minimal-agent 0.11.0 → 0.13.0

package/skills/skill-creator/scripts/run_eval.py

@@ -20,14 +20,15 @@ from scripts.utils import parse_skill_md
 
 
 def find_project_root() -> Path:
-    """Find the project root by walking up from cwd looking for .claude/.
+    """Find the project root by walking up from cwd looking for .kakadeai/
+    (this repo's project-config dir) or .claude/ (upstream convention).
 
     Mimics how Claude Code discovers its project root, so the command file
     we create ends up where claude -p will look for it.
     """
     current = Path.cwd()
     for parent in [current, *current.parents]:
-        if (parent / ".claude").is_dir():
+        if (parent / ".kakadeai").is_dir() or (parent / ".claude").is_dir():
             return parent
     return current
 

package/src/cli/print.js

@@ -27,7 +27,9 @@ import { updateSessionMetadata } from '../context/sessionRegistry.js';
 import { workspaceIndexService } from '../context/workspaceIndex.js';
 import { normalizeImageDetail } from '../multimodal.js';
 import { runWithPlugins } from '../plugins/pluginRunner.js';
-import { buildTransientPromptBundle } from '../prompts/system.js';
+import { buildBaseSystemPrompt, buildTransientPromptBundle } from '../prompts/system.js';
+import { getOrCreateSessionState, sessionStateKey } from '../session/sessionState.js';
+import { hydrateSessionWorktree } from '../tools/worktree/sessionWorktree.js';
 import { ALL_TOOLS } from '../tools/index.js';
 import { createPromptCacheContext } from '../llm/promptCache.js';
 import { createAgentRuntimeContext, filterToolsForRuntime, memoryContextFromRuntime, } from '../runtime/agentRuntimeContext.js';
@@ -106,8 +108,16 @@ export function extractImageOptions(args) {
  *   stop_reason = end_turn | max_turns | stalled | error | interrupted | config_error
  *   num_turns   = assistant_message 事件计数
  *   error       = 失败文案（成功时 null）
+ *   warnings    = 本轮所有非致命警告（与 stderr 同源；宿主要结构化的那一份）
  *   session_file（可选）= 当前对话落盘路径；拿得到才填，拿不到省略该键
  *
+ *   completionStatus = 本层从 stop_reason + deliverables 重算：跑完了没、产物验没验
+ *   delivery_status（可选）= 内核 turn_done 的交付判定，**回答的是另一个问题**：
+ *     completed / delivery_incomplete / delivery_unverified / claims_unverified。
+ *     两者并列而非替代 —— `ok:true` 配 `delivery_status:'claims_unverified'` 是
+ *     完全合法的一行：Agent 确实正常收尾了，但它宣称的「测试已通过」没有证据支撑。
+ *     判断「任务是否真做成」的宿主应当读它，而不是只看 ok。
+ *
  * 导出供单测断言。
  */
 export function emitJsonResult(result, sessionFile) {
@@ -134,6 +144,7 @@ export function emitJsonResult(result, sessionFile) {
         num_turns: result.numTurns,
         error: result.error,
         completionStatus,
+        ...(result.deliveryStatus ? { delivery_status: result.deliveryStatus } : {}),
         deliverables,
         verification: allVerified ? 'passed' : deliverables.length > 0 ? 'not_run' : 'none',
         warnings,
@@ -232,10 +243,16 @@ export async function runPrintMode(provider, args, initialHistory, options) {
         void finalizeOnce().finally(() => process.exit(130));
     });
     try {
-        const cwd = options.workspace?.cwd ?? process.cwd();
-        const workspace = await workspaceIndexService.findByPath(cwd).catch(() => null);
+        const workspaceRoot = options.workspace?.cwd ?? process.cwd();
+        const workspace = await workspaceIndexService.findByPath(workspaceRoot).catch(() => null);
         const sessionId = options.workspace?.sessionId ?? 'default';
-        const conversationKey = `workspace:${workspace?.id ?? cwd}:session:${sessionId}`;
+        const sessionState = getOrCreateSessionState(sessionStateKey({
+            surface: 'cli', workspaceId: workspace?.id,
+            cwd: workspace?.id ? undefined : workspaceRoot, sessionId,
+        }));
+        await hydrateSessionWorktree({ sessionState, workspaceId: workspace?.id, sessionId, cwd: workspaceRoot });
+        const cwd = sessionState.activeWorktree?.path ?? workspaceRoot;
+        const conversationKey = `workspace:${workspace?.id ?? workspaceRoot}:session:${sessionId}`;
         const computerRun = options.computerUseMode && options.computerUseMode !== 'disabled'
             ? createCliComputerRun({
                 mode: options.computerUseMode,
@@ -259,6 +276,11 @@ export async function runPrintMode(provider, args, initialHistory, options) {
             computer: computerRun,
         });
         const turnTools = filterToolsForRuntime(ALL_TOOLS, runtimeContext);
+        const stablePrompt = await buildBaseSystemPrompt(cwd, turnTools);
+        if (history[0]?.role === 'system')
+            history[0] = { ...history[0], content: stablePrompt };
+        else
+            history.unshift({ role: 'system', content: stablePrompt });
         const transientBundle = await buildTransientPromptBundle({
             cwd,
             tools: turnTools,
@@ -266,6 +288,7 @@ export async function runPrintMode(provider, args, initialHistory, options) {
                 surface: 'cli',
                 cwd,
                 workspaceId: workspace?.id,
+                sessionId,
                 interactiveAsk: false,
                 memoryAccess: runtimeContext.memoryAccess,
                 currentUserInput: prompt,
@@ -282,6 +305,7 @@ export async function runPrintMode(provider, args, initialHistory, options) {
             imageMaxLongSidePixel: imageOptions.imageMaxLongSidePixel,
             allowedTools: turnTools.map((tool) => tool.name),
             runtimeContext,
+            sessionState,
             ...(transientBundle ? { transientBundle } : {}),
             promptCache: createPromptCacheContext(`workspace:${workspace?.id ?? cwd}:session:${sessionId}`),
             memoryContext: memoryContextFromRuntime(runtimeContext),
@@ -384,11 +408,19 @@ export function handleEvent(event, result, verbose, outputFormat = 'text') {
             // 通用非致命警告（如 workflow 多余参数、pause 被自动跳过等）始终走 stderr，
             // 不受 verbose 控制 —— 这是用户必须看到的信号，避免静默丢弃语义被忽略。
             process.stderr.write(`${event.message}\n`);
+            // 同时收进结局契约：stderr 是给人看的，spawn 它的宿主要的是结构化的一条。
+            // （此前 PrintResult.warnings 声明了、读了、也输出了，就是没有任何地方写它 ——
+            //  json 那行的 warnings 恒为空数组。）
+            (result.warnings ??= []).push(event.message);
             return {};
         case 'turn_done':
             // text 模式：非 verbose 时这里一次性把累积答案打到 stdout（旧行为，逐字节不变）。
             // json 模式：只记 stopReason='end_turn'，stdout 保持静默。
             result.stopReason = 'end_turn';
+            // 内核的交付判定必须过进程边界：模型谎称「已验证」时，若不记这一笔，
+            // 宿主收到的 JSON 与一切正常逐字节相同，唯一线索只剩 stderr 上一行中文。
+            if (event.completionStatus)
+                result.deliveryStatus = event.completionStatus;
             if (!isJson && !verbose)
                 console.log(result.buffer);
             return {};

