@orquentra/codebuddy 2.14.0 → 2.15.1

package/dist/chunk-K4XNTYV7.js

@@ -0,0 +1,136 @@
+import {
+  readConfigFile,
+  setConfigBlock
+} from "./chunk-Y4QQML24.js";
+import {
+  LLAMA_RUNTIME,
+  LOCAL_MODELS,
+  checkLocalModelResources,
+  downloadLocalModel,
+  formatBytes,
+  freeMemory,
+  installedLocalModel,
+  localModelPath,
+  localRuntimeAvailable,
+  removeLocalModels
+} from "./chunk-NNJSAYHY.js";
+
+// src/cli/commands/local-llm.ts
+import { spawn } from "child_process";
+import { confirm, isCancel, spinner } from "@clack/prompts";
+import pc from "picocolors";
+var ok = (text) => `${pc.green("\u2713")} ${text}`;
+var warn = (text) => `${pc.yellow("!")} ${text}`;
+var row = (label, value) => `  ${label.padEnd(10)} ${value}`;
+var LOCAL_LLM_SIZES = Object.keys(LOCAL_MODELS);
+async function addLocalLlm(options = {}) {
+  const spec = LOCAL_MODELS[options.size ?? "default"];
+  const interactive = options.interactive ?? process.stdin.isTTY === true;
+  console.log(
+    `${pc.bold("Local LLM")} ${spec.id} ${pc.dim(`(${formatBytes(spec.bytes)} download, ~${formatBytes(spec.ramBytes)} RAM while carding)`)}`
+  );
+  const resources = await checkLocalModelResources(spec);
+  if (!resources.ok) {
+    console.log(warn(resources.reason));
+    return false;
+  }
+  if (freeMemory() < spec.ramBytes) {
+    console.log(warn(`only ${formatBytes(freeMemory())} RAM free right now; carding may be slow`));
+  }
+  if (!await localRuntimeAvailable()) {
+    const install = options.yes ? true : interactive ? await confirm({
+      message: `Install the local runtime (${LLAMA_RUNTIME}, ~20\u201335 MB prebuilt) with npm -g?`,
+      initialValue: true
+    }) : false;
+    if (isCancel(install) || !install) {
+      console.log(warn(`Runtime missing. Install it with: npm install -g ${LLAMA_RUNTIME}`));
+      return false;
+    }
+    const code = await new Promise((resolve) => {
+      const child = spawn("npm", ["install", "-g", LLAMA_RUNTIME], { stdio: "inherit" });
+      child.on("error", () => resolve(1));
+      child.on("close", (exit) => resolve(exit ?? 1));
+    });
+    if (code !== 0 || !await localRuntimeAvailable()) {
+      console.log(warn(`Runtime install failed. Run: npm install -g ${LLAMA_RUNTIME}`));
+      return false;
+    }
+  }
+  if ((await installedLocalModel(spec.size))?.id !== spec.id) {
+    const progress = spinner();
+    progress.start(`Downloading ${spec.file}`);
+    let shown = 0;
+    try {
+      await downloadLocalModel(spec, (received, total) => {
+        const percent = Math.floor(received / total * 100);
+        if (percent !== shown) {
+          shown = percent;
+          progress.message(`Downloading ${spec.file} ${percent}%`);
+        }
+      });
+      progress.stop(ok(`${spec.id} verified ${pc.dim(localModelPath(spec))}`));
+    } catch (error) {
+      progress.stop(warn(error.message));
+      return false;
+    }
+  } else {
+    console.log(ok(`${spec.id} already installed`));
+  }
+  const llm = (await readConfigFile()).llm;
+  if (llm === void 0 || llm.provider === "local" || llm.enabled === false) {
+    await setConfigBlock("llm", { ...llm, enabled: true, provider: "local", size: spec.size });
+  } else {
+    console.log(
+      pc.dim(
+        `  .codebuddy/config.json uses ${llm.provider}; set llm.provider to "local" to use it.`
+      )
+    );
+  }
+  console.log(
+    pc.dim(
+      "  Cards are written in the background by `codebuddy serve` or `codebuddy index --cards`."
+    )
+  );
+  return true;
+}
+async function removeLocalLlm() {
+  const removed = await removeLocalModels();
+  const llm = (await readConfigFile()).llm;
+  if (llm?.provider === "local") await setConfigBlock("llm", { ...llm, enabled: false });
+  console.log(
+    removed.length > 0 ? ok(`removed ${removed.join(", ")}; CodeBuddy falls back to 2.14 behaviour`) : pc.dim("no local LLM installed")
+  );
+}
+async function printLocalLlmStatus(cards) {
+  const llm = (await readConfigFile()).llm;
+  const installed = await installedLocalModel(llm?.size);
+  console.log(pc.bold("Local LLM"));
+  if (!installed) {
+    console.log(row("model", pc.dim("not installed (optional: codebuddy models add local-llm)")));
+    return;
+  }
+  const runtime = await localRuntimeAvailable();
+  console.log(row("model", ok(`${installed.id} ${pc.dim(formatBytes(installed.bytes))}`)));
+  console.log(
+    row("runtime", runtime ? ok(LLAMA_RUNTIME) : warn(`missing: npm install -g ${LLAMA_RUNTIME}`))
+  );
+  if (llm?.enabled === false)
+    console.log(row("status", warn("disabled in .codebuddy/config.json")));
+  if (!cards) return;
+  const pass = cards.lastPass;
+  const detail = cards.writer ? `being written by pid ${cards.writer}` : pass && `last pass: ${pass.summarized} written${pass.timedOut ? `, ${pass.timedOut} timed out` : ""}, ${pass.remaining} left (${pass.stoppedBy})`;
+  console.log(
+    row(
+      "cards",
+      `${cards.carded} / ${cards.eligible} source files${detail ? pc.dim(` (${detail})`) : ""}`
+    )
+  );
+}
+
+export {
+  LOCAL_LLM_SIZES,
+  addLocalLlm,
+  removeLocalLlm,
+  printLocalLlmStatus
+};
+//# sourceMappingURL=chunk-K4XNTYV7.js.map

