@tea-agent/loop-agent 0.39.0-beta.4 → 0.39.0-beta.6

package/dist/workflows/dag/frontend-test-standard-scenarios.js

@@ -1,5 +1,6 @@
 import { copyFile, mkdir, readFile, writeFile } from "node:fs/promises";
 import path from "node:path";
+import { defaultFrontendTestLayout, frontendTestStandardScenariosRel, } from "./frontend-test-layout.js";
 export const FRONTEND_TEST_STANDARD_SCENARIOS_FILENAME = "frontend-test-standard-scenarios.v1.json";
 export const FRONTEND_TEST_STANDARD_SCENARIOS_DEST = "testcase/frontend/rag/standard-scenarios.v1.json";
 const MINIMAL_STANDARD_SCENARIOS = {
@@ -47,7 +48,8 @@ async function readGovernanceRoot(workspaceRoot) {
     }
 }
 export async function copyFrontendTestStandardScenarios(input) {
-    const destRel = FRONTEND_TEST_STANDARD_SCENARIOS_DEST;
+    const layout = input.layout ?? defaultFrontendTestLayout();
+    const destRel = frontendTestStandardScenariosRel(layout);
     const destAbs = path.join(input.workspaceRoot, destRel);
     const candidates = listFrontendTestStandardScenarioCandidates({
         governanceRoot: await readGovernanceRoot(input.workspaceRoot),

package/dist/workflows/dag/frontend-typed-event-store.js

@@ -113,6 +113,18 @@ export const contractBlockingOwnerSchema = z.enum([
  * `finalize_plan` terminal fact.
  */
 export const PLAN_TERMINAL_FACT_KINDS = ["finalize_plan"];
+/**
+ * A+B: incremental payload records for `frontend-plan-pi` (requirements,
+ * verification targets, evidence gaps). Each is committed by one
+ * `record_plan_*` call so a large plan never exceeds a single model output
+ * budget; `assemblePlanPatchFromCommittedFacts` aggregates them from the
+ * ledger in commit order.
+ */
+export const PLAN_RECORD_FACT_KINDS = [
+    "plan-requirement",
+    "plan-verification-target",
+    "plan-evidence-gap",
+];
 /**
  * A+B (AC-009): typed issue category shared by review and design change
  * requests. The five-value enum replaces free-form issueCategory strings.
@@ -143,6 +155,7 @@ const ALL_TYPED_EVENT_FACT_KIND_VALUES = [
         ...FRONTEND_SHAPE_FACT_KINDS,
         ...CONTRACT_FACT_KINDS,
         ...PLAN_TERMINAL_FACT_KINDS,
+        ...PLAN_RECORD_FACT_KINDS,
     ]),
 ];
 export const typedEventFactKindSchema = z.enum(ALL_TYPED_EVENT_FACT_KIND_VALUES);

