mocode-ai 1.1.5 → 1.1.7

package/README.md

@@ -10,6 +10,19 @@ MoCode explores your code, reads/writes/edits files, runs shell commands, and se
 
 ## Architecture
 
+## Engineering discipline
+
+MoCode encodes "how to take coding seriously" into the agent's own behavior, not just into the prompt:
+
+- **Plan → Build → Verify → Fix four-phase discipline** — Injected fresh each turn into `buildBasePrompt`, with per-model-family light adaptation. The agent must restate the task, plan, and acceptance signal before touching anything; changes must pass an automatic validation gate; on failure the agent enters a Fix phase and feeds real command output back as a fresh observation. Evidence: `src/agent/work-discipline.ts` + `evals/work-discipline.ts` (6 assertions). Basis: [`docs/coding-harness-quality-roadmap.md` §4.1 PROMPT-01](docs/coding-harness-quality-roadmap.md).
+- **Pre-Completion Checklist middleware** — `finish`/`stop` is blocked when `mutation > 0 && no tool call && validation !== 'passed'`. Simple read-only tasks deliberately bypass it to avoid noise. Evidence: `src/agent/middleware/checklist.ts` + `evals/checklist.ts` (6 assertions).
+- **Reflective retry + thrash throttling** — Errors are classified into 6 categories (`retry-classifier`); the same tool with the same args ≥3 times appends a hint to switch strategy; failed traces stay in context but receive a targeted reflection prompt instead of a blind retry. Evidence: `src/tools/retry.ts` + `src/agent/retry-classifier.ts` + `evals/retry-classifier.ts` (9 assertions).
+- **`ask_human` as a deliberate de-escalation** — No blind guessing. On whitelisted scenarios (sandbox deviation, ambiguous params, conflicting user instructions) the agent prefers "disclose rather than guess" and will explicitly call `ask_human` to pop a panel for your decision (with a call budget). Evidence: ASK_WHITELIST_SECTION in `src/agent/work-discipline.ts` + `evals/ask-budget.ts` (6 assertions).
+- **Verification cascade V0 → V3 with content fingerprint cache** — Changes pass through file post-conditions → scoped tsc/eslint → targeted unit tests → affected npm scripts, in that order; the first actionable failure stops the cascade; a SHA-256 content cache skips repeated work on unchanged files. Evidence: `src/validators/` plus the VER chapters in the main roadmap.
+- **Five-zone context controls + token self-calibration** — Five independent dials (`autoCompact` / `contextOptimize` / `contextRelprune` / `contextLifecycle` / `contextBudget`), each with its own `MOCODE_*=false` kill switch; even with all five off, observations still age through the lifecycle. Token estimation uses EWMA to self-calibrate against real provider usage rather than trusting the estimator. Evidence: the five modules under `src/context/` plus the `MOCODE_*` switches in `src/config/index.ts`.
+
+## Architecture
+
 MoCode is organized as a layered runtime: the terminal experience drives an autonomous core, the core reaches capabilities through a guarded execution plane, and a persistent intelligence layer keeps long-running work coherent.
 
 <p align="center"><img src="./assets/architecture/system-overview.svg" alt="MoCode layered system architecture" width="100%"></p>

package/README.zh-CN.md

@@ -64,6 +64,19 @@ MoCode 是一个分层的自治运行时：终端交互层驱动 Agent 内核，
 
 ## 为什么用 mocode
 
+## 工程化纪律
+
+mocode 把"如何认真写代码"这件事也写进了 agent 自身的行为准则，而不是只靠 prompt 教：
+
+- **Plan → Build → Verify → Fix 四阶段纪律** — 每轮 prompt 现拼现读注入 `buildBasePrompt`,并按模型家族做轻量适配;agent 必须先复述任务、规划与验收信号,再动手,改动必经自动验证门,失败时进入修复阶段并把真实命令输出当新观察反馈。证据:`src/agent/work-discipline.ts` + `evals/work-discipline.ts`(6 块断言)。依据:[`docs/coding-harness-quality-roadmap.md` §4.1 PROMPT-01](docs/coding-harness-quality-roadmap.md)。
+- **Pre-Completion Checklist 硬关卡** — 在 `mutation > 0 && no tool call && validation !== 'passed'` 三个条件同时成立前,`finish`/`stop` 不会被放行;简单无改动的任务刻意不触发,避免噪音。证据:`src/agent/middleware/checklist.ts` + `evals/checklist.ts`(6 块断言)。
+- **反思式重试 + thrash 节流** — 错误按 6 类分类(`retry-classifier`),同一工具同参数 ≥3 次追加 hint 提醒换策略;失败 trace 留在上下文中但有针对性反思 prompt 注入,而不是盲目重试。证据:`src/tools/retry.ts` + `src/agent/retry-classifier.ts` + `evals/retry-classifier.ts`(9 块断言)。
+- **ask_human 卡点降级** — 不盲猜:遇到 sandbox 偏差、参数二义、用户指令冲突等白名单场景时,agent 倾向"披露而不是瞎猜",必要时显式调用 `ask_human` 弹面板让你拍板(带调用预算)。证据:`src/agent/work-discipline.ts` 的 ASK_WHITELIST_SECTION + `evals/ask-budget.ts`(6 块断言)。
+- **验证瀑布 V0 → V3 + 内容指纹缓存** — 改动先走文件后置条件 → 受限 tsc/eslint → 定向单测 → 受影响 npm 脚本,首个可操作失败立刻停;SHA-256 文件指纹缓存保证未改动文件不重复劳动。证据:`src/validators/` + 主路线图 VER 章节。
+- **五区上下文控制 + token 自校准** — 五个独立开关各自把控一档(`autoCompact` / `contextOptimize` / `contextRelprune` / `contextLifecycle` / `contextBudget`),即使全关观察结果也按生命周期老化;token 估算走 EWMA 自动校准真实 provider 用量,而不是死信估算函数。证据:`src/context/` 五模块 + `src/config/index.ts` 的 `MOCODE_*` 开关。
+
+## 为什么用 mocode
+
 mocode 不是一个套壳聊天框,而是一个能真正动手干活的 agent:
 
