@skyramp/mcp 0.3.4 → 0.3.6-rc.1

package/build/tools/submitReportTool.js

@@ -3,23 +3,27 @@ import { logger } from "../utils/logger.js";
 import * as fs from "fs/promises";
 import * as path from "path";
 import { AnalyticsService } from "../services/AnalyticsService.js";
-import { TEST_CATEGORIES, externalCategory } from "../types/TestRecommendation.js";
+import { TEST_CATEGORIES, externalCategory, } from "../types/TestRecommendation.js";
 import { TestType, HttpMethod } from "../types/TestTypes.js";
 import { DriftAction } from "../types/TestAnalysis.js";
 import { TestExecutionStatus } from "../types/TestExecution.js";
 import { IssueFoundCategory } from "../types/TestbotReport.js";
-import { StateManager, runArtifactDir } from "../utils/AnalysisStateManager.js";
+import { StateManager, runArtifactDir, getTestsRepoDir, } from "../utils/AnalysisStateManager.js";
 import { toolError, testFileMatches } from "../utils/utils.js";
 import { matchesApprovedPlan } from "../utils/planMatchKeys.js";
 import { isTestbotEnabled } from "../utils/featureFlags.js";
-import { findUnbackedClaims, listChangedFiles } from "../utils/reportVerification.js";
-import { rederiveReuseOutcome } from "./code-refactor/reuse-state.js";
+import { findInvalidSourceCitations, findUnchangedFileClaims, listChangedFiles, listChangedFilesAcross, } from "../utils/reportVerification.js";
+import { getReportLanguage, isEnforcedReportLanguage, findLanguageViolations, findLanguageNearMisses, reportLanguageDisplayName, } from "../utils/reportLanguage.js";
+import { rederiveReuseOutcome, reuseChainSkipped, } from "./code-refactor/reuse-state.js";
 // SKYR-3879 Path B: which testTypes the register-plan checkpoint gates. Mirrors
 // the generation tools actually wired to planGuard (batch-scenario/integration,
 // contract) — UI and E2E are on a separate blueprint-grounded pipeline and are
 // NOT gated at generation time (see planGuard.ts wiring), so they are excluded
 // here too rather than surprising the agent with a report-time-only gate.
-const PLAN_GATED_TEST_TYPES = new Set([TestType.CONTRACT, TestType.INTEGRATION]);
+const PLAN_GATED_TEST_TYPES = new Set([
+    TestType.CONTRACT,
+    TestType.INTEGRATION,
+]);
 /**
  * Filename of the report, written beside the state file. SKYR-4147: the report path is
  * derived here rather than accepted as a parameter. The caller builds the state file and
@@ -45,7 +49,9 @@ function planNameCandidates(testId, testType) {
     if (!id)
         return [];
     const prefix = `${testType}-`;
-    return id.toLowerCase().startsWith(prefix) ? [id, id.slice(prefix.length)] : [id];
+    return id.toLowerCase().startsWith(prefix)
+        ? [id, id.slice(prefix.length)]
+        : [id];
 }
 /**
  * Split an `endpoint` field into one {method, path} per endpoint it names.
@@ -54,14 +60,20 @@ function planNameCandidates(testId, testType) {
  * unmatchable path, so such an entry could match nothing (SKYR-4123).
  */
 function parseEndpointField(endpoint) {
-    const parts = (endpoint ?? "").split(",").map((p) => p.trim()).filter(Boolean);
+    const parts = (endpoint ?? "")
+        .split(",")
+        .map((p) => p.trim())
+        .filter(Boolean);
     if (parts.length === 0)
         return [{}];
     return parts.map((part) => {
         const spaceIdx = part.indexOf(" ");
         if (spaceIdx <= 0)
             return { path: part };
-        return { method: part.slice(0, spaceIdx), path: part.slice(spaceIdx + 1).trim() };
+        return {
+            method: part.slice(0, spaceIdx),
+            path: part.slice(spaceIdx + 1).trim(),
+        };
     });
 }
 // Drift actions that actually modify a test file. VERIFY and IGNORE are
@@ -88,23 +100,33 @@ const repositoryField = z
  *  (downstream consumers treat absence as "the primary repo"). */
 function normalizeRepository(item) {
     const trimmed = item.repository?.trim();
-    return trimmed ? { ...item, repository: trimmed } : { ...item, repository: undefined };
+    return trimmed
+        ? { ...item, repository: trimmed }
+        : { ...item, repository: undefined };
 }
 // videoPath is deliberately absent from this input contract: it is attached server-side
 // from the run's execution records (see attachVideoPath), and zod strips any the model
 // supplies anyway. SKYR-4156 is what happens when the agent owns that field instead.
 const testResultSchema = z.object({
-    testType: z.nativeEnum(TestType).describe("Type of test. Do not include priority or other metadata in this field."),
-    endpoint: z.string().describe("HTTP verb and path, e.g. 'GET /api/v1/products'"),
+    testType: z
+        .nativeEnum(TestType)
+        .describe("Type of test. Do not include priority or other metadata in this field."),
+    endpoint: z
+        .string()
+        .describe("HTTP verb and path, e.g. 'GET /api/v1/products'"),
     status: z.enum(["Pass", "Fail", "Skipped"]).describe("Test execution result"),
-    details: z.string().describe("One sentence — no embedded newlines, no markdown. e.g. '10.8s, products_contract_test.py' or 'failed: <one-line error summary>, products_contract_test.py'"),
+    details: z
+        .string()
+        .describe("One sentence — no embedded newlines, no markdown. e.g. '10.8s, products_contract_test.py' or 'failed: <one-line error summary>, products_contract_test.py'"),
     // Required for every row: each one reports a specific test file the agent ran, so it
     // can always name it. It is what identifies the row server-side — `endpoint` cannot,
     // since several tests routinely exercise one endpoint — and for ui/e2e it is what
     // attaches the recorded video (SKYR-4156). Not included in the report itself.
     testFilePath: z
         .string()
-        .refine((p) => path.isAbsolute(p), { message: "testFilePath must be an absolute path" })
+        .refine((p) => path.isAbsolute(p), {
+        message: "testFilePath must be an absolute path",
+    })
         .describe("Absolute path of the test file this result is for — the same path you passed to skyramp_execute_test's testFile param. Consumers basename it for display."),
     repository: repositoryField,
 });
@@ -116,21 +138,47 @@ const testResultSchema = z.object({
 // parsing. `reasoning` is free-form prose constrained by the Blueprint
 // Citation Invariant — every element cited must appear in `targetElements`.
 // See testbot prompt step 4 for the full populate-and-render rules.
+/**
+ * Accept an omitted key as an explicit `null`.
