opencode-translate 0.3.4 → 0.3.5

package/README.md

@@ -1,5 +1,20 @@
 # opencode-translate
 
+![banner](https://github.com/user-attachments/assets/0bb6739c-abc5-4c3e-837e-2aaf5533a359)
+
+## Why
+
+LLMs are worse in non-English. Benchmarks confirm it on every frontier model.
+
+- **Claude Opus 4.7 / GPT-5.5 / Gemini 3.1 Pro**: identical coding tasks scored **5/5 in English** but dropped to **0–1/5 in Arabic and Korean** ([LILT, 2025](https://lilt.com/blog/multilingual-ai-coding-gap-non-english-developers)).
+- **Anthropic's own numbers**: Japanese 96.9%, Korean 96.6%, Yoruba 80.3% — English is always the baseline ([Anthropic docs](https://platform.claude.com/docs/en/build-with-claude/multilingual-support)).
+- **Translation Barrier Hypothesis** (JHU, IJCNLP 2025): LLMs reason in an English-dominant latent space, then translate at the last layers — every translation step is a source of error.
+- **Token tax**: Korean ~1.25×, Japanese ~1.25×, Arabic ~3× more tokens per equivalent content. Higher cost, smaller effective context.
+
+This plugin lets you write in your language while the model works in English — best of both worlds.
+
+> Full research write-up: [docs/why.en.md](./docs/why.en.md)
+
 ## Install
 
 ```bash

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "opencode-translate",
-  "version": "0.3.4",
+  "version": "0.3.5",
   "description": "OpenCode plugin that lets the user chat in a configured language while the main chat loop only sees English.",
   "type": "module",
   "main": "src/index.ts",

package/src/activation/messages-transform.ts

@@ -1,6 +1,7 @@
 import type { Hooks } from "@opencode-ai/plugin"
 import { isTextPart, type MessageWithPartsLike } from "../constants"
 import { extractEnglishHistoryText } from "../formatting"
+import { getDisplayLanguageLabel } from "../labels"
 import { logError } from "./logging"
 import { isTranslatedUserDisplayPart } from "./metadata"
 import { resolveSessionState } from "./state"
@@ -18,6 +19,11 @@ export function createMessagesTransformHook(ctx: HookContext): MessagesTransform
       const activeState = resolved.state
       if (!activeState) return
 
+      const extractContext = {
+        nonce: activeState.translate_nonce,
+        label: getDisplayLanguageLabel(activeState.translate_user_lang),
+      }
+
       for (const message of output.messages as MessageWithPartsLike[]) {
         if (message.info.role === "user") {
           for (const part of message.parts) {
@@ -28,7 +34,7 @@ export function createMessagesTransformHook(ctx: HookContext): MessagesTransform
 
         if (message.info.role !== "assistant") continue
         for (const part of message.parts) {
-          if (isTextPart(part)) part.text = extractEnglishHistoryText(part.text, activeState.translate_nonce)
+          if (isTextPart(part)) part.text = extractEnglishHistoryText(part.text, extractContext)
         }
       }
     } catch (error) {

package/src/activation/parts.ts

@@ -9,7 +9,7 @@ import { createSyntheticPartID } from "../translator"
 
 export function createActivationBannerText(options: ResolvedTranslateOptions): string {
   const { modelID } = parseTranslatorModel(options.model)
-  return `✓ Translation mode enabled · translator: ${modelID} · language: ${options.lang}`
+  return `✓ Translation mode enabled · model: ${modelID} · language: ${options.lang}`
 }
 
 export function createLlmOnlyTextPart(

package/src/activation/text-complete.ts

@@ -38,10 +38,9 @@ export function createTextCompleteHook(ctx: HookContext): TextCompleteHook {
           output.text,
           getDisplayLanguageLabel(activeState.translate_user_lang),
           translated,
-          activeState.translate_nonce,
         )
       } catch (error) {
-        output.text = composeTranslationFailureText(output.text, activeState.translate_nonce)
+        output.text = composeTranslationFailureText(output.text)
         await logError(ctx.client, error)
       }
     } catch (error) {

package/src/formatting.ts

@@ -1,35 +1,111 @@
 import { FAILURE_NOTICE } from "./constants"
 
-function startMarker(nonce: string) {
-  return `<!-- oc-translate:${nonce}:start -->`
+// Visible bilingual trailer structure (no invisible delimiters):
+//
+//   <english>
+//
+//   ---
+//
+//   **<label>:**
+//
+//   <translated>
+//
+// Failure variant:
+//
+//   <english>
+//
+//   ---
+//
+//   _Translation unavailable for this segment._
+//
+// The structure renders cleanly under every Markdown front-end we ship to
+// (web `marked`, OpenTUI `<markdown>`, plain text). The history transform
+// recognises it by walking the trailing `---` separator backwards through the
+// stored text and matching the exact label (or failure notice) the plugin
+// emitted for the active session.
