@tea-agent/loop-agent 0.42.0-next.9 → 0.43.0-next.1

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

@@ -14,7 +14,7 @@ import { DEFAULT_OPENSPEC_GOVERNANCE_ROOT, DEFAULT_FRONTEND_SPEC_ROOTS, extractT
 import { readFrontendSpecRegistry, scoreFrontendSpecCandidate, } from "../../task/frontend-project-capability.js";
 import { BASELINE_FORBIDDEN_PATHS } from "./governance-constants.js";
 import { buildDecisionEnvelopePromptContract } from "./decision-envelope.js";
-import { DEFAULT_READ_ONLY_PI_RETRY_POLICY, FRONTEND_SCOUT_COMPLETENESS_RETRY_POLICY, PLANNER_OUTPUT_LIMIT_RETRY_POLICY, PROTOCOL_AWARE_PI_RETRY_POLICY, BACKEND_TEST_WRITER_COMPLETENESS_RETRY_POLICY, BACKEND_TEST_MARKDOWN_BINDING_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 { DEFAULT_READ_ONLY_PI_RETRY_POLICY, FRONTEND_SCOUT_COMPLETENESS_RETRY_POLICY, PLANNER_OUTPUT_LIMIT_RETRY_POLICY, PROTOCOL_AWARE_PI_RETRY_POLICY, BACKEND_TEST_MD_PLAN_RETRY_POLICY, BACKEND_TEST_WRITER_COMPLETENESS_RETRY_POLICY, BACKEND_TEST_MARKDOWN_BINDING_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";
@@ -40,7 +40,7 @@ import { buildBackendTestIntakeContext } from "./backend-test-intake-context.js"
 import { buildFrontendTestOutcomeGateShellSnippet } from "./frontend-test-result-contract.js";
 import { classifyFrontendRisk, } from "./frontend-risk.js";
 import { discoverFrontendProjectCapability, resolveFrontendSpecRootAliases, } from "./frontend-project-capability.js";
-import { buildFrontendImplementationContractSkeleton, FRONTEND_IMPLEMENTATION_CONTRACT_SCHEMA_ID, } from "./frontend-implementation-contract.js";
+import { buildFrontendImplementationContractSkeleton, buildFrontendVerifyCommandDirectory, classifyFrontendVerifyCommandText, FRONTEND_IMPLEMENTATION_CONTRACT_SCHEMA_ID, isRecognizedFrontendVerifyCommandText, } from "./frontend-implementation-contract.js";
 import { computeFrontendShapeSourceDigest, parseFrontendShapeTransitionCapsule, resolveFrontendTaskShape, } from "./frontend-shape.js";
 import { discoverLatestCommittedFrontendShapeCapsule } from "./frontend-shape-capsule-store.js";
 import { FRONTEND_NO_VERIFICATION_MARKER_TEXT } from "./frontend-verification-trace.js";
@@ -1007,11 +1007,19 @@ function extractFrontendVerifyCommandsFromMarkdown(input) {
         const verifyCommand = markdownVerifyCommand(input.repoRoot, command);
         if (!verifyCommand)
             continue;
-        if (/\b(typecheck|lint|eslint|tsc|build)\b/i.test(command)) {
-            staticCommands.push(verifyCommand);
+        // Markdown is untrusted prose: admit only commands with known verification
+        // semantics before using the shared classifier. Explicit task verifyCommands
+        // are operator-authorized separately and do not pass through this filter.
+        if (!isRecognizedFrontendVerifyCommandText(command))
             continue;
+        // Single shared classifier: never fork lane regexes here. The old
+        // inline static-first regex misclassified behavior commands whose
+        // arguments mention build paths (e.g. `npx vitest run
+        // test/build.test.ts`).
+        if (classifyFrontendVerifyCommandText(command) === "static") {
+            staticCommands.push(verifyCommand);
         }
-        if (/\b(test|vitest|jest|playwright|cypress|e2e)\b/i.test(command)) {
+        else {
             behaviorCommands.push(verifyCommand);
         }
     }
@@ -1051,10 +1059,13 @@ function chooseFrontendVerifyCommands(input) {
     }
     return { commandSource: "inline" };
 }
+/**
+ * Generation-time lane split for frontend verify commands. The mode itself
+ * is owned by `classifyFrontendVerifyCommandText` (contract module); this
+ * wrapper only preserves the local call shape.
