deepccc 0.1.25 → 0.1.27

package/README.md

@@ -75,6 +75,7 @@ $env:DEEPCCC_STREAMING="true"
   "apiKey": "sk-...",
   "baseURL": "https://api.deepseek.com/v1",
   "model": "deepseek-v4-pro",
+  "subModel": "",
   "effort": "",
   "streaming": true,
   "contextWindow": 1048576,
@@ -102,6 +103,16 @@ $env:DEEPCCC_STREAMING="true"
 API 拒绝（context length exceeded），实际窗口以模型与所用服务端为准（如 litellm 的
 `max_input_tokens`）。
 
+`subModel` 是子模型（选填），默认 `""`（留空跟随主模型）。配置后，DeepCCC 内部的轻量
+环节——上下文压缩摘要生成、`task` 子代理任务——使用子模型执行，主对话仍用主模型。
+典型用法：主模型用 pro 承担复杂推理，子模型用 flash 做高频廉价的摘要与子任务。
+可通过 `DEEPCCC_SUB_MODEL` 环境变量或命令行 `--sub-model` 覆盖。
+
+`task` 子代理工具：主模型可把边界清晰的独立子任务（仓库调研、长文档阅读、独立模块生成）
+委派给子代理执行——子代理使用子模型、独立上下文，不污染主对话上下文；结果截断回传。
+子代理**不能再次委派**（禁止嵌套），单轮最多 20 个工具步，超时与主会话压缩超时一致。
+仅在配置了子模型时建议使用（未配置时子代理跟随主模型，节省有限）。
+
 `rawStreamLogs.enabled` 默认 `true`，通过 `DEEPCCC_RAW_STREAM_LOGS` 环境变量或配置 JSON 关闭。
 开启时，每次对话的原始流按 gzip JSONL 落到 `~/.deepccc/raw-stream-logs/`，供
 `session_search` 工具在会话被压缩后找回被压缩消息的精确原文（检索时设置

package/dist/cli.js

@@ -57,6 +57,10 @@ function parseArgs(argv = process.argv.slice(2)) {
             config.model = next;
             i++;
         }
+        else if (arg === "--sub-model" && next !== undefined) {
+            config.subModel = next;
+            i++;
+        }
         else if (arg === "--effort" && next !== undefined) {
             config.effort = next;
             i++;
@@ -125,6 +129,7 @@ function printHelp(appConfig) {
         "Options:",
         `  --provider <name>    API protocol: openai or anthropic (current default ${appConfig.provider})`,
         `  --model <name>       Model name (current default ${appConfig.model})`,
+        `  --sub-model <name>   Sub-model for lightweight steps (compaction/task; empty = follow main model)`,
         `  --effort <level>     Reasoning effort: none/minimal/low/medium/high/xhigh/max (overrides config.effort)`,
         `  --base-url <url>     Provider API base URL (current default ${appConfig.baseURL})`,
         "  --api-key <key>      API key",

package/dist/config.d.ts

@@ -5,6 +5,11 @@ export interface DeepCccConfig {
     apiKey: string;
     baseURL: string;
     model: string;
+    /**
+     * 子模型（可选）：用于内部轻量环节（上下文压缩摘要生成、task 子代理任务）。
+     * 留空（""）时跟随主模型，行为与旧版完全一致。
+     */
+    subModel: string;
     /** Reasoning effort（none/minimal/low/medium/high/xhigh/max），留空不传 reasoning_effort */
     effort: string;
     /** 主对话是否使用流式请求；默认开启 */

package/dist/config.js

@@ -12,6 +12,7 @@ export const DEFAULT_CONFIG = {
     apiKey: "",
     baseURL: "https://api.deepseek.com/v1",
     model: "deepseek-v4-pro",
+    subModel: "",
     effort: "",
     streaming: true,
     contextWindow: 1_048_576,
@@ -66,6 +67,7 @@ function loadConfig() {
         apiKey: env("DEEPCCC_API_KEY") ?? env("DEEPSEEK_API_KEY") ?? file.apiKey ?? DEFAULT_CONFIG.apiKey,
         baseURL: env("DEEPCCC_BASE_URL") ?? env("DEEPSEEK_BASE_URL") ?? file.baseURL ?? DEFAULT_CONFIG.baseURL,
         model: env("DEEPCCC_MODEL") ?? env("DEEPSEEK_MODEL") ?? file.model ?? DEFAULT_CONFIG.model,
+        subModel: env("DEEPCCC_SUB_MODEL") ?? file.subModel ?? DEFAULT_CONFIG.subModel,
         effort: env("DEEPCCC_EFFORT") ?? env("DEEPSEEK_EFFORT") ?? file.effort ?? DEFAULT_CONFIG.effort,
         streaming: boolEnv("DEEPCCC_STREAMING") ?? file.streaming ?? DEFAULT_CONFIG.streaming,
         contextWindow: numberEnv("DEEPCCC_CONTEXT_WINDOW") ?? file.contextWindow ?? DEFAULT_CONFIG.contextWindow,

