open-agents-ai 0.187.532 → 0.187.533

package/dist/index.js

@@ -526818,6 +526818,85 @@ Pick the SMALLEST concrete deliverable from the spec — typically the project e
           }
         }
       }
+      /**
+       * REG-66 Debug-Loop Detection (root-cause from batch535-midi, 2026-05-04).
+       *
+       * Empirical: midi run had 11x `npm run build 2>&1` + same 5 files re-read
+       * 5-6 times each + 22 BFC-61.G coercion BLOCKS — and ZERO of those blocks
+       * resulted in a creative edit. The agent was rationally stuck: it
+       * believed it needed to read more to debug, the build command kept
+       * giving the same error, and the standard "issue an edit" directive
+       * gave no traction because the agent had no concrete edit hypothesis.
+       *
+       * This method analyzes toolCallLog for the debug-loop signature:
+       *   - Same shell command stem repeated ≥5 times in the trailing window, OR
+       *   - Same file_read path re-read ≥4 times in the trailing window.
+       * Both indicate the agent is reading/running the same things hoping for
+       * different output. Without this signal we'd just keep telling the
+       * agent to "make an edit" — which is exactly what it can't think of.
+       *
+       * When detected, the BFC-61.G block message swaps to a PERTURB-strategy
+       * directive: stop reading, change ONE thing in the most-likely-culprit
+       * file even if you're uncertain, and let the new error signal guide
+       * the next iteration. This is real human debugging strategy ("perturb
+       * to disambiguate"), NOT reward-hacking — the agent still has to
+       * produce a real edit and the success criteria (todos done + build
+       * passing) are unchanged.
+       *
+       * @returns Detection result. `detected=false` → use standard message.
+       *          `detected=true` → use REG-66 perturb-strategy message;
+       *          `repeatedSample` carries the offending command/path for the
+       *          message body so the agent sees the specific pattern called out.
+       */
+      _detectDebugLoop(toolCallLog) {
+        if (process.env["OA_DISABLE_REG66"] === "1")
+          return { detected: false };
+        const WINDOW = 20;
+        const SHELL_REPEAT_THRESHOLD = 5;
+        const READ_REPEAT_THRESHOLD = 4;
+        const window2 = toolCallLog.slice(-WINDOW);
+        if (window2.length < SHELL_REPEAT_THRESHOLD)
+          return { detected: false };
+        const _editClasses = /* @__PURE__ */ new Set(["file_write", "file_edit", "batch_edit", "file_patch"]);
+        for (const c9 of window2) {
+          if (_editClasses.has(c9.name) && c9.success !== false)
+            return { detected: false };
+        }
+        const shellCounts = /* @__PURE__ */ new Map();
+        const readCounts = /* @__PURE__ */ new Map();
+        for (const c9 of window2) {
+          if (c9.name === "shell") {
+            const m2 = c9.argsKey.match(/(?:^|,)command=([^,]+)/);
+            if (m2 && m2[1]) {
+              const stem = m2[1].trim();
+              shellCounts.set(stem, (shellCounts.get(stem) ?? 0) + 1);
+            }
+          } else if (c9.name === "file_read" || c9.name === "file_explore") {
+            const m2 = c9.argsKey.match(/(?:^|,)path=([^,]+)/);
+            if (m2 && m2[1]) {
+              const stem = m2[1].trim();
+              readCounts.set(stem, (readCounts.get(stem) ?? 0) + 1);
+            }
+          }
+        }
+        let bestShell = null;
+        for (const [k, n2] of shellCounts) {
+          if (n2 >= SHELL_REPEAT_THRESHOLD && (!bestShell || n2 > bestShell[1]))
+            bestShell = [k, n2];
+        }
+        let bestRead = null;
+        for (const [k, n2] of readCounts) {
+          if (n2 >= READ_REPEAT_THRESHOLD && (!bestRead || n2 > bestRead[1]))
+            bestRead = [k, n2];
+        }
+        if (bestShell) {
+          return { detected: true, repeatedSample: bestShell[0], count: bestShell[1], kind: "shell" };
+        }
+        if (bestRead) {
+          return { detected: true, repeatedSample: bestRead[0], count: bestRead[1], kind: "read" };
+        }
+        return { detected: false };
+      }
       readSessionTodos() {
         try {
           const sid = process.env["OA_SESSION_ID"] || this._sessionId || "default";
@@ -530749,7 +530828,32 @@ ${memoryLines.join("\n")}`
                   turn,
                   timestamp: (/* @__PURE__ */ new Date()).toISOString()
                 });
-                const reg61BlockMsg = [
+                const _dbgLoop = this._detectDebugLoop(toolCallLog);
+                const _debugLoopSampleSafe = (_dbgLoop.repeatedSample ?? "").slice(0, 120);
+                const reg61BlockMsg = _dbgLoop.detected ? [
+                  `[BLOCKED — REG-61 directive in effect — REG-66 DEBUG-LOOP detected]`,
+                  ``,
+                  `Pattern: ${_dbgLoop.kind === "shell" ? "shell command" : "file"} "${_debugLoopSampleSafe}" was used ${_dbgLoop.count}× in the trailing window with ZERO creative edits landing. You are stuck in a debug loop where re-running / re-reading is producing no new information.`,
+                  ``,
+                  `STOP DEBUGGING. PERTURB.`,
+                  ``,
+                  `Strategy when stuck like this (real human debuggers do this):`,
+                  `  1. Pick the source file most likely implicated by the recurring failure (probably in src/, the one most-imported by failing tests).`,
+                  `  2. Pick ONE plausible cause — most-recently-modified line, most-complex function, most-likely-misnamed import, most-likely off-by-one.`,
+                  `  3. Make a SPECULATIVE edit that changes that thing — even if you are NOT certain it'll fix the bug. The point is to get a NEW error signal that disambiguates.`,
+                  `  4. Re-run the failing command. If the error CHANGED, you've learned something. If it's identical, you've ruled out one hypothesis.`,
+                  ``,
+                  `This is NOT random guessing — it's targeted hypothesis falsification. Reading the same files 5+ times has already proven uninformative; only a state change will move the system.`,
+                  ``,
+                  `Issue EXACTLY ONE of: file_write / file_edit / batch_edit / file_patch on a single concrete change. The exact CHOICE of edit matters less than NOT continuing to re-read.`,
+                  ``,
+                  `Allowed bypasses (will not be blocked but will not clear the directive either):`,
+                  `  • web_search    — search the EXACT recurring error string`,
+                  `  • task_complete — exit if you genuinely cannot identify any plausible perturbation`,
+                  `  • ask_user      — escalate to human (if available)`,
+                  ``,
+                  `Once you make a real edit, the directive clears and you'll see the new test result.`
+                ].join("\n") : [
                   `[BLOCKED — REG-61 directive in effect]`,
                   ``,
                   `A REG-61 FIRST-EDIT NUDGE was issued earlier and has not yet been satisfied. The directive: your next tool call MUST be a creative edit. You issued '${tc.name}' instead, which is a read/explore/shell call. This call has been BLOCKED.`,
@@ -530777,12 +530881,12 @@ ${memoryLines.join("\n")}`
                 });
                 this.emit({
                   type: "status",
-                  content: `REG-61 COERCION BLOCK — rejected '${tc.name}' at turn ${turn}; gate stays active until creative edit dispatches`,
+                  content: `REG-61 COERCION BLOCK — rejected '${tc.name}' at turn ${turn}; gate stays active until creative edit dispatches${_dbgLoop.detected ? `; REG-66 debug-loop variant (${_dbgLoop.kind} "${_debugLoopSampleSafe.slice(0, 60)}" ${_dbgLoop.count}×)` : ""}`,
                   timestamp: (/* @__PURE__ */ new Date()).toISOString()
                 });
                 this._tagSyntheticFailure({
                   mode: "step_repetition",
-                  rationale: `REG-61 perpetual coercion block on '${tc.name}' — agent ignored FIRST-EDIT NUDGE`
+                  rationale: `REG-61 perpetual coercion block on '${tc.name}' — agent ignored FIRST-EDIT NUDGE${_dbgLoop.detected ? " (debug-loop variant)" : ""}`
                 });
                 return { tc, output: reg61BlockMsg };
               }
@@ -599758,6 +599862,15 @@ async function handleRequest(req2, res, ollamaUrl, verbose) {
       });
       return;
     }
+    if (pathname === "/favicon.ico" && method === "GET") {
+      const svg = '<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 16 16"><rect width="16" height="16" rx="3" fill="#b2920a"/><text x="50%" y="55%" font-size="11" font-family="monospace" font-weight="700" text-anchor="middle" dominant-baseline="middle" fill="#0b0b0b">oa</text></svg>';
+      res.writeHead(200, {
+        "Content-Type": "image/svg+xml",
+        "Cache-Control": "public, max-age=86400"
+      });
+      res.end(svg);
+      return;
+    }
     if (pathname === "/" && method === "GET" && req2.headers.accept?.includes("text/html")) {
       res.writeHead(200, {
         "Content-Type": "text/html; charset=utf-8",

