@latimer-woods-tech/llm 0.6.1 → 0.8.1

package/CHANGELOG.md

@@ -1,5 +1,62 @@
 # Changelog
 
+## 0.8.1 — 2026-09-02
+
+### Fixed — Qwen 27B visible-output contract
+
+- Workbench requests now send OpenAI-compatible `reasoning_effort: "none"`.
+  Ollama otherwise enables Qwen thinking by default, so a short request could
+  return HTTP 200 while exhausting `max_tokens` in hidden reasoning and leaving
+  `message.content` empty. The existing empty-content guard remains the
+  fail-closed backstop.
+
+## 0.8.0 — 2026-08-23
+
+### Added (#4204 — the counter half of "LLM chain DEGRADED is unreachable without reading logs")
+
+- `LLMResult.degraded: boolean` — true when a fallback provider served instead
+  of the tier's intended primary. Computed from `legIndex > 0` at the point the
+  route is walked (`[route.primary, route.fallback]`), so it is a direct read
+  of "servedBy != intended" rather than a proxy. `attempts > 1` already
+  existed but conflates a same-provider retry with an actual fallback to a
+  different provider — `degraded` is the distinct signal.
+- Threaded into `LLMRecordRow.degraded` so every metering consumer (starting
+  with `@latimer-woods-tech/llm-meter`) gets it automatically via the existing
+  `onRecord` callback — no new dependency inside this package, no Sentry/
+  PostHog call added here.
+- `completionStream()`'s direct-Anthropic path sets `degraded` from whether
+  the resolved `streamLeg` is `route.fallback` (the grok-no-key special case);
+  every other stream branch already delegates to `complete()` and inherits its
+  `degraded` value.
+- No routing/fallback decision logic changed — purely additive observability.
+
+## 0.7.0 — 2026-08-23
+
+### Fixed (metering seam)
+
+- `LLMOptions` carries `project`/`actor`/`runId`/`workload` at the TOP level and
+  again inside `ledger`. Both shapes typecheck, but only `ledger` was read by the
+  `onRecord` metering callback, so choosing the wrong one silently produced logs
+  and no ledger row. Measured across the estate on 2026-08-23: **8 of 8** call
+  sites supplying metering context used the top-level shape and **0** used
+  `ledger`. A 0% adoption rate on one of two identical-looking APIs is a design
+  defect, not eight independent mistakes.
+- New exported `resolveLedgerContext(opts)` reconciles the two: an explicit
+  `ledger` always wins, otherwise a context is derived from top-level
+  `project` + `actor`. **Both** are required — a partial context would write a
+  row that cannot be attributed, which is worse than no row.
+- No behaviour change for callers already passing `ledger`.
+
+### Added (cost observability)
+
+- `llm.complete` / `llm.completionStream` log lines now carry **`tierExplicit`**.
+  `false` means no caller named a tier, so the call took the package default —
+  `balanced`, i.e. claude-sonnet-5 at $3/$15 per 1M, the second most expensive
+  of five tiers. Selecting the expensive route by silence is invisible at the
+  call site: `apps/video-studio/scripts/generate-script.mjs` made SIX untiered
+  `complete()` calls while `CLAUDE.md` described that step as Haiku ($1/$5).
+  A defaulted tier is now greppable in logs. No behaviour change.
+
 ## 0.6.1 — 2026-08-18
 
 ### Fixed — ships the routing and sampling fixes from #3946, which never reached npm

package/README.md

@@ -25,7 +25,8 @@ falls through to Grok, never directly to Anthropic.
 
 Set `LLM_LOCAL_WORKBENCH=true` to make `qwen3.6:27b` the `workbench` primary. The 27B model
 retains the OpenAI-compatible `custom-local-gpu/v1/chat/completions` route and normalized tool
-protocol, with DeepSeek as its independent fallback. Both routes require `GPU_LLM_API_TOKEN`;
+protocol, sends `reasoning_effort: "none"` so bounded jobs cannot return a blank answer after
+spending their budget on hidden reasoning, and retains DeepSeek as its independent fallback. Both routes require `GPU_LLM_API_TOKEN`;
 the Cloudflare Access client id and secret are forwarded when supplied.
 
 ## Usage

package/dist/index.d.mts

@@ -255,6 +255,8 @@ interface LLMResult {
      * (non-empty) when `stopReason === 'tool_use'`.
      */
     toolCalls?: LLMToolCall[];
+    /** True when a fallback provider served instead of the tier's intended primary — see #4204. */
+    degraded: boolean;
 }
 /**
  * Environment bindings required by {@link complete}.
@@ -351,6 +353,8 @@ interface LLMRecordRow extends LLMRecordContext {
     model: string;
     provider: LLMProvider;
     tier: LLMTier;
+    /** True when a fallback provider served instead of the tier's intended primary — see #4204. */
+    degraded: boolean;
     inputTokens: number;
     outputTokens: number;
     cacheReadTokens: number;
@@ -359,6 +363,29 @@ interface LLMRecordRow extends LLMRecordContext {
     costUsd: number;
     yyyyMm: string;
 }
+/**
+ * Resolve the metering context for a call.
+ *
+ * `LLMOptions` carries the same four field names in TWO places: top-level
+ * `project`/`actor`/`runId`/`workload` (documented for logs and cost-policy
+ * call sites) and the nested {@link LLMOptions.ledger}, which is the ONLY one
+ * the `onRecord` metering callback reads. Both shapes typecheck, so picking
+ * the wrong one is silent — you get logging and no ledger row.
+ *
+ * Every caller picked the wrong one. Measured across the estate 2026-08-23:
+ * **8 of 8** call sites that supply metering context set it at the top level
+ * (`admin-studio` test-analyst + ai, `daily-brief` wisdom + insights,
+ * `inbound-oracle`, `supervisor` reflect, `lead-gen`, `linkedin-publisher`)
+ * and **0** use `ledger`. A 0% adoption rate on one of two identical-looking
+ * APIs is a design defect, not eight independent mistakes — so the fix belongs
+ * here rather than in eight call sites that already read as correct.
+ *
+ * Explicit `ledger` always wins; the top-level fields are a fallback. A call
+ * needs BOTH `project` and `actor` to be meterable, because those two are the
+ * required columns on the ledger row — a partial context would write a row
+ * that cannot be attributed, which is worse than no row.
+ */
+declare function resolveLedgerContext(opts: LLMOptions): LLMRecordContext | undefined;
 /**
  * Optional dependencies for {@link complete}.
  */
@@ -383,7 +410,7 @@ declare const MODELS: {
         readonly smart: "gemini-2.5-flash";
     };
     readonly groq: {
-        readonly verifier: "llama-3.3-70b-versatile";
+        readonly verifier: "openai/gpt-oss-120b";
     };
     readonly grok: {
         readonly fast: "grok-4.3";
@@ -504,4 +531,4 @@ declare function completionStream(messages: LLMMessage[], env: LLMEnv, opts?: LL
  */
 declare function assertGrounding(response: string, sources: string[]): boolean;
 
-export { type AiBinding, BASE_BACKOFF_MS, type CostKvStore, DEFAULT_EMBEDDING_MODEL, type EmbedResult, type EmbeddingModel, type LLMContentBlock, type LLMDeps, type LLMEnv, type LLMMessage, type LLMOptions, type LLMProvider, type LLMRecordContext, type LLMRecordRow, type LLMResult, type LLMTier, type LLMTool, type LLMToolCall, LOCAL_EMBEDDING_MODEL, type LocalEmbedEnv, MODELS, MODEL_PRICE_PER_1M, PROVIDER_COOLDOWN_MS, assertGrounding, clearGcpTokenCache, clearProviderCooldown, complete, completionStream, embed, embedLocal, isProviderCoolingDown, markProviderCoolingDown, mintGcpAccessToken, serviceAccountProjectId };
+export { type AiBinding, BASE_BACKOFF_MS, type CostKvStore, DEFAULT_EMBEDDING_MODEL, type EmbedResult, type EmbeddingModel, type LLMContentBlock, type LLMDeps, type LLMEnv, type LLMMessage, type LLMOptions, type LLMProvider, type LLMRecordContext, type LLMRecordRow, type LLMResult, type LLMTier, type LLMTool, type LLMToolCall, LOCAL_EMBEDDING_MODEL, type LocalEmbedEnv, MODELS, MODEL_PRICE_PER_1M, PROVIDER_COOLDOWN_MS, assertGrounding, clearGcpTokenCache, clearProviderCooldown, complete, completionStream, embed, embedLocal, isProviderCoolingDown, markProviderCoolingDown, mintGcpAccessToken, resolveLedgerContext, serviceAccountProjectId };

package/dist/index.mjs

@@ -199,6 +199,14 @@ function systemText(opts, messages) {
   const c = messages.find((m) => m.role === "system")?.content;
   return c === void 0 ? void 0 : contentToText(c);
 }
+function resolveLedgerContext(opts) {
+  if (opts.ledger) return opts.ledger;
+  if (!opts.project || !opts.actor) return void 0;
+  const ctx = { project: opts.project, actor: opts.actor };
+  if (opts.runId !== void 0) ctx.runId = opts.runId;
+  if (opts.workload !== void 0) ctx.workload = opts.workload;
+  return ctx;
+}
 var MODELS = {
   anthropic: {
     // `claude-haiku-4-20250514` was never a model Anthropic served — it 404s
@@ -232,12 +240,13 @@ var MODELS = {
     smart: "gemini-2.5-flash"
   },
   groq: {
-    // `llama-4-maverick` was NOT a model Groq serves — every `verifier` call
-    // 404'd (the tier has no fallback, so it was hard-broken), and it was also
-    // the `workbench` fallback. Groq's llama-4 offering is `scout`, which is
-    // blocked at our org level; `llama-3.3-70b-versatile` is served and
-    // unblocked (verified against /v1/models and a live completion).
-    verifier: "llama-3.3-70b-versatile"
+    // Groq deprecated `llama-3.3-70b-versatile` on 2026-08-16 — every
+    // `verifier` call 404'd `model_not_found` from then on (the tier has no
+    // fallback, so it was hard-broken; second time this tier has died to a
+    // model retirement). `openai/gpt-oss-120b` is Groq's recommended
+    // migration target and is served to our org (verified against /v1/models
+    // and a live completion, 2026-08-23).
+    verifier: "openai/gpt-oss-120b"
   },
   grok: {
     fast: "grok-4.3"
@@ -313,8 +322,11 @@ var MODEL_PRICE_PER_1M = {
   "gemini-2.5-flash": { input: 0.3, output: 2.5, cacheRead: 0.075, cacheWrite: 0.3 },
   // Gemini 2.5 Pro — retained for historical ledger rows (was the routed model).
   "gemini-2.5-pro": { input: 1.25, output: 10, cacheRead: 0.31, cacheWrite: 4.5 },
-  // Groq Llama 3.3 70B Versatile (`verifier` tier). Groq has no prompt caching,
-  // so cache rates are 0.00 — same convention as grok-4.3 below.
+  // Groq GPT-OSS 120B (`verifier` tier). Groq bills cached input at $0.075/1M
+  // with no separate cache-write charge (implicit caching), so cacheWrite is 0.
+  "openai/gpt-oss-120b": { input: 0.15, output: 0.6, cacheRead: 0.075, cacheWrite: 0 },
+  // Groq Llama 3.3 70B Versatile — retained for historical ledger rows
+  // (routed `verifier` until Groq retired the model on 2026-08-16).
   "llama-3.3-70b-versatile": { input: 0.59, output: 0.79, cacheRead: 0, cacheWrite: 0 },
   // Grok 4.3
   "grok-4.3": { input: 1.25, output: 2.5, cacheRead: 0, cacheWrite: 0 },
@@ -656,6 +668,10 @@ function buildLocalRequest(model, messages, opts, env) {
     headers,
     body: JSON.stringify({
       model,
+      // Ollama enables Qwen thinking by default.  Without this OpenAI-
+      // compatible control, short bounded calls can spend every token in the
+      // hidden `reasoning` field and return HTTP 200 with empty visible text.
+      reasoning_effort: "none",
       max_tokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,
       temperature: opts.temperature ?? DEFAULT_TEMPERATURE,
       messages: toOpenAiMessages(messages, sys),
@@ -1015,6 +1031,12 @@ async function complete(messages, env, opts = {}, deps = {}) {
         throw { provider: leg.provider, status: 200, retryable: false, message: "empty content" };
       }
       logger?.info?.("llm.complete", {
+        // `false` means NO caller named a tier, so this call took the
+        // package default (`balanced` — claude-sonnet-5, the second most
+        // expensive route). Emitted so a defaulted tier is greppable rather
+        // than invisible: an expensive route chosen by silence is the thing
+        // that made the video pipeline bill Sonnet while its docs said Haiku.
+        tierExplicit: opts.tier !== void 0,
         provider: leg.provider,
         model: leg.model,
         tier,
@@ -1025,6 +1047,7 @@ async function complete(messages, env, opts = {}, deps = {}) {
         actor: opts.actor,
         workload: opts.workload
       });
+      const degraded = legIndex > 0;
       const llmResult = {
         content: result.parsed.content,
         provider: leg.provider,
@@ -1040,7 +1063,8 @@ async function complete(messages, env, opts = {}, deps = {}) {
         attempts: result.attempts,
         gatewayRequestId: result.gatewayRequestId,
         stopReason: result.parsed.stopReason,
-        toolCalls: result.parsed.toolCalls
+        toolCalls: result.parsed.toolCalls,
+        degraded
       };
       const costUsd = estimateCostUsd(llmResult.tokens, llmResult.model);
       if (opts.maxCostUsd !== void 0 && costUsd > opts.maxCostUsd) {
@@ -1059,12 +1083,14 @@ async function complete(messages, env, opts = {}, deps = {}) {
       if (kv && (opts.dailyCapUsd !== void 0 || opts.monthlyCapUsd !== void 0)) {
         await recordOrgCostUsage(kv, todayKey, monthKey, costUsd, opts);
       }
-      if (deps.onRecord && opts.ledger) {
+      const ledgerCtx = resolveLedgerContext(opts);
+      if (deps.onRecord && ledgerCtx) {
         const row = {
-          ...opts.ledger,
+          ...ledgerCtx,
           model: llmResult.model,
           provider: llmResult.provider,
           tier: llmResult.tier,
+          degraded: llmResult.degraded,
           inputTokens: llmResult.tokens.input,
           outputTokens: llmResult.tokens.output,
           cacheReadTokens: llmResult.tokens.cacheRead ?? 0,
@@ -1244,6 +1270,12 @@ async function* completionStream(messages, env, opts = {}) {
   }
   clearProviderCooldown(streamLeg.provider);
   logger?.info?.("llm.completionStream", {
+    // `false` means NO caller named a tier, so this call took the
+    // package default (`balanced` — claude-sonnet-5, the second most
+    // expensive route). Emitted so a defaulted tier is greppable rather
+    // than invisible: an expensive route chosen by silence is the thing
+    // that made the video pipeline bill Sonnet while its docs said Haiku.
+    tierExplicit: opts.tier !== void 0,
     provider: streamLeg.provider,
     model: streamLeg.model,
     tier,
@@ -1268,7 +1300,10 @@ async function* completionStream(messages, env, opts = {}) {
     attempts: 1,
     gatewayRequestId,
     stopReason: streamStopReason,
-    toolCalls: toolCalls.length > 0 ? toolCalls : void 0
+    toolCalls: toolCalls.length > 0 ? toolCalls : void 0,
+    // streamLeg is only ever route.fallback in the grok-no-key special case above —
+    // same "servedBy != intended" signal as the non-streaming path, see #4204.
