skyloom 1.13.6 → 1.13.8

package/src/core/estimate.ts

@@ -1,104 +1,104 @@
-/**
- * 资源估算模块 — Token & time budget estimation for task planning.
- *
- * Helps Snow and other planning agents estimate the cost of
- * proposed sub-tasks before committing to execution.
- */
-
-import type { Task } from "./agent";
-
-/* ═══════════════════════════════════════
-   Token estimation
-   ═══════════════════════════════════════ */
-const CJK_REGEX = /[一-鿿぀-ゟ가-힯㐀-䶿]/g;
-
-/** Estimate tokens for a given text (CJK ~2 each, ASCII ~4 chars each). */
-export function estimateTokens(text: string): number {
-  const cjk = (text.match(CJK_REGEX) || []).length;
-  return cjk * 2 + Math.ceil((text.length - cjk) / 4);
-}
-
-/* ═══════════════════════════════════════
-   Per-task-type cost estimates
-   ═══════════════════════════════════════ */
-const TASK_TYPE_PATTERNS: Array<[RegExp, number, number]> = [
-  // [pattern, estimated tokens, estimated tools]
-  [/read|read_file|grep|search|查|搜索|list/i, 2000, 2],
-  [/write|write_file|生成|写|create|implement/i, 4000, 5],
-  [/edit|edit_file|改|修改|fix|修复/i, 3000, 3],
-  [/delete|delete_file|删|rm/i, 1500, 2],
-  [/deploy|部署|publish|发布|release/i, 8000, 8],
-  [/review|审查|audit|审计|scan|扫描/i, 5000, 4],
-  [/test|测试|run_test|coverage/i, 3000, 3],
-  [/research|研究|调研|analyze|分析/i, 6000, 4],
-  [/orchestrate|编排|multi-step|多步/i, 12000, 10],
-];
-
-/** Estimate cost for a single task description. */
-export function estimateTaskCost(description: string): { tokens: number; tools: number; timeSeconds: number } {
-  let tokens = 2000; // base
-  let tools = 2;     // base
-  for (const [pattern, t, tc] of TASK_TYPE_PATTERNS) {
-    if (pattern.test(description)) { tokens = Math.max(tokens, t); tools = Math.max(tools, tc); }
-  }
-
-  // Time estimate: ~0.5s per tool call + 2s per 1k tokens
-  const timeSeconds = (tokens / 1000) * 2 + tools * 0.5 + 2;
-  return { tokens, tools, timeSeconds };
-}
-
-/* ═══════════════════════════════════════
-   Task plan cost summary
-   ═══════════════════════════════════════ */
-export interface PlanEstimate {
-  totalTokens: number;
-  totalTools: number;
-  totalTimeSeconds: number;
-  perTask: Array<{ id: string; tokens: number; tools: number; time: number }>;
-  warnings: string[];
-}
-
-export function estimateTaskPlan(tasks: Task[]): PlanEstimate {
-  const perTask: PlanEstimate["perTask"] = [];
-  let totalTokens = 500; // system prompt overhead
-  let totalTools = 0;
-  let totalTime = 5; // init overhead
-  const warnings: string[] = [];
-
-  for (const t of tasks) {
-    const est = estimateTaskCost(t.description);
-    perTask.push({ id: t.id, tokens: est.tokens, tools: est.tools, time: est.timeSeconds });
-    totalTokens += est.tokens;
-    totalTools += est.tools;
-    totalTime += est.timeSeconds;
-
-    if (est.timeSeconds > 60) warnings.push(`Task ${t.id} may take >${Math.round(est.timeSeconds)}s`);
-    if (est.tools > 10) warnings.push(`Task ${t.id} uses many tool calls (${est.tools})`);
-  }
-
-  if (totalTokens > 64000) warnings.push(`Total token estimate (${totalTokens}) exceeds typical context window`);
-  if (totalTime > 120) warnings.push(`Estimated total time (${Math.round(totalTime)}s) is significant`);
-  if (tasks.length > 6) warnings.push(`Large number of sub-tasks (${tasks.length}) — consider merging simpler ones`);
-
-  return { totalTokens, totalTools, totalTimeSeconds: Math.round(totalTime), perTask, warnings };
-}
-
-/* ═══════════════════════════════════════
-   Format estimate for display
-   ═══════════════════════════════════════ */
-export function formatPlanEstimate(est: PlanEstimate): string {
-  const lines: string[] = [
-    `## Plan Estimate`,
-    `| Task | Tokens | Tools | Time |`,
-    `|------|--------|-------|------|`,
-    ...est.perTask.map(t => `| ${t.id} | ${t.tokens} | ${t.tools} | ${t.time.toFixed(0)}s |`),
-    `| **Total** | **${est.totalTokens}** | **${est.totalTools}** | **${est.totalTimeSeconds}s** |`,
-  ];
-
-  if (est.warnings.length > 0) {
-    lines.push("", "### Warnings");
-    for (const w of est.warnings) lines.push(`- ⚠ ${w}`);
-  }
-
-  return lines.join("\n");
-}
+/**
+ * 资源估算模块 — Token & time budget estimation for task planning.
