From bbba54c28c127b7ea6883e51d5cdc34c1e9140ed Mon Sep 17 00:00:00 2001
From: LockhartKZ <omer56kazzaz56@gmail.com>
Date: Sat, 4 Jul 2026 18:56:08 +0300
Subject: [PATCH] Translate from preserved source, retry transient errors, add
 parallel requests

- Cue.source keeps the original text: retranslate (per-line or full re-run)
  always starts from the source, never from a previous translation.
- Transient LLM errors (network/timeout/429/5xx) retry with backoff instead
  of killing a long run; cancel mid-request now shows the cancel toast.
- Progress refresh coalesced to one per animation frame (was quadratic on
  large files).
- New "Parallel requests" option (default 1 = old sequential behaviour):
  batches translated concurrently, big speedup on cloud endpoints.
- Bump version to 0.3.0.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---
 README.md                   |   5 +-
 index.html                  |   4 +
 package.json                |   2 +-
 src-tauri/Cargo.lock        |   2 +-
 src-tauri/Cargo.toml        |   2 +-
 src-tauri/tauri.conf.json   |   2 +-
 src/lib/llm/client.ts       |  14 ++++
 src/lib/settings.ts         |   3 +
 src/lib/subtitle/ass.ts     |   2 +-
 src/lib/subtitle/srt.ts     |   2 +-
 src/lib/subtitle/types.ts   |   7 ++
 src/lib/translate/engine.ts | 149 +++++++++++++++++++++++++++---------
 src/main.ts                 |  29 +++++--
 tests/engine.test.ts        |  80 ++++++++++++++++++-
 14 files changed, 251 insertions(+), 52 deletions(-)

diff --git a/README.md b/README.md
index 4deb023..22ec90e 100644
--- a/README.md
+++ b/README.md
@@ -90,8 +90,9 @@ need FUSE 2 once: `sudo apt install libfuse2t64` (older distros: `libfuse2`).
 4. Open **Model & connection**, choose your endpoint preset, click **Test connection**
    (this also lists the available models), pick a model.
 5. Optionally adjust **Translation options** — *Lines per batch* (how many lines go to
-   the model at once) and *Context lines* (how many already-translated lines to show it
-   for continuity).
+   the model at once), *Context lines* (how many already-translated lines to show it
+   for continuity), and *Parallel requests* (batches translated at once — a big speedup
+   on cloud endpoints; keep it at 1 for local servers and best context continuity).
 6. Optionally fill in **Context / notes for the model** — a short description of the
    movie/show so the model uses the right terminology. e.g. for a chess film: *"keep
    chess terms accurate (grandmaster, gambit) and leave move notation like Nf3, O-O
diff --git a/index.html b/index.html
index 78a1023..1c6193e 100644
--- a/index.html
+++ b/index.html
@@ -77,6 +77,10 @@ <h2>Translation options</h2>
             <input id="contextLines" type="number" min="0" max="40" />
             <small class="muted">How many already-translated lines to show the model for continuity (tone, names, gender). 0 disables context.</small>
           </label>
+          <label>Parallel requests
+            <input id="parallelRequests" type="number" min="1" max="8" />
+            <small class="muted">Batches translated at once. 1 keeps full context continuity; higher is much faster on cloud endpoints. Most local servers handle one request at a time anyway.</small>
+          </label>
           <label>Tone / formality
             <select id="tone"></select>
             <small class="muted">Register to aim for. Formal/Informal also nudge the polite vs familiar "you" (vous/Sie vs tu/du) where the language has one.</small>
diff --git a/package.json b/package.json
index 116ff1d..46e97a9 100644
--- a/package.json
+++ b/package.json
@@ -1,7 +1,7 @@
 {
   "name": "subllminal",
   "private": true,
-  "version": "0.2.0",
+  "version": "0.3.0",
   "type": "module",
   "description": "Desktop app that translates .srt/.ass subtitles via a local-first (OpenAI-compatible) LLM, preserving timing, styling, and RTL languages.",
   "scripts": {
diff --git a/src-tauri/Cargo.lock b/src-tauri/Cargo.lock
index 0818067..fb25b04 100644
--- a/src-tauri/Cargo.lock
+++ b/src-tauri/Cargo.lock
@@ -77,7 +77,7 @@ checksum = "7f202df86484c868dbad7eaa557ef785d5c66295e41b460ef922eca0723b842c"
 
