@yemi33/minions 0.1.1104 → 0.1.1106

package/CHANGELOG.md

@@ -1,11 +1,15 @@
 # Changelog
 
-## 0.1.1104 (2026-04-18)
+## 0.1.1106 (2026-04-18)
 
 ### Features
 - seed realActivityMap at spawn time, stamp pid in live-output (#1200)
 
 ### Fixes
+- preserve buildErrorLog through transient build states (#1232) (#1274)
+- annotate fast-exit empty-output failures with diagnostic hint (#1276)
+- invalidate PRD cache on pr-links.json change + guard aggregate PRs (#1220) (#1272)
+- pass --add-dir for minions + ~/.claude to agents (#1271)
 - preserve VERDICT marker by tail-slicing agent output (#1234) (#1270)
 - PRD info cache staleness and aggregate PR bleed-through (#1222)
 -  avoid no-op work item writes

package/engine/ado.js

@@ -465,6 +465,11 @@ async function pollPrStatus(config) {
       } catch (e) { log('warn', `ADO build query for ${pr.id}: ${e.message}`); }
     }
 
+    // Record actual poll time — makes lastBuildCheck reflect when the engine last
+    // talked to ADO, not when the agent was dispatched. Issue #1233.
+    pr.lastBuildCheck = ts();
+    updated = true;
+
     if (pr.buildStatus !== buildStatus) {
       log('info', `PR ${pr.id} build: ${pr.buildStatus || 'none'} → ${buildStatus}${buildFailReason ? ' (' + buildFailReason + ')' : ''}`);
       pr.buildStatus = buildStatus;
@@ -475,9 +480,17 @@ async function pollPrStatus(config) {
       if (buildStatus === 'failing') delete pr._autoCompleted;
       if (buildStatus !== 'failing') {
         delete pr._buildFailNotified;
-        delete pr.buildErrorLog;
-        // Reset build fix retry counter on recovery — allows fresh auto-fix cycles if build breaks again
-        if (pr.buildFixAttempts) { delete pr.buildFixAttempts; delete pr.buildFixEscalated; }
+        // Preserve buildErrorLog + buildFixAttempts through transient 'none'/'running'
+        // transitions — only clear on confirmed 'passing' recovery. Issue #1232: 'none'
+        // can also occur when ADO recomputes the merge commit after a target-branch
+        // update but no new builds have been triggered yet (filter by sourceVersion
+        // returns []), which previously wiped the last known error log and caused
+        // fix agents to be dispatched blind.
+        if (buildStatus === 'passing') {
+          delete pr.buildErrorLog;
+          // Reset build fix retry counter on recovery — allows fresh auto-fix cycles if build breaks again
+          if (pr.buildFixAttempts) { delete pr.buildFixAttempts; delete pr.buildFixEscalated; }
+        }
       }
       updated = true;
 

package/engine/github.js

@@ -422,6 +422,10 @@ async function pollPrStatus(config) {
     if (prData.state === 'open' && prData.head?.sha) {
       const checksData = await ghApi(`/commits/${prData.head.sha}/check-runs`, slug);
       if (checksData && checksData.check_runs) {
+        // Record actual poll time — makes lastBuildCheck reflect when the engine last
+        // talked to GitHub, not when the agent was dispatched. Issue #1233.
+        pr.lastBuildCheck = ts();
+        updated = true;
         const runs = checksData.check_runs;
         let buildStatus = 'none';
         let buildFailReason = '';
@@ -452,9 +456,15 @@ async function pollPrStatus(config) {
           if (buildStatus === 'failing') delete pr._autoCompleted; // allow re-merge after fix
           if (buildStatus !== 'failing') {
             delete pr._buildFailNotified;
-            delete pr.buildErrorLog;
-            // Reset build fix retry counter on recovery — allows fresh auto-fix cycles if build breaks again
-            if (pr.buildFixAttempts) { delete pr.buildFixAttempts; delete pr.buildFixEscalated; }
+            // Preserve buildErrorLog + buildFixAttempts through transient 'none'/'running'
+            // transitions — only clear on confirmed 'passing' recovery. Issue #1232:
+            // clearing on every non-failing transition blinded the next fix agent
+            // while a queued build was still running.
+            if (buildStatus === 'passing') {
+              delete pr.buildErrorLog;
+              // Reset build fix retry counter on recovery — allows fresh auto-fix cycles if build breaks again
+              if (pr.buildFixAttempts) { delete pr.buildFixAttempts; delete pr.buildFixEscalated; }
+            }
           }
           updated = true;
 

