@vibgrate/cli 2026.722.2 → 2026.727.2

package/dist/chunk-PXHNTC6D.js

@@ -0,0 +1,382 @@
+import { acquireHostSession, generateOnSession } from './chunk-K7REHDQS.js';
+import { annotateUnknownIdentifiers } from './chunk-JIBOJTGS.js';
+import { vibgrateRuntimeDir } from './chunk-BWPBB45M.js';
+import * as path from 'path';
+import * as fs from 'fs';
+import * as net from 'net';
+
+// src/runtime/llm-host/binding.ts
+async function generateWithLlamaBinding(input) {
+  const lib = input.lib;
+  if (!lib || typeof lib.getLlama !== "function") {
+    throw new Error("llm-host: invalid node-llama-cpp module (getLlama missing)");
+  }
+  const llama = await lib.getLlama();
+  const model = await llama.loadModel({ modelPath: input.modelPath });
+  const ctx = await model.createContext();
+  const { LlamaChatSession } = lib;
+  if (typeof LlamaChatSession !== "function") {
+    throw new Error("llm-host: LlamaChatSession missing from node-llama-cpp module");
+  }
+  const chat = new LlamaChatSession({ contextSequence: ctx.getSequence() });
+  const prompt = input.messages.map((m) => `${m.role}: ${m.content}`).join("\n\n");
+  const temperature = input.opts?.temperature ?? 0;
+  const maxTokens = input.opts?.maxTokens;
+  const promptOpts = { temperature };
+  if (typeof maxTokens === "number") promptOpts.maxTokens = maxTokens;
+  if (input.opts?.grammar) {
+    promptOpts.grammar = input.opts.grammar;
+  }
+  const text = await chat.prompt(prompt, promptOpts);
+  return typeof text === "string" ? text : String(text ?? "");
+}
+
+// src/runtime/llm-host/embedded.ts
+var EmbeddedLlmHost = class {
+  isolation = "embedded";
+  modelRef = null;
+  loaded = false;
+  binding = null;
+  preferGrammar = false;
+  lastReport = null;
+  /** Attach the dynamically loaded node-llama-cpp module. */
+  setBinding(lib) {
+    this.binding = lib;
+  }
+  /** Hint from MEP: attempt constrained decoding when grammar is supplied. */
+  setPreferGrammar(on) {
+    this.preferGrammar = on;
+  }
+  /** Last generate metrics (KV plan, draft accept, warm start). */
+  lastGenerateReport() {
+    return this.lastReport;
+  }
+  async load(modelRef) {
+    this.modelRef = modelRef;
+    this.loaded = true;
+    if (this.binding) {
+      await acquireHostSession(this.binding, modelRef);
+    }
+  }
+  async unload() {
+    this.modelRef = null;
+    this.loaded = false;
+    this.lastReport = null;
+  }
+  supportsConstrainedDecoding() {
+    return this.preferGrammar && this.binding != null;
+  }
+  memoryReport() {
+    return {
+      isolation: this.isolation,
+      modelRef: this.modelRef,
+      loaded: this.loaded
+    };
+  }
+  async generate(messages, opts = {}) {
+    if (!this.loaded || !this.modelRef) {
+      throw new Error("llm-host: no model loaded \u2014 orchestrator must call load() after resolve/install");
+    }
+    if (!this.binding) {
+      throw new Error(
+        "llm-host: node-llama-cpp binding not attached \u2014 run `vg models install` for the llama-cpp provider, then retry"
+      );
+    }
+    try {
+      const session = await acquireHostSession(this.binding, this.modelRef);
+      const report = await generateOnSession(session, messages, {
+        grammar: opts.grammar,
+        requireGrammar: opts.requireGrammar,
+        identifierTrie: opts.identifierTrie,
+        annotateUnknownIdentifiers: opts.annotateUnknownIdentifiers,
+        draftCandidates: opts.draftCandidates,
+        maxTokens: opts.maxTokens,
+        temperature: opts.temperature,
+        promptSegments: opts.promptSegments
+      });
+      this.lastReport = report;
+      return {
+        text: report.text,
+        model: report.model,
+        isolation: report.isolation,
+        constrained: report.constrained || !!(opts.grammar && this.preferGrammar),
+        unknownIdentifiers: report.unknownIdentifiers
+      };
+    } catch {
+      let text = await generateWithLlamaBinding({
+        lib: this.binding,
+        modelPath: this.modelRef,
+        messages,
+        opts
+      });
+      let unknownIdentifiers;
+      if (opts.identifierTrie && opts.annotateUnknownIdentifiers !== false) {
+        const ann = annotateUnknownIdentifiers(text, opts.identifierTrie);
+        text = ann.text;
+        unknownIdentifiers = ann.unknown.length ? ann.unknown : void 0;
+      }
+      return {
+        text,
+        model: this.modelRef,
+        isolation: this.isolation,
+        constrained: !!(opts.grammar && this.supportsConstrainedDecoding()),
+        unknownIdentifiers
+      };
+    }
+  }
+};
+var LLM_HOST_IPC_SCHEMA = "llm-host-ipc/0";
+function startProcessHostServer(options) {
+  const state = {
+    modelRef: null,
+    loaded: false,
+    binding: options.binding ?? null
+  };
+  try {
+    if (fs.existsSync(options.socketPath)) fs.unlinkSync(options.socketPath);
+  } catch {
+  }
+  fs.mkdirSync(path.dirname(options.socketPath), { recursive: true });
+  return new Promise((resolve, reject) => {
+    const server = net.createServer((socket) => {
+      let buf = "";
+      socket.setEncoding("utf8");
+      socket.on("data", (chunk) => {
+        buf += chunk;
+        let nl;
+        while ((nl = buf.indexOf("\n")) >= 0) {
+          const line = buf.slice(0, nl).trim();
+          buf = buf.slice(nl + 1);
+          if (!line) continue;
+          void handleLine(line, state, (resp) => {
+            socket.write(`${JSON.stringify(resp)}
+`);
+          });
+        }
+      });
+    });
+    server.once("error", reject);
+    server.listen(options.socketPath, () => {
+      options.onListen?.(options.socketPath);
+      resolve({
+        close: () => new Promise((res, rej) => {
+          server.close((err) => {
+            try {
+              if (fs.existsSync(options.socketPath)) fs.unlinkSync(options.socketPath);
+            } catch {
+            }
+            if (err) rej(err);
+            else res();
+          });
+        })
+      });
+    });
+  });
+}
+async function handleLine(line, state, reply) {
+  let req;
+  try {
+    req = JSON.parse(line);
+  } catch {
+    reply({ v: LLM_HOST_IPC_SCHEMA, id: "?", ok: false, error: "invalid JSON" });
+    return;
+  }
+  if (req.v !== LLM_HOST_IPC_SCHEMA) {
+    reply({ v: LLM_HOST_IPC_SCHEMA, id: req.id, ok: false, error: `unsupported schema ${String(req.v)}` });
+    return;
+  }
+  try {
+    switch (req.op) {
+      case "ping":
+        reply({ v: LLM_HOST_IPC_SCHEMA, id: req.id, ok: true, op: "ping" });
+        return;
+      case "load":
+        state.modelRef = req.modelRef;
+        state.loaded = true;
+        reply({ v: LLM_HOST_IPC_SCHEMA, id: req.id, ok: true, op: "load" });
+        return;
+      case "unload":
+        state.modelRef = null;
+        state.loaded = false;
+        reply({ v: LLM_HOST_IPC_SCHEMA, id: req.id, ok: true, op: "unload" });
+        return;
+      case "memory":
+        reply({
+          v: LLM_HOST_IPC_SCHEMA,
+          id: req.id,
+          ok: true,
+          op: "memory",
