mocode-ai 1.7.10 → 1.8.0

package/dist/agent/index.js

@@ -5,7 +5,7 @@
 import { ui } from '../ui/theme.js';
 import { Spinner } from '../ui/spinner.js';
 import { summarizeToolCall, summarizeToolResult, truncateDisplay, fmtElapsed } from '../ui/render.js';
-import { renderFileChange } from '../ui/diff.js';
+import { renderFileChange, renderMultiFileChange } from '../ui/diff.js';
 import * as layout from '../ui/layout.js';
 import * as batch from '../ui/batch.js';
 import { config } from '../config/index.js';
@@ -150,7 +150,7 @@ function writeToolHeader(tc, toolRuntime) {
 }
 /** 渲染工具结果:mutation 成功走 diff 块(行号 + 语法高亮);其余走一行 preview。
  *  同 writeToolHeader,改为累积到 BatchRenderer(只缓存字符串,不写屏)。 */
-function writeToolResult(tc, output, parsed, preWriteOld, editStartLine, toolRuntime) {
+function writeToolResult(tc, output, parsed, preWriteOld, editStartLine, toolRuntime, multiHunks) {
     // 优先按 tool_call id 反查所属批（并行 / sub-agent 组时 currentBatchId 已不是它）。
     const batchId = batch.batchIdForCall(tc.id) ?? currentBatchId;
     if (!batchId)
@@ -173,19 +173,55 @@ function writeToolResult(tc, output, parsed, preWriteOld, editStartLine, toolRun
     }
     let diff = null;
     if ((tc.name === 'edit_file' || tc.name === 'write_file') && parsed && !isToolErrorOutput(output)) {
-        const oldText = tc.name === 'edit_file' ? String(parsed.old_string ?? '') : preWriteOld;
-        const newText = String((tc.name === 'edit_file' ? parsed.new_string : parsed.content) ?? '');
-        diff = renderFileChange({
-            path: String(parsed.path ?? ''),
-            kind: tc.name === 'edit_file' ? 'edit' : 'write',
-            oldStr: tc.name === 'edit_file' ? String(parsed.old_string ?? '') : preWriteOld,
-            newStr: String((tc.name === 'edit_file' ? parsed.new_string : parsed.content) ?? ''),
-            startLine: tc.name === 'edit_file' ? editStartLine : 1,
-        });
+        const filePath = String(parsed.path ?? '');
+        const isEdit = tc.name === 'edit_file';
+        // 多替换模式:每个 old/new 段一个 hunk,渲染成一个 Update 块内的多个 hunk。
+        const multi = isEdit && multiHunks && multiHunks.length > 0 ? multiHunks : null;
+        const oldText = isEdit
+            ? multi
+                ? multi.map((h) => h.oldStr).join('\n')
+                : String(parsed.old_string ?? '')
+            : preWriteOld;
+        const newText = isEdit
+            ? multi
+                ? multi.map((h) => h.newStr).join('\n')
+                : String(parsed.new_string ?? '')
+            : String(parsed.content ?? '');
+        if (multi) {
+            diff = renderMultiFileChange({
+                path: filePath,
+                hunks: multi.map((h) => ({ oldStr: h.oldStr, newStr: h.newStr, startLine: h.startLine })),
+            });
+            // 渲染器对超大 diff 返回 null(无法做 LCS)时退回工具结果摘要。
+            if (!diff)
+                diff = renderFileChange({
+                    path: filePath,
+                    kind: 'edit',
+                    oldStr: oldText,
+                    newStr: newText,
+                    startLine: editStartLine,
+                });
+        }
+        else {
+            diff = renderFileChange({
+                path: filePath,
+                kind: isEdit ? 'edit' : 'write',
+                oldStr: isEdit ? String(parsed.old_string ?? '') : preWriteOld,
+                newStr: String((isEdit ? parsed.new_string : parsed.content) ?? ''),
+                startLine: isEdit ? editStartLine : 1,
+            });
+        }
+        // 回合改动统计:多替换按各 hunk 独立计算后累加(拼接不相邻段再 diff 会把跨段行误算成增删)。
+        const delta = multi
+            ? multi.reduce((acc, h) => {
+                const d = lineDelta(h.oldStr, h.newStr);
+                return { added: acc.added + d.added, removed: acc.removed + d.removed };
+            }, { added: 0, removed: 0 })
+            : lineDelta(oldText, newText);
         turnFileChanges.push({
-            path: String(parsed.path ?? ''),
+            path: filePath,
             kind: tc.name === 'write_file' && preWriteOld == null ? 'A' : 'M',
-            ...lineDelta(oldText, newText),
+            ...delta,
         });
     }
     const preview = diff ? '' : summarizeToolResult(tc.name, output);
@@ -406,7 +442,7 @@ runtimeOrContext = defaultRuntime) {
         onToolStart: (name) => spinner.start(t('agent.executing', { tool: name })),
         onToolOutput: (tc, chunk) => batch.recordOutput(tc.id, chunk, layout),
         onToolDone: () => spinner.stop(),
-        onToolResult: (tc, output, parsed, preWriteOld, editStartLine) => writeToolResult(tc, output, parsed, preWriteOld, editStartLine, runtime.context.toolRuntime),
+        onToolResult: (tc, output, parsed, preWriteOld, editStartLine, multiHunks) => writeToolResult(tc, output, parsed, preWriteOld, editStartLine, runtime.context.toolRuntime, multiHunks),
         onToolBatchEnd: () => {
             // 一次工具轮次结束不再切 UI batch；下一轮若仍无正文，继续复用 currentBatchId。
         },

