diff --git a/CHANGELOG.md b/CHANGELOG.md
index cb30e92..6460caf 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,6 +1,7 @@
 # Changelog
 
 ## Unreleased (master, since v0.1.38)
+- **fix(reasoning): 闭合判定改按回合证据——无用户消息的长 agent 会话不再永久保留 compress thinking（closes #348）** — 原门控“compress 调用之后存在真实用户消息才算闭合”在长 agent 会话不可达（整个会话只有开头 1–2 条用户消息，后续 30 个 compress 全部被永久视为活跃回合，观察会话 0 次触发，thinking 地板 20.6K/8.4K/10.6K 字符全部滞留）。现在闭合判定改为：消息内**每个** compress toolCall 的 toolResult（role `toolResult`、`toolCallId` 匹配）已出现在更晚位置，且其后至少还有一条消息（回合已实际推进）。安全门不变：结果未返回或结果仍是最后一条消息（在飞中）绝不动；nudge 在 drop 之后才注入，不可能光当“结果后的消息”闭合在飞回合；per-provider `compress.providers.<name>.reasoning.drop=false` 逃生阀保留（GLM 等 reasoning 回显模型）。测试重写 + 新增 #348 场景（无用户消息的助手链闭合、result 悬置、result 在 call 之前、多 toolCall 部分闭合、误 id 不闭合）
 - **fix(overflow): output headroom 预留按窗口比例封顶，默认 25%（closes #207）** — `reserveOutputHeadroom` 原按模型注册表 maxTokens **全额**预留输出预算：maxTokens 占窗口比例大的模型（qwen3.8-27b：262144 窗口 / 131072 maxTokens）输入预算被砍半，kernel 75% 强制压缩带在完整窗口 ~37% 处触发（host pct 仅 ~34%，两个口径不同加剧误导）。现在预留量 = min(maxTokens, `outputHeadroomMaxPct` × window)：新增 acp.json 配置键 `outputHeadroomMaxPct`（默认 0.25，接受比例或 `"N%"`；0 完全禁用预留，≥1 恢复旧的全额行为）。小预留不受影响（同窗口 int4 版 32K maxTokens 保持原样），超出预留的超长回复溢出一次后由既有 overflow self-heal（learned window + armed emergency）下一轮恢复。可观测性：`[turn]` 日志新增 `fullWindow` 字段（仅当 limit 被预留削减时出现，= 本轮 recenter 后的完整窗口），消除 pct（完整窗口口径）vs limit（预留后口径）混淆；`output-headroom` 事件日志新增 `cap` 字段；`/acp` 面板与 `acp_status` 分母经 `applyOutputHeadroom` 同步使用同一封顶值（#267 统一口径不回归）
 - **fix(state): in-memory（无 session 文件）会话压缩后状态不再丢失（closes #322）** — `src/state.ts` 的 cache 更新原本在 no-file early return 之后，导致没有 session 文件的 in-memory 会话压缩后状态不被保存（恢复/重启时丢失）。现在将 cache 更新移到 early return 之前，确保 in-memory 会话也持久化压缩状态。测试：`tests/state.test.ts` + `tests/compress-tool.test.ts` 扩展
 - **fix(nudge): 增长基线刻度混用复位（closes #267）** — token 计量在估算值 ↔ provider 真实 usage 锚点间翻转（hostFloor 激活/失活）时，增长增量跨翻转相减是伪度量：旧刻度基线对新刻度 tokenCount 相减，凭空多出 "+35K 假增长"，绕过 cadence 节流乱发 nudge（或反向漏发）。现在翻转时统一重置全部增长基线：kernel 侧 `lastNudgeShownTokens` / `lastPerMessageNudgeTokens` / `lastShownByTier`（0.0.55 per-tier cadence 基线）+ 扩展侧 `clearNudgeTokenStamps()`（#316 引入的同轮 re-inject 刻度戳，旧刻度戳同样会伪造满地板增长触发伪重注入）；usage bands 的 floor-stale 行为保留不动。附带：② `/acp` 面板与 `acp_status` 百分比分母统一为与实时仲裁相同的 `window − maxOutput`（新增 `src/overflow-selfheal.ts` `applyOutputHeadroom`，此前面板按全窗口报告而 nudge bands 按 headroom 后窗口仲裁）；③ auto-update 限流与只读停止标记按安装位置 sha256 隔离（多副本互不干扰）。实现：`src/runtime.ts` `noteTokenScale`/`dropTokenScale`，`src/index.ts` 翻转检测重置，`src/update.ts` `locationKey`；测试 `tests/growth-scale-flip.test.ts` + `tests/update.test.ts` 扩展
diff --git a/src/index.ts b/src/index.ts
index 8724e21..875e278 100644
