skyloom 1.13.6 → 1.13.8

package/src/core/tool_router.ts

@@ -1,193 +1,193 @@
-/**
- * Tool-subset selection for LLM calls.
- *
- * Without filtering, every chat turn ships ~50 tool schemas (built-ins + MCP +
- * skill-required + delegation) to the model. That dilutes attention (the LLM
- * picks plausible-but-wrong tools more often) and burns 8-15k input tokens per
- * turn. This module narrows the active tool set to ~12 by lightweight scoring
- * against the user's latest message.
- *
- * The router intentionally avoids embeddings / LLM calls — it must run in <1ms
- * on every turn, before the real LLM call. A coarse keyword/substring score is
- * good enough to keep the right tools in and bad enough to be cheap.
- */
-
-import type { ToolDefinition, ToolRegistry } from './tool';
-
-// Infrastructure tools that are useful on most turns.
-const INFRA_TOOLS: ReadonlySet<string> = new Set([
-  'delegate_to',
-  'list_skills',
-  'use_skill',
-  'recall_facts',
-  'remember_fact',
-]);
-
-const MUTATING_TOOLS: ReadonlySet<string> = new Set([
-  'write_file',
-  'edit_file',
-  'move_file',
-  'copy_file',
-  'delete_file',
-  'git_add',
-  'git_commit',
-  'git_checkout',
-  'shell_exec',
-  'http_post',
-  'mcp_add_server',
-  'mcp_remove_server',
-  'mcp_scaffold_server',
-  'kill_process',
-  'package_manager',
-  'service_control',
-]);
-
-const TOKEN_RE = /[A-Za-z][A-Za-z0-9_]*|[一-鿿]+/g;
-
-const STOPWORDS: ReadonlySet<string> = new Set([
-  'a', 'an', 'the', 'is', 'are', 'was', 'were', 'and', 'or', 'but',
-  'to', 'for', 'of', 'in', 'on', 'at', 'with', 'by', 'do', 'did',
-  'does', 'i', 'me', 'my', 'you', 'your', 'it', 'this', 'that',
-  'what', 'how', 'can', 'could', 'would', 'should', 'please',
-  'tell', 'show', 'help', 'ok', 'yes', 'no',
-  '好', '的', '是', '我', '你', '他', '她', '它', '这', '那',
-  '什么', '怎么', '请', '帮', '麻烦',
-]);
-
-function tokenize(text: string): Set<string> {
-  const tokens = new Set<string>();
-  const matches = text?.toLowerCase().match(TOKEN_RE) || [];
-  for (const t of matches) {
-    if (t.length >= 2 && !STOPWORDS.has(t)) {
-      tokens.add(t);
-    }
-  }
-  return tokens;
-}
-
-function scoreTool(tool: ToolDefinition, queryTokens: Set<string>, queryLc: string): number {
-  if (queryTokens.size === 0) return 0;
-  let score = 0;
-
-  // Tool name tokens carry the strongest signal
-  const nameTokens = tokenize(tool.name.replace(/_/g, ' '));
-  for (const qt of queryTokens) {
-    if (nameTokens.has(qt)) {
-      score += 5;
-    } else if (tool.name.toLowerCase().includes(qt)) {
-      score += 3;
-    }
-  }
-
-  // Description tokens are weaker
-  const descTokens = tokenize(tool.description);
-  for (const qt of queryTokens) {
-    if (descTokens.has(qt)) {
-      score += 1;
-    }
-  }
-
-  // Intent boosts for common Chinese/English tasks
-  const name = tool.name;
-  if (['read_file', 'list_directory', 'tree', 'file_search', 'code_search', 'grep'].includes(name) &&
-    ['read', 'file', 'inspect', 'search', 'grep', '看', '读', '查', '搜索', '文件', '代码'].some(k => queryLc.includes(k))) {
-    score += 4;
-  }
-  if (['write_file', 'edit_file', 'move_file', 'copy_file', 'delete_file'].includes(name) &&
-    ['write', 'edit', 'modify', 'fix', 'save', '生成', '写', '改', '修复', '保存', '删除'].some(k => queryLc.includes(k))) {
-    score += 4;
-  }
-  if (name.startsWith('git_') &&
-    ['git', 'commit', 'diff', 'branch', '提交', '分支', '差异'].some(k => queryLc.includes(k))) {
-    score += 4;
-  }
-  if (['web_search', 'fetch_page', 'http_get'].includes(name) &&
-    ['web', 'url', 'http', 'research', '搜索', '网页', '联网', '资料'].some(k => queryLc.includes(k))) {
-    score += 4;
-  }
-  if (['list_skills', 'use_skill'].includes(name) &&
-    ['skill', '能力', '技能', 'ppt', 'pdf', 'excel', 'xlsx', 'docx'].some(k => queryLc.includes(k))) {
-    score += 4;
-  }
-
-  return score;
-}
-
-/**
- * Return up to ~topK tool names ordered by relevance to the query.