+    degraded: streamLeg !== route.primary
   };
 }
 function assertGrounding(response, sources) {
@@ -1308,6 +1343,7 @@ export {
   isProviderCoolingDown,
   markProviderCoolingDown,
   mintGcpAccessToken,
+  resolveLedgerContext,
   serviceAccountProjectId
 };
 //# sourceMappingURL=index.mjs.map

package/dist/index.mjs.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/index.ts","../src/gcp-token.ts","../src/embed.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';\nimport { fetchAdcAccessToken, mintGcpAccessToken, serviceAccountProjectId } from './gcp-token.js';\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 Flash 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' | 'local';\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   * Cost optimization: when true, `fast`-tier calls route to the self-hosted GPU\n   * (qwen3:8b via the `custom-local-fast-gpu` AI Gateway provider) FIRST, with the normal\n   * fast route (Grok→Haiku) as automatic fallback on any error. Off by default →\n   * fully dormant. Requires {@link LLMEnv.GPU_LLM_API_TOKEN}.\n   */\n  LLM_LOCAL_FIRST?: boolean;\n  /**\n   * Agentic-work opt-in: route the `workbench` tier to local Qwen 27B first,\n   * with DeepSeek as the independent cloud fallback.\n   */\n  LLM_LOCAL_WORKBENCH?: boolean;\n  /** Bearer for the local GPU proxy (`GPU_LLM_API_TOKEN` in Secret Manager). Required to enable local. */\n  GPU_LLM_API_TOKEN?: string;\n  /** Cloudflare Access service-token id for the local GPU app; forwarded by the gateway to the origin. */\n  GPU_LLM_ACCESS_CLIENT_ID?: string;\n  /** Cloudflare Access service-token secret for the local GPU app. */\n  GPU_LLM_ACCESS_CLIENT_SECRET?: string;\n  /**\n   * GCP service-account JSON key (base64-encoded or raw) with `roles/aiplatform.user`.\n   * The credential for the `gemini` leg in environments WITHOUT a metadata server\n   * (Cloudflare Workers): the package exchanges it for a fresh access token per\n   * isolate and caches it (see `./gcp-token.ts`).\n   *\n   * Optional. When neither this nor {@link LLMEnv.VERTEX_ACCESS_TOKEN} is set, the\n   * leg falls back to Application Default Credentials — the GCP metadata server's\n   * ambient service-account token (keyless; the default on Cloud Run / GCE). Off\n   * GCP with no static credential, the metadata probe fails and the chain falls\n   * through to its next leg.\n   */\n  GCP_SA_KEY?: string;\n  /**\n   * Pre-minted Google Cloud access token with `aiplatform.endpoints.predict`.\n   * Explicit override, used only when {@link LLMEnv.GCP_SA_KEY} is absent and\n   * before ADC is attempted. Valid for ~1 hour from minting, so it cannot be a\n   * durable Worker secret.\n   */\n  VERTEX_ACCESS_TOKEN?: string;\n  /** Vertex project. Defaults to the `GCP_SA_KEY`'s own `project_id`. */\n  VERTEX_PROJECT?: string;\n  /** Vertex location. Defaults to `us-central1`. */\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    // `claude-haiku-4-20250514` was never a model Anthropic served — it 404s\n    // (verified live against /v1/messages). Date suffixes are never appended to\n    // an alias; the alias is `claude-haiku-4-5`, which resolves server-side to\n    // `claude-haiku-4-5-20251001`. This is the `fast` tier's FALLBACK, so the\n    // 404 stayed invisible for as long as the Grok primary held.\n    fast: 'claude-haiku-4-5',\n    // Current-generation Sonnet/Opus (2026-08 refresh). The previous pins —\n    // `claude-sonnet-4-6` / `claude-opus-4-7` — were a generation behind, and\n    // the opus-4-7 leg was HARD-BROKEN: Claude 4.7+ rejects `temperature`\n    // (400 \"`temperature` is deprecated for this model\", verified live\n    // 2026-08-14) and buildAnthropicRequest always sent one, so every `smart`\n    // call 400'd on its Anthropic primary and silently served Gemini Flash.\n    // Sonnet 5 / Opus 5 cost the same list price or less than the models they\n    // replace ($3/$15, $5/$25 per 1M). The sampling/thinking request-shape\n    // differences these models introduce are handled in buildAnthropicRequest.\n    balanced: 'claude-sonnet-5',\n    smart: 'claude-opus-5',\n  },\n  gemini: {\n    // `gemini-2.5-flash`, not `-pro`: the leg's job here is a fast, reliable,\n    // JSON-returning fallback. Gemini 2.5 *Pro* mandates a thinking budget of\n    // ≥128 tokens that is drawn from `maxOutputTokens` and CANNOT be disabled\n    // (thinkingBudget=0 is rejected). On the render-runner's large judge /\n    // generation prompts — and any low-`maxTokens` call (headline uses 40) —\n    // the thinking phase exhausts the whole budget and Vertex returns 200 with\n    // an empty candidate (finishReason MAX_TOKENS, no text), which the router\n    // treats as a failed leg. Flash supports `thinkingBudget: 0` (set in\n    // buildGeminiRequest), so text is always emitted. Both are Vertex-served.\n    smart: 'gemini-2.5-flash',\n  },\n  groq: {\n    // `llama-4-maverick` was NOT a model Groq serves — every `verifier` call\n    // 404'd (the tier has no fallback, so it was hard-broken), and it was also\n    // the `workbench` fallback. Groq's llama-4 offering is `scout`, which is\n    // blocked at our org level; `llama-3.3-70b-versatile` is served and\n    // unblocked (verified against /v1/models and a live completion).\n    verifier: 'llama-3.3-70b-versatile',\n  },\n  grok: {\n    fast: 'grok-4.3',\n  },\n  deepseek: {\n    workbench: 'deepseek-chat',\n  },\n  local: {\n    // Self-hosted qwen3:8b on the dedicated fast GPU, reached through the\n    // `custom-local-fast-gpu` AI Gateway provider's native Ollama chat route.\n    fast: 'qwen3:8b',\n    workbench: 'qwen3.6:27b',\n  },\n} as const;\n\nconst DEFAULT_MAX_TOKENS = 1024;\nconst DEFAULT_TEMPERATURE = 0.7;\n/** Vertex region used when the caller does not pin one. */\nconst DEFAULT_VERTEX_LOCATION = 'us-central1';\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 4.5 — `claude-haiku-4-5` is the routed alias; the dated id\n  // is what the API echoes back in `response.model`, so both must price.\n  // (Rates corrected 2026-08: the previous $0.80/$4.00 rows UNDERSTATED the\n  // actual $1/$5 list price.)\n  'claude-haiku-4-5': { input: 1.00, output: 5.00, cacheRead: 0.10, cacheWrite: 1.25 },\n  'claude-haiku-4-5-20251001': { input: 1.00, output: 5.00, cacheRead: 0.10, cacheWrite: 1.25 },\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 Sonnet 5 — the routed `balanced` primary. List $3/$15; intro\n  // pricing ($2/$10) runs through 2026-08-31 — priced at list here, the\n  // conservative upper bound (table convention).\n  'claude-sonnet-5': { input: 3.00, output: 15.00, cacheRead: 0.30, cacheWrite: 3.75 },\n  // Anthropic Opus 4 (dated id) — genuinely the $15/$75 era.\n  'claude-opus-4-20250514': { input: 15.00, output: 75.00, cacheRead: 1.50, cacheWrite: 18.75 },\n  // Opus 4.7 was NEVER $15/$75 — it launched at $5/$25. The old row copied the\n  // Opus 4 rate and overstated every smart-tier ledger entry 3x.\n  'claude-opus-4-7': { input: 5.00, output: 25.00, cacheRead: 0.50, cacheWrite: 6.25 },\n  // Anthropic Opus 5 — the routed `smart` primary. Same $5/$25 as Opus 4.8/4.7.\n  'claude-opus-5': { input: 5.00, output: 25.00, cacheRead: 0.50, cacheWrite: 6.25 },\n  // Gemini 2.5 Flash — the routed `smart`/long-context model (JSON fallback leg).\n  'gemini-2.5-flash': { input: 0.30, output: 2.50, cacheRead: 0.075, cacheWrite: 0.30 },\n  // Gemini 2.5 Pro — retained for historical ledger rows (was the routed model).\n  'gemini-2.5-pro': { input: 1.25, output: 10.00, cacheRead: 0.31, cacheWrite: 4.50 },\n  // Groq Llama 3.3 70B Versatile (`verifier` tier). Groq has no prompt caching,\n  // so cache rates are 0.00 — same convention as grok-4.3 below.\n  'llama-3.3-70b-versatile': { input: 0.59, output: 0.79, cacheRead: 0.00, cacheWrite: 0.00 },\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  // Self-hosted qwen3 on the GPU box — zero marginal cost (electricity aside).\n  'qwen3:8b': { input: 0.0, output: 0.0, cacheRead: 0.0, cacheWrite: 0.0 },\n  'qwen3.6:27b': { input: 0.0, output: 0.0, cacheRead: 0.0, cacheWrite: 0.0 },\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\n/**\n * Claude 4.7-and-later models reject the classic sampling params: sending\n * `temperature` (or `top_p`/`top_k`) returns 400 \"`temperature` is deprecated\n * for this model\" (verified live against /v1/messages, 2026-08-14). Models\n * matching this pattern must not receive a `temperature` field at all — this\n * is exactly what hard-broke the `smart` tier while it pinned opus-4-7.\n */\nconst ANTHROPIC_SAMPLING_REMOVED = /^claude-(opus-4-[78]|opus-5|sonnet-5|fable-5|mythos-5)/;\n\n/**\n * Models that run ADAPTIVE THINKING by default when the `thinking` field is\n * omitted. Thinking tokens draw from `max_tokens`, and this package's budgets\n * are small (DEFAULT_MAX_TOKENS 1024; the headline call passes 40) — left on,\n * the model can spend the whole budget thinking and truncate the visible\n * text. Same failure mode as Gemini 2.5 Pro's mandatory thinking budget (see\n * the MODELS.gemini comment), same fix: turn it off explicitly.\n * `{type:'disabled'}` is accepted on both models at the default effort\n * (verified live 2026-08-14). Do NOT add `claude-fable-5`/`claude-mythos-5`\n * here — those models reject an explicit `disabled` (thinking is always on).\n */\nconst ANTHROPIC_THINKING_DEFAULT_ON = /^claude-(opus-5|sonnet-5)$/;\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    messages: filtered.map((m) => ({ role: m.role, content: m.content })),\n  };\n  if (!ANTHROPIC_SAMPLING_REMOVED.test(model)) {\n    body.temperature = opts.temperature ?? DEFAULT_TEMPERATURE;\n  }\n  if (ANTHROPIC_THINKING_DEFAULT_ON.test(model)) {\n    body.thinking = { type: 'disabled' };\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\n/** A resolved Vertex credential: a bearer token plus (for ADC) the ambient project id. */\ninterface VertexAuth {\n  token: string;\n  /** Project id discovered from the metadata server (ADC path only). */\n  project?: string;\n}\n\n/**\n * Resolve a Vertex credential, in priority order:\n *   1. `GCP_SA_KEY` — mint a fresh token from the service-account key (the\n *      credential Cloudflare Workers supply, where there is no metadata server).\n *   2. `VERTEX_ACCESS_TOKEN` — a pre-minted token (legacy / CI override).\n *   3. Application Default Credentials — the GCP metadata server's ambient\n *      service-account token. This is the KEYLESS default on Cloud Run / GCE:\n *      neither env var need be set, mirroring the post-WIF-migration pattern\n *      the render-runner already uses for every other GCP API call.\n *\n * @throws ValidationError when no static credential is set AND the metadata\n * server is unreachable (i.e. off-GCP) — the leg fails before the request so\n * the router falls through to the next provider, never calling Vertex\n * unauthenticated.\n */\nasync function resolveVertexAuth(env: LLMEnv, fetchImpl: typeof fetch): Promise<VertexAuth> {\n  if (env.GCP_SA_KEY) {\n    try {\n      return { token: await mintGcpAccessToken(env.GCP_SA_KEY, fetchImpl) };\n    } catch (error) {\n      // A pre-minted token is worth trying if minting broke. With nothing to fall\n      // back to, surface the minting failure rather than swallowing it.\n      if (!env.VERTEX_ACCESS_TOKEN) throw error;\n    }\n  }\n  if (env.VERTEX_ACCESS_TOKEN) return { token: env.VERTEX_ACCESS_TOKEN };\n  // Keyless default: Application Default Credentials from the GCP metadata\n  // server. Throws (→ chain falls through) when off-GCP.\n  return fetchAdcAccessToken(fetchImpl);\n}\n\nfunction buildGeminiRequest(\n  model: string,\n  messages: LLMMessage[],\n  opts: LLMOptions,\n  env: LLMEnv,\n  accessToken: string,\n  adcProject?: string,\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      // Disable \"thinking\": Gemini 2.5 draws thinking tokens from\n      // maxOutputTokens, so on a large prompt (or a small token budget) the\n      // model can spend the entire budget thinking and return an empty\n      // candidate (finishReason MAX_TOKENS). This leg wants deterministic text\n      // out, so thinking is turned off (supported by gemini-2.5-flash).\n      thinkingConfig: { thinkingBudget: 0 },\n    },\n  };\n  if (sys) {\n    body.systemInstruction = { parts: [{ text: sys }] };\n  }\n  // Project resolution: an explicit VERTEX_PROJECT wins; else the service\n  // account's own project (GCP_SA_KEY path); else the ambient project id the\n  // metadata server reported (ADC path). So a Cloud Run job needs no Vertex\n  // config at all — the runner's own project is used.\n  const project =\n    env.VERTEX_PROJECT ||\n    (env.GCP_SA_KEY ? serviceAccountProjectId(env.GCP_SA_KEY) : (adcProject ?? ''));\n  const location = env.VERTEX_LOCATION || DEFAULT_VERTEX_LOCATION;\n  const path = `v1/projects/${project}/locations/${location}/publishers/google/models/${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 ${accessToken}`,\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/**\n * Converts normalized messages to Ollama's native chat shape. Native tool calls\n * carry object arguments (rather than OpenAI's JSON string), and tool results\n * identify the function by name rather than by call id.\n */\nfunction toOllamaMessages(messages: LLMMessage[], sys: string | undefined): Array<Record<string, unknown>> {\n  const toolNames = new Map<string, string>();\n  for (const message of messages) {\n    if (typeof message.content === 'string') continue;\n    for (const block of message.content) {\n      if (block.type === 'tool_use') toolNames.set(block.id, block.name);\n    }\n  }\n\n  const out: Array<Record<string, unknown>> = [];\n  if (sys) out.push({ role: 'system', content: sys });\n  for (const message of messages) {\n    if (message.role === 'system') continue;\n    if (typeof message.content === 'string') {\n      out.push({ role: message.role, content: message.content });\n      continue;\n    }\n\n    let text = '';\n    const toolCalls: Array<Record<string, unknown>> = [];\n    const results: Array<{ toolUseId: string; content: string }> = [];\n    for (const block of message.content) {\n      if (block.type === 'text') text += block.text;\n      else if (block.type === 'tool_use') {\n        toolCalls.push({\n          type: 'function',\n          function: { index: toolCalls.length, name: block.name, arguments: block.input },\n        });\n      } else if (block.type === 'tool_result') {\n        results.push({ toolUseId: block.tool_use_id, content: block.content });\n      }\n    }\n    if (results.length > 0) {\n      for (const result of results) {\n        out.push({\n          role: 'tool',\n          tool_name: toolNames.get(result.toolUseId) ?? result.toolUseId,\n          content: result.content,\n        });\n      }\n      if (text) out.push({ role: 'user', content: text });\n    } else if (toolCalls.length > 0) {\n      out.push({ role: 'assistant', content: text, tool_calls: toolCalls });\n    } else {\n      out.push({ role: message.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/**\n * Builds a local Qwen request behind Cloudflare AI Gateway. Qwen 8B uses the\n * dedicated fast origin's native Ollama contract so `think:false` is enforced;\n * Qwen 27B retains the OpenAI-compatible agentic/tool-call route.