@tea-agent/loop-agent 0.39.0-beta.5 → 0.39.0-beta.7

package/dist/workflows/dag/report.js

@@ -125,6 +125,8 @@ const dagNodeArtifactsReportSchema = z
     humanEscalationMd: dagArtifactRefSchema,
     humanApproval: dagArtifactRefSchema,
     humanRejection: dagArtifactRefSchema,
+    changeManifest: dagArtifactRefSchema,
+    changeManifestChangedFiles: z.array(z.string()).optional(),
 })
     .strict();
 const dagNodeReportRowSchema = z
@@ -259,6 +261,20 @@ export async function probeArtifact(filePath) {
 export async function buildNodeArtifactsReport(runDir, nodeId) {
     const nodeDir = path.join(runDir, nodeId);
     const artifact = (name) => probeArtifact(path.join(nodeDir, name));
+    const changeManifestRef = await artifact("change-manifest.json");
+    let changeManifestChangedFiles;
+    if (changeManifestRef.exists) {
+        try {
+            const manifest = JSON.parse(await readFile(path.join(nodeDir, "change-manifest.json"), "utf-8"));
+            changeManifestChangedFiles = Array.isArray(manifest.changedFiles)
+                ? manifest.changedFiles
+                    .filter((item) => typeof item === "string")
+                : undefined;
+        }
+        catch {
+            // unreadable manifest → leave undefined
+        }
+    }
     return {
         resultSummary: await artifact("result.summary.md"),
         promptRedacted: await artifact("prompt.redacted.md"),
@@ -270,6 +286,8 @@ export async function buildNodeArtifactsReport(runDir, nodeId) {
         humanEscalationMd: await artifact("human-escalation.md"),
         humanApproval: await artifact("human-approval.json"),
         humanRejection: await artifact("human-rejection.json"),
+        changeManifest: changeManifestRef,
+        changeManifestChangedFiles,
     };
 }
 function taskById(spec) {
@@ -726,6 +744,12 @@ function formatArtifactFlags(artifacts) {
         flags.push("approval");
     if (artifacts.humanRejection.exists)
         flags.push("rejection");
+    if (artifacts.changeManifest.exists) {
+        const count = artifacts.changeManifestChangedFiles?.length;
+        flags.push(count !== undefined
+            ? `changed:${count}`
+            : "change-manifest");
+    }
     return flags.length > 0 ? flags.join(",") : "-";
 }
 function formatTimestamp(value) {
@@ -749,7 +773,19 @@ function summarizeNodeStatuses(nodes) {
     return `${total} total / ${finished} finished / ${failed} failed / ${skipped} skipped / ${pending} pending`;
 }
 function hasArtifact(artifacts) {
-    return Object.values(artifacts).some((artifact) => artifact.exists);
+    return [
+        artifacts.resultSummary,
+        artifacts.promptRedacted,
+        artifacts.skills,
+        artifacts.nodeRecord,
+        artifacts.decisionEnvelope,
+        artifacts.decisionEnvelopeValidation,
+        artifacts.humanEscalationJson,
+        artifacts.humanEscalationMd,
+        artifacts.humanApproval,
+        artifacts.humanRejection,
+        artifacts.changeManifest,
+    ].some((artifact) => artifact.exists);
 }
 function failureRows(run) {
     return run.nodes.filter((node) => node.status === "ERROR" ||

