from \"./types.js\";\n\nexport interface EnsembleOptions {\n  provider: InferenceProvider;\n  system: string;\n  user: string;\n  /** Ensemble size; default 3. */\n  runs?: number;\n  /** Default 0. Nonzero adds noise to verdicts and warns once. */\n  temperature?: number;\n  /**\n   * Verdict schema. Defaults to the canonical one; pass your own to keep\n   * domain-specific field descriptions (they measurably steer the model).\n   * Must still produce objects matching `JudgeVerdict`.\n   */\n  schema?: Record<string, unknown>;\n  /** Optional result cache. Requires `cacheKey`. */\n  cache?: JsonCache<JudgeRun[]>;\n  /** Content-addressed key; build it with `buildCacheKey`. */\n  cacheKey?: string;\n  /** Prefix for warnings, e.g. your tool's name. */\n  label?: string;\n}\n\nlet warnedTemperature = false;\n\n/**\n * Run the ensemble and return every run. Cached ensembles replay identically,\n * with each run flagged `cached: true` so cost accounting skips them.\n */\nexport async function runEnsemble(\n  options: EnsembleOptions,\n): Promise<JudgeRun[]> {\n  const {\n    provider,\n    system,\n    user,\n    runs: runCount = 3,\n    temperature = 0,\n    schema = VERDICT_SCHEMA,\n    cache,\n    cacheKey,\n    label = \"inference\",\n  } = options;\n\n  if (temperature > 0 && !warnedTemperature) {\n    warnedTemperature = true;\n    console.warn(\n      `${label}: judge temperature is ${temperature} — nonzero temperature adds noise to verdicts; 0 is strongly recommended.`,\n    );\n  }\n\n  if (cache && cacheKey) {\n    const hit = cache.get(cacheKey);\n    // A cache file can be valid JSON and still be the wrong shape — a\n    // truncated write, or an entry from an older cache generation. JsonCache\n    // only guards parse failures, so the shape check belongs here; a bad entry\n    // is a miss, never a crash.\n    if (Array.isArray(hit)) return hit.map((r) => ({ ...r, cached: true }));\n  }\n\n  // Compile the schema once for the whole ensemble rather than per run.\n  const validate = validatorFor(schema);\n\n  const results: JudgeRun[] = [];\n  for (let i = 0; i < runCount; i++) {\n    const run = await completeValidatedJSON<JudgeVerdict>({\n      provider,\n      system,\n      user,\n      schema,\n      temperature,\n      validate,\n    });\n    // `result` is the generic name at the completion layer; the judge layer\n    // calls it `verdict`, which is what consumers persist in their caches.\n    const { result, ...rest } = run;\n    results.push(result === undefined ? rest : { ...rest, verdict: result });\n  }\n\n  if (cache && cacheKey) cache.set(cacheKey, results);\n  return results;\n}\n\n/** Run the ensemble and aggregate it into a zoned consensus. */\nexport async function judge(\n  options: EnsembleOptions & { zones?: ZoneThresholds },\n): Promise<ConsensusResult> {\n  const runs = await runEnsemble(options);\n  const base = computeConsensus(runs);\n  return { ...base, zone: zoneFor(base, options.zones ?? DEFAULT_ZONES) };\n}\n\n/** Test seam: reset the once-per-process temperature warning. */\nexport function resetTemperatureWarning(): void {\n  warnedTemperature = false;\n}\n"],"mappings":";AAKO,IAAM,iBAAN,cAA6B,MAAM;AAAA,EACxC,YAAY,SAAiB;AAC3B,UAAM,OAAO;AACb,SAAK,OAAO;AAAA,EACd;AACF;;;ACYO,IAAM,qBAAqB;AAElC,IAAI,oBAAoB;AAQjB,SAAS,sBACd,UAAkB,QAAQ,SAAS,MAC7B;AACN,MAAI,kBAAmB;AACvB,QAAM,QAAQ,OAAO,SAAS,SAAS,EAAE;AAGzC,MAAI,CAAC,OAAO,UAAU,KAAK,KAAK,SAAS,mBAAoB;AAC7D,sBAAoB;AACpB,UAAQ;AAAA,IACN,8BAA8B,OAAO,yBAChC,kBAAkB;AAAA,EAGzB;AACF;AAGO,SAAS,0BAAgC;AAC9C,sBAAoB;AACtB;;;AChDA,OAAO,eAAe;AAQtB,IAAM,oBAAoB;AAcnB,IAAM,oBAAN,MAAqD;AAAA,EAM1D,YACmB,OACjB,WACA,UAAoC,CAAC,GACrC;AAHiB;AAIjB,UAAM,SAAS,QAAQ,IAAI,SAAS;AACpC,QAAI,CAAC,QAAQ;AACX,YAAM,IAAI;AAAA,QACR,4BAA4B,SAAS;AAAA,MACvC;AAAA,IACF;AACA,SAAK,SAAS,IAAI,UAAU,EAAE,OAAO,CAAC;AACtC,SAAK,WAAW,QAAQ,YAAY;AACpC,SAAK,kBACH,QAAQ,mBAAmB;AAC7B,SAAK,YAAY,QAAQ,aAAa;AAAA,EACxC;AAAA,EAfmB;AAAA,EANF;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAoBjB,WAAmB;AACjB,WAAO;AAAA,EACT;AAAA,EAEA,YAAoB;AAClB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAM,aAAa,KAAyD;AAC1E,UAAM,WAAW,MAAM,KAAK,OAAO,SAAS,OAAO;AAAA,MACjD,OAAO,KAAK;AAAA,MACZ,YAAY,KAAK;AAAA,MACjB,aAAa,IAAI;AAAA,MACjB,QAAQ,IAAI;AAAA,MACZ,UAAU,CAAC,EAAE,MAAM,QAAQ,SAAS,IAAI,KAAK,CAAC;AAAA,MAC9C,OAAO;AAAA,QACL;AAAA,UACE,MAAM,KAAK;AAAA,UACX,aAAa,KAAK;AAAA,UAClB,cAAc,IAAI;AAAA,QACpB;AAAA,MACF;AAAA,MACA,aAAa,EAAE,MAAM,QAAQ,MAAM,KAAK,SAAS;AAAA,IACnD,CAAC;AAMD,QAAI,SAAS,gBAAgB,cAAc;AACzC,YAAM,IAAI;AAAA,QACR,sCAAsC,KAAK,SAAS;AAAA,MACtD;AAAA,IACF;AAEA,UAAM,UAAU,SAAS,QAAQ;AAAA,MAC/B,CAAC,UAA2C,MAAM,SAAS;AAAA,IAC7D;AACA,QAAI,CAAC,SAAS;AACZ,YAAM,IAAI,MAAM,gDAAgD;AAAA,IAClE;AACA,WAAO;AAAA,MACL,MAAM,QAAQ;AAAA,MACd,OAAO;AAAA,QACL,aAAa,SAAS,MAAM;AAAA,QAC5B,cAAc,SAAS,MAAM;AAAA,MAC/B;AAAA,IACF;AAAA,EACF;AACF;;;AC/EO,SAAS,YAAY,SAA0B;AACpD,QAAM,UAAU,QACb,QAAQ,wBAAwB,EAAE,EAClC,QAAQ,cAAc,EAAE,EACxB,KAAK;AACR,MAAI;AACF,WAAO,KAAK,MAAM,OAAO;AAAA,EAC3B,QAAQ;AACN,UAAM,QAAQ,QAAQ,QAAQ,GAAG;AACjC,UAAM,MAAM,QAAQ,YAAY,GAAG;AACnC,QAAI,SAAS,KAAK,MAAM,OAAO;AAC7B,aAAO,KAAK,MAAM,QAAQ,MAAM,OAAO,MAAM,CAAC,CAAC;AAAA,IACjD;AACA,UAAM,IAAI,MAAM,6CAA6C;AAAA,EAC/D;AACF;AAQO,SAAS,eACd,QACyB;AACzB,QAAM,QAAQ,gBAAgB,MAAM;AACpC,QAAM,OAAO,CAAC,SAAwB;AACpC,QAAI,SAAS,QAAQ,OAAO,SAAS,SAAU;AAC/C,QAAI,MAAM,QAAQ,IAAI,GAAG;AACvB,iBAAW,QAAQ,KAAM,MAAK,IAAI;AAClC;AAAA,IACF;AACA,UAAM,MAAM;AACZ,WAAO,IAAI,WAAW;AACtB,WAAO,IAAI,aAAa;AACxB,UAAM,aAAa,IAAI,YAAY;AACnC,QAAI,cAAc,OAAO,eAAe,UAAU;AAChD,UAAI,UAAU,IAAI,OAAO,KAAK,UAAU;AAKxC,UAAI,sBAAsB,IAAI;AAC9B,iBAAW,QAAQ,OAAO,OAAO,UAAqC,GAAG;AACvE,aAAK,IAAI;AACT,YAAI,QAAQ,OAAO,SAAS,YAAY,CAAC,MAAM,QAAQ,IAAI,GAAG;AAC5D,gBAAM,IAAI;AACV,cAAI,OAAO,EAAE,MAAM,MAAM,YAAY,EAAE,MAAM,MAAM,QAAQ;AACzD,cAAE,MAAM,IAAI,CAAC,EAAE,MAAM,GAAG,MAAM;AAAA,UAChC;AAAA,QACF;AAAA,MACF;AAAA,IACF;AACA,SAAK,IAAI,OAAO,CAAC;AAAA,EACnB;AACA,OAAK,KAAK;AACV,SAAO;AACT;AAGO,SAAS,WAAW,OAAyB;AAClD,MAAI,UAAU,QAAQ,OAAO,UAAU,YAAY,MAAM,QAAQ,KAAK,GAAG;AACvE,WAAO;AAAA,EACT;AACA,SAAO,OAAO;AAAA,IACZ,OAAO,QAAQ,KAAgC,EAAE;AAAA,MAC/C,CAAC,CAAC,EAAE,CAAC,MAAM,MAAM;AAAA,IACnB;AAAA,EACF;AACF;AAOO,IAAM,uBAAN,MAAwD;AAAA,EAI7D,YACmB,SACA,OACjB,WACiB,SAA6B,QAAQ,IAAI,SAAS,GACnE,UAAuC,CAAC,GACxC;AALiB;AACA;AAEA;AAIjB,QAAI,CAAC,KAAK,UAAU,QAAQ,SAAS,gBAAgB,GAAG;AACtD,YAAM,IAAI;AAAA,QACR,yBAAyB,SAAS;AAAA,MACpC;AAAA,IACF;AACA,SAAK,aAAa,QAAQ,cAAc;AAAA,EAC1C;AAAA,EAbmB;AAAA,EACA;AAAA,EAEA;AAAA,EAPX,qBAAqB;AAAA,EACZ;AAAA,EAkBjB,WAAmB;AACjB,WAAO;AAAA,EACT;AAAA,EAEA,YAAoB;AAClB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAc,KAAK,MAAsD;AACvE,UAAM,WAAW,MAAM;AAAA,MACrB,GAAG,KAAK,QAAQ,QAAQ,OAAO,EAAE,CAAC;AAAA,MAClC;AAAA,QACE,QAAQ;AAAA,QACR,SAAS;AAAA,UACP,gBAAgB;AAAA,UAChB,GAAI,KAAK,SAAS,EAAE,eAAe,UAAU,KAAK,MAAM,GAAG,IAAI,CAAC;AAAA,QAClE;AAAA,QACA,MAAM,KAAK,UAAU,IAAI;AAAA,MAC3B;AAAA,IACF;AACA,UAAM,OAAQ,MAAM,SAAS,KAAK,EAAE,MAAM,OAAO,CAAC,EAAE;AACpD,QAAI,CAAC,SAAS,IAAI;AAChB,YAAM,UAAU,KAAK,OAAO,WAAW,QAAQ,SAAS,MAAM;AAC9D,YAAM,IAAI,MAAM,GAAG,OAAO,EAAE;AAAA,IAC9B;AACA,WAAO;AAAA,EACT;AAAA,EAEA,MAAM,aAAa,KAAyD;AAC1E,UAAM,OAAO;AAAA,MACX,OAAO,KAAK;AAAA,MACZ,aAAa,IAAI;AAAA,MACjB,UAAU;AAAA,QACR,EAAE,MAAM,UAAU,SAAS,IAAI,OAAO;AAAA,QACtC,EAAE,MAAM,QAAQ,SAAS,IAAI,KAAK;AAAA,MACpC;AAAA,IACF;AAEA,QAAI;AACJ,QAAI,KAAK,oBAAoB;AAC3B,UAAI;AACF,mBAAW,MAAM,KAAK,KAAK;AAAA,UACzB,GAAG;AAAA,UACH,iBAAiB;AAAA,YACf,MAAM;AAAA,YACN,aAAa;AAAA,cACX,MAAM,KAAK;AAAA,cACX,QAAQ;AAAA,cACR,QAAQ,eAAe,IAAI,MAAM;AAAA,YACnC;AAAA,UACF;AAAA,QACF,CAAC;AAAA,MACH,SAAS,GAAG;AACV,cAAM,UAAU,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC;AAGzD,YACE,CAAC,sCAAsC,KAAK,OAAO,KACnD,YAAY,YACZ;AACA,gBAAM;AAAA,QACR;AACA,aAAK,qBAAqB;AAC1B,mBAAW,MAAM,KAAK,mBAAmB,MAAM,GAAG;AAAA,MACpD;AAAA,IACF,OAAO;AACL,iBAAW,MAAM,KAAK,mBAAmB,MAAM,GAAG;AAAA,IACpD;AAEA,UAAM,UAAU,SAAS,UAAU,CAAC,GAAG,SAAS;AAChD,QAAI,CAAC,QAAS,OAAM,IAAI,MAAM,2BAA2B;AACzD,WAAO;AAAA,MACL,MAAM,WAAW,YAAY,OAAO,CAAC;AAAA,MACrC,OACE,SAAS,OAAO,iBAAiB,OAC7B;AAAA,QACE,aAAa,SAAS,MAAM,iBAAiB;AAAA,QAC7C,cAAc,SAAS,MAAM,qBAAqB;AAAA,MACpD,IACA;AAAA,IACR;AAAA,EACF;AAAA,EAEQ,mBACN,MACA,KACuB;AACvB,WAAO,KAAK,KAAK;AAAA,MACf,GAAG;AAAA,MACH,UAAU;AAAA,QACR;AAAA,UACE,MAAM;AAAA,UACN,SAAS,GAAG,IAAI,MAAM;AAAA;AAAA;AAAA,EAAwE,KAAK,UAAU,IAAI,MAAM,CAAC;AAAA,QAC1H;AAAA,QACA,EAAE,MAAM,QAAQ,SAAS,IAAI,KAAK;AAAA,MACpC;AAAA,MACA,iBAAiB,EAAE,MAAM,cAAc;AAAA,IACzC,CAAC;AAAA,EACH;AACF;;;AClNA,OAAO,WAAW;AAGX,IAAM,WAAmB,CAAC,KAAK,OAAO,CAAC,MAAM;AAClD,QAAM,CAAC,KAAK,GAAG,IAAI,IAAI;AACvB,MAAI,CAAC,KAAK;AACR,WAAO,QAAQ,QAAoB;AAAA,MACjC,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,QAAQ;AAAA,MACR,UAAU;AAAA,MACV,YAAY;AAAA,IACd,CAAC;AAAA,EACH;AACA,SAAO,IAAI,QAAoB,CAAC,mBAAmB;AACjD,UAAM,QAAQ,MAAM,KAAK,MAAM;AAAA,MAC7B,KAAK,KAAK;AAAA,MACV,KAAK,EAAE,GAAG,QAAQ,KAAK,GAAI,KAAK,OAAO,CAAC,EAAG;AAAA,MAC3C,OAAO,CAAC,KAAK,SAAS,OAAO,SAAS,UAAU,QAAQ,MAAM;AAAA,IAChE,CAAC;AAED,QAAI,SAAS;AACb,QAAI,SAAS;AACb,QAAI,WAAW;AACf,QAAI,UAAU;AAEd,UAAM,YAAY,KAAK,aAAa;AACpC,UAAM,QAAQ,WAAW,MAAM;AAC7B,iBAAW;AACX,YAAM,KAAK;AAKX,aAAO,EAAE,MAAM,MAAM,QAAQ,QAAQ,UAAU,KAAK,CAAC;AAAA,IACvD,GAAG,SAAS;AAEZ,UAAM,SAAS,CAAC,WAA6B;AAC3C,UAAI,QAAS;AACb,gBAAU;AACV,mBAAa,KAAK;AAClB,qBAAe,MAAM;AAAA,IACvB;AAEA,QAAI,KAAK,SAAS,QAAQ,MAAM,OAAO;AAErC,YAAM,MAAM,GAAG,SAAS,MAAM;AAAA,MAAC,CAAC;AAChC,YAAM,MAAM,IAAI,KAAK,KAAK;AAAA,IAC5B;AAKA,UAAM,QAAQ,YAAY,MAAM;AAChC,UAAM,QAAQ,YAAY,MAAM;AAChC,UAAM,QAAQ,GAAG,QAAQ,CAAC,MAAe,UAAU,CAAE;AACrD,UAAM,QAAQ,GAAG,QAAQ,CAAC,MAAe,UAAU,CAAE;AACrD,UAAM;AAAA,MAAG;AAAA,MAAS,CAAC,MACjB,OAAO,EAAE,MAAM,MAAM,QAAQ,QAAQ,UAAU,YAAY,EAAE,QAAQ,CAAC;AAAA,IACxE;AACA,UAAM,GAAG,SAAS,CAAC,SAAS,OAAO,EAAE,MAAM,QAAQ,QAAQ,SAAS,CAAC,CAAC;AAAA,EACxE,CAAC;AACH;;;ACtDO,IAAM,oBAAN,MAAqD;AAAA,EAC1D,YACmB,OACA,UAAkB,UAClB,OAAe,UACf,YAAoB,MACrC;AAJiB;AACA;AACA;AACA;AAAA,EAChB;AAAA,EAJgB;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAGnB,WAAmB;AACjB,WAAO;AAAA,EACT;AAAA,EAEA,YAAoB;AAClB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAM,aAAa,KAAyD;AAC1E,UAAM,SAAS;AAAA,MACb,IAAI;AAAA,MACJ;AAAA,MACA;AAAA,MACA;AAAA,MACA,KAAK,UAAU,IAAI,MAAM;AAAA,IAC3B,EAAE,KAAK,IAAI;AAIX,UAAM,SAAS,MAAM,KAAK;AAAA,MACxB;AAAA,QACE,KAAK;AAAA,QACL;AAAA,QACA;AAAA,QACA,IAAI;AAAA,QACJ;AAAA,QACA;AAAA,QACA;AAAA,QACA,KAAK;AAAA,MACP;AAAA,MACA,EAAE,WAAW,KAAK,WAAW,OAAO,OAAO;AAAA,IAC7C;AAEA,QAAI,OAAO,YAAY;AACrB,YAAM,IAAI;AAAA,QACR,iBAAiB,KAAK,OAAO,KAAK,OAAO,UAAU;AAAA,MACrD;AAAA,IACF;AACA,QAAI,OAAO,SAAU,OAAM,IAAI,MAAM,sBAAsB;AAC3D,QAAI,OAAO,SAAS,GAAG;AACrB,YAAM,IAAI;AAAA,QACR,qBAAqB,OAAO,IAAI,KAAK,OAAO,OAAO,KAAK,EAAE,MAAM,IAAI,CAAC;AAAA,MACvE;AAAA,IACF;AAWA,QAAI;AACJ,QAAI;AACF,gBAAU,KAAK,MAAM,OAAO,MAAM;AAAA,IACpC,QAAQ;AACN,YAAM,UAAU,OAAO,OAAO,KAAK,EAAE,QAAQ,QAAQ,GAAG,EAAE,MAAM,GAAG,GAAG;AACtE,YAAM,IAAI;AAAA,QACR,0DAA0D,WAAW,aAAa;AAAA,MACpF;AAAA,IACF;AACA,QAAI,OAAO,QAAQ,WAAW,UAAU;AACtC,YAAM,IAAI,MAAM,qCAAqC;AAAA,IACvD;AACA,WAAO,EAAE,MAAM,YAAY,QAAQ,MAAM,EAAE;AAAA,EAC7C;AACF;;;AC1EO,IAAM,eAAN,MAAgD;AAAA,EAKrD,YACmB,WACA,QAAQ,cACzB;AAFiB;AACA;AAEjB,QAAI,UAAU,WAAW,GAAG;AAC1B,YAAM,IAAI,MAAM,mDAAmD;AAAA,IACrE;AAAA,EACF;AAAA,EANmB;AAAA,EACA;AAAA,EANX,QAAQ;AAAA;AAAA,EAEA,WAAkC,CAAC;AAAA,EAWnD,WAAmB;AACjB,WAAO;AAAA,EACT;AAAA,EAEA,YAAoB;AAClB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,aAAa,KAAyD;AACpE,SAAK,SAAS,KAAK,GAAG;AACtB,UAAM,WAAW,KAAK,UAAU,KAAK,QAAQ,KAAK,UAAU,MAAM;AAClE,SAAK,SAAS;AACd,QAAI,WAAW,UAAU;AACvB,aAAO,QAAQ,OAAO,IAAI,MAAM,SAAS,KAAK,CAAC;AAAA,IACjD;AACA,WAAO,QAAQ,QAAQ;AAAA,MACrB,MAAM,SAAS;AAAA,MACf,OAAO,SAAS,SAAS,EAAE,aAAa,KAAK,cAAc,IAAI;AAAA,IACjE,CAAC;AAAA,EACH;AACF;AAGO,SAAS,YACd,OACA,YACA,YAIK,CAAC,GACa;AACnB,SAAO;AAAA,IACL,MAAM;AAAA,MACJ,OAAO,UAAU,SAAS;AAAA,MAC1B,UAAU,UAAU,YAAY;AAAA,MAChC;AAAA,MACA;AAAA,MACA,WAAW,UAAU,aAAa;AAAA,IACpC;AAAA,EACF;AACF;;;AC/DA,SAAS,kBAAkB;AAC3B,SAAS,YAAY,WAAW,cAAc,qBAAqB;AACnE,SAAS,YAAY;AAEd,SAAS,OAAO,MAAsB;AAC3C,SAAO,WAAW,QAAQ,EAAE,OAAO,MAAM,MAAM,EAAE,OAAO,KAAK;AAC/D;AAOO,SAAS,cAAc,OAAyB;AAIrD,SAAO,OAAO,MAAM,IAAI,CAAC,MAAM,GAAG,EAAE,MAAM,IAAI,CAAC,EAAE,EAAE,KAAK,GAAG,CAAC;AAC9D;AAEO,IAAM,YAAN,MAAmB;AAAA,EAIxB,YACmB,KACA,UAAmB,MAEnB,QAAgB,aACjC;AAJiB;AACA;AAEA;AAAA,EAChB;AAAA,EAJgB;AAAA,EACA;AAAA,EAEA;AAAA;AAAA,EANX,SAAS;AAAA,EASjB,IAAI,KAA4B;AAC9B,QAAI,CAAC,KAAK,QAAS,QAAO;AAC1B,UAAM,OAAO,KAAK,KAAK,KAAK,GAAG,GAAG,OAAO;AACzC,QAAI,CAAC,WAAW,IAAI,EAAG,QAAO;AAC9B,QAAI;AACF,aAAO,KAAK,MAAM,aAAa,MAAM,MAAM,CAAC;AAAA,IAC9C,QAAQ;AACN,aAAO;AAAA,IACT;AAAA,EACF;AAAA,EAEA,IAAI,KAAa,OAAgB;AAC/B,QAAI,CAAC,KAAK,QAAS;AAInB,QAAI;AACF,gBAAU,KAAK,KAAK,EAAE,WAAW,KAAK,CAAC;AACvC;AAAA,QACE,KAAK,KAAK,KAAK,GAAG,GAAG,OAAO;AAAA,QAC5B,KAAK,UAAU,OAAO,MAAM,CAAC;AAAA,MAC/B;AAAA,IACF,SAAS,GAAG;AACV,UAAI,CAAC,KAAK,QAAQ;AAChB,aAAK,SAAS;AACd,gBAAQ;AAAA,UACN,GAAG,KAAK,KAAK,kCAAkC,KAAK,GAAG,KACrD,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC,CAC3C;AAAA,QACF;AAAA,MACF;AAAA,IACF;AAAA,EACF;AACF;;;ACrDA,SAAS,mBAAmB;AAC5B,SAAS,eAAe;AACxB,SAAS,QAAAA,aAAY;AAcd,SAAS,8BAAsC;AACpD,SACE,QAAQ,IAAI,sBAAsB,KAClCC,MAAK,QAAQ,GAAG,wBAAwB,QAAQ;AAEpD;AAGO,IAAM,cAAc,CAAC,QAAQ,YAAY,SAAS;AAIlD,IAAM,kBAAkB,CAAC,QAAQ,GAAG,WAAW;AAuB/C,IAAM,eACX,kBAAkB;AAAA,EAChB,eAAe;AAAA,IACb,KAAK;AAAA,IACL,WAAW;AAAA,IACX,SAAS;AAAA,IACT,MAAM;AAAA,IACN,OAAO;AAAA,EACT;AAAA,EACA,eAAe;AAAA,IACb,KAAK;AAAA,IACL,WAAW;AAAA,IACX,SAAS;AAAA,IACT,MAAM;AAAA,IACN,OAAO;AAAA,EACT;AAAA,EACA,eAAe;AAAA,IACb,KAAK;AAAA,IACL,WAAW;AAAA,IACX,SAAS;AAAA,IACT,MAAM;AAAA,IACN,OAAO;AAAA,EACT;AAAA,EACA,mBAAmB;AAAA,IACjB,KAAK;AAAA,IACL,WAAW;AAAA,IACX,SAAS;AAAA,IACT,OACE;AAAA,EACJ;AAAA,EACA,kBAAkB;AAAA,IAChB,KAAK;AAAA,IACL,WAAW;AAAA,IACX,SAAS;AAAA,IACT,OAAO;AAAA,EACT;AACF,CAAC;AAEH,SAAS,kBACP,SACa;AACb,aAAW,SAAS,OAAO,OAAO,OAAO,EAAG,QAAO,OAAO,KAAK;AAC/D,SAAO,OAAO,OAAO,OAAO;AAC9B;AAGA,IAAM,aAAwC;AAAA,EAC5C,MAAM;AAAA,EACN,UAAU;AAAA,EACV,SAAS;AACX;AAEO,SAAS,gBAAgB,OAAuC;AACrE,SAAQ,gBAAsC,SAAS,KAAK;AAC9D;AAQA,IAAM,kBAAkB;AAUjB,SAAS,cAAc,aAAgC;AAC5D,MAAI,SAAoB;AACxB,aAAW,QAAQ,aAAa;AAC9B,UAAM,QAAQ,aAAa,WAAW,IAAI,CAAC;AAC3C,QAAI,MAAM,YAAY,mBAAmB,YAAa,UAAS;AAAA,EACjE;AAGA,SAAO;AACT;AAMO,SAAS,aAAa,MAAyB;AACpD,SAAO,WAAW,IAAI;AACxB;AAGO,SAAS,WAAW,MAAyB;AAClD,SAAO,aAAa,WAAW,IAAI,CAAC,EAAG;AACzC;AAWO,SAAS,qBAAqB,OAAuB;AAC1D,MAAI,gBAAgB,KAAK,GAAG;AAC1B,UAAM,IAAI;AAAA,MACR,oBAAoB,KAAK;AAAA,IAG3B;AAAA,EACF;AACA,QAAM,QAAQ,aAAa,KAAK;AAChC,MAAI,MAAO,QAAO,MAAM;AACxB,MAAI,iBAAiB,KAAK,EAAG,QAAO;AACpC,QAAM,IAAI;AAAA,IACR,4BAA4B,KAAK,sBAAsB,gBAAgB;AAAA,MACrE;AAAA,IACF,CAAC,uBAAuB,OAAO,KAAK,YAAY,EAAE;AAAA,MAChD;AAAA,IACF,CAAC;AAAA,EACH;AACF;AASO,SAAS,YAAY,OAAuB;AACjD,QAAM,MAAM,qBAAqB,KAAK,EAAE,MAAM,GAAG,EAAE,CAAC;AACpD,SAAO,IAAI,MAAM,OAAO,EAAE,IAAI;AAChC;AAQO,SAAS,iBAAiB,OAAe,UAA2B;AACzE,QAAM,OAAO,MAAM,QAAQ,YAAY,EAAE;AACzC,MAAI,KAAK,SAAS,QAAQ,EAAG,QAAO;AAGpC,QAAM,OAAO,SAAS,QAAQ,WAAW,EAAE;AAC3C,SAAO,IAAI,OAAO,GAAG,aAAa,IAAI,CAAC,2BAA2B,EAAE,KAAK,IAAI;AAC/E;AAEA,SAAS,aAAa,MAAsB;AAC1C,SAAO,KAAK,QAAQ,uBAAuB,MAAM;AACnD;AASO,SAAS,kBAAkB,OAAe,WAA4B;AAC3E,QAAM,WAAW,YAAY,KAAK;AAClC,SAAO,mBAAmB,SAAS,EAAE;AAAA,IACnC,CAAC,UAAU,CAAC,MAAM,SAAS,QAAQ,KAAK,iBAAiB,OAAO,QAAQ;AAAA,EAC1E;AACF;AAGO,SAAS,mBAAmB,WAA6B;AAC9D,MAAI;AACF,WAAO,YAAY,WAAW,EAAE,eAAe,KAAK,CAAC,EAClD,OAAO,CAAC,UAAU,MAAM,OAAO,CAAC,EAChC,IAAI,CAAC,UAAU,MAAM,IAAI;AAAA,EAC9B,QAAQ;AAEN,WAAO,CAAC;AAAA,EACV;AACF;AAYA,SAAS,iBAAiB,OAAwB;AAChD,SACE,sBAAsB,KAAK,KAAK,KAChC,gBAAgB,KAAK,KAAK,KAC1B,2BAA2B,KAAK,KAAK,KACrC,MAAM,SAAS,OAAO;AAE1B;;;AC7PA;AAAA,EACE,cAAAC;AAAA,EACA,aAAAC;AAAA,EACA;AAAA,EACA;AAAA,EACA,iBAAAC;AAAA,OACK;AACP,SAAS,WAAAC,gBAAe;AACxB,SAAS,QAAAC,aAAY;AACrB,SAAS,qBAAqB;AAM9B,IAAM,eAAe;AAMrB,IAAM,OAAO;AACb,IAAM,OAAO;AAOb,IAAM,qBAAqB;AAG3B,IAAM,eAAe;AAErB,IAAM,gBAAgB,qBAAqB;AA6BpC,SAAS,6BACd,MAA0C,QAAQ,KAC1C;AACR,SACE,IAAI,uBAAuB,KAC3BC,MAAKC,SAAQ,GAAG,wBAAwB,SAAS;AAErD;AAYO,SAAS,iBAAiB,GAAqB;AACpD,QAAM,OAAQ,GAAyC;AACvD,SAAO,SAAS,0BAA0B,SAAS;AACrD;AAEA,SAAS,SAAS,GAAoB;AACpC,SAAO,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC;AAClD;AAoBA,eAAsB,mBACpB,UAAiC,CAAC,GACV;AACxB,QAAM,MAAM,QAAQ,OAAO,QAAQ;AACnC,QAAM,YAAY,QAAQ,aAAa,6BAA6B,GAAG;AAEvE,QAAM,cACJ,QAAQ,gBAAgB,MAAwB,OAAO,gBAAgB;AACzE,MAAI;AACF,UAAM,YAAY;AAClB,WAAO,EAAE,OAAO,UAAU;AAAA,EAC5B,SAAS,GAAG;AAKV,QAAI,CAAC,iBAAiB,CAAC,GAAG;AACxB,aAAO;AAAA,QACL,OAAO;AAAA,QACP,QAAQ,mDAAmD,SAAS,CAAC,CAAC;AAAA,MACxE;AAAA,IACF;AAAA,EAEF;AACA,MAAIC,YAAWF,MAAK,WAAW,IAAI,CAAC,EAAG,QAAO,EAAE,OAAO,UAAU;AAEjE,OAAK,IAAI,2BAA2B,KAAK,QAAQ,IAAI;AACnD,WAAO;AAAA,MACL,OAAO;AAAA,MACP,QAAQ;AAAA,IACV;AAAA,EACF;AACA,SAAO,EAAE,OAAO,eAAe,UAAU;AAC3C;AAOA,IAAM,WAAW,oBAAI,IAA8B;AAEnD,IAAI,gBAAgB;AAGb,SAAS,sBAA4B;AAC1C,WAAS,MAAM;AACf,kBAAgB;AAClB;AASO,SAAS,mBACd,UAAiC,CAAC,GAChB;AAClB,QAAM,MAAM,QAAQ,OAAO,QAAQ;AACnC,QAAM,YAAY,QAAQ,aAAa,6BAA6B,GAAG;AAEvE,QAAM,WAAW,SAAS,IAAI,SAAS;AACvC,MAAI,SAAU,QAAO;AAErB,QAAM,UAAU,WAAW,WAAW,KAAK,OAAO;AAIlD,QAAM,UAAU,QAAQ,MAAM,CAAC,MAAe;AAC5C,QAAI,SAAS,IAAI,SAAS,MAAM,QAAS,UAAS,OAAO,SAAS;AAClE,UAAM;AAAA,EACR,CAAC;AACD,WAAS,IAAI,WAAW,OAAO;AAC/B,SAAO;AACT;AAEA,eAAe,WACb,WACA,KACA,SACkB;AAClB,QAAM,aACJ,QAAQ,eAAe,CAAC,QAAkC,OAAO;AACnE,QAAM,OAAOA,MAAK,WAAW,IAAI;AAEjC,MAAIE,YAAW,IAAI,EAAG,QAAO,WAAW,cAAc,IAAI,EAAE,IAAI;AAEhE,OAAK,IAAI,2BAA2B,KAAK,QAAQ,IAAI;AACnD,UAAM,IAAI;AAAA,MACR,iHACwC,YAAY,+DACI,SAAS;AAAA,IAEnE;AAAA,EACF;AAEA,EAAAC,WAAU,WAAW,EAAE,WAAW,KAAK,CAAC;AACxC,QAAM,SAAS,WAAW,YAAY;AAEpC,QAAID,YAAW,IAAI,EAAG;AACtB,mBAAe,SAAS;AACxB,UAAM,WAAW,WAAW,KAAK,OAAO;AAExC,IAAAE,eAAc,MAAM;AAAA,GAAqC,MAAM;AAAA,EACjE,CAAC;AAED,SAAO,WAAW,cAAc,IAAI,EAAE,IAAI;AAC5C;AAEA,eAAe,WACb,WACA,KACA,SACe;AAIf,QAAM,WAAWJ,MAAK,WAAW,cAAc;AAC/C,MAAI,CAACE,YAAW,QAAQ,GAAG;AACzB,IAAAE;AAAA,MACE;AAAA,MACA,GAAG,KAAK;AAAA,QACN;AAAA,UACE,MAAM;AAAA,UACN,SAAS;AAAA,UACT,SAAS;AAAA,UACT,aACE;AAAA,QACJ;AAAA,QACA;AAAA,QACA;AAAA,MACF,CAAC;AAAA;AAAA,MACD;AAAA,IACF;AAAA,EACF;AAEA,QAAM,OAAO,QAAQ,QAAQ;AAC7B,QAAM,SAAS,MAAM;AAAA,IACnB;AAAA,MACE;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,IACA,EAAE,WAAW,QAAQ,aAAa,oBAAoB,IAAI;AAAA,EAC5D;AAEA,MAAI,OAAO,cAAc,MAAM;AAC7B,UAAM,IAAI;AAAA,MACR,gDAAgD,OAAO,UAAU,sCAC5B,YAAY;AAAA,IAEnD;AAAA,EACF;AACA,MAAI,OAAO,UAAU;AACnB,UAAM,IAAI;AAAA,MACR,kCAAkC,SAAS,mHAE1B,YAAY;AAAA,IAC/B;AAAA,EACF;AACA,MAAI,OAAO,SAAS,GAAG;AACrB,UAAM,IAAI;AAAA,MACR,kCAAkC,SAAS,iBAAiB;AAAA,QAC1D,OAAO;AAAA,MACT,CAAC;AAAA,EAAO,KAAK,OAAO,UAAU,OAAO,MAAM,CAAC;AAAA,kCACP,YAAY;AAAA,IAEnD;AAAA,EACF;AACF;AAGA,SAAS,KAAK,QAAgB,QAAQ,IAAY;AAChD,SAAO,OAAO,QAAQ,EAAE,MAAM,OAAO,EAAE,MAAM,CAAC,KAAK,EAAE,KAAK,IAAI;AAChE;AAMA,SAAS,eAAe,WAAyB;AAC/C,MAAI,cAAe;AACnB,kBAAgB;AAChB,UAAQ;AAAA,IACN,sEAAiE,SAAS;AAAA,EAI5E;AACF;AASA,eAAe,SACb,WACA,IACe;AACf,QAAM,OAAOJ,MAAK,WAAW,IAAI;AACjC,QAAM,WAAW,KAAK,IAAI,IAAI;AAE9B,aAAS;AACP,QAAI;AACF,MAAAI,eAAc,MAAM,OAAO,QAAQ,GAAG,GAAG,EAAE,MAAM,KAAK,CAAC;AACvD;AAAA,IACF,SAAS,GAAG;AACV,UAAK,EAA4B,SAAS,SAAU,OAAM;AAC1D,UAAI,MAAM,IAAI,IAAI,eAAe;AAG/B,eAAO,MAAM,EAAE,OAAO,KAAK,CAAC;AAC5B;AAAA,MACF;AACA,UAAIF,YAAWF,MAAK,WAAW,IAAI,CAAC,EAAG;AACvC,UAAI,KAAK,IAAI,IAAI,UAAU;AACzB,cAAM,IAAI;AAAA,UACR,wEACK,SAAS,wCAAwC,IAAI;AAAA,QAC5D;AAAA,MACF;AACA,YAAM,MAAM,GAAG;AAAA,IACjB;AAAA,EACF;AAEA,MAAI;AACF,UAAM,GAAG;AAAA,EACX,UAAE;AACA,WAAO,MAAM,EAAE,OAAO,KAAK,CAAC;AAAA,EAC9B;AACF;AAGA,SAAS,MAAM,MAAsB;AACnC,MAAI;AACF,WAAO,KAAK,IAAI,IAAI,SAAS,IAAI,EAAE;AAAA,EACrC,QAAQ;AACN,WAAO,OAAO;AAAA,EAChB;AACF;AAEA,SAAS,MAAM,IAA2B;AACxC,SAAO,IAAI,QAAQ,CAAC,YAAY,WAAW,SAAS,EAAE,CAAC;AACzD;;;AC/QA,IAAM,eAAe,oBAAI,IAAuC;AAShE,eAAsB,qBAAoC;AACxD,QAAM,UAAU,CAAC,GAAG,aAAa,OAAO,CAAC;AACzC,eAAa,MAAM;AACnB,QAAM,QAAQ;AAAA,IACZ,QAAQ,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC,MAAM,EAAE,QAAQ,CAAC,EAAE,MAAM,MAAM,MAAS,CAAC;AAAA,EACtE;AACF;AAEO,IAAM,mBAAN,MAAoD;AAAA,EAczD,YACmB,OACjB,UAAmC,CAAC,GACpC;AAFiB;AAGjB,QAAI,gBAAgB,KAAK,GAAG;AAC1B,YAAM,IAAI;AAAA,QACR,oBAAoB,KAAK;AAAA,MAG3B;AAAA,IACF;AACA,SAAK,MAAM,qBAAqB,KAAK;AACrC,SAAK,UAAU,QAAQ,WAAW,oBAAoB;AACtD,SAAK,gBAAgB,QAAQ,iBAAiB;AAC9C,SAAK,YAAY,QAAQ;AACzB,SAAK,kBACH,QAAQ,mBAAmB,4BAA4B;AACzD,SAAK,WAAW,cAAc,CAAC,KAAK,iBAAiB,KAAK,GAAG,CAAC;AAAA,EAChE;AAAA,EAjBmB;AAAA,EAdF;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA;AAAA,EAsBjB,WAAmB;AACjB,WAAO;AAAA,EACT;AAAA,EAEA,YAAoB;AAClB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAM,aAAa,KAAyD;AAC1E,UAAM,QAAQ,MAAM,KAAK,KAAK;AAG9B,UAAM,UAAU,MAAM,MAAM,cAAc,gBAAgB,GAAG,CAAC;AAC9D,QAAI;AACF,YAAM,SAAS,MAAM,QAAQ,OAAO,IAAI,MAAM;AAAA,QAC5C,QAAQ,IAAI;AAAA,QACZ,aAAa,IAAI;AAAA,QACjB,eAAe,KAAK;AAAA,QACpB,GAAI,KAAK,aAAa,OAAO,EAAE,WAAW,KAAK,UAAU,IAAI,CAAC;AAAA,MAChE,CAAC;AAMD,UAAI,OAAO,eAAe,aAAa;AACrC,cAAM,IAAI;AAAA,UACR,sEACK,KAAK,aAAa,OAAO,gBAAgB,KAAK,SAAS,MAAM,EAAE;AAAA,QAEtE;AAAA,MACF;AACA,aAAO,EAAE,MAAM,YAAY,OAAO,IAAI,GAAG,OAAO,OAAO,MAAM;AAAA,IAC/D,UAAE;AACA,YAAM,QAAQ,QAAQ,EAAE,MAAM,MAAM,MAAS;AAAA,IAC/C;AAAA,EACF;AAAA,EAEQ,OAAkC;AACxC,UAAM,WAAW,aAAa,IAAI,KAAK,QAAQ;AAC/C,QAAI,SAAU,QAAO;AACrB,UAAM,WAAW,YAAY;AAC3B,YAAM,OAAO,MAAM,KAAK,QAAQ;AAAA,QAC9B,KAAK;AAAA,QACL,KAAK;AAAA,MACP;AACA,aAAO,KAAK,QAAQ,UAAU,IAAI;AAAA,IACpC,GAAG;AAKH,UAAM,UAAU,QAAQ,MAAM,CAAC,MAAe;AAC5C,UAAI,aAAa,IAAI,KAAK,QAAQ,MAAM,SAAS;AAC/C,qBAAa,OAAO,KAAK,QAAQ;AAAA,MACnC;AACA,YAAM;AAAA,IACR,CAAC;AACD,iBAAa,IAAI,KAAK,UAAU,OAAO;AACvC,WAAO;AAAA,EACT;AACF;AASA,SAAS,gBAAgB,KAAkC;AACzD,SAAO,GAAG,IAAI,MAAM;AAAA;AAAA;AAAA,EAAwE,KAAK;AAAA,IAC/F,IAAI;AAAA,EACN,CAAC;AACH;AAGA,IAAI;AAOG,SAAS,sBAAoC;AAClD,QAAM,OAAO;AAAA;AAAA;AAAA;AAAA;AAAA,IAKV,mBAAmB,iBAAiB,EAAE,MAAM,CAAC,MAAe;AAC3D,uBAAiB;AACjB,YAAM;AAAA,IACR,CAAC;AAAA;AACH,SAAO;AAAA,IACL,kBAAkB,CAAC,KAAK,cACtB,KAAK,EAAE,KAAK,CAAC,MAAM,EAAE,iBAAiB,KAAK,SAAS,CAAC;AAAA,IACvD,WAAW,CAAC,SAAS,KAAK,EAAE,KAAK,CAAC,MAAM,EAAE,UAAU,IAAI,CAAC;AAAA,IACzD,sBAAsB,MAAM,KAAK,EAAE,KAAK,CAAC,MAAM,EAAE,qBAAqB,CAAC;AAAA,EACzE;AACF;AAEA,eAAe,mBAA0C;AACvD,MAAI;AACJ,MAAI;AACF,UAAM,MAAM,OAAO,gBAAgB;AAAA,EACrC,SAAS,GAAG;AAKV,QAAI,CAAC,iBAAiB,CAAC,GAAG;AACxB,YAAM,IAAI;AAAA,QACR,mDACE,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC,CAC3C;AAAA,MAGF;AAAA,IACF;AAMA,UAAO,MAAM,mBAAmB;AAAA,EAClC;AAEA,QAAM,EAAE,UAAU,kBAAkB,kBAAkB,WAAW,IAAI;AACrE,QAAM,QAAQ,MAAM,SAAS;AAE7B,SAAO;AAAA;AAAA;AAAA,IAGL,kBAAkB,CAAC,KAAK,cAAc,iBAAiB,KAAK,EAAE,UAAU,CAAC;AAAA,IAEzE,MAAM,UAAU,MAAM;AACpB,YAAM,QAAQ,MAAM,MAAM,UAAU,EAAE,WAAW,KAAK,CAAC;AACvD,aAAO;AAAA,QACL,MAAM,cAAc,cAAc;AAChC,gBAAM,UAAU,MAAM,MAAM,cAAc;AAC1C,gBAAM,WAAW,QAAQ,YAAY;AACrC,gBAAM,UAAU,IAAI,iBAAiB;AAAA,YACnC,iBAAiB;AAAA,YACjB;AAAA,UACF,CAAC;AACD,iBAAO;AAAA,YACL,MAAM,OAAO,MAAM,SAAS;AAC1B,oBAAM,UAAU,MAAM,MAAM;AAAA,gBAC1B,QAAQ;AAAA,cAGV;AACA,oBAAM,SAAS,SAAS,WAAW,SAAS;AAC5C,oBAAM,SAAS,MAAM,QAAQ,eAAe,MAAM;AAAA,gBAChD;AAAA,gBACA,aAAa,QAAQ;AAAA,gBACrB,SAAS,EAAE,eAAe,QAAQ,cAAc;AAAA,gBAChD,GAAI,QAAQ,aAAa,OACrB,EAAE,WAAW,QAAQ,UAAU,IAC/B,CAAC;AAAA,cACP,CAAC;AAED,oBAAM,OAAO,WAAW,KAAK,SAAS,YAAY,MAAM;AACxD,qBAAO;AAAA,gBACL,MAAM,OAAO;AAAA,gBACb,YAAY,OAAO;AAAA,gBACnB,OAAO;AAAA,kBACL,aAAa,KAAK;AAAA,kBAClB,cAAc,KAAK;AAAA,gBACrB;AAAA,cACF;AAAA,YACF;AAAA,YACA,MAAM,UAAU;AACd,oBAAM,QAAQ,QAAQ;AAAA,YACxB;AAAA,UACF;AAAA,QACF;AAAA,QACA,MAAM,UAAU;AACd,gBAAM,MAAM,QAAQ;AAAA,QACtB;AAAA,MACF;AAAA,IACF;AAAA,IAEA,MAAM,uBAAuB;AAC3B,YAAM,EAAE,SAAS,IAAI,MAAM,OAAO,IAAS;AAG3C,YAAM,YAAY,SAAS,IAAI;AAC/B,UAAI;AACF,cAAM,OAAO,MAAM,MAAM,aAAa;AAKtC,eAAO,KAAK,IAAI,KAAK,MAAM,SAAS;AAAA,MACtC,QAAQ;AAEN,eAAO;AAAA,MACT;AAAA,IACF;AAAA,EACF;AACF;;;AC9UO,IAAM,kBAA2C;AAAA,EACtD;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEA,IAAM,kBAAyD;AAAA,EAC7D,WAAW;AAAA,EACX,QAAQ;AACV;AAqBA,SAAS,OAAO,UAA2C;AAEzD,UAAQ,QAAQ,IAAI,gBAAgB,QAAQ,CAAE,KAAK,QAAQ;AAC7D;AAWA,IAAM,YAAY,oBAAI,IAA8B;AAG7C,SAAS,sBAA4B;AAC1C,YAAU,MAAM;AAClB;AAEA,SAAS,eAAe,MAAsC;AAC5D,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,UAAU,KAAK,WAAW;AAChC,QAAM,SAAS,UAAU,IAAI,OAAO;AACpC,MAAI,OAAQ,QAAO;AACnB,QAAM,UAAU,KAAK,CAAC,SAAS,WAAW,GAAG,EAAE,WAAW,IAAO,CAAC,EAC/D,KAAK,CAAC,MAAM,EAAE,SAAS,KAAK,CAAC,EAAE,YAAY,EAAE,cAAc,IAAI,EAC/D,MAAM,MAAM,KAAK;AACpB,YAAU,IAAI,SAAS,OAAO;AAC9B,SAAO;AACT;AAEA,eAAe,cAAc,MAAoC;AAC/D,QAAM,WAAW,KAAK,gBAAgB,KAAK,UAAU;AACrD,MAAI,SAAU,QAAO,YAAY,QAAQ;AAMzC,QAAM,SAAS,MAAM,mBAAmB;AACxC,MAAI,OAAO,UAAU,WAAW;AAC9B,WAAO,EAAE,WAAW,OAAO,QAAQ,OAAO,OAAO;AAAA,EACnD;AACA,MAAI,OAAO,UAAU,eAAe;AAGlC,WAAO,EAAE,WAAW,KAAK;AAAA,EAC3B;AAGA,SAAO,YAAY,oBAAoB,CAAC;AAC1C;AAMA,SAAS,YAAY,SAAuC;AAC1D,SAAO,QAAQ,qBAAqB,EAAE;AAAA,IACpC,OAAO,EAAE,WAAW,KAAK;AAAA,IACzB,CAAC,OAAgB;AAAA,MACf,WAAW;AAAA,MACX,QACE,aAAa,SAAS,iBAAiB,KAAK,EAAE,OAAO,IACjD,2DACA,mCACE,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC,CAC3C;AAAA,IACR;AAAA,EACF;AACF;AAEA,eAAe,MACb,UACA,MACgB;AAChB,UAAQ,UAAU;AAAA,IAChB,KAAK;AACH,aAAO,OAAO,WAAW,IACrB,EAAE,WAAW,KAAK,IAClB;AAAA,QACE,WAAW;AAAA,QACX,QAAQ;AAAA,MACV;AAAA,IACN,KAAK;AAGH,aAAO,OAAO,QAAQ,KAAK,KAAK,UAC5B,EAAE,WAAW,KAAK,IAClB;AAAA,QACE,WAAW;AAAA,QACX,QAAQ;AAAA,MACV;AAAA,IACN,KAAK;AACH,aAAQ,MAAM,eAAe,IAAI,IAC7B,EAAE,WAAW,KAAK,IAClB;AAAA,QACE,WAAW;AAAA,QACX,QAAQ,mBAAmB,KAAK,WAAW,QAAQ;AAAA,MACrD;AAAA,IACN,KAAK;AACH,aAAO,cAAc,IAAI;AAAA,IAC3B;AACE,aAAO,EAAE,WAAW,OAAO,QAAQ,sBAAsB;AAAA,EAC7D;AACF;AAQA,eAAsB,mBACpB,OAAqB,CAAC,GACG;AACzB,QAAM,SAAS,MAAM,QAAQ;AAAA,IAC3B,gBAAgB,IAAI,CAAC,SAAS,MAAM,MAAM,IAAI,CAAC;AAAA,EACjD;AACA,SAAO,gBAAgB,OAAO,CAAC,GAAG,MAAM,OAAO,CAAC,EAAG,SAAS;AAC9D;AASA,eAAsB,eACpB,OAAqB,CAAC,GACC;AAQvB,QAAM,UAAoB,CAAC;AAC3B,aAAW,QAAQ,iBAAiB;AAClC,UAAM,SAAS,MAAM,MAAM,MAAM,IAAI;AACrC,QAAI,OAAO,WAAW;AACpB,mBAAa,MAAM,KAAK,aAAa,MAAM;AAC3C,aAAO;AAAA,IACT;AACA,YAAQ,KAAK,KAAK,KAAK,OAAO,EAAE,CAAC,WAAM,OAAO,MAAM,EAAE;AAAA,EACxD;AACA,QAAM,IAAI;AAAA,IACR;AAAA,EAA+C,QAAQ,KAAK,IAAI,CAAC;AAAA;AAAA,EAEnE;AACF;AAEA,IAAI,kBAAkB;AACtB,IAAI,iBAAiB;AAGd,SAAS,gCAAsC;AACpD,oBAAkB;AAClB,mBAAiB;AACnB;AAOA,SAAS,aAAa,UAAwB,iBAAgC;AAC5E,MAAI,gBAAiB;AACrB,oBAAkB;AAGlB,QAAM,UAAU,kBAAkB,oBAAoB;AACtD,UAAQ;AAAA,IACN,cAAc,OAAO,0BAAqB,QAAQ;AAAA,EAEpD;AACF;AAMO,SAAS,oBAAoB,OAAe,WAAyB;AAC1E,MAAI,eAAgB;AACpB,mBAAiB;AACjB,UAAQ;AAAA,IACN,eAAe,KAAK,6DACb,YAAY,KAAK,QAAQ,CAAC,CAAC;AAAA,EAEpC;AACF;;;AC7KO,IAAM,iBAA+C;AAAA,EAC1D,WAAW;AAAA,EACX,QAAQ;AAAA,EACR,cAAc;AAAA,EACd,MAAM;AAAA;AAAA;AAAA;AAAA,EAIN,aAAa;AACf;AAEA,IAAM,sBAA8C;AAAA,EAClD,WAAW;AAAA,EACX,QAAQ;AACV;AAEO,IAAM,0BAA0B;AAmBhC,SAAS,wBAAwB,MAAsC;AAC5E,MAAI,KAAK,YAAY,QAAQ,KAAK,aAAa,QAAQ;AACrD,UAAM,IAAI;AAAA,MACR,8NAGM,OAAO,KAAK,cAAc,EAAE,KAAK,IAAI,CAAC;AAAA,IAC9C;AAAA,EACF;AACA,QAAM,QAAQ,KAAK,SAAS,eAAe,KAAK,QAAQ,KAAK;AAC7D,MAAI,KAAK,aAAa,eAAe,gBAAgB,KAAK,GAAG;AAC3D,UAAM,IAAI;AAAA,MACR,oBAAoB,KAAK;AAAA,IAI3B;AAAA,EACF;AACA,SAAO,EAAE,UAAU,KAAK,UAAU,MAAM;AAC1C;AASA,eAAsB,6BACpB,MAC2B;AAM3B,OACG,KAAK,YAAY,QAAQ,KAAK,aAAa,WAC5C,KAAK,SAAS,MACd;AACA,UAAM,IAAI;AAAA,MACR,UAAU,KAAK,KAAK,gHAEd,OAAO,KAAK,cAAc,EAAE,KAAK,IAAI,CAAC;AAAA,IAE9C;AAAA,EACF;AAGA,QAAM,WACJ,KAAK,YAAY,QAAQ,KAAK,aAAa,SACvC,MAAM,eAAe,IAAI,IACzB,KAAK;AACX,QAAM,WAAyB,EAAE,GAAG,MAAM,SAAS;AAEnD,QAAM,QAAQ,KAAK,SAAS,eAAe,QAAQ,KAAK;AACxD,MAAI,aAAa,eAAe,CAAC,gBAAgB,KAAK,GAAG;AACvD,WAAO,wBAAwB,QAAQ;AAAA,EACzC;AACA,QAAM,OACJ,UAAU,SAAS,MAAM,UAAU,gBAAgB,IAAI,CAAC,IAAI;AAC9D,SAAO,EAAE,UAAU,OAAO,aAAa,IAAI,EAAE;AAC/C;AAWA,SAAS,sBAAsB,MAAoB,OAAqB;AACtE,QAAM,QAAQ,aAAa,KAAK;AAChC,MAAI,CAAC,MAAO;AACZ,QAAM,YACJ,KAAK,UAAU,mBAAmB,4BAA4B;AAChE,MAAI,CAAC,kBAAkB,OAAO,SAAS,GAAG;AACxC,wBAAoB,OAAO,MAAM,SAAS;AAAA,EAC5C;AACF;AAQA,SAAS,gBAAgB,MAA8C;AACrE,SAAO,KAAK,gBAAgB,KAAK,UAAU;AAC7C;AAEA,eAAe,UAAU,SAAuD;AAC9E,QAAM,SAAS,WAAW,oBAAoB;AAC9C,SAAO,cAAc,MAAM,OAAO,qBAAqB,CAAC;AAC1D;AAEO,SAAS,aAAa,MAAuC;AAIlE,wBAAsB;AACtB,QAAM,EAAE,MAAM,IAAI,wBAAwB,IAAI;AAE9C,UAAQ,KAAK,UAAU;AAAA,IACrB,KAAK;AACH,aAAO,IAAI;AAAA,QACT;AAAA,QACA,KAAK,aAAa,oBAAoB,WAAW;AAAA,QACjD,KAAK,aAAa,CAAC;AAAA,MACrB;AAAA,IACF,KAAK;AACH,aAAO,IAAI;AAAA,QACT,KAAK,WAAW;AAAA,QAChB;AAAA,QACA,KAAK,aAAa,oBAAoB,QAAQ;AAAA,QAC9C;AAAA,QACA,KAAK,UAAU,CAAC;AAAA,MAClB;AAAA,IACF,KAAK;AACH,aAAO,IAAI;AAAA,QACT;AAAA,QACA,KAAK,WAAW;AAAA,QAChB,KAAK;AAAA,QACL,KAAK;AAAA,MACP;AAAA,IACF,KAAK;AAEH,aAAO,IAAI,aAAa,KAAK,iBAAiB,CAAC,EAAE,MAAM,CAAC,EAAE,CAAC,GAAG,KAAK;AAAA,IACrE,KAAK;AACH,aAAO,IAAI,iBAAiB,OAAO;AAAA,QACjC,GAAI,KAAK,YAAY,CAAC;AAAA,QACtB,GAAI,KAAK,eAAe,EAAE,SAAS,KAAK,aAAa,IAAI,CAAC;AAAA,MAC5D,CAAC;AAAA,IACH;AACE,YAAM,IAAI;AAAA,QACR,qBAAqB,OAAO,KAAK,QAAQ,CAAC,iBAAiB,OAAO;AAAA,UAChE;AAAA,QACF,EAAE,KAAK,IAAI,CAAC;AAAA,MACd;AAAA,EACJ;AACF;AASA,eAAsB,kBACpB,MAC4B;AAG5B,QAAM,EAAE,UAAU,MAAM,IAAI,MAAM,6BAA6B,IAAI;AACnE,MAAI,aAAa,YAAa,uBAAsB,MAAM,KAAK;AAC/D,SAAO,aAAa,EAAE,GAAG,MAAM,UAAU,MAAM,CAAC;AAClD;;;ACtQA,SAAS,UAAAK,SAAQ,YAAAC,iBAAgB;AACjC,SAAS,QAAAC,aAAY;AAqCrB,eAAsB,iBACpB,UAAmC,CAAC,GACH;AACjC,QAAM,YAAY,QAAQ,aAAa,4BAA4B;AACnE,QAAM,SAAS,QAAQ,UAAU;AAGjC,QAAM,SAAS,QAAQ,QAAQ,IAAI,WAAW;AAE9C,MAAI,CAAC,OAAQ,OAAM,mBAAmB;AAEtC,QAAM,QAA4B,CAAC;AACnC,aAAW,SAAS,mBAAmB,SAAS,GAAG;AAGjD,QAAI,CAAC,YAAY,KAAK,EAAG;AACzB,QAAI,UAAU,CAAC,OAAO,KAAK,CAAC,SAAS,iBAAiB,OAAO,IAAI,CAAC,GAAG;AACnE;AAAA,IACF;AACA,UAAM,OAAOC,MAAK,WAAW,KAAK;AAClC,UAAM,YAAY,OAAO,IAAI;AAC7B,QAAI,cAAc,OAAW;AAC7B,QAAI,CAAC,QAAQ;AACX,UAAI;AACF,QAAAC,QAAO,IAAI;AAAA,MACb,QAAQ;AAGN;AAAA,MACF;AAAA,IACF;AACA,UAAM,KAAK,EAAE,MAAM,UAAU,CAAC;AAAA,EAChC;AAEA,SAAO;AAAA,IACL;AAAA,IACA,YAAY,MAAM,OAAO,CAAC,OAAO,SAAS,QAAQ,KAAK,WAAW,CAAC;AAAA,IACnE;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,YAAY,OAAwB;AAC3C,SAAO,MAAM,SAAS,OAAO,KAAK,MAAM,SAAS,aAAa;AAChE;AAEA,SAAS,OAAO,MAAkC;AAChD,MAAI;AACF,WAAOC,UAAS,IAAI,EAAE;AAAA,EACxB,QAAQ;AACN,WAAO;AAAA,EACT;AACF;;;AC3FA,SAAS,eAAe;AAmCxB,IAAM,iBAAiB,oBAAI,QAAkC;AAGtD,SAAS,aACd,QACkB;AAClB,QAAM,SAAS,eAAe,IAAI,MAAM;AACxC,MAAI,OAAQ,QAAO;AAOnB,QAAM,WAAW,IAAI,QAAQ,EAAE,WAAW,KAAK,CAAC,EAAE,QAAQ,MAAM;AAChE,iBAAe,IAAI,QAAQ,QAAQ;AACnC,SAAO;AACT;AAEA,eAAsB,sBACpB,SAC0B;AAG1B,wBAAsB;AACtB,QAAM;AAAA,IACJ;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,cAAc;AAAA,IACd,WAAW;AAAA,EACb,IAAI;AACJ,QAAM,WAAW,QAAQ,YAAY,aAAa,MAAM;AAExD,QAAM,QAAQ,KAAK,IAAI;AACvB,QAAM,OAAO;AAAA,IACX,UAAU,SAAS,SAAS;AAAA,IAC5B,OAAO,SAAS,UAAU;AAAA,IAC1B,QAAQ;AAAA,EACV;AAEA,MAAI,YAAY;AAChB,WAAS,UAAU,GAAG,UAAU,UAAU,WAAW;AACnD,QAAI;AACF,YAAM,WAAW,MAAM,SAAS,aAAa;AAAA,QAC3C;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,MACF,CAAC;AACD,UAAI,SAAS,SAAS,IAAI,GAAG;AAC3B,eAAO;AAAA,UACL,GAAG;AAAA,UACH,QAAQ,SAAS;AAAA,UACjB,OAAO,SAAS;AAAA,UAChB,YAAY,KAAK,IAAI,IAAI;AAAA,QAC3B;AAAA,MACF;AACA,kBAAY,uCAAuC,SAAS,UAAU,CAAC,GACpE,IAAI,CAAC,MAAM,GAAG,EAAE,YAAY,IAAI,EAAE,OAAO,EAAE,EAC3C,KAAK,IAAI,CAAC;AAAA,IACf,SAAS,GAAG;AACV,kBAAY,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC;AAAA,IACvD;AAAA,EACF;AACA,SAAO,EAAE,GAAG,MAAM,OAAO,WAAW,YAAY,KAAK,IAAI,IAAI,MAAM;AACrE;;;AC9FO,IAAM,cAAuC;AAAA,EAClD,qBAAqB,EAAE,cAAc,GAAG,eAAe,GAAG;AAAA,EAC1D,qBAAqB,EAAE,cAAc,GAAG,eAAe,GAAG;AAAA,EAC1D,oBAAoB,EAAE,cAAc,GAAG,eAAe,EAAE;AAAA,EACxD,mBAAmB,EAAE,cAAc,IAAI,eAAe,GAAG;AAAA,EACzD,eAAe,EAAE,cAAc,MAAM,eAAe,IAAI;AAAA,EACxD,UAAU,EAAE,cAAc,KAAK,eAAe,GAAG;AACnD;AAEO,SAAS,WACd,OACA,UACqB;AACrB,MAAI,SAAU,QAAO;AACrB,MAAI,YAAY,KAAK,EAAG,QAAO,YAAY,KAAK;AAGhD,QAAM,OAAO,OAAO,KAAK,WAAW,EACjC,OAAO,CAAC,MAAM,MAAM,WAAW,CAAC,CAAC,EACjC,KAAK,CAAC,GAAG,MAAM,EAAE,SAAS,EAAE,MAAM,EAAE,CAAC;AACxC,SAAO,OAAO,YAAY,IAAI,IAAI;AACpC;AAEO,SAAS,YACd,OACA,SACQ;AACR,MAAI,CAAC,SAAS,CAAC,QAAS,QAAO;AAC/B,SACG,MAAM,cAAc,MAAa,QAAQ,eACzC,MAAM,eAAe,MAAa,QAAQ;AAE/C;AAGO,SAAS,WACd,MACA,SACQ;AACR,MAAI,CAAC,QAAS,QAAO;AACrB,MAAI,MAAM;AACV,aAAW,OAAO,MAAM;AACtB,QAAI,CAAC,IAAI,SAAS,IAAI,OAAQ;AAC9B,WAAO,YAAY,IAAI,OAAO,OAAO;AAAA,EACvC;AACA,SAAO;AACT;;;AC9DA;AAAA,EACE,SAAW;AAAA,EACX,KAAO;AAAA,EACP,OAAS;AAAA,EACT,MAAQ;AAAA,EACR,UAAY,CAAC,SAAS,YAAY,SAAS,cAAc,WAAW;AAAA,EACpE,YAAc;AAAA,IACZ,OAAS;AAAA,MACP,MAAQ;AAAA,MACR,aAAe;AAAA,IACjB;AAAA,IACA,UAAY;AAAA,MACV,MAAQ;AAAA,MACR,aAAe;AAAA,IACjB;AAAA,IACA,OAAS;AAAA,MACP,MAAQ,CAAC,QAAQ,QAAQ,SAAS;AAAA,MAClC,aAAe;AAAA,IACjB;AAAA,IACA,YAAc;AAAA,MACZ,MAAQ;AAAA,MACR,SAAW;AAAA,MACX,SAAW;AAAA,MACX,aAAe;AAAA,IACjB;AAAA,IACA,WAAa;AAAA,MACX,MAAQ;AAAA,MACR,aAAe;AAAA,IACjB;AAAA,EACF;AAAA,EACA,sBAAwB;AAC1B;;;ACtBO,IAAM,iBAAiB;;;ACFvB,SAAS,iBACd,MAC+B;AAC/B,QAAM,QAAQ,EAAE,MAAM,GAAG,MAAM,GAAG,SAAS,GAAG,OAAO,EAAE;AACvD,MAAI,gBAAgB;AACpB,MAAI,kBAAkB;AACtB,aAAW,OAAO,MAAM;AACtB,QAAI,CAAC,IAAI,SAAS;AAChB,YAAM,SAAS;AACf;AAAA,IACF;AACA,UAAM,IAAI,QAAQ,KAAK,KAAK;AAC5B,qBAAiB,IAAI,QAAQ;AAC7B,uBAAmB;AAAA,EACrB;AAEA,QAAM,YAAY,MAAM;AACxB,QAAM,YAAY,MAAM,OAAO,MAAM;AAErC,QAAM,UAAiB,YAAY,YAAY,SAAS;AACxD,QAAM,SAAS,YAAY;AAC3B,QAAM,YAAY,SAAS,IAAI,KAAK,IAAI,WAAW,SAAS,IAAI,SAAS;AACzE,QAAM,iBACJ,kBAAkB,IAAI,gBAAgB,kBAAkB;AAE1D,SAAO,EAAE,MAAM,OAAO,SAAS,WAAW,eAAe;AAC3D;;;ACtBO,IAAM,gBAAgC,EAAE,UAAU,KAAK,UAAU,IAAI;AAErE,SAAS,QACd,WACA,aAA6B,eACvB;AACN,QAAM,EAAE,OAAO,eAAe,IAAI;AAClC,QAAM,gBACJ,MAAM,OAAO,KACb,MAAM,SAAS,KACf,MAAM,YAAY,KAClB,MAAM,UAAU;AAClB,QAAM,gBACJ,MAAM,SAAS,KAAK,MAAM,UAAU,KAAK,MAAM,OAAO,MAAM,UAAU;AAExE,MAAI,iBAAiB,kBAAkB,WAAW,SAAU,QAAO;AACnE,MAAI,iBAAiB,kBAAkB,WAAW,SAAU,QAAO;AACnE,SAAO;AACT;;;ACcA,IAAI,oBAAoB;AAMxB,eAAsB,YACpB,SACqB;AACrB,QAAM;AAAA,IACJ;AAAA,IACA;AAAA,IACA;AAAA,IACA,MAAM,WAAW;AAAA,IACjB,cAAc;AAAA,IACd,SAAS;AAAA,IACT;AAAA,IACA;AAAA,IACA,QAAQ;AAAA,EACV,IAAI;AAEJ,MAAI,cAAc,KAAK,CAAC,mBAAmB;AACzC,wBAAoB;AACpB,YAAQ;AAAA,MACN,GAAG,KAAK,0BAA0B,WAAW;AAAA,IAC/C;AAAA,EACF;AAEA,MAAI,SAAS,UAAU;AACrB,UAAM,MAAM,MAAM,IAAI,QAAQ;AAK9B,QAAI,MAAM,QAAQ,GAAG,EAAG,QAAO,IAAI,IAAI,CAAC,OAAO,EAAE,GAAG,GAAG,QAAQ,KAAK,EAAE;AAAA,EACxE;AAGA,QAAM,WAAW,aAAa,MAAM;AAEpC,QAAM,UAAsB,CAAC;AAC7B,WAAS,IAAI,GAAG,IAAI,UAAU,KAAK;AACjC,UAAM,MAAM,MAAM,sBAAoC;AAAA,MACpD;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AAGD,UAAM,EAAE,QAAQ,GAAG,KAAK,IAAI;AAC5B,YAAQ,KAAK,WAAW,SAAY,OAAO,EAAE,GAAG,MAAM,SAAS,OAAO,CAAC;AAAA,EACzE;AAEA,MAAI,SAAS,SAAU,OAAM,IAAI,UAAU,OAAO;AAClD,SAAO;AACT;AAGA,eAAsB,MACpB,SAC0B;AAC1B,QAAM,OAAO,MAAM,YAAY,OAAO;AACtC,QAAM,OAAO,iBAAiB,IAAI;AAClC,SAAO,EAAE,GAAG,MAAM,MAAM,QAAQ,MAAM,QAAQ,SAAS,aAAa,EAAE;AACxE;AAGO,SAAS,0BAAgC;AAC9C,sBAAoB;AACtB;","names":["join","join","existsSync","mkdirSync","writeFileSync","homedir","join","join","homedir","existsSync","mkdirSync","writeFileSync","rmSync","statSync","join","join","rmSync","statSync"]}
+{"version":3,"sources":["../src/types.ts","../src/runtime.ts","../src/providers/anthropic.ts","../src/providers/openai-compat.ts","../src/exec.ts","../src/providers/claude-cli.ts","../src/providers/mock.ts","../src/cache.ts","../src/providers/llama-models.ts","../src/providers/llama-install.ts","../src/providers/llama-cpp.ts","../src/providers/detect.ts","../src/providers/index.ts","../src/providers/llama-clean.ts","../src/complete.ts","../src/cost.ts","../src/judge/verdict-schema.json","../src/judge/types.ts","../src/judge/consensus.ts","../src/judge/zones.ts","../src/judge/ensemble.ts"],"sourcesContent":["/**\n * Shared error type. Consumers catch this to distinguish an inference-layer\n * operational failure (missing API key, unknown provider) from their own\n * domain errors, and typically map it to their own exit code.\n */\nexport class InferenceError extends Error {\n  constructor(message: string) {\n    super(message);\n    this.name = \"InferenceError\";\n  }\n}\n","/**\n * The one thing this library says about the runtime it was loaded into.\n *\n * `engines.node` is `>=24`, but npm treats an `engines` mismatch as an\n * `EBADENGINE` **warning**, not a refusal — scrolled past in CI, invisible in a\n * transitive install. So an older Node installs cleanly, runs, and then fails\n * somewhere unrelated with an error that never mentions the Node version.\n *\n * This is a warning rather than a throw on purpose. Nothing in `src/` uses a\n * Node-24-only API — the imports are `node:crypto`, `node:fs`, `node:os` and\n * `node:path`, the newest globals are `fetch` (18+) and `structuredClone`\n * (17+), and `tsconfig` targets ES2022 — and the strictest dependency floor is\n * `node-llama-cpp` at `>=20`. Node 24 is this package's *support* policy, not a\n * technical impossibility, and a library has no business refusing to run on a\n * consumer's behalf when it can still do the work. It says so once and\n * continues, the same way the four existing warn-once paths do.\n */\n\n/**\n * The major version `engines.node` declares. `test/unit/runtime.test.ts` pins\n * this against `package.json`, so the two cannot drift apart.\n */\nexport const MINIMUM_NODE_MAJOR = 24;\n\nlet warnedNodeVersion = false;\n\n/**\n * Warn once if the running Node is older than the declared minimum.\n *\n * The version is a parameter so this is testable in-process — no subprocess on\n * a second Node install just to see the string.\n */\nexport function warnIfUnsupportedNode(\n  version: string = process.versions.node,\n): void {\n  if (warnedNodeVersion) return;\n  const major = Number.parseInt(version, 10);\n  // An unreadable version is not evidence of an old one. Same rule as an\n  // unknown model price: never guess.\n  if (!Number.isInteger(major) || major >= MINIMUM_NODE_MAJOR) return;\n  warnedNodeVersion = true;\n  console.warn(\n    `inference: running on Node ${version}, older than the Node ` +\n      `${MINIMUM_NODE_MAJOR} this package requires. npm only warns about that ` +\n      `at install time (EBADENGINE), so nothing has stopped you yet — upgrade ` +\n      `Node, or expect failures this library cannot explain.`,\n  );\n}\n\n/** Test seam: reset the once-per-process Node version warning. */\nexport function resetNodeVersionWarning(): void {\n  warnedNodeVersion = false;\n}\n","/**\n * Anthropic provider: structured output via a single forced tool call whose\n * input schema is the caller's schema — the model cannot answer any other way.\n */\nimport Anthropic from \"@anthropic-ai/sdk\";\nimport { InferenceError } from \"../types.js\";\nimport type {\n  CompleteJSONRequest,\n  CompleteJSONResponse,\n  InferenceProvider,\n} from \"./types.js\";\n\nconst DEFAULT_TOOL_NAME = \"record_result\";\n\nexport interface AnthropicProviderOptions {\n  /**\n   * Name of the forced tool. Purely cosmetic to the model, but a descriptive\n   * name (\"record_verdict\", \"record_proposal\") measurably steers output, so\n   * consumers may set their own.\n   */\n  toolName?: string;\n  /** Tool description shown to the model. */\n  toolDescription?: string;\n  maxTokens?: number;\n}\n\nexport class AnthropicProvider implements InferenceProvider {\n  private readonly client: Anthropic;\n  private readonly toolName: string;\n  private readonly toolDescription: string;\n  private readonly maxTokens: number;\n\n  constructor(\n    private readonly model: string,\n    apiKeyEnv: string,\n    options: AnthropicProviderOptions = {},\n  ) {\n    const apiKey = process.env[apiKeyEnv];\n    if (!apiKey) {\n      throw new InferenceError(\n        `Anthropic provider needs ${apiKeyEnv} set (or choose another provider)`,\n      );\n    }\n    this.client = new Anthropic({ apiKey });\n    this.toolName = options.toolName ?? DEFAULT_TOOL_NAME;\n    this.toolDescription =\n      options.toolDescription ?? \"Record the structured result.\";\n    this.maxTokens = options.maxTokens ?? 1024;\n  }\n\n  provider(): string {\n    return \"anthropic\";\n  }\n\n  modelName(): string {\n    return this.model;\n  }\n\n  async completeJSON(req: CompleteJSONRequest): Promise<CompleteJSONResponse> {\n    const response = await this.client.messages.create({\n      model: this.model,\n      max_tokens: this.maxTokens,\n      temperature: req.temperature,\n      system: req.system,\n      messages: [{ role: \"user\", content: req.user }],\n      tools: [\n        {\n          name: this.toolName,\n          description: this.toolDescription,\n          input_schema: req.schema as Anthropic.Tool[\"input_schema\"],\n        },\n      ],\n      tool_choice: { type: \"tool\", name: this.toolName },\n    });\n\n    // A truncated tool call still arrives as a tool_use block, just with\n    // partial input. Without this check it fails schema validation instead,\n    // burning the retry and reporting a misleading validation error rather\n    // than the actionable \"raise maxTokens\".\n    if (response.stop_reason === \"max_tokens\") {\n      throw new Error(\n        `Anthropic response hit max_tokens (${this.maxTokens}) before completing the tool call — raise the anthropic.maxTokens option.`,\n      );\n    }\n\n    const toolUse = response.content.find(\n      (block): block is Anthropic.ToolUseBlock => block.type === \"tool_use\",\n    );\n    if (!toolUse) {\n      throw new Error(\"Anthropic response contained no tool_use block\");\n    }\n    return {\n      json: toolUse.input,\n      usage: {\n        inputTokens: response.usage.input_tokens,\n        outputTokens: response.usage.output_tokens,\n      },\n    };\n  }\n}\n","/**\n * OpenAI-compatible provider: works against OpenAI, Azure, Ollama, Groq,\n * Together, or any server speaking /chat/completions. Prefers strict\n * json_schema response_format; falls back to json_object + schema-in-prompt\n * when the server rejects it (older Ollama, some proxies).\n */\nimport { InferenceError } from \"../types.js\";\nimport type {\n  CompleteJSONRequest,\n  CompleteJSONResponse,\n  InferenceProvider,\n} from \"./types.js\";\n\ninterface ChatResponse {\n  choices?: { message?: { content?: string } }[];\n  usage?: { prompt_tokens?: number; completion_tokens?: number };\n  error?: { message?: string };\n}\n\n/** Extract a JSON object from content that may carry markdown fences. */\nexport function extractJson(content: string): unknown {\n  const trimmed = content\n    .replace(/^\\s*```(?:json)?\\s*/i, \"\")\n    .replace(/\\s*```\\s*$/, \"\")\n    .trim();\n  try {\n    return JSON.parse(trimmed);\n  } catch {\n    const start = trimmed.indexOf(\"{\");\n    const end = trimmed.lastIndexOf(\"}\");\n    if (start >= 0 && end > start) {\n      return JSON.parse(trimmed.slice(start, end + 1));\n    }\n    throw new Error(\"Response contained no parseable JSON object\");\n  }\n}\n\n/**\n * OpenAI strict mode requires `required` to list EVERY property (optionality\n * is expressed as a `null` type union) and rejects keywords outside its\n * subset (minLength, uniqueItems). Transform an all-optional schema into a\n * strict-compatible equivalent; null values are stripped from the response.\n */\nexport function toStrictSchema(\n  schema: Record<string, unknown>,\n): Record<string, unknown> {\n  const clone = structuredClone(schema);\n  const walk = (node: unknown): void => {\n    if (node === null || typeof node !== \"object\") return;\n    if (Array.isArray(node)) {\n      for (const item of node) walk(item);\n      return;\n    }\n    const obj = node as Record<string, unknown>;\n    delete obj[\"minLength\"];\n    delete obj[\"uniqueItems\"];\n    const properties = obj[\"properties\"];\n    if (properties && typeof properties === \"object\") {\n      obj[\"required\"] = Object.keys(properties);\n      // Strict mode requires additionalProperties:false on EVERY object, not\n      // just the root. A nested object without it is rejected as a schema\n      // error, which permanently downgrades this provider to the weaker\n      // json_object fallback for the rest of its life.