\n */\nfunction buildLocalRequest(\n  model: string,\n  messages: LLMMessage[],\n  opts: LLMOptions,\n  env: LLMEnv,\n): { url: string; headers: Record<string, string>; body: string } {\n  if (!env.GPU_LLM_API_TOKEN) {\n    throw new ValidationError('GPU_LLM_API_TOKEN required for the local provider');\n  }\n  const sys = systemText(opts, messages);\n  const fastMode = model === MODELS.local.fast;\n  const headers: Record<string, string> = {\n    'content-type': 'application/json',\n    authorization: `Bearer ${env.GPU_LLM_API_TOKEN}`,\n  };\n  if (env.GPU_LLM_ACCESS_CLIENT_ID) headers['CF-Access-Client-Id'] = env.GPU_LLM_ACCESS_CLIENT_ID;\n  if (env.GPU_LLM_ACCESS_CLIENT_SECRET) headers['CF-Access-Client-Secret'] = env.GPU_LLM_ACCESS_CLIENT_SECRET;\n  const tools = openAiTools(opts);\n  if (fastMode) {\n    if (typeof opts.toolChoice === 'object') {\n      throw new ValidationError('named tool choice is not supported by the native Qwen 8B route');\n    }\n    return {\n      url: `${env.AI_GATEWAY_BASE_URL}/custom-local-fast-gpu/api/chat`,\n      headers,\n      body: JSON.stringify({\n        model,\n        stream: false,\n        think: false,\n        messages: toOllamaMessages(messages, sys),\n        options: {\n          num_predict: opts.maxTokens ?? DEFAULT_MAX_TOKENS,\n          temperature: opts.temperature ?? DEFAULT_TEMPERATURE,\n        },\n        ...(tools && opts.toolChoice !== 'none' ? { tools } : {}),\n      }),\n    };\n  }\n  return {\n    url: `${env.AI_GATEWAY_BASE_URL}/custom-local-gpu/v1/chat/completions`,\n    headers,\n    body: JSON.stringify({\n      model,\n      max_tokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,\n      temperature: opts.temperature ?? DEFAULT_TEMPERATURE,\n      messages: toOpenAiMessages(messages, sys),\n      ...(tools\n        ? { tools, tool_choice: openAiToolChoice(opts.toolChoice ?? 'auto') }\n        : {}),\n    }),\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\ninterface OllamaChatResponse {\n  model?: string;\n  message?: {\n    content?: string;\n    thinking?: string;\n    tool_calls?: Array<{\n      function?: { name?: string; arguments?: Record<string, unknown> | string };\n    }>;\n  };\n  done?: boolean;\n  done_reason?: string;\n  prompt_eval_count?: number;\n  eval_count?: number;\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/** Parses Ollama's native non-streaming `/api/chat` response for Qwen 8B. */\nfunction parseOllamaChat(json: unknown): {\n  content: string;\n  input: number;\n  output: number;\n  model?: string;\n  toolCalls?: LLMToolCall[];\n  stopReason?: LLMResult['stopReason'];\n} {\n  const response = json as OllamaChatResponse;\n  const toolCalls: LLMToolCall[] = (response.message?.tool_calls ?? [])\n    .filter((call) => typeof call.function?.name === 'string')\n    .map((call, index) => ({\n      id: `call_${index}`,\n      name: call.function!.name!,\n      arguments:\n        typeof call.function?.arguments === 'string'\n          ? parseToolArgs(call.function.arguments)\n          : (call.function?.arguments ?? {}),\n    }));\n  let stopReason: LLMResult['stopReason'];\n  if (toolCalls.length > 0) stopReason = 'tool_use';\n  else if (response.done_reason === 'length') stopReason = 'max_tokens';\n  else if (response.done) stopReason = 'end';\n  else if (response.done_reason) stopReason = 'other';\n  return {\n    // Deliberately exclude message.thinking: callers receive only the bounded\n    // classification answer even if a non-conforming origin returns a trace.\n    content: response.message?.content ?? '',\n    input: response.prompt_eval_count ?? 0,\n    output: response.eval_count ?? 0,\n    model: response.model,\n    toolCalls: toolCalls.length > 0 ? toolCalls : undefined,\n    stopReason,\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', 'local']);\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    if (m.startsWith('qwen')) return { primary: { provider: 'local', 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      // Credential resolution is a network call (mint from key, or the metadata\n      // server for ADC), so it must happen here rather than inside the\n      // (synchronous) request builders. Cached per isolate, so it is almost\n      // always a no-op after the first gemini call.\n      const vertexAuth = await resolveVertexAuth(env, fetchImpl);\n      req = buildGeminiRequest(leg.model, messages, opts, env, vertexAuth.token, vertexAuth.project);\n      break;\n    }\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    case 'local':\n      req = buildLocalRequest(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    case 'local':\n      return {\n        parsed: leg.model === MODELS.local.fast ? parseOllamaChat(json) : parseOpenAi(json),\n        gatewayRequestId,\n        attempts,\n      };\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 Flash if `longContextThreshold` exceeded\n *   - `smart`    → Anthropic Opus; Gemini 2.5 Flash 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  let route = plan(tier, opts, tokenEstimate);\n  // Cost optimization: prefer the dedicated Qwen 8B GPU ($0) for cheap `fast`-tier\n  // work, with the original fast route (Grok→Haiku) as automatic fallback on any\n  // error. Dormant unless LLM_LOCAL_FIRST + GPU_LLM_API_TOKEN are set.\n  if (\n    env.LLM_LOCAL_FIRST &&\n    tier === 'fast' &&\n    !opts.model &&\n    env.GPU_LLM_API_TOKEN\n  ) {\n    route = { primary: { provider: 'local', model: MODELS.local.fast }, fallback: route.primary };\n  }\n  // Qwen 27B owns local agentic work. DeepSeek remains the independent fallback;\n  // explicit model overrides always win.\n  if (\n    env.LLM_LOCAL_WORKBENCH &&\n    tier === 'workbench' &&\n    !opts.model &&\n    env.GPU_LLM_API_TOKEN\n  ) {\n    route = { primary: { provider: 'local', model: MODELS.local.workbench }, fallback: route.primary };\n  }\n\n  // ── Org-level daily / monthly cap pre-check ──────────────────────────────\n  const kv = env.LLM_COST_KV;\n  const todayKey = `llm:daily-cost:${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; partial_json?: string; stop_reason?: string };\n  content_block?: { type?: string; id?: string; name?: 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  // Tool-use blocks arrive as content_block_start (id/name) then a sequence of\n  // input_json_delta fragments that concatenate into the arguments JSON string.\n  const toolBlocks = new Map<number, { id: string; name: string; json: string }>();\n  let streamStopReason: LLMResult['stopReason'];\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_start':\n            if (\n              event.content_block?.type === 'tool_use' &&\n              typeof event.index === 'number' &&\n              typeof event.content_block.id === 'string' &&\n              typeof event.content_block.name === 'string'\n            ) {\n              toolBlocks.set(event.index, { id: event.content_block.id, name: event.content_block.name, json: '' });\n            }\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            } else if (\n              event.delta?.type === 'input_json_delta' &&\n              typeof event.delta.partial_json === 'string' &&\n              typeof event.index === 'number'\n            ) {\n              const block = toolBlocks.get(event.index);\n              if (block) block.json += event.delta.partial_json;\n            }\n            break;\n          case 'message_delta':\n            outputTokens = event.usage?.output_tokens ?? outputTokens;\n            if (event.delta?.stop_reason) streamStopReason = normalizeAnthropicStop(event.delta.stop_reason);\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  const toolCalls: LLMToolCall[] = [...toolBlocks.values()].map((b) => ({\n    id: b.id,\n    name: b.name,\n    arguments: parseToolArgs(b.json),\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    stopReason: streamStopReason,\n    toolCalls: toolCalls.length > 0 ? toolCalls : undefined,\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\n// ─── Embeddings ────────────────────────────────────────────────────────────\n\nexport { embed, embedLocal, DEFAULT_EMBEDDING_MODEL, LOCAL_EMBEDDING_MODEL } from './embed.js';\nexport type { AiBinding, EmbedResult, EmbeddingModel, LocalEmbedEnv } from './embed.js';\n\nexport { mintGcpAccessToken, clearGcpTokenCache, serviceAccountProjectId } from './gcp-token.js';\n","/**\n * Mint short-lived Google Cloud OAuth2 access tokens from a service-account key,\n * inside a Cloudflare Worker.\n *\n * The `gemini` leg authenticates to Vertex AI with a GCP access token, which\n * expires after one hour. A token therefore cannot be a durable Worker secret —\n * any stored value is stale almost immediately. Callers supply the long-lived\n * service-account key instead, and this module exchanges it for a fresh token\n * via the JWT-bearer flow, caching the result per isolate.\n *\n * Web Crypto only — no Node.js built-ins, no `Buffer`. Lifted from\n * `apps/admin-studio/src/lib/gcp-secrets.ts`, where this exact flow has been\n * running in production; hoisted here so every consumer of the `gemini` leg\n * gets a working credential rather than just the one app that hand-rolled it.\n */\nimport { InternalError, ValidationError } from '@latimer-woods-tech/errors';\n\n/** The subset of a GCP service-account JSON key needed for the JWT-bearer flow. */\ninterface ServiceAccountKey {\n  project_id: string;\n  private_key: string;\n  client_email: string;\n  token_uri: string;\n}\n\ninterface OAuth2Response {\n  access_token: string;\n  expires_in: number;\n}\n\n/** Refresh this long before expiry so an in-flight request never races the clock. */\nconst EXPIRY_SKEW_MS = 300_000;\n\n/**\n * Per-isolate token cache, keyed by service-account email.\n *\n * Caching a *string* across requests is safe. Caching an I/O object (a socket, a\n * database client, a `Response`) is not — Workers rejects reuse of I/O created\n * on behalf of a different request. This holds only the token text and its\n * expiry, so it is reused freely.\n */\nconst tokenCache = new Map<string, { token: string; expiresAt: number }>();\n\n/** Decode a base64url-encoded service-account key, tolerating raw JSON too. */\nfunction parseServiceAccountKey(gcpSaKey: string): ServiceAccountKey {\n  const text = gcpSaKey.trimStart().startsWith('{') ? gcpSaKey : atob(gcpSaKey);\n  const key = JSON.parse(text) as ServiceAccountKey;\n  if (!key.client_email || !key.private_key || !key.token_uri) {\n    throw new ValidationError('GCP_SA_KEY is missing client_email, private_key, or token_uri');\n  }\n  return key;\n}\n\n/** Encode bytes as base64url (RFC 4648 §5): `-`/`_` for `+`/`/`, no padding. */\nfunction base64UrlEncode(bytes: Uint8Array): string {\n  let binary = '';\n  for (let i = 0; i < bytes.length; i++) binary += String.fromCharCode(bytes[i]!);\n  return btoa(binary).replace(/\\+/g, '-').replace(/\\//g, '_').replace(/=+$/g, '');\n}\n\n/** Import a PEM-encoded PKCS#8 private key as a Web Crypto signing key. */\nasync function importPrivateKey(pem: string): Promise<CryptoKey> {\n  const body = pem\n    .split('\\n')\n    .filter((line) => !line.startsWith('-----'))\n    .join('');\n  const binary = atob(body);\n  const bytes = new Uint8Array(binary.length);\n  for (let i = 0; i < binary.length; i++) bytes[i] = binary.charCodeAt(i);\n  return crypto.subtle.importKey(\n    'pkcs8',\n    bytes.buffer,\n    { name: 'RSASSA-PKCS1-v1_5', hash: 'SHA-256' },\n    false,\n    ['sign'],\n  );\n}\n\n/** Build a service-account JWT assertion signed with RS256, per GCP's spec. */\nasync function createAssertion(key: ServiceAccountKey, nowSeconds: number): Promise<string> {\n  const encoder = new TextEncoder();\n  const header = base64UrlEncode(encoder.encode(JSON.stringify({ alg: 'RS256', typ: 'JWT' })));\n  const payload = base64UrlEncode(\n    encoder.encode(\n      JSON.stringify({\n        iss: key.client_email,\n        scope: 'https://www.googleapis.com/auth/cloud-platform',\n        aud: key.token_uri,\n        exp: nowSeconds + 3600,\n        iat: nowSeconds,\n      }),\n    ),\n  );\n  const signingInput = `${header}.${payload}`;\n  const cryptoKey = await importPrivateKey(key.private_key);\n  const signature = await crypto.subtle.sign(\n    'RSASSA-PKCS1-v1_5',\n    cryptoKey,\n    encoder.encode(signingInput).buffer,\n  );\n  return `${signingInput}.${base64UrlEncode(new Uint8Array(signature))}`;\n}\n\n/**\n * Exchange a service-account key for a cloud-platform access token, reusing a\n * cached token until it is within {@link EXPIRY_SKEW_MS} of expiring.\n *\n * @param gcpSaKey - Service-account JSON key, base64-encoded or raw.\n * @param fetchImpl - Fetch implementation (injectable for tests).\n * @returns A bearer token valid for Vertex AI.\n * @throws ValidationError when the key is malformed; InternalError when the exchange fails.\n */\nexport async function mintGcpAccessToken(gcpSaKey: string, fetchImpl: typeof fetch): Promise<string> {\n  const key = parseServiceAccountKey(gcpSaKey);\n\n  const now = Date.now();\n  const cached = tokenCache.get(key.client_email);\n  if (cached && cached.expiresAt - EXPIRY_SKEW_MS > now) return cached.token;\n\n  const assertion = await createAssertion(key, Math.floor(now / 1000));\n  const response = await fetchImpl(key.token_uri, {\n    method: 'POST',\n    headers: { 'content-type': 'application/x-www-form-urlencoded' },\n    body: new URLSearchParams({\n      grant_type: 'urn:ietf:params:oauth:grant-type:jwt-bearer',\n      assertion,\n    }).toString(),\n  }).catch((cause: unknown) => {\n    throw new InternalError(`GCP token exchange failed: ${String(cause)}`);\n  });\n\n  if (!response.ok) {\n    throw new InternalError(`GCP token exchange returned ${response.status}`);\n  }\n\n  const data = (await response.json()) as OAuth2Response;\n  if (!data.access_token) {\n    throw new InternalError('GCP token exchange returned no access_token');\n  }\n\n  // Trust the server's TTL rather than assuming the 3600s default.\n  const ttlMs = (data.expires_in > 0 ? data.expires_in : 3600) * 1000;\n  tokenCache.set(key.client_email, { token: data.access_token, expiresAt: now + ttlMs });\n  return data.access_token;\n}\n\n// ─── Application Default Credentials (GCP metadata server) ───────────────────\n/**\n * GCE / Cloud Run metadata server base. Reachable only from GCP compute; the\n * hostname does not resolve elsewhere, so an off-GCP probe fails fast.\n */\nconst METADATA_BASE = 'http://metadata.google.internal/computeMetadata/v1';\n/** Required header for every metadata-server request (anti-SSRF guard). */\nconst METADATA_HEADER = { 'Metadata-Flavor': 'Google' } as const;\n/**\n * Bound the metadata probe so a non-GCP environment (where the host may hang\n * rather than refuse) falls through to the next provider quickly.\n */\nconst METADATA_TIMEOUT_MS = 2_000;\n\n/** Per-isolate cache for the ADC (metadata-server) credential. */\nlet adcCache: { token: string; project?: string; expiresAt: number } | undefined;\n\ninterface AdcCredential {\n  token: string;\n  /** The instance's own GCP project id, as reported by the metadata server. */\n  project?: string;\n}\n\n/**\n * Obtain an access token (and project id) from the GCP metadata server using\n * Application Default Credentials — the keyless credential every Cloud Run /\n * GCE workload already has via its attached service account. No key, no env\n * var: the instance SA's cloud-platform token is served by the metadata\n * server. This mirrors the render-runner's own `_meta_token` shell helper.\n *\n * Fails (throws {@link InternalError}) when the metadata server is unreachable\n * or returns no token — i.e. off-GCP — so the caller falls through to the next\n * provider leg. Never throws a bare `AbortError` (which the router treats as a\n * hard cancellation): an internal timeout is surfaced as an `InternalError`.\n *\n * @param fetchImpl - Fetch implementation (injectable for tests).\n * @returns A bearer token valid for Vertex AI plus the ambient project id.\n */\nexport async function fetchAdcAccessToken(fetchImpl: typeof fetch): Promise<AdcCredential> {\n  const now = Date.now();\n  if (adcCache && adcCache.expiresAt - EXPIRY_SKEW_MS > now) {\n    return { token: adcCache.token, project: adcCache.project };\n  }\n\n  const controller = new AbortController();\n  const timer = setTimeout(() => controller.abort(), METADATA_TIMEOUT_MS);\n  try {\n    const tokenRes = await fetchImpl(\n      `${METADATA_BASE}/instance/service-accounts/default/token`,\n      { headers: METADATA_HEADER, signal: controller.signal },\n    );\n    if (!tokenRes.ok) {\n      throw new InternalError(`GCP metadata token endpoint returned ${tokenRes.status}`);\n    }\n    const data = (await tokenRes.json()) as OAuth2Response;\n    if (!data.access_token) {\n      throw new InternalError('GCP metadata token endpoint returned no access_token');\n    }\n\n    // Project id is best-effort: a caller may pin VERTEX_PROJECT instead.