+          report: { isolation: "process", modelRef: state.modelRef, loaded: state.loaded }
+        });
+        return;
+      case "capabilities":
+        reply({
+          v: LLM_HOST_IPC_SCHEMA,
+          id: req.id,
+          ok: true,
+          op: "capabilities",
+          constrainedDecoding: state.binding != null,
+          isolation: "process"
+        });
+        return;
+      case "generate": {
+        if (!state.loaded || !state.modelRef) {
+          reply({ v: LLM_HOST_IPC_SCHEMA, id: req.id, ok: false, error: "no model loaded" });
+          return;
+        }
+        if (!state.binding) {
+          reply({
+            v: LLM_HOST_IPC_SCHEMA,
+            id: req.id,
+            ok: false,
+            error: "binding not loaded in host process \u2014 start host after vg models install"
+          });
+          return;
+        }
+        let result;
+        try {
+          const session = await acquireHostSession(state.binding, state.modelRef);
+          const report = await generateOnSession(session, req.messages, {
+            grammar: req.opts?.grammar,
+            requireGrammar: req.opts?.requireGrammar,
+            identifierTrie: req.opts?.identifierTrie,
+            annotateUnknownIdentifiers: req.opts?.annotateUnknownIdentifiers,
+            draftCandidates: req.opts?.draftCandidates,
+            maxTokens: req.opts?.maxTokens,
+            temperature: req.opts?.temperature,
+            promptSegments: req.opts?.promptSegments
+          });
+          result = {
+            text: report.text,
+            model: state.modelRef,
+            isolation: "process",
+            constrained: report.constrained,
+            unknownIdentifiers: report.unknownIdentifiers
+          };
+        } catch {
+          const text = await generateWithLlamaBinding({
+            lib: state.binding,
+            modelPath: state.modelRef,
+            messages: req.messages,
+            opts: req.opts
+          });
+          result = {
+            text,
+            model: state.modelRef,
+            isolation: "process",
+            constrained: !!(req.opts?.grammar && state.binding)
+          };
+        }
+        reply({ v: LLM_HOST_IPC_SCHEMA, id: req.id, ok: true, op: "generate", result });
+        return;
+      }
+      default:
+        reply({ v: LLM_HOST_IPC_SCHEMA, id: req.id, ok: false, error: "unknown op" });
+    }
+  } catch (e) {
+    reply({
+      v: LLM_HOST_IPC_SCHEMA,
+      id: req.id,
+      ok: false,
+      error: e instanceof Error ? e.message : String(e)
+    });
+  }
+}
+var ProcessLlmHost = class {
+  constructor(socketPath) {
+    this.socketPath = socketPath;
+  }
+  socketPath;
+  isolation = "process";
+  seq = 0;
+  getEndpoint() {
+    return this.socketPath;
+  }
+  supportsConstrainedDecoding() {
+    return true;
+  }
+  async load(modelRef) {
+    await this.rpc({ op: "load", modelRef });
+  }
+  async unload() {
+    await this.rpc({ op: "unload" });
+  }
+  async generate(messages, opts) {
+    const res = await this.rpc({ op: "generate", messages, opts });
+    if (!res.ok || res.op !== "generate") {
+      throw new Error(!res.ok ? res.error : "unexpected response");
+    }
+    return res.result;
+  }
+  memoryReport() {
+    return { isolation: "process", modelRef: null, loaded: false };
+  }
+  async memoryReportAsync() {
+    const res = await this.rpc({ op: "memory" });
+    if (!res.ok || res.op !== "memory") {
+      throw new Error(!res.ok ? res.error : "unexpected response");
+    }
+    return res.report;
+  }
+  rpc(body) {
+    const id = `r${++this.seq}`;
+    const req = { v: LLM_HOST_IPC_SCHEMA, id, ...body };
+    return new Promise((resolve, reject) => {
+      const socket = net.createConnection(this.socketPath);
+      let buf = "";
+      const timer = setTimeout(() => {
+        socket.destroy();
+        reject(new Error(`llm-host IPC timeout (${this.socketPath})`));
+      }, 12e4);
+      socket.setEncoding("utf8");
+      socket.on("connect", () => {
+        socket.write(`${JSON.stringify(req)}
+`);
+      });
+      socket.on("data", (chunk) => {
+        buf += chunk;
+        const nl = buf.indexOf("\n");
+        if (nl < 0) return;
+        const line = buf.slice(0, nl).trim();
+        clearTimeout(timer);
+        socket.end();
+        try {
+          resolve(JSON.parse(line));
+        } catch (e) {
+          reject(e);
+        }
+      });
+      socket.on("error", (e) => {
+        clearTimeout(timer);
+        reject(e);
+      });
+    });
+  }
+};
+var ProcessBridgeLlmHost = class extends ProcessLlmHost {
+};
+
+// src/runtime/llm-host/index.ts
+function createLlmHost(options = {}) {
+  const isolation = options.isolation ?? "embedded";
+  if (isolation === "process") {
+    return new ProcessLlmHost(options.endpoint ?? defaultProcessEndpoint());
+  }
+  const host = new EmbeddedLlmHost();
+  if (options.binding) host.setBinding(options.binding);
+  if (options.preferGrammar) host.setPreferGrammar(true);
+  return host;
+}
+function defaultProcessEndpoint(env = process.env) {
+  return path.join(vibgrateRuntimeDir(env), "llm-host.sock");
+}
+function resolveInferenceIsolation(explicit, env = process.env) {
+  if (explicit === "embedded" || explicit === "process") return explicit;
+  const v = (env.VIBGRATE_INFERENCE_ISOLATION || env.VG_INFERENCE_ISOLATION || "").toLowerCase();
+  if (v === "process" || v === "isolated") return "process";
+  return "embedded";
+}
+
+export { EmbeddedLlmHost, LLM_HOST_IPC_SCHEMA, ProcessBridgeLlmHost, ProcessLlmHost, createLlmHost, defaultProcessEndpoint, generateWithLlamaBinding, resolveInferenceIsolation, startProcessHostServer };
+//# sourceMappingURL=chunk-PXHNTC6D.js.map
+//# sourceMappingURL=chunk-PXHNTC6D.js.map

package/dist/chunk-PXHNTC6D.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/runtime/llm-host/binding.ts","../src/runtime/llm-host/embedded.ts","../src/runtime/llm-host/process-bridge.ts","../src/runtime/llm-host/index.ts"],"names":["path2"],"mappings":";;;;;;;;AAoBA,eAAsB,yBAAyB,KAAA,EAAmD;AAChG,EAAA,MAAM,MAAW,KAAA,CAAM,GAAA;AACvB,EAAA,IAAI,CAAC,GAAA,IAAO,OAAO,GAAA,CAAI,aAAa,UAAA,EAAY;AAC9C,IAAA,MAAM,IAAI,MAAM,4DAA4D,CAAA;AAAA,EAC9E;AACA,EAAA,MAAM,KAAA,GAAQ,MAAM,GAAA,CAAI,QAAA,EAAS;AACjC,EAAA,MAAM,KAAA,GAAQ,MAAM,KAAA,CAAM,SAAA,CAAU,EAAE,SAAA,EAAW,KAAA,CAAM,WAAW,CAAA;AAClE,EAAA,MAAM,GAAA,GAAM,MAAM,KAAA,CAAM,aAAA,EAAc;AACtC,EAAA,MAAM,EAAE,kBAAiB,GAAI,GAAA;AAC7B,EAAA,IAAI,OAAO,qBAAqB,UAAA,EAAY;AAC1C,IAAA,MAAM,IAAI,MAAM,+DAA+D,CAAA;AAAA,EACjF;AACA,EAAA,MAAM,IAAA,GAAO,IAAI,gBAAA,CAAiB,EAAE,iBAAiB,GAAA,CAAI,WAAA,IAAe,CAAA;AACxE,EAAA,MAAM,MAAA,GAAS,KAAA,CAAM,QAAA,CAAS,GAAA,CAAI,CAAC,CAAA,KAAM,CAAA,EAAG,CAAA,CAAE,IAAI,KAAK,CAAA,CAAE,OAAO,CAAA,CAAE,CAAA,CAAE,KAAK,MAAM,CAAA;AAC/E,EAAA,MAAM,WAAA,GAAc,KAAA,CAAM,IAAA,EAAM,WAAA,IAAe,CAAA;AAC/C,EAAA,MAAM,SAAA,GAAY,MAAM,IAAA,EAAM,SAAA;AAE9B,EAAA,MAAM,UAAA,GAAsC,EAAE,WAAA,EAAY;AAC1D,EAAA,IAAI,OAAO,SAAA,KAAc,QAAA,EAAU,UAAA,CAAW,SAAA,GAAY,SAAA;AAC1D,EAAA,IAAI,KAAA,CAAM,MAAM,OAAA,EAAS;AAEvB,IAAA,UAAA,CAAW,OAAA,GAAU,MAAM,IAAA,CAAK,