@tea-agent/loop-agent 0.39.0-next.30 → 0.39.0-next.31

package/dist/workflows/dag/frontend-test-result-contract.js

@@ -7,6 +7,25 @@ import { readPlaywrightCliReceipts, summarizePlaywrightCliReceipts, } from "../.
 import { validateFrontendCaseContent } from "./frontend-test-case-quality.js";
 import { markdownH1Title, markdownList, markdownSection, } from "./frontend-test-markdown.js";
 export const FRONTEND_TEST_RESULT_SCHEMA_ID = "frontend-test-result-v1";
+/**
+ * Truncated acceptance-id shapes observed in generator output: a bare
+ * numeric segment ("AC-002" dropped its AC-FE-<FEATURE> prefix) or an
+ * AC-FE-<FEATURE> literal without the trailing -NNN sequence number
+ * ("AC-FE-CHAT"). Generic short ids like "AC-1"/"AC-FE-001" stay valid;
+ * declared-source cross-checks (unknown-ac) remain the primary guard when a
+ * sourceBinding exists.
+ */
+export function isTruncatedFrontendAcId(acId) {
+    return /^AC-FE-[A-Z]+$/i.test(acId.trim());
+}
+/**
+ * Bare numeric ids ("AC-002") only indicate truncation when no declared AC
+ * list exists to adjudicate them: with a sourceBinding they are ordinary
+ * unknown-ac findings ("AC-999" may be a legitimately declared id style).
+ */
+export function isTruncatedWhenUndeclaredFrontendAcId(acId) {
+    return /^AC-\d{3}$/i.test(acId.trim()) || isTruncatedFrontendAcId(acId);
+}
 const safeRelativePathSchema = z.string().min(1).refine((value) => !path.posix.isAbsolute(value) &&
     !path.win32.isAbsolute(value) &&
     !value.includes("\\") &&
@@ -57,6 +76,8 @@ export const frontendTestResultContractSchema = z.object({
         passed: z.number().int().min(0),
         failed: z.number().int().min(0),
         blocked: z.number().int().min(0),
+        /** Stability metric: passed on attempt 0 (no rerun/self-heal retry). */
+        firstTryPassed: z.number().int().min(0).optional(),
     }).strict(),
     acceptanceCoverage: z.object({
         covered: z.array(z.string().min(1)),
@@ -110,7 +131,13 @@ export async function validateFrontendCaseEvidence(input) {
     catch {
         throw new Error("missing testcase/frontend/cases/manifest.json");
     }
-    const manifest = JSON.parse(await readFile(manifestPath, "utf8"));
+    let manifest;
+    try {
+        manifest = JSON.parse(await readFile(manifestPath, "utf8"));
+    }
+    catch {
+        throw new Error("invalid frontend case manifest: malformed JSON");
+    }
     if (!Array.isArray(manifest.cases))
         throw new Error("invalid frontend case manifest");
     const statuses = new Set(["passed", "failed", "blocked"]);
@@ -195,11 +222,17 @@ export function evaluatePassedCaseBrowserReceipts(summary, caseId) {
     }
     return { ok: true };
 }
-async function loadCaseBrowserReceiptSummary(input) {
+async function loadCaseBrowserReceiptsWithSummary(input) {
     try {
         const entries = await readdir(input.runDir, { withFileTypes: true });
         // A passed chain must belong to exactly one browser child stream. Never
-        // stitch receipts across DAG nodes or accept ambiguous duplicate authority.
+        // stitch receipts across DAG nodes or accept ambiguous duplicate
+        // authority — with one exception: bounded rerun creates a second,
+        // later stream for the same case (the rerun overwrites the disk
+        // case-result, so the final attempt is the authoritative one). When
+        // multiple streams authorize, prefer the highest rerun round
+        // (rerun-frontend-case-rN-<idx> directories); ties within the same
+        // round remain ambiguous and fail closed.
         const authorizingStreams = [];
         for (const entry of entries) {
             if (!entry.isDirectory())
@@ -207,17 +240,99 @@ async function loadCaseBrowserReceiptSummary(input) {
             const nodeReceipts = (await readPlaywrightCliReceipts(input.runDir, entry.name))
                 .filter((item) => item.caseId === input.caseId);
             const summary = summarizePlaywrightCliReceipts(nodeReceipts);
-            if (summary.hasOrderedPassedReceiptChain)
-                authorizingStreams.push(summary);
+            if (summary.hasOrderedPassedReceiptChain) {
+                const roundMatch = /rerun-frontend-case-r(\d+)-/i.exec(entry.name);
+                authorizingStreams.push({
+                    receipts: nodeReceipts,
+                    summary,
+                    round: roundMatch ? Number(roundMatch[1]) : 0,
+                });
+            }
+        }
+        if (authorizingStreams.length === 0) {
+            return { receipts: [], summary: summarizePlaywrightCliReceipts([]) };
         }
-        return authorizingStreams.length === 1
-            ? authorizingStreams[0]
-            : summarizePlaywrightCliReceipts([]);
+        const maxRound = Math.max(...authorizingStreams.map((s) => s.round));
+        const finalists = authorizingStreams.filter((s) => s.round === maxRound);
+        return finalists.length === 1
+            ? finalists[0]
+            : { receipts: [], summary: summarizePlaywrightCliReceipts([]) };
+    }
+    catch {
+        return { receipts: [], summary: summarizePlaywrightCliReceipts([]) };
+    }
+}
+/**
+ * Extract `find "literal" >=N` / `find "literal"` declarations from a case
+ * document. Only quoted literals are binding assertions — the count suffix
+ * (`>=N`, `==N`) is the expected minimum/exact match count from the ledger.
+ * Placeholder literals ("...", "…", "<字面>", "eX") appear in prose discussing
+ * the syntax itself; they are documentation, never declarations.
