ccqa 1.4.0 → 1.6.0

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
-${diffPatch}
-\`\`\`
-` : `## Source changes
-
-No diff context is available (the base ref could not be resolved, or there are no changes). Classify from the failure log, the spec, and what you can read in the repository — and be correspondingly more conservative: prefer UNKNOWN over a confident SPEC_CHANGE/PRODUCT_BUG call without diff evidence.
-`}
-${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:
+${executionBlock}
 
-${driftIssues.map((i) => `- [${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
 		};
 	} };
@@ -5831,7 +5890,6 @@ async function runLiveSpecs(specs, opts) {
 	const userPromptSuffix = userPromptBundle?.text ?? null;
 	const diffProvider = opts.diffProvider ?? null;
 	const failureAnalysisEnabled = diffProvider != null;
-	const driftAuditEnabled = failureAnalysisEnabled && opts.driftAudit !== false;
 	const auth = failureAnalysisEnabled ? driftAuthAvailable() : {
 		ok: false,
 		reason: "disabled"
@@ -5859,8 +5917,7 @@ async function runLiveSpecs(specs, opts) {
 			const row = await buildLiveReportRow(outcome, {
 				auth,
 				diffProvider,
-				reportDir,
-				driftAuditEnabled
+				reportDir
 			}, opts, cwd);
 			await opts.report?.upsert(row);
 			return {
@@ -5894,7 +5951,7 @@ async function buildLiveReportRow(r, ctx, opts, cwd) {
 		result: r.result,
 		reportDir: ctx.reportDir
 	});
-	const driftForSpec = ctx.driftAuditEnabled && r.result.status === "failed" ? await runDriftAuditOne(r, opts, cwd) : null;
+	const driftForSpec = ctx.diffProvider && r.result.status === "failed" ? await runDriftAuditOne(r, opts, cwd) : null;
 	const analysis = ctx.diffProvider && r.result.status === "failed" ? await analyzeOneLiveFailure(r, ctx.diffProvider, driftForSpec, ctx.auth, opts, cwd) : void 0;
 	return {
 		...base,
@@ -6146,6 +6203,8 @@ 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 } : {},
@@ -9494,7 +9553,6 @@ async function executeRun(targets, opts) {
 		...typeof opts.retry === "number" ? { retry: opts.retry } : {},
 		concurrency: opts.concurrency ?? 1,
 		...opts.profile ? { profile: opts.profile } : {},
-		...opts.driftAudit !== false ? { driftAudit: true } : {},
 		diffProvider,
 		hubContext: hubCtx,
 		customPrompt,
@@ -9724,7 +9782,6 @@ function failedSpec(s) {
 */
 async function analyzeDeterministicSummaries(summaries, opts, cwd, reportDir, customPrompt, triageUserPrompt, diffProvider) {
 	const failureAnalysisEnabled = diffProvider != null;
-	const driftAuditEnabled = failureAnalysisEnabled && opts.driftAudit !== false;
 	const auth = failureAnalysisEnabled ? driftAuthAvailable() : {
 		ok: false,
 		reason: "skipped by flags"
@@ -9735,7 +9792,7 @@ async function analyzeDeterministicSummaries(summaries, opts, cwd, reportDir, cu
 	const specInfoByKey = new Map(tree.flatMap((f) => f.specs.map((sp) => [`${f.featureName}/${sp.specName}`, sp])));
 	const findSpecInfo = (s) => specInfoByKey.get(`${s.featureName}/${s.specName}`) ?? null;
 	let driftResults = [];
-	if (driftAuditEnabled && auth.ok && failed.length > 0) {
+	if (failureAnalysisEnabled && auth.ok && failed.length > 0) {
 		const targets = failed.map((s) => {
 			const spec = findSpecInfo(s);
 			if (!spec) return null;
@@ -9823,6 +9880,8 @@ 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 +9947,7 @@ function buildReportEnvelope(args) {
 		},
 		model: opts.model ?? null,
 		language: opts.language ?? null,
-		promptVersion: "5",
+		promptVersion: "6",
 		customPromptVersion,
 		...triageUserPromptHash !== null ? { triageUserPromptHash } : {}
 	};
@@ -10212,7 +10271,7 @@ function installTeardownSignalHandlers(teardown) {
 }
 //#endregion
 //#region src/cli/run.ts
-const runCommand = addHubOptions(addProfileOption(addLanguageOption(new Command("run").argument("[targets...]", "Specs to run, space-separated: each '<feature>/<spec>', '<feature>', or omit for all. Duplicates are de-duped.").description("Run specs. Each spec's execution mode comes from its spec.yaml `mode:` field (default deterministic; set `mode: live` to have Claude drive agent-browser live per step). Deterministic specs replay the recorded test.spec.ts under vitest. A structured report (report.json + evidence) is always written; use --push-report to also stream it to a hub.").option("--report [dir]", `Directory for the structured run results (report.json + evidence PNGs) that are always written. Default: ${DEFAULT_REPORT_DIR}/. Pass this only to change the location.