diff --git a/CHANGELOG.md b/CHANGELOG.md
index a4209ba9..d1df2a90 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,5 +1,10 @@
 # 印度占星 Skill 更新日志
 
+## 2026-10-28 — 生时校正每一轮少等一次模型
+
+- 生时校正每一轮不再先让模型花一次请求去读案例，由服务器直接读好交给模型，模型一开始就回应你。估计每轮快 3～8 秒；问答内容和流程不变。
+- Skill 版本不变（BUG-1491，issue #136）。
+
 ## 2026-10-28 — 修正排盘服务同时算两张盘时偶发报错
 
 - 上一版「同一份代码只留一份」在两张盘同时算（如合盘全档案）时，偶尔会拿到还没载入完的代码而报错；现在会等它载入完再用。结果不变。
diff --git a/docs/BUG_HISTORY.md b/docs/BUG_HISTORY.md
index b5638765..dbe27cf0 100644
--- a/docs/BUG_HISTORY.md
+++ b/docs/BUG_HISTORY.md
@@ -20370,3 +20370,18 @@
 - 修复版本：分支 `codex/upstream-sync9-20261028`。
 
 > 第九轮同步（issue #141）原用 BUG-1472～1479，与同日另一会话合入 staging 的编号冲突，合并时整体顺延 11 号为 BUG-1483～1490；分支上的提交信息仍写旧号。
+
+## BUG-1491 | 生时校正每一轮先白调一次模型，只为读一下案例
+
+- 状态：resolved（代码 + 回归测试；待部署）
+- 首次发现 / 最近更新：2026-10-28 / 2026-10-28（issue #118 代码体检第 7 条 → #136；09-26 记为 D6「另议」，产品 10-28 同意）
+- 影响面：生时校正每一轮；`frontend/src/lib/rectification-agentic/v9/agent-run-attempt.ts`、新模块 `agent-run-opening-read.ts`、`agent-run-messages.ts`、`frontend/src/mastra/agentic-rectification.ts`、`frontend/src/mastra/rectification-v9-tools.ts`（read-case 的工具说明）。
+- 现象：每一轮都固定先发一次模型请求，第 0 步只允许调 `rectification-read-case`，模型唯一能做的就是读案例；读完再发下一次请求才开始真正的回应。估每轮多 3～8 秒（体检估计）。
+- 触发条件：任何一轮生时校正（开场、证据、选择题等）。
+- 根因：「第一步必须读案例」用模型的第 0 步来保证，而默认读取（`turn_decision` 投影、不恢复）的结果服务器自己就能算。
+- 修复：运行器在绑定 Skill 之后、调模型之前，直接执行同一个 read-case 工具（同一个工具对象、同一份回执、同一个投影），把这次调用和结果作为对话开头交给模型；模型从第 1 步起就是真正的回应，第 0 步也拿到全部工具。read-case 仍可再调：暂停的 Case 要按用户意思恢复（`resume=true`）、最终报告要 `full_diagnostics` 时由模型再调。服务器读取的 `tool-call` / `tool-result` 走原来同一条处理路径，所以 `case.loaded`、工具活动、Skill 回执各只出现一次；它们不计入模型步数和步耗时。思考模式下 DeepSeek 要求历史里的工具调用带回思考内容，空的会被丢掉、缺了会 HTTP 400，所以合成的那次调用带一句「服务器已读取本轮案例。」。服务器读取失败时本次 attempt 记为 `case_not_loaded`、不调模型（与以前模型第 0 步读失败的结果相同）。拿不到 read-case 工具的 agent（测试里的假 agent）保持原来的第 0 步读取。提示词三处同步：系统引导、两套 agent 指令的第 1 条、read-case 的工具说明。Skill 包本身没有「第一步读案例」的原文（只把 read-case 列为可用工具），版本不变。
+- 验证：`frontend/tests/rectification-v9-agent.test.ts` 新增 3 条：服务器只读一次、用默认投影；第 0 步已有全部工具且仍含 read-case；交给模型的历史里有带思考内容的合成调用和原样投影；引导语换成「本轮案例已由服务器读取」；read-case 的工具活动「started」与 `case.loaded` 各 1 次。服务器读取失败 → `case_not_loaded`、不调模型。没有 `listTools` 的 agent 仍走第 0 步读取、仍是原引导语。另用真实 Mastra `Agent` 验证 `listTools()` 拿到的工具能直接执行；用产品同一通道（Mastra OpenAI 兼容通道 + `deepseek-flash` 思考模式）实测带合成调用的请求被接受并按读取结果作答。全量前端测试与 staging 逐条比对（本机让 `python3` 指向项目 venv，见 ERR-117）：只多出新增的 3 条，没有新失败（合盘全档案那条在本分支上失败，是 BUG-1482 的线程问题，staging `a614895a` 已修，合入后消失）。
+- 防复发：上述测试。服务器能确定的工具调用不要让模型花一步去做。
+- 关联：BUG-282 / BUG-937（思考模式与工具选择）、BUG-368（重复工具调用）。
+- 复发自：无
+- 修复版本：分支 `codex/rectification-preread-case-20261028`。
diff --git a/docs/testing/code-health-118-20261028.md b/docs/testing/code-health-118-20261028.md
index 38cd3589..e2842cd5 100644
--- a/docs/testing/code-health-118-20261028.md
