open-agents-ai 0.187.468 → 0.187.470

package/dist/index.js

@@ -250490,6 +250490,11 @@ import { execSync as execSync20, exec } from "node:child_process";
 import { existsSync as existsSync24, writeFileSync as writeFileSync9, mkdirSync as mkdirSync8 } from "node:fs";
 import { join as join37, dirname as dirname11 } from "node:path";
 import { homedir as homedir9 } from "node:os";
+function pushBounded(buf, item) {
+  buf.push(item);
+  if (buf.length > MAX_BUFFER)
+    buf.splice(0, buf.length - MAX_BUFFER);
+}
 async function ensurePlaywright() {
   if (pw)
     return null;
@@ -250529,28 +250534,78 @@ async function ensureBrowser() {
         userAgent: "Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/120.0.0.0 Safari/537.36"
       });
       page = await context.newPage();
-      page.on("dialog", async (dialog) => {
-        try {
-          await dialog.dismiss();
-        } catch {
-        }
-      });
+      attachDiagnosticListeners(page);
     } catch (err) {
       return `Failed to launch browser: ${err instanceof Error ? err.message : String(err)}`;
     }
   }
   if (!page || page.isClosed?.()) {
     page = await context.newPage();
+    attachDiagnosticListeners(page);
   }
   return null;
 }
+function attachDiagnosticListeners(p2) {
+  p2.on("dialog", async (dialog) => {
+    try {
+      await dialog.dismiss();
+    } catch {
+    }
+  });
+  p2.on("console", (msg) => {
+    try {
+      const loc = msg.location?.();
+      pushBounded(consoleBuffer, {
+        ts: Date.now(),
+        type: String(msg.type?.() ?? "log"),
+        text: String(msg.text?.() ?? "").slice(0, 600),
+        loc: loc?.url ? `${loc.url}:${loc.lineNumber ?? 0}:${loc.columnNumber ?? 0}` : void 0
+      });
+    } catch {
+    }
+  });
+  p2.on("pageerror", (err) => {
+    pushBounded(errorBuffer, {
+      ts: Date.now(),
+      message: String(err?.message ?? err).slice(0, 600),
+      stack: typeof err?.stack === "string" ? err.stack.slice(0, 1500) : void 0
+    });
+  });
+  p2.on("requestfailed", (req2) => {
+    pushBounded(networkBuffer, {
+      ts: Date.now(),
+      method: String(req2.method?.() ?? "GET"),
+      url: String(req2.url?.() ?? "").slice(0, 600),
+      failure: String(req2.failure?.()?.errorText ?? "request failed").slice(0, 200),
+      ok: false
+    });
+  });
+  p2.on("response", (resp) => {
+    try {
+      const status = Number(resp.status?.() ?? 0);
+      pushBounded(networkBuffer, {
+        ts: Date.now(),
+        method: String(resp.request?.()?.method?.() ?? "GET"),
+        url: String(resp.url?.() ?? "").slice(0, 600),
+        status,
+        ok: status >= 200 && status < 400
+      });
+    } catch {
+    }
+  });
+}
+function clearDiagnosticBuffers() {
+  consoleBuffer = [];
+  networkBuffer = [];
+  errorBuffer = [];
+}
 function ok(output, start2) {
   return { success: true, output, durationMs: Date.now() - start2 };
 }
 function fail(error, start2) {
   return { success: false, output: "", error, durationMs: Date.now() - start2 };
 }
-var pw, browser2, context, page, dynamicImport, PlaywrightBrowserTool;
+var pw, browser2, context, page, MAX_BUFFER, consoleBuffer, networkBuffer, errorBuffer, dynamicImport, PlaywrightBrowserTool;
 var init_playwright_browser = __esm({
   "packages/execution/dist/tools/playwright-browser.js"() {
     "use strict";
@@ -250558,10 +250613,14 @@ var init_playwright_browser = __esm({
     browser2 = null;
     context = null;
     page = null;
+    MAX_BUFFER = 200;
+    consoleBuffer = [];
+    networkBuffer = [];
+    errorBuffer = [];
     dynamicImport = new Function("mod", "return import(mod)");
     PlaywrightBrowserTool = class {
       name = "playwright_browser";
-      description = "Full-scope Playwright browser automation. Launches a persistent headless Chromium session for interactive web tasks. Use this to: navigate pages, click elements, fill forms, take screenshots, evaluate JavaScript, wait for elements, extract content, and run end-to-end test assertions against live websites. Auto-installs Playwright + Chromium on first use. Actions: navigate, click, fill, type, press, screenshot, evaluate, content, select, check, wait, goto, title, url, pdf, close. Start with navigate, then use content/screenshot to see the page, then click/fill/type to interact.";
+      description = "Full-scope Playwright browser automation + diagnostic capture. Launches a persistent headless Chromium session. Beyond navigation/interaction, this tool buffers everything the running app emits (console messages, network requests, JS exceptions, accessibility tree) so the agent can verify what is ACTUALLY happening — not just what the build/test reports. Auto-installs Playwright + Chromium on first use. Diagnostic actions: console_logs, network_log, page_errors, a11y_snapshot, bounding_box, query_all, performance, cookies, storage, viewport, clear_diagnostics. Interaction actions: navigate, click, fill, type, press, select, check, hover. Capture actions: screenshot, pdf, content, innerText, innerHTML, getAttribute, evaluate. Workflow: navigate → screenshot or content → console_logs/network_log/page_errors to verify state → click/fill/type to drive UI → repeat.";