package/dist/context.d.ts

@@ -111,5 +111,6 @@ export declare class BuiltinContextManager {
     applyCompaction(summary: string, plan: BuiltinCompactionPlan): void;
     reset(): void;
     save(): void;
+    private modelSafeMessages;
     private load;
 }

package/dist/context.js

@@ -2,6 +2,7 @@ import { createHash, randomBytes } from "node:crypto";
 import { existsSync, mkdirSync, readdirSync, readFileSync, renameSync, writeFileSync } from "node:fs";
 import { homedir } from "node:os";
 import { join } from "node:path";
+import { hasMalformedToolProtocolText } from "./tool-protocol.js";
 export const DEFAULT_BUILTIN_CONTEXT_DIR = join(homedir(), ".deepccc", "sessions");
 /** 默认模型上下文窗口：1M（DeepSeek V4 Pro/Flash 原生规格）。 */
 export const DEFAULT_CONTEXT_WINDOW_TOKENS = 1_048_576;
@@ -290,14 +291,19 @@ export class BuiltinContextManager {
                 ].join("\n"),
             });
         }
-        messages.push(...this.state.messages);
+        // Keep malformed provider output on disk for diagnosis, but quarantine it
+        // from future prompts so one protocol failure cannot teach the model to
+        // repeat the same invalid tool syntax on every later turn.
+        messages.push(...this.modelSafeMessages());
         return messages;
     }
     planCompaction() {
-        const estimated = estimateBuiltinContextTokens(this.state.summary, this.state.messages);
+        // The compaction model is still a model input: exclude quarantined protocol
+        // leaks here too, otherwise a later summary could reintroduce the bad syntax.
+        const messages = this.modelSafeMessages();
+        const estimated = estimateBuiltinContextTokens(this.state.summary, messages);
         if (estimated <= this.compactAtTokens)
             return null;
-        const messages = this.state.messages;
         if (messages.length <= 1)
             return null;
         const earliestAllowed = Math.max(0, messages.length - this.keepRecentMessages);
@@ -340,6 +346,9 @@ export class BuiltinContextManager {
         writeFileSync(tmp, content, "utf8");
         renameSync(tmp, this.contextFilePath);
     }
+    modelSafeMessages() {
+        return this.state.messages.filter((message) => message.role !== "assistant" || !hasMalformedToolProtocolText(message.content));
+    }
     load() {
         if (!this.persist || !existsSync(this.contextFilePath))
             return emptyState(this.sessionId, this.cwd);

package/dist/file-tools.d.ts

@@ -1,5 +1,7 @@
 import { type ToolSet } from "ai";
 import { type PermissionGate } from "./permissions.js";
+/** task 子代理工具：子任务结果回传主会话前的最大字符数（防止子代理长输出撑爆主上下文） */
+export declare const MAX_TASK_OUTPUT_CHARS = 32000;
 export interface ReadFileInput {
     path: string;
     startLine?: number;
@@ -155,5 +157,21 @@ export interface BuiltinFileToolsOptions {
         contextDir?: string;
         rawLogsDir?: string;
     };
+    /**
+     * task 子代理工具的执行回调：把独立子任务委派给子会话（使用子模型、独立上下文）。
+     * 未提供时 task 工具不可用（抛出明确错误），主会话之外的工具集不会意外开启子代理。
+     */
+    runTask?: TaskRunner;
+}
+export interface TaskRunnerInput {
+    /** 子任务描述：包含目标、约束与交付物，子代理将以此作为唯一指令。 */
+    description: string;
+    /** 可选子任务工作目录（相对主会话 cwd 或绝对路径），默认继承主会话工作目录。 */
+    cwd?: string;
+}
+export interface TaskRunnerOutput {
+    /** 子代理最终文本回复（已截断到 MAX_TASK_OUTPUT_CHARS，已过隐私替换）。 */
+    result: string;
 }
+export type TaskRunner = (input: TaskRunnerInput, signal?: AbortSignal) => Promise<string>;
 export declare function createBuiltinFileTools(cwd: string, options?: BuiltinFileToolsOptions): ToolSet;