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

@@ -11,7 +11,7 @@ import { pathMatchesPattern } from "../../shared/git-progress.js";
 import { DEFAULT_OPENSPEC_GOVERNANCE_ROOT, DEFAULT_FRONTEND_SPEC_ROOTS, extractTaskSourceFrontendSpecPaths, } from "../../shared/openspec-spec.js";
 import { BASELINE_FORBIDDEN_PATHS } from "./governance-constants.js";
 import { buildDecisionEnvelopePromptContract } from "./decision-envelope.js";
-import { DEFAULT_READ_ONLY_PI_RETRY_POLICY, PLANNER_OUTPUT_LIMIT_RETRY_POLICY, PROTOCOL_AWARE_PI_RETRY_POLICY, BACKEND_TEST_WRITER_COMPLETENESS_RETRY_POLICY, WRITER_TRANSPORT_RETRY_POLICY, FRONTEND_PLAN_LADDER_RETRY_POLICY, TARGET_TEMPLATE_TRANSIENT_RETRY_PROFILE, TARGET_TEMPLATE_WRITER_TRANSPORT_RETRY_POLICY, isCanonicalFinalVerifyShellRetryCandidate, isSafeReadOnlyPiRetryCandidate, isTargetTemplateImplementPi, isWriterTransportRetryCandidate, } from "./retry-policy.js";
+import { DEFAULT_READ_ONLY_PI_RETRY_POLICY, PLANNER_OUTPUT_LIMIT_RETRY_POLICY, PROTOCOL_AWARE_PI_RETRY_POLICY, BACKEND_TEST_WRITER_COMPLETENESS_RETRY_POLICY, WRITER_TRANSPORT_RETRY_POLICY, FRONTEND_PLAN_LADDER_RETRY_POLICY, FRONTEND_REVIEW_TERMINAL_RETRY_POLICY, TARGET_TEMPLATE_TRANSIENT_RETRY_PROFILE, TARGET_TEMPLATE_WRITER_TRANSPORT_RETRY_POLICY, isCanonicalFinalVerifyShellRetryCandidate, isSafeReadOnlyPiRetryCandidate, isTargetTemplateImplementPi, isWriterTransportRetryCandidate, } from "./retry-policy.js";
 import { REVIEW_JSON_VERDICT_OUTPUT_PROTOCOL, REVIEW_VERDICT_OUTPUT_PROTOCOL, } from "./output-protocol.js";
 import { resolveAdapter } from "../../adapters/index.js";
 import { loadHarnessManifest } from "../../governance/harness.js";
@@ -28,6 +28,7 @@ import { resolveExecutorModelMatrices } from "../../executors/model-routing.js";
 import { normalizeTaskRequirementText, resolveTaskDagTemplateSelection, } from "./task-demand-routing.js";
 import { BACKEND_TEST_EXECUTION_DEFAULT_TEST_ROOT, buildBackendTestExecutionPreflightShellSnippet, } from "./backend-test-execution-contract.js";
 import { resolveBackendTestLayout, } from "./backend-test-layout.js";
+import { applyFrontendTestLayoutToText, resolveFrontendTestLayout, } from "./frontend-test-layout.js";
 import { buildBackendTestOutcomeGateShellSnippet } from "./backend-test-result-contract.js";
 import { buildBackendTestIntakeContext } from "./backend-test-intake-context.js";
 import { buildFrontendTestOutcomeGateShellSnippet } from "./frontend-test-result-contract.js";
@@ -2651,7 +2652,7 @@ async function buildFrontendHybridDagFromTask(sources) {
             '{"strategy":"native","productionDefaultOff":true,"activation":"VITE_ENABLE_MOCK=true","endpoints":[{"method":"GET","path":"/api/users","fixture":"mocks/fixtures/users.json","consumer":"src/api/users.ts"}]}',
             "",
             "### optional plan fields - GOOD (all optional; omit when absent):",
-            '{"implementationSteps":["confirm contract","sync tests"],"stylingStrategy":"reuse existing design tokens","dependencyPolicy":"no new runtime deps","residualRisks":["browser a11y not-run"],"realIntegrationGap":"FE-TEST owns live HTTP"}',
+            '{"stylingStrategy":"reuse existing design tokens","dependencyPolicy":"no new runtime deps","residualRisks":["browser a11y not-run"],"realIntegrationGap":"FE-TEST owns live HTTP"}',
             "",
             "### optional plan fields - BAD (present-but-empty strings are rejected):",
             '{"stylingStrategy":"","dependencyPolicy":""}  <-- REJECTED: optional string fields must be non-empty when present; omit them instead',
@@ -2676,8 +2677,8 @@ async function buildFrontendHybridDagFromTask(sources) {
         "- targets: files, routes, publicApiChanges",
         "- mockApi: strategy, productionDefaultOff, activation, endpoints[]",
         "- verificationTargets[]: id, type, commandLabel, file, symbol, requirementIds, uiStates",
-        "- designEvidence: source, paths, conflicts; evidenceGaps[]",
-        "- optional: implementationSteps[], stylingStrategy, uiComponentChoices[], dependencyPolicy, residualRisks[], realIntegrationGap",
+        "- designEvidence: source, paths, conflicts; evidenceGaps[] (optional)",
+        "- optional: stylingStrategy, uiComponentChoices[], dependencyPolicy, residualRisks[], realIntegrationGap",
         "- uiComponentChoices[]: purpose, component, decision (specified|reuse-existing|new), specReference { path, section, line } | null, rationale",
         "Do not require or read a separate plan prose section; the contract JSON is the only plan surface.",
     ].join("\n");
@@ -2982,13 +2983,14 @@ async function buildFrontendHybridDagFromTask(sources) {
                     retryOnInvalid: true,
                     skeleton: frontendContractSkeleton,
                 },
-                outputContract: "Short Markdown narrative plus incremental typed plan record tools (record_target_surface / record_component_choice / record_state_flow / record_data_flow / record_mock_api / record_design_deviation / record_dependency) and exactly one finalize_plan terminal call. finalize_plan assembles the canonical editable patch from committed facts (requirements / verificationTargets / evidenceGaps / implementationSteps / residualRisks / realIntegrationGap). Omit protected fields: schemaVersion, sourceBinding, riskLevel, targets.files, and mockApi.productionDefaultOff. The runtime compiles facts ⊕ skeleton into canonical full-contract JSON for downstream review. No file writes.",
+                outputContract: "Short Markdown narrative plus incremental typed plan record tools (record_target_surface / record_component_choice / record_state_flow / record_data_flow / record_mock_api / record_design_deviation / record_dependency / record_plan_requirement / record_plan_verification_target / record_plan_evidence_gap) and exactly one finalize_plan terminal call. finalize_plan assembles the canonical editable patch from the committed ledger facts (requirements / verificationTargets / evidenceGaps / residualRisks / realIntegrationGap). Commit requirements and verification targets incrementally — one entry per record call — so the plan never needs a single large output; evidence gaps are optional and only needed for genuine gaps. Implementation steps are NOT part of the plan — the implementer designs its own ordering. Omit protected fields: schemaVersion, sourceBinding, riskLevel, targets.files, and mockApi.productionDefaultOff. The runtime compiles facts ⊕ skeleton into canonical full-contract JSON for downstream review. No file writes.",
                 subtask_prompt: [
                     "Use frontend-contract-pi typed requirement facts (stable REQ/BR/AC identifiers, dispositions, evidence expectations, and frontend-test handoff intents), frontend-scout-pi, task sources, and the generation-time Mock capability evidence to fill the runtime-owned frontend contract skeleton. Record the plan decision ledger through the incremental record_* tools, then call finalize_plan exactly once. The runtime already owns schemaVersion, sourceBinding, riskLevel, targets.files, and mockApi.productionDefaultOff; omit those protected paths even when their values look obvious.",
+                    "Commit requirements with record_plan_requirement (one entry per call) and verification targets with record_plan_verification_target (one entry per call). Omit expectedOutcome in requirements — the runtime derives it from the contract; keep each entry to id + targets + verificationTargetIds. Evidence gaps are optional — call record_plan_evidence_gap only for genuine gaps. Never bundle these into finalize_plan arguments — finalize_plan only closes the plan. CRITICAL: emit exactly ONE record tool call per assistant message — never batch multiple record_* calls in the same message, even though they look independent. Parallel-batching them makes a single message as large as the old full-contract output and re-introduces the truncation bug. One call per message, many messages, then finalize_plan exactly once at the end. Do NOT plan implementation steps: the implementer decides ordering itself.",
                     ...(requiresOpenspecClassification ? ["Additionally consume the frontend-contract-pi openspec classification. Successfully read every required selection (including mandatory paths) and declare any used component specReference in the typed decision ledger; relevant/irrelevant selections are not forced into the ledger unless the plan actually uses them."] : []),
