@latimer-woods-tech/llm 0.4.1 → 0.4.3

package/CHANGELOG.md

@@ -1,5 +1,54 @@
 # Changelog
 
+## 0.4.3 — 2026-06-03
+
+### Added — tool-calling for OpenAI-style providers (Phase 1b: Grok, DeepSeek)
+
+- **Grok and DeepSeek now support tool-calling.** Requests carry OpenAI-format
+  `tools` + `tool_choice`; responses parse `tool_calls` + `finish_reason` into
+  the same normalized `LLMResult.toolCalls` / `stopReason` as Anthropic.
+- Anthropic-shaped tool blocks are converted to OpenAI wire format:
+  `tool_use` → an assistant message with `tool_calls`; `tool_result` → a
+  standalone `tool` message keyed by `tool_call_id`.
+- **`TOOL_CAPABLE_PROVIDERS`** widened to `anthropic, grok, deepseek`, so the
+  `fast` (Grok→Haiku) and `workbench` (DeepSeek) tiers support tool loops while
+  still failing closed for `verifier` (Groq Llama).
+- Malformed tool-call argument JSON is tolerated (billed as `{}`, never throws).
+
+### Not yet
+
+- **Gemini** tool-calling is deferred to a focused follow-up — its
+  `tool_use_id`↔function-name correlation and schema constraints need dedicated
+  handling. Gemini stays out of `TOOL_CAPABLE_PROVIDERS` until then.
+
+---
+
+## 0.4.2 — 2026-06-03
+
+### Added — tool-calling (Agent Runtime Phase 1a; Anthropic)
+
+- **`LLMOptions.tools`** (`LLMTool[]`) and **`LLMOptions.toolChoice`**
+  (`'auto' | 'none' | { name }`). When `tools` is set, routing **fails closed**
+  to tool-capable providers — failover never silently falls back to one that
+  can't honour the tool schema (1a: Anthropic only; others land in 1b).
+- **`LLMResult.toolCalls`** (`LLMToolCall[]`) and **`LLMResult.stopReason`**
+  (`'end' | 'tool_use' | 'max_tokens' | 'other'`), normalized across providers.
+- **`LLMMessage.content`** now accepts `string | LLMContentBlock[]` — structured
+  `text` / `tool_use` / `tool_result` blocks for multi-turn tool conversations.
+  Backwards-compatible: existing `string` content is unchanged; providers without
+  tool support receive the text projection.
+- New exported types: `LLMTool`, `LLMToolCall`, `LLMContentBlock`.
+
+### Notes
+
+- A `tool_use` response with no text content is no longer treated as an empty
+  (failed) completion.
+- Anthropic `tool_use` blocks pass through unchanged (the block shapes mirror the
+  Messages wire format). Other providers' tool formats are normalized in 1a's
+  follow-ups (1b: Grok/DeepSeek/Gemini; 1c: streaming tool-call accumulation).
+
+---
+
 ## 0.4.1 — 2026-06-03
 
 ### Added (no breaking changes)

package/dist/index.d.mts

@@ -1,12 +1,56 @@
 import { FactoryResponse } from '@latimer-woods-tech/errors';
 import { Logger } from '@latimer-woods-tech/logger';
 
+/**
+ * A tool the model may call. `parameters` is a JSON Schema object describing
+ * the tool's input. Provider-agnostic; normalized per provider at request time.
+ */
+interface LLMTool {
+    name: string;
+    description?: string;
+    /** JSON Schema for the tool's input arguments. */
+    parameters: Record<string, unknown>;
+}
+/**
+ * A tool invocation requested by the model, normalized across providers.
+ */
+interface LLMToolCall {
+    /** Provider-assigned call id; echo it back in the matching tool_result. */
+    id: string;
+    name: string;
+    /** Parsed argument object the model passed to the tool. */
+    arguments: Record<string, unknown>;
+}
+/**
+ * Structured content block for tool-calling conversations. The field shapes
+ * mirror the Anthropic Messages wire format so they pass through unchanged.
+ */
+type LLMContentBlock = {
+    type: 'text';
+    text: string;
+} | {
+    type: 'tool_use';
+    id: string;
+    name: string;
+    input: Record<string, unknown>;
+} | {
+    type: 'tool_result';
+    tool_use_id: string;
+    content: string;
+    is_error?: boolean;
+};
 /**
  * Single chat message exchanged with an LLM provider.
+ *
+ * `content` is a plain string in the common case. For tool-calling
+ * conversations it may be an array of {@link LLMContentBlock}s (e.g. an
+ * assistant turn carrying `tool_use` blocks, or a user turn carrying
+ * `tool_result` blocks). Providers that don't support tool-calling receive
+ * the text projection of the content (see `contentToText`).
  */
 interface LLMMessage {
     role: 'user' | 'assistant' | 'system';
-    content: string;
+    content: string | LLMContentBlock[];
 }
 /**
  * Quality tier selected by the caller. Routing is workload-split:
@@ -74,6 +118,19 @@ interface LLMOptions {
      * is emitted after every successful completion. Errors are swallowed.
      */
     ledger?: LLMRecordContext;
+    /**
+     * Tools the model may call. When present, routing **fails closed** to
+     * tool-capable providers — failover never falls back to a provider that
+     * can't honour the tool schema. See {@link LLMResult.toolCalls}.
+     */
+    tools?: LLMTool[];
+    /**
+     * Tool-selection policy. `'auto'` (default when `tools` is set) lets the
+     * model decide; `'none'` forbids tool use; `{ name }` forces a specific tool.
+     */
+    toolChoice?: 'auto' | 'none' | {
+        name: string;
+    };
 }
 /**
  * Provider that produced an LLM response.
@@ -98,6 +155,16 @@ interface LLMResult {
     attempts: number;
     /** Monotonic request id from AI Gateway, if present in headers. */
     gatewayRequestId?: string;
+    /**
+     * Why generation stopped, normalized across providers. `'tool_use'` means
+     * the model is requesting one or more tool calls (see {@link toolCalls}).
+     */
+    stopReason?: 'end' | 'tool_use' | 'max_tokens' | 'other';
+    /**
+     * Tool calls the model requested, normalized across providers. Present
+     * (non-empty) when `stopReason === 'tool_use'`.
+     */
+    toolCalls?: LLMToolCall[];
 }
 /**
  * Environment bindings required by {@link complete}.
@@ -309,4 +376,4 @@ declare function completionStream(messages: LLMMessage[], env: LLMEnv, opts?: LL
  */
 declare function assertGrounding(response: string, sources: string[]): boolean;
 
-export { BASE_BACKOFF_MS, type CostKvStore, type LLMDeps, type LLMEnv, type LLMMessage, type LLMOptions, type LLMProvider, type LLMRecordContext, type LLMRecordRow, type LLMResult, type LLMTier, MODELS, MODEL_PRICE_PER_1M, PROVIDER_COOLDOWN_MS, assertGrounding, clearProviderCooldown, complete, completionStream, isProviderCoolingDown, markProviderCoolingDown };
+export { BASE_BACKOFF_MS, type CostKvStore, type LLMContentBlock, type LLMDeps, type LLMEnv, type LLMMessage, type LLMOptions, type LLMProvider, type LLMRecordContext, type LLMRecordRow, type LLMResult, type LLMTier, type LLMTool, type LLMToolCall, MODELS, MODEL_PRICE_PER_1M, PROVIDER_COOLDOWN_MS, assertGrounding, clearProviderCooldown, complete, completionStream, isProviderCoolingDown, markProviderCoolingDown };

package/dist/index.mjs

@@ -5,6 +5,15 @@ import {
   ValidationError,
   toErrorResponse
 } from "@latimer-woods-tech/errors";
+function contentToText(content) {
+  if (typeof content === "string") return content;
+  return content.map((b) => b.type === "text" ? b.text : b.type === "tool_result" ? b.content : "").join("");
+}
+function systemText(opts, messages) {
+  if (opts.system !== void 0) return opts.system;
+  const c = messages.find((m) => m.role === "system")?.content;
+  return c === void 0 ? void 0 : contentToText(c);
+}
 var MODELS = {
   anthropic: {
     fast: "claude-haiku-4-20250514",
@@ -102,7 +111,7 @@ function isRetryableForBackoff(status) {
 }
 function estimateTokens(messages, system) {
   let chars = system?.length ?? 0;
-  for (const m of messages) chars += m.content.length;
+  for (const m of messages) chars += contentToText(m.content).length;
   return Math.ceil(chars / 4);
 }
 function sleep(ms, signal) {
@@ -123,7 +132,7 @@ function computeBackoffMs(attempt) {
   return Math.min(BACKOFF_BASE_MS * Math.pow(2, attempt) + jitter, BACKOFF_CAP_MS);
 }
 function buildAnthropicRequest(model, messages, opts, env, streaming = false) {
-  const sys = opts.system ?? messages.find((m) => m.role === "system")?.content;
+  const sys = systemText(opts, messages);
   const filtered = messages.filter((m) => m.role !== "system");
   const body = {
     model,
@@ -138,6 +147,15 @@ function buildAnthropicRequest(model, messages, opts, env, streaming = false) {
     const cache = opts.promptCache ?? sys.length >= 4096;
     body.system = cache ? [{ type: "text", text: sys, cache_control: { type: "ephemeral" } }] : sys;
   }
+  if (opts.tools && opts.tools.length > 0) {
+    body.tools = opts.tools.map((t) => ({
+      name: t.name,
+      description: t.description ?? "",
+      input_schema: t.parameters
+    }));
+    const tc = opts.toolChoice ?? "auto";
+    body.tool_choice = tc === "auto" ? { type: "auto" } : tc === "none" ? { type: "none" } : { type: "tool", name: tc.name };
+  }
   return {
     url: `${env.AI_GATEWAY_BASE_URL}/anthropic/v1/messages`,
     headers: {
@@ -150,10 +168,10 @@ function buildAnthropicRequest(model, messages, opts, env, streaming = false) {
   };
 }
 function buildGeminiRequest(model, messages, opts, env) {
-  const sys = opts.system ?? messages.find((m) => m.role === "system")?.content;
+  const sys = systemText(opts, messages);
   const contents = messages.filter((m) => m.role !== "system").map((m) => ({
     role: m.role === "assistant" ? "model" : "user",
-    parts: [{ text: m.content }]
+    parts: [{ text: contentToText(m.content) }]
   }));
   const body = {
     contents,
@@ -175,11 +193,52 @@ function buildGeminiRequest(model, messages, opts, env) {
     body: JSON.stringify(body)
   };
 }
+function toOpenAiMessages(messages, sys) {
+  const out = [];
+  if (sys) out.push({ role: "system", content: sys });
+  for (const m of messages) {
+    if (m.role === "system") continue;
+    if (typeof m.content === "string") {
+      out.push({ role: m.role, content: m.content });
+      continue;
+    }
+    let text = "";
+    const toolCalls = [];
+    const results = [];
+    for (const b of m.content) {
+      if (b.type === "text") text += b.text;
+      else if (b.type === "tool_use")
+        toolCalls.push({ id: b.id, type: "function", function: { name: b.name, arguments: JSON.stringify(b.input) } });
+      else if (b.type === "tool_result") results.push({ tool_use_id: b.tool_use_id, content: b.content });
+    }
+    if (results.length > 0) {
+      for (const r of results) out.push({ role: "tool", tool_call_id: r.tool_use_id, content: r.content });
+      if (text) out.push({ role: "user", content: text });
+    } else if (toolCalls.length > 0) {
+      out.push({ role: "assistant", content: text || null, tool_calls: toolCalls });
+    } else {
+      out.push({ role: m.role, content: text });
+    }
+  }
+  return out;
+}
+function openAiTools(opts) {
+  if (!opts.tools || opts.tools.length === 0) return void 0;
+  return opts.tools.map((t) => ({
+    type: "function",
+    function: { name: t.name, description: t.description ?? "", parameters: t.parameters }
+  }));
+}
+function openAiToolChoice(tc) {
+  if (tc === void 0) return void 0;
+  if (tc === "auto" || tc === "none") return tc;
+  return { type: "function", function: { name: tc.name } };
+}
 function buildGroqRequest(model, messages, opts, env) {
-  const sys = opts.system ?? messages.find((m) => m.role === "system")?.content;
+  const sys = systemText(opts, messages);
   const merged = [];
   if (sys) merged.push({ role: "system", content: sys });
-  for (const m of messages) if (m.role !== "system") merged.push(m);
+  for (const m of messages) if (m.role !== "system") merged.push({ role: m.role, content: contentToText(m.content) });
   return {
     url: `${env.AI_GATEWAY_BASE_URL}/groq/openai/v1/chat/completions`,
     headers: {
@@ -198,16 +257,19 @@ function buildGrokRequest(model, messages, opts, env) {
   if (!env.GROK_API_KEY) {
     throw new ValidationError("GROK_API_KEY required for grok-* model override");
   }
-  const sys = opts.system ?? messages.find((m) => m.role === "system")?.content;
-  const merged = [];
-  if (sys) merged.push({ role: "system", content: sys });
-  for (const m of messages) if (m.role !== "system") merged.push(m);
+  const sys = systemText(opts, messages);
   const body = {
     model,
     max_tokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,
     temperature: opts.temperature ?? DEFAULT_TEMPERATURE,
-    messages: merged
+    messages: toOpenAiMessages(messages, sys)
   };
+  const tools = openAiTools(opts);
+  if (tools) {
+    body.tools = tools;
+    const tc = openAiToolChoice(opts.toolChoice ?? "auto");
+    if (tc !== void 0) body.tool_choice = tc;
+  }
   if (model === MODELS.grok.fast) {
     body.reasoning_effort = opts.reasoningEffort ?? "none";
   }
@@ -224,33 +286,53 @@ function buildDeepSeekRequest(model, messages, opts, env) {
   if (!env.DEEPSEEK_API_KEY) {
     throw new ValidationError("DEEPSEEK_API_KEY required for workbench tier or deepseek-* model override");
   }
-  const sys = opts.system ?? messages.find((m) => m.role === "system")?.content;
-  const merged = [];
-  if (sys) merged.push({ role: "system", content: sys });
-  for (const m of messages) if (m.role !== "system") merged.push(m);
+  const sys = systemText(opts, messages);
+  const body = {
+    model,
+    max_tokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,
+    temperature: opts.temperature ?? DEFAULT_TEMPERATURE,
+    messages: toOpenAiMessages(messages, sys)
+  };
+  const tools = openAiTools(opts);
+  if (tools) {
+    body.tools = tools;
+    const tc = openAiToolChoice(opts.toolChoice ?? "auto");
+    if (tc !== void 0) body.tool_choice = tc;
+  }
   return {
     url: `${env.AI_GATEWAY_BASE_URL}/deepseek/chat/completions`,
     headers: {
       "content-type": "application/json",
       authorization: `Bearer ${env.DEEPSEEK_API_KEY}`
     },
-    body: JSON.stringify({
-      model,
-      max_tokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,
-      temperature: opts.temperature ?? DEFAULT_TEMPERATURE,
-      messages: merged
-    })
+    body: JSON.stringify(body)
   };
 }
+function normalizeAnthropicStop(reason) {
+  switch (reason) {
+    case "end_turn":
+    case "stop_sequence":
+      return "end";
+    case "tool_use":
+      return "tool_use";
+    case "max_tokens":
+      return "max_tokens";
+    default:
+      return reason ? "other" : void 0;
+  }
+}
 function parseAnthropic(json) {
   const r = json;
+  const toolCalls = (r.content ?? []).filter((c) => c.type === "tool_use" && typeof c.id === "string" && typeof c.name === "string").map((c) => ({ id: c.id, name: c.name, arguments: c.input ?? {} }));
   return {
     content: r.content?.find((c) => c.type === "text")?.text ?? "",
     input: r.usage?.input_tokens ?? 0,
     output: r.usage?.output_tokens ?? 0,
     cacheRead: r.usage?.cache_read_input_tokens ?? 0,
     cacheWrite: r.usage?.cache_creation_input_tokens ?? 0,
-    model: r.model
+    model: r.model,
+    toolCalls: toolCalls.length > 0 ? toolCalls : void 0,
+    stopReason: normalizeAnthropicStop(r.stop_reason)
   };
 }
 function parseGemini(json) {
@@ -271,6 +353,45 @@ function parseGroq(json) {
     model: r.model
   };
 }
+function normalizeOpenAiStop(reason) {
+  switch (reason) {
+    case "stop":
+      return "end";
+    case "tool_calls":
+    case "function_call":
+      return "tool_use";
+    case "length":
+      return "max_tokens";
+    default:
+      return reason ? "other" : void 0;
+  }
+}
+function parseToolArgs(raw) {
+  if (!raw) return {};
+  try {
+    const v = JSON.parse(raw);
+    return typeof v === "object" && v !== null ? v : {};
+  } catch {
+    return {};
+  }
+}
+function parseOpenAi(json) {
+  const r = json;
+  const choice = r.choices?.[0];
+  const toolCalls = (choice?.message?.tool_calls ?? []).filter((c) => typeof c.function?.name === "string").map((c, i) => ({
+    id: c.id ?? `call_${i}`,
+    name: c.function.name,
+    arguments: parseToolArgs(c.function?.arguments)
+  }));
+  return {
+    content: choice?.message?.content ?? "",
+    input: r.usage?.prompt_tokens ?? 0,
+    output: r.usage?.completion_tokens ?? 0,
+    model: r.model,
+    toolCalls: toolCalls.length > 0 ? toolCalls : void 0,
+    stopReason: normalizeOpenAiStop(choice?.finish_reason)
+  };
+}
 async function callWithBackoff(provider, request, fetchImpl, signal, logger, nowFn) {
   function exhaustAndThrow(err) {
     markProviderCoolingDown(provider, nowFn ?? Date.now);
@@ -334,6 +455,7 @@ async function callWithBackoff(provider, request, fetchImpl, signal, logger, now
 function isProviderError(err) {
   return typeof err === "object" && err !== null && typeof err.status === "number" && typeof err.message === "string" && typeof err.provider === "string";
 }
+var TOOL_CAPABLE_PROVIDERS = /* @__PURE__ */ new Set(["anthropic", "grok", "deepseek"]);
 function plan(tier, opts, tokenEstimate) {
   if (opts.model) {
     const m = opts.model;
@@ -421,9 +543,9 @@ async function callOne(leg, messages, opts, env, fetchImpl, logger, nowFn) {
     case "groq":
       return { parsed: parseGroq(json), gatewayRequestId, attempts };
     case "grok":
-      return { parsed: parseGroq(json), gatewayRequestId, attempts };
+      return { parsed: parseOpenAi(json), gatewayRequestId, attempts };
     case "deepseek":
-      return { parsed: parseGroq(json), gatewayRequestId, attempts };
+      return { parsed: parseOpenAi(json), gatewayRequestId, attempts };
   }
 }
 async function complete(messages, env, opts = {}, deps = {}) {
@@ -438,7 +560,7 @@ async function complete(messages, env, opts = {}, deps = {}) {
   const logger = deps.logger;
   const startedAt = now();
   const tier = opts.tier ?? "balanced";
-  const system = opts.system ?? messages.find((m) => m.role === "system")?.content;
+  const system = systemText(opts, messages);
   const tokenEstimate = estimateTokens(messages, system);
   const route = plan(tier, opts, tokenEstimate);
   const kv = env.LLM_COST_KV;
@@ -471,7 +593,15 @@ async function complete(messages, env, opts = {}, deps = {}) {
     }
   }
   const attemptLog = [];
-  const routeLegs = [route.primary, route.fallback].filter(Boolean);
+  let routeLegs = [route.primary, route.fallback].filter(Boolean);
+  if (opts.tools && opts.tools.length > 0) {
+    routeLegs = routeLegs.filter((l) => TOOL_CAPABLE_PROVIDERS.has(l.provider));
+    if (routeLegs.length === 0) {
+      throw new ValidationError(
+        `tool-calling requires a tool-capable provider (${[...TOOL_CAPABLE_PROVIDERS].join(", ")}); tier '${tier}' has none \u2014 use tier fast/balanced/smart or a claude-* model override`
+      );
+    }
+  }
   for (const [legIndex, leg] of routeLegs.entries()) {
     if (isProviderCoolingDown(leg.provider, now)) {
       logger?.warn?.("llm.provider.coolingDown", { provider: leg.provider });
@@ -485,7 +615,7 @@ async function complete(messages, env, opts = {}, deps = {}) {
     }
     try {
       const result = await callOne(leg, messages, opts, env, fetchImpl, logger, now);
-      if (!result.parsed.content) {
+      if (!result.parsed.content && !(result.parsed.toolCalls && result.parsed.toolCalls.length > 0)) {
         throw { provider: leg.provider, status: 200, retryable: false, message: "empty content" };
       }
       logger?.info?.("llm.complete", {
@@ -512,7 +642,9 @@ async function complete(messages, env, opts = {}, deps = {}) {
         },
         latency: now() - startedAt,
         attempts: result.attempts,
-        gatewayRequestId: result.gatewayRequestId
+        gatewayRequestId: result.gatewayRequestId,
+        stopReason: result.parsed.stopReason,
+        toolCalls: result.parsed.toolCalls
       };
       const costUsd = estimateCostUsd(llmResult.tokens, llmResult.model);
       if (opts.maxCostUsd !== void 0 && costUsd > opts.maxCostUsd) {
@@ -586,7 +718,7 @@ async function* completionStream(messages, env, opts = {}) {
   const logger = deps.logger;
   const startedAt = now();
   const tier = opts.tier ?? "balanced";
-  const system = opts.system ?? messages.find((m) => m.role === "system")?.content;
+  const system = systemText(opts, messages);
   const tokenEstimate = estimateTokens(messages, system);
   const route = plan(tier, opts, tokenEstimate);
   const streamLeg = route.primary.provider === "grok" && !env.GROK_API_KEY && route.fallback?.provider === "anthropic" ? route.fallback : route.primary;

