ccqa 1.4.0 → 1.5.0

package/dist/bin/ccqa.mjs

@@ -3161,12 +3161,45 @@ function hashTriageUserPrompt(text) {
 */
 const CHANGED_FILE_DIFF_TOOL = "mcp__diff__changed_file_diff";
 function buildFailureAnalysisPrompt(input) {
-	const { script, specYaml, failureLog, liveTranscriptExcerpt, diffPatch, changedFiles, baseRef, driftIssues, outputLanguage = "auto", triageUserPrompt, customPrompt } = input;
+	const { script, specYaml, failureLog, liveTranscriptExcerpt, diffPatch, changedFiles, baseRef, baseSource = null, range = null, driftIssues, outputLanguage = "auto", triageUserPrompt, customPrompt } = input;
+	const lastGreen = baseSource === "last-green";
 	const triageUserPromptBlock = buildTriageUserPromptBlock(triageUserPrompt);
 	const customPromptBlock = buildCustomPromptBlock(customPrompt);
-	return `You are analyzing a failing E2E regression test right after a source change landed. Your job is a root-cause CALL, not a fix: decide which of three categories explains the failure, using the source diff as your primary context.
+	const languageBlock = outputLanguageBlock(outputLanguage, "`reasoning`, `detail`", "label names (TEST_DRIFT, etc.)");
+	const executionBlock = buildExecutionEvidenceBlock(script, failureLog, liveTranscriptExcerpt);
+	const baseLabel = lastGreen ? `this spec's last passing commit${baseRef && baseRef !== "last-green" ? ` (${baseRef})` : ""}` : baseRef ?? "base";
+	const rangeNote = range ? ` — spans ${range.commitCount} commit${range.commitCount === 1 ? "" : "s"} over ${range.days} day${range.days === 1 ? "" : "s"}` : "";
+	let diffBlock;
+	if (diffPatch === null) diffBlock = `## Source changes
 
-${outputLanguageBlock(outputLanguage, "`reasoning`, `detail`", "label names (TEST_DRIFT, etc.)")}## The three categories
+No diff context is available (the base ref could not be resolved, or there are no changes). Classify from the failure log, the spec, and what you can read in the repository — and be correspondingly more conservative: prefer UNKNOWN over a confident SPEC_CHANGE/PRODUCT_BUG call without diff evidence.
+`;
+	else if (diffPatch.length === 0) diffBlock = `## Source changes since ${baseLabel}${rangeNote}
+
+### Changed files (name-status)
+${changedFiles && changedFiles.length > 0 ? changedFiles : "(no changes in range)"}
+
+No changed file matches this spec's relatedPaths, so no hunks are inlined. "No related change" is a real signal — but before concluding, scan the name-status list for anything that could plausibly reach this spec and fetch its hunk with \`${CHANGED_FILE_DIFF_TOOL}\`.
+`;
+	else diffBlock = `## Source changes since ${baseLabel}${rangeNote} (git diff, scoped to this spec's relatedPaths, may be truncated)
+
+### Changed files (name-status)
+${changedFiles ?? "(unavailable)"}
+
+### Patch
+\`\`\`diff
+${diffPatch}
+\`\`\`
+`;
+	const driftBlock = driftIssues && driftIssues.length > 0 ? `## Spec↔code drift audit findings
+
+A separate read-only audit compared the spec against the current source. Treat these as hints, not verdicts:
+
+${driftIssues.map((i) => `- [${i.severity}] (${DRAFT_CATEGORY_LABEL[i.category]}${i.stepId ? `, step ${i.stepId}` : ""}) ${i.message}${i.detail ? ` — ${i.detail}` : ""}`).join("\n")}
+` : "";
+	return `You are analyzing a failing E2E regression test against the source changes since a known-good baseline. Your job is a root-cause CALL, not a fix: decide which of three categories explains the failure, using the source diff as your primary context.
+
+${languageBlock}## The three categories
 