- *
- * Always-included infrastructure tools and mustInclude (e.g. active
- * skill required_tools) are appended regardless of score. When the candidate
- * set is already small (<= topK + |mustInclude|), no filtering is applied.
- *
- * A short or empty query means we have no signal to filter — returning the
- * full candidate set is correct in that case.
- */
-export function selectRelevantTools(
-  registry: ToolRegistry,
-  candidateNames: string[],
-  query: string,
-  options?: {
-    topK?: number;
-    mustInclude?: Set<string>;
-  }
-): string[] {
-  const topK = options?.topK ?? 12;
-  const explicitMust = new Set(options?.mustInclude ?? []);
-
-  const infraPresent = candidateNames.filter(n => INFRA_TOOLS.has(n) && !explicitMust.has(n));
-  const mustPresent = candidateNames.filter(n => explicitMust.has(n));
-  const remaining = candidateNames.filter(n => !explicitMust.has(n) && !INFRA_TOOLS.has(n));
-
-  const queryTokens = tokenize(query);
-  const queryLc = (query || '').toLowerCase();
-  const smallSurface = candidateNames.length <= topK + mustPresent.length + infraPresent.length;
-  const lowSignal = queryTokens.size < 2 && queryLc.length < 8;
-
-  // No filtering when the set is already small or the query is too short
-  if (smallSurface || lowSignal) {
-    return [...mustPresent, ...infraPresent, ...remaining];
-  }
-
-  const scored: Array<{ name: string; score: number; penalty: number }> = [];
-  const allScoredNames = [...infraPresent, ...remaining];
-
-  for (const name of allScoredNames) {
-    const tool = registry.get(name);
-    if (!tool) continue;
-    const score = scoreTool(tool, queryTokens, queryLc);
-    const penalty = MUTATING_TOOLS.has(name) && score === 0 ? 1 : 0;
-    scored.push({ name, score, penalty });
-  }
-
-  // Stable sort by descending score; zero-score tools only fill spare slots
-  scored.sort((a, b) => {
-    if (a.score !== b.score) return b.score - a.score;
-    return a.penalty - b.penalty;
-  });
-
-  const budget = Math.max(0, topK - mustPresent.length);
-  const picked: string[] = [];
-  const pickedSet = new Set<string>();
-
-  for (const item of scored) {
-    if (picked.length >= budget) break;
-    if (item.score > 0) {
-      picked.push(item.name);
-      pickedSet.add(item.name);
-    }
-  }
-
-  // Fill remaining slots with zero-score tools
-  if (picked.length < budget) {
-    for (const item of scored) {
-      if (picked.length >= budget) break;
-      if (!pickedSet.has(item.name)) {
-        picked.push(item.name);
-        pickedSet.add(item.name);
-      }
-    }
-  }
-
-  return [...mustPresent, ...picked];
-}
+/**
+ * Tool-subset selection for LLM calls.