+++ b/docs/testing/code-health-118-20261028.md
@@ -46,3 +46,9 @@
 22. 在档案页按刷新：列表和以前一样；改一个人的资料、添加一个人后，列表马上是新的。
 23. 在合盘页按刷新：选人区、余额、历史报告照常；合盘服务关着时仍只显示「合盘服务暂未开放」，不闪选人区。
 
+## #136 生时校正每一轮少等一次模型（BUG-1491）
+
+24. 从首页进入生时校正，走完开场：开场的话和第一道题照常出来，比以前快一截（以前出字前会多停几秒）。
+25. 说一件人生大事（带年月）：照常回「记下了」并出下一题；进度、选项都正常。
+26. 先说「先这样」暂停，再回来说「继续」：能接着校正，不会卡在「已暂停」。
+
diff --git a/frontend/src/lib/rectification-agentic/v9/agent-run-attempt.ts b/frontend/src/lib/rectification-agentic/v9/agent-run-attempt.ts
index 7f587c5f..73789f85 100644
--- a/frontend/src/lib/rectification-agentic/v9/agent-run-attempt.ts
+++ b/frontend/src/lib/rectification-agentic/v9/agent-run-attempt.ts
@@ -73,6 +73,7 @@ import {
   v9TurnReceipts,
 } from "./agent-run-support";
 import { buildAgentMessages } from "./agent-run-messages";
+import { openingCaseReader, withOpeningRead, type OpeningRead } from "./agent-run-opening-read";
 import { productDeliveryForDomain } from "./product-delivery.ts";
 import type { V9TimedTurn } from "./agent-run-prepare";
 
@@ -133,6 +134,8 @@ export async function streamV9Attempt(
       loadCompute: () => loadV9CaseCompute(accounting, userId, caseId),
     });
   }
+  // issue #136: the server reads the case itself instead of spending step 0 on it.
+  const readCase = await openingCaseReader(agent);
   const messages = buildAgentMessages(
     options,
     attemptNumber,
@@ -141,6 +144,7 @@ export async function streamV9Attempt(
     birthTimeClue,
     previousErrorCode,
     openingBoundarySentence(openingFacts),
+    readCase !== null,
   );
   const maxSteps = resolveRectificationStepBudget(action);
   const abortController = new AbortController();
@@ -215,6 +219,14 @@ export async function streamV9Attempt(
     await recordPhase("skill.bound");
     await publish({ type: "skill.bound" });
     emittedKeys.add("event:skill.bound::");
+    let openingRead: OpeningRead | null = null;
+    if (readCase) {
+      try {
+        openingRead = await readCase(caseId);
+      } catch {
+        return failedAttempt(attemptId, "case_not_loaded");
+      }
+    }
 
     const generation = agentGenerationSettings(options.generationModel, {
       thinking: "enabled",
@@ -249,14 +261,15 @@ export async function streamV9Attempt(
         }>;
         totalUsage?: Promise<Record<string, unknown>>;
       }>;
-    }).stream(messages, {
+    }).stream(openingRead ? [...messages, ...openingRead.messages] : messages, {
       maxSteps,
       abortSignal: abortController.signal,
       ...generation,
       // Thinking-mode providers reject named/required tool_choice. Restrict
       // the first step to read-case and keep tool_choice auto; the runner
-      // still refuses any other public tool before case.loaded.
