@tea-agent/loop-agent 0.38.3 → 0.38.4

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

@@ -11,7 +11,7 @@ import { pathMatchesPattern } from "../../shared/git-progress.js";
 import { extractTaskSourceOpenspecPaths } 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, STRUCTURED_REQUIRED_PI_RETRY_POLICY, BACKEND_TEST_WRITER_COMPLETENESS_RETRY_POLICY, WRITER_TRANSPORT_RETRY_POLICY, isSafeReadOnlyPiRetryCandidate, isWriterTransportRetryCandidate, } from "./retry-policy.js";
+import { DEFAULT_READ_ONLY_PI_RETRY_POLICY, PLANNER_OUTPUT_LIMIT_RETRY_POLICY, PROTOCOL_AWARE_PI_RETRY_POLICY, STRUCTURED_REQUIRED_PI_RETRY_POLICY, BACKEND_TEST_WRITER_COMPLETENESS_RETRY_POLICY, WRITER_TRANSPORT_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";
@@ -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;
     }
@@ -1719,6 +1769,7 @@ export function buildStandardHybridDagFromTask(sources) {
                 forbiddenPaths,
                 outputContract: "Archived final shell verification stdout/stderr with exit codes; no worktree writes.",
                 subtask_prompt: "Run the adapter-resolved final verification commands before read-only verification review.",
+                transientRetryProfile: TARGET_TEMPLATE_TRANSIENT_RETRY_PROFILE,
                 shell: {
                     commands: verifyShellCommands,
                     verifyEvidence: buildVerifyEvidence({
@@ -1839,6 +1890,7 @@ export function buildStandardHybridDagFromTask(sources) {
                 allowedPaths: implementPaths.allowedPaths,
                 forbiddenPaths,
                 writerOutcomePolicy: { type: "implementation-outcome-v1" },
+                transientRetryProfile: TARGET_TEMPLATE_TRANSIENT_RETRY_PROFILE,
                 subtask_prompt: [
                     "Implement the approved plan with minimal focused changes.",
                     "Stay within writeSet. Do not write root artifacts/** unless artifacts paths are explicitly declared in writeSet.",
@@ -1898,6 +1950,7 @@ export function buildStandardHybridDagFromTask(sources) {
         ],
     };
     applySddEmbeddedEnhancements(spec, sources.sddEmbeddedSkills ?? new Set());
+    stampTargetTemplateTransientRetryProfile(spec);
     applyDefaultReadOnlyRetryPolicy(spec);
     parseDagSpec(spec);
     assertValidDagSpec(spec);
@@ -2527,7 +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,
@@ -2626,6 +2686,11 @@ async function buildFrontendHybridDagFromTask(sources) {
         ...behaviorVerifyEvidence.commandLabels.map((command) => `  - ${JSON.stringify(command)}`),
     ].join("\n");
     const advisories = [];
+    if (!hasDeclaredFrontendVerification &&
+        fallbackVerifyCommands.staticCommands.length === 0 &&
+        fallbackVerifyCommands.behaviorCommands.length === 0) {
+        advisories.push("未发现可用的前端验证命令：目标项目没有可读取的 package.json scripts，也未探测到本地 TypeScript，verify 节点将没有静态/行为命令可执行。请在 task.json --verify 或执行约束.md 的验证约束中显式声明命令（例如 node --check src/app.js），然后重新生成 DAG。");
+    }
     if (frontendSourceMentionsMock(frontendSources) &&
         frontendMockStrategyMustBeNotNeeded(frontendSources)) {
         advisories.push("auto 模式已将 Mock 策略收窄为 not-needed：任务源提到接口/API/Mock 需求，但仓库无确认 Mock 能力或无确定性 Mock 验证命令。若项目规范要求 Mock，请声明 frontendMock.verifyCommands 或 policy:required 后重新生成 DAG。");
@@ -4759,6 +4824,13 @@ function buildFrontendTestHybridDag(sources) {
     };
     const declaredRequirementIds = buildDagSourceBinding(sources).requirementIds;
     const declaredAcIds = declaredRequirementIds.filter((id) => /^AC(?:-[A-Z0-9]+)+$/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;
@@ -4799,6 +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-*'});",
@@ -4817,6 +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'});",
         "  }",
         " }",
@@ -4848,6 +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);",
@@ -4855,7 +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)); }",
@@ -4980,6 +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):",
@@ -5139,7 +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"),
@@ -5162,7 +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({
@@ -5230,6 +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"),
                 },
@@ -6442,7 +6521,12 @@ async function buildHybridDagForTemplate(sources, template, options = {}) {
     spec.sourceBinding = buildDagSourceBinding(sources);
     spec.taskContractBinding = taskContractBinding;
     assertNoGovernanceFlagOnDisallowedTemplate(spec, template);
-    parseDagSpec(spec);
+    if (template === "standard-dag" ||
+        template === "review-gated-dag" ||
+        template === "supervised-implementation") {
+        stampTargetTemplateTransientRetryProfile(spec);
+    }
+    spec = parseDagSpec(spec);
     assertValidDagSpec(spec);
     return spec;
 }
@@ -6493,14 +6577,6 @@ export async function buildHybridDagFromTask(sources, options = {}) {
 function cloneTask(task, patch = {}) {
     return { ...task, ...patch };
 }
-/**
- * Apply default Pi retry policies to generated DAG nodes:
- * - safe read-only planner/scout/reviewer/verifier/supervisor/closeout
- * - exclusive implementers get a single clean-timeout transport retry
- *   (provider stall with zero attributed writes only)
- * Dynamic, shell, static, and decision-gate nodes are skipped. Idempotent:
- * never overwrites an explicit retryPolicy a task already declares.
