@yeaft/webchat-agent 1.0.242 → 1.0.243

package/local-runtime/version.json

@@ -1 +1 @@
-{"version":"1.0.242"}
+{"version":"1.0.243"}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yeaft/webchat-agent",
-  "version": "1.0.242",
+  "version": "1.0.243",
   "description": "Remote worker agent for Yeaft Web Code Agent — connects the native Yeaft engine, CLI providers, and workbench tools",
   "main": "index.js",
   "type": "module",

package/yeaft/templates/plan-instruction.md

@@ -17,7 +17,7 @@ You have just entered **planning mode** for the topic below. Your job is to thin
 
 1. Write a short prose plan: problem, approach, risks.
 2. Call `TodoWrite` with the ordered steps. Mark exactly one item as `in_progress`.
-3. Start executing that first step immediately in the same turn.
+3. If the first-step tools and arguments are already known, emit `TodoWrite` and those independent tool calls in the same assistant response. Otherwise wait only for a result that genuinely determines the next action.
 
 If the first step is to ask the user a blocking question, ask it and stop. Otherwise keep moving.
 
@@ -40,6 +40,6 @@ If the first step is to ask the user a blocking question, ask it and stop. Other
 
 1. 写一段简短计划：问题、方案、风险。
 2. 调用 `TodoWrite` 写入有序步骤，并且只能把一个条目标记为 `in_progress`。
-3. 在同一轮立即开始执行第一步。
+3. 如果第一步所需的工具和参数已经确定，应在同一个 assistant response 中发出 `TodoWrite` 和这些彼此独立的工具调用；否则只在确实需要某个结果来决定下一动作时等待。
 
 如果第一步是向用户询问阻塞问题，那就提问并停下。否则继续推进。

package/yeaft/templates/tool-guidance.md

@@ -61,6 +61,14 @@ How to use it:
 - `content` is the imperative form (e.g. "Run tests"); `activeForm` is
   the present-continuous form shown during execution (e.g. "Running
   tests").
+- Avoid an intermediate `TodoWrite`-only model round when the next work tool
+  and its arguments are already known. Emit `TodoWrite` and those independent
+  work tool calls in the same assistant response.
+- This is batching, not speculative progress: mark work completed only after
+  evidence, and keep calls separate when a pending result can change the next
+  action, its arguments, or its safety. A standalone `TodoWrite` remains valid
+  when no work tool should follow, including final completion or a blocking
+  user question.
 
 Do **not** use TodoWrite for single trivial edits, single command runs,
 or pure conversational/question turns — the checklist becomes noise.
@@ -125,6 +133,11 @@ or pure conversational/question turns — the checklist becomes noise.
 - 任何时刻最多只能有 **一个** `"in_progress"`。
 - `content` 是命令式（如 "Run tests"）；`activeForm` 是执行中展示的进行
   时（如 "Running tests"）。
+- 如果下一项工作所用的工具和参数已经确定，不要让中间状态的 `TodoWrite` 单独占一个
+  模型回合；应在同一个 assistant response 中发出 `TodoWrite` 和这些彼此独立的工作工具调用。
+- 这是合批，不是提前宣告进度：只有已有证据时才能把工作标记为完成。如果待返回结果可能改变
+  下一动作、参数或安全性，就必须分开调用。没有工作工具应继续执行时（包括记录最终完成态或
+  询问阻塞问题），`TodoWrite` 仍可单独调用。
 
 **不要**为单条琐碎修改、单次命令执行、纯对话/问题使用 TodoWrite——清单
 反而成了噪音。

package/yeaft/tools/start-plan.js

@@ -46,7 +46,7 @@ export default defineTool({
 This tool returns a planning instruction. Use it to land a structured plan, then keep working in the same turn. The expected flow is:
 1. Produce a short prose plan (problem, approach, risks).
 2. Call \`TodoWrite\` with the ordered steps. Mark the first concrete step "in_progress", the rest "pending".
-3. Start executing that first step — call whatever tools the work needs. Do not end the turn just because the plan is written; the plan is the runway.
+3. When the first-step tools and arguments are already known, emit \`TodoWrite\` and those independent tool calls in the same assistant response. Wait only when a result genuinely determines the next action.
 