+ *
+ * SKYR-4208. The blueprint capture no longer carries a key whose value is null,
+ * so an element the agent lifts verbatim simply has no `testId`, `stableId` or
+ * `contextText`. Consumers of the report still expect all three keys, so the
+ * missing one is filled in here rather than asked for in the prompt.
+ */
+function nullWhenAbsent(inner) {
+    return inner.nullish().transform((v) => v ?? null);
+}
 export const targetElementSchema = z.object({
-    role: z.string().describe("ARIA role (e.g. 'button', 'heading', 'textbox', 'link'). Lifted verbatim from the captured blueprint element's `role` field."),
-    accessibleName: z.string().describe("Computed accessible name (e.g. 'Save changes', 'Order Details'). Lifted verbatim from the captured blueprint element's `accessibleName` field. Must match the bolded name in `reasoning` character-for-character."),
-    testId: z.string().nullable().describe("data-testid attribute value (preferred locator handle), or null when the element has no data-testid. Lifted from the blueprint element's `testId`."),
-    stableId: z.string().nullable().describe("Unique HTML id attribute value (fallback locator when no testId), or null. Lifted from the blueprint element's `stableId`."),
-    contextText: z.array(z.string()).nullable().describe("Disambiguating row text for elements inside repeating sections (table rows, list items): the row's surrounding non-interactive text. Lifted from the blueprint repeatingElement's items[].contextText. null for non-repeating elements."),
-    mutability: z.enum(["mutable", "immutable", "unknown"]).optional().describe("Whether the element's content/state is expected to change between captures. 'mutable' elements are behavioral-test targets; 'immutable' are smoke-test targets. Copied from the blueprint element's `mutability`."),
-    widgetType: z.enum(["native", "custom", "unknown"]).optional().describe("'native' = HTML built-ins (button, input). 'custom' or 'unknown' = non-standard composites; fall back to snapshot-driven trial clicks for interaction. Copied from the blueprint element's `widgetType`."),
+    role: z
+        .string()
+        .describe("ARIA role (e.g. 'button', 'heading', 'textbox', 'link'). Lifted verbatim from the captured blueprint element's `role` field."),
+    accessibleName: z
+        .string()
+        .describe("Computed accessible name (e.g. 'Save changes', 'Order Details'). Lifted verbatim from the captured blueprint element's `accessibleName` field. Must match the bolded name in `reasoning` character-for-character."),
+    testId: nullWhenAbsent(z.string()).describe("data-testid attribute value (preferred locator handle), or null when the element has no data-testid. Lifted from the blueprint element's `testId`."),
+    stableId: nullWhenAbsent(z.string()).describe("Unique HTML id attribute value (fallback locator when no testId), or null. Lifted from the blueprint element's `stableId`."),
+    contextText: nullWhenAbsent(z.array(z.string())).describe("Disambiguating row text for elements inside repeating sections (table rows, list items): the row's surrounding non-interactive text. Lifted from the blueprint repeatingElement's items[].contextText. null for non-repeating elements."),
+    mutability: z
+        .enum(["mutable", "immutable", "unknown"])
+        .optional()
+        .describe("Whether the element's content/state is expected to change between captures. 'mutable' elements are behavioral-test targets; 'immutable' are smoke-test targets. Copied from the blueprint element's `mutability`."),
+    widgetType: z
+        .enum(["native", "custom", "unknown"])
+        .optional()
+        .describe("'native' = HTML built-ins (button, input). 'custom' or 'unknown' = non-standard composites; fall back to snapshot-driven trial clicks for interaction. Copied from the blueprint element's `widgetType`."),
 });
 // Page metadata for a UI recommendation. Lifted from the BlueprintCapture
 // the agent used to populate `targetElements`. Codegen reads `url` to drive
 // page.goto(); the verifier uses `pageHash` to detect stale captures.
 export const pageContextSchema = z.object({
-    url: z.string().describe("URL of the page where the test runs. Lifted from BlueprintCapture.url."),
-    pageHash: z.string().optional().describe("Opaque hash of the captured page state (BlueprintCapture.pageHash). Lets the verifier confirm the recommendation was grounded in a still-current capture."),
+    url: z
+        .string()
+        .describe("URL of the page where the test runs. Lifted from BlueprintCapture.url."),
+    pageHash: z
+        .string()
+        .optional()
+        .describe("Opaque hash of the captured page state (BlueprintCapture.pageHash). Lets the verifier confirm the recommendation was grounded in a still-current capture."),
 });
 /**
  * SKYR-4193: LLMs habitually emit every key a schema declares, using a
@@ -164,21 +212,53 @@ function stripNullGroundingFields(val) {
 // interface that adds an `implemented: boolean` field. Both describe the same
 // concept (a test recommendation) — the only difference is whether it was
 // generated in this run or left for later. Tracked per Archit's review comment.
