@ai-setting/roy-task-plugin 0.7.8 → 0.7.10

package/dist/cli.js

@@ -409,7 +409,12 @@ class NdjsonAssembler {
       this.toolCalls += 1;
       const name = parsed.part?.tool ?? parsed.tool;
       if (typeof name === "string" && name) {
-        this.tools.push(name);
+        const partId = parsed.part?.id;
+        this.tools.push({
+          name,
+          timestamp: typeof parsed.timestamp === "number" ? new Date(parsed.timestamp).toISOString() : undefined,
+          id: typeof partId === "string" && partId ? partId : undefined
+        });
       }
       if (this.createdTask === undefined && typeof name === "string") {
         const part = parsed.part;
@@ -445,110 +450,31 @@ class NdjsonAssembler {
 }
 
 // ../agent/src/roy-prompt-core.ts
+import {
+  TASK_DISCIPLINE_COMPLETE,
+  TASK_DISCIPLINE_GOALS,
+  TASK_DISCIPLINE_LIFECYCLE,
+  TASK_DISCIPLINE_OPERATION
+} from "@ai-setting/roy-agent-core/env/task/prompt";
 var ROY_PROMPT_COMMON = `# Roy — task-first-class-citizen agent
 
 You are **Roy**, an agent whose defining rule is simple:
 
 > **Task is a first-class citizen. No task, no work.**
 
-You are not a chat assistant that occasionally writes notes. You are a worker
-whose output is a *visible, trackable* stream of tasks. Every non-trivial
-request (more than one tool call / more than a couple of minutes of work) MUST
-become a tracked task before you start doing it, and MUST be closed out
-explicitly when you finish.
-
-## The lifecycle (non-negotiable)
-
-### 1. Orient — \`task_tree(depth=1)\` (or \`depth=2\` for bigger trees)
-Before creating anything, look at the existing tree. Decide:
-- does this work belong to an existing root (attach via \`parent_task_id\`)?
-- or is it genuinely new?
-
-Keep **root tasks ≤ 10**. Prefer attaching to an existing thematic root over
-spawning a new root. Never hardcode ids — read them from \`task_tree\`.
-
-### 2. Create — \`task_create\`
-Create the task **before** doing the work. Pass:
-- \`title\` (imperative, specific), \`description\`, \`goals_and_expected_deliverables\`
-- \`priority\` (low / medium / high), \`type\` (normal / cycle / longterm)
-- \`project_path\`, \`context\` (JSON string: worktree_path / branch / type …)
-- \`parent_task_id\` when the tree said so
-
-#### ⚠️ Goals must be numbered sub-items
-\`goals_and_expected_deliverables\` **MUST** be written as **numbered
-sub-items** (编号子项) that can each be verified **one by one**
-(逐条验收) — for \`type=normal\` tasks:
-
-  1. 实现 X（含文件路径）
-  2. 测试 Y 覆盖 Z 场景（pass/fail 计数）
-  3. 文档 W 落盘到 <path>
-
-Why: the completion hook judges the goal **item by item** against the
-operation-record chain, using each sub-item as fine-grained evidence. A
-vague one-line goal (一句话模糊目标，如「做好 X」「完成调研」) gives the
-judge nothing to check and reliably triggers a completion hint.
-
-Reference the four atomic assertion forms when writing each sub-item:
-- **文件断言** (file assertion) — file exists / size / content grep
-- **进程断言** (process assertion) — command exit code / stdout match
-- **测试断言** (test assertion) — pass/fail counts, no NEW fail
-- **状态断言** (state assertion) — version number / HTTP 200 / status field
-
-For \`type=longterm\` / \`type=cycle\` tasks there is no per-item completion
-acceptance — for those, split goals into staged, individually checkable
-sub-items instead of one vague sentence.
+You are a worker agent: you act through the tools, and your output is what
+the user sees.
+
+${TASK_DISCIPLINE_LIFECYCLE}
+
+${TASK_DISCIPLINE_GOALS}
 
 The tool returns a **numeric task id** — that number is your handle for the
 rest of the conversation.
 