 - **自主多步推进** — 一次对话里连续多步:读代码、改代码、跑测试、根据报错再改……agent 自己决定下一步,中途不用你反复催。遇到卡点会调 `ask_human` 弹面板问你(阻塞到回应)。

package/dist/agent/core.js

@@ -201,6 +201,34 @@ export function askHumanBudgetAnnotation(askHumanCountThisTurn, status) {
     // 第 3+ 次(超过预算):强硬提示,鼓励模型停下问自己是否还有意义。
     return `\n\n[ask budget EXCEEDED] This is ask_human call #${askHumanCountThisTurn} this turn (budget = ${ASK_HUMAN_PER_TURN_BUDGET}). Stop asking; choose a default, implement, and disclose the choice in your final reply. Continuing to ask is more harmful than a documented guess.`;
 }
+/** NARR-01: 工具轮旁白(assistant 消息同时带 content + tool_calls)的软预算。
+ * 超出即追加回压提示;未超只 emit trace(可度量,不打扰)。
+ * 单位是 code point 而非字节——中文一字一 point,英文一字母一 point,
+ * 对两种语言都是"一句话大约多长"的直觉量级。 */
+export const NARRATION_CHAR_BUDGET = 120;
+/** 分类工具轮旁白。返回 null 表示这一轮没有旁白(纯工具调用,理想情况)。
+ * hint 仅在超预算时非 null:prompt 里的"工具轮保持静默"是软约束,
+ * 这里给出机制层面的回压,同时让 trace 能统计旁白率。 */
+export function classifyNarration(content, toolCallCount) {
+    const text = content?.trim() ?? '';
+    if (!text)
+        return null;
+    const chars = [...text].length;
+    if (chars <= NARRATION_CHAR_BUDGET) {
+        return { chars, overBudget: false, hint: null };
+    }
+    const plural = toolCallCount === 1 ? '' : 's';
+    return {
+        chars,
+        overBudget: true,
+        hint: `\n\n[narration] Your previous message emitted ${chars} characters of prose alongside ` +
+            `${toolCallCount} tool call${plural} (soft budget = ${NARRATION_CHAR_BUDGET}). ` +
+            'Interstitial commentary costs the user screen space and costs you context on every ' +
+            'later step. For the rest of this turn, emit prose mid-turn ONLY to (a) flag a genuine ' +
+            'decision fork needing user input, (b) disclose an error or risk, or (c) deliver the ' +
+            'final answer once tools are done. Otherwise call the next tool with no preamble.',
+    };
+}
 function pushToolResult(history, tc, output, pruner, lifecycle, _scheduler, runtimeContextState = contextState, succeededOverride) {
     const succeeded = succeededOverride ?? isToolResultSuccess(output);
     const ageAware = config.contextOptimize ? ageAwareStateFor(history) : null;
@@ -243,6 +271,15 @@ export async function runAgentCore(opts) {
     // 预算超过时,工具结果尾部追加 ask budget 提示,不直接拒绝调用(更轻量,
     // 也避免和现有 permission 系统的"拒绝"语义重叠)。
     let askHumanCountThisTurn = 0;
+    // NARR-01: 上一条 assistant 消息(带 tool_calls)超出旁白预算时,待注入的回压提示。
+    // 只挂在本批次的第一条工具结果上并立即清空——重复注入会变成新的噪声源。
+    let pendingNarrationHint = null;
+    /** 取出并清空待注入的旁白回压提示(每批工具只消费一次)。 */
+    const takeNarrationHint = () => {
+        const hint = pendingNarrationHint;
+        pendingNarrationHint = null;
+        return hint;
+    };
     // PROMPT-02: 解析 preCompletionChecklist 选项。undefined = 启用默认 middleware;
     // false = opt-out(checklist 自身调试用);function = 调用方自定义。
     const _checklistMiddleware = createPreCompletionChecklistMiddleware();
@@ -556,6 +593,22 @@ export async function runAgentCore(opts) {
                             function: { name: tc.name, arguments: tc.arguments },