+ */
 function classifyFrontendVerifyCommand(command) {
-    return /\b(typecheck|check-types|lint|eslint|tsc|build|check)\b/i.test(command) || /\bscripts[\\/]+ci(?:-tests)?\.sh\b/i.test(command)
-        ? "static"
-        : "behavior";
+    return classifyFrontendVerifyCommandText(command);
 }
 function buildExplicitFrontendVerifyCommands(taskConfig, repoRoot) {
     const staticCommands = [];
@@ -2548,8 +2559,8 @@ function resolveFrontendMockContextBlock(sources) {
     if (mode === "not-required") {
         parts.push("Generation-time evidence does not require Mock. The assessment must still use contract/scout evidence: select not-needed when Mock is intentionally skipped, or select a safe Mock strategy if project evidence supports one.");
         if (frontendMockStrategyMustBeNotNeeded(sources)) {
-            parts.push('Auto mode has no confirmed project Mock capability or no deterministic Mock verification command. The structured contract must set mockApi.strategy to "not-needed". Do not add Mock files or dependencies; keep the real request path as default and record any unproved backend behavior as Real Integration Gap.');
-            parts.push('HARD CONSTRAINT (frozen at generation time): this DAG allows only mockApi.strategy "not-needed"; the prewrite gate rejects any other strategy. If project governance (openspec / ai_workspace / decision records, e.g. a DEC rule requiring native) demands Mock-backed verification, that is a generation-time contract gap, not a plan-revision defect: declare frontendMock.verifyCommands (or policy: "required") in task.json and regenerate the DAG. Do not emit any mockApi.strategy outside the allowlist and do not add Mock files or dependencies within this run.');
+            parts.push('Auto mode has no confirmed project Mock capability or no deterministic Mock verification command. The structured contract must set mockApi.strategy to "not-needed". Keep the real request path as the default, record any unproved backend behavior as Real Integration Gap, and do not add Mock files or dependencies within this run.');
+            parts.push('HARD CONSTRAINT (frozen at generation time): this DAG allows only mockApi.strategy "not-needed"; the prewrite gate rejects any other strategy. If project governance (openspec / ai_workspace / decision records, e.g. a DEC rule requiring native) demands Mock-backed verification, that is a generation-time contract gap, not a plan-revision defect: declare frontendMock.verifyCommands (or policy: "required") in task.json and regenerate the DAG.');
         }
     }
     if (mode === "blocked") {
@@ -3034,7 +3045,7 @@ async function buildFrontendHybridDagFromTask(sources) {
         "- interactions[]: name, trigger, expectedBehavior, implementationTargets, verificationTargetIds",
         "- targets: routes, publicApiChanges (files are runtime-owned)",
         "- mockApi: strategy, productionDefaultOff, activation, endpoints[]",
-        "- verificationTargets[]: id (stable test-title trace token for non-static targets), type, commandLabel, file, requirementIds, uiStates",
+        "- verificationTargets[]: id (stable test-title trace token for behavior targets), commandId (frozen directory key), mode + commandLabel (runtime-resolved), file, requirementIds, uiStates, scope? (display-only)",
         "- 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",
@@ -3080,7 +3091,7 @@ async function buildFrontendHybridDagFromTask(sources) {
         "Frontend planning must consume the read-only Mock assessment strategy produced after scouting; MOCK_STRATEGY: blocked must not pass the deterministic Mock contract gate.",
         "Mock implementations must preserve the real request path as the default, require explicit test/dev activation, and never rely on commenting out the real request.",
         "Mock-backed behavior evidence proves only the documented frontend contract, never real API integration.",
-        "frontend-implementation DAGs must complete deterministic static verification before final review. Behavior verification is also required when the task declares a behavior entrypoint or the implementation contract contains a non-static verification target; static-only contracts must map every target to the declared static entrypoint.",
+        "frontend-implementation DAGs must complete deterministic static verification before final review. Behavior verification is also required when the task declares a behavior entrypoint or the implementation contract contains a behavior-mode verification target; static-only contracts must map every target to the declared static entrypoint.",
         "Frontend closeout renders only from committed facts; a weak status (failed / not-run / baseline-debt / mock-backed / pending) can never be rewritten into a stronger one (passed / real-integrated).",
         `Frontend risk classification: ${frontendRisk.selectedRisk} — ${frontendRisk.reason}`,
         frontendRisk.forceFullGates
@@ -3197,13 +3208,25 @@ async function buildFrontendHybridDagFromTask(sources) {
         : undefined;
     const mockShellCommands = mockVerifyTemplate?.shell?.commands ?? [];
     const mockVerifyEvidence = mockVerifyTemplate?.shell?.verifyEvidence;
+    // Contract v2: the frozen command directory is the only command reference
+    // the plan may use. Modes are assigned here once from the generation-time
+    // lane split; downstream materialization resolves the same directory from
+    // the run spec, so plan and runtime can never disagree on a mode.