-### 3. Track — \`task_update\`
-Report progress as you go, not only at the end:
-- \`status\`: todo → active → (paused / cancelled) → completed
-- \`progress\`: 0–100
-- \`current_status\`: one human sentence about where you are right now
-
-If the scope changes, \`task_update\` the description/goals instead of silently
-doing something else.
-
-### 4. Record milestones — \`task_operation_create\` — **as you work**
-Call this tool the moment a stage is reached — **not after the fact**,
-not in a batch at the end. Finished a chunk of code? A test just went
-green? Hit a blocker? Made a decision? Record it **immediately**, then
-keep working. The record is the running log of the work, written while
-it happens.
-
-This tool has **two orthogonal dimensions** and you must keep them distinct:
-- **state / milestone dimension** — \`milestone_type\` + \`milestone_title\` +
-  \`milestone_description\`: *what* was reached (e.g. "GREEN: all tests pass").
-- **process dimension** — \`process_description\`: *how* you got there —
-  the operations just performed and why, written in the moment,
-  **100–500 characters**. Too short = rejected (no value);
-  too long = rejected (that is a log, not a summary).
-
-Rules: do **not** use the process field to restate the milestone, and do not
-use milestones as a substitute for progress. Milestones mark *state
-transitions*: RED→GREEN, blocker found, decision made, review passed.
-
-### 5. Close — \`task_complete\`
-Before completing:
-1. \`task_get(task_id)\` and re-read goal + delivery.
-2. **Accept the goal item-by-item** (按子项逐条验收): check **each** numbered
-   sub-item against the operation-record chain and the deliverable evidence.
-   If a sub-item is not met, either extend the task (\`task_update\`) or say
-   so honestly.
-3. Call \`task_complete\` with \`deliverable_storage_path\` — a locator for the
-   work product: a file path, a git branch/worktree, a doc URL, a commit URL.
-   The path is persisted into the task context and as an operation record.
-
-#### ⚠️ The completion hint is a verify duty, not a notification
-If the task context carries a \`[TASK_COMPLETION_HINT]\`, you **MUST** verify
-it (核实义务) — never ignore it (不允许无视). Re-read the task via
-\`task_get\`, then do **exactly one** of:
-- (a) the goal is **not** met → \`task_update(task_id=<id>, status=todo)\` to
-  reopen and continue the work; or
-- (b) the goal **is** met → \`task_operation_create\`
-  (milestone_type=decision) stating why the sub-items are actually met
-  (说明理由 — "state why").
+${TASK_DISCIPLINE_OPERATION}
+
+${TASK_DISCIPLINE_COMPLETE}
 
 #### Several deliverables in one field
 \`deliverable_storage_path\` may name **multiple paths**, space-separated
@@ -685,6 +611,7 @@ class CodexStreamDecoder {
   lines = new LineSplitter;
   state = { tools: [], errors: [], turnCompleted: false };
   mcpSeen = new Set;
+  cmdSeen = new Set;
   pushChunk(chunk) {
     for (const line of this.lines.pushChunk(chunk))
       this.consume(line);
@@ -742,8 +669,14 @@ class CodexStreamDecoder {
     }
     if (item.type === CODEX_ITEM_TYPES.COMMAND_EXECUTION) {
       const id = item.id;
-      if (typeof id === "string" && id && this.state.tools.includes(id)) {} else {
-        this.state.tools.push(typeof id === "string" ? id : `cmd-${this.state.tools.length}`);
+      if (typeof id === "string" && id && this.cmdSeen.has(id)) {} else {
+        if (typeof id === "string" && id)
+          this.cmdSeen.add(id);
+        this.state.tools.push({
+          name: typeof id === "string" ? id : `cmd-${this.state.tools.length}`,
+          timestamp: new Date().toISOString(),
+          id: typeof id === "string" ? id : undefined
+        });
       }
       if (completed && this.state.createdTask === undefined) {
         this.state.createdTask = extractCreatedTask(item.aggregated_output);
@@ -757,7 +690,11 @@ class CodexStreamDecoder {
         this.mcpSeen.add(id);
       if (!seen) {
         const name = typeof item.tool === "string" && item.tool ? item.tool : undefined;
-        this.state.tools.push(name ?? (typeof id === "string" && id ? id : `mcp-${this.state.tools.length}`));
+        this.state.tools.push({
+          name: name ?? (typeof id === "string" && id ? id : `mcp-${this.state.tools.length}`),
+          timestamp: new Date().toISOString(),
+          id: typeof id === "string" && id ? id : undefined
+        });
       }
       if (completed && this.state.createdTask === undefined && isSuccessfulMcpStatus(item.status) && typeof item.tool === "string" && isCreationTool(item.tool) && isRoyTaskServer(item.server)) {
         const text = mcpResultText(item.result);
@@ -939,7 +876,11 @@ class ClaudeStreamDecoder {
         this.state.assistantText.push(b.text);
       } else if (b.type === "tool_use") {
         if (typeof b.name === "string") {
-          this.state.tools.push(b.name);
+          this.state.tools.push({
+            name: b.name,
+            timestamp: new Date().toISOString(),
+            id: typeof b.id === "string" && b.id ? b.id : undefined
+          });
           if (isTaskCreateTool(b.name) && typeof b.id === "string") {
             const input = b.input;
             const title = input?.["title"];
@@ -1141,6 +1082,18 @@ function errMsg(err) {
   return String(err);
 }
 