package/dist/workflows/dag/rerun-feedback.js

@@ -2,6 +2,7 @@ import { createHash } from "node:crypto";
 import { readFile } from "node:fs/promises";
 import path from "node:path";
 import { parseJsonReviewVerdict } from "./output-protocol.js";
+import { FRONTEND_REVIEW_NODE_ID, readCommittedReviewRequestFact, } from "./frontend-review-findings.js";
 export const MAX_RERUN_FEEDBACK_CHARS = 12_000;
 const RERUN_FEEDBACK_TRUNCATION_MARKER = "\n\n...[上一轮反馈已按长度上限截断]...\n\n";
 const MAX_EVIDENCE_REFS = 8;
@@ -35,6 +36,44 @@ function requestRevision(text) {
         return (emphasized?.[1] ?? trimmed).trim() === "VERDICT: request-revision";
     });
 }
+/**
+ * Deterministically render a typed `request_review_changes` fact into feedback
+ * text. Unlike the legacy text channel (VERDICT: request-revision), the
+ * frontend review terminal is a committed typed fact; the findings are the
+ * actionable body the next round must consume.
+ */
+export function formatTypedReviewRequestFeedback(fact) {
+    const lines = [];
+    lines.push(`frontend review committed request_review_changes (issueCategory=${typeof fact.issueCategory === "string" ? fact.issueCategory : "unknown"})`);
+    if (Array.isArray(fact.findings) && fact.findings.length > 0) {
+        lines.push("review findings:");
+        for (const finding of fact.findings) {
+            if (!finding || typeof finding !== "object")
+                continue;
+            const record = finding;
+            const severity = typeof record.severity === "string" ? record.severity : "";
+            const issue = typeof record.issue === "string" ? record.issue : "";
+            const location = [
+                typeof record.file === "string" ? record.file : "",
+                typeof record.line === "number" ? String(record.line) : "",
+            ]
+                .filter(Boolean)
+                .join(":");
+            const prefix = location ? `[${location}] ` : "";
+            const requiredChange = typeof record.requiredChange === "string" &&
+                record.requiredChange.length > 0
+                ? ` (requiredChange: ${record.requiredChange})`
+                : "";
+            lines.push(`- ${prefix}${severity}: ${issue}${requiredChange}`);
+        }
+    }
+    if (Array.isArray(fact.evidenceRefs) && fact.evidenceRefs.length > 0) {
+        lines.push(`evidenceRefs: ${fact.evidenceRefs
+            .filter((ref) => typeof ref === "string")
+            .join(", ")}`);
+    }
+    return lines.join("\n");
+}
 function pathInside(root, candidate) {
     const relative = path.relative(root, candidate);
     return (relative === "" ||
@@ -127,6 +166,23 @@ function evidenceRefs(repoRoot, runDir, pass, sourceNodeId) {
     }
     return result;
 }
+/**
+ * Read the committed frontend review request fact (typed terminal channel).
+ * Returns { nodeId, fact } when the review node committed `request_review_changes`.
+ */
+async function readTypedReviewRequestFact(runDir) {
+    const fact = await readCommittedReviewRequestFact({ runDir });
+    if (!fact)
+        return undefined;
+    return {
+        nodeId: FRONTEND_REVIEW_NODE_ID,
+        fact: {
+            issueCategory: fact.issueCategory,
+            evidenceRefs: fact.evidenceRefs,
+            findings: fact.findings,
+        },
+    };
+}
 /**
  * Derive only the latest unresolved verify/review feedback from canonical
  * parent-run facts. Provider-only failures intentionally return undefined.
@@ -144,8 +200,52 @@ export async function deriveDagRerunFeedback(input) {
             break;
         }
     }
-    if (!sourceNodeId || !sourceNode)
+    if (!sourceNodeId || !sourceNode) {
+        // Typed terminal channel: the frontend review commits its verdict as a
+        // typed fact (request_review_changes) instead of VERDICT text, and the
+        // frontend DAG has no convergence passHistory. Fall back to reading the
+        // review node's typed event store so findings still flow into the next
+        // round as unresolved feedback.
+        const typedFact = await readTypedReviewRequestFact(input.runDir);
+        if (typedFact) {
+            const parentSourceBindingHash = sourceBindingHash(input.spec);
+            return withDigest({
+                schemaVersion: 1,
+                kind: "unresolved-terminal-feedback",
+                parentRunId: input.parentRunId,
+                taskId: input.taskId,
+                sourceNodeId: typedFact.nodeId,
+                verdict: "request-revision",
+                failureCategory: "review-request-changes",
+                feedbackText: scrubAndBoundFeedbackText(formatTypedReviewRequestFeedback(typedFact.fact), input.runDir, input.state.cwd),
+                evidenceRefs: [
+                    {
+                        nodeId: typedFact.nodeId,
+                        preservedNodeRecordPath: `${typedFact.nodeId}.json`,
+                    },
+                ],
+                attemptSummary: {
+                    completedPasses: 1,
+                    maxPasses: 1,
+                    ...(input.state.terminalReason
+                        ? { terminalReason: input.state.terminalReason }
+                        : {}),
+                },
+                binding: {
+                    ...(parentSourceBindingHash
+                        ? { sourceCanonicalHash: parentSourceBindingHash }
+                        : {}),
+                    ...(input.spec.taskContractBinding?.canonicalHash
+                        ? {
+                            taskContractCanonicalHash: input.spec.taskContractBinding.canonicalHash,
+                        }
+                        : {}),
+                    status: "unknown",
+                },
+            });
+        }
         return undefined;
+    }
     const completedPasses = pass?.pass ?? Math.max(1, input.state.convergence?.currentPass ?? 1);
     const maxPasses = Math.max(completedPasses, input.state.convergence?.maxPasses ?? 1);
     const parentSourceBindingHash = sourceBindingHash(input.spec);