package/src/cli/print.ts

@@ -29,7 +29,9 @@ import { workspaceIndexService } from '../context/workspaceIndex.js';
 import { normalizeImageDetail } from '../multimodal.js';
 import type { DeliverableManifest, FileActivityMetadata, HarnessDiagnostics, ImageInputSpec, LoopEvent, Message, Provider } from '../types.js';
 import { runWithPlugins } from '../plugins/pluginRunner.js';
-import { buildTransientPromptBundle } from '../prompts/system.js';
+import { buildBaseSystemPrompt, buildTransientPromptBundle } from '../prompts/system.js';
+import { getOrCreateSessionState, sessionStateKey } from '../session/sessionState.js';
+import { hydrateSessionWorktree } from '../tools/worktree/sessionWorktree.js';
 import { ALL_TOOLS } from '../tools/index.js';
 import { createPromptCacheContext } from '../llm/promptCache.js';
 import {
@@ -96,6 +98,15 @@ export interface PrintResult {
   deliverables?: DeliverableManifest['files'];
   warnings?: string[];
   harnessDiagnostics?: HarnessDiagnostics;
+  /**
+   * 内核 `turn_done` 给出的交付判定（loop.ts 的四态词表）。
+   *
+   * 与下面 json 里那个从 stopReason+deliverables 重算的 `completionStatus` **不是**
+   * 一回事：那个只回答「跑完了没、产物验没验」，这个回答「模型宣称的事到底做没做」。
+   * 尤其 `claims_unverified` —— 模型说「测试已全部通过」但本轮证据台账里没有一条
+   * 成功的验证记录 —— 是重不重要全看它能不能过进程边界。
+   */
+  deliveryStatus?: string;
 }
 