                     "The incremental record_* facts become the complete implementation plan after deterministic merge with the protected skeleton. Do not treat leftover JSON in the narrative as the compile authority.",
                     "Select the Mock / API strategy only through record_mock_api. Encode endpoint/fixture mapping, explicit activation, verification commands, and Real Integration Gap in schema-defined fields; productionDefaultOff comes from the protected skeleton and there is no second plan output.",
-                    "Encode ordered steps (implementationSteps), target files, UI state handling, styling/component strategy (stylingStrategy), interaction notes, Mock/API strategy, dependency policy (dependencyPolicy), deterministic verification entrypoints, Real Integration Gap (realIntegrationGap), and residual risks (residualRisks) into the ledger payload. Use only the fixed entrypoints below; implementation may add tests behind them but cannot replace them.",
+                    "Encode target files, UI state handling, styling/component strategy (stylingStrategy), interaction notes, Mock/API strategy, dependency policy (dependencyPolicy), deterministic verification entrypoints, Real Integration Gap (realIntegrationGap), and residual risks (residualRisks) into the ledger payload. Use only the fixed entrypoints below; implementation may add tests behind them but cannot replace them.",
                     "Every target file and verification target must be selected from the current target workspace and task scope. Do not reuse paths or symbols from examples, prior tasks, or loop-agent itself; if the project uses app/, packages/, spec/, __tests__, or another layout, preserve that layout.",
                     "Consume the Scout target surface and design evidence before selecting files. Preserve the discovered existing entrypoint and data source. If implementationPaths or testPaths are outside task allowedPaths, record a blocking scope conflict; do not substitute a new page or silently broaden the writeSet.",
                     "Output a short Markdown narrative, record the incremental facts, then call finalize_plan exactly once. Do NOT emit protected skeleton fields or a full contract as the authority.",
@@ -3234,6 +3236,7 @@ async function buildFrontendHybridDagFromTask(sources) {
                 executor: "pi",
                 complexity: "HIGH",
                 writePolicy: "read-only",
+                retryPolicy: FRONTEND_REVIEW_TERMINAL_RETRY_POLICY,
                 allowedPaths: readOnlyPaths,
                 forbiddenPaths,
                 skills: FRONTEND_REVIEW_SKILLS,
@@ -4860,11 +4863,57 @@ function applyBackendTestLayoutToDagSpec(spec, layout) {
     spec.successCriteria = spec.successCriteria?.map(rewrite);
     spec.backendTestLayout = layout;
 }
+/** Rewrite layout-dependent strings across a compiled frontend-test DAG spec. */
+function applyFrontendTestLayoutToDagSpec(spec, layout) {
+    if (layout.isDefault) {
+        spec.frontendTestLayout = layout;
+        return;
+    }
+    const rewrite = (value) => applyFrontendTestLayoutToText(value, layout);
+    for (const task of spec.tasks) {
+        task.subtask_prompt = rewrite(task.subtask_prompt);
+        if (task.outputContract)
+            task.outputContract = rewrite(task.outputContract);
+        task.writeSet = task.writeSet?.map(rewrite);
+        task.allowedPaths = task.allowedPaths?.map(rewrite);
+        task.forbiddenPaths = task.forbiddenPaths?.map(rewrite);
+        if (task.dynamicExpansion?.childTask) {
+            const child = task.dynamicExpansion.childTask;
+            if (typeof child.subtaskPromptTemplate === "string") {
+                child.subtaskPromptTemplate = rewrite(child.subtaskPromptTemplate);
+            }
+            if (typeof child.outputContract === "string") {
+                child.outputContract = rewrite(child.outputContract);
+            }
+            if (Array.isArray(child.allowedPaths)) {
+                child.allowedPaths = child.allowedPaths.map((entry) => typeof entry === "string" ? rewrite(entry) : entry);
+            }
+            if (Array.isArray(child.writeSet)) {
+                child.writeSet = child.writeSet.map((entry) => typeof entry === "string" ? rewrite(entry) : entry);
+            }
+            if (Array.isArray(child.forbiddenPaths)) {
+                child.forbiddenPaths = child.forbiddenPaths.map((entry) => typeof entry === "string" ? rewrite(entry) : entry);
+            }
+        }
+        if (task.shell?.commands) {
+            task.shell.commands = task.shell.commands.map(rewrite);
+        }
+    }
+    spec.globalConstraints = spec.globalConstraints?.map(rewrite);
+    spec.successCriteria = spec.successCriteria?.map(rewrite);
+    spec.frontendTestLayout = layout;
+}
+function allowedPathsCoverFrontendTestRoot(allowedPaths, testRoot) {
+    const probe = `${testRoot}/_layout_probe_`;
+    return allowedPaths.some((pattern) => pathMatchesPattern(testRoot, pattern) ||
+        pathMatchesPattern(probe, pattern));
+}
 // ---------------------------------------------------------------------------
 // Frontend browser-test RAG DAG template
 // ---------------------------------------------------------------------------
 function buildFrontendTestHybridDag(sources) {
     const rawFrontendTest = sources.taskConfig.frontendTest;
+    const layout = resolveFrontendTestLayout(rawFrontendTest);
     const config = {
         // Default 32: common FE suites cover ~24 AC with multi-dimension cases; 20 caused map maxExpandedNodes failures.
         maxCasesPerBatch: rawFrontendTest?.maxCasesPerBatch ?? 32,
@@ -4889,9 +4938,9 @@ function buildFrontendTestHybridDag(sources) {
     // 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")));
+        existsSync(path.join(sources.repoRoot, layout.ragDir, "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."
+        ? `Also read ${layout.ragDir}/ui-anchors.md (UI anchor ledger). Every passing find assertion in generated cases must quote a ledger row whose 状态 is 有效 for the matching page x state section; rows marked 不可断言/失效 must not be used as passing assertions. When an AC names a control absent from the ledger section for that state, emit a blocked case note (blockedReason unique-control-unavailable) instead of exploratory find steps, and follow ledger 备注 alternatives (split-node short literals) exactly.`
         : "";
     const reviewMode = config.reviewMode;
     const blockingReview = reviewMode === "blocking";
@@ -4899,22 +4948,19 @@ function buildFrontendTestHybridDag(sources) {
     const maxRerunAttempts = config.maxRerunAttempts;
     const enableRetrospect = config.reports?.retrospect === true;
     const enableL5Report = config.reports?.l5 !== false;
-    const hasFrontendTestWriteScope = sources.taskConfig.allowedPaths.some((pattern) => pattern === "testcase/frontend/**" ||
-        pattern === "testcase/**" ||
-        pattern === "**");
-    if (!hasFrontendTestWriteScope) {
-        throw new Error('frontend-test requires task.json allowedPaths to include "testcase/frontend/**" (or an explicit containing glob).');
+    if (!allowedPathsCoverFrontendTestRoot(sources.taskConfig.allowedPaths, layout.testRoot)) {
+        throw new Error(`frontend-test requires task.json allowedPaths to cover "${layout.testRoot}/**" (or an explicit containing glob).`);
     }
     const forbidden = commonForbiddenPaths(sources);
-    const ragWriteSet = ["testcase/frontend/rag/**"];
+    const ragWriteSet = [`${layout.ragDir}/**`];
     const caseDraftWriteSet = [
-        "testcase/frontend/cases/FE-*.md",
-        "testcase/frontend/cases/index.md",
-        "testcase/frontend/cases/manifest.draft.json",
+        `${layout.casesDir}/FE-*.md`,
+        `${layout.casesDir}/index.md`,
+        `${layout.casesDir}/manifest.draft.json`,
     ];
     // The shell materializer alone owns the final manifest boundary.
-    const casesWriteSet = ["testcase/frontend/cases/**"];
-    const evidenceRoot = "testcase/frontend/evidence";
+    const casesWriteSet = [`${layout.casesDir}/**`];
+    const evidenceRoot = layout.evidenceDir;
     const declaredAcIdsLiteral = JSON.stringify(declaredAcIds);
     const maxCasesPerBatchLiteral = String(config.maxCasesPerBatch);
     const checklistScript = [
@@ -5506,6 +5552,7 @@ function buildFrontendTestHybridDag(sources) {
         verifyStrategy: resolveDagVerifyStrategy(sources.taskConfig),
         tasks,
     };
+    applyFrontendTestLayoutToDagSpec(spec, layout);
     applyDefaultReadOnlyRetryPolicy(spec);
     parseDagSpec(spec);
     assertValidDagSpec(spec);