`).option("--push-report", "Incrementally push the run report to the hub as the run progresses (open → patch per spec → finalize). Requires --hub-url/--hub-token (or CCQA_HUB_URL/CCQA_HUB_TOKEN). Without it, hub credentials are used only to fetch variables/sessions/prompts, not to push.").option("--changed [base]", "Restrict execution to specs whose relatedPaths intersect the git diff against [base]. Without a value the base comes from $GITHUB_BASE_REF (pull_request CI); elsewhere pass it explicitly (e.g. --changed=origin/main). Cannot be combined with an explicit spec id.").option("--failure-analysis [base]", "Classify each failure (TEST_DRIFT / SPEC_CHANGE / PRODUCT_BUG) against the source diff since [base]. Without a value the base comes from $GITHUB_BASE_REF (pull_request CI); elsewhere pass it explicitly (e.g. --failure-analysis=origin/main), or pass 'last-green' to diff each spec against the commit where it last passed (per-spec baselines from the hub; requires a hub connection). Off by default — no Claude calls without it.").option("--no-drift-audit", "With --failure-analysis: skip the spec↔code drift audit shown in the report.").option("--cwd <path>", "Working directory containing the .ccqa/ tree (monorepo support). Defaults to the current directory.").option("--format <fmt>", "Additional output format alongside HTML when --report is set: 'text' (default), 'json' (writes report.json), 'github' (GitHub Actions annotations on stdout).", (raw) => {
+const runCommand = addHubOptions(addProfileOption(addLanguageOption(new Command("run").argument("[targets...]", "Specs to run, space-separated: each '<feature>/<spec>', '<feature>', or omit for all. Duplicates are de-duped.").description("Run specs. Each spec's execution mode comes from its spec.yaml `mode:` field (default deterministic; set `mode: live` to have Claude drive agent-browser live per step). Deterministic specs replay the recorded test.spec.ts under vitest. A structured report (report.json + evidence) is always written; use --push-report to also stream it to a hub.").option("--report [dir]", `Directory for the structured run results (report.json + evidence PNGs) that are always written. Default: ${DEFAULT_REPORT_DIR}/. Pass this only to change the location.`).option("--push-report", "Incrementally push the run report to the hub as the run progresses (open → patch per spec → finalize). Requires --hub-url/--hub-token (or CCQA_HUB_URL/CCQA_HUB_TOKEN). Without it, hub credentials are used only to fetch variables/sessions/prompts, not to push.").option("--changed [base]", "Restrict execution to specs whose relatedPaths intersect the git diff against [base]. Without a value the base comes from $GITHUB_BASE_REF (pull_request CI); elsewhere pass it explicitly (e.g. --changed=origin/main). Cannot be combined with an explicit spec id.").option("--failure-analysis [base]", "Classify each failure (TEST_DRIFT / SPEC_CHANGE / PRODUCT_BUG) against the source diff since [base]. Without a value the base comes from $GITHUB_BASE_REF (pull_request CI); elsewhere pass it explicitly (e.g. --failure-analysis=origin/main), or pass 'last-green' to diff each spec against the commit where it last passed (per-spec baselines from the hub; requires a hub connection). Off by default — no Claude calls without it.").option("--cwd <path>", "Working directory containing the .ccqa/ tree (monorepo support). Defaults to the current directory.").option("--format <fmt>", "Additional output format alongside HTML when --report is set: 'text' (default), 'json' (writes report.json), 'github' (GitHub Actions annotations on stdout).", (raw) => {
 	if (REPORT_FORMATS.includes(raw)) return raw;
 	throw new Error(`--format must be one of ${REPORT_FORMATS.join(" | ")}`);
 }, "text").option("-m, --model <name>", "Claude model alias ('sonnet'|'opus'|'haiku') or full ID. Overrides CCQA_MODEL.").option("--no-evidence", `(deterministic only) Skip step-boundary evidence capture (PNG + meta JSON written to ${DEFAULT_REPORT_DIR}/${EVIDENCE_SUBDIR}/ by default).`).option("--retry <n>", "(live only) Retry each failed step up to N more times before recording failure. Default 0.", (raw) => {
@@ -15041,21 +15100,25 @@ const HTML_BODY = `
         <div class="rd-head" id="rd-head"></div>
 