 /** handleEvent 返回的结构 */
@@ -189,8 +200,16 @@ export function extractImageOptions(args: string[]): {
  *   stop_reason = end_turn | max_turns | stalled | error | interrupted | config_error
  *   num_turns   = assistant_message 事件计数
  *   error       = 失败文案（成功时 null）
+ *   warnings    = 本轮所有非致命警告（与 stderr 同源；宿主要结构化的那一份）
  *   session_file（可选）= 当前对话落盘路径；拿得到才填，拿不到省略该键
  *
+ *   completionStatus = 本层从 stop_reason + deliverables 重算：跑完了没、产物验没验
+ *   delivery_status（可选）= 内核 turn_done 的交付判定，**回答的是另一个问题**：
+ *     completed / delivery_incomplete / delivery_unverified / claims_unverified。
+ *     两者并列而非替代 —— `ok:true` 配 `delivery_status:'claims_unverified'` 是
+ *     完全合法的一行：Agent 确实正常收尾了，但它宣称的「测试已通过」没有证据支撑。
+ *     判断「任务是否真做成」的宿主应当读它，而不是只看 ok。
+ *
  * 导出供单测断言。
  */
 export function emitJsonResult(result: PrintResult, sessionFile?: string): void {
@@ -218,6 +237,7 @@ export function emitJsonResult(result: PrintResult, sessionFile?: string): void
       num_turns: result.numTurns,
       error: result.error,
       completionStatus,
+      ...(result.deliveryStatus ? { delivery_status: result.deliveryStatus } : {}),
       deliverables,
       verification: allVerified ? 'passed' : deliverables.length > 0 ? 'not_run' : 'none',
       warnings,
@@ -325,10 +345,16 @@ export async function runPrintMode(
   });
 
   try {
-    const cwd = options.workspace?.cwd ?? process.cwd();
-    const workspace = await workspaceIndexService.findByPath(cwd).catch(() => null);
+    const workspaceRoot = options.workspace?.cwd ?? process.cwd();
+    const workspace = await workspaceIndexService.findByPath(workspaceRoot).catch(() => null);
     const sessionId = options.workspace?.sessionId ?? 'default';
-    const conversationKey = `workspace:${workspace?.id ?? cwd}:session:${sessionId}`;
+    const sessionState = getOrCreateSessionState(sessionStateKey({
+      surface: 'cli', workspaceId: workspace?.id,
+      cwd: workspace?.id ? undefined : workspaceRoot, sessionId,
+    }));
+    await hydrateSessionWorktree({ sessionState, workspaceId: workspace?.id, sessionId, cwd: workspaceRoot });
+    const cwd = sessionState.activeWorktree?.path ?? workspaceRoot;
+    const conversationKey = `workspace:${workspace?.id ?? workspaceRoot}:session:${sessionId}`;
     const computerRun = options.computerUseMode && options.computerUseMode !== 'disabled'
       ? createCliComputerRun({
           mode: options.computerUseMode,
@@ -352,6 +378,9 @@ export async function runPrintMode(
       computer: computerRun,
     });
     const turnTools = filterToolsForRuntime(ALL_TOOLS, runtimeContext);
+    const stablePrompt = await buildBaseSystemPrompt(cwd, turnTools);
+    if (history[0]?.role === 'system') history[0] = { ...history[0], content: stablePrompt };
+    else history.unshift({ role: 'system', content: stablePrompt });
     const transientBundle = await buildTransientPromptBundle({
       cwd,
       tools: turnTools,
@@ -359,6 +388,7 @@ export async function runPrintMode(
         surface: 'cli',
         cwd,
         workspaceId: workspace?.id,
+        sessionId,
         interactiveAsk: false,
         memoryAccess: runtimeContext.memoryAccess,
         currentUserInput: prompt,
@@ -375,6 +405,7 @@ export async function runPrintMode(
       imageMaxLongSidePixel: imageOptions.imageMaxLongSidePixel,
       allowedTools: turnTools.map((tool) => tool.name),
       runtimeContext,
+      sessionState,
       ...(transientBundle ? { transientBundle } : {}),
       promptCache: createPromptCacheContext(
         `workspace:${workspace?.id ?? cwd}:session:${sessionId}`,
@@ -496,12 +527,19 @@ export function handleEvent(
       // 通用非致命警告（如 workflow 多余参数、pause 被自动跳过等）始终走 stderr，
       // 不受 verbose 控制 —— 这是用户必须看到的信号，避免静默丢弃语义被忽略。
       process.stderr.write(`${event.message}\n`);
+      // 同时收进结局契约：stderr 是给人看的，spawn 它的宿主要的是结构化的一条。
+      // （此前 PrintResult.warnings 声明了、读了、也输出了，就是没有任何地方写它 ——
+      //  json 那行的 warnings 恒为空数组。）
+      (result.warnings ??= []).push(event.message);
       return {};
 