                         })),
                     });
+                    // NARR-01: 这条 assistant 消息同时带正文 + tool_calls,即"工具轮旁白"。
+                    // prompt 里的"工具轮保持静默"只是软约束,这里补上机制层:
+                    //   - 永远 emit trace(旁白率可度量,不再靠肉眼感觉);
+                    //   - 超预算时把回压提示挂到本批第一条工具结果尾部(复用 thrash/ask
+                    //     budget 已验证的注入接缝,不新增消息、不打断工具流)。
+                    const narration = classifyNarration(result.content, result.toolCalls.length);
+                    if (narration) {
+                        emitTrace('narration', {
+                            chars: narration.chars,
+                            toolCalls: result.toolCalls.length,
+                            budget: NARRATION_CHAR_BUDGET,
+                            overBudget: narration.overBudget,
+                            step,
+                        });
+                        pendingNarrationHint = narration.hint;
+                    }
                     // 工具分组执行(保 tool_calls 原顺序)：safe parallel 工具照常并发；连续
                     // resource-locked mutation 先按序完成权限预检，再按 canonical resource lock 启动。
                     // registry 对所有真实资源访问统一持锁，所以不同 Agent 间的 read/write/process 也不会竞态。
@@ -685,7 +738,9 @@ export async function runAgentCore(opts) {
                                     }, tc.id ? { providerToolCallId: tc.id } : {});
                                 }
                                 const askBudget = askHumanBudgetAnnotation(askHumanCountThisTurn, outcome.status);
-                                const finalOutput = askBudget ? `${annotated}${askBudget}` : annotated;
+                                // NARR-01: 旁白回压提示只挂本批第一条结果(takeNarrationHint 自清空)。
+                                const narrationHint = takeNarrationHint();
+                                const finalOutput = `${annotated}${askBudget ?? ''}${narrationHint ?? ''}`;
                                 pushToolResult(history, tc, finalOutput, relprune, lifecycle, scheduler, runtimeContextState, outcome.status === 'success');
                             }
                             hooks.onToolDone?.();
@@ -874,7 +929,9 @@ export async function runAgentCore(opts) {
                                 }, tc.id ? { providerToolCallId: tc.id } : {});
                             }
                             const askBudget = askHumanBudgetAnnotation(askHumanCountThisTurn, outcome.status);
-                            const finalOutput = askBudget ? `${annotated}${askBudget}` : annotated;
+                            // NARR-01: 旁白回压提示只挂本批第一条结果(takeNarrationHint 自清空)。
+                            const narrationHint = takeNarrationHint();
+                            const finalOutput = `${annotated}${askBudget ?? ''}${narrationHint ?? ''}`;
                             pushToolResult(history, tc, finalOutput, relprune, lifecycle, scheduler, runtimeContextState, outcome.status === 'success');
                             const invalidatedFiles = [...new Set([
                                     ...(outcome.changedFiles ?? []),