+ *
+ * Helps Snow and other planning agents estimate the cost of
+ * proposed sub-tasks before committing to execution.
+ */
+
+import type { Task } from "./agent";
+
+/* ═══════════════════════════════════════
+   Token estimation
+   ═══════════════════════════════════════ */
+const CJK_REGEX = /[一-鿿぀-ゟ가-힯㐀-䶿]/g;
+
+/** Estimate tokens for a given text (CJK ~2 each, ASCII ~4 chars each). */
+export function estimateTokens(text: string): number {
+  const cjk = (text.match(CJK_REGEX) || []).length;
+  return cjk * 2 + Math.ceil((text.length - cjk) / 4);
+}
+
+/* ═══════════════════════════════════════
+   Per-task-type cost estimates
+   ═══════════════════════════════════════ */
+const TASK_TYPE_PATTERNS: Array<[RegExp, number, number]> = [
+  // [pattern, estimated tokens, estimated tools]
+  [/read|read_file|grep|search|查|搜索|list/i, 2000, 2],
+  [/write|write_file|生成|写|create|implement/i, 4000, 5],
+  [/edit|edit_file|改|修改|fix|修复/i, 3000, 3],
+  [/delete|delete_file|删|rm/i, 1500, 2],
+  [/deploy|部署|publish|发布|release/i, 8000, 8],
+  [/review|审查|audit|审计|scan|扫描/i, 5000, 4],
+  [/test|测试|run_test|coverage/i, 3000, 3],
+  [/research|研究|调研|analyze|分析/i, 6000, 4],
+  [/orchestrate|编排|multi-step|多步/i, 12000, 10],
+];
+
+/** Estimate cost for a single task description. */
+export function estimateTaskCost(description: string): { tokens: number; tools: number; timeSeconds: number } {
+  let tokens = 2000; // base
+  let tools = 2;     // base
+  for (const [pattern, t, tc] of TASK_TYPE_PATTERNS) {
+    if (pattern.test(description)) { tokens = Math.max(tokens, t); tools = Math.max(tools, tc); }
+  }
+
+  // Time estimate: ~0.5s per tool call + 2s per 1k tokens
+  const timeSeconds = (tokens / 1000) * 2 + tools * 0.5 + 2;
+  return { tokens, tools, timeSeconds };
+}
+
+/* ═══════════════════════════════════════
+   Task plan cost summary
+   ═══════════════════════════════════════ */
+export interface PlanEstimate {
+  totalTokens: number;
+  totalTools: number;
+  totalTimeSeconds: number;
+  perTask: Array<{ id: string; tokens: number; tools: number; time: number }>;
+  warnings: string[];
+}
+
+export function estimateTaskPlan(tasks: Task[]): PlanEstimate {
+  const perTask: PlanEstimate["perTask"] = [];
+  let totalTokens = 500; // system prompt overhead
+  let totalTools = 0;
+  let totalTime = 5; // init overhead
+  const warnings: string[] = [];
+
+  for (const t of tasks) {
+    const est = estimateTaskCost(t.description);
+    perTask.push({ id: t.id, tokens: est.tokens, tools: est.tools, time: est.timeSeconds });
+    totalTokens += est.tokens;
+    totalTools += est.tools;
+    totalTime += est.timeSeconds;
+
+    if (est.timeSeconds > 60) warnings.push(`Task ${t.id} may take >${Math.round(est.timeSeconds)}s`);
+    if (est.tools > 10) warnings.push(`Task ${t.id} uses many tool calls (${est.tools})`);
+  }
+
+  if (totalTokens > 64000) warnings.push(`Total token estimate (${totalTokens}) exceeds typical context window`);
+  if (totalTime > 120) warnings.push(`Estimated total time (${Math.round(totalTime)}s) is significant`);
+  if (tasks.length > 6) warnings.push(`Large number of sub-tasks (${tasks.length}) — consider merging simpler ones`);
+
+  return { totalTokens, totalTools, totalTimeSeconds: Math.round(totalTime), perTask, warnings };
+}
+
+/* ═══════════════════════════════════════
