tewip-ai 0.70.7 → 0.70.9

package/dist/engine/pursue/pursue-command.js

@@ -11,6 +11,7 @@ import { datasetFiles, readJsonl, readTriageRecords } from '../../reporters/stor
 import { alreadyTried, cannotAct, planFor } from './act.js';
 import { decide } from './policy.js';
 import { CANNOT_FAIL } from '../vouch/vouch.js';
+import { applyCorrection, correctionFor, loadCorrections } from '../corrections.js';
 import { vouchAll } from '../vouch/vouch-command.js';
 const words = (text) => (text.toLowerCase().match(/[a-z][a-z0-9]{2,}/g) ?? []).filter((w) => !['the', 'and', 'for', 'this', 'that', 'with', 'keep', 'green', 'sprint', 'flow', 'test', 'tests'].includes(w));
 /** Does this test or scenario belong to the goal? Deterministic, so a person can predict it. */
@@ -25,6 +26,7 @@ export function observe(root, goal, only) {
     const dir = join(root, '.tewip', 'history');
     const runs = [];
     const records = [];
+    const corrections = [];
     if (existsSync(dir)) {
         for (const project of only ? [only] : readdirSync(dir)) {
             const where = join(dir, project);
@@ -32,16 +34,34 @@ export function observe(root, goal, only) {
                 continue;
             runs.push(...readJsonl(datasetFiles(where).runs));
             records.push(...readTriageRecords(datasetFiles(where).triage));
+            corrections.push(...loadCorrections(where));
         }
     }
     runs.sort((a, b) => (a.finishedAt > b.finishedAt ? 1 : -1));
     const newest = runs.at(-1);
     const wanted = words(goal);
+    // A person's correction decides the category, as it does everywhere else.
+    //
+    // This read `r.diagnosis` straight off the record, so a correction never reached the decision.
+    // Field-tested: a person marked a failure REAL_BUG, and `pursue` went on calling it LOCATOR_DRIFT
+    // and concluding "the goal is failing and Tewip cannot say why". That matters more here than in a
+    // report, because this is the command that **acts** - and its own documented order is "a real bug
+    // first... it will not touch a real bug", a safety rule that depends on recognising one. The
+    // strongest possible signal that something is a real bug is a person saying so.
+    //
+    // Phase 74 fixed this inside `buildItems` for every command that goes through it. This one does
+    // not, so it applies them itself.
     const failing = records
         .filter((r) => r.runId === newest?.runId && about(wanted, r.testId))
         .map((r) => {
-        const d = r.diagnosis;
-        return { testId: r.testId, category: String(d?.category ?? 'UNKNOWN'), confidence: Number(d?.confidence ?? 0) };
+        const correction = corrections.length > 0 ? correctionFor(r, corrections) : null;
+        const own = r.diagnosis;
+        const d = correction && own ? applyCorrection(own, correction) : own;
+        return {
+            testId: r.testId,
+            category: String((correction?.category ?? d?.category) ?? 'UNKNOWN'),
+            confidence: Number(d?.confidence ?? (correction ? 1 : 0)),
+        };
     });
     // A test that failed in more than one earlier run and is not failing now has come and gone.
     const failedBefore = new Map();