package/dist/agent/work-discipline.js

@@ -33,30 +33,30 @@ export function inferModelFamily(model) {
  * 4 阶段核心纪律(英文)。4 个 model family 共用此文本,只在首句与标题
  * 标签上做轻量变体。保持短小，详细的完成检查由动态 checklist 按需注入。
  */
-const CORE_SECTION = `## Working discipline — coding tasks (Build-and-Self-Verify)
-
-Treat "verification" as a first-class part of the task, not an afterthought. Use the smallest evidence-driven loop below.
-
-### Phase 1 — Plan & Discover
-- Open with a one-sentence restatement of your interpretation of the request; if a materially different reading exists, name it briefly before proceeding. This catches misunderstanding before any work is wasted.
-- State the goal and a concrete acceptance signal, then inspect the relevant code before changing it.
-- Ask only when an unresolved choice is high-impact or user-owned; otherwise follow repository evidence and proceed.
-
-### Phase 2 — Build
-- Make the smallest coherent change; avoid unrelated refactors.
-- Add or update a focused test when behavior changes and the project has an applicable test suite.
-- Re-read only when a dependent edit needs fresh exact content or state may be stale.
-
-### Phase 3 — Verify
-- Run the smallest executable check that proves the requested behavior, then read its complete result.
-- Compare evidence with the user's request, not merely with the diff.
-
-### Phase 4 — Fix
-- Diagnose the root cause, make a focused correction, and rerun the relevant check.
-- After two identical failures, change the approach instead of repeating the same call.
-
-**Hard rule (non-negotiable):** "I read the code and it looks right" is not a completion signal. Report the verification performed, or state clearly why it could not be run.
-
+const CORE_SECTION = `## Working discipline — coding tasks (Build-and-Self-Verify)
+
+Treat "verification" as a first-class part of the task, not an afterthought. Use the smallest evidence-driven loop below.
+
+### Phase 1 — Plan & Discover
+- Restate the request in one sentence ONLY when it admits two or more materially different readings; name the reading you picked and move on. An unambiguous request gets no restatement — start working. This is the single exception to staying silent during tool-calling turns.
+- Settle on the goal and a concrete acceptance signal before inspecting the relevant code; keep them internal unless the user has to weigh in.
+- Ask only when an unresolved choice is high-impact or user-owned; otherwise follow repository evidence and proceed.
+
+### Phase 2 — Build
+- Make the smallest coherent change; avoid unrelated refactors.
+- Add or update a focused test when behavior changes and the project has an applicable test suite.
+- Re-read only when a dependent edit needs fresh exact content or state may be stale.
+
+### Phase 3 — Verify
+- Run the smallest executable check that proves the requested behavior, then read its complete result.
+- Compare evidence with the user's request, not merely with the diff.
+
+### Phase 4 — Fix
+- Diagnose the root cause, make a focused correction, and rerun the relevant check.
+- After two identical failures, change the approach instead of repeating the same call.
+
+**Hard rule (non-negotiable):** "I read the code and it looks right" is not a completion signal. Report the verification performed, or state clearly why it could not be run.
+
 **Hard rule (non-negotiable):** Never invent file paths, APIs, config keys, flags, or behavior. Every claim about the codebase must trace to tool output in this conversation; explicitly label anything you have not verified as an assumption.`;
 /**
  * 把核心段适配到指定 model family:只替换首行(语序 / 强动词),段标题
@@ -68,16 +68,16 @@ function adapt(_model, opener) {
     return CORE_SECTION.replace('Treat "verification" as a first-class part of the task, not an afterthought.', opener);
 }
 /** ASK-01: only user-owned, high-impact choices should interrupt autonomous execution. */
-const ASK_WHITELIST_SECTION = `## When to ask instead of guess
-
-Call \`ask_human\` before coding only when repository evidence cannot resolve a user-owned, high-impact choice:
-1. irreversible deletion, migration, security, permission, or external side effect;
-2. public API compatibility (keep, deprecate, rename, or remove);
-3. multiple reasonable options that materially change product behavior;
-4. the request itself admits two or more materially different readings that lead to different deliverables (do not silently pick one and guess).
-
-For naming, implementation detail, and verification commands, follow repository precedent and choose the safest reversible default. Disclose any consequential assumption.
-
+const ASK_WHITELIST_SECTION = `## When to ask instead of guess
+
+Call \`ask_human\` before coding only when repository evidence cannot resolve a user-owned, high-impact choice:
+1. irreversible deletion, migration, security, permission, or external side effect;
+2. public API compatibility (keep, deprecate, rename, or remove);
+3. multiple reasonable options that materially change product behavior;
+4. the request itself admits two or more materially different readings that lead to different deliverables (do not silently pick one and guess).
+
+For naming, implementation detail, and verification commands, follow repository precedent and choose the safest reversible default. Disclose any consequential assumption.
+
 Budget: at most 2 \`ask_human\` calls per turn. Beyond that, use the safest reversible default and disclose it in the final reply.`;
 /**
  * 拼出纪律段 + ASK-01 卡点白名单。返回完整段(两段用 \`\\n\\n\` 隔开);

package/dist/config/index.js

@@ -200,6 +200,7 @@ ${buildWorkDisciplineSection(inferModelFamily(config.model))}
 ${buildCodegraphSection()}
 
 ## Tool use
+- During tool-calling turns, stay silent unless something important enough must reach the user — otherwise just call the tool and let it run.
 - Go directly to a known path or symbol; use discovery tools only when the location is unknown.
 - Edit against a FRESH read: before any edit_file/write_file, call read_file on the exact path and copy both its latest hash and the exact target text. Never reconstruct old_string from a grep/summary/diff — those lose whitespace and indentation and cause edit failures.
 - A read_file hash from before a compaction, session resume, edit conflict, or external change is STALE and will be rejected — re-read rather than reuse an old hash.