package/dist/chunk-K4XNTYV7.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/cli/commands/local-llm.ts"],"sourcesContent":["import { spawn } from \"node:child_process\";\nimport { confirm, isCancel, spinner } from \"@clack/prompts\";\nimport pc from \"picocolors\";\nimport { readConfigFile, setConfigBlock } from \"../../core/config-file.js\";\nimport {\n  checkLocalModelResources,\n  downloadLocalModel,\n  formatBytes,\n  freeMemory,\n  installedLocalModel,\n  LLAMA_RUNTIME,\n  LOCAL_MODELS,\n  type LocalModelSize,\n  localModelPath,\n  localRuntimeAvailable,\n  removeLocalModels,\n} from \"../../llm/local.js\";\nimport type { cardProgress } from \"../../llm/summaries.js\";\n\nconst ok = (text: string) => `${pc.green(\"✓\")} ${text}`;\nconst warn = (text: string) => `${pc.yellow(\"!\")} ${text}`;\nconst row = (label: string, value: string) => `  ${label.padEnd(10)} ${value}`;\n\nexport const LOCAL_LLM_SIZES = Object.keys(LOCAL_MODELS) as LocalModelSize[];\n\n/**\n * `codebuddy models add local-llm`: the explicit opt-in for the local small\n * model. Shows the size, checks disk and RAM, installs the node-llama-cpp\n * runtime when missing (after confirmation), downloads the pinned GGUF with a\n * checksum check, and records the size in `.codebuddy/config.json`.\n */\nexport async function addLocalLlm(\n  options: { size?: LocalModelSize; yes?: boolean; interactive?: boolean } = {},\n): Promise<boolean> {\n  const spec = LOCAL_MODELS[options.size ?? \"default\"];\n  const interactive = options.interactive ?? process.stdin.isTTY === true;\n  console.log(\n    `${pc.bold(\"Local LLM\")} ${spec.id} ${pc.dim(`(${formatBytes(spec.bytes)} download, ~${formatBytes(spec.ramBytes)} RAM while carding)`)}`,\n  );\n  const resources = await checkLocalModelResources(spec);\n  if (!resources.ok) {\n    console.log(warn(resources.reason));\n    return false;\n  }\n  if (freeMemory() < spec.ramBytes) {\n    console.log(warn(`only ${formatBytes(freeMemory())} RAM free right now; carding may be slow`));\n  }\n  if (!(await localRuntimeAvailable())) {\n    const install = options.yes\n      ? true\n      : interactive\n        ? await confirm({\n            message: `Install the local runtime (${LLAMA_RUNTIME}, ~20–35 MB prebuilt) with npm -g?`,\n            initialValue: true,\n          })\n        : false;\n    if (isCancel(install) || !install) {\n      console.log(warn(`Runtime missing. Install it with: npm install -g ${LLAMA_RUNTIME}`));\n      return false;\n    }\n    const code = await new Promise<number>((resolve) => {\n      const child = spawn(\"npm\", [\"install\", \"-g\", LLAMA_RUNTIME], { stdio: \"inherit\" });\n      child.on(\"error\", () => resolve(1));\n      child.on(\"close\", (exit) => resolve(exit ?? 1));\n    });\n    if (code !== 0 || !(await localRuntimeAvailable())) {\n      console.log(warn(`Runtime install failed. Run: npm install -g ${LLAMA_RUNTIME}`));\n      return false;\n    }\n  }\n  if ((await installedLocalModel(spec.size))?.id !== spec.id) {\n    const progress = spinner();\n    progress.start(`Downloading ${spec.file}`);\n    let shown = 0;\n    try {\n      await downloadLocalModel(spec, (received, total) => {\n        const percent = Math.floor((received / total) * 100);\n        if (percent !== shown) {\n          shown = percent;\n          progress.message(`Downloading ${spec.file} ${percent}%`);\n        }\n      });\n      progress.stop(ok(`${spec.id} verified ${pc.dim(localModelPath(spec))}`));\n    } catch (error) {\n      progress.stop(warn((error as Error).message));\n      return false;\n    }\n  } else {\n    console.log(ok(`${spec.id} already installed`));\n  }\n  const llm = (await readConfigFile()).llm;\n  // A BYOK block stays as it is (the local model is then unused until it is switched).\n  if (llm === undefined || llm.provider === \"local\" || llm.enabled === false) {\n    await setConfigBlock(\"llm\", { ...llm, enabled: true, provider: \"local\", size: spec.size });\n  } else {\n    console.log(\n      pc.dim(\n        `  .codebuddy/config.json uses ${llm.provider}; set llm.provider to \"local\" to use it.`,\n      ),\n    );\n  }\n  console.log(\n    pc.dim(\n      \"  Cards are written in the background by `codebuddy serve` or `codebuddy index --cards`.\",\n    ),\n  );\n  return true;\n}\n\nexport async function removeLocalLlm(): Promise<void> {\n  const removed = await removeLocalModels();\n  const llm = (await readConfigFile()).llm;\n  if (llm?.provider === \"local\") await setConfigBlock(\"llm\", { ...llm, enabled: false });\n  console.log(\n    removed.length > 0\n      ? ok(`removed ${removed.join(\", \")}; CodeBuddy falls back to 2.14 behaviour`)\n      : pc.dim(\"no local LLM installed\"),\n  );\n}\n\nexport async function printLocalLlmStatus(\n  cards: Awaited<ReturnType<typeof cardProgress>> | null,\n): Promise<void> {\n  const llm = (await readConfigFile()).llm;\n  const installed = await installedLocalModel(llm?.size);\n  console.log(pc.bold(\"Local LLM\"));\n  if (!installed) {\n    console.log(row(\"model\", pc.dim(\"not installed (optional: codebuddy models add local-llm)\")));\n    return;\n  }\n  const runtime = await localRuntimeAvailable();\n  console.log(row(\"model\", ok(`${installed.id} ${pc.dim(formatBytes(installed.bytes))}`)));\n  console.log(\n    row(\"runtime\", runtime ? ok(LLAMA_RUNTIME) : warn(`missing: npm install -g ${LLAMA_RUNTIME}`)),\n  );\n  if (llm?.enabled === false)\n    console.log(row(\"status\", warn(\"disabled in .codebuddy/config.json\")));\n  if (!cards) return;\n  const pass = cards.lastPass;\n  const detail = cards.writer\n    ? `being written by pid ${cards.writer}`\n    : pass &&\n      `last pass: ${pass.summarized} written${pass.timedOut ? `, ${pass.timedOut} timed out` : \"\"}, ${pass.remaining} left (${pass.stoppedBy})`;\n  console.log(\n    row(\n      \"cards\",\n      `${cards.carded} / ${cards.eligible} source files${detail ? pc.dim(` (${detail})`) : \"\"}`,\n    ),\n  );\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;AAAA,SAAS,aAAa;AACtB,SAAS,SAAS,UAAU,eAAe;AAC3C,OAAO,QAAQ;AAiBf,IAAM,KAAK,CAAC,SAAiB,GAAG,GAAG,MAAM,QAAG,CAAC,IAAI,IAAI;AACrD,IAAM,OAAO,CAAC,SAAiB,GAAG,GAAG,OAAO,GAAG,CAAC,IAAI,IAAI;AACxD,IAAM,MAAM,CAAC,OAAe,UAAkB,KAAK,MAAM,OAAO,EAAE,CAAC,IAAI,KAAK;AAErE,IAAM,kBAAkB,OAAO,KAAK,YAAY;AAQvD,eAAsB,YACpB,UAA2E,CAAC,GAC1D;AAClB,QAAM,OAAO,aAAa,QAAQ,QAAQ,SAAS;AACnD,QAAM,cAAc,QAAQ,eAAe,QAAQ,MAAM,UAAU;AACnE,UAAQ;AAAA,IACN,GAAG,GAAG,KAAK,WAAW,CAAC,IAAI,KAAK,EAAE,IAAI,GAAG,IAAI,IAAI,YAAY,KAAK,KAAK,CAAC,eAAe,YAAY,KAAK,QAAQ,CAAC,qBAAqB,CAAC;AAAA,EACzI;AACA,QAAM,YAAY,MAAM,yBAAyB,IAAI;AACrD,MAAI,CAAC,UAAU,IAAI;AACjB,YAAQ,IAAI,KAAK,UAAU,MAAM,CAAC;AAClC,WAAO;AAAA,EACT;AACA,MAAI,WAAW,IAAI,KAAK,UAAU;AAChC,YAAQ,IAAI,KAAK,QAAQ,YAAY,WAAW,CAAC,CAAC,0CAA0C,CAAC;AAAA,EAC/F;AACA,MAAI,CAAE,MAAM,sBAAsB,GAAI;AACpC,UAAM,UAAU,QAAQ,MACpB,OACA,cACE,MAAM,QAAQ;AAAA,MACZ,SAAS,8BAA8B,aAAa;AAAA,MACpD,cAAc;AAAA,IAChB,CAAC,IACD;AACN,QAAI,SAAS,OAAO,KAAK,CAAC,SAAS;AACjC,cAAQ,IAAI,KAAK,oDAAoD,aAAa,EAAE,CAAC;AACrF,aAAO;AAAA,IACT;AACA,UAAM,OAAO,MAAM,IAAI,QAAgB,CAAC,YAAY;AAClD,YAAM,QAAQ,MAAM,OAAO,CAAC,WAAW,MAAM,aAAa,GAAG,EAAE,OAAO,UAAU,CAAC;AACjF,YAAM,GAAG,SAAS,MAAM,QAAQ,CAAC,CAAC;AAClC,YAAM,GAAG,SAAS,CAAC,SAAS,QAAQ,QAAQ,CAAC,CAAC;AAAA,IAChD,CAAC;AACD,QAAI,SAAS,KAAK,CAAE,MAAM,sBAAsB,GAAI;AAClD,cAAQ,IAAI,KAAK,+CAA+C,aAAa,EAAE,CAAC;AAChF,aAAO;AAAA,IACT;AAAA,EACF;AACA,OAAK,MAAM,oBAAoB,KAAK,IAAI,IAAI,OAAO,KAAK,IAAI;AAC1D,UAAM,WAAW,QAAQ;AACzB,aAAS,MAAM,eAAe,KAAK,IAAI,EAAE;AACzC,QAAI,QAAQ;AACZ,QAAI;AACF,YAAM,mBAAmB,MAAM,CAAC,UAAU,UAAU;AAClD,cAAM,UAAU,KAAK,MAAO,WAAW,QAAS,GAAG;AACnD,YAAI,YAAY,OAAO;AACrB,kBAAQ;AACR,mBAAS,QAAQ,eAAe,KAAK,IAAI,IAAI,OAAO,GAAG;AAAA,QACzD;AAAA,MACF,CAAC;AACD,eAAS,KAAK,GAAG,GAAG,KAAK,EAAE,aAAa,GAAG,IAAI,eAAe,IAAI,CAAC,CAAC,EAAE,CAAC;AAAA,IACzE,SAAS,OAAO;AACd,eAAS,KAAK,KAAM,MAAgB,OAAO,CAAC;AAC5C,aAAO;AAAA,IACT;AAAA,EACF,OAAO;AACL,YAAQ,IAAI,GAAG,GAAG,KAAK,EAAE,oBAAoB,CAAC;AAAA,EAChD;AACA,QAAM,OAAO,MAAM,eAAe,GAAG;AAErC,MAAI,QAAQ,UAAa,IAAI,aAAa,WAAW,IAAI,YAAY,OAAO;AAC1E,UAAM,eAAe,OAAO,EAAE,GAAG,KAAK,SAAS,MAAM,UAAU,SAAS,MAAM,KAAK,KAAK,CAAC;AAAA,EAC3F,OAAO;AACL,YAAQ;AAAA,MACN,GAAG;AAAA,QACD,iCAAiC,IAAI,QAAQ;AAAA,MAC/C;AAAA,IACF;AAAA,EACF;AACA,UAAQ;AAAA,IACN,GAAG;AAAA,MACD;AAAA,IACF;AAAA,EACF;AACA,SAAO;AACT;AAEA,eAAsB,iBAAgC;AACpD,QAAM,UAAU,MAAM,kBAAkB;AACxC,QAAM,OAAO,MAAM,eAAe,GAAG;AACrC,MAAI,KAAK,aAAa,QAAS,OAAM,eAAe,OAAO,EAAE,GAAG,KAAK,SAAS,MAAM,CAAC;AACrF,UAAQ;AAAA,IACN,QAAQ,SAAS,IACb,GAAG,WAAW,QAAQ,KAAK,IAAI,CAAC,0CAA0C,IAC1E,GAAG,IAAI,wBAAwB;AAAA,EACrC;AACF;AAEA,eAAsB,oBACpB,OACe;AACf,QAAM,OAAO,MAAM,eAAe,GAAG;AACrC,QAAM,YAAY,MAAM,oBAAoB,KAAK,IAAI;AACrD,UAAQ,IAAI,GAAG,KAAK,WAAW,CAAC;AAChC,MAAI,CAAC,WAAW;AACd,YAAQ,IAAI,IAAI,SAAS,GAAG,IAAI,0DAA0D,CAAC,CAAC;AAC5F;AAAA,EACF;AACA,QAAM,UAAU,MAAM,sBAAsB;AAC5C,UAAQ,IAAI,IAAI,SAAS,GAAG,GAAG,UAAU,EAAE,IAAI,GAAG,IAAI,YAAY,UAAU,KAAK,CAAC,CAAC,EAAE,CAAC,CAAC;AACvF,UAAQ;AAAA,IACN,IAAI,WAAW,UAAU,GAAG,aAAa,IAAI,KAAK,2BAA2B,aAAa,EAAE,CAAC;AAAA,EAC/F;AACA,MAAI,KAAK,YAAY;AACnB,YAAQ,IAAI,IAAI,UAAU,KAAK,oCAAoC,CAAC,CAAC;AACvE,MAAI,CAAC,MAAO;AACZ,QAAM,OAAO,MAAM;AACnB,QAAM,SAAS,MAAM,SACjB,wBAAwB,MAAM,MAAM,KACpC,QACA,cAAc,KAAK,UAAU,WAAW,KAAK,WAAW,KAAK,KAAK,QAAQ,eAAe,EAAE,KAAK,KAAK,SAAS,UAAU,KAAK,SAAS;AAC1I,UAAQ;AAAA,IACN;AAAA,MACE;AAAA,MACA,GAAG,MAAM,MAAM,MAAM,MAAM,QAAQ,gBAAgB,SAAS,GAAG,IAAI,KAAK,MAAM,GAAG,IAAI,EAAE;AAAA,IACzF;AAAA,EACF;AACF;","names":[]}

package/dist/{chunk-ESZXCZIN.js → chunk-MNSMW4MZ.js}

@@ -1,9 +1,12 @@
 import {
   redactLlmKeys
-} from "./chunk-SFMIMLXR.js";
+} from "./chunk-RGIVHKBH.js";
 import {
   writeAtomic
 } from "./chunk-CG5B3LYG.js";
+import {
+  localChatJson
+} from "./chunk-NNJSAYHY.js";
 
 // src/llm/budget.ts
 import { mkdir, readFile } from "fs/promises";
@@ -64,6 +67,7 @@ async function storeResponse(root, key, value) {
 var DEFAULT_COOLDOWN_SECONDS = 60;
 async function chatJson(config, request) {
   const started = Date.now();
+  if (config.local) return localCall(config.local, request);
   try {
     return await call(config, request, started);
   } catch (error) {
@@ -121,7 +125,9 @@ async function call(config, request, started) {
         ],
         temperature: 0,
         max_tokens: request.maxTokens,
-        response_format: { type: "json_object" }
+        response_format: { type: "json_object" },
+        // Qwen3 reasons by default on some hosts; cards and reranks need none.