package/dist/engine/spec/spec.js

@@ -1,14 +1,14 @@
 export const SPEC_PROMPT_VERSION = 'spec-1';
-const SYSTEM = `You are Tewip's specification writer, a senior product engineer who works from the code.
-
-Write an epic, its user stories, and acceptance criteria for the feature described. Use only the facts given: the feature description, the change, the pages of the running app, the tests that already exist, the source files read, and this project's failure history.
-
-Rules:
-- Every statement about what the system ALREADY does must be backed by a quote: evidence is a file path from the facts, then ": ", then a piece of that file copied character for character (a line of code, a test name, a comment). If you cannot quote it, do not claim it - put the uncertainty in open_questions.
-- Acceptance criteria are Given / When / Then, each one checkable by a person or a test. No criterion may say "works correctly", "is fast" or "is user-friendly": say the observable thing.
-- A story is one user-visible outcome, written as "As a <role>, I want <outcome> so that <benefit>". Split anything larger.
-- Prefer what the failure history and the existing tests show is risky or already broken.
-- Never write estimates, story points, dates, budgets, market claims or business cases. You cannot check those, so they do not belong here.
+const SYSTEM = `You are Tewip's specification writer, a senior product engineer who works from the code.
+
+Write an epic, its user stories, and acceptance criteria for the feature described. Use only the facts given: the feature description, the change, the pages of the running app, the tests that already exist, the source files read, and this project's failure history.
+
+Rules:
+- Every statement about what the system ALREADY does must be backed by a quote: evidence is a file path from the facts, then ": ", then a piece of that file copied character for character (a line of code, a test name, a comment). If you cannot quote it, do not claim it - put the uncertainty in open_questions.
+- Acceptance criteria are Given / When / Then, each one checkable by a person or a test. No criterion may say "works correctly", "is fast" or "is user-friendly": say the observable thing.
+- A story is one user-visible outcome, written as "As a <role>, I want <outcome> so that <benefit>". Split anything larger.
+- Prefer what the failure history and the existing tests show is risky or already broken.
+- Never write estimates, story points, dates, budgets, market claims or business cases. You cannot check those, so they do not belong here.
 - out_of_scope is what this epic deliberately does not do. open_questions are the decisions a person must make.`;
 const CRITERION = {
     type: 'object',
@@ -64,9 +64,29 @@ function readable(facts) {
     return [...(facts.testSources ?? []), ...(facts.codeSources ?? [])].map((s) => ({ file: s.file, content: s.content }));
 }
 /**
- * A quote is real when the file is one Tewip read and the distinctive parts of the quote are in it.
- * Whitespace and quote style differ between a model's copy and the file, so the comparison is on
- * the words and symbols, not on the exact spacing.
+ * A quote is real when the file is one Tewip read and the quote appears in it **as one piece**.
+ *
+ * This used to collect the quote's "distinctive parts" - runs of `[A-Za-z0-9_$.]{3,}` - and require
+ * each to appear somewhere in the file, in any order. Three fabrications got through a real check
+ * against a real file, and the worst was this, offered as evidence that a 20% discount cap is
+ * enforced:
+ *
+ *     const rate = Math.min(percent, 20) / 100;     // the file says 100, not 20
+ *
+ * Every part of that - `const`, `rate`, `Math.min`, `percent`, `100` - is in the file. **`20` is
+ * not a part at all**, because `{3,}` cannot see a two-character number. So the one token that
+ * carried the entire meaning was the one token not checked, and a reader of "What the code already
+ * does - each line is quoted from a file Tewip read" would have concluded the bug was fixed.
+ *
+ * The comment that used to sit here saw the risk and named it exactly - *"a quote that is mostly
+ * real but wrong where it matters... 'slice(0, 50)' and 'slice(0, 10000)' differ by one word"* -
+ * and `50` was never a word it could compare. A line assembled from real words in an order the file
+ * has never contained got through for the same reason.
+ *
+ * So the quote has to be in the file contiguously. Whitespace is squashed and quote characters are
+ * normalised, because those are the two differences between a model's copy and the file that carry
+ * no meaning; nothing else is forgiven. A claim that fails this becomes an open question, which is
+ * the right place for it.
  */
 export function quoteHolds(evidence, sources) {
     const at = evidence.indexOf(': ');
@@ -77,13 +97,8 @@ export function quoteHolds(evidence, sources) {
     const source = sources.find((s) => s.file === file || s.file.endsWith(`/${file}`) || file.endsWith(s.file));
     if (!source || quote.length < 3)
         return false;
-    const parts = quote.match(/[A-Za-z0-9_$.]{3,}/g) ?? [];
-    if (parts.length === 0)
-        return false;
-    const body = source.content;
-    // Every distinctive part must be in the file. Anything less lets a quote that is mostly real but
-    // wrong where it matters through - "slice(0, 50)" and "slice(0, 10000)" differ by one word.
-    return parts.every((part) => body.includes(part));
+    const squash = (text) => text.replace(/["'`]/g, '"').replace(/\s+/g, ' ').trim();
+    return squash(source.content).includes(squash(quote));
 }
 /**
  * Checks every claim about what already exists, and demotes the ones Tewip cannot quote into open

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "tewip-ai",
-  "version": "0.70.7",
+  "version": "0.70.9",
   "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": [