+   Format estimate for display
+   ═══════════════════════════════════════ */
+export function formatPlanEstimate(est: PlanEstimate): string {
+  const lines: string[] = [
+    `## Plan Estimate`,
+    `| Task | Tokens | Tools | Time |`,
+    `|------|--------|-------|------|`,
+    ...est.perTask.map(t => `| ${t.id} | ${t.tokens} | ${t.tools} | ${t.time.toFixed(0)}s |`),
+    `| **Total** | **${est.totalTokens}** | **${est.totalTools}** | **${est.totalTimeSeconds}s** |`,
+  ];
+
+  if (est.warnings.length > 0) {
+    lines.push("", "### Warnings");
+    for (const w of est.warnings) lines.push(`- ⚠ ${w}`);
+  }
+
+  return lines.join("\n");
+}

package/src/core/evolve.ts

@@ -1,191 +1,191 @@
-/**
- * 自我进化模块 — Prompt self-optimization via failure analysis.
- *
- * When an agent repeatedly fails at similar tasks, this module analyzes
- * the failure patterns and suggests targeted improvements to the agent's
- * System Prompt. The agent can then apply these suggestions to improve
- * future performance.
- *
- * Architecture:
- *    Failure log → Pattern analysis → Prompt diff → Agent.applyDiff()
- */
-
-import * as fs from "fs";
-import * as path from "path";
-import { USER_CONFIG_DIR } from "./config";
-import { getLogger } from "./logger";
-
-const log = getLogger("evolve");
-
-/* ═══════════════════════════════════════
-   Prompt diff — a suggested change
-   ═══════════════════════════════════════ */
-export interface PromptDiff {
-  id: string;
-  ts: string;
-  agent: string;
-  reason: string;           // Why this change is needed
-  before: string;           // Old prompt fragment
-  after: string;            // New prompt fragment
-  applied: boolean;
-  improvement?: string;     // Measured improvement after applying
-}
-
-/* ═══════════════════════════════════════
-   Failure analysis
-   ═══════════════════════════════════════ */
-export interface FailureAnalysis {
-  agent: string;
-  period: string;
-  totalCalls: number;
-  failureCount: number;
-  topFailures: Array<{ pattern: string; count: number }>;
-  suggestedDiffs: PromptDiff[];
-}
-
-const evolveDir = path.join(USER_CONFIG_DIR, "evolve");
-function ensureDir() { if (!fs.existsSync(evolveDir)) fs.mkdirSync(evolveDir, { recursive: true }); }
-
-/** Analyze recent failures from the learning module and suggest prompt improvements. */
-export function analyzeFailures(
-  agent: string,
-  experiences: Array<{ pattern: string; solution: string; frequency: number; lastSeen: string }>,
-  systemPrompt: string
-): FailureAnalysis {
-  const recent = experiences.filter(e => {
-    try { return new Date(e.lastSeen).getTime() > Date.now() - 7 * 86400000; }
-    catch { return false; }
-  });
-
-  const topFailures = recent
-    .sort((a, b) => b.frequency - a.frequency)
-    .slice(0, 5)
-    .map(e => ({ pattern: e.pattern, count: e.frequency }));
-
-  const suggestedDiffs: PromptDiff[] = [];
-
-  // Rule-based suggestions from failure patterns
-  for (const f of topFailures) {
-    const lower = f.pattern.toLowerCase();
-
-    // Search storm → add search budget rule
-    if ((lower.includes("search") || lower.includes("web_search")) && f.count >= 3) {
-      const rule = `- 搜索不超过 5 轮。5 轮后直接基于已有信息综合回答。`;
-      if (!systemPrompt.includes("搜索不超过")) {
-        suggestedDiffs.push({
-          id: Math.random().toString(36).slice(2, 8),
-          ts: new Date().toISOString(), agent,
