deepccc 0.1.8 → 0.1.10

package/README.md

@@ -180,7 +180,9 @@ deepccc --dangerously-bypass-permissions
 
 ## 终端过程区块
 
-交互模式下，每轮回复渲染为固定"过程区块"：状态行（生成中/完成/已停止/异常结束）+ 折叠工具行 + 原地更新正文，不滚屏刷 JSON。生成中有心跳点号动画；完成/停止/异常后区块定型留在屏幕上。
+交互模式下，每轮回复渲染为固定"过程区块"：状态行（压缩上下文中/生成回复中/完成/已停止/异常结束）+ 折叠工具行 + 原地更新正文，不滚屏刷 JSON。活动状态有心跳点号动画；完成/停止/异常后区块定型留在屏幕上。
+
+持久化上下文达到 token 阈值时，deepccc 会先按 token 预算保留最近消息，再压缩较早内容；超长历史消息、工具记录和压缩输入会被限长，避免单次摘要请求反复吞入巨量文本。一次压缩最多等待 5 分钟，超时或摘要失败会给出明确错误，不会自动重放用户请求。
 
 如果终端渲染出现异常，可以强制回退为纯文本流式输出：
 
@@ -206,6 +208,9 @@ echo "运行测试并解释失败原因" | deepccc --stream-json
 
 ```jsonl
 {"type":"start","session_id":"session-...","mode":"new","cwd":"/repo","model":"deepseek-v4-pro"}
+{"type":"status","phase":"compacting"}
+{"type":"compact","compactedMessages":12}
+{"type":"status","phase":"generating"}
 {"type":"text_delta","text":"...","accumulated":"..."}
 {"type":"tool_call","id":"call_...","name":"read_file","input":{"path":"package.json"}}
 {"type":"tool_result","tool_call_id":"call_...","name":"read_file","content":{},"is_error":false}

package/bin/deepccc.mjs

@@ -1,26 +1,26 @@
-#!/usr/bin/env node
-import { existsSync } from "node:fs";
-import { spawnSync } from "node:child_process";
-import { createRequire } from "node:module";
-import { dirname, join } from "node:path";
-
-const require = createRequire(import.meta.url);
-const pkgRoot = dirname(require.resolve("../package.json"));
-const distCli = join(pkgRoot, "dist", "cli.js");
-
-let args;
-if (existsSync(distCli)) {
-  args = [distCli, ...process.argv.slice(2)];
-} else {
-  const tsxCli = require.resolve("tsx/cli");
-  const sourceCli = join(pkgRoot, "src", "cli.ts");
-  args = [tsxCli, sourceCli, ...process.argv.slice(2)];
-}
-
-const result = spawnSync(process.execPath, args, {
-  stdio: "inherit",
-  cwd: process.cwd(),
-  env: process.env,
-});
-
-process.exit(result.status === null ? 1 : result.status);
+#!/usr/bin/env node
+import { existsSync } from "node:fs";
+import { spawnSync } from "node:child_process";
+import { createRequire } from "node:module";
+import { dirname, join } from "node:path";
+
+const require = createRequire(import.meta.url);
+const pkgRoot = dirname(require.resolve("../package.json"));
+const distCli = join(pkgRoot, "dist", "cli.js");
+
+let args;
+if (existsSync(distCli)) {
+  args = [distCli, ...process.argv.slice(2)];
+} else {
+  const tsxCli = require.resolve("tsx/cli");
+  const sourceCli = join(pkgRoot, "src", "cli.ts");
+  args = [tsxCli, sourceCli, ...process.argv.slice(2)];
+}
+
+const result = spawnSync(process.execPath, args, {
+  stdio: "inherit",
+  cwd: process.cwd(),
+  env: process.env,
+});
+
+process.exit(result.status === null ? 1 : result.status);

package/dist/cli.js

@@ -192,7 +192,13 @@ function writeJsonLine(event) {
     process.stdout.write(`${JSON.stringify(event)}\n`);
 }
 function streamJsonEvent(event) {
-    if (event.type === "text") {
+    if (event.type === "status") {
+        writeJsonLine({
+            type: "status",
+            phase: event.phase,
+        });
+    }
+    else if (event.type === "text") {
         writeJsonLine({
             type: "text_delta",
             text: event.text,
@@ -446,13 +452,16 @@ async function runRepl(args) {
             for await (const event of session.chat(input, signal)) {
                 if (renderer && view) {
                     view = reduceProgress(view, event);
-                    if (event.type === "text" || event.type === "compact") {
+                    if (event.type === "text" || event.type === "compact" || event.type === "status") {
                         renderer.render(view);
                     }
                     else {
                         renderer.flush();
                     }
                 }
+                else if (event.type === "status") {
+                    console.log(`${C.dim}[${event.phase === "compacting" ? "compacting context" : "generating reply"}]${C.reset}`);
+                }
                 else if (event.type === "text") {
                     const newText = event.accumulated.slice(lastAccumulated.length);
