fix(web): treat duplicate rectification tool calls as idempotent

Aborting the turn on a second identical public tool-call failed staging
after evidence and compare had already succeeded. Skip the duplicate
receipt instead and let maxSteps bound real loops.

Co-authored-by: Cursor <cursoragent@cursor.com>

docs/BUG_HISTORY.md

@@ -5531,6 +5531,22 @@
 - 复发自：无
 - 修复版本：未修复
 
+## BUG-372 | 生时纠正同一轮重复工具调用把已成功写入打成失败
+
+- 状态：resolved
+- 首次发现：2026-08-25
+- 最近更新：2026-08-25
+- 影响面：`POST /api/rectification/agent`、`runV9AgentTurn` 观察器、公开 NDJSON
+- 用户现象：同一轮已读盘、写入证据、跑完诊断和候选比较后，流以 `run.failed` / `repeated_tool_call` 结束，文案「本轮没有完成，状态已记录。」，没有助手叙述。公开事件里每个工具只出现一次 started/completed。
+- 触发条件：`action=message` 的自由文本经历回合；模型在 `rectification-compare-candidates`（或同类只含 caseId 的公开工具）完成后，又发出一次相同工具名与相同参数的 `tool-call`。
+- 根因：BUG-368 P0-3 把「相同工具参数」做成观察器硬上限 1，第二次相同 `toolName + args` 在发布 `tool.activity` started 之前抛错。该码不在自动重试集合，`recoverable=false`。证据与比较已经落库，只是本轮被标失败且跳过 dossier 叙述器。`rectification-set-focus` 已从 Agent 工具列表移除，这条防线不再对应真实循环。
+- 修复：相同公开工具调用视为幂等，跳过重复的 started/phase 收据，不 abort、不 `attempt.reset`。真循环仍由 `maxSteps` 与超时约束。无模型正文时走既有服务器叙述。不改已哈希 Skill `10.0.11`。
+- 验证：`frontend/tests/rectification-v9-agent.test.ts` 锁定重复 read-case 完成叙述且无 `run.failed`；`frontend/tests/rectification-v9-stream.test.ts` 锁定重复 set-focus 继续出回答、诊断后重复 compare 走叙述且无 `attempt.reset`。
+- 防复发：禁止对第二次相同公开 `tool-call` 抛 `repeated_tool_call` 或失败整轮。禁止把幂等工具调用放进自动重试。禁止用观察器 abort 代替 `maxSteps` / timeout。
+- 相关记录：BUG-368、BUG-367
+- 复发自：BUG-368（去掉模型驱动 set-focus 后，仍用 abort-on-second-identical-call 防循环，误杀正常比较重发）
+- 修复版本：待发布
+
 ## BUG-371 | staging publish 的 next build 找不到 createServerSupabaseClient
 