- */
 /**
  * Backend-implementation Pi extension buckets (plan 2026-08-21 D3):
  * - read side: every non-writer, non-closeout Pi node gets navigation
@@ -6523,6 +6599,22 @@ function applyBackendPiExtensionBuckets(spec) {
             task.toolProfile === "write" ? ["pi-codegraph"] : ["pi-codegraph", "pi-lens"];
     }
 }
+function stampTargetTemplateTransientRetryProfile(spec) {
+    for (const task of spec.tasks) {
+        if (isTargetTemplateImplementPi(task) ||
+            isCanonicalFinalVerifyShellRetryCandidate(task)) {
+            task.transientRetryProfile = TARGET_TEMPLATE_TRANSIENT_RETRY_PROFILE;
+        }
+    }
+}
+/**
+ * Apply default Pi retry policies to generated DAG nodes:
+ * - safe read-only planner/scout/reviewer/verifier/supervisor/closeout
+ * - exclusive implementers get a single clean-timeout transport retry
+ *   (provider stall with zero attributed writes only)
+ * Dynamic, shell, static, and decision-gate nodes are skipped. Idempotent:
+ * never overwrites an explicit retryPolicy a task already declares.
+ */
 function applyDefaultReadOnlyRetryPolicy(spec) {
     for (const task of spec.tasks) {
         if (task.retryPolicy !== undefined)
@@ -6535,6 +6627,10 @@ function applyDefaultReadOnlyRetryPolicy(spec) {
                     : DEFAULT_READ_ONLY_PI_RETRY_POLICY;
             continue;
         }
+        if (isTargetTemplateImplementPi(task)) {
+            task.retryPolicy = TARGET_TEMPLATE_WRITER_TRANSPORT_RETRY_POLICY;
+            continue;
+        }
         if (isWriterTransportRetryCandidate(task)) {
             task.retryPolicy = WRITER_TRANSPORT_RETRY_POLICY;
         }
@@ -6795,6 +6891,7 @@ function buildReviewGatedHybridDag(standard, sources) {
     }));
     spec.tasks.splice(spec.tasks.length - 1, 0, buildReviewNode(sources), buildReviewVerdictRecoveryNode(sources), buildReviewGateNode(sources));
     applySddEmbeddedEnhancements(spec, sources.sddEmbeddedSkills ?? new Set());
+    stampTargetTemplateTransientRetryProfile(spec);
     applyDefaultReadOnlyRetryPolicy(spec);
     parseDagSpec(spec);
     assertValidDagSpec(spec);
@@ -7126,6 +7223,9 @@ async function buildHardVerifyNode(sources) {
         forbiddenPaths: commonForbiddenPaths(sources),
         outputContract: "Archived hard verification stdout/stderr with exit codes; no worktree writes.",
         subtask_prompt: "Run hard verification after repair round.",
+        transientRetryProfile: sources.verifyCommands && commands.length > 0
+            ? TARGET_TEMPLATE_TRANSIENT_RETRY_PROFILE
+            : undefined,
         shell: {
             commands,
             verifyEvidence: buildVerifyEvidence({
@@ -7296,6 +7396,7 @@ async function buildSupervisedHybridDag(standard, sources) {
         ],
     };
     applySddEmbeddedEnhancements(spec, sources.sddEmbeddedSkills ?? new Set());
+    stampTargetTemplateTransientRetryProfile(spec);
     applyDefaultReadOnlyRetryPolicy(spec);
     parseDagSpec(spec);
     assertValidDagSpec(spec);

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

@@ -10,7 +10,7 @@ import { resolveContextPolicy } from "./context-policy.js";
 import { buildDagNodePromptEnvelope, formatConvergenceFeedbackBlock, } from "./prompt.js";
 import { persistLongNodeOutputArtifacts } from "./upstream-artifacts.js";
 import { buildOutputLimitRecoverySection, loadBackendTestWriterProgressForRetry, } from "./backend-test-writer-completeness.js";
-import { computeBackoffDelayMs, isRetryablePiFailureCategory, isSafeReadOnlyPiRetryCandidate, isWriterEmptyDiffRetryCandidate, isWriterTransportRetryCandidate, } from "./retry-policy.js";
+import { computeBackoffDelayMs, isCanonicalFinalVerifyShellRetryCandidate, isRetryablePiFailureCategory, isSafeReadOnlyPiRetryCandidate, isTargetTemplateTransientRetryNode, isWriterEmptyDiffRetryCandidate, isWriterTransportRetryCandidate, VERIFY_SHELL_RETRY_POLICY, } from "./retry-policy.js";
 import { applyNodeActivity, evaluateNodeLiveness, resolveLivenessPolicy, } from "./liveness-policy.js";