-export const newTestSchema = z.preprocess(stripNullGroundingFields, z.object({
-    testId: z.string().describe("Human-readable kebab-case identifier, e.g. 'contract-get-products' or 'integration-users-orders-workflow'. Format: '<testType>-<method>-<resource>' for single-endpoint tests or '<testType>-<scenario-slug>' for multi-step tests. Must be unique within the report."),
-    testType: z.nativeEnum(TestType).describe("Type of test created. Do not include priority or other metadata in this field."),
-    category: z.preprocess((val) => externalCategory(val), z.enum(TEST_CATEGORIES)).describe("Test category — critical categories (security_boundary, business_rule, data_integrity, breaking_change) get generation priority over workflow"),
-    endpoint: z.string().describe("HTTP verb and path, e.g. 'GET /api/v1/products'"),
+export const newTestSchema = z.preprocess(stripNullGroundingFields, z
+    .object({
+    testId: z
+        .string()
+        .describe("Human-readable kebab-case identifier, e.g. 'contract-get-products' or 'integration-users-orders-workflow'. Format: '<testType>-<method>-<resource>' for single-endpoint tests or '<testType>-<scenario-slug>' for multi-step tests. Must be unique within the report."),
+    testType: z
+        .nativeEnum(TestType)
+        .describe("Type of test created. Do not include priority or other metadata in this field."),
+    category: z
+        .preprocess((val) => externalCategory(val), z.enum(TEST_CATEGORIES))
+        .describe("Test category — critical categories (security_boundary, business_rule, data_integrity, breaking_change) get generation priority over workflow"),
+    endpoint: z
+        .string()
+        .describe("HTTP verb and path, e.g. 'GET /api/v1/products'"),
     fileName: z.string().describe("Name of the generated test file"),
-    description: z.string().trim().min(1).describe("What the test does — the steps and assertions, not the bugs it finds. e.g. 'Creates a collection, adds a link, then verifies the link exists'. Do NOT describe expected failures or bugs here — those belong in issuesFound."),
-    scenarioFile: z.string().optional().describe("Path to the scenario JSON file if one was generated (e.g. 'tests/scenario_collections-links.json')"),
-    traceFile: z.string().optional().describe("Path to the backend trace file if used or created"),
-    frontendTrace: z.string().optional().describe("Path to the Playwright/UI trace file if used or created"),
-    reasoning: z.string().describe("Why this test was created: what production risk it mitigates, what code pattern it targets, or what coverage gap it fills"),
+    description: z
+        .string()
+        .trim()
+        .min(1)
+        .describe("What the test does — the steps and assertions, not the bugs it finds. e.g. 'Creates a collection, adds a link, then verifies the link exists'. Do NOT describe expected failures or bugs here — those belong in issuesFound."),
+    scenarioFile: z
+        .string()
+        .optional()
+        .describe("Path to the scenario JSON file if one was generated (e.g. 'tests/scenario_collections-links.json')"),
+    traceFile: z
+        .string()
+        .optional()
+        .describe("Path to the backend trace file if used or created"),
+    frontendTrace: z
+        .string()
+        .optional()
+        .describe("Path to the Playwright/UI trace file if used or created"),
+    reasoning: z
+        .string()
+        .describe("Why this test was created: what production risk it mitigates, what code pattern it targets, or what coverage gap it fills"),
     repository: repositoryField,
-    targetElements: z.array(targetElementSchema).min(1).nullable().optional().describe("UI tests only: structured grounding for one or more elements the test targets. Most tests target a single element (array length 1); render-state and multi-step UI tests target several (array length 2+). Each entry must be lifted verbatim from a captured blueprint element. Set to null when blueprint capture failed (also requires '[no-blueprint-data]' marker in BOTH description and reasoning). See Blueprint Citation Invariant in testbot prompt."),
-    pageContext: pageContextSchema.optional().describe("UI tests only: page metadata for the test. Lifted from the BlueprintCapture used during grounding."),
-}).superRefine((rec, ctx) => {
+    targetElements: z
+        .array(targetElementSchema)
+        .min(1)
+        .nullable()
+        .optional()
+        .describe("UI tests only: structured grounding for one or more elements the test targets. Most tests target a single element (array length 1); render-state and multi-step UI tests target several (array length 2+). Each entry must be lifted verbatim from a captured blueprint element. Set to null when blueprint capture failed (also requires '[no-blueprint-data]' marker in BOTH description and reasoning). See Blueprint Citation Invariant in testbot prompt."),
+    pageContext: pageContextSchema
+        .optional()
+        .describe("UI tests only: page metadata for the test. Lifted from the BlueprintCapture used during grounding."),
+})
+    .superRefine((rec, ctx) => {
     // targetElements / pageContext are UI-only.
     if (rec.testType !== TestType.UI) {
         for (const field of ["targetElements", "pageContext"]) {
@@ -209,7 +289,8 @@ export const newTestSchema = z.preprocess(stripNullGroundingFields, z.object({
             });
         }
         // pageContext is required when targetElements is an array (grounded), forbidden when null.
-        if (Array.isArray(rec.targetElements) && rec.pageContext === undefined) {
+        if (Array.isArray(rec.targetElements) &&
+            rec.pageContext === undefined) {
             ctx.addIssue({
                 code: z.ZodIssueCode.custom,
                 path: ["pageContext"],
@@ -232,7 +313,8 @@ export const newTestSchema = z.preprocess(stripNullGroundingFields, z.object({
                     message: "reasoning must contain '[no-blueprint-data]' when targetElements is null.",
                 });
             }
-            if (rec.description && !rec.description.includes("[no-blueprint-data]")) {
+            if (rec.description &&
+                !rec.description.includes("[no-blueprint-data]")) {
                 ctx.addIssue({
                     code: z.ZodIssueCode.custom,
                     path: ["description"],
@@ -242,8 +324,18 @@ export const newTestSchema = z.preprocess(stripNullGroundingFields, z.object({
         }
     }
 }));
-const issueFoundSchema = z.object({
-    description: z.string().describe("One-line description. Do NOT prefix with the severity level — severity is a separate field. Include code logic bugs from the diff, test generation/execution failures, and environment misconfiguration."),
+/** A citation string the submit-time check verifies and the report then ships.
+ *  Normalized HERE, once, so the value the verifier resolved is byte-identical to
+ *  the value written: the check trimmed its input, so a padded " src/x.py " used to
+ *  validate against the real path and ship the unusable padded one. A
+ *  blank/whitespace-only value means "no citation", the same rule `repository` uses. */
+const citationString = z.preprocess((v) => (typeof v === "string" ? v.trim() || undefined : v), z.string().optional());
+const issueFoundSchema = z
+    .object({
+    description: z
+        .string()
+        .describe("One-line description. Do NOT prefix with the severity level — severity is a separate field. Include code logic bugs from the diff, test generation/execution failures, and environment misconfiguration. " +
+        "When sourceFile or sourceSymbol is set, quote the offending line inside backticks in this description — the exact code, copied from the file, so review tools can find it."),
     severity: z
         .enum(["critical", "high", "medium", "low"])
         .optional()
@@ -255,36 +347,130 @@ const issueFoundSchema = z.object({
         .describe("Issue classification. bug = a product/code defect, e.g. found by a test or in the diff. " +
         "lint = a linter or formatter finding (eslint, flake8, prettier). " +
         "type = a type-check failure (tsc, mypy). " +
-        "config = environment or tooling misconfiguration (wrong workspace auth type, missing env var, setup command failure). " +
+        "config = environment or tooling misconfiguration (wrong workspace auth type, missing env var, setup command failure), " +
+        "and also every Skyramp tool or environment failure — a generation or execution tool error, an unreachable app, a missing credential, a failed capture. " +
         "The report renders lint/type/config entries in a separate 'Configuration Errors' section so product bugs stay prominent under 'Issues Found'."),
     repository: repositoryField,
+    sourceFile: citationString.describe("Path of the application file whose code is missing or wrong, relative to the repository root (e.g. 'src/crud/products.py'). " +
+        "REQUIRED when category is 'bug'. " +
+        "Cite only a file you actually opened this run: this tool rejects the report if the path does not exist in the repository."),
+    sourceSymbol: citationString.describe("The function, method, or identifier inside sourceFile whose code is missing or wrong (e.g. 'delete_product'). " +
+        "Optional — a config file or template has no symbol to name. Set it whenever the file has one. " +
+        "This tool rejects the report if the text does not appear in the cited file."),
+    sourceLine: z
+        .number()
+        .int()
+        .positive()
+        .optional()
+        .describe("1-based line number in sourceFile. Optional and advisory. " +
+        "Give the line of the offending statement itself, not the line of the enclosing function's signature — review views place the finding on exactly this line. " +
+        "The tool does not check that the line still holds the cited code, because line numbers drift."),
+})
+    .superRefine((issue, ctx) => {
+    // A `bug` entry asserts that product code is wrong, so it has to name the code:
+    // without a file the entry is prose a reviewer cannot act on, and nothing
+    // downstream can place it in the diff. Only the file is required — a config file
+    // or a template legitimately has no symbol to name, and line numbers are
+    // advisory. The other categories are tooling findings that often have no single
+    // source location (an unreachable app, a missing env var), so the requirement is
+    // scoped to `bug` alone.