package/dist/llm/index.js

@@ -141,8 +141,11 @@ function logRetry(attempt, err, waitMs) {
     const e = err;
     const tag = e.status ? `HTTP ${e.status}` : e.name || 'Error';
     const msg = e.message ?? '未知错误';
-    // stderr 而非 stdout —— 不污染流式正文;行首换行防止黏在上一行尾巴。
-    process.stderr.write(`\n[llm] 第 ${attempt}/${RETRY_MAX_ATTEMPTS} 次失败(${tag}: ${msg}),${(waitMs / 1000).toFixed(1)}s 后重试…\n`);
+    // 走 console.error 而非 process.stderr.write:TUI active 时 layout.installConsoleGuard
+    // 已把 console.* 劫持到 contentWrite(内容区),错误会落在 agent 输出区,不污染输入框;
+    // 非 TTY(管道 / CI / 启动早期 TUI 未启)降级走原生 console.error → stderr,行为与改造前一致。
+    // 不写前导 \n —— contentWrite 由续写位管位置,前置换行会留空行;结尾 \n 由劫持逻辑补。
+    console.error(`[llm] 第 ${attempt}/${RETRY_MAX_ATTEMPTS} 次失败(${tag}: ${msg}),${(waitMs / 1000).toFixed(1)}s 后重试…`);
 }
 let createImplOverride = null;
 /** 仅供单测用:覆盖 chat() 内部实际调用的 create 桩。生产代码不要碰。 */

package/dist/repl/index.js

@@ -342,6 +342,7 @@ const emitter = stdin;
 // ── 运行态交互(typeahead 输入 + 滚动回看 + Ctrl+C 中断)──
 // 只在 await runAgent() 期间挂载;/resume /rollback /compact 等走 askLine(cooked readline)的分支不挂(避免抢 stdin)。
 let runningInput = ''; // 运行中已打字缓冲(单行;agent 结束后预填下一轮 INPUT 态)
+let runningCursor = 0; // 缓冲内光标字符索引(0..len);运行态支持任意位置编辑,与空闲态一致
 let runningPlaceholder = '';
 let currentAbort = null;
 let pendingPrefill = null; // /rollback 选中后预填的 user 输入(下轮 INPUT 态消费)
@@ -391,7 +392,8 @@ function onRunningKey(_str, key) {
     if (key.ctrl && key.name === 'c') {
         if (runningInput.length > 0) {
             runningInput = '';
-            layout.paintRunningInputEcho(runningInput, runningPlaceholder);
+            runningCursor = 0;
+            layout.paintRunningInput(runningInput, runningCursor, runningPlaceholder);
         }
         else if (currentAbort && !currentAbort.signal.aborted) {
             appendCurrentSessionRuntimeEvent('abort', { phase: 'requested', source: 'keyboard' });
@@ -400,16 +402,46 @@ function onRunningKey(_str, key) {
         return;
     }
     const s = key.sequence ?? '';
+    // 光标移动(单行 typeahead,光标可任意位置,与空闲态一致)
+    if (key.name === 'left') {
+        runningCursor = Math.max(0, runningCursor - 1);
+        layout.paintRunningInput(runningInput, runningCursor, runningPlaceholder);
+        return;
+    }
+    if (key.name === 'right') {
+        runningCursor = Math.min(runningInput.length, runningCursor + 1);
+        layout.paintRunningInput(runningInput, runningCursor, runningPlaceholder);
+        return;
+    }
+    if (key.name === 'home' || (key.ctrl && key.name === 'a')) {
+        runningCursor = 0;
+        layout.paintRunningInput(runningInput, runningCursor, runningPlaceholder);
+        return;
+    }
+    if (key.name === 'end' || (key.ctrl && key.name === 'e')) {
+        runningCursor = runningInput.length;
+        layout.paintRunningInput(runningInput, runningCursor, runningPlaceholder);
+        return;
+    }
     if (key.name === 'backspace') {
-        if (runningInput.length > 0) {
-            runningInput = runningInput.slice(0, -1);
-            layout.paintRunningInputEcho(runningInput, runningPlaceholder);
+        if (runningCursor > 0) {
+            runningInput = runningInput.slice(0, runningCursor - 1) + runningInput.slice(runningCursor);
+            runningCursor--;
+            layout.paintRunningInput(runningInput, runningCursor, runningPlaceholder);
+        }
+        return;