\n      obj[\"additionalProperties\"] = false;\n      for (const prop of Object.values(properties as Record<string, unknown>)) {\n        walk(prop);\n        if (prop && typeof prop === \"object\" && !Array.isArray(prop)) {\n          const p = prop as Record<string, unknown>;\n          if (typeof p[\"type\"] === \"string\" && p[\"type\"] !== \"null\") {\n            p[\"type\"] = [p[\"type\"], \"null\"];\n          }\n        }\n      }\n    }\n    walk(obj[\"items\"]);\n  };\n  walk(clone);\n  return clone;\n}\n\n/** Remove null-valued keys (strict-mode \"omitted\" marker) from a response object. */\nexport function stripNulls(value: unknown): unknown {\n  if (value === null || typeof value !== \"object\" || Array.isArray(value)) {\n    return value;\n  }\n  return Object.fromEntries(\n    Object.entries(value as Record<string, unknown>).filter(\n      ([, v]) => v !== null,\n    ),\n  );\n}\n\nexport interface OpenAICompatProviderOptions {\n  /** `json_schema.name` sent to the server. Cosmetic; steers some models. */\n  schemaName?: string;\n}\n\nexport class OpenAICompatProvider implements InferenceProvider {\n  private supportsJsonSchema = true;\n  private readonly schemaName: string;\n\n  constructor(\n    private readonly baseUrl: string,\n    private readonly model: string,\n    apiKeyEnv: string,\n    private readonly apiKey: string | undefined = process.env[apiKeyEnv],\n    options: OpenAICompatProviderOptions = {},\n  ) {\n    // Local servers (Ollama) often need no key; only insist for api.openai.com.\n    if (!this.apiKey && baseUrl.includes(\"api.openai.com\")) {\n      throw new InferenceError(\n        `OpenAI provider needs ${apiKeyEnv} set (or point baseUrl at a local server)`,\n      );\n    }\n    this.schemaName = options.schemaName ?? \"result\";\n  }\n\n  provider(): string {\n    return \"openai\";\n  }\n\n  modelName(): string {\n    return this.model;\n  }\n\n  private async chat(body: Record<string, unknown>): Promise<ChatResponse> {\n    const response = await fetch(\n      `${this.baseUrl.replace(/\\/$/, \"\")}/chat/completions`,\n      {\n        method: \"POST\",\n        headers: {\n          \"content-type\": \"application/json\",\n          ...(this.apiKey ? { authorization: `Bearer ${this.apiKey}` } : {}),\n        },\n        body: JSON.stringify(body),\n      },\n    );\n    const json = (await response.json().catch(() => ({}))) as ChatResponse;\n    if (!response.ok) {\n      const message = json.error?.message ?? `HTTP ${response.status}`;\n      throw new Error(`${message}`);\n    }\n    return json;\n  }\n\n  async completeJSON(req: CompleteJSONRequest): Promise<CompleteJSONResponse> {\n    const base = {\n      model: this.model,\n      temperature: req.temperature,\n      messages: [\n        { role: \"system\", content: req.system },\n        { role: \"user\", content: req.user },\n      ],\n    };\n\n    let response: ChatResponse;\n    if (this.supportsJsonSchema) {\n      try {\n        response = await this.chat({\n          ...base,\n          response_format: {\n            type: \"json_schema\",\n            json_schema: {\n              name: this.schemaName,\n              strict: true,\n              schema: toStrictSchema(req.schema),\n            },\n          },\n        });\n      } catch (e) {\n        const message = e instanceof Error ? e.message : String(e);\n        // Fall back on schema/format complaints AND on opaque 400s (gateways\n        // that reject response_format without a parseable error body).\n        if (\n          !/response_format|json_schema|schema/i.test(message) &&\n          message !== \"HTTP 400\"\n        ) {\n          throw e;\n        }\n        this.supportsJsonSchema = false;\n        response = await this.jsonObjectFallback(base, req);\n      }\n    } else {\n      response = await this.jsonObjectFallback(base, req);\n    }\n\n    const content = response.choices?.[0]?.message?.content;\n    if (!content) throw new Error(\"Empty completion response\");\n    return {\n      json: stripNulls(extractJson(content)),\n      usage:\n        response.usage?.prompt_tokens != null\n          ? {\n              inputTokens: response.usage.prompt_tokens ?? 0,\n              outputTokens: response.usage.completion_tokens ?? 0,\n            }\n          : undefined,\n    };\n  }\n\n  private jsonObjectFallback(\n    base: Record<string, unknown>,\n    req: CompleteJSONRequest,\n  ): Promise<ChatResponse> {\n    return this.chat({\n      ...base,\n      messages: [\n        {\n          role: \"system\",\n          content: `${req.system}\\n\\nRespond with ONLY a JSON object conforming to this JSON Schema:\\n${JSON.stringify(req.schema)}`,\n        },\n        { role: \"user\", content: req.user },\n      ],\n      response_format: { type: \"json_object\" },\n    });\n  }\n}\n","/**\n * Process execution wrapper for subprocess-backed providers. Uses cross-spawn\n * so npm shims resolve on Windows without `shell: true` and its quoting\n * hazards. Large payloads go through `input` (piped stdin) rather than argv —\n * Windows caps the command line at ~32K characters.\n */\nimport spawn from \"cross-spawn\";\nimport type { ExecFn, ExecResult } from \"./providers/types.js\";\n\nexport const realExec: ExecFn = (cmd, opts = {}) => {\n  const [bin, ...args] = cmd;\n  if (!bin) {\n    return Promise.resolve<ExecResult>({\n      code: null,\n      stdout: \"\",\n      stderr: \"\",\n      timedOut: false,\n      spawnError: \"Empty command\",\n    });\n  }\n  return new Promise<ExecResult>((resolvePromise) => {\n    const child = spawn(bin, args, {\n      cwd: opts.cwd,\n      env: { ...process.env, ...(opts.env ?? {}) },\n      stdio: [opts.input != null ? \"pipe\" : \"ignore\", \"pipe\", \"pipe\"],\n    });\n\n    let stdout = \"\";\n    let stderr = \"\";\n    let timedOut = false;\n    let settled = false;\n\n    const timeoutMs = opts.timeoutMs ?? 60000;\n    const timer = setTimeout(() => {\n      timedOut = true;\n      child.kill();\n      // Settle on the timeout itself rather than waiting for 'close'. On POSIX\n      // a child that handles SIGTERM survives kill(), so 'close' would never\n      // fire and the caller would hang forever with no timeout error. (On\n      // Windows kill() is forceful, so this is belt-and-braces there.)\n      settle({ code: null, stdout, stderr, timedOut: true });\n    }, timeoutMs);\n\n    const settle = (result: ExecResult): void => {\n      if (settled) return;\n      settled = true;\n      clearTimeout(timer);\n      resolvePromise(result);\n    };\n\n    if (opts.input != null && child.stdin) {\n      // EPIPE from a child that exits before reading is not our failure.\n      child.stdin.on(\"error\", () => {});\n      child.stdin.end(opts.input);\n    }\n\n    // setEncoding routes chunks through a StringDecoder, so multi-byte UTF-8\n    // characters straddling pipe-chunk boundaries decode correctly; a raw\n    // per-chunk Buffer.toString() would corrupt them nondeterministically.\n    child.stdout?.setEncoding(\"utf8\");\n    child.stderr?.setEncoding(\"utf8\");\n    child.stdout?.on(\"data\", (d: string) => (stdout += d));\n    child.stderr?.on(\"data\", (d: string) => (stderr += d));\n    child.on(\"error\", (e) =>\n      settle({ code: null, stdout, stderr, timedOut, spawnError: e.message }),\n    );\n    child.on(\"close\", (code) => settle({ code, stdout, stderr, timedOut }));\n  });\n};\n","/**\n * Claude CLI provider: shells out to the `claude` CLI using its local\n * authentication — no API key needed. Token usage is not reported, so cost\n * shows as unknown.\n */\nimport { realExec } from \"../exec.js\";\nimport { extractJson } from \"./openai-compat.js\";\nimport type {\n  CompleteJSONRequest,\n  CompleteJSONResponse,\n  ExecFn,\n  InferenceProvider,\n} from \"./types.js\";\n\nexport class ClaudeCliProvider implements InferenceProvider {\n  constructor(\n    private readonly model: string,\n    private readonly command: string = \"claude\",\n    private readonly exec: ExecFn = realExec,\n    private readonly timeoutMs: number = 180000,\n  ) {}\n\n  provider(): string {\n    return \"claude-cli\";\n  }\n\n  modelName(): string {\n    return this.model;\n  }\n\n  async completeJSON(req: CompleteJSONRequest): Promise<CompleteJSONResponse> {\n    const prompt = [\n      req.user,\n      \"\",\n      \"Respond with ONLY a JSON object conforming to this JSON Schema — no\",\n      \"prose, no markdown fences:\",\n      JSON.stringify(req.schema),\n    ].join(\"\\n\");\n\n    // The prompt is piped via stdin: user content routinely exceeds the ~32K\n    // Windows command-line limit when passed as an argument.\n    const result = await this.exec(\n      [\n        this.command,\n        \"-p\",\n        \"--append-system-prompt\",\n        req.system,\n        \"--output-format\",\n        \"json\",\n        \"--model\",\n        this.model,\n      ],\n      { timeoutMs: this.timeoutMs, input: prompt },\n    );\n\n    if (result.spawnError) {\n      throw new Error(\n        `Failed to run ${this.command}: ${result.spawnError} (is the Claude CLI installed?)`,\n      );\n    }\n    if (result.timedOut) throw new Error(\"Claude CLI timed out\");\n    if (result.code !== 0) {\n      throw new Error(\n        `Claude CLI exited ${result.code}: ${result.stderr.trim().slice(-300)}`,\n      );\n    }\n\n    // --output-format json wraps the answer: { result: \"...\", ... }\n    //\n    // When it isn't JSON at all the CLI is talking to a human, not to us — an\n    // update banner, a login prompt, a proxy interception page. A bare\n    // `SyntaxError: Unexpected token 'W'` would reach the end user of a\n    // consuming CLI naming neither the culprit nor the fix, so say who printed\n    // it and quote enough for a login prompt to be recognisable on sight.\n    // Collapsed to one line and capped, like the stderr tail above, so a\n    // megabyte of HTML cannot become the error message.\n    let wrapper: { result?: string };\n    try {\n      wrapper = JSON.parse(result.stdout) as { result?: string };\n    } catch {\n      const excerpt = result.stdout.trim().replace(/\\s+/g, \" \").slice(0, 200);\n      throw new Error(\n        `Claude CLI printed non-JSON output (is it logged in?): ${excerpt || \"(no output)\"}`,\n      );\n    }\n    if (typeof wrapper.result !== \"string\") {\n      throw new Error(\"Claude CLI returned no result field\");\n    }\n    return { json: extractJson(wrapper.result) };\n  }\n}\n","/**\n * Mock provider for tests and offline development. Responds with scripted\n * results in order, cycling when exhausted. Exported from the public API so\n * downstream consumers can test their own pipelines without a live provider.\n */\nimport type {\n  CompleteJSONRequest,\n  CompleteJSONResponse,\n  InferenceProvider,\n  TokenUsage,\n} from \"./types.js\";\n\nexport type MockResponse =\n  | { json: unknown; usage?: TokenUsage }\n  | { error: string };\n\nexport class MockProvider implements InferenceProvider {\n  private calls = 0;\n  /** Every request seen, in order — assert against this in tests. */\n  public readonly requests: CompleteJSONRequest[] = [];\n\n  constructor(\n    private readonly responses: MockResponse[],\n    private readonly model = \"mock-model\",\n  ) {\n    if (responses.length === 0) {\n      throw new Error(\"MockProvider needs at least one scripted response\");\n    }\n  }\n\n  provider(): string {\n    return \"mock\";\n  }\n\n  modelName(): string {\n    return this.model;\n  }\n\n  completeJSON(req: CompleteJSONRequest): Promise<CompleteJSONResponse> {\n    this.requests.push(req);\n    const response = this.responses[this.calls % this.responses.length]!;\n    this.calls += 1;\n    if (\"error\" in response) {\n      return Promise.reject(new Error(response.error));\n    }\n    return Promise.resolve({\n      json: response.json,\n      usage: response.usage ?? { inputTokens: 500, outputTokens: 100 },\n    });\n  }\n}\n\n/** Convenience: a scripted response shaped like the canonical judge verdict. */\nexport function mockVerdict(\n  match: \"pass\" | \"fail\" | \"partial\",\n  confidence: number,\n  overrides: Partial<{\n    claim: string;\n    observed: string;\n    reasoning: string;\n  }> = {},\n): { json: unknown } {\n  return {\n    json: {\n      claim: overrides.claim ?? \"The assertion under test\",\n      observed: overrides.observed ?? \"Observed content\",\n      match,\n      confidence,\n      reasoning: overrides.reasoning ?? \"Scripted mock reasoning\",\n    },\n  };\n}\n","/**\n * Content-addressed JSON cache for inference results, so a cached run replays\n * identically instead of re-billing a nondeterministic call.\n *\n * Key composition stays with the caller: each consumer has a different notion\n * of what should invalidate an entry (page body, prompt version, ensemble\n * size, requested fields). `buildCacheKey` just hashes the parts you name.\n */\nimport { createHash } from \"node:crypto\";\nimport { existsSync, mkdirSync, readFileSync, writeFileSync } from \"node:fs\";\nimport { join } from \"node:path\";\n\nexport function sha256(text: string): string {\n  return createHash(\"sha256\").update(text, \"utf8\").digest(\"hex\");\n}\n\n/**\n * Hash an ordered list of key parts into a cache key. Long parts (page bodies,\n * rendered traces) should be pre-hashed with `sha256` by the caller so the\n * joined string stays small.\n */\nexport function buildCacheKey(parts: string[]): string {\n  // Length-prefixed, not plain-joined: a part that itself contains the\n  // separator would otherwise let two different compositions — [\"a|b\", \"c\"]\n  // and [\"a\", \"b|c\"] — share one cached result.\n  return sha256(parts.map((p) => `${p.length}:${p}`).join(\"|\"));\n}\n\nexport class JsonCache<T> {\n  /** Cache-write failures warn once per process, not once per entry. */\n  private warned = false;\n\n  constructor(\n    private readonly dir: string,\n    private readonly enabled: boolean = true,\n    /** Prefix for the one-time write-failure warning. */\n    private readonly label: string = \"inference\",\n  ) {}\n\n  get(key: string): T | undefined {\n    if (!this.enabled) return undefined;\n    const path = join(this.dir, `${key}.json`);\n    if (!existsSync(path)) return undefined;\n    try {\n      return JSON.parse(readFileSync(path, \"utf8\")) as T;\n    } catch {\n      return undefined; // Corrupt cache entry — treat as a miss.\n    }\n  }\n\n  set(key: string, value: T): void {\n    if (!this.enabled) return;\n    // The cache is an optimization: a write failure (read-only workspace, full\n    // disk, long path) must never abort a run whose inference already\n    // succeeded and was already paid for.\n    try {\n      mkdirSync(this.dir, { recursive: true });\n      writeFileSync(\n        join(this.dir, `${key}.json`),\n        JSON.stringify(value, null, 2),\n      );\n    } catch (e) {\n      if (!this.warned) {\n        this.warned = true;\n        console.warn(\n          `${this.label}: could not write the cache at ${this.dir} (${\n            e instanceof Error ? e.message : String(e)\n          }). Continuing without caching.`,\n        );\n      }\n    }\n  }\n}\n","/**\n * Curated GGUF models for the in-process `llama-cpp` provider, plus the\n * selector-to-model resolution that sits in front of them.\n *\n * The three tiers were chosen by measurement, not by published benchmark:\n * 26 real documentation pages with their `title` and `description` removed,\n * refilled through this library under a JSON Schema, and scored against the\n * human-written originals. See ADR 01009 for the numbers and the method.\n *\n * Two results from that run shape this catalog:\n *\n * 1. **Quality barely tracks size on schema-constrained extraction.** Every\n *    model between 1.4 GB and 5.5 GB scored within one standard error of the\n *    others. The tiers therefore buy latency and headroom for richer schemas\n *    than that corpus exercises — they do not buy a linear quality gain, and\n *    should not be described as if they do.\n * 2. **Latency stability does not track size either, and matters more.** Two\n *    builds ran away on long pages, generating until they hit a timeout rather\n *    than closing the JSON: Gemma 4 E2B at Q2 (6 of 12 pages unfinished at\n *    120s, one still going at 400s) and Granite 4.1 8B at Q4 (6 of 26 pages\n *    over 20s, three over 115s). Both are excluded from the tiers for that\n *    reason alone. A tier entry has to terminate.\n *\n * That is why the tiers are no longer one family. Chat templates now differ\n * across tiers, which is fine — node-llama-cpp reads the template out of each\n * GGUF — but it does mean a prompt tuned against one tier is not automatically\n * tuned against the next.\n *\n * The superseded Gemma aliases are kept, untiered, because consumers pin them\n * by name; dropping an alias is a hard failure rather than a slower download.\n *\n * Entries pin an exact blob path rather than a `:QUANT` tag. That is a hard\n * requirement, not a style preference, and it holds for every entry:\n *\n * - A tag can be re-pointed upstream, which silently changes the weights\n *   behind a cache key that already names the model. A pinned path cannot.\n *   This reason alone is sufficient, and it applies to all repos.\n * - Most of these repos carry `mmproj-*.gguf` (the ~1 GB vision projector)\n *   and some carry `mtp-*.gguf` beside the weights, which text-only judging\n *   must not download.\n * - On the Gemma QAT repos a tag would not resolve at all: they ship NO\n *   Q4_K_M, only UD-Q4_K_XL and UD-Q2_K_XL.\n *\n * Note the last point is specific to the Gemma QAT repos, which now back only\n * untiered entries. The Qwen3.5 and Granite repos behind the three tiers DO\n * publish Q4_K_M, so a tag would resolve there — and would still be wrong, for\n * the first reason. Do not read \"the tag resolves\" as \"the tag is allowed\".\n */\nimport { readdirSync } from \"node:fs\";\nimport { homedir } from \"node:os\";\nimport { join } from \"node:path\";\nimport { InferenceError } from \"../types.js\";\n\n/**\n * Where this library downloads weights — its OWN directory, not\n * node-llama-cpp's global `~/.node-llama-cpp/models`.\n *\n * That default is shared: node-llama-cpp's CLI writes there, as does anything\n * else on the machine using it. Owning a directory outright means clearing it\n * can never destroy a model this library did not download, and it keeps one\n * copy shared across every consumer of this package on the machine.\n *\n * `INFERENCE_MODELS_DIR` overrides it — useful for CI or a shared volume.\n */\nexport function defaultLlamaModelsDirectory(): string {\n  return (\n    process.env[\"INFERENCE_MODELS_DIR\"] ||\n    join(homedir(), \".hawkeyexl-inference\", \"models\")\n  );\n}\n\n/** Size tiers, smallest first. Order is load-bearing for `tierForBudget`. */\nexport const LLAMA_TIERS = [\"fast\", \"balanced\", \"quality\"] as const;\nexport type LlamaTier = (typeof LLAMA_TIERS)[number];\n\n/** Model selectors — resolved against hardware, never used as a cache key. */\nexport const LLAMA_SELECTORS = [\"auto\", ...LLAMA_TIERS] as const;\nexport type LlamaSelector = (typeof LLAMA_SELECTORS)[number];\n\nexport interface LlamaModelEntry {\n  /** `hf:` URI pinned to one blob, handed to `resolveModelFile` as-is. */\n  readonly uri: string;\n  /** Size of that blob in bytes, as reported by the Hugging Face API. */\n  readonly sizeBytes: number;\n  readonly license: string;\n  /** Absent for entries that are selectable by alias but not by tier. */\n  readonly tier?: LlamaTier;\n  /** Human note for `LLAMA_MODELS` readers deciding what to download. */\n  readonly notes: string;\n}\n\n/**\n * Frozen per entry, not just at the top level.\n *\n * A shallow freeze leaves the entries writable, and this catalog is exported\n * for consumers to read: a stray write to `sizeBytes` silently re-points\n * `tierForBudget` process-wide, and a write to `uri` defeats the pinned-blob\n * invariant the whole catalog exists to hold (ADR 01003).\n */\nexport const LLAMA_MODELS: Readonly<Record<string, LlamaModelEntry>> =\n  deepFreezeEntries({\n    \"granite-4.1-3b-q2\": {\n      uri: \"hf:unsloth/granite-4.1-3b-GGUF/granite-4.1-3b-UD-Q2_K_XL.gguf\",\n      sizeBytes: 1_414_548_800,\n      license: \"Apache-2.0\",\n      tier: \"fast\",\n      notes:\n        \"Smallest tier and the quickest measured (4.8s/page). Scores level \" +\n        \"with models three times its size on schema-constrained extraction.\",\n    },\n    \"qwen3.5-4b\": {\n      uri: \"hf:unsloth/Qwen3.5-4B-GGUF/Qwen3.5-4B-UD-Q4_K_XL.gguf\",\n      sizeBytes: 2_912_109_728,\n      license: \"Apache-2.0\",\n      notes:\n        \"The default for most machines. Smaller and faster than the Gemma 4 \" +\n        \"E4B it replaces, at indistinguishable measured quality.\",\n      tier: \"balanced\",\n    },\n    \"qwen3.5-9b\": {\n      uri: \"hf:unsloth/Qwen3.5-9B-GGUF/Qwen3.5-9B-UD-Q4_K_XL.gguf\",\n      sizeBytes: 5_966_095_584,\n      license: \"Apache-2.0\",\n      tier: \"quality\",\n      notes:\n        \"Best measured of everything tried, and still smaller and faster \" +\n        \"than the Gemma 4 12B it replaces. Wants a GPU or plenty of RAM.