+    //
+    // The exit when the code cannot be named is NOT to invent a citation: a failure
+    // the agent cannot localize is already reported by its `testResults` row, and
+    // a tool or environment failure belongs under `config`.
+    if (issue.category !== IssueFoundCategory.Bug)
+        return;
+    if (issue.sourceFile !== undefined)
+        return;
+    ctx.addIssue({
+        code: z.ZodIssueCode.custom,
+        path: ["sourceFile"],
+        message: `sourceFile is required when category is 'bug' — a bug entry must name the code it is about. ` +
+            `Open the file, then set sourceFile (path relative to the repository root). Also set sourceSymbol (the function or identifier inside it) when the file has one, and sourceLine (the line of the offending statement) when you have it — both are optional. ` +
+            `If you cannot point at the code: a test failure is already reported by its testResults entry, and a tool or environment failure belongs under category 'config' — do not guess a citation.`,
+    });
 });
 const scenarioStepSchema = z.object({
-    method: z.nativeEnum(HttpMethod).optional().describe("HTTP method. Required for API steps, omit for UI/E2E actions."),
-    path: z.string().optional().describe("Endpoint or page path (e.g. '/api/v1/products' or '/products'). Required for API steps, omit for UI actions."),
-    description: z.string().describe("What this step does, e.g. 'Create a product' or 'Click checkout button and verify confirmation'"),
-    expectedStatusCode: z.number().optional().describe("Expected HTTP status code, e.g. 200, 201, 404"),
-    requestBody: z.record(z.any()).optional().describe("Example request body with realistic field values"),
-    responseBody: z.record(z.any()).optional().describe("Key response fields to verify, e.g. { id: 'number', name: 'string', in_stock: 'boolean?' }"),
+    method: z
+        .nativeEnum(HttpMethod)
+        .optional()
+        .describe("HTTP method. Required for API steps, omit for UI/E2E actions."),
+    path: z
+        .string()
+        .optional()
+        .describe("Endpoint or page path (e.g. '/api/v1/products' or '/products'). Required for API steps, omit for UI actions."),
+    description: z
+        .string()
+        .describe("What this step does, e.g. 'Create a product' or 'Click checkout button and verify confirmation'"),
+    expectedStatusCode: z
+        .number()
+        .optional()
+        .describe("Expected HTTP status code, e.g. 200, 201, 404"),
+    requestBody: z
+        .record(z.any())
+        .optional()
+        .describe("Example request body with realistic field values"),
+    responseBody: z
+        .union([z.record(z.any()), z.array(z.any())])
+        .optional()
+        .describe("Key response fields to verify, e.g. { id: 'number', name: 'string', in_stock: 'boolean?' }. An array for a collection/list endpoint that returns a JSON array."),
 });
-export const additionalRecommendationSchema = z.preprocess(stripNullGroundingFields, z.object({
-    testId: z.string().describe("Human-readable kebab-case identifier, e.g. 'integration-products-orders-workflow' or 'e2e-checkout-flow'. Format: '<testType>-<scenario-slug>'. Must be unique within the report."),
-    testType: z.nativeEnum(TestType).describe("Type of test. Do not include priority or other metadata in this field."),
-    category: z.preprocess((val) => externalCategory(val), z.enum(TEST_CATEGORIES)).describe("Test category — critical categories get generation priority over workflow"),
-    primaryEndpoint: z.string().optional().describe("The focal endpoint this test targets, e.g. 'PATCH /api/v1/orders/{order_id}'. Required for single-step contract tests. For multi-step integration or E2E scenarios, omit — the steps array is the authoritative source of all endpoints involved."),
-    scenarioName: z.string().optional().describe("Proposed scenario name for future generation, e.g. 'products-orders-workflow'. No file exists yet — this is a suggestion only. Omit if not applicable."),
+export const additionalRecommendationSchema = z.preprocess(stripNullGroundingFields, z
+    .object({
+    testId: z
+        .string()
+        .describe("Human-readable kebab-case identifier, e.g. 'integration-products-orders-workflow' or 'e2e-checkout-flow'. Format: '<testType>-<scenario-slug>'. Must be unique within the report."),
+    testType: z
+        .nativeEnum(TestType)
+        .describe("Type of test. Do not include priority or other metadata in this field."),
+    category: z
+        .preprocess((val) => externalCategory(val), z.enum(TEST_CATEGORIES))
+        .describe("Test category — critical categories get generation priority over workflow"),
+    primaryEndpoint: z
+        .string()
+        .optional()
+        .describe("The focal endpoint this test targets, e.g. 'PATCH /api/v1/orders/{order_id}'. Required for single-step contract tests. For multi-step integration or E2E scenarios, omit — the steps array is the authoritative source of all endpoints involved."),
+    scenarioName: z
+        .string()
+        .optional()
+        .describe("Proposed scenario name for future generation, e.g. 'products-orders-workflow'. No file exists yet — this is a suggestion only. Omit if not applicable."),
     // TODO: replace text with max(3) and check for regression
-    steps: z.array(scenarioStepSchema).describe("Ordered sequence of API/UI steps in this test scenario (at most 3). Each step must include method and path so the endpoints are explicit. Omit requestBody and responseBody from steps."),
-    description: z.string().describe("Walkthrough of what the test does — the steps and assertions. For multi-step scenarios, list the endpoints involved. The 'why it is valuable' belongs in reasoning."),
-    priority: z.preprocess((val) => (typeof val === "string" ? val.toLowerCase() : val), z.enum(["high", "medium", "low"])).describe("Priority level: high, medium, or low. First check diff relevance — does the test target an endpoint changed in this PR? HIGH: diff-relevant security/auth/error tests, cross-resource isolation for diff endpoints, CRUD lifecycle for NEW endpoints in the diff. MEDIUM: diff-relevant business-rule happy paths, multi-resource workflows involving diff endpoints, security/error tests for NON-diff endpoints. LOW: tests targeting only unchanged endpoints, trivially discoverable happy paths duplicating generated tests."),
-    openApiSpec: z.string().optional().describe("Path to OpenAPI/Swagger spec file if available, e.g. 'openapi.yaml'"),
-    backendTrace: z.string().optional().describe("Path to backend trace file if available, e.g. 'tests/skyramp-traces.json'. Used by integration and E2E tests."),
-    frontendTrace: z.string().optional().describe("Path to Playwright/UI trace file if available, e.g. 'tests/skyramp-playwright.zip'. UI tests need this; E2E tests need both frontend and backend traces."),
-    reasoning: z.string().describe("Why this test is recommended: the specific production risk, business rule, or security boundary it would validate"),
+    steps: z
+        .array(scenarioStepSchema)
+        .describe("Ordered sequence of API/UI steps in this test scenario (at most 3). Each API step must include method and path so the endpoints are explicit; a UI/E2E action step omits them. Include requestBody and responseBody only where they carry something a reader needs — the concrete values a claim rests on, or a list endpoint's array response; omit them otherwise."),