OAAA;AAAA,EAClC;AACA,EAAA,MAAM,IAAA,GAAe,MAAM,IAAA,CAAK,MAAA,CAAO,QAAQ,UAAU,CAAA;AACzD,EAAA,OAAO,OAAO,IAAA,KAAS,QAAA,GAAW,IAAA,GAAO,MAAA,CAAO,QAAQ,EAAE,CAAA;AAC5D;;;ACvBO,IAAM,kBAAN,MAAyC;AAAA,EACrC,SAAA,GAAY,UAAA;AAAA,EACb,QAAA,GAA0B,IAAA;AAAA,EAC1B,MAAA,GAAS,KAAA;AAAA,EACT,OAAA,GAA0B,IAAA;AAAA,EAC1B,aAAA,GAAgB,KAAA;AAAA,EAChB,UAAA,GAAwC,IAAA;AAAA;AAAA,EAGhD,WAAW,GAAA,EAAoB;AAC7B,IAAA,IAAA,CAAK,OAAA,GAAU,GAAA;AAAA,EACjB;AAAA;AAAA,EAGA,iBAAiB,EAAA,EAAmB;AAClC,IAAA,IAAA,CAAK,aAAA,GAAgB,EAAA;AAAA,EACvB;AAAA;AAAA,EAGA,kBAAA,GAAgD;AAC9C,IAAA,OAAO,IAAA,CAAK,UAAA;AAAA,EACd;AAAA,EAEA,MAAM,KAAK,QAAA,EAAiC;AAC1C,IAAA,IAAA,CAAK,QAAA,GAAW,QAAA;AAChB,IAAA,IAAA,CAAK,MAAA,GAAS,IAAA;AAEd,IAAA,IAAI,KAAK,OAAA,EAAS;AAChB,MAAA,MAAM,kBAAA,CAAmB,IAAA,CAAK,OAAA,EAAS,QAAQ,CAAA;AAAA,IACjD;AAAA,EACF;AAAA,EAEA,MAAM,MAAA,GAAwB;AAC5B,IAAA,IAAA,CAAK,QAAA,GAAW,IAAA;AAChB,IAAA,IAAA,CAAK,MAAA,GAAS,KAAA;AACd,IAAA,IAAA,CAAK,UAAA,GAAa,IAAA;AAAA,EACpB;AAAA,EAEA,2BAAA,GAAuC;AACrC,IAAA,OAAO,IAAA,CAAK,aAAA,IAAiB,IAAA,CAAK,OAAA,IAAW,IAAA;AAAA,EAC/C;AAAA,EAEA,YAAA,GAAgC;AAC9B,IAAA,OAAO;AAAA,MACL,WAAW,IAAA,CAAK,SAAA;AAAA,MAChB,UAAU,IAAA,CAAK,QAAA;AAAA,MACf,QAAQ,IAAA,CAAK;AAAA,KACf;AAAA,EACF;AAAA,EAEA,MAAM,QAAA,CAAS,QAAA,EAA4B,IAAA,GAA2B,EAAC,EAA+B;AACpG,IAAA,IAAI,CAAC,IAAA,CAAK,MAAA,IAAU,CAAC,KAAK,QAAA,EAAU;AAClC,MAAA,MAAM,IAAI,MAAM,sFAAiF,CAAA;AAAA,IACnG;AACA,IAAA,IAAI,CAAC,KAAK,OAAA,EAAS;AACjB,MAAA,MAAM,IAAI,KAAA;AAAA,QACR;AAAA,OACF;AAAA,IACF;AAGA,IAAA,IAAI;AACF,MAAA,MAAM,UAAU,MAAM,kBAAA,CAAmB,IAAA,CAAK,OAAA,EAAS,KAAK,QAAQ,CAAA;AACpE,MAAA,MAAM,MAAA,GAAS,MAAM,iBAAA,CAAkB,OAAA,EAAS,QAAA,EAAU;AAAA,QACxD,SAAS,IAAA,CAAK,OAAA;AAAA,QACd,gBAAgB,IAAA,CAAK,cAAA;AAAA,QACrB,gBAAgB,IAAA,CAAK,cAAA;AAAA,QACrB,4BAA4B,IAAA,CAAK,0BAAA;AAAA,QACjC,iBAAiB,IAAA,CAAK,eAAA;AAAA,QACtB,WAAW,IAAA,CAAK,SAAA;AAAA,QAChB,aAAa,IAAA,CAAK,WAAA;AAAA,QAClB,gBAAgB,IAAA,CAAK;AAAA,OACtB,CAAA;AACD,MAAA,IAAA,CAAK,UAAA,GAAa,MAAA;AAClB,MAAA,OAAO;AAAA,QACL,MAAM,MAAA,CAAO,IAAA;AAAA,QACb,OAAO,MAAA,CAAO,KAAA;AAAA,QACd,WAAW,MAAA,CAAO,SAAA;AAAA,QAClB,aAAa,MAAA,CAAO,WAAA,IAAe,CAAC,EAAE,IAAA,CAAK,WAAW,IAAA,CAAK,aAAA,CAAA;AAAA,QAC3D,oBAAoB,MAAA,CAAO;AAAA,OAC7B;AAAA,IACF,CAAA,CAAA,MAAQ;AAEN,MAAA,IAAI,IAAA,GAAO,MAAM,wBAAA,CAAyB;AAAA,QACxC,KAAK,IAAA,CAAK,OAAA;AAAA,QACV,WAAW,IAAA,CAAK,QAAA;AAAA,QAChB,QAAA;AAAA,QACA;AAAA,OACD,CAAA;AACD,MAAA,IAAI,kBAAA;AACJ,MAAA,IAAI,IAAA,CAAK,cAAA,IAAkB,IAAA,CAAK,0BAAA,KAA+B,KAAA,EAAO;AACpE,QAAA,MAAM,GAAA,GAAM,0BAAA,CAA2B,IAAA,EAAM,IAAA,CAAK,cAAc,CAAA;AAChE,QAAA,IAAA,GAAO,GAAA,CAAI,IAAA;AACX,QAAA,kBAAA,GAAqB,GAAA,CAAI,OAAA,CAAQ,MAAA,GAAS,GAAA,CAAI,OAAA,GAAU,MAAA;AAAA,MAC1D;AACA,MAAA,OAAO;AAAA,QACL,IAAA;AAAA,QACA,OAAO,IAAA,CAAK,QAAA;AAAA,QACZ,WAAW,IAAA,CAAK,SAAA;AAAA,QAChB,aAAa,CAAC,EAAE,IAAA,CAAK,OAAA,IAAW,KAAK,2BAAA,EAA4B,CAAA;AAAA,QACjE;AAAA,OACF;AAAA,IACF;AAAA,EACF;AACF;AC1GO,IAAM,mBAAA,GAAsB;AA0C5B,SAAS,uBAAuB,OAAA,EAA4E;AACjH,EAAA,MAAM,KAAA,GAAQ;AAAA,IACZ,QAAA,EAAU,IAAA;AAAA,IACV,MAAA,EAAQ,KAAA;AAAA,IACR,OAAA,EAAS,QAAQ,OAAA,IAAW;AAAA,GAC9B;AAEA,EAAA,IAAI;AACF,IAAA,IAAO,cAAW,OAAA,CAAQ,UAAU,GAAM,EAAA,CAAA,UAAA,CAAW,QAAQ,UAAU,CAAA;AAAA,EACzE,CAAA,CAAA,MAAQ;AAAA,EAER;AACA,EAAG,EAAA,CAAA,SAAA,CAAe,aAAQ,OAAA,CAAQ,UAAU,GAAG,EAAE,SAAA,EAAW,MAAM,CAAA;AAElE,EAAA,OAAO,IAAI,OAAA,CAAQ,CAAC,OAAA,EAAS,MAAA,KAAW;AACtC,IAAA,MAAM,MAAA,GAAa,GAAA,CAAA,YAAA,CAAa,CAAC,MAAA,KAAW;AAC1C,MAAA,IAAI,GAAA,GAAM,EAAA;AACV,MAAA,MAAA,CAAO,YAAY,MAAM,CAAA;AACzB,MAAA,MAAA,CAAO,EAAA,CAAG,MAAA,EAAQ,CAAC,KAAA,KAAkB;AACnC,QAAA,GAAA,IAAO,KAAA;AACP,QAAA,IAAI,EAAA;AACJ,QAAA,OAAA,CAAQ,EAAA,GAAK,GAAA,CAAI,OAAA,CAAQ,IAAI,MAAM,CAAA,EAAG;AACpC,UAAA,MAAM,OAAO,GAAA,CAAI,KAAA,CAAM,CAAA,EAAG,EAAE,EAAE,IAAA,EAAK;AACnC,UAAA,GAAA,GAAM,GAAA,CAAI,KAAA,CAAM,EAAA,GAAK,CAAC,CAAA;AACtB,UAAA,IAAI,CAAC,IAAA,EAAM;AACX,UAAA,KAAK,UAAA,CAAW,IAAA,EAAM,KAAA,EAAO,CAAC,IAAA,KAAS;AACrC,YAAA,MAAA,CAAO,KAAA,CAAM,CAAA,EAAG,IAAA,CAAK,SAAA,CAAU,IAAI,CAAC;AAAA,CAAI,CAAA;AAAA,UAC1C,CAAC,CAAA;AAAA,QACH;AAAA,MACF,CAAC,CAAA;AAAA,IACH,CAAC,CAAA;AAED,IAAA,MAAA,CAAO,IAAA,CAAK,SAAS,MAAM,CAAA;AAC3B,IAAA,MAAA,CAAO,MAAA,CAAO,OAAA,CAAQ,UAAA,EAAY,MAAM;AACtC,MAAA,OAAA,CAAQ,QAAA,GAAW,QAAQ,UAAU,CAAA;AACrC,MAAA,OAAA,CAAQ;AAAA,QACN,OAAO,MACL,IAAI,OAAA,CAAQ,CAAC,KAAK,GAAA,KAAQ;AACxB,UAAA,MAAA,CAAO,KAAA,CAAM,CAAC,GAAA,KAAQ;AACpB,YAAA,IAAI;AACF,cAAA,IAAO,cAAW,OAAA,CAAQ,UAAU,GAAM,EAAA,CAAA,UAAA,CAAW,QAAQ,UAAU,CAAA;AAAA,YACzE,CAAA,CAAA,MAAQ;AAAA,YAER;AACA,YAAA,IAAI,GAAA,MAAS,GAAG,CAAA;AAAA,iBACX,GAAA,EAAI;AAAA,UACX,CAAC,CAAA;AAAA,QACH,CAAC;AAAA,OACJ,CAAA;AAAA,IACH,CAAC,CAAA;AAAA,EACH,CAAC,CAAA;AACH;AASA,eAAe,UAAA,CACb,IAAA,EACA,KAAA,EACA,KAAA,EACe;AACf,EAAA,IAAI,GAAA;AACJ,EAAA,IAAI;AACF,IAAA,GAAA,GAAM,IAAA,CAAK,MAAM,IAAI,CAAA;AAAA,EACvB,CAAA,CAAA,MAAQ;AACN,IAAA,KAAA,CAAM,EAAE,GAAG,mBAAA,EAAqB,EAAA,EAAI,KAAK,EAAA,EAAI,KAAA,EAAO,KAAA,EAAO,cAAA,EAAgB,CAAA;AAC3E,IAAA;AAAA,EACF;AACA,EAAA,IAAI,GAAA,CAAI,MAAM,mBAAA,EAAqB;AACjC,IAAA,KAAA,CAAM,EAAE,CAAA,EAAG,mBAAA,EAAqB,EAAA,EAAI,IAAI,EAAA,EAAI,EAAA,EAAI,KAAA,EAAO,KAAA,EAAO,sBAAsB,MAAA,CAAQ,GAAA,CAAuB,CAAC,CAAC,IAAI,CAAA;AACzH,IAAA;AAAA,EACF;AAEA,EAAA,IAAI;AACF,IAAA,QAAQ,IAAI,EAAA;AAAI,MACd,KAAK,MAAA;AACH,QAAA,KAAA,CAAM,EAAE,CAAA,EAAG,mBAAA,EAAqB,EAAA,EAAI,GAAA,CAAI,IAAI,EAAA,EAAI,IAAA,EAAM,EAAA,EAAI,MAAA,EAAQ,CAAA;AAClE,QAAA;AAAA,MACF,KAAK,MAAA;AACH,QAAA,KAAA,CAAM,WAAW,GAAA,CAAI,QAAA;AACrB,QAAA,KAAA,CAAM,MAAA,GAAS,IAAA;AACf,QAAA,KAAA,CAAM,EAAE,CAAA,EAAG,mBAAA,EAAqB,EAAA,EAAI,GAAA,CAAI,IAAI,EAAA,EAAI,IAAA,EAAM,EAAA,EAAI,MAAA,EAAQ,CAAA;AAClE,QAAA;AAAA,MACF,KAAK,QAAA;AACH,QAAA,KAAA,CAAM,QAAA,GAAW,IAAA;AACjB,QAAA,KAAA,CAAM,MAAA,GAAS,KAAA;AACf,QAAA,KAAA,CAAM,EAAE,CAAA,EAAG,mBAAA,EAAqB,EAAA,EAAI,GAAA,CAAI,IAAI,EAAA,EAAI,IAAA,EAAM,EAAA,EAAI,QAAA,EAAU,CAAA;AACpE,QAAA;AAAA,MACF,KAAK,QAAA;AACH,QAAA,KAAA,CAAM;AAAA,UACJ,CAAA,EAAG,mBAAA;AAAA,UACH,IAAI,GAAA,CAAI,EAAA;AAAA,UACR,EAAA,EAAI,IAAA;AAAA,UACJ,EAAA,EAAI,QAAA;AAAA,UACJ,MAAA,EAAQ,EAAE,SAAA,EAAW,SAAA,EAAW,UAAU,KAAA,CAAM,QAAA,EAAU,MAAA,EAAQ,KAAA,CAAM,MAAA;AAAO,SAChF,CAAA;AACD,QAAA;AAAA,MACF,KAAK,cAAA;AACH,QAAA,KAAA,CAAM;AAAA,UACJ,CAAA,EAAG,mBAAA;AAAA,UACH,IAAI,GAAA,CAAI,EAAA;AAAA,UACR,EAAA,EAAI,IAAA;AAAA,UACJ,EAAA,EAAI,cAAA;AAAA,UACJ,mBAAA,EAAqB,MAAM,OAAA,IAAW,IAAA;AAAA,UACtC,SAAA,EAAW;AAAA,SACZ,CAAA;AACD,QAAA;AAAA,MACF,KAAK,UAAA,EAAY;AACf,QAAA,IAAI,CAAC,KAAA,CAAM,MAAA,IAAU,CAAC,MAAM,QAAA,EAAU;AACpC,UAAA,KAAA,CAAM,EAAE,CAAA,EAAG,mBAAA,EAAqB,EAAA,EAAI,GAAA,CAAI,IAAI,EAAA,EAAI,KAAA,EAAO,KAAA,EAAO,iBAAA,EAAmB,CAAA;AACjF,UAAA;AAAA,QACF;AACA,QAAA,IAAI,CAAC,MAAM,OAAA,EAAS;AAClB,UAAA,KAAA,CAAM;AAAA,YACJ,CAAA,EAAG,mBAAA;AAAA,YACH,IAAI,GAAA,CAAI,EAAA;AAAA,YACR,EAAA,EAAI,KAAA;AAAA,YACJ,KAAA,EAAO;AAAA,WACR,CAAA;AACD,UAAA;AAAA,QACF;AAEA,QAAA,IAAI,MAAA;AACJ,QAAA,IAAI;AACF,UAAA,MAAM,UAAU,MAAM,kBAAA,CAAmB,KAAA,CAAM,OAAA,EAAS,MAAM,QAAQ,CAAA;AACtE,UAAA,MAAM,MAAA,GAAS,MAAM,iBAAA,CAAkB,OAAA,EAAS,IAAI,QAAA,EAAU;AAAA,YAC5D,OAAA,EAAS,IAAI,IAAA,EAAM,OAAA;AAAA,YACnB,cAAA,EAAgB,IAAI,IAAA,EAAM,cAAA;AAAA,YAC1B,cAAA,EAAgB,IAAI,IAAA,EAAM,cAAA;AAAA,YAC1B,0BAAA,EAA4B,IAAI,IAAA,EAAM,0BAAA;AAAA,YACtC,eAAA,EAAiB,IAAI,IAAA,EAAM,eAAA;AAAA,YAC3B,SAAA,EAAW,IAAI,IAAA,EAAM,SAAA;AAAA,YACrB,WAAA,EAAa,IAAI,IAAA,EAAM,WAAA;AAAA,YACvB,cAAA,EAAgB,IAAI,IAAA,EAAM;AAAA,WAC3B,CAAA;AACD,UAAA,MAAA,GAAS;AAAA,YACP,MAAM,MAAA,CAAO,IAAA;AAAA,YACb