 [[package]]
 name = "app"
-version = "0.2.0"
+version = "0.3.0"
 dependencies = [
  "chardetng",
  "encoding_rs",
diff --git a/src-tauri/Cargo.toml b/src-tauri/Cargo.toml
index 94e6572..3709e8b 100644
--- a/src-tauri/Cargo.toml
+++ b/src-tauri/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "app"
-version = "0.2.0"
+version = "0.3.0"
 description = "Desktop app that translates .srt/.ass subtitles via a local-first (OpenAI-compatible) LLM, preserving timing, styling, and RTL languages."
 authors = ["LockhartKZ"]
 license = "MIT"
diff --git a/src-tauri/tauri.conf.json b/src-tauri/tauri.conf.json
index e132e96..873d8c0 100644
--- a/src-tauri/tauri.conf.json
+++ b/src-tauri/tauri.conf.json
@@ -1,7 +1,7 @@
 {
   "$schema": "../node_modules/@tauri-apps/cli/config.schema.json",
   "productName": "SubLLMinal",
-  "version": "0.2.0",
+  "version": "0.3.0",
   "identifier": "com.omer.subllminal",
   "build": {
     "frontendDist": "../dist",
diff --git a/src/lib/llm/client.ts b/src/lib/llm/client.ts
index 3c0a6f5..6f0ba65 100644
--- a/src/lib/llm/client.ts
+++ b/src/lib/llm/client.ts
@@ -59,6 +59,20 @@ export class LlmError extends Error {
   }
 }
 
+/**
+ * Errors worth retrying automatically: network blips, timeouts, rate limits
+ * (429) and server errors (5xx). Auth/4xx errors and cancellations are not —
+ * retrying those just repeats the same failure.
+ */
+export function isTransientLlmError(err: unknown): boolean {
+  return (
+    err instanceof LlmError &&
+    (err.kind === "network" ||
+      err.kind === "timeout" ||
+      (err.kind === "http" && (err.status === 429 || (err.status ?? 0) >= 500)))
+  );
+}
+
 function joinUrl(baseUrl: string, path: string): string {
   return baseUrl.replace(/\/+$/, "") + path;
 }
diff --git a/src/lib/settings.ts b/src/lib/settings.ts
index 41d135a..90a0488 100644
--- a/src/lib/settings.ts
+++ b/src/lib/settings.ts
@@ -10,6 +10,8 @@ export interface AppSettings {
   targetLang: string;
   batchSize: number;
   contextLines: number;
+  /** Batches sent in parallel (1 = sequential; >1 mainly speeds up cloud endpoints). */
+  parallelRequests: number;
   temperature: number;
   /** Free-text notes about the material (genre, terminology, names) for the model. */
   contextNote: string;
@@ -34,6 +36,7 @@ export const DEFAULT_SETTINGS: AppSettings = {
   targetLang: "ar",
   batchSize: 10,
   contextLines: 6,
+  parallelRequests: 1,
   temperature: 0.2,
   contextNote: "",
   glossary: "",
diff --git a/src/lib/subtitle/ass.ts b/src/lib/subtitle/ass.ts
index 4c23399..f2aec6b 100644
--- a/src/lib/subtitle/ass.ts
+++ b/src/lib/subtitle/ass.ts
@@ -84,7 +84,7 @@ export function parseAss(input: string): ParsedSubtitle {
     // Only Dialogue lines are translated; Comment lines are left untouched.
     if (section === "[events]" && /^dialogue\s*:/i.test(trimmed)) {
       const { prefix, text } = splitAtTextField(raw, textIndex);
-      const cue: Cue = { id: String(ordinal++), text };
+      const cue: Cue = { id: String(ordinal++), text, source: text };
       // Fields before Text never contain commas, so a naive split is safe here.
       const cols = raw.slice(raw.indexOf(":") + 1).split(",");
       const at = (name: string) => {
diff --git a/src/lib/subtitle/srt.ts b/src/lib/subtitle/srt.ts
index 6592ec0..4cae181 100644
--- a/src/lib/subtitle/srt.ts
+++ b/src/lib/subtitle/srt.ts
@@ -41,7 +41,7 @@ export function parseSrt(input: string): ParsedSubtitle {
     const timing = lines[1] ?? "";
     const text = lines.slice(2).join("\n");
     const [rawStart, rawEnd] = timing.split("-->");
-    const cue: Cue = { id: String(i), text };
+    const cue: Cue = { id: String(i), text, source: text };
     const start = rawStart ? parseSrtTime(rawStart) : undefined;
     const end = rawEnd ? parseSrtTime(rawEnd) : undefined;
     if (start !== undefined) cue.start = start;
diff --git a/src/lib/subtitle/types.ts b/src/lib/subtitle/types.ts
index ebf1865..663df28 100644
--- a/src/lib/subtitle/types.ts
+++ b/src/lib/subtitle/types.ts
@@ -14,6 +14,13 @@ export interface Cue {
   /** Stable id, unique within the file (a stringified ordinal). */
   id: string;
   text: string;
+  /**
+   * The original source-language text, set at parse time and never mutated.
+   * The engine translates from this (falling back to `text` when absent), so
+   * retranslating always starts from the source — even after `text` holds a
+   * previous translation or a manual edit.
+   */
+  source?: string;
   /** Read-only timing metadata in milliseconds, when the format provides it. */
   start?: number;
   end?: number;
diff --git a/src/lib/translate/engine.ts b/src/lib/translate/engine.ts
index c59b900..734c56e 100644
--- a/src/lib/translate/engine.ts
+++ b/src/lib/translate/engine.ts
@@ -1,4 +1,4 @@
-import type { ChatBackend } from "../llm/client";
+import { isTransientLlmError, type ChatBackend, type ChatMessage } from "../llm/client";
 import type { Cue } from "../subtitle/types";
 import {
   maskTags,
@@ -29,6 +29,19 @@ export interface TranslateParams {
   batchSize?: number; // default 10
   contextLines?: number; // default 6
   maxRetries?: number; // default 2
+  /**
+   * Batches translated in parallel (default 1 = fully sequential). Values >1
+   * speed up cloud endpoints considerably; context continuity becomes
+   * best-effort (a batch sees whatever earlier lines have finished). Local
+   * single-slot servers (LM Studio, llama.cpp default) serialize requests
+   * anyway, so >1 buys little there.
+   */
+  concurrency?: number;
+  /**
+   * Base delay between transient-error retries (network/timeout/429/5xx),
+   * default 1000ms; grows linearly per attempt. Tests pass 0.
+   */
+  retryDelayMs?: number;
   /** Free-text background about the material, injected into the system prompt. */
   contextNote?: string;
   /** Term mappings the model must honour; injected into the system prompt. */
@@ -72,10 +85,14 @@ export async function translateCues(params: TranslateParams): Promise<void> {
   const batchSize = Math.max(1, params.batchSize ?? 10);
   const contextLines = Math.max(0, params.contextLines ?? 6);
   const maxRetries = Math.max(0, params.maxRetries ?? 2);
+  const concurrency = Math.max(1, params.concurrency ?? 1);
+  const retryDelayMs = params.retryDelayMs ?? 1000;
   const tone = toneInstruction(params.tone ?? "");
 