         <!-- Triage first: grading + learning is the most important action, so it
-             sits above the spec list rather than being buried below it. -->
+             sits above the spec list rather than being buried below it. The
+             heading row carries the graded counter; one card below it holds
+             the confusion matrix (or its empty state) with the learn CTA as
+             the card footer — grades are the learning job's input, so
+             grade → tally → learn reads top to bottom. -->
         <div class="triage-head" id="triage-head">
           <h3 style="font-size:14px" data-i18n="detail.triage">Triage</h3>
           <span class="triage-summary" id="triage-summary"></span>
         </div>
-        <div class="card" id="matrix-card"></div>
-        <p class="muted" id="triage-progress" style="font-size:12.5px;margin-top:8px"></p>
-
-        <div class="learn-cta" id="learn-cta" hidden>
-          <div class="learn-cta-text">
-            <div class="t" data-i18n="learn.cta.title">Learn from these grades</div>
-            <div class="d" data-i18n="learn.cta.desc">Turn the graded cases into a custom prompt that calibrates future failure classification.</div>
-          </div>
-          <div class="learn-cta-actions">
-            <button class="btn primary sm" id="learn-run" data-i18n="learn.cta.run">Learn</button>
+        <div class="card triage-card" id="triage-card">
+          <div id="matrix-card"></div>
+          <div class="learn-cta" id="learn-cta" hidden>
+            <div class="learn-cta-text">
+              <div class="t" data-i18n="learn.cta.title">Learn from these grades</div>
+              <div class="d" data-i18n="learn.cta.desc">Turn the graded cases into a custom prompt that calibrates future failure classification.</div>
+            </div>
+            <div class="learn-cta-actions">
+              <button class="btn primary sm" id="learn-run" data-i18n="learn.cta.run">Learn</button>
+            </div>
           </div>
         </div>
 