+
+const SEPARATOR_LINE = "---"
+
+interface ExtractContext {
+  /** Activation nonce. Used only by the legacy marker fallback. */
+  nonce: string
+  /** Display language label used when composing assistant text. */
+  label: string
 }
 
-function endMarker(nonce: string) {
-  return `<!-- oc-translate:${nonce}:end -->`
+export function composeTranslatedAssistantText(english: string, label: string, translated: string): string {
+  return `${english}\n\n${SEPARATOR_LINE}\n\n**${label}:**\n\n${translated}`
 }
 
-function failedMarker(nonce: string) {
-  return `<!-- oc-translate:${nonce}:status:failed -->`
+export function composeTranslationFailureText(english: string): string {
+  return `${english}\n\n${SEPARATOR_LINE}\n\n${FAILURE_NOTICE}`
 }
 
-export function composeTranslatedAssistantText(
-  english: string,
-  label: string,
-  translated: string,
-  nonce: string,
-): string {
-  return `${english}\n\n${startMarker(nonce)}\n---\n\n**${label}:**\n\n${translated}\n${endMarker(nonce)}`
+export function extractEnglishHistoryText(text: string, ctx: ExtractContext): string {
+  const legacy = extractLegacyMarkerTrailer(text, ctx.nonce)
+  if (legacy !== null) return legacy
+
+  const structural = extractStructuralTrailer(text, ctx.label)
+  if (structural !== null) return structural
+
+  return text
 }
 
-export function composeTranslationFailureText(english: string, nonce: string): string {
-  return `${english}\n\n${startMarker(nonce)}\n${failedMarker(nonce)}\n---\n\n${FAILURE_NOTICE}\n\n${endMarker(nonce)}`
+function extractStructuralTrailer(text: string, label: string): string | null {
+  const labelLine = `**${label}:**`
+  const lines = text.split("\n")
+
+  // Ignore trailing blank lines so a final `\n` (or several) does not throw
+  // off the structural match.
+  let endLine = lines.length - 1
+  while (endLine >= 0 && lines[endLine] === "") endLine -= 1
+
+  // Smallest valid failure trailer is 5 lines: english, "", ---, "", FAILURE.
+  if (endLine < 4) return null
+
+  // Walk backwards: the trailer's `---` is always preceded by exactly one
+  // blank line and a non-empty English half, and followed by exactly one
+  // blank line plus either the label line or the failure notice extending
+  // to the end of `text`.
+  for (let i = endLine; i >= 2; i -= 1) {
+    if (lines[i] !== SEPARATOR_LINE) continue
+    if (lines[i - 1] !== "") continue
+    if (i - 2 < 0) continue
+    if (i + 2 > endLine) continue
+    if (lines[i + 1] !== "") continue
+
+    const headLine = lines[i + 2]
+
+    if (headLine === labelLine) {
+      // Success trailer: `**label:**\n\n<translated...>` extending to end.
+      if (i + 3 > endLine) continue
+      if (lines[i + 3] !== "") continue
+      // Translated content occupies lines i+4..endLine and must be non-empty.