package/dist/workflows/dag/node-execution.js

@@ -228,6 +228,26 @@ function buildAttemptPrompt(task, basePrompt, attemptNumber, previousFailureCate
             buildProtocolRetryInstruction(task.outputProtocol, previousProtocolReason),
         ].join("\n");
     }
+    if (previousFailureCategory === "review-terminal-missing") {
+        return [
+            basePrompt,
+            "",
+            "<retry_instruction>",
+            "The review emitted a verdict in response text but never committed the authoritative typed terminal tool call (approve_review / request_review_changes). The response text is NOT the authority: no branch or gate reads it.",
+            "Call exactly one typed terminal tool to finish: approve_review (implementation passes, no Critical/Important findings) or request_review_changes (with typed issueCategory, at least one evidenceRef, and non-empty findings). Do not repeat the review analysis; commit the terminal tool once and stop.",
+            "</retry_instruction>",
+        ].join("\n");
+    }
+    if (previousFailureCategory === "read-burst") {
+        return [
+            basePrompt,
+            "",
+            "<retry_instruction>",
+            "Previous plan attempt issued too many read-only tool calls (read/grep/ls/find) and blew up the context window. Trust the upstream frontend-contract-pi typed requirement facts and frontend-scout-pi target surface already provided — do NOT re-read contract/scout stdout, PRD/source files, or component sources you already inspected.",
+            "Minimize discovery reads: only read what you genuinely need, once. Commit record_plan_requirement / record_plan_verification_target / record_* facts directly from the facts already in context (one tool call per message), then call finalize_plan exactly once.",
+            "</retry_instruction>",
+        ].join("\n");
+    }
     if (previousFailureCategory === "invalid-output" &&
         task.structuredContractOutput &&
         previousProtocolReason) {
@@ -269,7 +289,7 @@ function buildAttemptPrompt(task, basePrompt, attemptNumber, previousFailureCate
                 "",
                 "<retry_instruction>",
                 "Previous attempt was truncated by the provider (stopReason=length) before the plan facts were fully committed.",
-                "Re-commit the missing record_* facts and call finalize_plan exactly once; the committed typed ledger is the only compile authority.",
+                "Re-commit the missing record_* facts and call finalize_plan exactly once; the committed typed ledger is the only compile authority. Do NOT re-read contract/scout outputs or source files — use the facts already in context. Commit record_* facts one tool call per message, then finalize_plan immediately.",
                 "Do not emit a fenced JSON contract, protected skeleton fields, or an openspec-citations block; narrative JSON is not validated and wastes the output budget.",
                 "</retry_instruction>",
             ].join("\n");