package/engine/lifecycle.js

@@ -2118,6 +2118,28 @@ function classifyFailure(code, stdout = '', stderr = '') {
   return FAILURE_CLASS.UNKNOWN;
 }
 
+/**
+ * When a claude CLI agent exits in <3s with code 1 and no output, the raw
+ * EMPTY_OUTPUT class tells us "no meaningful output" but nothing about WHY.
+ * This helper detects the fast-exit pattern and returns a diagnostic hint
+ * listing likely root causes (machine sleep, network loss, auth failure).
+ *
+ * The hint is propagated as the dispatch `reason` and into the work-item
+ * `failReason`, so humans and other agents can triage without digging into
+ * `agents/*\/output-<id>.log`.
+ *
+ * @param {string|undefined} failureClass — one of FAILURE_CLASS values
+ * @param {number} code — process exit code
+ * @param {number} elapsedMs — milliseconds between spawn and close
+ * @returns {string|null} — annotated hint, or null when the pattern doesn't match
+ */
+function diagnoseEmptyOutput(failureClass, code, elapsedMs) {
+  if (failureClass !== FAILURE_CLASS.EMPTY_OUTPUT) return null;
+  if (code !== 1) return null;
+  if (!Number.isFinite(elapsedMs) || elapsedMs < 0 || elapsedMs >= 3000) return null;
+  return `[empty-output: process exited in ${elapsedMs}ms \u2014 possible causes: machine sleep, network unavailability, auth failure]`;
+}
+
 module.exports = {
   checkPlanCompletion,
   archivePlan,
@@ -2142,6 +2164,7 @@ module.exports = {
   resolveWorkItemPath,
   isItemCompleted,
   classifyFailure,
+  diagnoseEmptyOutput,
   processPendingRebases,
   findDependentActivePrs,
 };

package/engine/queries.js

@@ -1028,7 +1028,7 @@ function getPrdInfo(config) {
     const pr = prById[prId];
     // Skip aggregate / E2E PRs from per-item mapping — they link to multiple items
     // (or are typed as verify) and would bleed through as duplicate entries on every
-    // constituent item. They are surfaced via renderE2eSection instead.
+    // constituent item. They are surfaced via renderE2eSection instead. (#1220)
     if ((itemIds || []).length > 1 || pr?.itemType === 'verify' || pr?.title?.startsWith('[E2E]')) continue;
     const project = projects.find(p => p.name === pr?._project) || projects[0] || null;
     const prNumber = shared.getPrNumber(pr || prId);

package/engine/spawn-agent.js

@@ -7,6 +7,7 @@
  */
 
 const fs = require('fs');
+const os = require('os');
 const path = require('path');
 const { exec, runFile, cleanChildEnv, killGracefully, killImmediate, ts, safeJson, safeWrite } = require('./shared');
 
@@ -103,16 +104,27 @@ if (!fs.existsSync(tmpDir)) fs.mkdirSync(tmpDir, { recursive: true });
 const debugPath = path.join(tmpDir, 'spawn-debug.log');
 fs.writeFile(debugPath, `spawn-agent.js at ${ts()}\nclaudeBin=${claudeBin || 'not found'}\nnative=${claudeIsNative}\nprompt=${promptFile}\nsysPrompt=${sysPromptFile}\nextraArgs=${extraArgs.join(' ')}\n`, () => {});
 