\",\n    },\n    // --- Superseded, kept resolvable by name. Untiered: nothing selects these\n    // unless a caller asks for one outright.\n    \"gemma-4-e2b\": {\n      uri: \"hf:unsloth/gemma-4-E2B-it-qat-GGUF/gemma-4-E2B-it-qat-UD-Q4_K_XL.gguf\",\n      sizeBytes: 2_620_370_976,\n      license: \"Apache-2.0\",\n      notes: \"IFEval 94.6. Former `fast` tier; sound, but larger than Granite.\",\n    },\n    \"gemma-4-e4b\": {\n      uri: \"hf:unsloth/gemma-4-E4B-it-qat-GGUF/gemma-4-E4B-it-qat-UD-Q4_K_XL.gguf\",\n      sizeBytes: 4_215_695_776,\n      license: \"Apache-2.0\",\n      notes: \"IFEval 96.7. Former `balanced` tier; sound, but larger.\",\n    },\n    \"gemma-4-12b\": {\n      uri: \"hf:unsloth/gemma-4-12B-it-qat-GGUF/gemma-4-12B-it-qat-UD-Q4_K_XL.gguf\",\n      sizeBytes: 6_716_356_800,\n      license: \"Apache-2.0\",\n      notes: \"IFEval 97.2. Former `quality` tier. Dense 12B; wants a GPU.\",\n    },\n    \"gemma-4-26b-a4b\": {\n      uri: \"hf:unsloth/gemma-4-26B-A4B-it-qat-GGUF/gemma-4-26B-A4B-it-qat-UD-Q4_K_XL.gguf\",\n      sizeBytes: 14_249_047_104,\n      license: \"Apache-2.0\",\n      notes:\n        \"MoE: 25.2B total, 3.8B active — infers near E4B speed if it fits in memory.\",\n    },\n    \"gemma-4-e2b-q2\": {\n      uri: \"hf:unsloth/gemma-4-E2B-it-qat-GGUF/gemma-4-E2B-it-qat-UD-Q2_K_XL.gguf\",\n      sizeBytes: 2_186_186_784,\n      license: \"Apache-2.0\",\n      notes:\n        \"AVOID. Smallest download, but it does not reliably terminate: 6 of \" +\n        \"12 pages unfinished at 120s, one still running at 400s, and the \" +\n        \"pages that did finish proposed identifiers as prose. Kept only so \" +\n        \"existing pins still resolve.\",\n    },\n  });\n\nfunction deepFreezeEntries<T extends Record<string, object>>(\n  catalog: T,\n): Readonly<T> {\n  for (const entry of Object.values(catalog)) Object.freeze(entry);\n  return Object.freeze(catalog);\n}\n\n/** The alias backing each tier, used by `auto` and the tier keywords. */\nconst TIER_ALIAS: Record<LlamaTier, string> = {\n  fast: \"granite-4.1-3b-q2\",\n  balanced: \"qwen3.5-4b\",\n  quality: \"qwen3.5-9b\",\n};\n\nexport function isLlamaSelector(model: string): model is LlamaSelector {\n  return (LLAMA_SELECTORS as readonly string[]).includes(model);\n}\n\n/**\n * Weights are only part of the cost — the KV cache at a real context length,\n * the OS, and whatever else the machine is doing all want memory too. Requiring\n * several times the file size keeps `auto` from picking a model that technically\n * loads and then thrashes.\n */\nconst MEMORY_HEADROOM = 3.5;\n\n/**\n * Largest tier whose weights fit the memory budget with headroom. Lands at\n * roughly: >=24 GB -> quality, >=15 GB -> balanced, else fast.\n *\n * Sized off the catalog's recorded bytes rather than parameter counts: Gemma\n * 4's E-series are per-layer-embedding models whose footprint does not track\n * \"effective params\" (E4B is 4.5B effective but 15 GB at BF16).\n */\nexport function tierForBudget(budgetBytes: number): LlamaTier {\n  let chosen: LlamaTier = \"fast\";\n  for (const tier of LLAMA_TIERS) {\n    const entry = LLAMA_MODELS[TIER_ALIAS[tier]]!;\n    if (entry.sizeBytes * MEMORY_HEADROOM <= budgetBytes) chosen = tier;\n  }\n  // Falls back to the smallest tier rather than refusing: a machine too small\n  // for `fast` will thrash, but that is the caller's call to make, not ours.\n  return chosen;\n}\n\n/**\n * Catalog alias backing a tier. Selectors resolve to this rather than to a raw\n * URI so the identity — and therefore the cache key — stays human-readable.\n */\nexport function aliasForTier(tier: LlamaTier): string {\n  return TIER_ALIAS[tier];\n}\n\n/** The pinned URI backing a tier keyword. */\nexport function uriForTier(tier: LlamaTier): string {\n  return LLAMA_MODELS[TIER_ALIAS[tier]]!.uri;\n}\n\n/**\n * Turn a concrete model reference — a curated alias, an `hf:` URI, or a local\n * path — into something `resolveModelFile` accepts.\n *\n * Selectors are rejected rather than guessed at: they need a hardware probe,\n * which is async, and this runs on the synchronous cache-key path. Resolving\n * one here from RAM alone would emit a key naming a model the provider then\n * did not load.\n */\nexport function resolveLlamaModelRef(model: string): string {\n  if (isLlamaSelector(model)) {\n    throw new InferenceError(\n      `llama-cpp model \"${model}\" is a selector and needs a hardware probe to ` +\n        `resolve. Use resolveProviderIdentityAsync/makeProviderAsync, or name a ` +\n        `concrete model (e.g. \"${TIER_ALIAS.balanced}\").`,\n    );\n  }\n  const entry = LLAMA_MODELS[model];\n  if (entry) return entry.uri;\n  if (isModelPathOrUri(model)) return model;\n  throw new InferenceError(\n    `Unknown llama-cpp model \"${model}\". Use a selector (${LLAMA_SELECTORS.join(\n      \", \",\n    )}), a curated alias (${Object.keys(LLAMA_MODELS).join(\n      \", \",\n    )}), an hf: URI, or a path to a .gguf file.`,\n  );\n}\n\n/**\n * The catalog's pinned filename for a model reference.\n *\n * Strips a `#branch` fragment (node-llama-cpp accepts\n * `hf:user/repo/file.gguf#branch`) and splits on both separators, so a Windows\n * path resolves too. Getting either wrong makes callers silently match nothing.\n */\nexport function blobNameFor(model: string): string {\n  const ref = resolveLlamaModelRef(model).split(\"#\")[0]!;\n  return ref.split(/[/\\\\]/).pop()!;\n}\n\n/**\n * Does an on-disk entry belong to this model?\n *\n * node-llama-cpp prefixes downloads with `hf_<user>_`, so match by SUFFIX\n * rather than equality — that survives a change to their naming scheme.\n */\nexport function matchesModelBlob(entry: string, blobName: string): boolean {\n  const base = entry.replace(/\\.ipull$/, \"\");\n  if (base.endsWith(blobName)) return true;\n  // Split models land as `<stem>-00001-of-00003.gguf`; every part belongs to\n  // the same model, so matching the stem covers the whole set.\n  const stem = blobName.replace(/\\.gguf$/, \"\");\n  return new RegExp(`${escapeRegExp(stem)}-\\\\d{5}-of-\\\\d{5}\\\\.gguf$`).test(base);\n}\n\nfunction escapeRegExp(text: string): string {\n  return text.replace(/[.*+?^${}()|[\\]\\\\]/g, \"\\\\$&\");\n}\n\n/**\n * Are this model's weights already on disk?\n *\n * A `.ipull` partial counts as NOT downloaded: it cannot be loaded, so\n * treating it as present would skip the download warning and then stall on a\n * download anyway.\n */\nexport function isModelDownloaded(model: string, directory: string): boolean {\n  const blobName = blobNameFor(model);\n  return listModelDirectory(directory).some(\n    (entry) => !entry.endsWith(\".ipull\") && matchesModelBlob(entry, blobName),\n  );\n}\n\n/** Top level only — a nested directory is not ours to walk. */\nexport function listModelDirectory(directory: string): string[] {\n  try {\n    return readdirSync(directory, { withFileTypes: true })\n      .filter((entry) => entry.isFile())\n      .map((entry) => entry.name);\n  } catch {\n    // No directory means nothing was ever downloaded — not an error.\n    return [];\n  }\n}\n\n/**\n * A bare unknown word is a typo'd alias, not a model — catch it early rather\n * than letting it reach the downloader as a doomed repo name.\n *\n * Accepts exactly what the error message in `resolveLlamaModelRef` promises: a\n * recognised URI scheme, or something that names a `.gguf` file. A bare\n * `user/repo` is deliberately NOT a model reference — it would otherwise slip\n * past this guard and fail deep inside the downloader with a far worse\n * message. Any real path to weights ends in `.gguf`, on every platform.\n */\nfunction isModelPathOrUri(model: string): boolean {\n  return (\n    /^(hf|huggingface):/i.test(model) ||\n    /^https?:\\/\\//i.test(model) ||\n    /^(hf|huggingface)\\.co\\//i.test(model) ||\n    model.endsWith(\".gguf\")\n  );\n}\n","/**\n * Installing the optional `node-llama-cpp` peer on demand.\n *\n * Detection ends at `llama-cpp` precisely because it needs no key and no\n * account — but that rung is only reachable if the native binding is present,\n * and npm does not install optional peers. Without this module a machine with\n * no credentials and no binding falls off the end of the chain, which is the\n * one case auto-detection exists to serve.\n *\n * The install goes into a directory this library OWNS — never the consumer's\n * `node_modules`, `package.json` or lockfile. That is the same reasoning ADR\n * 01003 applied to weights: owning a directory removes the hazard instead of\n * defending against it. A consumer's dependency manifest is theirs, and a\n * library that edits it has broken reproducible installs for everyone\n * downstream.\n */\nimport {\n  existsSync,\n  mkdirSync,\n  rmSync,\n  statSync,\n  writeFileSync,\n} from \"node:fs\";\nimport { homedir } from \"node:os\";\nimport { join } from \"node:path\";\nimport { pathToFileURL } from \"node:url\";\nimport { InferenceError } from \"../types.js\";\nimport { realExec } from \"../exec.js\";\nimport type { ExecFn } from \"./types.js\";\n\n/** Pinned to the peer range: below 3.19.0 nothing in the catalog loads. */\nconst PACKAGE_SPEC = \"node-llama-cpp@^3.19.0\";\n\n/**\n * The shim doubles as the readiness marker — it is written only after npm\n * exits 0, so its presence means \"this prefix is complete\".\n */\nconst SHIM = \"loader.mjs\";\nconst LOCK = \".install.lock\";\n\n/**\n * Generous by necessity. Prebuilt binaries cover win32-x64, darwin-arm64 and\n * linux-x64; anything else falls back to a CMake build that genuinely takes\n * minutes. `realExec` defaults to 60s, which would kill it half-built.\n */\nconst INSTALL_TIMEOUT_MS = 900_000;\n\n/** How long to wait for another process's install before giving up. */\nconst LOCK_WAIT_MS = INSTALL_TIMEOUT_MS;\n/** A lock older than this belonged to a process that died holding it. */\nconst LOCK_STALE_MS = INSTALL_TIMEOUT_MS + 60_000;\n\nexport interface RuntimeInstallOptions {\n  /** Defaults to this library's own runtime directory. */\n  directory?: string;\n  /** Injected for tests; defaults to the real process runner. */\n  exec?: ExecFn;\n  /** Injected for tests; defaults to `process.env`. */\n  env?: Record<string, string | undefined>;\n  timeoutMs?: number;\n  /** Test seam: how the shim is imported once it exists. */\n  importShim?: (url: string) => Promise<unknown>;\n  /**\n   * Test seam: how the consumer's own copy is looked for.\n   *\n   * Whether an optional peer is installed is a property of the machine, and the\n   * behaviour that matters most here — that probing installs nothing — is\n   * unobservable on a machine that already has it. Injecting the lookup is the\n   * only way to assert it deterministically.\n   */\n  probeImport?: () => Promise<unknown>;\n}\n\n/**\n * Where this library installs the binding — its OWN directory, beside the\n * models directory and for the same reason.\n *\n * `INFERENCE_RUNTIME_DIR` overrides it, mirroring `INFERENCE_MODELS_DIR`.\n */\nexport function defaultLlamaRuntimeDirectory(\n  env: Record<string, string | undefined> = process.env,\n): string {\n  return (\n    env[\"INFERENCE_RUNTIME_DIR\"] ||\n    join(homedir(), \".hawkeyexl-inference\", \"runtime\")\n  );\n}\n\n/**\n * Is this import failure \"the package is not here\", as opposed to \"the package\n * is here and broken\"?\n *\n * The distinction decides whether installing can possibly help. `ERR_MODULE_NOT_FOUND`\n * is what Node raises for a missing bare specifier; `MODULE_NOT_FOUND` is its\n * CJS spelling. Anything else — a failed `dlopen`, an unsupported Node, a\n * package whose `exports` do not match — means the package resolved and then\n * failed, and no amount of reinstalling changes that.\n */\nexport function isModuleNotFound(e: unknown): boolean {\n  const code = (e as NodeJS.ErrnoException | undefined)?.code;\n  return code === \"ERR_MODULE_NOT_FOUND\" || code === \"MODULE_NOT_FOUND\";\n}\n\nfunction describe(e: unknown): string {\n  return e instanceof Error ? e.message : String(e);\n}\n\n/** Whether the binding can be used, and at what cost, WITHOUT installing it. */\nexport type RuntimeStatus =\n  /** Importable right now — either the consumer's own copy or a filled prefix. */\n  | { state: \"present\" }\n  /** Absent, but an install is permitted. Using it will fetch a native module. */\n  | { state: \"installable\"; directory: string }\n  /** Absent and installing is refused, so this provider cannot be used. */\n  | { state: \"refused\"; reason: string };\n\n/**\n * Can the local runtime be used, and would using it install anything?\n *\n * Detection needs this rather than simply calling `getMemoryBudgetBytes`:\n * that goes through `importNodeLlamaCpp` and would install, which turns\n * `availableProviders()` — a function whose entire job is to REPORT what is\n * usable — into one that changes what is usable. Probing must not have the side\n * effect it is probing for.\n */\nexport async function nodeLlamaCppStatus(\n  options: RuntimeInstallOptions = {},\n): Promise<RuntimeStatus> {\n  const env = options.env ?? process.env;\n  const directory = options.directory ?? defaultLlamaRuntimeDirectory(env);\n\n  const probeImport =\n    options.probeImport ?? ((): Promise<unknown> => import(\"node-llama-cpp\"));\n  try {\n    await probeImport();\n    return { state: \"present\" };\n  } catch (e) {\n    // Only a genuine \"no such package\" means absent. An ABI mismatch, a missing\n    // system library, or an unsupported Node all fail here too, and installing\n    // over them would fetch the same broken package again while burying the\n    // real cause under a download.\n    if (!isModuleNotFound(e)) {\n      return {\n        state: \"refused\",\n        reason: `node-llama-cpp is installed but failed to load (${describe(e)})`,\n      };\n    }\n    // Genuinely absent — fall through to our own prefix.\n  }\n  if (existsSync(join(directory, SHIM))) return { state: \"present\" };\n\n  if ((env[\"INFERENCE_NO_AUTO_INSTALL\"] ?? \"\") !== \"\") {\n    return {\n      state: \"refused\",\n      reason: `node-llama-cpp is not installed and INFERENCE_NO_AUTO_INSTALL is set`,\n    };\n  }\n  return { state: \"installable\", directory };\n}\n\n/**\n * In-flight installs, keyed by directory. docmeta and the ensemble runner both\n * work several items at once, so without this every worker that misses the\n * binding would spawn its own npm against one prefix.\n */\nconst installs = new Map<string, Promise<unknown>>();\n\nlet warnedInstall = false;\n\n/** Test seam: forget in-flight installs and re-arm the one-time warning. */\nexport function resetRuntimeInstall(): void {\n  installs.clear();\n  warnedInstall = false;\n}\n\n/**\n * Import `node-llama-cpp` from the library-owned prefix, installing it first if\n * it is not there.\n *\n * Callers reach this only after a plain `import(\"node-llama-cpp\")` has already\n * failed — a consumer who installed the peer themselves never gets here.\n */\nexport function importNodeLlamaCpp(\n  options: RuntimeInstallOptions = {},\n): Promise<unknown> {\n  const env = options.env ?? process.env;\n  const directory = options.directory ?? defaultLlamaRuntimeDirectory(env);\n\n  const existing = installs.get(directory);\n  if (existing) return existing;\n\n  const pending = fromPrefix(directory, env, options);\n  // Drop a failed attempt so the next call retries: a download killed by a\n  // flaky network must not poison the runtime for the rest of the process —\n  // the same rule `load()` applies to weights.\n  const guarded = pending.catch((e: unknown) => {\n    if (installs.get(directory) === guarded) installs.delete(directory);\n    throw e;\n  });\n  installs.set(directory, guarded);\n  return guarded;\n}\n\nasync function fromPrefix(\n  directory: string,\n  env: Record<string, string | undefined>,\n  options: RuntimeInstallOptions,\n): Promise<unknown> {\n  const importShim =\n    options.importShim ?? ((url: string): Promise<unknown> => import(url));\n  const shim = join(directory, SHIM);\n\n  if (existsSync(shim)) return importShim(pathToFileURL(shim).href);\n\n  if ((env[\"INFERENCE_NO_AUTO_INSTALL\"] ?? \"\") !== \"\") {\n    throw new InferenceError(\n      `The llama-cpp provider needs node-llama-cpp, and INFERENCE_NO_AUTO_INSTALL ` +\n        `is set. Install it yourself (npm i ${PACKAGE_SPEC}), unset ` +\n        `INFERENCE_NO_AUTO_INSTALL to allow installing into ${directory}, or name ` +\n        `a different provider.`,\n    );\n  }\n\n  mkdirSync(directory, { recursive: true });\n  await withLock(directory, async () => {\n    // Another process may have finished while we waited for the lock.\n    if (existsSync(shim)) return;\n    warnInstalling(directory);\n    await runInstall(directory, env, options);\n    // Only now is the prefix complete, so only now does it get its marker.\n    writeFileSync(shim, `export * from \"node-llama-cpp\";\\n`, \"utf8\");\n  });\n\n  return importShim(pathToFileURL(shim).href);\n}\n\nasync function runInstall(\n  directory: string,\n  env: Record<string, string | undefined>,\n  options: RuntimeInstallOptions,\n): Promise<void> {\n  // Anchor npm to this directory. Without a manifest here npm walks UP looking\n  // for one, and would install into whatever project happens to be above us —\n  // the exact mutation this module exists to avoid.\n  const manifest = join(directory, \"package.json\");\n  if (!existsSync(manifest)) {\n    writeFileSync(\n      manifest,\n      `${JSON.stringify(\n        {\n          name: \"hawkeyexl-inference-runtime\",\n          version: \"0.0.0\",\n          private: true,\n          description:\n            \"Auto-installed runtime for @hawkeyexl/inference. Safe to delete.\",\n        },\n        null,\n        2,\n      )}\\n`,\n      \"utf8\",\n    );\n  }\n\n  const exec = options.exec ?? realExec;\n  const result = await exec(\n    [\n      \"npm\",\n      \"install\",\n      \"--prefix\",\n      directory,\n      PACKAGE_SPEC,\n      \"--no-audit\",\n      \"--no-fund\",\n    ],\n    { timeoutMs: options.timeoutMs ?? INSTALL_TIMEOUT_MS, env },\n  );\n\n  if (result.spawnError != null) {\n    throw new InferenceError(\n      `Could not run npm to install node-llama-cpp (${result.spawnError}). ` +\n        `Install it yourself with: npm i ${PACKAGE_SPEC}, or name a provider ` +\n        `that does not need it.`,\n    );\n  }\n  if (result.timedOut) {\n    throw new InferenceError(\n      `Installing node-llama-cpp into ${directory} timed out. A source build ` +\n        `can take a while — retry, raise the timeout, or install it yourself ` +\n        `with: npm i ${PACKAGE_SPEC}.`,\n    );\n  }\n  if (result.code !== 0) {\n    throw new InferenceError(\n      `Installing node-llama-cpp into ${directory} failed (exit ${String(\n        result.code,\n      )}).\\n${tail(result.stderr || result.stdout)}\\n` +\n        `Install it yourself with: npm i ${PACKAGE_SPEC}, or name a provider ` +\n        `that does not need it.`,\n    );\n  }\n}\n\n/** Enough npm output to diagnose the failure, not enough to bury the advice. */\nfunction tail(output: string, lines = 12): string {\n  return output.trimEnd().split(/\\r?\\n/).slice(-lines).join(\"\\n\");\n}\n\n/**\n * Pulling a native module is not free, and the weights that follow are far less\n * free. Say so before it starts, not after a build has already stalled on it.\n */\nfunction warnInstalling(directory: string): void {\n  if (warnedInstall) return;\n  warnedInstall = true;\n  console.warn(\n    `inference: node-llama-cpp is not installed — fetching it into ${directory} ` +\n      `so the local model can run. This is a one-time native install; set ` +\n      `INFERENCE_NO_AUTO_INSTALL=1 to refuse it, or name a provider that does ` +\n      `not need it.`,\n  );\n}\n\n/**\n * Cross-process guard around one prefix.\n *\n * The in-process memo covers a worker pool inside one run; this covers two runs\n * started at once, where two npm processes writing one `node_modules` is how a\n * prefix ends up half-written.\n */\nasync function withLock(\n  directory: string,\n  fn: () => Promise<void>,\n): Promise<void> {\n  const lock = join(directory, LOCK);\n  const deadline = Date.now() + LOCK_WAIT_MS;\n\n  for (;;) {\n    try {\n      writeFileSync(lock, String(process.pid), { flag: \"wx\" });\n      break;\n    } catch (e) {\n      if ((e as NodeJS.ErrnoException).code !== \"EEXIST\") throw e;\n      if (ageOf(lock) > LOCK_STALE_MS) {\n        // Whoever held this died. Reclaiming a stale lock is safer than\n        // blocking forever on a process that will never return.\n        rmSync(lock, { force: true });\n        continue;\n      }\n      if (existsSync(join(directory, SHIM))) return;\n      if (Date.now() > deadline) {\n        throw new InferenceError(\n          `Timed out waiting for another process to install node-llama-cpp into ` +\n            `${directory}. If nothing else is running, remove ${lock} and retry.