@thanh01.pmt/domain-kit 0.6.0 → 0.6.1

package/dist/utils/llmClient.cjs

@@ -0,0 +1,344 @@
+'use strict';
+
+// src/utils/llmTransportResult.ts
+function transportResultFrom(provider, model, options) {
+  return {
+    provider,
+    model,
+    framingStripped: !!options?.framingStripped,
+    upstreamError: options?.upstreamError,
+    usage: options?.usage
+  };
+}
+
+// src/utils/llmClient.ts
+function summarizeProvenance(callLog) {
+  const providers = [];
+  const models = [];
+  let failed = 0;
+  for (const r of callLog) {
+    if (!providers.includes(r.provider)) providers.push(r.provider);
+    if (!models.includes(r.model)) models.push(r.model);
+    if (!r.ok) failed++;
+  }
+  return { providers, models, calls: callLog.length, failed_calls: failed, degraded: failed > 0 };
+}
+async function responseJsonCompat(response) {
+  const text = await response.text();
+  try {
+    return JSON.parse(text);
+  } catch {
+    const start = text.indexOf("{");
+    if (start < 0) {
+      throw new LlmClientError(`LLM API returned a non-JSON body: ${text.slice(0, 200)}`);
+    }
+    let depth = 0;
+    let inString = false;
+    let escape = false;
+    for (let i = start; i < text.length; i++) {
+      const ch = text[i];
+      if (escape) {
+        escape = false;
+        continue;
+      }
+      if (ch === "\\") {
+        escape = true;
+        continue;
+      }
+      if (ch === '"') {
+        inString = !inString;
+        continue;
+      }
+      if (inString) continue;
+      if (ch === "{") depth++;
+      if (ch === "}") {
+        depth--;
+        if (depth === 0) {
+          return JSON.parse(text.slice(start, i + 1));
+        }
+      }
+    }
+    throw new LlmClientError(`LLM API returned an unterminated JSON body: ${text.slice(0, 200)}`);
+  }
+}
+var LlmClientError = class extends Error {
+  constructor(message, status, cause) {
+    super(message);
+    this.status = status;
+    this.cause = cause;
+    this.name = "LlmClientError";
+  }
+};
+function build9routerProvider(model) {
+  const apiKey = process.env.NINEROUTER_API_KEY;
+  if (!apiKey) return null;
+  const baseUrl = (process.env.NINEROUTER_BASE_URL || "https://ai-router.orchable.app/v1").replace(/^["']|["']$/g, "").trim();
+  const label = "9router";
+  const modelName = (model || process.env.NINEROUTER_MODEL || "laguna-s-2.1").trim();
+  return { label, baseUrl, apiKey, model: modelName };
+}
+function resolve9routerChain(configModel) {
+  const primary = build9routerProvider(configModel);
+  if (!primary) return [];
+  const alt = process.env.NINEROUTER_MODEL_ALT && process.env.NINEROUTER_MODEL_ALT.trim();
+  if (!alt) return [primary];
+  const secondary = build9routerProvider(alt);
+  if (!secondary) return [primary];
+  return primary.model === secondary.model ? [primary] : [primary, secondary];
+}
+var MIN_REQUEST_INTERVAL_MS = 2500;
+var requestIntervalMs = Number(process.env.LLM_MIN_REQUEST_INTERVAL_MS || MIN_REQUEST_INTERVAL_MS);
+var lastRequestAt = 0;
+async function pacedDelay() {
+  const wait = lastRequestAt + requestIntervalMs - Date.now();
+  if (wait > 0) await new Promise((resolve) => setTimeout(resolve, wait));
+  lastRequestAt = Date.now();
+}
+function createLlmClient(config) {
+  const chain = resolve9routerChain(config?.model);
+  const primary = chain[0];
+  if (!primary) {
+    throw new LlmClientError(
+      "No 9router provider configured. Set NINEROUTER_API_KEY and NINEROUTER_MODEL (or the NINEROUTER_* env defaults expected by the kit)."
+    );
+  }
+  primary.apiKey;
+  primary.baseUrl;
+  const model = primary.model;
+  if (config?.minRequestIntervalMs !== void 0) {
+    requestIntervalMs = Math.max(0, config.minRequestIntervalMs);
+  }
+  const callLog = [];
+  async function chat(messages, options) {
+    const temperature = options?.temperature ?? config?.temperature ?? 0.1;
+    const maxTokens = options?.maxTokens ?? config?.maxTokens ?? parseInt(process.env.LLM_MAX_TOKENS || "65536", 10);
+    const attempts = [];
+    let lastError = null;
+    const TRANSIENT = /* @__PURE__ */ new Set([408, 429, 500, 502, 503, 504]);
+    const REQUEST_TIMEOUT_MS = Number(process.env.LLM_REQUEST_TIMEOUT_MS || 3e5);
+    const MAX_CHAIN_ROUNDS = 3;
+    for (let round = 1; round <= MAX_CHAIN_ROUNDS; round++) {
+      if (round > 1) {
+        attempts.push("round " + (round - 1) + " failed \u2014 backing off 20s before rewalking the chain");
+        await new Promise((resolve) => setTimeout(resolve, 2e4));
+      }
+      for (const provider of chain) {
+        const recIdx = callLog.push({ provider: provider.label, model: provider.model, ok: false }) - 1;
+        try {
+          const headers = {
+            "Content-Type": "application/json",
+            "Authorization": `Bearer ${provider.apiKey}`
+          };
+          const payload = {
+            model: provider.model,
+            messages,
+            temperature,
+            max_tokens: maxTokens
+          };
+          await pacedDelay();
+          const doFetch = () => fetch(`${provider.baseUrl}/chat/completions`, {
+            method: "POST",
+            headers: {
+              "Content-Type": "application/json",
+              "Authorization": `Bearer ${provider.apiKey}`
+            },
+            body: JSON.stringify({
+              model: provider.model,
+              messages,
+              temperature,
+              max_tokens: maxTokens
+            }),
+            signal: AbortSignal.timeout(REQUEST_TIMEOUT_MS)
+          });
+          let response;
+          try {
+            response = await doFetch();
+          } catch (netErr) {
+            attempts.push(`${provider.label}: network error (${netErr instanceof Error ? netErr.message : String(netErr)}) \u2014 retrying once`);
+            await new Promise((r) => setTimeout(r, 5e3));
+            lastRequestAt = Date.now();
+            response = await doFetch();
+          }
+          if (!response.ok) {
+            const body = await response.text().catch(() => "");
+            const err = new LlmClientError(
+              `[${provider.label}] LLM API error: ${response.status} ${response.statusText} \u2014 ${body.slice(0, 200)}`,
+              response.status
+            );
+            if (response.status === 429) {
+              const retryAfterRaw = response.headers.get("retry-after");
+              const retryAfterMs = Math.min(
+                6e4,
+                Math.max(15e3, (Number.isFinite(Number(retryAfterRaw)) ? Number(retryAfterRaw) : 20) * 1e3)
+              );
+              attempts.push(`${provider.label}: 429 rate-limited \u2014 backing off ${Math.round(retryAfterMs / 1e3)}s`);
+              await new Promise((resolve) => setTimeout(resolve, retryAfterMs));
