@yeaft/webchat-agent 0.1.774 → 0.1.776

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yeaft/webchat-agent",
-  "version": "0.1.774",
+  "version": "0.1.776",
   "description": "Remote agent for Yeaft WebChat — connects worker machines to the central server",
   "main": "index.js",
   "type": "module",

package/unify/engine.js

@@ -1230,8 +1230,8 @@ export class Engine {
     // `turnStartIdx` is where the current user message lives; the arc
     // we may collapse spans (arcStartIdx .. last assistant/tool).
     //
-    // Periodic-T1 fix: T1 must fire EVERY TOOL_BATCH_SIZE (13) tool
-    // calls, not just the first 13. So instead of a one-shot boolean,
+    // Periodic-T1 fix: T1 must fire EVERY TOOL_BATCH_SIZE (30) tool
+    // calls, not just the first batch. So instead of a one-shot boolean,
     // track:
     //   • `lastT1AtToolCount` — toolCount snapshot at the last T1
     //     ATTEMPT (success OR error). Trigger when
@@ -1797,6 +1797,7 @@ export class Engine {
               originalUserMsg: prompt,
               toolPairs: pairs,
               assistantText,
+              language: this.#config.language,
               signal,
             });
             // Detach: never await. The promise outlives this query() and
@@ -2034,9 +2035,9 @@ export class Engine {
       }
 
       // PR-L: T1 in-turn (synchronous) reflection. Fires once per
-      // adapter loop iteration where ≥ TOOL_BATCH_SIZE (13) tool
+      // adapter loop iteration where ≥ TOOL_BATCH_SIZE (30) tool
       // calls have accumulated since the last T1 firing — not just
-      // the first 13 of the query(). Generates a markdown reflection
+      // the first batch of the query(). Generates a markdown reflection
       // over the assistant+tool arc since the last T1 firing (or
       // since the user prompt for the first batch) and rewrites that
       // range to a SINGLE synthetic user message before the next
@@ -2044,10 +2045,10 @@ export class Engine {
       //
       // Loop semantics:
       //   - First batch: arcStartIdx = turnStartIdx + 1, fires when
-      //     queryToolCount reaches 13.
+      //     queryToolCount reaches TOOL_BATCH_SIZE.
       //   - Each subsequent batch: arcStartIdx is updated to the slot
       //     right after the just-inserted reflection message; fires
-      //     again whenever 13 more tools have run since
+      //     again whenever TOOL_BATCH_SIZE more tools have run since
       //     lastT1AtToolCount.
       //   - The dedup Set key includes `lastT1AtToolCount` so each
       //     batch within the same query gets a distinct entry — without
@@ -2084,6 +2085,7 @@ export class Engine {
             originalUserMsg: prompt,
             toolPairs: pairs,
             assistantText,
+            language: this.#config.language,
             signal,
           });
           const next = collapseRangeToReflection(
@@ -2129,8 +2131,8 @@ export class Engine {
           };
           // Advance lastT1AtToolCount past this batch so we don't
           // tight-loop on a hiccuping reflector. The next attempt is
-          // 13 tools from now, not immediately. arcStartIdx is left
-          // alone because history wasn't rewritten — the tail still
+          // TOOL_BATCH_SIZE tools from now, not immediately. arcStartIdx is
+          // left alone because history wasn't rewritten — the tail still
           // begins where it did. The trade-off: the next batch's
           // reflection will cover the tools that just failed too,
           // which is fine (they're still in conversationMessages).

package/unify/tool-folding/index.js

@@ -9,14 +9,25 @@
  *   - Duplicate-reminder text formatter
  *
  * The constants are NOT config-driven — V7 design freezes them in code.
+ *
+ * TOOL_BATCH_SIZE history: was 13 originally; raised to 30 (2026-05-15)
+ * after user feedback that 13 fired too often inside a single task and
+ * fragmented otherwise-coherent tool arcs into multiple reflections. 30
+ * keeps the periodic-reflection contract (it still fires every N tools,
+ * not just once) but gives a single task arc room to breathe before the
+ * arc gets collapsed.
  */
 