package/dist/index.mjs.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/index.ts"],"sourcesContent":["import {\n  InternalError,\n  RateLimitError,\n  ValidationError,\n  toErrorResponse,\n  type FactoryResponse,\n} from '@latimer-woods-tech/errors';\nimport type { Logger } from '@latimer-woods-tech/logger';\n\n/**\n * Single chat message exchanged with an LLM provider.\n */\nexport interface LLMMessage {\n  role: 'user' | 'assistant' | 'system';\n  content: string;\n}\n\n/**\n * Quality tier selected by the caller. Routing is workload-split:\n *  - `fast`      → Grok 4.3 with Anthropic Haiku fallback (routine drafts/small jobs)\n *  - `balanced`  → Anthropic Sonnet (default)\n *  - `smart`     → Anthropic Opus OR Gemini 2.5 Pro if input is long-context (>150k tokens estimated)\n *  - `verifier`  → Groq Llama (cheap second opinion; only used from verifier code path)\n *  - `workbench` → DeepSeek Chat with Groq fallback (boring, reviewable, non-sensitive batch work)\n */\nexport type LLMTier = 'fast' | 'balanced' | 'smart' | 'verifier' | 'workbench';\n\n/**\n * Options that influence LLM completion behaviour.\n */\nexport interface LLMOptions {\n  /** Quality tier; see {@link LLMTier}. Defaults to `balanced`. */\n  tier?: LLMTier;\n  /** Explicit model override. Takes precedence over tier. */\n  model?: string;\n  maxTokens?: number;\n  temperature?: number;\n  system?: string;\n  /** Token budget above which we force long-context routing (Gemini). */\n  longContextThreshold?: number;\n  /** Per-call cancellation signal. Aborts the in-flight provider request. */\n  signal?: AbortSignal;\n  /** Optional run identifier stamped on ledger rows + logs. */\n  runId?: string;\n  /** Optional project identifier stamped on ledger rows + logs. */\n  project?: string;\n  /** Optional actor identifier (supervisor / worker / human). */\n  actor?: string;\n  /** Optional workload label used in logs and cost-policy call sites. */\n  workload?: string;\n  /** Grok reasoning effort. Defaults to `none` for cost-controlled fast/draft calls. */\n  reasoningEffort?: 'none' | 'low' | 'medium' | 'high';\n  /** Anthropic prompt-cache control. Defaults to `true` for `system` prompts ≥ 1024 tokens. */\n  promptCache?: boolean;\n  /**\n   * Maximum estimated cost in USD for this completion.\n   * This cap is enforced after the provider returns because it uses actual\n   * response token counts to compute the final cost.\n   * If the post-call estimated cost exceeds this cap, `complete` returns a\n   * {@link RateLimitError} with code `LLM_COST_CAP_EXCEEDED` and\n   * `completionStream` throws the same error.\n   * Pricing is based on {@link MODEL_PRICE_PER_1M}; unknown models default to\n   * Opus rates (conservative upper bound).\n   */\n  maxCostUsd?: number;\n  /**\n   * Org-level daily cost cap in USD. Requires `env.LLM_COST_KV` to be set.\n   * When today's cumulative spend read from KV is >= this value, `complete`\n   * returns a {@link RateLimitError} with code `LLM_DAILY_CAP_EXCEEDED`\n   * without making any provider call. After a successful call the daily\n   * accumulator is updated in KV (TTL: 48 h).\n   */\n  dailyCapUsd?: number;\n  /**\n   * Org-level monthly cost cap in USD. Requires `env.LLM_COST_KV` to be set.\n   * Same enforcement pattern as {@link dailyCapUsd} but keyed by YYYY-MM.\n   * KV TTL: 40 days.\n   */\n  monthlyCapUsd?: number;\n  /**\n   * Metering context. When supplied and `deps.onRecord` is set, a {@link LLMRecordRow}\n   * is emitted after every successful completion. Errors are swallowed.\n   */\n  ledger?: LLMRecordContext;\n}\n\n/**\n * Provider that produced an LLM response.\n */\nexport type LLMProvider = 'anthropic' | 'gemini' | 'groq' | 'grok' | 'deepseek';\n\n/**\n * Result returned by a successful completion.\n */\nexport interface LLMResult {\n  content: string;\n  provider: LLMProvider;\n  model: string;\n  tier: LLMTier;\n  tokens: { input: number; output: number; cacheRead?: number; cacheWrite?: number };\n  latency: number;\n  /** Number of attempts before success (1 = primary succeeded). */\n  attempts: number;\n  /** Monotonic request id from AI Gateway, if present in headers. */\n  gatewayRequestId?: string;\n}\n\n/**\n * Environment bindings required by {@link complete}.\n *\n * `AI_GATEWAY_BASE_URL` is REQUIRED in 0.3.0. All provider calls flow through the\n * Cloudflare AI Gateway for unified logging, rate limiting, and cost telemetry.\n * In test/dev the caller may pass a custom fetch impl that short-circuits this.\n */\nexport interface LLMEnv {\n  AI_GATEWAY_BASE_URL: string;\n  ANTHROPIC_API_KEY: string;\n  GROQ_API_KEY: string;\n  /** Optional — only required for `{ tier: 'workbench' }` or `deepseek-*` model overrides. */\n  DEEPSEEK_API_KEY?: string;\n  /** Optional — only required when caller passes `{ model: 'grok-*' }` override. */\n  GROK_API_KEY?: string;\n  /**\n   * Google Cloud short-lived access token with `aiplatform.endpoints.predict`.\n   * Callers mint this via the JWT-bearer flow (service account → token exchange);\n   * see `docs/runbooks/rotate-gcp-sa.md`. Token must be valid for ≥ 5 minutes.\n   */\n  VERTEX_ACCESS_TOKEN: string;\n  VERTEX_PROJECT: string;\n  VERTEX_LOCATION: string;\n  /**\n   * Optional KV store for org-level daily/monthly cost tracking and enforcement.\n   * When provided alongside {@link LLMOptions.dailyCapUsd} or {@link LLMOptions.monthlyCapUsd},\n   * `complete` will block calls that would exceed the declared cap.\n   * Any KV-like store satisfying `get`/`put` works (e.g. Cloudflare KV, in-memory stub).\n   */\n  LLM_COST_KV?: CostKvStore;\n}\n\n/**\n * Minimal KV store interface for org-level LLM cost tracking.\n * Cloudflare KV satisfies this. An in-memory stub is sufficient for tests.\n */\nexport interface CostKvStore {\n  get(key: string): Promise<string | null>;\n  put(key: string, value: string, options?: { expirationTtl?: number }): Promise<void>;\n}\n\n/**\n * Caller-supplied context stamped on every metering row.\n * Mirrors the `LLMRecordContext` in `@latimer-woods-tech/llm-meter`; kept inline\n * to avoid a circular dependency (llm-meter imports llm).\n */\nexport interface LLMRecordContext {\n  project: string;\n  actor: string;\n  runId?: string;\n  workload?: string;\n  tenantId?: string;\n}\n\n/**\n * Row shape passed to the optional {@link LLMDeps.onRecord} callback.\n * Callers can wire this directly to `recordCall` from `@latimer-woods-tech/llm-meter`.\n */\nexport interface LLMRecordRow extends LLMRecordContext {\n  model: string;\n  provider: LLMProvider;\n  tier: LLMTier;\n  inputTokens: number;\n  outputTokens: number;\n  cacheReadTokens: number;\n  cacheWriteTokens: number;\n  latencyMs: number;\n  costUsd: number;\n  yyyyMm: string;\n}\n\n/**\n * Optional dependencies for {@link complete}.\n */\nexport interface LLMDeps {\n  fetch?: typeof fetch;\n  logger?: Logger;\n  now?: () => number;\n  /**\n   * Optional metering callback. Called after every successful completion.\n   * Errors are swallowed so metering never blocks the caller.\n   * Wire to `recordCall` from `@latimer-woods-tech/llm-meter`.\n   */\n  onRecord?: (row: LLMRecordRow) => Promise<void>;\n}\n\n// Model catalogue — keep in sync with docs/architecture/FACTORY_V1.md § LLM substrate.\nconst MODELS = {\n  anthropic: {\n    fast: 'claude-haiku-4-20250514',\n    balanced: 'claude-sonnet-4-6',\n    smart: 'claude-opus-4-7',\n  },\n  gemini: {\n    smart: 'gemini-2.5-pro',\n  },\n  groq: {\n    verifier: 'llama-4-maverick',\n  },\n  grok: {\n    fast: 'grok-4.3',\n  },\n  deepseek: {\n    workbench: 'deepseek-chat',\n  },\n} as const;\n\nconst DEFAULT_MAX_TOKENS = 1024;\nconst DEFAULT_TEMPERATURE = 0.7;\nconst DEFAULT_LONG_CONTEXT_THRESHOLD = 150_000; // tokens\n\n// ─── Per-provider exponential backoff constants ────────────────────────────\n/** Base delay in ms for the first retry. */\nconst BACKOFF_BASE_MS = 500;\n/** Maximum backoff cap in ms. */\nconst BACKOFF_CAP_MS = 8_000;\n/** Max random jitter added to each backoff delay, in ms. */\nconst BACKOFF_JITTER_MAX_MS = 250;\n/** Maximum number of attempts per provider (1 initial + 2 retries). */\nconst PER_PROVIDER_MAX_ATTEMPTS = 3;\n\n// ─── Per-provider cooldown state (module-level) ────────────────────────────\n/**\n * Tracks when a provider's cooldown period expires.\n * Keyed by {@link LLMProvider}; value is the `Date.now()` epoch ms at which\n * the cooldown expires. Absent key means \"not cooling down\".\n */\nconst providerCooldownUntil: Map<LLMProvider, number> = new Map();\n\n/** Cooldown duration in ms after a provider exhausts all retries. */\nconst PROVIDER_COOLDOWN_MS = 30_000;\n\n/**\n * Returns `true` if the provider is currently in its cooldown window.\n * Uses the injected `now` function (or `Date.now`) for testability.\n */\nfunction isProviderCoolingDown(provider: LLMProvider, now: () => number = Date.now): boolean {\n  const until = providerCooldownUntil.get(provider);\n  if (until === undefined) return false;\n  return now() < until;\n}\n\n/** Returns `YYYY-MM-DD` from a Unix timestamp (ms). Used for daily KV cost keys. */\nfunction isoDate(nowMs: number): string {\n  return new Date(nowMs).toISOString().slice(0, 10);\n}\n\n/**\n * Record actual call spend in org-level daily/monthly KV buckets.\n * This is intentionally best-effort: Cloudflare KV does not provide an atomic\n * compare-and-swap, so concurrent requests can race and undercount spend.\n */\nasync function recordOrgCostUsage(\n  kv: CostKvStore,\n  todayKey: string,\n  monthKey: string,\n  costUsd: number,\n  opts: LLMOptions,\n): Promise<void> {\n  if (opts.dailyCapUsd !== undefined) {\n    const raw = await kv.get(todayKey).catch(() => null);\n    const spent = parseFloat(raw ?? '0');\n    await kv.put(todayKey, String(spent + costUsd), { expirationTtl: 172_800 /* 48 h */ }).catch(() => undefined);\n  }\n  if (opts.monthlyCapUsd !== undefined) {\n    const raw = await kv.get(monthKey).catch(() => null);\n    const spent = parseFloat(raw ?? '0');\n    await kv.put(monthKey, String(spent + costUsd), { expirationTtl: 3_456_000 /* 40 d */ }).catch(() => undefined);\n  }\n}\n\n/**\n * USD cost per 1 million tokens for each model.\n * Source: Anthropic / Google / xAI pricing pages as of 2026-05.\n * Keep these model names in sync with the default routing constants in\n * {@link MODELS}; unknown models fall back to Opus rates (conservative upper bound).\n *\n * CANONICAL pricing source for the platform. `@latimer-woods-tech/llm-meter`\n * derives its cents-denominated rates from this table and a drift-guard test\n * there fails CI if they diverge — make all rate changes here.\n */\nexport const MODEL_PRICE_PER_1M: Record<string, { input: number; output: number; cacheRead: number; cacheWrite: number }> = {\n  // Anthropic Haiku 4\n  'claude-haiku-4-20250514': { input: 0.80, output: 4.00, cacheRead: 0.08, cacheWrite: 1.00 },\n  'claude-haiku-4-5-20251001': { input: 0.80, output: 4.00, cacheRead: 0.08, cacheWrite: 1.00 },\n  // Anthropic Sonnet 4\n  'claude-sonnet-4-20250514': { input: 3.00, output: 15.00, cacheRead: 0.30, cacheWrite: 3.75 },\n  'claude-sonnet-4-6': { input: 3.00, output: 15.00, cacheRead: 0.30, cacheWrite: 3.75 },\n  // Anthropic Opus 4\n  'claude-opus-4-20250514': { input: 15.00, output: 75.00, cacheRead: 1.50, cacheWrite: 18.75 },\n  'claude-opus-4-7': { input: 15.00, output: 75.00, cacheRead: 1.50, cacheWrite: 18.75 },\n  // Gemini 2.5 Pro\n  'gemini-2.5-pro': { input: 1.25, output: 10.00, cacheRead: 0.31, cacheWrite: 4.50 },\n  // Groq Llama 4 Maverick\n  'llama-4-maverick': { input: 0.50, output: 0.77, cacheRead: 0.05, cacheWrite: 0.50 },\n  // Grok 4.3\n  'grok-4.3': { input: 1.25, output: 2.50, cacheRead: 0.00, cacheWrite: 0.00 },\n  // DeepSeek API pricing as of 2026-05: cache-write conservatively uses cache-miss input pricing.\n  'deepseek-chat': { input: 0.27, output: 1.10, cacheRead: 0.07, cacheWrite: 0.27 },\n  'deepseek-reasoner': { input: 0.55, output: 2.19, cacheRead: 0.14, cacheWrite: 0.55 },\n  // Deprecated aliases retained for historical ledger rows.\n  'grok-4-fast': { input: 1.25, output: 2.50, cacheRead: 0.00, cacheWrite: 0.00 },\n  'grok-3-mini-latest': { input: 1.25, output: 2.50, cacheRead: 0.00, cacheWrite: 0.00 },\n};\n\n/** Fallback pricing used for unrecognised models (Opus rates — conservative upper bound). */\nconst PRICE_FALLBACK = MODEL_PRICE_PER_1M['claude-opus-4-7']!;\n\n/**\n * Estimates the USD cost of a single LLM completion from token counts.\n * Returns 0 for zero-token results. Uses {@link MODEL_PRICE_PER_1M} with\n * {@link PRICE_FALLBACK} for unknown models.\n */\nfunction estimateCostUsd(\n  tokens: { input: number; output: number; cacheRead?: number; cacheWrite?: number },\n  model: string,\n): number {\n  const price = MODEL_PRICE_PER_1M[model] ?? PRICE_FALLBACK;\n  return (\n    (tokens.input * price.input +\n      tokens.output * price.output +\n      (tokens.cacheRead ?? 0) * price.cacheRead +\n      (tokens.cacheWrite ?? 0) * price.cacheWrite) /\n    1_000_000\n  );\n}\n\n/** Returns `YYYY-MM` from a Unix timestamp (ms). Used for monthly KV cost keys. */\nfunction isoMonth(nowMs: number): string {\n  return new Date(nowMs).toISOString().slice(0, 7);\n}\n\n/**\n * Marks a provider as cooling down for {@link PROVIDER_COOLDOWN_MS} milliseconds.\n */\nfunction markProviderCoolingDown(provider: LLMProvider, now: () => number = Date.now): void {\n  providerCooldownUntil.set(provider, now() + PROVIDER_COOLDOWN_MS);\n}\n\n/**\n * Clears the cooldown state for a provider after a successful call.\n */\nfunction clearProviderCooldown(provider: LLMProvider): void {\n  providerCooldownUntil.delete(provider);\n}\n\n// ─── Legacy backoff constant (kept for the existing callWithBackoff signature) ─\nconst BASE_BACKOFF_MS = 250;\n\ninterface ProviderError {\n  provider: LLMProvider;\n  status: number;\n  retryable: boolean;\n  message: string;\n}\n\n/**\n * Returns `true` for status codes that should trigger a retry.\n * Only 429 and 5xx (transient server errors) qualify; other 4xx are terminal.\n */\nfunction isRetryableForBackoff(status: number): boolean {\n  return status === 429 || (status >= 500 && status < 600);\n}\n\nfunction estimateTokens(messages: LLMMessage[], system?: string): number {\n  // Cheap estimator: ~4 chars/token. Good enough for threshold routing.\n  let chars = system?.length ?? 0;\n  for (const m of messages) chars += m.content.length;\n  return Math.ceil(chars / 4);\n}\n\nfunction sleep(ms: number, signal?: AbortSignal): Promise<void> {\n  return new Promise((resolve, reject) => {\n    const t = setTimeout(resolve, ms);\n    if (signal) {\n      const onAbort = () => {\n        clearTimeout(t);\n        reject(new DOMException('Aborted', 'AbortError'));\n      };\n      if (signal.aborted) onAbort();\n      else signal.addEventListener('abort', onAbort, { once: true });\n    }\n  });\n}\n\n/**\n * Computes the exponential backoff delay for a given attempt with jitter.\n *\n * Formula: `Math.min(base * 2^attempt + jitter, cap)`\n * where `jitter` is a random value in `[0, BACKOFF_JITTER_MAX_MS)`.\n *\n * @param attempt - Zero-based attempt index (0 = first retry after initial failure).\n */\nfunction computeBackoffMs(attempt: number): number {\n  const jitter = Math.floor(Math.random() * BACKOFF_JITTER_MAX_MS);\n  return Math.min(BACKOFF_BASE_MS * Math.pow(2, attempt) + jitter, BACKOFF_CAP_MS);\n}\n\n// ─── Provider request builders ─────────────────────────────────────────────\n\nfunction buildAnthropicRequest(\n  model: string,\n  messages: LLMMessage[],\n  opts: LLMOptions,\n  env: LLMEnv,\n  streaming = false,\n): { url: string; headers: Record<string, string>; body: string } {\n  const sys = opts.system ?? messages.find((m) => m.role === 'system')?.content;\n  const filtered = messages.filter((m) => m.role !== 'system');\n  const body: Record<string, unknown> = {\n    model,\n    max_tokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,\n    temperature: opts.temperature ?? DEFAULT_TEMPERATURE,\n    messages: filtered.map((m) => ({ role: m.role, content: m.content })),\n  };\n  if (streaming) {\n    body.stream = true;\n  }\n  if (sys) {\n    const cache = opts.promptCache ?? sys.length >= 4096;\n    body.system = cache\n      ? [{ type: 'text', text: sys, cache_control: { type: 'ephemeral' } }]\n      : sys;\n  }\n  return {\n    url: `${env.AI_GATEWAY_BASE_URL}/anthropic/v1/messages`,\n    headers: {\n      'content-type': 'application/json',\n      'x-api-key': env.ANTHROPIC_API_KEY,\n      'anthropic-version': '2023-06-01',\n      'anthropic-beta': 'prompt-caching-2024-07-31',\n    },\n    body: JSON.stringify(body),\n  };\n}\n\nfunction buildGeminiRequest(\n  model: string,\n  messages: LLMMessage[],\n  opts: LLMOptions,\n  env: LLMEnv,\n): { url: string; headers: Record<string, string>; body: string } {\n  const sys = opts.system ?? messages.find((m) => m.role === 'system')?.content;\n  const contents = messages\n    .filter((m) => m.role !== 'system')\n    .map((m) => ({\n      role: m.role === 'assistant' ? 'model' : 'user',\n      parts: [{ text: m.content }],\n    }));\n  const body: Record<string, unknown> = {\n    contents,\n    generationConfig: {\n      maxOutputTokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,\n      temperature: opts.temperature ?? DEFAULT_TEMPERATURE,\n    },\n  };\n  if (sys) {\n    body.systemInstruction = { parts: [{ text: sys }] };\n  }\n  const path = `v1/projects/${env.VERTEX_PROJECT}/locations/${env.VERTEX_LOCATION}/publishers/google/models/${model}:generateContent`;\n  return {\n    url: `${env.AI_GATEWAY_BASE_URL}/google-vertex-ai/${path}`,\n    headers: {\n      'content-type': 'application/json',\n      authorization: `Bearer ${env.VERTEX_ACCESS_TOKEN}`,\n    },\n    body: JSON.stringify(body),\n  };\n}\n\nfunction buildGroqRequest(\n  model: string,\n  messages: LLMMessage[],\n  opts: LLMOptions,\n  env: LLMEnv,\n): { url: string; headers: Record<string, string>; body: string } {\n  const sys = opts.system ?? messages.find((m) => m.role === 'system')?.content;\n  const merged: LLMMessage[] = [];\n  if (sys) merged.push({ role: 'system', content: sys });\n  for (const m of messages) if (m.role !== 'system') merged.push(m);\n  return {\n    url: `${env.AI_GATEWAY_BASE_URL}/groq/openai/v1/chat/completions`,\n    headers: {\n      'content-type': 'application/json',\n      authorization: `Bearer ${env.GROQ_API_KEY}`,\n    },\n    body: JSON.stringify({\n      model,\n      max_tokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,\n      temperature: opts.temperature ?? DEFAULT_TEMPERATURE,\n      messages: merged,\n    }),\n  };\n}\n\nfunction buildGrokRequest(\n  model: string,\n  messages: LLMMessage[],\n  opts: LLMOptions,\n  env: LLMEnv,\n): { url: string; headers: Record<string, string>; body: string } {\n  if (!env.GROK_API_KEY) {\n    throw new ValidationError('GROK_API_KEY required for grok-* model override');\n  }\n  const sys = opts.system ?? messages.find((m) => m.role === 'system')?.content;\n  const merged: LLMMessage[] = [];\n  if (sys) merged.push({ role: 'system', content: sys });\n  for (const m of messages) if (m.role !== 'system') merged.push(m);\n  const body: Record<string, unknown> = {\n    model,\n    max_tokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,\n    temperature: opts.temperature ?? DEFAULT_TEMPERATURE,\n    messages: merged,\n  };\n  if (model === MODELS.grok.fast) {\n    body.reasoning_effort = opts.reasoningEffort ?? 'none';\n  }\n  return {\n    url: `${env.AI_GATEWAY_BASE_URL}/grok/v1/chat/completions`,\n    headers: {\n      'content-type': 'application/json',\n      authorization: `Bearer ${env.GROK_API_KEY}`,\n    },\n    body: JSON.stringify(body),\n  };\n}\n\nfunction buildDeepSeekRequest(\n  model: string,\n  messages: LLMMessage[],\n  opts: LLMOptions,\n  env: LLMEnv,\n): { url: string; headers: Record<string, string>; body: string } {\n  if (!env.DEEPSEEK_API_KEY) {\n    throw new ValidationError('DEEPSEEK_API_KEY required for workbench tier or deepseek-* model override');\n  }\n  const sys = opts.system ?? messages.find((m) => m.role === 'system')?.content;\n  const merged: LLMMessage[] = [];\n  if (sys) merged.push({ role: 'system', content: sys });\n  for (const m of messages) if (m.role !== 'system') merged.push(m);\n  return {\n    url: `${env.AI_GATEWAY_BASE_URL}/deepseek/chat/completions`,\n    headers: {\n      'content-type': 'application/json',\n      authorization: `Bearer ${env.DEEPSEEK_API_KEY}`,\n    },\n    body: JSON.stringify({\n      model,\n      max_tokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,\n      temperature: opts.temperature ?? DEFAULT_TEMPERATURE,\n      messages: merged,\n    }),\n  };\n}\n\n// ─── Response parsers ──────────────────────────────────────────────────────\n\ninterface AnthropicResponse {\n  content?: Array<{ type: string; text?: string }>;\n  usage?: {\n    input_tokens?: number;\n    output_tokens?: number;\n    cache_read_input_tokens?: number;\n    cache_creation_input_tokens?: number;\n  };\n  model?: string;\n}\n\ninterface GeminiResponse {\n  candidates?: Array<{ content?: { parts?: Array<{ text?: string }> } }>;\n  usageMetadata?: {\n    promptTokenCount?: number;\n    candidatesTokenCount?: number;\n  };\n}\n\ninterface GroqResponse {\n  choices?: Array<{ message?: { content?: string } }>;\n  usage?: { prompt_tokens?: number; completion_tokens?: number };\n  model?: string;\n}\n\nfunction parseAnthropic(\n  json: unknown,\n): { content: string; input: number; output: number; cacheRead: number; cacheWrite: number; model?: string } {\n  const r = json as AnthropicResponse;\n  return {\n    content: r.content?.find((c) => c.type === 'text')?.text ?? '',\n    input: r.usage?.input_tokens ?? 0,\n    output: r.usage?.output_tokens ?? 0,\n    cacheRead: r.usage?.cache_read_input_tokens ?? 0,\n    cacheWrite: r.usage?.cache_creation_input_tokens ?? 0,\n    model: r.model,\n  };\n}\n\nfunction parseGemini(json: unknown): { content: string; input: number; output: number } {\n  const r = json as GeminiResponse;\n  const text =\n    r.candidates?.[0]?.content?.parts?.map((p) => p.text ?? '').join('') ?? '';\n  return {\n    content: text,\n    input: r.usageMetadata?.promptTokenCount ?? 0,\n    output: r.usageMetadata?.candidatesTokenCount ?? 0,\n  };\n}\n\nfunction parseGroq(json: unknown): { content: string; input: number; output: number; model?: string } {\n  const r = json as GroqResponse;\n  return {\n    content: r.choices?.[0]?.message?.content ?? '',\n    input: r.usage?.prompt_tokens ?? 0,\n    output: r.usage?.completion_tokens ?? 0,\n    model: r.model,\n  };\n}\n\n// ─── Core call with backoff ────────────────────────────────────────────────\n\n/**\n * Calls a provider with per-provider exponential backoff.\n *\n * Retries up to {@link PER_PROVIDER_MAX_ATTEMPTS} times on 429 or transient 5xx.\n * Other 4xx codes are treated as terminal and not retried.\n * AbortError is never retried — it bubbles immediately.\n *\n * @param provider - Provider name, used for error tagging.\n * @param request - Pre-built HTTP request descriptor.\n * @param fetchImpl - Fetch implementation (injectable for tests).