 - 状态：resolved
@@ -5588,10 +5604,10 @@
 - 用户现象：用户说「2020年4月开始实习、6月转正、10月离职」后，界面先刷出残缺英文（`Let me`、`_probe`、`_gain`）。`rectification-set-focus` 连续失败后整轮重跑，已成功的证据再提交一次，只落地实习、另外两条被引文拒。随后聊天里出现一条用户从未输入的「2002 年发生什么了」。
 - 触发条件：自由文本经历回合；Agent 在调用工具前输出规划文本；`set-focus` 因重复探针或零信息增益被拒；失败运行的推荐问题被写进历史。
 - 根因：三组独立回归。(1) 禁止 `thinking.delta` 后，把每个 step 的 `text-delta` 都公开成 `answer.delta`，工具前规划变成用户正文；再对碎片做英文过滤，句子被剪成残片。(2) `compare-candidates` 已算出下一问，仍让模型自己调 `set-focus`；确定性校验失败后 `attempt.reset` 重放已成功的 Evidence 写入。(3) 未完成运行的 suggestion / 角色映射把「2002 年发生什么了」写成 user 消息。2002 不是引擎从 2020 算出来的。
-- 修复：按 step 缓冲，只发布无工具且 `stop`/`length` 的终端文本；`reasoning-delta` 不下发。Agent 工具列表去掉 `set-focus`，由 compare/read-case 在服务端持久化 `open_question`。`duplicate_focus` 等域错误不整轮重试，相同工具参数最多一次。用户消息记录 `origin`/`clientActionId`/`content_hash`。Evidence quote 用源消息 offset，按条返回 created/already_exists/quote_mismatch。推荐问题只在 `run.completed` 后解析，纠正 UI 仍无 suggestion chip。不改已哈希 Skill `10.0.11`。
+- 修复：按 step 缓冲，只发布无工具且 `stop`/`length` 的终端文本；`reasoning-delta` 不下发。Agent 工具列表去掉 `set-focus`，由 compare/read-case 在服务端持久化 `open_question`。`duplicate_focus` 等域错误不整轮重试。同一流里相同公开工具参数视为幂等，不得 abort 整轮；循环由 `maxSteps` / timeout 约束。用户消息记录 `origin`/`clientActionId`/`content_hash`。Evidence quote 用源消息 offset，按条返回 created/already_exists/quote_mismatch。推荐问题只在 `run.completed` 后解析，纠正 UI 仍无 suggestion chip。不改已哈希 Skill `10.0.11`。
 - 验证：`frontend/tests/rectification-step-answer.test.ts`、`frontend/tests/rectification-v9-stream.test.ts`、`frontend/tests/rectification-server-focus.test.ts`、`frontend/tests/rectification-evidence-quote.test.ts`、`frontend/tests/rectification-agentic-entry.test.ts`、`frontend/tests/rectification-v10-tool-contract.test.ts`。
-- 防复发：禁止把含工具调用的 step 的 `text-delta` 发给浏览器。禁止把 `thinking.delta` 改名为 `answer.delta`。禁止模型驱动 `set-focus`。禁止对 `duplicate_focus` / `quote_mismatch` / `zero_information_gain` 做 `attempt.reset`。禁止未完成运行解析或自动提交 suggestion。禁止模型改写 Evidence quote。
-- 相关记录：BUG-345、BUG-354、BUG-357、BUG-367
+- 防复发：禁止把含工具调用的 step 的 `text-delta` 发给浏览器。禁止把 `thinking.delta` 改名为 `answer.delta`。禁止模型驱动 `set-focus`。禁止对 `duplicate_focus` / `quote_mismatch` / `zero_information_gain` 做 `attempt.reset`。禁止对第二次相同公开工具调用 abort 整轮（见 BUG-372）。禁止未完成运行解析或自动提交 suggestion。禁止模型改写 Evidence quote。
+- 相关记录：BUG-345、BUG-354、BUG-357、BUG-367、BUG-372
 - 复发自：BUG-367（关掉公开 `thinking.delta` 后，中间 step 的 `text-delta` 被整段当成回答）
 - 修复版本：fe87a9ecdb71eae4eb79664664cc20c534177c19
 

frontend/src/lib/rectification-agentic/v9/agent-run.ts

@@ -109,7 +109,6 @@ type AttemptOutcome = Readonly<{
   attemptId: string;
 }>;
 
-const REPEATED_TOOL_CALL_LIMIT = 1;
 const MAX_ATTEMPTS = 2;
 const RETRYABLE_ERROR_CODES = new Set([
   "stream_aborted",
@@ -137,6 +136,22 @@ function first(value: unknown): unknown {
   return value;
 }
 
+/**
+ * Identity for a public tool-call chunk. Mastra may put the model input on
+ * `args`, `input`, or omit it; missing input collapses to `{}` so a second
+ * call of the same tool name still looks identical.
+ */
+function publicToolCallKey(toolName: string, payload: unknown): string {
+  if (!payload || typeof payload !== "object") return `${toolName}:{}`;
+  const record = payload as Record<string, unknown>;
+  const args = record.args ?? record.input ?? record.toolArgs ?? {};
+  try {
+    return `${toolName}:${JSON.stringify(args)}`;
+  } catch {
+    return `${toolName}:{}`;
+  }
+}
+
 async function rpcOf(
   accounting: RectificationRpcClient,
   fn: string,
@@ -622,6 +637,12 @@ export async function runV9AgentTurn(options: V9AgentRunOptions): Promise<V9Agen
 
