evrex-mcp 0.8.1 → 0.9.0

package/dist/index.js

@@ -63,7 +63,7 @@ var init_client = __esm({
       // Evidence-only retrieval (BM25 + embedding), no LLM synthesis call — see
       // apps/backend/src/query/query.service.ts#search. Used by evrex_search,
       // which wants ranked hits fast, not a synthesized paragraph.
-      search: (repoPath, text, filePaths) => post("/search", { repoPath, text, filePaths }),
+      search: (repoPath, text, filePaths, boost) => post("/search", { repoPath, text, filePaths, boost }),
       // Everything that happened in a repo, newest first, bounded by days — the
       // same query the desktop Timeline screen makes. Sessions and commits
       // interleaved, each with the handle evrex_expand takes.
@@ -354,6 +354,17 @@ import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
 import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
 import { z } from "zod";
 
+// src/subcommands.ts
+var SUBCOMMANDS = {
+  "evrex-hook": "hook.js",
+  "evrex-capture": "capture.js",
+  "evrex-import": "import.js",
+  "evrex-tickets": "tickets.js",
+  "evrex-pretool": "pretool.js",
+  "evrex-continuity": "continuity.js",
+  "evrex-account": "account.js"
+};
+
 // src/tools.ts
 init_client();
 import { execFileSync as execFileSync2 } from "node:child_process";
@@ -453,7 +464,7 @@ var EXTRACTION_SYSTEM_PROMPT = [
 // ../../packages/llm-core/src/synthesis.ts
 var MIN_SENTENCES = 1;
 var MAX_SENTENCES = 4;
-var SYNTHESIS_SYSTEM_PROMPT = `You answer questions about why code changed, using ONLY the evidence provided below. Never state a claim that is not directly supported by the evidence text. Respond with a JSON object: {"sentences": string[], "sentenceEvidence": number[][], "noRecordedReasoning": boolean}. Write 2-4 sentences when the evidence supports an answer; each sentence's entry in "sentenceEvidence" lists only the evidence indices (0-based) that directly support it. If the evidence is too thin or unrelated to actually answer the question, set "noRecordedReasoning": true and return empty arrays for "sentences" and "sentenceEvidence" \u2014 do not speculate or hedge.`;
+var SYNTHESIS_SYSTEM_PROMPT = `You answer questions about why code changed, using ONLY the evidence provided below. The evidence spans the project's whole history and the system it describes keeps changing: when items disagree, the most recently dated evidence describes the CURRENT state \u2014 answer as of the newest evidence, frame older accounts as how it used to work, and never present a superseded design as current. Never state a claim that is not directly supported by the evidence text. Respond with a JSON object: {"sentences": string[], "sentenceEvidence": number[][], "noRecordedReasoning": boolean}. Write 2-4 sentences when the evidence supports an answer; each sentence's entry in "sentenceEvidence" lists only the evidence indices (0-based) that directly support it. If the evidence is too thin or unrelated to actually answer the question, set "noRecordedReasoning": true and return empty arrays for "sentences" and "sentenceEvidence" \u2014 do not speculate or hedge.`;
 var SYNTHESIS_SCHEMA = {
   type: "object",
   additionalProperties: false,
@@ -469,7 +480,7 @@ var SYNTHESIS_SCHEMA = {
 };
 function buildSynthesisPrompt(question, evidence) {
   const list = evidence.map(
-    (e, i) => `[${i}] (${e.kind}, confidence ${(e.confidence ?? 1).toFixed(2)}): ${e.excerpt}`
+    (e, i) => `[${i}] (${e.kind}${e.at ? `, ${e.at.slice(0, 10)}` : ""}, confidence ${(e.confidence ?? 1).toFixed(2)}): ${e.excerpt}`
   ).join("\n");
   return `Question: ${question}
 
@@ -525,6 +536,29 @@ async function synthesizeAnswer(question, evidence, client2, options = {}) {
   };
 }
 