-      prepareStep: ({ stepNumber }) => stepNumber === 0
+      // still refuses any other public tool before case.loaded. When the
+      // server already read the case, every step has the full tool set.
+      prepareStep: ({ stepNumber }) => stepNumber === 0 && !openingRead
         ? {
             activeTools: ["rectification-read-case"],
             toolChoice: "auto",
@@ -329,8 +342,9 @@ export async function streamV9Attempt(
     };
 
     try {
-    for await (const chunk of result.fullStream) {
-      recordStepChunk(stepTimings, chunk, Date.now() - attemptStartedAt, isPublicRectificationToolName);
+    for await (const { chunk, synthetic } of withOpeningRead(openingRead, result.fullStream)) {
+      // The server's read-case is not a model step: no step timing, no step answer.
+      if (!synthetic) recordStepChunk(stepTimings, chunk, Date.now() - attemptStartedAt, isPublicRectificationToolName);
       const progressStage = turnProgressForChunk(chunk as never);
       if (progressStage) reportTurnProgress(progressStage);
       const rawToolName = typeof chunk.payload?.toolName === "string" ? chunk.payload.toolName : "";
@@ -355,11 +369,9 @@ export async function streamV9Attempt(
         }
       }
 
-      const stepEffect = applyStepAnswerChunk(
-        stepAnswer,
-        chunk,
-        isPublicRectificationToolName,
-      );
+      const stepEffect = synthetic
+        ? { kind: "none" as const }
+        : applyStepAnswerChunk(stepAnswer, chunk, isPublicRectificationToolName);
       if (stepEffect.kind === "live") await publishSpokenStep([stepEffect.text], true);
       if (stepEffect.kind === "publish") await publishSpokenStep(stepEffect.pieces);
       if (stepEffect.kind === "retract") await retractSpoken();
diff --git a/frontend/src/lib/rectification-agentic/v9/agent-run-messages.ts b/frontend/src/lib/rectification-agentic/v9/agent-run-messages.ts
index e6cd25e0..9fb8f07a 100644
--- a/frontend/src/lib/rectification-agentic/v9/agent-run-messages.ts
+++ b/frontend/src/lib/rectification-agentic/v9/agent-run-messages.ts
@@ -64,6 +64,7 @@ export function buildAgentMessages(
   birthTimeClue: string | null = null,
   previousErrorCode: string | null = null,
   boundarySentence: string | null = null,
+  caseReadByServer = false,
 ): unknown[] {
   const timeContext = options.timeContext
     ?? `服务端当前时间（权威）：${new Date().toISOString()}。涉及“现在、今天、今年、未来几个月”等相对时间时，以此为准。`;
@@ -72,7 +73,10 @@ export function buildAgentMessages(
   // the Skill sections it uses; other actions keep the whole bound body.
   const boundSkill = sliceRectificationSkill(skillInstructions, options.action).text;
   const bootstrapContent = [
-    "【服务器已绑定当前 Case 的精确 Skill】运行器已在本 attempt 内加载并核验下列指令；不要重复调用 skill。第一步必须调用 rectification-read-case。",
+    `【服务器已绑定当前 Case 的精确 Skill】运行器已在本 attempt 内加载并核验下列指令；不要重复调用 skill。${caseReadByServer
+      // issue #136: the turn opens with the server's read-case call and result.