+    }
+    if (key.name === 'delete') {
+        if (runningCursor < runningInput.length) {
+            runningInput = runningInput.slice(0, runningCursor) + runningInput.slice(runningCursor + 1);
+            layout.paintRunningInput(runningInput, runningCursor, runningPlaceholder);
         }
         return;
     }
     if (key.name === 'escape') {
         runningInput = '';
-        layout.paintRunningInputEcho(runningInput, runningPlaceholder);
+        runningCursor = 0;
+        layout.paintRunningInput(runningInput, runningCursor, runningPlaceholder);
         return;
     }
     // Enter / Ctrl+J:运行中 no-op(单行 typeahead;agent 结束后预填,用户在 INPUT 态按 Enter 提交)
@@ -418,21 +450,25 @@ function onRunningKey(_str, key) {
         (key.ctrl && key.name === 'j')) {
         return;
     }
-    // 可打印字符(>= 空格,非 ctrl/meta)→ 追加 + dim 回显
+    // 可打印字符(>= 空格,非 ctrl/meta)→ 光标处插入 + 非 dim 回显(与空闲态一致)
     if (s && s >= ' ' && !key.ctrl && !key.meta) {
-        runningInput += s;
-        layout.paintRunningInputEcho(runningInput, runningPlaceholder);
+        runningInput = runningInput.slice(0, runningCursor) + s + runningInput.slice(runningCursor);
+        runningCursor += s.length;
+        layout.paintRunningInput(runningInput, runningCursor, runningPlaceholder);
     }
 }
-/** 鼠标右键单击输入框(未拖动)时 layout 读剪贴板后回调:追加到 typeahead 缓冲(简单粘贴,不分长短)。 */
+/** 鼠标右键单击输入框(未拖动)时 layout 读剪贴板后回调:在光标处插入 typeahead 缓冲(单行,换行折为空格)。 */
 function onRunningMousePaste(text) {
-    runningInput += text.replace(/\r\n/g, '\n').replace(/\r/g, '\n');
-    layout.paintRunningInputEcho(runningInput, runningPlaceholder);
+    const flat = text.replace(/[\r\n]+/g, ' ');
+    runningInput = runningInput.slice(0, runningCursor) + flat + runningInput.slice(runningCursor);
+    runningCursor += flat.length;
+    layout.paintRunningInput(runningInput, runningCursor, runningPlaceholder);
 }
 /** 进入运行态:挂 keypress 监听 + raw mode + 新建 abort 控制器,返回其 signal。在 await runAgent 前、enterRunningMode 后调。 */
 function startRunningListener(placeholder) {
     runningPlaceholder = placeholder;
     runningInput = '';
+    runningCursor = 0;
     emitKeypressEvents(stdin); // 幂等:首轮 prompt 已永久挂解析器,这里防御性再调
     try {
         stdin.setRawMode(true);
@@ -762,6 +798,7 @@ export async function startRepl(initialHistory, sessionId, sandboxRootOverride,
         renderHistory(history);
         // 强制回尾:同 /resume 命令,renderHistory 展开详情会设 scrollOffset>0,需复位避免闪烁。
         layout.resetScroll();
+        layout.repaintViewport();
     }
     else {
         layout.writeBanner(bannerLines(banner()));
@@ -1047,11 +1084,12 @@ export async function startRepl(initialHistory, sessionId, sandboxRootOverride,
         lastTurnUsage = undefined; // 续接:旧会话的 token 累计已无意义,清空等下轮覆写
         layout.clearContent();
         renderHistory(history);
-        // 强制回尾:renderHistory 展开 mutation 工具详情会经 contentInsertAfter 设置 scrollOffset>0;
-        // 若不复位，后续 showLiveBatch 在冻结视口下 repaintViewport 会导致工具信息闪烁/滚动消失。
-        layout.resetScroll();
         // 末尾 \n\n:与后续用户消息(❯ bubble)之间空一行。
         layout.contentWrite(`${ui.dim}${t('repl.resumed', { id: loaded.id })}${ui.reset}\n\n`);
+        // 强制回尾:renderHistory 展开 mutation 工具详情会经 contentInsertAfter 设置 scrollOffset>0;
+        // 先把"已续接会话"提示写入缓冲，再统一回尾重画，避免切换后视口停留在历史顶部。
+        layout.resetScroll();
+        layout.repaintViewport();
     }
     let hasSubmittedInput = false;
     while (true) {