@@ -15379,8 +15442,6 @@ const CSS = `
   .badge-det { background: var(--surface-3); color: var(--muted); border-color: var(--border); }
   /* which generation target ran the spec (agent-browser / playwright / runn) */
   .badge-target { display: inline-flex; align-items: center; padding: 2px 8px; border-radius: var(--radius-sm); font-size: 11px; font-family: var(--mono); background: var(--surface-3); color: var(--muted); border: 1px solid var(--border); }
-  .badge.drift-warn { background: var(--amber-bg); color: var(--amber); border-color: var(--amber-border); }
-  .badge.drift-warn .d { background: var(--amber); }
   .badge-drift { background: var(--violet-bg); color: var(--violet); border-color: var(--violet-border); }
   .chip { display: inline-flex; align-items: center; padding: 1px 8px; border-radius: 6px; background: var(--surface-3); border: 1px solid var(--border); color: var(--fg-dim); font-size: 12px; font-family: var(--mono); }
   /* Below .chip in source order so these override its background/border/color
@@ -15423,11 +15484,18 @@ const CSS = `
   .spec-card-head .spacer { flex: 1; }
   .spec-card-body { padding: 0 20px 16px; }
   /* Tier2 verdict block */
-  .analysis-box { display: flex; flex-direction: column; gap: 12px; padding-bottom: 4px; }
+  /* The diagnosis card: a bordered sub-surface so the model's verdict + the
+     grading zone read as one unit, distinct from the execution details
+     (steps/assertions) below it. */
+  .analysis-box { display: flex; flex-direction: column; gap: 12px; border: 1px solid var(--border); border-radius: var(--radius-md); background: var(--surface-2); padding: 14px 16px; }
+  .analysis-box .acc > summary:hover { background: var(--surface-3); }
   .analysis-head { display: flex; align-items: center; gap: 8px; flex-wrap: wrap; }
-  .analysis-headline { font-size: 14px; font-weight: 600; color: var(--fg); line-height: 1.5; }
-  .analysis-rec { font-size: 13px; color: var(--fg-dim); background: var(--surface-2); border: 1px solid var(--border); border-left: 2px solid var(--muted); border-radius: var(--radius-sm); padding: 10px 12px; line-height: 1.55; }
-  .analysis-rec .rec-k { display: block; font-size: 10.5px; text-transform: uppercase; letter-spacing: 0.05em; color: var(--muted); font-weight: 600; margin-bottom: 4px; }
+  .analysis-kv { display: grid; grid-template-columns: auto 1fr; gap: 6px 14px; font-size: 13.5px; }
+  .analysis-kv .k { font-size: 11px; font-weight: 600; color: var(--muted); padding-top: 3px; white-space: nowrap; }
+  .analysis-kv .v { color: var(--fg-dim); line-height: 1.55; }
+  .analysis-kv .v.headline { color: var(--fg); font-weight: 600; }
+  /* Model-evidence rows reuse the drift-row list shape; only the file ref needs its own style. */
+  .ev-file { font-size: 12px; color: var(--fg-dim); }
   .analysis-reasoning { font-size: 13px; color: var(--fg-dim); white-space: pre-wrap; line-height: 1.6; }
   .analysis-inline-reason { font-size: 13px; color: var(--fg-dim); line-height: 1.55; }
   /* Tier3 accordion (real header bar + rotating chevron, replaces the tiny ▸) */
@@ -15493,11 +15561,12 @@ const CSS = `
 
   /* triage grading — an explicit question + a segmented single-select, framed
      as an action ("tell us the real cause"), not a data readout. */
-  .grade { margin-top: 4px; padding: 14px; border: 1px solid var(--border); border-radius: var(--radius-md); background: var(--surface-2); }
+  /* Embedded at the bottom of the diagnosis card: a divider separates the
+     human's grading zone from the model's output above it, without breaking
+     the two out of the shared context. */
+  .grade { margin-top: 2px; padding: 12px 0 0; border-top: 1px solid var(--border); }
   .grade-top { display: flex; align-items: center; gap: 10px; flex-wrap: wrap; margin-bottom: 10px; }
   .grade-q { font-size: 13px; font-weight: 600; color: var(--fg); }
-  .grade-pred { display: inline-flex; align-items: center; gap: 6px; margin-left: auto; font-size: 12px; color: var(--muted); }
-  .grade-pred .grade-arrow { color: var(--muted-2); }
   .grade-bottom { display: flex; align-items: center; gap: 12px; flex-wrap: wrap; }
   .grade-seg { display: inline-flex; border: 1px solid var(--border-strong); border-radius: var(--radius-md); overflow: hidden; background: var(--surface); }
   .grade-seg .seg { height: 34px; padding: 0 14px; border: 0; background: transparent; color: var(--muted); font-size: 13px; font-weight: 500; border-right: 1px solid var(--border); display: inline-flex; align-items: center; gap: 5px; }
@@ -15635,7 +15704,7 @@ const CSS = `
   .ro-tag { font-size: 10.5px; font-weight: 600; text-transform: uppercase; letter-spacing: 0.04em;
     color: var(--muted-2); background: var(--surface-3); border: 1px solid var(--border); border-radius: 999px; padding: 1px 8px; margin-left: 4px; }
 
