tewip-ai 0.70.6 → 0.70.8

package/dist/engine/quality/write-command.js

@@ -62,19 +62,52 @@ ${noModelHelp()}`);
         const fileName = framework.fileFor(scenario.id);
         const file = join(framework.testDir, fileName);
         if (existsSync(file) || (cmd.moveTo && existsSync(join(cmd.moveTo, fileName)))) {
-            results.push({ id: scenario.id, title: scenario.title, file, outcome: { kind: 'rejected', reason: `${fileName} already exists; left untouched`, repairs: 0, passedReliably: false } });
+            // Every file this command writes carries "tewip" in its name, so one already sitting here is
+            // one of its own - an unverified draft from a run that did not finish, not a test the team
+            // wrote. Calling that "left untouched" sounds like a decision and hides that the file is in
+            // the test folder, where the suite is running it.
+            //
+            // Matched case-insensitively on the whole name rather than as a prefix, because only the
+            // JavaScript frameworks produce `tewip-ls-002.spec.ts`: pytest gives `test_tewip_ls_002.py`
+            // and JUnit `TewipLs002Test.java`, and a prefix check would have handed Python and Java teams
+            // the misleading message this fix exists to remove.
+            const mine = /tewip/i.test(fileName);
+            results.push({
+                id: scenario.id,
+                title: scenario.title,
+                file,
+                outcome: {
+                    kind: 'rejected',
+                    reason: mine
+                        ? `${fileName} is already here from a run that did not finish, and was never verified. It is in your test folder, so your suite is running it: delete it, or move it to .tewip/drafts/, then ask again.`
+                        : `${fileName} already exists and was written by somebody else; left untouched`,
+                    repairs: 0,
+                    passedReliably: false,
+                },
+            });
             continue;
         }
         console.error(`  ${scenario.id} ${scenario.title}`);
-        // Stopped part-way (Ctrl+C), a draft that was never verified must not stay in the project.
+        // A draft that was never verified must not stay in the project, however the run ended.
+        //
+        // SIGINT and SIGTERM were covered; nothing else was. A run of this interrupted another way -
+        // a closed pipe, an uncaught throw, a kill - left the draft in the **test folder**, where the
+        // project's own suite then runs it. Seen for real: a two-scenario run was cut short by a broken
+        // shell command, the suite went from green to red, and the next `write` reported the leftover as
+        // "already exists; left untouched", which reads like a decision rather than a casualty.
+        //
+        // `renameSync` is synchronous, so an `exit` handler can still do it: that covers every way out.
+        const rescue = () => {
+            if (!existsSync(file))
+                return;
+            mkdirSync(drafts, { recursive: true });
+            renameSync(file, join(drafts, fileName));
+        };
         const leaveNothing = () => {
-            if (existsSync(file)) {
-                mkdirSync(drafts, { recursive: true });
-                renameSync(file, join(drafts, fileName));
-            }
+            rescue();
             process.exit(130);
         };
-        process.once('SIGINT', leaveNothing).once('SIGTERM', leaveNothing);
+        process.once('SIGINT', leaveNothing).once('SIGTERM', leaveNothing).once('exit', rescue);
         let outcome;
         try {
             ({ outcome } = await writeAndVerify(model, { framework, fileName, pages, feature: saved.feature, code, houseRules: guidanceFor(loadStandards(cmd.root), fileName) }, scenario, file, { runs: cmd.runs, repairs: cmd.repairs, mutateSource: cmd.mutateSource, log: (l) => console.error(l) }));
@@ -83,7 +116,7 @@ ${noModelHelp()}`);
             outcome = { kind: 'rejected', reason: `stopped: ${error.message}`, repairs: 0, passedReliably: false };
         }
         finally {
-            process.off('SIGINT', leaveNothing).off('SIGTERM', leaveNothing);
+            process.off('SIGINT', leaveNothing).off('SIGTERM', leaveNothing).off('exit', rescue);
         }
         // Some runners take minutes per run (Angular rebuilds the app for every `ng test`): say how long
         // the rest will take, so a slow run is not mistaken for a stuck one.

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": "0.70.6",
+  "version": "0.70.8",
   "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": [