package/dist/ui/intervention.js

@@ -205,7 +205,6 @@ export async function promptIntervention(req) {
             cursorLine: 0,
             cursorCol: 0,
             menu: null,
-            dim: true,
         });
         if (layout.isScrolled())
             layout.resetScroll();

package/dist/ui/layout.js

@@ -127,8 +127,11 @@ const esc = {
     //    落在内容区 → 复制当前选区(若有)到剪贴板(clipboard.ts),静默不弹提示。
     //  - 滚轮报表(button&64)转 scrollBy。
     // 代价:终端原生框选被鼠标捕获接管——想用终端原生选区可按住 Shift(多数终端放行)。
-    mouseOn: '\x1B[?1000h\x1B[?1002h\x1B[?1006h',
-    mouseOff: '\x1B[?1006l\x1B[?1002l\x1B[?1000l', // 关:反序
+    // 1007l=关闭 Alternate Scroll Mode:VSCode xterm.js 进 alt screen(?1049)后默认开 1007,
+    // 滚轮发 Up/Down 方向键而非 SGR 鼠标报表;不关的话滚轮会绕过 mouse.swallow 直接触发
+    // prompt 的历史导航(recallPrevious/recallNext)→ "输入框打字时滚轮跳到上一条历史消息"。
+    mouseOn: '\x1B[?1007l\x1B[?1000h\x1B[?1002h\x1B[?1006h',
+    mouseOff: '\x1B[?1006l\x1B[?1002l\x1B[?1000l\x1B[?1007h', // 关:反序;末尾恢复 1007(交还终端默认)
     cursorShow: '\x1B[?25h',
     cursorHide: '\x1B[?25l',
     clearLine: '\x1B[2K',
@@ -172,13 +175,23 @@ export function setRegion(fh) {
         contentRow = g.contentBottom; // 底栏撑高挤掉内容:钳到新区底
     return g;
 }
-/** 运行态真光标(隐藏)的归位点 = 输入框光标位:行 = 动态 contentBottom+4(resize 安全),列对齐 renderDimInputRow 的假光标(空→3,有字→❯+截断文本宽+1)。供 IME 锚定。 */
+/** 运行态真光标(隐藏)的归位点 = 输入框光标位:行 = 动态 contentBottom+4(resize 安全)。
+ *  非 dim 运行态(与空闲态同色、可任意位置编辑)→ 归当前编辑位,供 IME 锚定气泡到光标处;
+ *  dim 占位(空 + placeholder)兼容态→ 归输入框起点。供 IME 锚定。 */
 function runningCaretPos() {
     const g = getGeo();
+    const row = g.contentBottom + 4;
+    if (lastView && !lastView.dim) {
+        const promptW = displayWidth(lastView.prompt);
+        const line = lastView.lines[lastView.cursorLine] ?? '';
+        const before = line.slice(0, lastView.cursorCol);
+        const col = Math.min(g.cols, promptW + displayWidth(before) + 1);
+        return { row, col };
+    }
     const text = lastView?.dim ? lastView.lines[0] ?? '' : '';
     const contentW = Math.max(0, g.cols - 3); // ❯ =2 + 光标=1
     const w = displayWidth(truncateDisplayHead(text, contentW));
-    return { row: g.contentBottom + 4, col: Math.min(g.cols, 2 + w + 1) };
+    return { row, col: Math.min(g.cols, 2 + w + 1) };
 }
 export function contentMode() {
     if (!active)
@@ -529,6 +542,11 @@ export function clearContent() {
     // /theme、/clear、/resume 等命令 clearContent 后紧接 writeBanner 的场景均依赖此重置。
     bannerH = 0;
     bannerRows = [];
+    // 清内容区必须同时作废旧菜单擦除坐标:picker(/resume /rollback /theme)把菜单画在内容区底部,
+    // 菜单行号缓存在 lastMenuStartRow/lastMenuRows;若不清零,后续 paintInput 会按旧坐标“擦菜单”,
+    // 把刚 renderHistory/contentWrite 写好的内容(如“已续接会话”提示)清掉,导致用户要滚动一下才刷新。
+    lastMenuStartRow = 0;
+    lastMenuRows = 0;
     content.reset();
     notifyContentReset(); // batch 渲染器同步重置(batch 摘要行索引全部失效)
     stdout.write(esc.home);
@@ -1842,34 +1860,67 @@ function renderDimInputRow(prompt, text, placeholder, cols) {
     return `${ui.dim}${prompt}${p}${ui.reset}`;
 }
 /**
- * 运行态 typeahead 回显:定向写输入行(底栏输入框),把 dim 占位换成已打字文本 + 反白块状光标(无打字时光标在起点)。
- * 只 cup 输入行 + clearLine + dim 文本 + 归位——不调 setRegion(运行中禁多行,避免 DECSTBM 抖动)、
- * 不重画状态行/contentBottom、不用 ED。同步 lastView 为 dim 视图(text 与 placeholder 拆开存),使 scrollBy/resize
- * 的 repaint 仍显当前回显 + 光标。真光标归续写位(滚动回看时归内容区底)——不入输入框,假光标已画在行内。
+ * 单行运行态滑窗:以光标为中心,向左右扩展填满 contentW-1(留 1 cell 给光标),
+ * 返回可见子串与光标在子串内的显示列。保证光标恒可见,且不软折行(运行态输入框恒单行,
+ * 不触发 setRegion/ED,避免流式期间底栏抖动——见 scripts/check-layout.ts 断言)。
+ */
+function windowSingleLine(text, cursor, contentW) {
+    const n = text.length;
+    if (n === 0)
+        return { shown: '', curDisp: 0 };
+    const totalDisp = displayWidth(text);
