diff --git a/.gitignore b/.gitignore
index 2bca90a4a..c885c4588 100644
--- a/.gitignore
+++ b/.gitignore
@@ -465,3 +465,4 @@ ts/packages/agentSdkWrapper/plans/
 /.pnpm-store
 /ts/packages/copilot-plugin/dist-exe
 /ts/msi-out
+ts/config.local.yaml.bak
diff --git a/ts/packages/agentServer/client/src/agentServerClient.ts b/ts/packages/agentServer/client/src/agentServerClient.ts
index ea686589c..ddff9d951 100644
--- a/ts/packages/agentServer/client/src/agentServerClient.ts
+++ b/ts/packages/agentServer/client/src/agentServerClient.ts
@@ -562,9 +562,17 @@ function spawnAgentServer(
                     "'" + s.replace(/'/g, "''") + "'";
                 const psCommand = `& node ${psQuote(serverPath)} --port ${port}${idleTimeout > 0 ? ` --idle-timeout ${idleTimeout}` : ""}`;
                 const psArgs = ["-NoExit", "-Command", psCommand];
+                // The first quoted argument to `start` is the new window's
+                // title. It MUST be non-empty: an empty title leaves the
+                // console window title blank, which trips a libuv bug on
+                // Windows — GetConsoleTitleW returns 0 with GetLastError()==0,
+                // libuv reads that as success but leaves process_title NULL,
+                // and the next process.title read aborts with
+                // "Assertion failed: process_title, file src\\win\\util.c".
+                const windowTitle = `TypeAgent Server (port ${port})`;
                 const child = spawn(
                     "cmd.exe",
-                    ["/c", "start", "", psExe, ...psArgs],
+                    ["/c", "start", windowTitle, psExe, ...psArgs],
                     {
                         detached: true,
                         stdio: "ignore",
diff --git a/ts/tools/scripts/setModelProvider.mjs b/ts/tools/scripts/setModelProvider.mjs
new file mode 100644
index 000000000..15fddb672
--- /dev/null
+++ b/ts/tools/scripts/setModelProvider.mjs
@@ -0,0 +1,345 @@
+#!/usr/bin/env node
+// Copyright (c) Microsoft Corporation.
+// Licensed under the MIT License.
+
+/**
+ * setModelProvider — switch an EXISTING repo config.local.yaml between LLM
+ * providers in place, for testing the self-host provider-choice flow on a
+ * machine that already has a (Key-Vault-downloaded) config.
+ *
+ * Unlike generate-selfhost-config.mjs (which writes a fresh minimal config on
+ * repo-less machines), this patches the chat `modelProvider` + `embedding`
+ * fields of an existing config while PRESERVING everything else (azureOpenAI
+ * endpoints, azureFoundry, maps, speech, etc.). That lets a developer flip to
+ * Copilot chat + local embeddings, validate, then flip back to AI Systems
+ * (`--provider aisystems`) or `--restore` the pristine backup.
+ *
+ * Provider governs CHAT only (routed via aiclient's provider-mode override);
+ * `embedding.provider` is independent (the Copilot SDK has no embeddings).
+ *
+ * Config path precedence (mirrors getKeys / the @typeagent/config loader):
+ *   --config
+ *   > TYPEAGENT_CONFIG_LOCAL
+ *   > <TYPEAGENT_CONFIG_DIR>/config.local.yaml
+ *   > <repo ts/>/config.local.yaml   (in-repo default)
+ *
+ * Usage:
+ *   node setModelProvider.mjs --provider copilot|ollama|aisystems [options]
+ *   node setModelProvider.mjs --restore            # restore <config>.bak
+ *
+ * Options:
+ *   --provider <mode>            copilot | ollama | aisystems (required unless --restore).
+ *   --embedding <mode>           local | ollama | openai | none | azure (default: local
+ *                                for copilot/ollama; azure for aisystems).
+ *   --copilot-model <name>       Copilot chat model (default claude-sonnet-4.5).