   const prepared: Prepared[] = cues.map((cue) => {
-    const { masked, map } = maskTags(cue.text);
+    // Translate from the immutable source (parsers always set it); `cue.text`
+    // may already hold a previous translation or a manual edit.
+    const { masked, map } = maskTags(cue.source ?? cue.text);
     return { cue, masked, map, translatable: hasTranslatableText(masked) };
   });
 
@@ -92,6 +109,22 @@ export async function translateCues(params: TranslateParams): Promise<void> {
     if (params.signal?.aborted) throw new DOMException("Translation cancelled", "AbortError");
   };
 
+  // One network blip or 429 must not kill a long run: retry transient errors
+  // a couple of times with a short growing delay, then give up for real.
+  const chat = async (messages: ChatMessage[]): Promise<string> => {
+    for (let attempt = 0; ; attempt++) {
+      try {
+        return await backend.chat(messages, params.signal);
+      } catch (err) {
+        if (attempt >= TRANSIENT_RETRIES || !isTransientLlmError(err) || params.signal?.aborted) {
+          throw err;
+        }
+        await new Promise((r) => setTimeout(r, retryDelayMs * (attempt + 1)));
+        ensureNotAborted();
+      }
+    }
+  };
+
   // Render a masked translation into the cue: fix RTL bidi on the MASKED text
   // (markup is still `⟦n⟧` tokens, so Latin inside restored tags is never touched,
   // and break tokens let us force base direction per visual line), then restore.
@@ -113,7 +146,7 @@ export async function translateCues(params: TranslateParams): Promise<void> {
   const runGroup = async (items: Prepared[], strict: boolean): Promise<Set<string>> => {
     const batch: BatchLine[] = items.map((p) => ({ id: p.cue.id, masked: p.masked }));
     const messages = buildMessages(sourceName, targetName, batch, context.slice(-contextLines), strict, params.contextNote, params.glossary, tone);
-    const reply = await backend.chat(messages, params.signal);
+    const reply = await chat(messages);
     const map = parseTranslations(reply);
     const failed = new Set<string>();
     for (const p of items) {
@@ -136,14 +169,13 @@ export async function translateCues(params: TranslateParams): Promise<void> {
       params.glossary,
       tone,
     );
-    const reply = await backend.chat(messages, params.signal);
+    const reply = await chat(messages);
     const parsed = parseTranslations(reply).get(p.cue.id) ?? lenientSingle(reply);
     commit(p, sameTokens(p.masked, parsed) ? parsed : reconcileTokens(p.masked, parsed));
   };
 