+    if (totalDisp <= contentW)
+        return { shown: text, curDisp: displayWidth(text.slice(0, cursor)) };
+    let i = cursor;
+    let j = cursor;
+    let w = 0;
+    const cw = (idx) => charWidth(text.codePointAt(idx) ?? 0);
+    while (i > 0 && w + cw(i - 1) <= contentW - 1) {
+        w += cw(i - 1);
+        i--;
+    }
+    while (j < n && w + cw(j) <= contentW - 1) {
+        w += cw(j);
+        j++;
+    }
+    return { shown: text.slice(i, j), curDisp: displayWidth(text.slice(i, cursor)) };
+}
+/**
+ * 运行态 typeahead 回显:定向写输入行(底栏输入框),非 dim(与空闲态同色)、光标可任意位置、单行。
+ * 只 cup 输入行 + clearLine + 文本 + cup 真光标到编辑位——不调 setRegion(运行中禁多行,避免 DECSTBM 抖动)、
+ * 不重画状态行/contentBottom、不用 ED。同步 lastView 为运行视图(非 dim),使 scrollBy/resize 的 repaint
+ * 仍显当前回显 + 光标。空且带 placeholder 时显 dim ghost(提示 agent 状态;非用户输入,不违反"打字不变灰")。
+ * 选区高亮由 paintInput 路径(paint 时鼠标拖选)承担,此处轻量不重复实现。
  */
-export function paintRunningInputEcho(text, placeholder) {
+export function paintRunningInput(text, cursor, placeholder) {
     if (!active || !base)
         return;
     const g = getGeo();
-    const inputRow = g.contentBottom + 4; // 运行态 footerH 恒 5:虚拟空(+1)+状态(+2)+上线(+3)+输入行(+4);下线在 rows
-    // 先同步 lastView(供 runningCaretPos 算真光标位 = 新文本末尾,与假光标同位)
+    const inputRow = g.contentBottom + 4; // 运行态 footerH 恒 6(单行,无 setRegion)
+    const promptW = displayWidth('❯ ');
+    const contentW = Math.max(1, g.cols - promptW);
+    let outLine;
+    let curCol; // 输入框内的显示列(不含 prompt)
+    if (text.length === 0 && placeholder) {
+        outLine = `${ui.dim}❯ ${truncateDisplay(placeholder, contentW)}${ui.reset}`;
+        curCol = 0;
+    }
+    else {
+        const { shown, curDisp } = windowSingleLine(text, cursor, contentW);
+        outLine = `❯ ${shown}`; // 正常色 —— 运行态打字与空闲态一致
+        curCol = curDisp;
+    }
     lastView = {
         prompt: '❯ ',
         lines: [text],
         placeholder,
         cursorLine: 0,
-        cursorCol: 0,
+        cursorCol: cursor,
         menu: null,
-        dim: true,
+        // 不置 dim:运行态输入框与空闲态同色、可任意位置编辑
     };
-    // 单次 write:cup 输入行 + clearLine + dim 文本/假光标 + cup 真光标到输入框(供 IME 锚定)。
-    // 滚动态也归输入框——旧设计滚动态归 contentBottom,致 IME 候选气泡锚到内容区底白块
-    // (conhost IME 不跟随 cup 后续移动,须让打字前光标已在输入框);拆两次 write 会暂留 contentBottom 显白块。
-    const p = runningCaretPos();
-    stdout.write(cup(inputRow, 1) +
-        esc.clearLine +
-        renderDimInputRow('❯ ', text, placeholder, g.cols) +
-        cup(p.row, p.col));
+    const cursorCol = Math.min(g.cols, promptW + curCol + 1);
+    stdout.write(cup(inputRow, 1) + esc.clearLine + outLine + cup(inputRow, cursorCol));
 }
 /** 重画当前视图(resize / 内部用)。 */
 export function repaint() {
@@ -1913,7 +1964,7 @@ export function enterInputMode(status = t('repl.idle')) {
         stdout.write(esc.cursorShow); // 回 INPUT 态:显真光标(运行态藏了)
     }
 }
-/** 进入运行态:底栏输入行改 dim 占位,光标回续写位。footerH 恒 6(虚拟空+spinner行+上线+输入+下线+model行)。新轮回尾(确保新内容可见)。 */
+/** 进入运行态:底栏输入行与空闲态同色(非 dim)、可任意位置编辑,cursor 留输入框。footerH 恒 6(虚拟空+spinner行+上线+输入+下线+model行)。新轮回尾(确保新内容可见)。 */
 export function enterRunningMode(status, placeholder) {
     mode = 'running';
     statusText = status;
@@ -1930,7 +1981,6 @@ export function enterRunningMode(status, placeholder) {
             cursorLine: 0,
             cursorCol: 0,
             menu: null,
-            dim: true,
         });
         startTurnTimer(); // 续刷状态行走时(流式期间 spinner 停转,由它兜底)
         contentMode();

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mocode-ai",
-  "version": "1.1.5",
+  "version": "1.1.7",
   "description": "终端编码 agent:LLM + tool-call 循环 + 流式输出(含思考)+ 16 个工具,接任意 OpenAI 兼容后端。",
   "type": "module",
   "bin": {