tewip-ai 0.64.0 → 0.65.0

package/dist/engine/llm/llm-diagnose.js

@@ -50,6 +50,28 @@ export const LLM_CONFIDENCE_CEILING = 0.45;
 export function contradictsHistory(category, context) {
     if (category !== 'REAL_BUG' && category !== 'LOCATOR_DRIFT')
         return null;
+    // A timeout is not evidence of a bug, and that is measured rather than argued.
+    //
+    // Phase 56 took 471 real labelled CI failures from WordPress/gutenberg and asked, for each signal
+    // a classifier can actually see when the test fails, how far it moves the label from its base
+    // rate. Three signals move it a long way:
+    //
+    //   error says Timeout/exceeded   ->  TIMING 91%   (37% when absent)
+    //   error is an assertion         ->  TIMING 93%   (ENV 86% when absent)
+    //   error mentions a locator      ->  TIMING 94%
+    //
+    // and REAL_BUG sits at 3-6% in **every single bucket** - its base rate, undisturbed. Nothing in
+    // the evidence available at failure time distinguishes a real bug from flakiness. That is the
+    // whole explanation for a model layer scoring 17.8%: it reaches for the one category the evidence
+    // cannot support, and is wrong about 94% of the time when it does.
+    //
+    // So a verdict of REAL_BUG resting on nothing but a timeout is refused. Six per cent of those
+    // really are bugs, and they are not lost: the case goes to a person with the model's reasoning
+    // attached, which is what this project does with every question it cannot answer.
+    const message = context.error?.message ?? '';
+    if (category === 'REAL_BUG' && /timeout|exceeded/i.test(message)) {
+        return 'refused: the only evidence is a timeout, and across 471 labelled failures a timeout was TIMING 91% of the time and REAL_BUG 6%';
+    }
     if (context.thisRun?.anotherAttemptPassed) {
         return `refused: another attempt of the same code passed in this run, so the failure is not deterministic and cannot be ${category}`;
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "tewip-ai",
-  "version": "0.64.0",
+  "version": "0.65.0",
   "mcpName": "io.github.hmamut39/tewip",
   "description": "Tewip finds why each failing test failed - real bug, flaky test, environment, test data or changed selector - for any stack, from the command line, CI, or your editor.",
   "keywords": [