+// ../../../packages/cli/src/daemon/discipline.ts
+import { TASK_DISCIPLINE_PROMPT } from "@ai-setting/roy-agent-core/env/task/prompt";
+var RUNTIMES_WITH_AGENT_SYSTEM_PROMPT = Object.freeze([
+  "opencode",
+  "claude"
+]);
+function needsDispatchInjection(runtime) {
+  if (!runtime)
+    return false;
+  return !RUNTIMES_WITH_AGENT_SYSTEM_PROMPT.includes(runtime);
+}
+
 // ../../../packages/cli/src/daemon/runners/codex.ts
 var DEFAULT_CODEX_TIMEOUT_MS = 10 * 60 * 1000;
 var CODEX_STDIO = Object.freeze([
@@ -1195,9 +1148,21 @@ function buildCodexArgs(input) {
   args.push("--skip-git-repo-check");
   args.push("-s", input.sandbox ?? "read-only");
   args.push(...input.mcpConfigArgs ?? []);
-  args.push(input.prompt);
+  args.push(withDiscipline(input.prompt, input.systemPrompt));
   return args;
 }
+function withDiscipline(prompt, systemPrompt) {
+  const contract = systemPrompt === undefined ? TASK_DISCIPLINE_PROMPT : systemPrompt;
+  if (!contract)
+    return prompt;
+  if (prompt.includes(contract))
+    return prompt;
+  return `${contract}
+
+---
+
+${prompt}`;
+}
 function buildCodexResumeArgs(sessionId, input) {
   const args = ["exec", "resume", "--json"];
   if (input.model)
@@ -1208,7 +1173,7 @@ function buildCodexResumeArgs(sessionId, input) {
   }
   args.push(...input.mcpConfigArgs ?? []);
   args.push(sessionId);
-  args.push(input.prompt);
+  args.push(withDiscipline(input.prompt, input.systemPrompt));
   return args;
 }
 function makeCodexRunner(opts = {}) {
@@ -1724,49 +1689,12 @@ function buildDispatchPrompt(input) {
   const eventId = event?.event_id ?? "(unknown)";
   const parts = [
     "你收到一条飞书（Lark）消息，请处理它。",
-    "",
-    "## 硬性规则",
-    "",
-    "1. **不要**自己调用 `lark-cli` 发送回复。守护进程会在你结束后，把你的",
-    "   **最终回答文本**用 `lark-cli im +messages-reply` 一次性发回飞书。",
-    "   你自己发会造成重复消息。",
-    "2. 把**最终回答**写成给发件人看的**普通散文 / Markdown**。",
-    "   不要输出 JSON 包裹、不要输出“我调用了某某工具”的过程流水账。",
-    "   这段文本会**原样**成为飞书里的回复正文。",
-    "3. 需要落任务就用 `task_*` 工具（task_create / task_update / …），",
-    "   这是你的主要职责；不要只在回答里“描述”你打算做什么。",
-    "",
-    "## 什么时候该建任务（Task #4433）",
-    "",
-    "任务由**你**决定建不建，守护进程不会替你建。判断标准：",
-    "",
-    "- **该建**：需要多步操作、要动文件/跑命令、要调研并产出结论、",
-    "  耗时预计超过一两分钟、或用户明确要求「跟踪 / 建个任务 / 后续跟进」。",
-    "  → 用 `task_create` 建任务，**然后**再开始干活（先立任务，再动手）。",
-    "- **不必建**：寒暄、谢谢、纯问答（「今天几号」「这个参数叫啥」）、",
-    "  一两句话就能答完且不需要任何工具的问题。",
-    "  → 直接回答，不要为了「显得勤快」而建一个空任务。",
-    "- **建完要更新**：动手过程中用 `task_update` 推进度 / 改 current_status，",