\n    let project: string | undefined;\n    try {\n      const projRes = await fetchImpl(`${METADATA_BASE}/project/project-id`, {\n        headers: METADATA_HEADER,\n        signal: controller.signal,\n      });\n      if (projRes.ok) project = (await projRes.text()).trim() || undefined;\n    } catch {\n      /* project id is optional — the token is still usable */\n    }\n\n    const ttlMs = (data.expires_in > 0 ? data.expires_in : 3600) * 1000;\n    adcCache = { token: data.access_token, project, expiresAt: now + ttlMs };\n    return { token: data.access_token, project };\n  } catch (cause) {\n    // Normalize an internal-timeout AbortError into an InternalError so the\n    // router falls through to the next leg instead of hard-cancelling the call.\n    if (cause instanceof InternalError) throw cause;\n    throw new InternalError(`GCP metadata credential unavailable: ${String(cause)}`);\n  } finally {\n    clearTimeout(timer);\n  }\n}\n\n/** Reset the per-isolate token cache. Tests only. */\nexport function clearGcpTokenCache(): void {\n  tokenCache.clear();\n  adcCache = undefined;\n}\n\n/** The project the service account belongs to — the Vertex project by default. */\nexport function serviceAccountProjectId(gcpSaKey: string): string {\n  return parseServiceAccountKey(gcpSaKey).project_id;\n}\n","/**\n * Workers AI embedding API.\n *\n * Pins to bge-base-en-v1.5 (768-dim cosine) — the platform standard.\n * @remarks Local-rail errors use @latimer-woods-tech/errors (observability ratchet).\n * model_version is returned with every result so callers can record it\n * for provenance-aware re-embed when the model is swapped.\n *\n * Contract: inject the Workers AI binding (env.AI) — never import a vendor SDK.\n */\n\nimport { InternalError } from '@latimer-woods-tech/errors';\n\n/** The platform-standard embedding model (768-dim cosine) used when no override is given. */\nexport const DEFAULT_EMBEDDING_MODEL = '@cf/baai/bge-base-en-v1.5';\n\n/** Supported embedding model identifiers — widen deliberately, since dims are part of the contract. */\nexport type EmbeddingModel = '@cf/baai/bge-base-en-v1.5';\n\n/** One embedding call's output: the vectors plus the provenance needed to re-embed later. */\nexport interface EmbedResult {\n  vectors: number[][];\n  model: string;\n  dims: number;\n}\n\n/** Minimal Workers AI binding shape needed for embeddings. */\nexport interface AiBinding {\n  run(\n    model: string,\n    inputs: { text: string | string[] },\n  ): Promise<{ data: number[][] }>;\n}\n\n/**\n * Embeds one or more text strings using the Workers AI binding.\n *\n * @param ai    - Workers AI binding (`env.AI`) — injected, not imported.\n * @param input - Single string or array of strings to embed.\n * @param opts  - Optional model override (must be a supported EmbeddingModel).\n * @returns Embedding vectors, model name, and dimension count.\n *\n * @throws When the AI binding call fails (let the caller decide whether to catch).\n */\nexport async function embed(\n  ai: AiBinding,\n  input: string | string[],\n  opts?: { model?: EmbeddingModel },\n): Promise<EmbedResult> {\n  const model = opts?.model ?? DEFAULT_EMBEDDING_MODEL;\n  const texts = Array.isArray(input) ? input : [input];\n  const result = await ai.run(model, { text: texts });\n  const vectors = result.data;\n  if (!vectors || vectors.length === 0) {\n    throw new Error(`embed(): Workers AI returned no vectors for model ${model}`);\n  }\n  return {\n    vectors,\n    model,\n    dims: vectors[0]!.length,\n  };\n}\n\n/**\n * Self-hosted GPU embedding model (nomic-embed-text, 768-dim), reached through\n * the same CF AI Gateway custom provider as the local chat provider.\n *\n * ⚠️ VECTOR-SPACE WARNING: nomic-embed-text and bge-base-en-v1.5 are BOTH 768-dim\n * but occupy DIFFERENT embedding spaces. Their vectors are NOT comparable. A store\n * built with one model must be QUERIED — and, when migrating, RE-EMBEDDED — with\n * the SAME model. Never write vectors from both models into one store.\n */\nexport const LOCAL_EMBEDDING_MODEL = 'nomic-embed-text';\n\n/** Env subset needed to reach the local embedding rail (shares GPU_LLM_* with the chat provider). */\nexport interface LocalEmbedEnv {\n  AI_GATEWAY_BASE_URL: string;\n  GPU_LLM_API_TOKEN?: string;\n  GPU_LLM_ACCESS_CLIENT_ID?: string;\n  GPU_LLM_ACCESS_CLIENT_SECRET?: string;\n}\n\n/**\n * Embeds text on the self-hosted GPU rail (nomic-embed-text) via the AI Gateway\n * custom provider — the $0 alternative to Workers AI {@link embed}.\n *\n * Unlike the local CHAT provider, this has NO cross-provider fallback ON PURPOSE:\n * falling back to Workers AI (bge) would silently write an incompatible-space\n * vector into a nomic store and corrupt retrieval. If the rail is unavailable\n * this THROWS and the caller decides (retry local, or defer the write) — it must\n * NOT substitute a different model.\n *\n * @param env   - Gateway base URL + GPU_LLM_* auth (same secrets as the chat provider).\n * @param input - Single string or array of strings to embed.\n * @returns Embedding vectors, model `nomic-embed-text`, and dimension count.\n * @throws  On any transport/HTTP/empty-response error (no silent fallback).\n */\nexport async function embedLocal(\n  env: LocalEmbedEnv,\n  input: string | string[],\n): Promise<EmbedResult> {\n  const texts = Array.isArray(input) ? input : [input];\n  const res = await fetch(\n    `${env.AI_GATEWAY_BASE_URL}/custom-local-gpu/v1/embeddings`,\n    {\n      method: 'POST',\n      headers: {\n        'Content-Type': 'application/json',\n        // CF Access service-token headers only sent when present (bearer-only rails still work).\n        ...(env.GPU_LLM_API_TOKEN ? { Authorization: `Bearer ${env.GPU_LLM_API_TOKEN}` } : {}),\n        ...(env.GPU_LLM_ACCESS_CLIENT_ID ? { 'CF-Access-Client-Id': env.GPU_LLM_ACCESS_CLIENT_ID } : {}),\n        ...(env.GPU_LLM_ACCESS_CLIENT_SECRET ? { 'CF-Access-Client-Secret': env.GPU_LLM_ACCESS_CLIENT_SECRET } : {}),\n      },\n      body: JSON.stringify({ model: LOCAL_EMBEDDING_MODEL, input: texts }),\n    },\n  );\n  if (!res.ok) {\n    throw new InternalError(\n      `embedLocal(): rail ${res.status}: ${(await res.text()).slice(0, 160)}`,\n    );\n  }\n  const json = (await res.json()) as { data?: Array<{ embedding: number[] }> };\n  const vectors = (json.data ?? []).map((d) => d.embedding);\n  if (vectors.length === 0) {\n    throw new InternalError(\n      `embedLocal(): rail returned no vectors for model ${LOCAL_EMBEDDING_MODEL}`,\n    );\n  }\n  return {\n    vectors,\n    model: LOCAL_EMBEDDING_MODEL,\n    dims: vectors[0]!.length,\n  };\n}\n"],"mappings":";AAAA;AAAA,EACE,iBAAAA;AAAA,EACA;AAAA,EACA,mBAAAC;AAAA,EACA;AAAA,OAEK;;;ACSP,SAAS,eAAe,uBAAuB;AAgB/C,IAAM,iBAAiB;AAUvB,IAAM,aAAa,oBAAI,IAAkD;AAGzE,SAAS,uBAAuB,UAAqC;AACnE,QAAM,OAAO,SAAS,UAAU,EAAE,WAAW,GAAG,IAAI,WAAW,KAAK,QAAQ;AAC5E,QAAM,MAAM,KAAK,MAAM,IAAI;AAC3B,MAAI,CAAC,IAAI,gBAAgB,CAAC,IAAI,eAAe,CAAC,IAAI,WAAW;AAC3D,UAAM,IAAI,gBAAgB,+DAA+D;AAAA,EAC3F;AACA,SAAO;AACT;AAGA,SAAS,gBAAgB,OAA2B;AAClD,MAAI,SAAS;AACb,WAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,IAAK,WAAU,OAAO,aAAa,MAAM,CAAC,CAAE;AAC9E,SAAO,KAAK,MAAM,EAAE,QAAQ,OAAO,GAAG,EAAE,QAAQ,OAAO,GAAG,EAAE,QAAQ,QAAQ,EAAE;AAChF;AAGA,eAAe,iBAAiB,KAAiC;AAC/D,QAAM,OAAO,IACV,MAAM,IAAI,EACV,OAAO,CAAC,SAAS,CAAC,KAAK,WAAW,OAAO,CAAC,EAC1C,KAAK,EAAE;AACV,QAAM,SAAS,KAAK,IAAI;AACxB,QAAM,QAAQ,IAAI,WAAW,OAAO,MAAM;AAC1C,WAAS,IAAI,GAAG,IAAI,OAAO,QAAQ,IAAK,OAAM,CAAC,IAAI,OAAO,WAAW,CAAC;AACtE,SAAO,OAAO,OAAO;AAAA,IACnB;AAAA,IACA,MAAM;AAAA,IACN,EAAE,MAAM,qBAAqB,MAAM,UAAU;AAAA,IAC7C;AAAA,IACA,CAAC,MAAM;AAAA,EACT;AACF;AAGA,eAAe,gBAAgB,KAAwB,YAAqC;AAC1F,QAAM,UAAU,IAAI,YAAY;AAChC,QAAM,SAAS,gBAAgB,QAAQ,OAAO,KAAK,UAAU,EAAE,KAAK,SAAS,KAAK,MAAM,CAAC,CAAC,CAAC;AAC3F,QAAM,UAAU;AAAA,IACd,QAAQ;AAAA,MACN,KAAK,UAAU;AAAA,QACb,KAAK,IAAI;AAAA,QACT,OAAO;AAAA,QACP,KAAK,IAAI;AAAA,QACT,KAAK,aAAa;AAAA,QAClB,KAAK;AAAA,MACP,CAAC;AAAA,IACH;AAAA,EACF;AACA,QAAM,eAAe,GAAG,MAAM,IAAI,OAAO;AACzC,QAAM,YAAY,MAAM,iBAAiB,IAAI,WAAW;AACxD,QAAM,YAAY,MAAM,OAAO,OAAO;AAAA,IACpC;AAAA,IACA;AAAA,IACA,QAAQ,OAAO,YAAY,EAAE;AAAA,EAC/B;AACA,SAAO,GAAG,YAAY,IAAI,gBAAgB,IAAI,WAAW,SAAS,CAAC,CAAC;AACtE;AAWA,eAAsB,mBAAmB,UAAkB,WAA0C;AACnG,QAAM,MAAM,uBAAuB,QAAQ;AAE3C,QAAM,MAAM,KAAK,IAAI;AACrB,QAAM,SAAS,WAAW,IAAI,IAAI,YAAY;AAC9C,MAAI,UAAU,OAAO,YAAY,iBAAiB,IAAK,QAAO,OAAO;AAErE,QAAM,YAAY,MAAM,gBAAgB,KAAK,KAAK,MAAM,MAAM,GAAI,CAAC;AACnE,QAAM,WAAW,MAAM,UAAU,IAAI,WAAW;AAAA,IAC9C,QAAQ;AAAA,IACR,SAAS,EAAE,gBAAgB,oCAAoC;AAAA,IAC/D,MAAM,IAAI,gBAAgB;AAAA,MACxB,YAAY;AAAA,MACZ;AAAA,IACF,CAAC,EAAE,SAAS;AAAA,EACd,CAAC,EAAE,MAAM,CAAC,UAAmB;AAC3B,UAAM,IAAI,cAAc,8BAA8B,OAAO,KAAK,CAAC,EAAE;AAAA,EACvE,CAAC;AAED,MAAI,CAAC,SAAS,IAAI;AAChB,UAAM,IAAI,cAAc,+BAA+B,SAAS,MAAM,EAAE;AAAA,EAC1E;AAEA,QAAM,OAAQ,MAAM,SAAS,KAAK;AAClC,MAAI,CAAC,KAAK,cAAc;AACtB,UAAM,IAAI,cAAc,6CAA6C;AAAA,EACvE;AAGA,QAAM,SAAS,KAAK,aAAa,IAAI,KAAK,aAAa,QAAQ;AAC/D,aAAW,IAAI,IAAI,cAAc,EAAE,OAAO,KAAK,cAAc,WAAW,MAAM,MAAM,CAAC;AACrF,SAAO,KAAK;AACd;AAOA,IAAM,gBAAgB;AAEtB,IAAM,kBAAkB,EAAE,mBAAmB,SAAS;AAKtD,IAAM,sBAAsB;AAG5B,IAAI;AAuBJ,eAAsB,oBAAoB,WAAiD;AACzF,QAAM,MAAM,KAAK,IAAI;AACrB,MAAI,YAAY,SAAS,YAAY,iBAAiB,KAAK;AACzD,WAAO,EAAE,OAAO,SAAS,OAAO,SAAS,SAAS,QAAQ;AAAA,EAC5D;AAEA,QAAM,aAAa,IAAI,gBAAgB;AACvC,QAAM,QAAQ,WAAW,MAAM,WAAW,MAAM,GAAG,mBAAmB;AACtE,MAAI;AACF,UAAM,WAAW,MAAM;AAAA,MACrB,GAAG,aAAa;AAAA,MAChB,EAAE,SAAS,iBAAiB,QAAQ,WAAW,OAAO;AAAA,IACxD;AACA,QAAI,CAAC,SAAS,IAAI;AAChB,YAAM,IAAI,cAAc,wCAAwC,SAAS,MAAM,EAAE;AAAA,IACnF;AACA,UAAM,OAAQ,MAAM,SAAS,KAAK;AAClC,QAAI,CAAC,KAAK,cAAc;AACtB,YAAM,IAAI,cAAc,sDAAsD;AAAA,IAChF;AAGA,QAAI;AACJ,QAAI;AACF,YAAM,UAAU,MAAM,UAAU,GAAG,aAAa,uBAAuB;AAAA,QACrE,SAAS;AAAA,QACT,QAAQ,WAAW;AAAA,MACrB,CAAC;AACD,UAAI,QAAQ,GAAI,YAAW,MAAM,QAAQ,KAAK,GAAG,KAAK,KAAK;AAAA,IAC7D,QAAQ;AAAA,IAER;AAEA,UAAM,SAAS,KAAK,aAAa,IAAI,KAAK,aAAa,QAAQ;AAC/D,eAAW,EAAE,OAAO,KAAK,cAAc,SAAS,WAAW,MAAM,MAAM;AACvE,WAAO,EAAE,OAAO,KAAK,cAAc,QAAQ;AAAA,EAC7C,SAAS,OAAO;AAGd,QAAI,iBAAiB,cAAe,OAAM;AAC1C,UAAM,IAAI,cAAc,wCAAwC,OAAO,KAAK,CAAC,EAAE;AAAA,EACjF,UAAE;AACA,iBAAa,KAAK;AAAA,EACpB;AACF;AAGO,SAAS,qBAA2B;AACzC,aAAW,MAAM;AACjB,aAAW;AACb;AAGO,SAAS,wBAAwB,UAA0B;AAChE,SAAO,uBAAuB,QAAQ,EAAE;AAC1C;;;ACpOA,SAAS,iBAAAC,sBAAqB;AAGvB,IAAM,0BAA0B;AA8BvC,eAAsB,MACpB,IACA,OACA,MACsB;AACtB,QAAM,QAAQ,MAAM,SAAS;AAC7B,QAAM,QAAQ,MAAM,QAAQ,KAAK,IAAI,QAAQ,CAAC,KAAK;AACnD,QAAM,SAAS,MAAM,GAAG,IAAI,OAAO,EAAE,MAAM,MAAM,CAAC;AAClD,QAAM,UAAU,OAAO;AACvB,MAAI,CAAC,WAAW,QAAQ,WAAW,GAAG;AACpC,UAAM,IAAI,MAAM,qDAAqD,KAAK,EAAE;AAAA,EAC9E;AACA,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,MAAM,QAAQ,CAAC,EAAG;AAAA,EACpB;AACF;AAWO,IAAM,wBAAwB;AAyBrC,eAAsB,WACpB,KACA,OACsB;AACtB,QAAM,QAAQ,MAAM,QAAQ,KAAK,IAAI,QAAQ,CAAC,KAAK;AACnD,QAAM,MAAM,MAAM;AAAA,IAChB,GAAG,IAAI,mBAAmB;AAAA,IAC1B;AAAA,MACE,QAAQ;AAAA,MACR,SAAS;AAAA,QACP,gBAAgB;AAAA;AAAA,QAEhB,GAAI,IAAI,oBAAoB,EAAE,eAAe,UAAU,IAAI,iBAAiB,GAAG,IAAI,CAAC;AAAA,QACpF,GAAI,IAAI,2BAA2B,EAAE,uBAAuB,IAAI,yBAAyB,IAAI,CAAC;AAAA,QAC9F,GAAI,IAAI,+BAA+B,EAAE,2BAA2B,IAAI,6BAA6B,IAAI,CAAC;AAAA,MAC5G;AAAA,MACA,MAAM,KAAK,UAAU,EAAE,OAAO,uBAAuB,OAAO,MAAM,CAAC;AAAA,IACrE;AAAA,EACF;AACA,MAAI,CAAC,IAAI,IAAI;AACX,UAAM,IAAIA;AAAA,MACR,sBAAsB,IAAI,MAAM,MAAM,MAAM,IAAI,KAAK,GAAG,MAAM,GAAG,GAAG,CAAC;AAAA,IACvE;AAAA,EACF;AACA,QAAM,OAAQ,MAAM,IAAI,KAAK;AAC7B,QAAM,WAAW,KAAK,QAAQ,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,SAAS;AACxD,MAAI,QAAQ,WAAW,GAAG;AACxB,UAAM,IAAIA;AAAA,MACR,oDAAoD,qBAAqB;AAAA,IAC3E;AAAA,EACF;AACA,SAAO;AAAA,IACL;AAAA,IACA,OAAO;AAAA,IACP,MAAM,QAAQ,CAAC,EAAG;AAAA,EACpB;AACF;;;AFzEA,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;AA0OA,IAAM,SAAS;AAAA,EACb,WAAW;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAMT,MAAM;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAUN,UAAU;AAAA,IACV,OAAO;AAAA,EACT;AAAA,EACA,QAAQ;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAUN,OAAO;AAAA,EACT;AAAA,EACA,MAAM;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAMJ,UAAU;AAAA,EACZ;AAAA,EACA,MAAM;AAAA,IACJ,MAAM;AAAA,EACR;AAAA,EACA,UAAU;AAAA,IACR,WAAW;AAAA,EACb;AAAA,EACA,OAAO;AAAA;AAAA;AAAA,IAGL,MAAM;AAAA,IACN,WAAW;AAAA,EACb;AACF;AAEA,IAAM,qBAAqB;AAC3B,IAAM,sBAAsB;AAE5B,IAAM,0BAA0B;AAChC,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;AAAA;AAAA;AAAA,EAK1H,oBAAoB,EAAE,OAAO,GAAM,QAAQ,GAAM,WAAW,KAAM,YAAY,KAAK;AAAA,EACnF,6BAA6B,EAAE,OAAO,GAAM,QAAQ,GAAM,WAAW,KAAM,YAAY,KAAK;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;AAAA;AAAA,EAIrF,mBAAmB,EAAE,OAAO,GAAM,QAAQ,IAAO,WAAW,KAAM,YAAY,KAAK;AAAA;AAAA,EAEnF,0BAA0B,EAAE,OAAO,IAAO,QAAQ,IAAO,WAAW,KAAM,YAAY,MAAM;AAAA;AAAA;AAAA,EAG5F,mBAAmB,EAAE,OAAO,GAAM,QAAQ,IAAO,WAAW,KAAM,YAAY,KAAK;AAAA;AAAA,EAEnF,iBAAiB,EAAE,OAAO,GAAM,QAAQ,IAAO,WAAW,KAAM,YAAY,KAAK;AAAA;AAAA,EAEjF,oBAAoB,EAAE,OAAO,KAAM,QAAQ,KAAM,WAAW,OAAO,YAAY,IAAK;AAAA;AAAA,EAEpF,kBAAkB,EAAE,OAAO,MAAM,QAAQ,IAAO,WAAW,MAAM,YAAY,IAAK;AAAA;AAAA;AAAA,EAGlF,2BAA2B,EAAE,OAAO,MAAM,QAAQ,MAAM,WAAW,GAAM,YAAY,EAAK;AAAA;AAAA,EAE1F,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;AAAA;AAAA,EAErF,YAAY,EAAE,OAAO,GAAK,QAAQ,GAAK,WAAW,GAAK,YAAY,EAAI;AAAA,EACvE,eAAe,EAAE,OAAO,GAAK,QAAQ,GAAK,WAAW,GAAK,YAAY,EAAI;AAC5E;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;AAWA,IAAM,6BAA6B;AAanC,IAAM,gCAAgC;AAEtC,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,UAAU,SAAS,IAAI,CAAC,OAAO,EAAE,MAAM,EAAE,MAAM,SAAS,EAAE,QAAQ,EAAE;AAAA,EACtE;AACA,MAAI,CAAC,2BAA2B,KAAK,KAAK,GAAG;AAC3C,SAAK,cAAc,KAAK,eAAe;AAAA,EACzC;AACA,MAAI,8BAA8B,KAAK,KAAK,GAAG;AAC7C,SAAK,WAAW,EAAE,MAAM,WAAW;AAAA,EACrC;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;AAwBA,eAAe,kBAAkB,KAAa,WAA8C;AAC1F,MAAI,IAAI,YAAY;AAClB,QAAI;AACF,aAAO,EAAE,OAAO,MAAM,mBAAmB,IAAI,YAAY,SAAS,EAAE;AAAA,IACtE,SAAS,OAAO;AAGd,UAAI,CAAC,IAAI,oBAAqB,OAAM;AAAA,IACtC;AAAA,EACF;AACA,MAAI,IAAI,oBAAqB,QAAO,EAAE,OAAO,IAAI,oBAAoB;AAGrE,SAAO,oBAAoB,SAAS;AACtC;AAEA,SAAS,mBACP,OACA,UACA,MACA,KACA,aACA,YACgE;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;AAAA;AAAA;AAAA;AAAA;AAAA,MAMjC,gBAAgB,EAAE,gBAAgB,EAAE;AAAA,IACtC;AAAA,EACF;AACA,MAAI,KAAK;AACP,SAAK,oBAAoB,EAAE,OAAO,CAAC,EAAE,MAAM,IAAI,CAAC,EAAE;AAAA,EACpD;AAKA,QAAM,UACJ,IAAI,mBACH,IAAI,aAAa,wBAAwB,IAAI,UAAU,IAAK,cAAc;AAC7E,QAAM,WAAW,IAAI,mBAAmB;AACxC,QAAM,OAAO,eAAe,OAAO,cAAc,QAAQ,6BAA6B,KAAK;AAC3F,SAAO;AAAA,IACL,KAAK,GAAG,IAAI,mBAAmB,qBAAqB,IAAI;AAAA,IACxD,SAAS;AAAA,MACP,gBAAgB;AAAA,MAChB,eAAe,UAAU,WAAW;AAAA,IACtC;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;AAOA,SAAS,iBAAiB,UAAwB,KAAyD;AACzG,QAAM,YAAY,oBAAI,IAAoB;AAC1C,aAAW,WAAW,UAAU;AAC9B,QAAI,OAAO,QAAQ,YAAY,SAAU;AACzC,eAAW,SAAS,QAAQ,SAAS;AACnC,UAAI,MAAM,SAAS,WAAY,WAAU,IAAI,MAAM,IAAI,MAAM,IAAI;AAAA,IACnE;AAAA,EACF;AAEA,QAAM,MAAsC,CAAC;AAC7C,MAAI,IAAK,KAAI,KAAK,EAAE,MAAM,UAAU,SAAS,IAAI,CAAC;AAClD,aAAW,WAAW,UAAU;AAC9B,QAAI,QAAQ,SAAS,SAAU;AAC/B,QAAI,OAAO,QAAQ,YAAY,UAAU;AACvC,UAAI,KAAK,EAAE,MAAM,QAAQ,MAAM,SAAS,QAAQ,QAAQ,CAAC;AACzD;AAAA,IACF;AAEA,QAAI,OAAO;AACX,UAAM,YAA4C,CAAC;AACnD,UAAM,UAAyD,CAAC;AAChE,eAAW,SAAS,QAAQ,SAAS;AACnC,UAAI,MAAM,SAAS,OAAQ,SAAQ,MAAM;AAAA,eAChC,MAAM,SAAS,YAAY;AAClC,kBAAU,KAAK;AAAA,UACb,MAAM;AAAA,UACN,UAAU,EAAE,OAAO,UAAU,QAAQ,MAAM,MAAM,MAAM,WAAW,MAAM,MAAM;AAAA,QAChF,CAAC;AAAA,MACH,WAAW,MAAM,SAAS,eAAe;AACvC,gBAAQ,KAAK,EAAE,WAAW,MAAM,aAAa,SAAS,MAAM,QAAQ,CAAC;AAAA,MACvE;AAAA,IACF;AACA,QAAI,QAAQ,SAAS,GAAG;AACtB,iBAAW,UAAU,SAAS;AAC5B,YAAI,KAAK;AAAA,UACP,MAAM;AAAA,UACN,WAAW,UAAU,IAAI,OAAO,SAAS,KAAK,OAAO;AAAA,UACrD,SAAS,OAAO;AAAA,QAClB,CAAC;AAAA,MACH;AACA,UAAI,KAAM,KAAI,KAAK,EAAE,MAAM,QAAQ,SAAS,KAAK,CAAC;AAAA,IACpD,WAAW,UAAU,SAAS,GAAG;AAC/B,UAAI,KAAK,EAAE,MAAM,aAAa,SAAS,MAAM,YAAY,UAAU,CAAC;AAAA,IACtE,OAAO;AACL,UAAI,KAAK,EAAE,MAAM,QAAQ,MAAM,SAAS,KAAK,CAAC;AAAA,IAChD;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,IAAIC,iBAAgB,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,IAAIA,iBAAgB,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;AAOA,SAAS,kBACP,OACA