+              lastRequestAt = Date.now();
+              const retry = await fetch(`${provider.baseUrl}/chat/completions`, {
+                method: "POST",
+                headers,
+                body: JSON.stringify(payload),
+                signal: AbortSignal.timeout(REQUEST_TIMEOUT_MS)
+              });
+              if (!retry.ok) throw err;
+              const retryData = await responseJsonCompat(retry);
+              const retryContent2 = retryData.choices?.[0]?.message?.content || "";
+              if (!retryContent2.trim()) throw err;
+              callLog[recIdx] = { ...callLog[recIdx], ok: true };
+              const transport2 = transportResultFrom(provider.label, retryData.model || provider.model, {
+                usage: retryData.usage,
+                framingStripped: true
+              });
+              return {
+                content: retryContent2,
+                model: retryData.model || provider.model,
+                provider: provider.label,
+                attempts: attempts.slice(),
+                usage: retryData.usage,
+                finishReason: retryData.choices?.[0]?.finish_reason,
+                _transport: transport2
+              };
+            }
+            if (TRANSIENT.has(response.status)) {
+              await new Promise((r) => setTimeout(r, 3e3));
+              const retry = await fetch(`${provider.baseUrl}/chat/completions`, {
+                signal: AbortSignal.timeout(REQUEST_TIMEOUT_MS),
+                method: "POST",
+                headers: {
+                  "Content-Type": "application/json",
+                  "Authorization": `Bearer ${provider.apiKey}`
+                },
+                body: JSON.stringify({
+                  model: provider.model,
+                  messages,
+                  temperature,
+                  max_tokens: maxTokens
+                })
+              });
+              if (!retry.ok) throw err;
+              const retryData = await responseJsonCompat(retry);
+              const retryContent = retryData.choices?.[0]?.message?.content || "";
+              callLog[recIdx] = { ...callLog[recIdx], ok: true };
+              const transport2 = transportResultFrom(provider.label, retryData.model || provider.model, {
+                usage: retryData.usage,
+                framingStripped: true
+              });
+              return {
+                content: retryContent,
+                model: retryData.model || provider.model,
+                provider: provider.label,
+                attempts: attempts.slice(),
+                usage: retryData.usage,
+                finishReason: retryData.choices?.[0]?.finish_reason,
+                _transport: transport2
+              };
+            }
+            throw err;
+          }
+          const data = await responseJsonCompat(response);
+          const wrappedError = data.error;
+          if (wrappedError) {
+            throw new LlmClientError("[" + provider.label + "] upstream error " + (wrappedError.code ?? "") + ": " + (wrappedError.message ?? "unknown"));
+          }
+          const content = data.choices?.[0]?.message?.content || "";
+          if (!content.trim()) {
+            throw new LlmClientError("[" + provider.label + "] empty completion returned");
+          }
+          const transport = transportResultFrom(provider.label, data.model || provider.model, {
+            usage: data.usage,
+            framingStripped: true
+          });
+          return {
+            content,
+            model: data.model || provider.model,
+            provider: provider.label,
+            attempts: attempts.slice(),
+            usage: data.usage,
+            finishReason: data.choices?.[0]?.finish_reason,
+            _transport: transport
+          };
+        } catch (err) {
+          const msg = err instanceof Error ? err.message : String(err);
+          callLog[recIdx] = { ...callLog[recIdx], ok: false, error: msg.slice(0, 200) };
+          attempts.push(`${provider.label}: ${msg}`);
+          lastError = err instanceof LlmClientError ? err : new LlmClientError(`[${provider.label}] ${msg}`);
+        }
+      }
+      if (round < MAX_CHAIN_ROUNDS) continue;
+      throw lastError ?? new LlmClientError("All LLM providers failed (empty chain).");
+    }
+    throw lastError ?? new LlmClientError("All LLM providers failed after " + MAX_CHAIN_ROUNDS + " rounds.");
+  }
+  return { chat, model, callLog };
+}
+function parseJsonFromLlm(text) {
+  let cleaned = text.trim();
+  if (cleaned.startsWith("```json")) {
+    cleaned = cleaned.slice(7);
+  } else if (cleaned.startsWith("```")) {
+    cleaned = cleaned.slice(3);
+  }
+  if (cleaned.endsWith("```")) {
+    cleaned = cleaned.slice(0, -3);
+  }
+  cleaned = cleaned.trim();
+  try {
+    return JSON.parse(cleaned);
+  } catch {
+    const start = cleaned.indexOf("{");
+    if (start >= 0) {
+      let depth = 0;
+      let inString = false;
+      let escape = false;
+      for (let i = start; i < cleaned.length; i++) {
+        const ch = cleaned[i];
+        if (escape) {
+          escape = false;
+          continue;
+        }
+        if (ch === "\\") {
+          escape = true;
+          continue;
+        }
+        if (ch === '"') {
+          inString = !inString;
+          continue;
+        }
+        if (inString) continue;
+        if (ch === "{") depth++;
+        if (ch === "}") {
+          depth--;
+          if (depth === 0) {
+            try {
+              return JSON.parse(cleaned.slice(start, i + 1));
+            } catch {
+            }
+          }
+        }
+      }
+    }
+    throw new LlmClientError(`Failed to parse JSON from LLM response: ${cleaned.slice(0, 200)}`);
+  }
+}
+async function llmChatJson(client, systemPrompt, userPrompt, options) {
+  const result = await client.chat(
+    [
+      { role: "system", content: systemPrompt },
+      { role: "user", content: userPrompt }
+    ],
+    options
+  );
+  return parseJsonFromLlm(result.content);
+}
+
+exports.LlmClientError = LlmClientError;
+exports.build9routerProvider = build9routerProvider;
+exports.createLlmClient = createLlmClient;
+exports.llmChatJson = llmChatJson;
+exports.parseJsonFromLlm = parseJsonFromLlm;
+exports.resolve9routerChain = resolve9routerChain;
+exports.responseJsonCompat = responseJsonCompat;
+exports.summarizeProvenance = summarizeProvenance;
+exports.transportResultFrom = transportResultFrom;
+//# sourceMappingURL=llmClient.cjs.map
+//# sourceMappingURL=llmClient.cjs.map