package/dist/workflows/dag/retry-policy.js

@@ -25,6 +25,21 @@ export const PROTOCOL_INVALID_RETRY_CATEGORY = "protocol-invalid";
 export const STRUCTURED_ARTIFACT_INVALID_RETRY_CATEGORY = "invalid-output";
 /** Provider stopReason=length truncated the response before the JSON contract completed. */
 export const STRUCTURED_OUTPUT_TRUNCATED_RETRY_CATEGORY = "structured-output-truncated";
+/**
+ * The review node emitted a verdict in response text (e.g. "VERDICT: pass")
+ * but never committed the authoritative typed terminal tool call
+ * (approve_review / request_review_changes). Read-only and safe to retry with
+ * a corrected instruction; the typed terminal is the only authority.
+ */
+export const REVIEW_TERMINAL_MISSING_RETRY_CATEGORY = "review-terminal-missing";
+/**
+ * The plan node issued too many read-only tool calls (read/grep/ls/find),
+ * blowing up the context window and eventually a 400 request-too-large. The
+ * plan should trust upstream typed facts instead of re-reading contract/scout
+ * outputs and source files repeatedly. Read-only and safe to retry with a
+ * reduced-reading instruction.
+ */
+export const READ_BURST_RETRY_CATEGORY = "read-burst";
 /** Retry only a proven no-op from an explicitly opt-in bounded Pi writer. */
 export const WRITER_EMPTY_DIFF_RETRY_CATEGORY = "writer-empty-diff";
 /** Retry when a backend-test writer finished but Completeness Gate found missing/broken targets. */
@@ -57,6 +72,8 @@ export const ALL_DAG_RETRY_CATEGORIES = [
     WRITER_EMPTY_DIFF_RETRY_CATEGORY,
     INCOMPLETE_WRITE_SET_RETRY_CATEGORY,
     WRITER_CLEAN_TIMEOUT_RETRY_CATEGORY,
+    REVIEW_TERMINAL_MISSING_RETRY_CATEGORY,
+    READ_BURST_RETRY_CATEGORY,
 ];
 const RETRY_SAFE_PI_ROLES = new Set([
     "planner",
@@ -143,6 +160,20 @@ export const PROTOCOL_AWARE_PI_RETRY_POLICY = {
     ...DEFAULT_READ_ONLY_PI_RETRY_POLICY,
     retryCategories: [...PROTOCOL_AWARE_DAG_RETRY_CATEGORIES],
 };
+/**
+ * frontend-review-pi retry policy. The review verdict must be committed
+ * through the typed terminal tools (approve_review / request_review_changes);
+ * a model that only echoes "VERDICT: pass" text fails as
+ * review-terminal-missing. That is read-only and safe to retry with a
+ * corrected instruction so a single model slip does not burn the whole run.
+ */
+export const FRONTEND_REVIEW_TERMINAL_RETRY_POLICY = {
+    ...DEFAULT_READ_ONLY_PI_RETRY_POLICY,
+    retryCategories: [
+        ...DEFAULT_DAG_RETRY_CATEGORIES,
+        REVIEW_TERMINAL_MISSING_RETRY_CATEGORY,
+    ],
+};
 /**
  * The sole writer retry policy for requireChangedFiles writers. It is
  * intentionally not included in any read-only default: a writer may retry only
@@ -501,5 +532,6 @@ export const FRONTEND_PLAN_LADDER_RETRY_POLICY = {
     retryCategories: [
         ...DEFAULT_DAG_RETRY_CATEGORIES,
         STRUCTURED_ARTIFACT_INVALID_RETRY_CATEGORY,
+        READ_BURST_RETRY_CATEGORY,
     ],
 };

package/dist/workflows/dag/types.js

@@ -1159,6 +1159,18 @@ export const dagSpecSchema = z
         isDefault: z.boolean(),
     })
         .optional(),
+    /** Frozen frontend-test artifact layout; absent = historical testcase/frontend/ layout. */
+    frontendTestLayout: z
+        .object({
+        testRoot: z.string(),
+        ragDir: z.string(),
+        casesDir: z.string(),
+        evidenceDir: z.string(),
+        reportsDir: z.string(),
+        fixturesDir: z.string(),
+        isDefault: z.boolean(),
+    })
+        .optional(),
     /** Bound user shared-setup document (plan B); absent = no shared setup references expected. */
     backendTestSharedSetup: z
         .object({