+    const frontendVerifyDirectory = buildFrontendVerifyCommandDirectory({
+        staticLabels: staticVerifyEvidence.commandLabels,
+        behaviorLabels: behaviorVerifyEvidence.commandLabels,
+        mockLabels: mockVerifyEvidence?.commandLabels ?? [],
+        staticCommandTexts: staticVerifyEvidence.commandTexts,
+        behaviorCommandTexts: behaviorVerifyEvidence.commandTexts,
+        mockCommandTexts: mockVerifyEvidence?.commandTexts ?? [],
+    });
     const fixedVerificationContext = [
         "## Fixed frontend verification entrypoints",
         "These shell entrypoints are fixed at DAG generation and are the only commands the static and behavior shell nodes execute. A strategy or plan may add tests behind an existing entrypoint inside writeSet, but must not invent or replace commands or assume subtask_prompt executes a command.",
+        "Reference frozen commands ONLY by commandId (record_plan_verification_target entry.commandId). The runtime resolves mode and label; never invent a mode or type.",
+        ...frontendVerifyDirectory.map((entry) => `  - ${entry.commandId} [${entry.mode}]: ${JSON.stringify(entry.label)}`),
         `- Static command source: ${staticVerifyEvidence.commandSource}`,
-        ...staticVerifyEvidence.commandLabels.map((command) => `  - ${JSON.stringify(command)}`),
         `- Behavior command source: ${behaviorVerifyEvidence.commandSource}`,
-        ...behaviorVerifyEvidence.commandLabels.map((command) => `  - ${JSON.stringify(command)}`),
     ].join("\n");
     const advisories = [];
     if (!hasDeclaredFrontendVerification &&
@@ -3319,13 +3342,16 @@ async function buildFrontendHybridDagFromTask(sources) {
                 allowedPaths: readOnlyPaths,
                 forbiddenPaths,
                 skills: FRONTEND_CONTRACT_SKILLS,
-                outputContract: "Typed requirement facts plus a concise Markdown contract. Submit through the incremental typed tools record_requirement / record_constraint / record_evidence_expectation / record_handoff_intent / record_open_question / record_split_proposal / record_openspec_selection, then call finalize_contract exactly once. Requirements use stable REQ/BR/AC identifiers with source spans and a disposition (explicit | repository-resolvable | assumption | blocking); each requirement registers evidence expectations across static/behavior/Mock/real-integration (required | optional | not-applicable), and UI-visible or interactive requirements register a non-blocking frontend-test handoff intent. End finalize_contract with a single contract disposition of ready | ready-with-assumptions | blocked. Cover scope, non-goals, acceptance criteria, UI states, target runtime environment, risks, and verification expectations; do not fix target files, components, or implementation methods as requirements. When OpenSpec candidates exist, classify only the ones you actually use: call record_openspec_selection once per required/relevant path; never enumerate irrelevant candidates (unmentioned defaults to irrelevant) and never emit a fenced selection JSON. No file writes.",
+                outputContract: "Typed requirement facts plus a concise Markdown contract. Submit through the incremental typed tools record_requirement / record_constraint / record_evidence_expectation / record_handoff_intent / record_open_question / record_split_proposal / record_ui_state / record_required_deliverables / record_openspec_selection, then call finalize_contract exactly once. record_requirement takes only the canonical ledger requirement id — the runtime owns the authoritative text, source spans, fragment bindings, and disposition. UI-visible or interactive requirements register a non-blocking frontend-test handoff intent, and any source-declared UI-state table is extracted verbatim through record_ui_state. End finalize_contract with a single contract disposition of ready | ready-with-assumptions | blocked. Cover scope, non-goals, acceptance criteria, UI states, target runtime environment, risks, and verification expectations; do not fix target files, components, or implementation methods as requirements. When OpenSpec candidates exist, classify only the ones you actually use: call record_openspec_selection once per required/relevant path; never enumerate irrelevant candidates (unmentioned defaults to irrelevant) and never emit a fenced selection JSON. No file writes.",
                 subtask_prompt: [