package/dist/file-tools.js

@@ -22,6 +22,8 @@ const SEARCH_TIMEOUT_MS = 15_000;
 const MAX_COMMAND_OUTPUT_BYTES = 256 * 1024;
 const DEFAULT_COMMAND_TIMEOUT_MS = 120_000;
 const MAX_COMMAND_TIMEOUT_MS = 900_000;
+/** task 子代理工具：子任务结果回传主会话前的最大字符数（防止子代理长输出撑爆主上下文） */
+export const MAX_TASK_OUTPUT_CHARS = 32_000;
 const requireFromHere = createRequire(import.meta.url);
 const FALLBACK_SKIPPED_DIRECTORIES = new Set([".git", "node_modules"]);
 /**
@@ -939,6 +941,7 @@ export async function applyPatchForTool(cwd, input) {
 }
 export function createBuiltinFileTools(cwd, options = {}) {
     const gate = options.permissionGate;
+    const runTask = options.runTask;
     /** 文件路径的候选匹配键：原始路径 + 绝对路径 + 相对 cwd 路径（正/反斜杠双版本，规则可任选其一） */
     const pathKeys = (p) => {
         const abs = resolveToolPath(cwd, p);
@@ -1023,6 +1026,25 @@ export function createBuiltinFileTools(cwd, options = {}) {
                 return runCommandForTool(cwd, input, options.abortSignal);
             },
         }),
+        task: tool({
+            description: "把独立子任务委派给子代理执行：子代理使用子模型、拥有独立上下文，不污染主对话上下文。适合边界清晰、可独立交付的调研/代码生成子任务（如扫描整个仓库、阅读长文档、生成独立模块）。子代理不能再次委派任务（禁止嵌套），结果会截断回传。",
+            inputSchema: jsonSchema({
+                type: "object",
+                additionalProperties: false,
+                properties: {
+                    description: { type: "string", description: "子任务描述：目标、约束与交付物。请写清楚子代理需要返回什么。" },
+                    cwd: { type: "string", description: "可选子任务工作目录（绝对路径或相对主会话 cwd），默认继承主会话工作目录。" },
+                },
+                required: ["description"],
+            }),
+            execute: async (input, execOptions) => {
+                if (!runTask) {
+                    throw new Error("task 工具不可用：当前环境未启用子代理执行器");
+                }
+                const result = await runTask(input, execOptions.abortSignal);
+                return { result };
+            },
+        }),
         edit_file: tool({
             description: "通过精确的 oldText -> newText 替换编辑现有 UTF-8 文本文件。可行时使用 SHA-256 前置条件以避免覆盖并发编辑。",
             inputSchema: jsonSchema({

package/dist/index.d.ts

@@ -28,6 +28,10 @@ export interface ChatSessionConfig {
     apiKey?: string;
     /** Model id. Defaults to DEEPCCC_MODEL/config. */
     model?: string;
+    /**
+     * 子模型 id（可选）：用于压缩摘要生成与 task 子代理任务。留空跟随主模型。
+     */
+    subModel?: string;
     /**
      * Reasoning effort (none/minimal/low/medium/high/xhigh/max);
      * overrides config.effort; empty omits the reasoning_effort request field.
@@ -81,6 +85,9 @@ export interface ChatSessionOptions {
 export type ChatEvent = {
     type: "status";
     phase: "compacting" | "generating";
+} | {
+    type: "progress";
+    phase: "reasoning";
 } | {
     type: "compact";
     compactedMessages: number;
@@ -99,6 +106,8 @@ export type ChatEvent = {
     type: "text";
     text: string;
     accumulated: string;
+} | {
+    type: "text_reset";
 } | {
     type: "done";
     text: string;
@@ -115,17 +124,30 @@ interface ChatMessage {
 }
 export declare class ChatSession {
     private model;
+    /** 子模型实例；未配置 subModel 时与主模型同一实例 */
+    private subModel;
     private provider;
+    private apiKey;
+    private baseURL;
+    private modelId;
+    private subModelId;
     private cwd;
     private context;
     private compactionTimeoutMs;
     private maxSteps?;
     private effort;
+    private permissionMode;
+    private permissionResolver?;
     private permissionGate;
     private skillDirs;
     private customSystemPrompt;
     /** 最近一次 chat() 使用的 system prompt（供 history 等读取） */
     private systemPrompt;
+    /**
+     * task 子代理工具执行器：用子模型开独立子会话（独立上下文、独立 cwd）执行子任务，
+     * 回传最终文本（截断 + 隐私替换）。单层：子会话的工具集不包含 runTask，天然禁止嵌套。
+     */
+    private runTask;
     constructor(overrides?: ChatSessionConfig, options?: ChatSessionOptions);
     /**
      * 组装系统提示词。顺序遵循“稳定性优先”原则（缓存命中友好）：

