glotfile 0.3.1 → 0.4.0

package/dist/server/cli.js

@@ -281,7 +281,7 @@ function defaultState() {
     keys: {}
   };
 }
-var CURRENT_VERSION, STATES, PLURAL_CATEGORIES, EXACT_SELECTOR_RE, LOCALE_CASES, PROVIDERS, GlotfileError;
+var CURRENT_VERSION, STATES, PLURAL_CATEGORIES, EXACT_SELECTOR_RE, LOCALE_CASES, PROVIDERS, PROMPT_STYLES, GlotfileError;
 var init_schema = __esm({
   "src/server/schema.ts"() {
     "use strict";
@@ -292,6 +292,7 @@ var init_schema = __esm({
     EXACT_SELECTOR_RE = /^=\d+$/;
     LOCALE_CASES = ["lower-hyphen", "lower-underscore", "bcp47-hyphen", "bcp47-underscore"];
     PROVIDERS = ["anthropic", "openai", "bedrock", "openrouter", "ollama", "claude-code"];
+    PROMPT_STYLES = ["default", "translategemma"];
     GlotfileError = class extends Error {
     };
   }
@@ -1563,8 +1564,8 @@ var init_export_run = __esm({
 });
 
 // src/server/ai/provider.ts
-function buildSystemPrompt() {
-  return [
+function buildSystemPrompt(hasPluralItems) {
+  const lines = [
     "You are a professional software localization engine for a UI string catalog.",
     "Your goal: translate each source UI string into its target locale accurately and idiomatically, as a native speaker would phrase it in a real app interface.",
     "",
@@ -1576,20 +1577,28 @@ function buildSystemPrompt() {
     "- Glossary: a term marked do-not-translate MUST appear unchanged in the translation. A term with a forced translation for the target locale MUST use that exact translation.",
     "- Respect the max length (characters) when given; prefer a shorter natural phrasing over exceeding it.",
     "- Match the register and capitalization conventions of the target language and of UI microcopy.",
-    "- Return ONLY the translated string for each item \u2014 no quotes, notes, or explanations.",
-    "",
-    'Plural items: an item with a `plural` field gives you the source plural FORMS (keyed by CLDR category) and the `categories` REQUIRED for the target language. Return a `forms` object with one idiomatic translation per REQUIRED category \u2014 including categories the source language does not have (infer them from meaning). Keep the count token shown in the source forms (e.g. {count}) in every form that states a quantity; the `zero`, `one`, and `two` forms MAY omit it when that is natural in the target language \u2014 e.g. "No files", "One file", or a dual form that encodes the count grammatically (Arabic \u0645\u0644\u0641\u0627\u0646). Never introduce a placeholder the source did not have. For these items return `forms` instead of `translation`.'
-  ].join("\n");
+    "- Return ONLY the translated string for each item \u2014 no quotes, notes, or explanations."
+  ];
+  if (hasPluralItems) {
+    lines.push(
+      "",
+      'Plural items: an item with a `plural` field gives you the source plural FORMS (keyed by CLDR category) and the `categories` REQUIRED for the target language. Return a `forms` object with one idiomatic translation per REQUIRED category \u2014 including categories the source language does not have (infer them from meaning). Keep the count token shown in the source forms (e.g. {count}) in every form that states a quantity; the `zero`, `one`, and `two` forms MAY omit it when that is natural in the target language \u2014 e.g. "No files", "One file", or a dual form that encodes the count grammatically (Arabic \u0645\u0644\u0641\u0627\u0646). Never introduce a placeholder the source did not have. For these items return `forms` instead of `translation`.'
+    );
+  }
+  return lines.join("\n");
 }
 function buildBatchPrompt(reqs) {
   const targetLocale = reqs[0]?.targetLocale ?? "";
+  const hasPluralItems = reqs.some((r) => r.plural !== void 0);
   const items = reqs.map((r) => {
     const base = {
       id: r.id,
       key: r.key,
       context: r.context ?? null,
       maxLength: r.maxLength ?? null,
-      placeholders: r.placeholders,
+      // Wrap in braces so the model sees "{site}" not "site" — makes the visual
+      // connection to the source string obvious and reduces rename errors.