-    "  干完用 `task_complete` 收尾。建了任务却从不更新，比不建更糟——",
-    "  看板上会留下一个卡住的任务。",
-    "",
-    "⚠️ **卡片与任务一一对应**：只有你建了任务，用户才会收到一张进展卡片；",
-    "没建任务就只有文字回复。所以**不要**为了「想要一张卡片」而去建任务，",
-    "也不要建了任务又不用它——卡片上显示的就是你建的那个任务的真 id 和标题。",
-    "",
-    "## 环境事实（必须使用下面这些值，不要自己猜）",
-    "",
-    `- lark-cli profile：\`${profile}\`（飞书 openId 是 per-app 隔离的，换 profile 会失败）`,
-    `- chat_id：\`${chatId}\``,
-    `- message_id：\`${messageId}\``,
-    `- event_id：\`${eventId}\``,
-    "",
-    "## 原始事件（完整 JSON）",
-    "",
-    "```json",
-    JSON.stringify(input.rawEvent, null, 2),
-    "```"
+    ""
   ];
+  if (needsDispatchInjection(input.runtime)) {
+    parts.push("## 任务纪律（Task #4889 — 来自 roy-agent core 单一真相源）", "", TASK_DISCIPLINE_PROMPT, "");
+  }
+  parts.push("## 硬性规则", "", "1. **不要**自己调用 `lark-cli` 发送回复。守护进程会在你结束后，把你的", "   **最终回答文本**用 `lark-cli im +messages-reply` 一次性发回飞书。", "   你自己发会造成重复消息。", "2. 把**最终回答**写成给发件人看的**普通散文 / Markdown**。", "   不要输出 JSON 包裹、不要输出“我调用了某某工具”的过程流水账。", "   这段文本会**原样**成为飞书里的回复正文。", "3. 需要落任务就用 `task_*` 工具（task_create / task_update / …），", "   这是你的主要职责；不要只在回答里“描述”你打算做什么。", "", "## 什么时候该建任务（Task #4433）", "", "任务由**你**决定建不建，守护进程不会替你建。判断标准：", "", "- **该建**：需要多步操作、要动文件/跑命令、要调研并产出结论、", "  耗时预计超过一两分钟、或用户明确要求「跟踪 / 建个任务 / 后续跟进」。", "  → 用 `task_create` 建任务，**然后**再开始干活（先立任务，再动手）。", "- **不必建**：寒暄、谢谢、纯问答（「今天几号」「这个参数叫啥」）、", "  一两句话就能答完且不需要任何工具的问题。", "  → 直接回答，不要为了「显得勤快」而建一个空任务。", "- **建完要更新**：动手过程中用 `task_update` 推进度 / 改 current_status，", "  干完用 `task_complete` 收尾。建了任务却从不更新，比不建更糟——", "  看板上会留下一个卡住的任务。", "", "⚠️ **卡片与任务一一对应**：只有你建了任务，用户才会收到一张进展卡片；", "没建任务就只有文字回复。所以**不要**为了「想要一张卡片」而去建任务，", "也不要建了任务又不用它——卡片上显示的就是你建的那个任务的真 id 和标题。", "", "## 环境事实（必须使用下面这些值，不要自己猜）", "", `- lark-cli profile：\`${profile}\`（飞书 openId 是 per-app 隔离的，换 profile 会失败）`, `- chat_id：\`${chatId}\``, `- message_id：\`${messageId}\``, `- event_id：\`${eventId}\``, "", "## 原始事件（完整 JSON）", "", "```json", JSON.stringify(input.rawEvent, null, 2), "```");
   if (input.recommendedAction) {
     parts.push("", "## 事件源附带的回复契约（背景资料，仅供你理解上下文）", "", "> ⚠️ 下面这段通常要求“由 agent 自己调 lark-cli 回复”。本次**不适用**：", "> 回复由守护进程负责。但其中关于内容编排的规范（单条、结构化、", "> 30KB 上限等）仍然值得遵守——你写的内容会被塞进那一条 post 里。", "", input.recommendedAction);
   }
@@ -2001,11 +1929,17 @@ var TOOL_LABELS = {
   task_complete: "完成任务",
   task_operation_create: "记录进展"
 };
-function describeToolUse(tool) {
+var toolTraceSeq = 0;
+function describeToolUse(tool, seq, scope) {
+  const use = typeof tool === "string" ? { name: tool } : tool;
+  const name = use.name;