package/dist/index.js

@@ -8,13 +8,14 @@ import { createAnthropic } from "@ai-sdk/anthropic";
 import { generateText, isLoopFinished, stepCountIs, streamText } from "ai";
 import { existsSync, readFileSync } from "node:fs";
 import { homedir } from "node:os";
-import { join } from "node:path";
+import { isAbsolute, join, resolve } from "node:path";
 import { fileURLToPath } from "node:url";
 import { config as appConfig, normalizeDeepCccProvider, RAW_STREAM_LOGS_DIR, } from "./config.js";
 import { createRawStreamLog, } from "./raw-stream-log.js";
 import { buildPersistedAssistantMessage, buildSummaryPrompt, BuiltinContextManager, defaultBuiltinSessionId, } from "./context.js";
-import { createBuiltinFileTools } from "./file-tools.js";
+import { createBuiltinFileTools, MAX_TASK_OUTPUT_CHARS } from "./file-tools.js";
 import { PermissionGate } from "./permissions.js";
+import { hasMalformedToolProtocolText, TOOL_PROTOCOL_RECOVERY_PROMPT, } from "./tool-protocol.js";
 import { buildDefaultSkillDirs, buildSkillsIndexPrompt, scanSkillsDirs, } from "./skills.js";
 import { applyPrivacy, applyPrivacyToJson } from "./privacy.js";
 // ---------------------------------------------------------------------------
@@ -76,6 +77,15 @@ const COMPACTION_RECOVERY_HINT_DISABLED = [
 ].join("\n");
 export const DEFAULT_COMPACTION_TIMEOUT_MS = 5 * 60 * 1000;
 const MAX_COMPACTION_OUTPUT_TOKENS = 16_384;
+/** task 子代理工具：子代理单轮对话的最大工具步数（防失控循环，结果收敛后即结束） */
+const TASK_MAX_STEPS = 20;
+/** 子代理最终输出截断：保留前 MAX_TASK_OUTPUT_CHARS 字符，尾部注明截断信息 */
+function truncateTaskOutput(text) {
+    if (text.length <= MAX_TASK_OUTPUT_CHARS)
+        return text;
+    return (text.slice(0, MAX_TASK_OUTPUT_CHARS) +
+        `\n…[子代理输出已截断，共 ${text.length} 字符，仅保留前 ${MAX_TASK_OUTPUT_CHARS} 字符]`);
+}
 const ANTHROPIC_TOOL_JSON_COMPATIBILITY_NOTE = [
     "[Protocol compatibility note]",
     "tool-call arguments use JSON encoding; the final reply does not need to be JSON unless the user requests it.",
@@ -217,17 +227,71 @@ function normalizeAnthropicBaseURL(baseURL) {
 }
 export class ChatSession {
     model;
+    /** 子模型实例；未配置 subModel 时与主模型同一实例 */
+    subModel;
     provider;
+    apiKey;
+    baseURL;
+    modelId;
+    subModelId;
     cwd;
     context;
     compactionTimeoutMs;
     maxSteps;
     effort;
+    permissionMode;
+    permissionResolver;
     permissionGate;
     skillDirs;
     customSystemPrompt;
     /** 最近一次 chat() 使用的 system prompt（供 history 等读取） */
     systemPrompt = "";
+    /**
+     * task 子代理工具执行器：用子模型开独立子会话（独立上下文、独立 cwd）执行子任务，
+     * 回传最终文本（截断 + 隐私替换）。单层：子会话的工具集不包含 runTask，天然禁止嵌套。
+     */
+    runTask = async (input, signal) => {
+        const rawCwd = input.cwd?.trim();
+        const taskCwd = rawCwd ? (isAbsolute(rawCwd) ? rawCwd : resolve(this.cwd, rawCwd)) : this.cwd;
+        const timeoutController = new AbortController();
+        const timeout = setTimeout(() => timeoutController.abort(), this.compactionTimeoutMs);
+        timeout.unref?.();
+        const taskSignal = signal
+            ? AbortSignal.any([signal, timeoutController.signal])
+            : timeoutController.signal;
+        try {
+            const child = new ChatSession({
+                provider: this.provider,
+                apiKey: this.apiKey,
+                baseURL: this.baseURL,
+                model: this.subModelId || this.modelId,
+                effort: this.effort,
+            }, {
+                cwd: taskCwd,
+                persist: false,
+                permissionMode: this.permissionMode,
+                permissionResolver: this.permissionResolver,
+                maxSteps: TASK_MAX_STEPS,
+                compactionTimeoutMs: this.compactionTimeoutMs,
+                skillsDirs: this.skillDirs.map((s) => s.dir),
+            });
+            let full = "";
+            for await (const event of child.chat(input.description, taskSignal)) {
+                if (event.type === "text") {
+                    full += event.text;
+                }
+                else if (event.type === "error") {
+                    throw new Error(`task 子代理执行失败: ${event.message}`);
+                }
+            }
+            if (!full.trim())
+                return "(子代理未返回文本内容)";
+            return applyPrivacy(truncateTaskOutput(full));
+        }
+        finally {
+            clearTimeout(timeout);
+        }
+    };
     constructor(overrides = {}, options = {}) {
         const apiKey = overrides.apiKey ?? appConfig.apiKey;
         if (!apiKey) {
@@ -235,28 +299,32 @@ export class ChatSession {
         }
         const baseURL = overrides.baseURL ?? appConfig.baseURL;
         const modelId = overrides.model ?? appConfig.model;
+        this.modelId = modelId;
+        this.subModelId = (overrides.subModel ?? appConfig.subModel ?? "").trim();