package/dist/utils/llmClient.cjs.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../src/utils/llmTransportResult.ts","../../src/utils/llmClient.ts"],"names":["transport"],"mappings":";;;AA8BO,SAAS,mBAAA,CACd,QAAA,EACA,KAAA,EACA,OAAA,EAKoB;AACpB,EAAA,OAAO;AAAA,IACL,QAAA;AAAA,IACA,KAAA;AAAA,IACA,eAAA,EAAiB,CAAC,CAAC,OAAA,EAAS,eAAA;AAAA,IAC5B,eAAe,OAAA,EAAS,aAAA;AAAA,IACxB,OAAO,OAAA,EAAS;AAAA,GAClB;AACF;;;ACFO,SAAS,oBAAoB,OAAA,EAAyC;AAC3E,EAAA,MAAM,YAAsB,EAAC;AAC7B,EAAA,MAAM,SAAmB,EAAC;AAC1B,EAAA,IAAI,MAAA,GAAS,CAAA;AACb,EAAA,KAAA,MAAW,KAAK,OAAA,EAAS;AACvB,IAAA,IAAI,CAAC,UAAU,QAAA,CAAS,CAAA,CAAE,QAAQ,CAAA,EAAG,SAAA,CAAU,IAAA,CAAK,CAAA,CAAE,QAAQ,CAAA;AAC9D,IAAA,IAAI,CAAC,OAAO,QAAA,CAAS,CAAA,CAAE,KAAK,CAAA,EAAG,MAAA,CAAO,IAAA,CAAK,CAAA,CAAE,KAAK,CAAA;AAClD,IAAA,IAAI,CAAC,EAAE,EAAA,EAAI,MAAA,EAAA;AAAA,EACb;AACA,EAAA,OAAO,EAAE,SAAA,EAAW,MAAA,EAAQ,KAAA,EAAO,OAAA,CAAQ,QAAQ,YAAA,EAAc,MAAA,EAAQ,QAAA,EAAU,MAAA,GAAS,CAAA,EAAE;AAChG;AASA,eAAsB,mBAAmB,QAAA,EAAkD;AACzF,EAAA,MAAM,IAAA,GAAO,MAAM,QAAA,CAAS,IAAA,EAAK;AACjC,EAAA,IAAI;AACF,IAAA,OAAO,IAAA,CAAK,MAAM,IAAI,CAAA;AAAA,EACxB,CAAA,CAAA,MAAQ;AACN,IAAA,MAAM,KAAA,GAAQ,IAAA,CAAK,OAAA,CAAQ,GAAG,CAAA;AAC9B,IAAA,IAAI,QAAQ,CAAA,EAAG;AACb,MAAA,MAAM,IAAI,eAAe,CAAA,kCAAA,EAAqC,IAAA,CAAK,MAAM,CAAA,EAAG,GAAG,CAAC,CAAA,CAAE,CAAA;AAAA,IACpF;AACA,IAAA,IAAI,KAAA,GAAQ,CAAA;AACZ,IAAA,IAAI,QAAA,GAAW,KAAA;AACf,IAAA,IAAI,MAAA,GAAS,KAAA;AACb,IAAA,KAAA,IAAS,CAAA,GAAI,KAAA,EAAO,CAAA,GAAI,IAAA,CAAK,QAAQ,CAAA,EAAA,EAAK;AACxC,MAAA,MAAM,EAAA,GAAK,KAAK,CAAC,CAAA;AACjB,MAAA,IAAI,MAAA,EAAQ;AAAE,QAAA,MAAA,GAAS,KAAA;AAAO,QAAA;AAAA,MAAU;AACxC,MAAA,IAAI,OAAO,IAAA,EAAM;AAAE,QAAA,MAAA,GAAS,IAAA;AAAM,QAAA;AAAA,MAAU;AAC5C,MAAA,IAAI,OAAO,GAAA,EAAK;AAAE,QAAA,QAAA,GAAW,CAAC,QAAA;AAAU,QAAA;AAAA,MAAU;AAClD,MAAA,IAAI,QAAA,EAAU;AACd,MAAA,IAAI,OAAO,GAAA,EAAK,KAAA,EAAA;AAChB,MAAA,IAAI,OAAO,GAAA,EAAK;AACd,QAAA,KAAA,EAAA;AACA,QAAA,IAAI,UAAU,CAAA,EAAG;AACf,UAAA,OAAO,KAAK,KAAA,CAAM,IAAA,CAAK,MAAM,KAAA,EAAO,CAAA,GAAI,CAAC,CAAC,CAAA;AAAA,QAC5C;AAAA,MACF;AAAA,IACF;AACA,IAAA,MAAM,IAAI,eAAe,CAAA,4CAAA,EAA+C,IAAA,CAAK,MAAM,CAAA,EAAG,GAAG,CAAC,CAAA,CAAE,CAAA;AAAA,EAC9F;AACF;AAwCO,IAAM,cAAA,GAAN,cAA6B,KAAA,CAAM;AAAA,EACxC,WAAA,CACE,OAAA,EACgB,MAAA,EACA,KAAA,EAChB;AACA,IAAA,KAAA,CAAM,OAAO,CAAA;AAHG,IAAA,IAAA,CAAA,MAAA,GAAA,MAAA;AACA,IAAA,IAAA,CAAA,KAAA,GAAA,KAAA;AAGhB,IAAA,IAAA,CAAK,IAAA,GAAO,gBAAA;AAAA,EACd;AACF;AAKO,SAAS,qBAAqB,KAAA,EAAqC;AACxE,EAAA,MAAM,MAAA,GAAS,QAAQ,GAAA,CAAI,kBAAA;AAC3B,EAAA,IAAI,CAAC,QAAQ,OAAO,IAAA;AACpB,EAAA,MAAM,OAAA,GAAA,CAAW,QAAQ,GAAA,CAAI,mBAAA,IAAuB,qCAAqC,OAAA,CAAQ,cAAA,EAAgB,EAAE,CAAA,CAAE,IAAA,EAAK;AAC1H,EAAA,MAAM,KAAA,GAAQ,SAAA;AACd,EAAA,MAAM,aAAa,KAAA,IAAS,OAAA,CAAQ,GAAA,CAAI,gBAAA,IAAoB,gBAAgB,IAAA,EAAK;AACjF,EAAA,OAAO,EAAE,KAAA,EAAO,OAAA,EAAS,MAAA,EAAQ,OAAO,SAAA,EAAU;AACpD;AASO,SAAS,oBAAoB,WAAA,EAAsC;AACxE,EAAA,MAAM,OAAA,GAAU,qBAAqB,WAAW,CAAA;AAChD,EAAA,IAAI,CAAC,OAAA,EAAS,OAAO,EAAC;AACtB,EAAA,MAAM,MAAM,OAAA,CAAQ,GAAA,CAAI,wBAAwB,OAAA,CAAQ,GAAA,CAAI,qBAAqB,IAAA,EAAK;AACtF,EAAA,IAAI,CAAC,GAAA,EAAK,OAAO,CAAC,OAAO,CAAA;AACzB,EAAA,MAAM,SAAA,GAAY,qBAAqB,GAAG,CAAA;AAC1C,EAAA,IAAI,CAAC,SAAA,EAAW,OAAO,CAAC,OAAO,CAAA;AAC/B,EAAA,OAAO,OAAA,CAAQ,UAAU,SAAA,CAAU,KAAA,GAAQ,CAAC,OAAO,CAAA,GAAI,CAAC,OAAA,EAAS,SAAS,CAAA;AAC5E;AAKA,IAAM,uBAAA,GAA0B,IAAA;AAChC,IAAI,iBAAA,GAAoB,MAAA,CAAO,OAAA,CAAQ,GAAA,CAAI,+BAA+B,uBAAuB,CAAA;AACjG,IAAI,aAAA,GAAgB,CAAA;AAEpB,eAAe,UAAA,GAA4B;AACzC,EAAA,MAAM,IAAA,GAAO,aAAA,GAAgB,iBAAA,GAAoB,IAAA,CAAK,GAAA,EAAI;AAC1D,EAAA,IAAI,IAAA,GAAO,CAAA,EAAG,MAAM,IAAI,OAAA,CAAQ,CAAC,OAAA,KAAY,UAAA,CAAW,OAAA,EAAS,IAAI,CAAC,CAAA;AACtE,EAAA,aAAA,GAAgB,KAAK,GAAA,EAAI;AAC3B;AAEO,SAAS,gBAAgB,MAAA,EAAqC;AACnE,EAAA,MAAM,KAAA,GAAQ,mBAAA,CAAoB,MAAA,EAAQ,KAAK,CAAA;AAC/C,EAAA,MAAM,OAAA,GAAU,MAAM,CAAC,CAAA;AACvB,EAAA,IAAI,CAAC,OAAA,EAAS;AACZ,IAAA,MAAM,IAAI,cAAA;AAAA,MACR;AAAA,KACF;AAAA,EACF;AACA,EAAe,OAAA,CAAQ;AACvB,EAAgB,OAAA,CAAQ;AACxB,EAAA,MAAM,QAAQ,OAAA,CAAQ,KAAA;AACtB,EAAA,IAAI,MAAA,EAAQ,yBAAyB,MAAA,EAAW;AAC9C,IAAA,iBAAA,GAAoB,IAAA,CAAK,GAAA,CAAI,CAAA,EAAG,MAAA,CAAO,oBAAoB,CAAA;AAAA,EAC7D;AAEA,EAAA,MAAM,UAA2B,EAAC;AAElC,EAAA,eAAe,IAAA,CACb,UACA,OAAA,EACwB;AACxB,IAAA,MAAM,WAAA,GAAc,OAAA,EAAS,WAAA,IAAe,MAAA,EAAQ,WAAA,IAAe,GAAA;AACnE,IAAA,MAAM,SAAA,GAAY,OAAA,EAAS,SAAA,IAAa,MAAA,EAAQ,SAAA,IAAa,SAAS,OAAA,CAAQ,GAAA,CAAI,cAAA,IAAkB,OAAA,EAAS,EAAE,CAAA;AAE/G,IAAA,MAAM,WAAqB,EAAC;AAC5B,IAAA,IAAI,SAAA,GAAmC,IAAA;AACvC,IAAA,MAAM,SAAA,mBAAY,IAAI,GAAA,CAAI,CAAC,GAAA,EAAK,KAAK,GAAA,EAAK,GAAA,EAAK,GAAA,EAAK,GAAG,CAAC,CAAA;AAKxD,IAAA,MAAM,kBAAA,GAAqB,MAAA,CAAO,OAAA,CAAQ,GAAA,CAAI,0BAA0B,GAAO,CAAA;AAC/E,IAAA,MAAM,gBAAA,GAAmB,CAAA;AACzB,IAAA,KAAA,IAAS,KAAA,GAAQ,CAAA,EAAG,KAAA,IAAS,gBAAA,EAAkB,KAAA,EAAA,EAAS;AACtD,MAAA,IAAI,QAAQ,CAAA,EAAG;AAGb,QAAA,QAAA,CAAS,IAAA,CAAK,QAAA,IAAY,KAAA,GAAQ,CAAA,CAAA,GAAK,2DAAsD,CAAA;AAC7F,QAAA,MAAM,IAAI,OAAA,CAAQ,CAAC,YAAY,UAAA,CAAW,OAAA,EAAS,GAAM,CAAC,CAAA;AAAA,MAC5D;AACA,MAAA,KAAA,MAAW,YAAY,KAAA,EAAO;AAE9B,QAAA,MAAM,MAAA,GAAS,OAAA,CAAQ,IAAA,CAAK,EAAE,QAAA,EAAU,QAAA,CAAS,KAAA,EAAO,KAAA,EAAO,QAAA,CAAS,KAAA,EAAO,EAAA,EAAI,KAAA,EAAO,CAAA,GAAI,CAAA;AAC9F,QAAA,IAAI;AACF,UAAA,MAAM,OAAA,GAAU;AAAA,YACd,cAAA,EAAgB,kBAAA;AAAA,YAChB,eAAA,EAAiB,CAAA,OAAA,EAAU,QAAA,CAAS,MAAM,CAAA;AAAA,WAC5C;AACA,UAAA,MAAM,OAAA,GAAU;AAAA,YACd,OAAO,QAAA,CAAS,KAAA;AAAA,YAChB,QAAA;AAAA,YACA,WAAA;AAAA,YACA,UAAA,EAAY;AAAA,WACd;AACA,UAAA,MAAM,UAAA,EAAW;AACjB,UAAA,MAAM,UAAU,MAAyB,KAAA,CAAM,CAAA,EAAG,QAAA,CAAS,OAAO,CAAA,iBAAA,CAAA,EAAqB;AAAA,YACrF,MAAA,EAAQ,MAAA;AAAA,YACR,OAAA,EAAS;AAAA,cACP,cAAA,EAAgB,kBAAA;AAAA,cAChB,eAAA,EAAiB,CAAA,OAAA,EAAU,QAAA,CAAS,MAAM,CAAA;AAAA,aAC5C;AAAA,YACA,IAAA,EAAM,KAAK,SAAA,CAAU;AAAA,cACnB,OAAO,QAAA,CAAS,KAAA;AAAA,cAChB,QAAA;AAAA,cACA,WAAA;AAAA,cACA,UAAA,EAAY;AAAA,aACb,CAAA;AAAA,YACD,MAAA,EAAQ,WAAA,CAAY,OAAA,CAAQ,kBAAkB;AAAA,WAC/C,CAAA;AAGD,UAAA,IAAI,QAAA;AACJ,UAAA,IAAI;AACF,YAAA,QAAA,GAAW,MAAM,OAAA,EAAQ;AAAA,UAC3B,SAAS,MAAA,EAAQ;AACf,YAAA,QAAA,CAAS,IAAA,CAAK,CAAA,EAAG,QAAA,CAAS,KAAK,CAAA,iBAAA,EAAoB,MAAA,YAAkB,KAAA,GAAQ,MAAA,CAAO,OAAA,GAAU,MAAA,CAAO,MAAM,CAAC,CAAA,sBAAA,CAAmB,CAAA;AAC/H,YAAA,MAAM,IAAI,OAAA,CAAQ,CAAC,MAAM,UAAA,CAAW,CAAA,EAAG,GAAK,CAAC,CAAA;AAC7C,YAAA,aAAA,GAAgB,KAAK,GAAA,EAAI;AACzB,YAAA,QAAA,GAAW,MAAM,OAAA,EAAQ;AAAA,UAC3B;AACA,UAAA,IAAI,CAAC,SAAS,EAAA,EAAI;AAChB,YAAA,MAAM,OAAO,MAAM,QAAA,CAAS,MAAK,CAAE,KAAA,CAAM,MAAM,EAAE,CAAA;AACjD,YAAA,MAAM,MAAM,IAAI,cAAA;AAAA,cACd,CAAA,CAAA,EAAI,QAAA,CAAS,KAAK,CAAA,iBAAA,EAAoB,SAAS,MAAM,CAAA,CAAA,EAAI,QAAA,CAAS,UAAU,CAAA,QAAA,EAAM,IAAA,CAAK,KAAA,CAAM,CAAA,EAAG,GAAG,CAAC,CAAA,CAAA;AAAA,cACpG,QAAA,CAAS;AAAA,aACX;AACA,YAAA,IAAI,QAAA,CAAS,WAAW,GAAA,EAAK;AAG3B,cAAA,MAAM,aAAA,GAAgB,QAAA,CAAS,OAAA,CAAQ,GAAA,CAAI,aAAa,CAAA;AACxD,cAAA,MAAM,eAAe,IAAA,CAAK,GAAA;AAAA,gBACxB,GAAA;AAAA,gBACA,IAAA,CAAK,GAAA,CAAI,IAAA,EAAA,CAAS,MAAA,CAAO,QAAA,CAAS,MAAA,CAAO,aAAa,CAAC,CAAA,GAAI,MAAA,CAAO,aAAa,CAAA,GAAI,MAAM,GAAK;AAAA,eAChG;AACA,cAAA,QAAA,CAAS,IAAA,CAAK,CAAA,EAAG,QAAA,CAAS,KAAK,CAAA,sCAAA,EAAoC,KAAK,KAAA,CAAM,YAAA,GAAe,GAAI,CAAC,CAAA,CAAA,CAAG,CAAA;AACrG,cAAA,MAAM,IAAI,OAAA,CAAQ,CAAC,YAAY,UAAA,CAAW,OAAA,EAAS,YAAY,CAAC,CAAA;AAChE,cAAA,aAAA,GAAgB,KAAK,GAAA,EAAI;AACzB,