+      placeholders: r.placeholders.map((p) => `{${p}}`),
       glossary: r.glossary ?? [],
       hasScreenshot: r.image !== void 0
     };
@@ -1598,10 +1607,24 @@ function buildBatchPrompt(reqs) {
     }
     return { ...base, source: r.source };
   });
+  const returnFormat = hasPluralItems ? 'For a scalar item (has `source`) return {"id","translation"}; for a plural item (has `plural`) return {"id","forms"} with one string per required category.' : 'Return {"id","translation"} for each item.';
   return `Translate every item below into the target locale: ${targetLocale}. All items share this one target language.
-Glossary entries are constraints you MUST apply. Items with hasScreenshot:true have a screenshot supplied as a separate image block above; use it for context. For a scalar item (has \`source\`) return {"id","translation"}; for a plural item (has \`plural\`) return {"id","forms"} with one string per required category. Return JSON {"items":[\u2026]}.
+Glossary entries are constraints you MUST apply. Items with hasScreenshot:true have a screenshot supplied as a separate image block above; use it for context. ${returnFormat} Return JSON {"items":[\u2026]}.
 ` + JSON.stringify(items, null, 2);
 }
+function buildTranslateGemmaSystemPrompt(sourceLocale, targetLocale) {
+  return `You are a professional ${sourceLocale} to ${targetLocale} translator. Your goal is to accurately convey the meaning and nuances of the original ${sourceLocale} text while adhering to ${targetLocale} grammar, vocabulary, and cultural sensitivities. Produce only the ${targetLocale} translation, without any additional explanations or commentary. Preserve every interpolation placeholder exactly as written (e.g. {site}, {count}, {name}) \u2014 do not translate, rename, or remove them. Preserve markdown formatting markers exactly as written (e.g. **bold**, *italic*, __underline__) \u2014 copy them into the translation in the same positions. Please translate the following ${sourceLocale} text into ${targetLocale}:`;
+}
+function buildTranslateGemmaUserPrompt(source) {
+  return `
+
+${source}`;
+}
+function parseTranslateGemmaResponse(text, id) {
+  const translation = text.trim();
+  if (!translation) return { id, error: "No translation returned" };
+  return { id, translation };
+}
 var BATCH_SCHEMA;
 var init_provider = __esm({
   "src/server/ai/provider.ts"() {
@@ -1766,7 +1789,7 @@ var init_anthropic = __esm({
         const res = await this.client.messages.create({
           model: this.config.model,
           max_tokens: 8192,
-          system: [{ type: "text", text: buildSystemPrompt(), cache_control: { type: "ephemeral" } }],
+          system: [{ type: "text", text: buildSystemPrompt(batch.some((r) => r.plural !== void 0)), cache_control: { type: "ephemeral" } }],
           output_config: { format: { type: "json_schema", schema: BATCH_SCHEMA } },
           messages: [{ role: "user", content }]
         }, { signal });
@@ -1864,7 +1887,7 @@ var init_openai = __esm({
           // via runBatched, so non-strict schema guidance is sufficient.
           response_format: { type: "json_schema", json_schema: { name: "translations", schema: BATCH_SCHEMA, strict: false } },
           messages: [
-            { role: "system", content: buildSystemPrompt() },
+            { role: "system", content: buildSystemPrompt(batch.some((r) => r.plural !== void 0)) },
             { role: "user", content: this.buildUserContent(batch) }
           ]
         }, { signal });
@@ -1976,7 +1999,7 @@ var init_bedrock = __esm({
       buildInput(batch) {
         const input = {
           modelId: this.config.model,
-          system: [{ text: buildSystemPrompt() }],
+          system: [{ text: buildSystemPrompt(batch.some((r) => r.plural !== void 0)) }],
           messages: [{ role: "user", content: this.buildContentBlocks(batch) }]
         };
         if (!this.isMeta()) {
@@ -2043,6 +2066,7 @@ var OLLAMA_BASE_URL, OllamaProvider;
 var init_ollama = __esm({
   "src/server/ai/ollama.ts"() {
     "use strict";
+    init_provider();
     init_openai();
     OLLAMA_BASE_URL = "http://localhost:11434/v1";
     OllamaProvider = class extends OpenAIProvider {
@@ -2052,6 +2076,26 @@ var init_ollama = __esm({
       supportsVision() {
         return this.config.vision === true;
       }
+      // translategemma expects a per-item system prompt, a plain-text user message
+      // (no JSON, no structured output), and returns a plain-text translation.