       for await (const chunk of result.fullStream) {
         const rawToolName = typeof chunk.payload?.toolName === "string" ? chunk.payload.toolName : "";
+        // Identical public tool-call + args are idempotent. Throwing
+        // `repeated_tool_call` (BUG-368 P0-3) aborted the turn after
+        // evidence / diagnostics / compare had already committed, because
+        // the model often re-issued compare with the same caseId. Bound
+        // loops with maxSteps / timeout instead; do not attempt.reset.
+        let skipDuplicateToolCallReceipt = false;
         if (chunk.type === "tool-call") {
           if (rawToolName === "rectification-read-case" && !skillBound) {
             throw new Error("skill_not_bound");
@@ -630,10 +651,10 @@ export async function runV9AgentTurn(options: V9AgentRunOptions): Promise<V9Agen
             throw new Error("case_not_loaded");
           }
           if (isPublicRectificationToolName(rawToolName)) {
-            const key = `${rawToolName}:${JSON.stringify(chunk.payload?.args ?? {})}`;
+            const key = publicToolCallKey(rawToolName, chunk.payload);
             const count = (repeatedCalls.get(key) ?? 0) + 1;
             repeatedCalls.set(key, count);
-            if (count > REPEATED_TOOL_CALL_LIMIT) throw new Error("repeated_tool_call");
+            skipDuplicateToolCallReceipt = count > 1;
           }
         }
 
@@ -644,7 +665,9 @@ export async function runV9AgentTurn(options: V9AgentRunOptions): Promise<V9Agen
         );
         if (stepEffect.kind === "publish") await publishSpokenStep(stepEffect.pieces);
 
-        const activityEvent = mapStreamChunkToActivity(chunk as never);
+        const activityEvent = skipDuplicateToolCallReceipt
+          ? null
+          : mapStreamChunkToActivity(chunk as never);
         if (activityEvent) {
           await publish(activityEvent);
           if (activityEvent.status === "started") {
@@ -658,7 +681,9 @@ export async function runV9AgentTurn(options: V9AgentRunOptions): Promise<V9Agen
             toolTerminalStatus.set(activityEvent.tool, activityEvent.status);
           }
         }
-        const phaseEvent = mapStreamChunkToPhase(chunk as never);
+        const phaseEvent = skipDuplicateToolCallReceipt
+          ? null
+          : mapStreamChunkToPhase(chunk as never);
         if (phaseEvent) {
           if (phaseEvent.type === "skill.bound" && !skillBound) {
             skillBound = true;

frontend/tests/rectification-v9-agent.test.ts

@@ -409,20 +409,32 @@ test("distinct evidence calls in one natural turn are not mistaken for a repeate
   ]);
 });
 
-test("a repeated identical tool call is detected and aborts the turn", async () => {
-  const { options, billing } = runOptions({
+test("a repeated identical tool call is treated as idempotent and does not abort", async () => {
+  const { options, emitted, billing } = runOptions({
     buildAgent: async () => fakeAgentStream([
       chunk("start"),
       chunk("tool-call", { toolName: "skill", args: { name: RECTIFICATION_SKILL_NAME } }),
       chunk("tool-result", { toolName: "skill" }),
-      ...Array.from({ length: 2 }, () => chunk("tool-call", { toolName: "rectification-read-case", args: { caseId: CASE_ID } })),
+      chunk("tool-call", { toolName: "rectification-read-case", args: { caseId: CASE_ID } }),
+      chunk("tool-result", { toolName: "rectification-read-case" }),
+      chunk("tool-call", { toolName: "rectification-read-case", args: { caseId: CASE_ID } }),
       chunk("finish"),
     ]) as never,
   });
   const result = await runV9AgentTurn(options);
-  assert.equal(result.ok, false);
-  assert.equal(result.errorCode, "repeated_tool_call");
-  assert.equal(billing.released, 1);
+  assert.equal(result.ok, true);
+  assert.equal(result.errorCode, null);
+  assert.match(result.answerText, /已经记下|请继续说下一件/);
+  assert.equal(billing.completed, 1);
+  assert.equal(billing.released, 0);
+  assert.equal(emitted.some((event) => event.type === "attempt.reset"), false);
+  assert.equal(emitted.some((event) => event.type === "run.failed"), false);
+  assert.equal(
+    emitted.filter((event) => event.type === "tool.activity"
+      && (event as { tool?: string; status?: string }).tool === "rectification-read-case"
+      && (event as { tool?: string; status?: string }).status === "started").length,
+    1,
+  );
 });
 
 test("a failed opening does not let the next turn skip the server Skill load gate", async () => {