-          reason: `搜索风暴 (${f.count} 次重复搜索)`,
-          before: "", after: rule, applied: false,
-        });
-      }
-    }
-
-    // Empty response → add deliverable checklist
-    if ((lower.includes("empty") || lower.includes("placeholder") || lower.includes("完成了")) && f.count >= 2) {
-      const rule = `- 完成任务后，必须输出实际产物（代码/文件路径/数据），禁止只说"完成了"而无产出。`;
-      if (!systemPrompt.includes("必须输出实际产物")) {
-        suggestedDiffs.push({
-          id: Math.random().toString(36).slice(2, 8),
-          ts: new Date().toISOString(), agent,
-          reason: `空响应/占位 (${f.count} 次)`,
-          before: "", after: rule, applied: false,
-        });
-      }
-    }
-
-    // Tool not found → add tool discovery to prompt
-    if (lower.includes("does not exist") || lower.includes("tool") && lower.includes("not found")) {
-      const rule = `- 使用不熟悉的工具前先调 list_skills 查看可用工具列表。`;
-      if (!systemPrompt.includes("list_skills")) {
-        suggestedDiffs.push({
-          id: Math.random().toString(36).slice(2, 8),
-          ts: new Date().toISOString(), agent,
-          reason: `工具不存在 (${f.count} 次)`,
-          before: "", after: rule, applied: false,
-        });
-      }
-    }
-
-    // File not found → add path verification rule
-    if (lower.includes("file not found") || lower.includes("directory not found")) {
-      const rule = `- 文件操作前先用 list_directory 或 read_file 确认路径存在。`;
-      if (!systemPrompt.includes("确认路径存在")) {
-        suggestedDiffs.push({
-          id: Math.random().toString(36).slice(2, 8),
-          ts: new Date().toISOString(), agent,
-          reason: `文件路径错误 (${f.count} 次)`,
-          before: "", after: rule, applied: false,
-        });
-      }
-    }
-  }
-
-  // General rule: if failure rate > 20%, suggest self-review
-  const recentExperiences = recent.filter(e => {
-    try { return new Date(e.lastSeen).getTime() > Date.now() - 3 * 86400000; }
-    catch { return false; }
-  });
-
-  // Deduplicate suggestions
-  const seen = new Set<string>();
-  const uniqueDiffs = suggestedDiffs.filter(d => {
-    const key = d.after.slice(0, 30);
-    if (seen.has(key)) return false;
-    seen.add(key); return true;
-  });
-
-  // Max 3 suggestions per analysis
-  return {
-    agent, period: "last 7 days",
-    totalCalls: 0, failureCount: 0,
-    topFailures,
-    suggestedDiffs: uniqueDiffs.slice(0, 3),
-  };
-}
-
-/* ═══════════════════════════════════════
-   Apply prompt diff to agent
-   ═══════════════════════════════════════ */
-export function applyPromptDiff(agent: any, diff: PromptDiff): boolean {
-  try {
-    const currentPrompt = agent.systemPrompt;
-    if (!diff.after || currentPrompt.includes(diff.after.slice(0, 20))) return false;
-
-    // Append the new rule after "## 行为守则" or "## Behavior" section
-    const marker = currentPrompt.includes("行为守则") ? "## 行为守则" : "## Behavior";
-    const idx = currentPrompt.indexOf(marker);
-    if (idx < 0) { agent.systemPrompt += "\n" + diff.after; }
-    else {
-      const insertPoint = currentPrompt.indexOf("\n", currentPrompt.indexOf("\n-", idx) + 1);
-      agent.systemPrompt = currentPrompt.slice(0, insertPoint) + "\n" + diff.after + "\n" + currentPrompt.slice(insertPoint);
-    }
-
-    diff.applied = true;
-    diff.improvement = "pending evaluation";
-
-    // Persist the diff
-    ensureDir();