 The question that separates them: **is the behavior the spec describes still what the product intends?**
 
@@ -3189,10 +3222,13 @@ You have **up to 12 tool turns**. Do NOT write, edit, run shell commands, or hit
 
 ## Decision guidance
 
+${lastGreen ? `The baseline is the commit where THIS spec last passed, so the range strictly covers the window in which it broke: the cause is either inside these changes or outside the code entirely (flaky timing, environment, an external service, test data). The range may mix several unrelated merges — most of the diff is noise; what matters is the specific change you can tie to the failing step.` : `The baseline is a fixed ref (typically the PR base): the spec is NOT guaranteed to have passed there, so the range is not guaranteed to contain the cause.`}
+
 - Diff touches only attributes/identifiers the test selects on (labels, testids, class names, timing) while the user-visible flow is intact → TEST_DRIFT.
 - Diff intentionally removes/reworks the UI or flow that a spec step verifies (component deleted, page restructured, copy redefined, feature flag flipped) → SPEC_CHANGE.
 - Diff UNINTENTIONALLY breaks behavior the spec still intends — e.g. a refactor that drops a side effect, an inverted condition, a regression hiding inside a cleanup commit — → PRODUCT_BUG, citing the diff hunk as evidence. A product bug is often introduced BY the diff; what separates it from SPEC_CHANGE is intent: does the change read as a deliberate redesign of what the spec verifies, or as collateral damage?
-- Diff is unrelated to the failing step (or there is no relevant diff) and the test was passing before → lean PRODUCT_BUG; first rule out timing/data flakiness and infrastructure errors (daemon not running, network down, missing credentials) — those read as UNKNOWN with low confidence, not PRODUCT_BUG.
+${lastGreen ? `- No change in the range explains the failing step (after checking the inline patch, the name-status list, and any hunks you fetched) → the cause is outside the code: answer UNKNOWN with low confidence and name the suspected external cause (flaky timing, environment, external service, test data). Do NOT default to PRODUCT_BUG here — under this baseline a product regression must be tied to an in-range change.` : `- Diff is unrelated to the failing step (or there is no relevant diff) and the test was passing before → lean PRODUCT_BUG; first rule out timing/data flakiness and infrastructure errors (daemon not running, network down, missing credentials) — those read as UNKNOWN with low confidence, not PRODUCT_BUG.`}${range ? `
+- This range spans ${range.commitCount} commit${range.commitCount === 1 ? "" : "s"} over ${range.days} day${range.days === 1 ? "" : "s"}. The wider the range, the more unrelated changes are mixed in: SPEC_CHANGE and TEST_DRIFT still require citing the specific hunk — do not infer intent from the bulk of a large diff, and lower confidence when the evidence is spread thin.` : ""}
 - The drift audit findings (when present) flag spec↔code mismatches; an ERROR there usually supports TEST_DRIFT or SPEC_CHANGE over PRODUCT_BUG.
 
 ## Sub-diagnosis vocabulary
@@ -3232,32 +3268,15 @@ Your **final** assistant message must start with \`{\` and end with \`}\` — a
 - 0.4-0.7: plausible but another category could explain it
 - < 0.4: answer UNKNOWN instead of guessing
 
-Evidence rules: TEST_DRIFT and SPEC_CHANGE require at least one concrete \`file\` reference (diff hunk or file:line you actually read). PRODUCT_BUG should explain why the diff does NOT account for the failure.
+Evidence rules: TEST_DRIFT and SPEC_CHANGE require at least one concrete \`file\` reference (diff hunk or file:line you actually read). PRODUCT_BUG should cite the in-range change that unintentionally broke the behavior when one exists; ${lastGreen ? "under this last-green baseline, if no in-range change explains the failure, that is UNKNOWN (external cause), not PRODUCT_BUG" : "when no such change exists, explain why the diff does NOT account for the failure"}.
 