\n * @param signal - Optional AbortSignal for cancellation.\n * @param logger - Optional logger for per-attempt warnings.\n * @param nowFn - Optional clock injection for testability.\n * @returns Parsed JSON body, optional AI Gateway request ID, and attempt count.\n */\nasync function callWithBackoff(\n  provider: LLMProvider,\n  request: { url: string; headers: Record<string, string>; body: string },\n  fetchImpl: typeof fetch,\n  signal: AbortSignal | undefined,\n  logger: Logger | undefined,\n  nowFn?: () => number,\n): Promise<{ json: unknown; gatewayRequestId?: string; attempts: number }> {\n  /**\n   * Helper: mark provider cooling down and then throw the error.\n   * Called whenever we determine we've exhausted all retries for the provider.\n   * AbortError is never counted as a provider exhaustion — it bypasses this.\n   */\n  function exhaustAndThrow(err: ProviderError): never {\n    markProviderCoolingDown(provider, nowFn ?? Date.now);\n    throw err;\n  }\n\n  let lastErr: ProviderError | undefined;\n  for (let attempt = 1; attempt <= PER_PROVIDER_MAX_ATTEMPTS; attempt++) {\n    try {\n      const response = await fetchImpl(request.url, {\n        method: 'POST',\n        headers: request.headers,\n        body: request.body,\n        signal,\n      });\n      if (!response.ok) {\n        const text = await response.text().catch(() => '');\n        const retryable = isRetryableForBackoff(response.status);\n        const err: ProviderError = {\n          provider,\n          status: response.status,\n          retryable,\n          message: `${provider} ${String(response.status)}: ${text.slice(0, 300)}`,\n        };\n        logger?.warn?.('llm.provider.error', { provider, status: response.status, attempt });\n        if (!err.retryable || attempt === PER_PROVIDER_MAX_ATTEMPTS) {\n          if (err.retryable) exhaustAndThrow(err); // retryable but exhausted\n          throw err; // terminal non-retryable error — no cooldown\n        }\n        lastErr = err;\n      } else {\n        const gatewayRequestId = response.headers.get('cf-aig-request-id') ?? undefined;\n        clearProviderCooldown(provider);\n        return { json: await response.json(), gatewayRequestId, attempts: attempt };\n      }\n    } catch (e) {\n      if (e instanceof DOMException && e.name === 'AbortError') throw e;\n      if (typeof e === 'object' && e !== null && 'retryable' in e) {\n        const err = e as ProviderError;\n        if (!err.retryable || attempt === PER_PROVIDER_MAX_ATTEMPTS) {\n          if (err.retryable) exhaustAndThrow(err); // retryable but exhausted\n          throw err; // terminal — no cooldown\n        }\n        lastErr = err;\n      } else {\n        const err: ProviderError = {\n          provider,\n          status: 0,\n          retryable: true,\n          message: e instanceof Error ? e.message : String(e),\n        };\n        if (attempt === PER_PROVIDER_MAX_ATTEMPTS) exhaustAndThrow(err);\n        lastErr = err;\n      }\n    }\n    // Exponential backoff with jitter: base=500ms, cap=8000ms, jitter up to 250ms\n    const backoffMs = computeBackoffMs(attempt - 1);\n    await sleep(backoffMs, signal);\n  }\n  // Fallthrough — should not be reached, but mark cooling down defensively.\n  markProviderCoolingDown(provider, nowFn ?? Date.now);\n  throw lastErr ?? ({ provider, status: 0, retryable: false, message: 'exhausted' } as ProviderError);\n}\n\nfunction isProviderError(err: unknown): err is ProviderError {\n  return (\n    typeof err === 'object' &&\n    err !== null &&\n    typeof (err as { status?: unknown }).status === 'number' &&\n    typeof (err as { message?: unknown }).message === 'string' &&\n    typeof (err as { provider?: unknown }).provider === 'string'\n  );\n}\n\n// ─── Routing ───────────────────────────────────────────────────────────────\n\ninterface RoutePlan {\n  primary: { provider: LLMProvider; model: string };\n  fallback?: { provider: LLMProvider; model: string };\n}\n\nfunction plan(tier: LLMTier, opts: LLMOptions, tokenEstimate: number): RoutePlan {\n  if (opts.model) {\n    // Explicit override — best-effort provider detection.\n    const m = opts.model;\n    if (m.startsWith('claude')) return { primary: { provider: 'anthropic', model: m } };\n    if (m.startsWith('gemini')) return { primary: { provider: 'gemini', model: m } };\n    if (m.startsWith('grok')) return { primary: { provider: 'grok', model: m } };\n    if (m.startsWith('deepseek')) return { primary: { provider: 'deepseek', model: m } };\n    return { primary: { provider: 'groq', model: m } };\n  }\n  const longContext = tokenEstimate >= (opts.longContextThreshold ?? DEFAULT_LONG_CONTEXT_THRESHOLD);\n  switch (tier) {\n    case 'workbench':\n      return {\n        primary: { provider: 'deepseek', model: MODELS.deepseek.workbench },\n        fallback: { provider: 'groq', model: MODELS.groq.verifier },\n      };\n    case 'verifier':\n      return { primary: { provider: 'groq', model: MODELS.groq.verifier } };\n    case 'smart':\n      return longContext\n        ? {\n            primary: { provider: 'gemini', model: MODELS.gemini.smart },\n            fallback: { provider: 'anthropic', model: MODELS.anthropic.smart },\n          }\n        : {\n            primary: { provider: 'anthropic', model: MODELS.anthropic.smart },\n            fallback: { provider: 'gemini', model: MODELS.gemini.smart },\n          };\n    case 'fast':\n      return {\n        primary: { provider: 'grok', model: MODELS.grok.fast },\n        fallback: { provider: 'anthropic', model: MODELS.anthropic.fast },\n      };\n    case 'balanced':\n    default:\n      return longContext\n        ? {\n            primary: { provider: 'gemini', model: MODELS.gemini.smart },\n            fallback: { provider: 'anthropic', model: MODELS.anthropic.balanced },\n          }\n        : {\n            primary: { provider: 'anthropic', model: MODELS.anthropic.balanced },\n            fallback: { provider: 'gemini', model: MODELS.gemini.smart },\n          };\n  }\n}\n\n/**\n * Build the `cf-aig-metadata` header value for the Cloudflare AI Gateway.\n *\n * Carries caller attribution (project / workload / actor / runId) so a single\n * shared gateway can be sliced per-app and per-feature in the AI Gateway\n * dashboard and logs. This replaces the per-app-gateway convention: rather than\n * one gateway per app (which has to be provisioned and silently 401s when it\n * isn't), one gateway tags every request with who made it.\n *\n * Returns `undefined` when no attribution fields are set (header omitted).\n * The CF AI Gateway accepts a JSON object of string/number/boolean values.\n */\nfunction buildAigMetadata(opts: LLMOptions): string | undefined {\n  const meta: Record<string, string> = {};\n  if (opts.project) meta.project = opts.project;\n  if (opts.workload) meta.workload = opts.workload;\n  if (opts.actor) meta.actor = opts.actor;\n  if (opts.runId) meta.runId = opts.runId;\n  return Object.keys(meta).length > 0 ? JSON.stringify(meta) : undefined;\n}\n\nasync function callOne(\n  leg: { provider: LLMProvider; model: string },\n  messages: LLMMessage[],\n  opts: LLMOptions,\n  env: LLMEnv,\n  fetchImpl: typeof fetch,\n  logger: Logger | undefined,\n  nowFn?: () => number,\n): Promise<{ parsed: { content: string; input: number; output: number; cacheRead?: number; cacheWrite?: number; model?: string }; gatewayRequestId?: string; attempts: number }> {\n  let req: { url: string; headers: Record<string, string>; body: string };\n  switch (leg.provider) {\n    case 'anthropic':\n      req = buildAnthropicRequest(leg.model, messages, opts, env);\n      break;\n    case 'gemini':\n      req = buildGeminiRequest(leg.model, messages, opts, env);\n      break;\n    case 'groq':\n      req = buildGroqRequest(leg.model, messages, opts, env);\n      break;\n    case 'grok':\n      req = buildGrokRequest(leg.model, messages, opts, env);\n      break;\n    case 'deepseek':\n      req = buildDeepSeekRequest(leg.model, messages, opts, env);\n      break;\n  }\n  // Attribution for the shared AI Gateway — one gateway, sliced per-app/feature.\n  const aigMetadata = buildAigMetadata(opts);\n  if (aigMetadata) req.headers['cf-aig-metadata'] = aigMetadata;\n  const { json, gatewayRequestId, attempts } = await callWithBackoff(\n    leg.provider,\n    req,\n    fetchImpl,\n    opts.signal,\n    logger,\n    nowFn,\n  );\n  switch (leg.provider) {\n    case 'anthropic':\n      return { parsed: parseAnthropic(json), gatewayRequestId, attempts };\n    case 'gemini':\n      return { parsed: parseGemini(json), gatewayRequestId, attempts };\n    case 'groq':\n      return { parsed: parseGroq(json), gatewayRequestId, attempts };\n    case 'grok':\n      return { parsed: parseGroq(json), gatewayRequestId, attempts };\n    case 'deepseek':\n      return { parsed: parseGroq(json), gatewayRequestId, attempts };\n  }\n}\n\n/**\n * Run a completion through the routing plan for the requested tier.\n *\n * Routing summary (0.3.0):\n *   - `fast`     → Grok 4.3; Anthropic Haiku fallback when Grok is unavailable\n *   - `balanced` → Anthropic Sonnet; Gemini 2.5 Pro if `longContextThreshold` exceeded\n *   - `smart`    → Anthropic Opus; Gemini 2.5 Pro if long-context\n *   - `verifier` → Groq Llama 3.3 70B (no fallback — verifier is inherently cheap/best-effort)\n *   - `workbench` → DeepSeek Chat; Groq fallback for boring/reviewable internal batch jobs\n *\n * All provider traffic flows through Cloudflare AI Gateway at `AI_GATEWAY_BASE_URL`.\n *\n * Per-provider reliability guarantees (0.4.0):\n *   - Exponential backoff with jitter on 429 / 5xx (base 500ms, cap 8s, up to 2 retries).\n *   - Provider cooldown: after exhausting retries the provider is marked cooling down\n *     for 30 seconds; subsequent calls skip it and go straight to the fallback leg.\n *\n * @param messages - Ordered chat history.\n * @param env - API key + gateway bindings.\n * @param opts - Optional tier/model/parameters override.\n * @param deps - Optional fetch/logger/clock injection (for testing).\n * @returns A {@link FactoryResponse} carrying either an {@link LLMResult} or\n *   an error (`LLM_ALL_PROVIDERS_FAILED`, `LLM_RATE_LIMITED`, or `INTERNAL_ERROR`).\n */\nexport async function complete(\n  messages: LLMMessage[],\n  env: LLMEnv,\n  opts: LLMOptions = {},\n  deps: LLMDeps = {},\n): Promise<FactoryResponse<LLMResult>> {\n  if (messages.length === 0) {\n    throw new ValidationError('messages must not be empty');\n  }\n  if (!env.AI_GATEWAY_BASE_URL) {\n    throw new ValidationError('AI_GATEWAY_BASE_URL is required in 0.3.0');\n  }\n  const fetchImpl = deps.fetch ?? fetch;\n  const now = deps.now ?? (() => Date.now());\n  const logger = deps.logger;\n  const startedAt = now();\n\n  const tier: LLMTier = opts.tier ?? 'balanced';\n  const system = opts.system ?? messages.find((m) => m.role === 'system')?.content;\n  const tokenEstimate = estimateTokens(messages, system);\n  const route = plan(tier, opts, tokenEstimate);\n\n  // ── Org-level daily / monthly cap pre-check ──────────────────────────────\n  const kv = env.LLM_COST_KV;\n  const todayKey = `llm:daily-cost:${isoDate(now())}`;\n  const monthKey = `llm:monthly-cost:${isoMonth(now())}`;\n  if (kv) {\n    if (opts.dailyCapUsd !== undefined) {\n      const raw = await kv.get(todayKey).catch(() => null);\n      const spent = parseFloat(raw ?? '0');\n      if (spent >= opts.dailyCapUsd) {\n        return toErrorResponse(\n          new RateLimitError('LLM_DAILY_CAP_EXCEEDED', {\n            spentUsd: spent,\n            dailyCapUsd: opts.dailyCapUsd,\n          }),\n        );\n      }\n    }\n    if (opts.monthlyCapUsd !== undefined) {\n      const raw = await kv.get(monthKey).catch(() => null);\n      const spent = parseFloat(raw ?? '0');\n      if (spent >= opts.monthlyCapUsd) {\n        return toErrorResponse(\n          new RateLimitError('LLM_MONTHLY_CAP_EXCEEDED', {\n            spentUsd: spent,\n            monthlyCapUsd: opts.monthlyCapUsd,\n          }),\n        );\n      }\n    }\n  }\n\n  const attemptLog: Array<{ provider: LLMProvider; status?: number; message: string }> = [];\n\n  const routeLegs = [route.primary, route.fallback].filter(Boolean) as Array<{ provider: LLMProvider; model: string }>;\n  for (const [legIndex, leg] of routeLegs.entries()) {\n    // Skip providers that are currently in their cooldown window.\n    if (isProviderCoolingDown(leg.provider, now)) {\n      logger?.warn?.('llm.provider.coolingDown', { provider: leg.provider });\n      attemptLog.push({ provider: leg.provider, message: 'skipped: cooling down' });\n      continue;\n    }\n    if (opts.signal?.aborted) {\n      return toErrorResponse(\n        new InternalError('llm call aborted', { provider: leg.provider, model: leg.model }),\n      );\n    }\n    try {\n      const result = await callOne(leg, messages, opts, env, fetchImpl, logger, now);\n      if (!result.parsed.content) {\n        throw { provider: leg.provider, status: 200, retryable: false, message: 'empty content' } satisfies ProviderError;\n      }\n      logger?.info?.('llm.complete', {\n        provider: leg.provider,\n        model: leg.model,\n        tier,\n        tokenEstimate,\n        attempts: result.attempts,\n        runId: opts.runId,\n        project: opts.project,\n        actor: opts.actor,\n        workload: opts.workload,\n      });\n      const llmResult: LLMResult = {\n        content: result.parsed.content,\n        provider: leg.provider,\n        model: result.parsed.model ?? leg.model,\n        tier,\n        tokens: {\n          input: result.parsed.input,\n          output: result.parsed.output,\n          cacheRead: result.parsed.cacheRead,\n          cacheWrite: result.parsed.cacheWrite,\n        },\n        latency: now() - startedAt,\n        attempts: result.attempts,\n        gatewayRequestId: result.gatewayRequestId,\n      };\n      const costUsd = estimateCostUsd(llmResult.tokens, llmResult.model);\n      if (opts.maxCostUsd !== undefined && costUsd > opts.maxCostUsd) {\n        if (kv && (opts.dailyCapUsd !== undefined || opts.monthlyCapUsd !== undefined)) {\n          await recordOrgCostUsage(kv, todayKey, monthKey, costUsd, opts);\n        }\n        return toErrorResponse(\n          new RateLimitError('LLM_COST_CAP_EXCEEDED', {\n            costUsd,\n            maxCostUsd: opts.maxCostUsd,\n            model: llmResult.model,\n            tokens: llmResult.tokens,\n          }),\n        );\n      }\n      // ── Update org-level cost accumulators in KV ─────────────────────────\n      // The KV writes are best-effort. Cloudflare KV does not support atomic\n      // compare-and-swap, so concurrent increments may undercount spend.\n      if (kv && (opts.dailyCapUsd !== undefined || opts.monthlyCapUsd !== undefined)) {\n        await recordOrgCostUsage(kv, todayKey, monthKey, costUsd, opts);\n      }\n      // ── Metering callback ────────────────────────────────────────────────\n      if (deps.onRecord && opts.ledger) {\n        const row: LLMRecordRow = {\n          ...opts.ledger,\n          model: llmResult.model,\n          provider: llmResult.provider,\n          tier: llmResult.tier,\n          inputTokens: llmResult.tokens.input,\n          outputTokens: llmResult.tokens.output,\n          cacheReadTokens: llmResult.tokens.cacheRead ?? 0,\n          cacheWriteTokens: llmResult.tokens.cacheWrite ?? 0,\n          latencyMs: llmResult.latency,\n          costUsd,\n          yyyyMm: isoMonth(now()),\n        };\n        deps.onRecord(row).catch((e: unknown) => {\n          logger?.warn?.('llm.onRecord.error', { message: e instanceof Error ? e.message : String(e) });\n        });\n      }\n      return { data: llmResult, error: null };\n    } catch (e) {\n      if (e instanceof DOMException && e.name === 'AbortError') {\n        return toErrorResponse(\n          new InternalError('llm call aborted', { provider: leg.provider, model: leg.model }),\n        );\n      }\n      if (isProviderError(e)) {\n        attemptLog.push({ provider: e.provider, status: e.status, message: e.message });\n        if (e.status === 429 && legIndex === routeLegs.length - 1) {\n          return toErrorResponse(\n            new RateLimitError(`llm rate limited on ${e.provider}`, { attempts: attemptLog }),\n          );\n        }\n        logger?.warn?.('llm.leg.failed', { provider: leg.provider, status: e.status });\n        continue;\n      }\n      attemptLog.push({ provider: leg.provider, message: e instanceof Error ? e.message : String(e) });\n    }\n  }\n\n  return toErrorResponse(\n    new InternalError('LLM_ALL_PROVIDERS_FAILED', { attempts: attemptLog, tier, tokenEstimate }),\n  );\n}\n\n// ─── Streaming ────────────────────────────────────────────────────────────\n\n/**\n * Anthropic server-sent event shapes used by the streaming parser.\n * Only the fields we consume are typed; the rest are ignored.\n */\ninterface AnthropicStreamEvent {\n  type: string;\n  index?: number;\n  delta?: { type?: string; text?: string };\n  message?: {\n    usage?: {\n      input_tokens?: number;\n      output_tokens?: number;\n      cache_read_input_tokens?: number;\n      cache_creation_input_tokens?: number;\n    };\n    model?: string;\n  };\n  usage?: {\n    input_tokens?: number;\n    output_tokens?: number;\n  };\n}\n\n/**\n * Streams a completion from the primary Anthropic provider, yielding text chunks\n * as they arrive. Falls back to the non-streaming {@link complete} function when\n * the provider does not support streaming (i.e. a non-Anthropic primary is selected).\n *\n * The generator's **return value** (accessible via `gen.return()` or by consuming\n * the full iteration) is an {@link LLMResult} with the same shape as {@link complete}.\n *\n * Usage pattern:\n * ```ts\n * const gen = completionStream(messages, env, opts);\n * for await (const chunk of gen) {\n *   // stream chunk to client\n * }\n * const result = (await gen.return(undefined)).value; // LLMResult\n * ```\n *\n * @param messages - Ordered chat history.\n * @param env - API key + gateway bindings.\n * @param opts - Optional tier/model/parameters override. Accepts `deps` as nested field.\n * @returns An async generator that yields `string` chunks and returns an {@link LLMResult}.\n */\nexport async function* completionStream(\n  messages: LLMMessage[],\n  env: LLMEnv,\n  opts: LLMOptions & { deps?: LLMDeps } = {},\n): AsyncGenerator<string, LLMResult, unknown> {\n  if (messages.length === 0) {\n    throw new ValidationError('messages must not be empty');\n  }\n  if (!env.AI_GATEWAY_BASE_URL) {\n    throw new ValidationError('AI_GATEWAY_BASE_URL is required in 0.3.0');\n  }\n\n  const deps: LLMDeps = opts.deps ?? {};\n  const fetchImpl = deps.fetch ?? fetch;\n  const now = deps.now ?? (() => Date.now());\n  const logger = deps.logger;\n  const startedAt = now();\n\n  const tier: LLMTier = opts.tier ?? 'balanced';\n  const system = opts.system ?? messages.find((m) => m.role === 'system')?.content;\n  const tokenEstimate = estimateTokens(messages, system);\n  const route = plan(tier, opts, tokenEstimate);\n  const streamLeg =\n    route.primary.provider === 'grok' && !env.GROK_API_KEY && route.fallback?.provider === 'anthropic'\n      ? route.fallback\n      : route.primary;\n\n  // Only Anthropic supports streaming in the current implementation.\n  // For all other primaries, fall back to non-streaming complete().\n  if (streamLeg.provider !== 'anthropic') {\n    const result = await complete(messages, env, opts, deps);\n    if (result.error !== null || result.data === null) {\n      throw new InternalError('LLM_ALL_PROVIDERS_FAILED', { error: result.error });\n    }\n    yield result.data.content;\n    return result.data;\n  }\n\n  // Check cooldown before attempting the streaming call.\n  if (isProviderCoolingDown(streamLeg.provider, now)) {\n    logger?.warn?.('llm.provider.coolingDown', { provider: streamLeg.provider });\n    // Fall back to non-streaming complete() which will handle the fallback leg.\n    const result = await complete(messages, env, opts, deps);\n    if (result.error !== null || result.data === null) {\n      throw new InternalError('LLM_ALL_PROVIDERS_FAILED', { error: result.error });\n    }\n    yield result.data.content;\n    return result.data;\n  }\n\n  const req = buildAnthropicRequest(streamLeg.model, messages, opts, env, true);\n  // Attribution for the shared AI Gateway (matches the non-streaming path).\n  const streamAigMetadata = buildAigMetadata(opts);\n  if (streamAigMetadata) req.headers['cf-aig-metadata'] = streamAigMetadata;\n\n  let response: Response;\n  try {\n    response = await fetchImpl(req.url, {\n      method: 'POST',\n      headers: req.headers,\n      body: req.body,\n      // Fall back to a 60 s default when the caller provides no signal — prevents\n      // a hung provider connection from consuming the Worker's wall-clock budget.\n      signal: opts.signal ?? AbortSignal.timeout(60_000),\n    });\n  } catch (e) {\n    if (e instanceof DOMException && e.name === 'AbortError') {\n      throw new InternalError('llm call aborted', {\n        provider: streamLeg.provider,\n        model: streamLeg.model,\n      });\n    }\n    throw new InternalError('llm stream fetch failed', {\n      message: e instanceof Error ? e.message : String(e),\n    });\n  }\n\n  if (!response.ok) {\n    const text = await response.text().catch(() => '');\n    const retryable = isRetryableForBackoff(response.status);\n    if (retryable && response.status === 429) {\n      markProviderCoolingDown(streamLeg.provider, now);\n    }\n    // Fall back to non-streaming complete() which will try the fallback leg.\n    const result = await complete(messages, env, opts, deps);\n    if (result.error !