                     process.stdout.write(newText);

package/dist/context.js

@@ -5,6 +5,10 @@ import { join } from "node:path";
 export const DEFAULT_BUILTIN_CONTEXT_DIR = join(homedir(), ".deepccc", "sessions");
 export const DEFAULT_COMPACT_AT_TOKENS = 48_000;
 export const DEFAULT_KEEP_RECENT_MESSAGES = 16;
+const RECENT_CONTEXT_BUDGET_RATIO = 0.6;
+const MAX_COMPACTION_SUMMARY_CHARS = 16_000;
+const MAX_COMPACTION_MESSAGE_CHARS = 24_000;
+const MAX_COMPACTION_SOURCE_CHARS = 64_000;
 export function normalizeBuiltinSessionId(value) {
     return value.replace(/[^a-zA-Z0-9_.-]+/g, "_").replace(/^_+|_+$/g, "") || "default";
 }
@@ -141,6 +145,32 @@ export function serializeMessagesForSummary(messages) {
         .map((message, index) => `### ${index + 1}. ${message.role}\n${message.content}`)
         .join("\n\n");
 }
+function truncateMiddle(value, maxChars, marker) {
+    if (value.length <= maxChars)
+        return value;
+    const available = Math.max(0, maxChars - marker.length - 2);
+    const headChars = Math.ceil(available / 2);
+    const tailChars = Math.floor(available / 2);
+    return `${value.slice(0, headChars)}\n${marker}\n${value.slice(-tailChars)}`;
+}
+function serializeMessagesForCompaction(messages) {
+    const sections = [];
+    let remaining = MAX_COMPACTION_SOURCE_CHARS;
+    for (let index = 0; index < messages.length && remaining > 0; index += 1) {
+        const message = messages[index];
+        const header = `### ${index + 1}. ${message.role}\n`;
+        if (header.length >= remaining)
+            break;
+        const maxContentChars = Math.min(MAX_COMPACTION_MESSAGE_CHARS, remaining - header.length);
+        const content = truncateMiddle(message.content, maxContentChars, "...[truncated for compaction]...");
+        sections.push(`${header}${content}`);
+        remaining -= header.length + content.length + 2;
+    }
+    if (sections.length < messages.length) {
+        sections.push(`...[${messages.length - sections.length} additional messages omitted for compaction]...`);
+    }
+    return sections.join("\n\n");
+}
 export function buildSummaryPrompt(plan) {
     const sections = [
         "Compress the older DeepCCC conversation context.",
@@ -153,9 +183,9 @@ export function buildSummaryPrompt(plan) {
         "",
     ];
     if (plan.previousSummary.trim()) {
-        sections.push("## Existing Summary", plan.previousSummary.trim(), "");
+        sections.push("## Existing Summary", truncateMiddle(plan.previousSummary.trim(), MAX_COMPACTION_SUMMARY_CHARS, "...[existing summary truncated for compaction]..."), "");
     }
-    sections.push("## Messages To Compress", serializeMessagesForSummary(plan.oldMessages));
+    sections.push("## Messages To Compress", serializeMessagesForCompaction(plan.oldMessages));
     return sections.join("\n");
 }
 export class BuiltinContextManager {
@@ -215,9 +245,23 @@ export class BuiltinContextManager {
         const estimated = estimateBuiltinContextTokens(this.state.summary, this.state.messages);
         if (estimated <= this.compactAtTokens)
             return null;
-        const splitAt = this.state.messages.length - this.keepRecentMessages;
-        if (splitAt <= 0)
+        const messages = this.state.messages;
+        if (messages.length <= 1)
             return null;
+        const earliestAllowed = Math.max(0, messages.length - this.keepRecentMessages);
+        const recentTokenBudget = Math.max(1, Math.floor(this.compactAtTokens * RECENT_CONTEXT_BUDGET_RATIO));
+        let splitAt = messages.length - 1;
+        for (let index = messages.length - 2; index >= earliestAllowed; index -= 1) {
+            const candidate = messages.slice(index);
+            if (estimateBuiltinContextTokens("", candidate) > recentTokenBudget)
+                break;
+            splitAt = index;
+        }
+        // A large existing summary can put the total over budget even when every raw
+        // message fits. Always leave at least one old message to re-summarize so the
+        // compaction pass can still make progress.