+        ...config.provider === "groq" && /^qwen\//.test(config.model) ? { reasoning_effort: "none" } : {}
       }),
       signal: AbortSignal.timeout(request.deadlineMs)
     });
@@ -176,7 +182,28 @@ async function call(config, request, started) {
     cached: false
   };
 }
-function parseJsonContent(content) {
+async function localCall(spec, request) {
+  const started = Date.now();
+  const cacheKey = request.cacheKey && `local:${spec.id}:${request.cacheKey}`;
+  if (cacheKey) {
+    const parsed = request.schema.safeParse(await cachedResponse(request.repositoryRoot, cacheKey));
+    if (parsed.success) {
+      return {
+        ok: true,
+        value: parsed.data,
+        model: spec.id,
+        latencyMs: Date.now() - started,
+        tokens: 0,
+        cached: true
+      };
+    }
+  }
+  const result = await localChatJson(spec, request);
+  if (result.ok && cacheKey) await storeResponse(request.repositoryRoot, cacheKey, result.value);
+  return result;
+}
+function parseJsonContent(raw) {
+  const content = raw.replace(/<think>[\s\S]*?<\/think>/g, "");
   const fenced = content.match(/```(?:json)?\s*([\s\S]*?)```/i)?.[1];
   try {
     return JSON.parse((fenced ?? content).trim());
@@ -188,7 +215,8 @@ function estimateTokens(request, content) {
   return Math.ceil((request.system.length + request.user.length + content.length) / 4);
 }
 function redact(config, text) {
-  return redactLlmKeys(text).split(config.apiKey).join("[REDACTED]");
+  const redacted = redactLlmKeys(text);
+  return config.apiKey ? redacted.split(config.apiKey).join("[REDACTED]") : redacted;
 }
 
 export {
@@ -196,4 +224,4 @@ export {
   coolingDown,
   chatJson
 };
-//# sourceMappingURL=chunk-ESZXCZIN.js.map
+//# sourceMappingURL=chunk-MNSMW4MZ.js.map

package/dist/chunk-MNSMW4MZ.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/llm/budget.ts","../src/llm/client.ts"],"sourcesContent":["import { mkdir, readFile } from \"node:fs/promises\";\nimport { join } from \"node:path\";\nimport { writeAtomic } from \"../core/markdown-store-fs.js\";\n\n/**\n * Free-tier hygiene for the LLM layer, all in `.codebuddy/cache/` (gitignored,\n * rebuildable): a per-day token ledger per provider:model, a cooldown after\n * 429s, and a response cache. Background work stops at 70% of the daily\n * budget so prompt-time calls keep a reserved 30%.\n *\n * ponytail: read-modify-write JSON without a lock; two hook processes can lose\n * one update. Fine for a soft budget; add the staging lock if it ever matters.\n */\nconst BACKGROUND_SHARE = 0.7;\nconst CACHE_LIMIT = 2_000;\n\nconst file = (root: string, name: string) => join(root, \".codebuddy\", \"cache\", name);\n\nasync function readJson<T>(path: string, fallback: T): Promise<T> {\n  try {\n    return JSON.parse(await readFile(path, \"utf8\")) as T;\n  } catch {\n    return fallback;\n  }\n}\n\nasync function writeJson(path: string, value: unknown): Promise<void> {\n  await mkdir(join(path, \"..\"), { recursive: true });\n  await writeAtomic(path, `${JSON.stringify(value)}\\n`);\n}\n\ntype Ledger = { day: string; used: Record<string, number> };\nconst utcDay = (now: Date) => now.toISOString().slice(0, 10);\n\nexport async function usageToday(root: string, now = new Date()): Promise<Record<string, number>> {\n  const ledger = await readJson<Ledger>(file(root, \"llm-usage.json\"), { day: \"\", used: {} });\n  return ledger.day === utcDay(now) ? ledger.used : {};\n}\n\nexport async function recordUsage(\n  root: string,\n  key: string,\n  tokens: number,\n  now = new Date(),\n): Promise<void> {\n  const used = await usageToday(root, now);\n  used[key] = (used[key] ?? 0) + tokens;\n  await writeJson(file(root, \"llm-usage.json\"), { day: utcDay(now), used });\n}\n\n/** Whether `estimate` more tokens fit today's share for this purpose. */\nexport async function reserveBudget(\n  root: string,\n  key: string,\n  purpose: \"prompt\" | \"background\",\n  estimate: number,\n  dailyBudget: number,\n  now = new Date(),\n): Promise<boolean> {\n  const limit = purpose === \"background\" ? dailyBudget * BACKGROUND_SHARE : dailyBudget;\n  return ((await usageToday(root, now))[key] ?? 0) + estimate <= limit;\n}\n\n/** Milliseconds of cooldown left for this provider:model (0 = none). */\nexport async function coolingDown(root: string, key: string, now = new Date()): Promise<number> {\n  const until = (await readJson<Record<string, number>>(file(root, \"llm-cooldown.json\"), {}))[key];\n  return until ? Math.max(0, until - now.getTime()) : 0;\n}\n\nexport async function startCooldown(\n  root: string,\n  key: string,\n  seconds: number,\n  now = new Date(),\n): Promise<void> {\n  const path = file(root, \"llm-cooldown.json\");\n  const cooldowns = await readJson<Record<string, number>>(path, {});\n  cooldowns[key] = now.getTime() + seconds * 1_000;\n  await writeJson(path, cooldowns);\n}\n\nexport async function cachedResponse(root: string, key: string): Promise<unknown | undefined> {\n  return (await readJson<Record<string, unknown>>(file(root, \"llm-responses.json\"), {}))[key];\n}\n\n/** Insertion-ordered; the oldest entries go once the cache holds CACHE_LIMIT. */\nexport async function storeResponse(root: string, key: string, value: unknown): Promise<void> {\n  const path = file(root, \"llm-responses.json\");\n  const cache = await readJson<Record<string, unknown>>(path, {});\n  delete cache[key];\n  cache[key] = value;\n  const keys = Object.keys(cache);\n  for (const old of keys.slice(0, Math.max(0, keys.length - CACHE_LIMIT))) delete cache[old];\n  await writeJson(path, cache);\n}\n","import type { z } from \"zod\";\nimport {\n  cachedResponse,\n  coolingDown,\n  recordUsage,\n  reserveBudget,\n  startCooldown,\n  storeResponse,\n} from \"./budget.js\";\nimport { type EnabledLlmConfig, redactLlmKeys } from \"./config.js\";\nimport { type LocalModelSpec, localChatJson } from \"./local.js\";\n\n/** Seconds to back off after a 429 without a usable retry-after header. */\nconst DEFAULT_COOLDOWN_SECONDS = 60;\n\n/**\n * One OpenAI-compatible chat call that must answer with JSON (Groq and\n * OpenRouter share the API). It never throws: every failure is a reason the\n * caller turns into today's deterministic behaviour plus an explain line.