+ *   --ollama-host <url>          Ollama base URL (default http://localhost:11434).
+ *   --chat-model <name>          Ollama chat model (default llama3.2).
+ *   --embedding-endpoint <url>   Embedding endpoint (openai mode; full path).
+ *   --embedding-model <name>     Embedding model name.
+ *   --local-embedding-model <n>  transformers.js model (default Xenova/all-MiniLM-L6-v2).
+ *   --openai-key <key>           API key for openai embedding mode.
+ *   --config <path>              Config file to patch (overrides env resolution).
+ *   --backup                     Write <config>.bak before editing (won't clobber an
+ *                                existing backup).
+ *   --restore                    Restore <config> from <config>.bak and exit.
+ *   --dry-run                    Print the patched YAML to stdout without writing.
+ *   --no-validate                Skip loadRuntimeConfigSync validation.
+ */
+
+import fs from "node:fs";
+import path from "node:path";
+import { fileURLToPath } from "node:url";
+import { createRequire } from "node:module";
+
+const __dirname = path.dirname(fileURLToPath(import.meta.url));
+const require = createRequire(import.meta.url);
+const yaml = require("js-yaml");
+
+const DEFAULT_OLLAMA_HOST = "http://localhost:11434";
+const DEFAULT_OLLAMA_CHAT_MODEL = "llama3.2";
+const DEFAULT_COPILOT_MODEL = "claude-sonnet-4.5";
+const DEFAULT_LOCAL_EMBEDDING_MODEL = "Xenova/all-MiniLM-L6-v2";
+const DEFAULT_OLLAMA_EMBEDDING_MODEL = "nomic-embed-text";
+
+function arg(name, fallback) {
+    const i = process.argv.indexOf(name);
+    return i !== -1 && i + 1 < process.argv.length
+        ? process.argv[i + 1]
+        : fallback;
+}
+
+function hasFlag(name) {
+    return process.argv.includes(name);
+}
+
+// Match getKeys.mjs / generate-selfhost-config.mjs config-path precedence.
+function resolveConfigPath() {
+    const explicit = arg("--config");
+    if (explicit) {
+        return path.resolve(explicit);
+    }
+    if (process.env.TYPEAGENT_CONFIG_LOCAL) {
+        return path.resolve(process.env.TYPEAGENT_CONFIG_LOCAL);
+    }
+    if (process.env.TYPEAGENT_CONFIG_DIR) {
+        return path.join(process.env.TYPEAGENT_CONFIG_DIR, "config.local.yaml");
+    }
+    return path.resolve(__dirname, "../../config.local.yaml");
+}
+
+// Strip a trailing chat path so we keep only the Ollama base URL (aiclient
+// appends /v1/chat/completions and /v1/embeddings itself).
+function ollamaBaseUrl(url) {
+    let base = (url ?? DEFAULT_OLLAMA_HOST).trim().replace(/\/+$/, "");
+    base = base.replace(/\/v1\/chat\/completions$/i, "");
+    base = base.replace(/\/v1$/i, "");
+    return base;
+}
+
+// Ensure a nested object exists at cfg[key] and return it.