== null || result.data === null) {\n      throw new InternalError('LLM_ALL_PROVIDERS_FAILED', {\n        streamError: `${streamLeg.provider} ${String(response.status)}: ${text.slice(0, 300)}`,\n        error: result.error,\n      });\n    }\n    yield result.data.content;\n    return result.data;\n  }\n\n  if (!response.body) {\n    throw new InternalError('llm stream response body is null', {\n      provider: streamLeg.provider,\n    });\n  }\n\n  // Stream SSE events from Anthropic.\n  const decoder = new TextDecoder();\n  let accumulatedText = '';\n  let inputTokens = 0;\n  let outputTokens = 0;\n  let cacheRead = 0;\n  let cacheWrite = 0;\n  let modelName: string | undefined;\n  const gatewayRequestId: string | undefined = response.headers.get('cf-aig-request-id') ?? undefined;\n\n  const reader = response.body.getReader();\n  let buffer = '';\n\n  try {\n    while (true) {\n      const { done, value } = await reader.read();\n      if (done) break;\n      buffer += decoder.decode(value, { stream: true });\n\n      // SSE lines are delimited by '\\n'. Events are separated by '\\n\\n'.\n      const lines = buffer.split('\\n');\n      // Keep the last (potentially incomplete) line in the buffer.\n      buffer = lines.pop() ?? '';\n\n      for (const line of lines) {\n        if (!line.startsWith('data: ')) continue;\n        const data = line.slice(6).trim();\n        if (data === '[DONE]') break;\n        let event: AnthropicStreamEvent;\n        try {\n          event = JSON.parse(data) as AnthropicStreamEvent;\n        } catch {\n          continue; // Skip malformed SSE lines.\n        }\n\n        switch (event.type) {\n          case 'message_start':\n            inputTokens = event.message?.usage?.input_tokens ?? 0;\n            cacheRead = event.message?.usage?.cache_read_input_tokens ?? 0;\n            cacheWrite = event.message?.usage?.cache_creation_input_tokens ?? 0;\n            modelName = event.message?.model;\n            break;\n          case 'content_block_delta':\n            if (event.delta?.type === 'text_delta' && typeof event.delta.text === 'string') {\n              accumulatedText += event.delta.text;\n              yield event.delta.text;\n            }\n            break;\n          case 'message_delta':\n            outputTokens = event.usage?.output_tokens ?? outputTokens;\n            break;\n          default:\n            break;\n        }\n      }\n    }\n  } finally {\n    reader.releaseLock();\n  }\n\n  clearProviderCooldown(streamLeg.provider);\n  logger?.info?.('llm.completionStream', {\n    provider: streamLeg.provider,\n    model: streamLeg.model,\n    tier,\n    tokenEstimate,\n    runId: opts.runId,\n    project: opts.project,\n    actor: opts.actor,\n    workload: opts.workload,\n  });\n\n  return {\n    content: accumulatedText,\n    provider: streamLeg.provider,\n    model: modelName ?? streamLeg.model,\n    tier,\n    tokens: { input: inputTokens, output: outputTokens, cacheRead, cacheWrite },\n    latency: now() - startedAt,\n    attempts: 1,\n    gatewayRequestId,\n  };\n}\n\n// ─── Grounding assertion ───────────────────────────────────────────────────\n\n/**\n * Returns `true` if `response` contains at least one verbatim phrase of at\n * least 5 consecutive whitespace-delimited tokens that also appears in one of\n * the `sources` strings.\n *\n * Returns `true` unconditionally when `sources` is empty (no grounding\n * documents means grounding cannot be violated).\n *\n * This is a lightweight guard for RAG pipelines — it detects obvious\n * hallucinations where the model generates content not present in any\n * retrieved source. It is NOT a semantic similarity check.\n *\n * @param response - The LLM-generated text to inspect.\n * @param sources - Retrieved source documents to check against.\n * @returns `true` if the response is grounded, `false` if hallucination detected.\n *\n * @example\n * ```ts\n * const grounded = assertGrounding(llmAnswer, retrievedDocs);\n * if (!grounded) {\n *   // flag or re-rank the response\n * }\n * ```\n */\nexport function assertGrounding(response: string, sources: string[]): boolean {\n  if (sources.length === 0) return true;\n\n  const WINDOW = 5;\n  const responseTokens = response.split(/\\s+/).filter((t) => t.length > 0);\n\n  if (responseTokens.length < WINDOW) return false;\n\n  // Build a set of all 5-token ngrams from each source for O(n) lookup.\n  const sourceNgrams = new Set<string>();\n  for (const source of sources) {\n    const tokens = source.split(/\\s+/).filter((t) => t.length > 0);\n    for (let i = 0; i <= tokens.length - WINDOW; i++) {\n      const ngram = tokens.slice(i, i + WINDOW).join(' ');\n      sourceNgrams.add(ngram);\n    }\n  }\n\n  if (sourceNgrams.size === 0) return false;\n\n  // Slide a window of WINDOW tokens over the response and check for a match.\n  for (let i = 0; i <= responseTokens.length - WINDOW; i++) {\n    const ngram = responseTokens.slice(i, i + WINDOW).join(' ');\n    if (sourceNgrams.has(ngram)) return true;\n  }\n\n  return false;\n}\n\n// ─── Exported helpers (kept for existing consumers) ───────────────────────\n\nexport { MODELS, isProviderCoolingDown, markProviderCoolingDown, clearProviderCooldown, PROVIDER_COOLDOWN_MS };\nexport { BASE_BACKOFF_MS };\n"],"mappings":";AAAA;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,OAEK;AA4LP,IAAM,SAAS;AAAA,EACb,WAAW;AAAA,IACT,MAAM;AAAA,IACN,UAAU;AAAA,IACV,OAAO;AAAA,EACT;AAAA,EACA,QAAQ;AAAA,IACN,OAAO;AAAA,EACT;AAAA,EACA,MAAM;AAAA,IACJ,UAAU;AAAA,EACZ;AAAA,EACA,MAAM;AAAA,IACJ,MAAM;AAAA,EACR;AAAA,EACA,UAAU;AAAA,IACR,WAAW;AAAA,EACb;AACF;AAEA,IAAM,qBAAqB;AAC3B,IAAM,sBAAsB;AAC5B,IAAM,iCAAiC;AAIvC,IAAM,kBAAkB;AAExB,IAAM,iBAAiB;AAEvB,IAAM,wBAAwB;AAE9B,IAAM,4BAA4B;AAQlC,IAAM,wBAAkD,oBAAI,IAAI;AAGhE,IAAM,uBAAuB;AAM7B,SAAS,sBAAsB,UAAuB,MAAoB,KAAK,KAAc;AAC3F,QAAM,QAAQ,sBAAsB,IAAI,QAAQ;AAChD,MAAI,UAAU,OAAW,QAAO;AAChC,SAAO,IAAI,IAAI;AACjB;AAGA,SAAS,QAAQ,OAAuB;AACtC,SAAO,IAAI,KAAK,KAAK,EAAE,YAAY,EAAE,MAAM,GAAG,EAAE;AAClD;AAOA,eAAe,mBACb,IACA,UACA,UACA,SACA,MACe;AACf,MAAI,KAAK,gBAAgB,QAAW;AAClC,UAAM,MAAM,MAAM,GAAG,IAAI,QAAQ,EAAE,MAAM,MAAM,IAAI;AACnD,UAAM,QAAQ,WAAW,OAAO,GAAG;AACnC,UAAM,GAAG,IAAI,UAAU,OAAO,QAAQ,OAAO,GAAG;AAAA,MAAE,eAAe;AAAA;AAAA,IAAmB,CAAC,EAAE,MAAM,MAAM,MAAS;AAAA,EAC9G;AACA,MAAI,KAAK,kBAAkB,QAAW;AACpC,UAAM,MAAM,MAAM,GAAG,IAAI,QAAQ,EAAE,MAAM,MAAM,IAAI;AACnD,UAAM,QAAQ,WAAW,OAAO,GAAG;AACnC,UAAM,GAAG,IAAI,UAAU,OAAO,QAAQ,OAAO,GAAG;AAAA,MAAE,eAAe;AAAA;AAAA,IAAqB,CAAC,EAAE,MAAM,MAAM,MAAS;AAAA,EAChH;AACF;AAYO,IAAM,qBAA+G;AAAA;AAAA,EAE1H,2BAA2B,EAAE,OAAO,KAAM,QAAQ,GAAM,WAAW,MAAM,YAAY,EAAK;AAAA,EAC1F,6BAA6B,EAAE,OAAO,KAAM,QAAQ,GAAM,WAAW,MAAM,YAAY,EAAK;AAAA;AAAA,EAE5F,4BAA4B,EAAE,OAAO,GAAM,QAAQ,IAAO,WAAW,KAAM,YAAY,KAAK;AAAA,EAC5F,qBAAqB,EAAE,OAAO,GAAM,QAAQ,IAAO,WAAW,KAAM,YAAY,KAAK;AAAA;AAAA,EAErF,0BAA0B,EAAE,OAAO,IAAO,QAAQ,IAAO,WAAW,KAAM,YAAY,MAAM;AAAA,EAC5F,mBAAmB,EAAE,OAAO,IAAO,QAAQ,IAAO,WAAW,KAAM,YAAY,MAAM;AAAA;AAAA,EAErF,kBAAkB,EAAE,OAAO,MAAM,QAAQ,IAAO,WAAW,MAAM,YAAY,IAAK;AAAA;AAAA,EAElF,oBAAoB,EAAE,OAAO,KAAM,QAAQ,MAAM,WAAW,MAAM,YAAY,IAAK;AAAA;AAAA,EAEnF,YAAY,EAAE,OAAO,MAAM,QAAQ,KAAM,WAAW,GAAM,YAAY,EAAK;AAAA;AAAA,EAE3E,iBAAiB,EAAE,OAAO,MAAM,QAAQ,KAAM,WAAW,MAAM,YAAY,KAAK;AAAA,EAChF,qBAAqB,EAAE,OAAO,MAAM,QAAQ,MAAM,WAAW,MAAM,YAAY,KAAK;AAAA;AAAA,EAEpF,eAAe,EAAE,OAAO,MAAM,QAAQ,KAAM,WAAW,GAAM,YAAY,EAAK;AAAA,EAC9E,sBAAsB,EAAE,OAAO,MAAM,QAAQ,KAAM,WAAW,GAAM,YAAY,EAAK;AACvF;AAGA,IAAM,iBAAiB,mBAAmB,iBAAiB;AAO3D,SAAS,gBACP,QACA,OACQ;AACR,QAAM,QAAQ,mBAAmB,KAAK,KAAK;AAC3C,UACG,OAAO,QAAQ,MAAM,QACpB,OAAO,SAAS,MAAM,UACrB,OAAO,aAAa,KAAK,MAAM,aAC/B,OAAO,cAAc,KAAK,MAAM,cACnC;AAEJ;AAGA,SAAS,SAAS,OAAuB;AACvC,SAAO,IAAI,KAAK,KAAK,EAAE,YAAY,EAAE,MAAM,GAAG,CAAC;AACjD;AAKA,SAAS,wBAAwB,UAAuB,MAAoB,KAAK,KAAW;AAC1F,wBAAsB,IAAI,UAAU,IAAI,IAAI,oBAAoB;AAClE;AAKA,SAAS,sBAAsB,UAA6B;AAC1D,wBAAsB,OAAO,QAAQ;AACvC;AAGA,IAAM,kBAAkB;AAaxB,SAAS,sBAAsB,QAAyB;AACtD,SAAO,WAAW,OAAQ,UAAU,OAAO,SAAS;AACtD;AAEA,SAAS,eAAe,UAAwB,QAAyB;AAEvE,MAAI,QAAQ,QAAQ,UAAU;AAC9B,aAAW,KAAK,SAAU,UAAS,EAAE,QAAQ;AAC7C,SAAO,KAAK,KAAK,QAAQ,CAAC;AAC5B;AAEA,SAAS,MAAM,IAAY,QAAqC;AAC9D,SAAO,IAAI,QAAQ,CAAC,SAAS,WAAW;AACtC,UAAM,IAAI,WAAW,SAAS,EAAE;AAChC,QAAI,QAAQ;AACV,YAAM,UAAU,MAAM;AACpB,qBAAa,CAAC;AACd,eAAO,IAAI,aAAa,WAAW,YAAY,CAAC;AAAA,MAClD;AACA,UAAI,OAAO,QAAS,SAAQ;AAAA,UACvB,QAAO,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;AAAA,IAC/D;AAAA,EACF,CAAC;AACH;AAUA,SAAS,iBAAiB,SAAyB;AACjD,QAAM,SAAS,KAAK,MAAM,KAAK,OAAO,IAAI,qBAAqB;AAC/D,SAAO,KAAK,IAAI,kBAAkB,KAAK,IAAI,GAAG,OAAO,IAAI,QAAQ,cAAc;AACjF;AAIA,SAAS,sBACP,OACA,UACA,MACA,KACA,YAAY,OACoD;AAChE,QAAM,MAAM,KAAK,UAAU,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,QAAQ,GAAG;AACtE,QAAM,WAAW,SAAS,OAAO,CAAC,MAAM,EAAE,SAAS,QAAQ;AAC3D,QAAM,OAAgC;AAAA,IACpC;AAAA,IACA,YAAY,KAAK,aAAa;AAAA,IAC9B,aAAa,KAAK,eAAe;AAAA,IACjC,UAAU,SAAS,IAAI,CAAC,OAAO,EAAE,MAAM,EAAE,MAAM,SAAS,EAAE,QAAQ,EAAE;AAAA,EACtE;AACA,MAAI,WAAW;AACb,SAAK,SAAS;AAAA,EAChB;AACA,MAAI,KAAK;AACP,UAAM,QAAQ,KAAK,eAAe,IAAI,UAAU;AAChD,SAAK,SAAS,QACV,CAAC,EAAE,MAAM,QAAQ,MAAM,KAAK,eAAe,EAAE,MAAM,YAAY,EAAE,CAAC,IAClE;AAAA,EACN;AACA,SAAO;AAAA,IACL,KAAK,GAAG,IAAI,mBAAmB;AAAA,IAC/B,SAAS;AAAA,MACP,gBAAgB;AAAA,MAChB,aAAa,IAAI;AAAA,MACjB,qBAAqB;AAAA,MACrB,kBAAkB;AAAA,IACpB;AAAA,IACA,MAAM,KAAK,UAAU,IAAI;AAAA,EAC3B;AACF;AAEA,SAAS,mBACP,OACA,UACA,MACA,KACgE;AAChE,QAAM,MAAM,KAAK,UAAU,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,QAAQ,GAAG;AACtE,QAAM,WAAW,SACd,OAAO,CAAC,MAAM,EAAE,SAAS,QAAQ,EACjC,IAAI,CAAC,OAAO;AAAA,IACX,MAAM,EAAE,SAAS,cAAc,UAAU;AAAA,IACzC,OAAO,CAAC,EAAE,MAAM,EAAE,QAAQ,CAAC;AAAA,EAC7B,EAAE;AACJ,QAAM,OAAgC;AAAA,IACpC;AAAA,IACA,kBAAkB;AAAA,MAChB,iBAAiB,KAAK,aAAa;AAAA,MACnC,aAAa,KAAK,eAAe;AAAA,IACnC;AAAA,EACF;AACA,MAAI,KAAK;AACP,SAAK,oBAAoB,EAAE,OAAO,CAAC,EAAE,MAAM,IAAI,CAAC,EAAE;AAAA,EACpD;AACA,QAAM,OAAO,eAAe,IAAI,cAAc,cAAc,IAAI,eAAe,6BAA6B,KAAK;AACjH,SAAO;AAAA,IACL,KAAK,GAAG,IAAI,mBAAmB,qBAAqB,IAAI;AAAA,IACxD,SAAS;AAAA,MACP,gBAAgB;AAAA,MAChB,eAAe,UAAU,IAAI,mBAAmB;AAAA,IAClD;AAAA,IACA,MAAM,KAAK,UAAU,IAAI;AAAA,EAC3B;AACF;AAEA,SAAS,iBACP,OACA,UACA,MACA,KACgE;AAChE,QAAM,MAAM,KAAK,UAAU,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,QAAQ,GAAG;AACtE,QAAM,SAAuB,CAAC;AAC9B,MAAI,IAAK,QAAO,KAAK,EAAE,MAAM,UAAU,SAAS,IAAI,CAAC;AACrD,aAAW,KAAK,SAAU,KAAI,EAAE,SAAS,SAAU,QAAO,KAAK,CAAC;AAChE,SAAO;AAAA,IACL,KAAK,GAAG,IAAI,mBAAmB;AAAA,IAC/B,SAAS;AAAA,MACP,gBAAgB;AAAA,MAChB,eAAe,UAAU,IAAI,YAAY;AAAA,IAC3C;AAAA,IACA,MAAM,KAAK,UAAU;AAAA,MACnB;AAAA,MACA,YAAY,KAAK,aAAa;AAAA,MAC9B,aAAa,KAAK,eAAe;AAAA,MACjC,UAAU;AAAA,IACZ,CAAC;AAAA,EACH;AACF;AAEA,SAAS,iBACP,OACA,UACA,MACA,KACgE;AAChE,MAAI,CAAC,IAAI,cAAc;AACrB,UAAM,IAAI,gBAAgB,iDAAiD;AAAA,EAC7E;AACA,QAAM,MAAM,KAAK,UAAU,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,QAAQ,GAAG;AACtE,QAAM,SAAuB,CAAC;AAC9B,MAAI,IAAK,QAAO,KAAK,EAAE,MAAM,UAAU,SAAS,IAAI,CAAC;AACrD,aAAW,KAAK,SAAU,KAAI,EAAE,SAAS,SAAU,QAAO,KAAK,CAAC;AAChE,QAAM,OAAgC;AAAA,IACpC;AAAA,IACA,YAAY,KAAK,aAAa;AAAA,IAC9B,aAAa,KAAK,eAAe;AAAA,IACjC,UAAU;AAAA,EACZ;AACA,MAAI,UAAU,OAAO,KAAK,MAAM;AAC9B,SAAK,mBAAmB,KAAK,mBAAmB;AAAA,EAClD;AACA,SAAO;AAAA,IACL,KAAK,GAAG,IAAI,mBAAmB;AAAA,IAC/B,SAAS;AAAA,MACP,gBAAgB;AAAA,MAChB,eAAe,UAAU,IAAI,YAAY;AAAA,IAC3C;AAAA,IACA,MAAM,KAAK,UAAU,IAAI;AAAA,EAC3B;AACF;AAEA,SAAS,qBACP,OACA,UACA,MACA,KACgE;AAChE,MAAI,CAAC,IAAI,kBAAkB;AACzB,UAAM,IAAI,gBAAgB,2EAA2E;AAAA,EACvG;AACA,QAAM,MAAM,KAAK,UAAU,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,QAAQ,GAAG;AACtE,QAAM,SAAuB,CAAC;AAC9B,MAAI,IAAK,QAAO,KAAK,EAAE,MAAM,UAAU,SAAS,IAAI,CAAC;AACrD,aAAW,KAAK,SAAU,KAAI,EAAE,SAAS,SAAU,QAAO,KAAK,CAAC;AAChE,SAAO;AAAA,IACL,KAAK,GAAG,IAAI,mBAAmB;AAAA,IAC/B,SAAS;AAAA,MACP,gBAAgB;AAAA,MAChB,eAAe,UAAU,IAAI,gBAAgB;AAAA,IAC/C;AAAA,IACA,MAAM,KAAK,UAAU;AAAA,MACnB;AAAA,MACA,YAAY,KAAK,aAAa;AAAA,MAC9B,aAAa,KAAK,eAAe;AAAA,MACjC,UAAU;AAAA,IACZ,CAAC;AAAA,EACH;AACF;AA6BA,SAAS,eACP,MAC2G;AAC3G,QAAM,IAAI;AACV,SAAO;AAAA,IACL,SAAS,EAAE,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,MAAM,GAAG,QAAQ;AAAA,IAC5D,OAAO,EAAE,OAAO,gBAAgB;AAAA,IAChC,QAAQ,EAAE,OAAO,iBAAiB;AAAA,IAClC,WAAW,EAAE,OAAO,2BAA2B;AAAA,IAC/C,YAAY,EAAE,OAAO,+BAA+B;AAAA,IACpD,OAAO,EAAE;AAAA,EACX;AACF;AAEA,SAAS,YAAY,MAAmE;AACtF,QAAM,IAAI;AACV,QAAM,OACJ,EAAE,aAAa,CAAC,GAAG,SAAS,OAAO,IAAI,CAAC,MAAM,EAAE,QAAQ,EAAE,EAAE,KAAK,EAAE,KAAK;AAC1E,SAAO;AAAA,IACL,SAAS;AAAA,IACT,OAAO,EAAE,eAAe,oBAAoB;AAAA,IAC5C,QAAQ,EAAE,eAAe,wBAAwB;AAAA,EACnD;AACF;AAEA,SAAS,UAAU,MAAmF;AACpG,QAAM,IAAI;AACV,SAAO;AAAA,IACL,SAAS,EAAE,UAAU,CAAC,GAAG,SAAS,WAAW;AAAA,IAC7C,OAAO,EAAE,OAAO,iBAAiB;AAAA,IACjC,QAAQ,EAAE,OAAO,qBAAqB;AAAA,IACtC,OAAO,EAAE;AAAA,EACX;AACF;AAmBA,eAAe,gBACb,UACA,SACA,WACA,QACA,QACA,OACyE;AAMzE,WAAS,gBAAgB,KAA2B;AAClD,4BAAwB,UAAU,SAAS,KAAK,GAAG;AACnD,UAAM;AAAA,EACR;AAEA,MAAI;AACJ,WAAS,UAAU,GAAG,WAAW,2BAA2B,WAAW;AACrE,QAAI;AACF,YAAM,WAAW,MAAM,UAAU,QAAQ,KAAK;AAAA,QAC5C,QAAQ;AAAA,QACR,SAAS,QAAQ;AAAA,QACjB,MAAM,QAAQ;AAAA,QACd;AAAA,MACF,CAAC;AACD,UAAI,CAAC,SAAS,IAAI;AAChB,cAAM,OAAO,MAAM,SAAS,KAAK,EAAE,MAAM,MAAM,EAAE;AACjD,cAAM,YAAY,sBAAsB,SAAS,MAAM;AACvD,cAAM,MAAqB;AAAA,UACzB;AAAA,UACA,QAAQ,SAAS;AAAA,UACjB;AAAA,UACA,SAAS,GAAG,QAAQ,IAAI,OAAO,SAAS,MAAM,CAAC,KAAK,KAAK,MAAM,GAAG,GAAG,CAAC;AAAA,QACxE;AACA,gBAAQ,OAAO,sBAAsB,EAAE,UAAU,QAAQ,SAAS,QAAQ,QAAQ,CAAC;AACnF,YAAI,CAAC,IAAI,aAAa,YAAY,2BAA2B;AAC3D,cAAI,IAAI,UAAW,iBAAgB,GAAG;AACtC,gBAAM;AAAA,QACR;AACA,kBAAU;AAAA,MACZ,OAAO;AACL,cAAM,mBAAmB,SAAS,QAAQ,IAAI,mBAAmB,KAAK;AACtE,8BAAsB,QAAQ;AAC9B,eAAO,EAAE,MAAM,MAAM,SAAS,KAAK,GAAG,kBAAkB,UAAU,QAAQ;AAAA,MAC5E;AAAA,IACF,SAAS,GAAG;AACV,UAAI,aAAa,gBAAgB,EAAE,SAAS,aAAc,OAAM;AAChE,UAAI,OAAO,MAAM,YAAY,MAAM,QAAQ,eAAe,GAAG;AAC3D,cAAM,MAAM;AACZ,YAAI,CAAC,IAAI,aAAa,YAAY,2BAA2B;AAC3D,cAAI,IAAI,UAAW,iBAAgB,GAAG;AACtC,gBAAM;AAAA,QACR;AACA,kBAAU;AAAA,MACZ,OAAO;AACL,cAAM,MAAqB;AAAA,UACzB;AAAA,UACA,QAAQ;AAAA,UACR,WAAW;AAAA,UACX,SAAS,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC;AAAA,QACpD;AACA,YAAI,YAAY,0BAA2B,iBAAgB,GAAG;AAC9D,kBAAU;AAAA,MACZ;AAAA,IACF;AAEA,UAAM,YAAY,iBAAiB,UAAU,CAAC;AAC9C,UAAM,MAAM,WAAW,MAAM;AAAA,EAC/B;AAEA,0BAAwB,UAAU,SAAS,KAAK,GAAG;AACnD,QAAM,WAAY,EAAE,UAAU,QAAQ,GAAG,WAAW,OAAO,SAAS,YAAY;AAClF;AAEA,SAAS,gBAAgB,KAAoC;AAC3D,SACE,OAAO,QAAQ,YACf,QAAQ,QACR,OAAQ,IAA6B,WAAW,YAChD,OAAQ,IAA8B,YAAY,YAClD,OAAQ,IAA+B,aAAa;AAExD;AASA,SAAS,KAAK,MAAe,MAAkB,eAAkC;AAC/E,MAAI,KAAK,OAAO;AAEd,UAAM,IAAI,KAAK;AACf,QAAI,EAAE,WAAW,QAAQ,EAAG,QAAO,EAAE,SAAS,EAAE,UAAU,aAAa,OAAO,EAAE,EAAE;AAClF,QAAI,EAAE,WAAW,QAAQ,EAAG,QAAO,EAAE,SAAS,EAAE,UAAU,UAAU,OAAO,EAAE,EAAE;AAC/E,QAAI,EAAE,WAAW,MAAM,EAAG,QAAO,EAAE,SAAS,EAAE,UAAU,QAAQ,OAAO,EAAE,EAAE;AAC3E,QAAI,EAAE,WAAW,UAAU,EAAG,QAAO,EAAE,SAAS,EAAE,UAAU,YAAY,OAAO,EAAE,EAAE;AACnF,WAAO,EAAE,SAAS,EAAE,UAAU,QAAQ,OAAO,EAAE,EAAE;AAAA,EACnD;AACA,QAAM,cAAc,kBAAkB,KAAK,wBAAwB;AACnE,UAAQ,MAAM;AAAA,IACZ,KAAK;AACH,aAAO;AAAA,QACL,SAAS,EAAE,UAAU,YAAY,OAAO,OAAO,SAAS,UAAU;AAAA,QAClE,UAAU,EAAE,UAAU,QAAQ,OAAO,OAAO,KAAK,SAAS;AAAA,MAC5D;AAAA,IACF,KAAK;AACH,aAAO,EAAE,SAAS,EAAE,UAAU,QAAQ,OAAO,OAAO,KAAK,SAAS,EAAE;AAAA,IACtE,KAAK;AACH,aAAO,cACH;AAAA,QACE,SAAS,EAAE,UAAU,UAAU,OAAO,OAAO,OAAO,MAAM;AAAA,QAC1D,UAAU,EAAE,UAAU,aAAa,OAAO,OAAO,UAAU,MAAM;AAAA,MACnE,IACA;AAAA,QACE,SAAS,EAAE,UAAU,aAAa,OAAO,OAAO,UAAU,MAAM;AAAA,QAChE,UAAU,EAAE,UAAU,UAAU,OAAO,OAAO,OAAO,MAAM;AAAA,MAC7D;AAAA,IACN,KAAK;AACH,aAAO;AAAA,QACL,SAAS,EAAE,UAAU,QAAQ,OAAO,OAAO,KAAK,KAAK;AAAA,QACrD,UAAU,EAAE,UAAU,aAAa,OAAO,OAAO,UAAU,KAAK;AAAA,MAClE;AAAA,IACF,KAAK;AAAA,IACL;AACE,aAAO,cACH;AAAA,QACE,SAAS,EAAE,UAAU,UAAU,OAAO,OAAO,OAAO,MAAM;AAAA,QAC1D,UAAU,EAAE,UAAU,aAAa,OAAO,OAAO,UAAU,SAAS;AAAA,MACtE,IACA;AAAA,QACE,SAAS,EAAE,UAAU,aAAa,OAAO,OAAO,UAAU,SAAS;AAAA,QACnE,UAAU,EAAE,UAAU,UAAU,OAAO,OAAO,OAAO,MAAM;AAAA,MAC7D;AAAA,EACR;AACF;AAcA,SAAS,iBAAiB,MAAsC;AAC9D,QAAM,OAA+B,CAAC;AACtC,MAAI,KAAK,QAAS,MAAK,UAAU,KAAK;AACtC,MAAI,KAAK,SAAU,MAAK,WAAW,KAAK;AACxC,MAAI,KAAK,MAAO,MAAK,QAAQ,KAAK;AAClC,MAAI,KAAK,MAAO,MAAK,QAAQ,KAAK;AAClC,SAAO,OAAO,KAAK,IAAI,EAAE,SAAS,IAAI,KAAK,UAAU,IAAI,IAAI;AAC/D;AAEA,eAAe,QACb,KACA,UACA,MACA,KACA,WACA,QACA,OAC+K;AAC/K,MAAI;AACJ,UAAQ,IAAI,UAAU;AAAA,IACpB,KAAK;AACH,YAAM,sBAAsB,IAAI,OAAO,UAAU,MAAM,GAAG;AAC1D;AAAA,IACF,KAAK;AACH,YAAM,mBAAmB,IAAI,OAAO,UAAU,MAAM,GAAG;AACvD;AAAA,IACF,KAAK;AACH,YAAM,iBAAiB,IAAI,OAAO,UAAU,MAAM,GAAG;AACrD;AAAA,IACF,KAAK;AACH,YAAM,iBAAiB,IAAI,OAAO,UAAU,MAAM,GAAG;AACrD;AAAA,IACF,KAAK;AACH,YAAM,qBAAqB,IAAI,OAAO,UAAU,MAAM,GAAG;AACzD;AAAA,EACJ;AAEA,QAAM,cAAc,iBAAiB,IAAI;AACzC,MAAI,YAAa,KAAI,QAAQ,iBAAiB,IAAI;AAClD,QAAM,EAAE,MAAM,kBAAkB,SAAS,IAAI,MAAM;AAAA,IACjD,IAAI;AAAA,IACJ;AAAA,IACA;AAAA,IACA,KAAK;AAAA,IACL;AAAA,IACA;AAAA,EACF;AACA,UAAQ,IAAI,UAAU;AAAA,IACpB,KAAK;AACH,aAAO,EAAE,QAAQ,eAAe,IAAI,GAAG,kBAAkB,SAAS;AAAA,IACpE,KAAK;AACH,aAAO,EAAE,QAAQ,YAAY,IAAI,GAAG,kBAAkB,SAAS;AAAA,IACjE,KAAK;AACH,aAAO,EAAE,QAAQ,UAAU,IAAI,GAAG,kBAAkB,SAAS;AAAA,IAC/D,KAAK;AACH,aAAO,EAAE,QAAQ,UAAU,IAAI,GAAG,kBAAkB,SAAS;AAAA,IAC/D,KAAK;AACH,aAAO,EAAE,QAAQ,UAAU,IAAI,GAAG,kBAAkB,SAAS;AAAA,EACjE;AACF;AA0BA,eAAsB,SACpB,UACA,KACA,OAAmB,CAAC,GACpB,OAAgB,CAAC,GACoB;AACrC,MAAI,SAAS,WAAW,GAAG;AACzB,UAAM,IAAI,gBAAgB,4BAA4B;AAAA,EACxD;AACA,MAAI,CAAC,IAAI,qBAAqB;AAC5B,UAAM,IAAI,gBAAgB,0CAA0C;AAAA,EACtE;AACA,QAAM,YAAY,KAAK,SAAS;AAChC,QAAM,MAAM,KAAK,QAAQ,MAAM,KAAK,IAAI;AACxC,QAAM,SAAS,KAAK;AACpB,QAAM,YAAY,IAAI;AAEtB,QAAM,OAAgB,KAAK,QAAQ;AACnC,QAAM,SAAS,KAAK,UAAU,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,QAAQ,GAAG;AACzE,QAAM,gBAAgB,eAAe,UAAU,MAAM;AACrD,QAAM,QAAQ,KAAK,MAAM,MAAM,aAAa;AAG5C,QAAM,KAAK,IAAI;AACf,QAAM,WAAW,kBAAkB,QAAQ,IAAI,CAAC,CAAC;AACjD,QAAM,WAAW,oBAAoB,SAAS,IAAI,CAAC,CAAC;AACpD,MAAI,IAAI;AACN,QAAI,KAAK,gBAAgB,QAAW;AAClC,YAAM,MAAM,MAAM,GAAG,IAAI,QAAQ,EAAE,MAAM,MAAM,IAAI;AACnD,YAAM,QAAQ,WAAW,OAAO,GAAG;AACnC,UAAI,SAAS,KAAK,aAAa;AAC7B,eAAO;AAAA,UACL,IAAI,eAAe,0BAA0B;AAAA,YAC3C,UAAU;AAAA,YACV,aAAa,KAAK;AAAA,UACpB,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF;AACA,QAAI,KAAK,kBAAkB,QAAW;AACpC,YAAM,MAAM,MAAM,GAAG,IAAI,QAAQ,EAAE,MAAM,MAAM,IAAI;AACnD,YAAM,QAAQ,WAAW,OAAO,GAAG;AACnC,UAAI,SAAS,KAAK,eAAe;AAC/B,eAAO;AAAA,UACL,IAAI,eAAe,4BAA4B;AAAA,YAC7C,UAAU;AAAA,YACV,eAAe,KAAK;AAAA,UACtB,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,QAAM,aAAiF,CAAC;AAExF,QAAM,YAAY,CAAC,MAAM,SAAS,MAAM,QAAQ,EAAE,OAAO,OAAO;AAChE,aAAW,CAAC,UAAU,GAAG,KAAK,UAAU,QAAQ,GAAG;AAEjD,QAAI,sBAAsB,IAAI,UAAU,GAAG,GAAG;AAC5C,cAAQ,OAAO,4BAA4B,EAAE,UAAU,IAAI,SAAS,CAAC;AACrE,iBAAW,KAAK,EAAE,UAAU,IAAI,UAAU,SAAS,wBAAwB,CAAC;AAC5E;AAAA,IACF;AACA,QAAI,KAAK,QAAQ,SAAS;AACxB,aAAO;AAAA,QACL,IAAI,cAAc,oBAAoB,EAAE,UAAU,IAAI,UAAU,OAAO,IAAI,MAAM,CAAC;AAAA,MACpF;AAAA,IACF;AACA,QAAI;AACF,YAAM,SAAS,MAAM,QAAQ,KAAK,UAAU,MAAM,KAAK,WAAW,QAAQ,GAAG;AAC7E,UAAI,CAAC,OAAO,OAAO,SAAS;AAC1B,cAAM,EAAE,UAAU,IAAI,UAAU,QAAQ,KAAK,WAAW,OAAO,SAAS,gBAAgB;AAAA,MAC1F;AACA,cAAQ,OAAO,gBAAgB;AAAA,QAC7B,UAAU,IAAI;AAAA,QACd,OAAO,IAAI;AAAA,QACX;AAAA,QACA;AAAA,QACA,UAAU,OAAO;AAAA,QACjB,OAAO,KAAK;AAAA,QACZ,SAAS,KAAK;AAAA,QACd,OAAO,KAAK;AAAA,QACZ,UAAU,KAAK;AAAA,MACjB,CAAC;AACD,YAAM,YAAuB;AAAA,QAC3B,SAAS,OAAO,OAAO;AAAA,QACvB,UAAU,IAAI;AAAA,QACd,OAAO,OAAO,OAAO,SAAS,IAAI;AAAA,QAClC;AAAA,QACA,QAAQ;AAAA,UACN,OAAO,OAAO,OAAO;AAAA,UACrB,QAAQ,OAAO,OAAO;AAAA,UACtB,WAAW,OAAO,OAAO;AAAA,UACzB,YAAY,OAAO,OAAO;AAAA,QAC5B;AAAA,QACA,SAAS,IAAI,IAAI;AAAA,QACjB,UAAU,OAAO;AAAA,QACjB,kBAAkB,OAAO;AAAA,MAC3B;AACA,YAAM,UAAU,gBAAgB,UAAU,QAAQ,UAAU,KAAK;AACjE,UAAI,KAAK,eAAe,UAAa,UAAU,KAAK,YAAY;AAC9D,YAAI,OAAO,KAAK,gBAAgB,UAAa,KAAK,kBAAkB,SAAY;AAC9E,gBAAM,mBAAmB,IAAI,UAAU,UAAU,SAAS,IAAI;AAAA,QAChE;AACA,eAAO;AAAA,UACL,IAAI,eAAe,yBAAyB;AAAA,YAC1C;AAAA,YACA,YAAY,KAAK;AAAA,YACjB,OAAO,UAAU;AAAA,YACjB,QAAQ,UAAU;AAAA,UACpB,CAAC;AAAA,QACH;AAAA,MACF;AAIA,UAAI,OAAO,KAAK,gBAAgB,UAAa,KAAK,kBAAkB,SAAY;AAC9E,cAAM,mBAAmB,IAAI,UAAU,UAAU,SAAS,IAAI;AAAA,MAChE;AAEA,UAAI,KAAK,YAAY,KAAK,QAAQ;AAChC,cAAM,MAAoB;AAAA,UACxB,GAAG,KAAK;AAAA,UACR,OAAO,UAAU;AAAA,UACjB,UAAU,UAAU;AAAA,UACpB,MAAM,UAAU;AAAA,UAChB,aAAa,UAAU,OAAO;AAAA,UAC9B,cAAc,UAAU,OAAO;AAAA,UAC/B,iBAAiB,UAAU,OAAO,aAAa;AAAA,UAC/C,kBAAkB,UAAU,OAAO,cAAc;AAAA,UACjD,WAAW,UAAU;AAAA,UACrB;AAAA,UACA,QAAQ,SAAS,IAAI,CAAC;AAAA,QACxB;AACA,aAAK,SAAS,GAAG,EAAE,MAAM,CAAC,MAAe;AACvC,kBAAQ,OAAO,sBAAsB,EAAE,SAAS,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC,EAAE,CAAC;AAAA,QAC9F,CAAC;AAAA,MACH;AACA,aAAO,EAAE,MAAM,WAAW,OAAO,KAAK;AAAA,IACxC,SAAS,GAAG;AACV,UAAI,aAAa,gBAAgB,EAAE,SAAS,cAAc;AACxD,eAAO;AAAA,UACL,IAAI,cAAc,oBAAoB,EAAE,UAAU,IAAI,UAAU,OAAO,IAAI,MAAM,CAAC;AAAA,QACpF;AAAA,MACF;AACA,UAAI,gBAAgB,CAAC,GAAG;AACtB,mBAAW,KAAK,EAAE,UAAU,EAAE,UAAU,QAAQ,EAAE,QAAQ,SAAS,EAAE,QAAQ,CAAC;AAC9E,YAAI,EAAE,WAAW,OAAO,aAAa,UAAU,SAAS,GAAG;AACzD,iBAAO;AAAA,YACL,IAAI,eAAe,uBAAuB,EAAE,QAAQ,IAAI,EAAE,UAAU,WAAW,CAAC;AAAA,UAClF;AAAA,QACF;AACA,gBAAQ,OAAO,kBAAkB,EAAE,UAAU,IAAI,UAAU,QAAQ,EAAE,OAAO,CAAC;AAC7E;AAAA,MACF;AACA,iBAAW,KAAK,EAAE,UAAU,IAAI,UAAU,SAAS,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC,EAAE,CAAC;AAAA,IACjG;AAAA,EACF;AAEA,SAAO;AAAA,IACL,IAAI,cAAc,4BAA4B,EAAE,UAAU,YAAY,MAAM,cAAc,CAAC;AAAA,EAC7F;AACF;AAiDA,gBAAuB,iBACrB,UACA,KACA,OAAwC,CAAC,GACG;AAC5C,MAAI,SAAS,WAAW,GAAG;AACzB,UAAM,IAAI,gBAAgB,4BAA4B;AAAA,EACxD;AACA,MAAI,CAAC,IAAI,qBAAqB;AAC5B,UAAM,IAAI,gBAAgB,0CAA0C;AAAA,EACtE;AAEA,QAAM,OAAgB,KAAK,QAAQ,CAAC;AACpC,QAAM,YAAY,KAAK,SAAS;AAChC,QAAM,MAAM,KAAK,QAAQ,MAAM,KAAK,IAAI;AACxC,QAAM,SAAS,KAAK;AACpB,QAAM,YAAY,IAAI;AAEtB,QAAM,OAAgB,KAAK,QAAQ;AACnC,QAAM,SAAS,KAAK,UAAU,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,QAAQ,GAAG;AACzE,QAAM,gBAAgB,eAAe,UAAU,MAAM;AACrD,QAAM,QAAQ,KAAK,MAAM,MAAM,aAAa;AAC5C,QAAM,YACJ,MAAM,QAAQ,aAAa,UAAU,CAAC,IAAI,gBAAgB,MAAM,UAAU,aAAa,cACnF,MAAM,WACN,MAAM;AAIZ,MAAI,UAAU,aAAa,aAAa;AACtC,UAAM,SAAS,MAAM,SAAS,UAAU,KAAK,MAAM,IAAI;AACvD,QAAI,OAAO,UAAU,QAAQ,OAAO,SAAS,MAAM;AACjD,YAAM,IAAI,cAAc,4BAA4B,EAAE,OAAO,OAAO,MAAM,CAAC;AAAA,IAC7E;AACA,UAAM,OAAO,KAAK;AAClB,WAAO,OAAO;AAAA,EAChB;AAGA,MAAI,sBAAsB,UAAU,UAAU,GAAG,GAAG;AAClD,YAAQ,OAAO,4BAA4B,EAAE,UAAU,UAAU,SAAS,CAAC;AAE3E,UAAM,SAAS,MAAM,SAAS,UAAU,KAAK,MAAM,IAAI;AACvD,QAAI,OAAO,UAAU,QAAQ,OAAO,SAAS,MAAM;AACjD,YAAM,IAAI,cAAc,4BAA4B,EAAE,OAAO,OAAO,MAAM,CAAC;AAAA,IAC7E;AACA,UAAM,OAAO,KAAK;AAClB,WAAO,OAAO;AAAA,EAChB;AAEA,QAAM,MAAM,sBAAsB,UAAU,OAAO,UAAU,MAAM,KAAK,IAAI;AAE5E,QAAM,oBAAoB,iBAAiB,IAAI;AAC/C,MAAI,kBAAmB,KAAI,QAAQ,iBAAiB,IAAI;AAExD,MAAI;AACJ,MAAI;AACF,eAAW,MAAM,UAAU,IAAI,KAAK;AAAA,MAClC,QAAQ;AAAA,MACR,SAAS,IAAI;AAAA,MACb,MAAM,IAAI;AAAA;AAAA;AAAA,MAGV,QAAQ,KAAK,UAAU,YAAY,QAAQ,GAAM;AAAA,IACnD,CAAC;AAAA,EACH,SAAS,GAAG;AACV,QAAI,aAAa,gBAAgB,EAAE,SAAS,cAAc;AACxD,YAAM,IAAI,cAAc,oBAAoB;AAAA,QAC1C,UAAU,UAAU;AAAA,QACpB,OAAO,UAAU;AAAA,MACnB,CAAC;AAAA,IACH;AACA,UAAM,IAAI,cAAc,2BAA2B;AAAA,MACjD,SAAS,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC;AAAA,IACpD,CAAC;AAAA,EACH;AAEA,MAAI,CAAC,SAAS,IAAI;AAChB,UAAM,OAAO,MAAM,SAAS,KAAK,EAAE,MAAM,MAAM,EAAE;AACjD,UAAM,YAAY,sBAAsB,SAAS,MAAM;AACvD,QAAI,aAAa,SAAS,WAAW,KAAK;AACxC,8BAAwB,UAAU,UAAU,GAAG;AAAA,IACjD;AAEA,UAAM,SAAS,MAAM,SAAS,UAAU,KAAK,MAAM,IAAI;AACvD,QAAI,OAAO,UAAU,QAAQ,OAAO,SAAS,MAAM;AACjD,YAAM,IAAI,cAAc,4BAA4B;AAAA,QAClD,aAAa,GAAG,UAAU,QAAQ,IAAI,OAAO,SAAS,MAAM,CAAC,KAAK,KAAK,MAAM,GAAG,GAAG,CAAC;AAAA,QACpF,OAAO,OAAO;AAAA,MAChB,CAAC;AAAA,IACH;AACA,UAAM,OAAO,KAAK;AAClB,WAAO,OAAO;AAAA,EAChB;AAEA,MAAI,CAAC,SAAS,MAAM;AAClB,UAAM,IAAI,cAAc,oCAAoC;AAAA,MAC1D,UAAU,UAAU;AAAA,IACtB,CAAC;AAAA,EACH;AAGA,QAAM,UAAU,IAAI,YAAY;AAChC,MAAI,kBAAkB;AACtB,MAAI,cAAc;AAClB,MAAI,eAAe;AACnB,MAAI,YAAY;AAChB,MAAI,aAAa;AACjB,MAAI;AACJ,QAAM,mBAAuC,SAAS,QAAQ,IAAI,mBAAmB,KAAK;AAE1F,QAAM,SAAS,SAAS,KAAK,UAAU;AACvC,MAAI,SAAS;AAEb,MAAI;AACF,WAAO,MAAM;AACX,YAAM,EAAE,MAAM,MAAM,IAAI,MAAM,OAAO,KAAK;AAC1C,UAAI,KAAM;AACV,gBAAU,QAAQ,OAAO,OAAO,EAAE,QAAQ,KAAK,CAAC;AAGhD,YAAM,QAAQ,OAAO,MAAM,IAAI;AAE/B,eAAS,MAAM,IAAI,KAAK;AAExB,iBAAW,QAAQ,OAAO;AACxB,YAAI,CAAC,KAAK,WAAW,QAAQ,EAAG;AAChC,cAAM,OAAO,KAAK,MAAM,CAAC,EAAE,KAAK;AAChC,YAAI,SAAS,SAAU;AACvB,YAAI;AACJ,YAAI;AACF,kBAAQ,KAAK,MAAM,IAAI;AAAA,QACzB,QAAQ;AACN;AAAA,QACF;AAEA,gBAAQ,MAAM,MAAM;AAAA,UAClB,KAAK;AACH,0BAAc,MAAM,SAAS,OAAO,gBAAgB;AACpD,wBAAY,MAAM,SAAS,OAAO,2BAA2B;AAC7D,yBAAa,MAAM,SAAS,OAAO,+BAA+B;AAClE,wBAAY,MAAM,SAAS;AAC3B;AAAA,UACF,KAAK;AACH,gBAAI,MAAM,OAAO,SAAS,gBAAgB,OAAO,MAAM,MAAM,SAAS,UAAU;AAC9E,iCAAmB,MAAM,MAAM;AAC/B,oBAAM,MAAM,MAAM;AAAA,YACpB;AACA;AAAA,UACF,KAAK;AACH,2BAAe,MAAM,OAAO,iBAAiB;AAC7C;AAAA,UACF;AACE;AAAA,QACJ;AAAA,MACF;AAAA,IACF;AAAA,EACF,UAAE;AACA,WAAO,YAAY;AAAA,EACrB;AAEA,wBAAsB,UAAU,QAAQ;AACxC,UAAQ,OAAO,wBAAwB;AAAA,IACrC,UAAU,UAAU;AAAA,IACpB,OAAO,UAAU;AAAA,IACjB;AAAA,IACA;AAAA,IACA,OAAO,KAAK;AAAA,IACZ,SAAS,KAAK;AAAA,IACd,OAAO,KAAK;AAAA,IACZ,UAAU,KAAK;AAAA,EACjB,CAAC;AAED,SAAO;AAAA,IACL,SAAS;AAAA,IACT,UAAU,UAAU;AAAA,IACpB,OAAO,aAAa,UAAU;AAAA,IAC9B;AAAA,IACA,QAAQ,EAAE,OAAO,aAAa,QAAQ,cAAc,WAAW,WAAW;AAAA,IAC1E,SAAS,IAAI,IAAI;AAAA,IACjB,UAAU;AAAA,IACV;AAAA,EACF;AACF;AA4BO,SAAS,gBAAgB,UAAkB,SAA4B;AAC5E,MAAI,QAAQ,WAAW,EAAG,QAAO;AAEjC,QAAM,SAAS;AACf,QAAM,iBAAiB,SAAS,MAAM,KAAK,EAAE,OAAO,CAAC,MAAM,EAAE,SAAS,CAAC;AAEvE,MAAI,eAAe,SAAS,OAAQ,QAAO;AAG3C,QAAM,eAAe,oBAAI,IAAY;AACrC,aAAW,UAAU,SAAS;AAC5B,UAAM,SAAS,OAAO,MAAM,KAAK,EAAE,OAAO,CAAC,MAAM,EAAE,SAAS,CAAC;AAC7D,aAAS,IAAI,GAAG,KAAK,OAAO,SAAS,QAAQ,KAAK;AAChD,YAAM,QAAQ,OAAO,MAAM,GAAG,IAAI,MAAM,EAAE,KAAK,GAAG;AAClD,mBAAa,IAAI,KAAK;AAAA,IACxB;AAAA,EACF;AAEA,MAAI,aAAa,SAAS,EAAG,QAAO;AAGpC,WAAS,IAAI,GAAG,KAAK,eAAe,SAAS,QAAQ,KAAK;AACxD,UAAM,QAAQ,eAAe,MAAM,GAAG,IAAI,MAAM,EAAE,KAAK,GAAG;AAC1D,QAAI,aAAa,IAAI,KAAK,EAAG,QAAO;AAAA,EACtC;AAEA,SAAO;AACT;","names":[]}