-  .learn-cta { display: flex; align-items: center; gap: 16px; margin-top: 18px; padding: 16px 18px; border: 1px solid var(--border-strong); border-radius: var(--radius-sm); background: var(--surface-2); }
+  .learn-cta { display: flex; align-items: center; gap: 16px; padding: 14px 20px; border-top: 1px solid var(--border); background: var(--surface-2); border-radius: 0 0 var(--radius-md) var(--radius-md); }
   .learn-cta-text { flex: 1; min-width: 0; }
   .learn-cta-text .t { font-size: 13.5px; font-weight: 600; color: var(--fg); }
   .learn-cta-text .d { font-size: 12px; color: var(--muted); margin-top: 3px; }
@@ -15756,26 +15825,25 @@ const CLIENT_JS = `
       "meta.branch": "Branch", "meta.specs": "Specs", "meta.prompt": "Prompt",
       "meta.created": "Created", "meta.passed": "passed", "meta.profile": "Profile",
       "meta.drift": "Drift",
-      "rec.title": "Recommendation",
+      "diag.cause": "Cause", "diag.fix": "Fix",
       "acc.reasoning": "Reasoning", "acc.evidence": "Evidence", "acc.steps": "Live run steps",
       "acc.assertions": "Assertions", "acc.drift": "Drift audit",
       "acc.artifacts": "Artifacts",
       "art.open": "Open", "art.loadFailed": "could not load (it may have been omitted from the push)",
       "acc.assertions.hint": "Test cases from the recorded spec run",
       "spec.kind.live": "Live", "spec.kind.det": "Deterministic",
-      "spec.driftWarn": "Spec drift", "det.steps": "Steps",
+      "det.steps": "Steps",
       "kind.run": "Test run", "kind.drift": "Drift audit",
       "drift.summary.issues": "Issues", "drift.summary.errors": "Errors",
       "drift.summary.warnings": "Warnings", "drift.summary.specsWithIssues": "Specs with issues",
       "drift.clean": "No drift issues",
-      "grade.question": "What was the real cause?", "grade.predicted": "predicted",
+      "grade.question": "What was the real cause?",
       "grade.ungraded": "ungraded", "grade.matches": "saved · matches",
       "grade.corrected": "saved · corrected", "grade.saving": "saving…",
       "grade.error": "couldn't save — retry",
-      "matrix.empty": "No graded cases yet. Grade a failed spec above to populate the confusion matrix.",
+      "matrix.empty": "No grades yet. Pick the real cause on a failed spec's diagnosis card below and it is tallied here.",
       "matrix.predicted": "predicted \\\\ actual", "matrix.accuracy": "Accuracy",
       "matrix.accSuffix": "of graded cases match the prediction", "matrix.graded": "graded",
-      "matrix.progress": "Recorded actual cause: {n} / {total} failing specs",
       "learn.cta.title": "Learn from these grades",
       "learn.cta.desc": "Learn from what you graded so ccqa classifies failure causes the same way next time.",
       "learn.cta.run": "Learn",
@@ -15846,26 +15914,25 @@ const CLIENT_JS = `
       "meta.branch": "ブランチ", "meta.specs": "スペック", "meta.prompt": "プロンプト",
       "meta.created": "作成", "meta.passed": "合格", "meta.profile": "プロファイル",
       "meta.drift": "ドリフト",
-      "rec.title": "推奨対応",
+      "diag.cause": "原因", "diag.fix": "対処",
       "acc.reasoning": "推論", "acc.evidence": "根拠", "acc.steps": "実行ステップ",
       "acc.assertions": "アサーション", "acc.drift": "ドリフト監査",
       "acc.artifacts": "成果物",
       "art.open": "開く", "art.loadFailed": "読み込めませんでした（push時に省略された可能性があります）",
       "acc.assertions.hint": "記録したスペック実行のテストケース",
       "spec.kind.live": "ライブ", "spec.kind.det": "決定的",
-      "spec.driftWarn": "仕様ドリフト", "det.steps": "ステップ",
+      "det.steps": "ステップ",
       "kind.run": "テスト実行", "kind.drift": "ドリフト監査",
       "drift.summary.issues": "問題数", "drift.summary.errors": "エラー",
       "drift.summary.warnings": "警告", "drift.summary.specsWithIssues": "問題のあるスペック",
       "drift.clean": "ドリフトの問題なし",
-      "grade.question": "実際の原因は何でしたか？", "grade.predicted": "予測",
+      "grade.question": "実際の原因は何でしたか？",
       "grade.ungraded": "未評価", "grade.matches": "保存済み · 一致",
       "grade.corrected": "保存済み · 修正", "grade.saving": "保存中…",
       "grade.error": "保存に失敗 — 再試行",
-      "matrix.empty": "まだ評価がありません。上の失敗スペックを評価すると混同行列に反映されます。",
+      "matrix.empty": "まだ採点がありません。下の失敗スペックの診断カードで実際の原因を選ぶと、ここに集計されます。",
       "matrix.predicted": "予測 \\\\ 実際", "matrix.accuracy": "正解率",
       "matrix.accSuffix": "件の採点が予測と一致", "matrix.graded": "採点済み",
-      "matrix.progress": "実際の原因を記録: {total} 件中 {n} 件の失敗スペック",
       "learn.cta.title": "この採点から学習",
       "learn.cta.desc": "採点した内容をもとに、ccqaが次回から同じように失敗の原因を分類できるよう学習します。",
       "learn.cta.run": "学習",
@@ -16465,27 +16532,61 @@ const CLIENT_JS = `
 