+ */
+export function extractDeclaredFindAssertions(caseMarkdown) {
+    const source = stripFencedCodeBlocks(caseMarkdown);
+    const out = [];
+    const seen = new Set();
+    const pattern = /find\s+"((?:[^"\\]|\\.)+)"\s*(>=|==)\s*(\d+)/g;
+    const isPlaceholder = (literal) => /^(?:\.{2,}|…+|<[^>]*>|e\d+)$/.test(literal.trim());
+    let match;
+    while ((match = pattern.exec(source)) !== null) {
+        const literal = match[1].replace(/\\(.)/g, "$1");
+        if (isPlaceholder(literal))
+            continue;
+        const operator = match[2];
+        const expected = Number(match[3]);
+        const key = `${operator}:${literal}`;
+        if (seen.has(key))
+            continue;
+        seen.add(key);
+        out.push({ literal, expected, operator });
+    }
+    return out;
+}
+function stripFencedCodeBlocks(markdown) {
+    return markdown.replace(/```[\s\S]*?```/g, "");
+}
+async function readCaseMarkdownForAudit(workspaceRoot, casePath) {
+    try {
+        return await readFile(path.join(workspaceRoot, casePath), "utf8");
     }
     catch {
-        return summarizePlaywrightCliReceipts([]);
+        return "";
     }
 }
+/**
+ * Controller-side assertion audit: every find-family declaration in the case
+ * document must be backed by a successful find receipt whose parsed match
+ * count satisfies the declared operator. This moves pass authority from
+ * "the model says it asserted" to "receipts prove it asserted".
+ */
+export function auditDeclaredFindAssertions(declared, receipts) {
+    if (declared.length === 0)
+        return { ok: true };
+    const successfulFinds = receipts.filter((receipt) => (receipt.command === "find" || receipt.command === "find-unique") &&
+        receipt.exitCode === 0 &&
+        !receipt.timedOut);
+    const missing = [];
+    for (const decl of declared) {
+        const backing = successfulFinds.some((receipt) => receipt.argsRedacted.some((arg) => arg === decl.literal || arg === `"${decl.literal}"`) &&
+            typeof receipt.matchCount === "number" &&
+            (decl.operator === ">="
+                ? receipt.matchCount >= decl.expected
+                : receipt.matchCount === decl.expected));
+        if (!backing) {
+            missing.push(`${decl.operator}${decl.expected} "${decl.literal.slice(0, 40)}"`);
+        }
+    }
+    if (missing.length > 0) {
+        return {
+            ok: false,
+            reason: "assertion-not-receipt-backed",
+            missing,
+        };
+    }
+    return { ok: true };
+}
 function sha256(content) {
     return createHash("sha256").update(content).digest("hex");
 }
@@ -324,7 +439,13 @@ export async function materializeFrontendTestResult(input) {
     }
     const manifestPath = "testcase/frontend/cases/manifest.json";
     const manifestAbsolute = path.resolve(input.workspaceRoot, manifestPath);
-    const manifest = JSON.parse(await readFile(manifestAbsolute, "utf-8"));
+    let manifest;
+    try {
+        manifest = JSON.parse(await readFile(manifestAbsolute, "utf-8"));
+    }
+    catch {
+        throw new Error("invalid or empty frontend case manifest: malformed JSON");
+    }
     if (manifest.schemaVersion !== 1 || !Array.isArray(manifest.cases) || manifest.cases.length === 0) {
         throw new Error("invalid or empty frontend case manifest");
     }
@@ -387,7 +508,7 @@ export async function materializeFrontendTestResult(input) {
             advisoryFindings.push({ ruleId: "passed-without-browser-evidence", caseId: item.caseId, detail: "passed case has no browser evidence" });
         let blockedReason = status === "blocked" && typeof resultRaw.blockedReason === "string" && resultRaw.blockedReason.trim() ? resultRaw.blockedReason.trim() : undefined;
         if (status === "passed") {
-            const receiptSummary = await loadCaseBrowserReceiptSummary({
+            const { receipts: caseReceipts, summary: receiptSummary } = await loadCaseBrowserReceiptsWithSummary({
                 runDir: input.runDir,
                 caseId: item.caseId,
             });
@@ -401,6 +522,24 @@ export async function materializeFrontendTestResult(input) {
                     detail: "passed case lacks controller-owned open + interaction/assertion + cleanup receipts",
                 });
             }
+            else {
+                // Controller-side assertion audit: every `find "literal" >=N` declared
+                // in the case document must be backed by a successful find receipt
+                // whose parsed matchCount satisfies the operator. Pass authority moves
+                // from model prose to receipt evidence.
+                const caseMarkdown = await readCaseMarkdownForAudit(input.workspaceRoot, item.casePath);
+                const declared = extractDeclaredFindAssertions(caseMarkdown);
+                const audit = auditDeclaredFindAssertions(declared, caseReceipts);
+                if (!audit.ok) {
+                    status = "blocked";
+                    blockedReason = audit.reason;
+                    advisoryFindings.push({
+                        ruleId: "assertion-not-receipt-backed",
+                        caseId: item.caseId,
+                        detail: `declared find assertions without receipt backing: ${audit.missing.join("; ")}`,
+                    });
+                }
+            }
         }
         const { caseContent, testPoints } = await readFrontendCaseContent(input.workspaceRoot, item.casePath);
         // Parse bounded fixed sections from execution.md. Missing optional
@@ -451,6 +590,7 @@ export async function materializeFrontendTestResult(input) {
         passed: cases.filter((item) => item.status === "passed").length,
         failed: cases.filter((item) => item.status === "failed").length,
         blocked: cases.filter((item) => item.status === "blocked").length,
+        firstTryPassed: cases.filter((item) => item.status === "passed" && !item.rerunAttempt).length,
     };
     const outcome = totals.failed > 0 ? "failed" : totals.blocked > 0 || missing.length > 0 ? "incomplete" : "passed";
     const result = frontendTestResultContractSchema.parse({

package/dist/workflows/dag/frontend-verification-trace.js

@@ -2,6 +2,18 @@ import { readFile, stat, writeFile, mkdir } from "node:fs/promises";
 import path from "node:path";
 import { FRONTEND_IMPLEMENTATION_CONTRACT_SCHEMA_ID, frontendImplementationContractSchema, } from "./frontend-implementation-contract.js";
 export const FRONTEND_VERIFICATION_TRACE_SCHEMA_ID = "frontend-verification-trace-v1";
+/**
+ * Shared marker for a project that exposes no usable verification command at
+ * generation time. The marker command runs (and logs not-run honestly), but it
+ * must never count as successful verification evidence: labels carrying this
+ * text are excluded from successful evidence and targets bound to it fail the
+ * trace gate instead of passing it.