                     "OUTPUT BUDGET DISCIPLINE (hard requirement, extreme-environment safe): the provider output window is small — NEVER attempt to emit the whole contract in one response; a single large JSON dump will be truncated and rejected. Incremental submission through the typed tools is the ONLY supported output mode. Start submitting with the FIRST tool call: after each read, call record_requirement for the requirements you have already confirmed, one or a few per call. Every tool-call round MUST make progress by submitting at least one record_* fact. Do not re-read the same source file that is already materialized in this session; read each file at most once.",
                     "Read task source and produce a concise frontend implementation contract as typed requirement facts plus narrative Markdown.",
-                    "Assign each requirement the SAME id as the ledger canonical requirement it covers (sourceBinding.requirementIds, e.g. AC-001) — do NOT invent new REQ/BR prefixed ids for canonical requirements: the compiled contract must match the ledger canonical requirement ids exactly or schema validation rejects it (unknown requirement id). Requirements use the canonical id with a source span (task-source section or repository file:line). Label each requirement's disposition as explicit | repository-resolvable | assumption | blocking; a blocking requirement must name its owner (human-decision or external-state) and evidence refs.",
-                    "Source fidelity ledger: when the DAG sourceBinding carries a requirement→fragment mapping (requirementToFragments, e.g. REQ-SRC-* ids from the managed ledger), each record_requirement MUST declare the fragments that requirement is bound to: set sourceFragmentIds to the mapped fragment ids (the authoritative provenance evidence the design policy verifies). sourceRefs (fragment→path display refs) are optional — declare them only when you have the exact path from the materialized source; otherwise omit them rather than inventing paths. Declare exactly what the ledger binds — do not invent ids, do not omit them, and do not re-derive them from prose. A requirement that the ledger binds but the contract omits (or fabricates) fails writer admission.",
+                    "Confirm each requirement by the SAME id as the ledger canonical requirement it covers (sourceBinding.requirementIds, e.g. AC-001) — do NOT invent new REQ/BR prefixed ids for canonical requirements: the compiled contract must match the ledger canonical requirement ids exactly or schema validation rejects it (unknown requirement id). record_requirement takes ONLY the canonical id; the runtime commits the authoritative text and sourceFragmentIds from the frozen ledger. Never pass text/statement/sourceFragmentIds yourself — model rewrites and JSON-stringified fragment arrays are rejected.",
+                    "Requirement semantics, source spans, dispositions, and fragment bindings are ledger/runtime-owned. If a canonical requirement is genuinely blocked, say so in the Markdown contract narrative and finalize with the matching disposition instead of trying to encode it in the requirement fact.",
                     "Register evidence expectations for each requirement across static, behavior, Mock, and real integration as required | optional | not-applicable; required must follow from user requirements, task risk, or project governance, never from model convenience. For UI-visible or interactive requirements, register a non-blocking frontend-test handoff intent.",
+                    'Use record_evidence_expectation with {requirementId,evidence:{static,behavior,mock,"real-integration"}}; every lane is required | optional | not-applicable. Requirement text and provenance remain runtime-owned.',
+                    'Before finalize_contract ready, call record_required_deliverables once with the complete {items:[{path,requirementId,sourceFragmentId}]} inventory, or {items:[]} when no file delivery is mandatory. Interpret the original source, including lists and tables: allowedPaths/only-allowed-to-modify is permission, not an obligation; do not promote prohibited files, examples or references into deliverables. Paths must occur exactly in a frozen source fragment bound to that canonical requirement. Correct the whole inventory before finalizing if needed.',
+                    "Authoritative UI states: when the task source declares a UI-state table (state id / trigger / observable outcome), extract it VERBATIM through record_ui_state, one call per state, using the source's own state ids. The planner must bind these ids later — do not rename, merge, or invent states.",
                     "Cover scope, non-goals, acceptance criteria, UI states, target runtime environment, risks, and verification expectations. Do not fix target files, components, styling, or implementation methods as requirements; leave those to Scout and Plan.",
                     "If the task is too large for one bounded writer, record a task split proposal instead of silently widening scope.",