 ## Test Spec (spec.yaml)
 ${specYaml}
 
-${buildExecutionEvidenceBlock(script, failureLog, liveTranscriptExcerpt)}
-
-${diffPatch ? `## Source changes since ${baseRef ?? "base"} (git diff, may be truncated)
-
-### Changed files (name-status)
-${changedFiles ?? "(unavailable)"}
-
-### Patch
-\`\`\`diff
-${diffPatch}
-\`\`\`
-` : `## Source changes
+${executionBlock}
 
-No diff context is available (the base ref could not be resolved, or there are no changes). Classify from the failure log, the spec, and what you can read in the repository — and be correspondingly more conservative: prefer UNKNOWN over a confident SPEC_CHANGE/PRODUCT_BUG call without diff evidence.
-`}
-${driftIssues && driftIssues.length > 0 ? `## Spec↔code drift audit findings
-
-A separate read-only audit compared the spec against the current source. Treat these as hints, not verdicts:
-
-${driftIssues.map((i) => `- [${i.severity}] (${DRAFT_CATEGORY_LABEL[i.category]}${i.stepId ? `, step ${i.stepId}` : ""}) ${i.message}${i.detail ? ` — ${i.detail}` : ""}`).join("\n")}
-` : ""}`;
+${diffBlock}
+${driftBlock}`;
 }
 /**
 * Render the execution-evidence section the model needs to classify the
@@ -3671,17 +3690,20 @@ function splitPatchByFile(patch) {
 /**
 * Scope a full patch down to the files a spec depends on, then truncate so
 * the analysis prompt stays bounded. `relatedPaths` null/empty means the
-* spec is unscoped — keep the whole patch (still truncated). Callers scoping
-* the same patch for many specs can pass pre-split sections instead.
+* spec is unscoped — keep the whole patch (still truncated). When
+* relatedPaths are declared but nothing in the diff matches, the result is
+* the empty string: "no related change" is itself a signal the prompt
+* renders explicitly, and the model can inspect any unmatched file's hunk
+* via the on-demand diff tool — inlining the full unrelated diff (the old
+* fallback) just burned the prompt budget, especially under wide last-green
+* baselines. Callers scoping the same patch for many specs can pass
+* pre-split sections instead.
 */
 function scopePatchForSpec(patch, relatedPaths, caps = {}) {
 	const perFile = caps.perFile ?? 8192;
 	const total = caps.total ?? 49152;
 	let sections = typeof patch === "string" ? splitPatchByFile(patch) : patch;
-	if (relatedPaths && relatedPaths.length > 0) {
-		const scoped = sections.filter((s) => isPathAffectedBy(s.path, relatedPaths));
-		if (scoped.length > 0) sections = scoped;
-	}
+	if (relatedPaths && relatedPaths.length > 0) sections = sections.filter((s) => isPathAffectedBy(s.path, relatedPaths));
 	const parts = [];
 	let used = 0;
 	let droppedFiles = 0;
@@ -3716,6 +3738,40 @@ function lookupFileDiff(sections, path) {
 	if (section.body.length <= 16384) return section.body;
 	return `${section.body.slice(0, FILE_DIFF_RESPONSE_CAP)}\n[truncated: ${section.body.length - FILE_DIFF_RESPONSE_CAP} more chars — Read the file for its full current state]`;
 }
+/**
+* Best-effort width of the base..HEAD range. Two-dot rev-list matches what
+* the three-dot diff shows: commits on the HEAD side since the merge base.
+*/
+async function measureRange(sha, cwd) {
+	try {
+		const [{ stdout: count }, { stdout: baseTime }, { stdout: headTime }] = await Promise.all([
+			execFileP("git", [
+				"rev-list",
+				"--count",
+				`${sha}..HEAD`
+			], { cwd }),
+			execFileP("git", [
+				"log",
+				"-1",