+ */
+export const FRONTEND_NO_VERIFICATION_MARKER_TEXT = "No frontend verification command configured";
+export function isFrontendNoVerificationMarkerLabel(label) {
+    return label.includes(FRONTEND_NO_VERIFICATION_MARKER_TEXT);
+}
+export const FRONTEND_VERIFICATION_UNAVAILABLE_MESSAGE = `frontend verification unavailable: the only frozen static command is the no-op not-run marker (${FRONTEND_NO_VERIFICATION_MARKER_TEXT}); declare verification commands via task.json --verify or 执行约束.md and regenerate the DAG`;
 async function readJsonFile(filePath) {
     return JSON.parse(await readFile(filePath, "utf8"));
 }
@@ -52,7 +64,14 @@ export async function evaluateVerificationTargets(input) {
     const hardIssues = [];
     for (const target of input.contract.verificationTargets) {
         const issues = [];
-        const owners = input.labelOwners.get(target.commandLabel);
+        if (isFrontendNoVerificationMarkerLabel(target.commandLabel)) {
+            // A target bound to the not-run marker has no verification behind it;
+            // passing it would let the run finish green without any check.
+            issues.push(`${target.id}: ${FRONTEND_VERIFICATION_UNAVAILABLE_MESSAGE}`);
+        }
+        const owners = isFrontendNoVerificationMarkerLabel(target.commandLabel)
+            ? []
+            : (input.labelOwners.get(target.commandLabel) ?? []);
         if (!owners?.length) {
             issues.push(`commandLabel not executed in current-run frozen evidence: ${target.commandLabel}`);
         }
@@ -189,11 +208,17 @@ export async function runFrontendVerificationTraceGate(input) {
         throw new Error("trace: selected Mock strategy requires successful Mock verification commands");
     }
     if (input.evidence) {
-        if (input.evidence.static.commandLabels.length === 0) {
-            throw new Error("trace: no successful static verification commands in current run");
+        // The no-op not-run marker never counts as successful verification
+        // evidence, even if it executed with exit code 0.
+        const effectiveStaticLabels = input.evidence.static.commandLabels.filter((label) => !isFrontendNoVerificationMarkerLabel(label));
+        const effectiveBehaviorLabels = input.evidence.behavior.commandLabels.filter((label) => !isFrontendNoVerificationMarkerLabel(label));
+        if (effectiveStaticLabels.length === 0) {
+            throw new Error(input.evidence.static.commandLabels.length > 0
+                ? `trace: ${FRONTEND_VERIFICATION_UNAVAILABLE_MESSAGE}`
+                : "trace: no successful static verification commands in current run");
         }
         if (requiresBehaviorVerification &&
-            input.evidence.behavior.commandLabels.length === 0) {
+            effectiveBehaviorLabels.length === 0) {
             throw new Error("trace: no successful behavior verification commands in current run");
         }
     }

package/dist/workflows/dag/init-hybrid.js

@@ -34,6 +34,7 @@ import { buildFrontendTestOutcomeGateShellSnippet } from "./frontend-test-result
 import { classifyFrontendRisk, } from "./frontend-risk.js";
 import { discoverFrontendProjectCapability, } from "./frontend-project-capability.js";
 import { buildFrontendImplementationContractSkeleton, FRONTEND_IMPLEMENTATION_CONTRACT_SCHEMA_ID, FRONTEND_IMPLEMENTATION_CONTRACT_PLAN_PATCH_SCHEMA_ID, loadFrontendImplementationContractJsonSchema, } from "./frontend-implementation-contract.js";
+import { FRONTEND_NO_VERIFICATION_MARKER_TEXT } from "./frontend-verification-trace.js";
 import { serializeDagTaskSourcePath } from "../../task/dag-source-paths.js";
 const REQUIREMENT_FILE = "需求.md";
 const CONSTRAINT_FILE = "执行约束.md";
@@ -1281,11 +1282,15 @@ function deriveParallelScoutPaths(taskConfig) {
                 : allowed,
     };
 }
+const FRONTEND_NO_STATIC_VERIFICATION_MARKER = `node -e "console.log('${FRONTEND_NO_VERIFICATION_MARKER_TEXT}; static/behavior verification not-run')"`;
 /**
  * Resolve frontend verification fallbacks from the target project's own
  * package scripts. The DAG builder is also used by unit fixtures without a
- * package.json, so those fixtures retain the historical generic fallback.
- * Real projects never inherit loop-agent's commands when package.json exists.
+ * repoRoot, so those fixtures retain the historical generic fallback. A real
+ * project (repoRoot set) without a readable package.json has no scripts to
+ * invoke; the generic fallback would only inject commands guaranteed to fail
+ * at verify time, so such projects fall through to the tsc probe and may end
+ * up with no fallback commands plus a generation-time advisory.