cAAA,MAAM,QAAQ,MAAM,KAAA,CAAM,CAAA,EAAG,QAAA,CAAS,OAAO,CAAA,iBAAA,CAAA,EAAqB;AAAA,gBAChE,MAAA,EAAQ,MAAA;AAAA,gBACR,OAAA;AAAA,gBACA,IAAA,EAAM,IAAA,CAAK,SAAA,CAAU,OAAO,CAAA;AAAA,gBAC5B,MAAA,EAAQ,WAAA,CAAY,OAAA,CAAQ,kBAAkB;AAAA,eAC/C,CAAA;AACD,cAAA,IAAI,CAAC,KAAA,CAAM,EAAA,EAAI,MAAM,GAAA;AACrB,cAAA,MAAM,SAAA,GAAY,MAAM,kBAAA,CAAmB,KAAK,CAAA;AAKhD,cAAA,MAAM,gBAAgB,SAAA,CAAU,OAAA,GAAU,CAAC,CAAA,EAAG,SAAS,OAAA,IAAW,EAAA;AAClE,cAAA,IAAI,CAAC,aAAA,CAAc,IAAA,EAAK,EAAG,MAAM,GAAA;AACjC,cAAA,OAAA,CAAQ,MAAM,IAAI,EAAE,GAAG,QAAQ,MAAM,CAAA,EAAG,IAAI,IAAA,EAAK;AACjD,cAAA,MAAMA,aAAY,mBAAA,CAAoB,QAAA,CAAS,OAAO,SAAA,CAAU,KAAA,IAAS,SAAS,KAAA,EAAO;AAAA,gBACvF,OAAO,SAAA,CAAU,KAAA;AAAA,gBACjB,eAAA,EAAiB;AAAA,eAClB,CAAA;AACD,cAAA,OAAO;AAAA,gBACL,OAAA,EAAS,aAAA;AAAA,gBACT,KAAA,EAAO,SAAA,CAAU,KAAA,IAAS,QAAA,CAAS,KAAA;AAAA,gBACnC,UAAU,QAAA,CAAS,KAAA;AAAA,gBACnB,QAAA,EAAU,SAAS,KAAA,EAAM;AAAA,gBACzB,OAAO,SAAA,CAAU,KAAA;AAAA,gBACjB,YAAA,EAAc,SAAA,CAAU,OAAA,GAAU,CAAC,CAAA,EAAG,aAAA;AAAA,gBACtC,UAAA,EAAYA;AAAA,eACd;AAAA,YACF;AACA,YAAA,IAAI,SAAA,CAAU,GAAA,CAAI,QAAA,CAAS,MAAM,CAAA,EAAG;AAClC,cAAA,MAAM,IAAI,OAAA,CAAQ,CAAC,MAAM,UAAA,CAAW,CAAA,EAAG,GAAI,CAAC,CAAA;AAC5C,cAAA,MAAM,QAAQ,MAAM,KAAA,CAAM,CAAA,EAAG,QAAA,CAAS,OAAO,CAAA,iBAAA,CAAA,EAAqB;AAAA,gBAChE,MAAA,EAAQ,WAAA,CAAY,OAAA,CAAQ,kBAAkB,CAAA;AAAA,gBAC9C,MAAA,EAAQ,MAAA;AAAA,gBACR,OAAA,EAAS;AAAA,kBACP,cAAA,EAAgB,kBAAA;AAAA,kBAChB,eAAA,EAAiB,CAAA,OAAA,EAAU,QAAA,CAAS,MAAM,CAAA;AAAA,iBAC5C;AAAA,gBACA,IAAA,EAAM,KAAK,SAAA,CAAU;AAAA,kBACnB,OAAO,QAAA,CAAS,KAAA;AAAA,kBAChB,QAAA;AAAA,kBACA,WAAA;AAAA,kBACA,UAAA,EAAY;AAAA,iBACb;AAAA,eACF,CAAA;AACD,cAAA,IAAI,CAAC,KAAA,CAAM,EAAA,EAAI,MAAM,GAAA;AACrB,cAAA,MAAM,SAAA,GAAY,MAAM,kBAAA,CAAmB,KAAK,CAAA;AAKhD,cAAA,MAAM,eAAe,SAAA,CAAU,OAAA,GAAU,CAAC,CAAA,EAAG,SAAS,OAAA,IAAW,EAAA;AACjE,cAAA,OAAA,CAAQ,MAAM,IAAI,EAAE,GAAG,QAAQ,MAAM,CAAA,EAAG,IAAI,IAAA,EAAK;AACjD,cAAA,MAAMA,aAAY,mBAAA,CAAoB,QAAA,CAAS,OAAO,SAAA,CAAU,KAAA,IAAS,SAAS,KAAA,EAAO;AAAA,gBACvF,OAAO,SAAA,CAAU,KAAA;AAAA,gBACjB,eAAA,EAAiB;AAAA,eAClB,CAAA;AACD,cAAA,OAAO;AAAA,gBACL,OAAA,EAAS,YAAA;AAAA,gBACT,KAAA,EAAO,SAAA,CAAU,KAAA,IAAS,QAAA,CAAS,KAAA;AAAA,gBACnC,UAAU,QAAA,CAAS,KAAA;AAAA,gBACnB,QAAA,EAAU,SAAS,KAAA,EAAM;AAAA,gBACzB,OAAO,SAAA,CAAU,KAAA;AAAA,gBACjB,YAAA,EAAc,SAAA,CAAU,OAAA,GAAU,CAAC,CAAA,EAAG,aAAA;AAAA,gBACtC,UAAA,EAAYA;AAAA,eACd;AAAA,YACF;AACA,YAAA,MAAM,GAAA;AAAA,UACR;AACA,UAAA,MAAM,IAAA,GAAO,MAAM,kBAAA,CAAmB,QAAQ,CAAA;AAK9C,UAAA,MAAM,eAAgB,IAAA,CAAyD,KAAA;AAC/E,UAAA,IAAI,YAAA,EAAc;AAGhB,YAAA,MAAM,IAAI,cAAA,CAAe,GAAA,GAAM,QAAA,CAAS,KAAA,GAAQ,mBAAA,IAAuB,YAAA,CAAa,IAAA,IAAQ,EAAA,CAAA,GAAM,IAAA,IAAQ,YAAA,CAAa,OAAA,IAAW,SAAA,CAAU,CAAA;AAAA,UAC9I;AACA,UAAA,MAAM,UAAU,IAAA,CAAK,OAAA,GAAU,CAAC,CAAA,EAAG,SAAS,OAAA,IAAW,EAAA;AACvD,UAAA,IAAI,CAAC,OAAA,CAAQ,IAAA,EAAK,EAAG;AAGnB,YAAA,MAAM,IAAI,cAAA,CAAe,GAAA,GAAM,QAAA,CAAS,QAAQ,6BAA6B,CAAA;AAAA,UAC/E;AACA,UAAA,MAAM,YAAY,mBAAA,CAAoB,QAAA,CAAS,OAAO,IAAA,CAAK,KAAA,IAAS,SAAS,KAAA,EAAO;AAAA,YAClF,OAAO,IAAA,CAAK,KAAA;AAAA,YACZ,eAAA,EAAiB;AAAA,WAClB,CAAA;AACD,UAAA,OAAO;AAAA,YACL,OAAA;AAAA,YACA,KAAA,EAAO,IAAA,CAAK,KAAA,IAAS,QAAA,CAAS,KAAA;AAAA,YAC9B,UAAU,QAAA,CAAS,KAAA;AAAA,YACnB,QAAA,EAAU,SAAS,KAAA,EAAM;AAAA,YACzB,OAAO,IAAA,CAAK,KAAA;AAAA,YACZ,YAAA,EAAc,IAAA,CAAK,OAAA,GAAU,CAAC,CAAA,EAAG,aAAA;AAAA,YACjC,UAAA,EAAY;AAAA,WACd;AAAA,QACF,SAAS,GAAA,EAAK;AACZ,UAAA,MAAM,MAAM,GAAA,YAAe,KAAA,GAAQ,GAAA,CAAI,OAAA,GAAU,OAAO,GAAG,CAAA;AAC3D,UAAA,OAAA,CAAQ,MAAM,CAAA,GAAI,EAAE,GAAG,QAAQ,MAAM,CAAA,EAAG,EAAA,EAAI,KAAA,EAAO,KAAA,EAAO,GAAA,CAAI,KAAA,CAAM,CAAA,EAAG,GAAG,CAAA,EAAE;AAC5E,UAAA,QAAA,CAAS,KAAK,CAAA,EAAG,QAAA,CAAS,KAAK,CAAA,EAAA,EAAK,GAAG,CAAA,CAAE,CAAA;AACzC,UAAA,SAAA,GAAY,GAAA,YAAe,cAAA,GAAiB,GAAA,GAAM,IAAI,cAAA,CAAe,IAAI,QAAA,CAAS,KAAK,CAAA,EAAA,EAAK,GAAG,CAAA,CAAE,CAAA;AAAA,QAEnG;AAAA,MACF;AACE,MAAA,IAAI,QAAQ,gBAAA,EAAkB;AAC9B,MAAA,MAAM,SAAA,IAAa,IAAI,cAAA,CAAe,yCAAyC,CAAA;AAAA,IACjF;AACA,IAAA,MAAM,SAAA,IAAa,IAAI,cAAA,CAAe,iCAAA,GAAoC,mBAAmB,UAAU,CAAA;AAAA,EACzG;AAEA,EAAA,OAAO,EAAE,IAAA,EAAM,KAAA,EAAO,OAAA,EAAQ;AAChC;AAKO,SAAS,iBAAiB,IAAA,EAAuC;AAEtE,EAAA,IAAI,OAAA,GAAU,KAAK,IAAA,EAAK;AACxB,EAAA,IAAI,OAAA,CAAQ,UAAA,CAAW,SAAS,CAAA,EAAG;AACjC,IAAA,OAAA,GAAU,OAAA,CAAQ,MAAM,CAAC,CAAA;AAAA,EAC3B,CAAA,MAAA,IAAW,OAAA,CAAQ,UAAA,CAAW,KAAK,CAAA,EAAG;AACpC,IAAA,OAAA,GAAU,OAAA,CAAQ,MAAM,CAAC,CAAA;AAAA,EAC3B;AACA,EAAA,IAAI,OAAA,CAAQ,QAAA,CAAS,KAAK,CAAA,EAAG;AAC3B,IAAA,OAAA,GAAU,OAAA,CAAQ,KAAA,CAAM,CAAA,EAAG,EAAE,CAAA;AAAA,EAC/B;AACA,EAAA,OAAA,GAAU,QAAQ,IAAA,EAAK;AAEvB,EAAA,IAAI;AACF,IAAA,OAAO,IAAA,CAAK,MAAM,OAAO,CAAA;AAAA,EAC3B,CAAA,CAAA,MAAQ;AAEN,IAAA,MAAM,KAAA,GAAQ,OAAA,CAAQ,OAAA,CAAQ,GAAG,CAAA;AACjC,IAAA,IAAI,SAAS,CAAA,EAAG;AACd,MAAA,IAAI,KAAA,GAAQ,CAAA;AACZ,MAAA,IAAI,QAAA,GAAW,KAAA;AACf,MAAA,IAAI,MAAA,GAAS,KAAA;AACb,MAAA,KAAA,IAAS,CAAA,GAAI,KAAA,EAAO,CAAA,GAAI,OAAA,CAAQ,QAAQ,CAAA,EAAA,EAAK;AAC3C,QAAA,MAAM,EAAA,GAAK,QAAQ,CAAC,CAAA;AACpB,QAAA,IAAI,MAAA,EAAQ;AAAE,UAAA,MAAA,GAAS,KAAA;AAAO,UAAA;AAAA,QAAU;AACxC,QAAA,IAAI,OAAO,IAAA,EAAM;AAAE,UAAA,MAAA,GAAS,IAAA;AAAM,UAAA;AAAA,QAAU;AAC5C,QAAA,IAAI,OAAO,GAAA,EAAK;AAAE,UAAA,QAAA,GAAW,CAAC,QAAA;AAAU,UAAA;AAAA,QAAU;AAClD,QAAA,IAAI,QAAA,EAAU;AACd,QAAA,IAAI,OAAO,GAAA,EAAK,KAAA,EAAA;AAChB,QAAA,IAAI,OAAO,GAAA,EAAK;AACd,UAAA,KAAA,EAAA;AACA,UAAA,IAAI,UAAU,CAAA,EAAG;AACf,YAAA,IAAI;AACF,cAAA,OAAO,KAAK,KAAA,CAAM,OAAA,CAAQ,MAAM,KAAA,EAAO,CAAA,GAAI,CAAC,CAAC,CAAA;AAAA,YAC/C,CAAA,CAAA,MAAQ;AAAA,YAER;AAAA,UACF;AAAA,QACF;AAAA,MACF;AAAA,IACF;AACA,IAAA,MAAM,IAAI,eAAe,CAAA,wCAAA,EAA2C,OAAA,CAAQ,MAAM,CAAA,EAAG,GAAG,CAAC,CAAA,CAAE,CAAA;AAAA,EAC7F;AACF;AAKA,eAAsB,WAAA,CACpB,MAAA,EACA,YAAA,EACA,UAAA,EACA,OAAA,EACkC;AAClC,EAAA,MAAM,MAAA,GAAS,MAAM,MAAA,CAAO,IAAA;AAAA,IAC1B;AAAA,MACE,EAAE,IAAA,EAAM,QAAA,EAAU,OAAA,EAAS,YAAA,EAAa;AAAA,MACxC,EAAE,IAAA,EAAM,MAAA,EAAQ,OAAA,EAAS,UAAA;AAAW,KACtC;AAAA,IACA;AAAA,GACF;AAEA,EAAA,OAAO,gBAAA,CAAiB,OAAO,OAAO,CAAA;AACxC","file":"llmClient.cjs","sourcesContent":["/**\n * Single source of truth for ONE LLM response through the kit.\n *\n * Every `LlmClient.chat` return path attaches one of these so downstream\n * code inspects the ACTUAL resolved provider/model and framing flags instead\n * of trusting the static client.model card. This is the contract behind the\n * \"no silent cross-gateway rerouting\" policy: each response carries its own\n * provenance and framing metadata.\n */\nexport interface LlmTransportResult {\n  /** Provider label that actually served this response (e.g. '9router'). */\n  provider: string;\n  /** Exact model string sent to and returned by that provider. */\n  model: string;\n  /** True when the response body was stripped of 9router SSE framing. */\n  framingStripped: boolean;\n  /** When present, the upstream returned an error envelope inside HTTP 200. */\n  upstreamError?: string;\n  /** When present, usage record attached to the response. */\n  usage?: {\n    prompt_tokens: number;\n    completion_tokens: number;\n    total_tokens: number;\n    finishReason?: string;\n  };\n}\n\n/**\n * Build the transport result that accompanies a successful chat completion.