package/dist/agent/run-coordinator.js

@@ -290,7 +290,7 @@ export async function runAgentCoreLegacy(opts, historyManager, stages) {
                                             break;
                                         }
                                         case 'result':
-                                            hooks.onToolResult?.(event.call, event.outcome.output, event.parsed, event.diff.preWriteOld, event.diff.editStartLine);
+                                            hooks.onToolResult?.(event.call, event.outcome.output, event.parsed, event.diff.preWriteOld, event.diff.editStartLine, event.diff.multiHunks);
                                             if (event.call.name === 'ask_human' && event.outcome.status === 'success') {
                                                 askHumanCountThisTurn += 1;
                                                 emitTrace('ask_human_call', {
@@ -326,6 +326,7 @@ export async function runAgentCoreLegacy(opts, historyManager, stages) {
                                     delegation: delegationForOrchestrator,
                                     readDedup,
                                     argumentErrorHint: (name) => argumentErrorHint(name, runtimeContextState),
+                                    ...(result.finishReason ? { finishReason: result.finishReason } : {}),
                                     onToolOutput: (call, chunk) => hooks.onToolOutput?.(call, chunk),
                                     ...(opts.toolPolicy
                                         ? {
@@ -401,6 +402,7 @@ export async function runAgentCoreLegacy(opts, historyManager, stages) {
                                                 allowedToolNames: currentAllowedToolNames(),
                                                 delegation: delegationForOrchestrator(),
                                                 readDedup,
+                                                ...(result.finishReason ? { finishReason: result.finishReason } : {}),
                                             });
                                         }
                                     }
@@ -572,6 +574,7 @@ export async function runAgentCoreLegacy(opts, historyManager, stages) {
                                             allowedToolNames: currentAllowedToolNames(),
                                             delegation: delegationForOrchestrator(),
                                             readDedup,
+                                            ...(result.finishReason ? { finishReason: result.finishReason } : {}),
                                         }));
                                         // 完成即发:usage/onToolOutcome/trace 在各 outcome 落地瞬间发出(1s 的
                                         // grep 不再被 60s 的 web_fetch 卡住可见性);history 回灌严格按原序。
@@ -671,6 +674,7 @@ export async function runAgentCoreLegacy(opts, historyManager, stages) {
                                                         delegation: delegationForOrchestrator(),
                                                         readDedup,
                                                         ...(hint ? { argumentErrorHint: hint } : {}),
+                                                        ...(result.finishReason ? { finishReason: result.finishReason } : {}),
                                                         onLockAcquired: (lockedArgs) => {
                                                             entry.diff = readDiffContext(entry.tc, lockedArgs, ctx.jailResolve);
                                                         },
@@ -685,7 +689,7 @@ export async function runAgentCoreLegacy(opts, historyManager, stages) {
                                         for (let offset = 0; offset < entries.length; offset++) {
                                             const entry = entries[offset];
                                             const outcome = await outcomes[offset];
-                                            hooks.onToolResult?.(entry.tc, outcome.output, entry.denied ? null : entry.parsed, entry.diff.preWriteOld, entry.diff.editStartLine);
+                                            hooks.onToolResult?.(entry.tc, outcome.output, entry.denied ? null : entry.parsed, entry.diff.preWriteOld, entry.diff.editStartLine, entry.diff.multiHunks);
