canary-test-cli 6.7.0 → 6.7.1

package/dist/reporters/testtracker.d.ts

@@ -73,7 +73,24 @@ export declare function runStatus(totals: {
     failed: number;
     flaky: number;
 }, fullResultStatus?: string): "passed" | "failed" | "flaky" | "cancelled";
-export declare function buildPayload(results: ResultEntry[], cfg: ResolvedConfig, timing: {
+/**
+ * Collapse results that share a `full_title` into one row.
+ *
+ * TestTracker's ingest holds a unique index on `(run_id, full_title)` and
+ * rejects the WHOLE run on a collision, so a duplicate title is not a cosmetic
+ * problem — it takes the suite dark. The reporter keys its in-flight map by
+ * `test.id`, which is genuinely unique, but a title is not: a Playwright
+ * `dependencies:` setup project runs in full in EVERY shard (dependencies are
+ * not sharded), so a `merge-reports` payload over a sharded matrix legitimately
+ * carries the same setup title once per shard. Capwell's `capwell-web` suite
+ * had every nightly rejected this way and never ingested a single run.
+ *
+ * Collapse rule keeps the merged row honest: worst status wins, the first real
+ * error is preserved (so the SDET still sees why it failed), and duration and
+ * retries take the max rather than the last-seen value.
+ */
+export declare function dedupeByFullTitle(results: ResultEntry[]): ResultEntry[];
+export declare function buildPayload(rawResults: ResultEntry[], cfg: ResolvedConfig, timing: {
     startedAt: string;
     finishedAt: string;
 }, env?: NodeJS.ProcessEnv, fullResultStatus?: string): IngestPayload;

package/dist/reporters/testtracker.js

@@ -8,6 +8,7 @@ exports.resolveTestStatus = resolveTestStatus;
 exports.resolveConfig = resolveConfig;
 exports.shouldPush = shouldPush;
 exports.runStatus = runStatus;
+exports.dedupeByFullTitle = dedupeByFullTitle;
 exports.buildPayload = buildPayload;
 const node_crypto_1 = __importDefault(require("node:crypto"));
 // Optional .env load — MUST NOT crash the suite if dotenv is absent.
@@ -92,7 +93,64 @@ function runStatus(totals, fullResultStatus) {
         return "flaky";
     return "passed";
 }
-function buildPayload(results, cfg, timing, env = process.env, fullResultStatus) {
+/** Worst-first, so a collapsed row can never look healthier than its parts. */
+const STATUS_SEVERITY = {
+    failed: 3,
+    flaky: 2,
+    passed: 1,
+    skipped: 0,
+};
+/**
+ * Collapse results that share a `full_title` into one row.
+ *
+ * TestTracker's ingest holds a unique index on `(run_id, full_title)` and
+ * rejects the WHOLE run on a collision, so a duplicate title is not a cosmetic
+ * problem — it takes the suite dark. The reporter keys its in-flight map by
+ * `test.id`, which is genuinely unique, but a title is not: a Playwright
+ * `dependencies:` setup project runs in full in EVERY shard (dependencies are
+ * not sharded), so a `merge-reports` payload over a sharded matrix legitimately
+ * carries the same setup title once per shard. Capwell's `capwell-web` suite
+ * had every nightly rejected this way and never ingested a single run.
+ *
+ * Collapse rule keeps the merged row honest: worst status wins, the first real
+ * error is preserved (so the SDET still sees why it failed), and duration and
+ * retries take the max rather than the last-seen value.
+ */
+function dedupeByFullTitle(results) {
+    const merged = new Map();
+    for (const r of results) {
+        const prior = merged.get(r.full_title);
+        merged.set(r.full_title, prior ? mergeEntries(prior, r) : { ...r, tags: [...r.tags] });
+    }
+    return [...merged.values()];
+}
+function worstOf(a, b) {
+    return (STATUS_SEVERITY[b] ?? 0) > (STATUS_SEVERITY[a] ?? 0) ? b : a;
+}
+/** `Math.max` treats a missing duration as 0; absent-on-both must stay absent. */
+function longerOf(a, b) {
+    if (a === undefined)
+        return b;
+    if (b === undefined)
+        return a;
+    return Math.max(a, b);
+}
+function mergeEntries(prior, next) {
+    return {
+        ...prior,
+        status: worstOf(prior.status, next.status),
+        error_message: prior.error_message ?? next.error_message,
+        error_stack: prior.error_stack ?? next.error_stack,
+        duration_ms: longerOf(prior.duration_ms, next.duration_ms),
+        retries: Math.max(prior.retries, next.retries),
+        tags: [...new Set([...prior.tags, ...next.tags])],
+    };
+}
+function buildPayload(rawResults, cfg, timing, env = process.env, fullResultStatus) {
+    // Deduped before the totals are counted, so `totals` always describes the
+    // rows actually sent — a payload whose totals disagree with `results.length`
+    // would misreport the run on the dashboard.
+    const results = dedupeByFullTitle(rawResults);
     const count = (s) => results.filter((r) => r.status === s).length;
     const passed = count("passed");
     const failed = count("failed");

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "canary-test-cli",
-  "version": "6.7.0",
+  "version": "6.7.1",
   "description": "Canary — AI-powered test automation agent",
   "license": "MIT",
   "repository": {