Skip to content

Voice-to-text dictation app with local (Nvidia Parakeet/Whisper) and cloud models (BYOK). Privacy-first and available cross-platform.

fix(reasoning): guard Anthropic response text extraction against non-text content (#1709)

Aditya Painuli84dd7902b6ab
diff --git a/src/helpers/anthropicResponse.js b/src/helpers/anthropicResponse.js
new file mode 100644
index 000000000..929bcb79e
--- /dev/null
+++ b/src/helpers/anthropicResponse.js
@@ -0,0 +1,27 @@
+// Text extraction for Anthropic Messages API responses. `content` is a block
+// array whose shape varies run to run: it can be empty (max_tokens hit before
+// any text), or lead with non-text blocks (thinking, tool_use), so
+// `content[0].text` is not a safe read (#1653).
+
+// Returns the trimmed concatenation of all text blocks, or null when the
+// response contains no text block at all.
+function extractAnthropicText(data) {
+ const blocks = Array.isArray(data?.content) ? data.content : [];
+ const parts = blocks
+ .filter((block) => block?.type === "text" && typeof block.text === "string")
+ .map((block) => block.text);
+ if (parts.length === 0) return null;
+ return parts.join("").trim();
+}
+
+function describeMissingAnthropicText(data) {
+ const stopReason = data?.stop_reason;
+ if (stopReason === "max_tokens") {
+ return "Anthropic returned no text before hitting the max_tokens limit";
+ }
+ return stopReason
+ ? `Anthropic response contained no text content (stop_reason: ${stopReason})`
+ : "Anthropic response contained no text content";
+}
+
+module.exports = { extractAnthropicText, describeMissingAnthropicText };
diff --git a/src/helpers/ipcHandlers.js b/src/helpers/ipcHandlers.js
index e041cd0dc..6c98268eb 100644
--- a/src/helpers/ipcHandlers.js
+++ b/src/helpers/ipcHandlers.js
@@ -19,6 +19,7 @@ const {
} = require("./workspacePolicyManager");
const { createEnterpriseIdentityManager } = require("./enterpriseIdentityManager");
const { createCloudConfigRequestHandler } = require("./cloudConfigRequest");
+const { extractAnthropicText, describeMissingAnthropicText } = require("./anthropicResponse");
const {
createPolicyResponseError,
readPolicyResponseError,
@@ -4744,7 +4745,11 @@ class IPCHandlers {
if (config?.requireCompleteOutput && data.stop_reason === "max_tokens") {
throw new Error("Model output was truncated before the selection edit completed");
}
- return { success: true, text: data.content[0].text.trim() };
+ const outputText = extractAnthropicText(data);
+ if (outputText === null) {
+ throw new Error(describeMissingAnthropicText(data));
+ }
+ return { success: true, text: outputText };
} catch (error) {
debugLogger.error("Anthropic reasoning error:", error);
return { success: false, error: error.message };
diff --git a/test/helpers/anthropicResponse.test.js b/test/helpers/anthropicResponse.test.js
new file mode 100644
index 000000000..60574c868
--- /dev/null
+++ b/test/helpers/anthropicResponse.test.js
@@ -0,0 +1,84 @@
+const { test, describe } = require("node:test");
+const assert = require("node:assert");
+const {
+ extractAnthropicText,
+ describeMissingAnthropicText,
+} = require("../../src/helpers/anthropicResponse.js");
+
+describe("extractAnthropicText", () => {
+ test("returns trimmed text for a normal single text block", () => {
+ const data = { content: [{ type: "text", text: " cleaned up dictation " }] };
+ assert.strictEqual(extractAnthropicText(data), "cleaned up dictation");
+ });
+
+ test("joins multiple text blocks", () => {
+ const data = {
+ content: [
+ { type: "text", text: "first " },
+ { type: "text", text: "second" },
+ ],
+ };
+ assert.strictEqual(extractAnthropicText(data), "first second");
+ });
+
+ test("skips leading thinking block and returns the text block (#1653)", () => {
+ const data = {
+ content: [
+ { type: "thinking", thinking: "let me consider..." },
+ { type: "text", text: "the answer" },
+ ],
+ };
+ assert.strictEqual(extractAnthropicText(data), "the answer");
+ });
+
+ test("returns null when content is empty (max_tokens before any text)", () => {
+ assert.strictEqual(extractAnthropicText({ content: [], stop_reason: "max_tokens" }), null);
+ });
+
+ test("returns null when content has only non-text blocks", () => {
+ const data = { content: [{ type: "thinking", thinking: "..." }] };
+ assert.strictEqual(extractAnthropicText(data), null);
+ });
+
+ test("returns null for missing or malformed content", () => {
+ assert.strictEqual(extractAnthropicText({}), null);
+ assert.strictEqual(extractAnthropicText(null), null);
+ assert.strictEqual(extractAnthropicText({ content: "nope" }), null);
+ assert.strictEqual(extractAnthropicText({ content: [{ type: "text" }] }), null);
+ });
+
+ test("preserves empty-string success for an all-whitespace text block", () => {
+ const data = { content: [{ type: "text", text: " " }] };
+ assert.strictEqual(extractAnthropicText(data), "");
+ });
+
+ test("pre-fix expression throws on the shapes this helper now handles", () => {
+ // The old handler read data.content[0].text.trim() directly; these are the
+ // response shapes from #1653 that made it throw the reported TypeError.
+ const shapes = [
+ { content: [], stop_reason: "max_tokens" },
+ { content: [{ type: "thinking", thinking: "..." }] },
+ ];
+ for (const data of shapes) {
+ assert.throws(() => data.content[0].text.trim(), TypeError);
+ assert.doesNotThrow(() => extractAnthropicText(data));
+ }
+ });
+});
+
+describe("describeMissingAnthropicText", () => {
+ test("names max_tokens explicitly", () => {
+ assert.match(describeMissingAnthropicText({ stop_reason: "max_tokens" }), /max_tokens limit/);
+ });
+
+ test("includes other stop reasons", () => {
+ assert.match(describeMissingAnthropicText({ stop_reason: "refusal" }), /refusal/);
+ });
+
+ test("falls back to a generic message", () => {
+ assert.strictEqual(
+ describeMissingAnthropicText({}),
+ "Anthropic response contained no text content"
+ );
+ });
+});