+      ? "本轮案例已由服务器读取（rectification-read-case 默认投影，就在对话末尾），不要再读同一份；只有暂停的 Case 要按用户意思恢复（resume=true）或最终报告需要 full_diagnostics 时才再调用它。"
+      : "第一步必须调用 rectification-read-case。"}`,
     boundSkill,
     ...(productDeliveryForDomain(dossier.case.rectificationDomain) ? [
       `【服务端持久产品身份】${JSON.stringify(productDeliveryForDomain(dossier.case.rectificationDomain))}。本轮按目标盘型交付，不能因摘要缺失退分钟优先；盘型事实只读工具返回的segment_summary，代表签名仅诊断，不能作实际采用盘。`,
diff --git a/frontend/src/lib/rectification-agentic/v9/agent-run-opening-read.ts b/frontend/src/lib/rectification-agentic/v9/agent-run-opening-read.ts
new file mode 100644
index 00000000..8929a5b2
--- /dev/null
+++ b/frontend/src/lib/rectification-agentic/v9/agent-run-opening-read.ts
@@ -0,0 +1,90 @@
+/**
+ * The server reads the case before the model's first step (issue #136).
+ *
+ * Every turn used to spend one model call whose only job was to call
+ * `rectification-read-case` with the default projection, a result the server
+ * can produce itself. The runner now calls the bound tool directly (same
+ * tool, same receipts, same projection) and hands the model the call and its
+ * result as the opening of the turn, so the model starts at the real work.
+ * The model can still call read-case again: `resume=true` for a paused case
+ * the user wants to continue, `projection=full_diagnostics` for the report.
+ */
+
+export const OPENING_READ_TOOL = "rectification-read-case";
+export const OPENING_READ_TOOL_CALL_ID = "server-read-case-0";
+/**
+ * Thinking-mode providers reject a prior assistant tool call without
+ * reasoning_content (DeepSeek: HTTP 400), and an empty reasoning part is
+ * dropped on the way out, so the synthetic call carries one short line.
+ */
+const OPENING_READ_REASONING = "服务器已读取本轮案例。";
+
+export type OpeningReadChunk = Readonly<{
+  type: "tool-call" | "tool-result";
+  payload: Readonly<{ toolCallId: string; toolName: string; args: { caseId: string }; result?: unknown }>;
+}>;
+
+export type OpeningRead = Readonly<{
+  chunks: readonly OpeningReadChunk[];
+  messages: readonly unknown[];
+}>;
+
+type ToolLike = { execute?: (input: { caseId: string }, context?: unknown) => Promise<unknown> };
+
+/**
+ * The agent's own read-case tool, or null when it cannot hand one out; the
+ * model then reads on step 0 as before. Looked up before the turn messages
+ * are built, run after the Skill is bound so receipts keep their order.
+ */
+export async function openingCaseReader(agent: unknown): Promise<((caseId: string) => Promise<OpeningRead>) | null> {
+  const listTools = (agent as { listTools?: () => Promise<Record<string, unknown>> } | null)?.listTools;
+  if (typeof listTools !== "function") return null;
+  let tool: ToolLike | undefined;
+  try {
+    tool = (await listTools.call(agent))?.[OPENING_READ_TOOL] as ToolLike | undefined;
+  } catch {
+    return null;
+  }
+  const execute = tool?.execute;
+  if (typeof execute !== "function") return null;
+  return async (caseId) => {
+    const args = { caseId };
+    // The tool writes its own started / completed (or failed) receipt; a throw
+    // here is the same failure the model's step-0 call would have hit.