         this.provider = normalizeDeepCccProvider(overrides.provider ?? appConfig.provider);
         this.effort = (overrides.effort ?? appConfig.effort ?? "").trim();
-        if (this.provider === "anthropic") {
-            const provider = createAnthropic({
+        this.apiKey = apiKey;
+        this.baseURL = baseURL;
+        const provider = this.provider === "anthropic"
+            ? createAnthropic({
                 baseURL: normalizeAnthropicBaseURL(baseURL),
                 apiKey,
-            });
-            this.model = provider(modelId);
-        }
-        else {
-            const provider = createOpenAICompatible({
+            })
+            : createOpenAICompatible({
                 name: "deepccc",
                 baseURL,
                 apiKey,
                 includeUsage: true,
             });
-            this.model = provider(modelId);
-        }
+        this.model = provider(modelId);
+        // 子模型：留空时与主模型共用同一实例（行为与旧版完全一致，零开销）
+        this.subModel = this.subModelId ? provider(this.subModelId) : this.model;
         this.cwd = options.cwd ?? process.cwd();
         this.maxSteps = normalizeMaxSteps(options.maxSteps);
         this.compactionTimeoutMs = Math.max(1, options.compactionTimeoutMs ?? DEFAULT_COMPACTION_TIMEOUT_MS);
         this.customSystemPrompt = options.systemPrompt ?? "";
+        this.permissionMode = options.permissionMode ?? "ask";
+        this.permissionResolver = options.permissionResolver;
         // 技能目录在构造时确定；技能内容在每次 chat() 前重新扫描（mtime 热加载），
         // 因此创建/修改技能后下一次对话自动生效，无需重启。
         this.skillDirs =
@@ -271,7 +339,7 @@ export class ChatSession {
             compactAtTokens: options.compactAtTokens,
             keepRecentMessages: options.keepRecentMessages,
         });
-        this.permissionGate = new PermissionGate(options.permissionMode ?? "ask", options.permissionResolver);
+        this.permissionGate = new PermissionGate(this.permissionMode, this.permissionResolver);
     }
     /**
      * 组装系统提示词。顺序遵循“稳定性优先”原则（缓存命中友好）：
@@ -333,7 +401,6 @@ export class ChatSession {
             catch (err) {
                 console.error(`[DeepCCC raw stream log] create failed: ${errorMessage(err)}`);
             }
-            const toolContext = [];
             const maxSteps = this.maxSteps;
             // 每次对话前重新扫描技能索引（并行 + mtime 缓存，开销极小）：
             // 新技能/修改的技能在下一次对话自动生效（热加载）。
@@ -356,97 +423,143 @@ export class ChatSession {
                     ? { deepseek: { reasoningEffort: this.effort } }
                     : { anthropic: { effort: this.effort } };
             }
-            const generationOptions = {
+            const baseGenerationOptions = {
                 model: this.model,
                 system,
-                messages: modelMessages,
-                tools: createBuiltinFileTools(this.cwd, { permissionGate: this.permissionGate }),
+                tools: createBuiltinFileTools(this.cwd, {
+                    permissionGate: this.permissionGate,
+                    runTask: this.runTask,
+                }),
                 stopWhen: maxSteps !== undefined ? stepCountIs(maxSteps) : isLoopFinished(),
                 abortSignal: signal,
                 ...(effortProviderOptions ? { providerOptions: effortProviderOptions } : {}),
             };
-            let stream;
-            if (appConfig.streaming) {
-                const result = streamText(generationOptions);
-                stream = result.fullStream ?? textStreamToFullStream(result.textStream);
-            }
-            else {
-                const result = await generateText(generationOptions);
-                stream = generateResultToFullStream(result);
-            }
-            for await (const part of stream) {
-                rawLog?.writeLine(safeRawStreamJson(part));
-                if (part.type === "text-delta") {
-                    fullText += part.text;
-                    // 隐私替换只在展示层：safeAccumulated 供事件消费者（终端/JSONL）使用，
-                    // fullText 原文用于持久化上下文，避免替换结果回流污染上下文。
-                    const safeText = applyPrivacy(part.text);
-                    safeAccumulated += safeText;
-                    yield { type: "text", text: safeText, accumulated: safeAccumulated };
+            for (let attempt = 0; attempt < 2; attempt += 1) {
+                fullText = "";
+                safeAccumulated = "";
+                const toolContext = [];
+                toolCallsById.clear();
+                toolCallOrder.length = 0;
+                let lastReasoningProgressAt;
+                const attemptMessages = attempt === 0
+                    ? modelMessages
+                    : [...modelMessages, { role: "user", content: TOOL_PROTOCOL_RECOVERY_PROMPT }];