`,\n        );\n      }\n      await delay(250);\n    }\n  }\n\n  try {\n    await fn();\n  } finally {\n    rmSync(lock, { force: true });\n  }\n}\n\n/** Infinity for a lock that vanished mid-check, so the caller retries. */\nfunction ageOf(path: string): number {\n  try {\n    return Date.now() - statSync(path).mtimeMs;\n  } catch {\n    return Number.POSITIVE_INFINITY;\n  }\n}\n\nfunction delay(ms: number): Promise<void> {\n  return new Promise((resolve) => setTimeout(resolve, ms));\n}\n","/**\n * In-process local inference over GGUF weights via `node-llama-cpp`.\n *\n * Unlike every other provider here, this one owns weights: it downloads them\n * from Hugging Face on first use and holds gigabytes of RAM once loaded. Two\n * consequences shape the design.\n *\n * First, `node-llama-cpp` is a native module with prebuilt binaries per\n * platform and a CMake fallback. It is an OPTIONAL peer dependency reached\n * through a dynamic `import()`, so the four repos consuming this library pay\n * nothing — install cost or toolchain risk — unless they ask for local models.\n *\n * Second, everything real happens behind `LlamaRuntime`. Tests inject a fake\n * and never touch the network, the filesystem, or a GPU (the same seam as\n * `ExecFn` for the Claude CLI provider).\n */\nimport { InferenceError } from \"../types.js\";\nimport { buildCacheKey } from \"../cache.js\";\nimport { extractJson } from \"./openai-compat.js\";\nimport {\n  aliasForTier,\n  defaultLlamaModelsDirectory,\n  isLlamaSelector,\n  resolveLlamaModelRef,\n} from \"./llama-models.js\";\nimport { importNodeLlamaCpp, isModuleNotFound } from \"./llama-install.js\";\nimport type {\n  CompleteJSONRequest,\n  CompleteJSONResponse,\n  InferenceProvider,\n  TokenUsage,\n} from \"./types.js\";\n\nexport interface LlamaPromptOptions {\n  /** JSON Schema converted to a GBNF grammar by the runtime. */\n  schema: Record<string, unknown>;\n  temperature: number;\n  /** Thinking budget; 0 disables it. See the note in `completeJSON`. */\n  thoughtTokens: number;\n  maxTokens?: number;\n}\n\nexport interface LlamaPromptResult {\n  text: string;\n  usage?: TokenUsage;\n  /**\n   * Why generation stopped. `\"maxTokens\"` means the output was cut off, so the\n   * text is almost certainly truncated JSON — see the guard in `completeJSON`.\n   */\n  stopReason?: string;\n}\n\nexport interface LlamaSession {\n  prompt(text: string, options: LlamaPromptOptions): Promise<LlamaPromptResult>;\n  dispose(): Promise<void>;\n}\n\nexport interface LlamaLoadedModel {\n  createSession(systemPrompt: string): Promise<LlamaSession>;\n  dispose(): Promise<void>;\n}\n\n/**\n * The whole of `node-llama-cpp` that this provider uses. Kept this narrow so a\n * test fake is a few lines and so the real adapter is the only place that\n * knows the upstream API shape.\n */\nexport interface LlamaRuntime {\n  /**\n   * Resolve an `hf:` URI or path to a local file inside `directory`,\n   * downloading if needed.\n   */\n  resolveModelFile(uri: string, directory: string): Promise<string>;\n  loadModel(path: string): Promise<LlamaLoadedModel>;\n  /** Memory available for weights, in bytes — VRAM if there is a GPU, else RAM. */\n  getMemoryBudgetBytes(): Promise<number>;\n}\n\nexport interface LlamaCppProviderOptions {\n  /** Injected for tests; defaults to the real `node-llama-cpp` adapter. */\n  runtime?: LlamaRuntime;\n  /**\n   * Thinking budget in tokens, default 0.\n   *\n   * Gemma 4 has a thinking mode, but a grammar constrains generation from\n   * token 0 — so an unbudgeted model starts reasoning and gets cut off\n   * mid-thought. Zero is the deterministic choice for judging; raise it if you\n   * want reasoning before the JSON.\n   */\n  thoughtTokens?: number;\n  maxTokens?: number;\n  /**\n   * Where to download and look for weights. Defaults to this library's own\n   * directory — see `defaultLlamaModelsDirectory`.\n   */\n  modelsDirectory?: string;\n}\n\n/**\n * Loaded weights, keyed by directory + URI.\n *\n * `runEnsemble` issues N sequential calls and a load costs seconds and\n * gigabytes, so this is process-wide rather than per-instance: two providers\n * naming the same model share one copy. Values are the in-flight promise so\n * concurrent first calls coalesce instead of loading twice. The directory is\n * part of the key because the same URI in two directories is two files.\n */\nconst loadedModels = new Map<string, Promise<LlamaLoadedModel>>();\n\n/**\n * Free every loaded model.\n *\n * A standalone function rather than a `dispose()` on `InferenceProvider`:\n * adding one to the contract would make all five providers carry a lifecycle\n * only this one has. Short-lived processes can skip it.\n */\nexport async function disposeLlamaModels(): Promise<void> {\n  const pending = [...loadedModels.values()];\n  loadedModels.clear();\n  await Promise.all(\n    pending.map((p) => p.then((m) => m.dispose()).catch(() => undefined)),\n  );\n}\n\nexport class LlamaCppProvider implements InferenceProvider {\n  private readonly uri: string;\n  private readonly runtime: LlamaRuntime;\n  private readonly thoughtTokens: number;\n  private readonly maxTokens: number | undefined;\n  private readonly modelsDirectory: string;\n  /**\n   * Loaded-model key: the same URI in two directories is two different files.\n   * Built with `buildCacheKey` so its parts are length-prefixed — a plain join\n   * would let two different (directory, uri) pairs collide and hand a provider\n   * back the wrong weights.\n   */\n  private readonly cacheKey: string;\n\n  constructor(\n    private readonly model: string,\n    options: LlamaCppProviderOptions = {},\n  ) {\n    if (isLlamaSelector(model)) {\n      throw new InferenceError(\n        `llama-cpp model \"${model}\" is a selector. Constructing a provider ` +\n          `directly needs a concrete model (e.g. \"${aliasForTier(\"balanced\")}\") — use ` +\n          `makeProviderAsync to resolve a selector against this machine.`,\n      );\n    }\n    this.uri = resolveLlamaModelRef(model);\n    this.runtime = options.runtime ?? defaultLlamaRuntime();\n    this.thoughtTokens = options.thoughtTokens ?? 0;\n    this.maxTokens = options.maxTokens;\n    this.modelsDirectory =\n      options.modelsDirectory ?? defaultLlamaModelsDirectory();\n    this.cacheKey = buildCacheKey([this.modelsDirectory, this.uri]);\n  }\n\n  provider(): string {\n    return \"llama-cpp\";\n  }\n\n  modelName(): string {\n    return this.model;\n  }\n\n  async completeJSON(req: CompleteJSONRequest): Promise<CompleteJSONResponse> {\n    const model = await this.load();\n    // A fresh session per call: the contract is single-shot, and reusing one\n    // would leak the previous run's turns into this one's context.\n    const session = await model.createSession(systemPromptFor(req));\n    try {\n      const result = await session.prompt(req.user, {\n        schema: req.schema,\n        temperature: req.temperature,\n        thoughtTokens: this.thoughtTokens,\n        ...(this.maxTokens != null ? { maxTokens: this.maxTokens } : {}),\n      });\n      // A run cut off at the token or context limit leaves truncated JSON.\n      // Without this it surfaces as \"failed schema validation\" — or worse,\n      // extractJson's brace-slicing fallback salvages a wrong-but-parseable\n      // object — and the retry burns another full local inference to fail the\n      // same way. Same guard as the Anthropic provider's max_tokens check.\n      if (result.stopReason === \"maxTokens\") {\n        throw new Error(\n          `llama-cpp generation hit the token limit before completing the JSON` +\n            `${this.maxTokens != null ? ` (maxTokens: ${this.maxTokens})` : \"\"}` +\n            ` — raise llamaCpp.maxTokens, or shorten the prompt if the context is full.`,\n        );\n      }\n      return { json: extractJson(result.text), usage: result.usage };\n    } finally {\n      await session.dispose().catch(() => undefined);\n    }\n  }\n\n  private load(): Promise<LlamaLoadedModel> {\n    const existing = loadedModels.get(this.cacheKey);\n    if (existing) return existing;\n    const pending = (async () => {\n      const path = await this.runtime.resolveModelFile(\n        this.uri,\n        this.modelsDirectory,\n      );\n      return this.runtime.loadModel(path);\n    })();\n    // Drop a failed load so the next call retries — a download interrupted by\n    // a flaky network must not poison the model for the rest of the process.\n    // Only evict our OWN entry: a dispose plus a re-load between the failure\n    // and this handler would otherwise orphan the newer model, leaking it.\n    const guarded = pending.catch((e: unknown) => {\n      if (loadedModels.get(this.cacheKey) === guarded) {\n        loadedModels.delete(this.cacheKey);\n      }\n      throw e;\n    });\n    loadedModels.set(this.cacheKey, guarded);\n    return guarded;\n  }\n}\n\n/**\n * The grammar constrains the SHAPE of the output, but `node-llama-cpp` never\n * shows the schema to the model — so `description` fields, which is where\n * consumers put their domain instructions (ADR 01001), would be invisible.\n * Restating the schema is the same fix the Claude CLI provider and the\n * OpenAI json_object fallback already use.\n */\nfunction systemPromptFor(req: CompleteJSONRequest): string {\n  return `${req.system}\\n\\nRespond with ONLY a JSON object conforming to this JSON Schema:\\n${JSON.stringify(\n    req.schema,\n  )}`;\n}\n\n/** Cached so repeated provider construction imports the native module once. */\nlet runtimePromise: Promise<LlamaRuntime> | undefined;\n\n/**\n * Lazy adapter over the real `node-llama-cpp`. Every method defers to the\n * dynamic import, so constructing a provider for a fully-cached run never\n * loads the native binary.\n */\nexport function defaultLlamaRuntime(): LlamaRuntime {\n  const real = (): Promise<LlamaRuntime> =>\n    // Drop a failed init so the next call retries. A GPU that failed to\n    // initialise, or a binary still being extracted by a concurrent install,\n    // must not poison the runtime for the rest of the process — the same rule\n    // `load()` applies to weights.\n    (runtimePromise ??= loadNodeLlamaCpp().catch((e: unknown) => {\n      runtimePromise = undefined;\n      throw e;\n    }));\n  return {\n    resolveModelFile: (uri, directory) =>\n      real().then((r) => r.resolveModelFile(uri, directory)),\n    loadModel: (path) => real().then((r) => r.loadModel(path)),\n    getMemoryBudgetBytes: () => real().then((r) => r.getMemoryBudgetBytes()),\n  };\n}\n\nasync function loadNodeLlamaCpp(): Promise<LlamaRuntime> {\n  let mod: typeof import(\"node-llama-cpp\");\n  try {\n    mod = await import(\"node-llama-cpp\");\n  } catch (e) {\n    // A package that resolved and then failed to load — ABI mismatch, missing\n    // system library, unsupported Node — is not a missing package. Installing\n    // over it would fetch the same broken thing again and replace a precise\n    // error with a download.\n    if (!isModuleNotFound(e)) {\n      throw new InferenceError(\n        `node-llama-cpp is installed but failed to load (${\n          e instanceof Error ? e.message : String(e)\n        }). This is the copy resolved from your own node_modules, so ` +\n          `reinstalling it here will not help — check the Node version and the ` +\n          `platform build.`,\n      );\n    }\n    // Genuinely absent, so fall back to the library's own prefix — installing\n    // it there if needed. npm does not install optional peers, and detection\n    // ends at this provider precisely because it needs no credentials, so\n    // refusing here would strand the one machine `auto` exists to serve.\n    // Resetting `runtimePromise` is the caller's job — see `defaultLlamaRuntime`.\n    mod = (await importNodeLlamaCpp()) as typeof import(\"node-llama-cpp\");\n  }\n\n  const { getLlama, resolveModelFile, LlamaChatSession, TokenMeter } = mod;\n  const llama = await getLlama();\n\n  return {\n    // `directory` is this library's own, not node-llama-cpp's global default —\n    // owning it is what makes `clearLlamaModels` safe.\n    resolveModelFile: (uri, directory) => resolveModelFile(uri, { directory }),\n\n    async loadModel(path) {\n      const model = await llama.loadModel({ modelPath: path });\n      return {\n        async createSession(systemPrompt) {\n          const context = await model.createContext();\n          const sequence = context.getSequence();\n          const session = new LlamaChatSession({\n            contextSequence: sequence,\n            systemPrompt,\n          });\n          return {\n            async prompt(text, options) {\n              const grammar = await llama.createGrammarForJsonSchema(\n                options.schema as Parameters<\n                  typeof llama.createGrammarForJsonSchema\n                >[0],\n              );\n              const before = sequence.tokenMeter.getState();\n              const result = await session.promptWithMeta(text, {\n                grammar,\n                temperature: options.temperature,\n                budgets: { thoughtTokens: options.thoughtTokens },\n                ...(options.maxTokens != null\n                  ? { maxTokens: options.maxTokens }\n                  : {}),\n              });\n              // promptWithMeta does not report usage; the sequence's meter does.\n              const diff = TokenMeter.diff(sequence.tokenMeter, before);\n              return {\n                text: result.responseText,\n                stopReason: result.stopReason,\n                usage: {\n                  inputTokens: diff.usedInputTokens,\n                  outputTokens: diff.usedOutputTokens,\n                },\n              };\n            },\n            async dispose() {\n              await context.dispose();\n            },\n          };\n        },\n        async dispose() {\n          await model.dispose();\n        },\n      };\n    },\n\n    async getMemoryBudgetBytes() {\n      const { totalmem } = await import(\"node:os\");\n      // Half of RAM is what a judge can reasonably claim on a shared machine;\n      // a GPU's free VRAM is usable outright.\n      const ramBudget = totalmem() / 2;\n      try {\n        const vram = await llama.getVramState();\n        // The LARGER of the two, not VRAM in preference to RAM: llama.cpp\n        // offloads the layers that fit onto the GPU and keeps the rest in\n        // system RAM, so a small GPU beside plenty of RAM still runs a big\n        // model. Sizing off VRAM alone would idle most of such a machine.\n        return Math.max(vram.free, ramBudget);\n      } catch {\n        // CPU-only builds and probe failures are normal, never fatal.\n        return ramBudget;\n      }\n    },\n  };\n}\n","/**\n * Which providers can this machine actually use, and which should it pick?\n *\n * Every probe goes through a seam that already exists — environment variables,\n * `ExecFn`, `LlamaRuntime` — so the whole matrix is exercisable offline with no\n * network, no subprocess, and no weights.\n *\n * Detection is async because two of the four probes are: running the Claude CLI\n * and loading the optional `node-llama-cpp` binding. That is why only\n * `resolveProviderIdentityAsync`/`makeProviderAsync` can resolve an `auto`\n * provider, and the synchronous twins throw instead of guessing.\n */\nimport { InferenceError } from \"../types.js\";\nimport { realExec } from \"../exec.js\";\nimport { defaultLlamaRuntime } from \"./llama-cpp.js\";\nimport { nodeLlamaCppStatus } from \"./llama-install.js\";\nimport type { LlamaRuntime } from \"./llama-cpp.js\";\nimport type { ProviderName, ProviderSpec } from \"./index.js\";\n\n/**\n * Priority order. `mock` is deliberately absent: it answers `{ json: {} }`\n * unless scripted, which would sail through as a non-error result — the exact\n * opposite of the \"an errored run is recorded, never coerced\" invariant the\n * consuming eval tools depend on. It must always be asked for by name.\n */\nexport const DETECTION_ORDER: readonly ProviderName[] = [\n  \"anthropic\",\n  \"openai\",\n  \"claude-cli\",\n  \"llama-cpp\",\n];\n\nconst DEFAULT_KEY_ENV: Partial<Record<ProviderName, string>> = {\n  anthropic: \"ANTHROPIC_API_KEY\",\n  openai: \"OPENAI_API_KEY\",\n};\n\ninterface Probe {\n  available: boolean;\n  /** Why not, phrased as advice — this is what the aggregate error prints. */\n  reason?: string;\n}\n\n/**\n * Probes the provider's DEFAULT key variable, deliberately ignoring\n * `spec.apiKeyEnv`.\n *\n * `apiKeyEnv` is one field shared by both API providers, and detection only\n * runs when no provider was named — so a custom name cannot say which provider\n * it belongs to. Honouring it here made a single custom variable satisfy both\n * probes, and anthropic then won on priority: a spec carrying an OpenAI key\n * under a custom name selected `anthropic` and 401'd at call time.\n *\n * A custom `apiKeyEnv` still applies in full once a provider is named — it just\n * cannot be what chooses one.\n */\nfunction hasKey(provider: \"anthropic\" | \"openai\"): boolean {\n  // An empty string is not a key; treating it as one produces a 401 later.\n  return (process.env[DEFAULT_KEY_ENV[provider]!] ?? \"\") !== \"\";\n}\n\n/**\n * Memoised **per command**: spawning a process costs ~150ms and detection may\n * run on every provider construction, but a spec naming a different executable\n * is a different question — memoising on one key would make a fallback to an\n * absolute path silently inherit the bare command's failure.\n *\n * Environment probes stay unmemoised: they are free, and a consumer may\n * legitimately set a key part-way through a process.\n */\nconst cliProbes = new Map<string, Promise<boolean>>();\n\n/** Test seam: forget the memoised Claude CLI probes. */\nexport function resetClaudeCliProbe(): void {\n  cliProbes.clear();\n}\n\nfunction probeClaudeCli(spec: ProviderSpec): Promise<boolean> {\n  const exec = spec.exec ?? realExec;\n  const command = spec.command ?? \"claude\";\n  const cached = cliProbes.get(command);\n  if (cached) return cached;\n  const probing = exec([command, \"--version\"], { timeoutMs: 10_000 })\n    .then((r) => r.code === 0 && !r.timedOut && r.spawnError == null)\n    .catch(() => false);\n  cliProbes.set(command, probing);\n  return probing;\n}\n\nasync function probeLlamaCpp(spec: ProviderSpec): Promise<Probe> {\n  const injected = spec.llamaRuntime ?? spec.llamaCpp?.runtime;\n  if (injected) return budgetProbe(injected);\n\n  // Ask whether the binding is usable BEFORE touching it. Going straight to\n  // `getMemoryBudgetBytes` would route through the auto-installer, so merely\n  // asking `availableProviders()` what this machine can do would fetch a native\n  // module — a query with a side effect, and a slow one.\n  const status = await nodeLlamaCppStatus();\n  if (status.state === \"refused\") {\n    return { available: false, reason: status.reason };\n  }\n  if (status.state === \"installable\") {\n    // Usable, at the cost of an install that happens when it is actually\n    // needed. `warnInstalling` announces it at that point.\n    return { available: true };\n  }\n  // Present: load it for real, which is also what catches a binding that is\n  // installed but whose backend fails to start.\n  return budgetProbe(defaultLlamaRuntime());\n}\n\n/**\n * The same call the `auto` MODEL selector makes, so choosing llama-cpp here\n * costs nothing extra: it is loaded either way.\n */\nfunction budgetProbe(runtime: LlamaRuntime): Promise<Probe> {\n  return runtime.getMemoryBudgetBytes().then(\n    () => ({ available: true }),\n    (e: unknown) => ({\n      available: false,\n      reason:\n        e instanceof Error && /node-llama-cpp/.test(e.message)\n          ? \"node-llama-cpp is not installed (npm i node-llama-cpp)\"\n          : `node-llama-cpp could not start (${\n              e instanceof Error ? e.message : String(e)\n            })`,\n    }),\n  );\n}\n\nasync function probe(\n  provider: ProviderName,\n  spec: ProviderSpec,\n): Promise<Probe> {\n  switch (provider) {\n    case \"anthropic\":\n      return hasKey(\"anthropic\")\n        ? { available: true }\n        : {\n            available: false,\n            reason: `ANTHROPIC_API_KEY is not set`,\n          };\n    case \"openai\":\n      // A local OpenAI-compatible server needs no key — same rule the\n      // OpenAICompatProvider constructor applies.\n      return hasKey(\"openai\") || spec.baseUrl\n        ? { available: true }\n        : {\n            available: false,\n            reason: `OPENAI_API_KEY is not set and no baseUrl was given`,\n          };\n    case \"claude-cli\":\n      return (await probeClaudeCli(spec))\n        ? { available: true }\n        : {\n            available: false,\n            reason: `could not run \\`${spec.command ?? \"claude\"}\\` (is the Claude CLI installed?)