   // ── run detail: spec cards ──────────────────────────────────────────
 
-  // Tier2 verdict block: label + confidence, headline, and the recommendation
-  // callout. Reasoning is NOT here — renderSpecCard places it as a Tier3
-  // accordion (or inline when it's too short to be worth folding).
+  // The diagnosis card: one surface for everything about a failure's cause.
+  // Verdict (label + confidence), then the cause→fix pair as labelled rows —
+  // headline and recommendation are one causal unit, so they read as one.
+  // subDiagnosis is deliberately NOT shown: it is a machine vocabulary for
+  // accuracy stratification and learning, not for humans. The caller appends
+  // the evidence/reasoning accordions and the grading zone into this box.
   function analysisSection(runId, r) {
     var wrap = el("div", "analysis-box");
     var a = r.analysis;
     var head = el("div", "analysis-head");
     head.appendChild(labelChip(a.label));
     head.appendChild(el("span", "conf", Math.round(a.confidence * 100) + "%"));
-    if (a.subDiagnosis && a.subDiagnosis !== "NONE") head.appendChild(el("span", "muted", a.subDiagnosis));
     wrap.appendChild(head);
-    if (a.headline) wrap.appendChild(el("div", "analysis-headline", a.headline));
+    var kv = el("div", "analysis-kv");
+    if (a.headline) {
+      kv.appendChild(el("div", "k", t("diag.cause")));
+      kv.appendChild(el("div", "v headline", a.headline));
+    }
     if (a.recommendation) {
-      var rec = el("div", "analysis-rec");
-      rec.appendChild(el("span", "rec-k", t("rec.title")));
-      rec.appendChild(document.createTextNode(a.recommendation));
-      wrap.appendChild(rec);
+      kv.appendChild(el("div", "k", t("diag.fix")));
+      kv.appendChild(el("div", "v", a.recommendation));
     }
+    if (kv.childNodes.length > 0) wrap.appendChild(kv);
     return wrap;
   }
 