+      async callBatch(batch, signal) {
+        if (this.config.promptStyle !== "translategemma") {
+          return super.callBatch(batch, signal);
+        }
+        const results = [];
+        for (const req of batch) {
+          const res = await this.client.chat.completions.create({
+            model: this.config.model,
+            messages: [
+              { role: "system", content: buildTranslateGemmaSystemPrompt(req.sourceLocale, req.targetLocale) },
+              { role: "user", content: buildTranslateGemmaUserPrompt(req.source) }
+            ]
+          }, { signal });
+          const text = res.choices?.[0]?.message?.content ?? "";
+          results.push(parseTranslateGemmaResponse(text, req.id));
+        }
+        return results;
+      }
     };
   }
 });
@@ -2138,7 +2182,7 @@ var init_claudecode = __esm({
         const prompt = buildBatchPrompt(batch);
         let result;
         try {
-          result = await this.spawnFn(prompt, buildSystemPrompt(), this.config.model);
+          result = await this.spawnFn(prompt, buildSystemPrompt(batch.some((r) => r.plural !== void 0)), this.config.model);
         } catch (err) {
           if (signal?.aborted) return [];
           throw err;
@@ -2226,18 +2270,36 @@ function coerceAi(raw) {
     endpoint: typeof a.endpoint === "string" ? a.endpoint : null,
     region: typeof a.region === "string" ? a.region : null,
     batchSize: typeof a.batchSize === "number" && a.batchSize > 0 ? a.batchSize : DEFAULT_AI.batchSize,
-    vision: typeof a.vision === "boolean" ? a.vision : void 0
+    concurrency: typeof a.concurrency === "number" && a.concurrency > 0 ? a.concurrency : void 0,
+    contextBatchSize: typeof a.contextBatchSize === "number" && a.contextBatchSize > 0 ? a.contextBatchSize : void 0,
+    contextConcurrency: typeof a.contextConcurrency === "number" && a.contextConcurrency > 0 ? a.contextConcurrency : void 0,
+    vision: typeof a.vision === "boolean" ? a.vision : void 0,
+    promptStyle: PROMPT_STYLES.includes(a.promptStyle) ? a.promptStyle : void 0
   };
 }
+function coerceProfiles(raw) {
+  if (!raw || typeof raw !== "object" || Array.isArray(raw)) return {};
+  const result = {};
+  for (const [k, v] of Object.entries(raw)) {
+    if (typeof k === "string" && k.trim()) result[k] = coerceAi(v);
+  }
+  return result;
+}
 function loadLocalSettings(projectRoot) {
   const raw = readJson(settingsPath(projectRoot));
-  return { ai: coerceAi(raw.ai), editor: isEditorId(raw.editor) ? raw.editor : DEFAULT_EDITOR };
+  const profiles = coerceProfiles(raw.profiles);
+  const activeProfile = typeof raw.activeProfile === "string" && raw.activeProfile in profiles ? raw.activeProfile : null;
+  const baseAi = coerceAi(raw.ai);
+  const ai = activeProfile ? profiles[activeProfile] : baseAi;
+  return { ai, editor: isEditorId(raw.editor) ? raw.editor : DEFAULT_EDITOR, profiles, activeProfile };
 }
 function saveLocalSettings(projectRoot, patch) {
   const path = settingsPath(projectRoot);
   const merged = { ...readJson(path) };
   if (patch.ai !== void 0) merged.ai = patch.ai;
   if (patch.editor !== void 0) merged.editor = patch.editor;
+  if (patch.profiles !== void 0) merged.profiles = patch.profiles;
+  if (patch.activeProfile !== void 0) merged.activeProfile = patch.activeProfile;
   ensureGlotfileDir(projectRoot);
   writeFileAtomic(path, JSON.stringify(merged, null, 2) + "\n");
 }
@@ -2313,6 +2375,7 @@ function selectRequests(state, opts) {
           id: String(id++),
           key,
           source: other,
+          sourceLocale: state.config.sourceLocale,
           context: entry.context,
           targetLocale: locale,
           maxLength: entry.maxLength,
@@ -2333,6 +2396,7 @@ function selectRequests(state, opts) {
         id: String(id++),
         key,
         source,
+        sourceLocale: state.config.sourceLocale,