+                const generationOptions = {
+                    ...baseGenerationOptions,
+                    messages: attemptMessages,
+                };
+                let stream;
+                if (appConfig.streaming) {
+                    const result = streamText(generationOptions);
+                    stream = result.fullStream ?? textStreamToFullStream(result.textStream);
                 }
-                else if (part.type === "tool-call") {
-                    toolContext.push(`tool_call ${part.toolName}: ${safeJson(part.input)}`);
-                    toolCallsById.set(part.toolCallId, { name: part.toolName, input: safeJson(part.input) });
-                    toolCallOrder.push(part.toolCallId);
-                    yield {
-                        type: "tool_use",
-                        id: part.toolCallId,
-                        name: part.toolName,
-                        input: applyPrivacyToJson(part.input),
-                    };
+                else {
+                    const result = await generateText(generationOptions);
+                    stream = generateResultToFullStream(result);
                 }
-                else if (part.type === "tool-result") {
-                    toolContext.push(`tool_result ${part.toolName}: ${truncateToolContext(safeJson(part.output))}`);
-                    const call = toolCallsById.get(part.toolCallId);
-                    if (call) {
-                        call.output = truncateToolContext(safeJson(part.output));
+                for await (const part of stream) {
+                    rawLog?.writeLine(safeRawStreamJson(part));
+                    if (part.type === "reasoning-start" || part.type === "reasoning-delta") {
+                        // Reasoning content remains private. A throttled heartbeat is enough
+                        // for ChatCCC to distinguish active inference from a stalled stream.
+                        const now = Date.now();
+                        if (lastReasoningProgressAt === undefined || now - lastReasoningProgressAt >= 1_000) {
+                            lastReasoningProgressAt = now;
+                            yield { type: "progress", phase: "reasoning" };
+                        }
                     }
-                    yield {
-                        type: "tool_result",
-                        tool_use_id: part.toolCallId,
-                        name: part.toolName,
-                        content: applyPrivacyToJson(part.output),
-                        is_error: false,
-                    };
-                }
-                else if (part.type === "tool-error") {
-                    const message = errorMessage(part.error);
-                    toolContext.push(`tool_error ${part.toolName}: ${message}`);
-                    const call = toolCallsById.get(part.toolCallId);
-                    if (call) {
-                        call.output = message;
-                        call.is_error = true;
+                    else if (part.type === "text-delta") {
+                        fullText += part.text;
+                        // 隐私替换只在展示层：safeAccumulated 供事件消费者（终端/JSONL）使用，
+                        // fullText 原文用于持久化上下文，避免替换结果回流污染上下文。
+                        const safeText = applyPrivacy(part.text);
+                        safeAccumulated += safeText;
+                        yield { type: "text", text: safeText, accumulated: safeAccumulated };
+                    }
+                    else if (part.type === "tool-call") {
+                        toolContext.push(`tool_call ${part.toolName}: ${safeJson(part.input)}`);
+                        toolCallsById.set(part.toolCallId, { name: part.toolName, input: safeJson(part.input) });
+                        toolCallOrder.push(part.toolCallId);
+                        yield {
+                            type: "tool_use",
+                            id: part.toolCallId,
+                            name: part.toolName,
+                            input: applyPrivacyToJson(part.input),
+                        };
+                    }
+                    else if (part.type === "tool-result") {
+                        toolContext.push(`tool_result ${part.toolName}: ${truncateToolContext(safeJson(part.output))}`);
+                        const call = toolCallsById.get(part.toolCallId);
+                        if (call)
+                            call.output = truncateToolContext(safeJson(part.output));
+                        yield {
+                            type: "tool_result",
+                            tool_use_id: part.toolCallId,
+                            name: part.toolName,
+                            content: applyPrivacyToJson(part.output),
+                            is_error: false,