                     "End the contract with a single disposition: ready, ready-with-assumptions (bounded assumptions that do not change product behavior), or blocked.",
@@ -3366,7 +3392,10 @@ async function buildFrontendHybridDagFromTask(sources) {
                 depends_on: ["frontend-contract-pi", "frontend-scout-pi"],
                 role: "planner",
                 executor: "pi",
-                complexity: "MED",
+                // Small topology has already proven a concentrated, no-remote scope;
+                // keep its bounded plan on the LOW model tier. Standard/High retain
+                // MED for broader contract-to-surface decisions.
+                complexity: frontendTaskShape.shape === "small" ? "LOW" : "MED",
                 writePolicy: "read-only",
                 retryPolicy: FRONTEND_PLAN_LADDER_RETRY_POLICY,
                 allowedPaths: readOnlyPaths,
@@ -3382,7 +3411,8 @@ async function buildFrontendHybridDagFromTask(sources) {
                     "Plan only the delta between the frozen frontend-contract-pi facts and frontend-scout-pi target surface. Do not reinterpret the task, repeat requirements, search the repository, or choose implementation order.",
                     "Record only: requirement-to-file/verification coverage; component/styling choices; applicable UI state and interaction behavior; data/Mock strategy; and a dependency policy or genuine evidence gap. Reuse Scout paths. If scope is missing, record a blocking gap instead of inventing a path.",
                     "Use the typed tool schemas as the field contract. Runtime owns schemaVersion, sourceBinding, riskLevel, targets.files, mockApi.productionDefaultOff, aliases, command allowlisting, path containment, and final validation; do not restate those rules or emit a full JSON contract.",
-                    `Cover each frozen requirement ID exactly once: ${requirementIds.join(", ") || "(none)"}. Bind every verification target to a listed frozen command and a Scout-confirmed file. Define behavior-level non-static targets: one target may cover multiple related requirementIds when one observable test behavior proves them together; do not mechanically create one target per requirement. A non-static target id is the stable machine trace token. Use static targets for project-wide commands. Static targets are traced by file and command only.`,
+                    `Cover each frozen requirement ID exactly once: ${requirementIds.join(", ") || "(none)"}. Bind every verification target to a frozen commandId from the directory above plus a Scout-confirmed file. Behavior commands prove observable behavior: one target may cover multiple related requirementIds when one test behavior proves them together; do not mechanically create one target per requirement. A behavior target id is the stable machine trace token and its file must be a test file. Static commands are project-wide checks traced by file and command only.`,
+                    "UX vocabulary protocol: record_state_registry FIRST with the full global vocabulary — one stable kebab-case behavior-domain name per UI state/interaction (e.g. planner-task-edit, focus-queue-move), never one name per AC number and never a rename of an already-recorded concept. Coverage slices by requirement; UX does not. Then record_state_flow entries whose names all come from that registry; uiState names must use the contract's declared authoritative ids (declaredUiStates in the plan input) when present. Retry attempts see committedUx in this input — reuse those exact names. Components: one choice may cover many state/interaction ids via covers; reuse-existing requires evidencePath naming an existing repo file (greenfield must be decision=new).",
                     ...(requiresOpenspecClassification ? ["When a component choice uses an OpenSpec selection, cite that selection; otherwise do not classify unrelated candidates."] : []),
                     "Call finalize_plan exactly once after the necessary typed facts. Return no Markdown narrative.",
                     "TOOL-ONLY PLAN: Do not read Contract/Scout stdout, task sources, or Scout-confirmed target files. Contract and Scout already own evidence discovery; use the injected upstream facts, record a genuine evidence gap when those facts are insufficient, and start committing record_* facts immediately. For decision=new, pass sourceRequirementIds to record_component_choice; runtime derives the exact PRD citation from the frozen ledger.",
@@ -3474,18 +3504,16 @@ async function buildFrontendHybridDagFromTask(sources) {
                     "Your authoritative terminal verdict is exactly one committed typed tool call: approve_design or request_design_changes. Call exactly one of them; after calling one, do not call the other.",