+    const result = await execute.call(tool, args, {});
+    const toolCall = { toolCallId: OPENING_READ_TOOL_CALL_ID, toolName: OPENING_READ_TOOL, args };
+    return {
+      chunks: [
+        { type: "tool-call", payload: toolCall },
+        { type: "tool-result", payload: { ...toolCall, result } },
+      ],
+      messages: [
+        {
+          role: "assistant",
+          content: [
+            { type: "reasoning", text: OPENING_READ_REASONING },
+            { type: "tool-call", toolCallId: OPENING_READ_TOOL_CALL_ID, toolName: OPENING_READ_TOOL, input: args },
+          ],
+        },
+        {
+          role: "tool",
+          content: [{
+            type: "tool-result",
+            toolCallId: OPENING_READ_TOOL_CALL_ID,
+            toolName: OPENING_READ_TOOL,
+            output: { type: "json", value: result },
+          }],
+        },
+      ],
+    };
+  };
+}
+
+/** The server's chunks first, then the model's stream; each marked so steps are only the model's. */
+export async function* withOpeningRead<T>(
+  opening: OpeningRead | null,
+  stream: AsyncIterable<T>,
+): AsyncGenerator<{ chunk: T; synthetic: boolean }> {
+  for (const chunk of opening?.chunks ?? []) yield { chunk: chunk as unknown as T, synthetic: true };
+  for await (const chunk of stream) yield { chunk, synthetic: false };
+}
diff --git a/frontend/src/mastra/agentic-rectification.ts b/frontend/src/mastra/agentic-rectification.ts
index f673f699..60d8a961 100644
--- a/frontend/src/mastra/agentic-rectification.ts
+++ b/frontend/src/mastra/agentic-rectification.ts
@@ -45,7 +45,7 @@ const RECTIFICATION_VOICE_HEAD = `你是 Jyotisha，只服务当前绑定 jyotis
 像一个懂行、说话温和的姐姐：先回应人，再展开方法；语气放软，但不含糊。用户想不起来、说不准时不催、不评判，「记不清也没关系」这类话最多一句；不说教，不写「呀」「哦」「抱抱」「宝」「亲爱的」，不叠字，不卖萌，不用表情。`;
 
 const agenticRectificationInstructions = `${RECTIFICATION_VOICE_HEAD}范围由系统说：范围变化、没重新比较、比较没跑成这几句由服务器接在你的正文后面；除出牌轮外你不写时刻（HH:MM）、时间区间和百分比；任何一句里的时刻或百分比与服务端当轮事实不符，整句不会落库。下一问是采集题时，先承接用户刚说的事；进度数字只用 collection_progress，没有该字段不说进度。不得写「范围在收窄」「继续收窄」这类没有数字的进度句。
-1. 第一步调用 rectification-read-case。服务器是事实、焦点、权限与终态的唯一权威。
+1. 每轮先要有 rectification-read-case 的结果：服务器已在对话里附上本轮的读取结果时直接用，不要再读；没有附上时第一步调用它（issue #136）。服务器是事实、焦点、权限与终态的唯一权威。
 2. 事实只能来自用户原话；复述日期必须用 display_date_label。不得虚构事件、候选或出生分钟。
 3. 新事件走 rectification-record-evidence-batch。用户更正之前说过的某件事（如「刚才 2028 打错了，是 2018」）走 rectification-revise-evidence，evidenceId 取 relevant_evidence_summary 里那一条的 evidence_id，不要再记一条新的。工具执行保持静默；思考用简体中文写在思维链；对用户说的话必须自己写在正文里，不叙述工具或内部状态。
 4. 每轮在记录证据后，用 rectification-set-focus 的 spokenPrompt 写出服务端给你的下一问：用自己的话、结合用户刚说的事，问出同一个年份/期间和同一个事件家族；不得改年份、不得改选项含义、不得合并两道题。正文只做承接，不提问、不复述题干、不预告选项——题干会作为同一条消息的下一段自动出现。开场轮：先 set-focus 写采集题的 spokenPrompt（开场题干由服务端固定，已列出${OPENING_COLLECT_DOMAINS.join("、")}和一个回答示例），正文三层大白话：要把出生时间缩小到更准的范围、现在先在哪段时间里找；做法是用户说几件人生大事和大概年月，拿去和星盘对照；限度是「${OPENING_LIMIT_BODY}」，不得丢掉「不保证」和「不准」。正文不用大运、盘面、分盘、候选、区间、代表分钟、精确到秒这类词，不重复题干里的例子，不提问；不得写具体年份，不得要求先准备材料。没有下一问（服务端返回 next_followup=null）时不要自拟问题。证据轮的复述由服务器写：「记下了 N 件事」（两件以内直接列出），清单挂在同一条消息里；你不逐条复述用户说的经历，不得评价价值或写「很有帮助 / 很有价值 / 很有分量 / 特别有用」。case.accepted_time 非空时，正文第一句要说明已按该时间采用、现在在核对。正文不得断言界面当前状态，不要写「界面上有下一问」「界面上出现了…」。choice 选项由服务端写入同一条消息，collect_spoken 只承接用户刚说的事实，不输出输入提示。点选与「先这样」由服务器处理。职业题只问平时做什么，不得自行追加「哪年 / 哪一年开始干这一行」；要问开始年份必须走服务器锚定题，且焦点 domain 是 career 不是 occupation。