       parameters = {
         type: "object",
         properties: {
@@ -250593,6 +250652,18 @@ var init_playwright_browser = __esm({
               "goBack",
               "goForward",
               "reload",
+              // FEAT-1: diagnostic capture actions
+              "console_logs",
+              "network_log",
+              "page_errors",
+              "a11y_snapshot",
+              "bounding_box",
+              "query_all",
+              "performance",
+              "cookies",
+              "storage",
+              "viewport",
+              "clear_diagnostics",
               "close"
             ],
             description: "Action to perform:\n- navigate: go to a URL\n- click: click element by selector\n- fill: clear input and type text (for form fields)\n- type: type text character by character (simulates real typing)\n- press: press a key (Enter, Tab, Escape, etc.)\n- screenshot: capture page screenshot (saves to file)\n- evaluate: run JavaScript in page context\n- content: get page text content (readable, stripped)\n- innerText: get innerText of a specific element\n- select: select dropdown option by value\n- check/uncheck: toggle checkbox\n- hover: hover over element\n- wait: wait for a selector to appear\n- waitForNavigation: wait for page navigation to complete\n- waitForSelector: wait for element matching selector\n- title: get page title\n- url: get current URL\n- getAttribute: get element attribute value\n- innerHTML: get element's innerHTML\n- textContent: get element's textContent\n- goBack/goForward/reload: browser navigation\n- pdf: save page as PDF\n- close: close browser session"
@@ -250641,6 +250712,7 @@ var init_playwright_browser = __esm({
           page = null;
           context = null;
           browser2 = null;
+          clearDiagnosticBuffers();
           return { success: true, output: "Browser closed.", durationMs: Date.now() - start2 };
         }
         const err = await ensureBrowser();
@@ -250831,6 +250903,171 @@ var init_playwright_browser = __esm({
               await page.pdf({ path: pdfPath, format: "A4" });
               return ok(`PDF saved to ${pdfPath}`, start2);
             }
+            // ── FEAT-1: Diagnostic capture ────────────────────────────────────
+            case "console_logs": {
+              const filter2 = (text || "").toLowerCase();
+              const entries = filter2 ? consoleBuffer.filter((e2) => e2.type.toLowerCase().includes(filter2)) : consoleBuffer;
+              const tail = entries.slice(-50);
+              if (tail.length === 0) {
+                return ok(`No console messages buffered yet${filter2 ? ` (filter="${filter2}")` : ""}. Buffer holds last 200 messages.`, start2);
+              }
+              const lines = tail.map((e2) => {
+                const at = new Date(e2.ts).toISOString().slice(11, 19);
+                return `[${at}] ${e2.type.toUpperCase()}: ${e2.text}${e2.loc ? `  (${e2.loc})` : ""}`;
+              });
+              return ok(`Console messages (${tail.length}/${entries.length} shown${filter2 ? `, filter="${filter2}"` : ""}):
+${lines.join("\n")}`, start2);
+            }
+            case "network_log": {
+              const mode = (text || "").toLowerCase();
+              let entries = networkBuffer;
+              if (mode === "failed" || mode === "errors") {
+                entries = entries.filter((e2) => e2.ok === false);
+              }
+              const tail = entries.slice(-50);
+              if (tail.length === 0) {
+                return ok(`No network entries buffered${mode ? ` (mode="${mode}")` : ""}. Navigate to a URL to populate.`, start2);
+              }
+              const lines = tail.map((e2) => {
+                const at = new Date(e2.ts).toISOString().slice(11, 19);
+                const tag = e2.failure ? `FAILED (${e2.failure})` : `${e2.status ?? "?"}`;
+                return `[${at}] ${e2.method} ${tag}  ${e2.url}`;
+              });
+              return ok(`Network log (${tail.length}/${entries.length} shown${mode ? `, mode="${mode}"` : ""}):
+${lines.join("\n")}`, start2);
+            }
+            case "page_errors": {
+              if (errorBuffer.length === 0) {
+                return ok("No uncaught page errors recorded.", start2);
+              }
+              const lines = errorBuffer.map((e2) => {