-    const file = path.join(evolveDir, `${diff.agent}_diffs.jsonl`);
-    fs.appendFileSync(file, JSON.stringify(diff) + "\n");
-
-    agent.rebuildSystemPrompt();
-    return true;
-  } catch (e) {
-    log.warn("apply_prompt_diff_failed", { agent: diff.agent, error: String(e) });
-    return false;
-  }
-}
-
-/** Get all applied diffs for an agent. */
-export function getAppliedDiffs(agent: string): PromptDiff[] {
-  const diffs: PromptDiff[] = [];
-  try {
-    const file = path.join(evolveDir, `${agent}_diffs.jsonl`);
-    if (fs.existsSync(file)) {
-      const lines = fs.readFileSync(file, "utf-8").split("\n").filter(Boolean);
-      for (const line of lines) { try { diffs.push(JSON.parse(line)); } catch { } }
-    }
-  } catch { /* ignore */ }
-  return diffs;
-}
+/**
+ * 自我进化模块 — Prompt self-optimization via failure analysis.
+ *
+ * When an agent repeatedly fails at similar tasks, this module analyzes
+ * the failure patterns and suggests targeted improvements to the agent's
+ * System Prompt. The agent can then apply these suggestions to improve
+ * future performance.
+ *
+ * Architecture:
+ *    Failure log → Pattern analysis → Prompt diff → Agent.applyDiff()
+ */
+
+import * as fs from "fs";
+import * as path from "path";
+import { USER_CONFIG_DIR } from "./config";
+import { getLogger } from "./logger";
+
+const log = getLogger("evolve");
+
+/* ═══════════════════════════════════════
+   Prompt diff — a suggested change
+   ═══════════════════════════════════════ */
+export interface PromptDiff {
+  id: string;
+  ts: string;
+  agent: string;
+  reason: string;           // Why this change is needed
+  before: string;           // Old prompt fragment
+  after: string;            // New prompt fragment
+  applied: boolean;
+  improvement?: string;     // Measured improvement after applying
+}
+
+/* ═══════════════════════════════════════
+   Failure analysis
+   ═══════════════════════════════════════ */
+export interface FailureAnalysis {
+  agent: string;
+  period: string;
+  totalCalls: number;
+  failureCount: number;
+  topFailures: Array<{ pattern: string; count: number }>;
+  suggestedDiffs: PromptDiff[];
+}
+
+const evolveDir = path.join(USER_CONFIG_DIR, "evolve");
+function ensureDir() { if (!fs.existsSync(evolveDir)) fs.mkdirSync(evolveDir, { recursive: true }); }
+
+/** Analyze recent failures from the learning module and suggest prompt improvements. */
+export function analyzeFailures(
+  agent: string,
+  experiences: Array<{ pattern: string; solution: string; frequency: number; lastSeen: string }>,
+  systemPrompt: string
+): FailureAnalysis {
+  const recent = experiences.filter(e => {
+    try { return new Date(e.lastSeen).getTime() > Date.now() - 7 * 86400000; }
+    catch { return false; }
+  });
+
+  const topFailures = recent
+    .sort((a, b) => b.frequency - a.frequency)
+    .slice(0, 5)
+    .map(e => ({ pattern: e.pattern, count: e.frequency }));
+
+  const suggestedDiffs: PromptDiff[] = [];
+
+  // Rule-based suggestions from failure patterns
+  for (const f of topFailures) {
+    const lower = f.pattern.toLowerCase();
+
+    // Search storm → add search budget rule
+    if ((lower.includes("search") || lower.includes("web_search")) && f.count >= 3) {
+      const rule = `- 搜索不超过 5 轮。5 轮后直接基于已有信息综合回答。`;
+      if (!systemPrompt.includes("搜索不超过")) {
+        suggestedDiffs.push({