                                             pushToolResult(history, entry.tc, outcome.output, relprune, lifecycle, scheduler, runtimeContextState, outcome.status === 'success');
                                             const invalidatedFiles = [
                                                 ...new Set([...(outcome.changedFiles ?? []), ...(outcome.staleFiles ?? [])]),
@@ -760,6 +764,7 @@ export async function runAgentCoreLegacy(opts, historyManager, stages) {
                                                 delegation: delegationForOrchestrator(),
                                                 readDedup,
                                                 ...(hint ? { argumentErrorHint: hint } : {}),
+                                                ...(result.finishReason ? { finishReason: result.finishReason } : {}),
                                                 onLockAcquired: (lockedArgs) => {
                                                     entry.diff = readDiffContext(entry.tc, lockedArgs, ctx.jailResolve);
                                                 },
@@ -776,7 +781,7 @@ export async function runAgentCoreLegacy(opts, historyManager, stages) {
                                         for (let k = 0; k < entries.length; k++) {
                                             const entry = entries[k];
                                             const outcome = await started[k];
-                                            hooks.onToolResult?.(entry.tc, outcome.output, entry.denied ? null : entry.parsed, entry.diff.preWriteOld, entry.diff.editStartLine);
+                                            hooks.onToolResult?.(entry.tc, outcome.output, entry.denied ? null : entry.parsed, entry.diff.preWriteOld, entry.diff.editStartLine, entry.diff.multiHunks);
                                             pushToolResult(history, entry.tc, outcome.output, relprune, lifecycle, scheduler, runtimeContextState, outcome.status === 'success');
                                             const invalidatedFiles = [
                                                 ...new Set([...(outcome.changedFiles ?? []), ...(outcome.staleFiles ?? [])]),
@@ -856,6 +861,7 @@ export async function runAgentCoreLegacy(opts, historyManager, stages) {
                                             delegation: delegationForOrchestrator(),
                                             readDedup,
                                             ...(serialHint ? { argumentErrorHint: serialHint } : {}),
+                                            ...(result.finishReason ? { finishReason: result.finishReason } : {}),
                                             onLockAcquired: (lockedArgs) => {
                                                 if (mutationParsed)
                                                     diff = readDiffContext(tc, lockedArgs, ctx.jailResolve);
@@ -866,7 +872,7 @@ export async function runAgentCoreLegacy(opts, historyManager, stages) {
                                         traceToolEnd(tc, i, outcome);
                                         const output = outcome.output;
                                         hooks.onToolDone?.();
-                                        hooks.onToolResult?.(tc, output, mutationParsed, diff.preWriteOld, diff.editStartLine);
+                                        hooks.onToolResult?.(tc, output, mutationParsed, diff.preWriteOld, diff.editStartLine, diff.multiHunks);
                                         if (tc.name === 'ask_human' && outcome.status === 'success') {
                                             askHumanCountThisTurn += 1;
                                             emitTrace('ask_human_call', {

package/dist/agent/stages/tool-dispatcher.js

@@ -49,6 +49,7 @@ class LegacyCompatibleToolDispatcher {
             delegation: request.delegation(),
             readDedup: request.readDedup,
             ...(hint ? { argumentErrorHint: hint } : {}),
+            ...(request.finishReason ? { finishReason: request.finishReason } : {}),
             ...(onLockAcquired ? { onLockAcquired } : {}),
         });
         const executionEvents = (index, parsed, outcome) => {