package/dist/workflows/dag/validate.js

@@ -5,6 +5,7 @@ import { resolveRepairTaskForGate } from "./repair-artifact.js";
 import { topoSortToRanks } from "./topo.js";
 import { isSafeReadOnlyPiRetryCandidate, isWriterEmptyDiffRetryCandidate, isWriterTransportRetryCandidate, INCOMPLETE_WRITE_SET_RETRY_CATEGORY, WRITER_CLEAN_TIMEOUT_RETRY_CATEGORY, WRITER_EMPTY_DIFF_RETRY_CATEGORY, } from "./retry-policy.js";
 import { isBackendTestCompletenessRetryCandidate } from "./backend-test-writer-completeness.js";
+import { defaultFrontendTestLayout, frontendTestLayoutFromSpec, } from "./frontend-test-layout.js";
 const GOVERNANCE_WARNING_TYPES = new Set([
     "read-only-missing-artifacts-forbidden",
     "read-only-prompt-mentions-artifact-writes",
@@ -690,22 +691,23 @@ function validateWriterOutcomePolicyTaskConfig(task, issues) {
         });
     }
 }
-function isFrontendEvidenceWriteSet(writeSet) {
+function isFrontendEvidenceWriteSet(writeSet, layout = defaultFrontendTestLayout()) {
     const entries = writeSet ?? [];
     if (entries.length === 0)
         return false;
+    const prefix = `${layout.evidenceDir}/`;
+    const glob = `${layout.evidenceDir}/**`;
     return entries.every((entry) => {
         const normalized = entry.trim().replace(/\\/g, "/").replace(/^\.\//, "");
-        return (normalized === "testcase/frontend/evidence/**" ||
-            normalized.startsWith("testcase/frontend/evidence/"));
+        return normalized === glob || normalized.startsWith(prefix);
     });
 }
-function isFrontendBrowserExecutorTask(task) {
+function isFrontendBrowserExecutorTask(task, layout = defaultFrontendTestLayout()) {
     const skills = task.skills ?? [];
     return (task.executor === "pi" &&
         task.toolProfile === "write" &&
         skills.includes("playwright-cli") &&
-        isFrontendEvidenceWriteSet(task.writeSet));
+        isFrontendEvidenceWriteSet(task.writeSet, layout));
 }
 function taskMentionsPlaywrightCliContract(task) {
     const prompt = `${task.subtask_prompt}\n${task.outputContract ?? ""}`;
@@ -717,10 +719,11 @@ function taskMentionsPlaywrightCliContract(task) {
  * - capability is granted without skill/prompt contract or outside evidence writeSet
  * - non-browser nodes receive command capability
  */
-function validateCommandPolicyTaskConfig(task, issues) {
+function validateCommandPolicyTaskConfig(task, issues, spec) {
+    const layout = frontendTestLayoutFromSpec(spec);
     const policy = resolveDagCommandPolicy(task.commandPolicy);
     const allowsPlaywright = dagCommandPolicyAllows(task.commandPolicy, "playwright-cli");
-    const isBrowserExecutor = isFrontendBrowserExecutorTask(task);
+    const isBrowserExecutor = isFrontendBrowserExecutorTask(task, layout);
     if (isBrowserExecutor && !allowsPlaywright) {
         issues.push({
             type: "invalid-command-policy",
@@ -751,10 +754,10 @@ function validateCommandPolicyTaskConfig(task, issues) {
                 message: `task ${task.id} playwright-cli command capability requires prompt/outputContract to reference playwright_cli`,
             });
         }
-        if (!isFrontendEvidenceWriteSet(task.writeSet)) {
+        if (!isFrontendEvidenceWriteSet(task.writeSet, layout)) {
             issues.push({
                 type: "invalid-command-policy",
-                message: `task ${task.id} playwright-cli command capability requires frontend evidence writeSet under testcase/frontend/evidence/`,
+                message: `task ${task.id} playwright-cli command capability requires frontend evidence writeSet under ${layout.evidenceDir}/`,
             });
         }
         if (task.toolProfile !== "write") {
@@ -765,7 +768,7 @@ function validateCommandPolicyTaskConfig(task, issues) {
         }
     }
 }
-function validateDynamicChildCommandPolicy(task, issues) {
+function validateDynamicChildCommandPolicy(task, issues, spec) {
     const expansion = task.dynamicExpansion;
     if (!expansion)
         return;
@@ -786,7 +789,7 @@ function validateDynamicChildCommandPolicy(task, issues) {
         writeSet: child.writeSet,
         outputContract: child.outputContract,
     };
-    validateCommandPolicyTaskConfig(synthetic, issues);
+    validateCommandPolicyTaskConfig(synthetic, issues, spec);
 }
 function validateDecisionGateTaskConfig(task, issues) {
     if (!task.decisionGate?.enabled) {
@@ -893,12 +896,13 @@ export function collectForbiddenExecutorIssues(spec, forbiddenExecutors) {
     }));
 }
 /** Revalidate a rendered dynamic child before it enters state or executes. */
-export function assertValidMaterializedDagTask(task) {
+export function assertValidMaterializedDagTask(task, parentSpec) {
     const issues = [];
-    // A minimal v2 carrier makes write-policy validation strict while avoiding
-    // unrelated parent graph/rank checks.  Dynamic templates are validated at
-    // init; this guards values introduced by {{item}} rendering.
-    const spec = { version: 2, tasks: [task] };
+    const spec = {
+        version: 2,
+        tasks: [task],
+        frontendTestLayout: parentSpec?.frontendTestLayout,
+    };
     validateTaskWritePolicy(task, spec, issues);
     validateShellTaskConfig(task, spec, issues);
     validateStaticTaskConfig(task, issues);
@@ -906,7 +910,7 @@ export function assertValidMaterializedDagTask(task) {
     validateRetryPolicyTaskConfig(task, issues);
     validateOutputProtocolTaskConfig(task, issues);
     validateWriterOutcomePolicyTaskConfig(task, issues);
-    validateCommandPolicyTaskConfig(task, issues);
+    validateCommandPolicyTaskConfig(task, issues, spec);
     if (issues.length > 0) {
         throw new Error(`invalid materialized dynamic child ${task.id}: ${issues.map((issue) => issue.message).join("; ")}`);
     }
@@ -962,8 +966,8 @@ export function validateDagSpec(spec) {
         validateRetryPolicyTaskConfig(task, issues);
         validateOutputProtocolTaskConfig(task, issues);
         validateWriterOutcomePolicyTaskConfig(task, issues);
-        validateCommandPolicyTaskConfig(task, issues);
-        validateDynamicChildCommandPolicy(task, issues);
+        validateCommandPolicyTaskConfig(task, issues, spec);
+        validateDynamicChildCommandPolicy(task, issues, spec);
         validateProjectGovernanceTaskConfig(task, spec, issues);
         validatePiExtensionsTaskConfig(task, issues);
         validateFailureAwareDependsOn(task, spec, issues);