         context: entry.context,
         targetLocale: locale,
         maxLength: entry.maxLength,
@@ -2909,7 +2973,7 @@ function buildContextBatchPrompt(reqs) {
 ${s.lines}
 \`\`\``;
     }).join("\n\n") : "(no code references found \u2014 infer from key path and source value)";
-    return { id: r.id, key: r.key, source: r.source, codeSnippets: snippetText, hasScreenshot: r.image !== void 0 };
+    return { id: r.id, key: r.key, source: r.source, codeSnippets: snippetText };
   });
   return 'Write a context note for each key. Return JSON {"items":[{"id","context"}]}.\n' + JSON.stringify(items, null, 2);
 }
@@ -3893,6 +3957,43 @@ function createApi(deps) {
     saveLocalSettings(projectRoot, patch);
     return c.json({ ok: true });
   });
+  app.get("/ai-profiles", (c) => {
+    const ls = loadLocalSettings(projectRoot);
+    return c.json({ profiles: ls.profiles, activeProfile: ls.activeProfile });
+  });
+  app.put("/ai-profiles/:name", async (c) => {
+    const name = c.req.param("name").trim();
+    if (!name) return c.json({ error: "name required" }, 400);
+    const body = await c.req.json().catch(() => ({}));
+    const err = aiConfigError(body);
+    if (err) return c.json({ error: err }, 400);
+    const ls = loadLocalSettings(projectRoot);
+    saveLocalSettings(projectRoot, { profiles: { ...ls.profiles, [name]: body } });
+    return c.json({ ok: true });
+  });
+  app.delete("/ai-profiles/:name", (c) => {
+    const name = c.req.param("name");
+    const ls = loadLocalSettings(projectRoot);
+    if (!(name in ls.profiles)) return c.json({ error: "profile not found" }, 404);
+    const profiles = { ...ls.profiles };
+    delete profiles[name];
+    const patch = { profiles };
+    if (ls.activeProfile === name) patch.activeProfile = null;
+    saveLocalSettings(projectRoot, patch);
+    return c.json({ ok: true });
+  });
+  app.post("/ai-profiles/active", async (c) => {
+    const { name } = await c.req.json().catch(() => ({}));
+    if (name !== null && name !== void 0) {
+      if (typeof name !== "string") return c.json({ error: "name must be a string or null" }, 400);
+      const ls = loadLocalSettings(projectRoot);
+      if (name !== "" && !(name in ls.profiles)) return c.json({ error: "profile not found" }, 404);
+      saveLocalSettings(projectRoot, { activeProfile: name || null });
+    } else {
+      saveLocalSettings(projectRoot, { activeProfile: null });
+    }
+    return c.json({ ok: true });
+  });
   app.get("/file", (c) => c.json({ path: deps.statePath, name: basename(deps.statePath), dir: projectRoot, project: basename(projectRoot) }));
   app.get("/files", (c) => {
     const found = /* @__PURE__ */ new Map();
@@ -4071,7 +4172,7 @@ function createApi(deps) {
     return c.json({ removed });
   });
   app.post("/keys/bulk-meta", async (c) => {
-    const { keys, addTags, removeTags, skipTranslate } = await c.req.json();
+    const { keys, addTags, removeTags, skipTranslate, clearContext } = await c.req.json();
     if (!Array.isArray(keys) || keys.length === 0) return c.json({ error: "keys must be a non-empty array" }, 400);
     const s = load();
     let updated = 0;
@@ -4089,6 +4190,11 @@ function createApi(deps) {
         if (skipTranslate) setMetadata(s, key, { skipTranslate: true });
         else delete entry.skipTranslate;
       }
+      if (clearContext === true) {
+        delete entry.context;
+        delete entry.contextSource;
+        delete entry.contextAt;
+      }
       updated++;
     }
     persist(s);
@@ -4369,7 +4475,7 @@ function createApi(deps) {
       console.log(`[translate] ${reqs.length} string(s) \u2192 ${aiCfg.model}`);
       let totalWritten = 0;
       const allErrors = [];
-      const system = buildSystemPrompt();
+      const system = buildSystemPrompt(reqs.some((r) => r.plural !== void 0));
       const reqById = new Map(reqs.map((r) => [r.id, r]));
