skyloom 1.13.6 → 1.13.8

package/src/core/longdoc.ts

@@ -1,155 +1,155 @@
-/**
- * 长文档处理策略 — sliding window + summary chain.
- *
- * When an input exceeds the agent's effective context window,
- * split into overlapping chunks, summarize each, then chain
- * summaries into a final digest.
- *
- * Architecture:
- *   Input → Chunk(sliding window) → Per-chunk Summary → Chain → Final Digest
- *
- * All summaries are generated by the calling agent's LLM, so quality
- * depends on the model in use. The chunker is pure text processing
- * and works without any LLM call.
- */
-
-import type { BaseAgent } from "./agent";
-
-/* ═══════════════════════════════════════
-   Chunker — split text into overlapping windows
-   ═══════════════════════════════════════ */
-export interface ChunkOptions {
-  /** Target chunk size in characters (default 6000) */
-  chunkSize?: number;
-  /** Overlap between consecutive chunks in characters (default 800) */
-  overlap?: number;
-  /** Minimum chunk size before we stop splitting (default 500) */
-  minChunk?: number;
-}
-
-export function chunkText(text: string, opts?: ChunkOptions): string[] {
-  const cs = opts?.chunkSize ?? 6000;
-  const ol = opts?.overlap ?? 800;
-  const min = opts?.minChunk ?? 500;
-  const chunks: string[] = [];
-
-  if (text.length <= cs + min) { chunks.push(text); return chunks; }
-
-  let start = 0;
-  while (start < text.length) {
-    let end = start + cs;
-    if (end >= text.length) { end = text.length; }
-    else {
-      // Try to break at paragraph boundary
-      const searchEnd = Math.min(end + 400, text.length);
-      const paraBreak = text.lastIndexOf("\n\n", searchEnd);
-      if (paraBreak > start + min) end = paraBreak;
-      else {
-        const lineBreak = text.lastIndexOf("\n", searchEnd);
-        if (lineBreak > start + min) end = lineBreak;
-        else {
-          const space = text.lastIndexOf(" ", searchEnd);
-          if (space > start + min) end = space;
-        }
-      }
-    }
-
-    chunks.push(text.slice(start, end).trim());
-    if (end >= text.length) break;
-    start = end - ol;
-    if (start < 0) start = 0;
-  }
-
-  return chunks;
-}
-
-/* ═══════════════════════════════════════
-   Summary chain — ask agent to summarize chunks then chain
-   ═══════════════════════════════════════ */
-export interface SummaryOptions {
-  /** Max total chars for the final digest (default 3000) */
-  maxDigestChars?: number;
-  /** Custom summarization prompt for each chunk */
-  chunkPrompt?: string;
-  /** Custom chain prompt for combining summaries */
-  chainPrompt?: string;
-}
-
-const DEFAULT_CHUNK_PROMPT = `Summarize the following text concisely. Keep all key facts, names, numbers, and code snippets. Output the summary directly without preamble. Limit to 300 words.
-
-Text:
-{text}`;
-
-const DEFAULT_CHAIN_PROMPT = `Combine the following section summaries into a single coherent digest. Preserve all key facts, remove redundancy. Output directly without preamble.
-
-{summaries}`;
-
-export async function summarizeLongDoc(
-  agent: BaseAgent,
-  text: string,
-  opts?: SummaryOptions
-): Promise<string> {
-  const maxDigest = opts?.maxDigestChars ?? 3000;
-  const chunks = chunkText(text);
-
-  // Single chunk — no summarization needed
-  if (chunks.length <= 1) {
-    if (text.length <= maxDigest) return text;
-    const prompt = (opts?.chunkPrompt || DEFAULT_CHUNK_PROMPT).replace("{text}", text);
-    return agent.chatOneshot(prompt, { maxTokens: maxDigest });
-  }
-
-  // Multi-chunk: summarize each, then chain
-  const summaries: string[] = [];
-  for (let i = 0; i < chunks.length; i++) {
-    const prompt = (opts?.chunkPrompt || DEFAULT_CHUNK_PROMPT).replace("{text}", chunks[i]);
-    try {
-      const s = await agent.chatOneshot(prompt, { maxTokens: 600 });
-      summaries.push(s);
-    } catch {
-      summaries.push(chunks[i].slice(0, 400) + "...");
-    }
-  }
-
-  // Chain summaries
-  if (summaries.length === 1) return summaries[0].slice(0, maxDigest);
-
-  const joined = summaries.map((s, i) => `## Section ${i + 1}\n${s}`).join("\n\n");
-  if (joined.length <= maxDigest) return joined;
-
-  const chainPrompt = (opts?.chainPrompt || DEFAULT_CHAIN_PROMPT).replace("{summaries}", joined);
-  const final = await agent.chatOneshot(chainPrompt, { maxTokens: maxDigest });
-  return final.slice(0, maxDigest);
-}
-
-/* ═══════════════════════════════════════
-   Structured data parsing helpers
-   ═══════════════════════════════════════ */
-export function parseStructuredInput(input: string): {
-  hasTable: boolean; hasJSON: boolean; hasCSV: boolean;
-  extractedJSON: string | null; extractedTable: string[][] | null;
-} {
-  const result = { hasTable: false, hasJSON: false, hasCSV: false, extractedJSON: null as string | null, extractedTable: null as string[][] | null };
-
-  // Detect JSON
-  const jsonMatch = input.match(/\{[\s\S]*\}|\[[\s\S]*\]/);
-  if (jsonMatch) {
-    try { JSON.parse(jsonMatch[0]); result.hasJSON = true; result.extractedJSON = jsonMatch[0]; }
-    catch { /* not valid JSON */ }
-  }
-
-  // Detect markdown table
-  const tableMatch = input.match(/\|[\s\S]*?\|/);
-  if (tableMatch) {
-    result.hasTable = true;
-    const lines = input.split("\n").filter(l => l.includes("|") && !l.startsWith("|---") && !l.startsWith("| --"));
-    result.extractedTable = lines.map(l => l.split("|").filter(c => c.trim()).map(c => c.trim()));
-  }
-
-  // Detect CSV
-  if (input.includes(",") && input.split("\n").filter(l => l.includes(",")).length >= 2) {
-    result.hasCSV = true;