+      if (i + 4 > endLine) continue
+      return lines.slice(0, i - 1).join("\n")
+    }
+
+    if (headLine === FAILURE_NOTICE) {
+      // Failure trailer ends exactly at the notice.
+      if (i + 2 !== endLine) continue
+      return lines.slice(0, i - 1).join("\n")
+    }
+  }
+
+  return null
 }
 
-export function extractEnglishHistoryText(text: string, nonce: string): string {
+// Legacy fallback. Earlier versions of the plugin stored bilingual assistant
+// text wrapped in `<!-- oc-translate:{nonce}:start -->` ... `<!-- oc-translate:{nonce}:end -->`
+// HTML comments. Those comments render cleanly in the web UI but show up as
+// literal text in the terminal UI, which is the bug this refactor fixes.
+// We keep parsing them so existing sessions continue to feed English-only
+// history to the LLM after the plugin is upgraded.
+function extractLegacyMarkerTrailer(text: string, nonce: string): string | null {
   const lines = text.split("\n")
-  const exactEnd = endMarker(nonce)
-  const exactStart = startMarker(nonce)
-  const exactFailed = failedMarker(nonce)
+  const exactStart = `<!-- oc-translate:${nonce}:start -->`
+  const exactEnd = `<!-- oc-translate:${nonce}:end -->`
+  const exactFailed = `<!-- oc-translate:${nonce}:status:failed -->`
 
   let lastNonEmpty = -1
   for (let index = lines.length - 1; index >= 0; index -= 1) {
@@ -38,10 +114,7 @@ export function extractEnglishHistoryText(text: string, nonce: string): string {
       break
     }
   }
-
-  if (lastNonEmpty < 0 || lines[lastNonEmpty] !== exactEnd) {
-    return text
-  }
+  if (lastNonEmpty < 0 || lines[lastNonEmpty] !== exactEnd) return null
 
   let endIndex = -1
   for (let index = lastNonEmpty; index >= 0; index -= 1) {
@@ -50,7 +123,7 @@ export function extractEnglishHistoryText(text: string, nonce: string): string {
       break
     }
   }
-  if (endIndex < 0) return text
+  if (endIndex < 0) return null
 
   let startIndex = -1
   for (let index = endIndex - 1; index >= 0; index -= 1) {
@@ -59,27 +132,26 @@ export function extractEnglishHistoryText(text: string, nonce: string): string {
       break
     }
   }
-  if (startIndex < 2) return text
+  if (startIndex < 2) return null
 
   let cursor = startIndex + 1
   const failed = lines[cursor] === exactFailed
   if (failed) cursor += 1
 
-  if (lines[cursor] !== "---") return text
-  if (lines[cursor + 1] !== "") return text
+  if (lines[cursor] !== SEPARATOR_LINE) return null
+  if (lines[cursor + 1] !== "") return null
 
   if (failed) {
-    if (lines[cursor + 2] !== FAILURE_NOTICE) return text
-    if (lines[cursor + 3] !== "") return text
-    if (cursor + 4 !== endIndex) return text
+    if (lines[cursor + 2] !== FAILURE_NOTICE) return null
+    if (lines[cursor + 3] !== "") return null
+    if (cursor + 4 !== endIndex) return null
   } else {
     const labelLine = lines[cursor + 2]
-    if (!/^\*\*.+:\*\*$/.test(labelLine)) return text
-    if (lines[cursor + 3] !== "") return text
-    if (cursor + 4 > endIndex) return text
+    if (!/^\*\*.+:\*\*$/.test(labelLine)) return null
+    if (lines[cursor + 3] !== "") return null
+    if (cursor + 4 > endIndex) return null
   }
 
-  if (lines[startIndex - 1] !== "") return text
-  const english = lines.slice(0, startIndex - 1).join("\n")
-  return english
+  if (lines[startIndex - 1] !== "") return null
+  return lines.slice(0, startIndex - 1).join("\n")
 }