support lmstudio

Author: FreakDev

src/agent/llm-model.test.ts

@@ -27,4 +27,26 @@ describe("createLlmModel", () => {
     assert.ok(model);
     assert.equal(typeof model, "object");
   });
+
+  it("returns a model for lmstudio provider", () => {
+    const model = createLlmModel({
+      ...base,
+      llm: { provider: "lmstudio", model: "qwen/qwen3-4b-2507" },
+    } as PqaConfig);
+    assert.ok(model);
+    assert.equal(typeof model, "object");
+  });
+
+  it("uses custom baseURL for lmstudio provider", () => {
+    const model = createLlmModel({
+      ...base,
+      llm: {
+        provider: "lmstudio",
+        model: "qwen/qwen3-4b-2507",
+        baseURL: "http://127.0.0.1:8080/v1",
+      },
+    } as PqaConfig);
+    assert.ok(model);
+    assert.equal(typeof model, "object");
+  });
 });

src/agent/llm-model.ts

@@ -1,11 +1,15 @@
 import { createAnthropic } from "@ai-sdk/anthropic";
 import { createFireworks } from "@ai-sdk/fireworks";
 import { createGoogleGenerativeAI } from "@ai-sdk/google";
+import { createOpenAICompatible } from "@ai-sdk/openai-compatible";
 import { createOpenAI } from "@ai-sdk/openai";
 import { createOpenRouter } from "@openrouter/ai-sdk-provider";
 import { ollama } from "ollama-ai-provider-v2";
 import type { LanguageModel } from "ai";
-import { PQA_LLM_API_KEY } from "../config/load.js";
+import {
+  DEFAULT_LMSTUDIO_BASE_URL,
+  PQA_LLM_API_KEY,
+} from "../config/load.js";
 import type { PqaConfig } from "../types/config.js";
 
 function llmProviderSettings():
@@ -15,6 +19,15 @@ function llmProviderSettings():
   return apiKey ? { apiKey } : {};
 }
 
+function createLmstudioProvider(config: PqaConfig) {
+  const settings = {
+    name: "lmstudio",
+    baseURL: config.llm.baseURL ?? DEFAULT_LMSTUDIO_BASE_URL,
+    ...llmProviderSettings(),
+  };
+  return createOpenAICompatible(settings);
+}
+
 /** Resolve the Vercel AI SDK language model for the configured LLM provider. */
 export function createLlmModel(config: PqaConfig): LanguageModel {
   const { provider, model } = config.llm;
@@ -31,6 +44,8 @@ export function createLlmModel(config: PqaConfig): LanguageModel {
       return createFireworks(settings)(model);
     case "ollama":
       return ollama(model);
+    case "lmstudio":
+      return createLmstudioProvider(config)(model);
     case "google":
       return createGoogleGenerativeAI(settings)(model);
     case "openrouter":

src/agent/provider-options.test.ts

@@ -181,6 +181,15 @@ describe("buildProviderOptions", () => {
     );
   });
 
+  it("maps thinking budget to lmstudio reasoning effort", () => {
+    assert.deepEqual(
+      buildProviderOptions(
+        config("lmstudio", { enabled: true, budgetTokens: 3_000 }),
+      ),
+      { lmstudio: { reasoningEffort: "low" } },
+    );
+  });
+
   it("enables openrouter reasoning with budget", () => {
     assert.deepEqual(
       buildProviderOptions(

src/agent/provider-options.ts

@@ -52,7 +52,7 @@ function anthropicEffort(
 
 /**
  * Provider-specific options for extended thinking / reasoning.
- * Honors `config.llm.thinking.enabled` across Anthropic, OpenAI, Fireworks, Google, OpenRouter, and Ollama.
+ * Honors `config.llm.thinking.enabled` across Anthropic, OpenAI, LM Studio, Fireworks, Google, OpenRouter, and Ollama.
  */
 export function buildProviderOptions(
   config: PqaConfig,
@@ -84,6 +84,12 @@ export function buildProviderOptions(
           reasoningEffort: resolveOpenAIReasoningEffort(config),
         },
       };
+    case "lmstudio":
+      return {
+        lmstudio: {
+          reasoningEffort: resolveOpenAIReasoningEffort(config),
+        },
+      };
     case "fireworks":
       return {
         fireworks: {

src/config/load.test.ts

@@ -11,6 +11,7 @@ import {
   missingLlmApiKey,
   missingLlmConfig,
   PQA_LLM_API_KEY,
+  PQA_LLM_BASE_URL,
   resolveAgentGuardConfig,
   resolveAgentParallel,
   resolveBrowserHeaded,
@@ -186,6 +187,38 @@ describe("loadConfig", () => {
     }
   });
 