frontend/tests/rectification-v9-stream.test.ts

@@ -1050,7 +1050,11 @@ test("does not retry set-focus with identical arguments", async () => {
     ...receiptHandlers,
     get_agentic_rectification_case_dossier: () => dossierFixture(),
     append_agentic_rectification_turn: () => ({ turn_id: TURN_ID }),
-    finalize_agentic_rectification_turn: () => ({ turn_id: TURN_ID, status: "failed", idempotent: false }),
+    finalize_agentic_rectification_turn: (_fn, args) => ({
+      turn_id: TURN_ID,
+      status: args.p_status,
+      idempotent: false,
+    }),
   });
   const { options, emitted, billing } = runOptions({
     accounting: accounting.client,
@@ -1071,10 +1075,66 @@ test("does not retry set-focus with identical arguments", async () => {
 
   const result = await runV9AgentTurn(options);
 
-  assert.equal(result.ok, false);
-  assert.equal(result.errorCode, "repeated_tool_call");
+  assert.equal(result.ok, true);
+  assert.equal(result.errorCode, null);
+  assert.equal(result.answerText, "主问题：请确认这段经历发生在哪个月？");
   assert.equal(emitted.some((event) => event.type === "attempt.reset"), false);
-  assert.deepEqual(billing, { reserved: 1, completed: 0, released: 1 });
+  assert.equal(emitted.some((event) => event.type === "run.failed"), false);
+  assert.deepEqual(billing, { reserved: 1, completed: 1, released: 0 });
+});
+
+test("duplicate compare after diagnostics still completes with server narration", async () => {
+  const executedMethods = [
+    "ashtakavarga",
+    "d1-rashi",
+    "d10-dashamsa",
+    "shadbala",
+    "functional-benefic-malefic",
+    "arudha-pada",
+  ];
+  const { options, emitted, billing } = runOptions({
+    buildAgent: async () => fakeAgentStream([
+      chunk("start"),
+      chunk("tool-call", { toolName: "skill", args: { name: RECTIFICATION_SKILL_NAME } }),
+      chunk("tool-result", { toolName: "skill" }),
+      chunk("tool-call", { toolName: "rectification-read-case", args: { caseId: CASE_ID } }),
+      chunk("tool-result", { toolName: "rectification-read-case" }),
+      chunk("tool-call", { toolName: "rectification-record-evidence-batch", args: { caseId: CASE_ID } }),
+      chunk("tool-result", { toolName: "rectification-record-evidence-batch" }),
+      chunk("tool-call", { toolName: "rectification-propose-evidence", args: { caseId: CASE_ID } }),
+      chunk("tool-result", { toolName: "rectification-propose-evidence" }),
+      chunk("tool-call", { toolName: "rectification-confirm-evidence", args: { caseId: CASE_ID } }),
+      chunk("tool-result", { toolName: "rectification-confirm-evidence" }),
+      chunk("tool-call", { toolName: "rectification-read-diagnostics", args: { caseId: CASE_ID } }),
+      chunk("tool-result", {
+        toolName: "rectification-read-diagnostics",
+        result: { executed_methods: executedMethods },
+      }),
+      chunk("tool-call", { toolName: "rectification-compare-candidates", args: { caseId: CASE_ID } }),
+      chunk("tool-result", {
+        toolName: "rectification-compare-candidates",
+        result: { executed_methods: executedMethods },
+      }),
+      chunk("tool-call", { toolName: "rectification-compare-candidates", args: { caseId: CASE_ID } }),
+      chunk("finish"),
+    ]) as never,
+  });
+
+  const result = await runV9AgentTurn(options);
+
+  assert.equal(result.ok, true);
+  assert.equal(result.errorCode, null);
+  assert.match(result.answerText, /已经记下|请继续说下一件/);
+  assert.equal(emitted.some((event) => event.type === "attempt.reset"), false);
+  assert.equal(emitted.some((event) => event.type === "run.failed"), false);
+  assert.equal(emitted.some((event) => event.type === "run.completed"), true);
+  assert.equal(
+    emitted.filter((event) => event.type === "tool.activity"
+      && (event as { tool?: string; status?: string }).tool === "rectification-compare-candidates"
+      && (event as { tool?: string; status?: string }).status === "started").length,
+    1,
+  );
+  assert.deepEqual(billing, { reserved: 1, completed: 1, released: 0 });
 });
 
 test("an unclaimed V10 attempt never starts the model", async () => {