                     "request_design_changes must carry a typed issueCategory, at least one evidenceRef, and non-empty findings.",
                     "Your verdict is consumed as deterministic data input by frontend-writer-admission-shell. approve_design permits admission; request_design_changes blocks writer admission until a recovery plan incorporates every Critical/Important finding.",
-                    "Request design changes when the Mock strategy is MOCK_STRATEGY: blocked, missing, unsupported by repository evidence, inconsistent with the API contract, outside authorized paths/dependencies, unable to prove production-default-off behavior with the fixed production/default-real-path static check, or missing deterministic behavior verification for a declared behavior target or selected Mock strategy. Mock strategies require Mock-backed evidence. A static-only contract is allowed only when every verification target is static and maps to a declared static entrypoint. not-needed otherwise requires applicable real/no-remote behavior evidence unless auto mode explicitly skipped Mock because no project Mock capability exists; in that case the plan must preserve the real request path and record the Real Integration Gap.",
+                    "Request design changes when the Mock strategy is blocked, missing, unsupported by repository evidence, inconsistent with the API contract, outside authorized paths/dependencies, unable to prove production-default-off behavior with the fixed production/default-real-path static check, or missing deterministic behavior verification for a declared behavior target or selected Mock strategy. Mock strategies require Mock-backed evidence; a static-only contract is allowed only when every verification target is static and maps to a declared static entrypoint; not-needed requires applicable real/no-remote behavior evidence unless auto mode explicitly skipped Mock because no project Mock capability exists, in which case the plan must preserve the real request path and record the Real Integration Gap.",
                     "Also request design changes for missing applicable UI states, unsupported dependency additions, design-system drift without reason, weak interaction coverage, broad scope, inline fake data, schema drift, or missing deterministic verification commands.",
-                    "Component selection conformance is a hard blocking condition: request_design_changes when the frontend spec (component/theme/rule.components bucket) already defines a component for a purpose but the plan selects another or self-invents one without a declared deviation; when uiComponentChoices is missing/empty for UI-visible work while the frozen component/theme bucket is non-empty; when a decision=specified specReference.path is missing a ledger OpenSpec reference or successful read event; or when a decision=new component lacks a traceable task-source/PRD specReference. A PRD reference for decision=new is not an OpenSpec citation and must not be rejected merely for lacking an OpenSpec read event.",
-                    "For uiComponentChoices, purpose is the stable coverage key and must match interaction.name or uiState.name. Responsibility is expressed by the matched expectedBehavior plus rationale; you must not reject it merely for matching an interaction or component identifier.",
-                    "You must NOT make authoritative assertions about the execution result of frozen verification commands (typecheck/test/build/lint/etc.). Predicting that a command will necessarily pass or fail, or declaring an acceptance criterion unreachable on that basis, is out of your authority: command results are deterministically established by frontend-verify-shell. Any concern about verification feasibility must be recorded only as a non-blocking verification concern in findings (severity must not be Critical, and it must never be the sole fatal basis for request_design_changes). Only semantic design defects (component selection, state flow, interaction contract, or conflicts with the specification) may be Critical. A pure command-will-fail prediction must not be classified as contract-requirement-gap.",
+                    "Component selection conformance is a hard blocking condition: request_design_changes when the frontend spec (component/theme/rule.components bucket) already defines a component for a purpose but the plan selects another or self-invents one without a declared deviation; when uiComponentChoices is missing/empty for UI-visible work while the frozen component/theme bucket is non-empty; when a decision=specified specReference.path is missing a ledger OpenSpec reference or successful read event; or when a decision=new component lacks a traceable task-source/PRD specReference. A PRD reference for decision=new is not an OpenSpec citation and must not be rejected merely for lacking an OpenSpec read event. For uiComponentChoices, purpose is the stable coverage key and must match interaction.name or uiState.name; responsibility is expressed by the matched expectedBehavior plus rationale, and you must not reject it merely for matching an interaction or component identifier.",