     case 'turn_done':
       // text 模式：非 verbose 时这里一次性把累积答案打到 stdout（旧行为，逐字节不变）。
       // json 模式：只记 stopReason='end_turn'，stdout 保持静默。
       result.stopReason = 'end_turn';
+      // 内核的交付判定必须过进程边界：模型谎称「已验证」时，若不记这一笔，
+      // 宿主收到的 JSON 与一切正常逐字节相同，唯一线索只剩 stderr 上一行中文。
+      if (event.completionStatus) result.deliveryStatus = event.completionStatus;
       if (!isJson && !verbose) console.log(result.buffer);
       return {};
 

package/src/cli/streamJson.js

@@ -28,7 +28,9 @@ import { saveContext } from '../context/persistContext.js';
 import { setAskHandler } from '../askBridge.js';
 import { normalizeImageDetail } from '../multimodal.js';
 import { runWithPlugins } from '../plugins/pluginRunner.js';
-import { buildTransientPromptBundle } from '../prompts/system.js';
+import { buildBaseSystemPrompt, buildTransientPromptBundle } from '../prompts/system.js';
+import { getOrCreateSessionState, sessionStateKey } from '../session/sessionState.js';
+import { hydrateSessionWorktree } from '../tools/worktree/sessionWorktree.js';
 import { ALL_TOOLS } from '../tools/index.js';
 import { createPromptCacheContext } from '../llm/promptCache.js';
 import { createAgentRuntimeContext, filterToolsForRuntime, memoryContextFromRuntime, } from '../runtime/agentRuntimeContext.js';
@@ -267,10 +269,16 @@ export async function runStreamJsonMode(provider, args, initialHistory, options)
         void finalize().finally(() => process.exit(130));
     });
     try {
-        const cwd = options.workspace?.cwd ?? getWorkingDir();
-        const workspace = await workspaceIndexService.findByPath(cwd).catch(() => null);
+        const workspaceRoot = options.workspace?.cwd ?? getWorkingDir();
+        const workspace = await workspaceIndexService.findByPath(workspaceRoot).catch(() => null);
         const sessionId = options.workspace?.sessionId ?? 'default';
-        const conversationKey = `workspace:${workspace?.id ?? cwd}:session:${sessionId}`;
+        const sessionState = getOrCreateSessionState(sessionStateKey({
+            surface: 'headless', workspaceId: workspace?.id,
+            cwd: workspace?.id ? undefined : workspaceRoot, sessionId,
+        }));
+        await hydrateSessionWorktree({ sessionState, workspaceId: workspace?.id, sessionId, cwd: workspaceRoot });
+        const cwd = sessionState.activeWorktree?.path ?? workspaceRoot;
+        const conversationKey = `workspace:${workspace?.id ?? workspaceRoot}:session:${sessionId}`;
         const computerRun = options.computerUseMode && options.computerUseMode !== 'disabled'
             ? createCliComputerRun({
                 mode: options.computerUseMode,
@@ -294,6 +302,11 @@ export async function runStreamJsonMode(provider, args, initialHistory, options)
             computer: computerRun,
         });
         const turnTools = filterToolsForRuntime(ALL_TOOLS, runtimeContext);
+        const stablePrompt = await buildBaseSystemPrompt(cwd, turnTools);
+        if (history[0]?.role === 'system')
+            history[0] = { ...history[0], content: stablePrompt };
+        else
+            history.unshift({ role: 'system', content: stablePrompt });
         const transientBundle = await buildTransientPromptBundle({
             cwd,
             tools: turnTools,
@@ -301,6 +314,7 @@ export async function runStreamJsonMode(provider, args, initialHistory, options)
                 surface: 'headless',
                 cwd,
                 workspaceId: workspace?.id,
+                sessionId,
                 interactiveAsk: control,
                 memoryAccess: runtimeContext.memoryAccess,
                 currentUserInput: prompt,
@@ -316,6 +330,7 @@ export async function runStreamJsonMode(provider, args, initialHistory, options)
             imageDetail: cliImageOptions.imageDetail,
             imageMaxLongSidePixel: cliImageOptions.imageMaxLongSidePixel,
             runtimeContext,
+            sessionState,
             allowedTools: turnTools.map((tool) => tool.name),
             ...(transientBundle ? { transientBundle } : {}),
             promptCache: createPromptCacheContext(`workspace:${workspace?.id ?? cwd}:session:${sessionId}`),