+          id: Math.random().toString(36).slice(2, 8),
+          ts: new Date().toISOString(), agent,
+          reason: `搜索风暴 (${f.count} 次重复搜索)`,
+          before: "", after: rule, applied: false,
+        });
+      }
+    }
+
+    // Empty response → add deliverable checklist
+    if ((lower.includes("empty") || lower.includes("placeholder") || lower.includes("完成了")) && f.count >= 2) {
+      const rule = `- 完成任务后，必须输出实际产物（代码/文件路径/数据），禁止只说"完成了"而无产出。`;
+      if (!systemPrompt.includes("必须输出实际产物")) {
+        suggestedDiffs.push({
+          id: Math.random().toString(36).slice(2, 8),
+          ts: new Date().toISOString(), agent,
+          reason: `空响应/占位 (${f.count} 次)`,
+          before: "", after: rule, applied: false,
+        });
+      }
+    }
+
+    // Tool not found → add tool discovery to prompt
+    if (lower.includes("does not exist") || lower.includes("tool") && lower.includes("not found")) {
+      const rule = `- 使用不熟悉的工具前先调 list_skills 查看可用工具列表。`;
+      if (!systemPrompt.includes("list_skills")) {
+        suggestedDiffs.push({
+          id: Math.random().toString(36).slice(2, 8),
+          ts: new Date().toISOString(), agent,
+          reason: `工具不存在 (${f.count} 次)`,
+          before: "", after: rule, applied: false,
+        });
+      }
+    }
+
+    // File not found → add path verification rule
+    if (lower.includes("file not found") || lower.includes("directory not found")) {
+      const rule = `- 文件操作前先用 list_directory 或 read_file 确认路径存在。`;
+      if (!systemPrompt.includes("确认路径存在")) {
+        suggestedDiffs.push({
+          id: Math.random().toString(36).slice(2, 8),
+          ts: new Date().toISOString(), agent,
+          reason: `文件路径错误 (${f.count} 次)`,
+          before: "", after: rule, applied: false,
+        });
+      }
+    }
+  }
+
+  // General rule: if failure rate > 20%, suggest self-review
+  const recentExperiences = recent.filter(e => {
+    try { return new Date(e.lastSeen).getTime() > Date.now() - 3 * 86400000; }
+    catch { return false; }
+  });
+
+  // Deduplicate suggestions
+  const seen = new Set<string>();
+  const uniqueDiffs = suggestedDiffs.filter(d => {
+    const key = d.after.slice(0, 30);
+    if (seen.has(key)) return false;
+    seen.add(key); return true;
+  });
+
+  // Max 3 suggestions per analysis
+  return {
+    agent, period: "last 7 days",
+    totalCalls: 0, failureCount: 0,
+    topFailures,
+    suggestedDiffs: uniqueDiffs.slice(0, 3),
+  };
+}
+
+/* ═══════════════════════════════════════
+   Apply prompt diff to agent
+   ═══════════════════════════════════════ */
+export function applyPromptDiff(agent: any, diff: PromptDiff): boolean {
+  try {
+    const currentPrompt = agent.systemPrompt;
+    if (!diff.after || currentPrompt.includes(diff.after.slice(0, 20))) return false;
+
+    // Append the new rule after "## 行为守则" or "## Behavior" section
+    const marker = currentPrompt.includes("行为守则") ? "## 行为守则" : "## Behavior";
+    const idx = currentPrompt.indexOf(marker);
+    if (idx < 0) { agent.systemPrompt += "\n" + diff.after; }
+    else {
+      const insertPoint = currentPrompt.indexOf("\n", currentPrompt.indexOf("\n-", idx) + 1);
+      agent.systemPrompt = currentPrompt.slice(0, insertPoint) + "\n" + diff.after + "\n" + currentPrompt.slice(insertPoint);
+    }
+
+    diff.applied = true;