package/dist/agent/tool-helpers.js

@@ -13,6 +13,8 @@ import { defaultToolRuntime } from '../tools/registry.js';
 import { jailResolve } from '../sandbox/index.js';
 import { contextState } from '../session/compact.js';
 import { capToolResultForHistory } from '../session/compact.js';
+import { getActiveSessionStore } from '../session/store.js';
+import { getCurrentSessionId } from '../session/state.js';
 import { recordArtifact, knownEditTargets } from '../context/index.js';
 import { isToolResultSuccess } from '../context/utils.js';
 import { appendGuiAction } from '../session/gui-actions.js';
@@ -126,31 +128,59 @@ export function unregisteredToolOutcome(name, toolRuntime) {
 /** mutation 执行前读旧内容供 diff:write_file 取整文件旧内容(不存在→null=新建),
  * edit_file 取 old_string 起始行号(供 diff 显示真实文件行号)。读不到则 diff 退化为相对行号。
  * 非 mutation 或参数非法返 { preWriteOld: null, editStartLine: 1 }。失败不阻断。 */
+const toLf = (s) => s.replace(/\r\n/g, '\n').replace(/\r/g, '\n');
 export function readDiffContext(tc, parsed, resolvePath = jailResolve) {
+    const empty = { preWriteOld: null, editStartLine: 1 };
     if (!parsed)
-        return { preWriteOld: null, editStartLine: 1 };
+        return empty;
     const p = String(parsed.path ?? '');
     if (!p)
-        return { preWriteOld: null, editStartLine: 1 };
+        return empty;
     if (tc.name === 'write_file') {
         try {
             // jailResolve:沙箱越界(../../、绝对外圈、软链出圈)抛错 → catch 兜底返 null,不泄露牢外内容(TOCTOU)
             return { preWriteOld: readFileSync(resolvePath(p), 'utf8'), editStartLine: 1 };
         }
         catch {
-            return { preWriteOld: null, editStartLine: 1 }; // 文件不存在(新建)、不可读 或 沙箱越界(不泄露)
+            return empty; // 文件不存在(新建)、不可读 或 沙箱越界(不泄露)
         }
     }
     if (tc.name === 'edit_file') {
         // 行尾归一化:LLM 生成的 old_string 用 LF(\n),但 Windows 文件可能是 CRLF(\r\n),
         // 不统一则 indexOf 必败、editStartLine 恒为 1。与 edit-file.ts 保持一致归一化为 LF。
-        const oldStr = String(parsed.old_string ?? '')
-            .replace(/\r\n/g, '\n')
-            .replace(/\r/g, '\n');
+        const replacements = Array.isArray(parsed.replacements) ? parsed.replacements : null;
         try {
             // jailResolve:同上,沙箱越界抛错 → catch 兜底,不泄露牢外内容
             const raw = readFileSync(resolvePath(p), 'utf8');
-            const data = raw.replace(/\r\n/g, '\n').replace(/\r/g, '\n');
+            const data = toLf(raw);
+            // 多替换模式:在「编辑前」的原文上为每个 old_string 定位唯一命中的起始行,
+            // 供多 hunk diff 显示真实文件行号。任一项非唯一/找不到就不带 hunk(退回工具结果摘要),
+            // 真正的原子校验仍以 edit-file.ts execute 为准,这里只做展示、失败不阻断。
+            if (replacements) {
+                const hunks = [];
+                for (const item of replacements) {
+                    if (!item || typeof item !== 'object')
+                        return empty;
+                    const rec = item;
+                    if (typeof rec.old_string !== 'string' || typeof rec.new_string !== 'string')
+                        return empty;
+                    const oldStr = toLf(rec.old_string);
+                    if (!oldStr)
+                        return empty;
+                    const first = data.indexOf(oldStr);
+                    if (first < 0 || data.indexOf(oldStr, first + oldStr.length) !== -1)
+                        return empty;
+                    hunks.push({
+                        oldStr: rec.old_string,
+                        newStr: rec.new_string,
+                        startLine: data.slice(0, first).split('\n').length,
+                    });
+                }
+                if (hunks.length === 0)
+                    return empty;
+                return { preWriteOld: null, editStartLine: hunks[0].startLine, multiHunks: hunks };
+            }
+            const oldStr = toLf(String(parsed.old_string ?? ''));
             const idx = oldStr ? data.indexOf(oldStr) : -1;
             return {
                 preWriteOld: null,
@@ -158,10 +188,10 @@ export function readDiffContext(tc, parsed, resolvePath = jailResolve) {
             };
         }
         catch {
-            return { preWriteOld: null, editStartLine: 1 }; // 读不到:diff 退化为相对行号(含沙箱越界)
+            return empty; // 读不到:diff 退化为相对行号(含沙箱越界)
         }
     }
-    return { preWriteOld: null, editStartLine: 1 };
+    return empty;
 }
 /** 回灌 tool 结果到 history。
  *  正常路径只做单条 hard cap;原始 output 同时供 TUI 展示,因此用户与模型
@@ -169,12 +199,27 @@ export function readDiffContext(tc, parsed, resolvePath = jailResolve) {
  *  不在这里改写旧正文;所有自动清理与压缩统一由 80% pressure scheduler 决定。 */
 export function pushToolResult(history, tc, output, pruner, lifecycle, _scheduler, runtimeContextState = contextState, succeededOverride) {
     const succeeded = succeededOverride ?? isToolResultSuccess(output);
+    const capped = capToolResultForHistory(tc.name, output);
+    // push-time 截断兜底:完整结果另存 compacted/result-<id>.txt,并在进 history 的
+    // 截断内容尾部留一行全文指针。仅主 agent(共享 contextState)落盘——子 agent 的
+    // 结果随其临时 history 消亡,不污染主会话。best-effort:存不上就只保留截断内容。
+    let content = capped;
+    if (capped.length < output.length && runtimeContextState === contextState) {
+        try {
+            const rel = getActiveSessionStore().archiveToolResult(getCurrentSessionId() ?? '', tc.id, output);
+            if (rel)
+                content = `${capped.replace(/\s+$/, '')}\n…[完整结果已归档,需要原文时 grep/read_file:${rel}]`;
+        }
+        catch {
+            // 归档失败不影响工具回灌。
+        }
+    }
     const msg = {
         role: 'tool',
         tool_call_id: tc.id,
         // Preserve evidence verbatim in normal operation; the hard per-result cap
         // remains solely as a request-size safety rail.
-        content: capToolResultForHistory(tc.name, output),
+        content,
     };
     history.push(msg);
     const messageIndex = history.length - 1;