\n */\nexport type LlmFailure =\n  | \"disabled\"\n  | \"timeout\"\n  | \"rate-limited\"\n  | \"cooldown\"\n  | \"budget\"\n  | \"http\"\n  | \"invalid\";\n\nexport type LlmResult<T> =\n  | { ok: true; value: T; model: string; latencyMs: number; tokens: number; cached: boolean }\n  | { ok: false; reason: LlmFailure; detail: string; latencyMs: number };\n\nexport type ChatJsonRequest<T> = {\n  system: string;\n  user: string;\n  schema: z.ZodType<T>;\n  maxTokens: number;\n  deadlineMs: number;\n  /** Prompt-path calls get the reserved budget share; background work stops earlier. */\n  purpose: \"prompt\" | \"background\";\n  /**\n   * Local backend only: a JSON schema for the generation grammar (with\n   * `maxItems`/`maxLength` bounds) instead of one derived from `schema`.\n   */\n  jsonSchema?: object;\n  /** Cache identity (inputs + prompt version); omitted = never cached. */\n  cacheKey?: string;\n  repositoryRoot: string;\n};\n\nexport async function chatJson<T>(\n  config: EnabledLlmConfig,\n  request: ChatJsonRequest<T>,\n): Promise<LlmResult<T>> {\n  const started = Date.now();\n  if (config.local) return localCall(config.local, request);\n  try {\n    return await call(config, request, started);\n  } catch (error) {\n    // Structural guarantee: an unexpected body shape is an invalid answer, not a crash.\n    return {\n      ok: false,\n      reason: \"invalid\",\n      detail: redact(config, String((error as Error)?.message ?? error)).slice(0, 300),\n      latencyMs: Date.now() - started,\n    };\n  }\n}\n\nasync function call<T>(\n  config: EnabledLlmConfig,\n  request: ChatJsonRequest<T>,\n  started: number,\n): Promise<LlmResult<T>> {\n  const fail = (reason: LlmFailure, detail: string): LlmResult<T> => ({\n    ok: false,\n    reason,\n    detail: redact(config, detail).slice(0, 300),\n    latencyMs: Date.now() - started,\n  });\n  const root = request.repositoryRoot;\n  const ledgerKey = `${config.provider}:${config.model}`;\n  const cooling = await coolingDown(root, ledgerKey);\n  if (cooling > 0) return fail(\"cooldown\", `rate limited; retry in ${Math.ceil(cooling / 1000)} s`);\n  const cacheKey = request.cacheKey && `${ledgerKey}:${request.cacheKey}`;\n  if (cacheKey) {\n    const parsed = request.schema.safeParse(await cachedResponse(root, cacheKey));\n    if (parsed.success) {\n      return {\n        ok: true,\n        value: parsed.data,\n        model: config.model,\n        latencyMs: Date.now() - started,\n        tokens: 0,\n        cached: true,\n      };\n    }\n  }\n  const estimate = estimateTokens(request, \"\") + request.maxTokens;\n  if (!(await reserveBudget(root, ledgerKey, request.purpose, estimate, config.dailyTokenBudget))) {\n    return fail(\"budget\", `daily ${request.purpose} budget for ${ledgerKey} is spent`);\n  }\n  let response: Response;\n  try {\n    response = await fetch(`${config.baseUrl}/chat/completions`, {\n      method: \"POST\",\n      headers: {\n        authorization: `Bearer ${config.apiKey}`,\n        \"content-type\": \"application/json\",\n        ...(config.provider === \"openrouter\" ? { \"x-title\": \"CodeBuddy\" } : {}),\n      },\n      body: JSON.stringify({\n        model: config.model,\n        messages: [\n          { role: \"system\", content: request.system },\n          { role: \"user\", content: request.user },\n        ],\n        temperature: 0,\n        max_tokens: request.maxTokens,\n        response_format: { type: \"json_object\" },\n        // Qwen3 reasons by default on some hosts; cards and reranks need none.\n        ...(config.provider === \"groq\" && /^qwen\\//.test(config.model)\n          ? { reasoning_effort: \"none\" }\n          : {}),\n      }),\n      signal: AbortSignal.timeout(request.deadlineMs),\n    });\n  } catch (error) {\n    const name = (error as Error)?.name;\n    if (name === \"TimeoutError\" || name === \"AbortError\") {\n      return fail(\"timeout\", `no answer within ${request.deadlineMs} ms`);\n    }\n    return fail(\"http\", (error as Error)?.message ?? String(error));\n  }\n  let body: string;\n  try {\n    body = await response.text();\n  } catch (error) {\n    const name = (error as Error)?.name;\n    return fail(\n      name === \"TimeoutError\" || name === \"AbortError\" ? \"timeout\" : \"http\",\n      String(error),\n    );\n  }\n  if (response.status === 429) {\n    const retryAfter = Number(response.headers.get(\"retry-after\"));\n    await startCooldown(\n      root,\n      ledgerKey,\n      Number.isFinite(retryAfter) && retryAfter > 0 ? retryAfter : DEFAULT_COOLDOWN_SECONDS,\n    );\n    return fail(\"rate-limited\", body);\n  }\n  if (!response.ok) return fail(\"http\", `HTTP ${response.status}: ${body}`);\n\n  let completion: {\n    model?: unknown;\n    choices?: Array<{ message?: { content?: unknown } }>;\n    usage?: { total_tokens?: unknown };\n  } | null;\n  try {\n    completion = JSON.parse(body);\n  } catch {\n    return fail(\"invalid\", \"response is not JSON\");\n  }\n  const raw = completion?.choices?.[0]?.message?.content;\n  if (typeof raw !== \"string\") return fail(\"invalid\", \"answer has no text content\");\n  const content = raw;\n  const reported = completion?.usage?.total_tokens;\n  const tokens =\n    typeof reported === \"number\" && Number.isFinite(reported)\n      ? reported\n      : estimateTokens(request, content);\n  await recordUsage(root, ledgerKey, tokens);\n  const parsed = request.schema.safeParse(parseJsonContent(content));\n  if (!parsed.success) return fail(\"invalid\", `output failed validation: ${content.slice(0, 120)}`);\n  if (cacheKey) await storeResponse(root, cacheKey, parsed.data);\n  return {\n    ok: true,\n    value: parsed.data,\n    model: typeof completion?.model === \"string\" ? completion.model : config.model,\n    latencyMs: Date.now() - started,\n    tokens,\n    cached: false,\n  };\n}\n\n/** Local generation: same response cache, no budget ledger or cooldown (no cost, no quota). */\nasync function localCall<T>(\n  spec: LocalModelSpec,\n  request: ChatJsonRequest<T>,\n): Promise<LlmResult<T>> {\n  const started = Date.now();\n  const cacheKey = request.cacheKey && `local:${spec.id}:${request.cacheKey}`;\n  if (cacheKey) {\n    const parsed = request.schema.safeParse(await cachedResponse(request.repositoryRoot, cacheKey));\n    if (parsed.success) {\n      return {\n        ok: true,\n        value: parsed.data,\n        model: spec.id,\n        latencyMs: Date.now() - started,\n        tokens: 0,\n        cached: true,\n      };\n    }\n  }\n  const result = await localChatJson(spec, request);\n  if (result.ok && cacheKey) await storeResponse(request.repositoryRoot, cacheKey, result.value);\n  return result;\n}\n\n/** Free models often wrap JSON in a ```json fence despite json_object mode. */\nfunction parseJsonContent(raw: string): unknown {\n  // A Qwen3 host may still prepend an (empty) reasoning block.