-  for (let i = 0; i < translatables.length; i += batchSize) {
+  const processBatch = async (slice: Prepared[]): Promise<void> => {
     ensureNotAborted();
-    const slice = translatables.slice(i, i + batchSize);
     let failed = await runGroup(slice, false);
 
     for (let attempt = 0; attempt < maxRetries && failed.size > 0; attempt++) {
@@ -156,47 +188,88 @@ export async function translateCues(params: TranslateParams): Promise<void> {
       ensureNotAborted();
       await translateSingle(p);
     }
+  };
+
+  const batches: Prepared[][] = [];
+  for (let i = 0; i < translatables.length; i += batchSize) {
+    batches.push(translatables.slice(i, i + batchSize));
   }
 
-  if (params.refine) await refinePass();
+  await runPool(concurrency, batches.length, (i) => processBatch(batches[i]));
 
   // Second pass: ask the model to improve each batch's draft. Best-effort —
-  // one call per batch, no retry/fallback; a refinement that fails id/token
+  // one call per batch, no fallback; a refinement that fails id/token
   // validation is dropped so the (already valid) first-pass draft is kept.
-  async function refinePass(): Promise<void> {
-    for (let i = 0; i < translatables.length; i += batchSize) {
-      ensureNotAborted();
-      const slice = translatables.slice(i, i + batchSize);
-      const batch: RefineLine[] = slice.map((p) => ({
-        id: p.cue.id,
-        source: p.masked,
-        draft: p.draftMasked ?? p.masked,
-      }));
-      const messages = buildRefineMessages(
-        sourceName,
-        targetName,
-        batch,
-        params.contextNote,
-        params.glossary,
-        tone,
-      );
-
-      let map: Map<string, string> | null = null;
+  async function refineBatch(slice: Prepared[]): Promise<void> {
+    ensureNotAborted();
+    const batch: RefineLine[] = slice.map((p) => ({
+      id: p.cue.id,
+      source: p.masked,
+      draft: p.draftMasked ?? p.masked,
+    }));
+    const messages = buildRefineMessages(
+      sourceName,
+      targetName,
+      batch,
+      params.contextNote,
+      params.glossary,
+      tone,
+    );
+
+    let map: Map<string, string> | null = null;
+    try {
+      map = parseTranslations(await chat(messages));
+    } catch (err) {
+      if (params.signal?.aborted) throw err;
+      // A flaky refine call must not discard good drafts: keep them, move on.
+    }
+
+    for (const p of slice) {
+      const cand = map?.get(p.cue.id);
+      if (cand !== undefined && sameTokens(p.masked, cand)) finalize(p, cand);
+      done++;
+      params.onProgress?.(done, total);
+    }
+  }
+
+  if (params.refine) {
+    await runPool(concurrency, batches.length, (i) => refineBatch(batches[i]));
+  }
+}
+
+/** Attempts beyond the first for a transient (network/timeout/429/5xx) error. */
+const TRANSIENT_RETRIES = 2;
+
+/**
+ * Run `jobs` async tasks with at most `limit` in flight. The first failure
+ * stops workers from picking up new jobs (batches already in flight finish
+ * and their results are kept) and is rethrown once all workers settle.
+ */
+async function runPool(
+  limit: number,
+  jobs: number,
+  run: (i: number) => Promise<void>,
+): Promise<void> {
+  let next = 0;
+  let failed = false;
+  let firstError: unknown;
+  const worker = async () => {
+    while (!failed && next < jobs) {
+      const i = next++;
       try {
-        map = parseTranslations(await backend.chat(messages, params.signal));
+        await run(i);
       } catch (err) {
-        if (params.signal?.aborted) throw err;
-        // A flaky refine call must not discard good drafts: keep them, move on.
-      }
-
-      for (const p of slice) {
-        const cand = map?.get(p.cue.id);
-        if (cand !== undefined && sameTokens(p.masked, cand)) finalize(p, cand);
-        done++;
-        params.onProgress?.(done, total);
+        if (!failed) {
+          failed = true;
+          firstError = err;
+        }
+        return;
       }
     }
-  }
+  };
+  const workers = Math.max(1, Math.min(limit, jobs));
+  await Promise.all(Array.from({ length: workers }, worker));
+  if (failed) throw firstError;
 }
 
