pi-web-voice 0.1.0 → 0.1.2

package/README.md

@@ -297,9 +297,14 @@ no restart, and any agent session you have running stays alive.
 
 ## Privacy
 
-Audio goes from your browser to your own pi-web origin, and from there to the
-speech backend you configured. Credentials stay on the server; the page is only told
-which provider is active. Nothing is written to disk and nothing else is contacted.
+Audio goes from your browser to your own pi-web origin, and from there to the speech
+backend you configured. The page is only told which provider is active, never the key.
+Nothing is written to disk and nothing else is contacted.
+
+`~/.pi/agent/voice.env` is parsed into a private object rather than merged into
+`process.env`. pi-web runs the agent's shell commands as children of its own process, so
+anything placed in that environment would be handed to every command the agent ever runs.
+Keeping the credentials out of it means only the transcription request sees them.
 
 ## Compatibility
 

package/lib/config.cjs

@@ -19,17 +19,39 @@ const path = require("node:path");
 
 const ENV_FILE = path.join(os.homedir(), ".pi", "agent", "voice.env");
 
-let loadedEnvFile = null;
+let fileEnv = null;
 
+/**
+ * Reads the key file into a private object.
+ *
+ * Node's own `process.loadEnvFile` would be shorter, but it merges into
+ * `process.env`, and pi-web spawns the agent's shell commands as children of
+ * this process. That would hand every command the agent ever runs a copy of
+ * the speech credentials. Keeping them in a local object means they are only
+ * ever read by the request that needs them.
+ */
 function loadEnvFile() {
-  if (loadedEnvFile !== null) return loadedEnvFile;
-  loadedEnvFile = "";
-  if (!fs.existsSync(ENV_FILE)) return loadedEnvFile;
+  if (fileEnv !== null) return fileEnv;
+  fileEnv = {};
+  if (!fs.existsSync(ENV_FILE)) return fileEnv;
 
   try {
-    // Node applies the same rule as --env-file: the real environment wins.
-    process.loadEnvFile(ENV_FILE);
-    loadedEnvFile = ENV_FILE;
+    for (const line of fs.readFileSync(ENV_FILE, "utf8").split("\n")) {
+      const trimmed = line.trim();
+      if (!trimmed || trimmed.startsWith("#")) continue;
+      const eq = trimmed.indexOf("=");
+      if (eq < 1) continue;
+      const key = trimmed.slice(0, eq).replace(/^export\s+/, "").trim();
+      let value = trimmed.slice(eq + 1).trim();
+      if (
+        value.length >= 2 &&
+        ((value.startsWith('"') && value.endsWith('"')) ||
+          (value.startsWith("'") && value.endsWith("'")))
+      ) {
+        value = value.slice(1, -1);
+      }
+      if (key) fileEnv[key] = value;
+    }
 
     const mode = fs.statSync(ENV_FILE).mode & 0o077;
     if (mode !== 0 && process.platform !== "win32") {
@@ -38,7 +60,12 @@ function loadEnvFile() {
   } catch (error) {
     console.error(`[pi-web-voice] cannot read ${ENV_FILE}: ${error.message}`);
   }
-  return loadedEnvFile;
+  return fileEnv;
+}
+
+/** Real environment first, then the key file. */
+function fromEnv(name) {
+  return process.env[name] ?? loadEnvFile()[name];
 }
 
 // Not worth configuring: one correct value each.
@@ -76,8 +103,14 @@ function azureEndpoint(value, suffix) {
 }
 
 function loadConfig() {
-  const envFile = loadEnvFile();
-  const env = process.env;
+  loadEnvFile();
+  const envFile = fs.existsSync(ENV_FILE) ? ENV_FILE : "";
+  const env = new Proxy(
+    {},
+    {
+      get: (_target, name) => (typeof name === "string" ? fromEnv(name) : undefined),
+    },
+  );
 
   const provider = text(
     env.PI_VOICE_PROVIDER,

package/lib/doctor.cjs

@@ -68,6 +68,9 @@ function describe(config) {
 
 /** Turns provider failures into the thing that is actually wrong. */
 function diagnose(message) {
+  if (/cannot go in an HTTP header|ByteString/i.test(message)) {
+    return "the key in ~/.pi/agent/voice.env is not a real key yet";
+  }
   if (/\b401\b|Unauthorized|Access denied/i.test(message)) {
     return "the key is wrong, or it belongs to a different resource";
   }

package/lib/providers.cjs

@@ -27,6 +27,22 @@ function audioBlob(audio) {
   return new Blob([audio], { type: "audio/wav" });
 }
 
+/**
+ * Credentials travel in HTTP headers, which only carry bytes. A key that is
+ * still the placeholder, or that picked up a stray quote or newline, would
+ * otherwise surface as "Cannot convert argument to a ByteString", which says
+ * nothing about what to go and fix.
+ */
+function checkKey(key, variable) {
+  if (!key) throw new Error(`${variable} is not set in ~/.pi/agent/voice.env`);
+  if (!/^[\x21-\x7e]+$/.test(key)) {
+    throw new Error(
+      `${variable} contains characters that cannot go in an HTTP header — ` +
+        `is the placeholder still in ~/.pi/agent/voice.env?`,
+    );
+  }
+}
+
 /**
  * Azure AI Speech — fast transcription, including MAI-Transcribe-2.
  * Phrase list gives real decode-time keyword biasing, and leaving `locales`
@@ -34,7 +50,8 @@ function audioBlob(audio) {
  */
 async function azureSpeech(audio, config, terms) {
   const { endpoint, key, model, apiVersion, style } = config.azureSpeech;
-  if (!endpoint || !key) throw new Error("azure-speech needs AZURE_SPEECH_ENDPOINT and AZURE_SPEECH_KEY");
+  if (!endpoint) throw new Error("azure-speech needs AZURE_SPEECH_ENDPOINT");
+  checkKey(key, "AZURE_SPEECH_KEY");
 
   const url = `${endpoint}/speechtotext/transcriptions:transcribe?api-version=${encodeURIComponent(apiVersion)}`;
 
@@ -102,7 +119,8 @@ function vocabularyPrompt(terms) {
  */
 async function azureOpenAI(audio, config, terms, languages) {
   const { endpoint, key, deployment, apiVersion } = config.azureOpenAI;
-  if (!endpoint || !key) throw new Error("azure-openai needs AZURE_OPENAI_ENDPOINT and AZURE_OPENAI_API_KEY");
+  if (!endpoint) throw new Error("azure-openai needs AZURE_OPENAI_ENDPOINT");
+  checkKey(key, "AZURE_OPENAI_API_KEY");
 
   const explicit = /\/audio\/transcriptions/.test(endpoint);
   const url = explicit
@@ -144,6 +162,7 @@ async function azureOpenAI(audio, config, terms, languages) {
 async function openAICompatible(audio, config, terms) {
   const { baseUrl, key, model } = config.openai;
   if (!baseUrl) throw new Error("openai needs PI_VOICE_OPENAI_BASE_URL");
+  if (key) checkKey(key, "OPENAI_API_KEY");
 
   const form = new FormData();
   form.append("file", audioBlob(audio), "clip.wav");

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-web-voice",
-  "version": "0.1.0",
+  "version": "0.1.2",
   "description": "Voice input for pi-web. A NODE_OPTIONS hook that injects a microphone button into the chat composer and transcribes speech with Azure AI Speech, Azure OpenAI, or any OpenAI-compatible endpoint. No fork, no patching, no rebuild.",
   "keywords": [
     "pi",