package/src/cli/streamJson.ts

@@ -33,7 +33,9 @@ export type { AskAnswer } from '../askBridge.js';
 import { normalizeImageDetail } from '../multimodal.js';
 import type { ImageInputSpec, Message, Provider } from '../types.js';
 import { runWithPlugins } from '../plugins/pluginRunner.js';
-import { buildTransientPromptBundle } from '../prompts/system.js';
+import { buildBaseSystemPrompt, buildTransientPromptBundle } from '../prompts/system.js';
+import { getOrCreateSessionState, sessionStateKey } from '../session/sessionState.js';
+import { hydrateSessionWorktree } from '../tools/worktree/sessionWorktree.js';
 import { ALL_TOOLS } from '../tools/index.js';
 import { createPromptCacheContext } from '../llm/promptCache.js';
 import {
@@ -344,10 +346,16 @@ export async function runStreamJsonMode(
   });
 
   try {
-    const cwd = options.workspace?.cwd ?? getWorkingDir();
-    const workspace = await workspaceIndexService.findByPath(cwd).catch(() => null);
+    const workspaceRoot = options.workspace?.cwd ?? getWorkingDir();
+    const workspace = await workspaceIndexService.findByPath(workspaceRoot).catch(() => null);
     const sessionId = options.workspace?.sessionId ?? 'default';
-    const conversationKey = `workspace:${workspace?.id ?? cwd}:session:${sessionId}`;
+    const sessionState = getOrCreateSessionState(sessionStateKey({
+      surface: 'headless', workspaceId: workspace?.id,
+      cwd: workspace?.id ? undefined : workspaceRoot, sessionId,
+    }));
+    await hydrateSessionWorktree({ sessionState, workspaceId: workspace?.id, sessionId, cwd: workspaceRoot });
+    const cwd = sessionState.activeWorktree?.path ?? workspaceRoot;
+    const conversationKey = `workspace:${workspace?.id ?? workspaceRoot}:session:${sessionId}`;
     const computerRun = options.computerUseMode && options.computerUseMode !== 'disabled'
       ? createCliComputerRun({
           mode: options.computerUseMode,
@@ -371,6 +379,9 @@ export async function runStreamJsonMode(
       computer: computerRun,
     });
     const turnTools = filterToolsForRuntime(ALL_TOOLS, runtimeContext);
+    const stablePrompt = await buildBaseSystemPrompt(cwd, turnTools);
+    if (history[0]?.role === 'system') history[0] = { ...history[0], content: stablePrompt };
+    else history.unshift({ role: 'system', content: stablePrompt });
     const transientBundle = await buildTransientPromptBundle({
       cwd,
       tools: turnTools,
@@ -378,6 +389,7 @@ export async function runStreamJsonMode(
         surface: 'headless',
         cwd,
         workspaceId: workspace?.id,
+        sessionId,
         interactiveAsk: control,
         memoryAccess: runtimeContext.memoryAccess,
         currentUserInput: prompt,
@@ -393,6 +405,7 @@ export async function runStreamJsonMode(
       imageDetail: cliImageOptions.imageDetail,
       imageMaxLongSidePixel: cliImageOptions.imageMaxLongSidePixel,
       runtimeContext,
+      sessionState,
       allowedTools: turnTools.map((tool) => tool.name),
       ...(transientBundle ? { transientBundle } : {}),
       promptCache: createPromptCacheContext(