@@ -55,7 +55,7 @@ const agenticRectificationInstructions = `${RECTIFICATION_VOICE_HEAD}范围由
 坏：「范围还在收。」好：「2016 年 9 月入学记下了。还有吗？比如第一份工作、搬到别的城市。」`;
 
 const segmentProductInstructions = `${RECTIFICATION_VOICE_HEAD}服务器持久产品身份是唯一交付口径；目标是判断用哪张盘解读，不承诺缩到某一分钟。没有摘要仍是新产品，不能退回分钟优先。
-1. 第一步调用 rectification-read-case。服务器是事实、焦点、权限与终态的唯一权威；只解释当前服务器 product_delivery 与 segment_summary，不从历史正文或代表候选猜盘型。
+1. 每轮先要有 rectification-read-case 的结果：服务器已在对话里附上本轮的读取结果时直接用，不要再读；没有附上时第一步调用它（issue #136）。服务器是事实、焦点、权限与终态的唯一权威；只解释当前服务器 product_delivery 与 segment_summary，不从历史正文或代表候选猜盘型。
 2. 事实只能来自用户原话；日期只用 display_date_label。新事件走 rectification-record-evidence-batch；工具执行保持静默，对用户说的话必须自己写在正文里。
 3. 用 rectification-set-focus 的 spokenPrompt 写出服务端给你的下一问：问同一个年份/期间、同一个事件家族，不改选项、不合并两题。正文只承接，不提问、不复述题干。没有下一问就不自拟问题。职业题只问平时做什么；开始年份只能走服务器 career 锚定题。
 4. 开场正文三层大白话：先判断现在的出生时间能不能选定用来解读的星盘；你说几件人生大事和大概年月，拿去和星盘对照；限度是「${OPENING_LIMIT_BODY}」，不得丢掉「不保证」和「不准」。搜索窗口只作次要信息。不说要把出生时间缩小到更准的范围，不写具体年份、不列例子、不要求准备材料；开场题干仍由服务器固定。
diff --git a/frontend/src/mastra/rectification-v9-tools.ts b/frontend/src/mastra/rectification-v9-tools.ts
index 315c4aa9..9ab23ec7 100644
--- a/frontend/src/mastra/rectification-v9-tools.ts
+++ b/frontend/src/mastra/rectification-v9-tools.ts
@@ -1217,7 +1217,7 @@ export function createRectificationV9Tools(ctx: RectificationV9Context) {
   const readCaseTool = createTool({
     id: "rectification-read-case",
     description:
-      "读取服务端 Case 的对话决策投影。默认 turn_decision，只含当前问题、探针、候选摘要和近期对话，不得把完整诊断当每轮输入。最终报告或审计才用 projection=full_diagnostics。任何一轮开始必须先调用本工具。仅当 paused Case 的用户语义明确表示继续校正或提交新证据时传 resume=true；询问当前结果、重复停止或只看结果不得恢复。",
+      "读取服务端 Case 的对话决策投影。默认 turn_decision，只含当前问题、探针、候选摘要和近期对话，不得把完整诊断当每轮输入。最终报告或审计才用 projection=full_diagnostics。任何一轮开始都必须先有本工具的结果；服务器已在对话里附上本轮结果时不要重复调用。仅当 paused Case 的用户语义明确表示继续校正或提交新证据时传 resume=true；询问当前结果、重复停止或只看结果不得恢复。",
     inputSchema: z.object({
       caseId: z.string().uuid(),
       projection: z.enum(["turn_decision", "full_diagnostics"]).optional(),
diff --git a/frontend/tests/rectification-v9-agent.test.ts b/frontend/tests/rectification-v9-agent.test.ts
index b8a30d73..3d5b0c33 100644
--- a/frontend/tests/rectification-v9-agent.test.ts
+++ b/frontend/tests/rectification-v9-agent.test.ts
@@ -1122,3 +1122,71 @@ test("an evidence turn that records twelve items is written 「记下了 12 件
   const finalize = accounting.calls.find((call) => call.fn === "finalize_agentic_rectification_turn");
   assert.equal(String(finalize?.args.p_assistant_message ?? ""), "记下了 12 件事。");
 });
+
+// issue #136: the server reads the case itself, so the model's first step is the real work.