`,\n          };\n    case \"llama-cpp\":\n      return probeLlamaCpp(spec);\n    default:\n      return { available: false, reason: \"not auto-selectable\" };\n  }\n}\n\n/**\n * Every provider this machine could use right now, in priority order.\n *\n * Useful for showing a picker or explaining a fallback; `detectProvider` is\n * the same sweep with the first hit returned.\n */\nexport async function availableProviders(\n  spec: ProviderSpec = {},\n): Promise<ProviderName[]> {\n  const probes = await Promise.all(\n    DETECTION_ORDER.map((name) => probe(name, spec)),\n  );\n  return DETECTION_ORDER.filter((_, i) => probes[i]!.available);\n}\n\n/**\n * The highest-priority provider this machine can use.\n *\n * Throws an `InferenceError` naming every provider and why each was\n * unavailable — far more actionable than the `Unknown provider \"undefined\"`\n * this replaces.\n */\nexport async function detectProvider(\n  spec: ProviderSpec = {},\n): Promise<ProviderName> {\n  // Sequential, not Promise.all: the probes get dramatically more expensive\n  // down the list, and the cheapest usually wins. Reading an environment\n  // variable costs microseconds, spawning the Claude CLI ~150ms, and loading\n  // the node-llama-cpp binding ~850ms — the last of which also initialises the\n  // llama backend and allocates GPU context. Probing eagerly would pay all of\n  // that on every construction just to pick `anthropic` off an env var, and\n  // would touch the GPU for a provider that is never used.\n  const reasons: string[] = [];\n  for (const name of DETECTION_ORDER) {\n    const result = await probe(name, spec);\n    if (result.available) {\n      warnSelected(name, spec.provider === \"auto\");\n      return name;\n    }\n    reasons.push(`  ${name.padEnd(10)} — ${result.reason}`);\n  }\n  throw new InferenceError(\n    `No inference provider is available. Tried:\\n${reasons.join(\"\\n\")}\\n` +\n      `Pass an explicit \\`provider\\`, set one of the keys above, or install node-llama-cpp.`,\n  );\n}\n\nlet warnedSelection = false;\nlet warnedDownload = false;\n\n/** Test seam: reset the once-per-process auto-detection warnings. */\nexport function resetProviderDetectionWarning(): void {\n  warnedSelection = false;\n  warnedDownload = false;\n}\n\n/**\n * These are eval tools: a run whose provider silently changed because an\n * environment variable moved is a run whose verdicts and cache are no longer\n * comparable to the last one. Say which provider was picked, once.\n */\nfunction warnSelected(provider: ProviderName, wasExplicitAuto: boolean): void {\n  if (warnedSelection) return;\n  warnedSelection = true;\n  // `provider: \"auto\"` IS a specification — saying otherwise sends someone\n  // hunting their config for a field they did set.\n  const because = wasExplicitAuto ? `provider \"auto\"` : \"no provider specified\";\n  console.warn(\n    `inference: ${because} — auto-selected \"${provider}\". ` +\n      `Pass an explicit \\`provider\\` to pin it.`,\n  );\n}\n\n/**\n * Falling back to a local model can mean pulling gigabytes. Say so before it\n * starts, not after a CI job has already stalled on it.\n */\nexport function warnPendingDownload(model: string, sizeBytes: number): void {\n  if (warnedDownload) return;\n  warnedDownload = true;\n  console.warn(\n    `inference: \"${model}\" is not downloaded yet — the first run will fetch ` +\n      `~${(sizeBytes / 1e9).toFixed(2)} GB. Pre-fetch it, or pass an explicit ` +\n      `\\`provider\\` to avoid the local model entirely.`,\n  );\n}\n","/**\n * Provider factory over a library-owned `ProviderSpec`.\n *\n * The spec is deliberately NOT any consumer's config type. Every consumer maps\n * its own config into this flat shape, so adding a provider here does not\n * require touching four config schemas, and no consumer has to model its\n * config on another's to reuse this layer (ADR 01000).\n */\nimport { InferenceError } from \"../types.js\";\nimport { warnIfUnsupportedNode } from \"../runtime.js\";\nimport type { Pricing } from \"../cost.js\";\nimport { AnthropicProvider } from \"./anthropic.js\";\nimport { OpenAICompatProvider } from \"./openai-compat.js\";\nimport { ClaudeCliProvider } from \"./claude-cli.js\";\nimport { MockProvider } from \"./mock.js\";\nimport { LlamaCppProvider, defaultLlamaRuntime } from \"./llama-cpp.js\";\nimport {\n  LLAMA_MODELS,\n  aliasForTier,\n  defaultLlamaModelsDirectory,\n  isLlamaSelector,\n  isModelDownloaded,\n  tierForBudget,\n} from \"./llama-models.js\";\nimport { detectProvider, warnPendingDownload } from \"./detect.js\";\nimport type { AnthropicProviderOptions } from \"./anthropic.js\";\nimport type { OpenAICompatProviderOptions } from \"./openai-compat.js\";\nimport type { MockResponse } from \"./mock.js\";\nimport type { LlamaCppProviderOptions, LlamaRuntime } from \"./llama-cpp.js\";\nimport type { LlamaTier } from \"./llama-models.js\";\nimport type { ExecFn, InferenceProvider } from \"./types.js\";\n\nexport type ProviderName =\n  | \"anthropic\"\n  | \"openai\"\n  | \"claude-cli\"\n  | \"mock\"\n  | \"llama-cpp\";\n\n/** A concrete provider, or `\"auto\"` to detect one. */\nexport type ProviderSelector = ProviderName | \"auto\";\n\nexport interface ProviderSpec {\n  /**\n   * Omitting this is identical to `\"auto\"`: the highest-priority provider this\n   * machine can actually use is detected, ending at the free local model.\n   * Resolving it needs `makeProviderAsync`/`resolveProviderIdentityAsync`.\n   */\n  provider?: ProviderSelector;\n  /** null/undefined selects the per-provider default. */\n  model?: string | null;\n  /** Env var NAME holding the API key; null/undefined selects the default. */\n  apiKeyEnv?: string | null;\n  /** openai only. */\n  baseUrl?: string;\n  /** claude-cli only: the executable to run. */\n  command?: string;\n  /** claude-cli only: subprocess timeout. */\n  timeoutMs?: number;\n  /**\n   * Pricing override for this model. Not used to construct the provider —\n   * carried here so a consumer passes one object to both `makeProvider` and\n   * `pricingFor`.\n   */\n  pricing?: Pricing;\n  /** Provider-specific tuning, ignored by the other providers. */\n  anthropic?: AnthropicProviderOptions;\n  openai?: OpenAICompatProviderOptions;\n  llamaCpp?: LlamaCppProviderOptions;\n  /** Test seam for the claude-cli provider. */\n  exec?: ExecFn;\n  /** Test seam for the llama-cpp provider. */\n  llamaRuntime?: LlamaRuntime;\n  /** Scripted responses for the mock provider; defaults to a single empty object. */\n  mockResponses?: MockResponse[];\n}\n\nexport const DEFAULT_MODELS: Record<ProviderName, string> = {\n  anthropic: \"claude-sonnet-4-5\",\n  openai: \"gpt-4o-mini\",\n  \"claude-cli\": \"claude-sonnet-4-5\",\n  mock: \"mock-model\",\n  // A selector, not a pinned model: which weights a tier points at is then a\n  // catalog change rather than an API change. Resolving it needs the async\n  // factory — see `resolveProviderIdentityAsync`.\n  \"llama-cpp\": \"auto\",\n};\n\nconst DEFAULT_API_KEY_ENV: Record<string, string> = {\n  anthropic: \"ANTHROPIC_API_KEY\",\n  openai: \"OPENAI_API_KEY\",\n};\n\nexport const DEFAULT_OPENAI_BASE_URL = \"https://api.openai.com/v1\";\n\nexport interface ProviderIdentity {\n  /** Always concrete — never `\"auto\"`, so it is safe as cache-key material. */\n  provider: ProviderName;\n  model: string;\n}\n\n/**\n * Resolve the provider name and model WITHOUT constructing the provider —\n * cache keys and pricing need the identity, but construction may require an\n * API key that a fully-cached run never uses.\n *\n * Throws for an unresolved `llama-cpp` selector. That is deliberate: picking a\n * tier weighs GPU VRAM, which needs an await, and returning the literal\n * \"auto\" as cache-key material would let a 2 GB and a 12 GB model share cached\n * verdicts — and give different results per machine under one key. Use\n * `resolveProviderIdentityAsync` for selectors.\n */\nexport function resolveProviderIdentity(spec: ProviderSpec): ProviderIdentity {\n  if (spec.provider == null || spec.provider === \"auto\") {\n    throw new InferenceError(\n      `No provider specified. Detecting one probes the environment, the Claude ` +\n        `CLI and the local model runtime, which cannot be done synchronously — ` +\n        `use resolveProviderIdentityAsync/makeProviderAsync, or name a provider ` +\n        `(${Object.keys(DEFAULT_MODELS).join(\", \")}).`,\n    );\n  }\n  const model = spec.model ?? DEFAULT_MODELS[spec.provider] ?? \"unknown\";\n  if (spec.provider === \"llama-cpp\" && isLlamaSelector(model)) {\n    throw new InferenceError(\n      `llama-cpp model \"${model}\" is a selector and cannot be resolved ` +\n        `synchronously — picking a tier probes GPU memory. Use ` +\n        `resolveProviderIdentityAsync/makeProviderAsync, or name a concrete ` +\n        `model (e.g. \"${aliasForTier(\"balanced\")}\").`,\n    );\n  }\n  return { provider: spec.provider, model };\n}\n\n/**\n * Selector-aware identity resolution. Returns the CONCRETE model a selector\n * resolved to, so the cache key names the weights that actually ran.\n *\n * Every other provider delegates to the synchronous form, so a consumer can\n * switch to this wholesale.\n */\nexport async function resolveProviderIdentityAsync(\n  spec: ProviderSpec,\n): Promise<ProviderIdentity> {\n  // A model name belongs to exactly one provider, so it cannot be carried into\n  // whichever provider detection happens to pick: `{ model: \"gpt-4o-mini\" }` on\n  // a machine with an Anthropic key selected `anthropic` and then 404'd at call\n  // time, after the caller had already paid for detection. `null` still means\n  // \"use the default\" — only a real name is ambiguous.\n  if (\n    (spec.provider == null || spec.provider === \"auto\") &&\n    spec.model != null\n  ) {\n    throw new InferenceError(\n      `Model \"${spec.model}\" was given without a provider, and a model name ` +\n        `does not say which provider owns it. Name the provider too ` +\n        `(${Object.keys(DEFAULT_MODELS).join(\", \")}), or drop the model to ` +\n        `take the detected provider's default.`,\n    );\n  }\n\n  // Provider first, then the model logic for whichever provider won.\n  const provider =\n    spec.provider == null || spec.provider === \"auto\"\n      ? await detectProvider(spec)\n      : spec.provider;\n  const resolved: ProviderSpec = { ...spec, provider };\n\n  const model = spec.model ?? DEFAULT_MODELS[provider] ?? \"unknown\";\n  if (provider !== \"llama-cpp\" || !isLlamaSelector(model)) {\n    return resolveProviderIdentity(resolved);\n  }\n  const tier: LlamaTier =\n    model === \"auto\" ? await probeTier(llamaRuntimeFor(spec)) : model;\n  return { provider, model: aliasForTier(tier) };\n}\n\n/**\n * Warn before a multi-gigabyte download, but only once the caller has actually\n * committed to running.\n *\n * Deliberately NOT in `resolveProviderIdentityAsync`: that resolves an identity\n * *without* constructing anything, which is exactly what a fully-cached run\n * does — and such a run downloads nothing, so warning there announces gigabytes\n * that never move.\n */\nfunction warnIfDownloadPending(spec: ProviderSpec, model: string): void {\n  const entry = LLAMA_MODELS[model];\n  if (!entry) return;\n  const directory =\n    spec.llamaCpp?.modelsDirectory ?? defaultLlamaModelsDirectory();\n  if (!isModelDownloaded(model, directory)) {\n    warnPendingDownload(model, entry.sizeBytes);\n  }\n}\n\n/**\n * A runtime can be injected either as `spec.llamaRuntime` or inside\n * `spec.llamaCpp` — `makeProvider` honours both, so selector resolution must\n * too. Missing one sends the probe to the real native module and throws for a\n * consumer whose whole point was to avoid it.\n */\nfunction llamaRuntimeFor(spec: ProviderSpec): LlamaRuntime | undefined {\n  return spec.llamaRuntime ?? spec.llamaCpp?.runtime;\n}\n\nasync function probeTier(runtime: LlamaRuntime | undefined): Promise<LlamaTier> {\n  const source = runtime ?? defaultLlamaRuntime();\n  return tierForBudget(await source.getMemoryBudgetBytes());\n}\n\nexport function makeProvider(spec: ProviderSpec): InferenceProvider {\n  // First use of the library, for anyone who goes through the factory — before\n  // the spec is even validated, so an unsupported Node is named ahead of any\n  // error it might be the real cause of.\n  warnIfUnsupportedNode();\n  const { model } = resolveProviderIdentity(spec);\n\n  switch (spec.provider) {\n    case \"anthropic\":\n      return new AnthropicProvider(\n        model,\n        spec.apiKeyEnv ?? DEFAULT_API_KEY_ENV[\"anthropic\"]!,\n        spec.anthropic ?? {},\n      );\n    case \"openai\":\n      return new OpenAICompatProvider(\n        spec.baseUrl ?? DEFAULT_OPENAI_BASE_URL,\n        model,\n        spec.apiKeyEnv ?? DEFAULT_API_KEY_ENV[\"openai\"]!,\n        undefined,\n        spec.openai ?? {},\n      );\n    case \"claude-cli\":\n      return new ClaudeCliProvider(\n        model,\n        spec.command ?? \"claude\",\n        spec.exec,\n        spec.timeoutMs,\n      );\n    case \"mock\":\n      // Offline smoke-testing seam: proposes nothing unless scripted.\n      return new MockProvider(spec.mockResponses ?? [{ json: {} }], model);\n    case \"llama-cpp\":\n      return new LlamaCppProvider(model, {\n        ...(spec.llamaCpp ?? {}),\n        ...(spec.llamaRuntime ? { runtime: spec.llamaRuntime } : {}),\n      });\n    default:\n      throw new InferenceError(\n        `Unknown provider \"${String(spec.provider)}\". Available: ${Object.keys(\n          DEFAULT_MODELS,\n        ).join(\", \")}.`,\n      );\n  }\n}\n\n/**\n * Selector-aware provider construction. Resolves a `llama-cpp` selector\n * against this machine first, so the returned provider's `modelName()` — and\n * therefore the cache key — names the weights it will actually load.\n *\n * Every other provider delegates to `makeProvider`.\n */\nexport async function makeProviderAsync(\n  spec: ProviderSpec,\n): Promise<InferenceProvider> {\n  // Both halves must be threaded through: passing only the model would leave a\n  // detected provider as `undefined` and throw in `makeProvider`.\n  const { provider, model } = await resolveProviderIdentityAsync(spec);\n  if (provider === \"llama-cpp\") warnIfDownloadPending(spec, model);\n  return makeProvider({ ...spec, provider, model });\n}\n","/**\n * Reclaiming disk space from downloaded GGUF weights.\n *\n * This is straightforward because the models directory belongs to this library\n * alone (see `defaultLlamaModelsDirectory`). node-llama-cpp's own global\n * directory is shared with its CLI and anything else on the machine, so\n * clearing THAT would destroy models this library never downloaded; owning a\n * directory removes the hazard rather than defending against it.\n */\nimport { rmSync, statSync } from \"node:fs\";\nimport { join } from \"node:path\";\nimport {\n  blobNameFor,\n  defaultLlamaModelsDirectory,\n  listModelDirectory,\n  matchesModelBlob,\n} from \"./llama-models.js\";\nimport { disposeLlamaModels } from \"./llama-cpp.js\";\n\nexport interface ClearedModelFile {\n  path: string;\n  sizeBytes: number;\n}\n\nexport interface ClearLlamaModelsResult {\n  /** Files removed, or that would be removed under `dryRun`. */\n  files: ClearedModelFile[];\n  freedBytes: number;\n  directory: string;\n  dryRun: boolean;\n}\n\nexport interface ClearLlamaModelsOptions {\n  /** Defaults to this library's own models directory. */\n  directory?: string;\n  /** Clear only these models (alias or `hf:` URI); default clears all. */\n  models?: string[];\n  /** Report what would be removed without deleting anything. */\n  dryRun?: boolean;\n}\n\n/**\n * Delete downloaded GGUF weights and interrupted partial downloads.\n *\n * Loaded models are disposed first: on Windows the weights are memory-mapped\n * while loaded, and deleting an open file fails with EBUSY/EPERM.\n */\nexport async function clearLlamaModels(\n  options: ClearLlamaModelsOptions = {},\n): Promise<ClearLlamaModelsResult> {\n  const directory = options.directory ?? defaultLlamaModelsDirectory();\n  const dryRun = options.dryRun ?? false;\n  // Resolve BEFORE touching disk so an unknown name fails without having\n  // deleted half the set.\n  const wanted = options.models?.map(blobNameFor);\n\n  if (!dryRun) await disposeLlamaModels();\n\n  const files: ClearedModelFile[] = [];\n  for (const entry of listModelDirectory(directory)) {\n    // Only ever weights — a stray config or log in this directory is safe, and\n    // so is anything else if `directory` was pointed somewhere shared.\n    if (!isModelBlob(entry)) continue;\n    if (wanted && !wanted.some((name) => matchesModelBlob(entry, name))) {\n      continue;\n    }\n    const path = join(directory, entry);\n    const sizeBytes = sizeOf(path);\n    if (sizeBytes === undefined) continue;\n    if (!dryRun) {\n      try {\n        rmSync(path);\n      } catch {\n        // A file held open by another process is not this call's business to\n        // force; report only what actually went away.\n        continue;\n      }\n    }\n    files.push({ path, sizeBytes });\n  }\n\n  return {\n    files,\n    freedBytes: files.reduce((total, file) => total + file.sizeBytes, 0),\n    directory,\n    dryRun,\n  };\n}\n\nfunction isModelBlob(entry: string): boolean {\n  return entry.endsWith(\".gguf\") || entry.endsWith(\".gguf.ipull\");\n}\n\nfunction sizeOf(path: string): number | undefined {\n  try {\n    return statSync(path).size;\n  } catch {\n    return undefined;\n  }\n}\n","/**\n * One schema-validated completion, with a single retry.\n *\n * The invariant every consumer depends on: a run that cannot produce\n * schema-valid JSON after the retry is recorded as an ERROR, not dropped and\n * not coerced. Downstream, an errored run counts against consensus — it can\n * push a result toward human review, but it can never produce a silent pass.\n */\nimport { Ajv2020 } from \"ajv/dist/2020.js\";\nimport type { ValidateFunction } from \"ajv\";\nimport { warnIfUnsupportedNode } from \"./runtime.js\";\nimport type { InferenceProvider, TokenUsage } from \"./providers/types.js\";\n\n/** One attempt at a schema-constrained completion. */\nexport interface InferenceRun<T = unknown> {\n  /** Absent when the run errored (invalid JSON after retry, API failure). */\n  result?: T;\n  error?: string;\n  provider: string;\n  model: string;\n  cached: boolean;\n  usage?: TokenUsage;\n  durationMs: number;\n}\n\nexport interface CompleteValidatedOptions {\n  provider: InferenceProvider;\n  system: string;\n  user: string;\n  schema: Record<string, unknown>;\n  temperature?: number;\n  /**\n   * Attempts before recording an error. Default 2 (one initial call plus one\n   * retry) — matches the behavior all three source projects shipped.\n   */\n  attempts?: number;\n  /**\n   * Pre-compiled validator. Compiling Ajv per call is wasteful in an ensemble\n   * loop, so `runEnsemble` compiles once and passes it down.\n   */\n  validate?: ValidateFunction;\n}\n\nconst validatorCache = new WeakMap<object, ValidateFunction>();\n\n/** Compile once per schema object identity — Ajv compilation is not cheap. */\nexport function validatorFor(\n  schema: Record<string, unknown>,\n): ValidateFunction {\n  const cached = validatorCache.get(schema);\n  if (cached) return cached;\n  // A fresh Ajv per distinct schema object, not one shared instance: Ajv keeps\n  // a registry keyed by `$id`, so a caller that rebuilds an equal schema object\n  // per call (spreading VERDICT_SCHEMA to override descriptions, say) misses\n  // the identity cache above and would hit \"schema with key or id ... already\n  // exists\" on the second compile. Instances are held only by this WeakMap, so\n  // they are collected with the schemas that own them.\n  const compiled = new Ajv2020({ allErrors: true }).compile(schema);\n  validatorCache.set(schema, compiled);\n  return compiled;\n}\n\nexport async function completeValidatedJSON<T = unknown>(\n  options: CompleteValidatedOptions,\n): Promise<InferenceRun<T>> {\n  // The other half of \"first use\": a consumer that constructs a provider\n  // directly never touches `makeProvider`, but everything still funnels here.