feat: add contextWindow and autoCompactWindow settings with related functionality

Author: qorzj

docs/configuration.md

@@ -27,6 +27,8 @@ Deep Code 使用 `settings.json` 设置文件进行持久化配置，支持两
 | 字段                 | 类型      | 说明                                                                |
 | -------------------- | --------- | ------------------------------------------------------------------- |
 | `env`                | object    | 环境变量分组（见下方子字段表）                                       |
+| `contextWindow`     | number/string | 上下文窗口上限，可使用精确 token 数或 `128K`、`1M` 等格式          |
+| `autoCompactWindow` | number/string | 自动压缩阈值，默认取最终上下文窗口的 50%                           |
 | `model`              | string    | 模型名称。优先级高于 `env.MODEL`                                    |
 | `thinkingEnabled`    | boolean   | 是否启用思考模式（DeepSeek V4 系列默认启用）                         |
 | `reasoningEffort`    | string    | 推理强度，可选 `"high"` 或 `"max"`（默认 `"max"`）                  |
@@ -53,6 +55,19 @@ Deep Code 使用 `settings.json` 设置文件进行持久化配置，支持两
 | `TELEMETRY_ENABLED`  | string | 是否启用匿名使用数据上报                                   |
 | `<其他任意KEY>` | string | 自定义环境变量 |
 
+#### 上下文窗口
+
+`contextWindow` 和 `autoCompactWindow` 是 `settings.json` 的顶层字段。number 必须是正整数，表示精确 token 数；string 使用大小写不敏感的 `K` 或 `M` 后缀，按 `1K = 1024`、`1M = 1024²` 换算：
+
+```json
+{
+  "contextWindow": "1M",
+  "autoCompactWindow": "512K"
+}
+```
+
+普通模型的默认上下文窗口为 `256K`，DeepSeek V4 系列为 `1M`。未设置自动压缩阈值时取最终上下文窗口的 50%；无效值会被忽略，自动压缩阈值超过上下文窗口时会限制为上下文窗口。
+
 #### `thinkingEnabled` — 思考模式
 
 是否启用 DeepSeek 思考模式。设置为 `true` 启用、`false` 禁用。

docs/configuration_en.md

@@ -27,6 +27,8 @@ The following are all the top-level fields supported in `settings.json`, along w
 | Field              | Type    | Description                                                                 |
 | ------------------ | ------- | --------------------------------------------------------------------------- |
 | `env`              | object  | Group of environment variables (see sub-field table below)                 |
+| `contextWindow`   | number/string | Context-window limit as an exact token count or `128K`/`1M` value   |
+| `autoCompactWindow` | number/string | Auto-compaction threshold; defaults to 50% of the final context window |
 | `model`            | string  | Model name. Takes precedence over `env.MODEL`                              |
 | `thinkingEnabled`  | boolean | Whether to enable thinking mode (enabled by default for DeepSeek V4 series)|
 | `reasoningEffort`  | string  | Reasoning intensity, either `"high"` or `"max"` (default `"max"`)          |
@@ -53,6 +55,19 @@ The following are all the top-level fields supported in `settings.json`, along w
 | `TELEMETRY_ENABLED`| string| Enable anonymous usage reporting                                |
 | `<any other KEY>` | string | Custom environment variable                                     |
 
+#### Context Windows
+
+`contextWindow` and `autoCompactWindow` are top-level `settings.json` fields. A number must be a positive integer and represents an exact token count. A string uses a case-insensitive `K` or `M` suffix, with `1K = 1024` and `1M = 1024²`:
+
+```json
+{
+  "contextWindow": "1M",
+  "autoCompactWindow": "512K"
+}
+```
+
+The default context window is `256K` for regular models and `1M` for DeepSeek V4 models. If the auto-compaction threshold is omitted, it is 50% of the final context window. Invalid values are ignored, and an auto-compaction threshold larger than the context window is capped at the context window.
+
 #### `thinkingEnabled` — Thinking Mode
 