 /** Force `candidate` to carry exactly the formatting tokens of `sourceMasked`. */
diff --git a/src/main.ts b/src/main.ts
index 8722644..0bf627b 100644
--- a/src/main.ts
+++ b/src/main.ts
@@ -3,7 +3,7 @@ import { fetch as tauriFetch } from "@tauri-apps/plugin-http";
 
 import { parseSubtitle, type ParsedSubtitle } from "./lib/subtitle";
 import type { Cue } from "./lib/subtitle/types";
-import { LlmClient, reasoningEffortFor, type FetchLike, type LlmConfig } from "./lib/llm/client";
+import { LlmClient, LlmError, reasoningEffortFor, type FetchLike, type LlmConfig } from "./lib/llm/client";
 import { ENDPOINT_PRESETS, groupedPresets } from "./lib/llm/presets";
 import { translateCues } from "./lib/translate/engine";
 import { parseGlossary } from "./lib/translate/glossary";
@@ -71,7 +71,11 @@ function msg(err: unknown): string {
 }
 
 function isAbort(err: unknown): boolean {
-  return err instanceof DOMException && err.name === "AbortError";
+  // Cancelling mid-request surfaces as LlmError("aborted"), not a DOMException.
+  return (
+    (err instanceof DOMException && err.name === "AbortError") ||
+    (err instanceof LlmError && err.kind === "aborted")
+  );
 }
 
 let toastTimer: number | undefined;
@@ -295,6 +299,7 @@ function applySettings(s: AppSettings): void {
   el<HTMLSelectElement>("targetLang").value = s.targetLang;
   el<HTMLInputElement>("batchSize").value = String(s.batchSize);
   el<HTMLInputElement>("contextLines").value = String(s.contextLines);
+  el<HTMLInputElement>("parallelRequests").value = String(s.parallelRequests);
   el<HTMLInputElement>("temperature").value = String(s.temperature);
   el<HTMLTextAreaElement>("contextNote").value = s.contextNote;
   el<HTMLTextAreaElement>("glossary").value = s.glossary;
@@ -320,6 +325,7 @@ function gatherSettings(): Omit<AppSettings, "recentFiles"> {
     targetLang: el<HTMLSelectElement>("targetLang").value,
     batchSize: int("batchSize", 10),
     contextLines: int("contextLines", 6),
+    parallelRequests: Math.max(1, int("parallelRequests", 1)),
     temperature: parseFloat(el<HTMLInputElement>("temperature").value) || 0,
     contextNote: el<HTMLTextAreaElement>("contextNote").value,
     glossary: el<HTMLTextAreaElement>("glossary").value,
@@ -418,7 +424,7 @@ function renderPreview(p: ParsedSubtitle): void {
 
     const orig = document.createElement("div");
     orig.className = "cell orig";
-    orig.textContent = cue.text; // source, captured before translation mutates it
+    orig.textContent = cue.source ?? cue.text;
 
     const trans = document.createElement("div");
     trans.className = "cell trans";
@@ -504,6 +510,18 @@ function refreshTranslationCells(): void {
   }
 }
 
+// Progress ticks fire once per translated line; refreshing the whole table on
+// each would be quadratic on large files. Coalesce to one refresh per frame.
+let refreshQueued = false;
+function queueRefreshTranslationCells(): void {
+  if (refreshQueued) return;
+  refreshQueued = true;
+  requestAnimationFrame(() => {
+    refreshQueued = false;
+    refreshTranslationCells();
+  });
+}
+
 // ---- actions ----------------------------------------------------------------
 
 function setBusy(busy: boolean): void {
@@ -594,6 +612,7 @@ async function translate(): Promise<void> {
       targetName: promptName(settings.targetLang),
       batchSize: settings.batchSize,
       contextLines: settings.contextLines,
+      concurrency: settings.parallelRequests,
       contextNote: settings.contextNote,
       glossary: parseGlossary(settings.glossary),
       tone: settings.tone,
@@ -659,7 +678,7 @@ function updateProgress(done: number, total: number): void {
     if (rate > 0) extra = ` · ${rate.toFixed(1)} lines/s · ~${fmtDuration((total - done) / rate)} left`;
   }
   el("progressText").textContent = `${done} / ${total} (${pct}%)${extra}`;
-  refreshTranslationCells();
+  queueRefreshTranslationCells();
 }
 