,OAAO,KAAA,CAAM,QAAA;AAAA,YACb,SAAA,EAAW,SAAA;AAAA,YACX,aAAa,MAAA,CAAO,WAAA;AAAA,YACpB,oBAAoB,MAAA,CAAO;AAAA,WAC7B;AAAA,QACF,CAAA,CAAA,MAAQ;AACN,UAAA,MAAM,IAAA,GAAO,MAAM,wBAAA,CAAyB;AAAA,YAC1C,KAAK,KAAA,CAAM,OAAA;AAAA,YACX,WAAW,KAAA,CAAM,QAAA;AAAA,YACjB,UAAU,GAAA,CAAI,QAAA;AAAA,YACd,MAAM,GAAA,CAAI;AAAA,WACX,CAAA;AACD,UAAA,MAAA,GAAS;AAAA,YACP,IAAA;AAAA,YACA,OAAO,KAAA,CAAM,QAAA;AAAA,YACb,SAAA,EAAW,SAAA;AAAA,YACX,aAAa,CAAC,EAAE,GAAA,CAAI,IAAA,EAAM,WAAW,KAAA,CAAM,OAAA;AAAA,WAC7C;AAAA,QACF;AACA,QAAA,KAAA,CAAM,EAAE,CAAA,EAAG,mBAAA,EAAqB,EAAA,EAAI,GAAA,CAAI,EAAA,EAAI,EAAA,EAAI,IAAA,EAAM,EAAA,EAAI,UAAA,EAAY,MAAA,EAAQ,CAAA;AAC9E,QAAA;AAAA,MACF;AAAA,MACA;AACE,QAAA,KAAA,CAAM,EAAE,CAAA,EAAG,mBAAA,EAAqB,EAAA,EAAK,GAAA,CAAuB,IAAI,EAAA,EAAI,KAAA,EAAO,KAAA,EAAO,YAAA,EAAc,CAAA;AAAA;AACpG,EACF,SAAS,CAAA,EAAG;AACV,IAAA,KAAA,CAAM;AAAA,MACJ,CAAA,EAAG,mBAAA;AAAA,MACH,IAAI,GAAA,CAAI,EAAA;AAAA,MACR,EAAA,EAAI,KAAA;AAAA,MACJ,OAAO,CAAA,YAAa,KAAA,GAAQ,CAAA,CAAE,OAAA,GAAU,OAAO,CAAC;AAAA,KACjD,CAAA;AAAA,EACH;AACF;AAGO,IAAM,iBAAN,MAAwC;AAAA,EAI7C,YAA6B,UAAA,EAAoB;AAApB,IAAA,IAAA,CAAA,UAAA,GAAA,UAAA;AAAA,EAAqB;AAAA,EAArB,UAAA;AAAA,EAHpB,SAAA,GAAY,SAAA;AAAA,EACb,GAAA,GAAM,CAAA;AAAA,EAId,WAAA,GAAsB;AACpB,IAAA,OAAO,IAAA,CAAK,UAAA;AAAA,EACd;AAAA,EAEA,2BAAA,GAAuC;AACrC,IAAA,OAAO,IAAA;AAAA,EACT;AAAA,EAEA,MAAM,KAAK,QAAA,EAAiC;AAC1C,IAAA,MAAM,KAAK,GAAA,CAAI,EAAE,EAAA,EAAI,MAAA,EAAQ,UAAU,CAAA;AAAA,EACzC;AAAA,EAEA,MAAM,MAAA,GAAwB;AAC5B,IAAA,MAAM,IAAA,CAAK,GAAA,CAAI,EAAE,EAAA,EAAI,UAAU,CAAA;AAAA,EACjC;AAAA,EAEA,MAAM,QAAA,CAAS,QAAA,EAA4B,IAAA,EAAuD;AAChG,IAAA,MAAM,GAAA,GAAM,MAAM,IAAA,CAAK,GAAA,CAAI,EAAE,EAAA,EAAI,UAAA,EAAY,QAAA,EAAU,IAAA,EAAM,CAAA;AAC7D,IAAA,IAAI,CAAC,GAAA,CAAI,EAAA,IAAM,GAAA,CAAI,OAAO,UAAA,EAAY;AACpC,MAAA,MAAM,IAAI,KAAA,CAAM,CAAC,IAAI,EAAA,GAAK,GAAA,CAAI,QAAQ,qBAAqB,CAAA;AAAA,IAC7D;AACA,IAAA,OAAO,GAAA,CAAI,MAAA;AAAA,EACb;AAAA,EAEA,YAAA,GAAgC;AAE9B,IAAA,OAAO,EAAE,SAAA,EAAW,SAAA,EAAW,QAAA,EAAU,IAAA,EAAM,QAAQ,KAAA,EAAM;AAAA,EAC/D;AAAA,EAEA,MAAM,iBAAA,GAA8C;AAClD,IAAA,MAAM,MAAM,MAAM,IAAA,CAAK,IAAI,EAAE,EAAA,EAAI,UAAU,CAAA;AAC3C,IAAA,IAAI,CAAC,GAAA,CAAI,EAAA,IAAM,GAAA,CAAI,OAAO,QAAA,EAAU;AAClC,MAAA,MAAM,IAAI,KAAA,CAAM,CAAC,IAAI,EAAA,GAAK,GAAA,CAAI,QAAQ,qBAAqB,CAAA;AAAA,IAC7D;AACA,IAAA,OAAO,GAAA,CAAI,MAAA;AAAA,EACb;AAAA,EAEQ,IAAI,IAAA,EAAmD;AAC7D,IAAA,MAAM,EAAA,GAAK,CAAA,CAAA,EAAI,EAAE,IAAA,CAAK,GAAG,CAAA,CAAA;AACzB,IAAA,MAAM,MAAM,EAAE,CAAA,EAAG,mBAAA,EAAqB,EAAA,EAAI,GAAG,IAAA,EAAK;AAClD,IAAA,OAAO,IAAI,OAAA,CAAQ,CAAC,OAAA,EAAS,MAAA,KAAW;AACtC,MAAA,MAAM,MAAA,GAAa,GAAA,CAAA,gBAAA,CAAiB,IAAA,CAAK,UAAU,CAAA;AACnD,MAAA,IAAI,GAAA,GAAM,EAAA;AACV,MAAA,MAAM,KAAA,GAAQ,WAAW,MAAM;AAC7B,QAAA,MAAA,CAAO,OAAA,EAAQ;AACf,QAAA,MAAA,CAAO,IAAI,KAAA,CAAM,CAAA,sBAAA,EAAyB,IAAA,CAAK,UAAU,GAAG,CAAC,CAAA;AAAA,MAC/D,GAAG,IAAO,CAAA;AACV,MAAA,MAAA,CAAO,YAAY,MAAM,CAAA;AACzB,MAAA,MAAA,CAAO,EAAA,CAAG,WAAW,MAAM;AACzB,QAAA,MAAA,CAAO,KAAA,CAAM,CAAA,EAAG,IAAA,CAAK,SAAA,CAAU,GAAG,CAAC;AAAA,CAAI,CAAA;AAAA,MACzC,CAAC,CAAA;AACD,MAAA,MAAA,CAAO,EAAA,CAAG,MAAA,EAAQ,CAAC,KAAA,KAAkB;AACnC,QAAA,GAAA,IAAO,KAAA;AACP,QAAA,MAAM,EAAA,GAAK,GAAA,CAAI,OAAA,CAAQ,IAAI,CAAA;AAC3B,QAAA,IAAI,KAAK,CAAA,EAAG;AACZ,QAAA,MAAM,OAAO,GAAA,CAAI,KAAA,CAAM,CAAA,EAAG,EAAE,EAAE,IAAA,EAAK;AACnC,QAAA,YAAA,CAAa,KAAK,CAAA;AAClB,QAAA,MAAA,CAAO,GAAA,EAAI;AACX,QAAA,IAAI;AACF,UAAA,OAAA,CAAQ,IAAA,CAAK,KAAA,CAAM,IAAI,CAAuB,CAAA;AAAA,QAChD,SAAS,CAAA,EAAG;AACV,UAAA,MAAA,CAAO,CAAC,CAAA;AAAA,QACV;AAAA,MACF,CAAC,CAAA;AACD,MAAA,MAAA,CAAO,EAAA,CAAG,OAAA,EAAS,CAAC,CAAA,KAAM;AACxB,QAAA,YAAA,CAAa,KAAK,CAAA;AAClB,QAAA,MAAA,CAAO,CAAC,CAAA;AAAA,MACV,CAAC,CAAA;AAAA,IACH,CAAC,CAAA;AAAA,EACH;AACF;AAGO,IAAM,oBAAA,GAAN,cAAmC,cAAA,CAAe;AAAC;;;ACpPnD,SAAS,aAAA,CAAc,OAAA,GAAgC,EAAC,EAAY;AACzE,EAAA,MAAM,SAAA,GAAY,QAAQ,SAAA,IAAa,UAAA;AACvC,EAAA,IAAI,cAAc,SAAA,EAAW;AAC3B,IAAA,OAAO,IAAI,cAAA,CAAe,OAAA,CAAQ,QAAA,IAAY,wBAAwB,CAAA;AAAA,EACxE;AACA,EAAA,MAAM,IAAA,GAAO,IAAI,eAAA,EAAgB;AACjC,EAAA,IAAI,OAAA,CAAQ,OAAA,EAAS,IAAA,CAAK,UAAA,CAAW,QAAQ,OAAO,CAAA;AACpD,EAAA,IAAI,OAAA,CAAQ,aAAA,EAAe,IAAA,CAAK,gBAAA,CAAiB,IAAI,CAAA;AACrD,EAAA,OAAO,IAAA;AACT;AAGO,SAAS,sBAAA,CAAuB,GAAA,GAAyB,OAAA,CAAQ,GAAA,EAAa;AACnF,EAAA,OAAYA,IAAA,CAAA,IAAA,CAAK,kBAAA,CAAmB,GAAG,CAAA,EAAG,eAAe,CAAA;AAC3D;AAGO,SAAS,yBAAA,CACd,QAAA,EACA,GAAA,GAAyB,OAAA,CAAQ,GAAA,EACf;AAClB,EAAA,IAAI,QAAA,KAAa,UAAA,IAAc,QAAA,KAAa,SAAA,EAAW,OAAO,QAAA;AAC9D,EAAA,MAAM,KAAK,GAAA,CAAI,4BAAA,IAAgC,GAAA,CAAI,sBAAA,IAA0B,IAAI,WAAA,EAAY;AAC7F,EAAA,IAAI,CAAA,KAAM,SAAA,IAAa,CAAA,KAAM,UAAA,EAAY,OAAO,SAAA;AAChD,EAAA,OAAO,UAAA;AACT","file":"chunk-PXHNTC6D.js","sourcesContent":["/**\n * node-llama-cpp binding adapter — pure glue so embedded + process hosts share\n * one code path. The module is injected (from ensurePackage); never imported\n * statically so the published CLI stays free of native deps.\n */\n\nimport type { LlmChatMessage, LlmGenerateOptions } from './types.js';\n\nexport interface LlamaBindingGenerateInput {\n  /** Loaded node-llama-cpp module (dynamic). */\n  lib: unknown;\n  modelPath: string;\n  messages: LlmChatMessage[];\n  opts?: LlmGenerateOptions;\n}\n\n/**\n * Run one chat completion via node-llama-cpp. Throws on binding/model errors\n * (callers wrap with redaction).\n */\nexport async function generateWithLlamaBinding(input: LlamaBindingGenerateInput): Promise<string> {\n  const lib: any = input.lib;\n  if (!lib || typeof lib.getLlama !== 'function') {\n    throw new Error('llm-host: invalid node-llama-cpp module (getLlama missing)');\n  }\n  const llama = await lib.getLlama();\n  const model = await llama.loadModel({ modelPath: input.modelPath });\n  const ctx = await model.createContext();\n  const { LlamaChatSession } = lib;\n  if (typeof LlamaChatSession !== 'function') {\n    throw new Error('llm-host: LlamaChatSession missing from node-llama-cpp module');\n  }\n  const chat = new LlamaChatSession({ contextSequence: ctx.getSequence() });\n  const prompt = input.messages.map((m) => `${m.role}: ${m.content}`).join('\\n\\n');\n  const temperature = input.opts?.temperature ?? 0;\n  const maxTokens = input.opts?.maxTokens;\n  // Grammar: pass through when binding supports it (future constrained PatchIR).\n  const promptOpts: Record<string, unknown> = { temperature };\n  if (typeof maxTokens === 'number') promptOpts.maxTokens = maxTokens;\n  if (input.opts?.grammar) {\n    // Best-effort: node-llama-cpp versions differ on the grammar API surface.\n    promptOpts.grammar = input.opts.grammar;\n  }\n  const text: string = await chat.prompt(prompt, promptOpts);\n  return typeof text === 'string' ? text : String(text ?? '');\n}\n","/**\n * Embedded llm-host (Approach B default) — same address space as the agent.\n *\n * Production path: warm session pool ({@link acquireHostSession}) so models are\n * not reloaded every chat turn. Binding is injected via {@link setBinding}.