+ *
+ * Without filtering, every chat turn ships ~50 tool schemas (built-ins + MCP +
+ * skill-required + delegation) to the model. That dilutes attention (the LLM
+ * picks plausible-but-wrong tools more often) and burns 8-15k input tokens per
+ * turn. This module narrows the active tool set to ~12 by lightweight scoring
+ * against the user's latest message.
+ *
+ * The router intentionally avoids embeddings / LLM calls — it must run in <1ms
+ * on every turn, before the real LLM call. A coarse keyword/substring score is
+ * good enough to keep the right tools in and bad enough to be cheap.
+ */
+
+import type { ToolDefinition, ToolRegistry } from './tool';
+
+// Infrastructure tools that are useful on most turns.
+const INFRA_TOOLS: ReadonlySet<string> = new Set([
+  'delegate_to',
+  'list_skills',
+  'use_skill',
+  'recall_facts',
+  'remember_fact',
+]);
+
+const MUTATING_TOOLS: ReadonlySet<string> = new Set([
+  'write_file',
+  'edit_file',
+  'move_file',
+  'copy_file',
+  'delete_file',
+  'git_add',
+  'git_commit',
+  'git_checkout',
+  'shell_exec',
+  'http_post',
+  'mcp_add_server',
+  'mcp_remove_server',
+  'mcp_scaffold_server',
+  'kill_process',
+  'package_manager',
+  'service_control',
+]);
+
+const TOKEN_RE = /[A-Za-z][A-Za-z0-9_]*|[一-鿿]+/g;
+
+const STOPWORDS: ReadonlySet<string> = new Set([
+  'a', 'an', 'the', 'is', 'are', 'was', 'were', 'and', 'or', 'but',
+  'to', 'for', 'of', 'in', 'on', 'at', 'with', 'by', 'do', 'did',
+  'does', 'i', 'me', 'my', 'you', 'your', 'it', 'this', 'that',
+  'what', 'how', 'can', 'could', 'would', 'should', 'please',
+  'tell', 'show', 'help', 'ok', 'yes', 'no',
+  '好', '的', '是', '我', '你', '他', '她', '它', '这', '那',
+  '什么', '怎么', '请', '帮', '麻烦',
+]);
+
+function tokenize(text: string): Set<string> {
+  const tokens = new Set<string>();
+  const matches = text?.toLowerCase().match(TOKEN_RE) || [];
+  for (const t of matches) {
+    if (t.length >= 2 && !STOPWORDS.has(t)) {
+      tokens.add(t);
+    }
+  }
+  return tokens;
+}
+
+function scoreTool(tool: ToolDefinition, queryTokens: Set<string>, queryLc: string): number {
+  if (queryTokens.size === 0) return 0;
+  let score = 0;
+
+  // Tool name tokens carry the strongest signal
+  const nameTokens = tokenize(tool.name.replace(/_/g, ' '));
+  for (const qt of queryTokens) {
+    if (nameTokens.has(qt)) {
+      score += 5;
+    } else if (tool.name.toLowerCase().includes(qt)) {
+      score += 3;
+    }
+  }
+
+  // Description tokens are weaker
+  const descTokens = tokenize(tool.description);
+  for (const qt of queryTokens) {
+    if (descTokens.has(qt)) {
+      score += 1;
+    }
+  }
+
+  // Intent boosts for common Chinese/English tasks
+  const name = tool.name;
+  if (['read_file', 'list_directory', 'tree', 'file_search', 'code_search', 'grep'].includes(name) &&
+    ['read', 'file', 'inspect', 'search', 'grep', '看', '读', '查', '搜索', '文件', '代码'].some(k => queryLc.includes(k))) {
+    score += 4;
+  }
+  if (['write_file', 'edit_file', 'move_file', 'copy_file', 'delete_file'].includes(name) &&
+    ['write', 'edit', 'modify', 'fix', 'save', '生成', '写', '改', '修复', '保存', '删除'].some(k => queryLc.includes(k))) {
+    score += 4;
+  }
+  if (name.startsWith('git_') &&
+    ['git', 'commit', 'diff', 'branch', '提交', '分支', '差异'].some(k => queryLc.includes(k))) {
+    score += 4;
+  }
+  if (['web_search', 'fetch_page', 'http_get'].includes(name) &&
+    ['web', 'url', 'http', 'research', '搜索', '网页', '联网', '资料'].some(k => queryLc.includes(k))) {
+    score += 4;
+  }
+  if (['list_skills', 'use_skill'].includes(name) &&
+    ['skill', '能力', '技能', 'ppt', 'pdf', 'excel', 'xlsx', 'docx'].some(k => queryLc.includes(k))) {
+    score += 4;
+  }
+
+  return score;
+}
+
+/**
+ * Return up to ~topK tool names ordered by relevance to the query.