 import { buildProtocolRetryInstruction, normalizeReviewVerdictAfterRetries, parseJsonReviewVerdict, validateOutputProtocol, } from "./output-protocol.js";
 import { getStructuredContractValidator } from "./contract-output-registry.js";
@@ -260,6 +260,16 @@ function buildAttemptPrompt(task, basePrompt, attemptNumber, previousFailureCate
             "</retry_instruction>",
         ].join("\n");
     }
+    if (previousFailureCategory === "writer-clean-timeout") {
+        return [
+            basePrompt,
+            "",
+            "<retry_instruction>",
+            "Previous attempt ended in a clean provider/transport failure with zero attributed workspace writes and zero write/edit tool calls.",
+            "Do not invent existing implementation progress. Continue from the original task only if the workspace is still unchanged.",
+            "</retry_instruction>",
+        ].join("\n");
+    }
     if (previousFailureCategory !== "output-too-large") {
         return basePrompt;
     }
@@ -696,12 +706,18 @@ export async function executeDagNode(input) {
             // Invalid or missing harness preserves the previous no-override behavior.
         }
     }
-    const retryPolicy = task.retryPolicy &&
+    const specRetryPolicy = task.retryPolicy &&
         (isSafeReadOnlyPiRetryCandidate(task) ||
             isWriterEmptyDiffRetryCandidate(task) ||
             isWriterTransportRetryCandidate(task))
         ? task.retryPolicy
         : undefined;
+    const implicitVerifyRetryPolicy = specRetryPolicy === undefined &&
+        isCanonicalFinalVerifyShellRetryCandidate(task) &&
+        isTargetTemplateTransientRetryNode(task)
+        ? VERIFY_SHELL_RETRY_POLICY
+        : undefined;
+    const retryPolicy = specRetryPolicy ?? implicitVerifyRetryPolicy;
     const maxAttempts = retryPolicy?.maxAttempts ?? 1;
     const attempts = [];
     let totalAttemptWallDurationMs = 0;
@@ -943,6 +959,7 @@ export async function executeDagNode(input) {
         const attemptFinishedAt = new Date().toISOString();
         const attemptWallDurationMs = durationBetween(attemptStartedAt, attemptFinishedAt);
         totalAttemptWallDurationMs += attemptWallDurationMs;
+        const rawFailureCategory = result.rawFailureCategory ?? result.failureCategory;
         const attemptRecord = {
             attempt: attemptNumber,
             startedAt: attemptStartedAt,
@@ -954,6 +971,7 @@ export async function executeDagNode(input) {
             stderr: result.stderr,
             assistantText: result.assistantText,
             failureCategory: result.failureCategory,
+            ...(rawFailureCategory ? { rawFailureCategory } : {}),
             backend: result.backend,
             sdkAttempted: result.sdkAttempted,
             tokensUsed: result.tokensUsed,

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

@@ -3,6 +3,7 @@ import { readFile } from "node:fs/promises";
 import { assertTaskContractBindingFresh } from "./task-contract-binding.js";
 import { isSafeReadOnlyPiRetryCandidate, isTransportCrashedExclusiveWriter } from "./retry-policy.js";
 import { resolveDagTaskSourcePath } from "../../task/dag-source-paths.js";
+import { composeDagPromptOverridePrompt, MAX_DAG_RERUN_PROMPT_OVERRIDE_CHARS, } from "../../shared/dag-prompt-override.js";
 const REVIEW_VERDICT_RECOVERY_NODE_ID = "review-verdict-recovery-pi";
 const SAFE_SHELL_GATE_KEYS = [
     "verdictGate",
@@ -381,6 +382,100 @@ export function assessResetClosureSafety(spec, resetNodeIds, options = {}) {
         reasonCodes: [...new Set(reasonCodes)],
     };
 }
+/**
+ * Experimental prompt-override admission (fail-closed):
+ * - requires the internal switch (see shared/dag-prompt-override.ts);
+ * - the target must be a failed (ERROR) node inside the reset closure, so an
+ *   override can only adjust how an already-failed node retries — never a
+ *   finished, skipped, or not-yet-started node;
+ * - only `subtask_prompt` semantics change: "append" keeps the frozen base
+ *   and adds a marked addendum block, "replace" swaps the whole prompt;
+ *   base/final sha256 are recorded for provenance and bound into planHash.