+// ../../packages/llm-core/src/retrieval-boost.ts
+var CORPUS = `The corpus is an engineering-memory index built from a team's coding-agent conversations and git history, retrieved with Postgres full-text OR-matching plus BM25 (exact word stems match; meanings do not). Searchable documents:
+- Conversation turns: prose by the person and the agent, plus tool-call lines like "[tool_call: Edit] apps/backend/src/query/query.service.ts" \u2014 so file paths, function names, error strings and identifiers appear verbatim.
+- Extracted insights: short titled records of decisions made, approaches REJECTED (with the reason), constraints, and procedures.
+- Commit messages: conventional-commit subjects and explanatory bodies.`;
+var EXPAND_SYSTEM = `You expand a search query into terms that will match an engineering record.
+
+${CORPUS}
+
+Given a person's question, produce the search terms most likely to appear VERBATIM in that corpus but absent from the question itself: the identifiers an engineer would have typed (function, constant, env var, file names), the error text the failure would have printed, and the words the same idea takes in commit-message register ("fix the flaky test" \u2192 "flaky", "deflake", "retry", "timeout"). Prefer specific over general \u2014 "RLS" and "row level security" over "database permissions". Never invent identifiers you are not confident exist in this kind of codebase; a wrong specific term matches nothing and costs nothing, but fill the list with your best evidence-seeking guesses.
+
+Respond with JSON only: {"terms": ["...", ...]} \u2014 3 to 8 terms, each 1 to 4 words, no duplicates of the question's own words.`;
+var RERANK_SYSTEM = `You order search results for an engineer's question about their own project's history.
+
+${CORPUS}
+
+You are given the question and a numbered list of candidate results that lexical retrieval fetched. Order the candidate ids best-first by how much each would actually help ANSWER the question:
+- Recorded reasoning outranks mechanical record: a decision, a rejected approach with its reason, a constraint, or an explanation beats a log line or a file listing that merely shares words with the question.
+- A candidate about the question's subject beats one that only quotes its vocabulary; watch for results that repeat the question's words while being about something else entirely.
+- Match recency to the question's intent: for "why does X exist / why was this decided", the origin of the idea is the answer; for "how does X work / what is the current pipeline", the system has usually changed since its founding \u2014 prefer the most recent account of the mechanism and treat older ones as history. Among near-duplicates prefer the more complete one.
+
+Respond with JSON only: {"ranked": ["id", ...]} \u2014 every listed id at most once, ids from the list only, best first. Include every id you consider relevant; omit only what would waste the reader's attention.`;
+
 // ../../packages/llm-core/src/cli-client.ts
 import { execFile } from "node:child_process";
 var TIMEOUT_MS = 12e4;
@@ -1635,6 +1669,28 @@ function proceduresBlock(procedures) {
   return "PROCEDURES (how this was done here, as steps that worked once \u2014 check they still apply before following them):\n" + procedures.slice(0, MAX_ITEMS).map((p) => `- ${p.title}
 ${p.steps.slice(0, 12).map((step, i) => `    ${i + 1}. ${step}`).join("\n")}`).join("\n");
 }
+function subagentsBlock(subagents) {
+  const budget = MAX_ITEMS * 2;
+  const lines = [];
+  let printed = 0;
+  let total = 0;
+  for (const s of subagents) {
+    const items = [
+      ...s.rejected.map((r) => `  - rejected: ${r.title} \u2014 ${r.reason}`),
+      ...s.constraints.map((c) => `  - constraint [${c.source}]: ${c.text}`)
+    ];
+    total += items.length;
+    const shown = items.slice(0, budget - printed);
+    if (shown.length === 0) continue;
+    const what = [s.agentType, s.description].filter(Boolean).join(": ");
+    lines.push(`- session:${s.id}${what ? ` (${what})` : ""}`, ...shown);
+    printed += shown.length;
+  }
+  if (lines.length === 0) return "";
+  if (total > printed) lines.push(`  \u2026 ${total - printed} more in their own records (expand the sub-agent handles).`);
+  return `FROM SUB-AGENTS (what the agents this session ran ruled out and were bound by):
+${lines.join("\n")}`;
+}
 async function evrexExpand(handles) {
   if (handles.length === 0) return "No handles given. Pass ids from evrex_search, e.g. commit:069a0b5.";
   const parts = [];
@@ -1688,6 +1744,8 @@ Open in Evrex: evrex://open?handle=commit:${id}`);
     }
     const procedures = proceduresBlock(session.procedures ?? []);
     if (procedures) block.push(procedures);
+    const fromSubagents = subagentsBlock(session.subagents ?? []);
+    if (fromSubagents) block.push(fromSubagents);
     parts.push(block.join("\n"));
   }
   parts.push(
@@ -1806,6 +1864,8 @@ ${truncate(commit.body, MAX_EXCERPT)}` : ""}`
           "CONSTRAINTS:\n" + session.constraints.slice(0, MAX_ITEMS).map((c) => `- [${c.source}] ${c.text}`).join("\n")
         );
       }
+      const fromSubagents = subagentsBlock(session.subagents ?? []);
+      if (fromSubagents) parts.push(fromSubagents);
       parts.push(`OUTCOME: ${truncate(session.outcome, MAX_EXCERPT)}`);
     }
   } else if (!commit.agentTrailers?.length) {
@@ -1962,10 +2022,22 @@ async function runEnrolment() {
   });
   process.exit(result.ok ? 0 : 1);
 }
+async function runSubcommand(file) {
+  const { spawnSync } = await import("node:child_process");
+  const { fileURLToPath } = await import("node:url");
+  const result = spawnSync(
+    process.execPath,
+    [fileURLToPath(new URL(`./${file}`, import.meta.url)), ...process.argv.slice(3)],
+    { stdio: "inherit" }
+  );
+  process.exit(result.status ?? 1);
+}
 async function main() {
   if (process.argv[2] === "enrol" || process.argv[2] === "enroll") {
     await runEnrolment();
   }
+  const subcommand = SUBCOMMANDS[process.argv[2] ?? ""];
+  if (subcommand) await runSubcommand(subcommand);
   assertConfigured();
   const transport = new StdioServerTransport();
   await server.connect(transport);