       const localeTotals = /* @__PURE__ */ new Map();
       for (const r of reqs) localeTotals.set(r.targetLocale, (localeTotals.get(r.targetLocale) ?? 0) + 1);
@@ -4398,7 +4504,7 @@ function createApi(deps) {
             system,
             items: batchResults.map((r) => {
               const req = reqById.get(r.id);
-              return { id: r.id, key: req?.key ?? "", source: req?.source ?? "", targetLocale: req?.targetLocale, context: req?.context, glossary: req?.glossary, screenshot: req ? s.keys[req.key]?.screenshot : void 0 };
+              return { id: r.id, key: req?.key ?? "", source: req?.source ?? "", targetLocale: req?.targetLocale, context: req?.context, glossary: req?.glossary, screenshot: req ? fresh.keys[req.key]?.screenshot : void 0 };
             }),
             results: batchResults
           });
@@ -4413,7 +4519,7 @@ function createApi(deps) {
         onLocaleDone: (locale) => {
           void stream.writeSSE({ event: "locale-done", data: JSON.stringify({ locale }) });
         }
-      }, void 0, signal);
+      }, aiCfg.concurrency, signal);
       if (!signal?.aborted) {
         console.log(`[translate] done \u2014 wrote ${totalWritten}, ${allErrors.length} error(s)`);
         await stream.writeSSE({ event: "done", data: JSON.stringify({ written: totalWritten, errors: allErrors }) });
@@ -4444,14 +4550,15 @@ function createApi(deps) {
       }
       const { skipped } = attachScreenshotsForProvider(toTranslate, s, projectRoot, provider.supportsVision());
       if (skipped) console.warn(`Model "${aiCfg.model}" has no vision support; ${skipped} screenshot(s) ignored.`);
-      const results = await runLocaleParallel(toTranslate, provider);
-      ({ written, errors } = applyResults(s, toTranslate, results, void 0, force));
+      const results = await runLocaleParallel(toTranslate, provider, {}, aiCfg.concurrency);
+      const latest = load();
+      ({ written, errors } = applyResults(latest, toTranslate, results, void 0, force));
       const entry = {
         at: (/* @__PURE__ */ new Date()).toISOString(),
         kind: "translate",
         summary: `Translated ${toTranslate.length} item(s)`,
         model: aiCfg.model,
-        system: buildSystemPrompt(),
+        system: buildSystemPrompt(toTranslate.some((r) => r.plural !== void 0)),
         // Log the screenshot PATH only — never the image bytes.
         items: toTranslate.map((r) => ({
           id: r.id,
@@ -4460,13 +4567,13 @@ function createApi(deps) {
           targetLocale: r.targetLocale,
           context: r.context,
           glossary: r.glossary,
-          screenshot: s.keys[r.key]?.screenshot
+          screenshot: latest.keys[r.key]?.screenshot
         })),
         results
       };
       appendLog(projectRoot, entry);
+      persist(latest);
     }
-    persist(s);
     return c.json({ requested: reqs.length, written, errors });
   }));
   app.get("/log", (c) => c.json(readLog(projectRoot, 100)));
@@ -4522,55 +4629,95 @@ function createApi(deps) {
     return c.json({ indexed: true, scannedAt: cache2.scannedAt, used: computeUsedKeys(load(), cache2) });
   });
   app.post("/context/build", async (c) => {
+    const signal = c.req.raw.signal;
     const body = await c.req.json().catch(() => ({}));
-    const s = load();
-    const cache2 = loadUsageCache(projectRoot);
-    if (!cache2) return c.json({ error: "No usage index found. Run 'glotfile scan' first." }, 400);
-    const targets = selectContextTargets(s, {
-      all: body.all,
-      keyGlob: body.keyGlob,
-      limit: body.limit,
-      since: body.since,
-      keys: body.keys
-    }, cache2, body.lastRunAt);
-    if (!targets.length) return c.json({ requested: 0, written: 0, errors: [] });
-    const aiCfg = loadLocalSettings(projectRoot).ai;
-    let provider;
-    try {
-      provider = deps.makeProvider ? deps.makeProvider() : makeProvider(aiCfg);