+                    "You must NOT make authoritative assertions about the execution result of frozen verification commands: command results are deterministically established by frontend-verify-shell. Record a verification-feasibility concern only as a non-blocking finding (severity must not be Critical, and it must never be the sole fatal basis for request_design_changes). Only semantic design defects (component selection, state flow, interaction contract, or conflicts with the specification) may be Critical; a pure command-will-fail prediction must not be classified as contract-requirement-gap.",
                     "Read-only: do not modify repository files.",
                     "LARGE-FILE AUDIT (avoid full reads): style/theme audit files can be large (e.g. styles.css is often hundreds of KB). Prefer grep to locate the exact rules/variables you must verify (e.g. grep the oc- class, is-* modifier, or --oc- theme variables with their line numbers), then read only the narrow line range when surrounding context is needed. Do not read a large style/test file in full — a single full read can exhaust the read budget and fail the attempt.",
                     "Canonical contract reading: frontend-design-policy-shell prints absolute paths for Contract, Contract index, and the non-blocking Capacity diagnostic. Read the capacity diagnostic first. When it recommends full-contract, read the exact Contract path. When it recommends indexed-sections, read the Contract index and its hash-bound section files instead of opening the full contract. Never resolve a bare contracts/... path against the repository root or hunt for substitutes. Implementation target files inside the writeSet are created later by the implement node: do not read them and do not treat their absence as a design defect.",
                     fixedVerificationContext,
                     sourceContexts.designReview,
                     scopedOpenspecContext,
-                    frontendContractFieldSummary,
                     mockContextBlock,
                 ].join("\n\n"),
             },
@@ -3553,8 +3581,8 @@ async function buildFrontendHybridDagFromTask(sources) {
                     "The canonical contract already contains the approved requirement, target-file, UI-state, verification, design, and Mock/API decisions. Do not re-open task sources, OpenSpec, AI workspace, plan/revision, or design-review prose, and do not repeat broad repository research. Inspect only contract target files and directly related local code needed to implement them.",
                     "Execute in fixed stages and report each in the delivery summary: (1) Contract confirm, (2) Tests sync, (3) Component/UI state implementation, (4) API/Mock wiring per contract.mockApi, (5) Focused checks behind frozen entrypoints only, (6) Diff cleanup.",
                     "Map every requirement id, expectedOutcome, interaction trigger/expectedBehavior, and applicable UI state from the contract to concrete files. Do not invent shell verification commands; only frozen static/behavior entrypoints will run.",
-                    "For every non-static verification target, treat target.id as a stable trace token and include that exact token in a real describe/it/test literal title (for example, it('[VT-DASHBOARD-SHELL] renders the dashboard', ...)). One test title may carry multiple target ids when it proves multiple grouped behaviors; comments and ordinary strings do not count as trace evidence.",
-                    "Before reporting Tests Changed as done, self-audit with the executor's rule: collect ONLY the string literals passed directly to describe(/it(/test( calls in each test file and confirm every non-static target id for that file appears inside one of those literals. A token in a comment, a variable, or a non-title string does not satisfy the trace check; if any id is missing from the literal titles, edit the title strings before finishing.",
+                    "For every behavior verification target, treat target.id as a stable trace token and include that exact token in a real describe/it/test literal title (for example, it('[VT-DASHBOARD-SHELL] renders the dashboard', ...)). One test title may carry multiple target ids when it proves multiple grouped behaviors; comments and ordinary strings do not count as trace evidence.",