package/docs/templates/frontend-test-case-checklist.md

@@ -15,7 +15,7 @@ LLM review (when `frontendTest.reviewMode=blocking`) must not invent blocking ru
 | `unknown-ac` | When task `sourceBinding.requirementIds` lists ACs, every `acIds` entry must be in that set |
 | `case-id-shape` | `caseId` matches `FE-<FEATURE>-<NNN>-...` (never `AC-FE-*`) |
 | `case-id-is-ac` | Do not use acceptance id as `caseId` / filename |
-| `case-path-mismatch` | `casePath === testcase/frontend/cases/{caseId}.md` |
+| `case-path-mismatch` | `casePath === {casesDir}/{caseId}.md`（默认 `testcase/frontend/cases/{caseId}.md`） |
 | `case-file-missing` | `casePath` exists |
 
 ## Tool guidance (non-blocking)
@@ -40,5 +40,5 @@ LLM review (when `frontendTest.reviewMode=blocking`) must not invent blocking ru
 
 ## Pipeline vs quality
 
-- **Pipeline acceptance**: final `testcase/frontend/reports/frontend-test-retrospect-*.md` exists after result materialize
+- **Pipeline acceptance**: final `{reportsDir}/frontend-test-retrospect-*.md` exists after result materialize (default `testcase/frontend/reports/`)
 - **Quality**: `frontend-test-result-v1.outcome=passed` with 0 blocked/failed (opt-in via `frontendTest.strictOutcomeGate`)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tea-agent/loop-agent",
-  "version": "0.39.0-beta.4",
+  "version": "0.39.0-beta.6",
   "type": "module",
   "bin": {
     "loop-agent": "bin/loop-agent.js",

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

@@ -1,6 +1,6 @@
 ---
 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。
+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 {testRoot}/rag/ui-anchors.md 锚点台账（默认 testcase/frontend/rag/ui-anchors.md）。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
@@ -13,7 +13,7 @@ references:
 ## 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、
+`{testRoot}/rag/ui-anchors.md`（默认 `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。
 
@@ -23,7 +23,7 @@ references:
 
 ## 输入
 
-- 被测 baseUrl（来自 `testcase/frontend/rag/context.md` 的 `environmentProbe=reachable` URL）
+- 被测 baseUrl（来自 `{testRoot}/rag/context.md` 的 `environmentProbe=reachable` URL；默认 `testcase/frontend/rag/context.md`）
 - 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）
 
@@ -56,7 +56,7 @@ references:
 
 ## 输出
 
-- `testcase/frontend/rag/ui-anchors.md`（增量更新）
+- `{testRoot}/rag/ui-anchors.md`（增量更新；默认 `testcase/frontend/rag/ui-anchors.md`）
 - 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

package/skills/fe-test-ui-scout/references/ledger-schema.md

@@ -2,7 +2,7 @@
 
 ## 文件位置与命名
 
-- 路径：`testcase/frontend/rag/ui-anchors.md`
+- 路径：`{testRoot}/rag/ui-anchors.md`（默认 `testcase/frontend/rag/ui-anchors.md`）
 - 编码：UTF-8，LF；表格用 GitHub Flavored Markdown。
 - 台账是**增量账本**：修订不删除历史行；失效行改状态列并注明原因。
 