--- a/src/index.ts
+++ b/src/index.ts
@@ -403,11 +403,12 @@ function wireContextTransform(pi: ExtensionAPI, runtime: AcpRuntime, standDownIf
     let rebuilt = coreOutToAgentMessages(turn.messages, originalById);
     // [#336] Request-time reasoning drop, aligned with opencode-acp #377:
     // compress calls are hard-exempt from compression, so their thinking
-    // rides along every request as an unreclaimable floor. Applied BEFORE the
-    // nudge push so a synthetic user-role nudge can never become the "last
-    // genuine user message" boundary and extend the closed zone over the
-    // active round. Persisted history is never modified — this only rewrites
-    // the outgoing view, rebuilt fresh from entries on every event.
+    // rides along every request as an unreclaimable floor. Round closure is
+    // judged on tool-result evidence [#348]; applied BEFORE the nudge push
+    // so a synthetic user-role nudge can never count as the trailing
+    // "message after the result" and close an in-flight round. Persisted
+    // history is never modified — this only rewrites the outgoing view,
+    // rebuilt fresh from entries on every event.
     const reasoningDrop = runtime.reasoningDropFor(ctx);
     const droppedThinking = dropCompressReasoning(rebuilt, reasoningDrop);
     if (droppedThinking !== rebuilt) {
diff --git a/src/reasoning-drop.ts b/src/reasoning-drop.ts
index 48503ed..26b2595 100644
--- a/src/reasoning-drop.ts
+++ b/src/reasoning-drop.ts
@@ -9,9 +9,11 @@ type AgentMessage = SessionMessageEntry["message"];
  *  the anchors that keep block summaries addressable), so their thinking
  *  rides along every request as an unreclaimable context floor. This pass
  *  removes `thinking` parts at request time (persisted history is never
- *  modified) from closed-turn compress messages whose total reasoning length
- *  exceeds `threshold`. The active round (from the last genuine user message
- *  onward) is never touched. */
+ *  modified) from closed-round compress messages whose total reasoning length
+ *  exceeds `threshold`. A round is closed on ROUND EVIDENCE, not on user
+ *  messages: the compress tool result must have arrived and at least one
+ *  message must exist after it. The in-flight round (result still missing or
+ *  still the last message) is never touched [#348]. */
 export interface CompressReasoningConfig {
   /** Master switch. Default: true. `drop: false` disables the pass entirely
    *  (kill-switch; also the recipe for providers whose thinking items are
@@ -54,6 +56,29 @@ function hasCompressCall(content: unknown): boolean {
   });
 }
 