+				"--format=%ct",
+				sha
+			], { cwd }),
+			execFileP("git", [
+				"log",
+				"-1",
+				"--format=%ct",
+				"HEAD"
+			], { cwd })
+		]);
+		const seconds = Number(headTime.trim()) - Number(baseTime.trim());
+		return {
+			commitCount: Number(count.trim()),
+			days: Math.max(0, Math.round(seconds / 86400))
+		};
+	} catch {
+		return null;
+	}
+}
 function createDiffProvider(args) {
 	const { resolveBase, cwd } = args;
 	const captures = /* @__PURE__ */ new Map();
@@ -3724,17 +3780,19 @@ function createDiffProvider(args) {
 		const cached = captures.get(sha);
 		if (cached) return cached;
 		const pending = (async () => {
-			const result = await capturePrDiff(sha, cwd);
+			const [result, range] = await Promise.all([capturePrDiff(sha, cwd), measureRange(sha, cwd)]);
 			if (!result.ok) return {
 				sections: null,
 				nameStatus: null,
-				error: result.error
+				error: result.error,
+				range
 			};
 			const { patch, nameStatus } = result.diff;
 			return {
 				sections: patch.length > 0 ? splitPatchByFile(patch) : [],
 				nameStatus,
-				error: null
+				error: null,
+				range
 			};
 		})();
 		captures.set(sha, pending);
@@ -3760,6 +3818,7 @@ function createDiffProvider(args) {
 			patch: sections ? scopePatchForSpec(sections, scope) : null,
 			nameStatus: captured.nameStatus,
 			error: captured.error,
+			range: captured.range,
 			fileDiff: (path) => sections ? lookupFileDiff(sections, path) : null
 		};
 	} };
@@ -6146,6 +6205,8 @@ async function analyzeOneLiveFailure(r, diffProvider, driftForSpec, auth, opts,
 		diffPatch: specDiff.patch,
 		changedFiles: specDiff.nameStatus,
 		baseRef: specDiff.base.ref,
+		baseSource: specDiff.base.source,
+		range: specDiff.range,
 		driftIssues: driftForSpec,
 		...opts.language ? { outputLanguage: opts.language } : {},
 		...opts.triageUserPrompt ? { triageUserPrompt: opts.triageUserPrompt } : {},
@@ -9823,6 +9884,8 @@ async function analyzeDeterministicSummaries(summaries, opts, cwd, reportDir, cu
 				diffPatch: diffExcerpt,
 				changedFiles: specDiffResult.nameStatus,
 				baseRef: specDiffResult.base.ref,
+				baseSource: specDiffResult.base.source,
+				range: specDiffResult.range,
 				driftIssues,
 				...opts.language ? { outputLanguage: opts.language } : {},
 				...triageUserPrompt ? { triageUserPrompt } : {},
@@ -9888,7 +9951,7 @@ function buildReportEnvelope(args) {
 		},
 		model: opts.model ?? null,
 		language: opts.language ?? null,
-		promptVersion: "5",
+		promptVersion: "6",
 		customPromptVersion,
 		...triageUserPromptHash !== null ? { triageUserPromptHash } : {}
 	};
@@ -18409,7 +18472,7 @@ function createLearningWorker(deps) {
 		const generatedAt = (/* @__PURE__ */ new Date()).toISOString();
 		const customPrompt = {
 			schemaVersion: 1,
-			basePromptVersion: "5",
+			basePromptVersion: "6",
 			customPromptVersion: `${generatedAt}-c${cases.length}`,
 			generatedAt,
 			guidance

package/dist/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ccqa",
-  "version": "1.4.0",
+  "version": "1.5.0",
   "type": "module",
   "description": "Browser test recorder powered by Claude Code and agent-browser",
   "repository": {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ccqa",
-  "version": "1.4.0",
+  "version": "1.5.0",
   "type": "module",
   "description": "Browser test recorder powered by Claude Code and agent-browser",
   "repository": {