\n */\n\nimport { annotateUnknownIdentifiers } from '../identifier-mask.js';\nimport { generateWithLlamaBinding } from './binding.js';\nimport {\n  acquireHostSession,\n  generateOnSession,\n  type HostGenerateReport,\n} from './session.js';\nimport type {\n  LlmChatMessage,\n  LlmGenerateOptions,\n  LlmGenerateResult,\n  LlmHost,\n  LlmMemoryReport,\n} from './types.js';\n\nexport class EmbeddedLlmHost implements LlmHost {\n  readonly isolation = 'embedded' as const;\n  private modelRef: string | null = null;\n  private loaded = false;\n  private binding: unknown | null = null;\n  private preferGrammar = false;\n  private lastReport: HostGenerateReport | null = null;\n\n  /** Attach the dynamically loaded node-llama-cpp module. */\n  setBinding(lib: unknown): void {\n    this.binding = lib;\n  }\n\n  /** Hint from MEP: attempt constrained decoding when grammar is supplied. */\n  setPreferGrammar(on: boolean): void {\n    this.preferGrammar = on;\n  }\n\n  /** Last generate metrics (KV plan, draft accept, warm start). */\n  lastGenerateReport(): HostGenerateReport | null {\n    return this.lastReport;\n  }\n\n  async load(modelRef: string): Promise<void> {\n    this.modelRef = modelRef;\n    this.loaded = true;\n    // Eagerly warm the session when binding is already available.\n    if (this.binding) {\n      await acquireHostSession(this.binding, modelRef);\n    }\n  }\n\n  async unload(): Promise<void> {\n    this.modelRef = null;\n    this.loaded = false;\n    this.lastReport = null;\n  }\n\n  supportsConstrainedDecoding(): boolean {\n    return this.preferGrammar && this.binding != null;\n  }\n\n  memoryReport(): LlmMemoryReport {\n    return {\n      isolation: this.isolation,\n      modelRef: this.modelRef,\n      loaded: this.loaded,\n    };\n  }\n\n  async generate(messages: LlmChatMessage[], opts: LlmGenerateOptions = {}): Promise<LlmGenerateResult> {\n    if (!this.loaded || !this.modelRef) {\n      throw new Error('llm-host: no model loaded — orchestrator must call load() after resolve/install');\n    }\n    if (!this.binding) {\n      throw new Error(\n        'llm-host: node-llama-cpp binding not attached — run `vg models install` for the llama-cpp provider, then retry',\n      );\n    }\n\n    // Production path: pooled warm session.\n    try {\n      const session = await acquireHostSession(this.binding, this.modelRef);\n      const report = await generateOnSession(session, messages, {\n        grammar: opts.grammar,\n        requireGrammar: opts.requireGrammar,\n        identifierTrie: opts.identifierTrie,\n        annotateUnknownIdentifiers: opts.annotateUnknownIdentifiers,\n        draftCandidates: opts.draftCandidates,\n        maxTokens: opts.maxTokens,\n        temperature: opts.temperature,\n        promptSegments: opts.promptSegments,\n      });\n      this.lastReport = report;\n      return {\n        text: report.text,\n        model: report.model,\n        isolation: report.isolation,\n        constrained: report.constrained || !!(opts.grammar && this.preferGrammar),\n        unknownIdentifiers: report.unknownIdentifiers,\n      };\n    } catch {\n      // Fallback one-shot (tests / minimal bindings).\n      let text = await generateWithLlamaBinding({\n        lib: this.binding,\n        modelPath: this.modelRef,\n        messages,\n        opts,\n      });\n      let unknownIdentifiers: string[] | undefined;\n      if (opts.identifierTrie && opts.annotateUnknownIdentifiers !== false) {\n        const ann = annotateUnknownIdentifiers(text, opts.identifierTrie);\n        text = ann.text;\n        unknownIdentifiers = ann.unknown.length ? ann.unknown : undefined;\n      }\n      return {\n        text,\n        model: this.modelRef,\n        isolation: this.isolation,\n        constrained: !!(opts.grammar && this.supportsConstrainedDecoding()),\n        unknownIdentifiers,\n      };\n    }\n  }\n}\n","/**\n * Process-isolation llm-host: client + server over a Unix/local socket.\n *\n * Management (Code Modes, fit, packs, install) stays in vgd / the CLI.\n * This process only loads weights and decodes (ADR-005).\n */\n\nimport * as fs from 'node:fs';\nimport * as net from 'node:net';\nimport * as path from 'node:path';\nimport { generateWithLlamaBinding } from './binding.js';\nimport { acquireHostSession, generateOnSession } from './session.js';\nimport type {\n  LlmChatMessage,\n  LlmGenerateOptions,\n  LlmGenerateResult,\n  LlmHost,\n  LlmMemoryReport,\n} from './types.js';\n\nexport const LLM_HOST_IPC_SCHEMA = 'llm-host-ipc/0' as const;\n\nexport type LlmHostIpcRequest =\n  | { v: typeof LLM_HOST_IPC_SCHEMA; id: string; op: 'load'; modelRef: string }\n  | { v: typeof LLM_HOST_IPC_SCHEMA; id: string; op: 'unload' }\n  | { v: typeof LLM_HOST_IPC_SCHEMA; id: string; op: 'generate'; messages: LlmChatMessage[]; opts?: LlmGenerateOptions }\n  | { v: typeof LLM_HOST_IPC_SCHEMA; id: string; op: 'memory' }\n  | { v: typeof LLM_HOST_IPC_SCHEMA; id: string; op: 'capabilities' }\n  | { v: typeof LLM_HOST_IPC_SCHEMA; id: string; op: 'ping' };\n\nexport type LlmHostIpcResponse =\n  | { v: typeof LLM_HOST_IPC_SCHEMA; id: string; ok: true; op: 'load' | 'unload' | 'ping' }\n  | { v: typeof LLM_HOST_IPC_SCHEMA; id: string; ok: true; op: 'generate'; result: LlmGenerateResult }\n  | { v: typeof LLM_HOST_IPC_SCHEMA; id: string; ok: true; op: 'memory'; report: LlmMemoryReport }\n  | {\n      v: typeof LLM_HOST_IPC_SCHEMA;\n      id: string;\n      ok: true;\n      op: 'capabilities';\n      constrainedDecoding: boolean;\n      isolation: 'process';\n    }\n  | { v: typeof LLM_HOST_IPC_SCHEMA; id: string; ok: false; error: string };\n\n/** Request body without envelope — distribute Omit over the union (keyof union is only common keys). */\nexport type LlmHostIpcBody = LlmHostIpcRequest extends infer R\n  ? R extends { v: typeof LLM_HOST_IPC_SCHEMA; id: string }\n    ? Omit<R, 'v' | 'id'>\n    : never\n  : never;\n\nexport interface ProcessHostServerOptions {\n  socketPath: string;\n  /** node-llama-cpp module (optional until first generate). */\n  binding?: unknown;\n  onListen?: (socketPath: string) => void;\n}\n\n/**\n * Serve llm-host IPC on a local socket. Resolves when the server is listening;\n * call the returned `close` to stop.\n */\nexport function startProcessHostServer(options: ProcessHostServerOptions): Promise<{ close: () => Promise<void> }> {\n  const state = {\n    modelRef: null as string | null,\n    loaded: false,\n    binding: options.binding ?? null,\n  };\n\n  try {\n    if (fs.existsSync(options.socketPath)) fs.unlinkSync(options.socketPath);\n  } catch {\n    /* ignore */\n  }\n  fs.mkdirSync(path.dirname(options.socketPath), { recursive: true });\n\n  return new Promise((resolve, reject) => {\n    const server = net.createServer((socket) => {\n      let buf = '';\n      socket.setEncoding('utf8');\n      socket.on('data', (chunk: string) => {\n        buf += chunk;\n        let nl: number;\n        while ((nl = buf.indexOf('\\n')) >= 0) {\n          const line = buf.slice(0, nl).trim();\n          buf = buf.slice(nl + 1);\n          if (!line) continue;\n          void handleLine(line, state, (resp) => {\n            socket.write(`${JSON.stringify(resp)}\\n`);\n          });\n        }\n      });\n    });\n\n    server.once('error', reject);\n    server.listen(options.socketPath, () => {\n      options.onListen?.