-export const TOOL_BATCH_SIZE = 13;
+export const TOOL_BATCH_SIZE = 30;
 export const TURN_SUMMARY_THRESHOLD = 5;
 export const DUP_TOOL_THRESHOLD = 3;
 
 export { ExecLog, buildEntry, argsHashOf } from './exec-log.js';
-export { buildReflectionPrompt, REFLECTION_TEMPLATE } from './reflection-prompt.js';
+export {
+  buildReflectionPrompt,
+  REFLECTION_TEMPLATE_EN,
+  REFLECTION_TEMPLATE_ZH,
+} from './reflection-prompt.js';
 export { runT1Reflection } from './t1-reflector.js';
 export { runT2Reflection } from './t2-reflector.js';
 export { buildFallbackStub } from './fallback-stub.js';

package/unify/tool-folding/reflection-prompt.js

@@ -5,9 +5,21 @@
  * The primary model is asked to REFLECT on a sequence of tool calls — not
  * just summarise. The output is markdown with five fixed sections so the
  * frontend ReflectionCard can render each independently.
+ *
+ * Bilingual: the prompt itself is rendered in the user's language (en/zh)
+ * so the model's output (which the user reads inside the ReflectionCard
+ * and which the next adapter loop sees as a synthetic user message) is
+ * in the matching language. Falls back to English for any other lang
+ * value or when `language` is omitted.
+ *
+ * Section headings stay English in both templates because the frontend
+ * ReflectionCard parses them by literal string match ("## What was
+ * attempted" et al.) and a localised heading would break rendering.
  */
 
-const TEMPLATE = `You are reviewing a sequence of {N} tool calls executed by an AI agent.
+import { isZhLanguage } from '../prompts.js';
+
+const TEMPLATE_EN = `You are reviewing a sequence of {N} tool calls executed by an AI agent.
 Your job is NOT just to summarize, but to REFLECT.
 
 Output as markdown with these exact sections:
@@ -39,19 +51,54 @@ User original request:
 Tool execution sequence:
 {toolCallsAndResults}`;
 
+const TEMPLATE_ZH = `你正在复盘一个 AI agent 刚刚执行的 {N} 次工具调用。
+你的任务不是简单总结，而是要"反思"。
+
+请用 markdown 输出，必须严格包含以下五个段落（标题保持英文，便于前端解析）：
+
+## What was attempted
+2-3 句话说明本批工具调用的目标和动作链路。
+
+## Key findings
+具体事实：路径、行号、ID、错误码（必须逐字保留原始值，不要改写）。
+
+## Direction check
+- 当前轨迹是否仍然贴合用户最初的请求？
+- 是否有偏离 / 范围漂移？
+- 是否有看起来重复或多余的工具调用？
+- 是否有迹象表明陷入了无效循环？
+
+## Suggested next direction
+下一轮 loop 应该聚焦什么？应该避免什么？
+
+## Tool execution log
+紧凑列表：<tool_name> × <count>（附上值得注意的参数）。
+
+注意：所有标识符、路径、URL、行号、错误消息必须逐字保留，不允许改写数据值。
+
+用户最初的请求：
+{originalUserMessage}
+
+工具执行序列：
+{toolCallsAndResults}`;
+
 /**
  * Render the prompt.
  *
- * @param {{ originalUserMsg: string, toolPairs: Array<{ name: string, input: any, output: string, isError: boolean }>, assistantText?: string }} p
+ * @param {{ originalUserMsg: string, toolPairs: Array<{ name: string, input: any, output: string, isError: boolean }>, assistantText?: string, language?: string }} p
  * @returns {string}
  */
-export function buildReflectionPrompt({ originalUserMsg, toolPairs, assistantText }) {
+export function buildReflectionPrompt({ originalUserMsg, toolPairs, assistantText, language }) {
   const N = toolPairs.length;
   const seq = toolPairs.map((p, i) => formatPair(i + 1, p)).join('\n\n');
+  const isZh = isZhLanguage(language);
   const head = assistantText && assistantText.trim()
-    ? `Assistant text emitted during this batch:\n${assistantText.trim()}\n\n`
+    ? (isZh
+      ? `本批工具调用期间助手输出的文本：\n${assistantText.trim()}\n\n`
+      : `Assistant text emitted during this batch:\n${assistantText.trim()}\n\n`)
     : '';
-  return TEMPLATE
+  const template = isZh ? TEMPLATE_ZH : TEMPLATE_EN;
+  return template
     .replace('{N}', String(N))
     .replace('{originalUserMessage}', String(originalUserMsg || '').slice(0, 4000))
     .replace('{toolCallsAndResults}', head + seq);
@@ -71,4 +118,8 @@ function formatPair(idx, p) {
   return `[${idx}] ${p.name}${status}\n  args: ${inputStr}\n  result: ${outputStr}`;
 }
 