  */
 async function discoverFrontendFallbackVerifyCommands(repoRoot) {
     const genericFallback = {
@@ -1302,10 +1307,8 @@ async function discoverFrontendFallbackVerifyCommands(repoRoot) {
         }
     }
     catch {
-        return genericFallback;
+        // No readable manifest: leave scripts unset and keep probing.
     }
-    if (!scripts)
-        return genericFallback;
     let packageManager = "npm";
     for (const [lockfile, manager] of [
         ["pnpm-lock.yaml", "pnpm"],
@@ -1382,19 +1385,66 @@ function toDagSourcePath(sources, absolutePath) {
         absolutePath,
     });
 }
-function extractExplicitRequirementIds(...markdownInputs) {
+/**
+ * Declaration-line shapes that assert a requirement id: list items
+ * ("- AC-1: ..."), numbered/ordered acceptance entries ("1. AC-1 ...",
+ * "1、AC-1"), table rows ("| AC-1 | ..."), and explicit key/value or
+ * heading declarations ("AC-1：...", "id: AC-1"). Ids that merely appear
+ * inside narrative prose (e.g. a fixture run id mentioned in background:
+ * "夹具 run 2026-08-23-req-f03535aa 已存在") are incidental mentions, not
+ * declared requirements, and must not pollute sourceBinding.requirementIds.
+ */
+const REQUIREMENT_ID_DECLARATION_LINE = /^\s*(?:[-+*]\s+|\d+[.、)）]\s*|\|\s*)?(?:[-A-Z0-9]+\s*[::]\s*)?(?:\*\*)?(?:REQ|BR|AC)-/i;
+function isRequirementIdDeclarationLine(line) {
+    return REQUIREMENT_ID_DECLARATION_LINE.test(line);
+}
+function extractExplicitRequirementIds(requirementMarkdown, ...fallbackMarkdown) {
     const ids = [];
     const seen = new Set();
-    for (const markdown of markdownInputs) {
+    const collect = (markdown, declarationsOnly) => {
         if (!markdown)
-            continue;
-        for (const match of markdown.matchAll(/\b(?:REQ|BR|AC)-[A-Z0-9]+(?:-[A-Z0-9]+)*\b/gi)) {
+            return;
+        const source = declarationsOnly
+            ? markdown
+                .split(/\r?\n/)
+                .filter((line) => isRequirementIdDeclarationLine(line))
+                .join("\n")
+            : markdown;
+        for (const match of source.matchAll(/\b(?:REQ|BR|AC)-[A-Z0-9]+(?:-[A-Z0-9]+)*\b/gi)) {
             const id = match[0].toUpperCase();
             if (!seen.has(id)) {
                 seen.add(id);
                 ids.push(id);
             }
         }
+    };
+    // The requirement document itself is human-authored acceptance prose:
+    // only declaration lines declare requirements there. If the document has
+    // no declaration-shaped lines at all (minimal free-form tasks like
+    // "covers AC-001"), fall back to its full text so bindings never go
+    // missing for unstructured input. References and constraints are bound
+    // documents (acceptance yaml, analysis docs) whose ids are authoritative
+    // wherever they appear — keep full scan for them.
+    const declarationLines = requirementMarkdown
+        .split(/\r?\n/)
+        .filter((line) => isRequirementIdDeclarationLine(line));
+    collect(requirementMarkdown, declarationLines.length > 0);
+    for (const markdown of fallbackMarkdown) {
+        if (markdown)
+            collect(markdown, false);
+    }
+    return ids;
+}
+/** Extract ids from an explicit authoritative section (full scan inside it). */
+function extractSectionRequirementIds(sectionMarkdown) {
+    const ids = [];
+    const seen = new Set();
+    for (const match of sectionMarkdown.matchAll(/\b(?:REQ|BR|AC)-[A-Z0-9]+(?:-[A-Z0-9]+)*\b/gi)) {
+        const id = match[0].toUpperCase();
+        if (!seen.has(id)) {
+            seen.add(id);
+            ids.push(id);
+        }
     }
     return ids;
 }
@@ -1408,7 +1458,7 @@ export function extractTaskScopedRequirementIds(requirementMarkdown, ...fallback
     // Without it, preserve legacy full-scan behavior (requirement + references).
     const sectionMatch = requirementMarkdown.match(/(?:^|\n)##\s*Acceptance References\s*\n([\s\S]*?)(?=\n##\s+|\n#\s+|$)/i);
     if (sectionMatch?.[1]) {
-        const fromSection = extractExplicitRequirementIds(sectionMatch[1]);
+        const fromSection = extractSectionRequirementIds(sectionMatch[1]);
         if (fromSection.length > 0)
             return fromSection;
     }
@@ -2530,7 +2580,14 @@ async function buildFrontendHybridDagFromTask(sources) {
         return buildBlockedFrontendMockDag(frontendSources, readOnlyPaths, forbiddenPaths, globalConstraints, blockedReason);
     }
     const fallbackVerifyCommands = await discoverFrontendFallbackVerifyCommands(sources.repoRoot);
-    const staticFallbackCommands = fallbackVerifyCommands.staticCommands;
+    // The verification bundle schema requires at least one static command. When
+    // a real project exposes no usable verification command at all, run an
+    // explicit no-op marker instead of a command that is guaranteed to fail:
+    // the trace then records not-run honestly and the advisory asks the task
+    // to declare verification commands and regenerate.