package/src/config/configDefaults.js

@@ -0,0 +1,25 @@
+/**
+ * 浏览器与 Node/Bun 都可安全加载的纯配置默认值。
+ * 不得在此模块引入 node:*、文件系统或运行时路径逻辑。
+ */
+export const DEFAULT_SUBAGENTS_CONFIG = {
+    enabled: true,
+    maxCallsPerTurn: 6,
+    maxConcurrent: 3,
+    defaultWriteIsolation: 'none',
+    timeoutMs: 1_800_000,
+    exploreMaxTurns: 50,
+    generalMaxTurns: 100,
+};
+export const DEFAULT_BROWSER_CONFIG = {
+    enabled: true,
+    profileId: 'default',
+    developerMode: false,
+    allowedCdpOrigins: [],
+};
+export const DEFAULT_COMPUTER_USE_CONFIG = {
+    enabled: true,
+    imagePolicy: 'single-adaptive',
+    maxImagePixels: 1_600_000,
+    observationFrames: 1,
+};

package/src/config/configDefaults.ts

@@ -0,0 +1,33 @@
+import type {
+  BrowserConfig,
+  ComputerUseConfig,
+  SubagentsConfig,
+} from './configFile.js';
+
+/**
+ * 浏览器与 Node/Bun 都可安全加载的纯配置默认值。
+ * 不得在此模块引入 node:*、文件系统或运行时路径逻辑。
+ */
+export const DEFAULT_SUBAGENTS_CONFIG: SubagentsConfig = {
+  enabled: true,
+  maxCallsPerTurn: 6,
+  maxConcurrent: 3,
+  defaultWriteIsolation: 'none',
+  timeoutMs: 1_800_000,
+  exploreMaxTurns: 50,
+  generalMaxTurns: 100,
+};
+
+export const DEFAULT_BROWSER_CONFIG: BrowserConfig = {
+  enabled: true,
+  profileId: 'default',
+  developerMode: false,
+  allowedCdpOrigins: [],
+};
+
+export const DEFAULT_COMPUTER_USE_CONFIG: ComputerUseConfig = {
+  enabled: true,
+  imagePolicy: 'single-adaptive',
+  maxImagePixels: 1_600_000,
+  observationFrames: 1,
+};

package/src/config/configFile.js

@@ -5,27 +5,8 @@ import { dirname, join } from 'node:path';
 import { execSiblingDir } from '../utils/greenRoot.js';
 import { normalizePromptCacheMode } from '../llm/promptCache.js';
 import { CONFIG_LIMITS, normalizeLimited } from './numericLimits.js';
-export const DEFAULT_SUBAGENTS_CONFIG = {
-    enabled: true,
-    maxCallsPerTurn: 6,
-    maxConcurrent: 3,
-    defaultWriteIsolation: 'none',
-    timeoutMs: 1_800_000,
-    exploreMaxTurns: 50,
-    generalMaxTurns: 100,
-};
-export const DEFAULT_BROWSER_CONFIG = {
-    enabled: true,
-    profileId: 'default',
-    developerMode: false,
-    allowedCdpOrigins: [],
-};
-export const DEFAULT_COMPUTER_USE_CONFIG = {
-    enabled: true,
-    imagePolicy: 'single-adaptive',
-    maxImagePixels: 1_600_000,
-    observationFrames: 1,
-};
+import { DEFAULT_BROWSER_CONFIG, DEFAULT_COMPUTER_USE_CONFIG, DEFAULT_SUBAGENTS_CONFIG, } from './configDefaults.js';
+export { DEFAULT_BROWSER_CONFIG, DEFAULT_COMPUTER_USE_CONFIG, DEFAULT_SUBAGENTS_CONFIG, } from './configDefaults.js';
 const DEFAULT_CONTEXT_WINDOW = 128_000;
 const LOCK_STALE_MS = 120_000;
 const LOCK_WAIT_MS = 15_000;

