sp-rag 0.6.4 → 0.6.6

package/README.md

@@ -17,7 +17,7 @@ CLI để setup nhanh SP-RAG theo hướng dev-friendly:
 ## Trạng thái package
 
 - package npm public: `sp-rag`
-- version đang publish: `0.6.4`
+- version đang publish: `0.6.6`
 - binary public: `sp-rag`
 
 ## Cài từ source trong monorepo
@@ -70,6 +70,8 @@ Ghi chú:
 
 - với `cursor` và `vscode` ở `scope project`, nếu Sếpp đang đứng sẵn trong repo thì có thể bỏ `--cwd`
 - CLI sẽ tự dùng thư mục hiện tại cho cả MCP lẫn skill
+- rule/agent mới đã được tăng độ ưu tiên MCP-first, giảm khả năng model nhảy thẳng sang grep/read file local
+- skill mới cũng dặn model tổng hợp từ `matched_passages`, `top_entities`, `top_relations`, `citations`; không bê nguyên `answer_brief`
 
 Ví dụ:
 

package/dist/cli.js

@@ -4,7 +4,7 @@ import { runEvaluationSuite } from './lib/eval.js';
 import { defaultBaseUrl, defaultMcpServerAlias, defaultMcpUrl, installMcpConfig, resolveMcpConfigPath, } from './lib/mcp-config.js';
 import { fetchJson, fetchText, runDoctor } from './lib/http.js';
 import { installSkill, resolveSkillInstallTarget, } from './lib/skill.js';
-const cliVersion = '0.6.4';
+const cliVersion = '0.6.6';
 function parseArgv(argv) {
     const positionals = [];
     const options = {};

package/dist/lib/skill.js

@@ -62,7 +62,7 @@ export function defaultSkillDir(client = 'codex', cwd, scope) {
 function renderSkillMarkdown(context) {
     return `---
 name: sp-rag
-description: Use SP-RAG whenever the user asks about this codebase, internal business domain, rendered docs, import inventory, or codegraph sync status. Prefer MCP-backed evidence before answering from memory.
+description: Use SP-RAG whenever the user asks about this codebase, internal business domain, rendered docs, import inventory, or codegraph sync status. You must call SP-RAG MCP tools first before answering from memory or reading local files.
 ---
 
 # SP-RAG
@@ -85,30 +85,36 @@ Docs URL: \`${context.docsUrl}\`
 3. Use \`get_rendered_docs\` for public, function, or dev docs that were already rendered from the latest graph.
 4. Use \`get_sync_status\`, \`get_sync_runs\`, or \`get_sync_metrics\` when you need to verify commit freshness, investigate failures, or inspect operational history.
 5. Only call \`trigger_code_graph_sync\` when the user explicitly asks to refresh the graph or when a stale graph is the confirmed blocker.
+6. After \`query_context\` returns, synthesize the final answer from \`matched_passages\`, \`top_entities\`, \`top_relations\`, and \`citations\`. Treat \`answer_brief\` as a hint only.
 
 ## Guardrails
 
+- You must call SP-RAG MCP tools first for codebase or domain questions before using local workspace search, grep, or file reads.
 - Prefer MCP-grounded answers before relying on memory.
+- Treat \`answer_brief\` as a hint only. Prefer the richer evidence in \`matched_passages\`, \`top_entities\`, \`top_relations\`, and \`citations\` when writing the final answer.
 - If the evidence may be stale, say so clearly and mention that the graph or docs may need a refresh.
 - Do not trigger sync or import actions unless the user asked for it or the workflow truly requires it.
 - When rendered docs already answer the question, cite or summarize those docs instead of rewriting everything from scratch.
+- Only fall back to local workspace search or file reads after MCP is unavailable or clearly lacks the needed evidence, and say that you are falling back.
 `;
 }
 function renderCursorRule(context) {
     return `---
 description: Use SP-RAG when the request is about this codebase, internal business workflows, rendered docs, import inventory, or codegraph sync status.
 globs:
-alwaysApply: false
+alwaysApply: true
 ---
 
 # SP-RAG
 
-- Prefer the \`${context.serverAlias}\` MCP server at \`${context.mcpUrl}\` before answering from memory.
+- You must call the \`${context.serverAlias}\` MCP server first before using local workspace search, grep, or file reads for codebase and domain questions.
 - Use rendered docs from \`${context.docsUrl}\` when documentation already answers the question.
 - For architecture, domain, entities, relations, and business workflow questions, query MCP first and only then synthesize the answer.
+- Treat \`answer_brief\` only as a hint. Build the final answer from \`matched_passages\`, \`top_entities\`, \`top_relations\`, and \`citations\`.
 - For freshness or operational questions, check sync status, recent runs, and metrics before assuming the graph is current.
 - Only trigger codegraph sync when the user explicitly asks for it or stale evidence is the confirmed blocker.
 - If the evidence may be stale, say so clearly.
+- Only fall back to local workspace search after MCP is unavailable or clearly lacks the needed evidence, and say that you are falling back.
 `;
 }
 function renderVsCodeAgent(context) {
@@ -122,6 +128,8 @@ You are the SP-RAG custom agent for this workspace.
 
 Use the \`${context.serverAlias}\` MCP server at \`${context.mcpUrl}\` and the rendered docs URL \`${context.docsUrl}\` as your grounded source of truth.
 
+You must call an SP-RAG MCP tool before using local workspace search, grep, or file reads for codebase and domain questions.
+
 ## Recommended Workflow
 
 1. Start with \`healthz\` if connectivity or freshness is uncertain.
@@ -129,13 +137,17 @@ Use the \`${context.serverAlias}\` MCP server at \`${context.mcpUrl}\` and the r
 3. Use \`get_rendered_docs\` when public, function, or dev docs may already answer the question.
 4. Use \`get_sync_status\`, \`get_sync_runs\`, or \`get_sync_metrics\` for freshness, incident review, or operational debugging.
 5. Trigger codegraph sync only when the user explicitly requests a refresh or stale evidence is the confirmed blocker.
+6. After \`query_context\` returns, synthesize the final answer from \`matched_passages\`, \`top_entities\`, \`top_relations\`, and \`citations\`. Treat \`answer_brief\` as a hint.
 
 ## Guardrails
 
 - Prefer MCP-grounded evidence before answering from memory.
+- Treat \`answer_brief\` as a hint. Prefer the richer evidence in \`matched_passages\`, \`top_entities\`, \`top_relations\`, and \`citations\` when writing the final answer.
+- Do not use local workspace search, grep, or file reads until SP-RAG MCP has been tried first for the current question.
 - When evidence may be stale, say so clearly and mention that a refresh might be needed.
 - Do not trigger sync or import actions unless the workflow truly requires it.
 - When rendered docs already answer the question, summarize those docs instead of rewriting everything from scratch.
+- Only fall back to local workspace search or file reads after MCP is unavailable or clearly lacks the needed evidence, and say that you are falling back.
 `;
 }
 function renderSkillArtifact(context) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "sp-rag",
-  "version": "0.6.4",
+  "version": "0.6.6",
   "description": "CLI cho setup MCP, codegraph GitNexus và skill của SP-RAG",
   "type": "module",
   "files": [