+                const at = new Date(e2.ts).toISOString().slice(11, 19);
+                return `[${at}] ${e2.message}${e2.stack ? `
+  ${e2.stack.split("\n").slice(0, 4).join("\n  ")}` : ""}`;
+              });
+              return ok(`Uncaught page errors (${errorBuffer.length}):
+${lines.join("\n\n")}`, start2);
+            }
+            case "a11y_snapshot": {
+              const root = selector ? await page.locator(selector).first().elementHandle() : null;
+              const tree2 = await page.accessibility.snapshot({ root: root ?? void 0, interestingOnly: true });
+              if (!tree2)
+                return ok("No accessibility tree (page may be empty).", start2);
+              const render2 = (n2, depth = 0) => {
+                const pad = "  ".repeat(depth);
+                const role = n2.role ?? "?";
+                const name10 = n2.name ? ` "${String(n2.name).slice(0, 80)}"` : "";
+                const value3 = n2.value ? ` value="${String(n2.value).slice(0, 60)}"` : "";
+                const checked = n2.checked != null ? ` checked=${n2.checked}` : "";
+                const lines = [`${pad}${role}${name10}${value3}${checked}`];
+                for (const child of (n2.children ?? []).slice(0, 30)) {
+                  lines.push(render2(child, depth + 1));
+                }
+                return lines.join("\n");
+              };
+              const out = render2(tree2).slice(0, 6e3);
+              return ok(`Accessibility snapshot:
+${out}`, start2);
+            }
+            case "bounding_box": {
+              if (!selector)
+                return fail("selector is required", start2);
+              const loc = page.locator(selector).first();
+              const box = await loc.boundingBox();
+              if (!box)
+                return fail(`Element ${selector} has no bounding box (not rendered or detached)`, start2);
+              return ok(JSON.stringify(box), start2);
+            }
+            case "query_all": {
+              if (!selector)
+                return fail("selector is required", start2);
+              const handles = await page.locator(selector).all();
+              if (handles.length === 0)
+                return ok(`No elements matched ${selector}.`, start2);
+              const max = Math.min(handles.length, 30);
+              const lines = [];
+              for (let i2 = 0; i2 < max; i2++) {
+                try {
+                  const h = handles[i2];
+                  const txt = (await h.textContent())?.trim().slice(0, 120) ?? "";
+                  const tag = await h.evaluate((el) => (el.tagName || "?").toLowerCase());
+                  const id = await h.evaluate((el) => el.id || "");
+                  const cls = await h.evaluate((el) => (el.className || "").toString().slice(0, 80));
+                  lines.push(`[${i2}] <${tag}${id ? ` id="${id}"` : ""}${cls ? ` class="${cls}"` : ""}> ${txt}`);
+                } catch (e2) {
+                  lines.push(`[${i2}] (error: ${e2 instanceof Error ? e2.message.slice(0, 80) : String(e2).slice(0, 80)})`);
+                }
+              }
+              if (handles.length > max)
+                lines.push(`... +${handles.length - max} more`);
+              return ok(`Matched ${handles.length} element(s) for ${selector}:
+${lines.join("\n")}`, start2);
+            }
+            case "performance": {
+              const data = await page.evaluate(`(() => {
+            const perf = window.performance;
+            const nav = (perf.getEntriesByType('navigation') || [])[0];
+            const paints = (perf.getEntriesByType('paint') || []);
+            return {
+              url: window.location.href,
+              domContentLoaded: nav ? nav.domContentLoadedEventEnd : null,
+              loadComplete: nav ? nav.loadEventEnd : null,
+              firstPaint: (paints.find(p => p.name === 'first-paint') || {}).startTime || null,
+              firstContentfulPaint: (paints.find(p => p.name === 'first-contentful-paint') || {}).startTime || null,
+              transferSize: nav ? nav.transferSize : null,
+              encodedBodySize: nav ? nav.encodedBodySize : null,
+              decodedBodySize: nav ? nav.decodedBodySize : null,
+            };
+          })()`);
+              return ok(`Performance metrics:
+${JSON.stringify(data, null, 2)}`, start2);
+            }
+            case "cookies": {
+              const ck = await context.cookies();
+              if (!ck || ck.length === 0)
+                return ok("No cookies set.", start2);
+              const lines = ck.slice(0, 50).map((c9) => `  ${c9.name}=${String(c9.value).slice(0, 80)}  (domain=${c9.domain}, path=${c9.path}${c9.httpOnly ? ", httpOnly" : ""}${c9.secure ? ", secure" : ""})`);