+ */
+function assessPromptOverrides(input) {
+    const blockingNodes = [];
+    const reasonCodes = [];
+    const tasks = taskById(input.spec);
+    const resetSet = new Set(input.resetNodeIds);
+    const seen = new Set();
+    const entries = [];
+    if (!input.enabled) {
+        reasonCodes.push("prompt-override-disabled");
+    }
+    for (const override of input.overrides) {
+        const nodeId = override.nodeId?.trim() ?? "";
+        if (seen.has(nodeId)) {
+            blockingNodes.push({
+                nodeId,
+                reasonCode: "prompt-override-duplicate-node",
+            });
+            reasonCodes.push("prompt-override-duplicate-node");
+            continue;
+        }
+        seen.add(nodeId);
+        const task = tasks.get(nodeId);
+        if (!task) {
+            blockingNodes.push({
+                nodeId,
+                reasonCode: "prompt-override-node-missing",
+            });
+            reasonCodes.push("prompt-override-node-missing");
+            continue;
+        }
+        if (!resetSet.has(nodeId)) {
+            blockingNodes.push({
+                nodeId,
+                reasonCode: "prompt-override-outside-reset-closure",
+            });
+            reasonCodes.push("prompt-override-outside-reset-closure");
+            continue;
+        }
+        if (input.state.nodes[nodeId]?.status !== "ERROR") {
+            blockingNodes.push({
+                nodeId,
+                reasonCode: "prompt-override-node-not-failed",
+            });
+            reasonCodes.push("prompt-override-node-not-failed");
+            continue;
+        }
+        const mode = override.mode === "append" ? "append" : "replace";
+        const prompt = override.prompt ?? "";
+        if (!prompt.trim() ||
+            prompt.length > MAX_DAG_RERUN_PROMPT_OVERRIDE_CHARS) {
+            blockingNodes.push({
+                nodeId,
+                reasonCode: "prompt-override-invalid-prompt",
+            });
+            reasonCodes.push("prompt-override-invalid-prompt");
+            continue;
+        }
+        const basePromptSha256 = createHash("sha256")
+            .update(task.subtask_prompt, "utf8")
+            .digest("hex");
+        const finalPrompt = composeDagPromptOverridePrompt({
+            basePrompt: task.subtask_prompt,
+            text: prompt,
+            mode,
+        });
+        const promptSha256 = createHash("sha256")
+            .update(finalPrompt, "utf8")
+            .digest("hex");
+        if (promptSha256 === basePromptSha256) {
+            blockingNodes.push({
+                nodeId,
+                reasonCode: "prompt-override-unchanged",
+            });
+            reasonCodes.push("prompt-override-unchanged");
+            continue;
+        }
+        entries.push({ nodeId, mode, basePromptSha256, promptSha256 });
+    }
+    return {
+        entries: entries.length > 0 ? entries : undefined,
+        blockingNodes,
+        reasonCodes: [...new Set(reasonCodes)],
+    };
+}
 export function computePlanHash(planWithoutHash) {
     const canonical = JSON.stringify(sortKeysDeep(planWithoutHash));
     return createHash("sha256").update(canonical, "utf8").digest("hex");
@@ -648,6 +743,22 @@ export async function evaluateDagRerunPlan(input) {
             blockingNodes.push({ nodeId, reasonCode: "parent-facts-invalid" });
         }
     }
+    let promptOverrideEntries;
+    if (input.promptOverrides !== undefined && input.promptOverrides.length > 0) {
+        const overrideAssessment = assessPromptOverrides({
+            spec: input.parentSpec,
+            state: input.parentState,
+            resetNodeIds,
+            overrides: input.promptOverrides,
+            enabled: input.promptOverrideEnabled === true,
+        });
+        blockingNodes.push(...overrideAssessment.blockingNodes);
+        for (const code of overrideAssessment.reasonCodes) {
+            reasonCodes.push(code);
+            blockedReasons.push(code);
+        }
+        promptOverrideEntries = overrideAssessment.entries;
+    }
     const uniqueReasonCodes = [...new Set(reasonCodes)];
     const uniqueBlockedReasons = [...new Set(blockedReasons)];
     const eligible = uniqueBlockedReasons.length === 0 &&
@@ -691,6 +802,7 @@ export async function evaluateDagRerunPlan(input) {
             parentLifecycle: input.parentLifecycle,
             workerManaged: input.workerManaged,
         }),
+        ...(promptOverrideEntries ? { promptOverrides: promptOverrideEntries } : {}),
     };
     return {
         ...planWithoutHash,