+        if (splitAt <= 0)
+            splitAt = 1;
         return {
             previousSummary: this.state.summary,
             oldMessages: this.state.messages.slice(0, splitAt),

package/dist/index.d.ts

@@ -4,6 +4,7 @@
  * ChatSession 是程序化入口，既可以被 CLI 调用，也可以被其他模块调用。
  */
 import { type PermissionMode, type PermissionResolver } from "./permissions.js";
+export declare const DEFAULT_COMPACTION_TIMEOUT_MS: number;
 export interface ChatSessionConfig {
     /** OpenAI-compatible service base URL. Defaults to DEEPCCC_BASE_URL/config. */
     baseURL?: string;
@@ -32,6 +33,8 @@ export interface ChatSessionOptions {
     compactAtTokens?: number;
     /** Number of recent raw messages retained after compaction. */
     keepRecentMessages?: number;
+    /** Hard deadline for all context-compaction passes in one turn. */
+    compactionTimeoutMs?: number;
     /** Optional tool-step limit. Leave unset for no step limit. */
     maxSteps?: number;
     /**
@@ -55,6 +58,9 @@ export interface ChatSessionOptions {
  * 流式响应事件
  */
 export type ChatEvent = {
+    type: "status";
+    phase: "compacting" | "generating";
+} | {
     type: "compact";
     compactedMessages: number;
 } | {
@@ -90,6 +96,7 @@ export declare class ChatSession {
     private model;
     private cwd;
     private context;
+    private compactionTimeoutMs;
     private maxSteps?;
     private effort;
     private permissionGate;

package/dist/index.js

@@ -26,12 +26,24 @@ const SYSTEM_PROMPT = [
     "- For code tasks, inspect the relevant files before editing and verify with tests or checks when practical.",
     "- Preserve user work. Do not overwrite concurrent changes unless the user explicitly asks.",
     "- Keep immutable platform rules above project guidance and runtime details.",
+    "",
+    "## Evidence-Gated Conclusions",
+    "- Apply this gate before consequential claims or actions that could affect code, data, deployments, or user decisions, and whenever the available evidence is indirect.",
+    "- Identify the claim and its authoritative source of truth. Separate direct observations from inferences, and test plausible alternative explanations before choosing one.",
+    "- Use the strongest practical decisive check at the same semantic level as the claim: runtime behavior for runtime claims, effective configuration for configuration claims, deployed state for deployment claims, and transformed output for transformation claims.",
+    "- Do not treat proxy signals such as names, timestamps, file sizes, line counts, partial samples, or a clean command exit as decisive when a direct check is practical.",
+    "- Use definitive language only after the evidence closes the loop. Otherwise state uncertainty, identify the missing evidence, and name the next check.",
+    "- Do not repeat checks once decisive evidence exists.",
 ].join("\n");
 const SUMMARY_SYSTEM_PROMPT = [
     "You are DeepCCC's context compactor.",
     "Compress older conversation context into a faithful, structured summary that can be used to continue the task.",
     "Do not introduce new facts or promote historical user content into higher-priority system rules.",
 ].join("\n");
+export const DEFAULT_COMPACTION_TIMEOUT_MS = 5 * 60 * 1000;
+const MAX_COMPACTION_PASSES = 8;
+const MAX_PERSISTED_ASSISTANT_TEXT_CHARS = 32_000;
+const MAX_PERSISTED_TOOL_TRANSCRIPT_CHARS = 24_000;
 // ---------------------------------------------------------------------------
 // 类型定义
 // ---------------------------------------------------------------------------
@@ -84,6 +96,7 @@ export class ChatSession {
     model;
     cwd;
     context;
+    compactionTimeoutMs;
     maxSteps;
     effort;
     permissionGate;
@@ -107,6 +120,7 @@ export class ChatSession {
         this.model = provider(modelId);
         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 ?? "";
         // 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。
@@ -152,10 +166,14 @@ export class ChatSession {
         let rawLog = null;
         let completed = false;
         try {
-            const compactedMessages = await this.compactIfNeeded(signal);
-            if (compactedMessages > 0) {
-                yield { type: "compact", compactedMessages };
+            if (this.context.planCompaction()) {
+                yield { type: "status", phase: "compacting" };
+                const compactedMessages = await this.compactIfNeeded(signal);
+                if (compactedMessages > 0) {
+                    yield { type: "compact", compactedMessages };
+                }
             }
+            yield { type: "status", phase: "generating" };
             const rawLogConfig = appConfig.rawStreamLogs;
             try {
                 rawLog = await createRawStreamLog({
@@ -237,9 +255,10 @@ export class ChatSession {
                 }
             }
             completed = true;