,UACA,MACA,KACgE;AAChE,MAAI,CAAC,IAAI,mBAAmB;AAC1B,UAAM,IAAIA,iBAAgB,mDAAmD;AAAA,EAC/E;AACA,QAAM,MAAM,WAAW,MAAM,QAAQ;AACrC,QAAM,WAAW,UAAU,OAAO,MAAM;AACxC,QAAM,UAAkC;AAAA,IACtC,gBAAgB;AAAA,IAChB,eAAe,UAAU,IAAI,iBAAiB;AAAA,EAChD;AACA,MAAI,IAAI,yBAA0B,SAAQ,qBAAqB,IAAI,IAAI;AACvE,MAAI,IAAI,6BAA8B,SAAQ,yBAAyB,IAAI,IAAI;AAC/E,QAAM,QAAQ,YAAY,IAAI;AAC9B,MAAI,UAAU;AACZ,QAAI,OAAO,KAAK,eAAe,UAAU;AACvC,YAAM,IAAIA,iBAAgB,gEAAgE;AAAA,IAC5F;AACA,WAAO;AAAA,MACL,KAAK,GAAG,IAAI,mBAAmB;AAAA,MAC/B;AAAA,MACA,MAAM,KAAK,UAAU;AAAA,QACnB;AAAA,QACA,QAAQ;AAAA,QACR,OAAO;AAAA,QACP,UAAU,iBAAiB,UAAU,GAAG;AAAA,QACxC,SAAS;AAAA,UACP,aAAa,KAAK,aAAa;AAAA,UAC/B,aAAa,KAAK,eAAe;AAAA,QACnC;AAAA,QACA,GAAI,SAAS,KAAK,eAAe,SAAS,EAAE,MAAM,IAAI,CAAC;AAAA,MACzD,CAAC;AAAA,IACH;AAAA,EACF;AACA,SAAO;AAAA,IACL,KAAK,GAAG,IAAI,mBAAmB;AAAA,IAC/B;AAAA,IACA,MAAM,KAAK,UAAU;AAAA,MACnB;AAAA,MACA,YAAY,KAAK,aAAa;AAAA,MAC9B,aAAa,KAAK,eAAe;AAAA,MACjC,UAAU,iBAAiB,UAAU,GAAG;AAAA,MACxC,GAAI,QACA,EAAE,OAAO,aAAa,iBAAiB,KAAK,cAAc,MAAM,EAAE,IAClE,CAAC;AAAA,IACP,CAAC;AAAA,EACH;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;AA8BA,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;AAGA,SAAS,gBAAgB,MAOvB;AACA,QAAM,WAAW;AACjB,QAAM,aAA4B,SAAS,SAAS,cAAc,CAAC,GAChE,OAAO,CAAC,SAAS,OAAO,KAAK,UAAU,SAAS,QAAQ,EACxD,IAAI,CAAC,MAAM,WAAW;AAAA,IACrB,IAAI,QAAQ,KAAK;AAAA,IACjB,MAAM,KAAK,SAAU;AAAA,IACrB,WACE,OAAO,KAAK,UAAU,cAAc,WAChC,cAAc,KAAK,SAAS,SAAS,IACpC,KAAK,UAAU,aAAa,CAAC;AAAA,EACtC,EAAE;AACJ,MAAI;AACJ,MAAI,UAAU,SAAS,EAAG,cAAa;AAAA,WAC9B,SAAS,gBAAgB,SAAU,cAAa;AAAA,WAChD,SAAS,KAAM,cAAa;AAAA,WAC5B,SAAS,YAAa,cAAa;AAC5C,SAAO;AAAA;AAAA;AAAA,IAGL,SAAS,SAAS,SAAS,WAAW;AAAA,IACtC,OAAO,SAAS,qBAAqB;AAAA,IACrC,QAAQ,SAAS,cAAc;AAAA,IAC/B,OAAO,SAAS;AAAA,IAChB,WAAW,UAAU,SAAS,IAAI,YAAY;AAAA,IAC9C;AAAA,EACF;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,YAAY,OAAO,CAAC;AAE9F,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,QAAI,EAAE,WAAW,MAAM,EAAG,QAAO,EAAE,SAAS,EAAE,UAAU,SAAS,OAAO,EAAE,EAAE;AAC5E,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,UAAU;AAKb,YAAM,aAAa,MAAM,kBAAkB,KAAK,SAAS;AACzD,YAAM,mBAAmB,IAAI,OAAO,UAAU,MAAM,KAAK,WAAW,OAAO,WAAW,OAAO;AAC7F;AAAA,IACF;AAAA,IACA,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,IACF,KAAK;AACH,YAAM,kBAAkB,IAAI,OAAO,UAAU,MAAM,GAAG;AACtD;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,IACjE,KAAK;AACH,aAAO;AAAA,QACL,QAAQ,IAAI,UAAU,OAAO,MAAM,OAAO,gBAAgB,IAAI,IAAI,YAAY,IAAI;AAAA,QAClF;AAAA,QACA;AAAA,MACF;AAAA,EACJ;AACF;AA0BA,eAAsB,SACpB,UACA,KACA,OAAmB,CAAC,GACpB,OAAgB,CAAC,GACoB;AACrC,MAAI,SAAS,WAAW,GAAG;AACzB,UAAM,IAAIA,iBAAgB,4BAA4B;AAAA,EACxD;AACA,MAAI,CAAC,IAAI,qBAAqB;AAC5B,UAAM,IAAIA,iBAAgB,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,MAAI,QAAQ,KAAK,MAAM,MAAM,aAAa;AAI1C,MACE,IAAI,mBACJ,SAAS,UACT,CAAC,KAAK,SACN,IAAI,mBACJ;AACA,YAAQ,EAAE,SAAS,EAAE,UAAU,SAAS,OAAO,OAAO,MAAM,KAAK,GAAG,UAAU,MAAM,QAAQ;AAAA,EAC9F;AAGA,MACE,IAAI,uBACJ,SAAS,eACT,CAAC,KAAK,SACN,IAAI,mBACJ;AACA,YAAQ,EAAE,SAAS,EAAE,UAAU,SAAS,OAAO,OAAO,MAAM,UAAU,GAAG,UAAU,MAAM,QAAQ;AAAA,EACnG;AAGA,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,IAAIA;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,IAAIC,eAAc,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,IAAIA,eAAc,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,IAAIA,eAAc,4BAA4B,EAAE,UAAU,YAAY,MAAM,cAAc,CAAC;AAAA,EAC7F;AACF;AAkDA,gBAAuB,iBACrB,UACA,KACA,OAAwC,CAAC,GACG;AAC5C,MAAI,SAAS,WAAW,GAAG;AACzB,UAAM,IAAID,iBAAgB,4BAA4B;AAAA,EACxD;AACA,MAAI,CAAC,IAAI,qBAAqB;AAC5B,UAAM,IAAIA,iBAAgB,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,IAAIC,eAAc,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,IAAIA,eAAc,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,IAAIA,eAAc,oBAAoB;AAAA,QAC1C,UAAU,UAAU;AAAA,QACpB,OAAO,UAAU;AAAA,MACnB,CAAC;AAAA,IACH;AACA,UAAM,IAAIA,eAAc,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,IAAIA,eAAc,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,IAAIA,eAAc,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;AAGJ,QAAM,aAAa,oBAAI,IAAwD;AAC/E,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,gBACE,MAAM,eAAe,SAAS,cAC9B,OAAO,MAAM,UAAU,YACvB,OAAO,MAAM,cAAc,OAAO,YAClC,OAAO,MAAM,cAAc,SAAS,UACpC;AACA,yBAAW,IAAI,MAAM,OAAO,EAAE,IAAI,MAAM,cAAc,IAAI,MAAM,MAAM,cAAc,MAAM,MAAM,GAAG,CAAC;AAAA,YACtG;AACA;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,WACE,MAAM,OAAO,SAAS,sBACtB,OAAO,MAAM,MAAM,iBAAiB,YACpC,OAAO,MAAM,UAAU,UACvB;AACA,oBAAM,QAAQ,WAAW,IAAI,MAAM,KAAK;AACxC,kBAAI,MAAO,OAAM,QAAQ,MAAM,MAAM;AAAA,YACvC;AACA;AAAA,UACF,KAAK;AACH,2BAAe,MAAM,OAAO,iBAAiB;AAC7C,gBAAI,MAAM,OAAO,YAAa,oBAAmB,uBAAuB,MAAM,MAAM,WAAW;AAC/F;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,QAAM,YAA2B,CAAC,GAAG,WAAW,OAAO,CAAC,EAAE,IAAI,CAAC,OAAO;AAAA,IACpE,IAAI,EAAE;AAAA,IACN,MAAM,EAAE;AAAA,IACR,WAAW,cAAc,EAAE,IAAI;AAAA,EACjC,EAAE;AAEF,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,IACA,YAAY;AAAA,IACZ,WAAW,UAAU,SAAS,IAAI,YAAY;AAAA,EAChD;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":["InternalError","ValidationError","InternalError","ValidationError","InternalError"]}
+{"version":3,"sources":["../src/index.ts","../src/gcp-token.ts","../src/embed.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';\nimport { fetchAdcAccessToken, mintGcpAccessToken, serviceAccountProjectId } from './gcp-token.js';\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 Flash 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' | 'local';\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  /** True when a fallback provider served instead of the tier's intended primary — see #4204. */\n  degraded: boolean;\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   * Cost optimization: when true, `fast`-tier calls route to the self-hosted GPU\n   * (qwen3:8b via the `custom-local-fast-gpu` AI Gateway provider) FIRST, with the normal\n   * fast route (Grok→Haiku) as automatic fallback on any error. Off by default →\n   * fully dormant. Requires {@link LLMEnv.GPU_LLM_API_TOKEN}.\n   */\n  LLM_LOCAL_FIRST?: boolean;\n  /**\n   * Agentic-work opt-in: route the `workbench` tier to local Qwen 27B first,\n   * with DeepSeek as the independent cloud fallback.\n   */\n  LLM_LOCAL_WORKBENCH?: boolean;\n  /** Bearer for the local GPU proxy (`GPU_LLM_API_TOKEN` in Secret Manager). Required to enable local. */\n  GPU_LLM_API_TOKEN?: string;\n  /** Cloudflare Access service-token id for the local GPU app; forwarded by the gateway to the origin. */\n  GPU_LLM_ACCESS_CLIENT_ID?: string;\n  /** Cloudflare Access service-token secret for the local GPU app. */\n  GPU_LLM_ACCESS_CLIENT_SECRET?: string;\n  /**\n   * GCP service-account JSON key (base64-encoded or raw) with `roles/aiplatform.user`.\n   * The credential for the `gemini` leg in environments WITHOUT a metadata server\n   * (Cloudflare Workers): the package exchanges it for a fresh access token per\n   * isolate and caches it (see `./gcp-token.ts`).\n   *\n   * Optional. When neither this nor {@link LLMEnv.VERTEX_ACCESS_TOKEN} is set, the\n   * leg falls back to Application Default Credentials — the GCP metadata server's\n   * ambient service-account token (keyless; the default on Cloud Run / GCE). Off\n   * GCP with no static credential, the metadata probe fails and the chain falls\n   * through to its next leg.\n   */\n  GCP_SA_KEY?: string;\n  /**\n   * Pre-minted Google Cloud access token with `aiplatform.endpoints.predict`.\n   * Explicit override, used only when {@link LLMEnv.GCP_SA_KEY} is absent and\n   * before ADC is attempted. Valid for ~1 hour from minting, so it cannot be a\n   * durable Worker secret.\n   */\n  VERTEX_ACCESS_TOKEN?: string;\n  /** Vertex project. Defaults to the `GCP_SA_KEY`'s own `project_id`. */\n  VERTEX_PROJECT?: string;\n  /** Vertex location. Defaults to `us-central1`. */\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  /** True when a fallback provider served instead of the tier's intended primary — see #4204. */\n  degraded: boolean;\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 * Resolve the metering context for a call.\n *\n * `LLMOptions` carries the same four field names in TWO places: top-level\n * `project`/`actor`/`runId`/`workload` (documented for logs and cost-policy\n * call sites) and the nested {@link LLMOptions.ledger}, which is the ONLY one\n * the `onRecord` metering callback reads. Both shapes typecheck, so picking\n * the wrong one is silent — you get logging and no ledger row.\n *\n * Every caller picked the wrong one. Measured across the estate 2026-08-23:\n * **8 of 8** call sites that supply metering context set it at the top level\n * (`admin-studio` test-analyst + ai, `daily-brief` wisdom + insights,\n * `inbound-oracle`, `supervisor` reflect, `lead-gen`, `linkedin-publisher`)\n * and **0** use `ledger`. A 0% adoption rate on one of two identical-looking\n * APIs is a design defect, not eight independent mistakes — so the fix belongs\n * here rather than in eight call sites that already read as correct.\n *\n * Explicit `ledger` always wins; the top-level fields are a fallback. A call\n * needs BOTH `project` and `actor` to be meterable, because those two are the\n * required columns on the ledger row — a partial context would write a row\n * that cannot be attributed, which is worse than no row.\n */\nexport function resolveLedgerContext(opts: LLMOptions): LLMRecordContext | undefined {\n  if (opts.ledger) return opts.ledger;\n  if (!opts.project || !opts.actor) return undefined;\n  const ctx: LLMRecordContext = { project: opts.project, actor: opts.actor };\n  if (opts.runId !== undefined) ctx.runId = opts.runId;\n  if (opts.workload !== undefined) ctx.workload = opts.workload;\n  return ctx;\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    // `claude-haiku-4-20250514` was never a model Anthropic served — it 404s\n    // (verified live against /v1/messages). Date suffixes are never appended to\n    // an alias; the alias is `claude-haiku-4-5`, which resolves server-side to\n    // `claude-haiku-4-5-20251001`. This is the `fast` tier's FALLBACK, so the\n    // 404 stayed invisible for as long as the Grok primary held.\n    fast: 'claude-haiku-4-5',\n    // Current-generation Sonnet/Opus (2026-08 refresh). The previous pins —\n    // `claude-sonnet-4-6` / `claude-opus-4-7` — were a generation behind, and\n    // the opus-4-7 leg was HARD-BROKEN: Claude 4.7+ rejects `temperature`\n    // (400 \"`temperature` is deprecated for this model\", verified live\n    // 2026-08-14) and buildAnthropicRequest always sent one, so every `smart`\n    // call 400'd on its Anthropic primary and silently served Gemini Flash.\n    // Sonnet 5 / Opus 5 cost the same list price or less than the models they\n    // replace ($3/$15, $5/$25 per 1M). The sampling/thinking request-shape\n    // differences these models introduce are handled in buildAnthropicRequest.\n    balanced: 'claude-sonnet-5',\n    smart: 'claude-opus-5',\n  },\n  gemini: {\n    // `gemini-2.5-flash`, not `-pro`: the leg's job here is a fast, reliable,\n    // JSON-returning fallback. Gemini 2.5 *Pro* mandates a thinking budget of\n    // ≥128 tokens that is drawn from `maxOutputTokens` and CANNOT be disabled\n    // (thinkingBudget=0 is rejected). On the render-runner's large judge /\n    // generation prompts — and any low-`maxTokens` call (headline uses 40) —\n    // the thinking phase exhausts the whole budget and Vertex returns 200 with\n    // an empty candidate (finishReason MAX_TOKENS, no text), which the router\n    // treats as a failed leg. Flash supports `thinkingBudget: 0` (set in\n    // buildGeminiRequest), so text is always emitted. Both are Vertex-served.\n    smart: 'gemini-2.5-flash',\n  },\n  groq: {\n    // Groq deprecated `llama-3.3-70b-versatile` on 2026-08-16 — every\n    // `verifier` call 404'd `model_not_found` from then on (the tier has no\n    // fallback, so it was hard-broken; second time this tier has died to a\n    // model retirement). `openai/gpt-oss-120b` is Groq's recommended\n    // migration target and is served to our org (verified against /v1/models\n    // and a live completion, 2026-08-23).\n    verifier: 'openai/gpt-oss-120b',\n  },\n  grok: {\n    fast: 'grok-4.3',\n  },\n  deepseek: {\n    workbench: 'deepseek-chat',\n  },\n  local: {\n    // Self-hosted qwen3:8b on the dedicated fast GPU, reached through the\n    // `custom-local-fast-gpu` AI Gateway provider's native Ollama chat route.\n    fast: 'qwen3:8b',\n    workbench: 'qwen3.6:27b',\n  },\n} as const;\n\nconst DEFAULT_MAX_TOKENS = 1024;\nconst DEFAULT_TEMPERATURE = 0.7;\n/** Vertex region used when the caller does not pin one. */\nconst DEFAULT_VERTEX_LOCATION = 'us-central1';\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 4.5 — `claude-haiku-4-5` is the routed alias; the dated id\n  // is what the API echoes back in `response.model`, so both must price.\n  // (Rates corrected 2026-08: the previous $0.80/$4.00 rows UNDERSTATED the\n  // actual $1/$5 list price.)\n  'claude-haiku-4-5': { input: 1.00, output: 5.00, cacheRead: 0.10, cacheWrite: 1.25 },\n  'claude-haiku-4-5-20251001': { input: 1.00, output: 5.00, cacheRead: 0.10, cacheWrite: 1.25 },\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 Sonnet 5 — the routed `balanced` primary. List $3/$15; intro\n  // pricing ($2/$10) runs through 2026-08-31 — priced at list here, the\n  // conservative upper bound (table convention).\n  'claude-sonnet-5': { input: 3.00, output: 15.00, cacheRead: 0.30, cacheWrite: 3.75 },\n  // Anthropic Opus 4 (dated id) — genuinely the $15/$75 era.\n  'claude-opus-4-20250514': { input: 15.00, output: 75.00, cacheRead: 1.50, cacheWrite: 18.75 },\n  // Opus 4.7 was NEVER $15/$75 — it launched at $5/$25. The old row copied the\n  // Opus 4 rate and overstated every smart-tier ledger entry 3x.\n  'claude-opus-4-7': { input: 5.00, output: 25.00, cacheRead: 0.50, cacheWrite: 6.25 },\n  // Anthropic Opus 5 — the routed `smart` primary. Same $5/$25 as Opus 4.8/4.7.\n  'claude-opus-5': { input: 5.00, output: 25.00, cacheRead: 0.50, cacheWrite: 6.25 },\n  // Gemini 2.5 Flash — the routed `smart`/long-context model (JSON fallback leg).\n  'gemini-2.5-flash': { input: 0.30, output: 2.50, cacheRead: 0.075, cacheWrite: 0.30 },\n  // Gemini 2.5 Pro — retained for historical ledger rows (was the routed model).\n  'gemini-2.5-pro': { input: 1.25, output: 10.00, cacheRead: 0.31, cacheWrite: 4.50 },\n  // Groq GPT-OSS 120B (`verifier` tier). Groq bills cached input at $0.075/1M\n  // with no separate cache-write charge (implicit caching), so cacheWrite is 0.\n  'openai/gpt-oss-120b': { input: 0.15, output: 0.60, cacheRead: 0.075, cacheWrite: 0.00 },\n  // Groq Llama 3.3 70B Versatile — retained for historical ledger rows\n  // (routed `verifier` until Groq retired the model on 2026-08-16).\n  'llama-3.3-70b-versatile': { input: 0.59, output: 0.79, cacheRead: 0.00, cacheWrite: 0.00 },\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  // Self-hosted qwen3 on the GPU box — zero marginal cost (electricity aside).\n  'qwen3:8b': { input: 0.0, output: 0.0, cacheRead: 0.0, cacheWrite: 0.0 },\n  'qwen3.6:27b': { input: 0.0, output: 0.0, cacheRead: 0.0, cacheWrite: 0.0 },\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\n/**\n * Claude 4.7-and-later models reject the classic sampling params: sending\n * `temperature` (or `top_p`/`top_k`) returns 400 \"`temperature` is deprecated\n * for this model\" (verified live against /v1/messages, 2026-08-14). Models\n * matching this pattern must not receive a `temperature` field at all — this\n * is exactly what hard-broke the `smart` tier while it pinned opus-4-7.\n */\nconst ANTHROPIC_SAMPLING_REMOVED = /^claude-(opus-4-[78]|opus-5|sonnet-5|fable-5|mythos-5)/;\n\n/**\n * Models that run ADAPTIVE THINKING by default when the `thinking` field is\n * omitted. Thinking tokens draw from `max_tokens`, and this package's budgets\n * are small (DEFAULT_MAX_TOKENS 1024; the headline call passes 40) — left on,\n * the model can spend the whole budget thinking and truncate the visible\n * text. Same failure mode as Gemini 2.5 Pro's mandatory thinking budget (see\n * the MODELS.gemini comment), same fix: turn it off explicitly.