+              return ok(`Cookies (${ck.length}):
+${lines.join("\n")}`, start2);
+            }
+            case "storage": {
+              const data = await page.evaluate(`(() => {
+            const dump = function(s) {
+              var o = {};
+              for (var i = 0; i < s.length; i++) {
+                var k = s.key(i);
+                if (k) o[k] = String(s.getItem(k) || '').slice(0, 200);
+              }
+              return o;
+            };
+            return {
+              localStorage: dump(window.localStorage),
+              sessionStorage: dump(window.sessionStorage),
+            };
+          })()`);
+              return ok(`Storage:
+${JSON.stringify(data, null, 2)}`, start2);
+            }
+            case "viewport": {
+              const v = (text || value2 || "").trim();
+              const m2 = v.match(/^(\d+)\s*[x×]\s*(\d+)$/);
+              if (m2) {
+                const w = parseInt(m2[1], 10), h = parseInt(m2[2], 10);
+                await page.setViewportSize({ width: w, height: h });
+                return ok(`Viewport set to ${w}x${h}`, start2);
+              }
+              const dev = pw?.devices?.[v];
+              if (dev) {
+                await page.setViewportSize(dev.viewport);
+                return ok(`Viewport set to ${v} (${dev.viewport.width}x${dev.viewport.height})`, start2);
+              }
+              return fail(`viewport: provide "WxH" (e.g. "375x667") or a Playwright device name (e.g. "iPhone 13")`, start2);
+            }
+            case "clear_diagnostics": {
+              const sizes = `console=${consoleBuffer.length} network=${networkBuffer.length} errors=${errorBuffer.length}`;
+              clearDiagnosticBuffers();
+              return ok(`Cleared diagnostic buffers (${sizes}).`, start2);
+            }
             default:
               return fail(`Unknown action: ${action}`, start2);
           }
@@ -518601,41 +518838,98 @@ ${body}`;
         if (!output || typeof output !== "string")
           return null;
         const text = output;
-        const tail = text.slice(-3e3);
-        if (/\bTS\d{3,5}\b.*?: /m.test(tail) && /\berror\s+TS\d{3,5}\b/i.test(tail)) {
+        const tail = text.slice(-4e3);
+        if (/\berror\s+TS\d{3,5}\b/i.test(tail)) {
           const m2 = tail.match(/error\s+TS\d{3,5}[^\n]{0,200}/i);
-          return `TypeScript build error detected: ${(m2?.[0] ?? "").slice(0, 200)}`;
+          return `Compile error: ${(m2?.[0] ?? "").slice(0, 240)}`;
+        }
+        if (/\berror\[E\d+\]/i.test(tail)) {
+          const m2 = tail.match(/error\[E\d+\][^\n]{0,200}/i);
+          return `Compile error: ${(m2?.[0] ?? "").slice(0, 240)}`;
+        }
+        if (/\berror\s+CS\d{3,5}\b/i.test(tail)) {
+          const m2 = tail.match(/error\s+CS\d{3,5}[^\n]{0,200}/i);
+          return `Compile error: ${(m2?.[0] ?? "").slice(0, 240)}`;
+        }
+        if (/(^|\n)\S+:\d+:\d+:\s+error:/i.test(tail)) {
+          const m2 = tail.match(/\S+:\d+:\d+:\s+error:[^\n]{0,200}/i);
+          return `Compile error: ${(m2?.[0] ?? "").slice(0, 240)}`;
         }
-        if (/\b(Failed to compile|Compilation failed|Build failed)\b/i.test(tail)) {
-          const m2 = tail.match(/(Failed to compile|Compilation failed|Build failed)[^\n]{0,300}/i);
-          return `Build failed: ${(m2?.[0] ?? "").slice(0, 200)}`;
+        if (/\bundefined\s+reference\s+to\b|\bld:\s+error:/i.test(tail)) {
+          const m2 = tail.match(/(undefined\s+reference\s+to|ld:\s+error:)[^\n]{0,200}/i);
+          return `Linker error: ${(m2?.[0] ?? "").slice(0, 240)}`;
         }
-        if (/\bError:\s+Build failed\b/i.test(tail)) {
-          return "Next.js build failed (webpack/rspack errors)";
+        if (/\b(Failed to compile|Compilation failed|Build failed|BUILD\s+FAIL(ED|URE))\b/i.test(tail)) {
+          const m2 = tail.match(/(Failed to compile|Compilation failed|Build failed|BUILD\s+FAIL(?:ED|URE))[^\n]{0,300}/i);
+          return `Build failed: ${(m2?.[0] ?? "").slice(0, 240)}`;
         }
-        if (/\b\d+\s+errors? found\b/i.test(tail) && /\b(error|failed)\b/i.test(tail)) {
-          const m2 = tail.match(/\d+\s+errors? found[^\n]{0,200}/i);
-          return `Build reported errors: ${(m2?.[0] ?? "").slice(0, 200)}`;
+        if (/\b\d+\s+errors?\s+(found|generated)\b/i.test(tail)) {
+          const m2 = tail.match(/\d+\s+errors?\s+(?:found|generated)[^\n]{0,200}/i);
+          return `Build reported errors: ${(m2?.[0] ?? "").slice(0, 240)}`;
         }
         if (/\bnpm\s+ERR!|\bnpm error code\b/i.test(tail)) {
           const m2 = tail.match(/(npm\s+ERR!|npm error code)[^\n]{0,300}/i);
-          return `npm error: ${(m2?.[0] ?? "").slice(0, 200)}`;