+    const staticFallbackCommands = fallbackVerifyCommands.staticCommands.length > 0
+        ? fallbackVerifyCommands.staticCommands
+        : [FRONTEND_NO_STATIC_VERIFICATION_MARKER];
     const behaviorFallbackCommands = fallbackVerifyCommands.behaviorCommands;
     const parsedFrontendVerifyCommands = extractFrontendVerifyCommandsFromMarkdown({
         repoRoot: sources.repoRoot,
@@ -2629,6 +2686,11 @@ async function buildFrontendHybridDagFromTask(sources) {
         ...behaviorVerifyEvidence.commandLabels.map((command) => `  - ${JSON.stringify(command)}`),
     ].join("\n");
     const advisories = [];
+    if (!hasDeclaredFrontendVerification &&
+        fallbackVerifyCommands.staticCommands.length === 0 &&
+        fallbackVerifyCommands.behaviorCommands.length === 0) {
+        advisories.push("未发现可用的前端验证命令：目标项目没有可读取的 package.json scripts，也未探测到本地 TypeScript，verify 节点将没有静态/行为命令可执行。请在 task.json --verify 或执行约束.md 的验证约束中显式声明命令（例如 node --check src/app.js），然后重新生成 DAG。");
+    }
     if (frontendSourceMentionsMock(frontendSources) &&
         frontendMockStrategyMustBeNotNeeded(frontendSources)) {
         advisories.push("auto 模式已将 Mock 策略收窄为 not-needed：任务源提到接口/API/Mock 需求，但仓库无确认 Mock 能力或无确定性 Mock 验证命令。若项目规范要求 Mock，请声明 frontendMock.verifyCommands 或 policy:required 后重新生成 DAG。");
@@ -4762,6 +4824,13 @@ function buildFrontendTestHybridDag(sources) {
     };
     const declaredRequirementIds = buildDagSourceBinding(sources).requirementIds;
     const declaredAcIds = declaredRequirementIds.filter((id) => /^AC(?:-[A-Z0-9]+)+$/i.test(id));
+    // Optional project-local UI anchor ledger: only mention it in prompts when it
+    // exists so ledger-less projects keep generating without noise.
+    const hasUiAnchorsLedger = Boolean(sources.repoRoot &&
+        existsSync(path.join(sources.repoRoot, "testcase/frontend/rag/ui-anchors.md")));
+    const uiAnchorsLedgerInstruction = hasUiAnchorsLedger
+        ? "Also read testcase/frontend/rag/ui-anchors.md (UI anchor ledger). Every passing find assertion in generated cases must quote a ledger row whose 状态 is 有效 for the matching page x state section; rows marked 不可断言/失效 must not be used as passing assertions. When an AC names a control absent from the ledger section for that state, emit a blocked case note (blockedReason unique-control-unavailable) instead of exploratory find steps, and follow ledger 备注 alternatives (split-node short literals) exactly."
+        : "";
     const reviewMode = config.reviewMode;
     const blockingReview = reviewMode === "blocking";
     const strictOutcomeGate = config.strictOutcomeGate;
@@ -4802,6 +4871,8 @@ function buildFrontendTestHybridDag(sources) {
         "const caseIdRe=/^FE-[A-Za-z0-9][A-Za-z0-9-]*$/;",
         ,
         "const acIdRe=/^AC(?:-[A-Z0-9]+)+$/i;",
+        "const truncatedAcRe=/^AC-FE-[A-Z]+$/i;",
+        "const truncatedUndeclaredAcRe=/^(?:AC-\\d{3}|AC-FE-[A-Z]+)$/i;",
         "for(const c of manifest.cases){",
         " const id=c&&c.caseId||'?';",
         " if(typeof c.caseId!=='string'||!caseIdRe.test(c.caseId))issues.push({ruleId:'case-id-shape',caseId:id,detail:'caseId must be FE-<FEATURE>-<NNN>-<dimension>, never AC-FE-*'});",
@@ -4820,6 +4891,7 @@ function buildFrontendTestHybridDag(sources) {
         "   if(typeof ac!=='string'){issues.push({ruleId:'ac-id-shape',caseId:id,detail:String(ac)+' must look like AC-FE-001'});continue;}",
         "   if(/^FE-/i.test(ac)){issues.push({ruleId:'ac-id-is-case',caseId:id,detail:ac+' looks like caseId; acIds must be AC-*'});continue;}",
         "   if(!acIdRe.test(ac)){issues.push({ruleId:'ac-id-shape',caseId:id,detail:ac+' must look like AC-FE-001'});continue;}",
+        "   if(truncatedAcRe.test(ac)||(declaredAc.size===0&&truncatedUndeclaredAcRe.test(ac))){issues.push({ruleId:'ac-id-truncated',caseId:id,detail:ac+' is a truncated acceptance id (missing feature prefix or sequence number)'});continue;}",
         "   if(declaredAc.size>0&&!declaredAc.has(ac))issues.push({ruleId:'unknown-ac',caseId:id,detail:ac+' not in task sourceBinding.requirementIds'});",
         "  }",
         " }",
@@ -4851,6 +4923,8 @@ function buildFrontendTestHybridDag(sources) {
             `const declaredAc=new Set(${declaredAcIdsLiteral});`,
             "const seen=new Set(); const seenCasePath=new Set(); const seenEvidenceDir=new Set();",
             "const acIdRe=/^AC(?:-[A-Z0-9]+)+$/i;",
+            "const truncatedAcRe=/^AC-FE-[A-Z]+$/i;",
+            "const truncatedUndeclaredAcRe=/^(?:AC-\\d{3}|AC-FE-[A-Z]+)$/i;",
             "for(const c of manifest.cases){",
             " if(!c||typeof c.caseId!=='string'||!/^FE-[A-Za-z0-9][A-Za-z0-9-]*$/.test(c.caseId)) fail('case-id-shape','caseId must be FE-*, never AC-FE-*: '+String(c&&c.caseId));",
             " if(/^AC-/i.test(c.caseId)) fail('case-id-is-ac','caseId must not be an acceptance id: '+c.caseId);",
@@ -4858,7 +4932,7 @@ function buildFrontendTestHybridDag(sources) {