\n */\nexport function transportResultFrom(\n  provider: string,\n  model: string,\n  options?: {\n    framingStripped?: boolean;\n    upstreamError?: string;\n    usage?: LlmTransportResult['usage'];\n  },\n): LlmTransportResult {\n  return {\n    provider,\n    model,\n    framingStripped: !!options?.framingStripped,\n    upstreamError: options?.upstreamError,\n    usage: options?.usage,\n  };\n}\n","/**\n * LLM client wrapper — provides a unified interface for LLM calls\n * used by the graph generation pipeline.\n *\n * Supports OpenAI-compatible APIs via environment variables:\n *   LLM_API_KEY, LLM_BASE_URL, LLM_MODEL (or LLM_TIER_FAST / LLM_TIER_STRONG)\n */\n\nimport { transportResultFrom, type LlmTransportResult } from './llmTransportResult';\nexport { transportResultFrom, type LlmTransportResult } from './llmTransportResult';\n\nexport interface LlmClientConfig {\n  apiKey?: string;\n  baseUrl?: string;\n  model?: string;\n  temperature?: number;\n  maxTokens?: number;\n  /**\n   * Minimum ms between requests, enforced process-wide so free tiers are not\n   * rate-limited by a burst. Defaults to LLM_MIN_REQUEST_INTERVAL_MS (2500).\n   * Set 0 to disable — correct when several pipelines run concurrently in one\n   * process and per-pipeline pacing is enough, or when the caller owns the\n   * rate limit itself (serverless).\n   */\n  minRequestIntervalMs?: number;\n}\n\n/** One provider attempt. Kept per client instance so provenance never leaks across requests. */\nexport interface LlmCallRecord {\n  provider: string;\n  model: string;\n  ok: boolean;\n  error?: string;\n}\n\nexport interface LlmProvenance {\n  providers: string[];\n  models: string[];\n  calls: number;\n  failed_calls: number;\n  /** true when at least one provider fell through to the next in the chain. */\n  degraded: boolean;\n}\n\nexport function summarizeProvenance(callLog: LlmCallRecord[]): LlmProvenance {\n  const providers: string[] = [];\n  const models: string[] = [];\n  let failed = 0;\n  for (const r of callLog) {\n    if (!providers.includes(r.provider)) providers.push(r.provider);\n    if (!models.includes(r.model)) models.push(r.model);\n    if (!r.ok) failed++;\n  }\n  return { providers, models, calls: callLog.length, failed_calls: failed, degraded: failed > 0 };\n}\n\n/**\n * Some OpenAI-compatible gateways (observed: 9router / ai-router.orchable.app)\n * paste SSE stream framing onto NON-stream responses: a leading blank line,\n * then the chat.completion JSON object, wrapped with a trailing `data: [DONE]`.\n * response.json() throws \"Unexpected non-whitespace character after JSON\", so\n * read the text and cut the body back to the first balanced JSON object.\n */\nexport async function responseJsonCompat(response: Response): Promise<Record<string, any>> {\n  const text = await response.text();\n  try {\n    return JSON.parse(text);\n  } catch {\n    const start = text.indexOf('{');\n    if (start < 0) {\n      throw new LlmClientError(`LLM API returned a non-JSON body: ${text.slice(0, 200)}`);\n    }\n    let depth = 0;\n    let inString = false;\n    let escape = false;\n    for (let i = start; i < text.length; i++) {\n      const ch = text[i];\n      if (escape) { escape = false; continue; }\n      if (ch === '\\\\') { escape = true; continue; }\n      if (ch === '\"') { inString = !inString; continue; }\n      if (inString) continue;\n      if (ch === '{') depth++;\n      if (ch === '}') {\n        depth--;\n        if (depth === 0) {\n          return JSON.parse(text.slice(start, i + 1));\n        }\n      }\n    }\n    throw new LlmClientError(`LLM API returned an unterminated JSON body: ${text.slice(0, 200)}`);\n  }\n}\n\n/** One callable provider endpoint in the fallback chain. */\nexport interface ProviderSpec {\n  label: string;\n  baseUrl: string;\n  apiKey: string;\n  model: string;\n}\n\nexport interface LlmClient {\n  chat(messages: LlmChatMessage[], options?: { temperature?: number; maxTokens?: number }): Promise<LlmChatResult>;\n  model: string;\n  /** Every provider attempt made through this client — the provenance of the graph it produced. */\n  callLog: LlmCallRecord[];\n}\n\nexport interface LlmChatMessage {\n  role: 'system' | 'user' | 'assistant';\n  content: string;\n}\n\nexport interface LlmChatResult {\n  content: string;\n  model: string;\n  /** Provider label that ultimately served the request. */\n  provider?: string;\n  /** What happened per provider when fallback occurred (empty when first try succeeded). */\n  attempts?: string[];\n  /** 'stop' when complete; 'length' means the response was TRUNCATED by max_tokens. */\n  finishReason?: string;\n  usage?: {\n    prompt_tokens: number;\n    completion_tokens: number;\n    total_tokens: number;\n    finishReason?: string;\n  };\n  _transport?: LlmTransportResult;\n}\n\nexport class LlmClientError extends Error {\n  constructor(\n    message: string,\n    public readonly status?: number,\n    public readonly cause?: Error,\n  ) {\n    super(message);\n    this.name = 'LlmClientError';\n  }\n}\n\n/**\n * Create an LLM client from environment variables.\n */\nexport function build9routerProvider(model?: string): ProviderSpec | null {\n  const apiKey = process.env.NINEROUTER_API_KEY;\n  if (!apiKey) return null;\n  const baseUrl = (process.env.NINEROUTER_BASE_URL || 'https://ai-router.orchable.app/v1').replace(/^[\"']|[\"']$/g, '').trim();\n  const label = '9router';\n  const modelName = (model || process.env.NINEROUTER_MODEL || 'laguna-s-2.1').trim();\n  return { label, baseUrl, apiKey, model: modelName };\n}\n\n/**\n * Synthesize a 9router-only chain from the configured primary model.\n * POLICY (user directive 2026-10-07): 9router free-tier ONLY — laguna-s-2.1\n * and nemotron-ultra ride the same pool from different upstream providers.\n * The returned chain is PRIMARY then SAME-POOL SECONDARY; the caller owns any\n * fallback semantics and MUST NOT silently branch to another gateway.\n */\nexport function resolve9routerChain(configModel?: string): ProviderSpec[] {\n  const primary = build9routerProvider(configModel);\n  if (!primary) return [];\n  const alt = process.env.NINEROUTER_MODEL_ALT && process.env.NINEROUTER_MODEL_ALT.trim();\n  if (!alt) return [primary];\n  const secondary = build9routerProvider(alt);\n  if (!secondary) return [primary];\n  return primary.model === secondary.model ? [primary] : [primary, secondary];\n}\n\n// Free-tier rate limits (NVIDIA NIM, OpenRouter) trip when a pipeline fires\n// many requests in a burst. Global pacing keeps us under the limit; a single\n// isolated call still goes through immediately.