-    } catch (e) {
-      return c.json({ error: e.message }, 400);
-    }
-    const fileCache = /* @__PURE__ */ new Map();
-    for (const target of targets) {
-      const allRefs = Object.entries(cache2.files).flatMap(
-        ([file, entry]) => entry.refs.filter((r) => r.key === target.key).map((r) => ({
-          key: r.key,
-          file,
-          line: r.line,
-          col: r.col,
-          scanner: r.scanner
-        }))
-      );
-      target.usageSnippets = extractSnippets(allRefs, projectRoot, fileCache);
-    }
-    const system = buildContextSystemPrompt();
-    const prompt = buildContextBatchPrompt(targets);
-    const raw = await provider.complete({ system, content: [{ type: "text", text: prompt }], schema: CONTEXT_BATCH_SCHEMA });
-    const batch = raw;
-    const { written, errors } = applyContext(s, targets, batch.items ?? []);
-    appendLog(projectRoot, {
-      at: (/* @__PURE__ */ new Date()).toISOString(),
-      kind: "context",
-      summary: `Generated context for ${targets.length} key(s)`,
-      model: aiCfg.model,
-      system,
-      items: targets.map((t) => ({ id: t.id, key: t.key, source: t.source })),
-      results: (batch.items ?? []).map((r) => ({ id: r.id, value: r.context, error: r.error }))
+    return streamSSE(c, async (stream) => {
+      const s = load();
+      const cache2 = loadUsageCache(projectRoot);
+      if (!cache2) {
+        await stream.writeSSE({ event: "error", data: JSON.stringify({ error: "No usage index found. Run 'glotfile scan' first." }) });
+        return;
+      }
+      const targets = selectContextTargets(s, {
+        all: body.all,
+        keyGlob: body.keyGlob,
+        limit: body.limit,
+        since: body.since,
+        keys: body.keys
+      }, cache2, body.lastRunAt);
+      if (!targets.length) {
+        await stream.writeSSE({ event: "done", data: JSON.stringify({ requested: 0, written: 0, errors: [] }) });
+        return;
+      }
+      const aiCfg = loadLocalSettings(projectRoot).ai;
+      let provider;
+      try {
+        provider = deps.makeProvider ? deps.makeProvider() : makeProvider(aiCfg);
+      } catch (e) {
+        await stream.writeSSE({ event: "error", data: JSON.stringify({ error: e.message }) });
+        return;
+      }
+      await stream.writeSSE({ event: "start", data: JSON.stringify({ total: targets.length }) });
+      const fileCache = /* @__PURE__ */ new Map();
+      for (const target of targets) {
+        const allRefs = Object.entries(cache2.files).flatMap(
+          ([file, entry]) => entry.refs.filter((r) => r.key === target.key).map((r) => ({
+            key: r.key,
+            file,
+            line: r.line,
+            col: r.col,
+            scanner: r.scanner
+          }))
+        );
+        target.usageSnippets = extractSnippets(allRefs, projectRoot, fileCache);
+      }
+      const system = buildContextSystemPrompt();
+      const batchSize = aiCfg.contextBatchSize ?? aiCfg.batchSize ?? 10;
+      const concurrency = aiCfg.contextConcurrency ?? aiCfg.concurrency ?? 3;
+      const chunks = [];
+      for (let i = 0; i < targets.length; i += batchSize) chunks.push(targets.slice(i, i + batchSize));
+      let totalWritten = 0;
+      let totalDone = 0;
+      const allErrors = [];
+      let next = 0;
+      async function worker() {
+        while (next < chunks.length) {
+          if (signal?.aborted) break;
+          const chunk2 = chunks[next++];
+          let raw;
+          try {
+            raw = await provider.complete({ system, content: [{ type: "text", text: buildContextBatchPrompt(chunk2) }], schema: CONTEXT_BATCH_SCHEMA });
+          } catch (e) {
+            totalDone += chunk2.length;
+            allErrors.push(...chunk2.map((t) => ({ key: t.key, error: e.message })));
+            void stream.writeSSE({ event: "progress", data: JSON.stringify({ done: totalDone, total: targets.length, written: totalWritten }) });