+                        };
+                    }
+                    else if (part.type === "tool-error") {
+                        const message = errorMessage(part.error);
+                        toolContext.push(`tool_error ${part.toolName}: ${message}`);
+                        const call = toolCallsById.get(part.toolCallId);
+                        if (call) {
+                            call.output = message;
+                            call.is_error = true;
+                        }
+                        yield {
+                            type: "tool_result",
+                            tool_use_id: part.toolCallId,
+                            name: part.toolName,
+                            content: applyPrivacy(message),
+                            is_error: true,
+                        };
+                    }
+                    else if (part.type === "error") {
+                        const message = errorMessage(part.error);
+                        yield { type: "error", message: applyPrivacy(message) };
+                        throw new Error(message);
                     }
-                    yield {
-                        type: "tool_result",
-                        tool_use_id: part.toolCallId,
-                        name: part.toolName,
-                        content: applyPrivacy(message),
-                        is_error: true,
-                    };
                 }
-                else if (part.type === "error") {
-                    const message = errorMessage(part.error);
-                    yield { type: "error", message: applyPrivacy(message) };
-                    throw new Error(message);
+                if (hasMalformedToolProtocolText(fullText)) {
+                    console.warn(`[DeepCCC] malformed DSML tool output detected for ${this.context.sessionId} `
+                        + `(attempt ${attempt + 1}/2, structuredToolCalls=${toolCallOrder.length})`);
+                    rawLog?.writeLine(safeRawStreamJson({
+                        type: "deepccc_tool_protocol_recovery",
+                        attempt: attempt + 1,
+                        structuredToolCalls: toolCallOrder.length,
+                    }));
+                    yield { type: "text_reset" };
+                    if (attempt === 0 && toolCallOrder.length === 0) {
+                        yield { type: "status", phase: "generating" };
+                        continue;
+                    }
+                    throw new Error(toolCallOrder.length > 0
+                        ? "工具调用协议异常：检测到混合的结构化与 DSML 文本调用，为避免重复执行工具，本轮已安全终止"
+                        : "工具调用协议异常：模型重试后仍输出了无效的 DSML 工具调用");
                 }
+                completed = true;
+                const collectedToolCalls = toolCallOrder
+                    .map((id) => toolCallsById.get(id))
+                    .filter((call) => call !== undefined);
+                this.context.appendMessage(buildPersistedAssistantMessage({
+                    fullText,
+                    transcriptLines: toolContext,
+                    toolCalls: collectedToolCalls,
+                }));
+                yield { type: "done", text: safeAccumulated };
+                return;
             }
-            completed = true;
-            const collectedToolCalls = toolCallOrder
-                .map((id) => toolCallsById.get(id))
-                .filter((call) => call !== undefined);
-            this.context.appendMessage(buildPersistedAssistantMessage({
-                fullText,
-                transcriptLines: toolContext,
-                toolCalls: collectedToolCalls,
-            }));
-            yield { type: "done", text: safeAccumulated };
         }
         catch (err) {
             const message = err instanceof Error ? err.message : String(err);
+            const malformedProtocolOutput = hasMalformedToolProtocolText(fullText);
+            if (malformedProtocolOutput)
+                yield { type: "text_reset" };
             if (err.name === "AbortError" || signal?.aborted) {
                 // 被中断时，不保存不完整的助手消息
-                if (fullText) {
+                if (fullText && !malformedProtocolOutput) {
                     this.context.appendMessage({ role: "assistant", content: `${fullText}\n[interrupted]` });
                 }
                 yield { type: "done", text: safeAccumulated };
@@ -500,7 +613,7 @@ export class ChatSession {
             if (!plan)
                 return 0;
             const result = await generateText({
-                model: this.model,
+                model: this.subModel,
                 system: SUMMARY_SYSTEM_PROMPT,
                 messages: [{ role: "user", content: buildSummaryPrompt(plan) }],
                 abortSignal: compactionSignal,