\nconst MIN_REQUEST_INTERVAL_MS = 2_500;\nlet requestIntervalMs = Number(process.env.LLM_MIN_REQUEST_INTERVAL_MS || MIN_REQUEST_INTERVAL_MS);\nlet lastRequestAt = 0;\n\nasync function pacedDelay(): Promise<void> {\n  const wait = lastRequestAt + requestIntervalMs - Date.now();\n  if (wait > 0) await new Promise((resolve) => setTimeout(resolve, wait));\n  lastRequestAt = Date.now();\n}\n\nexport function createLlmClient(config?: LlmClientConfig): LlmClient {\n  const chain = resolve9routerChain(config?.model);\n  const primary = chain[0];\n  if (!primary) {\n    throw new LlmClientError(\n      'No 9router provider configured. Set NINEROUTER_API_KEY and NINEROUTER_MODEL (or the NINEROUTER_* env defaults expected by the kit).',\n    );\n  }\n  const apiKey = primary.apiKey;\n  const baseUrl = primary.baseUrl;\n  const model = primary.model;\n  if (config?.minRequestIntervalMs !== undefined) {\n    requestIntervalMs = Math.max(0, config.minRequestIntervalMs);\n  }\n  // Per-client, so concurrent pipelines never inherit each other's provenance.\n  const callLog: LlmCallRecord[] = [];\n\n  async function chat(\n    messages: LlmChatMessage[],\n    options?: { temperature?: number; maxTokens?: number },\n  ): Promise<LlmChatResult> {\n    const temperature = options?.temperature ?? config?.temperature ?? 0.1;\n    const maxTokens = options?.maxTokens ?? config?.maxTokens ?? parseInt(process.env.LLM_MAX_TOKENS || '65536', 10);\n\n    const attempts: string[] = [];\n    let lastError: LlmClientError | null = null;\n    const TRANSIENT = new Set([408, 429, 500, 502, 503, 504]);\n    // Hard per-request cap so a hanging provider degrades to a transient error\n    // and the chain falls through instead of blocking the pipeline forever.\n    // Reasoning presets (e.g. deepseek-v4-flash) can legitimately think for\n    // minutes on large structured outputs — 5 minutes default, env-overridable.\n    const REQUEST_TIMEOUT_MS = Number(process.env.LLM_REQUEST_TIMEOUT_MS || 300_000);\n    const MAX_CHAIN_ROUNDS = 3;\n    for (let round = 1; round <= MAX_CHAIN_ROUNDS; round++) {\n      if (round > 1) {\n        // Whole chain failed transiently (rate limits / upstream 5xx) — back\n        // off and rewalk the chain from the primary instead of dying.\n        attempts.push('round ' + (round - 1) + ' failed — backing off 20s before rewalking the chain');\n        await new Promise((resolve) => setTimeout(resolve, 20_000));\n      }\n      for (const provider of chain) {\n      // Provenance: opened as failed, flipped to ok on the paths that return.\n      const recIdx = callLog.push({ provider: provider.label, model: provider.model, ok: false }) - 1;\n      try {\n        const headers = {\n          'Content-Type': 'application/json',\n          'Authorization': `Bearer ${provider.apiKey}`,\n        };\n        const payload = {\n          model: provider.model,\n          messages,\n          temperature,\n          max_tokens: maxTokens,\n        };\n        await pacedDelay();\n        const doFetch = (): Promise<Response> => fetch(`${provider.baseUrl}/chat/completions`, {\n          method: 'POST',\n          headers: {\n            'Content-Type': 'application/json',\n            'Authorization': `Bearer ${provider.apiKey}`,\n          },\n          body: JSON.stringify({\n            model: provider.model,\n            messages,\n            temperature,\n            max_tokens: maxTokens,\n          }),\n          signal: AbortSignal.timeout(REQUEST_TIMEOUT_MS),\n        });\n        // Transient network failures (DNS blip, TLS reset, \"fetch failed\") get\n        // one same-provider retry before falling through the chain.\n        let response: Response;\n        try {\n          response = await doFetch();\n        } catch (netErr) {\n          attempts.push(`${provider.label}: network error (${netErr instanceof Error ? netErr.message : String(netErr)}) — retrying once`);\n          await new Promise((r) => setTimeout(r, 5_000));\n          lastRequestAt = Date.now();\n          response = await doFetch();\n        }\n        if (!response.ok) {\n          const body = await response.text().catch(() => '');\n          const err = new LlmClientError(\n            `[${provider.label}] LLM API error: ${response.status} ${response.statusText} — ${body.slice(0, 200)}`,\n            response.status,\n          );\n          if (response.status === 429) {\n            // Rate-limited: honor Retry-After (bounded) and retry once on the\n            // same provider before falling through to the next one.\n            const retryAfterRaw = response.headers.get('retry-after');\n            const retryAfterMs = Math.min(\n              60_000,\n              Math.max(15_000, (Number.isFinite(Number(retryAfterRaw)) ? Number(retryAfterRaw) : 20) * 1_000),\n            );\n            attempts.push(`${provider.label}: 429 rate-limited — backing off ${Math.round(retryAfterMs / 1000)}s`);\n            await new Promise((resolve) => setTimeout(resolve, retryAfterMs));\n            lastRequestAt = Date.now();\n            const retry = await fetch(`${provider.baseUrl}/chat/completions`, {\n              method: 'POST',\n              headers,\n              body: JSON.stringify(payload),\n              signal: AbortSignal.timeout(REQUEST_TIMEOUT_MS),\n            });\n            if (!retry.ok) throw err;\n            const retryData = await responseJsonCompat(retry) as {\n              choices?: Array<{ message?: { content?: string }; finish_reason?: string }>;\n              model?: string;\n              usage?: { prompt_tokens: number; completion_tokens: number; total_tokens: number };\n            };\n            const retryContent2 = retryData.choices?.[0]?.message?.content || '';\n            if (!retryContent2.trim()) throw err;\n            callLog[recIdx] = { ...callLog[recIdx], ok: true };\n            const transport = transportResultFrom(provider.label, retryData.model || provider.model, {\n              usage: retryData.usage,\n              framingStripped: true,\n            });\n            return {\n              content: retryContent2,\n              model: retryData.model || provider.model,\n              provider: provider.label,\n              attempts: attempts.slice(),\n              usage: retryData.usage,\n              finishReason: retryData.choices?.