tewip-ai 0.69.5 → 0.69.6

package/dist/engine/rules.js

@@ -157,6 +157,19 @@ const envServerError = {
  * Declines when the server failed during the test: the wrong value may come from the
  * backend (ENV) or from a test that provokes errors on purpose, which only the LLM layer,
  * seeing the whole trace, can tell apart (rules-3).
+ *
+ * **Do not remove the history requirement.** Phase 64 found that `tewip bugs` files almost nothing
+ * from one CI run - 1 ticket out of 1,964 real failures - because every test in a single run has
+ * `runs < 2` and this rule declines. The proposed fix was to let a plain value mismatch stand on its
+ * own: "expected 4, received 3 needs no model" is an obvious-sounding argument, and it was measured
+ * in Phase 65 over **2,767 hand-labelled real failures from nine projects**:
+ *
+ *     293 value mismatches -> TIMING 265 (90.4%), REAL_BUG 26 (8.9%), ENV 2 (0.7%)
+ *
+ * Per project it is 0%, 0%, 0%, 4%, 5%, 6%, 7%, 11%, 39%. **A flaky test fails its assertion too -
+ * that is what failing looks like.** The mismatch is not the signal; the history around it is, which
+ * is what the streak conditions below are reading and what one CI run cannot supply. Re-derive with
+ * `node eval/mismatch-precision.ts` before changing any of this.
  */
 const realBugAssertion = {
     id: 'real-bug-assertion',

package/dist/engine/work/bug-command.js

@@ -60,8 +60,13 @@ export function whyNothingWasFiled(items, unsure) {
     return [
         `${items.length} failure(s). ${unsure.length} of them Tewip reads as a real bug, and none of those reaches the ${CONFIDENT_ENOUGH.toFixed(2)} confidence it requires before opening a ticket in somebody's name - the highest is ${highest.toFixed(2)}.`,
         byModel > 0
-            ? `${byModel === unsure.length ? 'All' : `${byModel}`} of them came from the AI layer, which is capped at 0.45 on purpose: a model's unsupported word is not enough to file on. A person's correction on the failure in CI, or a rule that agrees, is what would make it enough.`
+            ? `${byModel === unsure.length ? 'All' : `${byModel}`} of them came from the AI layer, which is capped at 0.45 on purpose: a model's unsupported word is not enough to file on.`
             : '',
+        // What a CI user can actually do about it. The rules decline a value mismatch without a stable
+        // failing streak, and on one run there is no streak to read - which is not over-caution:
+        // measured over 2,767 labelled real failures, a value mismatch is TIMING 90% of the time and a
+        // real bug 9% of the time (`node eval/mismatch-precision.ts`). History is what separates them.
+        'What would change this: history. The rules will not call a wrong value a bug without a stable failing streak behind it, because a flaky test produces the same wrong value - on real labelled data a value mismatch is a flaky test nine times out of ten. Keep this project\'s runs (`--name <project>`, or let the Action keep them) and the same failure in a second and third run is what the rules read. A person\'s correction on the failure in CI settles it immediately.',
         `Nothing was filed. These are listed in \`tewip explain\` with their evidence${kinds.length > 0 ? `, alongside ${kinds.join(', ')} failures that are deliberately never filed` : ''}.`,
     ]
         .filter(Boolean)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "tewip-ai",
-  "version": "0.69.5",
+  "version": "0.69.6",
   "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": [