(options.socketPath);\n      resolve({\n        close: () =>\n          new Promise((res, rej) => {\n            server.close((err) => {\n              try {\n                if (fs.existsSync(options.socketPath)) fs.unlinkSync(options.socketPath);\n              } catch {\n                /* ignore */\n              }\n              if (err) rej(err);\n              else res();\n            });\n          }),\n      });\n    });\n  });\n}\n\n/** Attach binding after server start (e.g. after ensurePackage). */\nexport type ProcessHostState = {\n  modelRef: string | null;\n  loaded: boolean;\n  binding: unknown | null;\n};\n\nasync function handleLine(\n  line: string,\n  state: ProcessHostState,\n  reply: (r: LlmHostIpcResponse) => void,\n): Promise<void> {\n  let req: LlmHostIpcRequest;\n  try {\n    req = JSON.parse(line) as LlmHostIpcRequest;\n  } catch {\n    reply({ v: LLM_HOST_IPC_SCHEMA, id: '?', ok: false, error: 'invalid JSON' });\n    return;\n  }\n  if (req.v !== LLM_HOST_IPC_SCHEMA) {\n    reply({ v: LLM_HOST_IPC_SCHEMA, id: req.id, ok: false, error: `unsupported schema ${String((req as { v?: string }).v)}` });\n    return;\n  }\n\n  try {\n    switch (req.op) {\n      case 'ping':\n        reply({ v: LLM_HOST_IPC_SCHEMA, id: req.id, ok: true, op: 'ping' });\n        return;\n      case 'load':\n        state.modelRef = req.modelRef;\n        state.loaded = true;\n        reply({ v: LLM_HOST_IPC_SCHEMA, id: req.id, ok: true, op: 'load' });\n        return;\n      case 'unload':\n        state.modelRef = null;\n        state.loaded = false;\n        reply({ v: LLM_HOST_IPC_SCHEMA, id: req.id, ok: true, op: 'unload' });\n        return;\n      case 'memory':\n        reply({\n          v: LLM_HOST_IPC_SCHEMA,\n          id: req.id,\n          ok: true,\n          op: 'memory',\n          report: { isolation: 'process', modelRef: state.modelRef, loaded: state.loaded },\n        });\n        return;\n      case 'capabilities':\n        reply({\n          v: LLM_HOST_IPC_SCHEMA,\n          id: req.id,\n          ok: true,\n          op: 'capabilities',\n          constrainedDecoding: state.binding != null,\n          isolation: 'process',\n        });\n        return;\n      case 'generate': {\n        if (!state.loaded || !state.modelRef) {\n          reply({ v: LLM_HOST_IPC_SCHEMA, id: req.id, ok: false, error: 'no model loaded' });\n          return;\n        }\n        if (!state.binding) {\n          reply({\n            v: LLM_HOST_IPC_SCHEMA,\n            id: req.id,\n            ok: false,\n            error: 'binding not loaded in host process — start host after vg models install',\n          });\n          return;\n        }\n        // Same warm session path as embedded (Approach B parity for enterprise isolation).\n        let result: LlmGenerateResult;\n        try {\n          const session = await acquireHostSession(state.binding, state.modelRef);\n          const report = await generateOnSession(session, req.messages, {\n            grammar: req.opts?.grammar,\n            requireGrammar: req.opts?.requireGrammar,\n            identifierTrie: req.opts?.identifierTrie,\n            annotateUnknownIdentifiers: req.opts?.annotateUnknownIdentifiers,\n            draftCandidates: req.opts?.draftCandidates,\n            maxTokens: req.opts?.maxTokens,\n            temperature: req.opts?.temperature,\n            promptSegments: req.opts?.promptSegments,\n          });\n          result = {\n            text: report.text,\n            model: state.modelRef,\n            isolation: 'process',\n            constrained: report.constrained,\n            unknownIdentifiers: report.unknownIdentifiers,\n          };\n        } catch {\n          const text = await generateWithLlamaBinding({\n            lib: state.binding,\n            modelPath: state.modelRef,\n            messages: req.messages,\n            opts: req.opts,\n          });\n          result = {\n            text,\n            model: state.modelRef,\n            isolation: 'process',\n            constrained: !!(req.opts?.grammar && state.binding),\n          };\n        }\n        reply({ v: LLM_HOST_IPC_SCHEMA, id: req.id, ok: true, op: 'generate', result });\n        return;\n      }\n      default:\n        reply({ v: LLM_HOST_IPC_SCHEMA, id: (req as { id: string }).id, ok: false, error: 'unknown op' });\n    }\n  } catch (e) {\n    reply({\n      v: LLM_HOST_IPC_SCHEMA,\n      id: req.id,\n      ok: false,\n      error: e instanceof Error ? e.message : String(e),\n    });\n  }\n}\n\n/** Client: full {@link LlmHost} over the process socket. */\nexport class ProcessLlmHost implements LlmHost {\n  readonly isolation = 'process' as const;\n  private seq = 0;\n\n  constructor(private readonly socketPath: string) {}\n\n  getEndpoint(): string {\n    return this.socketPath;\n  }\n\n  supportsConstrainedDecoding(): boolean {\n    return true; // host may apply grammar when binding is present\n  }\n\n  async load(modelRef: string): Promise<void> {\n    await this.rpc({ op: 'load', modelRef });\n  }\n\n  async unload(): Promise<void> {\n    await this.rpc({ op: 'unload' });\n  }\n\n  async generate(messages: LlmChatMessage[], opts?: LlmGenerateOptions): Promise<LlmGenerateResult> {\n    const res = await this.rpc({ op: 'generate', messages, opts });\n    if (!res.ok || res.op !== 'generate') {\n      throw new Error(!res.ok ? res.error : 'unexpected response');\n    }\n    return res.result;\n  }\n\n  memoryReport(): LlmMemoryReport {\n    // Sync surface required by LlmHost; use last known or empty (async memory via rpc elsewhere).\n    return { isolation: 'process', modelRef: null, loaded: false };\n  }\n\n  async memoryReportAsync(): Promise<LlmMemoryReport> {\n    const res = await this.rpc({ op: 'memory' });\n    if (!res.ok || res.op !== 'memory') {\n      throw new Error(!res.ok ? res.error : 'unexpected response');\n    }\n    return res.report;\n  }\n\n  private rpc(body: LlmHostIpcBody): Promise<LlmHostIpcResponse> {\n    const id = `r${++this.seq}`;\n    const req = { v: LLM_HOST_IPC_SCHEMA, id, ...body } as LlmHostIpcRequest;\n    return new Promise((resolve, reject) => {\n      const socket = net.createConnection(this.socketPath);\n      let buf = '';\n      const timer = setTimeout(() => {\n        socket.destroy();\n        reject(new Error(`llm-host IPC timeout (${this.socketPath})`));\n      }, 120_000);\n      socket.setEncoding('utf8');\n      socket.on('connect', () => {\n        socket.write(`${JSON.stringify(req)}\\n`);\n      });\n      socket.on('data', (chunk: string) => {\n        buf += chunk;\n        const nl = buf.indexOf('\\n');\n        if (nl < 0) return;\n        const line = buf.slice(0, nl).trim();\n        clearTimeout(timer);\n        socket.end();\n        try {\n          resolve(JSON.parse(line) as LlmHostIpcResponse);\n        } catch (e) {\n          reject(e);\n        }\n      });\n      socket.on('error', (e) => {\n        clearTimeout(timer);\n        reject(e);\n      });\n    });\n  }\n}\n\n/** @deprecated use ProcessLlmHost */\nexport class ProcessBridgeLlmHost extends ProcessLlmHost {}\n","/**\n * Factory for the dual-mode llm-host (ADR-005).