+  // 根拠: the model's evidence items (file + what it proves) merged with the
+  // drift-audit findings. The audit is an input hint TO the classifier, so
+  // its findings belong here as supporting evidence — not as a sibling
+  // section that reads like an independent feature.
+  function analysisEvidenceSection(r) {
+    var wrap = el("div");
+    var count = 0;
+    var items = r.analysis && r.analysis.evidence ? r.analysis.evidence : [];
+    items.forEach(function (e) {
+      // Same list shape as the drift findings below, so the merged 根拠 list
+      // reads as one.
+      var row = el("div", "drift-row");
+      if (e.file) {
+        var head = el("div", "drift-head");
+        head.appendChild(el("code", "ev-file", e.file));
+        row.appendChild(head);
+      }
+      row.appendChild(el("div", "drift-msg", e.detail));
+      wrap.appendChild(row);
+      count++;
+    });
+    if (r.driftIssues && r.driftIssues.length > 0) {
+      wrap.appendChild(el("div", "section-label", t("acc.drift")));
+      wrap.appendChild(driftSection(r.driftIssues));
+      count += r.driftIssues.length;
+    }
+    return { node: wrap, count: count };
+  }
+
   function evidenceSection(runId, evidence) {
     var grid = el("div", "evidence-grid");
     evidence.forEach(function (e) {
@@ -16717,25 +16818,22 @@ const CLIENT_JS = `
     return det;
   }
 
-  // The grading action: an explicit question ("What was the real cause?") with
-  // the model's guess as muted context, a segmented single-select over the
-  // failure labels, and a status chip (ungraded / saved·matches / saved·
-  // corrected). One tap grades it. Optimistic PUT with rollback; on success it
-  // refreshes the confusion matrix. The English label value is what's sent and
-  // stored; the segment just shows its localized name.
+  // The grading action: an explicit question ("What was the real cause?"), a
+  // segmented single-select over the failure labels, and a status chip
+  // (ungraded / saved·matches / saved·corrected). One tap grades it.
+  // Optimistic PUT with rollback; on success it refreshes the confusion
+  // matrix. The English label value is what's sent and stored; the segment
+  // just shows its localized name.
   function triageGradeControl(runId, r, triageState) {
     var key = r.feature + "/" + r.spec;
     var predicted = r.analysis ? r.analysis.label : "UNKNOWN";
 
     var wrap = el("div", "grade");
 
+    // No "predicted →" chip here: the control lives inside the diagnosis
+    // card, directly under the prediction it grades — repeating it is noise.
     var top = el("div", "grade-top");
     top.appendChild(el("span", "grade-q", t("grade.question")));
-    var pred = el("span", "grade-pred");
-    pred.appendChild(document.createTextNode(t("grade.predicted")));
-    pred.appendChild(el("span", "grade-arrow", "→"));
-    pred.appendChild(labelChip(predicted));
-    top.appendChild(pred);
     wrap.appendChild(top);
 
     var bottom = el("div", "grade-bottom");
@@ -16819,13 +16917,6 @@ const CLIENT_JS = `
     else if (r.liveRun) head.appendChild(el("span", "badge-live", t("spec.kind.live")));
     else if (!external) head.appendChild(el("span", "badge-det", t("spec.kind.det")));
     head.appendChild(statusBadge(r.status));
-    var hasDriftError = r.driftIssues && r.driftIssues.some(function (d) { return d.severity === "ERROR"; });
-    if (hasDriftError) {
-      var w = el("span", "badge drift-warn");
-      w.appendChild(el("span", "d"));
-      w.appendChild(document.createTextNode(" " + t("spec.driftWarn")));
-      head.appendChild(w);
-    }
     card.appendChild(head);
 
     var body = el("div", "spec-card-body");
@@ -16836,17 +16927,23 @@ const CLIENT_JS = `
       any = true;
     }
 