+ *
+ * Always-included infrastructure tools and mustInclude (e.g. active
+ * skill required_tools) are appended regardless of score. When the candidate
+ * set is already small (<= topK + |mustInclude|), no filtering is applied.
+ *
+ * A short or empty query means we have no signal to filter — returning the
+ * full candidate set is correct in that case.
+ */
+export function selectRelevantTools(
+  registry: ToolRegistry,
+  candidateNames: string[],
+  query: string,
+  options?: {
+    topK?: number;
+    mustInclude?: Set<string>;
+  }
+): string[] {
+  const topK = options?.topK ?? 12;
+  const explicitMust = new Set(options?.mustInclude ?? []);
+
+  const infraPresent = candidateNames.filter(n => INFRA_TOOLS.has(n) && !explicitMust.has(n));
+  const mustPresent = candidateNames.filter(n => explicitMust.has(n));
+  const remaining = candidateNames.filter(n => !explicitMust.has(n) && !INFRA_TOOLS.has(n));
+
+  const queryTokens = tokenize(query);
+  const queryLc = (query || '').toLowerCase();
+  const smallSurface = candidateNames.length <= topK + mustPresent.length + infraPresent.length;
+  const lowSignal = queryTokens.size < 2 && queryLc.length < 8;
+
+  // No filtering when the set is already small or the query is too short
+  if (smallSurface || lowSignal) {
+    return [...mustPresent, ...infraPresent, ...remaining];
+  }
+
+  const scored: Array<{ name: string; score: number; penalty: number }> = [];
+  const allScoredNames = [...infraPresent, ...remaining];
+
+  for (const name of allScoredNames) {
+    const tool = registry.get(name);
+    if (!tool) continue;
+    const score = scoreTool(tool, queryTokens, queryLc);
+    const penalty = MUTATING_TOOLS.has(name) && score === 0 ? 1 : 0;
+    scored.push({ name, score, penalty });
+  }
+
+  // Stable sort by descending score; zero-score tools only fill spare slots
+  scored.sort((a, b) => {
+    if (a.score !== b.score) return b.score - a.score;
+    return a.penalty - b.penalty;
+  });
+
+  const budget = Math.max(0, topK - mustPresent.length);
+  const picked: string[] = [];
+  const pickedSet = new Set<string>();
+
+  for (const item of scored) {
+    if (picked.length >= budget) break;
+    if (item.score > 0) {
+      picked.push(item.name);
+      pickedSet.add(item.name);
+    }
+  }
+
+  // Fill remaining slots with zero-score tools
+  if (picked.length < budget) {
+    for (const item of scored) {
+      if (picked.length >= budget) break;
+      if (!pickedSet.has(item.name)) {
+        picked.push(item.name);
+        pickedSet.add(item.name);
+      }
+    }
+  }
+
+  return [...mustPresent, ...picked];
+}