 Whether to enable DeepSeek thinking mode. Set to `true` to enable, `false` to disable.

packages/cli/src/tests/exec-runner.test.ts

@@ -25,6 +25,8 @@ function createSettings(
     env: {},
     baseURL: "https://example.invalid",
     model: "test-model",
+    contextWindow: 256 * 1024,
+    autoCompactWindow: 128 * 1024,
     thinkingEnabled: false,
     reasoningEffort: "high",
     debugLogEnabled: false,

packages/cli/src/tests/ui-utils.test.ts

@@ -0,0 +1,54 @@
+import { test } from "node:test";
+import assert from "node:assert/strict";
+import type { SessionEntry } from "@vegamo/deepcode-core";
+import { buildStatusLine, formatContextUsage, formatTokenCount } from "../ui/utils/index";
+
+test("formatTokenCount uses binary K/M units with at most one decimal", () => {
+  assert.equal(formatTokenCount(0), "0");
+  assert.equal(formatTokenCount(1023), "1023");
+  assert.equal(formatTokenCount(1126), "1.1K");
+  assert.equal(formatTokenCount(1024 * 1024), "1M");
+  assert.equal(formatTokenCount(1.25 * 1024 * 1024), "1.3M");
+});
+
+test("formatContextUsage rounds a ten-cell bar to whole blocks", () => {
+  assert.equal(formatContextUsage(0, 1000), "0/1000 [░░░░░░░░░░] 0%");
+  assert.equal(formatContextUsage(20, 1000), "20/1000 [░░░░░░░░░░] 2%");
+  assert.equal(formatContextUsage(200, 1000), "200/1000 [▓▓░░░░░░░░] 20%");
+  assert.equal(formatContextUsage(550, 1000), "550/1000 [▓▓▓▓▓▓░░░░] 55%");
+});
+
+test("formatContextUsage caps the bar and percentage at 100 percent", () => {
+  assert.equal(formatContextUsage(1200, 1000), "1.2K/1000 [▓▓▓▓▓▓▓▓▓▓] 100%");
+});
+
+test("buildStatusLine replaces the tokens label while preserving status and failure", () => {
+  const entry: SessionEntry = {
+    id: "session-1",
+    summary: null,
+    assistantReply: null,
+    assistantThinking: null,
+    assistantRefusal: null,
+    toolCalls: null,
+    status: "failed",
+    failReason: "boom",
+    usage: null,
+    usagePerModel: null,
+    activeTokens: 20,
+    createTime: "2026-01-01T00:00:00.000Z",
+    updateTime: "2026-01-01T00:00:01.000Z",
+    processes: null,
+  };
+
+  assert.equal(buildStatusLine(entry, 1000), "status: failed · 20/1000 [░░░░░░░░░░] 2% · fail: boom");
+});
+
+test("buildStatusLine omits context usage when no active tokens exist", () => {
+  const entry = {
+    status: "pending",
+    activeTokens: 0,
+    failReason: null,
+  } as SessionEntry;
+
+  assert.equal(buildStatusLine(entry, 1024 * 1024), "status: pending");
+});

packages/cli/src/ui/utils/index.ts

@@ -77,11 +77,36 @@ export function isCurrentSessionEmpty(sessionManager: SessionManager): boolean {
   return !activeSessionId || !sessionManager.getSession(activeSessionId);
 }
 