+    description: z
+        .string()
+        .describe("Walkthrough of what the test does — the steps and assertions. For multi-step scenarios, list the endpoints involved. The 'why it is valuable' belongs in reasoning."),
+    priority: z
+        .preprocess((val) => (typeof val === "string" ? val.toLowerCase() : val), z.enum(["high", "medium", "low"]))
+        .describe("Priority level: high, medium, or low. First check diff relevance — does the test target an endpoint changed in this PR? HIGH: diff-relevant security/auth/error tests, cross-resource isolation for diff endpoints, CRUD lifecycle for NEW endpoints in the diff. MEDIUM: diff-relevant business-rule happy paths, multi-resource workflows involving diff endpoints, security/error tests for NON-diff endpoints. LOW: tests targeting only unchanged endpoints, trivially discoverable happy paths duplicating generated tests."),
+    openApiSpec: z
+        .string()
+        .optional()
+        .describe("Path to OpenAPI/Swagger spec file if available, e.g. 'openapi.yaml'"),
+    backendTrace: z
+        .string()
+        .optional()
+        .describe("Path to backend trace file if available, e.g. 'tests/skyramp-traces.json'. Used by integration and E2E tests."),
+    frontendTrace: z
+        .string()
+        .optional()
+        .describe("Path to Playwright/UI trace file if available, e.g. 'tests/skyramp-playwright.zip'. UI tests need this; E2E tests need both frontend and backend traces."),
+    reasoning: z
+        .string()
+        .describe("Why this test is recommended: the specific production risk, business rule, or security boundary it would validate"),
     repository: repositoryField,
-    targetElements: z.array(targetElementSchema).min(1).nullable().optional().describe("UI tests only: structured grounding for one or more elements the test targets. Most tests target a single element (array length 1); render-state and multi-step UI tests target several (array length 2+). Each entry must be lifted verbatim from a captured blueprint element. Set to null when blueprint capture failed (also requires '[no-blueprint-data]' marker in BOTH description and reasoning). See Blueprint Citation Invariant in testbot prompt."),
-    pageContext: pageContextSchema.optional().describe("UI tests only: page metadata for the test. Lifted from the BlueprintCapture used during grounding."),
-}).superRefine((rec, ctx) => {
+    targetElements: z
+        .array(targetElementSchema)
+        .min(1)
+        .nullable()
+        .optional()
+        .describe("UI tests only: structured grounding for one or more elements the test targets. Most tests target a single element (array length 1); render-state and multi-step UI tests target several (array length 2+). Each entry must be lifted verbatim from a captured blueprint element. Set to null when blueprint capture failed (also requires '[no-blueprint-data]' marker in BOTH description and reasoning). See Blueprint Citation Invariant in testbot prompt."),
+    pageContext: pageContextSchema
+        .optional()
+        .describe("UI tests only: page metadata for the test. Lifted from the BlueprintCapture used during grounding."),
+})
+    .superRefine((rec, ctx) => {
     if (rec.testType === TestType.CONTRACT && !rec.primaryEndpoint) {
         ctx.addIssue({
             code: z.ZodIssueCode.custom,
@@ -312,7 +498,8 @@ export const additionalRecommendationSchema = z.preprocess(stripNullGroundingFie
                 message: "targetElements is required for testType: 'ui'. Use null when blueprint capture failed (and add '[no-blueprint-data]' to both description and reasoning).",
             });
         }
-        if (Array.isArray(rec.targetElements) && rec.pageContext === undefined) {
+        if (Array.isArray(rec.targetElements) &&
+            rec.pageContext === undefined) {
             ctx.addIssue({
                 code: z.ZodIssueCode.custom,
                 path: ["pageContext"],
@@ -347,19 +534,29 @@ export const additionalRecommendationSchema = z.preprocess(stripNullGroundingFie
 // TODO(multi-repo maintenance): no `repository` field yet — see readData() TODO below.
 const testMaintenanceSchema = z.object({
     testType: z.nativeEnum(TestType).describe("Type of test."),
-    endpoint: z.string().describe("HTTP verb and path, e.g. 'GET /api/v1/products'"),
+    endpoint: z
+        .string()
+        .describe("HTTP verb and path, e.g. 'GET /api/v1/products'"),
     testFilePath: z
         .string()
-        .refine((p) => path.isAbsolute(p), { message: "testFilePath must be an absolute path" })
+        .refine((p) => path.isAbsolute(p), {
+        message: "testFilePath must be an absolute path",
+    })
         .describe("Absolute path of the test file that was maintained, e.g. '/repo/tests/products_smoke_test.py' — the same path you passed to skyramp_execute_test's testFile param. Consumers should basename this for display."),
-    action: z.nativeEnum(DriftAction).describe("The drift action assigned to this test during maintenance triage."),
+    action: z
+        .nativeEnum(DriftAction)
+        .describe("The drift action assigned to this test during maintenance triage."),
     description: z.string().describe("What was changed and why"),
-    beforeDetails: z.string().describe("One line only — no embedded newlines, no raw HTTP headers or JSON blobs. " +
+    beforeDetails: z
+        .string()
+        .describe("One line only — no embedded newlines, no raw HTTP headers or JSON blobs. " +
         "For passing runs: count and timing, e.g. '4 passed in 15.09s'. " +
         "For failing runs: failure name and one-line root cause, e.g. " +
         "'FAILED test_foo — assert 403 got 200, auth middleware not enforced'. " +
         "Empty string for VERIFY/IGNORE entries where no before-execution was run."),
-    afterDetails: z.string().describe("One line only — no embedded newlines, no raw HTTP headers or JSON blobs. " +
+    afterDetails: z
+        .string()
+        .describe("One line only — no embedded newlines, no raw HTTP headers or JSON blobs. " +
         "For passing runs: count and timing, e.g. '5 passed in 10.96s'. " +
         "For failing runs: failure name and one-line root cause, e.g. " +
         "'FAILED test_foo — check_schema fails, order_id=1 has discount from prior PATCH test'. " +
@@ -404,7 +601,9 @@ function computeReportMetrics(params) {
     const recommendations = params.additionalRecommendations ?? [];
     const countBy = (items, pred) => items.filter(pred).length;
     const changedMaintenance = (params.testMaintenance ?? []).filter(isMaintenanceChange);
-    const maintenanceRecovered = countBy(changedMaintenance, (m) => (m.beforeStatus === TestExecutionStatus.Fail || m.beforeStatus === TestExecutionStatus.Error) && m.afterStatus === TestExecutionStatus.Pass);
+    const maintenanceRecovered = countBy(changedMaintenance, (m) => (m.beforeStatus === TestExecutionStatus.Fail ||
+        m.beforeStatus === TestExecutionStatus.Error) &&
+        m.afterStatus === TestExecutionStatus.Pass);
     return {
         testsGenerated: String(params.newTestsCreated.length),
         testsMaintained: String(changedMaintenance.length),
@@ -436,17 +635,35 @@ function computeReportMetrics(params) {
  * state: the execution fix-up can restore `<testFile>.raw.bak` over it afterwards
  * by plain `cp`, which this tool never sees. See rederiveReuseOutcome.