             " seen.add(c.caseId);",
             " if(typeof c.dimension!=='string'||!dims.has(c.dimension)) fail('invalid-dimension',String(c.dimension));",
             " if(!Array.isArray(c.acIds)||c.acIds.length===0||c.acIds.some(a=>typeof a!=='string'||!a.trim())) fail('ac-mapping','invalid acIds for '+c.caseId);",
-            " for(const ac of c.acIds){ if(!acIdRe.test(ac)) fail('ac-id-shape','acIds entry must be AC-* acceptance id, not caseId: '+ac); if(declaredAc.size>0&&!declaredAc.has(ac)) fail('unknown-ac',ac+' not in sourceBinding; repair generator input or AC list'); }",
+            " for(const ac of c.acIds){ if(!acIdRe.test(ac)) fail('ac-id-shape','acIds entry must be AC-* acceptance id, not caseId: '+ac); if(truncatedAcRe.test(ac)||(declaredAc.size===0&&truncatedUndeclaredAcRe.test(ac))) fail('ac-id-truncated','acIds entry is a truncated acceptance id (missing feature prefix or sequence number): '+ac); if(declaredAc.size>0&&!declaredAc.has(ac)) fail('unknown-ac',ac+' not in sourceBinding; repair generator input or AC list'); }",
             " c.casePath='testcase/frontend/cases/'+c.caseId+'.md';",
             " c.evidenceDir='testcase/frontend/evidence/'+c.caseId+'/';",
             " for(const k of ['casePath','evidenceDir']){ const v=c[k]; if(typeof v!=='string'||path.isAbsolute(v)||v.includes('..')) fail('unsafe-path',k+': '+String(v)); }",
@@ -4983,6 +5057,7 @@ function buildFrontendTestHybridDag(sources) {
             subtask_prompt: [
                 "Use skill playwright-cli-case-generator.",
                 "Read testcase/frontend/rag/standard-scenarios.v1.json and cover priority=must scenarios (or record GAP in coverage-map). Include ## 测试点 and ## 测试步骤 in each case.",
+                ...(uiAnchorsLedgerInstruction ? [uiAnchorsLedgerInstruction] : []),
                 "Read only testcase/frontend/rag/context.md, testcase/frontend/rag/coverage-map.md, and the draft case paths testcase/frontend/cases/FE-*.md, testcase/frontend/cases/index.md, and testcase/frontend/cases/manifest.draft.json. Write only those same draft paths. Do not write testcase/frontend/cases/manifest.json.",
                 "Generate Markdown cases, index.md and manifest.draft.json (schemaVersion 1; cases[] with caseId, casePath, dimension, acIds, evidenceDir).",
                 "HARD ID CONTRACT (do not confuse these):",
@@ -5142,7 +5217,7 @@ function buildFrontendTestHybridDag(sources) {
                 subtaskPromptTemplate: [
                     "Primary job: EXECUTE case {{case.caseId}} from {{case.casePath}} with skill playwright-cli (fresh Pi session; do not use /new). playwright-cli-only: never bare Playwright CLI/API/test runner; no fallback.",
                     "Use the structured playwright_cli tool for every browser action. Do not request or search for bash. Do not execute raw shell commands. Translate each playwright-cli line in the case Markdown into one playwright_cli tool call ({command, args?, timeoutSeconds?}).",
-                    "1) Read the concrete baseUrl from testcase/frontend/rag/context.md; it has already passed the environment shell safety/reachability gate. Start browser ONLY via playwright_cli command=open with args [--browser=chrome, <that-concrete-baseUrl>] (default session only; no -s=). 2) Dynamic refs: eX/eY in case Markdown are documentation placeholders, never tool args. Immediately before every structured playwright_cli call that references an element, parse the current real eNN from the immediately preceding latest snapshot and pass only that real eNN; never send literal `eX`/`eY`. A new snapshot invalidates prior refs, so never reuse stale refs. File outputs are canonical: screenshot args [--filename, final.png] (or [e5, --filename, final.png] for a real target), PDF args [--filename, final.pdf], and snapshot writes a file only with [--filename, snapshot.txt]; snapshot without filename is response-only. Never use --path, --output, --file, or any output path as a positional target. Follow case steps with snapshot before element refs using only playwright_cli. A passed case requires this same child receipt order: successful open → successful find → controller post-execution cleanup. Only successful find is a meaningful assertion; snapshot, goto, screenshot, request/console, click/fill and other ordinary interactions cannot establish passed authority. 3) Only when preflight or playwright_cli tool explicitly fails may you write blocked evidence (blockedReason playwright-cli-unavailable | frontend-base-url-unreachable); never invent CLI-unavailable solely because bash is absent. 4) For U/D: enforce current-user ownership / create-or-mock-or-blocked; never mutate other users' data.",