+// Skill discovery dirs — agents run with CWD set to an external repo worktree,
+// so skills in the minions repo and the user's global ~/.claude dir are otherwise
+// invisible. Pass them via --add-dir on every invocation (resume included, since
+// each resume spawns a fresh CLI process with the same CWD). See issue #1231.
+const minionsDir = path.resolve(__dirname, '..');
+const userClaudeDir = path.join(os.homedir(), '.claude');
+const addDirArgs = ['--add-dir', minionsDir];
+if (fs.existsSync(userClaudeDir) && path.resolve(userClaudeDir) !== path.resolve(minionsDir)) {
+  addDirArgs.push('--add-dir', userClaudeDir);
+}
+
 // When resuming a session, skip system prompt (it's baked into the session)
 const isResume = extraArgs.includes('--resume');
 const sysTmpPath = sysPromptFile + '.tmp';
 let cliArgs;
 if (isResume) {
-  cliArgs = ['-p', ...extraArgs];
+  cliArgs = ['-p', ...addDirArgs, ...extraArgs];
 } else {
   // Pass system prompt via file to avoid ENAMETOOLONG on Windows (32KB arg limit)
   fs.writeFileSync(sysTmpPath, sysPrompt);
-  cliArgs = ['-p', '--system-prompt-file', sysTmpPath, ...extraArgs];
+  cliArgs = ['-p', '--system-prompt-file', sysTmpPath, ...addDirArgs, ...extraArgs];
 }
 
 if (!claudeBin) {

package/engine.js

@@ -138,7 +138,7 @@ const { renderPlaybook, validatePlaybookVars, PLAYBOOK_REQUIRED_VARS,
 const { runPostCompletionHooks, updateWorkItemStatus, syncPrdItemStatus, reconcilePrdStatuses, handlePostMerge, checkPlanCompletion,
   syncPrsFromOutput, updatePrAfterReview, updatePrAfterFix, checkForLearnings, extractSkillsFromOutput,
   updateAgentHistory, updateMetrics, createReviewFeedbackForAuthor, parseAgentOutput, syncPrdFromPrs,
-  isItemCompleted, classifyFailure, processPendingRebases, resolveWorkItemPath } = require('./engine/lifecycle');
+  isItemCompleted, classifyFailure, diagnoseEmptyOutput, processPendingRebases, resolveWorkItemPath } = require('./engine/lifecycle');
 
 // ─── Agent Spawner ──────────────────────────────────────────────────────────
 
@@ -1231,6 +1231,13 @@ async function spawnAgent(dispatchItem, config) {
     let errorReason = '';
     if (effectiveResult === DISPATCH_RESULT.ERROR) {
       errorReason = stderr.split('\n').filter(l => l.trim()).slice(-5).join(' | ').trim().slice(0, 300);
+      // W-mo3zul9pirjb — when claude CLI exits in <3s with code 1 and no output (the
+      // "silent crash" pattern seen during scheduled tasks when the box went to sleep
+      // or lost network), prepend a diagnostic hint so failReason/dispatch log carry
+      // the actual elapsed time and likely root causes rather than a bare "[empty-output]".
+      const elapsedMs = Date.now() - Date.parse(startedAt);
+      const hint = diagnoseEmptyOutput(failureClass, code, elapsedMs);
+      if (hint) errorReason = errorReason ? `${hint} ${errorReason}` : hint;
     }
     completeDispatch(id, effectiveResult, errorReason, resultSummary, completeOpts);
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@yemi33/minions",
-  "version": "0.1.1104",
+  "version": "0.1.1106",
   "description": "Multi-agent AI dev team that runs from ~/.minions/ — five autonomous agents share a single engine, dashboard, and knowledge base",
   "bin": {
     "minions": "bin/minions.js"