\n *\n * `createLlmHost({ isolation: 'embedded' })` — default Approach B.\n * `createLlmHost({ isolation: 'process', endpoint })` — enterprise shell.\n */\n\nimport * as path from 'node:path';\nimport { vibgrateRuntimeDir } from '../paths.js';\nimport { EmbeddedLlmHost } from './embedded.js';\nimport { ProcessLlmHost } from './process-bridge.js';\nimport type { LlmHost, LlmHostIsolation } from './types.js';\n\nexport type {\n  LlmHost,\n  LlmHostConfig,\n  LlmHostIsolation,\n  LlmChatMessage,\n  LlmGenerateResult,\n  LlmGenerateOptions,\n  LlmMemoryReport,\n} from './types.js';\nexport { EmbeddedLlmHost } from './embedded.js';\nexport {\n  ProcessLlmHost,\n  ProcessBridgeLlmHost,\n  LLM_HOST_IPC_SCHEMA,\n  startProcessHostServer,\n} from './process-bridge.js';\nexport type { LlmHostIpcRequest, LlmHostIpcResponse, ProcessHostServerOptions } from './process-bridge.js';\nexport { generateWithLlamaBinding } from './binding.js';\nexport {\n  acquireHostSession,\n  generateOnSession,\n  clearHostSessionPool,\n  hostSessionPoolSize,\n  listHostSessions,\n  dropHostSession,\n  evictIdleSessions,\n  sessionKey,\n} from './session.js';\nexport type { HostGenerateExtras, HostGenerateReport, HostSessionInfo } from './session.js';\nexport { attachGrammar, assertGrammarOrThrow, allowGrammarStringFallback } from './grammar.js';\nexport type { GrammarAttachResult, GrammarAttachMethod } from './grammar.js';\nexport {\n  buildIdentifierTokenBias,\n  planNextCharMask,\n  tokenCompatibleWithTrie,\n  trieFingerprint,\n} from './logit-mask.js';\nexport type { LogitMaskBuildResult, SamplerHookMethod, BuildLogitMaskOptions } from './logit-mask.js';\nexport {\n  createIdentStreamState,\n  stepIdentStream,\n  filterTokenTexts,\n  attachDynamicIdentFilter,\n  dynamicAllowedNextChars,\n} from './identifier-sampler.js';\nexport type { IdentStreamState, IdentStepResult } from './identifier-sampler.js';\nexport { probeBindingCapabilities, supportsDynamicIdentSampler, supportsKvPrefixEvaluate } from './binding-capabilities.js';\nexport type { BindingCapabilities, ProbeBindingOptions } from './binding-capabilities.js';\nexport { composePromptConstraints } from './compose-constraints.js';\nexport { rankDraftCandidates, pickBestDraft } from './speculative.js';\nexport type { RankedDraft } from './speculative.js';\n\nexport interface CreateLlmHostOptions {\n  isolation?: LlmHostIsolation;\n  /** Socket path when isolation is `process`. Defaults under vibgrate runtime dir. */\n  endpoint?: string;\n  /** Pre-attach binding for embedded hosts (avoids a second ensure). */\n  binding?: unknown;\n  preferGrammar?: boolean;\n}\n\nexport function createLlmHost(options: CreateLlmHostOptions = {}): LlmHost {\n  const isolation = options.isolation ?? 'embedded';\n  if (isolation === 'process') {\n    return new ProcessLlmHost(options.endpoint ?? defaultProcessEndpoint());\n  }\n  const host = new EmbeddedLlmHost();\n  if (options.binding) host.setBinding(options.binding);\n  if (options.preferGrammar) host.setPreferGrammar(true);\n  return host;\n}\n\n/** Default IPC socket path for the isolated llm-host process. */\nexport function defaultProcessEndpoint(env: NodeJS.ProcessEnv = process.env): string {\n  return path.join(vibgrateRuntimeDir(env), 'llm-host.sock');\n}\n\n/** Resolve isolation from env / explicit override (MEP / config later). */\nexport function resolveInferenceIsolation(\n  explicit?: LlmHostIsolation,\n  env: NodeJS.ProcessEnv = process.env,\n): LlmHostIsolation {\n  if (explicit === 'embedded' || explicit === 'process') return explicit;\n  const v = (env.VIBGRATE_INFERENCE_ISOLATION || env.VG_INFERENCE_ISOLATION || '').toLowerCase();\n  if (v === 'process' || v === 'isolated') return 'process';\n  return 'embedded';\n}\n"]}

package/dist/{chunk-I7GY7MVN.js → chunk-RNP3SMTV.js}

@@ -1,9 +1,13 @@
-import { usageError, buildGraph, UsageError, ResourceLimitError, CliError, ExitCode, writeArtifacts, writeSnapshot, refreshInstalledInstructions, SMALL_REPO_FILES, epistemicBreakdown, verifyDeterminism, renderHtml, renderReport, discoverModels, loadGraph } from './chunk-3MXNNBLI.js';
-import { VERSION, brandProgressBar, resolveHead, workingTreeDirty } from './chunk-5BDHQ7DD.js';
-import { MockProvider, OPENAI_COMPATIBLE, OllamaProvider, LocalLlamaProvider, OpenAiCompatibleProvider } from './chunk-CS37OBE3.js';
-import { resolveEmbedModel, isModelReady, countPending, serializeGraph } from './chunk-WNIIKCNF.js';
-import { canonicalize } from './chunk-JFGYT6BI.js';
+import { resolveGgufPath } from './chunk-NDMDSPFX.js';
+import { MockProvider, OPENAI_COMPATIBLE, OllamaProvider, LocalLlamaProvider, OpenAiCompatibleProvider } from './chunk-OFUC6RYN.js';
 import { c, info, json, out } from './chunk-CT5BS56U.js';
+import { usageError, buildGraph, UsageError, ResourceLimitError, CliError, ExitCode, writeSnapshot, refreshInstalledInstructions, SMALL_REPO_FILES, epistemicBreakdown, verifyDeterminism } from './chunk-KNELIQJO.js';
+import { VERSION, brandProgressBar, resolveHead, workingTreeDirty } from './chunk-WGUL45AV.js';
+import { discoverModels } from './chunk-OOIG3RUM.js';
+import { resolveEmbedModel, isModelReady, countPending } from './chunk-6ABXHZRN.js';
+import { writeArtifacts, renderHtml, renderReport, loadGraph } from './chunk-2XQDMLKB.js';
+import { serializeGraph } from './chunk-K3MZIVZW.js';
+import { canonicalize } from './chunk-JFGYT6BI.js';
 import * as path from 'path';
 import * as fs from 'fs';
 import { spawn, execFile } from 'child_process';
@@ -357,7 +361,7 @@ function readGlobal(cmd) {
 
 // src/commands/build.ts
 function registerBuild(program) {
-  const cmd = program.command("build").description("build / update the code map (incremental, deterministic)").argument("[paths...]", "folders or files to map (default: current folder)").option("--only <langs>", "restrict to languages, e.g. ts,py,go").option("--exclude <glob>", "extra ignore glob (repeatable)", collect, []).option("--no-html", "do not write graph.html").option("--no-report", "do not write GRAPH_REPORT.md").option("--no-ground", "do not attach grounding (Phase 2)").option("--jobs <n>", "worker count (1 = single-threaded)").option("--scip <file>", "ingest a SCIP index for precise resolution (default: auto-detect index.scip)").option("--no-scip", "ignore any SCIP index").option("--no-tsc", "skip the in-process TypeScript resolver (heuristic floor only)").option("--no-warm", "do not warm the semantic index in the background after building").option("--grammars <dir>", "directory of grammar .wasm files (offline / air-gapped)").option("-o, --export <file>", "also write the map to a file (format inferred)").option("--attest", "sign the built graph \u2192 .vibgrate/attestation.intoto.jsonl").option("--verify", "verify a committed attestation against the graph (no rebuild)").option("--attest-key <path>", "signing key PEM (else $VG_ATTEST_KEY, else .vibgrate/attest-key.pem)").option("--attestation <file>", "attestation path (out for --attest, in for --verify)").option("--pub <path>", "public key PEM to pin the signer (with --verify)").action(async function(paths, opts) {