+function ensureObject(cfg, key) {
+    if (
+        cfg[key] === undefined ||
+        cfg[key] === null ||
+        typeof cfg[key] !== "object" ||
+        Array.isArray(cfg[key])
+    ) {
+        cfg[key] = {};
+    }
+    return cfg[key];
+}
+
+function applyEmbedding(cfg, options) {
+    const { embedding } = options;
+    switch (embedding) {
+        case "local": {
+            cfg.embedding = {
+                provider: "local",
+                model:
+                    options.localEmbeddingModel ||
+                    DEFAULT_LOCAL_EMBEDDING_MODEL,
+            };
+            break;
+        }
+        case "ollama": {
+            const openAI = ensureObject(cfg, "openAI");
+            openAI.apiKey ??= "ollama";
+            openAI.endpointEmbedding = `${ollamaBaseUrl(options.ollamaHost)}/v1/embeddings`;
+            openAI.modelEmbedding =
+                options.embeddingModel || DEFAULT_OLLAMA_EMBEDDING_MODEL;
+            cfg.embedding = { provider: "openai" };
+            break;
+        }
+        case "openai": {
+            const openAI = ensureObject(cfg, "openAI");
+            if (options.openaiKey) {
+                openAI.apiKey = options.openaiKey;
+            }
+            openAI.apiKey ??= "sk-REPLACE_ME";
+            openAI.endpointEmbedding =
+                options.embeddingEndpoint ||
+                "https://api.openai.com/v1/embeddings";
+            openAI.modelEmbedding =
+                options.embeddingModel || "text-embedding-3-small";
+            cfg.embedding = { provider: "openai" };
+            break;
+        }
+        case "azure": {
+            // Restore Azure/AI-Systems embeddings: drop the override so the
+            // loader falls back to the azureOpenAI embedding deployments.
+            delete cfg.embedding;
+            break;
+        }
+        case "none": {
+            cfg.embedding = { provider: "none" };
+            break;
+        }
+        default:
+            throw new Error(`Unknown embedding mode: ${embedding}`);
+    }
+}
+
+function patchConfig(cfg, options) {
+    const { provider } = options;
+
+    if (provider === "aisystems") {
+        // Return chat routing to the azureOpenAI deployments already present.
+        delete cfg.modelProvider;
+        applyEmbedding(cfg, options);
+        return cfg;
+    }
+
+    cfg.modelProvider = provider;
+
+    if (provider === "copilot") {
+        const copilot = ensureObject(cfg, "copilot");
+        copilot.defaultModel = options.copilotModel || DEFAULT_COPILOT_MODEL;
+    } else if (provider === "ollama") {
+        const openAI = ensureObject(cfg, "openAI");
+        openAI.apiKey ??= "ollama";
+        const local = ensureObject(openAI, "local");
+        local.apiKey = "None";
+        local.endpoint = ollamaBaseUrl(options.ollamaHost);
+        local.model = options.chatModel || DEFAULT_OLLAMA_CHAT_MODEL;
+    }
+
+    applyEmbedding(cfg, options);
+    return cfg;
+}
+
+// Best-effort validation through the real loader. Non-fatal if the built
+// @typeagent/config dist isn't available (e.g. before `pnpm run build`).
+function validate(outPath) {
+    const distCandidates = [
+        path.resolve(__dirname, "../../packages/config/dist/index.js"),
+    ];
+    const distPath = distCandidates.find((p) => fs.existsSync(p));
+    if (!distPath) {
+        console.warn(
+            "  (skipping loader validation: packages/config/dist not built)",
+        );
+        return Promise.resolve(true);
+    }
+    return import(pathToFileUrl(distPath))
+        .then((mod) => {
+            const { config } = mod.loadRuntimeConfigSync({
+                localPath: outPath,
+                populateProcessEnv: false,
+            });
+            console.log(
+                `  Validated: modelProvider=${config.modelProvider ?? "(azure default)"}, ` +
+                    `embedding.provider=${config.embedding?.provider ?? "(azure default)"}`,
+            );
+            return true;
+        })
+        .catch((e) => {
+            console.error(`  Validation FAILED: ${e.message}`);
+            return false;
+        });
+}
+
+function pathToFileUrl(p) {
+    return new URL(`file://${path.resolve(p).replace(/\\/g, "/")}`).href;