-export const REFLECTION_TEMPLATE = TEMPLATE;
+// Section headings stay English in both templates because ReflectionCard
+// parses them by literal string match. Exported separately so a debug panel
+// or test can compare against either dictionary directly.
+export const REFLECTION_TEMPLATE_EN = TEMPLATE_EN;
+export const REFLECTION_TEMPLATE_ZH = TEMPLATE_ZH;

package/unify/tool-folding/t1-reflector.js

@@ -1,13 +1,17 @@
 /**
  * t1-reflector.js — V7 in-turn (synchronous) reflection (PR-L).
  *
- * Triggered when the current turn has accumulated TOOL_BATCH_SIZE (13) tool
+ * Triggered when the current turn has accumulated TOOL_BATCH_SIZE (30) tool
  * results and the engine is about to loop back into adapter.stream(). Calls
  * the PRIMARY model — never the fast model — to generate a markdown
  * reflection over the batch.
  *
  * On success: returns { content, durationMs }.
  * On failure: throws (engine catches and leaves history unchanged).
+ *
+ * The `language` param is forwarded to buildReflectionPrompt so the prompt
+ * (and therefore the model's response) renders in the user's language. The
+ * engine reads this from config.language and passes it through.
  */
 
 import { buildReflectionPrompt } from './reflection-prompt.js';
@@ -19,13 +23,14 @@ import { buildReflectionPrompt } from './reflection-prompt.js';
  *   originalUserMsg: string,
  *   toolPairs: Array<{ name: string, input: any, output: string, isError: boolean }>,
  *   assistantText?: string,
+ *   language?: string,
  *   signal?: AbortSignal,
  * }} p
  * @returns {Promise<{ content: string, durationMs: number }>}
  */
-export async function runT1Reflection({ adapter, model, originalUserMsg, toolPairs, assistantText, signal }) {
+export async function runT1Reflection({ adapter, model, originalUserMsg, toolPairs, assistantText, language, signal }) {
   const t0 = Date.now();
-  const prompt = buildReflectionPrompt({ originalUserMsg, toolPairs, assistantText });
+  const prompt = buildReflectionPrompt({ originalUserMsg, toolPairs, assistantText, language });
   const result = await adapter.call({
     model,
     system: prompt,

package/unify/tool-folding/t2-reflector.js

@@ -17,13 +17,14 @@ import { buildReflectionPrompt } from './reflection-prompt.js';
  *   originalUserMsg: string,
  *   toolPairs: Array<{ name: string, input: any, output: string, isError: boolean }>,
  *   assistantText?: string,
+ *   language?: string,
  *   signal?: AbortSignal,
  * }} p
  * @returns {Promise<{ content: string, durationMs: number }>}
  */
-export async function runT2Reflection({ adapter, model, originalUserMsg, toolPairs, assistantText, signal }) {
+export async function runT2Reflection({ adapter, model, originalUserMsg, toolPairs, assistantText, language, signal }) {
   const t0 = Date.now();
-  const prompt = buildReflectionPrompt({ originalUserMsg, toolPairs, assistantText });
+  const prompt = buildReflectionPrompt({ originalUserMsg, toolPairs, assistantText, language });
   const result = await adapter.call({
     model,
     system: prompt,