+function compressCallIds(content: unknown): string[] {
+  if (!Array.isArray(content)) return [];
+  return content
+    .filter((p) => {
+      const b = p as { type?: string; name?: string };
+      return b?.type === "toolCall" && b.name === "compress";
+    })
+    .map((p) => (p as { id?: unknown }).id)
+    .filter((id): id is string => typeof id === "string");
+}
+
+/** toolCallId -> index of the message carrying its tool result. Pi gives
+ *  tool results their own `toolResult` role with a top-level `toolCallId`. */
+function resultIndexByCallId(messages: AgentMessage[]): Map<string, number> {
+  const map = new Map<string, number>();
+  for (let i = 0; i < messages.length; i++) {
+    const msg = messages[i] as { role?: string; toolCallId?: unknown };
+    if (msg?.role !== "toolResult" || typeof msg.toolCallId !== "string") continue;
+    if (!map.has(msg.toolCallId)) map.set(msg.toolCallId, i);
+  }
+  return map;
+}
+
 function reasoningLength(content: unknown): number {
   if (!Array.isArray(content)) return 0;
   return content.reduce((n, p) => (isThinking(p) ? n + p.thinking.length : n), 0);
@@ -61,9 +86,13 @@ function reasoningLength(content: unknown): number {
 
 /** Request-time pass aligned with opencode-acp #377: remove `thinking` parts
  *  from a message only when ALL gates hold —
- *  1. closed turn: the message is strictly before the last genuine user
- *     message (pi gives tool results their own `toolResult` role, so every
- *     `user` message is genuine);
+ *  1. closed round [#348]: EVERY `compress` toolCall in the message has its
+ *     tool-result message (role `toolResult`, matching `toolCallId`) at a
+ *     later index, and at least one message exists after that result (the
+ *     round has demonstrably moved on). A compress call without a result, or
+ *     whose result is still the last message, is in flight and never touched
+ *     — no user message is required, so long agentic sessions do close
+ *     rounds; a synthetic nudge pushed later cannot retroactively close one;
  *  2. selector: the message carries a `toolCall` part with name "compress"
  *     (only compress; other protected tools would need their own explicit
  *     config);
@@ -75,17 +104,20 @@ export function dropCompressReasoning(messages: AgentMessage[], cfg?: CompressRe
   const { drop, threshold } = resolveReasoningDrop(cfg);
   if (!drop || messages.length === 0) return messages;
   try {
-    let lastUser = -1;
-    for (let i = 0; i < messages.length; i++) {
-      if ((messages[i] as { role?: string }).role === "user") lastUser = i;
-    }
-    if (lastUser < 0) return messages;
+    const last = messages.length - 1;
+    const resultAt = resultIndexByCallId(messages);
     let changed = false;
     const out = messages.slice();
-    for (let i = 0; i < lastUser; i++) {
+    for (let i = 0; i <= last; i++) {
       const msg = messages[i] as { role?: string; content?: unknown };
       if (msg.role !== "assistant" || !Array.isArray(msg.content)) continue;
       if (!hasCompressCall(msg.content)) continue;
+      const ids = compressCallIds(msg.content);
+      const closed = ids.length > 0 && ids.every((id) => {
+        const ri = resultAt.get(id);
+        return ri !== undefined && ri > i && ri < last;
+      });
+      if (!closed) continue;
       if (reasoningLength(msg.content) <= threshold) continue;
       out[i] = {
         ...(msg as object),
diff --git a/tests/reasoning-drop.test.ts b/tests/reasoning-drop.test.ts
index eb1d523..ac87ab9 100644
--- a/tests/reasoning-drop.test.ts
+++ b/tests/reasoning-drop.test.ts
@@ -14,6 +14,10 @@ function assistant(parts: unknown[]): AgentMessage {
   return { role: "assistant", content: parts, timestamp: 0 } as unknown as AgentMessage;
 }
 
+function toolResult(toolCallId: string): AgentMessage {
+  return { role: "toolResult", toolCallId, content: [{ type: "text", text: "ok" }], timestamp: 0 } as unknown as AgentMessage;
+}
+
 function thinking(len: number, extra: Record<string, unknown> = {}): { type: "thinking"; thinking: string } & Record<string, unknown> {
   return { type: "thinking", thinking: "x".repeat(len), ...extra };
 }
@@ -41,23 +45,53 @@ test("invalid threshold falls back to default, invalid drop is truthy", () => {
   assert.equal(resolveReasoningDrop({ drop: "yes" as unknown as boolean }).drop, true);
 });
 