package/dist/config/file.js

@@ -22,11 +22,19 @@ export function readConfigFile() {
 }
 /**
  * 读-合并-写:把 entries 覆盖进 ~/.mocode/config,保留文件里其它键。
- * mkdirSync 兜底(首跑无 ~/.mocode)。不吞 I/O 异常——向导路径让其抛(与历史行为一致);
- * UI 偏好路径用 `updateConfigKey`(自带 try/catch 静默)。
+ * 值为 null 表示从配置文件删除该键(用于切到「不带某可选值」的预设时清除残留,
+ * 否则旧值会一直盖过新预设)。mkdirSync 兜底(首跑无 ~/.mocode)。不吞 I/O 异常——
+ * 向导路径让其抛(与历史行为一致);UI 偏好路径用 `updateConfigKey`(自带 try/catch 静默)。
  */
 export function writeConfigKeys(entries) {
-    const merged = { ...readConfigFile(), ...entries };
+    const existing = readConfigFile();
+    const merged = { ...existing };
+    for (const [k, v] of Object.entries(entries)) {
+        if (v === null)
+            delete merged[k];
+        else
+            merged[k] = v;
+    }
     const body = Object.entries(merged)
         .map(([k, v]) => `${k}=${v}`)
         .join('\n');

package/dist/config/index.js

@@ -57,6 +57,7 @@ const LLM_ENV_KEYS = [
     'LLM_API_KEY',
     'LLM_MODEL',
     'CONTEXT_WINDOW_TOKENS',
+    'MAX_TOKENS',
     'ANTHROPIC_PROMPT_CACHE',
     'REASONING_EFFORT',
 ];
@@ -485,7 +486,7 @@ ${buildWorkDisciplineSection(inferModelFamily(config.model))}
 - Silent Execution: invoke tools directly without preamble. Output visible text ONLY for the final answer and critical mid-task findings. Strictly no step-by-step narration (no "let me…", "让我先…", "now checking…" between calls).
 - Go directly to a known path or symbol; use discovery tools only when the location is unknown.
 - Edit against a FRESH read: before editing, call read_file on the exact path and copy the exact target text. Never reconstruct old_string from a grep/summary/diff — those lose whitespace and indentation and cause edit failures.
-- Pick the right file tool: use edit_file for partial changes to an existing file (expected_hash is OPTIONAL in old_string mode, REQUIRED in line-range mode); use write_file only to create a file, to rewrite most of a file (replacing needs the latest read_file hash), or to append (append=true needs no read_file and no hash). Never write a throwaway script (tmp_patch*.cjs, fix_*.py, ...) to modify source files: scripts run outside ChangeSet tracking, so rollback and diff cannot see them.
+- Pick the right file tool: for an EXISTING file, DEFAULT to edit_file — one old_string for a single change, replacements[] for many separate changes in one atomic call, or a line range for a big block. Do NOT overwrite an existing file with write_file just to tweak parts of it (re-sending the whole file wastes tokens and its long JSON is truncation-prone). Use write_file only to create a new file, to append (append=true needs no read_file and no hash), or to rewrite essentially the WHOLE file (a full rewrite needs the latest read_file hash). expected_hash is OPTIONAL in edit_file old_string/replacements mode, REQUIRED in line-range mode. Never write a throwaway script (tmp_patch*.cjs, fix_*.py, ...) to modify source files: scripts run outside ChangeSet tracking, so rollback and diff cannot see them.
 - A read_file hash from before a compaction, session resume, edit conflict, or external change is STALE and will be rejected if you pass it. In edit_file old_string mode just omit expected_hash; otherwise re-read rather than reuse an old hash.