+    diff.improvement = "pending evaluation";
+
+    // Persist the diff
+    ensureDir();
+    const file = path.join(evolveDir, `${diff.agent}_diffs.jsonl`);
+    fs.appendFileSync(file, JSON.stringify(diff) + "\n");
+
+    agent.rebuildSystemPrompt();
+    return true;
+  } catch (e) {
+    log.warn("apply_prompt_diff_failed", { agent: diff.agent, error: String(e) });
+    return false;
+  }
+}
+
+/** Get all applied diffs for an agent. */
+export function getAppliedDiffs(agent: string): PromptDiff[] {
+  const diffs: PromptDiff[] = [];
+  try {
+    const file = path.join(evolveDir, `${agent}_diffs.jsonl`);
+    if (fs.existsSync(file)) {
+      const lines = fs.readFileSync(file, "utf-8").split("\n").filter(Boolean);
+      for (const line of lines) { try { diffs.push(JSON.parse(line)); } catch { } }
+    }
+  } catch { /* ignore */ }
+  return diffs;
+}

package/src/core/factory.ts

@@ -76,6 +76,13 @@ export class SystemContext {
 export function createSystemContext(): SystemContext {
   const config = loadConfig();
 
+  // session_start hooks — user-configured shell commands (see core/hooks)
+  try {
+    const { loadHooks, runSessionStartHooks } = require('./hooks');
+    const hooks = loadHooks(config);
+    if (hooks.sessionStart.length > 0) runSessionStartHooks(hooks);
+  } catch { /* hooks must never block startup */ }
+
   let workspacePath = '';
   try {
     const { resolveWorkspacePath, initWorkspace } = require('./workspace');
@@ -121,11 +128,17 @@ export function createSystemContext(): SystemContext {
   // Configure MCP manager
   let mcpManager: any = null;
   try {
-    const { MCPManager, loadPersistedServers } = require('./mcp');
+    const { MCPManager, loadPersistedServers, loadProjectMcpJson } = require('./mcp');
     mcpManager = new MCPManager(baseToolRegistry);
     const persisted = loadPersistedServers();
     const mcpServers = (config as any).mcp?.servers || [];
-    const allServers = [...mcpServers, ...persisted];
+    const projectServers = loadProjectMcpJson(); // Claude Code 标准 .mcp.json
+    // dedupe by name — project .mcp.json wins over runtime-added over config
+    const byName = new Map<string, any>();
+    for (const s of [...mcpServers, ...persisted, ...projectServers]) {
+      if (s?.name) byName.set(s.name, s);
+    }
+    const allServers = [...byName.values()];
     if (allServers.length > 0) {
       mcpManager.configure(allServers);
     }
@@ -182,6 +195,22 @@ export function createSystemContext(): SystemContext {
         log.warn('delegate_tool_not_available', { agent: name, error: String(e) });
       }
 
+      // Register model self-service tools (list_models / set_my_model)
+      try {
+        const { createModelTools } = require('../tools/model_tool');
+        for (const t of createModelTools(name, config)) agentRegistry.register(t);
+      } catch (e) {
+        log.warn('model_tools_not_available', { agent: name, error: String(e) });
+      }
+
+      // Register the task-checklist tool (todo_write)
+      try {
+        const { createTodoTool } = require('../tools/todo');
+        agentRegistry.register(createTodoTool(agent));
+      } catch (e) {
+        log.warn('todo_tool_not_available', { agent: name, error: String(e) });
+      }
+
       agents.set(name, agent);
     } catch (e) {
       log.warn('agent_creation_failed', { agent: name, error: String(e) });