+                    "1) Read the concrete baseUrl from testcase/frontend/rag/context.md; it has already passed the environment shell safety/reachability gate. Start browser ONLY via playwright_cli command=open with args [--browser=chrome, <that-concrete-baseUrl>] (default session only; no -s=). 2) Dynamic refs: eX/eY in case Markdown are documentation placeholders, never tool args. Immediately before every structured playwright_cli call that references an element, parse the current real eNN from the immediately preceding latest snapshot and pass only that real eNN; never send literal `eX`/`eY`. A new snapshot invalidates prior refs, so never reuse stale refs. File outputs are canonical: screenshot args [--filename, final.png] (or [e5, --filename, final.png] for a real target), PDF args [--filename, final.pdf], and snapshot writes a file only with [--filename, snapshot.txt]; snapshot without filename is response-only. Never use --path, --output, --file, or any output path as a positional target. Follow case steps with snapshot before element refs using only playwright_cli. A passed case requires this same child receipt order: successful open → successful find → controller post-execution cleanup. Only successful find is a meaningful assertion; snapshot, goto, screenshot, request/console, click/fill and other ordinary interactions cannot establish passed authority. 3) Only when preflight or playwright_cli tool explicitly fails may you write blocked evidence (blockedReason playwright-cli-unavailable | frontend-base-url-unreachable); never invent CLI-unavailable solely because bash is absent. Unique-control early stop: when an AC names a specific control that is absent from the first post-navigation snapshot after reaching the required state (and no setup-required gate appeared), do ONE find with the case's literal as evidence; if it returns 0 matches, record status=blocked with blockedReason unique-control-unavailable citing that single find - do NOT re-explore with alternative literals, repeated snapshots, navigation detours, or rerun loops. 4) For U/D: enforce current-user ownership / create-or-mock-or-blocked; never mutate other users' data.",
                     "Always write {{case.evidenceDir}}execution.md and {{case.evidenceDir}}case-result.json with caseId={{case.caseId}}, status passed|failed|blocked, evidencePaths (relative under evidenceDir). blocked needs non-empty blockedReason. After writing, self-check the same contract; if self-check fails, rewrite both files as status=blocked blockedReason=invalid-evidence-shape (never leave missing/malformed evidence).",
                     "Business failed/blocked is a recorded result, not a node failure. Close browser via playwright_cli command=close. Return compact JSON (<=1200 chars): {caseId,status,evidencePaths,errorSummary,tokens}.",
                 ].join("\n\n"),
@@ -5165,7 +5240,7 @@ function buildFrontendTestHybridDag(sources) {
             "const manifestPath='testcase/frontend/cases/manifest.json';",
             "if(!fs.existsSync(manifestPath)){process.stdout.write(JSON.stringify({cases:[]}));process.exit(0);}",
             "const manifest=JSON.parse(fs.readFileSync(manifestPath,'utf8'));const cases=[];",
-            "for(const c of (manifest.cases||[])){const evidenceDir=(c.evidenceDir||('testcase/frontend/evidence/'+c.caseId+'/')).replace(/\\/+$/,'')+'/';const resultPath=path.join(evidenceDir,'case-result.json');const execPath=path.join(evidenceDir,'execution.md');let reason=null;let attempt=0;let missing=false;if(!fs.existsSync(resultPath)){missing=true;reason='missing-result-files';}else{try{const r=JSON.parse(fs.readFileSync(resultPath,'utf8'));attempt=Number(r.rerunAttempt||0)||0;if(r.status==='blocked')reason='blocked';if(!r.status){missing=true;reason='missing-result-files';}}catch(_){missing=true;reason='missing-result-files';}}if(!fs.existsSync(execPath)&&reason!=='blocked'){missing=true;reason=reason||'missing-result-files';}const should=(reason==='blocked'||missing)&&attempt<" + String(round) + ";if(should){cases.push({caseId:c.caseId,casePath:c.casePath||('testcase/frontend/cases/'+c.caseId+'.md'),evidenceDir,dimension:c.dimension||'core',acIds:c.acIds||[],rerunAttempt:attempt+1,reason:reason||'blocked'});}}",
+            "for(const c of (manifest.cases||[])){const evidenceDir=(c.evidenceDir||('testcase/frontend/evidence/'+c.caseId+'/')).replace(/\\/+$/,'')+'/';const resultPath=path.join(evidenceDir,'case-result.json');const execPath=path.join(evidenceDir,'execution.md');let reason=null;let attempt=0;let missing=false;if(!fs.existsSync(resultPath)){missing=true;reason='missing-result-files';}else{try{const r=JSON.parse(fs.readFileSync(resultPath,'utf8'));attempt=Number(r.rerunAttempt||0)||0;if(r.status==='blocked')reason='blocked';if(r.status==='failed')reason='failed-retry';if(!r.status){missing=true;reason='missing-result-files';}}catch(_){missing=true;reason='missing-result-files';}}if(!fs.existsSync(execPath)&&reason!=='blocked'){missing=true;reason=reason||'missing-result-files';}const should=(reason==='blocked'||reason==='failed-retry'||missing)&&attempt<" + String(round) + ";if(should){cases.push({caseId:c.caseId,casePath:c.casePath||('testcase/frontend/cases/'+c.caseId+'.md'),evidenceDir,dimension:c.dimension||'core',acIds:c.acIds||[],rerunAttempt:attempt+1,reason:reason||'blocked'});}}",