\n  warnIfUnsupportedNode();\n  const {\n    provider,\n    system,\n    user,\n    schema,\n    temperature = 0,\n    attempts = 2,\n  } = options;\n  const validate = options.validate ?? validatorFor(schema);\n\n  const start = Date.now();\n  const base = {\n    provider: provider.provider(),\n    model: provider.modelName(),\n    cached: false,\n  };\n\n  let lastError = \"unknown error\";\n  for (let attempt = 0; attempt < attempts; attempt++) {\n    try {\n      const response = await provider.completeJSON({\n        system,\n        user,\n        schema,\n        temperature,\n      });\n      if (validate(response.json)) {\n        return {\n          ...base,\n          result: response.json as T,\n          usage: response.usage,\n          durationMs: Date.now() - start,\n        };\n      }\n      lastError = `Response failed schema validation: ${(validate.errors ?? [])\n        .map((e) => `${e.instancePath} ${e.message}`)\n        .join(\"; \")}`;\n    } catch (e) {\n      lastError = e instanceof Error ? e.message : String(e);\n    }\n  }\n  return { ...base, error: lastError, durationMs: Date.now() - start };\n}\n","/**\n * Cost tracking: token usage priced from a small static table, overridable per\n * model by the caller. Unknown models cost 0 (unknown), never a guess — a\n * fabricated price is worse than an absent one when a budget gate depends on it.\n */\nimport type { TokenUsage } from \"./providers/types.js\";\n\nexport interface Pricing {\n  inputPerMTok: number;\n  outputPerMTok: number;\n}\n\n/**\n * USD per million tokens. Entries are base names; pinned variants\n * (`claude-sonnet-4-5-20250929`) resolve by prefix.\n */\nexport const PRICE_TABLE: Record<string, Pricing> = {\n  \"claude-sonnet-4-5\": { inputPerMTok: 3, outputPerMTok: 15 },\n  \"claude-sonnet-4-6\": { inputPerMTok: 3, outputPerMTok: 15 },\n  \"claude-haiku-4-5\": { inputPerMTok: 1, outputPerMTok: 5 },\n  \"claude-opus-4-8\": { inputPerMTok: 15, outputPerMTok: 75 },\n  \"gpt-4o-mini\": { inputPerMTok: 0.15, outputPerMTok: 0.6 },\n  \"gpt-4o\": { inputPerMTok: 2.5, outputPerMTok: 10 },\n};\n\nexport function pricingFor(\n  model: string,\n  override?: Pricing,\n): Pricing | undefined {\n  if (override) return override;\n  if (PRICE_TABLE[model]) return PRICE_TABLE[model];\n  // Match pinned variants like claude-sonnet-4-5-20250929. Longest prefix\n  // wins so `claude-sonnet-4-5` never shadows a more specific future entry.\n  const base = Object.keys(PRICE_TABLE)\n    .filter((k) => model.startsWith(k))\n    .sort((a, b) => b.length - a.length)[0];\n  return base ? PRICE_TABLE[base] : undefined;\n}\n\nexport function costOfUsage(\n  usage: TokenUsage | undefined,\n  pricing: Pricing | undefined,\n): number {\n  if (!usage || !pricing) return 0;\n  return (\n    (usage.inputTokens / 1_000_000) * pricing.inputPerMTok +\n    (usage.outputTokens / 1_000_000) * pricing.outputPerMTok\n  );\n}\n\n/** Sum the cost of a set of runs. Cached runs cost nothing — they made no call. */\nexport function costOfRuns(\n  runs: { usage?: TokenUsage; cached?: boolean }[],\n  pricing: Pricing | undefined,\n): number {\n  if (!pricing) return 0;\n  let usd = 0;\n  for (const run of runs) {\n    if (!run.usage || run.cached) continue;\n    usd += costOfUsage(run.usage, pricing);\n  }\n  return usd;\n}\n","{\n  \"$schema\": \"https://json-schema.org/draft/2020-12/schema\",\n  \"$id\": \"inference:verdict:0.1\",\n  \"title\": \"LLM-as-judge verdict\",\n  \"type\": \"object\",\n  \"required\": [\"claim\", \"observed\", \"match\", \"confidence\", \"reasoning\"],\n  \"properties\": {\n    \"claim\": {\n      \"type\": \"string\",\n      \"description\": \"The specific assertion under evaluation.\"\n    },\n    \"observed\": {\n      \"type\": \"string\",\n      \"description\": \"What was actually observed in the subject under evaluation, quoting evidence where possible.\"\n    },\n    \"match\": {\n      \"enum\": [\"pass\", \"fail\", \"partial\"],\n      \"description\": \"pass only if the assertion is fully satisfied; partial for partial compliance.\"\n    },\n    \"confidence\": {\n      \"type\": \"number\",\n      \"minimum\": 0,\n      \"maximum\": 1,\n      \"description\": \"Self-reported confidence in this verdict.\"\n    },\n    \"reasoning\": {\n      \"type\": \"string\",\n      \"description\": \"Why this conclusion was reached.\"\n    }\n  },\n  \"additionalProperties\": false\n}\n","/**\n * LLM-as-judge types. The verdict shape is the one all consuming projects\n * already share; the canonical schema lives in verdict-schema.json and is\n * overridable per consumer so domain-specific field descriptions survive\n * (ADR 01001).\n */\nimport verdictSchemaJson from \"./verdict-schema.json\" with { type: \"json\" };\nimport type { TokenUsage } from \"../providers/types.js\";\n\nexport const VERDICT_SCHEMA = verdictSchemaJson as Record<string, unknown>;\n\nexport type Match = \"pass\" | \"fail\" | \"partial\";\n\n/** Confidence-zone routing for LLM-judged evals. */\nexport type Zone = \"auto-pass\" | \"auto-fail\" | \"human-review\";\n\nexport interface JudgeVerdict {\n  /** The specific assertion under evaluation. */\n  claim: string;\n  /** What the judge actually observed. */\n  observed: string;\n  match: Match;\n  /** 0.0–1.0 self-reported confidence. */\n  confidence: number;\n  reasoning: string;\n}\n\n/** One run within an ensemble. */\nexport interface JudgeRun {\n  /** Absent when the run errored (invalid JSON after retry, API failure). */\n  verdict?: JudgeVerdict;\n  error?: string;\n  provider: string;\n  model: string;\n  cached: boolean;\n  usage?: TokenUsage;\n  durationMs: number;\n}\n\n/** Aggregated outcome of an ensemble of judge runs for one subject. */\nexport interface ConsensusResult {\n  runs: JudgeRun[];\n  votes: { pass: number; fail: number; partial: number; error: number };\n  /** Majority verdict; `partial` counts as fail for the binary outcome. */\n  verdict: Match;\n  /** Fraction of non-errored runs agreeing with the majority verdict. */\n  agreement: number;\n  /** Mean confidence across non-errored runs. */\n  meanConfidence: number;\n  zone: Zone;\n}\n","/**\n * Ensemble consensus math. `partial` counts as fail for the binary outcome\n * but stays visible in the vote counts. Errored runs count against consensus —\n * they can only push a result toward human review, never toward a silent pass.\n */\nimport type { ConsensusResult, JudgeRun, Match } from \"./types.js\";\n\nexport function computeConsensus(\n  runs: JudgeRun[],\n): Omit<ConsensusResult, \"zone\"> {\n  const votes = { pass: 0, fail: 0, partial: 0, error: 0 };\n  let confidenceSum = 0;\n  let confidenceCount = 0;\n  for (const run of runs) {\n    if (!run.verdict) {\n      votes.error += 1;\n      continue;\n    }\n    votes[run.verdict.match] += 1;\n    confidenceSum += run.verdict.confidence;\n    confidenceCount += 1;\n  }\n\n  const passVotes = votes.pass;\n  const failVotes = votes.fail + votes.partial;\n  // Binary majority; a tie is not a pass.\n  const verdict: Match = passVotes > failVotes ? \"pass\" : \"fail\";\n  const graded = passVotes + failVotes;\n  const agreement = graded > 0 ? Math.max(passVotes, failVotes) / graded : 0;\n  const meanConfidence =\n    confidenceCount > 0 ? confidenceSum / confidenceCount : 0;\n\n  return { runs, votes, verdict, agreement, meanConfidence };\n}\n","/**\n * Confidence-zone routing: only unanimous, high-confidence ensembles\n * auto-resolve; everything else goes to a human.\n */\nimport type { ConsensusResult, Zone } from \"./types.js\";\n\nexport interface ZoneThresholds {\n  autoPass: number;\n  autoFail: number;\n}\n\nexport const DEFAULT_ZONES: ZoneThresholds = { autoPass: 0.8, autoFail: 0.8 };\n\nexport function zoneFor(\n  consensus: Omit<ConsensusResult, \"zone\">,\n  thresholds: ZoneThresholds = DEFAULT_ZONES,\n): Zone {\n  const { votes, meanConfidence } = consensus;\n  const unanimousPass =\n    votes.pass > 0 &&\n    votes.fail === 0 &&\n    votes.partial === 0 &&\n    votes.error === 0;\n  const unanimousFail =\n    votes.pass === 0 && votes.error === 0 && votes.fail + votes.partial > 0;\n\n  if (unanimousPass && meanConfidence >= thresholds.autoPass) return \"auto-pass\";\n  if (unanimousFail && meanConfidence >= thresholds.autoFail) return \"auto-fail\";\n  return \"human-review\";\n}\n","/**\n * The ensemble judge: N independent runs for one subject, each a fresh request\n * with no shared context (eval isolation), aggregated by consensus and routed\n * through confidence zones.\n *\n * Runs within one ensemble stay sequential on purpose — they are meant to be\n * independent samples, and interleaving them buys nothing. Concurrency belongs\n * one level up, across subjects, where the consumer owns the pool.\n */\nimport { completeValidatedJSON, validatorFor } from \"../complete.js\";\nimport type { JsonCache } from \"../cache.js\";\nimport type { InferenceProvider } from \"../providers/types.js\";\nimport { computeConsensus } from \"./consensus.js\";\nimport { DEFAULT_ZONES, zoneFor, type ZoneThresholds } from \"./zones.js\";\nimport {\n  VERDICT_SCHEMA,\n  type ConsensusResult,\n  type JudgeRun,\n  type JudgeVerdict,\n} from \"./types.js\";\n\nexport interface EnsembleOptions {\n  provider: InferenceProvider;\n  system: string;\n  user: string;\n  /** Ensemble size; default 3. */\n  runs?: number;\n  /** Default 0. Nonzero adds noise to verdicts and warns once. */\n  temperature?: number;\n  /**\n   * Verdict schema. Defaults to the canonical one; pass your own to keep\n   * domain-specific field descriptions (they measurably steer the model).\n   * Must still produce objects matching `JudgeVerdict`.\n   */\n  schema?: Record<string, unknown>;\n  /** Optional result cache. Requires `cacheKey`. */\n  cache?: JsonCache<JudgeRun[]>;\n  /** Content-addressed key; build it with `buildCacheKey`. */\n  cacheKey?: string;\n  /** Prefix for warnings, e.g. your tool's name. */\n  label?: string;\n}\n\nlet warnedTemperature = false;\n\n/**\n * Run the ensemble and return every run. Cached ensembles replay identically,\n * with each run flagged `cached: true` so cost accounting skips them.\n */\nexport async function runEnsemble(\n  options: EnsembleOptions,\n): Promise<JudgeRun[]> {\n  const {\n    provider,\n    system,\n    user,\n    runs: runCount = 3,\n    temperature = 0,\n    schema = VERDICT_SCHEMA,\n    cache,\n    cacheKey,\n    label = \"inference\",\n  } = options;\n\n  if (temperature > 0 && !warnedTemperature) {\n    warnedTemperature = true;\n    console.warn(\n      `${label}: judge temperature is ${temperature} — nonzero temperature adds noise to verdicts; 0 is strongly recommended.`,\n    );\n  }\n\n  if (cache && cacheKey) {\n    const hit = cache.get(cacheKey);\n    // A cache file can be valid JSON and still be the wrong shape — a\n    // truncated write, or an entry from an older cache generation. JsonCache\n    // only guards parse failures, so the shape check belongs here; a bad entry\n    // is a miss, never a crash.\n    if (Array.isArray(hit)) return hit.map((r) => ({ ...r, cached: true }));\n  }\n\n  // Compile the schema once for the whole ensemble rather than per run.\n  const validate = validatorFor(schema);\n\n  const results: JudgeRun[] = [];\n  for (let i = 0; i < runCount; i++) {\n    const run = await completeValidatedJSON<JudgeVerdict>({\n      provider,\n      system,\n      user,\n      schema,\n      temperature,\n      validate,\n    });\n    // `result` is the generic name at the completion layer; the judge layer\n    // calls it `verdict`, which is what consumers persist in their caches.\n    const { result, ...rest } = run;\n    results.push(result === undefined ? rest : { ...rest, verdict: result });\n  }\n\n  if (cache && cacheKey) cache.set(cacheKey, results);\n  return results;\n}\n\n/** Run the ensemble and aggregate it into a zoned consensus. */\nexport async function judge(\n  options: EnsembleOptions & { zones?: ZoneThresholds },\n): Promise<ConsensusResult> {\n  const runs = await runEnsemble(options);\n  const base = computeConsensus(runs);\n  return { ...base, zone: zoneFor(base, options.zones ?? DEFAULT_ZONES) };\n}\n\n/** Test seam: reset the once-per-process temperature warning. */\nexport function resetTemperatureWarning(): void {\n  warnedTemperature = false;\n}\n"],"mappings":";AAKO,IAAM,iBAAN,cAA6B,MAAM;AAAA,EACxC,YAAY,SAAiB;AAC3B,UAAM,OAAO;AACb,SAAK,OAAO;AAAA,EACd;AACF;;;ACYO,IAAM,qBAAqB;AAElC,IAAI,oBAAoB;AAQjB,SAAS,sBACd,UAAkB,QAAQ,SAAS,MAC7B;AACN,MAAI,kBAAmB;AACvB,QAAM,QAAQ,OAAO,SAAS,SAAS,EAAE;AAGzC,MAAI,CAAC,OAAO,UAAU,KAAK,KAAK,SAAS,mBAAoB;AAC7D,sBAAoB;AACpB,UAAQ;AAAA,IACN,8BAA8B,OAAO,yBAChC,kBAAkB;AAAA,EAGzB;AACF;AAGO,SAAS,0BAAgC;AAC9C,sBAAoB;AACtB;;;AChDA,OAAO,eAAe;AAQtB,IAAM,oBAAoB;AAcnB,IAAM,oBAAN,MAAqD;AAAA,EAM1D,YACmB,OACjB,WACA,UAAoC,CAAC,GACrC;AAHiB;AAIjB,UAAM,SAAS,QAAQ,IAAI,SAAS;AACpC,QAAI,CAAC,QAAQ;AACX,YAAM,IAAI;AAAA,QACR,4BAA4B,SAAS;AAAA,MACvC;AAAA,IACF;AACA,SAAK,SAAS,IAAI,UAAU,EAAE,OAAO,CAAC;AACtC,SAAK,WAAW,QAAQ,YAAY;AACpC,SAAK,kBACH,QAAQ,mBAAmB;AAC7B,SAAK,YAAY,QAAQ,aAAa;AAAA,EACxC;AAAA,EAfmB;AAAA,EANF;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAoBjB,WAAmB;AACjB,WAAO;AAAA,EACT;AAAA,EAEA,YAAoB;AAClB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAM,aAAa,KAAyD;AAC1E,UAAM,WAAW,MAAM,KAAK,OAAO,SAAS,OAAO;AAAA,MACjD,OAAO,KAAK;AAAA,MACZ,YAAY,KAAK;AAAA,MACjB,aAAa,IAAI;AAAA,MACjB,QAAQ,IAAI;AAAA,MACZ,UAAU,CAAC,EAAE,MAAM,QAAQ,SAAS,IAAI,KAAK,CAAC;AAAA,MAC9C,OAAO;AAAA,QACL;AAAA,UACE,MAAM,KAAK;AAAA,UACX,aAAa,KAAK;AAAA,UAClB,cAAc,IAAI;AAAA,QACpB;AAAA,MACF;AAAA,MACA,aAAa,EAAE,MAAM,QAAQ,MAAM,KAAK,SAAS;AAAA,IACnD,CAAC;AAMD,QAAI,SAAS,gBAAgB,cAAc;AACzC,YAAM,IAAI;AAAA,QACR,sCAAsC,KAAK,SAAS;AAAA,MACtD;AAAA,IACF;AAEA,UAAM,UAAU,SAAS,QAAQ;AAAA,MAC/B,CAAC,UAA2C,MAAM,SAAS;AAAA,IAC7D;AACA,QAAI,CAAC,SAAS;AACZ,YAAM,IAAI,MAAM,gDAAgD;AAAA,IAClE;AACA,WAAO;AAAA,MACL,MAAM,QAAQ;AAAA,MACd,OAAO;AAAA,QACL,aAAa,SAAS,MAAM;AAAA,QAC5B,cAAc,SAAS,MAAM;AAAA,MAC/B;AAAA,IACF;AAAA,EACF;AACF;;;AC/EO,SAAS,YAAY,SAA0B;AACpD,QAAM,UAAU,QACb,QAAQ,wBAAwB,EAAE,EAClC,QAAQ,cAAc,EAAE,EACxB,KAAK;AACR,MAAI;AACF,WAAO,KAAK,MAAM,OAAO;AAAA,EAC3B,QAAQ;AACN,UAAM,QAAQ,QAAQ,QAAQ,GAAG;AACjC,UAAM,MAAM,QAAQ,YAAY,GAAG;AACnC,QAAI,SAAS,KAAK,MAAM,OAAO;AAC7B,aAAO,KAAK,MAAM,QAAQ,MAAM,OAAO,MAAM,CAAC,CAAC;AAAA,IACjD;AACA,UAAM,IAAI,MAAM,6CAA6C;AAAA,EAC/D;AACF;AAQO,SAAS,eACd,QACyB;AACzB,QAAM,QAAQ,gBAAgB,MAAM;AACpC,QAAM,OAAO,CAAC,SAAwB;AACpC,QAAI,SAAS,QAAQ,OAAO,SAAS,SAAU;AAC/C,QAAI,MAAM,QAAQ,IAAI,GAAG;AACvB,iBAAW,QAAQ,KAAM,MAAK,IAAI;AAClC;AAAA,IACF;AACA,UAAM,MAAM;AACZ,WAAO,IAAI,WAAW;AACtB,WAAO,IAAI,aAAa;AACxB,UAAM,aAAa,IAAI,YAAY;AACnC,QAAI,cAAc,OAAO,eAAe,UAAU;AAChD,UAAI,UAAU,IAAI,OAAO,KAAK,UAAU;AAKxC,UAAI,sBAAsB,IAAI;AAC9B,iBAAW,QAAQ,OAAO,OAAO,UAAqC,GAAG;AACvE,aAAK,IAAI;AACT,YAAI,QAAQ,OAAO,SAAS,YAAY,CAAC,MAAM,QAAQ,IAAI,GAAG;AAC5D,gBAAM,IAAI;AACV,cAAI,OAAO,EAAE,MAAM,MAAM,YAAY,EAAE,MAAM,MAAM,QAAQ;AACzD,cAAE,MAAM,IAAI,CAAC,EAAE,MAAM,GAAG,MAAM;AAAA,UAChC;AAAA,QACF;AAAA,MACF;AAAA,IACF;AACA,SAAK,IAAI,OAAO,CAAC;AAAA,EACnB;AACA,OAAK,KAAK;AACV,SAAO;AACT;AAGO,SAAS,WAAW,OAAyB;AAClD,MAAI,UAAU,QAAQ,OAAO,UAAU,YAAY,MAAM,QAAQ,KAAK,GAAG;AACvE,WAAO;AAAA,EACT;AACA,SAAO,OAAO;AAAA,IACZ,OAAO,QAAQ,KAAgC,EAAE;AAAA,MAC/C,CAAC,CAAC,EAAE,CAAC,MAAM,MAAM;AAAA,IACnB;AAAA,EACF;AACF;AAOO,IAAM,uBAAN,MAAwD;AAAA,EAI7D,YACmB,SACA,OACjB,WACiB,SAA6B,QAAQ,IAAI,SAAS,GACnE,UAAuC,CAAC,GACxC;AALiB;AACA;AAEA;AAIjB,QAAI,CAAC,KAAK,UAAU,QAAQ,SAAS,gBAAgB,GAAG;AACtD,YAAM,IAAI;AAAA,QACR,yBAAyB,SAAS;AAAA,MACpC;AAAA,IACF;AACA,SAAK,aAAa,QAAQ,cAAc;AAAA,EAC1C;AAAA,EAbmB;AAAA,EACA;AAAA,EAEA;AAAA,EAPX,qBAAqB;AAAA,EACZ;AAAA,EAkBjB,WAAmB;AACjB,WAAO;AAAA,EACT;AAAA,EAEA,YAAoB;AAClB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAc,KAAK,MAAsD;AACvE,UAAM,WAAW,MAAM;AAAA,MACrB,GAAG,KAAK,QAAQ,QAAQ,OAAO,EAAE,CAAC;AAAA,MAClC;AAAA,QACE,QAAQ;AAAA,QACR,SAAS;AAAA,UACP,gBAAgB;AAAA,UAChB,GAAI,KAAK,SAAS,EAAE,eAAe,UAAU,KAAK,MAAM,GAAG,IAAI,CAAC;AAAA,QAClE;AAAA,QACA,MAAM,KAAK,UAAU,IAAI;AAAA,MAC3B;AAAA,IACF;AACA,UAAM,OAAQ,MAAM,SAAS,KAAK,EAAE,MAAM,OAAO,CAAC,EAAE;AACpD,QAAI,CAAC,SAAS,IAAI;AAChB,YAAM,UAAU,KAAK,OAAO,WAAW,QAAQ,SAAS,MAAM;AAC9D,YAAM,IAAI,MAAM,GAAG,OAAO,EAAE;AAAA,IAC9B;AACA,WAAO;AAAA,EACT;AAAA,EAEA,MAAM,aAAa,KAAyD;AAC1E,UAAM,OAAO;AAAA,MACX,OAAO,KAAK;AAAA,MACZ,aAAa,IAAI;AAAA,MACjB,UAAU;AAAA,QACR,EAAE,MAAM,UAAU,SAAS,IAAI,OAAO;AAAA,QACtC,EAAE,MAAM,QAAQ,SAAS,IAAI,KAAK;AAAA,MACpC;AAAA,IACF;AAEA,QAAI;AACJ,QAAI,KAAK,oBAAoB;AAC3B,UAAI;AACF,mBAAW,MAAM,KAAK,KAAK;AAAA,UACzB,GAAG;AAAA,UACH,iBAAiB;AAAA,YACf,MAAM;AAAA,YACN,aAAa;AAAA,cACX,MAAM,KAAK;AAAA,cACX,QAAQ;AAAA,cACR,QAAQ,eAAe,IAAI,MAAM;AAAA,YACnC;AAAA,UACF;AAAA,QACF,CAAC;AAAA,MACH,SAAS,GAAG;AACV,cAAM,UAAU,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC;AAGzD,YACE,CAAC,sCAAsC,KAAK,OAAO,KACnD,YAAY,YACZ;AACA,gBAAM;AAAA,QACR;AACA,aAAK,qBAAqB;AAC1B,mBAAW,MAAM,KAAK,mBAAmB,MAAM,GAAG;AAAA,MACpD;AAAA,IACF,OAAO;AACL,iBAAW,MAAM,KAAK,mBAAmB,MAAM,GAAG;AAAA,IACpD;AAEA,UAAM,UAAU,SAAS,UAAU,CAAC,GAAG,SAAS;AAChD,QAAI,CAAC,QAAS,OAAM,IAAI,MAAM,2BAA2B;AACzD,WAAO;AAAA,MACL,MAAM,WAAW,YAAY,OAAO,CAAC;AAAA,MACrC,OACE,SAAS,OAAO,iBAAiB,OAC7B;AAAA,QACE,aAAa,SAAS,MAAM,iBAAiB;AAAA,QAC7C,cAAc,SAAS,MAAM,qBAAqB;AAAA,MACpD,IACA;AAAA,IACR;AAAA,EACF;AAAA,EAEQ,mBACN,MACA,KACuB;AACvB,WAAO,KAAK,KAAK;AAAA,MACf,GAAG;AAAA,MACH,UAAU;AAAA,QACR;AAAA,UACE,MAAM;AAAA,UACN,SAAS,GAAG,IAAI,MAAM;AAAA;AAAA;AAAA,EAAwE,KAAK,UAAU,IAAI,MAAM,CAAC;AAAA,QAC1H;AAAA,QACA,EAAE,MAAM,QAAQ,SAAS,IAAI,KAAK;AAAA,MACpC;AAAA,MACA,iBAAiB,EAAE,MAAM,cAAc;AAAA,IACzC,CAAC;AAAA,EACH;AACF;;;AClNA,OAAO,WAAW;AAGX,IAAM,WAAmB,CAAC,KAAK,OAAO,CAAC,MAAM;AAClD,QAAM,CAAC,KAAK,GAAG,IAAI,IAAI;AACvB,MAAI,CAAC,KAAK;AACR,WAAO,QAAQ,QAAoB;AAAA,MACjC,MAAM;AAAA,MACN,QAAQ;AAAA,MACR,QAAQ;AAAA,MACR,UAAU;AAAA,MACV,YAAY;AAAA,IACd,CAAC;AAAA,EACH;AACA,SAAO,IAAI,QAAoB,CAAC,mBAAmB;AACjD,UAAM,QAAQ,MAAM,KAAK,MAAM;AAAA,MAC7B,KAAK,KAAK;AAAA,MACV,KAAK,EAAE,GAAG,QAAQ,KAAK,GAAI,KAAK,OAAO,CAAC,EAAG;AAAA,MAC3C,OAAO,CAAC,KAAK,SAAS,OAAO,SAAS,UAAU,QAAQ,MAAM;AAAA,IAChE,CAAC;AAED,QAAI,SAAS;AACb,QAAI,SAAS;AACb,QAAI,WAAW;AACf,QAAI,UAAU;AAEd,UAAM,YAAY,KAAK,aAAa;AACpC,UAAM,QAAQ,WAAW,MAAM;AAC7B,iBAAW;AACX,YAAM,KAAK;AAKX,aAAO,EAAE,MAAM,MAAM,QAAQ,QAAQ,UAAU,KAAK,CAAC;AAAA,IACvD,GAAG,SAAS;AAEZ,UAAM,SAAS,CAAC,WAA6B;AAC3C,UAAI,QAAS;AACb,gBAAU;AACV,mBAAa,KAAK;AAClB,qBAAe,MAAM;AAAA,IACvB;AAEA,QAAI,KAAK,SAAS,QAAQ,MAAM,OAAO;AAErC,YAAM,MAAM,GAAG,SAAS,MAAM;AAAA,MAAC,CAAC;AAChC,YAAM,MAAM,IAAI,KAAK,KAAK;AAAA,IAC5B;AAKA,UAAM,QAAQ,YAAY,MAAM;AAChC,UAAM,QAAQ,YAAY,MAAM;AAChC,UAAM,QAAQ,GAAG,QAAQ,CAAC,MAAe,UAAU,CAAE;AACrD,UAAM,QAAQ,GAAG,QAAQ,CAAC,MAAe,UAAU,CAAE;AACrD,UAAM;AAAA,MAAG;AAAA,MAAS,CAAC,MACjB,OAAO,EAAE,MAAM,MAAM,QAAQ,QAAQ,UAAU,YAAY,EAAE,QAAQ,CAAC;AAAA,IACxE;AACA,UAAM,GAAG,SAAS,CAAC,SAAS,OAAO,EAAE,MAAM,QAAQ,QAAQ,SAAS,CAAC,CAAC;AAAA,EACxE,CAAC;AACH;;;ACtDO,IAAM,oBAAN,MAAqD;AAAA,EAC1D,YACmB,OACA,UAAkB,UAClB,OAAe,UACf,YAAoB,MACrC;AAJiB;AACA;AACA;AACA;AAAA,EAChB;AAAA,EAJgB;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EAGnB,WAAmB;AACjB,WAAO;AAAA,EACT;AAAA,EAEA,YAAoB;AAClB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAM,aAAa,KAAyD;AAC1E,UAAM,SAAS;AAAA,MACb,IAAI;AAAA,MACJ;AAAA,MACA;AAAA,MACA;AAAA,MACA,KAAK,UAAU,IAAI,MAAM;AAAA,IAC3B,EAAE,KAAK,IAAI;AAIX,UAAM,SAAS,MAAM,KAAK;AAAA,MACxB;AAAA,QACE,KAAK;AAAA,QACL;AAAA,QACA;AAAA,QACA,IAAI;AAAA,QACJ;AAAA,QACA;AAAA,QACA;AAAA,QACA,KAAK;AAAA,MACP;AAAA,MACA,EAAE,WAAW,KAAK,WAAW,OAAO,OAAO;AAAA,IAC7C;AAEA,QAAI,OAAO,YAAY;AACrB,YAAM,IAAI;AAAA,QACR,iBAAiB,KAAK,OAAO,KAAK,OAAO,UAAU;AAAA,MACrD;AAAA,IACF;AACA,QAAI,OAAO,SAAU,OAAM,IAAI,MAAM,sBAAsB;AAC3D,QAAI,OAAO,SAAS,GAAG;AACrB,YAAM,IAAI;AAAA,QACR,qBAAqB,OAAO,IAAI,KAAK,OAAO,OAAO,KAAK,EAAE,MAAM,IAAI,CAAC;AAAA,MACvE;AAAA,IACF;AAWA,QAAI;AACJ,QAAI;AACF,gBAAU,KAAK,MAAM,OAAO,MAAM;AAAA,IACpC,QAAQ;AACN,YAAM,UAAU,OAAO,OAAO,KAAK,EAAE,QAAQ,QAAQ,GAAG,EAAE,MAAM,GAAG,GAAG;AACtE,YAAM,IAAI;AAAA,QACR,0DAA0D,WAAW,aAAa;AAAA,MACpF;AAAA,IACF;AACA,QAAI,OAAO,QAAQ,WAAW,UAAU;AACtC,YAAM,IAAI,MAAM,qCAAqC;AAAA,IACvD;AACA,WAAO,EAAE,MAAM,YAAY,QAAQ,MAAM,EAAE;AAAA,EAC7C;AACF;;;AC1EO,IAAM,eAAN,MAAgD;AAAA,EAKrD,YACmB,WACA,QAAQ,cACzB;AAFiB;AACA;AAEjB,QAAI,UAAU,WAAW,GAAG;AAC1B,YAAM,IAAI,MAAM,mDAAmD;AAAA,IACrE;AAAA,EACF;AAAA,EANmB;AAAA,EACA;AAAA,EANX,QAAQ;AAAA;AAAA,EAEA,WAAkC,CAAC;AAAA,EAWnD,WAAmB;AACjB,WAAO;AAAA,EACT;AAAA,EAEA,YAAoB;AAClB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,aAAa,KAAyD;AACpE,SAAK,SAAS,KAAK,GAAG;AACtB,UAAM,WAAW,KAAK,UAAU,KAAK,QAAQ,KAAK,UAAU,MAAM;AAClE,SAAK,SAAS;AACd,QAAI,WAAW,UAAU;AACvB,aAAO,QAAQ,OAAO,IAAI,MAAM,SAAS,KAAK,CAAC;AAAA,IACjD;AACA,WAAO,QAAQ,QAAQ;AAAA,MACrB,MAAM,SAAS;AAAA,MACf,OAAO,SAAS,SAAS,EAAE,aAAa,KAAK,cAAc,IAAI;AAAA,IACjE,CAAC;AAAA,EACH;AACF;AAGO,SAAS,YACd,OACA,YACA,YAIK,CAAC,GACa;AACnB,SAAO;AAAA,IACL,MAAM;AAAA,MACJ,OAAO,UAAU,SAAS;AAAA,MAC1B,UAAU,UAAU,YAAY;AAAA,MAChC;AAAA,MACA;AAAA,MACA,WAAW,UAAU,aAAa;AAAA,IACpC;AAAA,EACF;AACF;;;AC/DA,SAAS,kBAAkB;AAC3B,SAAS,YAAY,WAAW,cAAc,qBAAqB;AACnE,SAAS,YAAY;AAEd,SAAS,OAAO,MAAsB;AAC3C,SAAO,WAAW,QAAQ,EAAE,OAAO,MAAM,MAAM,EAAE,OAAO,KAAK;AAC/D;AAOO,SAAS,cAAc,OAAyB;AAIrD,SAAO,OAAO,MAAM,IAAI,CAAC,MAAM,GAAG,EAAE,MAAM,IAAI,CAAC,EAAE,EAAE,KAAK,GAAG,CAAC;AAC9D;AAEO,IAAM,YAAN,MAAmB;AAAA,EAIxB,YACmB,KACA,UAAmB,MAEnB,QAAgB,aACjC;AAJiB;AACA;AAEA;AAAA,EAChB;AAAA,EAJgB;AAAA,EACA;AAAA,EAEA;AAAA;AAAA,EANX,SAAS;AAAA,EASjB,IAAI,KAA4B;AAC9B,QAAI,CAAC,KAAK,QAAS,QAAO;AAC1B,UAAM,OAAO,KAAK,KAAK,KAAK,GAAG,GAAG,OAAO;AACzC,QAAI,CAAC,WAAW,IAAI,EAAG,QAAO;AAC9B,QAAI;AACF,aAAO,KAAK,MAAM,aAAa,MAAM,MAAM,CAAC;AAAA,IAC9C,QAAQ;AACN,aAAO;AAAA,IACT;AAAA,EACF;AAAA,EAEA,IAAI,KAAa,OAAgB;AAC/B,QAAI,CAAC,KAAK,QAAS;AAInB,QAAI;AACF,gBAAU,KAAK,KAAK,EAAE,WAAW,KAAK,CAAC;AACvC;AAAA,QACE,KAAK,KAAK,KAAK,GAAG,GAAG,OAAO;AAAA,QAC5B,KAAK,UAAU,OAAO,MAAM,CAAC;AAAA,MAC/B;AAAA,IACF,SAAS,GAAG;AACV,UAAI,CAAC,KAAK,QAAQ;AAChB,aAAK,SAAS;AACd,gBAAQ;AAAA,UACN,GAAG,KAAK,KAAK,kCAAkC,KAAK,GAAG,KACrD,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC,CAC3C;AAAA,QACF;AAAA,MACF;AAAA,IACF;AAAA,EACF;