\n * `{type:'disabled'}` is accepted on both models at the default effort\n * (verified live 2026-08-14). Do NOT add `claude-fable-5`/`claude-mythos-5`\n * here — those models reject an explicit `disabled` (thinking is always on).\n */\nconst ANTHROPIC_THINKING_DEFAULT_ON = /^claude-(opus-5|sonnet-5)$/;\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    messages: filtered.map((m) => ({ role: m.role, content: m.content })),\n  };\n  if (!ANTHROPIC_SAMPLING_REMOVED.test(model)) {\n    body.temperature = opts.temperature ?? DEFAULT_TEMPERATURE;\n  }\n  if (ANTHROPIC_THINKING_DEFAULT_ON.test(model)) {\n    body.thinking = { type: 'disabled' };\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\n/** A resolved Vertex credential: a bearer token plus (for ADC) the ambient project id. */\ninterface VertexAuth {\n  token: string;\n  /** Project id discovered from the metadata server (ADC path only). */\n  project?: string;\n}\n\n/**\n * Resolve a Vertex credential, in priority order:\n *   1. `GCP_SA_KEY` — mint a fresh token from the service-account key (the\n *      credential Cloudflare Workers supply, where there is no metadata server).\n *   2. `VERTEX_ACCESS_TOKEN` — a pre-minted token (legacy / CI override).\n *   3. Application Default Credentials — the GCP metadata server's ambient\n *      service-account token. This is the KEYLESS default on Cloud Run / GCE:\n *      neither env var need be set, mirroring the post-WIF-migration pattern\n *      the render-runner already uses for every other GCP API call.\n *\n * @throws ValidationError when no static credential is set AND the metadata\n * server is unreachable (i.e. off-GCP) — the leg fails before the request so\n * the router falls through to the next provider, never calling Vertex\n * unauthenticated.\n */\nasync function resolveVertexAuth(env: LLMEnv, fetchImpl: typeof fetch): Promise<VertexAuth> {\n  if (env.GCP_SA_KEY) {\n    try {\n      return { token: await mintGcpAccessToken(env.GCP_SA_KEY, fetchImpl) };\n    } catch (error) {\n      // A pre-minted token is worth trying if minting broke. With nothing to fall\n      // back to, surface the minting failure rather than swallowing it.\n      if (!env.VERTEX_ACCESS_TOKEN) throw error;\n    }\n  }\n  if (env.VERTEX_ACCESS_TOKEN) return { token: env.VERTEX_ACCESS_TOKEN };\n  // Keyless default: Application Default Credentials from the GCP metadata\n  // server. Throws (→ chain falls through) when off-GCP.\n  return fetchAdcAccessToken(fetchImpl);\n}\n\nfunction buildGeminiRequest(\n  model: string,\n  messages: LLMMessage[],\n  opts: LLMOptions,\n  env: LLMEnv,\n  accessToken: string,\n  adcProject?: string,\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      // Disable \"thinking\": Gemini 2.5 draws thinking tokens from\n      // maxOutputTokens, so on a large prompt (or a small token budget) the\n      // model can spend the entire budget thinking and return an empty\n      // candidate (finishReason MAX_TOKENS). This leg wants deterministic text\n      // out, so thinking is turned off (supported by gemini-2.5-flash).\n      thinkingConfig: { thinkingBudget: 0 },\n    },\n  };\n  if (sys) {\n    body.systemInstruction = { parts: [{ text: sys }] };\n  }\n  // Project resolution: an explicit VERTEX_PROJECT wins; else the service\n  // account's own project (GCP_SA_KEY path); else the ambient project id the\n  // metadata server reported (ADC path). So a Cloud Run job needs no Vertex\n  // config at all — the runner's own project is used.\n  const project =\n    env.VERTEX_PROJECT ||\n    (env.GCP_SA_KEY ? serviceAccountProjectId(env.GCP_SA_KEY) : (adcProject ?? ''));\n  const location = env.VERTEX_LOCATION || DEFAULT_VERTEX_LOCATION;\n  const path = `v1/projects/${project}/locations/${location}/publishers/google/models/${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 ${accessToken}`,\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/**\n * Converts normalized messages to Ollama's native chat shape. Native tool calls\n * carry object arguments (rather than OpenAI's JSON string), and tool results\n * identify the function by name rather than by call id.\n */\nfunction toOllamaMessages(messages: LLMMessage[], sys: string | undefined): Array<Record<string, unknown>> {\n  const toolNames = new Map<string, string>();\n  for (const message of messages) {\n    if (typeof message.content === 'string') continue;\n    for (const block of message.content) {\n      if (block.type === 'tool_use') toolNames.set(block.id, block.name);\n    }\n  }\n\n  const out: Array<Record<string, unknown>> = [];\n  if (sys) out.push({ role: 'system', content: sys });\n  for (const message of messages) {\n    if (message.role === 'system') continue;\n    if (typeof message.content === 'string') {\n      out.push({ role: message.role, content: message.content });\n      continue;\n    }\n\n    let text = '';\n    const toolCalls: Array<Record<string, unknown>> = [];\n    const results: Array<{ toolUseId: string; content: string }> = [];\n    for (const block of message.content) {\n      if (block.type === 'text') text += block.text;\n      else if (block.type === 'tool_use') {\n        toolCalls.push({\n          type: 'function',\n          function: { index: toolCalls.length, name: block.name, arguments: block.input },\n        });\n      } else if (block.type === 'tool_result') {\n        results.push({ toolUseId: block.tool_use_id, content: block.content });\n      }\n    }\n    if (results.length > 0) {\n      for (const result of results) {\n        out.push({\n          role: 'tool',\n          tool_name: toolNames.get(result.toolUseId) ?? result.toolUseId,\n          content: result.content,\n        });\n      }\n      if (text) out.push({ role: 'user', content: text });\n    } else if (toolCalls.length > 0) {\n      out.push({ role: 'assistant', content: text, tool_calls: toolCalls });\n    } else {\n      out.push({ role: message.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/**\n * Builds a local Qwen request behind Cloudflare AI Gateway. Qwen 8B uses the\n * dedicated fast origin's native Ollama contract so `think:false` is enforced;\n * Qwen 27B retains the OpenAI-compatible agentic/tool-call route.\n */\nfunction buildLocalRequest(\n  model: string,\n  messages: LLMMessage[],\n  opts: LLMOptions,\n  env: LLMEnv,\n): { url: string; headers: Record<string, string>; body: string } {\n  if (!env.GPU_LLM_API_TOKEN) {\n    throw new ValidationError('GPU_LLM_API_TOKEN required for the local provider');\n  }\n  const sys = systemText(opts, messages);\n  const fastMode = model === MODELS.local.fast;\n  const headers: Record<string, string> = {\n    'content-type': 'application/json',\n    authorization: `Bearer ${env.GPU_LLM_API_TOKEN}`,\n  };\n  if (env.GPU_LLM_ACCESS_CLIENT_ID) headers['CF-Access-Client-Id'] = env.GPU_LLM_ACCESS_CLIENT_ID;\n  if (env.GPU_LLM_ACCESS_CLIENT_SECRET) headers['CF-Access-Client-Secret'] = env.GPU_LLM_ACCESS_CLIENT_SECRET;\n  const tools = openAiTools(opts);\n  if (fastMode) {\n    if (typeof opts.toolChoice === 'object') {\n      throw new ValidationError('named tool choice is not supported by the native Qwen 8B route');\n    }\n    return {\n      url: `${env.AI_GATEWAY_BASE_URL}/custom-local-fast-gpu/api/chat`,\n      headers,\n      body: JSON.stringify({\n        model,\n        stream: false,\n        think: false,\n        messages: toOllamaMessages(messages, sys),\n        options: {\n          num_predict: opts.maxTokens ?? DEFAULT_MAX_TOKENS,\n          temperature: opts.temperature ?? DEFAULT_TEMPERATURE,\n        },\n        ...(tools && opts.toolChoice !== 'none' ? { tools } : {}),\n      }),\n    };\n  }\n  return {\n    url: `${env.AI_GATEWAY_BASE_URL}/custom-local-gpu/v1/chat/completions`,\n    headers,\n    body: JSON.stringify({\n      model,\n      // Ollama enables Qwen thinking by default.  Without this OpenAI-\n      // compatible control, short bounded calls can spend every token in the\n      // hidden `reasoning` field and return HTTP 200 with empty visible text.\n      reasoning_effort: 'none',\n      max_tokens: opts.maxTokens ?? DEFAULT_MAX_TOKENS,\n      temperature: opts.temperature ?? DEFAULT_TEMPERATURE,\n      messages: toOpenAiMessages(messages, sys),\n      ...(tools\n        ? { tools, tool_choice: openAiToolChoice(opts.toolChoice ?? 'auto') }\n        : {}),\n    }),\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\ninterface OllamaChatResponse {\n  model?: string;\n  message?: {\n    content?: string;\n    thinking?: string;\n    tool_calls?: Array<{\n      function?: { name?: string; arguments?: Record<string, unknown> | string };\n    }>;\n  };\n  done?: boolean;\n  done_reason?: string;\n  prompt_eval_count?: number;\n  eval_count?: number;\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/** Parses Ollama's native non-streaming `/api/chat` response for Qwen 8B. */\nfunction parseOllamaChat(json: unknown): {\n  content: string;\n  input: number;\n  output: number;\n  model?: string;\n  toolCalls?: LLMToolCall[];\n  stopReason?: LLMResult['stopReason'];\n} {\n  const response = json as OllamaChatResponse;\n  const toolCalls: LLMToolCall[] = (response.message?.tool_calls ?? [])\n    .filter((call) => typeof call.function?.name === 'string')\n    .map((call, index) => ({\n      id: `call_${index}`,\n      name: call.function!.name!,\n      arguments:\n        typeof call.function?.arguments === 'string'\n          ? parseToolArgs(call.function.arguments)\n          : (call.function?.arguments ?? {}),\n    }));\n  let stopReason: LLMResult['stopReason'];\n  if (toolCalls.length > 0) stopReason = 'tool_use';\n  else if (response.done_reason === 'length') stopReason = 'max_tokens';\n  else if (response.done) stopReason = 'end';\n  else if (response.done_reason) stopReason = 'other';\n  return {\n    // Deliberately exclude message.thinking: callers receive only the bounded\n    // classification answer even if a non-conforming origin returns a trace.\n    content: response.message?.content ?? '',\n    input: response.prompt_eval_count ?? 0,\n    output: response.eval_count ?? 0,\n    model: response.model,\n    toolCalls: toolCalls.length > 0 ? toolCalls : undefined,\n    stopReason,\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', 'local']);\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    if (m.startsWith('qwen')) return { primary: { provider: 'local', 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      // Credential resolution is a network call (mint from key, or the metadata\n      // server for ADC), so it must happen here rather than inside the\n      // (synchronous) request builders. Cached per isolate, so it is almost\n      // always a no-op after the first gemini call.\n      const vertexAuth = await resolveVertexAuth(env, fetchImpl);\n      req = buildGeminiRequest(leg.model, messages, opts, env, vertexAuth.token, vertexAuth.project);\n      break;\n    }\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    case 'local':\n      req = buildLocalRequest(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    case 'local':\n      return {\n        parsed: leg.model === MODELS.local.fast ? parseOllamaChat(json) : parseOpenAi(json),\n        gatewayRequestId,\n        attempts,\n      };\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 Flash if `longContextThreshold` exceeded\n *   - `smart`    → Anthropic Opus; Gemini 2.5 Flash 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  let route = plan(tier, opts, tokenEstimate);\n  // Cost optimization: prefer the dedicated Qwen 8B GPU ($0) for cheap `fast`-tier\n  // work, with the original fast route (Grok→Haiku) as automatic fallback on any\n  // error. Dormant unless LLM_LOCAL_FIRST + GPU_LLM_API_TOKEN are set.\n  if (\n    env.LLM_LOCAL_FIRST &&\n    tier === 'fast' &&\n    !opts.model &&\n    env.GPU_LLM_API_TOKEN\n  ) {\n    route = { primary: { provider: 'local', model: MODELS.local.fast }, fallback: route.primary };\n  }\n  // Qwen 27B owns local agentic work. DeepSeek remains the independent fallback;\n  // explicit model overrides always win.\n  if (\n    env.LLM_LOCAL_WORKBENCH &&\n    tier === 'workbench' &&\n    !opts.model &&\n    env.GPU_LLM_API_TOKEN\n  ) {\n    route = { primary: { provider: 'local', model: MODELS.local.workbench }, fallback: route.primary };\n  }\n\n  // ── Org-level daily / monthly cap pre-check ──────────────────────────────\n  const kv = env.LLM_COST_KV;\n  const todayKey = `llm:daily-cost:${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        // `false` means NO caller named a tier, so this call took the\n        // package default (`balanced` — claude-sonnet-5, the second most\n        // expensive route). Emitted so a defaulted tier is greppable rather\n        // than invisible: an expensive route chosen by silence is the thing\n        // that made the video pipeline bill Sonnet while its docs said Haiku.\n        tierExplicit: opts.tier !== undefined,\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      // The tier's route is [primary, fallback] in order (see `routeLegs` above) — a\n      // fallback served iff this leg is anything past index 0. This is the\n      // \"servedBy != intended\" signal from #4204: cheap to compute here, and\n      // distinct from `attempts` (which conflates retries on the SAME provider\n      // with actually falling back to a DIFFERENT one).\n      const degraded = legIndex > 0;\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        degraded,\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      const ledgerCtx = resolveLedgerContext(opts);\n      if (deps.onRecord && ledgerCtx) {\n        const row: LLMRecordRow = {\n          ...ledgerCtx,\n          model: llmResult.model,\n          provider: llmResult.provider,\n          tier: llmResult.tier,\n          degraded: llmResult.degraded,\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; partial_json?: string; stop_reason?: string };\n  content_block?: { type?: string; id?: string; name?: 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  // Tool-use blocks arrive as content_block_start (id/name) then a sequence of\n  // input_json_delta fragments that concatenate into the arguments JSON string.\n  const toolBlocks = new Map<number, { id: string; name: string; json: string }>();\n  let streamStopReason: LLMResult['stopReason'];\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_start':\n            if (\n              event.content_block?.type === 'tool_use' &&\n              typeof event.index === 'number' &&\n              typeof event.content_block.id === 'string' &&\n              typeof event.content_block.name === 'string'\n            ) {\n              toolBlocks.set(event.index, { id: event.content_block.id, name: event.content_block.name, json: '' });\n            }\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            } else if (\n              event.delta?.type === 'input_json_delta' &&\n              typeof event.delta.partial_json === 'string' &&\n              typeof event.index === 'number'\n            ) {\n              const block = toolBlocks.get(event.index);\n              if (block) block.json += event.delta.partial_json;\n            }\n            break;\n          case 'message_delta':\n            outputTokens = event.usage?.output_tokens ?? outputTokens;\n            if (event.delta?.stop_reason) streamStopReason = normalizeAnthropicStop(event.delta.stop_reason);\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        // `false` means NO caller named a tier, so this call took the\n        // package default (`balanced` — claude-sonnet-5, the second most\n        // expensive route). Emitted so a defaulted tier is greppable rather\n        // than invisible: an expensive route chosen by silence is the thing\n        // that made the video pipeline bill Sonnet while its docs said Haiku.