+  it("applies PQA_LLM_BASE_URL when llm.baseURL is unset", async () => {
+    const cwd = mkdtempSync(path.join(tmpdir(), "pqa-config-"));
+    const prevBaseUrl = process.env[PQA_LLM_BASE_URL];
+    process.env[PQA_LLM_BASE_URL] = "http://127.0.0.1:8080/v1";
+    try {
+      const config = await loadConfig(undefined, cwd);
+      assert.equal(config.llm.baseURL, "http://127.0.0.1:8080/v1");
+    } finally {
+      if (prevBaseUrl === undefined) delete process.env[PQA_LLM_BASE_URL];
+      else process.env[PQA_LLM_BASE_URL] = prevBaseUrl;
+    }
+  });
+
+  it("prefers llm.baseURL from pqa.config.json over env var", async () => {
+    const cwd = mkdtempSync(path.join(tmpdir(), "pqa-config-"));
+    writeFileSync(
+      path.join(cwd, "pqa.config.json"),
+      JSON.stringify({
+        llm: { baseURL: "http://custom.local:1234/v1" },
+      }),
+    );
+    const prevBaseUrl = process.env[PQA_LLM_BASE_URL];
+    process.env[PQA_LLM_BASE_URL] = "http://127.0.0.1:8080/v1";
+    try {
+      const config = await loadConfig(undefined, cwd);
+      assert.equal(config.llm.baseURL, "http://custom.local:1234/v1");
+    } finally {
+      if (prevBaseUrl === undefined) delete process.env[PQA_LLM_BASE_URL];
+      else process.env[PQA_LLM_BASE_URL] = prevBaseUrl;
+    }
+  });
+
   it("deep-merges browser.lightpanda overrides", async () => {
     const cwd = mkdtempSync(path.join(tmpdir(), "pqa-config-"));
     writeFileSync(
@@ -222,6 +255,14 @@ describe("missingLlmApiKey", () => {
     };
     assert.equal(missingLlmApiKey(config), undefined);
   });
+
+  it("does not require an API key for lmstudio", () => {
+    const config = {
+      ...minimalConfig("chrome"),
+      llm: { provider: "lmstudio" as const, model: "qwen/qwen3-4b-2507" },
+    };
+    assert.equal(missingLlmApiKey(config), undefined);
+  });
 });
 
 describe("resolveSensitiveEnvVars", () => {

src/config/load.ts

@@ -193,12 +193,14 @@ type LlmProvider = NonNullable<PqaConfig["llm"]["provider"]>;
 function applyLlmEnvOverrides(config: PqaConfig): PqaConfig {
   const provider = process.env.PQA_LLM_PROVIDER as LlmProvider | undefined;
   const model = process.env.PQA_LLM_MODEL;
+  const baseURL = process.env[PQA_LLM_BASE_URL];
   return {
     ...config,
     llm: {
       ...config.llm,
       ...(provider && !config.llm.provider ? { provider } : {}),
       ...(model && !config.llm.model ? { model } : {}),
+      ...(baseURL && !config.llm.baseURL ? { baseURL } : {}),
     },
   };
 }
@@ -334,6 +336,8 @@ export function resolveHealingConfig(config: PqaConfig): Required<
 }
 
 export const PQA_LLM_API_KEY = "PQA_LLM_API_KEY";
+export const PQA_LLM_BASE_URL = "PQA_LLM_BASE_URL";
+export const DEFAULT_LMSTUDIO_BASE_URL = "http://localhost:1234/v1";
 
 const LLM_PROVIDERS_REQUIRING_API_KEY = new Set<
   NonNullable<PqaConfig["llm"]["provider"]>

src/redact/env-secrets.test.ts

@@ -146,4 +146,13 @@ describe("resolveSensitiveEnvVars", () => {
     });
     assert.deepEqual(names, ["PQA_TEST_PASSWORD"]);
   });
+
+  it("omits LLM API key env var for lmstudio", () => {
+    const names = resolveSensitiveEnvVars({
+      ...baseConfig,
+      llm: { provider: "lmstudio", model: "qwen/qwen3-4b-2507" },
+      envVars: ["PQA_TEST_PASSWORD"],
+    });
+    assert.deepEqual(names, ["PQA_TEST_PASSWORD"]);
+  });
 });

src/types/config.ts

@@ -100,13 +100,16 @@ export interface PqaConfig {
       | "openai"
       | "fireworks"
       | "ollama"
+      | "lmstudio"
       | "google"
       | "openrouter";
     model?: string;
+    /** LM Studio OpenAI-compatible server URL. Default: http://localhost:1234/v1 */
+    baseURL?: string;
     /**
      * Extended thinking / reasoning (opt-in).
      * `budgetTokens`: Anthropic, Fireworks & OpenRouter reasoning budget (ignored for Google).
-     * `reasoningEffort`: OpenAI reasoning effort; Anthropic effort; Google thinking level (mapped); OpenRouter reasoning effort.
+     * `reasoningEffort`: OpenAI reasoning effort; Anthropic effort; Google thinking level (mapped); OpenRouter reasoning effort; LM Studio reasoning effort.
      * Ollama: `think` only (other fields ignored).
      */
     thinking?: {