-  }
-
-  return result;
-}
+/**
+ * 长文档处理策略 — sliding window + summary chain.
+ *
+ * When an input exceeds the agent's effective context window,
+ * split into overlapping chunks, summarize each, then chain
+ * summaries into a final digest.
+ *
+ * Architecture:
+ *   Input → Chunk(sliding window) → Per-chunk Summary → Chain → Final Digest
+ *
+ * All summaries are generated by the calling agent's LLM, so quality
+ * depends on the model in use. The chunker is pure text processing
+ * and works without any LLM call.
+ */
+
+import type { BaseAgent } from "./agent";
+
+/* ═══════════════════════════════════════
+   Chunker — split text into overlapping windows
+   ═══════════════════════════════════════ */
+export interface ChunkOptions {
+  /** Target chunk size in characters (default 6000) */
+  chunkSize?: number;
+  /** Overlap between consecutive chunks in characters (default 800) */
+  overlap?: number;
+  /** Minimum chunk size before we stop splitting (default 500) */
+  minChunk?: number;
+}
+
+export function chunkText(text: string, opts?: ChunkOptions): string[] {
+  const cs = opts?.chunkSize ?? 6000;
+  const ol = opts?.overlap ?? 800;
+  const min = opts?.minChunk ?? 500;
+  const chunks: string[] = [];
+
+  if (text.length <= cs + min) { chunks.push(text); return chunks; }
+
+  let start = 0;
+  while (start < text.length) {
+    let end = start + cs;
+    if (end >= text.length) { end = text.length; }
+    else {
+      // Try to break at paragraph boundary
+      const searchEnd = Math.min(end + 400, text.length);
+      const paraBreak = text.lastIndexOf("\n\n", searchEnd);
+      if (paraBreak > start + min) end = paraBreak;
+      else {
+        const lineBreak = text.lastIndexOf("\n", searchEnd);
+        if (lineBreak > start + min) end = lineBreak;
+        else {
+          const space = text.lastIndexOf(" ", searchEnd);
+          if (space > start + min) end = space;
+        }
+      }
+    }
+
+    chunks.push(text.slice(start, end).trim());
+    if (end >= text.length) break;
+    start = end - ol;
+    if (start < 0) start = 0;
+  }
+
+  return chunks;
+}
+
+/* ═══════════════════════════════════════
+   Summary chain — ask agent to summarize chunks then chain
+   ═══════════════════════════════════════ */
+export interface SummaryOptions {
+  /** Max total chars for the final digest (default 3000) */
+  maxDigestChars?: number;
+  /** Custom summarization prompt for each chunk */
+  chunkPrompt?: string;
+  /** Custom chain prompt for combining summaries */
+  chainPrompt?: string;
+}
+
+const DEFAULT_CHUNK_PROMPT = `Summarize the following text concisely. Keep all key facts, names, numbers, and code snippets. Output the summary directly without preamble. Limit to 300 words.
+
+Text:
+{text}`;
+
+const DEFAULT_CHAIN_PROMPT = `Combine the following section summaries into a single coherent digest. Preserve all key facts, remove redundancy. Output directly without preamble.
+
+{summaries}`;
+
+export async function summarizeLongDoc(
+  agent: BaseAgent,
+  text: string,
+  opts?: SummaryOptions
+): Promise<string> {
+  const maxDigest = opts?.maxDigestChars ?? 3000;
+  const chunks = chunkText(text);
+
+  // Single chunk — no summarization needed
+  if (chunks.length <= 1) {
+    if (text.length <= maxDigest) return text;
+    const prompt = (opts?.chunkPrompt || DEFAULT_CHUNK_PROMPT).replace("{text}", text);
+    return agent.chatOneshot(prompt, { maxTokens: maxDigest });
+  }
+
+  // Multi-chunk: summarize each, then chain
+  const summaries: string[] = [];
+  for (let i = 0; i < chunks.length; i++) {
+    const prompt = (opts?.chunkPrompt || DEFAULT_CHUNK_PROMPT).replace("{text}", chunks[i]);
+    try {
+      const s = await agent.chatOneshot(prompt, { maxTokens: 600 });
+      summaries.push(s);
+    } catch {
+      summaries.push(chunks[i].slice(0, 400) + "...");
+    }
+  }
+
+  // Chain summaries
+  if (summaries.length === 1) return summaries[0].slice(0, maxDigest);
+
+  const joined = summaries.map((s, i) => `## Section ${i + 1}\n${s}`).join("\n\n");
+  if (joined.length <= maxDigest) return joined;
+
+  const chainPrompt = (opts?.chainPrompt || DEFAULT_CHAIN_PROMPT).replace("{summaries}", joined);
+  const final = await agent.chatOneshot(chainPrompt, { maxTokens: maxDigest });
+  return final.slice(0, maxDigest);
+}
+
+/* ═══════════════════════════════════════
+   Structured data parsing helpers
+   ═══════════════════════════════════════ */
+export function parseStructuredInput(input: string): {
+  hasTable: boolean; hasJSON: boolean; hasCSV: boolean;
+  extractedJSON: string | null; extractedTable: string[][] | null;
+} {
+  const result = { hasTable: false, hasJSON: false, hasCSV: false, extractedJSON: null as string | null, extractedTable: null as string[][] | null };
+
+  // Detect JSON
+  const jsonMatch = input.match(/\{[\s\S]*\}|\[[\s\S]*\]/);
+  if (jsonMatch) {
+    try { JSON.parse(jsonMatch[0]); result.hasJSON = true; result.extractedJSON = jsonMatch[0]; }
+    catch { /* not valid JSON */ }
+  }
+
+  // Detect markdown table
+  const tableMatch = input.match(/\|[\s\S]*?\|/);
+  if (tableMatch) {
+    result.hasTable = true;
+    const lines = input.split("\n").filter(l => l.includes("|") && !l.startsWith("|---") && !l.startsWith("| --"));
+    result.extractedTable = lines.map(l => l.split("|").filter(c => c.trim()).map(c => c.trim()));
+  }
+
+  // Detect CSV
+  if (input.includes(",") && input.split("\n").filter(l => l.includes(",")).length >= 2) {
+    result.hasCSV = true;
+  }
+
+  return result;
+}