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "open-agents-ai",
-  "version": "0.187.532",
+  "version": "0.187.533",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "open-agents-ai",
-      "version": "0.187.532",
+      "version": "0.187.533",
       "hasInstallScript": true,
       "license": "CC-BY-NC-4.0",
       "dependencies": {
@@ -3132,9 +3132,9 @@
       }
     },
     "node_modules/express-rate-limit": {
-      "version": "8.4.1",
-      "resolved": "https://registry.npmjs.org/express-rate-limit/-/express-rate-limit-8.4.1.tgz",
-      "integrity": "sha512-NGVYwQSAyEQgzxX1iCM978PP9AdO/hW93gMcF6ZwQCm+rFvLsBH6w4xcXWTcliS8La5EPRN3p9wzItqBwJrfNw==",
+      "version": "8.5.0",
+      "resolved": "https://registry.npmjs.org/express-rate-limit/-/express-rate-limit-8.5.0.tgz",
+      "integrity": "sha512-XKhFohWaSBdVJNTi5TaHziqnPkv04I9UQV6q1Wy7Ui6GGQZVW12ojDFwqer14EvCXxjvPG0CyWXx7cAXpALB4Q==",
       "license": "MIT",
       "dependencies": {
         "ip-address": "10.1.0"

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "open-agents-ai",
-  "version": "0.187.532",
+  "version": "0.187.533",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",