  *
- * Non-UI tests, tests the reuse tool never ran for, and specs whose outcome cannot
- * be re-derived all get nothing — which consumers already treat as "no reuse
- * summary". */
-async function attachReuseOutcome(test, outcomes) {
-    if (test.testType !== TestType.UI || !outcomes)
-        return test;
-    const found = outcomes[path.basename(test.fileName)];
-    if (!found)
+ * Tests the reuse tool never ran for and specs whose outcome cannot be re-derived
+ * get nothing — which consumers already treat as "no reuse summary" — with one
+ * exception: a test whose generation handed off a reuse step that was never taken
+ * gets `chainSkipped: true` (SKYR-4220), the one claim that is exactly about the
+ * reuse tool NOT having run. Applies to every test type: UI tests carry the POM
+ * fields, API tests the shared-helper ones. */
+async function attachReuseOutcome(test, outcomes, handOffs) {
+    // POM records describe browser specs; a basename collision with an API test's
+    // fileName must not attach them there. A utils-path record is attachable anywhere.
+    const record = outcomes?.[path.basename(test.fileName)];
+    // A utils record carries the test type its verify ran under: a mismatch is a
+    // basename collision with another test, not this row's outcome.
+    const typeMatches = !record?.utils?.testType || record.utils.testType === test.testType;
+    const found = record && typeMatches && (test.testType === TestType.UI || record.utils)
+        ? record
+        : undefined;
+    const derived = found ? await rederiveReuseOutcome(found) : undefined;
+    // The RAW record: a colliding record of any kind means reuse ran for this basename,
+    // and the chain claim must not be made on the filtered view.
+    const chainSkipped = await reuseChainSkipped(test.fileName, test.testType, record, handOffs);
+    const reuse = derived || chainSkipped
+        ? { ...(derived ?? {}), ...(chainSkipped ? { chainSkipped } : {}) }
+        : undefined;
+    if (!reuse)
         return test;
-    const reuse = await rederiveReuseOutcome(found);
-    return reuse ? { ...test, reuse } : test;
+    // A record that re-derives to nothing (e.g. a utils path that wrote no file) must
+    // not attach an empty object — consumers treat presence as "reuse ran".
+    const hasContent = Object.values(reuse).some((v) => v !== undefined);
+    return hasContent ? { ...test, reuse } : test;
 }
 /**
  * Attach the video recorded for this execution, matched by the row's testFilePath
@@ -460,7 +677,10 @@ async function attachReuseOutcome(test, outcomes) {
  * consumers already treat as "no recording for this test".
  */
 function attachVideoPath(row, videos) {
-    return { ...row, videoPath: videos?.[path.basename(row.testFilePath)]?.videoPath };
+    return {
+        ...row,
+        videoPath: videos?.[path.basename(row.testFilePath)]?.videoPath,
+    };
 }
 function deduplicateById(items) {
     const seen = new Set();
@@ -517,8 +737,8 @@ export function registerSubmitReportTool(server) {
                 .optional()
                 .default([])
                 .describe("Actionable follow-ups for the PR author. Each entry must be a single-line string (no embedded newlines). " +
-                "Include a next step for every critical/high severity issue in issuesFound. No next steps for low-severity issues. " +
-                "Also include a next step for every external test (Playwright, Cypress, RTL) assigned REGENERATE or DELETE in testMaintenance — the developer must act on these manually since they cannot be auto-applied (e.g. 'Regenerate frontend/tests/cart_pom.spec.ts — CartLine structure changed, all selectors need re-recording' or 'Delete frontend/tests/homepage.spec.ts — /cart route removed'). " +
+                "Do NOT add a next step that repeats an entry in issuesFound. The report renders every issue with its description and source location, and the consumer also posts an inline review comment on the cited line, so a repeated next step shows the same sentence to the author three times. Add a next step for an issue only when it names an action the issue text does not already state. " +
+                "Include a next step for every external test (Playwright, Cypress, RTL) assigned REGENERATE or DELETE in testMaintenance — the developer must act on these manually since they cannot be auto-applied (e.g. 'Regenerate frontend/tests/cart_pom.spec.ts — CartLine structure changed, all selectors need re-recording' or 'Delete frontend/tests/homepage.spec.ts — /cart route removed'). " +
                 "If multiple tests fail with 404 or connection refused: suggest checking targetSetupCommand/targetReadyCheckCommand. " +
                 "If 401/403 on auth endpoints: suggest authTokenCommand. " +
                 "When referencing code, use file name and relevant code pattern — no line numbers unless certain."),
@@ -543,6 +763,10 @@ export function registerSubmitReportTool(server) {
     }, async (params) => {
         const startTime = Date.now();
         let errorResult;
+        // Residual-leak telemetry (SKYR-4185): filled by the language guardrail
+        // below when a report language is enforced, spread into the analytics
+        // event in the finally. Empty (no properties) for non-enforced runs.
+        const languageTelemetry = {};
         // The report goes next to the state file, so in a Testbot run the state file has to
         // be in the run directory — that is where the action reads the report back from.
         // Two ways that breaks, both ending in a report nobody reads while this tool says
@@ -567,8 +791,105 @@ export function registerSubmitReportTool(server) {
                     `earlier run or one typed by hand.`);
             }
         }
+        // SKYR-4185: report-language guardrail. The language was captured when the
+        // testbot prompt was rendered (see reportLanguage.ts) — the instruction to
+        // write free-text fields in it is prose the agent sometimes ignores
+        // (letsramp/api-insight#215 shipped English testResults[].details in an
+        // otherwise-Japanese report). Reject rather than ship a half-translated
+        // report; the listed paths tell the agent exactly what to rewrite. Fields
+        // the agent does not author here (testMaintenance description from
+        // skyramp_actions verdicts, the server default commitMessage) are exempt —
+        // rejecting those would loop, since resubmitting cannot change them.