 - Emit multiple independent tool calls in ONE assistant message so they run concurrently — e.g. several read_file regions, a grep plus a glob, or several web_fetch calls. One lookup per message wastes a full model round-trip each time. Place parallel-safe calls consecutively; keep any call that depends on their results (e.g. an edit) for the next message.
 - Never batch a read with an edit that depends on it; do not repeat overlapping reads or unchanged failed calls.
@@ -518,7 +519,10 @@ ${buildVoiceSection()}
     dynamicParts.push(`## Session state (\`.mocode/sessions/${sessionId ?? '<id>'}/notes.md\`)\n` +
         'For tasks with 3+ steps or context-loss risk, keep an execution plan with `plan_update`; record non-obvious findings/decisions/open questions/risks with `note_append`. ' +
         'Both write notes.md, which survives compaction — run read_file on it to recover context after compaction. ' +
-        'Their formats and rules live in each tool\u2019s description; do not hand-edit the files.');
+        'Their formats and rules live in each tool\u2019s description; do not hand-edit the files. ' +
+        'When compaction drops older conversation, its verbatim text (user messages, your earlier decisions, full tool args/results) is archived under ' +
+        `\`.mocode/sessions/${sessionId ?? '<id>'}/compacted/\` (see index.md there): the summary line ends with an [archived: …] pointer. ` +
+        'Before re-running a command or guessing what an old output said, grep that directory with glob narrowed to `.mocode/sessions/**` (required: .mocode is gitignored), then read_file the segment at the hit lines.');
     // 日期段:模型需要知道今天才能判断 freshness(web 搜索、版本时效)。只随天变化,
     // 不破坏会话内前缀缓存;置于 '## Session state' 之前,子 agent 切片仍保留(无害且有用)。
     const todaySection = buildTodaySection();
@@ -578,7 +582,8 @@ export const config = {
     model: llmKeysFromShell.includes('LLM_MODEL')
         ? process.env.LLM_MODEL || 'gpt-4o-mini'
         : (__activePreset?.model ?? process.env.LLM_MODEL ?? 'gpt-4o-mini'),
-    maxTokens: process.env.MAX_TOKENS ? Number(process.env.MAX_TOKENS) : undefined,
+    // 优先级:shell env > 激活预设(目录添加时按 limit.output 预填)> undefined=不下发吃后端默认。
+    maxTokens: process.env.MAX_TOKENS ? Number(process.env.MAX_TOKENS) : (__activePreset?.maxTokens ?? undefined),
     // 用 getter 而非 buildBasePrompt() 立即求值：对象字面量初始化期间 buildBasePrompt 会读取
     // config.model 等字段，直接调用会触发 TDZ。Getter 也让语言、persona 等运行时变化在下一轮生效。
     get systemPrompt() {
@@ -724,6 +729,14 @@ export function updateModelConfig(opts) {
         config.contextWindowTokens = opts.contextWindowTokens;
         process.env.CONTEXT_WINDOW_TOKENS = String(opts.contextWindowTokens);
     }
+    if (opts.maxTokens !== undefined) {
+        // null = 显式清除(切到无该值的预设);数字 = 设置。同步 env 保证本窗口立即按新值/不发该字段。
+        config.maxTokens = opts.maxTokens === null ? undefined : opts.maxTokens;
+        if (opts.maxTokens === null)
+            delete process.env.MAX_TOKENS;
+        else
+            process.env.MAX_TOKENS = String(opts.maxTokens);
+    }
     if (opts.anthropicPromptCache !== undefined) {
         config.anthropicPromptCache = opts.anthropicPromptCache;
         process.env.ANTHROPIC_PROMPT_CACHE = opts.anthropicPromptCache ? 'true' : 'false';