-export function buildStatusLine(entry: SessionEntry): string {
+const CONTEXT_BAR_WIDTH = 10;
+
+export function formatTokenCount(tokens: number): string {
+  if (!Number.isFinite(tokens) || tokens <= 0) {
+    return "0";
+  }
+  if (tokens < 1024) {
+    return String(Math.round(tokens));
+  }
+
+  const unit = tokens >= 1024 * 1024 ? "M" : "K";
+  const divisor = unit === "M" ? 1024 * 1024 : 1024;
+  return `${Number((tokens / divisor).toFixed(1))}${unit}`;
+}
+
+export function formatContextUsage(activeTokens: number, contextWindow: number): string {
+  const safeActiveTokens = Number.isFinite(activeTokens) ? Math.max(0, activeTokens) : 0;
+  const ratio = Number.isFinite(contextWindow) && contextWindow > 0 ? safeActiveTokens / contextWindow : 0;
+  const cappedRatio = Math.min(1, ratio);
+  const filledBlocks = Math.round(cappedRatio * CONTEXT_BAR_WIDTH);
+  const bar = `${"▓".repeat(filledBlocks)}${"░".repeat(CONTEXT_BAR_WIDTH - filledBlocks)}`;
+  const percent = Math.min(100, Math.round(ratio * 100));
+  return `${formatTokenCount(safeActiveTokens)}/${formatTokenCount(contextWindow)} [${bar}] ${percent}%`;
+}
+
+export function buildStatusLine(entry: SessionEntry, contextWindow: number): string {
   const parts: string[] = [];
   parts.push(`status: ${entry.status}`);
   if (typeof entry.activeTokens === "number" && entry.activeTokens > 0) {
-    parts.push(`tokens: ${entry.activeTokens}`);
+    parts.push(formatContextUsage(entry.activeTokens, contextWindow));
   }
   if (entry.failReason) {
     parts.push(`fail: ${entry.failReason}`);

packages/cli/src/ui/views/App.tsx

@@ -48,7 +48,6 @@ import type {
   UserPromptContent,
 } from "@vegamo/deepcode-core";
 import { SessionManager } from "@vegamo/deepcode-core";
-import { getCompactPromptTokenThreshold } from "@vegamo/deepcode-core";
 import { writeStdout, writeStdoutLine } from "../../utils/stdio-helpers";
 
 type View = "chat" | "session-list" | "undo" | "mcp-status";
@@ -154,7 +153,7 @@ function App({ projectRoot, initialPrompt, resumeSessionId, onRestart }: AppProp
         }
       },
       onSessionEntryUpdated: (entry) => {
-        setStatusLine(buildStatusLine(entry));
+        setStatusLine(buildStatusLine(entry, resolveCurrentSettings(projectRoot).contextWindow));
         setRunningProcesses(entry.processes);
         setActiveStatus(entry.status);
         setActiveAskPermissions(entry.askPermissions);
@@ -545,7 +544,7 @@ function App({ projectRoot, initialPrompt, resumeSessionId, onRestart }: AppProp
       // Clear first so <Static> resets its index to 0.
       await resetStaticView(loadVisibleMessages(sessionManager, sessionId), { clearScreen: true });
       const session = sessionManager.getSession(sessionId);
-      setStatusLine(session ? buildStatusLine(session) : "");
+      setStatusLine(session ? buildStatusLine(session, resolveCurrentSettings(projectRoot).contextWindow) : "");
       setRunningProcesses(session?.processes ?? null);
       setActiveStatus(session?.status ?? null);
       setActiveAskPermissions(session?.askPermissions);
@@ -556,7 +555,7 @@ function App({ projectRoot, initialPrompt, resumeSessionId, onRestart }: AppProp
       }
       await refreshSkills(sessionId);
     },
-    [sessionManager, resetStaticView, pendingPermissionReply, refreshSkills]
+    [sessionManager, resetStaticView, pendingPermissionReply, projectRoot, refreshSkills]
   );
 
   /**
@@ -733,7 +732,7 @@ function App({ projectRoot, initialPrompt, resumeSessionId, onRestart }: AppProp
     const model = settings.model || "";
     const thinkingEnabled = settings.thinkingEnabled;
     const reasoningEffort = settings.reasoningEffort;
-    const maxContextTokens = getCompactPromptTokenThreshold(model);
+    const maxContextTokens = settings.contextWindow;
     if (!activeSessionId) {
       return {
         activeSessionId: null,

packages/core/src/index.ts

@@ -14,6 +14,8 @@ export {
   modelConfigKey,
   getUserSettingsPath,
   getProjectSettingsPath,
+  getDefaultContextWindow,
+  getDefaultAutoCompactWindow,
   DEFAULT_MODEL,
   DEFAULT_BASE_URL,
 } from "./settings";