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

@@ -32,6 +32,22 @@ export const STRUCTURED_OUTPUT_TRUNCATED_RETRY_CATEGORY = "structured-output-tru
  * 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";
+/**
+ * Provider returned a context-overflow error (400 request-too-large / context
+ * window exceeded). The SDK compacts and auto-retries once internally; when it
+ * still surfaces, the executor classifies this so retry policies and reports
+ * treat it as overflow (not empty-output) and can retry with a compacted
+ * session / fresh context.
+ */
+export const CONTEXT_OVERFLOW_RETRY_CATEGORY = "context-overflow";
 /** 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. */
@@ -65,6 +81,8 @@ export const ALL_DAG_RETRY_CATEGORIES = [
     INCOMPLETE_WRITE_SET_RETRY_CATEGORY,
     WRITER_CLEAN_TIMEOUT_RETRY_CATEGORY,
     REVIEW_TERMINAL_MISSING_RETRY_CATEGORY,
+    READ_BURST_RETRY_CATEGORY,
+    CONTEXT_OVERFLOW_RETRY_CATEGORY,
 ];
 const RETRY_SAFE_PI_ROLES = new Set([
     "planner",
@@ -187,7 +205,10 @@ export const WRITER_TRANSPORT_RETRY_POLICY = {
     backoff: "exponential",
     initialDelayMs: 5000,
     maxDelayMs: 30000,
-    retryCategories: [WRITER_CLEAN_TIMEOUT_RETRY_CATEGORY],
+    retryCategories: [
+        WRITER_CLEAN_TIMEOUT_RETRY_CATEGORY,
+        CONTEXT_OVERFLOW_RETRY_CATEGORY,
+    ],
 };
 /**
  * Implicit runtime retry for canonical final verify shells. Not a DagRetryPolicy:
@@ -210,14 +231,20 @@ export const TARGET_TEMPLATE_TRANSIENT_RETRY_PROFILE = "standard-reviewed-superv
 /**
  * Target-template exclusive implement-pi transport retry. Same schema category
  * as frontend (`writer-clean-timeout`) so DagSpec validation stays unchanged;
- * only the delay cap is wider. Do not mutate {@link WRITER_TRANSPORT_RETRY_POLICY}.
+ * only the delay cap is wider. Context overflow is added so a writer whose
+ * session hit the context window (compact+retry inside the SDK) can retry once
+ * more at the DAG level with a fresh attempt. Do not mutate
+ * {@link WRITER_TRANSPORT_RETRY_POLICY}.
  */
 export const TARGET_TEMPLATE_WRITER_TRANSPORT_RETRY_POLICY = {
     maxAttempts: 2,
     backoff: "exponential",
     initialDelayMs: 5000,
     maxDelayMs: 60000,
-    retryCategories: [WRITER_CLEAN_TIMEOUT_RETRY_CATEGORY],
+    retryCategories: [
+        WRITER_CLEAN_TIMEOUT_RETRY_CATEGORY,
+        CONTEXT_OVERFLOW_RETRY_CATEGORY,
+    ],
 };
 /**
  * Backend-test generation writers: empty-diff once, plus bounded incomplete-write-set
@@ -523,5 +550,6 @@ 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/runner.js