[0]?.finish_reason,\n              _transport: transport,\n            };\n          }\n          if (TRANSIENT.has(response.status)) {\n            await new Promise((r) => setTimeout(r, 3000));\n            const retry = await fetch(`${provider.baseUrl}/chat/completions`, {\n              signal: AbortSignal.timeout(REQUEST_TIMEOUT_MS),\n              method: 'POST',\n              headers: {\n                'Content-Type': 'application/json',\n                'Authorization': `Bearer ${provider.apiKey}`,\n              },\n              body: JSON.stringify({\n                model: provider.model,\n                messages,\n                temperature,\n                max_tokens: maxTokens,\n              }),\n            });\n            if (!retry.ok) throw err; // still transient-failing — fall through to next provider\n            const retryData = await responseJsonCompat(retry) as {\n              choices?: Array<{ message?: { content?: string }; finish_reason?: string }>;\n              model?: string;\n              usage?: { prompt_tokens: number; completion_tokens: number; total_tokens: number };\n            };\n            const retryContent = retryData.choices?.[0]?.message?.content || '';\n            callLog[recIdx] = { ...callLog[recIdx], ok: true };\n            const transport = transportResultFrom(provider.label, retryData.model || provider.model, {\n              usage: retryData.usage,\n              framingStripped: true,\n            });\n            return {\n              content: retryContent,\n              model: retryData.model || provider.model,\n              provider: provider.label,\n              attempts: attempts.slice(),\n              usage: retryData.usage,\n              finishReason: retryData.choices?.[0]?.finish_reason,\n              _transport: transport,\n            };\n          }\n          throw err;\n        }\n        const data = await responseJsonCompat(response) as {\n          choices?: Array<{ message?: { content?: string }; finish_reason?: string }>;\n          model?: string;\n          usage?: { prompt_tokens: number; completion_tokens: number; total_tokens: number };\n        };\n        const wrappedError = (data as { error?: { message?: string; code?: number } }).error;\n        if (wrappedError) {\n          // OpenRouter proxies some upstream failures as HTTP 200 with an\n          // {\"error\": {...}} body — treat as transient and fall through.\n          throw new LlmClientError('[' + provider.label + '] upstream error ' + (wrappedError.code ?? '') + ': ' + (wrappedError.message ?? 'unknown'));\n        }\n        const content = data.choices?.[0]?.message?.content || '';\n        if (!content.trim()) {\n          // Empty completion (provider outage / content filter) is transient —\n          // fall through to the next provider instead of returning unusable ''.\n          throw new LlmClientError('[' + provider.label + '] empty completion returned');\n        }\n        const transport = transportResultFrom(provider.label, data.model || provider.model, {\n          usage: data.usage,\n          framingStripped: true,\n        });\n        return {\n          content,\n          model: data.model || provider.model,\n          provider: provider.label,\n          attempts: attempts.slice(),\n          usage: data.usage,\n          finishReason: data.choices?.[0]?.finish_reason,\n          _transport: transport,\n        };\n      } catch (err) {\n        const msg = err instanceof Error ? err.message : String(err);\n        callLog[recIdx] = { ...callLog[recIdx], ok: false, error: msg.slice(0, 200) };\n        attempts.push(`${provider.label}: ${msg}`);\n        lastError = err instanceof LlmClientError ? err : new LlmClientError(`[${provider.label}] ${msg}`);\n        // try next provider\n      }\n    }\n      if (round < MAX_CHAIN_ROUNDS) continue; // rewalk the chain\n      throw lastError ?? new LlmClientError('All LLM providers failed (empty chain).');\n    }\n    throw lastError ?? new LlmClientError('All LLM providers failed after ' + MAX_CHAIN_ROUNDS + ' rounds.');\n  }\n\n  return { chat, model, callLog };\n}\n\n/**\n * Parse JSON from LLM response, handling markdown code fences.\n */\nexport function parseJsonFromLlm(text: string): Record<string, unknown> {\n  // Strip markdown code fences if present\n  let cleaned = text.trim();\n  if (cleaned.startsWith('```json')) {\n    cleaned = cleaned.slice(7);\n  } else if (cleaned.startsWith('```')) {\n    cleaned = cleaned.slice(3);\n  }\n  if (cleaned.endsWith('```')) {\n    cleaned = cleaned.slice(0, -3);\n  }\n  cleaned = cleaned.trim();\n\n  try {\n    return JSON.parse(cleaned) as Record<string, unknown>;\n  } catch {\n    // Try to find first complete JSON object using brace-depth tracking\n    const start = cleaned.indexOf('{');\n    if (start >= 0) {\n      let depth = 0;\n      let inString = false;\n      let escape = false;\n      for (let i = start; i < cleaned.length; i++) {\n        const ch = cleaned[i];\n        if (escape) { escape = false; continue; }\n        if (ch === '\\\\') { escape = true; continue; }\n        if (ch === '\"') { inString = !inString; continue; }\n        if (inString) continue;\n        if (ch === '{') depth++;\n        if (ch === '}') {\n          depth--;\n          if (depth === 0) {\n            try {\n              return JSON.parse(cleaned.slice(start, i + 1)) as Record<string, unknown>;\n            } catch {\n              // keep looking\n            }\n          }\n        }\n      }\n    }\n    throw new LlmClientError(`Failed to parse JSON from LLM response: ${cleaned.slice(0, 200)}`);\n  }\n}\n\n/**\n * Call LLM and parse JSON response.\n */\nexport async function llmChatJson(\n  client: ReturnType<typeof createLlmClient>,\n  systemPrompt: string,\n  userPrompt: string,\n  options?: { temperature?: number; maxTokens?: number },\n): Promise<Record<string, unknown>> {\n  const result = await client.chat(\n    [\n      { role: 'system', content: systemPrompt },\n      { role: 'user', content: userPrompt },\n    ],\n    options,\n  );\n\n  return parseJsonFromLlm(result.content);\n}\n"]}

package/dist/utils/llmClient.d.cts

@@ -0,0 +1,118 @@
+import { LlmTransportResult } from './llmTransportResult.cjs';
+export { transportResultFrom } from './llmTransportResult.cjs';
+
+/**
+ * LLM client wrapper — provides a unified interface for LLM calls
+ * used by the graph generation pipeline.
+ *
+ * Supports OpenAI-compatible APIs via environment variables:
+ *   LLM_API_KEY, LLM_BASE_URL, LLM_MODEL (or LLM_TIER_FAST / LLM_TIER_STRONG)
+ */
+
+interface LlmClientConfig {
+    apiKey?: string;
+    baseUrl?: string;
+    model?: string;
+    temperature?: number;
+    maxTokens?: number;
+    /**
+     * Minimum ms between requests, enforced process-wide so free tiers are not
+     * rate-limited by a burst. Defaults to LLM_MIN_REQUEST_INTERVAL_MS (2500).
+     * Set 0 to disable — correct when several pipelines run concurrently in one
+     * process and per-pipeline pacing is enough, or when the caller owns the
+     * rate limit itself (serverless).