package/skills/playwright-cli/SKILL.md

@@ -55,7 +55,7 @@ request
 
 ## 推荐执行流程
 
-使用 `testcase/frontend/rag/context.md` 中已由 environment shell 标记为 `reachable` 的非生产 `baseUrl` 和默认浏览器 session；不得创建 named session。交互前先获取 snapshot，并只使用 snapshot 中可见的 ref 或已知安全 locator。
+使用 `{testRoot}/rag/context.md`（默认 `testcase/frontend/rag/context.md`）中已由 environment shell 标记为 `reachable` 的非生产 `baseUrl` 和默认浏览器 session；不得创建 named session。交互前先获取 snapshot，并只使用 snapshot 中可见的 ref 或已知安全 locator。截图写入当前 case 的 `{evidenceDir}/<caseId>/`，canonical 文件名仍是 `final.png`。
 
 ```text
 playwright-cli open --browser=chrome http://localhost:5173

package/skills/playwright-cli-case-generator/SKILL.md

@@ -13,10 +13,12 @@ description: 根据 FE-test RAG 知识包生成可由受限 playwright_cli runti
 
 ## 输入与边界
 
-- 只读取 `testcase/frontend/rag/context.md`、`coverage-map.md` 与已有 `testcase/frontend/cases/`。
-- 若存在 `testcase/frontend/rag/ui-anchors.md` 锚点台账，必须一并读取并作为 UI 断言的唯一事实源；台账缺失时按 RAG 知识包生成，但不得引用台账外的具体控件字面（缺失信息标 `blocked`）。
-- 只写 `testcase/frontend/cases/**`；不得回读 PRD、读取 `.harness/`，或写
-  `testcase/frontend/evidence/**`。
+默认产物根是 `testcase/frontend/`。若任务配置了 `frontendTest.testRoot`，把下列路径中的 `testcase/frontend/` 整段替换为该根（不要只替换 `testcase/`）。
+
+- 只读取 `{testRoot}/rag/context.md`、`coverage-map.md` 与已有 `{testRoot}/cases/`。
+- 若存在 `{testRoot}/rag/ui-anchors.md` 锚点台账，必须一并读取并作为 UI 断言的唯一事实源；台账缺失时按 RAG 知识包生成，但不得引用台账外的具体控件字面（缺失信息标 `blocked`）。
+- 只写 `{testRoot}/cases/**`；不得回读 PRD、读取 `.harness/`，或写
+  `{testRoot}/evidence/**`。
 - 所有 API、字段限制、状态流转、数据来源、SLA、URL 与账号要求必须能在 RAG 知识包中追溯。
   缺失信息标记 `blocked` 或“需人工确认”，不得猜测。
 
@@ -38,10 +40,10 @@ description: 根据 FE-test RAG 知识包生成可由受限 playwright_cli runti
     （例：`FE-LOGIN-001-core`）。禁止把验收标准写成 caseId。
   - `acIds` = **验收标准 ID 列表**，形态 `AC-FE-*` / `AC-*`
     （例：`["AC-FE-001"]`）。禁止把用例 ID 放进 acIds。
-  - `casePath` 必须等于 `testcase/frontend/cases/<caseId>.md`；
-    `evidenceDir` 必须等于 `testcase/frontend/evidence/<caseId>/`。
+  - `casePath` 必须等于 `{casesDir}/<caseId>.md`（默认 `testcase/frontend/cases/<caseId>.md`）；
+    `evidenceDir` 必须等于 `{evidenceDir}/<caseId>/`（默认 `testcase/frontend/evidence/<caseId>/`）。
 - `manifest` 使用 `schemaVersion: 1`，每项只含 `caseId`、`casePath`、`dimension`、`acIds`、
-  `evidenceDir`。所有 ID、路径和 evidenceDir 必须唯一，并位于 `testcase/frontend/` 内。
+  `evidenceDir`。所有 ID、路径和 evidenceDir 必须唯一，并位于 `{testRoot}/` 内。
 - `index.md` 按功能点列出 case、维度、AC、数据依赖、API 映射和预期执行状态。
 
 每个 case 必须包含：
@@ -73,7 +75,7 @@ description: 根据 FE-test RAG 知识包生成可由受限 playwright_cli runti
 7. 明确的 UI/API 预期与数据清理结果；无法满足的环境或数据依赖必须写为 `blocked`。
 
 所有文件型输出均由 controller 绑定到执行节点提供的
-`testcase/frontend/evidence/<case-id>/` 工作目录。截图一律使用 canonical 语法
+`{evidenceDir}/<case-id>/` 工作目录（默认 `testcase/frontend/evidence/<case-id>/`）。截图一律使用 canonical 语法
 `playwright-cli screenshot --filename final.png`；如确有从最新 snapshot 解析出的真实元素 ref，写为
 `playwright-cli screenshot e5 --filename final.png`。`pdf` 必须写为
 `playwright-cli pdf --filename final.pdf`。`snapshot` 无 filename 时仅返回响应；需要文件时使用

package/dist/worker/console/static/assets/channel-BuqaGKAT.js

@@ -1 +0,0 @@
-import{ag as o,ah as n}from"./mermaid.core-C-D3ZRUa.js";const t=(a,r)=>o.lang.round(n.parse(a)[r]);export{t as c};

package/dist/worker/console/static/assets/classDiagram-JCYQIIEL-DNdSHjH9.js

@@ -1 +0,0 @@
-import{s as a,c as s,a as e,C as t}from"./chunk-GF5L2VYU-DB2G4H8V.js";import{_ as i}from"./mermaid.core-C-D3ZRUa.js";import"./chunk-5VM5RSS4-SLBibhf5.js";import"./chunk-XXDRQBXY-BKkkC4VA.js";import"./chunk-KBJHAD2P-CG78FGkn.js";import"./chunk-2GRJ4B5K-DF_YSvOS.js";import"./index-D5zGX6fG.js";var n={parser:e,get db(){return new t},renderer:s,styles:a,init:i(r=>{r.class||(r.class={}),r.class.arrowMarkerAbsolute=r.arrowMarkerAbsolute},"init")};export{n as diagram};

package/dist/worker/console/static/assets/classDiagram-v2-OCEON4UE-DNdSHjH9.js

@@ -1 +0,0 @@
-import{s as a,c as s,a as e,C as t}from"./chunk-GF5L2VYU-DB2G4H8V.js";import{_ as i}from"./mermaid.core-C-D3ZRUa.js";import"./chunk-5VM5RSS4-SLBibhf5.js";import"./chunk-XXDRQBXY-BKkkC4VA.js";import"./chunk-KBJHAD2P-CG78FGkn.js";import"./chunk-2GRJ4B5K-DF_YSvOS.js";import"./index-D5zGX6fG.js";var n={parser:e,get db(){return new t},renderer:s,styles:a,init:i(r=>{r.class||(r.class={}),r.class.arrowMarkerAbsolute=r.arrowMarkerAbsolute},"init")};export{n as diagram};