package/dist/progress/reducer.js

@@ -57,6 +57,10 @@ export function reduceProgress(prev, event) {
         case "text":
             // accumulated 是全文累积，直接全量替换，天然幂等
             return withProgressView(prev, { text: event.accumulated });
+        case "progress":
+            return withProgressView(prev, { headerTitle: "思考中..." });
+        case "text_reset":
+            return withProgressView(prev, { text: "", tools: [] });
         case "tool_use": {
             const tool = {
                 id: event.id ?? `tool-${prev.tools.length + 1}`,

package/dist/tool-protocol.d.ts

@@ -0,0 +1,2 @@
+export declare function hasMalformedToolProtocolText(text: string): boolean;
+export declare const TOOL_PROTOCOL_RECOVERY_PROMPT: string;

package/dist/tool-protocol.js

@@ -0,0 +1,17 @@
+/**
+ * Detects provider output that leaked DeepSeek's internal DSML tool-call syntax
+ * into normal assistant text. A trailing closing tag is intentionally required:
+ * mentioning DSML in prose or a code sample must not trigger recovery.
+ */
+const MALFORMED_DSML_TRAILER = /<\/[|｜]{2}DSML[|｜]{2}(?:parameter|invoke)>\s*$/iu;
+const RENDERED_TOOL_CALL = /\[调用\s+[A-Za-z_][\w.-]*\]/u;
+const DSML_INVOKE_TAG = /<[|｜]{2}DSML[|｜]{2}invoke\b/iu;
+export function hasMalformedToolProtocolText(text) {
+    return MALFORMED_DSML_TRAILER.test(text)
+        && (RENDERED_TOOL_CALL.test(text) || DSML_INVOKE_TAG.test(text));
+}
+export const TOOL_PROTOCOL_RECOVERY_PROMPT = [
+    "[系统恢复提示] 上一次响应泄漏了内部工具调用协议，因此已被丢弃。",
+    "请重新完成当前用户请求。需要调用工具时，只能使用 API 提供的结构化工具调用；不要把 DSML 或工具参数作为普通文本输出。",
+    "不要提及本恢复提示，也不要假装工具已经执行。",
+].join("\n");

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "deepccc",
-  "version": "0.1.25",
+  "version": "0.1.27",
   "description": "A lightweight coding agent with OpenAI-compatible and Anthropic Messages API support.",
   "license": "Apache-2.0",
   "keywords": [