+        const reportLanguage = getReportLanguage();
+        if (reportLanguage && !isEnforcedReportLanguage(reportLanguage)) {
+            logger.warning(`Report language '${reportLanguage}' has no enforcement rules — skipping the language guardrail`);
+        }
+        if (isEnforcedReportLanguage(reportLanguage)) {
+            const textFields = [
+                { path: "businessCaseAnalysis", text: params.businessCaseAnalysis },
+            ];
+            params.newTestsCreated.forEach((t, i) => {
+                textFields.push({
+                    path: `newTestsCreated[${i}].description`,
+                    text: t.description,
+                });
+                textFields.push({
+                    path: `newTestsCreated[${i}].reasoning`,
+                    text: t.reasoning,
+                });
+            });
+            (params.additionalRecommendations ?? []).forEach((r, i) => {
+                textFields.push({
+                    path: `additionalRecommendations[${i}].description`,
+                    text: r.description,
+                });
+                textFields.push({
+                    path: `additionalRecommendations[${i}].reasoning`,
+                    text: r.reasoning,
+                });
+                r.steps.forEach((s, j) => {
+                    textFields.push({
+                        path: `additionalRecommendations[${i}].steps[${j}].description`,
+                        text: s.description,
+                    });
+                });
+            });
+            params.testResults.forEach((t, i) => {
+                textFields.push({
+                    path: `testResults[${i}].details`,
+                    text: t.details,
+                });
+            });
+            params.issuesFound.forEach((f, i) => {
+                textFields.push({
+                    path: `issuesFound[${i}].description`,
+                    text: f.description,
+                });
+            });
+            (params.nextSteps ?? []).forEach((s, i) => {
+                textFields.push({ path: `nextSteps[${i}]`, text: s });
+            });
+            (params.testMaintenanceDetails ?? []).forEach((d, i) => {
+                textFields.push({
+                    path: `testMaintenanceDetails[${i}].beforeDetails`,
+                    text: d.beforeDetails,
+                });
+                textFields.push({
+                    path: `testMaintenanceDetails[${i}].afterDetails`,
+                    text: d.afterDetails,
+                });
+            });
+            if (params.commitMessage &&
+                params.commitMessage !== DEFAULT_COMMIT_MESSAGE) {
+                textFields.push({
+                    path: "commitMessage",
+                    text: params.commitMessage,
+                });
+            }
+            const languageViolations = findLanguageViolations(textFields, reportLanguage);
+            if (languageViolations.length > 0) {
+                const languageName = reportLanguageDisplayName(reportLanguage);
+                errorResult = toolError(`This run's report language is ${languageName}, but ${languageViolations.length} free-text field(s) are written in English:\n` +
+                    languageViolations.map((v) => `  - ${v}`).join("\n") +
+                    `\nRewrite ONLY the listed fields in ${languageName} and resubmit the report. ` +
+                    `Keep code identifiers, endpoint paths, file names, test IDs, enum values ` +
+                    `(Pass/Fail/Skipped, severity values, testType), and anything inside backticks untranslated.`);
+                return errorResult;
+            }
+            // The sub-threshold band (a single stray English word) is accepted by
+            // design; count what ships so the leniency's real-world leak rate is
+            // visible in Amplitude before anyone tightens the threshold.
+            const nearMisses = findLanguageNearMisses(textFields, reportLanguage);
+            languageTelemetry.reportLanguage = reportLanguage;
+            languageTelemetry.languageNearMissCount = String(nearMisses.length);
+            if (nearMisses.length > 0) {
+                logger.info(`Report language ${reportLanguage}: accepting ${nearMisses.length} field(s) with a single stray English word`, { fields: nearMisses });
+            }
+        }
         const dedupedNewTests = deduplicateById([...params.newTestsCreated]);
-        const dedupedRecommendations = deduplicateById([...(params.additionalRecommendations ?? [])]);
+        const dedupedRecommendations = deduplicateById([
+            ...(params.additionalRecommendations ?? []),
+        ]);
         const stateManager = StateManager.fromStatePath(params.stateFile);
         let stateData;
         try {
@@ -589,7 +910,8 @@ export function registerSubmitReportTool(server) {
         // than silently reporting an empty testMaintenance section indistinguishable from a
         // real "nothing to do". With zero existing tests there was nothing to lose, so
         // undefined is safe there — no round-trip through skyramp_actions required.
-        if (stateData.maintenanceVerdicts === undefined && (stateData.existingTests?.length ?? 0) > 0) {
+        if (stateData.maintenanceVerdicts === undefined &&
+            (stateData.existingTests?.length ?? 0) > 0) {
             errorResult = toolError("stateFile has existingTests but no maintenanceVerdicts — skyramp_actions was not called for this run. " +
                 "Call skyramp_actions (with recommendations: [] if no existing tests needed action) before skyramp_submit_report.");
             return errorResult;
@@ -613,19 +935,24 @@ export function registerSubmitReportTool(server) {
                     return false;
                 const endpoints = parseEndpointField(t.endpoint);
                 const names = planNameCandidates(t.testId, t.testType);
-                return !(names.length > 0 ? names : [undefined]).some((scenarioName) => endpoints.some(({ method, path }) => matchesApprovedPlan(approvedPlan, { scenarioName, testType: t.testType, method, path })));
+                return !(names.length > 0 ? names : [undefined]).some((scenarioName) => endpoints.some(({ method, path }) => matchesApprovedPlan(approvedPlan, {
+                    scenarioName,
+                    testType: t.testType,
+                    method,
+                    path,
+                })));
             });
             if (unapproved.length > 0) {
                 const approvedList = approvedPlan.generate.length > 0
-                    // Show each item's endpoint keys: without them a rejection that turns on
-                    // the endpoint reads as self-contradicting, because the entry's own name
-                    // is printed in this same list (SKYR-4123).
-                    ? approvedPlan.generate
-                        .map((item) => {
-                        const eps = item.matchKeys.filter((k) => k.startsWith("ep:"));
-                        return `[${item.testType}] ${item.scenarioName}${eps.length > 0 ? ` covering ${eps.join(", ")}` : ""}`;
-                    })
-                        .join("; ")
+                    ? // Show each item's endpoint keys: without them a rejection that turns on
+                        // the endpoint reads as self-contradicting, because the entry's own name
+                        // is printed in this same list (SKYR-4123).