package/dist/config/presets.js

@@ -88,6 +88,14 @@ export function parsePreset(raw) {
     if (typeof contextWindow !== 'number' || !Number.isFinite(contextWindow) || contextWindow <= 0) {
         throw new Error(`预设 ${name}: contextWindow 必须为正数`);
     }
+    // maxTokens 可选:缺省=不下发(吃后端默认);给了就必须是正整数。
+    let maxTokens;
+    if (obj.maxTokens !== undefined && obj.maxTokens !== null) {
+        if (typeof obj.maxTokens !== 'number' || !Number.isFinite(obj.maxTokens) || obj.maxTokens <= 0) {
+            throw new Error(`预设 ${name}: maxTokens 必须为正数`);
+        }
+        maxTokens = Math.floor(obj.maxTokens);
+    }
     const provider = normalizePresetProvider(obj.provider);
     const anthropicPromptCache = provider === 'anthropic' && obj.anthropicPromptCache !== false;
     // reasoningEffort 缺省 = auto;显式写了非法值按现有校验风格报错(坏文件由 listPresets 跳过)。
@@ -132,6 +140,7 @@ export function parsePreset(raw) {
         apiKey,
         model,
         contextWindow: Math.floor(contextWindow),
+        ...(maxTokens ? { maxTokens } : {}),
         anthropicPromptCache,
         ...(reasoningEffort ? { reasoningEffort } : {}),
         ...extra,

package/dist/headless.js

@@ -23,6 +23,8 @@ import { initializeAllMcp, getMcpTools, closeAllMcp } from './mcp/index.js';
 import { config, buildBasePrompt, isMemoryEnabled } from './config/index.js';
 import { effectiveSystemPrompt } from './skills/index.js';
 import { buildMemoryIndexSection } from './memory/store.js';
+import { buildAboutYouSection, initPersona } from './memory/persona.js';
+import { buildProjectGraphHint } from './memory/recall.js';
 import { defaultRuntime, Runtime } from './runtime/index.js';
 import { runAgentCore } from './agent/core.js';
 import { summarizeToolCall, summarizeToolResult } from './ui/render.js';
@@ -154,7 +156,12 @@ export async function runHeadless(opts) {
         }
     }
     // 系统提示：Bot 岗位角色在前，基础约定在后；白名单存在时补一句工具边界。
-    const baseConventions = buildBasePrompt(sessionId) + buildMemoryIndexSection(isMemoryEnabled());
+    const memoryOn = isMemoryEnabled();
+    initPersona(memoryOn);
+    const baseConventions = buildBasePrompt(sessionId) +
+        buildAboutYouSection(memoryOn) +
+        buildMemoryIndexSection(memoryOn) +
+        buildProjectGraphHint(memoryOn);
     const systemContent = botInfo
         ? `# Role\n${botInfo.bot.systemPrompt}\n\n` +
             `${botInfo.bot.description ? `**Position:** ${botInfo.bot.description}\n\n` : ''}` +