+}
+
+async function main() {
+    const configPath = resolveConfigPath();
+    const backupPath = `${configPath}.bak`;
+
+    if (hasFlag("--restore")) {
+        if (!fs.existsSync(backupPath)) {
+            console.error(`No backup to restore: ${backupPath} not found.`);
+            return 1;
+        }
+        fs.copyFileSync(backupPath, configPath);
+        console.log(`Restored ${configPath} from ${backupPath}`);
+        return 0;
+    }
+
+    const provider = (arg("--provider") ?? "").toLowerCase();
+    const validProviders = ["copilot", "ollama", "aisystems"];
+    if (!validProviders.includes(provider)) {
+        console.error(
+            `setModelProvider: --provider must be one of ${validProviders.join(", ")} ` +
+                "(or pass --restore).",
+        );
+        return 1;
+    }
+
+    const embeddingDefault = provider === "aisystems" ? "azure" : "local";
+    const embedding = (arg("--embedding") ?? embeddingDefault).toLowerCase();
+    const validEmbedding = ["local", "ollama", "openai", "none", "azure"];
+    if (!validEmbedding.includes(embedding)) {
+        console.error(
+            `setModelProvider: --embedding must be one of ${validEmbedding.join(", ")}.`,
+        );
+        return 1;
+    }
+
+    if (!fs.existsSync(configPath)) {
+        console.error(
+            `Config not found: ${configPath}\n` +
+                "This tool patches an EXISTING config. For repo-less machines use " +
+                "generate-selfhost-config.mjs instead.",
+        );
+        return 1;
+    }
+
+    const options = {
+        provider,
+        embedding,
+        copilotModel: arg("--copilot-model"),
+        ollamaHost: arg("--ollama-host", DEFAULT_OLLAMA_HOST),
+        chatModel: arg("--chat-model"),
+        embeddingEndpoint: arg("--embedding-endpoint"),
+        embeddingModel: arg("--embedding-model"),
+        localEmbeddingModel: arg("--local-embedding-model"),
+        openaiKey: arg("--openai-key"),
+    };
+
+    const raw = fs.readFileSync(configPath, "utf8");
+    const cfg = yaml.load(raw) ?? {};
+    if (typeof cfg !== "object" || Array.isArray(cfg)) {
+        console.error(`Config at ${configPath} is not a YAML mapping.`);
+        return 1;
+    }
+
+    patchConfig(cfg, options);
+
+    const body = yaml.dump(cfg, { indent: 2, lineWidth: -1, noRefs: true });
+    const out =
+        "# Copyright (c) Microsoft Corporation.\n" +
+        "# Licensed under the MIT License.\n" +
+        "#\n" +
+        `# Patched by setModelProvider.mjs (chat=${provider}, embedding=${embedding}).\n` +
+        "# Restore the original with: node setModelProvider.mjs --restore\n" +
+        "#\n" +
+        body;
+
+    if (hasFlag("--dry-run")) {
+        process.stdout.write(out);
+        return 0;
+    }
+
+    if (hasFlag("--backup")) {
+        if (fs.existsSync(backupPath)) {
+            console.log(
+                `  Backup already exists (${backupPath}); keeping it as the pristine copy.`,
+            );
+        } else {
+            fs.writeFileSync(backupPath, raw, "utf8");
+            console.log(`  Backed up original to ${backupPath}`);
+        }
+    }
+
+    fs.writeFileSync(configPath, out, "utf8");
+    console.log(
+        `Patched ${configPath} (${provider} chat, ${embedding} embeddings)`,
+    );
+    if (provider === "ollama") {
+        console.log(
+            `  Prereq: 'ollama serve' running with '${options.chatModel || DEFAULT_OLLAMA_CHAT_MODEL}' pulled.`,
+        );
+    } else if (provider === "copilot") {
+        console.log("  Prereq: an authenticated 'copilot' CLI (github login).");
+    }
+    if (embedding === "local") {
+        console.log(
+            "  Embeddings: bundled CPU-only local model (downloads weights on first use).",
+        );
+    } else if (embedding === "none") {
+        console.log(
+            "  Embeddings: disabled — semantic search / fuzzy features degrade gracefully.",
+        );
+    }
+
+    if (!hasFlag("--no-validate")) {
+        const ok = await validate(configPath);
+        if (!ok) {
+            return 1;
+        }
+    }
+    return 0;
+}
+
+main().then((code) => process.exit(code));