\n  const content = raw.replace(/<think>[\\s\\S]*?<\\/think>/g, \"\");\n  const fenced = content.match(/```(?:json)?\\s*([\\s\\S]*?)```/i)?.[1];\n  try {\n    return JSON.parse((fenced ?? content).trim());\n  } catch {\n    return undefined;\n  }\n}\n\nfunction estimateTokens(request: { system: string; user: string }, content: string): number {\n  return Math.ceil((request.system.length + request.user.length + content.length) / 4);\n}\n\nfunction redact(config: EnabledLlmConfig, text: string): string {\n  const redacted = redactLlmKeys(text);\n  return config.apiKey ? redacted.split(config.apiKey).join(\"[REDACTED]\") : redacted;\n}\n"],"mappings":";;;;;;;;;;;AAAA,SAAS,OAAO,gBAAgB;AAChC,SAAS,YAAY;AAYrB,IAAM,mBAAmB;AACzB,IAAM,cAAc;AAEpB,IAAM,OAAO,CAAC,MAAc,SAAiB,KAAK,MAAM,cAAc,SAAS,IAAI;AAEnF,eAAe,SAAY,MAAc,UAAyB;AAChE,MAAI;AACF,WAAO,KAAK,MAAM,MAAM,SAAS,MAAM,MAAM,CAAC;AAAA,EAChD,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAEA,eAAe,UAAU,MAAc,OAA+B;AACpE,QAAM,MAAM,KAAK,MAAM,IAAI,GAAG,EAAE,WAAW,KAAK,CAAC;AACjD,QAAM,YAAY,MAAM,GAAG,KAAK,UAAU,KAAK,CAAC;AAAA,CAAI;AACtD;AAGA,IAAM,SAAS,CAAC,QAAc,IAAI,YAAY,EAAE,MAAM,GAAG,EAAE;AAE3D,eAAsB,WAAW,MAAc,MAAM,oBAAI,KAAK,GAAoC;AAChG,QAAM,SAAS,MAAM,SAAiB,KAAK,MAAM,gBAAgB,GAAG,EAAE,KAAK,IAAI,MAAM,CAAC,EAAE,CAAC;AACzF,SAAO,OAAO,QAAQ,OAAO,GAAG,IAAI,OAAO,OAAO,CAAC;AACrD;AAEA,eAAsB,YACpB,MACA,KACA,QACA,MAAM,oBAAI,KAAK,GACA;AACf,QAAM,OAAO,MAAM,WAAW,MAAM,GAAG;AACvC,OAAK,GAAG,KAAK,KAAK,GAAG,KAAK,KAAK;AAC/B,QAAM,UAAU,KAAK,MAAM,gBAAgB,GAAG,EAAE,KAAK,OAAO,GAAG,GAAG,KAAK,CAAC;AAC1E;AAGA,eAAsB,cACpB,MACA,KACA,SACA,UACA,aACA,MAAM,oBAAI,KAAK,GACG;AAClB,QAAM,QAAQ,YAAY,eAAe,cAAc,mBAAmB;AAC1E,WAAS,MAAM,WAAW,MAAM,GAAG,GAAG,GAAG,KAAK,KAAK,YAAY;AACjE;AAGA,eAAsB,YAAY,MAAc,KAAa,MAAM,oBAAI,KAAK,GAAoB;AAC9F,QAAM,SAAS,MAAM,SAAiC,KAAK,MAAM,mBAAmB,GAAG,CAAC,CAAC,GAAG,GAAG;AAC/F,SAAO,QAAQ,KAAK,IAAI,GAAG,QAAQ,IAAI,QAAQ,CAAC,IAAI;AACtD;AAEA,eAAsB,cACpB,MACA,KACA,SACA,MAAM,oBAAI,KAAK,GACA;AACf,QAAM,OAAO,KAAK,MAAM,mBAAmB;AAC3C,QAAM,YAAY,MAAM,SAAiC,MAAM,CAAC,CAAC;AACjE,YAAU,GAAG,IAAI,IAAI,QAAQ,IAAI,UAAU;AAC3C,QAAM,UAAU,MAAM,SAAS;AACjC;AAEA,eAAsB,eAAe,MAAc,KAA2C;AAC5F,UAAQ,MAAM,SAAkC,KAAK,MAAM,oBAAoB,GAAG,CAAC,CAAC,GAAG,GAAG;AAC5F;AAGA,eAAsB,cAAc,MAAc,KAAa,OAA+B;AAC5F,QAAM,OAAO,KAAK,MAAM,oBAAoB;AAC5C,QAAM,QAAQ,MAAM,SAAkC,MAAM,CAAC,CAAC;AAC9D,SAAO,MAAM,GAAG;AAChB,QAAM,GAAG,IAAI;AACb,QAAM,OAAO,OAAO,KAAK,KAAK;AAC9B,aAAW,OAAO,KAAK,MAAM,GAAG,KAAK,IAAI,GAAG,KAAK,SAAS,WAAW,CAAC,EAAG,QAAO,MAAM,GAAG;AACzF,QAAM,UAAU,MAAM,KAAK;AAC7B;;;ACjFA,IAAM,2BAA2B;AAsCjC,eAAsB,SACpB,QACA,SACuB;AACvB,QAAM,UAAU,KAAK,IAAI;AACzB,MAAI,OAAO,MAAO,QAAO,UAAU,OAAO,OAAO,OAAO;AACxD,MAAI;AACF,WAAO,MAAM,KAAK,QAAQ,SAAS,OAAO;AAAA,EAC5C,SAAS,OAAO;AAEd,WAAO;AAAA,MACL,IAAI;AAAA,MACJ,QAAQ;AAAA,MACR,QAAQ,OAAO,QAAQ,OAAQ,OAAiB,WAAW,KAAK,CAAC,EAAE,MAAM,GAAG,GAAG;AAAA,MAC/E,WAAW,KAAK,IAAI,IAAI;AAAA,IAC1B;AAAA,EACF;AACF;AAEA,eAAe,KACb,QACA,SACA,SACuB;AACvB,QAAM,OAAO,CAAC,QAAoB,YAAkC;AAAA,IAClE,IAAI;AAAA,IACJ;AAAA,IACA,QAAQ,OAAO,QAAQ,MAAM,EAAE,MAAM,GAAG,GAAG;AAAA,IAC3C,WAAW,KAAK,IAAI,IAAI;AAAA,EAC1B;AACA,QAAM,OAAO,QAAQ;AACrB,QAAM,YAAY,GAAG,OAAO,QAAQ,IAAI,OAAO,KAAK;AACpD,QAAM,UAAU,MAAM,YAAY,MAAM,SAAS;AACjD,MAAI,UAAU,EAAG,QAAO,KAAK,YAAY,0BAA0B,KAAK,KAAK,UAAU,GAAI,CAAC,IAAI;AAChG,QAAM,WAAW,QAAQ,YAAY,GAAG,SAAS,IAAI,QAAQ,QAAQ;AACrE,MAAI,UAAU;AACZ,UAAMA,UAAS,QAAQ,OAAO,UAAU,MAAM,eAAe,MAAM,QAAQ,CAAC;AAC5E,QAAIA,QAAO,SAAS;AAClB,aAAO;AAAA,QACL,IAAI;AAAA,QACJ,OAAOA,QAAO;AAAA,QACd,OAAO,OAAO;AAAA,QACd,WAAW,KAAK,IAAI,IAAI;AAAA,QACxB,QAAQ;AAAA,QACR,QAAQ;AAAA,MACV;AAAA,IACF;AAAA,EACF;AACA,QAAM,WAAW,eAAe,SAAS,EAAE,IAAI,QAAQ;AACvD,MAAI,CAAE,MAAM,cAAc,MAAM,WAAW,QAAQ,SAAS,UAAU,OAAO,gBAAgB,GAAI;AAC/F,WAAO,KAAK,UAAU,SAAS,QAAQ,OAAO,eAAe,SAAS,WAAW;AAAA,EACnF;AACA,MAAI;AACJ,MAAI;AACF,eAAW,MAAM,MAAM,GAAG,OAAO,OAAO,qBAAqB;AAAA,MAC3D,QAAQ;AAAA,MACR,SAAS;AAAA,QACP,eAAe,UAAU,OAAO,MAAM;AAAA,QACtC,gBAAgB;AAAA,QAChB,GAAI,OAAO,aAAa,eAAe,EAAE,WAAW,YAAY,IAAI,CAAC;AAAA,MACvE;AAAA,MACA,MAAM,KAAK,UAAU;AAAA,QACnB,OAAO,OAAO;AAAA,QACd,UAAU;AAAA,UACR,EAAE,MAAM,UAAU,SAAS,QAAQ,OAAO;AAAA,UAC1C,EAAE,MAAM,QAAQ,SAAS,QAAQ,KAAK;AAAA,QACxC;AAAA,QACA,aAAa;AAAA,QACb,YAAY,QAAQ;AAAA,QACpB,iBAAiB,EAAE,MAAM,cAAc;AAAA;AAAA,QAEvC,GAAI,OAAO,aAAa,UAAU,UAAU,KAAK,OAAO,KAAK,IACzD,EAAE,kBAAkB,OAAO,IAC3B,CAAC;AAAA,MACP,CAAC;AAAA,MACD,QAAQ,YAAY,QAAQ,QAAQ,UAAU;AAAA,IAChD,CAAC;AAAA,EACH,SAAS,OAAO;AACd,UAAM,OAAQ,OAAiB;AAC/B,QAAI,SAAS,kBAAkB,SAAS,cAAc;AACpD,aAAO,KAAK,WAAW,oBAAoB,QAAQ,UAAU,KAAK;AAAA,IACpE;AACA,WAAO,KAAK,QAAS,OAAiB,WAAW,OAAO,KAAK,CAAC;AAAA,EAChE;AACA,MAAI;AACJ,MAAI;AACF,WAAO,MAAM,SAAS,KAAK;AAAA,EAC7B,SAAS,OAAO;AACd,UAAM,OAAQ,OAAiB;AAC/B,WAAO;AAAA,MACL,SAAS,kBAAkB,SAAS,eAAe,YAAY;AAAA,MAC/D,OAAO,KAAK;AAAA,IACd;AAAA,EACF;AACA,MAAI,SAAS,WAAW,KAAK;AAC3B,UAAM,aAAa,OAAO,SAAS,QAAQ,IAAI,aAAa,CAAC;AAC7D,UAAM;AAAA,MACJ;AAAA,MACA;AAAA,MACA,OAAO,SAAS,UAAU,KAAK,aAAa,IAAI,aAAa;AAAA,IAC/D;AACA,WAAO,KAAK,gBAAgB,IAAI;AAAA,EAClC;AACA,MAAI,CAAC,SAAS,GAAI,QAAO,KAAK,QAAQ,QAAQ,SAAS,MAAM,KAAK,IAAI,EAAE;AAExE,MAAI;AAKJ,MAAI;AACF,iBAAa,KAAK,MAAM,IAAI;AAAA,EAC9B,QAAQ;AACN,WAAO,KAAK,WAAW,sBAAsB;AAAA,EAC/C;AACA,QAAM,MAAM,YAAY,UAAU,CAAC,GAAG,SAAS;AAC/C,MAAI,OAAO,QAAQ,SAAU,QAAO,KAAK,WAAW,4BAA4B;AAChF,QAAM,UAAU;AAChB,QAAM,WAAW,YAAY,OAAO;AACpC,QAAM,SACJ,OAAO,aAAa,YAAY,OAAO,SAAS,QAAQ,IACpD,WACA,eAAe,SAAS,OAAO;AACrC,QAAM,YAAY,MAAM,WAAW,MAAM;AACzC,QAAM,SAAS,QAAQ,OAAO,UAAU,iBAAiB,OAAO,CAAC;AACjE,MAAI,CAAC,OAAO,QAAS,QAAO,KAAK,WAAW,6BAA6B,QAAQ,MAAM,GAAG,GAAG,CAAC,EAAE;AAChG,MAAI,SAAU,OAAM,cAAc,MAAM,UAAU,OAAO,IAAI;AAC7D,SAAO;AAAA,IACL,IAAI;AAAA,IACJ,OAAO,OAAO;AAAA,IACd,OAAO,OAAO,YAAY,UAAU,WAAW,WAAW,QAAQ,OAAO;AAAA,IACzE,WAAW,KAAK,IAAI,IAAI;AAAA,IACxB;AAAA,IACA,QAAQ;AAAA,EACV;AACF;AAGA,eAAe,UACb,MACA,SACuB;AACvB,QAAM,UAAU,KAAK,IAAI;AACzB,QAAM,WAAW,QAAQ,YAAY,SAAS,KAAK,EAAE,IAAI,QAAQ,QAAQ;AACzE,MAAI,UAAU;AACZ,UAAM,SAAS,QAAQ,OAAO,UAAU,MAAM,eAAe,QAAQ,gBAAgB,QAAQ,CAAC;AAC9F,QAAI,OAAO,SAAS;AAClB,aAAO;AAAA,QACL,IAAI;AAAA,QACJ,OAAO,OAAO;AAAA,QACd,OAAO,KAAK;AAAA,QACZ,WAAW,KAAK,IAAI,IAAI;AAAA,QACxB,QAAQ;AAAA,QACR,QAAQ;AAAA,MACV;AAAA,IACF;AAAA,EACF;AACA,QAAM,SAAS,MAAM,cAAc,MAAM,OAAO;AAChD,MAAI,OAAO,MAAM,SAAU,OAAM,cAAc,QAAQ,gBAAgB,UAAU,OAAO,KAAK;AAC7F,SAAO;AACT;AAGA,SAAS,iBAAiB,KAAsB;AAE9C,QAAM,UAAU,IAAI,QAAQ,6BAA6B,EAAE;AAC3D,QAAM,SAAS,QAAQ,MAAM,+BAA+B,IAAI,CAAC;AACjE,MAAI;AACF,WAAO,KAAK,OAAO,UAAU,SAAS,KAAK,CAAC;AAAA,EAC9C,QAAQ;AACN,WAAO;AAAA,EACT;AACF;AAEA,SAAS,eAAe,SAA2C,SAAyB;AAC1F,SAAO,KAAK,MAAM,QAAQ,OAAO,SAAS,QAAQ,KAAK,SAAS,QAAQ,UAAU,CAAC;AACrF;AAEA,SAAS,OAAO,QAA0B,MAAsB;AAC9D,QAAM,WAAW,cAAc,IAAI;AACnC,SAAO,OAAO,SAAS,SAAS,MAAM,OAAO,MAAM,EAAE,KAAK,YAAY,IAAI;AAC5E;","names":["parsed"]}