 WHEN TO USE:
 - Multi-step implementation (3+ steps), refactor, or open-ended investigation.
@@ -65,7 +65,7 @@ The tool takes the topic plus optional guiding fields (stuckAt, userProblem, exp
 此工具返回规划指令。用它产出一份结构化计划，然后在同一个 turn 中继续工作。预期流程是：
 1. 产出简短文字计划（问题、方法、风险）。
 2. 调用 TodoWrite 写出有序步骤。将第一个具体步骤标记为 "in_progress"，其余标记为 "pending"。
-3. 开始执行第一个步骤——调用工作所需的任何工具。不要因为计划写完了就结束 turn；计划是跑道。
+3. 如果第一步所需的工具和参数已经确定，应在同一个 assistant response 中发出 TodoWrite 和这些彼此独立的工具调用；只有某个结果确实决定下一动作时才等待。
 
 何时使用：
 - 多步骤实现（3+ 步）、重构或开放式调查。
@@ -172,7 +172,10 @@ The tool takes the topic plus optional guiding fields (stuckAt, userProblem, exp
       lines.push('</guiding-context>');
     }
     lines.push('');
-    lines.push('Next: produce the plan, call `TodoWrite` to land the ordered steps, then keep going — start executing the first step in this same turn. Only stop after the plan if the first step is "ask the user".');
+    const nextInstruction = String(language).toLowerCase().startsWith('zh')
+      ? '下一步：产出计划并调用 `TodoWrite`。如果第一步的工具和参数已经确定，在同一个 assistant response 中一并发出这些彼此独立的工具调用。只有第一步必须询问用户时才在计划后停下。'
+      : 'Next: produce the plan and call `TodoWrite`. If the first-step tools and arguments are already known, emit those independent tool calls in the same assistant response. Stop after the plan only when the first step must ask the user.';
+    lines.push(nextInstruction);
 
     return lines.join('\n');
   },

package/yeaft/tools/todo-write.js

@@ -39,6 +39,11 @@ HOW TO USE:
 - AT MOST one item may be "in_progress" at any time.
 - \`content\` is the imperative form ("Run tests"); \`activeForm\` is the present-continuous shown during execution ("Running tests").
 
+BATCH WITH WORK:
+- Avoid an intermediate TodoWrite-only model round when the next work tool and its arguments are already known. Emit TodoWrite and those independent work tool calls in the same assistant response.
+- This is batching, not speculative progress: mark work completed only after evidence. Keep calls separate when a pending result can change the next action, its arguments, or its safety.
+- A standalone TodoWrite remains valid when no work tool should follow, including final completion or a blocking user question.
+
 WHEN NOT TO USE:
 - Single trivial change, single command run, pure conversation/question.`,
     zh: `用 checklist 跟踪多步骤任务进度，用户可以实时看到勾选。
@@ -54,6 +59,11 @@ WHEN NOT TO USE:
 - 任何时候最多只能有一个 "in_progress"。
 - content 是祈使形式（如 "Run tests"）；activeForm 是执行时显示的进行时态（如 "Running tests"）。
 
+和工作工具合批：
+- 如果下一项工作所用的工具和参数已经确定，不要让中间状态的 TodoWrite 单独占一个模型回合；应在同一个 assistant response 中发出 TodoWrite 和这些彼此独立的工作工具调用。
+- 这是合批，不是提前宣告进度：只有已有证据时才能把工作标记为完成。如果待返回结果可能改变下一动作、参数或安全性，就必须分开调用。
+- 没有工作工具应继续执行时（包括记录最终完成态或询问阻塞问题），TodoWrite 仍可单独调用。
+
 何时不使用：
 - 单个琐碎改动、单个命令执行、纯对话/提问。`
   },