+     */
+    minRequestIntervalMs?: number;
+}
+/** One provider attempt. Kept per client instance so provenance never leaks across requests. */
+interface LlmCallRecord {
+    provider: string;
+    model: string;
+    ok: boolean;
+    error?: string;
+}
+interface LlmProvenance {
+    providers: string[];
+    models: string[];
+    calls: number;
+    failed_calls: number;
+    /** true when at least one provider fell through to the next in the chain. */
+    degraded: boolean;
+}
+declare function summarizeProvenance(callLog: LlmCallRecord[]): LlmProvenance;
+/**
+ * Some OpenAI-compatible gateways (observed: 9router / ai-router.orchable.app)
+ * paste SSE stream framing onto NON-stream responses: a leading blank line,
+ * then the chat.completion JSON object, wrapped with a trailing `data: [DONE]`.
+ * response.json() throws "Unexpected non-whitespace character after JSON", so
+ * read the text and cut the body back to the first balanced JSON object.
+ */
+declare function responseJsonCompat(response: Response): Promise<Record<string, any>>;
+/** One callable provider endpoint in the fallback chain. */
+interface ProviderSpec {
+    label: string;
+    baseUrl: string;
+    apiKey: string;
+    model: string;
+}
+interface LlmClient {
+    chat(messages: LlmChatMessage[], options?: {
+        temperature?: number;
+        maxTokens?: number;
+    }): Promise<LlmChatResult>;
+    model: string;
+    /** Every provider attempt made through this client — the provenance of the graph it produced. */
+    callLog: LlmCallRecord[];
+}
+interface LlmChatMessage {
+    role: 'system' | 'user' | 'assistant';
+    content: string;
+}
+interface LlmChatResult {
+    content: string;
+    model: string;
+    /** Provider label that ultimately served the request. */
+    provider?: string;
+    /** What happened per provider when fallback occurred (empty when first try succeeded). */
+    attempts?: string[];
+    /** 'stop' when complete; 'length' means the response was TRUNCATED by max_tokens. */
+    finishReason?: string;
+    usage?: {
+        prompt_tokens: number;
+        completion_tokens: number;
+        total_tokens: number;
+        finishReason?: string;
+    };
+    _transport?: LlmTransportResult;
+}
+declare class LlmClientError extends Error {
+    readonly status?: number | undefined;
+    readonly cause?: Error | undefined;
+    constructor(message: string, status?: number | undefined, cause?: Error | undefined);
+}
+/**
+ * Create an LLM client from environment variables.
+ */
+declare function build9routerProvider(model?: string): ProviderSpec | null;
+/**
+ * Synthesize a 9router-only chain from the configured primary model.
+ * POLICY (user directive 2026-10-07): 9router free-tier ONLY — laguna-s-2.1
+ * and nemotron-ultra ride the same pool from different upstream providers.
+ * The returned chain is PRIMARY then SAME-POOL SECONDARY; the caller owns any
+ * fallback semantics and MUST NOT silently branch to another gateway.
+ */
+declare function resolve9routerChain(configModel?: string): ProviderSpec[];
+declare function createLlmClient(config?: LlmClientConfig): LlmClient;
+/**
+ * Parse JSON from LLM response, handling markdown code fences.
+ */
+declare function parseJsonFromLlm(text: string): Record<string, unknown>;
+/**
+ * Call LLM and parse JSON response.
+ */
+declare function llmChatJson(client: ReturnType<typeof createLlmClient>, systemPrompt: string, userPrompt: string, options?: {
+    temperature?: number;
+    maxTokens?: number;
+}): Promise<Record<string, unknown>>;
+
+export { type LlmCallRecord, type LlmChatMessage, type LlmChatResult, type LlmClient, type LlmClientConfig, LlmClientError, type LlmProvenance, LlmTransportResult, type ProviderSpec, build9routerProvider, createLlmClient, llmChatJson, parseJsonFromLlm, resolve9routerChain, responseJsonCompat, summarizeProvenance };

package/dist/utils/llmClient.d.ts

@@ -0,0 +1,118 @@
+import { LlmTransportResult } from './llmTransportResult.js';
+export { transportResultFrom } from './llmTransportResult.js';
+
+/**
+ * LLM client wrapper — provides a unified interface for LLM calls
+ * used by the graph generation pipeline.
+ *
+ * Supports OpenAI-compatible APIs via environment variables:
+ *   LLM_API_KEY, LLM_BASE_URL, LLM_MODEL (or LLM_TIER_FAST / LLM_TIER_STRONG)
+ */
+
+interface LlmClientConfig {
+    apiKey?: string;
+    baseUrl?: string;
+    model?: string;
+    temperature?: number;
+    maxTokens?: number;
+    /**
+     * Minimum ms between requests, enforced process-wide so free tiers are not
+     * rate-limited by a burst. Defaults to LLM_MIN_REQUEST_INTERVAL_MS (2500).
+     * Set 0 to disable — correct when several pipelines run concurrently in one
+     * process and per-pipeline pacing is enough, or when the caller owns the
+     * rate limit itself (serverless).
+     */
+    minRequestIntervalMs?: number;
+}
+/** One provider attempt. Kept per client instance so provenance never leaks across requests. */
+interface LlmCallRecord {
+    provider: string;
+    model: string;
+    ok: boolean;
+    error?: string;
+}
+interface LlmProvenance {
+    providers: string[];
+    models: string[];
+    calls: number;
+    failed_calls: number;
+    /** true when at least one provider fell through to the next in the chain. */
+    degraded: boolean;
+}
+declare function summarizeProvenance(callLog: LlmCallRecord[]): LlmProvenance;
+/**
+ * Some OpenAI-compatible gateways (observed: 9router / ai-router.orchable.app)
+ * paste SSE stream framing onto NON-stream responses: a leading blank line,
+ * then the chat.completion JSON object, wrapped with a trailing `data: [DONE]`.
+ * response.json() throws "Unexpected non-whitespace character after JSON", so
+ * read the text and cut the body back to the first balanced JSON object.
+ */
+declare function responseJsonCompat(response: Response): Promise<Record<string, any>>;
+/** One callable provider endpoint in the fallback chain. */
+interface ProviderSpec {
+    label: string;
+    baseUrl: string;
+    apiKey: string;
+    model: string;
+}
+interface LlmClient {
+    chat(messages: LlmChatMessage[], options?: {
+        temperature?: number;
+        maxTokens?: number;
+    }): Promise<LlmChatResult>;
+    model: string;
+    /** Every provider attempt made through this client — the provenance of the graph it produced. */
+    callLog: LlmCallRecord[];
+}
+interface LlmChatMessage {
+    role: 'system' | 'user' | 'assistant';
+    content: string;
+}
+interface LlmChatResult {
+    content: string;
+    model: string;
+    /** Provider label that ultimately served the request. */
+    provider?: string;
+    /** What happened per provider when fallback occurred (empty when first try succeeded). */
+    attempts?: string[];
+    /** 'stop' when complete; 'length' means the response was TRUNCATED by max_tokens. */
+    finishReason?: string;
+    usage?: {
+        prompt_tokens: number;
+        completion_tokens: number;
+        total_tokens: number;
+        finishReason?: string;
+    };
+    _transport?: LlmTransportResult;
+}
+declare class LlmClientError extends Error {
+    readonly status?: number | undefined;
+    readonly cause?: Error | undefined;
+    constructor(message: string, status?: number | undefined, cause?: Error | undefined);
+}
+/**
+ * Create an LLM client from environment variables.
+ */
+declare function build9routerProvider(model?: string): ProviderSpec | null;
+/**
+ * Synthesize a 9router-only chain from the configured primary model.
+ * POLICY (user directive 2026-10-07): 9router free-tier ONLY — laguna-s-2.1
+ * and nemotron-ultra ride the same pool from different upstream providers.
+ * The returned chain is PRIMARY then SAME-POOL SECONDARY; the caller owns any
+ * fallback semantics and MUST NOT silently branch to another gateway.
+ */
+declare function resolve9routerChain(configModel?: string): ProviderSpec[];
+declare function createLlmClient(config?: LlmClientConfig): LlmClient;
+/**
+ * Parse JSON from LLM response, handling markdown code fences.
+ */
+declare function parseJsonFromLlm(text: string): Record<string, unknown>;
+/**
+ * Call LLM and parse JSON response.
+ */
+declare function llmChatJson(client: ReturnType<typeof createLlmClient>, systemPrompt: string, userPrompt: string, options?: {
+    temperature?: number;
+    maxTokens?: number;
+}): Promise<Record<string, unknown>>;
+
+export { type LlmCallRecord, type LlmChatMessage, type LlmChatResult, type LlmClient, type LlmClientConfig, LlmClientError, type LlmProvenance, LlmTransportResult, type ProviderSpec, build9routerProvider, createLlmClient, llmChatJson, parseJsonFromLlm, resolve9routerChain, responseJsonCompat, summarizeProvenance };

package/dist/utils/llmClient.mjs

@@ -0,0 +1,4 @@
+export { LlmClientError, build9routerProvider, createLlmClient, llmChatJson, parseJsonFromLlm, resolve9routerChain, responseJsonCompat, summarizeProvenance } from '../chunk-7A2SG466.mjs';
+export { transportResultFrom } from '../chunk-IBVXXVGE.mjs';
+//# sourceMappingURL=llmClient.mjs.map
+//# sourceMappingURL=llmClient.mjs.map

package/dist/utils/llmClient.mjs.map

@@ -0,0 +1 @@
+{"version":3,"sources":[],"names":[],"mappings":"","file":"llmClient.mjs"}