             `fs.mkdirSync('testcase/frontend/evidence',{recursive:true});fs.writeFileSync('testcase/frontend/evidence/${candidateArtifact}',JSON.stringify({schemaVersion:1,cases},null,2)+'\\n');process.stdout.write(JSON.stringify({cases}));`,
         ].join("");
         tasks.push({
@@ -5233,6 +5308,7 @@ function buildFrontendTestHybridDag(sources) {
                         "RERUN attempt {{case.rerunAttempt}} for {{case.caseId}} (reason={{case.reason}}). Rewrite the authoritative execution.md and case-result.json; its final status replaces the earlier case result.",
                         "Primary job: EXECUTE case {{case.caseId}} from {{case.casePath}} with skill playwright-cli (fresh Pi session). Use structured playwright_cli only; headless open.",
                         "Read the concrete baseUrl from testcase/frontend/rag/context.md and start via playwright_cli command=open with args [--browser=chrome, <that-concrete-baseUrl>]. Passed requires open → find → cleanup receipts.",
+                        "Unique-control early stop: if the first post-navigation snapshot lacks the control an AC names (and no setup-required gate appeared), one find returning 0 matches is sufficient evidence - write status=blocked blockedReason unique-control-unavailable instead of exploratory retries.",
                         "Always write {{case.evidenceDir}}execution.md and {{case.evidenceDir}}case-result.json with caseId, status, evidencePaths, rerunAttempt={{case.rerunAttempt}}. Write fixed sections `### 执行摘要` and `### 实际执行步骤` to execution.md when available.",
                     ].join("\n\n"),
                 },

package/harness.json

@@ -83,8 +83,8 @@
     "pi": {
       "description": "Pi 负责规划、评审、诊断；当 DAG toolProfile=write 时也可做有界写入。模型按复杂度三档配置，支持 provider/model 字符串或带 thinking 的对象；斜杠前为 Pi provider，后为 modelId，勿只写裸 modelId。",
       "LOW": "wizard-local/minimax-m3",
-      "MED": "wizard-local/grok-4.6",
-      "HIGH": "wizard-local/gpt-5.6-sol"
+      "MED": "wizard-local/minimax-m3",
+      "HIGH": "wizard-local/minimax-m3"
     }
   }
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tea-agent/loop-agent",
-  "version": "0.39.0-next.30",
+  "version": "0.39.0-next.31",
   "type": "module",
   "bin": {
     "loop-agent": "bin/loop-agent.js",

package/skills/fe-test-ui-scout/SKILL.md

@@ -0,0 +1,65 @@
+---
+name: fe-test-ui-scout
+description: 在写 frontend-test Wave PRD/回归目录之前，对真实运行的被测页面做只读锚点侦察，产出并维护 testcase/frontend/rag/ui-anchors.md 锚点台账。用 playwright-cli snapshot/find 逐字面验证命中数与唯一性，防止 PRD 验收标准引用不存在或不唯一的文案。触发词：frontend-test、锚点、台账、ui-anchors、侦察、写 PRD 前、snapshot/find 校准、Wave PRD。
+references:
+  - path: references/ledger-schema.md
+    required: true
+  - path: references/recon-protocol.md
+    required: true
+---
+
+# FE-Test UI Scout（锚点侦察）
+
+## 定位
+
+写 Wave PRD / 校准回归目录 AC **之前**的前置侦察。产出锚点台账
+`testcase/frontend/rag/ui-anchors.md`：每个「页面 × 状态」下真实可见、
+`find` 可命中的字面清单。PRD 的每条 `find "..."` 断言必须能在台账找到对应行；
+台账没有的字面禁止写进验收标准。
+
+本 skill **不生成用例、不执行用例、不点任何写控件**。生成由
+`playwright-cli-case-generator`（DAG generate 节点）负责，执行由
+`playwright-cli`（DAG execute 节点）负责；本 skill 只在 DAG 之外的主会话使用。
+
+## 输入
+
+- 被测 baseUrl（来自 `testcase/frontend/rag/context.md` 的 `environmentProbe=reachable` URL）
+- 目标 Wave 的 AC 清单（回归目录中本 Wave 范围的行）
+- 既有台账（增量更新，不整表重写）
+
+## 流程
+
+1. **读取范围**：从回归目录取本 Wave AC 的「页面 + 测试点 + 验收标准」，
+   提取所有候选字面（引号内文案、aria-label、控件名）。
+2. **逐状态侦察**：对每个涉及页面/状态组合：
+   `playwright-cli open --browser=chrome <baseUrl>/#/<route>` → `snapshot` →
+   逐候选 `find "<字面>"`，记录命中数与命中位置。
+3. **记录陷阱**：拆节点（`<kbd>`+`<span>`、文本+`<code>`）、同名多控件、
+   aria-label 与可见文本不一致、状态前置（no-session / has-session / empty、
+   setup-required）都要写进备注列。
+4. **增量写台账**：按 [references/ledger-schema.md](references/ledger-schema.md)
+   的表结构与分节规则追加/修订行，保留既有行历史（用备注标注失效原因，
+   不删行）。
+5. **报告**：侦察完成后输出小结（新增/修订/失效行数、命中率、给 PRD 的
+   可用锚点清单与禁用清单）。
+
+## 硬规则
+
+- 只读：不点提交/保存/发送类控件、不创建会话以外的任何状态（「新建空白
+  对话」CTA 允许点击一次以进入 has-session 状态，属于台账明确的状态夹具）。
+- 命中 0 或命中数 >1 且无法用状态/作用域区分的字面，标记 `不可断言`，
+  禁止进入 PRD `find` 断言。
+- 台账每行必须带验证时间与当时 baseUrl；页面改版后旧行标 `失效` 而非删除。
+- 不猜测：没跑过 `find` 的字面不进台账。
+- PRD 正文（含背景/夹具段）不得出现会被需求 id 扫描捕获的形态：大写或小写的 REQ/BR/AC 前缀 + 连字符 + hash/序号（如夹具任务 id 里的 req 段、或自示范用的「req 加连字符 hash」字面）。夹具引用只写日期 + 短 hash（如 f03535aa），规则说明文字用中文描述或代码块包裹仍可能被扫——一律避免连写。
+- PRD 验收标准只用**目录 AC id**（如 AC-FE-INS-005）作条目 id；不要写本地序号前缀（如 AC-003 加粗目录 id 的双 id 形态）——双 id 会被同时提取，生成器只引用其一，另一个落入 acceptanceCoverage.missing。
+
+## 输出
+
+- `testcase/frontend/rag/ui-anchors.md`（增量更新）
+- 侦察小结（对话内输出，不落盘；结论回写 PRD 草稿的锚点引用）
+
+## References
+
+- [references/ledger-schema.md](references/ledger-schema.md) — 台账表结构与分节规则
+- [references/recon-protocol.md](references/recon-protocol.md) — 逐状态侦察操作协议与陷阱清单