\n        tierExplicit: opts.tier !== undefined,\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  const toolCalls: LLMToolCall[] = [...toolBlocks.values()].map((b) => ({\n    id: b.id,\n    name: b.name,\n    arguments: parseToolArgs(b.json),\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    stopReason: streamStopReason,\n    toolCalls: toolCalls.length > 0 ? toolCalls : undefined,\n    // streamLeg is only ever route.fallback in the grok-no-key special case above —\n    // same \"servedBy != intended\" signal as the non-streaming path, see #4204.\n    degraded: streamLeg !== route.primary,\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\n// ─── Embeddings ────────────────────────────────────────────────────────────\n\nexport { embed, embedLocal, DEFAULT_EMBEDDING_MODEL, LOCAL_EMBEDDING_MODEL } from './embed.js';\nexport type { AiBinding, EmbedResult, EmbeddingModel, LocalEmbedEnv } from './embed.js';\n\nexport { mintGcpAccessToken, clearGcpTokenCache, serviceAccountProjectId } from './gcp-token.js';\n","/**\n * Mint short-lived Google Cloud OAuth2 access tokens from a service-account key,\n * inside a Cloudflare Worker.\n *\n * The `gemini` leg authenticates to Vertex AI with a GCP access token, which\n * expires after one hour. A token therefore cannot be a durable Worker secret —\n * any stored value is stale almost immediately. Callers supply the long-lived\n * service-account key instead, and this module exchanges it for a fresh token\n * via the JWT-bearer flow, caching the result per isolate.\n *\n * Web Crypto only — no Node.js built-ins, no `Buffer`. Lifted from\n * `apps/admin-studio/src/lib/gcp-secrets.ts`, where this exact flow has been\n * running in production; hoisted here so every consumer of the `gemini` leg\n * gets a working credential rather than just the one app that hand-rolled it.\n */\nimport { InternalError, ValidationError } from '@latimer-woods-tech/errors';\n\n/** The subset of a GCP service-account JSON key needed for the JWT-bearer flow. */\ninterface ServiceAccountKey {\n  project_id: string;\n  private_key: string;\n  client_email: string;\n  token_uri: string;\n}\n\ninterface OAuth2Response {\n  access_token: string;\n  expires_in: number;\n}\n\n/** Refresh this long before expiry so an in-flight request never races the clock. */\nconst EXPIRY_SKEW_MS = 300_000;\n\n/**\n * Per-isolate token cache, keyed by service-account email.\n *\n * Caching a *string* across requests is safe. Caching an I/O object (a socket, a\n * database client, a `Response`) is not — Workers rejects reuse of I/O created\n * on behalf of a different request. This holds only the token text and its\n * expiry, so it is reused freely.\n */\nconst tokenCache = new Map<string, { token: string; expiresAt: number }>();\n\n/** Decode a base64url-encoded service-account key, tolerating raw JSON too. */\nfunction parseServiceAccountKey(gcpSaKey: string): ServiceAccountKey {\n  const text = gcpSaKey.trimStart().startsWith('{') ? gcpSaKey : atob(gcpSaKey);\n  const key = JSON.parse(text) as ServiceAccountKey;\n  if (!key.client_email || !key.private_key || !key.token_uri) {\n    throw new ValidationError('GCP_SA_KEY is missing client_email, private_key, or token_uri');\n  }\n  return key;\n}\n\n/** Encode bytes as base64url (RFC 4648 §5): `-`/`_` for `+`/`/`, no padding. */\nfunction base64UrlEncode(bytes: Uint8Array): string {\n  let binary = '';\n  for (let i = 0; i < bytes.length; i++) binary += String.fromCharCode(bytes[i]!);\n  return btoa(binary).replace(/\\+/g, '-').replace(/\\//g, '_').replace(/=+$/g, '');\n}\n\n/** Import a PEM-encoded PKCS#8 private key as a Web Crypto signing key. */\nasync function importPrivateKey(pem: string): Promise<CryptoKey> {\n  const body = pem\n    .split('\\n')\n    .filter((line) => !line.startsWith('-----'))\n    .join('');\n  const binary = atob(body);\n  const bytes = new Uint8Array(binary.length);\n  for (let i = 0; i < binary.length; i++) bytes[i] = binary.charCodeAt(i);\n  return crypto.subtle.importKey(\n    'pkcs8',\n    bytes.buffer,\n    { name: 'RSASSA-PKCS1-v1_5', hash: 'SHA-256' },\n    false,\n    ['sign'],\n  );\n}\n\n/** Build a service-account JWT assertion signed with RS256, per GCP's spec. */\nasync function createAssertion(key: ServiceAccountKey, nowSeconds: number): Promise<string> {\n  const encoder = new TextEncoder();\n  const header = base64UrlEncode(encoder.encode(JSON.stringify({ alg: 'RS256', typ: 'JWT' })));\n  const payload = base64UrlEncode(\n    encoder.encode(\n      JSON.stringify({\n        iss: key.client_email,\n        scope: 'https://www.googleapis.com/auth/cloud-platform',\n        aud: key.token_uri,\n        exp: nowSeconds + 3600,\n        iat: nowSeconds,\n      }),\n    ),\n  );\n  const signingInput = `${header}.${payload}`;\n  const cryptoKey = await importPrivateKey(key.private_key);\n  const signature = await crypto.subtle.sign(\n    'RSASSA-PKCS1-v1_5',\n    cryptoKey,\n    encoder.encode(signingInput).buffer,\n  );\n  return `${signingInput}.${base64UrlEncode(new Uint8Array(signature))}`;\n}\n\n/**\n * Exchange a service-account key for a cloud-platform access token, reusing a\n * cached token until it is within {@link EXPIRY_SKEW_MS} of expiring.\n *\n * @param gcpSaKey - Service-account JSON key, base64-encoded or raw.\n * @param fetchImpl - Fetch implementation (injectable for tests).\n * @returns A bearer token valid for Vertex AI.\n * @throws ValidationError when the key is malformed; InternalError when the exchange fails.\n */\nexport async function mintGcpAccessToken(gcpSaKey: string, fetchImpl: typeof fetch): Promise<string> {\n  const key = parseServiceAccountKey(gcpSaKey);\n\n  const now = Date.now();\n  const cached = tokenCache.get(key.client_email);\n  if (cached && cached.expiresAt - EXPIRY_SKEW_MS > now) return cached.token;\n\n  const assertion = await createAssertion(key, Math.floor(now / 1000));\n  const response = await fetchImpl(key.token_uri, {\n    method: 'POST',\n    headers: { 'content-type': 'application/x-www-form-urlencoded' },\n    body: new URLSearchParams({\n      grant_type: 'urn:ietf:params:oauth:grant-type:jwt-bearer',\n      assertion,\n    }).toString(),\n  }).catch((cause: unknown) => {\n    throw new InternalError(`GCP token exchange failed: ${String(cause)}`);\n  });\n\n  if (!response.ok) {\n    throw new InternalError(`GCP token exchange returned ${response.status}`);\n  }\n\n  const data = (await response.json()) as OAuth2Response;\n  if (!data.access_token) {\n    throw new InternalError('GCP token exchange returned no access_token');\n  }\n\n  // Trust the server's TTL rather than assuming the 3600s default.\n  const ttlMs = (data.expires_in > 0 ? data.expires_in : 3600) * 1000;\n  tokenCache.set(key.client_email, { token: data.access_token, expiresAt: now + ttlMs });\n  return data.access_token;\n}\n\n// ─── Application Default Credentials (GCP metadata server) ───────────────────\n/**\n * GCE / Cloud Run metadata server base. Reachable only from GCP compute; the\n * hostname does not resolve elsewhere, so an off-GCP probe fails fast.\n */\nconst METADATA_BASE = 'http://metadata.google.internal/computeMetadata/v1';\n/** Required header for every metadata-server request (anti-SSRF guard). */\nconst METADATA_HEADER = { 'Metadata-Flavor': 'Google' } as const;\n/**\n * Bound the metadata probe so a non-GCP environment (where the host may hang\n * rather than refuse) falls through to the next provider quickly.\n */\nconst METADATA_TIMEOUT_MS = 2_000;\n\n/** Per-isolate cache for the ADC (metadata-server) credential. */\nlet adcCache: { token: string; project?: string; expiresAt: number } | undefined;\n\ninterface AdcCredential {\n  token: string;\n  /** The instance's own GCP project id, as reported by the metadata server. */\n  project?: string;\n}\n\n/**\n * Obtain an access token (and project id) from the GCP metadata server using\n * Application Default Credentials — the keyless credential every Cloud Run /\n * GCE workload already has via its attached service account. No key, no env\n * var: the instance SA's cloud-platform token is served by the metadata\n * server. This mirrors the render-runner's own `_meta_token` shell helper.\n *\n * Fails (throws {@link InternalError}) when the metadata server is unreachable\n * or returns no token — i.e. off-GCP — so the caller falls through to the next\n * provider leg. Never throws a bare `AbortError` (which the router treats as a\n * hard cancellation): an internal timeout is surfaced as an `InternalError`.\n *\n * @param fetchImpl - Fetch implementation (injectable for tests).\n * @returns A bearer token valid for Vertex AI plus the ambient project id.\n */\nexport async function fetchAdcAccessToken(fetchImpl: typeof fetch): Promise<AdcCredential> {\n  const now = Date.now();\n  if (adcCache && adcCache.expiresAt - EXPIRY_SKEW_MS > now) {\n    return { token: adcCache.token, project: adcCache.project };\n  }\n\n  const controller = new AbortController();\n  const timer = setTimeout(() => controller.abort(), METADATA_TIMEOUT_MS);\n  try {\n    const tokenRes = await fetchImpl(\n      `${METADATA_BASE}/instance/service-accounts/default/token`,\n      { headers: METADATA_HEADER, signal: controller.signal },\n    );\n    if (!tokenRes.ok) {\n      throw new InternalError(`GCP metadata token endpoint returned ${tokenRes.status}`);\n    }\n    const data = (await tokenRes.json()) as OAuth2Response;\n    if (!data.access_token) {\n      throw new InternalError('GCP metadata token endpoint returned no access_token');\n    }\n\n    // Project id is best-effort: a caller may pin VERTEX_PROJECT instead.\n    let project: string | undefined;\n    try {\n      const projRes = await fetchImpl(`${METADATA_BASE}/project/project-id`, {\n        headers: METADATA_HEADER,\n        signal: controller.signal,\n      });\n      if (projRes.ok) project = (await projRes.text()).trim() || undefined;\n    } catch {\n      /* project id is optional — the token is still usable */\n    }\n\n    const ttlMs = (data.expires_in > 0 ? data.expires_in : 3600) * 1000;\n    adcCache = { token: data.access_token, project, expiresAt: now + ttlMs };\n    return { token: data.access_token, project };\n  } catch (cause) {\n    // Normalize an internal-timeout AbortError into an InternalError so the\n    // router falls through to the next leg instead of hard-cancelling the call.\n    if (cause instanceof InternalError) throw cause;\n    throw new InternalError(`GCP metadata credential unavailable: ${String(cause)}`);\n  } finally {\n    clearTimeout(timer);\n  }\n}\n\n/** Reset the per-isolate token cache. Tests only. */\nexport function clearGcpTokenCache(): void {\n  tokenCache.clear();\n  adcCache = undefined;\n}\n\n/** The project the service account belongs to — the Vertex project by default. */\nexport function serviceAccountProjectId(gcpSaKey: string): string {\n  return parseServiceAccountKey(gcpSaKey).project_id;\n}\n","/**\n * Workers AI embedding API.\n *\n * Pins to bge-base-en-v1.5 (768-dim cosine) — the platform standard.\n * @remarks Local-rail errors use @latimer-woods-tech/errors (observability ratchet).\n * model_version is returned with every result so callers can record it\n * for provenance-aware re-embed when the model is swapped.\n *\n * Contract: inject the Workers AI binding (env.AI) — never import a vendor SDK.\n */\n\nimport { InternalError } from '@latimer-woods-tech/errors';\n\n/** The platform-standard embedding model (768-dim cosine) used when no override is given. */\nexport const DEFAULT_EMBEDDING_MODEL = '@cf/baai/bge-base-en-v1.5';\n\n/** Supported embedding model identifiers — widen deliberately, since dims are part of the contract. */\nexport type EmbeddingModel = '@cf/baai/bge-base-en-v1.5';\n\n/** One embedding call's output: the vectors plus the provenance needed to re-embed later. */\nexport interface EmbedResult {\n  vectors: number[][];\n  model: string;\n  dims: number;\n}\n\n/** Minimal Workers AI binding shape needed for embeddings. */\nexport interface AiBinding {\n  run(\n    model: string,\n    inputs: { text: string | string[] },\n  ): Promise<{ data: number[][] }>;\n}\n\n/**\n * Embeds one or more text strings using the Workers AI binding.\n *\n * @param ai    - Workers AI binding (`env.AI`) — injected, not imported.\n * @param input - Single string or array of strings to embed.\n * @param opts  - Optional model override (must be a supported EmbeddingModel).\n * @returns Embedding vectors, model name, and dimension count.\n *\n * @throws When the AI binding call fails (let the caller decide whether to catch).\n */\nexport async function embed(\n  ai: AiBinding,\n  input: string | string[],\n  opts?: { model?: EmbeddingModel },\n): Promise<EmbedResult> {\n  const model = opts?.model ?? DEFAULT_EMBEDDING_MODEL;\n  const texts = Array.isArray(input) ? input : [input];\n  const result = await ai.run(model, { text: texts });\n  const vectors = result.data;\n  if (!vectors || vectors.length === 0) {\n    throw new Error(`embed(): Workers AI returned no vectors for model ${model}`);\n  }\n  return {\n    vectors,\n    model,\n    dims: vectors[0]!.length,\n  };\n}\n\n/**\n * Self-hosted GPU embedding model (nomic-embed-text, 768-dim), reached through\n * the same CF AI Gateway custom provider as the local chat provider.\n *\n * ⚠️ VECTOR-SPACE WARNING: nomic-embed-text and bge-base-en-v1.5 are BOTH 768-dim\n * but occupy DIFFERENT embedding spaces. Their vectors are NOT comparable. A store\n * built with one model must be QUERIED — and, when migrating, RE-EMBEDDED — with\n * the SAME model. Never write vectors from both models into one store.\n */\nexport const LOCAL_EMBEDDING_MODEL = 'nomic-embed-text';\n\n/** Env subset needed to reach the local embedding rail (shares GPU_LLM_* with the chat provider). */\nexport interface LocalEmbedEnv {\n  AI_GATEWAY_BASE_URL: string;\n  GPU_LLM_API_TOKEN?: string;\n  GPU_LLM_ACCESS_CLIENT_ID?: string;\n  GPU_LLM_ACCESS_CLIENT_SECRET?: string;\n}\n\n/**\n * Embeds text on the self-hosted GPU rail (nomic-embed-text) via the AI Gateway\n * custom provider — the $0 alternative to Workers AI {@link embed}.\n *\n * Unlike the local CHAT provider, this has NO cross-provider fallback ON PURPOSE:\n * falling back to Workers AI (bge) would silently write an incompatible-space\n * vector into a nomic store and corrupt retrieval. If the rail is unavailable\n * this THROWS and the caller decides (retry local, or defer the write) — it must\n * NOT substitute a different model.\n *\n * @param env   - Gateway base URL + GPU_LLM_* auth (same secrets as the chat provider).\n * @param input - Single string or array of strings to embed.\n * @returns Embedding vectors, model `nomic-embed-text`, and dimension count.\n * @throws  On any transport/HTTP/empty-response error (no silent fallback).\n */\nexport async function embedLocal(\n  env: LocalEmbedEnv,\n  input: string | string[],\n): Promise<EmbedResult> {\n  const texts = Array.isArray(input) ? input : [input];\n  const res = await fetch(\n    `${env.AI_GATEWAY_BASE_URL}/custom-local-gpu/v1/embeddings`,\n    {\n      method: 'POST',\n      headers: {\n        'Content-Type': 'application/json',\n        // CF Access service-token headers only sent when present (bearer-only rails still work).\n        ...(env.GPU_LLM_API_TOKEN ? { Authorization: `Bearer ${env.GPU_LLM_API_TOKEN}` } : {}),\n        ...(env.GPU_LLM_ACCESS_CLIENT_ID ? { 'CF-Access-Client-Id': env.GPU_LLM_ACCESS_CLIENT_ID } : {}),\n        ...