packages/core/src/session.ts

@@ -7,7 +7,6 @@ import ejs from "ejs";
 import type { ChatCompletionMessageParam } from "openai/resources/chat/completions";
 import { launchNotifyScript } from "./common/notify";
 import { buildThinkingRequestOptions } from "./common/openai-thinking";
-import { DEEPSEEK_V4_MODELS } from "./common/model-capabilities";
 import { readTextFileWithMetadata } from "./common/file-utils";
 import {
   buildSkillDocumentsPrompt,
@@ -29,7 +28,12 @@ import {
   type ToolExecutionHooks,
 } from "./tools/executor";
 import { McpManager } from "./mcp/mcp-manager";
-import type { McpServerConfig, PermissionScope, PermissionSettings } from "./settings";
+import {
+  getDefaultAutoCompactWindow,
+  type McpServerConfig,
+  type PermissionScope,
+  type PermissionSettings,
+} from "./settings";
 import { logApiError } from "./common/error-logger";
 import { logOpenAIChatCompletionDebug, normalizeDebugError } from "./common/debug-logger";
 import { describeLlmError, getLlmErrorDetails } from "./common/llm-error";
@@ -66,8 +70,6 @@ const MAX_SESSION_ENTRIES = 50;
 const MAX_PROJECT_CODE_LENGTH = 64;
 const PROJECT_CODE_HASH_LENGTH = 16;
 const BACKGROUND_FAILURE_LOG_TAIL_CHARS = 4000;
-const DEFAULT_COMPACT_PROMPT_TOKEN_THRESHOLD = 128 * 1024;
-const DEEPSEEK_V4_COMPACT_PROMPT_TOKEN_THRESHOLD = 512 * 1024;
 const PLAN_MODE_ON_STATUS_MESSAGE = "  └ Set Plan Mode on. Awaiting <proposed_plan>.";
 const PLAN_MODE_OFF_STATUS_MESSAGE = "  └ Set Plan Mode off.";
 const PLAN_MODE_FORCE_ASK_SCOPES = [
@@ -86,9 +88,7 @@ type ChatCompletionDebugOptions = {
 };
 
 export function getCompactPromptTokenThreshold(model: string): number {
-  return DEEPSEEK_V4_MODELS.has(model)
-    ? DEEPSEEK_V4_COMPACT_PROMPT_TOKEN_THRESHOLD
-    : DEFAULT_COMPACT_PROMPT_TOKEN_THRESHOLD;
+  return getDefaultAutoCompactWindow(model);
 }
 
 // Keep project storage paths short enough for Git's internal files on Windows.
@@ -309,6 +309,8 @@ export type SessionManagerOptions = {
   createOpenAIClient: CreateOpenAIClient;
   getResolvedSettings: () => {
     model: string;
+    contextWindow?: number;
+    autoCompactWindow?: number;
     webSearchTool?: string;
     mcpServers?: Record<string, McpServerConfig>;
     permissions?: Required<PermissionSettings>;
@@ -337,6 +339,8 @@ export class SessionManager {
   private readonly createOpenAIClient: CreateOpenAIClient;
   private readonly getResolvedSettings: () => {
     model: string;
+    contextWindow?: number;
+    autoCompactWindow?: number;
     webSearchTool?: string;
     mcpServers?: Record<string, McpServerConfig>;
     permissions?: Required<PermissionSettings>;
@@ -1362,7 +1366,8 @@ ${agentInstructions}
           }
         }
 