+          return `Package install error: ${(m2?.[0] ?? "").slice(0, 240)}`;
         }
-        if (/\bTests:\s+\d+\s+failed\b/i.test(tail) || /\b\d+\s+failed\b.*\bof\s+\d+\b/i.test(tail)) {
-          const m2 = tail.match(/Tests:[^\n]{0,200}|\d+\s+failed[^\n]{0,200}/i);
-          return `Test failures detected: ${(m2?.[0] ?? "").slice(0, 200)}`;
+        if (/\bERROR:\s+Could not (install|find|build)\b/i.test(tail)) {
+          const m2 = tail.match(/ERROR:\s+Could not[^\n]{0,200}/i);
+          return `Package install error: ${(m2?.[0] ?? "").slice(0, 240)}`;
         }
-        if (/^Traceback \(most recent call last\):/m.test(tail) && /\bError:|Exception:/m.test(tail)) {
-          const m2 = tail.match(/(\w+(?:Error|Exception)):[^\n]{0,200}/);
-          return `Python exception: ${(m2?.[0] ?? "").slice(0, 200)}`;
+        if (/\b(E:\s+(Unable to|Could not)|apt-get:\s+error|dnf:\s+(?:Error|Failed)|pacman:\s+error)/i.test(tail)) {
+          const m2 = tail.match(/(E:\s+|apt-get:\s+error|dnf:\s+|pacman:\s+error)[^\n]{0,200}/i);
+          return `System package error: ${(m2?.[0] ?? "").slice(0, 240)}`;
         }
-        if (/\berror\[E\d+\]:/i.test(tail) || /\b(\.\/[\w./-]+):\d+:\d+:\s+error:/i.test(tail)) {
-          const m2 = tail.match(/error\[E\d+\][^\n]{0,200}|\.\/[\w./-]+:\d+:\d+:\s+error:[^\n]{0,200}/i);
-          return `Compiler error: ${(m2?.[0] ?? "").slice(0, 200)}`;
+        if (/\bTests?:\s+\d+\s+failed\b/i.test(tail)) {
+          const m2 = tail.match(/Tests?:[^\n]{0,200}/i);
+          return `Test failures: ${(m2?.[0] ?? "").slice(0, 240)}`;
+        }
+        if (/^FAILED\s+\S+::/m.test(tail) || /=+\s*\d+\s+failed/i.test(tail)) {
+          const m2 = tail.match(/(FAILED\s+\S+|=+\s*\d+\s+failed)[^\n]{0,200}/i);
+          return `Test failures: ${(m2?.[0] ?? "").slice(0, 240)}`;
+        }
+        if (/\b\d+\s+failing\b/i.test(tail) && /\bpassing\b/i.test(tail)) {
+          const m2 = tail.match(/\d+\s+failing[^\n]{0,200}/i);
+          return `Test failures: ${(m2?.[0] ?? "").slice(0, 240)}`;
+        }
+        if (/^---\s+FAIL:/m.test(tail) || /^FAIL\t\S+/m.test(tail)) {
+          const m2 = tail.match(/(---\s+FAIL:[^\n]{0,200}|FAIL\t\S+[^\n]{0,200})/i);
+          return `Test failures: ${(m2?.[0] ?? "").slice(0, 240)}`;
+        }
+        if (/test\s+result:\s+FAILED/i.test(tail)) {
+          const m2 = tail.match(/test\s+result:\s+FAILED[^\n]{0,200}/i);
+          return `Test failures: ${(m2?.[0] ?? "").slice(0, 240)}`;
+        }
+        if (/\b\d+\s+failed\s+(of|out\s+of)\s+\d+\b/i.test(tail)) {
+          const m2 = tail.match(/\d+\s+failed\s+(?:of|out\s+of)\s+\d+[^\n]{0,200}/i);
+          return `Test failures: ${(m2?.[0] ?? "").slice(0, 240)}`;
+        }
+        if (/^Traceback \(most recent call last\):/m.test(tail) && /\w+(?:Error|Exception):\s/.test(tail)) {
+          const m2 = tail.match(/\w+(?:Error|Exception):[^\n]{0,200}/);
+          return `Runtime exception: ${(m2?.[0] ?? "").slice(0, 240)}`;
+        }
+        if (/\b(UnhandledPromiseRejection|throw\s+new\s+\w+|Uncaught\s+\w+Error)\b/i.test(tail)) {
+          const m2 = tail.match(/(UnhandledPromiseRejection|throw\s+new\s+\w+|Uncaught[^\n]{0,200})/i);
+          return `Runtime exception: ${(m2?.[0] ?? "").slice(0, 240)}`;
+        }
+        if (/\b(Exception\s+in\s+thread|java\.\w+(?:Error|Exception):)/i.test(tail)) {
+          const m2 = tail.match(/(Exception\s+in\s+thread[^\n]{0,200}|java\.\w+(?:Error|Exception):[^\n]{0,200})/i);
+          return `Runtime exception: ${(m2?.[0] ?? "").slice(0, 240)}`;
+        }
+        if (/^[\w./:-]+:\d+:in\s+`[^']*':\s+/m.test(tail) && /\b\w+Error\b/.test(tail)) {
+          const m2 = tail.match(/\b\w+Error[^\n]{0,200}/);
+          return `Runtime exception: ${(m2?.[0] ?? "").slice(0, 240)}`;
+        }
+        if (/\bthread\s+'[^']*'\s+panicked\s+at/i.test(tail)) {
+          const m2 = tail.match(/thread\s+'[^']*'\s+panicked\s+at[^\n]{0,200}/i);
+          return `Runtime panic: ${(m2?.[0] ?? "").slice(0, 240)}`;
         }
         if (/^(fatal|FATAL|panic):\s/m.test(tail)) {
           const m2 = tail.match(/^(fatal|FATAL|panic):\s[^\n]{0,200}/m);
-          return `Fatal error: ${(m2?.[0] ?? "").slice(0, 200)}`;
+          return `Fatal error: ${(m2?.[0] ?? "").slice(0, 240)}`;
+        }
+        if (/\b(Segmentation\s+fault|core\s+dumped|Abort(?:ed)?\s+\(core dumped\)|Killed)\b/i.test(tail)) {
+          const m2 = tail.match(/(Segmentation\s+fault|core\s+dumped|Abort(?:ed)?\s+\(core dumped\)|Killed)[^\n]{0,80}/i);
+          return `Process terminated abnormally: ${(m2?.[0] ?? "").slice(0, 240)}`;
         }
         return null;
       }
@@ -518692,6 +518986,7 @@ ${body}`;
         if (/(^|[^&\d])(>|>>)\s*\S/.test(cmd))