-test("gate: closed turn — messages from the last genuine user message onward are untouched", () => {
+test("gate: closed round — result arrived and a later message exists", () => {
   const big = thinking(4096);
-  const msgs = [
-    assistant([{ type: "text", text: "old" }, thinking(4096), compressCall()]),
-    user("go"),
-    assistant([{ type: "text", text: "active" }, thinking(4096), compressCall()]),
-  ];
+  const closed = assistant([{ type: "text", text: "old" }, thinking(4096), compressCall("c1")]);
+  const inFlight = assistant([{ type: "text", text: "active" }, thinking(4096), compressCall("c2")]);
+  const msgs = [closed, toolResult("c1"), user("go"), inFlight, toolResult("c2")];
   const out = dropCompressReasoning(msgs, { drop: true, threshold: 0 });
   assert.equal((out[0]!.content as unknown[]).includes(big), false);
-  assert.deepEqual(out[2]!.content, msgs[2]!.content);
+  assert.deepEqual(out[3]!.content, msgs[3]!.content); // result is last message → in flight
+});
+
+test("#348: closes WITHOUT any user message — assistant continuation is round evidence", () => {
+  const closed = assistant([{ type: "text", text: "agentic" }, thinking(4096), compressCall("c1")]);
+  const continuing = assistant([{ type: "text", text: "carrying on" }]);
+  const out = dropCompressReasoning([closed, toolResult("c1"), continuing], { drop: true, threshold: 0 });
+  assert.deepEqual(out[0]!.content, [{ type: "text", text: "agentic" }, compressCall("c1")]);
+  assert.deepEqual(out[2]!.content, continuing.content);
+});
+
+test("gate: result pending (call without any result) is never touched", () => {
+  const pending = assistant([{ type: "text", text: "hi" }, thinking(4096), compressCall("c1")]);
+  const msgs = [pending, assistant([{ type: "text", text: "next" }])];
+  assert.equal(dropCompressReasoning(msgs, { drop: true, threshold: 0 }), msgs);
+});
+
+test("gate: result for a different call id does not close the round", () => {
+  const msg = assistant([{ type: "text", text: "hi" }, thinking(4096), compressCall("c1")]);
+  const msgs = [msg, toolResult("other"), user("go")];
+  assert.equal(dropCompressReasoning(msgs, { drop: true, threshold: 0 }), msgs);
+});
+
+test("gate: result BEFORE the call message does not close the round", () => {
+  const msg = assistant([{ type: "text", text: "hi" }, thinking(4096), compressCall("c1")]);
+  const msgs = [toolResult("c1"), msg, user("go")];
+  assert.equal(dropCompressReasoning(msgs, { drop: true, threshold: 0 }), msgs);
+});
+
+test("gate: mixed message closes on the compress result alone (other calls' results are irrelevant)", () => {
+  const msg = assistant([{ type: "text", text: "hi" }, thinking(4096), compressCall("c1"), otherCall("read")]);
+  const msgs = [msg, toolResult("c1"), user("go")]; // read result never arrives
+  const out = dropCompressReasoning(msgs, { drop: true, threshold: 0 });
+  assert.deepEqual(out[0]!.content, [{ type: "text", text: "hi" }, compressCall("c1"), otherCall("read")]);
 });
 
 test("gate: selector — only messages carrying a compress toolCall part are touched", () => {
-  const think = thinking(4096);
   const textOnly = assistant([{ type: "text", text: "hi" }, thinking(4096)]);
   const otherTool = assistant([{ type: "text", text: "hi" }, thinking(4096), otherCall()]);
-  const msgs = [textOnly, otherTool, user("go")];
+  const msgs = [textOnly, otherTool, toolResult("t1"), user("go")];
   const out = dropCompressReasoning(msgs, { drop: true, threshold: 0 });
   assert.deepEqual(out[0]!.content, textOnly.content);
   assert.deepEqual(out[1]!.content, otherTool.content);
@@ -67,26 +101,33 @@ test("gate: selector — only messages carrying a compress toolCall part are tou
 test("gate: size — strictly exceeds threshold, summed across parts of the same message only", () => {
   const at = assistant([{ type: "text", text: "hi" }, thinking(1024), thinking(1024), compressCall()]);
   const above = assistant([{ type: "text", text: "hi" }, thinking(1025), thinking(1025), compressCall()]);
-  const splitKept = [assistant([{ type: "text", text: "hi" }, thinking(1500), compressCall()]), assistant([{ type: "text", text: "hi" }, thinking(1500), compressCall()]), user("go")];
-  const out = dropCompressReasoning([at, above, user("go")], { drop: true, threshold: 2048 });
+  const splitKept = [
+    assistant([{ type: "text", text: "hi" }, thinking(1500), compressCall("a")]),
+    toolResult("a"),
+    assistant([{ type: "text", text: "mid" }]),
+    assistant([{ type: "text", text: "hi" }, thinking(1500), compressCall("b")]),
+    toolResult("b"),
+    user("go"),
+  ];
+  const out = dropCompressReasoning([at, above, toolResult("c1"), user("go")], { drop: true, threshold: 2048 });
   assert.deepEqual(out[0]!.content, at.content); // 2048 == threshold → kept
   assert.equal((out[1]!.content as unknown[]).some((p) => p === (above.content as unknown[])[1]), false); // 2050 > 2048 → dropped
   const out2 = dropCompressReasoning(splitKept, { drop: true, threshold: 2048 });
   assert.deepEqual(out2[0]!.content, splitKept[0]!.content); // lengths not accumulated across messages
-  assert.deepEqual(out2[1]!.content, splitKept[1]!.content);
+  assert.deepEqual(out2[3]!.content, splitKept[3]!.content);
 });
 