+                    "Before reporting Tests Changed as done, self-audit with the executor's rule: collect ONLY the string literals passed directly to describe(/it(/test( calls in each test file and confirm every behavior target id for that file appears inside one of those literals. A token in a comment, a variable, or a non-title string does not satisfy the trace check; if any id is missing from the literal titles, edit the title strings before finishing.",
                     "Begin implementation after the contract and its target files are confirmed. Do not spend the turn collecting optional context. If the canonical contract lacks behavior needed to edit safely, stop and state the blocking reason in the summary instead of reopening broad discovery.",
                     "Your implementation status is derived by the executor from mechanical facts (persisted write-tool events, run delta, write guard, requirement coverage, focused-check failures), never from any IMPLEMENTATION_OUTCOME first line. Do not emit an IMPLEMENTATION_OUTCOME first line.",
                     "The node runs a bounded micro-loop: after each write attempt the executor re-runs frozen focused checks and records a per-round diff checkpoint; the write guard stays active every round. Only repair local issues attributable to the current diff (syntax/type/import/format/unit-assert/obvious omission). Never change requirements, design, writeSet, or verification strictness inside the loop.",
@@ -5074,10 +5102,11 @@ async function buildBackendTestHybridDag(sources) {
             ".harness/dag-runs/**/reports/backend-test-environment.md",
         ],
         forbiddenPaths: forbidden,
+        retryPolicy: BACKEND_TEST_MD_PLAN_RETRY_POLICY,
         outputContract: "Return a Chinese, human-readable Markdown-first plan with Coverage Scope, Coverage Matrix, Scenario Partitions when applicable, and a machine-parseable Module Index. The runtime persists it as a run-owned Harness artifact; do not write testcase/md/README.md, execute pytest, or modify project files.",
         subtask_prompt: [
-            "This is a required plan-generation node. Read only the strict read set and return the complete Markdown plan in the assistant response. The runtime persists the response as a run-owned Harness artifact named generate-backend-md-plan-pi/plan.md. Do not write testcase/md/README.md or any project file; module case cards are written by downstream sharded nodes.",
-            "Output budget protocol (hard, max output <=16K per turn): Never paste full Matrix, case bodies, or source text into assistant chat. README holds only Scope+Matrix+module index; never inline full case bodies. If a Completeness Gate / OUTPUT_LIMIT_RECOVERY retry is injected, continue only listed target paths.",
+            "This is a required plan-generation node. Read only the strict read set and return the complete Markdown plan in the assistant response. Start the response with the final Markdown artifact immediately; never narrate analysis, reasoning, source summaries, or plans for producing the plan. The runtime persists the response as a run-owned Harness artifact named generate-backend-md-plan-pi/plan.md. Do not write testcase/md/README.md or any project file; module case cards are written by downstream sharded nodes.",
+            "Output budget protocol (hard, max output <=16K per turn): Emit the complete required section skeleton before filling long table rows, including exactly one Module Index table with at least one module row. Never paste full Matrix, case bodies, source text, analysis or reasoning into assistant chat. README holds only Scope+Matrix+module index; never inline full case bodies. If a Completeness Gate / OUTPUT_LIMIT_RECOVERY retry is injected, continue only listed target paths.",
             "Return the complete plan as plain Markdown. Do not emit JSON or code fences. The plan must contain the exact English protocol headings ## Coverage Scope, ## Coverage Matrix, ## Scenario Partitions when applicable, and ## Module Index. Never translate those headings into 覆盖范围/覆盖矩阵/场景分区/模块索引.",
             "Read the upstream environment report only through the strict read set. Generate the Markdown-first backend test plan; it will be persisted under the current DAG run's Harness artifacts, not under testcase/md/.",
             "Write human-readable content in Simplified Chinese by default. Keep English protocol literals exact: section headings, table headers, Partition IDs, TP IDs, Case IDs, HTTP methods, paths, field names, enum values, commands, filenames, code symbols and source citations. Never translate ## Module Index into ## 模块索引.",