           return false;
         const MUTATE_BINS = [
+          // POSIX file/process mutators
           "rm",
           "mv",
           "cp",
@@ -518699,56 +518994,158 @@ ${body}`;
           "rmdir",
           "chmod",
           "chown",
+          "chgrp",
           "touch",
           "tee",
           "dd",
           "truncate",
           "ln",
+          "install",
           "kill",
           "pkill",
           "killall",
           "reboot",
           "shutdown",
+          "halt",
+          "poweroff",
           "fakeroot",
           "sudo",
+          "doas",
           "nohup",
           "setsid",
+          "su",
+          // Build orchestrators (always trigger compile/test side-effects)
           "make",
+          "gmake",
+          "ninja",
           "gradle",
+          "gradlew",
           "mvn",
+          "mvnw",
+          "ant",
+          "bazel",
+          "buck",
+          "buck2",
+          "cmake",
+          "meson",
+          "scons",
+          "rake",
+          "leiningen",
+          "lein",
+          "sbt",
+          "stack",
+          // Configuration-management / orchestration
           "ansible",
-          "systemd-run"
+          "ansible-playbook",
+          "puppet",
+          "chef-client",
+          "salt",
+          "salt-call",
+          "systemd-run",
+          "systemd-analyze",
+          // System package managers
+          "apt",
+          "apt-get",
+          "yum",
+          "dnf",
+          "rpm",
+          "pacman",
+          "zypper",
+          "emerge",
+          "snap",
+          "flatpak",
+          "brew",
+          "port",
+          "pkg",
+          "apk",
+          "choco",
+          "scoop",
+          "winget",
+          // Helm / k8s-flavored
+          "helm",
+          "helmfile",
+          "kustomize"
         ];
         const mutateBinsRe = new RegExp(`\\b(${MUTATE_BINS.join("|")})\\b`, "i");
         if (mutateBinsRe.test(cmd))
           return false;
         if (/\bsed\s+(-i|--in-place)\b/.test(cmd))
           return false;
-        if (/\bsystemctl\s+(?!status\b|show\b|is-)/i.test(cmd))
+        if (/\bsystemctl\s+(?!status\b|show\b|is-|cat\b|list-|get-)/i.test(cmd))
           return false;
         if (/\bservice\s+\S+\s+(?!status\b)/i.test(cmd))
           return false;
         if (/\bcrontab\s+-(e|d|r)\b/.test(cmd))
           return false;
-        if (/\bnpm\s+(install|uninstall|update|run|test|exec|publish|init|link|unlink|version|cache\s+clean|ci|audit\s+fix)\b/i.test(cmd))
+        if (/\bnpm\s+(install|i\b|uninstall|un\b|update|up\b|run\s|test\b|exec|publish|init|link|unlink|version|cache\s+clean|ci\b|audit\s+fix|prune|rebuild|adduser|login|logout|deprecate)\b/i.test(cmd))
           return false;
-        if (/\bpnpm\s+(install|update|add|remove|run|test|exec|publish|init|link|unlink|version)\b/i.test(cmd))
+        if (/\bpnpm\s+(install|i\b|update|up\b|add|remove|rm\b|run\s|test\b|exec|publish|init|link|unlink|version|prune|rebuild)\b/i.test(cmd))
           return false;
-        if (/\byarn\s+(install|add|remove|upgrade|run|test|exec|publish|init|link|unlink|version)\b/i.test(cmd))
-          return false;
-        if (/\bpip\s+(install|uninstall|wheel)\b/i.test(cmd))
+        if (/\byarn\s+(install|add|remove|upgrade|run\s|test\b|exec|publish|init|link|unlink|version)\b/i.test(cmd))
           return false;
         if (/\bnpx\b/.test(cmd))
           return false;
-        if (/\bcargo\s+(build|run|test|update|publish|install|uninstall|fmt|fix)\b/i.test(cmd))
+        if (/\bpip3?\s+(install|uninstall|wheel|download)\b/i.test(cmd))
+          return false;
+        if (/\bpipenv\s+(install|uninstall|run|update|sync)\b/i.test(cmd))
+          return false;
+        if (/\bpoetry\s+(add|install|run|update|remove|build|publish|init|new|export|lock)\b/i.test(cmd))
+          return false;
+        if (/\bconda\s+(install|remove|update|create|env\s+(create|update|remove))\b/i.test(cmd))
+          return false;
+        if (/\buv\s+(add|remove|sync|install|run|build|publish|pip\s+install)\b/i.test(cmd))
+          return false;
+        if (/\bbundle\s+(install|update|exec|add|remove|init)\b/i.test(cmd))
+          return false;
+        if (/\bgem\s+(install|uninstall|update|build|push|owner)\b/i.test(cmd))
+          return false;
+        if (/\bcpan\b|\bcpanm\b|\bperl\s+-MCPAN\b/i.test(cmd))
+          return false;
+        if (/\bcomposer\s+(install|update|require|remove|create-project|dump-autoload|run-script)\b/i.test(cmd))
+          return false;
+        if (/\bdotnet\s+(build|run|test|publish|pack|restore|add|remove|new|tool\s+install)\b/i.test(cmd))
+          return false;