+            continue;
+          }
+          if (signal?.aborted) break;
+          const batch = raw;
+          const fresh = load();
+          const { written, errors } = applyContext(fresh, chunk2, batch.items ?? []);
+          appendLog(projectRoot, {
+            at: (/* @__PURE__ */ new Date()).toISOString(),
+            kind: "context",
+            summary: `Generated context for ${chunk2.length} key(s)`,
+            model: aiCfg.model,
+            system,
+            items: chunk2.map((t) => ({ id: t.id, key: t.key, source: t.source })),
+            results: (batch.items ?? []).map((r) => ({ id: r.id, value: r.context, error: r.error }))
+          });
+          persist(fresh);
+          totalWritten += written;
+          totalDone += chunk2.length;
+          allErrors.push(...errors);
+          void stream.writeSSE({ event: "progress", data: JSON.stringify({ done: totalDone, total: targets.length, written: totalWritten }) });
+        }
+      }
+      await Promise.all(Array.from({ length: Math.min(concurrency, chunks.length) }, worker));
+      if (signal?.aborted) return;
+      console.log(`[context] ${totalWritten} context(s) written${allErrors.length ? `, ${allErrors.length} error(s)` : ""}`);
+      await stream.writeSSE({ event: "done", data: JSON.stringify({ requested: targets.length, written: totalWritten, errors: allErrors }) });
     });
-    persist(s);
-    console.log(`[context] ${written} context(s) written${errors.length ? `, ${errors.length} error(s)` : ""}`);
-    return c.json({ requested: targets.length, written, errors });
   });
   app.onError(
     (err, c) => c.json({ error: err.message }, err instanceof GlotfileError ? 400 : 500)
@@ -5297,7 +5444,7 @@ async function runTranslate(args) {
       kind: "translate",
       summary: `Translated ${toTranslate.length} item(s)`,
       model: ai.model,
-      system: buildSystemPrompt(),
+      system: buildSystemPrompt(toTranslate.some((r) => r.plural !== void 0)),
       items: toTranslate.map((r) => ({
         id: r.id,
         key: r.key,
@@ -5420,10 +5567,32 @@ async function runBuildContext(args) {
     target.usageSnippets = extractSnippets(refs, projectRoot, fileCache);
   }
   const system = buildContextSystemPrompt();
-  const prompt = buildContextBatchPrompt(targets);
-  const raw = await provider.complete({ system, content: [{ type: "text", text: prompt }], schema: CONTEXT_BATCH_SCHEMA });
-  const batch = raw;
-  const { written, errors } = applyContext(state, targets, batch.items ?? []);
+  const aiCfg = loadLocalSettings(projectRoot).ai;
+  const batchSize = aiCfg.contextBatchSize ?? aiCfg.batchSize ?? 10;
+  const concurrency = aiCfg.contextConcurrency ?? aiCfg.concurrency ?? 3;
+  const chunks = [];
+  for (let i = 0; i < targets.length; i += batchSize) chunks.push(targets.slice(i, i + batchSize));
+  let written = 0;
+  const errors = [];
+  let next = 0;
+  async function worker() {
+    while (next < chunks.length) {
+      const chunk2 = chunks[next++];
+      let raw;
+      try {
+        raw = await provider.complete({ system, content: [{ type: "text", text: buildContextBatchPrompt(chunk2) }], schema: CONTEXT_BATCH_SCHEMA });
+      } catch (e2) {
+        errors.push(...chunk2.map((t) => ({ key: t.key, error: e2.message })));
+        continue;
+      }
+      const batch = raw;
+      const { written: w, errors: e } = applyContext(state, chunk2, batch.items ?? []);
+      written += w;
+      errors.push(...e);
+      console.log(`[${next * batchSize}/${targets.length}] wrote ${w}`);
+    }
+  }
+  await Promise.all(Array.from({ length: Math.min(concurrency, chunks.length) }, worker));
   saveState(args.statePath, state);
   console.log(`Wrote context for ${written} key(s).`);
   for (const e of errors) console.warn(`skip ${e.key}: ${e.error}`);