@infinitedusky/indusk-mcp 1.11.0 → 1.11.1

package/dist/lib/eval/judge-runner.js

@@ -5,7 +5,7 @@
  * The judge is a detached child process so the calling hook can exit immediately.
  * Results appear asynchronously in `.indusk/eval/results.log`.
  */
-import { execSync, spawn } from "node:child_process";
+import { spawn } from "node:child_process";
 import { join } from "node:path";
 import { getProjectGroupId } from "../config.js";
 import { EvalLogWriter } from "./log-writer.js";
@@ -14,14 +14,6 @@ import { V1_RUBRIC } from "./rubric.js";
 function getEvalLogPath(projectRoot) {
     return join(projectRoot, ".indusk", "eval", "results.log");
 }
-function getDiff(changeId) {
-    try {
-        return execSync(`jj diff -r ${changeId}`, { encoding: "utf8", maxBuffer: 10 * 1024 * 1024 });
-    }
-    catch {
-        return "(diff unavailable)";
-    }
-}
 async function postTelemetry(endpoint, scorecard) {
     try {
         const controller = new AbortController();
@@ -46,13 +38,11 @@ async function postTelemetry(endpoint, scorecard) {
  * If anything fails, logs an error entry instead of silently dropping.
  */
 export function runJudgeBackground(opts) {
-    const diff = getDiff(opts.changeId);
     const projectGroup = getProjectGroupId(opts.projectRoot);
     const prompt = buildJudgePrompt({
         rubric: V1_RUBRIC,
         changeId: opts.changeId,
         transcriptPath: opts.transcriptPath,
-        diff,
         mode: opts.mode,
         projectGroup,
     });
@@ -75,16 +65,18 @@ export function runJudgeBackground(opts) {
         "--permission-mode",
         "acceptEdits",
         "--allowed-tools",
-        ...allowedTools,
-        prompt,
+        allowedTools.join(","),
     ];
+    // Not detached — the eval-trigger hook already spawns this in a separate
+    // node process. Detaching + unref causes the close handler to never fire.
     const child = spawn("claude", args, {
         cwd: opts.projectRoot,
-        stdio: ["ignore", "pipe", "pipe"],
-        detached: true,
+        stdio: ["pipe", "pipe", "pipe"],
         env: { ...process.env },
     });
-    child.unref();
+    // Pipe the prompt via stdin (too large for CLI arg)
+    child.stdin?.write(prompt);
+    child.stdin?.end();
     let stdout = "";
     let stderr = "";
     child.stdout?.on("data", (chunk) => {
@@ -140,13 +132,11 @@ export function runJudgeBackground(opts) {
  * Returns the scorecard or error entry.
  */
 export async function runJudgeSync(opts) {
-    const diff = getDiff(opts.changeId);
     const projectGroup = getProjectGroupId(opts.projectRoot);
     const prompt = buildJudgePrompt({
         rubric: V1_RUBRIC,
         changeId: opts.changeId,
         transcriptPath: opts.transcriptPath,
-        diff,
         mode: opts.mode,
         projectGroup,
     });
@@ -169,15 +159,16 @@ export async function runJudgeSync(opts) {
         "--permission-mode",
         "acceptEdits",
         "--allowed-tools",
-        ...allowedTools,
-        prompt,
+        allowedTools.join(","),
     ];
     return new Promise((resolve) => {
         const child = spawn("claude", args, {
             cwd: opts.projectRoot,
-            stdio: ["ignore", "pipe", "pipe"],
+            stdio: ["pipe", "pipe", "pipe"],
             env: { ...process.env },
         });
+        child.stdin?.write(prompt);
+        child.stdin?.end();
         let stdout = "";
         let stderr = "";
         child.stdout?.on("data", (chunk) => {

package/dist/lib/eval/prompt-builder.d.ts

@@ -2,15 +2,17 @@
  * Builds the judge agent's system prompt.
  *
  * The prompt instructs the judge to: do catchup, read the transcript, read the
- * diff, answer each rubric question, write findings to Graphiti (eval mode
- * only), and output a JSON scorecard.
+ * diff itself via jj, answer each rubric question, write findings to Graphiti
+ * (eval mode only), and output a JSON scorecard.
+ *
+ * The diff is NOT embedded in the prompt — the judge reads it via tool calls.
+ * This keeps the prompt small regardless of commit size.
  */
 import type { RubricQuestion } from "./types.js";
 export interface PromptBuilderOptions {
     rubric: RubricQuestion[];
     changeId: string;
     transcriptPath: string;
-    diff: string;
     mode: "eval" | "baseline";
     projectGroup: string;
 }

package/dist/lib/eval/prompt-builder.js

@@ -2,8 +2,11 @@
  * Builds the judge agent's system prompt.
  *
  * The prompt instructs the judge to: do catchup, read the transcript, read the
- * diff, answer each rubric question, write findings to Graphiti (eval mode
- * only), and output a JSON scorecard.
+ * diff itself via jj, answer each rubric question, write findings to Graphiti
+ * (eval mode only), and output a JSON scorecard.
+ *
+ * The diff is NOT embedded in the prompt — the judge reads it via tool calls.
+ * This keeps the prompt small regardless of commit size.
  */
 export function buildJudgePrompt(opts) {
     const questionsBlock = opts.rubric
@@ -54,13 +57,9 @@ This is the JSONL record of the working agent's session. Read it to understand:
 
 ### Step 3: Read the diff
 
-Here is the diff of the committed work:
-
-\`\`\`
-${opts.diff}
-\`\`\`
+Run \`jj diff -r ${opts.changeId}\` to see what was committed. This is the work being evaluated.
 
-This is what was actually built. Cross-reference with the transcript to understand the journey from task to result.
+Then read the specific files that were changed to understand the full context — not just the diff lines, but the surrounding code.
 
 ### Step 4: Answer the evaluation questions
 

package/hooks/eval-trigger.js

@@ -90,17 +90,18 @@ const transcriptPath =
 	"(transcript unavailable)";
 
 // Spawn the judge runner as a detached background process.
-// We use a small inline node script that imports and calls runJudgeBackground.
-// This avoids needing the compiled dist/ to exist at hook time.
+// Spawn a detached node process that calls runJudgeSync (which awaits completion).
+// runJudgeSync keeps the process alive until claude --print finishes and logs the result.
 const judgeScript = `
 import("${resolve(projectRoot, "apps/indusk-mcp/dist/lib/eval/judge-runner.js")}")
-  .then(m => m.runJudgeBackground({
+  .then(m => m.runJudgeSync({
     projectRoot: ${JSON.stringify(projectRoot)},
     changeId: ${JSON.stringify(changeId)},
     transcriptPath: ${JSON.stringify(transcriptPath)},
     mode: "eval",
     evalEndpoint: ${JSON.stringify(evalConfig.endpoint)},
   }))
+  .then(() => process.exit(0))
   .catch(err => {
     const fs = require("fs");
     const path = require("path");
@@ -115,6 +116,7 @@ import("${resolve(projectRoot, "apps/indusk-mcp/dist/lib/eval/judge-runner.js")}
       message: err.message || String(err),
     });
     fs.appendFileSync(logPath, entry + "\\n", "utf8");
+    process.exit(1);
   });
 `;
 

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@infinitedusky/indusk-mcp",
-	"version": "1.11.0",
+	"version": "1.11.1",
 	"description": "InDusk development system — skills, MCP tools, and CLI for structured AI-assisted development",
 	"type": "module",
 	"files": [