+  const localId = use.id ?? `t-${seq ?? toolTraceSeq++}`;
   return {
-    title: TOOL_LABELS[normaliseMcpToolName(tool)] ?? tool,
-    timestamp: new Date().toISOString(),
-    category: tool
+    title: TOOL_LABELS[normaliseMcpToolName(name)] ?? name,
+    timestamp: use.timestamp ?? new Date().toISOString(),
+    category: name,
+    source: "tool-trace",
+    id: scope !== undefined ? `${scope}#${localId}` : localId
   };
 }
 function normaliseMcpToolName(tool) {
@@ -2210,7 +2144,8 @@ class Dispatcher {
           chatId,
           profile,
           tools: result.tools,
-          status: result.error ? "failed" : "running"
+          status: result.error ? "failed" : "running",
+          scope: messageId
         });
       }
       if (result.error) {
@@ -2346,7 +2281,7 @@ class Dispatcher {
         profile: input.profile,
         status: input.status,
         ...handle.meta ? { meta: handle.meta } : {},
-        processSteps: tools.map(describeToolUse)
+        processSteps: tools.map((t, i) => describeToolUse(t, i, input.scope))
       });
     } catch (err) {
       handle.error ??= errorMessage(err);
@@ -2377,7 +2312,8 @@ class Dispatcher {
       prompt: buildDispatchPrompt({
         rawEvent: input.raw,
         profile: input.profile,
-        recommendedAction: input.recommendedAction
+        recommendedAction: input.recommendedAction,
+        runtime: this.opts.runtime
       }),
       chatId: input.chatId,
       messageId: input.messageId,
@@ -2640,7 +2576,11 @@ class CursorStreamDecoder {
     const subtype = ev["subtype"];
     if (subtype === "started") {
       if (tool) {
-        this.state.tools.push(tool);
+        this.state.tools.push({
+          name: tool,
+          timestamp: new Date().toISOString(),
+          id: callId ?? undefined
+        });
         if (callId && isTaskCreateTool(tool)) {
           this.state.pendingCreates.set(callId, {
             tool,
@@ -3220,7 +3160,8 @@ async function startServe(options, deps) {
   }, {
     dryRun: options.dryRun,
     replyEnabled: options.replyEnabled,
-    agent: options.agent
+    agent: options.agent,
+    runtime: options.runtime
   });
   const listener = (deps.createListener ?? createEventListener)({
     profile: options.profile,
@@ -3465,7 +3406,7 @@ function formatStatus(report) {
 }
 
 // ../../../packages/cli/src/main.ts
-var VERSION = "0.7.8";
+var VERSION = "0.7.9";
 async function main(argv) {
   const parsed = parseArgs(argv);
   switch (parsed.command) {