package/src/config/configFile.ts

@@ -7,6 +7,17 @@ import { execSiblingDir } from '../utils/greenRoot.js';
 import type { PromptCacheConfig } from '../types.js';
 import { normalizePromptCacheMode } from '../llm/promptCache.js';
 import { CONFIG_LIMITS, normalizeLimited } from './numericLimits.js';
+import {
+  DEFAULT_BROWSER_CONFIG,
+  DEFAULT_COMPUTER_USE_CONFIG,
+  DEFAULT_SUBAGENTS_CONFIG,
+} from './configDefaults.js';
+
+export {
+  DEFAULT_BROWSER_CONFIG,
+  DEFAULT_COMPUTER_USE_CONFIG,
+  DEFAULT_SUBAGENTS_CONFIG,
+} from './configDefaults.js';
 
 export type ImageProtocol = 'auto' | 'chat' | 'images';
 export type ApiFormat = 'openai' | 'anthropic';
@@ -95,30 +106,6 @@ export interface MinimalAgentConfig {
   savedAt: number;
 }
 
-export const DEFAULT_SUBAGENTS_CONFIG: SubagentsConfig = {
-  enabled: true,
-  maxCallsPerTurn: 6,
-  maxConcurrent: 3,
-  defaultWriteIsolation: 'none',
-  timeoutMs: 1_800_000,
-  exploreMaxTurns: 50,
-  generalMaxTurns: 100,
-};
-
-export const DEFAULT_BROWSER_CONFIG: BrowserConfig = {
-  enabled: true,
-  profileId: 'default',
-  developerMode: false,
-  allowedCdpOrigins: [],
-};
-
-export const DEFAULT_COMPUTER_USE_CONFIG: ComputerUseConfig = {
-  enabled: true,
-  imagePolicy: 'single-adaptive',
-  maxImagePixels: 1_600_000,
-  observationFrames: 1,
-};
-
 /**
  * 加模型是提交式弹窗、顾问改字段都是有界枚举/数字——不会出现「输入中途非法值」，
  * 所以 v3 时代的 provider/advisor 草稿桶随 schema 一起删除，只剩 tools（imageGeneration/

package/src/config/configManagement.js

@@ -1,4 +1,4 @@
-import { DEFAULT_BROWSER_CONFIG, DEFAULT_COMPUTER_USE_CONFIG, } from './configFile.js';
+import { DEFAULT_BROWSER_CONFIG, DEFAULT_COMPUTER_USE_CONFIG, } from './configDefaults.js';
 /**
  * CLI 和 Desktop 共用的脱敏配置视图。这里不返回任何凭据，也不读取 Runtime
  * 状态；它只是把 V4 权威配置稳定地投影成可展示、可比较的管理模型。

package/src/config/configManagement.ts

@@ -1,8 +1,8 @@
 import {
   DEFAULT_BROWSER_CONFIG,
   DEFAULT_COMPUTER_USE_CONFIG,
-  type ModelConfig,
-} from './configFile.js';
+} from './configDefaults.js';
+import type { ModelConfig } from './configFile.js';
 import type { ConfigSnapshot } from './configService.js';
 
 export type ManagementStatus = 'unconfigured' | 'configured' | 'incomplete';

package/src/config/configService.js

@@ -373,6 +373,30 @@ export class ConfigService {
             return null;
         return { ...advisor, provider: resolveProvider(entry), supportsVision: entry.supportsVision !== false };
     }
+    /**
+     * 同步版 advisor 可用性探测——仅用于 resolveAvailability 这类工具列表探测场景。
+     * 读当前内存快照（可能略滞后于磁盘最新值），但调用频率高、对实时性要求不高，
+     * 下一轮 LLM 请求前 configService.reload() 会刷新。
+     *
+     * 为什么不用 advisor() 的 fresh reload：
+     * resolveAvailability 是在主循环构造工具集时同步调用的，不能 await；
+     * 而且每次工具集构造都读盘的话开销太大（主循环每轮都可能构造一次）。
+     */
+    advisorAvailabilitySync() {
+        const config = this.snapshot.config;
+        const advisor = config?.advisor;
+        if (!advisor?.modelId)
+            return { enabled: false, mode: 'auto', modelId: null, configured: false };
+        const entry = config?.models?.find((m) => m.id === advisor.modelId);
+        if (!entry?.apiKey)
+            return { enabled: false, mode: 'auto', modelId: null, configured: false };
+        return {
+            enabled: advisor.enabled !== false,
+            mode: advisor.mode === 'manual' ? 'manual' : 'auto',
+            modelId: advisor.modelId,
+            configured: true,
+        };
+    }
     async subagents() {
         return (await this.reload()).config?.subagents ?? DEFAULT_SUBAGENTS_CONFIG;
     }