(env.GPU_LLM_ACCESS_CLIENT_SECRET ? { 'CF-Access-Client-Secret': env.GPU_LLM_ACCESS_CLIENT_SECRET } : {}),\n      },\n      body: JSON.stringify({ model: LOCAL_EMBEDDING_MODEL, input: texts }),\n    },\n  );\n  if (!res.ok) {\n    throw new InternalError(\n      `embedLocal(): rail ${res.status}: ${(await res.text()).slice(0, 160)}`,\n    );\n  }\n  const json = (await res.json()) as { data?: Array<{ embedding: number[] }> };\n  const vectors = (json.data ?? []).map((d) => d.embedding);\n  if (vectors.length === 0) {\n    throw new InternalError(\n      `embedLocal(): rail returned no vectors for model ${LOCAL_EMBEDDING_MODEL}`,\n    );\n  }\n  return {\n    vectors,\n    model: LOCAL_EMBEDDING_MODEL,\n    dims: vectors[0]!.length,\n  };\n}\n"],"mappings":";AAAA;AAAA,EACE,iBAAAA;AAAA,EACA;AAAA,EACA,mBAAAC;AAAA,EACA;AAAA,OAEK;;;ACSP,SAAS,eAAe,uBAAuB;AAgB/C,IAAM,iBAAiB;AAUvB,IAAM,aAAa,oBAAI,IAAkD;AAGzE,SAAS,uBAAuB,UAAqC;AACnE,QAAM,OAAO,SAAS,UAAU,EAAE,WAAW,GAAG,IAAI,WAAW,KAAK,QAAQ;AAC5E,QAAM,MAAM,KAAK,MAAM,IAAI;AAC3B,MAAI,CAAC,IAAI,gBAAgB,CAAC,IAAI,eAAe,CAAC,IAAI,WAAW;AAC3D,UAAM,IAAI,gBAAgB,+DAA+D;AAAA,EAC3F;AACA,SAAO;AACT;AAGA,SAAS,gBAAgB,OAA2B;AAClD,MAAI,SAAS;AACb,WAAS,IAAI,GAAG,IAAI,MAAM,QAAQ,IAAK,WAAU,OAAO,aAAa,MAAM,CAAC,CAAE;AAC9E,SAAO,KAAK,MAAM,EAAE,QAAQ,OAAO,GAAG,EAAE,QAAQ,OAAO,GAAG,EAAE,QAAQ,QAAQ,EAAE;AAChF;AAGA,eAAe,iBAAiB,KAAiC;AAC/D,QAAM,OAAO,IACV,MAAM,IAAI,EACV,OAAO,CAAC,SAAS,CAAC,KAAK,WAAW,OAAO,CAAC,EAC1C,KAAK,EAAE;AACV,QAAM,SAAS,KAAK,IAAI;AACxB,QAAM,QAAQ,IAAI,WAAW,OAAO,MAAM;AAC1C,WAAS,IAAI,GAAG,IAAI,OAAO,QAAQ,IAAK,OAAM,CAAC,IAAI,OAAO,WAAW,CAAC;AACtE,SAAO,OAAO,OAAO;AAAA,IACnB;AAAA,IACA,MAAM;AAAA,IACN,EAAE,MAAM,qBAAqB,MAAM,UAAU;AAAA,IAC7C;AAAA,IACA,CAAC,MAAM;AAAA,EACT;AACF;AAGA,eAAe,gBAAgB,KAAwB,YAAqC;AAC1F,QAAM,UAAU,IAAI,YAAY;AAChC,QAAM,SAAS,gBAAgB,QAAQ,OAAO,KAAK,UAAU,EAAE,KAAK,SAAS,KAAK,MAAM,CAAC,CAAC,CAAC;AAC3F,QAAM,UAAU;AAAA,IACd,QAAQ;AAAA,MACN,KAAK,UAAU;AAAA,QACb,KAAK,IAAI;AAAA,QACT,OAAO;AAAA,QACP,KAAK,IAAI;AAAA,QACT,KAAK,aAAa;AAAA,QAClB,KAAK;AAAA,MACP,CAAC;AAAA,IACH;AAAA,EACF;AACA,QAAM,eAAe,GAAG,MAAM,IAAI,OAAO;AACzC,QAAM,YAAY,MAAM,iBAAiB,IAAI,WAAW;AACxD,QAAM,YAAY,MAAM,OAAO,OAAO;AAAA,IACpC;AAAA,IACA;AAAA,IACA,QAAQ,OAAO,YAAY,EAAE;AAAA,EAC/B;AACA,SAAO,GAAG,YAAY,IAAI,gBAAgB,IAAI,WAAW,SAAS,CAAC,CAAC;AACtE;AAWA,eAAsB,mBAAmB,UAAkB,WAA0C;AACnG,QAAM,MAAM,uBAAuB,QAAQ;AAE3C,QAAM,MAAM,KAAK,IAAI;AACrB,QAAM,SAAS,WAAW,IAAI,IAAI,YAAY;AAC9C,MAAI,UAAU,OAAO,YAAY,iBAAiB,IAAK,QAAO,OAAO;AAErE,QAAM,YAAY,MAAM,gBAAgB,KAAK,KAAK,MAAM,MAAM,GAAI,CAAC;AACnE,QAAM,WAAW,MAAM,UAAU,IAAI,WAAW;AAAA,IAC9C,QAAQ;AAAA,IACR,SAAS,EAAE,gBAAgB,oCAAoC;AAAA,IAC/D,MAAM,IAAI,gBAAgB;AAAA,MACxB,YAAY;AAAA,MACZ;AAAA,IACF,CAAC,EAAE,SAAS;AAAA,EACd,CAAC,EAAE,MAAM,CAAC,UAAmB;AAC3B,UAAM,IAAI,cAAc,8BAA8B,OAAO,KAAK,CAAC,EAAE;AAAA,EACvE,CAAC;AAED,MAAI,CAAC,SAAS,IAAI;AAChB,UAAM,IAAI,cAAc,+BAA+B,SAAS,MAAM,EAAE;AAAA,EAC1E;AAEA,QAAM,OAAQ,MAAM,SAAS,KAAK;AAClC,MAAI,CAAC,KAAK,cAAc;AACtB,UAAM,IAAI,cAAc,6CAA6C;AAAA,EACvE;AAGA,QAAM,SAAS,KAAK,aAAa,IAAI,KAAK,aAAa,QAAQ;AAC/D,aAAW,IAAI,IAAI,cAAc,EAAE,OAAO,KAAK,cAAc,WAAW,MAAM,MAAM,CAAC;AACrF,SAAO,KAAK;AACd;AAOA,IAAM,gBAAgB;AAEtB,IAAM,kBAAkB,EAAE,mBAAmB,SAAS;AAKtD,IAAM,sBAAsB;AAG5B,IAAI;AAuBJ,eAAsB,oBAAoB,WAAiD;AACzF,QAAM,MAAM,KAAK,IAAI;AACrB,MAAI,YAAY,SAAS,YAAY,iBAAiB,KAAK;AACzD,WAAO,EAAE,OAAO,SAAS,OAAO,SAAS,SAAS,QAAQ;AAAA,EAC5D;AAEA,QAAM,aAAa,IAAI,gBAAgB;AACvC,QAAM,QAAQ,WAAW,MAAM,WAAW,MAAM,GAAG,mBAAmB;AACtE,MAAI;AACF,UAAM,WAAW,MAAM;AAAA,MACrB,GAAG,aAAa;AAAA,MAChB,EAAE,SAAS,iBAAiB,QAAQ,WAAW,OAAO;AAAA,IACxD;AACA,QAAI,CAAC,SAAS,IAAI;AAChB,YAAM,IAAI,cAAc,wCAAwC,SAAS,MAAM,EAAE;AAAA,IACnF;AACA,UAAM,OAAQ,MAAM,SAAS,KAAK;AAClC,QAAI,CAAC,KAAK,cAAc;AACtB,YAAM,IAAI,cAAc,sDAAsD;AAAA,IAChF;AAGA,QAAI;AACJ,QAAI;AACF,YAAM,UAAU,MAAM,UAAU,GAAG,aAAa,uBAAuB;AAAA,QACrE,SAAS;AAAA,QACT,QAAQ,WAAW;AAAA,MACrB,CAAC;AACD,UAAI,QAAQ,GAAI,YAAW,MAAM,QAAQ,KAAK,GAAG,KAAK,KAAK;AAAA,IAC7D,QAAQ;AAAA,IAER;AAEA,UAAM,SAAS,KAAK,aAAa,IAAI,KAAK,aAAa,QAAQ;AAC/D,eAAW,EAAE,OAAO,KAAK,cAAc,SAAS,WAAW,MAAM,MAAM;AACvE,WAAO,EAAE,OAAO,KAAK,cAAc,QAAQ;AAAA,EAC7C,SAAS,OAAO;AAGd,QAAI,iBAAiB,cAAe,OAAM;AAC1C,UAAM,IAAI,cAAc,wCAAwC,OAAO,KAAK,CAAC,EAAE;AAAA,EACjF,UAAE;AACA,iBAAa,KAAK;AAAA,EACpB;AACF;AAGO,SAAS,qBAA2B;AACzC,aAAW,MAAM;AACjB,aAAW;AACb;AAGO,SAAS,wBAAwB,UAA0B;AAChE,SAAO,uBAAuB,QAAQ,EAAE;AAC1C;;;ACpOA,SAAS,iBAAAC,sBAAqB;AAGvB,IAAM,0BAA0B;AA8BvC,eAAsB,MACpB,IACA,OACA,MACsB;AACtB,QAAM,QAAQ,MAAM,SAAS;AAC7B,QAAM,QAAQ,MAAM,QAAQ,KAAK,IAAI,QAAQ,CAAC,KAAK;AACnD,QAAM,SAAS,MAAM,GAAG,IAAI,OAAO,EAAE,MAAM,MAAM,CAAC;AAClD,QAAM,UAAU,OAAO;AACvB,MAAI,CAAC,WAAW,QAAQ,WAAW,GAAG;AACpC,UAAM,IAAI,MAAM,qDAAqD,KAAK,EAAE;AAAA,EAC9E;AACA,SAAO;AAAA,IACL;AAAA,IACA;AAAA,IACA,MAAM,QAAQ,CAAC,EAAG;AAAA,EACpB;AACF;AAWO,IAAM,wBAAwB;AAyBrC,eAAsB,WACpB,KACA,OACsB;AACtB,QAAM,QAAQ,MAAM,QAAQ,KAAK,IAAI,QAAQ,CAAC,KAAK;AACnD,QAAM,MAAM,MAAM;AAAA,IAChB,GAAG,IAAI,mBAAmB;AAAA,IAC1B;AAAA,MACE,QAAQ;AAAA,MACR,SAAS;AAAA,QACP,gBAAgB;AAAA;AAAA,QAEhB,GAAI,IAAI,oBAAoB,EAAE,eAAe,UAAU,IAAI,iBAAiB,GAAG,IAAI,CAAC;AAAA,QACpF,GAAI,IAAI,2BAA2B,EAAE,uBAAuB,IAAI,yBAAyB,IAAI,CAAC;AAAA,QAC9F,GAAI,IAAI,+BAA+B,EAAE,2BAA2B,IAAI,6BAA6B,IAAI,CAAC;AAAA,MAC5G;AAAA,MACA,MAAM,KAAK,UAAU,EAAE,OAAO,uBAAuB,OAAO,MAAM,CAAC;AAAA,IACrE;AAAA,EACF;AACA,MAAI,CAAC,IAAI,IAAI;AACX,UAAM,IAAIA;AAAA,MACR,sBAAsB,IAAI,MAAM,MAAM,MAAM,IAAI,KAAK,GAAG,MAAM,GAAG,GAAG,CAAC;AAAA,IACvE;AAAA,EACF;AACA,QAAM,OAAQ,MAAM,IAAI,KAAK;AAC7B,QAAM,WAAW,KAAK,QAAQ,CAAC,GAAG,IAAI,CAAC,MAAM,EAAE,SAAS;AACxD,MAAI,QAAQ,WAAW,GAAG;AACxB,UAAM,IAAIA;AAAA,MACR,oDAAoD,qBAAqB;AAAA,IAC3E;AAAA,EACF;AACA,SAAO;AAAA,IACL;AAAA,IACA,OAAO;AAAA,IACP,MAAM,QAAQ,CAAC,EAAG;AAAA,EACpB;AACF;;;AFzEA,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;AAoPO,SAAS,qBAAqB,MAAgD;AACnF,MAAI,KAAK,OAAQ,QAAO,KAAK;AAC7B,MAAI,CAAC,KAAK,WAAW,CAAC,KAAK,MAAO,QAAO;AACzC,QAAM,MAAwB,EAAE,SAAS,KAAK,SAAS,OAAO,KAAK,MAAM;AACzE,MAAI,KAAK,UAAU,OAAW,KAAI,QAAQ,KAAK;AAC/C,MAAI,KAAK,aAAa,OAAW,KAAI,WAAW,KAAK;AACrD,SAAO;AACT;AAkBA,IAAM,SAAS;AAAA,EACb,WAAW;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAMT,MAAM;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAUN,UAAU;AAAA,IACV,OAAO;AAAA,EACT;AAAA,EACA,QAAQ;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAUN,OAAO;AAAA,EACT;AAAA,EACA,MAAM;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,IAOJ,UAAU;AAAA,EACZ;AAAA,EACA,MAAM;AAAA,IACJ,MAAM;AAAA,EACR;AAAA,EACA,UAAU;AAAA,IACR,WAAW;AAAA,EACb;AAAA,EACA,OAAO;AAAA;AAAA;AAAA,IAGL,MAAM;AAAA,IACN,WAAW;AAAA,EACb;AACF;AAEA,IAAM,qBAAqB;AAC3B,IAAM,sBAAsB;AAE5B,IAAM,0BAA0B;AAChC,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;AAAA;AAAA;AAAA,EAK1H,oBAAoB,EAAE,OAAO,GAAM,QAAQ,GAAM,WAAW,KAAM,YAAY,KAAK;AAAA,EACnF,6BAA6B,EAAE,OAAO,GAAM,QAAQ,GAAM,WAAW,KAAM,YAAY,KAAK;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;AAAA;AAAA,EAIrF,mBAAmB,EAAE,OAAO,GAAM,QAAQ,IAAO,WAAW,KAAM,YAAY,KAAK;AAAA;AAAA,EAEnF,0BAA0B,EAAE,OAAO,IAAO,QAAQ,IAAO,WAAW,KAAM,YAAY,MAAM;AAAA;AAAA;AAAA,EAG5F,mBAAmB,EAAE,OAAO,GAAM,QAAQ,IAAO,WAAW,KAAM,YAAY,KAAK;AAAA;AAAA,EAEnF,iBAAiB,EAAE,OAAO,GAAM,QAAQ,IAAO,WAAW,KAAM,YAAY,KAAK;AAAA;AAAA,EAEjF,oBAAoB,EAAE,OAAO,KAAM,QAAQ,KAAM,WAAW,OAAO,YAAY,IAAK;AAAA;AAAA,EAEpF,kBAAkB,EAAE,OAAO,MAAM,QAAQ,IAAO,WAAW,MAAM,YAAY,IAAK;AAAA;AAAA;AAAA,EAGlF,uBAAuB,EAAE,OAAO,MAAM,QAAQ,KAAM,WAAW,OAAO,YAAY,EAAK;AAAA;AAAA;AAAA,EAGvF,2BAA2B,EAAE,OAAO,MAAM,QAAQ,MAAM,WAAW,GAAM,YAAY,EAAK;AAAA;AAAA,EAE1F,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;AAAA;AAAA,EAErF,YAAY,EAAE,OAAO,GAAK,QAAQ,GAAK,WAAW,GAAK,YAAY,EAAI;AAAA,EACvE,eAAe,EAAE,OAAO,GAAK,QAAQ,GAAK,WAAW,GAAK,YAAY,EAAI;AAC5E;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;AAWA,IAAM,6BAA6B;AAanC,IAAM,gCAAgC;AAEtC,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,UAAU,SAAS,IAAI,CAAC,OAAO,EAAE,MAAM,EAAE,MAAM,SAAS,EAAE,QAAQ,EAAE;AAAA,EACtE;AACA,MAAI,CAAC,2BAA2B,KAAK,KAAK,GAAG;AAC3C,SAAK,cAAc,KAAK,eAAe;AAAA,EACzC;AACA,MAAI,8BAA8B,KAAK,KAAK,GAAG;AAC7C,SAAK,WAAW,EAAE,MAAM,WAAW;AAAA,EACrC;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;AAwBA,eAAe,kBAAkB,KAAa,WAA8C;AAC1F,MAAI,IAAI,YAAY;AAClB,QAAI;AACF,aAAO,EAAE,OAAO,MAAM,mBAAmB,IAAI,YAAY,SAAS,EAAE;AAAA,IACtE,SAAS,OAAO;AAGd,UAAI,CAAC,IAAI,oBAAqB,OAAM;AAAA,IACtC;AAAA,EACF;AACA,MAAI,IAAI,oBAAqB,QAAO,EAAE,OAAO,IAAI,oBAAoB;AAGrE,SAAO,oBAAoB,SAAS;AACtC;AAEA,SAAS,mBACP,OACA,UACA,MACA,KACA,aACA,YACgE;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;AAAA;AAAA;AAAA;AAAA;AAAA,MAMjC,gBAAgB,EAAE,gBAAgB,EAAE;AAAA,IACtC;AAAA,EACF;AACA,MAAI,KAAK;AACP,SAAK,oBAAoB,EAAE,OAAO,CAAC,EAAE,MAAM,IAAI,CAAC,EAAE;AAAA,EACpD;AAKA,QAAM,UACJ,IAAI,mBACH,IAAI,aAAa,wBAAwB,IAAI,UAAU,IAAK,cAAc;AAC7E,QAAM,WAAW,IAAI,mBAAmB;AACxC,QAAM,OAAO,eAAe,OAAO,cAAc,QAAQ,6BAA6B,KAAK;AAC3F,SAAO;AAAA,IACL,KAAK,GAAG,IAAI,mBAAmB,qBAAqB,IAAI;AAAA,IACxD,SAAS;AAAA,MACP,gBAAgB;AAAA,MAChB,eAAe,UAAU,WAAW;AAAA,IACtC;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;AAOA,SAAS,iBAAiB,UAAwB,KAAyD;AACzG,QAAM,YAAY,oBAAI,IAAoB;AAC1C,aAAW,WAAW,UAAU;AAC9B,QAAI,OAAO,QAAQ,YAAY,SAAU;AACzC,eAAW,SAAS,QAAQ,SAAS;AACnC,UAAI,MAAM,SAAS,WAAY,WAAU,IAAI,MAAM,IAAI,MAAM,IAAI;AAAA,IACnE;AAAA,EACF;AAEA,QAAM,MAAsC,CAAC;AAC7C,MAAI,IAAK,KAAI,KAAK,EAAE,MAAM,UAAU,SAAS,IAAI,CAAC;AAClD,aAAW,WAAW,UAAU;AAC9B,QAAI,QAAQ,SAAS,SAAU;AAC/B,QAAI,OAAO,QAAQ,YAAY,UAAU;AACvC,UAAI,KAAK,EAAE,MAAM,QAAQ,MAAM,SAAS,QAAQ,QAAQ,CAAC;AACzD;AAAA,IACF;AAEA,QAAI,OAAO;AACX,UAAM,YAA4C,CAAC;AACnD,UAAM,UAAyD,CAAC;AAChE,eAAW,SAAS,QAAQ,SAAS;AACnC,UAAI,MAAM,SAAS,OAAQ,SAAQ,MAAM;AAAA,eAChC,MAAM,SAAS,YAAY;AAClC,kBAAU,KAAK;AAAA,UACb,MAAM;AAAA,UACN,UAAU,EAAE,OAAO,UAAU,QAAQ,MAAM,MAAM,MAAM,WAAW,MAAM,MAAM;AAAA,QAChF,CAAC;AAAA,MACH,WAAW,MAAM,SAAS,eAAe;AACvC,gBAAQ,KAAK,EAAE,WAAW,MAAM,aAAa,SAAS,MAAM,QAAQ,CAAC;AAAA,MACvE;AAAA,IACF;AACA,QAAI,QAAQ,SAAS,GAAG;AACtB,iBAAW,UAAU,SAAS;AAC5B,YAAI,KAAK;AAAA,UACP,MAAM;AAAA,UACN,WAAW,UAAU,IAAI,OAAO,SAAS,KAAK,OAAO;AAAA,UACrD,SAAS,OAAO;AAAA,QAClB,CAAC;AAAA,MACH;AACA,UAAI,KAAM,KAAI,KAAK,EAAE,MAAM,QAAQ,SAAS,KAAK,CAAC;AAAA,IACpD,WAAW,UAAU,SAAS,GAAG;AAC/B,UAAI,KAAK,EAAE,MAAM,aAAa,SAAS,MAAM,YAAY,UAAU,CAAC;AAAA,IACtE,OAAO;AACL,UAAI,KAAK,EAAE,MAAM,QAAQ,MAAM,SAAS,KAAK,CAAC;AAAA,IAChD;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,IAAIC,iBAAgB,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,IAAIA,iBAAgB,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;AAOA,SAAS,kBACP,OACA,UACA,MACA,KACgE;AAChE,MAAI,CAAC,IAAI,mBAAmB;AAC1B,UAAM,IAAIA,iBAAgB,mDAAmD;AAAA,EAC/E;AACA,QAAM,MAAM,WAAW,MAAM,QAAQ;AACrC,QAAM,WAAW,UAAU,OAAO,MAAM;AACxC,QAAM,UAAkC;AAAA,IACtC,gBAAgB;AAAA,IAChB,eAAe,UAAU,IAAI,iBAAiB;AAAA,EAChD;AACA,MAAI,IAAI,yBAA0B,SAAQ,qBAAqB,IAAI,IAAI;AACvE,MAAI,IAAI,6BAA8B,SAAQ,yBAAyB,IAAI,IAAI;AAC/E,QAAM,QAAQ,YAAY,IAAI;AAC9B,MAAI,UAAU;AACZ,QAAI,OAAO,KAAK,eAAe,UAAU;AACvC,YAAM,IAAIA,iBAAgB,gEAAgE;AAAA,IAC5F;AACA,WAAO;AAAA,MACL,KAAK,GAAG,IAAI,mBAAmB;AAAA,MAC/B;AAAA,MACA,MAAM,KAAK,UAAU;AAAA,QACnB;AAAA,QACA,QAAQ;AAAA,QACR,OAAO;AAAA,QACP,UAAU,iBAAiB,UAAU,GAAG;AAAA,QACxC,SAAS;AAAA,UACP,aAAa,KAAK,aAAa;AAAA,UAC/B,aAAa,KAAK,eAAe;AAAA,QACnC;AAAA,QACA,GAAI,SAAS,KAAK,eAAe,SAAS,EAAE,MAAM,IAAI,CAAC;AAAA,MACzD,CAAC;AAAA,IACH;AAAA,EACF;AACA,SAAO;AAAA,IACL,KAAK,GAAG,IAAI,mBAAmB;AAAA,IAC/B;AAAA,IACA,MAAM,KAAK,UAAU;AAAA,MACnB;AAAA;AAAA;AAAA;AAAA,MAIA,kBAAkB;AAAA,MAClB,YAAY,KAAK,aAAa;AAAA,MAC9B,aAAa,KAAK,eAAe;AAAA,MACjC,UAAU,iBAAiB,UAAU,GAAG;AAAA,MACxC,GAAI,QACA,EAAE,OAAO,aAAa,iBAAiB,KAAK,cAAc,MAAM,EAAE,IAClE,CAAC;AAAA,IACP,CAAC;AAAA,EACH;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;AA8BA,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;AAGA,SAAS,gBAAgB,MAOvB;AACA,QAAM,WAAW;AACjB,QAAM,aAA4B,SAAS,SAAS,cAAc,CAAC,GAChE,OAAO,CAAC,SAAS,OAAO,KAAK,UAAU,SAAS,QAAQ,EACxD,IAAI,CAAC,MAAM,WAAW;AAAA,IACrB,IAAI,QAAQ,KAAK;AAAA,IACjB,MAAM,KAAK,SAAU;AAAA,IACrB,WACE,OAAO,KAAK,UAAU,cAAc,WAChC,cAAc,KAAK,SAAS,SAAS,IACpC,KAAK,UAAU,aAAa,CAAC;AAAA,EACtC,EAAE;AACJ,MAAI;AACJ,MAAI,UAAU,SAAS,EAAG,cAAa;AAAA,WAC9B,SAAS,gBAAgB,SAAU,cAAa;AAAA,WAChD,SAAS,KAAM,cAAa;AAAA,WAC5B,SAAS,YAAa,cAAa;AAC5C,SAAO;AAAA;AAAA;AAAA,IAGL,SAAS,SAAS,SAAS,WAAW;AAAA,IACtC,OAAO,SAAS,qBAAqB;AAAA,IACrC,QAAQ,SAAS,cAAc;AAAA,IAC/B,OAAO,SAAS;AAAA,IAChB,WAAW,UAAU,SAAS,IAAI,YAAY;AAAA,IAC9C;AAAA,EACF;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,YAAY,OAAO,CAAC;AAE9F,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,QAAI,EAAE,WAAW,MAAM,EAAG,QAAO,EAAE,SAAS,EAAE,UAAU,SAAS,OAAO,EAAE,EAAE;AAC5E,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,UAAU;AAKb,YAAM,aAAa,MAAM,kBAAkB,KAAK,SAAS;AACzD,YAAM,mBAAmB,IAAI,OAAO,UAAU,MAAM,KAAK,WAAW,OAAO,WAAW,OAAO;AAC7F;AAAA,IACF;AAAA,IACA,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,IACF,KAAK;AACH,YAAM,kBAAkB,IAAI,OAAO,UAAU,MAAM,GAAG;AACtD;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,IACjE,KAAK;AACH,aAAO;AAAA,QACL,QAAQ,IAAI,UAAU,OAAO,MAAM,OAAO,gBAAgB,IAAI,IAAI,YAAY,IAAI;AAAA,QAClF;AAAA,QACA;AAAA,MACF;AAAA,EACJ;AACF;AA0BA,eAAsB,SACpB,UACA,KACA,OAAmB,CAAC,GACpB,OAAgB,CAAC,GACoB;AACrC,MAAI,SAAS,WAAW,GAAG;AACzB,UAAM,IAAIA,iBAAgB,4BAA4B;AAAA,EACxD;AACA,MAAI,CAAC,IAAI,qBAAqB;AAC5B,UAAM,IAAIA,iBAAgB,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,MAAI,QAAQ,KAAK,MAAM,MAAM,aAAa;AAI1C,MACE,IAAI,mBACJ,SAAS,UACT,CAAC,KAAK,SACN,IAAI,mBACJ;AACA,YAAQ,EAAE,SAAS,EAAE,UAAU,SAAS,OAAO,OAAO,MAAM,KAAK,GAAG,UAAU,MAAM,QAAQ;AAAA,EAC9F;AAGA,MACE,IAAI,uBACJ,SAAS,eACT,CAAC,KAAK,SACN,IAAI,mBACJ;AACA,YAAQ,EAAE,SAAS,EAAE,UAAU,SAAS,OAAO,OAAO,MAAM,UAAU,GAAG,UAAU,MAAM,QAAQ;AAAA,EACnG;AAGA,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,IAAIA;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,IAAIC,eAAc,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;AAAA;AAAA;AAAA;AAAA;AAAA,QAM7B,cAAc,KAAK,SAAS;AAAA,QAC5B,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;AAMD,YAAM,WAAW,WAAW;AAC5B,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,QACzB;AAAA,MACF;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,YAAM,YAAY,qBAAqB,IAAI;AAC3C,UAAI,KAAK,YAAY,WAAW;AAC9B,cAAM,MAAoB;AAAA,UACxB,GAAG;AAAA,UACH,OAAO,UAAU;AAAA,UACjB,UAAU,UAAU;AAAA,UACpB,MAAM,UAAU;AAAA,UAChB,UAAU,UAAU;AAAA,UACpB,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,IAAIA,eAAc,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,IAAIA,eAAc,4BAA4B,EAAE,UAAU,YAAY,MAAM,cAAc,CAAC;AAAA,EAC7F;AACF;AAkDA,gBAAuB,iBACrB,UACA,KACA,OAAwC,CAAC,GACG;AAC5C,MAAI,SAAS,WAAW,GAAG;AACzB,UAAM,IAAID,iBAAgB,4BAA4B;AAAA,EACxD;AACA,MAAI,CAAC,IAAI,qBAAqB;AAC5B,UAAM,IAAIA,iBAAgB,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,IAAIC,eAAc,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,IAAIA,eAAc,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,IAAIA,eAAc,oBAAoB;AAAA,QAC1C,UAAU,UAAU;AAAA,QACpB,OAAO,UAAU;AAAA,MACnB,CAAC;AAAA,IACH;AACA,UAAM,IAAIA,eAAc,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,IAAIA,eAAc,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,IAAIA,eAAc,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;AAGJ,QAAM,aAAa,oBAAI,IAAwD;AAC/E,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,gBACE,MAAM,eAAe,SAAS,cAC9B,OAAO,MAAM,UAAU,YACvB,OAAO,MAAM,cAAc,OAAO,YAClC,OAAO,MAAM,cAAc,SAAS,UACpC;AACA,yBAAW,IAAI,MAAM,OAAO,EAAE,IAAI,MAAM,cAAc,IAAI,MAAM,MAAM,cAAc,MAAM,MAAM,GAAG,CAAC;AAAA,YACtG;AACA;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,WACE,MAAM,OAAO,SAAS,sBACtB,OAAO,MAAM,MAAM,iBAAiB,YACpC,OAAO,MAAM,UAAU,UACvB;AACA,oBAAM,QAAQ,WAAW,IAAI,MAAM,KAAK;AACxC,kBAAI,MAAO,OAAM,QAAQ,MAAM,MAAM;AAAA,YACvC;AACA;AAAA,UACF,KAAK;AACH,2BAAe,MAAM,OAAO,iBAAiB;AAC7C,gBAAI,MAAM,OAAO,YAAa,oBAAmB,uBAAuB,MAAM,MAAM,WAAW;AAC/F;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;AAAA;AAAA;AAAA;AAAA;AAAA,IAMjC,cAAc,KAAK,SAAS;AAAA,IAChC,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,QAAM,YAA2B,CAAC,GAAG,WAAW,OAAO,CAAC,EAAE,IAAI,CAAC,OAAO;AAAA,IACpE,IAAI,EAAE;AAAA,IACN,MAAM,EAAE;AAAA,IACR,WAAW,cAAc,EAAE,IAAI;AAAA,EACjC,EAAE;AAEF,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,IACA,YAAY;AAAA,IACZ,WAAW,UAAU,SAAS,IAAI,YAAY;AAAA;AAAA;AAAA,IAG9C,UAAU,cAAc,MAAM;AAAA,EAChC;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":["InternalError","ValidationError","InternalError","ValidationError","InternalError"]}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@latimer-woods-tech/llm",
-  "version": "0.6.1",
+  "version": "0.8.1",
   "license": "MIT",
   "private": false,
   "repository": {
@@ -37,12 +37,13 @@
   "devDependencies": {
     "@latimer-woods-tech/monitoring": "^0.2.0",
     "@types/node": "^25.6.0",
-    "typescript": "^5.4.0",
-    "tsup": "^8.1.0",
-    "vitest": "^3.2.4",
+    "@typescript-eslint/eslint-plugin": "^7.0.0",
+    "@typescript-eslint/parser": "^7.0.0",
     "@vitest/coverage-v8": "^3.2.4",
     "eslint": "^8.57.0",
-    "@typescript-eslint/eslint-plugin": "^7.0.0",
-    "@typescript-eslint/parser": "^7.0.0"
+    "fast-check": "^3.23.2",
+    "tsup": "^8.1.0",
+    "typescript": "^5.4.0",
+    "vitest": "^3.2.4"
   }
 }