package/dist/pretool.js

@@ -49,7 +49,7 @@ var evrexApi = {
   // Evidence-only retrieval (BM25 + embedding), no LLM synthesis call — see
   // apps/backend/src/query/query.service.ts#search. Used by evrex_search,
   // which wants ranked hits fast, not a synthesized paragraph.
-  search: (repoPath, text, filePaths) => post("/search", { repoPath, text, filePaths }),
+  search: (repoPath, text, filePaths, boost) => post("/search", { repoPath, text, filePaths, boost }),
   // Everything that happened in a repo, newest first, bounded by days — the
   // same query the desktop Timeline screen makes. Sessions and commits
   // interleaved, each with the handle evrex_expand takes.
@@ -86,7 +86,34 @@ function repoMatchForCwd(repos, cwd) {
       }
     }
   }
-  return best;
+  if (best) return best;
+  const hop = worktreeHop(cwd);
+  if (!hop) return null;
+  for (const repo of repos) {
+    for (const path of repo.localPaths ?? []) {
+      if (!path) continue;
+      if (hop.main === path || hop.main.startsWith(`${path}/`)) {
+        return { id: repo.id, root: hop.top };
+      }
+    }
+  }
+  return null;
+}
+function worktreeHop(cwd) {
+  try {
+    const raw = execFileSync(
+      "git",
+      ["-C", cwd, "rev-parse", "--git-common-dir", "--show-toplevel"],
+      { stdio: ["ignore", "pipe", "ignore"], timeout: 1500 }
+    ).toString("utf-8").trim().split("\n");
+    const [gitDir, top] = [raw[0] ?? "", raw[1] ?? ""];
+    if (!gitDir.startsWith("/") || !gitDir.endsWith("/.git") || !top) return null;
+    const main2 = gitDir.slice(0, -"/.git".length);
+    if (!main2 || main2 === top) return null;
+    return { main: main2, top };
+  } catch {
+    return null;
+  }
 }
 function toRepoRelative(filePath, root) {
   if (!root || !filePath.startsWith("/")) return filePath;

package/dist/tickets.js

@@ -63,7 +63,7 @@ var init_client = __esm({
       // Evidence-only retrieval (BM25 + embedding), no LLM synthesis call — see
       // apps/backend/src/query/query.service.ts#search. Used by evrex_search,
       // which wants ranked hits fast, not a synthesized paragraph.
-      search: (repoPath, text, filePaths) => post("/search", { repoPath, text, filePaths }),
+      search: (repoPath, text, filePaths, boost) => post("/search", { repoPath, text, filePaths, boost }),
       // Everything that happened in a repo, newest first, bounded by days — the
       // same query the desktop Timeline screen makes. Sessions and commits
       // interleaved, each with the handle evrex_expand takes.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "evrex-mcp",
-  "version": "0.8.1",
+  "version": "0.9.0",
   "description": "MCP server that gives coding agents the recorded reasoning behind a repo: prior decisions, hard constraints, and approaches already rejected.",
   "keywords": [
     "mcp",
@@ -36,28 +36,27 @@
   "engines": {
     "node": ">=20"
   },
-  "scripts": {
-    "build": "tsc -p tsconfig.json",
-    "bundle": "node scripts/bundle.mjs",
-    "dev": "tsc -p tsconfig.json --watch",
-    "lint": "eslint .",
-    "check-types": "tsc --noEmit",
-    "test": "tsc -p tsconfig.json && node --test dist/*.test.js",
-    "prepack": "pnpm run bundle"
-  },
   "dependencies": {
     "@anthropic-ai/sdk": "^0.72.0",
     "@modelcontextprotocol/sdk": "^1.20.1",
     "zod": "^3.24.1"
   },
   "devDependencies": {
-    "@repo/eslint-config": "workspace:*",
-    "@repo/ingest-core": "workspace:*",
-    "@repo/llm-core": "workspace:*",
-    "@repo/typescript-config": "workspace:*",
     "@types/node": "^24.0.0",
     "esbuild": "^0.25.12",
     "eslint": "^9.39.1",
-    "typescript": "^5.9.2"
+    "typescript": "^5.9.2",
+    "@repo/eslint-config": "0.0.0",
+    "@repo/ingest-core": "0.0.0",
+    "@repo/typescript-config": "0.0.0",
+    "@repo/llm-core": "0.0.0"
+  },
+  "scripts": {
+    "build": "tsc -p tsconfig.json",
+    "bundle": "node scripts/bundle.mjs",
+    "dev": "tsc -p tsconfig.json --watch",
+    "lint": "eslint .",
+    "check-types": "tsc --noEmit",
+    "test": "tsc -p tsconfig.json && node --test dist/*.test.js"
   }
-}
+}