package/dist/chunk-NNJSAYHY.js

@@ -0,0 +1,372 @@
+import {
+  modelCacheDir
+} from "./chunk-3QGW53GD.js";
+
+// src/llm/local.ts
+import { createHash } from "crypto";
+import { createWriteStream } from "fs";
+import { mkdir, rename, rm, stat, statfs, utimes, writeFile } from "fs/promises";
+import { availableParallelism, freemem, totalmem } from "os";
+import { dirname, join } from "path";
+import { Readable } from "stream";
+import { pipeline } from "stream/promises";
+import { z } from "zod";
+var GB = 1024 ** 3;
+var LOCAL_MODELS = {
+  default: {
+    size: "default",
+    id: "qwen3-0.6b-q4_k_m",
+    repo: "unsloth/Qwen3-0.6B-GGUF",
+    revision: "50968a4468ef4233ed78cd7c3de230dd1d61a56b",
+    file: "Qwen3-0.6B-Q4_K_M.gguf",
+    bytes: 396705472,
+    sha256: "ac2d97712095a558e31573f62f466a3f9d93990898b0ec79d7c974c1780d524a",
+    ramBytes: 1 * GB,
+    noThink: true
+  },
+  better: {
+    size: "better",
+    id: "qwen3-1.7b-q4_k_m",
+    repo: "unsloth/Qwen3-1.7B-GGUF",
+    revision: "d7f544eead698dbd1f15126ef60b45a1e1933222",
+    file: "Qwen3-1.7B-Q4_K_M.gguf",
+    bytes: 1107409472,
+    sha256: "b139949c5bd74937ad8ed8c8cf3d9ffb1e99c866c823204dc42c0d91fa181897",
+    ramBytes: 2.2 * GB,
+    noThink: true
+  }
+};
+var LLAMA_RUNTIME = "node-llama-cpp@3.22.3";
+function localModelPath(spec) {
+  return join(modelCacheDir(), "gguf", spec.file);
+}
+async function installedLocalModel(preferred) {
+  if (process.env.CODEBUDDY_LOCAL_LLM === "0") return null;
+  const order = [...preferred ? [preferred] : [], "default", "better"];
+  for (const size of new Set(order)) {
+    const spec = LOCAL_MODELS[size];
+    const info = await stat(localModelPath(spec)).catch(() => null);
+    if (info?.isFile() && info.size === spec.bytes) return spec;
+  }
+  return null;
+}
+async function localRuntimeAvailable() {
+  try {
+    await importRuntime();
+    return true;
+  } catch {
+    return false;
+  }
+}
+function machineCanCard() {
+  if (totalmem() < 4 * GB) {
+    return { ok: false, reason: `${formatBytes(totalmem())} RAM (setup installs it from 4 GB)` };
+  }
+  if (availableParallelism() < 2) return { ok: false, reason: "a single CPU core" };
+  return { ok: true };
+}
+async function checkLocalModelResources(spec) {
+  const directory = modelCacheDir();
+  await mkdir(directory, { recursive: true });
+  const fs = await statfs(directory).catch(() => null);
+  if (fs && fs.bavail * fs.bsize < spec.bytes * 1.1) {
+    return {
+      ok: false,
+      reason: `needs ${formatBytes(spec.bytes * 1.1)} free in ${directory}, ${formatBytes(fs.bavail * fs.bsize)} available`
+    };
+  }
+  if (totalmem() < spec.ramBytes * 2) {
+    return {
+      ok: false,
+      reason: `needs about ${formatBytes(spec.ramBytes)} of RAM while loaded; this machine has ${formatBytes(totalmem())} in total`
+    };
+  }
+  return { ok: true };
+}
+async function downloadLocalModel(spec, onProgress, fetchImpl = fetch) {
+  const target = localModelPath(spec);
+  await mkdir(dirname(target), { recursive: true });
+  const partial = `${target}.partial`;
+  const url = `https://huggingface.co/${spec.repo}/resolve/${spec.revision}/${spec.file}`;
+  const response = await fetchImpl(url);
+  if (!response.ok || !response.body) {
+    throw new Error(`download failed: HTTP ${response.status} for ${url}`);
+  }
+  const hash = createHash("sha256");
+  let received = 0;
+  const source = Readable.fromWeb(response.body);
+  source.on("data", (chunk) => {
+    hash.update(chunk);
+    received += chunk.length;
+    onProgress?.(received, spec.bytes);
+  });
+  try {
+    await pipeline(source, createWriteStream(partial));
+    const digest = hash.digest("hex");
+    if (received !== spec.bytes || digest !== spec.sha256) {
+      throw new Error(
+        `checksum mismatch for ${spec.file} (got ${received} bytes, sha256 ${digest.slice(0, 12)}\u2026); not installed`
+      );
+    }
+    await rename(partial, target);
+  } finally {
+    await rm(partial, { force: true });
+  }
+  return target;
+}
+async function removeLocalModels() {
+  const removed = [];
+  for (const spec of Object.values(LOCAL_MODELS)) {
+    const path = localModelPath(spec);
+    if (await stat(path).catch(() => null)) {
+      await rm(path, { force: true });
+      removed.push(spec.id);
+    }
+  }
+  return removed;
+}
+var loading = null;
+var loaded = null;
+var queue = Promise.resolve();
+var grammars = /* @__PURE__ */ new Map();
+async function importRuntime() {
+  const specifier = process.env.CODEBUDDY_LLAMA_RUNTIME ?? "node-llama-cpp";
+  return await import(specifier);
+}
+var CONTEXT_TOKENS = 4096;
+function generationThreads() {
+  const configured = Number(process.env.CODEBUDDY_LLM_THREADS);
+  if (Number.isInteger(configured) && configured > 0) return configured;
+  return Math.max(1, availableParallelism() - 1);
+}
+async function load(spec) {
+  if (loaded?.spec.id === spec.id) return loaded;
+  if (!loading) {
+    loading = (async () => {
+      const runtime = await importRuntime();
+      const llama = await runtime.getLlama({ gpu: "auto" });
+      const model = await llama.loadModel({ modelPath: localModelPath(spec) });
+      const context = await model.createContext({
+        contextSize: CONTEXT_TOKENS,
+        threads: generationThreads()
+      });
+      loaded = { spec, llama, model, context, sequence: context.getSequence(), runtime };
+      return loaded;
+    })().finally(() => {
+      loading = null;
+    });
+  }
+  return loading;
+}
+var IDLE_UNLOAD_MS = 6e4;
+var idleTimer = null;
+var running = 0;
+function scheduleUnload() {
+  if (idleTimer) clearTimeout(idleTimer);
+  idleTimer = setTimeout(() => void unloadLocalModel(), IDLE_UNLOAD_MS);