@@ -26,6 +26,7 @@ import { runConvergencePassController, } from "./convergence/controller.js";
 import { executeDagRanksOnce, FRONTEND_PREWRITE_RESULT_SOURCE_ARTIFACT, FRONTEND_WRITER_NODE_IDS, isConditionSkippedReason, readFrontendPrewriteResult, } from "./scheduler.js";
 import { topoSortToRanks } from "./topo.js";
 import { isFrontendWriterTransientPartialWrite } from "./frontend-writer-recovery.js";
+import { classifyReviewIssueCategoryFailureOwner, readCommittedReviewRequestFact, writeFrontendReviewFindingsPrd, } from "./frontend-review-findings.js";
 import { bindFrontendShapeStateToRunTopology, commitFrontendShapeTransitionCas, commitRuntimeFrontendShapeFact, isCanonicalFrontendShapeDag, } from "./frontend-shape-facts.js";
 import { assertFrontendTopologyBinding, buildFrontendTopologyBinding, computeFrontendShapeSourceDigest, } from "./frontend-shape.js";
 import { ensureFrontendShapeCapsuleRecord } from "./frontend-shape-capsule-store.js";
@@ -1151,6 +1152,44 @@ export async function finalizeTerminalRunStatus(state, taskCount, runDir, cwd) {
             }
         }
     }
+    // Phase 7: frontend review "request changes" → not passed. When the review
+    // node FINISHED but committed `request_review_changes` (typed terminal
+    // verdict), the implementation was reviewed and rejected. The run must not
+    // surface as finished: terminal aggregation below stays partial_failed, a
+    // terminal reason is recorded, the committed findings are rendered into a
+    // downstream-task PRD draft (run-owned), and a recovery trigger carries the
+    // issueCategory-derived owner so a later DAG round consumes the findings
+    // instead of repeating the previous flow.
+    const reviewRequestFact = await readCommittedReviewRequestFact({
+        runDir,
+    });
+    if (reviewRequestFact) {
+        state.terminalReason = "review-request-changes";
+        try {
+            await writeFrontendReviewFindingsPrd({
+                runDir,
+                parentRunId: state.runId,
+                fact: reviewRequestFact,
+            });
+        }
+        catch (error) {
+            // The PRD draft is additive; a write failure must not mask the
+            // not-passed verdict. The terminal aggregation below still applies.
+            console.warn(`[run-dag] warning: failed to write frontend review findings PRD: ${error instanceof Error ? error.message : String(error)}`);
+        }
+        if (!recoveryTrigger && !skipFrontendRecovery) {
+            const continuationCount = state.frontendRecoveryState?.continuationCount ?? 0;
+            if (continuationCount < 1) {
+                recoveryTrigger = {
+                    requestId: state.runId,
+                    failureSource: "frontend-review-pi",
+                    failureOwner: classifyReviewIssueCategoryFailureOwner(reviewRequestFact.issueCategory),
+                    protocolFailureReason: `frontend review request_review_changes (issueCategory=${reviewRequestFact.issueCategory})`,
+                    recoveryMode: "new-dag",
+                };
+            }
+        }
+    }
     const finishedCount = Object.values(state.nodes).filter((n) => n.status === "FINISHED").length;
     const { supersededIds, supersededFailures } = collectSupersededIntermediateFailures(state);
     if (state.convergence && supersededFailures.length > 0) {
@@ -1166,7 +1205,13 @@ export async function finalizeTerminalRunStatus(state, taskCount, runDir, cwd) {
     if (effectiveErrorCount === 0 &&
         successTerminal &&
         finishedCount + conditionSkippedCount + supersededErrorCount === taskCount) {
-        state.status = "finished";
+        // Phase 7 guard: a committed `request_review_changes` means the review
+        // rejected the implementation; the run must never surface as finished
+        // even though every node technically FINISHED.
+        state.status =
+            state.terminalReason === "review-request-changes"
+                ? "partial_failed"
+                : "finished";
     }
     else if (finishedCount > 0) {
         state.status = "partial_failed";