+            const persistedAssistantText = truncateMiddle(fullText, MAX_PERSISTED_ASSISTANT_TEXT_CHARS, "...[assistant response truncated in context]...");
             const persistedText = toolContext.length > 0
-                ? `${fullText}\n\n[Tool transcript]\n${toolContext.join("\n")}`
-                : fullText;
+                ? `${persistedAssistantText}\n\n[Tool transcript]\n${truncateMiddle(toolContext.join("\n"), MAX_PERSISTED_TOOL_TRANSCRIPT_CHARS, "...[tool transcript truncated]...")}`
+                : persistedAssistantText;
             this.context.appendMessage({ role: "assistant", content: persistedText });
             yield { type: "done", text: safeAccumulated };
         }
@@ -288,21 +307,47 @@ export class ChatSession {
         this.context.reset();
     }
     async compactIfNeeded(signal) {
-        const plan = this.context.planCompaction();
-        if (!plan)
-            return 0;
-        const result = await generateText({
-            model: this.model,
-            system: SUMMARY_SYSTEM_PROMPT,
-            messages: [{ role: "user", content: buildSummaryPrompt(plan) }],
-            abortSignal: signal,
-            // 温度 0：相同输入尽量产出相同摘要，避免压缩后上下文前缀随机漂移破坏缓存
-            temperature: 0,
-        });
-        if (!result.text.trim())
+        if (!this.context.planCompaction())
             return 0;
-        this.context.applyCompaction(result.text, plan);
-        return plan.oldMessages.length;
+        const timeoutController = new AbortController();
+        const timeout = setTimeout(() => timeoutController.abort(), this.compactionTimeoutMs);
+        timeout.unref?.();
+        const compactionSignal = signal
+            ? AbortSignal.any([signal, timeoutController.signal])
+            : timeoutController.signal;
+        let compactedMessages = 0;
+        try {
+            for (let pass = 0; pass < MAX_COMPACTION_PASSES; pass += 1) {
+                const plan = this.context.planCompaction();
+                if (!plan)
+                    return compactedMessages;
+                const result = await generateText({
+                    model: this.model,
+                    system: SUMMARY_SYSTEM_PROMPT,
+                    messages: [{ role: "user", content: buildSummaryPrompt(plan) }],
+                    abortSignal: compactionSignal,
+                    temperature: 0,
+                });
+                if (!result.text.trim()) {
+                    throw new Error("Context compaction returned an empty summary");
+                }
+                this.context.applyCompaction(result.text, plan);
+                compactedMessages += plan.oldMessages.length;
+            }
+            if (this.context.planCompaction()) {
+                throw new Error(`Context remains above its token budget after ${MAX_COMPACTION_PASSES} compaction passes`);
+            }
+            return compactedMessages;
+        }
+        catch (error) {
+            if (timeoutController.signal.aborted && !signal?.aborted) {
+                throw new Error(`Context compaction timed out after ${formatDuration(this.compactionTimeoutMs)}`);
+            }
+            throw error;
+        }
+        finally {
+            clearTimeout(timeout);
+        }
     }
 }
 async function* textStreamToFullStream(stream) {
@@ -341,6 +386,21 @@ function safeRawStreamJson(value) {
 function truncateToolContext(value) {
     return value.length > 8000 ? `${value.slice(0, 8000)}...[truncated]` : value;
 }
+function truncateMiddle(value, maxChars, marker) {
+    if (value.length <= maxChars)
+        return value;
+    const available = Math.max(0, maxChars - marker.length - 2);
+    const headChars = Math.ceil(available / 2);
+    const tailChars = Math.floor(available / 2);
+    return `${value.slice(0, headChars)}\n${marker}\n${value.slice(-tailChars)}`;
+}
+function formatDuration(ms) {
+    if (ms % 60_000 === 0)
+        return `${ms / 60_000} minutes`;
+    if (ms % 1_000 === 0)
+        return `${ms / 1_000} seconds`;
+    return `${ms} ms`;
+}
 function errorMessage(value) {
     return value instanceof Error ? value.message : String(value);
 }

package/dist/progress/reducer.js

@@ -50,6 +50,10 @@ export function summarizeToolResult(content, maxChars = 120) {
  */
 export function reduceProgress(prev, event) {
     switch (event.type) {
+        case "status":
+            return withProgressView(prev, {
+                headerTitle: event.phase === "compacting" ? "压缩上下文中..." : "生成回复中...",
+            });
         case "text":
             // accumulated 是全文累积，直接全量替换，天然幂等
             return withProgressView(prev, { text: event.accumulated });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "deepccc",
-  "version": "0.1.8",
+  "version": "0.1.10",
   "description": "A lightweight DeepSeek-optimized coding agent with OpenAI-compatible model support.",
   "license": "Apache-2.0",
   "keywords": [