+  idleTimer.unref();
+}
+async function unloadLocalModel() {
+  if (idleTimer) clearTimeout(idleTimer);
+  idleTimer = null;
+  const current = loaded;
+  if (!current || running > 0) return;
+  loaded = null;
+  grammars.clear();
+  await current.context.dispose().catch(() => {
+  });
+  await current.model.dispose().catch(() => {
+  });
+}
+function isLocalModelWarm() {
+  return loaded !== null;
+}
+async function warmLocalModel(spec) {
+  try {
+    await load(spec);
+    return true;
+  } catch {
+    return false;
+  }
+}
+async function localChatJson(spec, request) {
+  const started = Date.now();
+  const fail = (reason, detail) => ({
+    ok: false,
+    reason,
+    detail: detail.slice(0, 300),
+    latencyMs: Date.now() - started
+  });
+  let model;
+  try {
+    model = await withDeadline(load(spec), request.deadlineMs);
+  } catch (error) {
+    return error.name === "TimeoutError" ? fail("timeout", "model still loading") : fail("disabled", `local model unavailable: ${error.message}`);
+  }
+  const run = async () => {
+    running++;
+    try {
+      return await generate();
+    } finally {
+      running--;
+      scheduleUnload();
+    }
+  };
+  const generate = async () => {
+    if (!loaded) model = await load(spec);
+    const remaining = request.deadlineMs - (Date.now() - started);
+    if (remaining <= 0) return fail("timeout", `no answer within ${request.deadlineMs} ms`);
+    const grammar = await grammarFor(model.llama, request.jsonSchema ?? request.schema);
+    const session = new model.runtime.LlamaChatSession({
+      contextSequence: model.sequence,
+      autoDisposeSequence: false,
+      systemPrompt: request.system
+    });
+    const controller = new AbortController();
+    const timer = setTimeout(() => controller.abort(), remaining);
+    try {
+      const text = await session.prompt(`${spec.noThink ? "/no_think\n" : ""}${request.user}`, {
+        grammar,
+        maxTokens: request.maxTokens,
+        temperature: 0,
+        ...spec.noThink ? { budgets: { thoughtTokens: 0 } } : {},
+        signal: controller.signal,
+        stopOnAbortSignal: true
+      });
+      if (controller.signal.aborted) {
+        return fail("timeout", `no answer within ${request.deadlineMs} ms`);
+      }
+      let value;
+      try {
+        value = JSON.parse(text);
+      } catch {
+        return fail("invalid", `output is not JSON: ${text.slice(0, 120)}`);
+      }
+      const parsed = request.schema.safeParse(value);
+      if (!parsed.success)
+        return fail("invalid", `output failed validation: ${text.slice(0, 120)}`);
+      return {
+        ok: true,
+        value: parsed.data,
+        model: spec.id,
+        latencyMs: Date.now() - started,
+        tokens: Math.ceil((request.system.length + request.user.length + text.length) / 4),
+        cached: false
+      };
+    } catch (error) {
+      return controller.signal.aborted ? fail("timeout", `no answer within ${request.deadlineMs} ms`) : fail("invalid", String(error?.message ?? error));
+    } finally {
+      clearTimeout(timer);
+      session.dispose({ disposeSequence: false });
+    }
+  };
+  const result = queue.then(run, run);
+  queue = result.catch(() => void 0);
+  return result;
+}
+async function grammarFor(llama, schema) {
+  let grammar = grammars.get(schema);
+  if (!grammar) {
+    grammar = (async () => {
+      try {
+        const json = schema instanceof z.ZodType ? z.toJSONSchema(schema, { io: "input", unrepresentable: "any" }) : schema;
+        return await llama.createGrammarForJsonSchema(stripSchemaMeta(json));
+      } catch {
+        return llama.getGrammarFor("json");
+      }
+    })();
+    grammars.set(schema, grammar);
+  }
+  return grammar;
+}
+function stripSchemaMeta(value) {
+  if (Array.isArray(value)) return value.map(stripSchemaMeta);
+  if (!value || typeof value !== "object") return value;
+  const out = {};
+  for (const [key, child] of Object.entries(value)) {
+    if (key === "$schema" || key === "default" || key === "additionalProperties") continue;
+    out[key] = stripSchemaMeta(child);
+  }
+  return out;
+}
+function withDeadline(promise, ms) {
+  return new Promise((resolve, reject) => {
+    const timer = setTimeout(
+      () => reject(Object.assign(new Error("deadline"), { name: "TimeoutError" })),
+      ms
+    );
+    promise.then(
+      (value) => {
+        clearTimeout(timer);
+        resolve(value);
+      },
+      (error) => {
+        clearTimeout(timer);
+        reject(error);
+      }
+    );
+  });
+}
+var FOREGROUND_WINDOW_MS = 2e3;
+var FOREGROUND_MAX_WAIT_MS = 15e3;
+function foregroundPath(root) {
+  return join(root, ".codebuddy", "cache", "foreground");
+}
+async function markForeground(root) {
+  const path = foregroundPath(root);
+  const now = /* @__PURE__ */ new Date();
+  try {
+    await utimes(path, now, now);
+  } catch {
+    await mkdir(dirname(path), { recursive: true }).catch(() => {
+    });
+    await writeFile(path, "").catch(() => {
+    });
+  }
+}
+async function waitForForeground(root, signal) {
+  const until = Date.now() + FOREGROUND_MAX_WAIT_MS;
+  while (Date.now() < until && !signal?.aborted) {
+    const info = await stat(foregroundPath(root)).catch(() => null);
+    if (!info || Date.now() - info.mtimeMs > FOREGROUND_WINDOW_MS) return;
+    await new Promise((resolve) => setTimeout(resolve, 250));
+  }
+}
+async function localModelStatus(preferred) {
+  const installed = await installedLocalModel(preferred);
+  return {
+    runtime: await localRuntimeAvailable(),
+    installed,
+    path: installed ? localModelPath(installed) : null
+  };
+}
+function formatBytes(bytes) {
+  return bytes >= GB ? `${(bytes / GB).toFixed(1)} GB` : `${Math.round(bytes / 1024 ** 2)} MB`;
+}
+function freeMemory() {
+  return freemem();
+}
+
+export {
+  LOCAL_MODELS,
+  LLAMA_RUNTIME,
+  localModelPath,
+  installedLocalModel,
+  localRuntimeAvailable,
+  machineCanCard,
+  checkLocalModelResources,
+  downloadLocalModel,
+  removeLocalModels,
+  generationThreads,
+  IDLE_UNLOAD_MS,
+  unloadLocalModel,
+  isLocalModelWarm,
+  warmLocalModel,
+  localChatJson,
+  markForeground,
+  waitForForeground,
+  localModelStatus,
+  formatBytes,
+  freeMemory
+};
+//# sourceMappingURL=chunk-NNJSAYHY.js.map