+{"version":3,"sources":["../src/index.ts"],"sourcesContent":["import {\n  InternalError,\n  RateLimitError,\n  ValidationError,\n  toErrorResponse,\n  type FactoryResponse,\n} from '@latimer-woods-tech/errors';\nimport type { Logger } from '@latimer-woods-tech/logger';\n\n/**\n * A tool the model may call. `parameters` is a JSON Schema object describing\n * the tool's input. Provider-agnostic; normalized per provider at request time.\n */\nexport interface LLMTool {\n  name: string;\n  description?: string;\n  /** JSON Schema for the tool's input arguments. */\n  parameters: Record<string, unknown>;\n}\n\n/**\n * A tool invocation requested by the model, normalized across providers.\n */\nexport interface LLMToolCall {\n  /** Provider-assigned call id; echo it back in the matching tool_result. */\n  id: string;\n  name: string;\n  /** Parsed argument object the model passed to the tool. */\n  arguments: Record<string, unknown>;\n}\n\n/**\n * Structured content block for tool-calling conversations. The field shapes\n * mirror the Anthropic Messages wire format so they pass through unchanged.\n */\nexport type LLMContentBlock =\n  | { type: 'text'; text: string }\n  | { type: 'tool_use'; id: string; name: string; input: Record<string, unknown> }\n  | { type: 'tool_result'; tool_use_id: string; content: string; is_error?: boolean };\n\n/**\n * Single chat message exchanged with an LLM provider.\n *\n * `content` is a plain string in the common case. For tool-calling\n * conversations it may be an array of {@link LLMContentBlock}s (e.g. an\n * assistant turn carrying `tool_use` blocks, or a user turn carrying\n * `tool_result` blocks). Providers that don't support tool-calling receive\n * the text projection of the content (see `contentToText`).\n */\nexport interface LLMMessage {\n  role: 'user' | 'assistant' | 'system';\n  content: string | LLMContentBlock[];\n}\n\n/**\n * Flattens message content to plain text for providers/paths that only handle\n * strings. `tool_use` blocks contribute nothing; `tool_result` blocks\n * contribute their textual content.\n */\nfunction contentToText(content: string | LLMContentBlock[]): string {\n  if (typeof content === 'string') return content;\n  return content\n    .map((b) => (b.type === 'text' ? b.text : b.type === 'tool_result' ? b.content : ''))\n    .join('');\n}\n\n/**\n * Resolves the system prompt: explicit `opts.system` wins, else the first\n * `system` message, flattened to text. Returns `undefined` when neither is set.\n */\nfunction systemText(opts: LLMOptions, messages: LLMMessage[]): string | undefined {\n  if (opts.system !== undefined) return opts.system;\n  const c = messages.find((m) => m.role === 'system')?.content;\n  return c === undefined ? undefined : contentToText(c);\n}\n\n/**\n * Quality tier selected by the caller. Routing is workload-split:\n *  - `fast`      → Grok 4.3 with Anthropic Haiku fallback (routine drafts/small jobs)\n *  - `balanced`  → Anthropic Sonnet (default)\n *  - `smart`     → Anthropic Opus OR Gemini 2.5 Pro if input is long-context (>150k tokens estimated)\n *  - `verifier`  → Groq Llama (cheap second opinion; only used from verifier code path)\n *  - `workbench` → DeepSeek Chat with Groq fallback (boring, reviewable, non-sensitive batch work)\n */\nexport type LLMTier = 'fast' | 'balanced' | 'smart' | 'verifier' | 'workbench';\n\n/**\n * Options that influence LLM completion behaviour.\n */\nexport interface LLMOptions {\n  /** Quality tier; see {@link LLMTier}. Defaults to `balanced`. */\n  tier?: LLMTier;\n  /** Explicit model override. Takes precedence over tier. */\n  model?: string;\n  maxTokens?: number;\n  temperature?: number;\n  system?: string;\n  /** Token budget above which we force long-context routing (Gemini). */\n  longContextThreshold?: number;\n  /** Per-call cancellation signal. Aborts the in-flight provider request. */\n  signal?: AbortSignal;\n  /** Optional run identifier stamped on ledger rows + logs. */\n  runId?: string;\n  /** Optional project identifier stamped on ledger rows + logs. */\n  project?: string;\n  /** Optional actor identifier (supervisor / worker / human). */\n  actor?: string;\n  /** Optional workload label used in logs and cost-policy call sites. */\n  workload?: string;\n  /** Grok reasoning effort. Defaults to `none` for cost-controlled fast/draft calls. */\n  reasoningEffort?: 'none' | 'low' | 'medium' | 'high';\n  /** Anthropic prompt-cache control. Defaults to `true` for `system` prompts ≥ 1024 tokens. */\n  promptCache?: boolean;\n  /**\n   * Maximum estimated cost in USD for this completion.\n   * This cap is enforced after the provider returns because it uses actual\n   * response token counts to compute the final cost.\n   * If the post-call estimated cost exceeds this cap, `complete` returns a\n   * {@link RateLimitError} with code `LLM_COST_CAP_EXCEEDED` and\n   * `completionStream` throws the same error.\n   * Pricing is based on {@link MODEL_PRICE_PER_1M}; unknown models default to\n   * Opus rates (conservative upper bound).\n   */\n  maxCostUsd?: number;\n  /**\n   * Org-level daily cost cap in USD. Requires `env.LLM_COST_KV` to be set.\n   * When today's cumulative spend read from KV is >= this value, `complete`\n   * returns a {@link RateLimitError} with code `LLM_DAILY_CAP_EXCEEDED`\n   * without making any provider call. After a successful call the daily\n   * accumulator is updated in KV (TTL: 48 h).\n   */\n  dailyCapUsd?: number;\n  /**\n   * Org-level monthly cost cap in USD. Requires `env.LLM_COST_KV` to be set.\n   * Same enforcement pattern as {@link dailyCapUsd} but keyed by YYYY-MM.\n   * KV TTL: 40 days.\n   */\n  monthlyCapUsd?: number;\n  /**\n   * Metering context. When supplied and `deps.onRecord` is set, a {@link LLMRecordRow}\n   * is emitted after every successful completion. Errors are swallowed.\n   */\n  ledger?: LLMRecordContext;\n  /**\n   * Tools the model may call. When present, routing **fails closed** to\n   * tool-capable providers — failover never falls back to a provider that\n   * can't honour the tool schema. See {@link LLMResult.toolCalls}.\n   */\n  tools?: LLMTool[];\n  /**\n   * Tool-selection policy. `'auto'` (default when `tools` is set) lets the\n   * model decide; `'none'` forbids tool use; `{ name }` forces a specific tool.\n   */\n  toolChoice?: 'auto' | 'none' | { name: string };\n}\n\n/**\n * Provider that produced an LLM response.\n */\nexport type LLMProvider = 'anthropic' | 'gemini' | 'groq' | 'grok' | 'deepseek';\n\n/**\n * Result returned by a successful completion.\n */\nexport interface LLMResult {\n  content: string;\n  provider: LLMProvider;\n  model: string;\n  tier: LLMTier;\n  tokens: { input: number; output: number; cacheRead?: number; cacheWrite?: number };\n  latency: number;\n  /** Number of attempts before success (1 = primary succeeded). */\n  attempts: number;\n  /** Monotonic request id from AI Gateway, if present in headers. */\n  gatewayRequestId?: string;\n  /**\n   * Why generation stopped, normalized across providers. `'tool_use'` means\n   * the model is requesting one or more tool calls (see {@link toolCalls}).\n   */\n  stopReason?: 'end' | 'tool_use' | 'max_tokens' | 'other';\n  /**\n   * Tool calls the model requested, normalized across providers. Present\n   * (non-empty) when `stopReason === 'tool_use'`.\n   */\n  toolCalls?: LLMToolCall[];\n}\n\n/**\n * Environment bindings required by {@link complete}.\n *\n * `AI_GATEWAY_BASE_URL` is REQUIRED in 0.3.0. All provider calls flow through the\n * Cloudflare AI Gateway for unified logging, rate limiting, and cost telemetry.\n * In test/dev the caller may pass a custom fetch impl that short-circuits this.\n */\nexport interface LLMEnv {\n  AI_GATEWAY_BASE_URL: string;\n  ANTHROPIC_API_KEY: string;\n  GROQ_API_KEY: string;\n  /** Optional — only required for `{ tier: 'workbench' }` or `deepseek-*` model overrides. */\n  DEEPSEEK_API_KEY?: string;\n  /** Optional — only required when caller passes `{ model: 'grok-*' }` override. */\n  GROK_API_KEY?: string;\n  /**\n   * Google Cloud short-lived access token with `aiplatform.endpoints.predict`.\n   * Callers mint this via the JWT-bearer flow (service account → token exchange);\n   * see `docs/runbooks/rotate-gcp-sa.md`. Token must be valid for ≥ 5 minutes.\n   */\n  VERTEX_ACCESS_TOKEN: string;\n  VERTEX_PROJECT: string;\n  VERTEX_LOCATION: string;\n  /**\n   * Optional KV store for org-level daily/monthly cost tracking and enforcement.\n   * When provided alongside {@link LLMOptions.dailyCapUsd} or {@link LLMOptions.monthlyCapUsd},\n   * `complete` will block calls that would exceed the declared cap.\n   * Any KV-like store satisfying `get`/`put` works (e.g. Cloudflare KV, in-memory stub).\n   */\n  LLM_COST_KV?: CostKvStore;\n}\n\n/**\n * Minimal KV store interface for org-level LLM cost tracking.\n * Cloudflare KV satisfies this. An in-memory stub is sufficient for tests.\n */\nexport interface CostKvStore {\n  get(key: string): Promise<string | null>;\n  put(key: string, value: string, options?: { expirationTtl?: number }): Promise<void>;\n}\n\n/**\n * Caller-supplied context stamped on every metering row.\n * Mirrors the `LLMRecordContext` in `@latimer-woods-tech/llm-meter`; kept inline\n * to avoid a circular dependency (llm-meter imports llm).\n */\nexport interface LLMRecordContext {\n  project: string;\n  actor: string;\n  runId?: string;\n  workload?: string;\n  tenantId?: string;\n}\n\n/**\n * Row shape passed to the optional {@link LLMDeps.onRecord} callback.\n * Callers can wire this directly to `recordCall` from `@latimer-woods-tech/llm-meter`.\n */\nexport interface LLMRecordRow extends LLMRecordContext {\n  model: string;\n  provider: LLMProvider;\n  tier: LLMTier;\n  inputTokens: number;\n  outputTokens: number;\n  cacheReadTokens: number;\n  cacheWriteTokens: number;\n  latencyMs: number;\n  costUsd: number;\n  yyyyMm: string;\n}\n\n/**\n * Optional dependencies for {@link complete}.\n */\nexport interface LLMDeps {\n  fetch?: typeof fetch;\n  logger?: Logger;\n  now?: () => number;\n  /**\n   * Optional metering callback. Called after every successful completion.\n   * Errors are swallowed so metering never blocks the caller.\n   * Wire to `recordCall` from `@latimer-woods-tech/llm-meter`.\n   */\n  onRecord?: (row: LLMRecordRow) => Promise<void>;\n}\n\n// Model catalogue — keep in sync with docs/architecture/FACTORY_V1.md § LLM substrate.\nconst MODELS = {\n  anthropic: {\n    fast: 'claude-haiku-4-20250514',\n    balanced: 'claude-sonnet-4-6',\n    smart: 'claude-opus-4-7',\n  },\n  gemini: {\n    smart: 'gemini-2.5-pro',\n  },\n  groq: {\n    verifier: 'llama-4-maverick',\n  },\n  grok: {\n    fast: 'grok-4.3',\n  },\n  deepseek: {\n    workbench: 'deepseek-chat',\n  },\n} as const;\n\nconst DEFAULT_MAX_TOKENS = 1024;\nconst DEFAULT_TEMPERATURE = 0.7;\nconst DEFAULT_LONG_CONTEXT_THRESHOLD = 150_000; // tokens\n\n// ─── Per-provider exponential backoff constants ────────────────────────────\n/** Base delay in ms for the first retry. */\nconst BACKOFF_BASE_MS = 500;\n/** Maximum backoff cap in ms. */\nconst BACKOFF_CAP_MS = 8_000;\n/** Max random jitter added to each backoff delay, in ms. */\nconst BACKOFF_JITTER_MAX_MS = 250;\n/** Maximum number of attempts per provider (1 initial + 2 retries). */\nconst PER_PROVIDER_MAX_ATTEMPTS = 3;\n\n// ─── Per-provider cooldown state (module-level) ────────────────────────────\n/**\n * Tracks when a provider's cooldown period expires.\n * Keyed by {@link LLMProvider}; value is the `Date.now()` epoch ms at which\n * the cooldown expires. Absent key means \"not cooling down\".\n */\nconst providerCooldownUntil: Map<LLMProvider, number> = new Map();\n\n/** Cooldown duration in ms after a provider exhausts all retries. */\nconst PROVIDER_COOLDOWN_MS = 30_000;\n\n/**\n * Returns `true` if the provider is currently in its cooldown window.\n * Uses the injected `now` function (or `Date.now`) for testability.\n */\nfunction isProviderCoolingDown(provider: LLMProvider, now: () => number = Date.now): boolean {\n  const until = providerCooldownUntil.get(provider);\n  if (until === undefined) return false;\n  return now() < until;\n}\n\n/** Returns `YYYY-MM-DD` from a Unix timestamp (ms). Used for daily KV cost keys. */\nfunction isoDate(nowMs: number): string {\n  return new Date(nowMs).toISOString().slice(0, 10);\n}\n\n/**\n * Record actual call spend in org-level daily/monthly KV buckets.\n * This is intentionally best-effort: Cloudflare KV does not provide an atomic\n * compare-and-swap, so concurrent requests can race and undercount spend.\n */\nasync function recordOrgCostUsage(\n  kv: CostKvStore,\n  todayKey: string,\n  monthKey: string,\n  costUsd: number,\n  opts: LLMOptions,\n): Promise<void> {\n  if (opts.dailyCapUsd !== undefined) {\n    const raw = await kv.get(todayKey).catch(() => null);\n    const spent = parseFloat(raw ?? '0');\n    await kv.put(todayKey, String(spent + costUsd), { expirationTtl: 172_800 /* 48 h */ }).catch(() => undefined);\n  }\n  if (opts.monthlyCapUsd !== undefined) {\n    const raw = await kv.get(monthKey).catch(() => null);\n    const spent = parseFloat(raw ?? '0');\n    await kv.put(monthKey, String(spent + costUsd), { expirationTtl: 3_456_000 /* 40 d */ }).catch(() => undefined);\n  }\n}\n\n/**\n * USD cost per 1 million tokens for each model.\n * Source: Anthropic / Google / xAI pricing pages as of 2026-05.\n * Keep these model names in sync with the default routing constants in\n * {@link MODELS}; unknown models fall back to Opus rates (conservative upper bound).\n *\n * CANONICAL pricing source for the platform. `@latimer-woods-tech/llm-meter`\n * derives its cents-denominated rates from this table and a drift-guard test\n * there fails CI if they diverge — make all rate changes here.\n */\nexport const MODEL_PRICE_PER_1M: Record<string, { input: number; output: number; cacheRead: number; cacheWrite: number }> = {\n  // Anthropic Haiku 4\n  'claude-haiku-4-20250514': { input: 0.80, output: 4.00, cacheRead: 0.08, cacheWrite: 1.00 },\n  'claude-haiku-4-5-20251001': { input: 0.80, output: 4.00, cacheRead: 0.08, cacheWrite: 1.00 },\n  // Anthropic Sonnet 4\n  'claude-sonnet-4-20250514': { input: 3.00, output: 15.00, cacheRead: 0.30, cacheWrite: 3.75 },\n  'claude-sonnet-4-6': { input: 3.00, output: 15.00, cacheRead: 0.30, cacheWrite: 3.75 },\n  // Anthropic Opus 4\n  'claude-opus-4-20250514': { input: 15.00, output: 75.00, cacheRead: 1.50, cacheWrite: 18.75 },\n  'claude-opus-4-7': { input: 15.00, output: 75.00, cacheRead: 1.50, cacheWrite: 18.75 },\n  // Gemini 2.5 Pro\n  'gemini-2.5-pro': { input: 1.25, output: 10.00, cacheRead: 0.31, cacheWrite: 4.50 },\n  // Groq Llama 4 Maverick\n  'llama-4-maverick': { input: 0.50, output: 0.77, cacheRead: 0.05, cacheWrite: 0.50 },\n  // Grok 4.3\n  'grok-4.3': { input: 1.25, output: 2.50, cacheRead: 0.00, cacheWrite: 0.00 },\n  // DeepSeek API pricing as of 2026-05: cache-write conservatively uses cache-miss input pricing.\n  'deepseek-chat': { input: 0.27, output: 1.10, cacheRead: 0.07, cacheWrite: 0.27 },\n  'deepseek-reasoner': { input: 0.55, output: 2.19, cacheRead: 0.14, cacheWrite: 0.55 },\n  // Deprecated aliases retained for historical ledger rows.\n  'grok-4-fast': { input: 1.25, output: 2.50, cacheRead: 0.00, cacheWrite: 0.00 },\n  'grok-3-mini-latest': { input: 1.25, output: 2.50, cacheRead: 0.00, cacheWrite: 0.00 },\n};\n\n/** Fallback pricing used for unrecognised models (Opus rates — conservative upper bound). */\nconst PRICE_FALLBACK = MODEL_PRICE_PER_1M['claude-opus-4-7']!;\n\n/**\n * Estimates the USD cost of a single LLM completion from token counts.\n * Returns 0 for zero-token results. Uses {@link MODEL_PRICE_PER_1M} with\n * {@link PRICE_FALLBACK} for unknown models.\n */\nfunction estimateCostUsd(\n  tokens: { input: number; output: number; cacheRead?: number; cacheWrite?: number },\n  model: string,\n): number {\n  const price = MODEL_PRICE_PER_1M[model] ?? PRICE_FALLBACK;\n  return (\n    (tokens.input * price.input +\n      tokens.output * price.output +\n      (tokens.cacheRead ?? 0) * price.cacheRead +\n      (tokens.cacheWrite ?? 0) * price.cacheWrite) /\n    1_000_000\n  );\n}\n\n/** Returns `YYYY-MM` from a Unix timestamp (ms). Used for monthly KV cost keys. */\nfunction isoMonth(nowMs: number): string {\n  return new Date(nowMs).toISOString().slice(0, 7);\n}\n\n/**\n * Marks a provider as cooling down for {@link PROVIDER_COOLDOWN_MS} milliseconds.\n */\nfunction markProviderCoolingDown(provider: LLMProvider, now: () => number = Date.now): void {\n  providerCooldownUntil.set(provider, now() + PROVIDER_COOLDOWN_MS);\n}\n\n/**\n * Clears the cooldown state for a provider after a successful call.\n */\nfunction clearProviderCooldown(provider: LLMProvider): void {\n  providerCooldownUntil.delete(provider);\n}\n\n// ─── Legacy backoff constant (kept for the existing callWithBackoff signature) ─\nconst BASE_BACKOFF_MS = 250;\n\ninterface ProviderError {\n  provider: LLMProvider;\n  status: number;\n  retryable: boolean;\n  message: string;\n}\n\n/**\n * Returns `true` for status codes that should trigger a retry.\n * Only 429 and 5xx (transient server errors) qualify; other 4xx are terminal.\n */\nfunction isRetryableForBackoff(status: number): boolean {\n  return status === 429 || (status >= 500 && status < 600);\n}\n\nfunction estimateTokens(messages: LLMMessage[], system?: string): number {\n  // Cheap estimator: ~4 chars/token. Good enough for threshold routing.\n  let chars = system?.length ?? 0;\n  for (const m of messages) chars += contentToText(m.content).length;\n  return Math.ceil(chars / 4);\n}\n\nfunction sleep(ms: number, signal?: AbortSignal): Promise<void> {\n  return new Promise((resolve, reject) => {\n    const t = setTimeout(resolve, ms);\n    if (signal) {\n      const onAbort = () => {\n        clearTimeout(t);\n        reject(new DOMException('Aborted', 'AbortError'));\n      };\n      if (signal.aborted) onAbort();\n      else signal.addEventListener('abort', onAbort, { once: true });\n    }\n  });\n}\n\n/**\n * Computes the exponential backoff delay for a given attempt with jitter.\n *\n * Formula: `Math.min(base * 2^attempt + jitter, cap)`\n * where `jitter` is a random value in `[0, BACKOFF_JITTER_MAX_MS)`.\n *\n * @param attempt - Zero-based attempt index (0 = first retry after initial failure).\n */\nfunction computeBackoffMs(attempt: number): number {\n  const jitter = Math.floor(Math.random() * BACKOFF_JITTER_MAX_MS);\n  return Math.min(BACKOFF_BASE_MS * Math.pow(2, attempt) + jitter, BACKOFF_CAP_MS);\n}\n\n// ─── Provider request builders ─────────────────────────────────────────────\n\nfunction buildAnthropicRequest(\n  model: string,\n  messages: LLMMessage[],\n  opts: LLMOptions,\n  env: LLMEnv,\n  streaming = false,\n): { url: string; headers: Record<string, string>; body: string } {\n  const sys = systemText(opts, messages);\n  const filtered = messages.filter((m) => m.role !