 function replaceAll(): void {
@@ -779,7 +798,7 @@ function wireEvents(): void {
     // A key just appeared/changed -> probe the key-gated endpoint now.
     if (canDiscoverModels()) void testConnection();
   });
-  for (const id of ["sourceLang", "model", "batchSize", "contextLines", "temperature", "contextNote", "glossary", "tone", "refine", "reasoning"]) {
+  for (const id of ["sourceLang", "model", "batchSize", "contextLines", "parallelRequests", "temperature", "contextNote", "glossary", "tone", "refine", "reasoning"]) {
     el(id).addEventListener("change", persist);
   }
   setupModelDropdown();
diff --git a/tests/engine.test.ts b/tests/engine.test.ts
index 203f4a6..de174d3 100644
--- a/tests/engine.test.ts
+++ b/tests/engine.test.ts
@@ -1,5 +1,5 @@
 import { describe, it, expect } from "vitest";
-import type { ChatBackend, ChatMessage } from "../src/lib/llm/client";
+import { LlmError, type ChatBackend, type ChatMessage } from "../src/lib/llm/client";
 import { parseTranslations } from "../src/lib/translate/prompt";
 import { translateCues, reconcileTokens } from "../src/lib/translate/engine";
 import { RLE, LRE, PDF } from "../src/lib/translate/bidi";
@@ -272,6 +272,84 @@ describe("translateCues", () => {
     expect(seen.at(-1)).toEqual([5, 5]);
   });
 
+  it("translates from cue.source, not the current (already translated) text", async () => {
+    const cues = [{ id: "0", text: "HOLA — a previous translation", source: "hello" }];
+    await translateCues({
+      cues,
+      backend: new MockBackend(),
+      sourceName: "English",
+      targetName: "Spanish",
+    });
+    expect(cues[0].text).toBe("HELLO");
+  });
+
+  it("retries transient errors instead of failing the run", async () => {
+    const inner = new MockBackend();
+    let failures = 2;
+    const backend: ChatBackend = {
+      async chat(messages: ChatMessage[]) {
+        if (failures-- > 0) throw new LlmError("connection reset", "network");
+        return inner.chat(messages);
+      },
+    };
+    const cues = [{ id: "0", text: "hello" }];
+    await translateCues({
+      cues,
+      backend,
+      sourceName: "English",
+      targetName: "Spanish",
+      retryDelayMs: 0,
+    });
+    expect(cues[0].text).toBe("HELLO");
+  });
+
+  it("does not retry non-transient errors (e.g. 401)", async () => {
+    let calls = 0;
+    const backend: ChatBackend = {
+      async chat() {
+        calls++;
+        throw new LlmError("Server returned 401 Unauthorized", "http", 401);
+      },
+    };
+    await expect(
+      translateCues({
+        cues: [{ id: "0", text: "hello" }],
+        backend,
+        sourceName: "English",
+        targetName: "Spanish",
+        retryDelayMs: 0,
+      }),
+    ).rejects.toThrow("401");
+    expect(calls).toBe(1);
+  });
+
+  it("runs batches in parallel when concurrency > 1 and translates all lines", async () => {
+    let inFlight = 0;
+    let maxInFlight = 0;
+    const inner = new MockBackend();
+    const backend: ChatBackend = {
+      async chat(messages: ChatMessage[]) {
+        inFlight++;
+        maxInFlight = Math.max(maxInFlight, inFlight);
+        await new Promise((r) => setTimeout(r, 5));
+        const reply = await inner.chat(messages);
+        inFlight--;
+        return reply;
+      },
+    };
+    const cues = Array.from({ length: 6 }, (_, i) => ({ id: String(i), text: `line ${i}` }));
+    await translateCues({
+      cues,
+      backend,
+      sourceName: "English",
+      targetName: "Spanish",
+      batchSize: 2,
+      concurrency: 3,
+    });
+    expect(maxInFlight).toBeGreaterThan(1);
+    expect(cues.map((c) => c.text)).toEqual(cues.map((_, i) => `LINE ${i}`));
+  });
+
   it("rejects when the signal is already aborted", async () => {
     const ac = new AbortController();
     ac.abort();