+  const cmd = program.command("build").description("build / update the code map (incremental, deterministic)").argument("[paths...]", "folders or files to map (default: current folder)").option("--only <langs>", "restrict to languages, e.g. ts,py,go").option("--exclude <glob>", "extra ignore glob (repeatable)", collect, []).option("--no-html", "do not write graph.html").option("--no-report", "do not write GRAPH_REPORT.md").option("--no-ground", "do not attach grounding (Phase 2)").option("--jobs <n>", "worker count (1 = single-threaded)").option("--scip <file>", "ingest a SCIP index for precise resolution (default: auto-detect index.scip)").option("--no-scip", "ignore any SCIP index").option("--no-tsc", "skip the in-process TypeScript resolver (heuristic floor only)").option("--fast", "skip precise tsc resolve (heuristic only \u2014 faster XL cold builds)").option("--no-index", "do not write the SQLite serve index under .vibgrate/cache/").option("--analysis-tier <tier>", "force analysis tier: full | large | xl (default: auto)").option("--no-warm", "do not warm the semantic index in the background after building").option("--grammars <dir>", "directory of grammar .wasm files (offline / air-gapped)").option("-o, --export <file>", "also write the map to a file (format inferred)").option("--attest", "sign the built graph \u2192 .vibgrate/attestation.intoto.jsonl").option("--verify", "verify a committed attestation against the graph (no rebuild)").option("--attest-key <path>", "signing key PEM (else $VG_ATTEST_KEY, else .vibgrate/attest-key.pem)").option("--attestation <file>", "attestation path (out for --attest, in for --verify)").option("--pub <path>", "public key PEM to pin the signer (with --verify)").action(async function(paths, opts) {
     await runBuild(paths, opts, readGlobal(this));
   });
   applyGlobalOptions(cmd);
@@ -376,6 +380,11 @@ async function runBuild(paths, opts, global) {
     throw usageError(`--jobs must be a positive integer (got "${opts.jobs}")`);
   }
   const only = opts.only ? opts.only.split(",").map((s) => s.trim()).filter(Boolean) : void 0;
+  const tierRaw = opts.analysisTier?.toLowerCase();
+  const analysisTier = tierRaw === "full" || tierRaw === "large" || tierRaw === "xl" ? tierRaw : void 0;
+  if (opts.analysisTier && !analysisTier) {
+    throw usageError(`--analysis-tier must be full, large, or xl (got "${opts.analysisTier}")`);
+  }
   let result;
   try {
     result = await buildGraph({
@@ -390,6 +399,9 @@ async function runBuild(paths, opts, global) {
       scip: typeof opts.scip === "string" ? opts.scip : void 0,
       noScip: opts.scip === false,
       noTsc: opts.tsc === false,
+      fast: opts.fast === true,
+      noIndex: opts.index === false,
+      analysisTier,
       generatedAt: global.generatedAt,
       onParseProgress: bar ? (done, total) => bar.update(done, total) : void 0,
       grammarsDir: opts.grammars
@@ -777,13 +789,29 @@ function resolveProviders(opts, deps = {}) {
 function buildExplicit(id, opts, env, discover) {
   if (id === "ollama") return new OllamaProvider(resolveLocalModel(opts, discover, "ollama"), env.OLLAMA_HOST);
   if (id === "llama-cpp") {
-    const modelPath = opts.modelPath || env.VG_CODE_MODEL_PATH;
-    if (!modelPath) throw new CliError("--provider llama-cpp needs a gguf path via --model-path or VG_CODE_MODEL_PATH (no weights are downloaded automatically; see `vg models`).", ExitCode.USAGE_ERROR);
-    return new LocalLlamaProvider(opts.model ?? basename2(modelPath), modelPath, void 0, !!opts.consent);
+    const resolved = resolveGgufPath({
+      modelPath: opts.modelPath || env.VG_CODE_MODEL_PATH,
+      modelRef: opts.model,
+      env,
+      discover
+    });
+    if (!resolved) {
+      throw new CliError(
+        "--provider llama-cpp needs a GGUF: pass --model-path, set VG_CODE_MODEL_PATH, or run `vg models install` so the weight store has a catalogued gguf. See `vg models`.",
+        ExitCode.USAGE_ERROR
+      );
+    }
+    return new LocalLlamaProvider(opts.model ?? basename2(resolved.path), resolved.path, void 0, !!opts.consent);
   }
   if (id === "lmstudio") return buildHosted("lmstudio", resolveHostedModel(opts, env, "lmstudio"));
+  if (id === "foundry-local" || id === "foundry") {
+    return buildHosted("foundry-local", resolveHostedModel(opts, env, "foundry-local"));
+  }
   if (HOSTED_IDS.has(id)) return buildHosted(id, resolveHostedModel(opts, env, id));
-  throw new CliError(`unknown --provider "${id}". Known: ollama, lmstudio, openrouter, litellm, openai, together, llama-cpp.`, ExitCode.USAGE_ERROR);
+  throw new CliError(
+    `unknown --provider "${id}". Known: ollama, lmstudio, foundry-local, openrouter, litellm, openai, together, llama-cpp.`,
+    ExitCode.USAGE_ERROR
+  );
 }
 function buildHosted(id, model) {
   const base = OPENAI_COMPATIBLE[id];
@@ -793,10 +821,30 @@ function buildHosted(id, model) {
 function localFallbacks(opts, env, discover, exclude) {
   const out2 = [];
   const models = safeDiscover(discover);
+  const preferOllama = env.VG_PREFER_OLLAMA === "1" || env.VIBGRATE_PREFER_OLLAMA === "1" || opts.preferEmbedded === false;
+  const gguf = exclude === "llama-cpp" ? null : resolveGgufPath({
+    modelPath: opts.modelPath || env.VG_CODE_MODEL_PATH,
+    modelRef: opts.model,
+    env,
+    discover: () => models
+  });
+  if (gguf && !preferOllama) {
+    out2.push(new LocalLlamaProvider(opts.model ?? basename2(gguf.path), gguf.path, void 0, !!opts.consent));
+  }
   const ollamaModel = opts.model ?? firstOfRuntime(models, "ollama");
   if (exclude !== "ollama" && ollamaModel) out2.push(new OllamaProvider(ollamaModel, env.OLLAMA_HOST));
   const lmModel = opts.model ?? firstOfRuntime(models, "lm-studio");
   if (exclude !== "lmstudio" && lmModel) out2.push(buildHosted("lmstudio", lmModel));
+  if (exclude !== "foundry-local" && exclude !== "foundry" && (env.FOUNDRY_LOCAL_BASE_URL || env.VG_CODE_PROVIDER === "foundry-local")) {
+    try {
+      const m = opts.model || env.VG_CODE_MODEL;
+      if (m) out2.push(buildHosted("foundry-local", m));
+    } catch {
+    }
+  }
+  if (gguf && !out2.some((p) => p.id === "llama-cpp")) {
+    out2.push(new LocalLlamaProvider(opts.model ?? basename2(gguf.path), gguf.path, void 0, !!opts.consent));
+  }
   return out2;
 }
 function resolveLocalModel(opts, discover, runtime) {
@@ -829,5 +877,5 @@ function basename2(p) {
 }
 
 export { DSSE_PAYLOAD_TYPE, ProgressBar, applyGlobalOptions, dssePae, fetchCatalog, generateKeypair, keyId, printCodeLogo, printLogo, readGlobal, registerBuild, resolveProviders, runBuild, sha256Hex, topProviders };
-//# sourceMappingURL=chunk-I7GY7MVN.js.map
-//# sourceMappingURL=chunk-I7GY7MVN.js.map
+//# sourceMappingURL=chunk-RNP3SMTV.js.map
+//# sourceMappingURL=chunk-RNP3SMTV.js.map