== 'system');\n  const body: Record<string, unknown> = {\n    model,\n    max_tokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,\n    temperature: opts.temperature ?? DEFAULT_TEMPERATURE,\n    messages: filtered.map((m) => ({ role: m.role, content: m.content })),\n  };\n  if (streaming) {\n    body.stream = true;\n  }\n  if (sys) {\n    const cache = opts.promptCache ?? sys.length >= 4096;\n    body.system = cache\n      ? [{ type: 'text', text: sys, cache_control: { type: 'ephemeral' } }]\n      : sys;\n  }\n  if (opts.tools && opts.tools.length > 0) {\n    body.tools = opts.tools.map((t) => ({\n      name: t.name,\n      description: t.description ?? '',\n      input_schema: t.parameters,\n    }));\n    const tc = opts.toolChoice ?? 'auto';\n    body.tool_choice =\n      tc === 'auto'\n        ? { type: 'auto' }\n        : tc === 'none'\n          ? { type: 'none' }\n          : { type: 'tool', name: tc.name };\n  }\n  return {\n    url: `${env.AI_GATEWAY_BASE_URL}/anthropic/v1/messages`,\n    headers: {\n      'content-type': 'application/json',\n      'x-api-key': env.ANTHROPIC_API_KEY,\n      'anthropic-version': '2023-06-01',\n      'anthropic-beta': 'prompt-caching-2024-07-31',\n    },\n    body: JSON.stringify(body),\n  };\n}\n\nfunction buildGeminiRequest(\n  model: string,\n  messages: LLMMessage[],\n  opts: LLMOptions,\n  env: LLMEnv,\n): { url: string; headers: Record<string, string>; body: string } {\n  const sys = systemText(opts, messages);\n  const contents = messages\n    .filter((m) => m.role !== 'system')\n    .map((m) => ({\n      role: m.role === 'assistant' ? 'model' : 'user',\n      parts: [{ text: contentToText(m.content) }],\n    }));\n  const body: Record<string, unknown> = {\n    contents,\n    generationConfig: {\n      maxOutputTokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,\n      temperature: opts.temperature ?? DEFAULT_TEMPERATURE,\n    },\n  };\n  if (sys) {\n    body.systemInstruction = { parts: [{ text: sys }] };\n  }\n  const path = `v1/projects/${env.VERTEX_PROJECT}/locations/${env.VERTEX_LOCATION}/publishers/google/models/${model}:generateContent`;\n  return {\n    url: `${env.AI_GATEWAY_BASE_URL}/google-vertex-ai/${path}`,\n    headers: {\n      'content-type': 'application/json',\n      authorization: `Bearer ${env.VERTEX_ACCESS_TOKEN}`,\n    },\n    body: JSON.stringify(body),\n  };\n}\n\n// ─── OpenAI-style (Grok / DeepSeek / Groq) tool-calling helpers ──────────────\n\n/**\n * Converts provider-agnostic messages to OpenAI chat-completions format,\n * translating the Anthropic-shaped tool blocks: `tool_use` → an assistant\n * message with `tool_calls`; `tool_result` → a standalone `tool` message keyed\n * by `tool_call_id`. Plain-string content passes through unchanged.\n */\nfunction toOpenAiMessages(messages: LLMMessage[], sys: string | undefined): Array<Record<string, unknown>> {\n  const out: Array<Record<string, unknown>> = [];\n  if (sys) out.push({ role: 'system', content: sys });\n  for (const m of messages) {\n    if (m.role === 'system') continue;\n    if (typeof m.content === 'string') {\n      out.push({ role: m.role, content: m.content });\n      continue;\n    }\n    let text = '';\n    const toolCalls: Array<Record<string, unknown>> = [];\n    const results: Array<{ tool_use_id: string; content: string }> = [];\n    for (const b of m.content) {\n      if (b.type === 'text') text += b.text;\n      else if (b.type === 'tool_use')\n        toolCalls.push({ id: b.id, type: 'function', function: { name: b.name, arguments: JSON.stringify(b.input) } });\n      else if (b.type === 'tool_result') results.push({ tool_use_id: b.tool_use_id, content: b.content });\n    }\n    if (results.length > 0) {\n      for (const r of results) out.push({ role: 'tool', tool_call_id: r.tool_use_id, content: r.content });\n      if (text) out.push({ role: 'user', content: text });\n    } else if (toolCalls.length > 0) {\n      out.push({ role: 'assistant', content: text || null, tool_calls: toolCalls });\n    } else {\n      out.push({ role: m.role, content: text });\n    }\n  }\n  return out;\n}\n\n/** Builds the OpenAI `tools` array from {@link LLMOptions.tools}, or undefined. */\nfunction openAiTools(opts: LLMOptions): Array<Record<string, unknown>> | undefined {\n  if (!opts.tools || opts.tools.length === 0) return undefined;\n  return opts.tools.map((t) => ({\n    type: 'function',\n    function: { name: t.name, description: t.description ?? '', parameters: t.parameters },\n  }));\n}\n\n/** Maps {@link LLMOptions.toolChoice} to the OpenAI `tool_choice` value. */\nfunction openAiToolChoice(tc: LLMOptions['toolChoice']): unknown {\n  if (tc === undefined) return undefined;\n  if (tc === 'auto' || tc === 'none') return tc;\n  return { type: 'function', function: { name: tc.name } };\n}\n\nfunction buildGroqRequest(\n  model: string,\n  messages: LLMMessage[],\n  opts: LLMOptions,\n  env: LLMEnv,\n): { url: string; headers: Record<string, string>; body: string } {\n  const sys = systemText(opts, messages);\n  const merged: LLMMessage[] = [];\n  if (sys) merged.push({ role: 'system', content: sys });\n  for (const m of messages) if (m.role !== 'system') merged.push({ role: m.role, content: contentToText(m.content) });\n  return {\n    url: `${env.AI_GATEWAY_BASE_URL}/groq/openai/v1/chat/completions`,\n    headers: {\n      'content-type': 'application/json',\n      authorization: `Bearer ${env.GROQ_API_KEY}`,\n    },\n    body: JSON.stringify({\n      model,\n      max_tokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,\n      temperature: opts.temperature ?? DEFAULT_TEMPERATURE,\n      messages: merged,\n    }),\n  };\n}\n\nfunction buildGrokRequest(\n  model: string,\n  messages: LLMMessage[],\n  opts: LLMOptions,\n  env: LLMEnv,\n): { url: string; headers: Record<string, string>; body: string } {\n  if (!env.GROK_API_KEY) {\n    throw new ValidationError('GROK_API_KEY required for grok-* model override');\n  }\n  const sys = systemText(opts, messages);\n  const body: Record<string, unknown> = {\n    model,\n    max_tokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,\n    temperature: opts.temperature ?? DEFAULT_TEMPERATURE,\n    messages: toOpenAiMessages(messages, sys),\n  };\n  const tools = openAiTools(opts);\n  if (tools) {\n    body.tools = tools;\n    const tc = openAiToolChoice(opts.toolChoice ?? 'auto');\n    if (tc !== undefined) body.tool_choice = tc;\n  }\n  if (model === MODELS.grok.fast) {\n    body.reasoning_effort = opts.reasoningEffort ?? 'none';\n  }\n  return {\n    url: `${env.AI_GATEWAY_BASE_URL}/grok/v1/chat/completions`,\n    headers: {\n      'content-type': 'application/json',\n      authorization: `Bearer ${env.GROK_API_KEY}`,\n    },\n    body: JSON.stringify(body),\n  };\n}\n\nfunction buildDeepSeekRequest(\n  model: string,\n  messages: LLMMessage[],\n  opts: LLMOptions,\n  env: LLMEnv,\n): { url: string; headers: Record<string, string>; body: string } {\n  if (!env.DEEPSEEK_API_KEY) {\n    throw new ValidationError('DEEPSEEK_API_KEY required for workbench tier or deepseek-* model override');\n  }\n  const sys = systemText(opts, messages);\n  const body: Record<string, unknown> = {\n    model,\n    max_tokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,\n    temperature: opts.temperature ?? DEFAULT_TEMPERATURE,\n    messages: toOpenAiMessages(messages, sys),\n  };\n  const tools = openAiTools(opts);\n  if (tools) {\n    body.tools = tools;\n    const tc = openAiToolChoice(opts.toolChoice ?? 'auto');\n    if (tc !== undefined) body.tool_choice = tc;\n  }\n  return {\n    url: `${env.AI_GATEWAY_BASE_URL}/deepseek/chat/completions`,\n    headers: {\n      'content-type': 'application/json',\n      authorization: `Bearer ${env.DEEPSEEK_API_KEY}`,\n    },\n    body: JSON.stringify(body),\n  };\n}\n\n// ─── Response parsers ──────────────────────────────────────────────────────\n\ninterface AnthropicResponse {\n  content?: Array<{\n    type: string;\n    text?: string;\n    id?: string;\n    name?: string;\n    input?: Record<string, unknown>;\n  }>;\n  stop_reason?: string;\n  usage?: {\n    input_tokens?: number;\n    output_tokens?: number;\n    cache_read_input_tokens?: number;\n    cache_creation_input_tokens?: number;\n  };\n  model?: string;\n}\n\n/** Maps a provider stop reason to the normalized {@link LLMResult.stopReason}. */\nfunction normalizeAnthropicStop(reason: string | undefined): LLMResult['stopReason'] {\n  switch (reason) {\n    case 'end_turn':\n    case 'stop_sequence':\n      return 'end';\n    case 'tool_use':\n      return 'tool_use';\n    case 'max_tokens':\n      return 'max_tokens';\n    default:\n      return reason ? 'other' : undefined;\n  }\n}\n\ninterface GeminiResponse {\n  candidates?: Array<{ content?: { parts?: Array<{ text?: string }> } }>;\n  usageMetadata?: {\n    promptTokenCount?: number;\n    candidatesTokenCount?: number;\n  };\n}\n\ninterface GroqResponse {\n  choices?: Array<{ message?: { content?: string } }>;\n  usage?: { prompt_tokens?: number; completion_tokens?: number };\n  model?: string;\n}\n\nfunction parseAnthropic(\n  json: unknown,\n): {\n  content: string;\n  input: number;\n  output: number;\n  cacheRead: number;\n  cacheWrite: number;\n  model?: string;\n  toolCalls?: LLMToolCall[];\n  stopReason?: LLMResult['stopReason'];\n} {\n  const r = json as AnthropicResponse;\n  const toolCalls: LLMToolCall[] = (r.content ?? [])\n    .filter((c) => c.type === 'tool_use' && typeof c.id === 'string' && typeof c.name === 'string')\n    .map((c) => ({ id: c.id!, name: c.name!, arguments: c.input ?? {} }));\n  return {\n    content: r.content?.find((c) => c.type === 'text')?.text ?? '',\n    input: r.usage?.input_tokens ?? 0,\n    output: r.usage?.output_tokens ?? 0,\n    cacheRead: r.usage?.cache_read_input_tokens ?? 0,\n    cacheWrite: r.usage?.cache_creation_input_tokens ?? 0,\n    model: r.model,\n    toolCalls: toolCalls.length > 0 ? toolCalls : undefined,\n    stopReason: normalizeAnthropicStop(r.stop_reason),\n  };\n}\n\nfunction parseGemini(json: unknown): { content: string; input: number; output: number } {\n  const r = json as GeminiResponse;\n  const text =\n    r.candidates?.[0]?.content?.parts?.map((p) => p.text ?? '').join('') ?? '';\n  return {\n    content: text,\n    input: r.usageMetadata?.promptTokenCount ?? 0,\n    output: r.usageMetadata?.candidatesTokenCount ?? 0,\n  };\n}\n\nfunction parseGroq(json: unknown): { content: string; input: number; output: number; model?: string } {\n  const r = json as GroqResponse;\n  return {\n    content: r.choices?.[0]?.message?.content ?? '',\n    input: r.usage?.prompt_tokens ?? 0,\n    output: r.usage?.completion_tokens ?? 0,\n    model: r.model,\n  };\n}\n\ninterface OpenAiResponse {\n  choices?: Array<{\n    message?: {\n      content?: string | null;\n      tool_calls?: Array<{ id?: string; function?: { name?: string; arguments?: string } }>;\n    };\n    finish_reason?: string;\n  }>;\n  usage?: { prompt_tokens?: number; completion_tokens?: number };\n  model?: string;\n}\n\n/** Maps an OpenAI `finish_reason` to the normalized {@link LLMResult.stopReason}. */\nfunction normalizeOpenAiStop(reason: string | undefined): LLMResult['stopReason'] {\n  switch (reason) {\n    case 'stop':\n      return 'end';\n    case 'tool_calls':\n    case 'function_call':\n      return 'tool_use';\n    case 'length':\n      return 'max_tokens';\n    default:\n      return reason ? 'other' : undefined;\n  }\n}\n\n/** Best-effort parse of an OpenAI tool-call arguments string; `{}` on failure. */\nfunction parseToolArgs(raw: string | undefined): Record<string, unknown> {\n  if (!raw) return {};\n  try {\n    const v = JSON.parse(raw) as unknown;\n    return typeof v === 'object' && v !== null ? (v as Record<string, unknown>) : {};\n  } catch {\n    return {};\n  }\n}\n\n/**\n * Parses an OpenAI chat-completions response (Grok, DeepSeek), extracting\n * normalized tool calls and a stop reason in addition to text + tokens.\n */\nfunction parseOpenAi(json: unknown): {\n  content: string;\n  input: number;\n  output: number;\n  model?: string;\n  toolCalls?: LLMToolCall[];\n  stopReason?: LLMResult['stopReason'];\n} {\n  const r = json as OpenAiResponse;\n  const choice = r.choices?.[0];\n  const toolCalls: LLMToolCall[] = (choice?.message?.tool_calls ?? [])\n    .filter((c) => typeof c.function?.name === 'string')\n    .map((c, i) => ({\n      id: c.id ?? `call_${i}`,\n      name: c.function!.name!,\n      arguments: parseToolArgs(c.function?.arguments),\n    }));\n  return {\n    content: choice?.message?.content ?? '',\n    input: r.usage?.prompt_tokens ?? 0,\n    output: r.usage?.completion_tokens ?? 0,\n    model: r.model,\n    toolCalls: toolCalls.length > 0 ? toolCalls : undefined,\n    stopReason: normalizeOpenAiStop(choice?.finish_reason),\n  };\n}\n\n// ─── Core call with backoff ────────────────────────────────────────────────\n\n/**\n * Calls a provider with per-provider exponential backoff.\n *\n * Retries up to {@link PER_PROVIDER_MAX_ATTEMPTS} times on 429 or transient 5xx.\n * Other 4xx codes are treated as terminal and not retried.\n * AbortError is never retried — it bubbles immediately.\n *\n * @param provider - Provider name, used for error tagging.\n * @param request - Pre-built HTTP request descriptor.\n * @param fetchImpl - Fetch implementation (injectable for tests).\n * @param signal - Optional AbortSignal for cancellation.\n * @param logger - Optional logger for per-attempt warnings.\n * @param nowFn - Optional clock injection for testability.\n * @returns Parsed JSON body, optional AI Gateway request ID, and attempt count.\n */\nasync function callWithBackoff(\n  provider: LLMProvider,\n  request: { url: string; headers: Record<string, string>; body: string },\n  fetchImpl: typeof fetch,\n  signal: AbortSignal | undefined,\n  logger: Logger | undefined,\n  nowFn?: () => number,\n): Promise<{ json: unknown; gatewayRequestId?: string; attempts: number }> {\n  /**\n   * Helper: mark provider cooling down and then throw the error.\n   * Called whenever we determine we've exhausted all retries for the provider.\n   * AbortError is never counted as a provider exhaustion — it bypasses this.\n   */\n  function exhaustAndThrow(err: ProviderError): never {\n    markProviderCoolingDown(provider, nowFn ?? Date.now);\n    throw err;\n  }\n\n  let lastErr: ProviderError | undefined;\n  for (let attempt = 1; attempt <= PER_PROVIDER_MAX_ATTEMPTS; attempt++) {\n    try {\n      const response = await fetchImpl(request.url, {\n        method: 'POST',\n        headers: request.headers,\n        body: request.body,\n        signal,\n      });\n      if (!response.ok) {\n        const text = await response.text().catch(() => '');\n        const retryable = isRetryableForBackoff(response.status);\n        const err: ProviderError = {\n          provider,\n          status: response.status,\n          retryable,\n          message: `${provider} ${String(response.status)}: ${text.slice(0, 300)}`,\n        };\n        logger?.warn?.('llm.provider.error', { provider, status: response.status, attempt });\n        if (!err.retryable || attempt === PER_PROVIDER_MAX_ATTEMPTS) {\n          if (err.retryable) exhaustAndThrow(err); // retryable but exhausted\n          throw err; // terminal non-retryable error — no cooldown\n        }\n        lastErr = err;\n      } else {\n        const gatewayRequestId = response.headers.get('cf-aig-request-id') ?? undefined;\n        clearProviderCooldown(provider);\n        return { json: await response.json(), gatewayRequestId, attempts: attempt };\n      }\n    } catch (e) {\n      if (e instanceof DOMException && e.name === 'AbortError') throw e;\n      if (typeof e === 'object' && e !== null && 'retryable' in e) {\n        const err = e as ProviderError;\n        if (!err.retryable || attempt === PER_PROVIDER_MAX_ATTEMPTS) {\n          if (err.retryable) exhaustAndThrow(err); // retryable but exhausted\n          throw err; // terminal — no cooldown\n        }\n        lastErr = err;\n      } else {\n        const err: ProviderError = {\n          provider,\n          status: 0,\n          retryable: true,\n          message: e instanceof Error ? e.message : String(e),\n        };\n        if (attempt === PER_PROVIDER_MAX_ATTEMPTS) exhaustAndThrow(err);\n        lastErr = err;\n      }\n    }\n    // Exponential backoff with jitter: base=500ms, cap=8000ms, jitter up to 250ms\n    const backoffMs = computeBackoffMs(attempt - 1);\n    await sleep(backoffMs, signal);\n  }\n  // Fallthrough — should not be reached, but mark cooling down defensively.\n  markProviderCoolingDown(provider, nowFn ?? Date.now);\n  throw lastErr ?? ({ provider, status: 0, retryable: false, message: 'exhausted' } as ProviderError);\n}\n\nfunction isProviderError(err: unknown): err is ProviderError {\n  return (\n    typeof err === 'object' &&\n    err !== null &&\n    typeof (err as { status?: unknown }).status === 'number' &&\n    typeof (err as { message?: unknown }).message === 'string' &&\n    typeof (err as { provider?: unknown }).provider === 'string'\n  );\n}\n\n// ─── Routing ───────────────────────────────────────────────────────────────\n\ninterface RoutePlan {\n  primary: { provider: LLMProvider; model: string };\n  fallback?: { provider: LLMProvider; model: string };\n}\n\n/**\n * Providers whose request builder + response parser support tool-calling.\n * When `opts.tools` is set, routing is restricted to this set and **fails\n * closed** rather than silently calling a provider that would ignore the\n * tools. Expanded in 1b as the other providers' tool formats are normalized.\n */\nconst TOOL_CAPABLE_PROVIDERS = new Set<LLMProvider>(['anthropic', 'grok', 'deepseek']);\n\nfunction plan(tier: LLMTier, opts: LLMOptions, tokenEstimate: number): RoutePlan {\n  if (opts.model) {\n    // Explicit override — best-effort provider detection.\n    const m = opts.model;\n    if (m.startsWith('claude')) return { primary: { provider: 'anthropic', model: m } };\n    if (m.startsWith('gemini')) return { primary: { provider: 'gemini', model: m } };\n    if (m.startsWith('grok')) return { primary: { provider: 'grok', model: m } };\n    if (m.startsWith('deepseek')) return { primary: { provider: 'deepseek', model: m } };\n    return { primary: { provider: 'groq', model: m } };\n  }\n  const longContext = tokenEstimate >= (opts.longContextThreshold ?? DEFAULT_LONG_CONTEXT_THRESHOLD);\n  switch (tier) {\n    case 'workbench':\n      return {\n        primary: { provider: 'deepseek', model: MODELS.deepseek.workbench },\n        fallback: { provider: 'groq', model: MODELS.groq.verifier },\n      };\n    case 'verifier':\n      return { primary: { provider: 'groq', model: MODELS.groq.verifier } };\n    case 'smart':\n      return longContext\n        ? {\n            primary: { provider: 'gemini', model: MODELS.gemini.smart },\n            fallback: { provider: 'anthropic', model: MODELS.anthropic.smart },\n          }\n        : {\n            primary: { provider: 'anthropic', model: MODELS.anthropic.smart },\n            fallback: { provider: 'gemini', model: MODELS.gemini.smart },\n          };\n    case 'fast':\n      return {\n        primary: { provider: 'grok', model: MODELS.grok.fast },\n        fallback: { provider: 'anthropic', model: MODELS.anthropic.fast },\n      };\n    case 'balanced':\n    default:\n      return longContext\n        ? {\n            primary: { provider: 'gemini', model: MODELS.gemini.smart },\n            fallback: { provider: 'anthropic', model: MODELS.anthropic.balanced },\n          }\n        : {\n            primary: { provider: 'anthropic', model: MODELS.anthropic.balanced },\n            fallback: { provider: 'gemini', model: MODELS.gemini.smart },\n          };\n  }\n}\n\n/**\n * Build the `cf-aig-metadata` header value for the Cloudflare AI Gateway.\n *\n * Carries caller attribution (project / workload / actor / runId) so a single\n * shared gateway can be sliced per-app and per-feature in the AI Gateway\n * dashboard and logs. This replaces the per-app-gateway convention: rather than\n * one gateway per app (which has to be provisioned and silently 401s when it\n * isn't), one gateway tags every request with who made it.\n *\n * Returns `undefined` when no attribution fields are set (header omitted).\n * The CF AI Gateway accepts a JSON object of string/number/boolean values.\n */\nfunction buildAigMetadata(opts: LLMOptions): string | undefined {\n  const meta: Record<string, string> = {};\n  if (opts.project) meta.project = opts.project;\n  if (opts.workload) meta.workload = opts.workload;\n  if (opts.actor) meta.actor = opts.actor;\n  if (opts.runId) meta.runId = opts.runId;\n  return Object.keys(meta).length > 0 ? JSON.stringify(meta) : undefined;\n}\n\nasync function callOne(\n  leg: { provider: LLMProvider; model: string },\n  messages: LLMMessage[],\n  opts: LLMOptions,\n  env: LLMEnv,\n  fetchImpl: typeof fetch,\n  logger: Logger | undefined,\n  nowFn?: () => number,\n): Promise<{ parsed: { content: string; input: number; output: number; cacheRead?: number; cacheWrite?: number; model?: string; toolCalls?: LLMToolCall[]; stopReason?: LLMResult['stopReason'] }; gatewayRequestId?: string; attempts: number }> {\n  let req: { url: string; headers: Record<string, string>; body: string };\n  switch (leg.provider) {\n    case 'anthropic':\n      req = buildAnthropicRequest(leg.model, messages, opts, env);\n      break;\n    case 'gemini':\n      req = buildGeminiRequest(leg.model, messages, opts, env);\n      break;\n    case 'groq':\n      req = buildGroqRequest(leg.model, messages, opts, env);\n      break;\n    case 'grok':\n      req = buildGrokRequest(leg.model, messages, opts, env);\n      break;\n    case 'deepseek':\n      req = buildDeepSeekRequest(leg.model, messages, opts, env);\n      break;\n  }\n  // Attribution for the shared AI Gateway — one gateway, sliced per-app/feature.