+        if (/\bcargo\s+(build|run|test|bench|update|publish|install|uninstall|fmt|fix|clippy\s+--fix|new|init|add|remove|generate-lockfile)\b/i.test(cmd))
+          return false;
+        if (/\bgo\s+(build|run|test|get|install|generate|mod\s+(tidy|download|init|edit|vendor)|work\s+)\b/i.test(cmd))
+          return false;
+        if (/\bdocker\s+(build|run|push|pull|exec|kill|stop|rm|rmi|tag|create|cp|commit|save|load|import|export|network\s+create|volume\s+create|compose\s+(up|down|build|run|exec|restart))\b/i.test(cmd))
+          return false;
+        if (/\bpodman\s+(build|run|push|pull|exec|kill|stop|rm|rmi|tag|create|commit)\b/i.test(cmd))
+          return false;
+        if (/\bkubectl\s+(apply|delete|create|edit|patch|scale|rollout|exec|cp|drain|cordon|uncordon|taint|label|annotate|set\s)\b/i.test(cmd))
+          return false;
+        if (/\bterraform\s+(apply|destroy|init|plan|import|taint|untaint|workspace\s+(new|delete)|state\s+(rm|mv|push|replace-provider))\b/i.test(cmd))
+          return false;
+        if (/\bpulumi\s+(up|destroy|new|stack\s+(rm|init)|config\s+set|policy\s+enable|import)\b/i.test(cmd))
           return false;
-        if (/\bgo\s+(build|run|test|get|install)\b/i.test(cmd))
+        if (/\baws\s+(?:\S+\s+)+(create|delete|put|update|run|start|stop|terminate|attach|detach|cp|sync|mv)\b/i.test(cmd))
           return false;
-        if (/\bdocker\s+(build|run|push|pull|exec|kill|stop|rm|rmi|tag)\b/i.test(cmd))
+        if (/\bgcloud\s+(?:\S+\s+)+(create|delete|update|deploy|apply|set|enable|disable|attach|detach|reset|move|sign-in|sign-out)\b/i.test(cmd))
           return false;
-        if (/\bkubectl\s+(apply|delete|create|edit|patch|scale|rollout|exec)\b/i.test(cmd))
+        if (/\baz\s+(?:\S+\s+)+(create|delete|update|deploy|set|attach|detach|reset|move)\b/i.test(cmd))
           return false;
-        if (/\bterraform\s+(apply|destroy|init|plan|import)\b/i.test(cmd))
+        if (/\beslint\s+[^|;&]*--fix\b/i.test(cmd))
+          return false;
+        if (/\bprettier\s+[^|;&]*--write\b/i.test(cmd))
+          return false;
+        if (/\bblack\s+(?!--check\b)/i.test(cmd))
+          return false;
+        if (/\bisort\s+(?!--check\b|--diff\b)/i.test(cmd))
+          return false;
+        if (/\bruff\s+(?:format\b|check\s+[^|;&]*--fix\b)/i.test(cmd))
+          return false;
+        if (/\bgofmt\s+-w\b/i.test(cmd))
+          return false;
+        if (/\brustfmt\s+(?!--check\b)/i.test(cmd))
+          return false;
+        if (/\b(node|python\d?|ruby|perl|bash|sh|zsh|dash|fish)\s+-(e|c)\b/.test(cmd))
+          return false;
+        if (/\bcurl\s+[^|]*\|\s*(bash|sh|zsh|fish|python\d?|ruby|node|perl)\b/i.test(cmd))
+          return false;
+        if (/\bgit\s+(add|commit|push|pull|fetch|clone|init|checkout|switch|restore|reset|rm|mv|merge|rebase|cherry-pick|revert|stash|apply|am|tag(?:\s+\S)?|notes|worktree\s+(add|remove)|gc|prune|repack|filter-branch|filter-repo|reflog\s+(delete|expire)|update-ref|update-index|symbolic-ref|hash-object\s+-w)\b/i.test(cmd))
+          return false;
+        if (/\bollama\s+(pull|push|run|create|rm|cp|serve)\b/i.test(cmd))
           return false;
         const READ_ONLY_BINS = /* @__PURE__ */ new Set([
           "cd",
@@ -518758,79 +519155,128 @@ ${body}`;
           "fgrep",
           "rg",
           "ag",
+          "ack",
           "cat",
           "head",
           "tail",
           "less",
           "more",
+          "bat",
+          "tac",
           "ls",
           "ll",
           "la",
+          "tree",
           "find",
-          // ALLOWED only if no -delete/-exec mutating action — pre-filtered above
+          "fd",
+          // mutating actions pre-filtered above
           "wc",
           "awk",
           "gawk",
+          "mawk",
+          "nawk",
           "sort",
           "uniq",
+          "shuf",
           "tr",
           "cut",
           "paste",
           "join",
           "comm",
+          "column",
+          "expand",
+          "unexpand",
           "diff",
           "cmp",
+          "patch",
+          // patch with -R or no-args could be mutating; --dry-run only is read
           "echo",
           "printf",
           "pwd",
           "which",
           "type",
           "command",
-          "node",
-          "python",
-          "python3",
-          "ruby",
-          "perl",
+          "whereis",
           "git",
-          // git log/show/diff/status are read; but git add/commit/push/pull are writes — pre-filtered above
+          // pre-filtered above for mutating subcommands
           "ollama",
-          // ollama show/list are read; ollama pull/run/create are writes — pre-filtered above
+          // pre-filtered above
           "cargo",
-          // pre-filtered above for build/run/etc.