-    if (r.status === "failed" && r.analysis) {
-      // Tier2: the verdict block (analysis) + the grading action, always shown.
-      body.appendChild(analysisSection(runId, r));
-      body.appendChild(triageGradeControl(runId, r, triageState));
-      // Reasoning: fold it as a Tier3 accordion, but only when it carries real
+    var hasAnalysis = r.status === "failed" && r.analysis;
+    if (hasAnalysis) {
+      // The diagnosis card: verdict + cause/fix, then evidence and reasoning
+      // as accordions, then the grading zone — one surface for the whole
+      // "why did this fail and was the call right" story.
+      var box = analysisSection(runId, r);
+      var ev = analysisEvidenceSection(r);
+      if (ev.count > 0) box.appendChild(detailsBlock(t("acc.evidence"), ev.count, ev.node));
+      // Reasoning: fold it as an accordion, but only when it carries real
       // content. A one-char/empty reasoning behind a disclosure reads as broken
       // (the old "▸ r"), so drop it entirely below the threshold.
       var reasoning = r.analysis.reasoning ? String(r.analysis.reasoning).trim() : "";
       if (reasoning.length > 2) {
-        body.appendChild(detailsBlock(t("acc.reasoning"), null, el("div", "analysis-reasoning", reasoning)));
+        box.appendChild(detailsBlock(t("acc.reasoning"), null, el("div", "analysis-reasoning", reasoning)));
       }
+      box.appendChild(triageGradeControl(runId, r, triageState));
+      body.appendChild(box);
       any = true;
     } else if (r.status === "failed" && r.analysisSkipped) {
       body.appendChild(el("div", "muted", "Analysis skipped: " + r.analysisSkipped));
@@ -16878,7 +16975,10 @@ const CLIENT_JS = `
       any = true;
     }
 
-    if (r.driftIssues && r.driftIssues.length > 0) {
+    // Drift findings render standalone only when they aren't already folded
+    // into the diagnosis card's evidence: drift-kind runs (the audit IS the
+    // content) and failed rows whose analysis was skipped.
+    if (!hasAnalysis && r.driftIssues && r.driftIssues.length > 0) {
       body.appendChild(detailsBlock(t("acc.drift"), r.driftIssues.length, driftSection(r.driftIssues)));
       any = true;
     }
@@ -16913,12 +17013,9 @@ const CLIENT_JS = `
     var cases = Object.keys(triageState.byKey).map(function (k) { return triageState.byKey[k]; })
       .filter(function (c) { return c.predicted && c.actual; });
 
-    // Recompute the progress line here so it stays in sync after each grade,
-    // not just on initial load.
+    // The "graded m / n" counter in the header is the single progress
+    // readout — recomputed here so it stays in sync after each grade.
     var total = typeof triageState.total === "number" ? triageState.total : Object.keys(triageState.byKey).length;
-    document.getElementById("triage-progress").textContent =
-      t("matrix.progress").replace("{n}", cases.length).replace("{total}", total);
-
     var summary = document.getElementById("triage-summary");
 
     if (cases.length === 0) {
@@ -16989,7 +17086,12 @@ const CLIENT_JS = `
       renderMatrix(triageState);
       onLoaded(triageState);
     }).catch(function (err) {
-      document.getElementById("triage-progress").textContent = "Error loading triage: " + err.message;
+      // Surface the load failure inside the triage card so it isn't silent.
+      var card = document.getElementById("matrix-card");
+      clear(card);
+      var wrap = el("div", "matrix-wrap");
+      wrap.appendChild(el("div", "muted", "Error loading triage: " + err.message));
+      card.appendChild(wrap);
       onLoaded({ byKey: {} });
     });
   }
@@ -17015,7 +17117,6 @@ const CLIENT_JS = `
     clear(document.getElementById("rd-head"));
     clear(document.getElementById("matrix-card"));
     document.getElementById("detail-spec-count").textContent = "";
-    document.getElementById("triage-progress").textContent = "";
     document.getElementById("triage-summary").textContent = "";
 
     apiFetch("/api/v1/runs/" + encodeURIComponent(runId)).then(function (run) {
@@ -17029,8 +17130,7 @@ const CLIENT_JS = `
       // as the other, and neither escapes its own catch.
       var isDrift = report.kind === "drift";
       document.getElementById("triage-head").hidden = isDrift;
-      document.getElementById("matrix-card").hidden = isDrift;
-      document.getElementById("triage-progress").hidden = isDrift;
+      document.getElementById("triage-card").hidden = isDrift;
       renderSpecCards(runId, report.results, { byKey: {} }, isDrift);
       if (isDrift) return; // drift runs have no triage: skip loadTriage entirely
       loadTriage(runId, function (triageState) {
@@ -18409,7 +18509,7 @@ 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": "1.6.0",
   "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": "1.6.0",
   "type": "module",
   "description": "Browser test recorder powered by Claude Code and agent-browser",
   "repository": {