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

@@ -3,8 +3,9 @@ import { resolveShellCommands } from "../../executors/shell-executor.js";
 import { pathMatchesPattern } from "../../shared/git-progress.js";
 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 { isSafeReadOnlyPiRetryCandidate, isWriterEmptyDiffRetryCandidate, isWriterTransportRetryCandidate, INCOMPLETE_WRITE_SET_RETRY_CATEGORY, WRITER_CLEAN_TIMEOUT_RETRY_CATEGORY, WRITER_EMPTY_DIFF_RETRY_CATEGORY, CONTEXT_OVERFLOW_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",
@@ -655,12 +656,18 @@ function validateRetryPolicyTaskConfig(task, issues) {
     }
     if (isWriterTransportCandidate) {
         const categories = new Set(task.retryPolicy.retryCategories);
+        const allowedWriterTransportCategories = new Set([
+            WRITER_CLEAN_TIMEOUT_RETRY_CATEGORY,
+            CONTEXT_OVERFLOW_RETRY_CATEGORY,
+        ]);
+        const allAllowed = [...categories].every((category) => allowedWriterTransportCategories.has(category));
         if (task.retryPolicy.maxAttempts !== 2 ||
-            categories.size !== 1 ||
+            categories.size === 0 ||
+            !allAllowed ||
             !categories.has(WRITER_CLEAN_TIMEOUT_RETRY_CATEGORY)) {
             issues.push({
                 type: "invalid-retry-policy",
-                message: `task ${task.id} writer transport retryPolicy must use exactly two total attempts and only ${WRITER_CLEAN_TIMEOUT_RETRY_CATEGORY}`,
+                message: `task ${task.id} writer transport retryPolicy must use exactly two total attempts and only writer-clean-timeout (plus optional context-overflow)`,
             });
         }
     }
@@ -690,22 +697,23 @@ function validateWriterOutcomePolicyTaskConfig(task, issues) {
         });
     }
 }
-function isFrontendEvidenceWriteSet(writeSet) {
+function isFrontendEvidenceWriteSet(writeSet, layout = defaultFrontendTestLayout()) {
     const entries = writeSet ?? [];
     if (entries.length === 0)
         return false;
+    const prefix = `${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 +725,11 @@ 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 +760,10 @@ 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 +774,7 @@ function validateCommandPolicyTaskConfig(task, issues) {
         }
     }
 }
-function validateDynamicChildCommandPolicy(task, issues) {
+function validateDynamicChildCommandPolicy(task, issues, spec) {
     const expansion = task.dynamicExpansion;
     if (!expansion)
         return;
@@ -786,7 +795,7 @@ 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 +902,13 @@ 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 +916,7 @@ 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 +972,8 @@ 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/agent-dag-report.schema.json

@@ -176,7 +176,8 @@
         "humanEscalationJson",
         "humanEscalationMd",
         "humanApproval",
-        "humanRejection"
+        "humanRejection",
+        "changeManifest"
       ],
       "properties": {
         "resultSummary": { "$ref": "#/$defs/dagArtifactRef" },
@@ -188,7 +189,12 @@
         "humanEscalationJson": { "$ref": "#/$defs/dagArtifactRef" },
         "humanEscalationMd": { "$ref": "#/$defs/dagArtifactRef" },
         "humanApproval": { "$ref": "#/$defs/dagArtifactRef" },
-        "humanRejection": { "$ref": "#/$defs/dagArtifactRef" }
+        "humanRejection": { "$ref": "#/$defs/dagArtifactRef" },
+        "changeManifest": { "$ref": "#/$defs/dagArtifactRef" },
+        "changeManifestChangedFiles": {
+          "type": "array",
+          "items": { "type": "string" }
+        }
       }
     },
     "repairArtifact": {

package/docs/templates/agent-dag.schema.json

@@ -466,7 +466,7 @@
         "retryCategories": {
           "type": "array",
           "items": {
-            "enum": ["timeout", "network", "rate-limit", "unavailable", "empty-output", "output-too-large", "protocol-invalid", "invalid-output", "writer-empty-diff", "incomplete-write-set", "writer-clean-timeout"]
+            "enum": ["timeout", "network", "rate-limit", "unavailable", "empty-output", "output-too-large", "protocol-invalid", "invalid-output", "writer-empty-diff", "incomplete-write-set", "writer-clean-timeout", "read-burst", "context-overflow"]
           },
           "default": ["timeout", "network", "rate-limit", "unavailable", "empty-output"],
           "description": "Failure categories eligible for retry. quota/auth/write-guard are never eligible. invalid-output is allowed only on explicit structured/protocol policies; writer-empty-diff and incomplete-write-set are reserved for requireChangedFiles generation/completeness writers; writer-clean-timeout is reserved for standard bounded writers after explicit zero-write-tool-call and empty-diff evidence."

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.5",
+  "version": "0.39.0-beta.7",
   "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