package/src/config/configService.ts

@@ -431,6 +431,29 @@ export class ConfigService {
     return { ...advisor, provider: resolveProvider(entry), supportsVision: entry.supportsVision !== false };
   }
 
+  /**
+   * 同步版 advisor 可用性探测——仅用于 resolveAvailability 这类工具列表探测场景。
+   * 读当前内存快照（可能略滞后于磁盘最新值），但调用频率高、对实时性要求不高，
+   * 下一轮 LLM 请求前 configService.reload() 会刷新。
+   *
+   * 为什么不用 advisor() 的 fresh reload：
+   * resolveAvailability 是在主循环构造工具集时同步调用的，不能 await；
+   * 而且每次工具集构造都读盘的话开销太大（主循环每轮都可能构造一次）。
+   */
+  advisorAvailabilitySync(): { enabled: boolean; mode: 'auto' | 'manual'; modelId: string | null; configured: boolean } {
+    const config = this.snapshot.config;
+    const advisor = config?.advisor;
+    if (!advisor?.modelId) return { enabled: false, mode: 'auto', modelId: null, configured: false };
+    const entry = config?.models?.find((m) => m.id === advisor.modelId);
+    if (!entry?.apiKey) return { enabled: false, mode: 'auto', modelId: null, configured: false };
+    return {
+      enabled: advisor.enabled !== false,
+      mode: advisor.mode === 'manual' ? 'manual' : 'auto',
+      modelId: advisor.modelId,
+      configured: true,
+    };
+  }
+
   async subagents(): Promise<SubagentsConfig> {
     return (await this.reload()).config?.subagents ?? DEFAULT_SUBAGENTS_CONFIG;
   }

package/src/context/compact.js

@@ -379,6 +379,10 @@ async function runCompaction(messages, provider, signal, chatFn = chat, promptCa
             ...(systemMsg ? [systemMsg] : []),
             {
                 role: 'user',
+                // ⚠️ 这里**故意不带** provider 原始报错：本条消息会进 history，既落盘也会在
+                //   后续轮次重新发给模型，错误体里可能夹带 URL / 请求体 / 凭据。
+                //   排障线索走 `compact_failed` 事件（只到 stderr / UI，不进 history）。
+                //   `compact.test.ts::reports snip fallback without exposing the raw LLM error` 锁定此约定。
                 content: '(LLM 摘要压缩失败，已使用规则式截断继续会话)'
                     + (failureLedger
                         ? `\n\n<user_messages>\n以下是被截掉的那部分对话里**用户的原话**（程序逐字抽取）。摘要不存在，这里就是唯一残留的历史要求。\n${failureLedger}\n</user_messages>`

package/src/context/compact.ts

@@ -491,6 +491,10 @@ async function runCompaction(
       ...(systemMsg ? [systemMsg] : []),
       {
         role: 'user',
+        // ⚠️ 这里**故意不带** provider 原始报错：本条消息会进 history，既落盘也会在
+        //   后续轮次重新发给模型，错误体里可能夹带 URL / 请求体 / 凭据。
+        //   排障线索走 `compact_failed` 事件（只到 stderr / UI，不进 history）。
+        //   `compact.test.ts::reports snip fallback without exposing the raw LLM error` 锁定此约定。
         content: '(LLM 摘要压缩失败，已使用规则式截断继续会话)'
           + (failureLedger
             ? `\n\n<user_messages>\n以下是被截掉的那部分对话里**用户的原话**（程序逐字抽取）。摘要不存在，这里就是唯一残留的历史要求。\n${failureLedger}\n</user_messages>`