package/dist/index.js

@@ -252,7 +252,7 @@ async function main() {
             }
             process.exit(0);
         }
-        const loaded = loadSession(id);
+        const loaded = loadSession(id, { restoreAssets: true });
         if (!loaded || !loaded.history.length) {
             console.error(t('cli.sessionMissing', { id }));
             process.exit(1);

package/dist/llm/index.js

@@ -740,6 +740,10 @@ async function chatOnce(messages, handlers, signal, toolsOverride, runtime = def
     let consumedAny = false;
     const thinkFilter = new ThinkTagFilter();
     let usage;
+    // finish_reason 只在最后一个带 choice 的 chunk 出现;取最后一次非空值。
+    // 'length' 意味着输出在 max_tokens 处被硬切断——长 tool_call 参数此时必为残缺 JSON,
+    // 工具层据此把笼统的 INVALID_JSON 升级成截断恢复指引。
+    let finishReason;
     const toolAcc = new Map();
     const emitVisible = (text) => {
         if (!text)
@@ -784,7 +788,11 @@ async function chatOnce(messages, handlers, signal, toolsOverride, runtime = def
                 cachedTokens: usage.cachedTokens,
             });
         }
-        const delta = chunk.choices?.[0]?.delta;
+        const choice0 = chunk.choices?.[0];
+        const delta = choice0?.delta;
+        if (typeof choice0?.finish_reason === 'string' && choice0.finish_reason) {
+            finishReason = choice0.finish_reason;
+        }
         if (!delta)
             continue; // 末尾 usage-only chunk 等无 delta
         if (delta.content) {
@@ -835,6 +843,7 @@ async function chatOnce(messages, handlers, signal, toolsOverride, runtime = def
         content: consumedAny ? visibleContent : null,
         toolCalls,
         usage,
+        finishReason,
     };
 }
 // ── token 估算(自包含,不依赖 ui / 外部 tokenizer)──────────────────────

package/dist/llm/providers/anthropic.js

@@ -315,6 +315,8 @@ export async function anthropicChatOnce(messages, handlers, signal, tools, runti
         await throwResponseError(response);
     let content = '';
     let usage;
+    // message_delta 的 delta.stop_reason 给出停止原因;max_tokens 归一到 OpenAI 的 'length'。
+    let finishReason;
     const toolAcc = new Map();
     const live = { cjk: 0, other: 0 };
     const reportLive = () => {
@@ -391,6 +393,12 @@ export async function anthropicChatOnce(messages, handlers, signal, tools, runti
             continue;
         }
         if (type === 'message_delta') {
+            // stop_reason 位于 message_delta 的 delta 子对象(与 usage 同级事件)。
+            const d = event.delta;
+            if (typeof d?.stop_reason === 'string' && d.stop_reason) {
+                // Anthropic: end_turn / max_tokens / stop_sequence / tool_use / tool_use 等。
+                finishReason = d.stop_reason === 'max_tokens' ? 'length' : d.stop_reason;
+            }
             const raw = event.usage;
             if (raw) {
                 const previous = usage ?? usageFromAnthropic({});
@@ -414,5 +422,5 @@ export async function anthropicChatOnce(messages, handlers, signal, tools, runti
     const toolCalls = [...toolAcc.entries()]
         .sort((a, b) => a[0] - b[0])
         .map(([, call]) => ({ ...call, arguments: call.arguments || '{}' }));
-    return { content: content || null, toolCalls, usage };
+    return { content: content || null, toolCalls, usage, finishReason };
 }

package/dist/llm/providers/vertex.js

@@ -407,5 +407,11 @@ export async function vertexChatOnce(messages, handlers, signal, tools, runtime,
         content += marker;
         handlers.onText?.(marker);
     }
-    return { content: content || null, toolCalls, usage };
+    return {
+        content: content || null,
+        toolCalls,
+        usage,
+        // Gemini MAX_TOKENS 归一到 OpenAI 的 'length',供工具层给出截断恢复指引。
+        ...(finishReason ? { finishReason: finishReason === 'MAX_TOKENS' ? 'length' : finishReason } : {}),
+    };
 }