ccqa 1.37.0 → 1.37.1

package/dist/bin/ccqa.mjs

@@ -3726,9 +3726,19 @@ function scrubOutcome(outcome, scrubMap) {
 		}
 	};
 }
+/**
+* Pause before the single retry of an errored classification call. Long enough
+* to ride out a transient network/model hiccup, short enough not to stall the
+* report when the error is persistent.
+*/
+const RETRY_DELAY_MS = 2e3;
+function sleep(ms) {
+	return new Promise((resolve) => setTimeout(resolve, ms));
+}
 async function classifyFailure(input, options) {
-	const { result: raw, isError } = await invokeClaudeStreaming({
-		prompt: buildFailureAnalysisPrompt(input),
+	const prompt = buildFailureAnalysisPrompt(input);
+	const invoke = () => invokeClaudeStreaming({
+		prompt,
 		allowedTools: [
 			"Read",
 			"Grep",
@@ -3741,11 +3751,22 @@ async function classifyFailure(input, options) {
 		...options.model ? { model: options.model } : {},
 		...options.cwd ? { cwd: options.cwd } : {}
 	}, () => {});
-	if (isError || !raw) return {
-		analysis: unknownAnalysis(isError ? "Claude returned an error result" : "Claude returned no output"),
-		raw: raw ?? "",
-		sdkError: isError
-	};
+	let { result: raw, isError } = await invoke();
+	let retried = false;
+	if (isError) {
+		warn("failure analysis: Claude invocation errored — retrying once");
+		await sleep(RETRY_DELAY_MS);
+		({result: raw, isError} = await invoke());
+		retried = true;
+	}
+	if (isError || !raw) {
+		const cause = isError ? "Claude returned an error result" : "Claude returned no output";
+		return {
+			analysis: unknownAnalysis(retried ? `${cause} (after 1 retry)` : cause),
+			raw: raw ?? "",
+			sdkError: isError
+		};
+	}
 	let sawParseableJson = false;
 	for (const candidate of extractJsonCandidates(raw)) {
 		let parsed;
@@ -8147,8 +8168,9 @@ const C$1 = {
 * diff — so report rows and CI logs look the same whichever target a
 * project's specs use.
 *
-* One Claude call per failing spec, which reads the source itself rather than
-* deferring to a drift audit run beforehand. It is still one *phase*:
+* One Claude call per failing spec (two when the first errors and is retried
+* once), which reads the source itself rather than deferring to a drift audit
+* run beforehand. It is still one *phase*:
 * `beginFailureAnalysis` hands back the state every path shares, so a mixed
 * run prints one `failure analysis` banner in one place rather than one per
 * execution path. It runs after every spec has executed, so no Claude turn is
@@ -9065,8 +9087,10 @@ z.object({ error: z.object({
 * One spec's record of a single run, as stored in a ledger bucket. Identical
 * to `LastGreenEntry` plus the commit the environment was running at the time
 * — without it a bucket entry can be ordered in wall-clock time but not
-* *positioned* against the deploy log, which is the only ordering re-run
-* selection may use (ADR-0010).
+* *positioned* against the deploy log, which is the only ordering the
+* staleness verdict may use (ADR-0010). Wall-clock order is fit only for
+* scheduling within an already-decided set, where a mis-ranking delays a
+* spec rather than excusing it.
 */
 const SpecLedgerEntrySchema = z.object({
 	gitHead: z.string(),
@@ -9651,7 +9675,10 @@ function selectSpecsNeedingRerun(specs, report) {
 		const entry = report.specs[specKey(spec)];
 		const verdict = entry?.verdict ?? "inProgress";
 		counts.set(verdict, (counts.get(verdict) ?? 0) + 1);
-		if (verdict === "rerunNeeded") selected.push(spec);
+		if (verdict === "rerunNeeded") selected.push({
+			spec,
+			lastRunAt: entry?.lastRun?.at ?? ""
+		});
 		else if (verdict === "inProgress") {
 			excludedInProgress++;
 			if (!entry) excludedUnknownToHub++;
@@ -9661,8 +9688,9 @@ function selectSpecsNeedingRerun(specs, report) {
 			}
 		}
 	}
+	selected.sort((a, b) => a.lastRunAt.localeCompare(b.lastRunAt));
 	return {
-		selected,
+		selected: selected.map((s) => s.spec),
 		summary: formatCounts(SUMMARY_ORDER$1, counts),
 		excludedInProgress,
 		excludedUnknownToHub,
@@ -11583,6 +11611,15 @@ ${stepsText}
 - Do not invent success when blocked: fail honestly with a short reason.
 - **Evidence discipline**: when the assertion target is a specific row / message / banner / URL, scroll it into view (or focus the relevant pane) before letting the step end. The "after" screenshot is captured for you automatically — your job is to make sure that screenshot shows the thing your STEP_RESULT line is talking about.
 
+### Waiting for asynchronous responses
+
+Some expected outcomes arrive asynchronously — an automated reply, a background job finishing, a list refreshing. Waiting for them is fine, but the wait has a budget:
+
+- Prefer bounded probes (\`agent-browser wait --text "..."\`, or a short pause followed by a fresh \`snapshot\`) over long blind sleeps, and keep a rough running total of how long you have waited within this step.
+- **The total wait within one step must not exceed 3 minutes**, unless the step's own instruction explicitly names a longer wait. Do not keep adding "one more" sleep past the budget.
+- When the budget is spent and the expected outcome has still not appeared, STOP waiting and emit \`STEP_RESULT|<stepId>|fail|...\`. **Never end your turn without a STEP_RESULT because you were still waiting** — a silent timeout is recorded as a protocol failure and hides the real cause from failure analysis.
+- The fail reason must state what you waited for, roughly how long in total, and what you observed instead (e.g. "waited ~3 min for a reply to appear after submitting; none appeared, the view still shows only the submitted item").
+
 ### Output contract (STRICT)
 
 Your final assistant message MUST contain exactly one line of the form:

package/dist/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ccqa",
-  "version": "1.37.0",
+  "version": "1.37.1",
   "type": "module",
   "description": "Browser test recorder powered by Claude Code and agent-browser",
   "repository": {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ccqa",
-  "version": "1.37.0",
+  "version": "1.37.1",
   "type": "module",
   "description": "Browser test recorder powered by Claude Code and agent-browser",
   "repository": {