-        const compactPromptTokenThreshold = getCompactPromptTokenThreshold(model);
+        const compactPromptTokenThreshold =
+          this.getResolvedSettings().autoCompactWindow ?? getCompactPromptTokenThreshold(model);
         if (session.activeTokens > compactPromptTokenThreshold) {
           const message = this.buildAssistantMessage(
             sessionId,

packages/core/src/settings.ts

@@ -1,4 +1,4 @@
-import { defaultsToThinkingMode } from "./common/model-capabilities";
+import { DEEPSEEK_V4_MODELS, defaultsToThinkingMode } from "./common/model-capabilities";
 import * as fs from "fs";
 import * as os from "os";
 import * as path from "path";
@@ -82,6 +82,8 @@ export type ResolvedStatusLineSettings = {
 
 export type DeepcodingSettings = {
   env?: DeepcodingEnv;
+  contextWindow?: number | string;
+  autoCompactWindow?: number | string;
   model?: string;
   temperature?: number;
   thinkingEnabled?: boolean;
@@ -101,6 +103,8 @@ export type ResolvedDeepcodingSettings = {
   apiKey?: string;
   baseURL: string;
   model: string;
+  contextWindow: number;
+  autoCompactWindow: number;
   temperature?: number;
   thinkingEnabled: boolean;
   reasoningEffort: ReasoningEffort;
@@ -122,6 +126,45 @@ export type ModelConfigSelection = {
 
 export type SettingsProcessEnv = Record<string, string | undefined>;
 
+const DEFAULT_CONTEXT_WINDOW = 256 * 1024;
+const DEEPSEEK_V4_CONTEXT_WINDOW = 1024 * 1024;
+
+export function getDefaultContextWindow(model: string): number {
+  return DEEPSEEK_V4_MODELS.has(model) ? DEEPSEEK_V4_CONTEXT_WINDOW : DEFAULT_CONTEXT_WINDOW;
+}
+
+export function getDefaultAutoCompactWindow(model: string): number {
+  return getDefaultContextWindow(model) / 2;
+}
+
+function parseTokenWindow(value: unknown): number | undefined {
+  if (typeof value === "number") {
+    return Number.isSafeInteger(value) && value > 0 ? value : undefined;
+  }
+  if (typeof value !== "string") {
+    return undefined;
+  }
+
+  const match = /^(\d+)([km])$/i.exec(value.trim());
+  if (!match) {
+    return undefined;
+  }
+  const amount = Number(match[1]);
+  const multiplier = match[2]?.toLowerCase() === "m" ? 1024 * 1024 : 1024;
+  const tokens = amount * multiplier;
+  return Number.isSafeInteger(tokens) && tokens > 0 ? tokens : undefined;
+}
+
+function firstTokenWindow(...values: unknown[]): number | undefined {
+  for (const value of values) {
+    const parsed = parseTokenWindow(value);
+    if (parsed !== undefined) {
+      return parsed;
+    }
+  }
+  return undefined;
+}
+
 function resolveReasoningEffort(value: unknown): ReasoningEffort | undefined {
   return value === "high" || value === "max" ? value : undefined;
 }
@@ -484,6 +527,17 @@ export function resolveSettingsSources(
     trimString(userEnv.MODEL) ||
     defaults.model;
 
+  const contextWindow =
+    firstTokenWindow(systemEnv.CONTEXT_WINDOW, projectSettings?.contextWindow, userSettings?.contextWindow) ??
+    getDefaultContextWindow(model);
+  const configuredAutoCompactWindow = firstTokenWindow(
+    systemEnv.AUTO_COMPACT_WINDOW,
+    projectSettings?.autoCompactWindow,
+    userSettings?.autoCompactWindow
+  );
+  const defaultAutoCompactWindow = Math.max(1, Math.floor(contextWindow / 2));
+  const autoCompactWindow = Math.min(configuredAutoCompactWindow ?? defaultAutoCompactWindow, contextWindow);
+
   const thinkingEnabled =
     parseBoolean(systemEnv.THINKING_ENABLED) ??
     parseBoolean(projectSettings?.thinkingEnabled) ??
@@ -536,6 +590,8 @@ export function resolveSettingsSources(
     apiKey: trimString(env.API_KEY) || undefined,
     baseURL: trimString(env.BASE_URL) || defaults.baseURL,
     model,
+    contextWindow,
+    autoCompactWindow,
     temperature,
     thinkingEnabled,
     reasoningEffort,