           "go",
-          // pre-filtered above for build/run/etc.
+          "rustc",
+          // bin-only forms are read; mutating subcommands pre-filtered
           "stat",
           "file",
           "du",
           "df",
+          "lsof",
+          "fuser",
           "date",
           "uname",
           "id",
           "whoami",
           "hostname",
           "uptime",
+          "tty",
           "env",
           "printenv",
+          "set",
           "test",
           "[",
           "true",
           "false",
+          "yes",
+          "seq",
           "tsc",
           "eslint",
           "prettier",
-          // these emit but mostly read
-          "head",
-          "tail",
+          "ruff",
+          "black",
+          "isort",
+          "rustfmt",
+          "gofmt",
+          // Linters/formatters: mutating modes pre-filtered above (--fix/--write/-w/-i),
+          // so reaching here means we have a check-only invocation.
           "jq",
           "yq",
           "xq",
+          "tomlq",
+          "fx",
           "base64",
           "md5sum",
           "sha256sum",
           "sha1sum",
+          "sha512sum",
           "tldr",
           "man",
-          "info"
+          "info",
+          "help",
+          "pip",
+          "pip3",
+          "pipenv",
+          "poetry",
+          "uv",
+          "conda",
+          // pre-filtered above for mutating subcommands
+          "npm",
+          "pnpm",
+          "yarn",
+          // pre-filtered above
+          "kubectl",
+          "terraform",
+          "pulumi",
+          "aws",
+          "gcloud",
+          "az",
+          // pre-filtered above
+          "docker",
+          "podman",
+          // pre-filtered above
+          "composer",
+          "dotnet",
+          "bundle",
+          "gem"
+          // pre-filtered above
         ]);
         if (/\bfind\b[\s\S]*?(-delete|-exec\s+(rm|mv|cp|chmod|chown|sed\s+-i)|--?ok\s+(rm|mv))/i.test(cmd))
           return false;
@@ -518839,11 +519285,42 @@ ${body}`;
         const segments = cmd.split(/(?:\|\||&&|;)/).map((s2) => s2.trim()).filter(Boolean);
         if (segments.length === 0)
           return false;
+        const SAFE_INTERP_FLAGS = /^(?:--version|--help|-V|-v|-h|--vers|version)$/;
+        const INTERPRETER_BINS = /* @__PURE__ */ new Set([
+          "node",
+          "python",
+          "python2",
+          "python3",
+          "ruby",
+          "perl",
+          "php",
+          "lua",
+          "bash",
+          "sh",
+          "zsh",
+          "dash",
+          "fish",
+          "ksh",
+          "tcsh",
+          "java",
+          "kotlin",
+          "scala",
+          "groovy"
+        ]);
         for (const seg of segments) {
           const stripped = seg.replace(/^cd\s+\S+\s*$/i, "true").replace(/^!/, "");
-          const firstTok = stripped.split(/\s+/)[0]?.replace(/^.*\//, "") || "";
+          const tokens = stripped.split(/\s+/).filter(Boolean);
+          const firstTok = (tokens[0] ?? "").replace(/^.*\//, "");
           if (!firstTok)
             continue;
+          if (INTERPRETER_BINS.has(firstTok)) {
+            const restToks = tokens.slice(1);
+            if (restToks.length === 0)
+              return false;
+            if (!restToks.every((t2) => SAFE_INTERP_FLAGS.test(t2)))
+              return false;
+            continue;
+          }
           if (!READ_ONLY_BINS.has(firstTok))
             return false;
         }
@@ -518852,10 +519329,11 @@ ${body}`;
       /**
        * REG-5: Render the recent-failures block so the agent SEES its own error
        * output before deciding what to do next. Detects same-fingerprint failure
-       * repetition and escalates the warning. Without this, the agent runs
-       * `npx next build`, gets a 200-line TypeScript error, ignores the specific
-       * error and blindly retries with `npm install --force`. Caching the failure
-       * + injecting it pre-LLM forces the model to confront what actually broke.
+       * repetition and escalates the warning. Without this, the agent runs a
+       * build/test/install command, gets a long error stream from the
+       * underlying compiler/runner, ignores the specific error, and blindly
+       * retries with a different flag combination. Caching the failure +
+       * injecting it pre-LLM forces the model to confront what actually broke.
        */
       _renderRecentFailuresBlock(turn) {
         const fails = this._recentFailures;

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "open-agents-ai",
-  "version": "0.187.468",
+  "version": "0.187.470",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "open-agents-ai",
-      "version": "0.187.468",
+      "version": "0.187.470",
       "hasInstallScript": true,
       "license": "CC-BY-NC-4.0",
       "dependencies": {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "open-agents-ai",
-  "version": "0.187.468",
+  "version": "0.187.470",
   "description": "AI coding agent powered by open-source models (Ollama/vLLM) — interactive TUI with agentic tool-calling loop",
   "type": "module",
   "main": "./dist/index.js",