AACF;;;ACxBA,SAAS,mBAAmB;AAC5B,SAAS,eAAe;AACxB,SAAS,QAAAA,aAAY;AAcd,SAAS,8BAAsC;AACpD,SACE,QAAQ,IAAI,sBAAsB,KAClCC,MAAK,QAAQ,GAAG,wBAAwB,QAAQ;AAEpD;AAGO,IAAM,cAAc,CAAC,QAAQ,YAAY,SAAS;AAIlD,IAAM,kBAAkB,CAAC,QAAQ,GAAG,WAAW;AAuB/C,IAAM,eACX,kBAAkB;AAAA,EAChB,qBAAqB;AAAA,IACnB,KAAK;AAAA,IACL,WAAW;AAAA,IACX,SAAS;AAAA,IACT,MAAM;AAAA,IACN,OACE;AAAA,EAEJ;AAAA,EACA,cAAc;AAAA,IACZ,KAAK;AAAA,IACL,WAAW;AAAA,IACX,SAAS;AAAA,IACT,OACE;AAAA,IAEF,MAAM;AAAA,EACR;AAAA,EACA,cAAc;AAAA,IACZ,KAAK;AAAA,IACL,WAAW;AAAA,IACX,SAAS;AAAA,IACT,MAAM;AAAA,IACN,OACE;AAAA,EAEJ;AAAA;AAAA;AAAA,EAGA,eAAe;AAAA,IACb,KAAK;AAAA,IACL,WAAW;AAAA,IACX,SAAS;AAAA,IACT,OAAO;AAAA,EACT;AAAA,EACA,eAAe;AAAA,IACb,KAAK;AAAA,IACL,WAAW;AAAA,IACX,SAAS;AAAA,IACT,OAAO;AAAA,EACT;AAAA,EACA,eAAe;AAAA,IACb,KAAK;AAAA,IACL,WAAW;AAAA,IACX,SAAS;AAAA,IACT,OAAO;AAAA,EACT;AAAA,EACA,mBAAmB;AAAA,IACjB,KAAK;AAAA,IACL,WAAW;AAAA,IACX,SAAS;AAAA,IACT,OACE;AAAA,EACJ;AAAA,EACA,kBAAkB;AAAA,IAChB,KAAK;AAAA,IACL,WAAW;AAAA,IACX,SAAS;AAAA,IACT,OACE;AAAA,EAIJ;AACF,CAAC;AAEH,SAAS,kBACP,SACa;AACb,aAAW,SAAS,OAAO,OAAO,OAAO,EAAG,QAAO,OAAO,KAAK;AAC/D,SAAO,OAAO,OAAO,OAAO;AAC9B;AAGA,IAAM,aAAwC;AAAA,EAC5C,MAAM;AAAA,EACN,UAAU;AAAA,EACV,SAAS;AACX;AAEO,SAAS,gBAAgB,OAAuC;AACrE,SAAQ,gBAAsC,SAAS,KAAK;AAC9D;AAQA,IAAM,kBAAkB;AAUjB,SAAS,cAAc,aAAgC;AAC5D,MAAI,SAAoB;AACxB,aAAW,QAAQ,aAAa;AAC9B,UAAM,QAAQ,aAAa,WAAW,IAAI,CAAC;AAC3C,QAAI,MAAM,YAAY,mBAAmB,YAAa,UAAS;AAAA,EACjE;AAGA,SAAO;AACT;AAMO,SAAS,aAAa,MAAyB;AACpD,SAAO,WAAW,IAAI;AACxB;AAGO,SAAS,WAAW,MAAyB;AAClD,SAAO,aAAa,WAAW,IAAI,CAAC,EAAG;AACzC;AAWO,SAAS,qBAAqB,OAAuB;AAC1D,MAAI,gBAAgB,KAAK,GAAG;AAC1B,UAAM,IAAI;AAAA,MACR,oBAAoB,KAAK,8IAEE,WAAW,QAAQ;AAAA,IAChD;AAAA,EACF;AACA,QAAM,QAAQ,aAAa,KAAK;AAChC,MAAI,MAAO,QAAO,MAAM;AACxB,MAAI,iBAAiB,KAAK,EAAG,QAAO;AACpC,QAAM,IAAI;AAAA,IACR,4BAA4B,KAAK,sBAAsB,gBAAgB;AAAA,MACrE;AAAA,IACF,CAAC,uBAAuB,OAAO,KAAK,YAAY,EAAE;AAAA,MAChD;AAAA,IACF,CAAC;AAAA,EACH;AACF;AASO,SAAS,YAAY,OAAuB;AACjD,QAAM,MAAM,qBAAqB,KAAK,EAAE,MAAM,GAAG,EAAE,CAAC;AACpD,SAAO,IAAI,MAAM,OAAO,EAAE,IAAI;AAChC;AAQO,SAAS,iBAAiB,OAAe,UAA2B;AACzE,QAAM,OAAO,MAAM,QAAQ,YAAY,EAAE;AACzC,MAAI,KAAK,SAAS,QAAQ,EAAG,QAAO;AAGpC,QAAM,OAAO,SAAS,QAAQ,WAAW,EAAE;AAC3C,SAAO,IAAI,OAAO,GAAG,aAAa,IAAI,CAAC,2BAA2B,EAAE,KAAK,IAAI;AAC/E;AAEA,SAAS,aAAa,MAAsB;AAC1C,SAAO,KAAK,QAAQ,uBAAuB,MAAM;AACnD;AASO,SAAS,kBAAkB,OAAe,WAA4B;AAC3E,QAAM,WAAW,YAAY,KAAK;AAClC,SAAO,mBAAmB,SAAS,EAAE;AAAA,IACnC,CAAC,UAAU,CAAC,MAAM,SAAS,QAAQ,KAAK,iBAAiB,OAAO,QAAQ;AAAA,EAC1E;AACF;AAGO,SAAS,mBAAmB,WAA6B;AAC9D,MAAI;AACF,WAAO,YAAY,WAAW,EAAE,eAAe,KAAK,CAAC,EAClD,OAAO,CAAC,UAAU,MAAM,OAAO,CAAC,EAChC,IAAI,CAAC,UAAU,MAAM,IAAI;AAAA,EAC9B,QAAQ;AAEN,WAAO,CAAC;AAAA,EACV;AACF;AAYA,SAAS,iBAAiB,OAAwB;AAChD,SACE,sBAAsB,KAAK,KAAK,KAChC,gBAAgB,KAAK,KAAK,KAC1B,2BAA2B,KAAK,KAAK,KACrC,MAAM,SAAS,OAAO;AAE1B;;;ACxTA;AAAA,EACE,cAAAC;AAAA,EACA,aAAAC;AAAA,EACA;AAAA,EACA;AAAA,EACA,iBAAAC;AAAA,OACK;AACP,SAAS,WAAAC,gBAAe;AACxB,SAAS,QAAAC,aAAY;AACrB,SAAS,qBAAqB;AAM9B,IAAM,eAAe;AAMrB,IAAM,OAAO;AACb,IAAM,OAAO;AAOb,IAAM,qBAAqB;AAG3B,IAAM,eAAe;AAErB,IAAM,gBAAgB,qBAAqB;AA6BpC,SAAS,6BACd,MAA0C,QAAQ,KAC1C;AACR,SACE,IAAI,uBAAuB,KAC3BC,MAAKC,SAAQ,GAAG,wBAAwB,SAAS;AAErD;AAYO,SAAS,iBAAiB,GAAqB;AACpD,QAAM,OAAQ,GAAyC;AACvD,SAAO,SAAS,0BAA0B,SAAS;AACrD;AAEA,SAAS,SAAS,GAAoB;AACpC,SAAO,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC;AAClD;AAoBA,eAAsB,mBACpB,UAAiC,CAAC,GACV;AACxB,QAAM,MAAM,QAAQ,OAAO,QAAQ;AACnC,QAAM,YAAY,QAAQ,aAAa,6BAA6B,GAAG;AAEvE,QAAM,cACJ,QAAQ,gBAAgB,MAAwB,OAAO,gBAAgB;AACzE,MAAI;AACF,UAAM,YAAY;AAClB,WAAO,EAAE,OAAO,UAAU;AAAA,EAC5B,SAAS,GAAG;AAKV,QAAI,CAAC,iBAAiB,CAAC,GAAG;AACxB,aAAO;AAAA,QACL,OAAO;AAAA,QACP,QAAQ,mDAAmD,SAAS,CAAC,CAAC;AAAA,MACxE;AAAA,IACF;AAAA,EAEF;AACA,MAAIC,YAAWF,MAAK,WAAW,IAAI,CAAC,EAAG,QAAO,EAAE,OAAO,UAAU;AAEjE,OAAK,IAAI,2BAA2B,KAAK,QAAQ,IAAI;AACnD,WAAO;AAAA,MACL,OAAO;AAAA,MACP,QAAQ;AAAA,IACV;AAAA,EACF;AACA,SAAO,EAAE,OAAO,eAAe,UAAU;AAC3C;AAOA,IAAM,WAAW,oBAAI,IAA8B;AAEnD,IAAI,gBAAgB;AAGb,SAAS,sBAA4B;AAC1C,WAAS,MAAM;AACf,kBAAgB;AAClB;AASO,SAAS,mBACd,UAAiC,CAAC,GAChB;AAClB,QAAM,MAAM,QAAQ,OAAO,QAAQ;AACnC,QAAM,YAAY,QAAQ,aAAa,6BAA6B,GAAG;AAEvE,QAAM,WAAW,SAAS,IAAI,SAAS;AACvC,MAAI,SAAU,QAAO;AAErB,QAAM,UAAU,WAAW,WAAW,KAAK,OAAO;AAIlD,QAAM,UAAU,QAAQ,MAAM,CAAC,MAAe;AAC5C,QAAI,SAAS,IAAI,SAAS,MAAM,QAAS,UAAS,OAAO,SAAS;AAClE,UAAM;AAAA,EACR,CAAC;AACD,WAAS,IAAI,WAAW,OAAO;AAC/B,SAAO;AACT;AAEA,eAAe,WACb,WACA,KACA,SACkB;AAClB,QAAM,aACJ,QAAQ,eAAe,CAAC,QAAkC,OAAO;AACnE,QAAM,OAAOA,MAAK,WAAW,IAAI;AAEjC,MAAIE,YAAW,IAAI,EAAG,QAAO,WAAW,cAAc,IAAI,EAAE,IAAI;AAEhE,OAAK,IAAI,2BAA2B,KAAK,QAAQ,IAAI;AACnD,UAAM,IAAI;AAAA,MACR,iHACwC,YAAY,+DACI,SAAS;AAAA,IAEnE;AAAA,EACF;AAEA,EAAAC,WAAU,WAAW,EAAE,WAAW,KAAK,CAAC;AACxC,QAAM,SAAS,WAAW,YAAY;AAEpC,QAAID,YAAW,IAAI,EAAG;AACtB,mBAAe,SAAS;AACxB,UAAM,WAAW,WAAW,KAAK,OAAO;AAExC,IAAAE,eAAc,MAAM;AAAA,GAAqC,MAAM;AAAA,EACjE,CAAC;AAED,SAAO,WAAW,cAAc,IAAI,EAAE,IAAI;AAC5C;AAEA,eAAe,WACb,WACA,KACA,SACe;AAIf,QAAM,WAAWJ,MAAK,WAAW,cAAc;AAC/C,MAAI,CAACE,YAAW,QAAQ,GAAG;AACzB,IAAAE;AAAA,MACE;AAAA,MACA,GAAG,KAAK;AAAA,QACN;AAAA,UACE,MAAM;AAAA,UACN,SAAS;AAAA,UACT,SAAS;AAAA,UACT,aACE;AAAA,QACJ;AAAA,QACA;AAAA,QACA;AAAA,MACF,CAAC;AAAA;AAAA,MACD;AAAA,IACF;AAAA,EACF;AAEA,QAAM,OAAO,QAAQ,QAAQ;AAC7B,QAAM,SAAS,MAAM;AAAA,IACnB;AAAA,MACE;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,IACA,EAAE,WAAW,QAAQ,aAAa,oBAAoB,IAAI;AAAA,EAC5D;AAEA,MAAI,OAAO,cAAc,MAAM;AAC7B,UAAM,IAAI;AAAA,MACR,gDAAgD,OAAO,UAAU,sCAC5B,YAAY;AAAA,IAEnD;AAAA,EACF;AACA,MAAI,OAAO,UAAU;AACnB,UAAM,IAAI;AAAA,MACR,kCAAkC,SAAS,mHAE1B,YAAY;AAAA,IAC/B;AAAA,EACF;AACA,MAAI,OAAO,SAAS,GAAG;AACrB,UAAM,IAAI;AAAA,MACR,kCAAkC,SAAS,iBAAiB;AAAA,QAC1D,OAAO;AAAA,MACT,CAAC;AAAA,EAAO,KAAK,OAAO,UAAU,OAAO,MAAM,CAAC;AAAA,kCACP,YAAY;AAAA,IAEnD;AAAA,EACF;AACF;AAGA,SAAS,KAAK,QAAgB,QAAQ,IAAY;AAChD,SAAO,OAAO,QAAQ,EAAE,MAAM,OAAO,EAAE,MAAM,CAAC,KAAK,EAAE,KAAK,IAAI;AAChE;AAMA,SAAS,eAAe,WAAyB;AAC/C,MAAI,cAAe;AACnB,kBAAgB;AAChB,UAAQ;AAAA,IACN,sEAAiE,SAAS;AAAA,EAI5E;AACF;AASA,eAAe,SACb,WACA,IACe;AACf,QAAM,OAAOJ,MAAK,WAAW,IAAI;AACjC,QAAM,WAAW,KAAK,IAAI,IAAI;AAE9B,aAAS;AACP,QAAI;AACF,MAAAI,eAAc,MAAM,OAAO,QAAQ,GAAG,GAAG,EAAE,MAAM,KAAK,CAAC;AACvD;AAAA,IACF,SAAS,GAAG;AACV,UAAK,EAA4B,SAAS,SAAU,OAAM;AAC1D,UAAI,MAAM,IAAI,IAAI,eAAe;AAG/B,eAAO,MAAM,EAAE,OAAO,KAAK,CAAC;AAC5B;AAAA,MACF;AACA,UAAIF,YAAWF,MAAK,WAAW,IAAI,CAAC,EAAG;AACvC,UAAI,KAAK,IAAI,IAAI,UAAU;AACzB,cAAM,IAAI;AAAA,UACR,wEACK,SAAS,wCAAwC,IAAI;AAAA,QAC5D;AAAA,MACF;AACA,YAAM,MAAM,GAAG;AAAA,IACjB;AAAA,EACF;AAEA,MAAI;AACF,UAAM,GAAG;AAAA,EACX,UAAE;AACA,WAAO,MAAM,EAAE,OAAO,KAAK,CAAC;AAAA,EAC9B;AACF;AAGA,SAAS,MAAM,MAAsB;AACnC,MAAI;AACF,WAAO,KAAK,IAAI,IAAI,SAAS,IAAI,EAAE;AAAA,EACrC,QAAQ;AACN,WAAO,OAAO;AAAA,EAChB;AACF;AAEA,SAAS,MAAM,IAA2B;AACxC,SAAO,IAAI,QAAQ,CAAC,YAAY,WAAW,SAAS,EAAE,CAAC;AACzD;;;AC9QA,IAAM,eAAe,oBAAI,IAAuC;AAShE,eAAsB,qBAAoC;AACxD,QAAM,UAAU,CAAC,GAAG,aAAa,OAAO,CAAC;AACzC,eAAa,MAAM;AACnB,QAAM,QAAQ;AAAA,IACZ,QAAQ,IAAI,CAAC,MAAM,EAAE,KAAK,CAAC,MAAM,EAAE,QAAQ,CAAC,EAAE,MAAM,MAAM,MAAS,CAAC;AAAA,EACtE;AACF;AAEO,IAAM,mBAAN,MAAoD;AAAA,EAczD,YACmB,OACjB,UAAmC,CAAC,GACpC;AAFiB;AAGjB,QAAI,gBAAgB,KAAK,GAAG;AAC1B,YAAM,IAAI;AAAA,QACR,oBAAoB,KAAK,mFACmB,aAAa,UAAU,CAAC;AAAA,MAEtE;AAAA,IACF;AACA,SAAK,MAAM,qBAAqB,KAAK;AACrC,SAAK,UAAU,QAAQ,WAAW,oBAAoB;AACtD,SAAK,gBAAgB,QAAQ,iBAAiB;AAC9C,SAAK,YAAY,QAAQ;AACzB,SAAK,kBACH,QAAQ,mBAAmB,4BAA4B;AACzD,SAAK,WAAW,cAAc,CAAC,KAAK,iBAAiB,KAAK,GAAG,CAAC;AAAA,EAChE;AAAA,EAjBmB;AAAA,EAdF;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOA;AAAA,EAsBjB,WAAmB;AACjB,WAAO;AAAA,EACT;AAAA,EAEA,YAAoB;AAClB,WAAO,KAAK;AAAA,EACd;AAAA,EAEA,MAAM,aAAa,KAAyD;AAC1E,UAAM,QAAQ,MAAM,KAAK,KAAK;AAG9B,UAAM,UAAU,MAAM,MAAM,cAAc,gBAAgB,GAAG,CAAC;AAC9D,QAAI;AACF,YAAM,SAAS,MAAM,QAAQ,OAAO,IAAI,MAAM;AAAA,QAC5C,QAAQ,IAAI;AAAA,QACZ,aAAa,IAAI;AAAA,QACjB,eAAe,KAAK;AAAA,QACpB,GAAI,KAAK,aAAa,OAAO,EAAE,WAAW,KAAK,UAAU,IAAI,CAAC;AAAA,MAChE,CAAC;AAMD,UAAI,OAAO,eAAe,aAAa;AACrC,cAAM,IAAI;AAAA,UACR,sEACK,KAAK,aAAa,OAAO,gBAAgB,KAAK,SAAS,MAAM,EAAE;AAAA,QAEtE;AAAA,MACF;AACA,aAAO,EAAE,MAAM,YAAY,OAAO,IAAI,GAAG,OAAO,OAAO,MAAM;AAAA,IAC/D,UAAE;AACA,YAAM,QAAQ,QAAQ,EAAE,MAAM,MAAM,MAAS;AAAA,IAC/C;AAAA,EACF;AAAA,EAEQ,OAAkC;AACxC,UAAM,WAAW,aAAa,IAAI,KAAK,QAAQ;AAC/C,QAAI,SAAU,QAAO;AACrB,UAAM,WAAW,YAAY;AAC3B,YAAM,OAAO,MAAM,KAAK,QAAQ;AAAA,QAC9B,KAAK;AAAA,QACL,KAAK;AAAA,MACP;AACA,aAAO,KAAK,QAAQ,UAAU,IAAI;AAAA,IACpC,GAAG;AAKH,UAAM,UAAU,QAAQ,MAAM,CAAC,MAAe;AAC5C,UAAI,aAAa,IAAI,KAAK,QAAQ,MAAM,SAAS;AAC/C,qBAAa,OAAO,KAAK,QAAQ;AAAA,MACnC;AACA,YAAM;AAAA,IACR,CAAC;AACD,iBAAa,IAAI,KAAK,UAAU,OAAO;AACvC,WAAO;AAAA,EACT;AACF;AASA,SAAS,gBAAgB,KAAkC;AACzD,SAAO,GAAG,IAAI,MAAM;AAAA;AAAA;AAAA,EAAwE,KAAK;AAAA,IAC/F,IAAI;AAAA,EACN,CAAC;AACH;AAGA,IAAI;AAOG,SAAS,sBAAoC;AAClD,QAAM,OAAO;AAAA;AAAA;AAAA;AAAA;AAAA,IAKV,mBAAmB,iBAAiB,EAAE,MAAM,CAAC,MAAe;AAC3D,uBAAiB;AACjB,YAAM;AAAA,IACR,CAAC;AAAA;AACH,SAAO;AAAA,IACL,kBAAkB,CAAC,KAAK,cACtB,KAAK,EAAE,KAAK,CAAC,MAAM,EAAE,iBAAiB,KAAK,SAAS,CAAC;AAAA,IACvD,WAAW,CAAC,SAAS,KAAK,EAAE,KAAK,CAAC,MAAM,EAAE,UAAU,IAAI,CAAC;AAAA,IACzD,sBAAsB,MAAM,KAAK,EAAE,KAAK,CAAC,MAAM,EAAE,qBAAqB,CAAC;AAAA,EACzE;AACF;AAEA,eAAe,mBAA0C;AACvD,MAAI;AACJ,MAAI;AACF,UAAM,MAAM,OAAO,gBAAgB;AAAA,EACrC,SAAS,GAAG;AAKV,QAAI,CAAC,iBAAiB,CAAC,GAAG;AACxB,YAAM,IAAI;AAAA,QACR,mDACE,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC,CAC3C;AAAA,MAGF;AAAA,IACF;AAMA,UAAO,MAAM,mBAAmB;AAAA,EAClC;AAEA,QAAM,EAAE,UAAU,kBAAkB,kBAAkB,WAAW,IAAI;AACrE,QAAM,QAAQ,MAAM,SAAS;AAE7B,SAAO;AAAA;AAAA;AAAA,IAGL,kBAAkB,CAAC,KAAK,cAAc,iBAAiB,KAAK,EAAE,UAAU,CAAC;AAAA,IAEzE,MAAM,UAAU,MAAM;AACpB,YAAM,QAAQ,MAAM,MAAM,UAAU,EAAE,WAAW,KAAK,CAAC;AACvD,aAAO;AAAA,QACL,MAAM,cAAc,cAAc;AAChC,gBAAM,UAAU,MAAM,MAAM,cAAc;AAC1C,gBAAM,WAAW,QAAQ,YAAY;AACrC,gBAAM,UAAU,IAAI,iBAAiB;AAAA,YACnC,iBAAiB;AAAA,YACjB;AAAA,UACF,CAAC;AACD,iBAAO;AAAA,YACL,MAAM,OAAO,MAAM,SAAS;AAC1B,oBAAM,UAAU,MAAM,MAAM;AAAA,gBAC1B,QAAQ;AAAA,cAGV;AACA,oBAAM,SAAS,SAAS,WAAW,SAAS;AAC5C,oBAAM,SAAS,MAAM,QAAQ,eAAe,MAAM;AAAA,gBAChD;AAAA,gBACA,aAAa,QAAQ;AAAA,gBACrB,SAAS,EAAE,eAAe,QAAQ,cAAc;AAAA,gBAChD,GAAI,QAAQ,aAAa,OACrB,EAAE,WAAW,QAAQ,UAAU,IAC/B,CAAC;AAAA,cACP,CAAC;AAED,oBAAM,OAAO,WAAW,KAAK,SAAS,YAAY,MAAM;AACxD,qBAAO;AAAA,gBACL,MAAM,OAAO;AAAA,gBACb,YAAY,OAAO;AAAA,gBACnB,OAAO;AAAA,kBACL,aAAa,KAAK;AAAA,kBAClB,cAAc,KAAK;AAAA,gBACrB;AAAA,cACF;AAAA,YACF;AAAA,YACA,MAAM,UAAU;AACd,oBAAM,QAAQ,QAAQ;AAAA,YACxB;AAAA,UACF;AAAA,QACF;AAAA,QACA,MAAM,UAAU;AACd,gBAAM,MAAM,QAAQ;AAAA,QACtB;AAAA,MACF;AAAA,IACF;AAAA,IAEA,MAAM,uBAAuB;AAC3B,YAAM,EAAE,SAAS,IAAI,MAAM,OAAO,IAAS;AAG3C,YAAM,YAAY,SAAS,IAAI;AAC/B,UAAI;AACF,cAAM,OAAO,MAAM,MAAM,aAAa;AAKtC,eAAO,KAAK,IAAI,KAAK,MAAM,SAAS;AAAA,MACtC,QAAQ;AAEN,eAAO;AAAA,MACT;AAAA,IACF;AAAA,EACF;AACF;;;AC/UO,IAAM,kBAA2C;AAAA,EACtD;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEA,IAAM,kBAAyD;AAAA,EAC7D,WAAW;AAAA,EACX,QAAQ;AACV;AAqBA,SAAS,OAAO,UAA2C;AAEzD,UAAQ,QAAQ,IAAI,gBAAgB,QAAQ,CAAE,KAAK,QAAQ;AAC7D;AAWA,IAAM,YAAY,oBAAI,IAA8B;AAG7C,SAAS,sBAA4B;AAC1C,YAAU,MAAM;AAClB;AAEA,SAAS,eAAe,MAAsC;AAC5D,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,UAAU,KAAK,WAAW;AAChC,QAAM,SAAS,UAAU,IAAI,OAAO;AACpC,MAAI,OAAQ,QAAO;AACnB,QAAM,UAAU,KAAK,CAAC,SAAS,WAAW,GAAG,EAAE,WAAW,IAAO,CAAC,EAC/D,KAAK,CAAC,MAAM,EAAE,SAAS,KAAK,CAAC,EAAE,YAAY,EAAE,cAAc,IAAI,EAC/D,MAAM,MAAM,KAAK;AACpB,YAAU,IAAI,SAAS,OAAO;AAC9B,SAAO;AACT;AAEA,eAAe,cAAc,MAAoC;AAC/D,QAAM,WAAW,KAAK,gBAAgB,KAAK,UAAU;AACrD,MAAI,SAAU,QAAO,YAAY,QAAQ;AAMzC,QAAM,SAAS,MAAM,mBAAmB;AACxC,MAAI,OAAO,UAAU,WAAW;AAC9B,WAAO,EAAE,WAAW,OAAO,QAAQ,OAAO,OAAO;AAAA,EACnD;AACA,MAAI,OAAO,UAAU,eAAe;AAGlC,WAAO,EAAE,WAAW,KAAK;AAAA,EAC3B;AAGA,SAAO,YAAY,oBAAoB,CAAC;AAC1C;AAMA,SAAS,YAAY,SAAuC;AAC1D,SAAO,QAAQ,qBAAqB,EAAE;AAAA,IACpC,OAAO,EAAE,WAAW,KAAK;AAAA,IACzB,CAAC,OAAgB;AAAA,MACf,WAAW;AAAA,MACX,QACE,aAAa,SAAS,iBAAiB,KAAK,EAAE,OAAO,IACjD,2DACA,mCACE,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC,CAC3C;AAAA,IACR;AAAA,EACF;AACF;AAEA,eAAe,MACb,UACA,MACgB;AAChB,UAAQ,UAAU;AAAA,IAChB,KAAK;AACH,aAAO,OAAO,WAAW,IACrB,EAAE,WAAW,KAAK,IAClB;AAAA,QACE,WAAW;AAAA,QACX,QAAQ;AAAA,MACV;AAAA,IACN,KAAK;AAGH,aAAO,OAAO,QAAQ,KAAK,KAAK,UAC5B,EAAE,WAAW,KAAK,IAClB;AAAA,QACE,WAAW;AAAA,QACX,QAAQ;AAAA,MACV;AAAA,IACN,KAAK;AACH,aAAQ,MAAM,eAAe,IAAI,IAC7B,EAAE,WAAW,KAAK,IAClB;AAAA,QACE,WAAW;AAAA,QACX,QAAQ,mBAAmB,KAAK,WAAW,QAAQ;AAAA,MACrD;AAAA,IACN,KAAK;AACH,aAAO,cAAc,IAAI;AAAA,IAC3B;AACE,aAAO,EAAE,WAAW,OAAO,QAAQ,sBAAsB;AAAA,EAC7D;AACF;AAQA,eAAsB,mBACpB,OAAqB,CAAC,GACG;AACzB,QAAM,SAAS,MAAM,QAAQ;AAAA,IAC3B,gBAAgB,IAAI,CAAC,SAAS,MAAM,MAAM,IAAI,CAAC;AAAA,EACjD;AACA,SAAO,gBAAgB,OAAO,CAAC,GAAG,MAAM,OAAO,CAAC,EAAG,SAAS;AAC9D;AASA,eAAsB,eACpB,OAAqB,CAAC,GACC;AAQvB,QAAM,UAAoB,CAAC;AAC3B,aAAW,QAAQ,iBAAiB;AAClC,UAAM,SAAS,MAAM,MAAM,MAAM,IAAI;AACrC,QAAI,OAAO,WAAW;AACpB,mBAAa,MAAM,KAAK,aAAa,MAAM;AAC3C,aAAO;AAAA,IACT;AACA,YAAQ,KAAK,KAAK,KAAK,OAAO,EAAE,CAAC,WAAM,OAAO,MAAM,EAAE;AAAA,EACxD;AACA,QAAM,IAAI;AAAA,IACR;AAAA,EAA+C,QAAQ,KAAK,IAAI,CAAC;AAAA;AAAA,EAEnE;AACF;AAEA,IAAI,kBAAkB;AACtB,IAAI,iBAAiB;AAGd,SAAS,gCAAsC;AACpD,oBAAkB;AAClB,mBAAiB;AACnB;AAOA,SAAS,aAAa,UAAwB,iBAAgC;AAC5E,MAAI,gBAAiB;AACrB,oBAAkB;AAGlB,QAAM,UAAU,kBAAkB,oBAAoB;AACtD,UAAQ;AAAA,IACN,cAAc,OAAO,0BAAqB,QAAQ;AAAA,EAEpD;AACF;AAMO,SAAS,oBAAoB,OAAe,WAAyB;AAC1E,MAAI,eAAgB;AACpB,mBAAiB;AACjB,UAAQ;AAAA,IACN,eAAe,KAAK,6DACb,YAAY,KAAK,QAAQ,CAAC,CAAC;AAAA,EAEpC;AACF;;;AC7KO,IAAM,iBAA+C;AAAA,EAC1D,WAAW;AAAA,EACX,QAAQ;AAAA,EACR,cAAc;AAAA,EACd,MAAM;AAAA;AAAA;AAAA;AAAA,EAIN,aAAa;AACf;AAEA,IAAM,sBAA8C;AAAA,EAClD,WAAW;AAAA,EACX,QAAQ;AACV;AAEO,IAAM,0BAA0B;AAmBhC,SAAS,wBAAwB,MAAsC;AAC5E,MAAI,KAAK,YAAY,QAAQ,KAAK,aAAa,QAAQ;AACrD,UAAM,IAAI;AAAA,MACR,8NAGM,OAAO,KAAK,cAAc,EAAE,KAAK,IAAI,CAAC;AAAA,IAC9C;AAAA,EACF;AACA,QAAM,QAAQ,KAAK,SAAS,eAAe,KAAK,QAAQ,KAAK;AAC7D,MAAI,KAAK,aAAa,eAAe,gBAAgB,KAAK,GAAG;AAC3D,UAAM,IAAI;AAAA,MACR,oBAAoB,KAAK,qLAGP,aAAa,UAAU,CAAC;AAAA,IAC5C;AAAA,EACF;AACA,SAAO,EAAE,UAAU,KAAK,UAAU,MAAM;AAC1C;AASA,eAAsB,6BACpB,MAC2B;AAM3B,OACG,KAAK,YAAY,QAAQ,KAAK,aAAa,WAC5C,KAAK,SAAS,MACd;AACA,UAAM,IAAI;AAAA,MACR,UAAU,KAAK,KAAK,gHAEd,OAAO,KAAK,cAAc,EAAE,KAAK,IAAI,CAAC;AAAA,IAE9C;AAAA,EACF;AAGA,QAAM,WACJ,KAAK,YAAY,QAAQ,KAAK,aAAa,SACvC,MAAM,eAAe,IAAI,IACzB,KAAK;AACX,QAAM,WAAyB,EAAE,GAAG,MAAM,SAAS;AAEnD,QAAM,QAAQ,KAAK,SAAS,eAAe,QAAQ,KAAK;AACxD,MAAI,aAAa,eAAe,CAAC,gBAAgB,KAAK,GAAG;AACvD,WAAO,wBAAwB,QAAQ;AAAA,EACzC;AACA,QAAM,OACJ,UAAU,SAAS,MAAM,UAAU,gBAAgB,IAAI,CAAC,IAAI;AAC9D,SAAO,EAAE,UAAU,OAAO,aAAa,IAAI,EAAE;AAC/C;AAWA,SAAS,sBAAsB,MAAoB,OAAqB;AACtE,QAAM,QAAQ,aAAa,KAAK;AAChC,MAAI,CAAC,MAAO;AACZ,QAAM,YACJ,KAAK,UAAU,mBAAmB,4BAA4B;AAChE,MAAI,CAAC,kBAAkB,OAAO,SAAS,GAAG;AACxC,wBAAoB,OAAO,MAAM,SAAS;AAAA,EAC5C;AACF;AAQA,SAAS,gBAAgB,MAA8C;AACrE,SAAO,KAAK,gBAAgB,KAAK,UAAU;AAC7C;AAEA,eAAe,UAAU,SAAuD;AAC9E,QAAM,SAAS,WAAW,oBAAoB;AAC9C,SAAO,cAAc,MAAM,OAAO,qBAAqB,CAAC;AAC1D;AAEO,SAAS,aAAa,MAAuC;AAIlE,wBAAsB;AACtB,QAAM,EAAE,MAAM,IAAI,wBAAwB,IAAI;AAE9C,UAAQ,KAAK,UAAU;AAAA,IACrB,KAAK;AACH,aAAO,IAAI;AAAA,QACT;AAAA,QACA,KAAK,aAAa,oBAAoB,WAAW;AAAA,QACjD,KAAK,aAAa,CAAC;AAAA,MACrB;AAAA,IACF,KAAK;AACH,aAAO,IAAI;AAAA,QACT,KAAK,WAAW;AAAA,QAChB;AAAA,QACA,KAAK,aAAa,oBAAoB,QAAQ;AAAA,QAC9C;AAAA,QACA,KAAK,UAAU,CAAC;AAAA,MAClB;AAAA,IACF,KAAK;AACH,aAAO,IAAI;AAAA,QACT;AAAA,QACA,KAAK,WAAW;AAAA,QAChB,KAAK;AAAA,QACL,KAAK;AAAA,MACP;AAAA,IACF,KAAK;AAEH,aAAO,IAAI,aAAa,KAAK,iBAAiB,CAAC,EAAE,MAAM,CAAC,EAAE,CAAC,GAAG,KAAK;AAAA,IACrE,KAAK;AACH,aAAO,IAAI,iBAAiB,OAAO;AAAA,QACjC,GAAI,KAAK,YAAY,CAAC;AAAA,QACtB,GAAI,KAAK,eAAe,EAAE,SAAS,KAAK,aAAa,IAAI,CAAC;AAAA,MAC5D,CAAC;AAAA,IACH;AACE,YAAM,IAAI;AAAA,QACR,qBAAqB,OAAO,KAAK,QAAQ,CAAC,iBAAiB,OAAO;AAAA,UAChE;AAAA,QACF,EAAE,KAAK,IAAI,CAAC;AAAA,MACd;AAAA,EACJ;AACF;AASA,eAAsB,kBACpB,MAC4B;AAG5B,QAAM,EAAE,UAAU,MAAM,IAAI,MAAM,6BAA6B,IAAI;AACnE,MAAI,aAAa,YAAa,uBAAsB,MAAM,KAAK;AAC/D,SAAO,aAAa,EAAE,GAAG,MAAM,UAAU,MAAM,CAAC;AAClD;;;ACtQA,SAAS,UAAAK,SAAQ,YAAAC,iBAAgB;AACjC,SAAS,QAAAC,aAAY;AAqCrB,eAAsB,iBACpB,UAAmC,CAAC,GACH;AACjC,QAAM,YAAY,QAAQ,aAAa,4BAA4B;AACnE,QAAM,SAAS,QAAQ,UAAU;AAGjC,QAAM,SAAS,QAAQ,QAAQ,IAAI,WAAW;AAE9C,MAAI,CAAC,OAAQ,OAAM,mBAAmB;AAEtC,QAAM,QAA4B,CAAC;AACnC,aAAW,SAAS,mBAAmB,SAAS,GAAG;AAGjD,QAAI,CAAC,YAAY,KAAK,EAAG;AACzB,QAAI,UAAU,CAAC,OAAO,KAAK,CAAC,SAAS,iBAAiB,OAAO,IAAI,CAAC,GAAG;AACnE;AAAA,IACF;AACA,UAAM,OAAOC,MAAK,WAAW,KAAK;AAClC,UAAM,YAAY,OAAO,IAAI;AAC7B,QAAI,cAAc,OAAW;AAC7B,QAAI,CAAC,QAAQ;AACX,UAAI;AACF,QAAAC,QAAO,IAAI;AAAA,MACb,QAAQ;AAGN;AAAA,MACF;AAAA,IACF;AACA,UAAM,KAAK,EAAE,MAAM,UAAU,CAAC;AAAA,EAChC;AAEA,SAAO;AAAA,IACL;AAAA,IACA,YAAY,MAAM,OAAO,CAAC,OAAO,SAAS,QAAQ,KAAK,WAAW,CAAC;AAAA,IACnE;AAAA,IACA;AAAA,EACF;AACF;AAEA,SAAS,YAAY,OAAwB;AAC3C,SAAO,MAAM,SAAS,OAAO,KAAK,MAAM,SAAS,aAAa;AAChE;AAEA,SAAS,OAAO,MAAkC;AAChD,MAAI;AACF,WAAOC,UAAS,IAAI,EAAE;AAAA,EACxB,QAAQ;AACN,WAAO;AAAA,EACT;AACF;;;AC3FA,SAAS,eAAe;AAmCxB,IAAM,iBAAiB,oBAAI,QAAkC;AAGtD,SAAS,aACd,QACkB;AAClB,QAAM,SAAS,eAAe,IAAI,MAAM;AACxC,MAAI,OAAQ,QAAO;AAOnB,QAAM,WAAW,IAAI,QAAQ,EAAE,WAAW,KAAK,CAAC,EAAE,QAAQ,MAAM;AAChE,iBAAe,IAAI,QAAQ,QAAQ;AACnC,SAAO;AACT;AAEA,eAAsB,sBACpB,SAC0B;AAG1B,wBAAsB;AACtB,QAAM;AAAA,IACJ;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA,cAAc;AAAA,IACd,WAAW;AAAA,EACb,IAAI;AACJ,QAAM,WAAW,QAAQ,YAAY,aAAa,MAAM;AAExD,QAAM,QAAQ,KAAK,IAAI;AACvB,QAAM,OAAO;AAAA,IACX,UAAU,SAAS,SAAS;AAAA,IAC5B,OAAO,SAAS,UAAU;AAAA,IAC1B,QAAQ;AAAA,EACV;AAEA,MAAI,YAAY;AAChB,WAAS,UAAU,GAAG,UAAU,UAAU,WAAW;AACnD,QAAI;AACF,YAAM,WAAW,MAAM,SAAS,aAAa;AAAA,QAC3C;AAAA,QACA;AAAA,QACA;AAAA,QACA;AAAA,MACF,CAAC;AACD,UAAI,SAAS,SAAS,IAAI,GAAG;AAC3B,eAAO;AAAA,UACL,GAAG;AAAA,UACH,QAAQ,SAAS;AAAA,UACjB,OAAO,SAAS;AAAA,UAChB,YAAY,KAAK,IAAI,IAAI;AAAA,QAC3B;AAAA,MACF;AACA,kBAAY,uCAAuC,SAAS,UAAU,CAAC,GACpE,IAAI,CAAC,MAAM,GAAG,EAAE,YAAY,IAAI,EAAE,OAAO,EAAE,EAC3C,KAAK,IAAI,CAAC;AAAA,IACf,SAAS,GAAG;AACV,kBAAY,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC;AAAA,IACvD;AAAA,EACF;AACA,SAAO,EAAE,GAAG,MAAM,OAAO,WAAW,YAAY,KAAK,IAAI,IAAI,MAAM;AACrE;;;AC9FO,IAAM,cAAuC;AAAA,EAClD,qBAAqB,EAAE,cAAc,GAAG,eAAe,GAAG;AAAA,EAC1D,qBAAqB,EAAE,cAAc,GAAG,eAAe,GAAG;AAAA,EAC1D,oBAAoB,EAAE,cAAc,GAAG,eAAe,EAAE;AAAA,EACxD,mBAAmB,EAAE,cAAc,IAAI,eAAe,GAAG;AAAA,EACzD,eAAe,EAAE,cAAc,MAAM,eAAe,IAAI;AAAA,EACxD,UAAU,EAAE,cAAc,KAAK,eAAe,GAAG;AACnD;AAEO,SAAS,WACd,OACA,UACqB;AACrB,MAAI,SAAU,QAAO;AACrB,MAAI,YAAY,KAAK,EAAG,QAAO,YAAY,KAAK;AAGhD,QAAM,OAAO,OAAO,KAAK,WAAW,EACjC,OAAO,CAAC,MAAM,MAAM,WAAW,CAAC,CAAC,EACjC,KAAK,CAAC,GAAG,MAAM,EAAE,SAAS,EAAE,MAAM,EAAE,CAAC;AACxC,SAAO,OAAO,YAAY,IAAI,IAAI;AACpC;AAEO,SAAS,YACd,OACA,SACQ;AACR,MAAI,CAAC,SAAS,CAAC,QAAS,QAAO;AAC/B,SACG,MAAM,cAAc,MAAa,QAAQ,eACzC,MAAM,eAAe,MAAa,QAAQ;AAE/C;AAGO,SAAS,WACd,MACA,SACQ;AACR,MAAI,CAAC,QAAS,QAAO;AACrB,MAAI,MAAM;AACV,aAAW,OAAO,MAAM;AACtB,QAAI,CAAC,IAAI,SAAS,IAAI,OAAQ;AAC9B,WAAO,YAAY,IAAI,OAAO,OAAO;AAAA,EACvC;AACA,SAAO;AACT;;;AC9DA;AAAA,EACE,SAAW;AAAA,EACX,KAAO;AAAA,EACP,OAAS;AAAA,EACT,MAAQ;AAAA,EACR,UAAY,CAAC,SAAS,YAAY,SAAS,cAAc,WAAW;AAAA,EACpE,YAAc;AAAA,IACZ,OAAS;AAAA,MACP,MAAQ;AAAA,MACR,aAAe;AAAA,IACjB;AAAA,IACA,UAAY;AAAA,MACV,MAAQ;AAAA,MACR,aAAe;AAAA,IACjB;AAAA,IACA,OAAS;AAAA,MACP,MAAQ,CAAC,QAAQ,QAAQ,SAAS;AAAA,MAClC,aAAe;AAAA,IACjB;AAAA,IACA,YAAc;AAAA,MACZ,MAAQ;AAAA,MACR,SAAW;AAAA,MACX,SAAW;AAAA,MACX,aAAe;AAAA,IACjB;AAAA,IACA,WAAa;AAAA,MACX,MAAQ;AAAA,MACR,aAAe;AAAA,IACjB;AAAA,EACF;AAAA,EACA,sBAAwB;AAC1B;;;ACtBO,IAAM,iBAAiB;;;ACFvB,SAAS,iBACd,MAC+B;AAC/B,QAAM,QAAQ,EAAE,MAAM,GAAG,MAAM,GAAG,SAAS,GAAG,OAAO,EAAE;AACvD,MAAI,gBAAgB;AACpB,MAAI,kBAAkB;AACtB,aAAW,OAAO,MAAM;AACtB,QAAI,CAAC,IAAI,SAAS;AAChB,YAAM,SAAS;AACf;AAAA,IACF;AACA,UAAM,IAAI,QAAQ,KAAK,KAAK;AAC5B,qBAAiB,IAAI,QAAQ;AAC7B,uBAAmB;AAAA,EACrB;AAEA,QAAM,YAAY,MAAM;AACxB,QAAM,YAAY,MAAM,OAAO,MAAM;AAErC,QAAM,UAAiB,YAAY,YAAY,SAAS;AACxD,QAAM,SAAS,YAAY;AAC3B,QAAM,YAAY,SAAS,IAAI,KAAK,IAAI,WAAW,SAAS,IAAI,SAAS;AACzE,QAAM,iBACJ,kBAAkB,IAAI,gBAAgB,kBAAkB;AAE1D,SAAO,EAAE,MAAM,OAAO,SAAS,WAAW,eAAe;AAC3D;;;ACtBO,IAAM,gBAAgC,EAAE,UAAU,KAAK,UAAU,IAAI;AAErE,SAAS,QACd,WACA,aAA6B,eACvB;AACN,QAAM,EAAE,OAAO,eAAe,IAAI;AAClC,QAAM,gBACJ,MAAM,OAAO,KACb,MAAM,SAAS,KACf,MAAM,YAAY,KAClB,MAAM,UAAU;AAClB,QAAM,gBACJ,MAAM,SAAS,KAAK,MAAM,UAAU,KAAK,MAAM,OAAO,MAAM,UAAU;AAExE,MAAI,iBAAiB,kBAAkB,WAAW,SAAU,QAAO;AACnE,MAAI,iBAAiB,kBAAkB,WAAW,SAAU,QAAO;AACnE,SAAO;AACT;;;ACcA,IAAI,oBAAoB;AAMxB,eAAsB,YACpB,SACqB;AACrB,QAAM;AAAA,IACJ;AAAA,IACA;AAAA,IACA;AAAA,IACA,MAAM,WAAW;AAAA,IACjB,cAAc;AAAA,IACd,SAAS;AAAA,IACT;AAAA,IACA;AAAA,IACA,QAAQ;AAAA,EACV,IAAI;AAEJ,MAAI,cAAc,KAAK,CAAC,mBAAmB;AACzC,wBAAoB;AACpB,YAAQ;AAAA,MACN,GAAG,KAAK,0BAA0B,WAAW;AAAA,IAC/C;AAAA,EACF;AAEA,MAAI,SAAS,UAAU;AACrB,UAAM,MAAM,MAAM,IAAI,QAAQ;AAK9B,QAAI,MAAM,QAAQ,GAAG,EAAG,QAAO,IAAI,IAAI,CAAC,OAAO,EAAE,GAAG,GAAG,QAAQ,KAAK,EAAE;AAAA,EACxE;AAGA,QAAM,WAAW,aAAa,MAAM;AAEpC,QAAM,UAAsB,CAAC;AAC7B,WAAS,IAAI,GAAG,IAAI,UAAU,KAAK;AACjC,UAAM,MAAM,MAAM,sBAAoC;AAAA,MACpD;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,CAAC;AAGD,UAAM,EAAE,QAAQ,GAAG,KAAK,IAAI;AAC5B,YAAQ,KAAK,WAAW,SAAY,OAAO,EAAE,GAAG,MAAM,SAAS,OAAO,CAAC;AAAA,EACzE;AAEA,MAAI,SAAS,SAAU,OAAM,IAAI,UAAU,OAAO;AAClD,SAAO;AACT;AAGA,eAAsB,MACpB,SAC0B;AAC1B,QAAM,OAAO,MAAM,YAAY,OAAO;AACtC,QAAM,OAAO,iBAAiB,IAAI;AAClC,SAAO,EAAE,GAAG,MAAM,MAAM,QAAQ,MAAM,QAAQ,SAAS,aAAa,EAAE;AACxE;AAGO,SAAS,0BAAgC;AAC9C,sBAAoB;AACtB;","names":["join","join","existsSync","mkdirSync","writeFileSync","homedir","join","join","homedir","existsSync","mkdirSync","writeFileSync","rmSync","statSync","join","join","rmSync","statSync"]}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@hawkeyexl/inference",
-  "version": "0.2.0",
+  "version": "0.3.0",
   "description": "Shared TypeScript LLM inference layer: schema-constrained completion across Anthropic, OpenAI-compatible, Claude CLI, and in-process local llama.cpp providers, with caching, cost accounting, and an LLM-as-judge ensemble.",
   "type": "module",
   "main": "dist/index.js",