\n  const aigMetadata = buildAigMetadata(opts);\n  if (aigMetadata) req.headers['cf-aig-metadata'] = aigMetadata;\n  const { json, gatewayRequestId, attempts } = await callWithBackoff(\n    leg.provider,\n    req,\n    fetchImpl,\n    opts.signal,\n    logger,\n    nowFn,\n  );\n  switch (leg.provider) {\n    case 'anthropic':\n      return { parsed: parseAnthropic(json), gatewayRequestId, attempts };\n    case 'gemini':\n      return { parsed: parseGemini(json), gatewayRequestId, attempts };\n    case 'groq':\n      return { parsed: parseGroq(json), gatewayRequestId, attempts };\n    case 'grok':\n      return { parsed: parseOpenAi(json), gatewayRequestId, attempts };\n    case 'deepseek':\n      return { parsed: parseOpenAi(json), gatewayRequestId, attempts };\n  }\n}\n\n/**\n * Run a completion through the routing plan for the requested tier.\n *\n * Routing summary (0.3.0):\n *   - `fast`     → Grok 4.3; Anthropic Haiku fallback when Grok is unavailable\n *   - `balanced` → Anthropic Sonnet; Gemini 2.5 Pro if `longContextThreshold` exceeded\n *   - `smart`    → Anthropic Opus; Gemini 2.5 Pro if long-context\n *   - `verifier` → Groq Llama 3.3 70B (no fallback — verifier is inherently cheap/best-effort)\n *   - `workbench` → DeepSeek Chat; Groq fallback for boring/reviewable internal batch jobs\n *\n * All provider traffic flows through Cloudflare AI Gateway at `AI_GATEWAY_BASE_URL`.\n *\n * Per-provider reliability guarantees (0.4.0):\n *   - Exponential backoff with jitter on 429 / 5xx (base 500ms, cap 8s, up to 2 retries).\n *   - Provider cooldown: after exhausting retries the provider is marked cooling down\n *     for 30 seconds; subsequent calls skip it and go straight to the fallback leg.\n *\n * @param messages - Ordered chat history.\n * @param env - API key + gateway bindings.\n * @param opts - Optional tier/model/parameters override.\n * @param deps - Optional fetch/logger/clock injection (for testing).\n * @returns A {@link FactoryResponse} carrying either an {@link LLMResult} or\n *   an error (`LLM_ALL_PROVIDERS_FAILED`, `LLM_RATE_LIMITED`, or `INTERNAL_ERROR`).\n */\nexport async function complete(\n  messages: LLMMessage[],\n  env: LLMEnv,\n  opts: LLMOptions = {},\n  deps: LLMDeps = {},\n): Promise<FactoryResponse<LLMResult>> {\n  if (messages.length === 0) {\n    throw new ValidationError('messages must not be empty');\n  }\n  if (!env.AI_GATEWAY_BASE_URL) {\n    throw new ValidationError('AI_GATEWAY_BASE_URL is required in 0.3.0');\n  }\n  const fetchImpl = deps.fetch ?? fetch;\n  const now = deps.now ?? (() => Date.now());\n  const logger = deps.logger;\n  const startedAt = now();\n\n  const tier: LLMTier = opts.tier ?? 'balanced';\n  const system = systemText(opts, messages);\n  const tokenEstimate = estimateTokens(messages, system);\n  const route = plan(tier, opts, tokenEstimate);\n\n  // ── Org-level daily / monthly cap pre-check ──────────────────────────────\n  const kv = env.LLM_COST_KV;\n  const todayKey = `llm:daily-cost:${isoDate(now())}`;\n  const monthKey = `llm:monthly-cost:${isoMonth(now())}`;\n  if (kv) {\n    if (opts.dailyCapUsd !== undefined) {\n      const raw = await kv.get(todayKey).catch(() => null);\n      const spent = parseFloat(raw ?? '0');\n      if (spent >= opts.dailyCapUsd) {\n        return toErrorResponse(\n          new RateLimitError('LLM_DAILY_CAP_EXCEEDED', {\n            spentUsd: spent,\n            dailyCapUsd: opts.dailyCapUsd,\n          }),\n        );\n      }\n    }\n    if (opts.monthlyCapUsd !== undefined) {\n      const raw = await kv.get(monthKey).catch(() => null);\n      const spent = parseFloat(raw ?? '0');\n      if (spent >= opts.monthlyCapUsd) {\n        return toErrorResponse(\n          new RateLimitError('LLM_MONTHLY_CAP_EXCEEDED', {\n            spentUsd: spent,\n            monthlyCapUsd: opts.monthlyCapUsd,\n          }),\n        );\n      }\n    }\n  }\n\n  const attemptLog: Array<{ provider: LLMProvider; status?: number; message: string }> = [];\n\n  let routeLegs = [route.primary, route.fallback].filter(Boolean) as Array<{ provider: LLMProvider; model: string }>;\n  // Tool-calling fails closed: never fall back to a provider that can't honour\n  // the tool schema. Narrow the route to tool-capable providers when tools are set.\n  if (opts.tools && opts.tools.length > 0) {\n    routeLegs = routeLegs.filter((l) => TOOL_CAPABLE_PROVIDERS.has(l.provider));\n    if (routeLegs.length === 0) {\n      throw new ValidationError(\n        `tool-calling requires a tool-capable provider (${[...TOOL_CAPABLE_PROVIDERS].join(', ')}); tier '${tier}' has none — use tier fast/balanced/smart or a claude-* model override`,\n      );\n    }\n  }\n  for (const [legIndex, leg] of routeLegs.entries()) {\n    // Skip providers that are currently in their cooldown window.\n    if (isProviderCoolingDown(leg.provider, now)) {\n      logger?.warn?.('llm.provider.coolingDown', { provider: leg.provider });\n      attemptLog.push({ provider: leg.provider, message: 'skipped: cooling down' });\n      continue;\n    }\n    if (opts.signal?.aborted) {\n      return toErrorResponse(\n        new InternalError('llm call aborted', { provider: leg.provider, model: leg.model }),\n      );\n    }\n    try {\n      const result = await callOne(leg, messages, opts, env, fetchImpl, logger, now);\n      // A tool_use turn legitimately has no text content — only treat a\n      // genuinely empty response (no text AND no tool calls) as a failure.\n      if (!result.parsed.content && !(result.parsed.toolCalls && result.parsed.toolCalls.length > 0)) {\n        throw { provider: leg.provider, status: 200, retryable: false, message: 'empty content' } satisfies ProviderError;\n      }\n      logger?.info?.('llm.complete', {\n        provider: leg.provider,\n        model: leg.model,\n        tier,\n        tokenEstimate,\n        attempts: result.attempts,\n        runId: opts.runId,\n        project: opts.project,\n        actor: opts.actor,\n        workload: opts.workload,\n      });\n      const llmResult: LLMResult = {\n        content: result.parsed.content,\n        provider: leg.provider,\n        model: result.parsed.model ?? leg.model,\n        tier,\n        tokens: {\n          input: result.parsed.input,\n          output: result.parsed.output,\n          cacheRead: result.parsed.cacheRead,\n          cacheWrite: result.parsed.cacheWrite,\n        },\n        latency: now() - startedAt,\n        attempts: result.attempts,\n        gatewayRequestId: result.gatewayRequestId,\n        stopReason: result.parsed.stopReason,\n        toolCalls: result.parsed.toolCalls,\n      };\n      const costUsd = estimateCostUsd(llmResult.tokens, llmResult.model);\n      if (opts.maxCostUsd !== undefined && costUsd > opts.maxCostUsd) {\n        if (kv && (opts.dailyCapUsd !== undefined || opts.monthlyCapUsd !== undefined)) {\n          await recordOrgCostUsage(kv, todayKey, monthKey, costUsd, opts);\n        }\n        return toErrorResponse(\n          new RateLimitError('LLM_COST_CAP_EXCEEDED', {\n            costUsd,\n            maxCostUsd: opts.maxCostUsd,\n            model: llmResult.model,\n            tokens: llmResult.tokens,\n          }),\n        );\n      }\n      // ── Update org-level cost accumulators in KV ─────────────────────────\n      // The KV writes are best-effort. Cloudflare KV does not support atomic\n      // compare-and-swap, so concurrent increments may undercount spend.\n      if (kv && (opts.dailyCapUsd !== undefined || opts.monthlyCapUsd !== undefined)) {\n        await recordOrgCostUsage(kv, todayKey, monthKey, costUsd, opts);\n      }\n      // ── Metering callback ────────────────────────────────────────────────\n      if (deps.onRecord && opts.ledger) {\n        const row: LLMRecordRow = {\n          ...opts.ledger,\n          model: llmResult.model,\n          provider: llmResult.provider,\n          tier: llmResult.tier,\n          inputTokens: llmResult.tokens.input,\n          outputTokens: llmResult.tokens.output,\n          cacheReadTokens: llmResult.tokens.cacheRead ?? 0,\n          cacheWriteTokens: llmResult.tokens.cacheWrite ?? 0,\n          latencyMs: llmResult.latency,\n          costUsd,\n          yyyyMm: isoMonth(now()),\n        };\n        deps.onRecord(row).catch((e: unknown) => {\n          logger?.warn?.('llm.onRecord.error', { message: e instanceof Error ? e.message : String(e) });\n        });\n      }\n      return { data: llmResult, error: null };\n    } catch (e) {\n      if (e instanceof DOMException && e.name === 'AbortError') {\n        return toErrorResponse(\n          new InternalError('llm call aborted', { provider: leg.provider, model: leg.model }),\n        );\n      }\n      if (isProviderError(e)) {\n        attemptLog.push({ provider: e.provider, status: e.status, message: e.message });\n        if (e.status === 429 && legIndex === routeLegs.length - 1) {\n          return toErrorResponse(\n            new RateLimitError(`llm rate limited on ${e.provider}`, { attempts: attemptLog }),\n          );\n        }\n        logger?.warn?.('llm.leg.failed', { provider: leg.provider, status: e.status });\n        continue;\n      }\n      attemptLog.push({ provider: leg.provider, message: e instanceof Error ? e.message : String(e) });\n    }\n  }\n\n  return toErrorResponse(\n    new InternalError('LLM_ALL_PROVIDERS_FAILED', { attempts: attemptLog, tier, tokenEstimate }),\n  );\n}\n\n// ─── Streaming ────────────────────────────────────────────────────────────\n\n/**\n * Anthropic server-sent event shapes used by the streaming parser.\n * Only the fields we consume are typed; the rest are ignored.\n */\ninterface AnthropicStreamEvent {\n  type: string;\n  index?: number;\n  delta?: { type?: string; text?: string };\n  message?: {\n    usage?: {\n      input_tokens?: number;\n      output_tokens?: number;\n      cache_read_input_tokens?: number;\n      cache_creation_input_tokens?: number;\n    };\n    model?: string;\n  };\n  usage?: {\n    input_tokens?: number;\n    output_tokens?: number;\n  };\n}\n\n/**\n * Streams a completion from the primary Anthropic provider, yielding text chunks\n * as they arrive. Falls back to the non-streaming {@link complete} function when\n * the provider does not support streaming (i.e. a non-Anthropic primary is selected).\n *\n * The generator's **return value** (accessible via `gen.return()` or by consuming\n * the full iteration) is an {@link LLMResult} with the same shape as {@link complete}.\n *\n * Usage pattern:\n * ```ts\n * const gen = completionStream(messages, env, opts);\n * for await (const chunk of gen) {\n *   // stream chunk to client\n * }\n * const result = (await gen.return(undefined)).value; // LLMResult\n * ```\n *\n * @param messages - Ordered chat history.\n * @param env - API key + gateway bindings.\n * @param opts - Optional tier/model/parameters override. Accepts `deps` as nested field.\n * @returns An async generator that yields `string` chunks and returns an {@link LLMResult}.\n */\nexport async function* completionStream(\n  messages: LLMMessage[],\n  env: LLMEnv,\n  opts: LLMOptions & { deps?: LLMDeps } = {},\n): AsyncGenerator<string, LLMResult, unknown> {\n  if (messages.length === 0) {\n    throw new ValidationError('messages must not be empty');\n  }\n  if (!env.AI_GATEWAY_BASE_URL) {\n    throw new ValidationError('AI_GATEWAY_BASE_URL is required in 0.3.0');\n  }\n\n  const deps: LLMDeps = opts.deps ?? {};\n  const fetchImpl = deps.fetch ?? fetch;\n  const now = deps.now ?? (() => Date.now());\n  const logger = deps.logger;\n  const startedAt = now();\n\n  const tier: LLMTier = opts.tier ?? 'balanced';\n  const system = systemText(opts, messages);\n  const tokenEstimate = estimateTokens(messages, system);\n  const route = plan(tier, opts, tokenEstimate);\n  const streamLeg =\n    route.primary.provider === 'grok' && !env.GROK_API_KEY && route.fallback?.provider === 'anthropic'\n      ? route.fallback\n      : route.primary;\n\n  // Only Anthropic supports streaming in the current implementation.\n  // For all other primaries, fall back to non-streaming complete().\n  if (streamLeg.provider !== 'anthropic') {\n    const result = await complete(messages, env, opts, deps);\n    if (result.error !== null || result.data === null) {\n      throw new InternalError('LLM_ALL_PROVIDERS_FAILED', { error: result.error });\n    }\n    yield result.data.content;\n    return result.data;\n  }\n\n  // Check cooldown before attempting the streaming call.\n  if (isProviderCoolingDown(streamLeg.provider, now)) {\n    logger?.warn?.('llm.provider.coolingDown', { provider: streamLeg.provider });\n    // Fall back to non-streaming complete() which will handle the fallback leg.\n    const result = await complete(messages, env, opts, deps);\n    if (result.error !== null || result.data === null) {\n      throw new InternalError('LLM_ALL_PROVIDERS_FAILED', { error: result.error });\n    }\n    yield result.data.content;\n    return result.data;\n  }\n\n  const req = buildAnthropicRequest(streamLeg.model, messages, opts, env, true);\n  // Attribution for the shared AI Gateway (matches the non-streaming path).\n  const streamAigMetadata = buildAigMetadata(opts);\n  if (streamAigMetadata) req.headers['cf-aig-metadata'] = streamAigMetadata;\n\n  let response: Response;\n  try {\n    response = await fetchImpl(req.url, {\n      method: 'POST',\n      headers: req.headers,\n      body: req.body,\n      // Fall back to a 60 s default when the caller provides no signal — prevents\n      // a hung provider connection from consuming the Worker's wall-clock budget.\n      signal: opts.signal ?? AbortSignal.timeout(60_000),\n    });\n  } catch (e) {\n    if (e instanceof DOMException && e.name === 'AbortError') {\n      throw new InternalError('llm call aborted', {\n        provider: streamLeg.provider,\n        model: streamLeg.model,\n      });\n    }\n    throw new InternalError('llm stream fetch failed', {\n      message: e instanceof Error ? e.message : String(e),\n    });\n  }\n\n  if (!response.ok) {\n    const text = await response.text().catch(() => '');\n    const retryable = isRetryableForBackoff(response.status);\n    if (retryable && response.status === 429) {\n      markProviderCoolingDown(streamLeg.provider, now);\n    }\n    // Fall back to non-streaming complete() which will try the fallback leg.\n    const result = await complete(messages, env, opts, deps);\n    if (result.error !== null || result.data === null) {\n      throw new InternalError('LLM_ALL_PROVIDERS_FAILED', {\n        streamError: `${streamLeg.provider} ${String(response.status)}: ${text.slice(0, 300)}`,\n        error: result.error,\n      });\n    }\n    yield result.data.content;\n    return result.data;\n  }\n\n  if (!response.body) {\n    throw new InternalError('llm stream response body is null', {\n      provider: streamLeg.provider,\n    });\n  }\n\n  // Stream SSE events from Anthropic.\n  const decoder = new TextDecoder();\n  let accumulatedText = '';\n  let inputTokens = 0;\n  let outputTokens = 0;\n  let cacheRead = 0;\n  let cacheWrite = 0;\n  let modelName: string | undefined;\n  const gatewayRequestId: string | undefined = response.headers.get('cf-aig-request-id') ?? undefined;\n\n  const reader = response.body.getReader();\n  let buffer = '';\n\n  try {\n    while (true) {\n      const { done, value } = await reader.read();\n      if (done) break;\n      buffer += decoder.decode(value, { stream: true });\n\n      // SSE lines are delimited by '\\n'. Events are separated by '\\n\\n'.\n      const lines = buffer.split('\\n');\n      // Keep the last (potentially incomplete) line in the buffer.\n      buffer = lines.pop() ?? '';\n\n      for (const line of lines) {\n        if (!line.startsWith('data: ')) continue;\n        const data = line.slice(6).trim();\n        if (data === '[DONE]') break;\n        let event: AnthropicStreamEvent;\n        try {\n          event = JSON.parse(data) as AnthropicStreamEvent;\n        } catch {\n          continue; // Skip malformed SSE lines.\n        }\n\n        switch (event.type) {\n          case 'message_start':\n            inputTokens = event.message?.usage?.input_tokens ?? 0;\n            cacheRead = event.message?.usage?.cache_read_input_tokens ?? 0;\n            cacheWrite = event.message?.usage?.cache_creation_input_tokens ?? 0;\n            modelName = event.message?.model;\n            break;\n          case 'content_block_delta':\n            if (event.delta?.type === 'text_delta' && typeof event.delta.text === 'string') {\n              accumulatedText += event.delta.text;\n              yield event.delta.text;\n            }\n            break;\n          case 'message_delta':\n            outputTokens = event.usage?.output_tokens ?? outputTokens;\n            break;\n          default:\n            break;\n        }\n      }\n    }\n  } finally {\n    reader.releaseLock();\n  }\n\n  clearProviderCooldown(streamLeg.provider);\n  logger?.info?.('llm.completionStream', {\n    provider: streamLeg.provider,\n    model: streamLeg.model,\n    tier,\n    tokenEstimate,\n    runId: opts.runId,\n    project: opts.project,\n    actor: opts.actor,\n    workload: opts.workload,\n  });\n\n  return {\n    content: accumulatedText,\n    provider: streamLeg.provider,\n    model: modelName ?? streamLeg.model,\n    tier,\n    tokens: { input: inputTokens, output: outputTokens, cacheRead, cacheWrite },\n    latency: now() - startedAt,\n    attempts: 1,\n    gatewayRequestId,\n  };\n}\n\n// ─── Grounding assertion ───────────────────────────────────────────────────\n\n/**\n * Returns `true` if `response` contains at least one verbatim phrase of at\n * least 5 consecutive whitespace-delimited tokens that also appears in one of\n * the `sources` strings.\n *\n * Returns `true` unconditionally when `sources` is empty (no grounding\n * documents means grounding cannot be violated).\n *\n * This is a lightweight guard for RAG pipelines — it detects obvious\n * hallucinations where the model generates content not present in any\n * retrieved source. It is NOT a semantic similarity check.\n *\n * @param response - The LLM-generated text to inspect.\n * @param sources - Retrieved source documents to check against.\n * @returns `true` if the response is grounded, `false` if hallucination detected.\n *\n * @example\n * ```ts\n * const grounded = assertGrounding(llmAnswer, retrievedDocs);\n * if (!grounded) {\n *   // flag or re-rank the response\n * }\n * ```\n */\nexport function assertGrounding(response: string, sources: string[]): boolean {\n  if (sources.length === 0) return true;\n\n  const WINDOW = 5;\n  const responseTokens = response.split(/\\s+/).filter((t) => t.length > 0);\n\n  if (responseTokens.length < WINDOW) return false;\n\n  // Build a set of all 5-token ngrams from each source for O(n) lookup.\n  const sourceNgrams = new Set<string>();\n  for (const source of sources) {\n    const tokens = source.split(/\\s+/).filter((t) => t.length > 0);\n    for (let i = 0; i <= tokens.length - WINDOW; i++) {\n      const ngram = tokens.slice(i, i + WINDOW).join(' ');\n      sourceNgrams.add(ngram);\n    }\n  }\n\n  if (sourceNgrams.size === 0) return false;\n\n  // Slide a window of WINDOW tokens over the response and check for a match.