package/src/core/mcp.ts

@@ -871,6 +871,54 @@ export class MCPManager {
   }
 }
 
+/**
+ * Project-level .mcp.json (Claude Code standard) — drop the same file you
+ * use with Claude Code into the repo root and Skyloom picks it up:
+ *
+ *   { "mcpServers": {
+ *       "github": { "type": "http", "url": "https://api.example.com/mcp/" },
+ *       "db": { "command": "npx", "args": ["-y", "@x/dbhub"],
+ *               "env": { "DB_URL": "${DB_URL}" } } } }
+ *
+ * `${VAR}` references expand from the environment, so secrets stay out of
+ * the committed file.
+ */
+
+/** Expand ${VAR} environment references (Claude Code .mcp.json convention). */
+export function expandEnvRefs(s: string): string {
+  return s.replace(/\$\{([A-Za-z0-9_]+)\}/g, (_, v) => process.env[v] ?? '');
+}
+
+/** Load and translate <cwd>/.mcp.json into Skyloom server configs. */
+export function loadProjectMcpJson(cwd: string = process.cwd()): MCPServerConfig[] {
+  const file = path.join(cwd, '.mcp.json');
+  if (!fs.existsSync(file)) return [];
+  try {
+    const data = JSON.parse(fs.readFileSync(file, 'utf-8'));
+    const servers = data?.mcpServers;
+    if (!servers || typeof servers !== 'object') return [];
+    const out: MCPServerConfig[] = [];
+    for (const [name, raw] of Object.entries<any>(servers)) {
+      if (!raw || typeof raw !== 'object') continue;
+      const cfg: MCPServerConfig = { name, enabled: true };
+      if (typeof raw.command === 'string') {
+        cfg.command = expandEnvRefs(raw.command);
+        if (Array.isArray(raw.args)) cfg.args = raw.args.map((a: any) => expandEnvRefs(String(a)));
+      }
+      if (typeof raw.url === 'string') cfg.url = expandEnvRefs(raw.url);
+      if (raw.env && typeof raw.env === 'object') {
+        cfg.env = {};
+        for (const [k, v] of Object.entries(raw.env)) cfg.env[k] = expandEnvRefs(String(v));
+      }
+      if (!cfg.command && !cfg.url) continue; // unsupported transport entry
+      out.push(cfg);
+    }
+    return out;
+  } catch {
+    return [];
+  }
+}
+
 /**
  * Persistence helpers for runtime-added MCP servers.
  */