 test("threshold 0 drops any non-empty reasoning; zero-length survives", () => {
   const nonEmpty = assistant([{ type: "text", text: "hi" }, thinking(3), compressCall()]);
-  const empty = assistant([{ type: "text", text: "hi" }, thinking(0), compressCall()]);
-  const out = dropCompressReasoning([nonEmpty, empty, user("go")], { drop: true, threshold: 0 });
+  const empty = assistant([{ type: "text", text: "hi" }, thinking(0), compressCall("c2")]);
+  const out = dropCompressReasoning([nonEmpty, toolResult("c1"), empty, toolResult("c2"), user("go")], { drop: true, threshold: 0 });
   assert.equal((out[0]!.content as unknown[]).length, 2);
-  assert.equal((out[1]!.content as unknown[]).length, 3);
+  assert.equal((out[2]!.content as unknown[]).length, 3);
 });
 
 test("purity and idempotence: input never mutated; second pass is a no-op", () => {
   const original = assistant([{ type: "text", text: "hi" }, thinking(4096), compressCall()]);
-  const msgs = [original, user("go")];
+  const msgs = [original, toolResult("c1"), user("go")];
   const out1 = dropCompressReasoning(msgs, { drop: true, threshold: 0 });
   assert.deepEqual(original.content, [{ type: "text", text: "hi" }, thinking(4096), compressCall()]); // input unmutated
   assert.notEqual(out1[0], msgs[0]); // rewritten message is a new object
@@ -105,11 +146,11 @@ test("fail-safe: malformed messages return the input unchanged", () => {
 });
 
 test("drop:false is a full kill-switch", () => {
-  const msgs = [assistant([{ type: "text", text: "hi" }, thinking(99999), compressCall()]), user("go")];
+  const msgs = [assistant([{ type: "text", text: "hi" }, thinking(99999), compressCall()]), toolResult("c1"), user("go")];
   assert.equal(dropCompressReasoning(msgs, { drop: false }), msgs);
 });
 
-test("no user message at all → nothing touched (all open round)", () => {
+test("no round evidence at all → nothing touched", () => {
   const msgs = [assistant([{ type: "text", text: "hi" }, thinking(4096), compressCall()])];
   assert.equal(dropCompressReasoning(msgs, { drop: true, threshold: 0 }), msgs);
 });
@@ -117,7 +158,7 @@ test("no user message at all → nothing touched (all open round)", () => {
 test("text and toolCall parts (incl. thoughtSignature) survive the drop", () => {
   const call = { ...compressCall(), thoughtSignature: "sig" };
   const out = dropCompressReasoning(
-    [assistant([{ type: "text", text: "keep" }, thinking(4096), call]), user("go")],
+    [assistant([{ type: "text", text: "keep" }, thinking(4096), call]), toolResult("c1"), user("go")],
     { drop: true, threshold: 0 },
   );
   const content = out[0]!.content as unknown[];