+function preReadAgent(chunks: StreamChunk[], execute: (input: unknown) => Promise<unknown>) {
+  const seen: { messages: unknown[]; step0Tools: string[] | null } = { messages: [], step0Tools: null };
+  const agent = {
+    ...fakeAgentStream(chunks),
+    listTools: async () => ({ "rectification-read-case": { execute } }),
+    stream: async (messages: unknown[], options: { prepareStep: (input: { stepNumber: number }) => { activeTools: string[] } }) => {
+      seen.messages = messages;
+      seen.step0Tools = options.prepareStep({ stepNumber: 0 }).activeTools;
+      return {
+        fullStream: (async function* () { for (const item of chunks) yield item; })(),
+        totalUsage: Promise.resolve({ inputTokens: 10, outputTokens: 20 }),
+      };
+    },
+  };
+  return { agent, seen };
+}
+
+test("the server reads the case before the model; the model starts at the real work (issue #136)", async () => {
+  const reads: unknown[] = [];
+  const projection = { case: { id: CASE_ID }, current_question: null };
+  const { agent, seen } = preReadAgent([
+    chunk("start"),
+    chunk("text-delta", { text: "已经记下。" }),
+    chunk("finish"),
+  ], async (input) => { reads.push(input); return projection; });
+  const { options, emitted } = runOptions({ buildAgent: async () => agent as never });
+  const result = await runV9AgentTurn(options);
+
+  assert.equal(result.ok, true, String(result.errorCode));
+  assert.deepEqual(reads, [{ caseId: CASE_ID }], "read once, with the default projection");
+  assert.ok(seen.step0Tools && seen.step0Tools.length > 1 && seen.step0Tools.includes("rectification-read-case"),
+    "step 0 already has the full tool set; read-case stays available for resume / full diagnostics");
+  const history = seen.messages as Array<{ role: string; content: unknown }>;
+  const call = history.find((message) => message.role === "assistant") as { content: Array<{ type: string; text?: string; toolName?: string }> };
+  assert.ok(call.content.some((part) => part.type === "reasoning" && Boolean(part.text)), "thinking providers need reasoning_content");
+  assert.ok(call.content.some((part) => part.type === "tool-call" && part.toolName === "rectification-read-case"));
+  const tool = history.find((message) => message.role === "tool") as { content: Array<{ output: { value: unknown } }> };
+  assert.deepEqual(tool.content[0].output.value, projection, "the model sees the projection unchanged");
+  assert.match(JSON.stringify(history[0]), /本轮案例已由服务器读取/);
+  const started = emitted.filter((event) => event.type === "tool.activity"
+    && (event as { tool?: string; status?: string }).tool === "rectification-read-case"
+    && (event as { status?: string }).status === "started");
+  assert.equal(started.length, 1);
+  assert.equal(emitted.filter((event) => event.type === "case.loaded").length, 1);
+  assert.ok(result.toolsUsed.includes("rectification-read-case"));
+});
+
+test("a failed server read fails the attempt as case_not_loaded and never streams (issue #136)", async () => {
+  let streams = 0;
+  const { agent } = preReadAgent([chunk("start"), chunk("finish")], async () => { throw new Error("db down"); });
+  const counting = { ...agent, stream: async (...args: Parameters<typeof agent.stream>) => { streams += 1; return agent.stream(...args); } };
+  const { options } = runOptions({ buildAgent: async () => counting as never });
+  const result = await runV9AgentTurn(options);
+  assert.equal(result.ok, false);
+  assert.equal(result.errorCode, "case_not_loaded");
+  assert.equal(streams, 0);
+});
+
+test("an agent without listTools keeps the old step-0 read (issue #136)", async () => {
+  const { agent, seen } = preReadAgent([], async () => ({}));
+  const legacy = { ...agent, listTools: undefined };
+  const { options } = runOptions({ buildAgent: async () => legacy as never });
+  await runV9AgentTurn(options);
+  assert.deepEqual(seen.step0Tools, ["rectification-read-case"]);
+  assert.match(JSON.stringify((seen.messages as unknown[])[0]), /第一步必须调用 rectification-read-case/);
+});