\n  for (let i = 0; i <= responseTokens.length - WINDOW; i++) {\n    const ngram = responseTokens.slice(i, i + WINDOW).join(' ');\n    if (sourceNgrams.has(ngram)) return true;\n  }\n\n  return false;\n}\n\n// ─── Exported helpers (kept for existing consumers) ───────────────────────\n\nexport { MODELS, isProviderCoolingDown, markProviderCoolingDown, clearProviderCooldown, PROVIDER_COOLDOWN_MS };\nexport { BASE_BACKOFF_MS };\n"],"mappings":";AAAA;AAAA,EACE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AAAA,OAEK;AAqDP,SAAS,cAAc,SAA6C;AAClE,MAAI,OAAO,YAAY,SAAU,QAAO;AACxC,SAAO,QACJ,IAAI,CAAC,MAAO,EAAE,SAAS,SAAS,EAAE,OAAO,EAAE,SAAS,gBAAgB,EAAE,UAAU,EAAG,EACnF,KAAK,EAAE;AACZ;AAMA,SAAS,WAAW,MAAkB,UAA4C;AAChF,MAAI,KAAK,WAAW,OAAW,QAAO,KAAK;AAC3C,QAAM,IAAI,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,QAAQ,GAAG;AACrD,SAAO,MAAM,SAAY,SAAY,cAAc,CAAC;AACtD;AAwMA,IAAM,SAAS;AAAA,EACb,WAAW;AAAA,IACT,MAAM;AAAA,IACN,UAAU;AAAA,IACV,OAAO;AAAA,EACT;AAAA,EACA,QAAQ;AAAA,IACN,OAAO;AAAA,EACT;AAAA,EACA,MAAM;AAAA,IACJ,UAAU;AAAA,EACZ;AAAA,EACA,MAAM;AAAA,IACJ,MAAM;AAAA,EACR;AAAA,EACA,UAAU;AAAA,IACR,WAAW;AAAA,EACb;AACF;AAEA,IAAM,qBAAqB;AAC3B,IAAM,sBAAsB;AAC5B,IAAM,iCAAiC;AAIvC,IAAM,kBAAkB;AAExB,IAAM,iBAAiB;AAEvB,IAAM,wBAAwB;AAE9B,IAAM,4BAA4B;AAQlC,IAAM,wBAAkD,oBAAI,IAAI;AAGhE,IAAM,uBAAuB;AAM7B,SAAS,sBAAsB,UAAuB,MAAoB,KAAK,KAAc;AAC3F,QAAM,QAAQ,sBAAsB,IAAI,QAAQ;AAChD,MAAI,UAAU,OAAW,QAAO;AAChC,SAAO,IAAI,IAAI;AACjB;AAGA,SAAS,QAAQ,OAAuB;AACtC,SAAO,IAAI,KAAK,KAAK,EAAE,YAAY,EAAE,MAAM,GAAG,EAAE;AAClD;AAOA,eAAe,mBACb,IACA,UACA,UACA,SACA,MACe;AACf,MAAI,KAAK,gBAAgB,QAAW;AAClC,UAAM,MAAM,MAAM,GAAG,IAAI,QAAQ,EAAE,MAAM,MAAM,IAAI;AACnD,UAAM,QAAQ,WAAW,OAAO,GAAG;AACnC,UAAM,GAAG,IAAI,UAAU,OAAO,QAAQ,OAAO,GAAG;AAAA,MAAE,eAAe;AAAA;AAAA,IAAmB,CAAC,EAAE,MAAM,MAAM,MAAS;AAAA,EAC9G;AACA,MAAI,KAAK,kBAAkB,QAAW;AACpC,UAAM,MAAM,MAAM,GAAG,IAAI,QAAQ,EAAE,MAAM,MAAM,IAAI;AACnD,UAAM,QAAQ,WAAW,OAAO,GAAG;AACnC,UAAM,GAAG,IAAI,UAAU,OAAO,QAAQ,OAAO,GAAG;AAAA,MAAE,eAAe;AAAA;AAAA,IAAqB,CAAC,EAAE,MAAM,MAAM,MAAS;AAAA,EAChH;AACF;AAYO,IAAM,qBAA+G;AAAA;AAAA,EAE1H,2BAA2B,EAAE,OAAO,KAAM,QAAQ,GAAM,WAAW,MAAM,YAAY,EAAK;AAAA,EAC1F,6BAA6B,EAAE,OAAO,KAAM,QAAQ,GAAM,WAAW,MAAM,YAAY,EAAK;AAAA;AAAA,EAE5F,4BAA4B,EAAE,OAAO,GAAM,QAAQ,IAAO,WAAW,KAAM,YAAY,KAAK;AAAA,EAC5F,qBAAqB,EAAE,OAAO,GAAM,QAAQ,IAAO,WAAW,KAAM,YAAY,KAAK;AAAA;AAAA,EAErF,0BAA0B,EAAE,OAAO,IAAO,QAAQ,IAAO,WAAW,KAAM,YAAY,MAAM;AAAA,EAC5F,mBAAmB,EAAE,OAAO,IAAO,QAAQ,IAAO,WAAW,KAAM,YAAY,MAAM;AAAA;AAAA,EAErF,kBAAkB,EAAE,OAAO,MAAM,QAAQ,IAAO,WAAW,MAAM,YAAY,IAAK;AAAA;AAAA,EAElF,oBAAoB,EAAE,OAAO,KAAM,QAAQ,MAAM,WAAW,MAAM,YAAY,IAAK;AAAA;AAAA,EAEnF,YAAY,EAAE,OAAO,MAAM,QAAQ,KAAM,WAAW,GAAM,YAAY,EAAK;AAAA;AAAA,EAE3E,iBAAiB,EAAE,OAAO,MAAM,QAAQ,KAAM,WAAW,MAAM,YAAY,KAAK;AAAA,EAChF,qBAAqB,EAAE,OAAO,MAAM,QAAQ,MAAM,WAAW,MAAM,YAAY,KAAK;AAAA;AAAA,EAEpF,eAAe,EAAE,OAAO,MAAM,QAAQ,KAAM,WAAW,GAAM,YAAY,EAAK;AAAA,EAC9E,sBAAsB,EAAE,OAAO,MAAM,QAAQ,KAAM,WAAW,GAAM,YAAY,EAAK;AACvF;AAGA,IAAM,iBAAiB,mBAAmB,iBAAiB;AAO3D,SAAS,gBACP,QACA,OACQ;AACR,QAAM,QAAQ,mBAAmB,KAAK,KAAK;AAC3C,UACG,OAAO,QAAQ,MAAM,QACpB,OAAO,SAAS,MAAM,UACrB,OAAO,aAAa,KAAK,MAAM,aAC/B,OAAO,cAAc,KAAK,MAAM,cACnC;AAEJ;AAGA,SAAS,SAAS,OAAuB;AACvC,SAAO,IAAI,KAAK,KAAK,EAAE,YAAY,EAAE,MAAM,GAAG,CAAC;AACjD;AAKA,SAAS,wBAAwB,UAAuB,MAAoB,KAAK,KAAW;AAC1F,wBAAsB,IAAI,UAAU,IAAI,IAAI,oBAAoB;AAClE;AAKA,SAAS,sBAAsB,UAA6B;AAC1D,wBAAsB,OAAO,QAAQ;AACvC;AAGA,IAAM,kBAAkB;AAaxB,SAAS,sBAAsB,QAAyB;AACtD,SAAO,WAAW,OAAQ,UAAU,OAAO,SAAS;AACtD;AAEA,SAAS,eAAe,UAAwB,QAAyB;AAEvE,MAAI,QAAQ,QAAQ,UAAU;AAC9B,aAAW,KAAK,SAAU,UAAS,cAAc,EAAE,OAAO,EAAE;AAC5D,SAAO,KAAK,KAAK,QAAQ,CAAC;AAC5B;AAEA,SAAS,MAAM,IAAY,QAAqC;AAC9D,SAAO,IAAI,QAAQ,CAAC,SAAS,WAAW;AACtC,UAAM,IAAI,WAAW,SAAS,EAAE;AAChC,QAAI,QAAQ;AACV,YAAM,UAAU,MAAM;AACpB,qBAAa,CAAC;AACd,eAAO,IAAI,aAAa,WAAW,YAAY,CAAC;AAAA,MAClD;AACA,UAAI,OAAO,QAAS,SAAQ;AAAA,UACvB,QAAO,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;AAAA,IAC/D;AAAA,EACF,CAAC;AACH;AAUA,SAAS,iBAAiB,SAAyB;AACjD,QAAM,SAAS,KAAK,MAAM,KAAK,OAAO,IAAI,qBAAqB;AAC/D,SAAO,KAAK,IAAI,kBAAkB,KAAK,IAAI,GAAG,OAAO,IAAI,QAAQ,cAAc;AACjF;AAIA,SAAS,sBACP,OACA,UACA,MACA,KACA,YAAY,OACoD;AAChE,QAAM,MAAM,WAAW,MAAM,QAAQ;AACrC,QAAM,WAAW,SAAS,OAAO,CAAC,MAAM,EAAE,SAAS,QAAQ;AAC3D,QAAM,OAAgC;AAAA,IACpC;AAAA,IACA,YAAY,KAAK,aAAa;AAAA,IAC9B,aAAa,KAAK,eAAe;AAAA,IACjC,UAAU,SAAS,IAAI,CAAC,OAAO,EAAE,MAAM,EAAE,MAAM,SAAS,EAAE,QAAQ,EAAE;AAAA,EACtE;AACA,MAAI,WAAW;AACb,SAAK,SAAS;AAAA,EAChB;AACA,MAAI,KAAK;AACP,UAAM,QAAQ,KAAK,eAAe,IAAI,UAAU;AAChD,SAAK,SAAS,QACV,CAAC,EAAE,MAAM,QAAQ,MAAM,KAAK,eAAe,EAAE,MAAM,YAAY,EAAE,CAAC,IAClE;AAAA,EACN;AACA,MAAI,KAAK,SAAS,KAAK,MAAM,SAAS,GAAG;AACvC,SAAK,QAAQ,KAAK,MAAM,IAAI,CAAC,OAAO;AAAA,MAClC,MAAM,EAAE;AAAA,MACR,aAAa,EAAE,eAAe;AAAA,MAC9B,cAAc,EAAE;AAAA,IAClB,EAAE;AACF,UAAM,KAAK,KAAK,cAAc;AAC9B,SAAK,cACH,OAAO,SACH,EAAE,MAAM,OAAO,IACf,OAAO,SACL,EAAE,MAAM,OAAO,IACf,EAAE,MAAM,QAAQ,MAAM,GAAG,KAAK;AAAA,EACxC;AACA,SAAO;AAAA,IACL,KAAK,GAAG,IAAI,mBAAmB;AAAA,IAC/B,SAAS;AAAA,MACP,gBAAgB;AAAA,MAChB,aAAa,IAAI;AAAA,MACjB,qBAAqB;AAAA,MACrB,kBAAkB;AAAA,IACpB;AAAA,IACA,MAAM,KAAK,UAAU,IAAI;AAAA,EAC3B;AACF;AAEA,SAAS,mBACP,OACA,UACA,MACA,KACgE;AAChE,QAAM,MAAM,WAAW,MAAM,QAAQ;AACrC,QAAM,WAAW,SACd,OAAO,CAAC,MAAM,EAAE,SAAS,QAAQ,EACjC,IAAI,CAAC,OAAO;AAAA,IACX,MAAM,EAAE,SAAS,cAAc,UAAU;AAAA,IACzC,OAAO,CAAC,EAAE,MAAM,cAAc,EAAE,OAAO,EAAE,CAAC;AAAA,EAC5C,EAAE;AACJ,QAAM,OAAgC;AAAA,IACpC;AAAA,IACA,kBAAkB;AAAA,MAChB,iBAAiB,KAAK,aAAa;AAAA,MACnC,aAAa,KAAK,eAAe;AAAA,IACnC;AAAA,EACF;AACA,MAAI,KAAK;AACP,SAAK,oBAAoB,EAAE,OAAO,CAAC,EAAE,MAAM,IAAI,CAAC,EAAE;AAAA,EACpD;AACA,QAAM,OAAO,eAAe,IAAI,cAAc,cAAc,IAAI,eAAe,6BAA6B,KAAK;AACjH,SAAO;AAAA,IACL,KAAK,GAAG,IAAI,mBAAmB,qBAAqB,IAAI;AAAA,IACxD,SAAS;AAAA,MACP,gBAAgB;AAAA,MAChB,eAAe,UAAU,IAAI,mBAAmB;AAAA,IAClD;AAAA,IACA,MAAM,KAAK,UAAU,IAAI;AAAA,EAC3B;AACF;AAUA,SAAS,iBAAiB,UAAwB,KAAyD;AACzG,QAAM,MAAsC,CAAC;AAC7C,MAAI,IAAK,KAAI,KAAK,EAAE,MAAM,UAAU,SAAS,IAAI,CAAC;AAClD,aAAW,KAAK,UAAU;AACxB,QAAI,EAAE,SAAS,SAAU;AACzB,QAAI,OAAO,EAAE,YAAY,UAAU;AACjC,UAAI,KAAK,EAAE,MAAM,EAAE,MAAM,SAAS,EAAE,QAAQ,CAAC;AAC7C;AAAA,IACF;AACA,QAAI,OAAO;AACX,UAAM,YAA4C,CAAC;AACnD,UAAM,UAA2D,CAAC;AAClE,eAAW,KAAK,EAAE,SAAS;AACzB,UAAI,EAAE,SAAS,OAAQ,SAAQ,EAAE;AAAA,eACxB,EAAE,SAAS;AAClB,kBAAU,KAAK,EAAE,IAAI,EAAE,IAAI,MAAM,YAAY,UAAU,EAAE,MAAM,EAAE,MAAM,WAAW,KAAK,UAAU,EAAE,KAAK,EAAE,EAAE,CAAC;AAAA,eACtG,EAAE,SAAS,cAAe,SAAQ,KAAK,EAAE,aAAa,EAAE,aAAa,SAAS,EAAE,QAAQ,CAAC;AAAA,IACpG;AACA,QAAI,QAAQ,SAAS,GAAG;AACtB,iBAAW,KAAK,QAAS,KAAI,KAAK,EAAE,MAAM,QAAQ,cAAc,EAAE,aAAa,SAAS,EAAE,QAAQ,CAAC;AACnG,UAAI,KAAM,KAAI,KAAK,EAAE,MAAM,QAAQ,SAAS,KAAK,CAAC;AAAA,IACpD,WAAW,UAAU,SAAS,GAAG;AAC/B,UAAI,KAAK,EAAE,MAAM,aAAa,SAAS,QAAQ,MAAM,YAAY,UAAU,CAAC;AAAA,IAC9E,OAAO;AACL,UAAI,KAAK,EAAE,MAAM,EAAE,MAAM,SAAS,KAAK,CAAC;AAAA,IAC1C;AAAA,EACF;AACA,SAAO;AACT;AAGA,SAAS,YAAY,MAA8D;AACjF,MAAI,CAAC,KAAK,SAAS,KAAK,MAAM,WAAW,EAAG,QAAO;AACnD,SAAO,KAAK,MAAM,IAAI,CAAC,OAAO;AAAA,IAC5B,MAAM;AAAA,IACN,UAAU,EAAE,MAAM,EAAE,MAAM,aAAa,EAAE,eAAe,IAAI,YAAY,EAAE,WAAW;AAAA,EACvF,EAAE;AACJ;AAGA,SAAS,iBAAiB,IAAuC;AAC/D,MAAI,OAAO,OAAW,QAAO;AAC7B,MAAI,OAAO,UAAU,OAAO,OAAQ,QAAO;AAC3C,SAAO,EAAE,MAAM,YAAY,UAAU,EAAE,MAAM,GAAG,KAAK,EAAE;AACzD;AAEA,SAAS,iBACP,OACA,UACA,MACA,KACgE;AAChE,QAAM,MAAM,WAAW,MAAM,QAAQ;AACrC,QAAM,SAAuB,CAAC;AAC9B,MAAI,IAAK,QAAO,KAAK,EAAE,MAAM,UAAU,SAAS,IAAI,CAAC;AACrD,aAAW,KAAK,SAAU,KAAI,EAAE,SAAS,SAAU,QAAO,KAAK,EAAE,MAAM,EAAE,MAAM,SAAS,cAAc,EAAE,OAAO,EAAE,CAAC;AAClH,SAAO;AAAA,IACL,KAAK,GAAG,IAAI,mBAAmB;AAAA,IAC/B,SAAS;AAAA,MACP,gBAAgB;AAAA,MAChB,eAAe,UAAU,IAAI,YAAY;AAAA,IAC3C;AAAA,IACA,MAAM,KAAK,UAAU;AAAA,MACnB;AAAA,MACA,YAAY,KAAK,aAAa;AAAA,MAC9B,aAAa,KAAK,eAAe;AAAA,MACjC,UAAU;AAAA,IACZ,CAAC;AAAA,EACH;AACF;AAEA,SAAS,iBACP,OACA,UACA,MACA,KACgE;AAChE,MAAI,CAAC,IAAI,cAAc;AACrB,UAAM,IAAI,gBAAgB,iDAAiD;AAAA,EAC7E;AACA,QAAM,MAAM,WAAW,MAAM,QAAQ;AACrC,QAAM,OAAgC;AAAA,IACpC;AAAA,IACA,YAAY,KAAK,aAAa;AAAA,IAC9B,aAAa,KAAK,eAAe;AAAA,IACjC,UAAU,iBAAiB,UAAU,GAAG;AAAA,EAC1C;AACA,QAAM,QAAQ,YAAY,IAAI;AAC9B,MAAI,OAAO;AACT,SAAK,QAAQ;AACb,UAAM,KAAK,iBAAiB,KAAK,cAAc,MAAM;AACrD,QAAI,OAAO,OAAW,MAAK,cAAc;AAAA,EAC3C;AACA,MAAI,UAAU,OAAO,KAAK,MAAM;AAC9B,SAAK,mBAAmB,KAAK,mBAAmB;AAAA,EAClD;AACA,SAAO;AAAA,IACL,KAAK,GAAG,IAAI,mBAAmB;AAAA,IAC/B,SAAS;AAAA,MACP,gBAAgB;AAAA,MAChB,eAAe,UAAU,IAAI,YAAY;AAAA,IAC3C;AAAA,IACA,MAAM,KAAK,UAAU,IAAI;AAAA,EAC3B;AACF;AAEA,SAAS,qBACP,OACA,UACA,MACA,KACgE;AAChE,MAAI,CAAC,IAAI,kBAAkB;AACzB,UAAM,IAAI,gBAAgB,2EAA2E;AAAA,EACvG;AACA,QAAM,MAAM,WAAW,MAAM,QAAQ;AACrC,QAAM,OAAgC;AAAA,IACpC;AAAA,IACA,YAAY,KAAK,aAAa;AAAA,IAC9B,aAAa,KAAK,eAAe;AAAA,IACjC,UAAU,iBAAiB,UAAU,GAAG;AAAA,EAC1C;AACA,QAAM,QAAQ,YAAY,IAAI;AAC9B,MAAI,OAAO;AACT,SAAK,QAAQ;AACb,UAAM,KAAK,iBAAiB,KAAK,cAAc,MAAM;AACrD,QAAI,OAAO,OAAW,MAAK,cAAc;AAAA,EAC3C;AACA,SAAO;AAAA,IACL,KAAK,GAAG,IAAI,mBAAmB;AAAA,IAC/B,SAAS;AAAA,MACP,gBAAgB;AAAA,MAChB,eAAe,UAAU,IAAI,gBAAgB;AAAA,IAC/C;AAAA,IACA,MAAM,KAAK,UAAU,IAAI;AAAA,EAC3B;AACF;AAuBA,SAAS,uBAAuB,QAAqD;AACnF,UAAQ,QAAQ;AAAA,IACd,KAAK;AAAA,IACL,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT;AACE,aAAO,SAAS,UAAU;AAAA,EAC9B;AACF;AAgBA,SAAS,eACP,MAUA;AACA,QAAM,IAAI;AACV,QAAM,aAA4B,EAAE,WAAW,CAAC,GAC7C,OAAO,CAAC,MAAM,EAAE,SAAS,cAAc,OAAO,EAAE,OAAO,YAAY,OAAO,EAAE,SAAS,QAAQ,EAC7F,IAAI,CAAC,OAAO,EAAE,IAAI,EAAE,IAAK,MAAM,EAAE,MAAO,WAAW,EAAE,SAAS,CAAC,EAAE,EAAE;AACtE,SAAO;AAAA,IACL,SAAS,EAAE,SAAS,KAAK,CAAC,MAAM,EAAE,SAAS,MAAM,GAAG,QAAQ;AAAA,IAC5D,OAAO,EAAE,OAAO,gBAAgB;AAAA,IAChC,QAAQ,EAAE,OAAO,iBAAiB;AAAA,IAClC,WAAW,EAAE,OAAO,2BAA2B;AAAA,IAC/C,YAAY,EAAE,OAAO,+BAA+B;AAAA,IACpD,OAAO,EAAE;AAAA,IACT,WAAW,UAAU,SAAS,IAAI,YAAY;AAAA,IAC9C,YAAY,uBAAuB,EAAE,WAAW;AAAA,EAClD;AACF;AAEA,SAAS,YAAY,MAAmE;AACtF,QAAM,IAAI;AACV,QAAM,OACJ,EAAE,aAAa,CAAC,GAAG,SAAS,OAAO,IAAI,CAAC,MAAM,EAAE,QAAQ,EAAE,EAAE,KAAK,EAAE,KAAK;AAC1E,SAAO;AAAA,IACL,SAAS;AAAA,IACT,OAAO,EAAE,eAAe,oBAAoB;AAAA,IAC5C,QAAQ,EAAE,eAAe,wBAAwB;AAAA,EACnD;AACF;AAEA,SAAS,UAAU,MAAmF;AACpG,QAAM,IAAI;AACV,SAAO;AAAA,IACL,SAAS,EAAE,UAAU,CAAC,GAAG,SAAS,WAAW;AAAA,IAC7C,OAAO,EAAE,OAAO,iBAAiB;AAAA,IACjC,QAAQ,EAAE,OAAO,qBAAqB;AAAA,IACtC,OAAO,EAAE;AAAA,EACX;AACF;AAeA,SAAS,oBAAoB,QAAqD;AAChF,UAAQ,QAAQ;AAAA,IACd,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AAAA,IACL,KAAK;AACH,aAAO;AAAA,IACT,KAAK;AACH,aAAO;AAAA,IACT;AACE,aAAO,SAAS,UAAU;AAAA,EAC9B;AACF;AAGA,SAAS,cAAc,KAAkD;AACvE,MAAI,CAAC,IAAK,QAAO,CAAC;AAClB,MAAI;AACF,UAAM,IAAI,KAAK,MAAM,GAAG;AACxB,WAAO,OAAO,MAAM,YAAY,MAAM,OAAQ,IAAgC,CAAC;AAAA,EACjF,QAAQ;AACN,WAAO,CAAC;AAAA,EACV;AACF;AAMA,SAAS,YAAY,MAOnB;AACA,QAAM,IAAI;AACV,QAAM,SAAS,EAAE,UAAU,CAAC;AAC5B,QAAM,aAA4B,QAAQ,SAAS,cAAc,CAAC,GAC/D,OAAO,CAAC,MAAM,OAAO,EAAE,UAAU,SAAS,QAAQ,EAClD,IAAI,CAAC,GAAG,OAAO;AAAA,IACd,IAAI,EAAE,MAAM,QAAQ,CAAC;AAAA,IACrB,MAAM,EAAE,SAAU;AAAA,IAClB,WAAW,cAAc,EAAE,UAAU,SAAS;AAAA,EAChD,EAAE;AACJ,SAAO;AAAA,IACL,SAAS,QAAQ,SAAS,WAAW;AAAA,IACrC,OAAO,EAAE,OAAO,iBAAiB;AAAA,IACjC,QAAQ,EAAE,OAAO,qBAAqB;AAAA,IACtC,OAAO,EAAE;AAAA,IACT,WAAW,UAAU,SAAS,IAAI,YAAY;AAAA,IAC9C,YAAY,oBAAoB,QAAQ,aAAa;AAAA,EACvD;AACF;AAmBA,eAAe,gBACb,UACA,SACA,WACA,QACA,QACA,OACyE;AAMzE,WAAS,gBAAgB,KAA2B;AAClD,4BAAwB,UAAU,SAAS,KAAK,GAAG;AACnD,UAAM;AAAA,EACR;AAEA,MAAI;AACJ,WAAS,UAAU,GAAG,WAAW,2BAA2B,WAAW;AACrE,QAAI;AACF,YAAM,WAAW,MAAM,UAAU,QAAQ,KAAK;AAAA,QAC5C,QAAQ;AAAA,QACR,SAAS,QAAQ;AAAA,QACjB,MAAM,QAAQ;AAAA,QACd;AAAA,MACF,CAAC;AACD,UAAI,CAAC,SAAS,IAAI;AAChB,cAAM,OAAO,MAAM,SAAS,KAAK,EAAE,MAAM,MAAM,EAAE;AACjD,cAAM,YAAY,sBAAsB,SAAS,MAAM;AACvD,cAAM,MAAqB;AAAA,UACzB;AAAA,UACA,QAAQ,SAAS;AAAA,UACjB;AAAA,UACA,SAAS,GAAG,QAAQ,IAAI,OAAO,SAAS,MAAM,CAAC,KAAK,KAAK,MAAM,GAAG,GAAG,CAAC;AAAA,QACxE;AACA,gBAAQ,OAAO,sBAAsB,EAAE,UAAU,QAAQ,SAAS,QAAQ,QAAQ,CAAC;AACnF,YAAI,CAAC,IAAI,aAAa,YAAY,2BAA2B;AAC3D,cAAI,IAAI,UAAW,iBAAgB,GAAG;AACtC,gBAAM;AAAA,QACR;AACA,kBAAU;AAAA,MACZ,OAAO;AACL,cAAM,mBAAmB,SAAS,QAAQ,IAAI,mBAAmB,KAAK;AACtE,8BAAsB,QAAQ;AAC9B,eAAO,EAAE,MAAM,MAAM,SAAS,KAAK,GAAG,kBAAkB,UAAU,QAAQ;AAAA,MAC5E;AAAA,IACF,SAAS,GAAG;AACV,UAAI,aAAa,gBAAgB,EAAE,SAAS,aAAc,OAAM;AAChE,UAAI,OAAO,MAAM,YAAY,MAAM,QAAQ,eAAe,GAAG;AAC3D,cAAM,MAAM;AACZ,YAAI,CAAC,IAAI,aAAa,YAAY,2BAA2B;AAC3D,cAAI,IAAI,UAAW,iBAAgB,GAAG;AACtC,gBAAM;AAAA,QACR;AACA,kBAAU;AAAA,MACZ,OAAO;AACL,cAAM,MAAqB;AAAA,UACzB;AAAA,UACA,QAAQ;AAAA,UACR,WAAW;AAAA,UACX,SAAS,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC;AAAA,QACpD;AACA,YAAI,YAAY,0BAA2B,iBAAgB,GAAG;AAC9D,kBAAU;AAAA,MACZ;AAAA,IACF;AAEA,UAAM,YAAY,iBAAiB,UAAU,CAAC;AAC9C,UAAM,MAAM,WAAW,MAAM;AAAA,EAC/B;AAEA,0BAAwB,UAAU,SAAS,KAAK,GAAG;AACnD,QAAM,WAAY,EAAE,UAAU,QAAQ,GAAG,WAAW,OAAO,SAAS,YAAY;AAClF;AAEA,SAAS,gBAAgB,KAAoC;AAC3D,SACE,OAAO,QAAQ,YACf,QAAQ,QACR,OAAQ,IAA6B,WAAW,YAChD,OAAQ,IAA8B,YAAY,YAClD,OAAQ,IAA+B,aAAa;AAExD;AAeA,IAAM,yBAAyB,oBAAI,IAAiB,CAAC,aAAa,QAAQ,UAAU,CAAC;AAErF,SAAS,KAAK,MAAe,MAAkB,eAAkC;AAC/E,MAAI,KAAK,OAAO;AAEd,UAAM,IAAI,KAAK;AACf,QAAI,EAAE,WAAW,QAAQ,EAAG,QAAO,EAAE,SAAS,EAAE,UAAU,aAAa,OAAO,EAAE,EAAE;AAClF,QAAI,EAAE,WAAW,QAAQ,EAAG,QAAO,EAAE,SAAS,EAAE,UAAU,UAAU,OAAO,EAAE,EAAE;AAC/E,QAAI,EAAE,WAAW,MAAM,EAAG,QAAO,EAAE,SAAS,EAAE,UAAU,QAAQ,OAAO,EAAE,EAAE;AAC3E,QAAI,EAAE,WAAW,UAAU,EAAG,QAAO,EAAE,SAAS,EAAE,UAAU,YAAY,OAAO,EAAE,EAAE;AACnF,WAAO,EAAE,SAAS,EAAE,UAAU,QAAQ,OAAO,EAAE,EAAE;AAAA,EACnD;AACA,QAAM,cAAc,kBAAkB,KAAK,wBAAwB;AACnE,UAAQ,MAAM;AAAA,IACZ,KAAK;AACH,aAAO;AAAA,QACL,SAAS,EAAE,UAAU,YAAY,OAAO,OAAO,SAAS,UAAU;AAAA,QAClE,UAAU,EAAE,UAAU,QAAQ,OAAO,OAAO,KAAK,SAAS;AAAA,MAC5D;AAAA,IACF,KAAK;AACH,aAAO,EAAE,SAAS,EAAE,UAAU,QAAQ,OAAO,OAAO,KAAK,SAAS,EAAE;AAAA,IACtE,KAAK;AACH,aAAO,cACH;AAAA,QACE,SAAS,EAAE,UAAU,UAAU,OAAO,OAAO,OAAO,MAAM;AAAA,QAC1D,UAAU,EAAE,UAAU,aAAa,OAAO,OAAO,UAAU,MAAM;AAAA,MACnE,IACA;AAAA,QACE,SAAS,EAAE,UAAU,aAAa,OAAO,OAAO,UAAU,MAAM;AAAA,QAChE,UAAU,EAAE,UAAU,UAAU,OAAO,OAAO,OAAO,MAAM;AAAA,MAC7D;AAAA,IACN,KAAK;AACH,aAAO;AAAA,QACL,SAAS,EAAE,UAAU,QAAQ,OAAO,OAAO,KAAK,KAAK;AAAA,QACrD,UAAU,EAAE,UAAU,aAAa,OAAO,OAAO,UAAU,KAAK;AAAA,MAClE;AAAA,IACF,KAAK;AAAA,IACL;AACE,aAAO,cACH;AAAA,QACE,SAAS,EAAE,UAAU,UAAU,OAAO,OAAO,OAAO,MAAM;AAAA,QAC1D,UAAU,EAAE,UAAU,aAAa,OAAO,OAAO,UAAU,SAAS;AAAA,MACtE,IACA;AAAA,QACE,SAAS,EAAE,UAAU,aAAa,OAAO,OAAO,UAAU,SAAS;AAAA,QACnE,UAAU,EAAE,UAAU,UAAU,OAAO,OAAO,OAAO,MAAM;AAAA,MAC7D;AAAA,EACR;AACF;AAcA,SAAS,iBAAiB,MAAsC;AAC9D,QAAM,OAA+B,CAAC;AACtC,MAAI,KAAK,QAAS,MAAK,UAAU,KAAK;AACtC,MAAI,KAAK,SAAU,MAAK,WAAW,KAAK;AACxC,MAAI,KAAK,MAAO,MAAK,QAAQ,KAAK;AAClC,MAAI,KAAK,MAAO,MAAK,QAAQ,KAAK;AAClC,SAAO,OAAO,KAAK,IAAI,EAAE,SAAS,IAAI,KAAK,UAAU,IAAI,IAAI;AAC/D;AAEA,eAAe,QACb,KACA,UACA,MACA,KACA,WACA,QACA,OACgP;AAChP,MAAI;AACJ,UAAQ,IAAI,UAAU;AAAA,IACpB,KAAK;AACH,YAAM,sBAAsB,IAAI,OAAO,UAAU,MAAM,GAAG;AAC1D;AAAA,IACF,KAAK;AACH,YAAM,mBAAmB,IAAI,OAAO,UAAU,MAAM,GAAG;AACvD;AAAA,IACF,KAAK;AACH,YAAM,iBAAiB,IAAI,OAAO,UAAU,MAAM,GAAG;AACrD;AAAA,IACF,KAAK;AACH,YAAM,iBAAiB,IAAI,OAAO,UAAU,MAAM,GAAG;AACrD;AAAA,IACF,KAAK;AACH,YAAM,qBAAqB,IAAI,OAAO,UAAU,MAAM,GAAG;AACzD;AAAA,EACJ;AAEA,QAAM,cAAc,iBAAiB,IAAI;AACzC,MAAI,YAAa,KAAI,QAAQ,iBAAiB,IAAI;AAClD,QAAM,EAAE,MAAM,kBAAkB,SAAS,IAAI,MAAM;AAAA,IACjD,IAAI;AAAA,IACJ;AAAA,IACA;AAAA,IACA,KAAK;AAAA,IACL;AAAA,IACA;AAAA,EACF;AACA,UAAQ,IAAI,UAAU;AAAA,IACpB,KAAK;AACH,aAAO,EAAE,QAAQ,eAAe,IAAI,GAAG,kBAAkB,SAAS;AAAA,IACpE,KAAK;AACH,aAAO,EAAE,QAAQ,YAAY,IAAI,GAAG,kBAAkB,SAAS;AAAA,IACjE,KAAK;AACH,aAAO,EAAE,QAAQ,UAAU,IAAI,GAAG,kBAAkB,SAAS;AAAA,IAC/D,KAAK;AACH,aAAO,EAAE,QAAQ,YAAY,IAAI,GAAG,kBAAkB,SAAS;AAAA,IACjE,KAAK;AACH,aAAO,EAAE,QAAQ,YAAY,IAAI,GAAG,kBAAkB,SAAS;AAAA,EACnE;AACF;AA0BA,eAAsB,SACpB,UACA,KACA,OAAmB,CAAC,GACpB,OAAgB,CAAC,GACoB;AACrC,MAAI,SAAS,WAAW,GAAG;AACzB,UAAM,IAAI,gBAAgB,4BAA4B;AAAA,EACxD;AACA,MAAI,CAAC,IAAI,qBAAqB;AAC5B,UAAM,IAAI,gBAAgB,0CAA0C;AAAA,EACtE;AACA,QAAM,YAAY,KAAK,SAAS;AAChC,QAAM,MAAM,KAAK,QAAQ,MAAM,KAAK,IAAI;AACxC,QAAM,SAAS,KAAK;AACpB,QAAM,YAAY,IAAI;AAEtB,QAAM,OAAgB,KAAK,QAAQ;AACnC,QAAM,SAAS,WAAW,MAAM,QAAQ;AACxC,QAAM,gBAAgB,eAAe,UAAU,MAAM;AACrD,QAAM,QAAQ,KAAK,MAAM,MAAM,aAAa;AAG5C,QAAM,KAAK,IAAI;AACf,QAAM,WAAW,kBAAkB,QAAQ,IAAI,CAAC,CAAC;AACjD,QAAM,WAAW,oBAAoB,SAAS,IAAI,CAAC,CAAC;AACpD,MAAI,IAAI;AACN,QAAI,KAAK,gBAAgB,QAAW;AAClC,YAAM,MAAM,MAAM,GAAG,IAAI,QAAQ,EAAE,MAAM,MAAM,IAAI;AACnD,YAAM,QAAQ,WAAW,OAAO,GAAG;AACnC,UAAI,SAAS,KAAK,aAAa;AAC7B,eAAO;AAAA,UACL,IAAI,eAAe,0BAA0B;AAAA,YAC3C,UAAU;AAAA,YACV,aAAa,KAAK;AAAA,UACpB,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF;AACA,QAAI,KAAK,kBAAkB,QAAW;AACpC,YAAM,MAAM,MAAM,GAAG,IAAI,QAAQ,EAAE,MAAM,MAAM,IAAI;AACnD,YAAM,QAAQ,WAAW,OAAO,GAAG;AACnC,UAAI,SAAS,KAAK,eAAe;AAC/B,eAAO;AAAA,UACL,IAAI,eAAe,4BAA4B;AAAA,YAC7C,UAAU;AAAA,YACV,eAAe,KAAK;AAAA,UACtB,CAAC;AAAA,QACH;AAAA,MACF;AAAA,IACF;AAAA,EACF;AAEA,QAAM,aAAiF,CAAC;AAExF,MAAI,YAAY,CAAC,MAAM,SAAS,MAAM,QAAQ,EAAE,OAAO,OAAO;AAG9D,MAAI,KAAK,SAAS,KAAK,MAAM,SAAS,GAAG;AACvC,gBAAY,UAAU,OAAO,CAAC,MAAM,uBAAuB,IAAI,EAAE,QAAQ,CAAC;AAC1E,QAAI,UAAU,WAAW,GAAG;AAC1B,YAAM,IAAI;AAAA,QACR,kDAAkD,CAAC,GAAG,sBAAsB,EAAE,KAAK,IAAI,CAAC,YAAY,IAAI;AAAA,MAC1G;AAAA,IACF;AAAA,EACF;AACA,aAAW,CAAC,UAAU,GAAG,KAAK,UAAU,QAAQ,GAAG;AAEjD,QAAI,sBAAsB,IAAI,UAAU,GAAG,GAAG;AAC5C,cAAQ,OAAO,4BAA4B,EAAE,UAAU,IAAI,SAAS,CAAC;AACrE,iBAAW,KAAK,EAAE,UAAU,IAAI,UAAU,SAAS,wBAAwB,CAAC;AAC5E;AAAA,IACF;AACA,QAAI,KAAK,QAAQ,SAAS;AACxB,aAAO;AAAA,QACL,IAAI,cAAc,oBAAoB,EAAE,UAAU,IAAI,UAAU,OAAO,IAAI,MAAM,CAAC;AAAA,MACpF;AAAA,IACF;AACA,QAAI;AACF,YAAM,SAAS,MAAM,QAAQ,KAAK,UAAU,MAAM,KAAK,WAAW,QAAQ,GAAG;AAG7E,UAAI,CAAC,OAAO,OAAO,WAAW,EAAE,OAAO,OAAO,aAAa,OAAO,OAAO,UAAU,SAAS,IAAI;AAC9F,cAAM,EAAE,UAAU,IAAI,UAAU,QAAQ,KAAK,WAAW,OAAO,SAAS,gBAAgB;AAAA,MAC1F;AACA,cAAQ,OAAO,gBAAgB;AAAA,QAC7B,UAAU,IAAI;AAAA,QACd,OAAO,IAAI;AAAA,QACX;AAAA,QACA;AAAA,QACA,UAAU,OAAO;AAAA,QACjB,OAAO,KAAK;AAAA,QACZ,SAAS,KAAK;AAAA,QACd,OAAO,KAAK;AAAA,QACZ,UAAU,KAAK;AAAA,MACjB,CAAC;AACD,YAAM,YAAuB;AAAA,QAC3B,SAAS,OAAO,OAAO;AAAA,QACvB,UAAU,IAAI;AAAA,QACd,OAAO,OAAO,OAAO,SAAS,IAAI;AAAA,QAClC;AAAA,QACA,QAAQ;AAAA,UACN,OAAO,OAAO,OAAO;AAAA,UACrB,QAAQ,OAAO,OAAO;AAAA,UACtB,WAAW,OAAO,OAAO;AAAA,UACzB,YAAY,OAAO,OAAO;AAAA,QAC5B;AAAA,QACA,SAAS,IAAI,IAAI;AAAA,QACjB,UAAU,OAAO;AAAA,QACjB,kBAAkB,OAAO;AAAA,QACzB,YAAY,OAAO,OAAO;AAAA,QAC1B,WAAW,OAAO,OAAO;AAAA,MAC3B;AACA,YAAM,UAAU,gBAAgB,UAAU,QAAQ,UAAU,KAAK;AACjE,UAAI,KAAK,eAAe,UAAa,UAAU,KAAK,YAAY;AAC9D,YAAI,OAAO,KAAK,gBAAgB,UAAa,KAAK,kBAAkB,SAAY;AAC9E,gBAAM,mBAAmB,IAAI,UAAU,UAAU,SAAS,IAAI;AAAA,QAChE;AACA,eAAO;AAAA,UACL,IAAI,eAAe,yBAAyB;AAAA,YAC1C;AAAA,YACA,YAAY,KAAK;AAAA,YACjB,OAAO,UAAU;AAAA,YACjB,QAAQ,UAAU;AAAA,UACpB,CAAC;AAAA,QACH;AAAA,MACF;AAIA,UAAI,OAAO,KAAK,gBAAgB,UAAa,KAAK,kBAAkB,SAAY;AAC9E,cAAM,mBAAmB,IAAI,UAAU,UAAU,SAAS,IAAI;AAAA,MAChE;AAEA,UAAI,KAAK,YAAY,KAAK,QAAQ;AAChC,cAAM,MAAoB;AAAA,UACxB,GAAG,KAAK;AAAA,UACR,OAAO,UAAU;AAAA,UACjB,UAAU,UAAU;AAAA,UACpB,MAAM,UAAU;AAAA,UAChB,aAAa,UAAU,OAAO;AAAA,UAC9B,cAAc,UAAU,OAAO;AAAA,UAC/B,iBAAiB,UAAU,OAAO,aAAa;AAAA,UAC/C,kBAAkB,UAAU,OAAO,cAAc;AAAA,UACjD,WAAW,UAAU;AAAA,UACrB;AAAA,UACA,QAAQ,SAAS,IAAI,CAAC;AAAA,QACxB;AACA,aAAK,SAAS,GAAG,EAAE,MAAM,CAAC,MAAe;AACvC,kBAAQ,OAAO,sBAAsB,EAAE,SAAS,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC,EAAE,CAAC;AAAA,QAC9F,CAAC;AAAA,MACH;AACA,aAAO,EAAE,MAAM,WAAW,OAAO,KAAK;AAAA,IACxC,SAAS,GAAG;AACV,UAAI,aAAa,gBAAgB,EAAE,SAAS,cAAc;AACxD,eAAO;AAAA,UACL,IAAI,cAAc,oBAAoB,EAAE,UAAU,IAAI,UAAU,OAAO,IAAI,MAAM,CAAC;AAAA,QACpF;AAAA,MACF;AACA,UAAI,gBAAgB,CAAC,GAAG;AACtB,mBAAW,KAAK,EAAE,UAAU,EAAE,UAAU,QAAQ,EAAE,QAAQ,SAAS,EAAE,QAAQ,CAAC;AAC9E,YAAI,EAAE,WAAW,OAAO,aAAa,UAAU,SAAS,GAAG;AACzD,iBAAO;AAAA,YACL,IAAI,eAAe,uBAAuB,EAAE,QAAQ,IAAI,EAAE,UAAU,WAAW,CAAC;AAAA,UAClF;AAAA,QACF;AACA,gBAAQ,OAAO,kBAAkB,EAAE,UAAU,IAAI,UAAU,QAAQ,EAAE,OAAO,CAAC;AAC7E;AAAA,MACF;AACA,iBAAW,KAAK,EAAE,UAAU,IAAI,UAAU,SAAS,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC,EAAE,CAAC;AAAA,IACjG;AAAA,EACF;AAEA,SAAO;AAAA,IACL,IAAI,cAAc,4BAA4B,EAAE,UAAU,YAAY,MAAM,cAAc,CAAC;AAAA,EAC7F;AACF;AAiDA,gBAAuB,iBACrB,UACA,KACA,OAAwC,CAAC,GACG;AAC5C,MAAI,SAAS,WAAW,GAAG;AACzB,UAAM,IAAI,gBAAgB,4BAA4B;AAAA,EACxD;AACA,MAAI,CAAC,IAAI,qBAAqB;AAC5B,UAAM,IAAI,gBAAgB,0CAA0C;AAAA,EACtE;AAEA,QAAM,OAAgB,KAAK,QAAQ,CAAC;AACpC,QAAM,YAAY,KAAK,SAAS;AAChC,QAAM,MAAM,KAAK,QAAQ,MAAM,KAAK,IAAI;AACxC,QAAM,SAAS,KAAK;AACpB,QAAM,YAAY,IAAI;AAEtB,QAAM,OAAgB,KAAK,QAAQ;AACnC,QAAM,SAAS,WAAW,MAAM,QAAQ;AACxC,QAAM,gBAAgB,eAAe,UAAU,MAAM;AACrD,QAAM,QAAQ,KAAK,MAAM,MAAM,aAAa;AAC5C,QAAM,YACJ,MAAM,QAAQ,aAAa,UAAU,CAAC,IAAI,gBAAgB,MAAM,UAAU,aAAa,cACnF,MAAM,WACN,MAAM;AAIZ,MAAI,UAAU,aAAa,aAAa;AACtC,UAAM,SAAS,MAAM,SAAS,UAAU,KAAK,MAAM,IAAI;AACvD,QAAI,OAAO,UAAU,QAAQ,OAAO,SAAS,MAAM;AACjD,YAAM,IAAI,cAAc,4BAA4B,EAAE,OAAO,OAAO,MAAM,CAAC;AAAA,IAC7E;AACA,UAAM,OAAO,KAAK;AAClB,WAAO,OAAO;AAAA,EAChB;AAGA,MAAI,sBAAsB,UAAU,UAAU,GAAG,GAAG;AAClD,YAAQ,OAAO,4BAA4B,EAAE,UAAU,UAAU,SAAS,CAAC;AAE3E,UAAM,SAAS,MAAM,SAAS,UAAU,KAAK,MAAM,IAAI;AACvD,QAAI,OAAO,UAAU,QAAQ,OAAO,SAAS,MAAM;AACjD,YAAM,IAAI,cAAc,4BAA4B,EAAE,OAAO,OAAO,MAAM,CAAC;AAAA,IAC7E;AACA,UAAM,OAAO,KAAK;AAClB,WAAO,OAAO;AAAA,EAChB;AAEA,QAAM,MAAM,sBAAsB,UAAU,OAAO,UAAU,MAAM,KAAK,IAAI;AAE5E,QAAM,oBAAoB,iBAAiB,IAAI;AAC/C,MAAI,kBAAmB,KAAI,QAAQ,iBAAiB,IAAI;AAExD,MAAI;AACJ,MAAI;AACF,eAAW,MAAM,UAAU,IAAI,KAAK;AAAA,MAClC,QAAQ;AAAA,MACR,SAAS,IAAI;AAAA,MACb,MAAM,IAAI;AAAA;AAAA;AAAA,MAGV,QAAQ,KAAK,UAAU,YAAY,QAAQ,GAAM;AAAA,IACnD,CAAC;AAAA,EACH,SAAS,GAAG;AACV,QAAI,aAAa,gBAAgB,EAAE,SAAS,cAAc;AACxD,YAAM,IAAI,cAAc,oBAAoB;AAAA,QAC1C,UAAU,UAAU;AAAA,QACpB,OAAO,UAAU;AAAA,MACnB,CAAC;AAAA,IACH;AACA,UAAM,IAAI,cAAc,2BAA2B;AAAA,MACjD,SAAS,aAAa,QAAQ,EAAE,UAAU,OAAO,CAAC;AAAA,IACpD,CAAC;AAAA,EACH;AAEA,MAAI,CAAC,SAAS,IAAI;AAChB,UAAM,OAAO,MAAM,SAAS,KAAK,EAAE,MAAM,MAAM,EAAE;AACjD,UAAM,YAAY,sBAAsB,SAAS,MAAM;AACvD,QAAI,aAAa,SAAS,WAAW,KAAK;AACxC,8BAAwB,UAAU,UAAU,GAAG;AAAA,IACjD;AAEA,UAAM,SAAS,MAAM,SAAS,UAAU,KAAK,MAAM,IAAI;AACvD,QAAI,OAAO,UAAU,QAAQ,OAAO,SAAS,MAAM;AACjD,YAAM,IAAI,cAAc,4BAA4B;AAAA,QAClD,aAAa,GAAG,UAAU,QAAQ,IAAI,OAAO,SAAS,MAAM,CAAC,KAAK,KAAK,MAAM,GAAG,GAAG,CAAC;AAAA,QACpF,OAAO,OAAO;AAAA,MAChB,CAAC;AAAA,IACH;AACA,UAAM,OAAO,KAAK;AAClB,WAAO,OAAO;AAAA,EAChB;AAEA,MAAI,CAAC,SAAS,MAAM;AAClB,UAAM,IAAI,cAAc,oCAAoC;AAAA,MAC1D,UAAU,UAAU;AAAA,IACtB,CAAC;AAAA,EACH;AAGA,QAAM,UAAU,IAAI,YAAY;AAChC,MAAI,kBAAkB;AACtB,MAAI,cAAc;AAClB,MAAI,eAAe;AACnB,MAAI,YAAY;AAChB,MAAI,aAAa;AACjB,MAAI;AACJ,QAAM,mBAAuC,SAAS,QAAQ,IAAI,mBAAmB,KAAK;AAE1F,QAAM,SAAS,SAAS,KAAK,UAAU;AACvC,MAAI,SAAS;AAEb,MAAI;AACF,WAAO,MAAM;AACX,YAAM,EAAE,MAAM,MAAM,IAAI,MAAM,OAAO,KAAK;AAC1C,UAAI,KAAM;AACV,gBAAU,QAAQ,OAAO,OAAO,EAAE,QAAQ,KAAK,CAAC;AAGhD,YAAM,QAAQ,OAAO,MAAM,IAAI;AAE/B,eAAS,MAAM,IAAI,KAAK;AAExB,iBAAW,QAAQ,OAAO;AACxB,YAAI,CAAC,KAAK,WAAW,QAAQ,EAAG;AAChC,cAAM,OAAO,KAAK,MAAM,CAAC,EAAE,KAAK;AAChC,YAAI,SAAS,SAAU;AACvB,YAAI;AACJ,YAAI;AACF,kBAAQ,KAAK,MAAM,IAAI;AAAA,QACzB,QAAQ;AACN;AAAA,QACF;AAEA,gBAAQ,MAAM,MAAM;AAAA,UAClB,KAAK;AACH,0BAAc,MAAM,SAAS,OAAO,gBAAgB;AACpD,wBAAY,MAAM,SAAS,OAAO,2BAA2B;AAC7D,yBAAa,MAAM,SAAS,OAAO,+BAA+B;AAClE,wBAAY,MAAM,SAAS;AAC3B;AAAA,UACF,KAAK;AACH,gBAAI,MAAM,OAAO,SAAS,gBAAgB,OAAO,MAAM,MAAM,SAAS,UAAU;AAC9E,iCAAmB,MAAM,MAAM;AAC/B,oBAAM,MAAM,MAAM;AAAA,YACpB;AACA;AAAA,UACF,KAAK;AACH,2BAAe,MAAM,OAAO,iBAAiB;AAC7C;AAAA,UACF;AACE;AAAA,QACJ;AAAA,MACF;AAAA,IACF;AAAA,EACF,UAAE;AACA,WAAO,YAAY;AAAA,EACrB;AAEA,wBAAsB,UAAU,QAAQ;AACxC,UAAQ,OAAO,wBAAwB;AAAA,IACrC,UAAU,UAAU;AAAA,IACpB,OAAO,UAAU;AAAA,IACjB;AAAA,IACA;AAAA,IACA,OAAO,KAAK;AAAA,IACZ,SAAS,KAAK;AAAA,IACd,OAAO,KAAK;AAAA,IACZ,UAAU,KAAK;AAAA,EACjB,CAAC;AAED,SAAO;AAAA,IACL,SAAS;AAAA,IACT,UAAU,UAAU;AAAA,IACpB,OAAO,aAAa,UAAU;AAAA,IAC9B;AAAA,IACA,QAAQ,EAAE,OAAO,aAAa,QAAQ,cAAc,WAAW,WAAW;AAAA,IAC1E,SAAS,IAAI,IAAI;AAAA,IACjB,UAAU;AAAA,IACV;AAAA,EACF;AACF;AA4BO,SAAS,gBAAgB,UAAkB,SAA4B;AAC5E,MAAI,QAAQ,WAAW,EAAG,QAAO;AAEjC,QAAM,SAAS;AACf,QAAM,iBAAiB,SAAS,MAAM,KAAK,EAAE,OAAO,CAAC,MAAM,EAAE,SAAS,CAAC;AAEvE,MAAI,eAAe,SAAS,OAAQ,QAAO;AAG3C,QAAM,eAAe,oBAAI,IAAY;AACrC,aAAW,UAAU,SAAS;AAC5B,UAAM,SAAS,OAAO,MAAM,KAAK,EAAE,OAAO,CAAC,MAAM,EAAE,SAAS,CAAC;AAC7D,aAAS,IAAI,GAAG,KAAK,OAAO,SAAS,QAAQ,KAAK;AAChD,YAAM,QAAQ,OAAO,MAAM,GAAG,IAAI,MAAM,EAAE,KAAK,GAAG;AAClD,mBAAa,IAAI,KAAK;AAAA,IACxB;AAAA,EACF;AAEA,MAAI,aAAa,SAAS,EAAG,QAAO;AAGpC,WAAS,IAAI,GAAG,KAAK,eAAe,SAAS,QAAQ,KAAK;AACxD,UAAM,QAAQ,eAAe,MAAM,GAAG,IAAI,MAAM,EAAE,KAAK,GAAG;AAC1D,QAAI,aAAa,IAAI,KAAK,EAAG,QAAO;AAAA,EACtC;AAEA,SAAO;AACT;","names":[]}

package/package.json

@@ -1,15 +1,13 @@
 {
   "name": "@latimer-woods-tech/llm",
-  "version": "0.4.1",
+  "version": "0.4.3",
   "private": false,
   "repository": {
     "type": "git",
     "url": "git+https://github.com/Latimer-Woods-Tech/Factory.git",
     "directory": "packages/llm"
   },
-  "publishConfig": {
-    "registry": "https://npm.pkg.github.com"
-  },
+  "publishConfig": {},
   "main": "./dist/index.mjs",
   "module": "./dist/index.mjs",
   "types": "./dist/index.d.mts",