+                        approvedPlan.generate
+                            .map((item) => {
+                            const eps = item.matchKeys.filter((k) => k.startsWith("ep:"));
+                            return `[${item.testType}] ${item.scenarioName}${eps.length > 0 ? ` covering ${eps.join(", ")}` : ""}`;
+                        })
+                            .join("; ")
                     : "(none)";
                 errorResult = toolError(`${unapproved.length} newTestsCreated entr${unapproved.length === 1 ? "y" : "ies"} not in the approved plan from ` +
                     `skyramp_register_test_plan (plan ${approvedPlan.planId}) — neither its GENERATE list nor its ADDITIONAL backfill pool: ` +
@@ -676,9 +1003,14 @@ export function registerSubmitReportTool(server) {
                 const recorded = stateData.existingTests?.find((t) => testFileMatches(t.testFile, m.testFilePath));
                 const detail = params.testMaintenanceDetails?.find((d) => d.testFilePath === m.testFilePath);
                 const displayName = path.basename(m.testFilePath);
-                const defaultBeforeStatus = MAINTENANCE_CHANGE_ACTIONS.has(m.action) ? TestExecutionStatus.Unknown : TestExecutionStatus.Skipped;
-                const defaultAfterStatus = m.action === DriftAction.Delete ? TestExecutionStatus.Skipped
-                    : MAINTENANCE_CHANGE_ACTIONS.has(m.action) ? TestExecutionStatus.Unknown : TestExecutionStatus.Skipped;
+                const defaultBeforeStatus = MAINTENANCE_CHANGE_ACTIONS.has(m.action)
+                    ? TestExecutionStatus.Unknown
+                    : TestExecutionStatus.Skipped;
+                const defaultAfterStatus = m.action === DriftAction.Delete
+                    ? TestExecutionStatus.Skipped
+                    : MAINTENANCE_CHANGE_ACTIONS.has(m.action)
+                        ? TestExecutionStatus.Unknown
+                        : TestExecutionStatus.Skipped;
                 const beforeStatus = recorded?.executionBefore?.status ?? defaultBeforeStatus;
                 const afterStatus = recorded?.executionAfter?.status ?? defaultAfterStatus;
                 // Trim before checking — a whitespace-only string is semantically blank and
@@ -690,7 +1022,13 @@ export function registerSubmitReportTool(server) {
                 if (recorded?.executionAfter && !afterDetails)
                     missingDetails.push(`${displayName} (afterDetails)`);
                 logger.info(`${displayName}: before=${beforeStatus} after=${afterStatus}`);
-                return { ...m, beforeDetails, afterDetails, beforeStatus, afterStatus };
+                return {
+                    ...m,
+                    beforeDetails,
+                    afterDetails,
+                    beforeStatus,
+                    afterStatus,
+                };
             });
         }
         if (missingDetails.length > 0) {
@@ -710,8 +1048,24 @@ export function registerSubmitReportTool(server) {
             const repoRoot = fullState?.metadata?.repositoryPath;
             if (repoRoot && repoRoot !== "unknown") {
                 try {
-                    const changedFiles = await listChangedFiles(repoRoot);
-                    const unbacked = findUnbackedClaims({
+                    // A run writes tests into up to three kinds of trees: the primary
+                    // checkout, the tests-repo checkout (testRepoPath), and related
+                    // repos. Scanning only the primary rejected a real, executed UI
+                    // test in the tests-repo checkout as "unbacked", forcing the agent
+                    // to demote it to additionalRecommendations (SKYR-4204 reopen,
+                    // run 32510028428).
+                    // The primary scan keeps its throwing form: an unscannable primary
+                    // degrades the whole check to a pass (SKYR-3883 semantics) via the
+                    // catch below. The extra trees are best-effort — a missing or
+                    // non-git one contributes nothing.
+                    const changedFiles = [
+                        ...(await listChangedFiles(repoRoot)),
+                        ...(await listChangedFilesAcross([
+                            getTestsRepoDir(),
+                            ...Object.values(fullState?.relatedRepos ?? {}).map((section) => section.repositoryPath),
+                        ])),
+                    ];
+                    const unbacked = findUnchangedFileClaims({
                         repoRoot,
                         changedFiles,
                         newTests: dedupedNewTests,
@@ -730,7 +1084,10 @@ export function registerSubmitReportTool(server) {
                         // SKYR-4129. Do NOT offer re-running skyramp_actions to restate the verdict:
                         // this branch only fires when the edit is genuinely absent, so rewriting the
                         // record to match that would be the tamper path, not the fix.
-                        const changedList = changedFiles.slice(0, 20).map((f) => `  - ${f}`).join("\n");
+                        const changedList = changedFiles
+                            .slice(0, 20)
+                            .map((f) => `  - ${f}`)
+                            .join("\n");
                         errorResult = toolError(`${unbacked.length} report claim(s) are not backed by any change in the working tree — ` +
                             `Testbot will not report file work that hasn't actually been made. Do NOT make a token edit to the claimed file just to satisfy this check.\n` +
                             `For a newTestsCreated claim: create the file, correct the claim's fileName to the file you actually created (see the changed files below), or remove the claim.\n` +
@@ -747,12 +1104,36 @@ export function registerSubmitReportTool(server) {
                 }
             }
         }
+        // Validated against the tool INPUT only, never the state file (a state-file
+        // guard here previously caused a rejection loop). Only entries carrying a
+        // citation are checked, so an ordinary run pays no extra cost.
+        // `repository` is normalized BEFORE the check so the attribution the check
+        // reconciles is the one the report ships, and so the stamp below survives
+        // into the written file — this array is what the report is built from.
+        const issuesFound = params.issuesFound.map(normalizeRepository);
+        if (issuesFound.some((issue) => issue.sourceFile?.trim())) {
+            const citations = await findInvalidSourceCitations({
+                checkouts: await stateManager.listRepoCheckouts(),
+                issues: issuesFound,
+            });
+            if (citations.invalid.length > 0) {
+                errorResult = toolError(citations.invalid.join("\n"));
+                return errorResult;
+            }
+            // The checkout the citation resolved in owns the finding, so the report
+            // states it. Left to the consumer, an unattributed related-repo path is
+            // resolved against the primary checkout and links the wrong repo's file.
+            citations.repository.forEach((repo, i) => {
+                if (repo)
+                    issuesFound[i] = { ...issuesFound[i], repository: repo };
+            });
+        }
         // Strip generation-artifact fields from newTestsCreated before writing.
         // scenarioFile, traceFile, frontendTrace are internal paths used during
         // generation — downstream scoring scripts don't expect them and fail if
         // they encounter these string fields while traversing the object.
         // Also normalize each item's `repository` (blank → undefined).
-        const sanitizedNewTests = await Promise.all(dedupedNewTests.map(({ scenarioFile: _sf, traceFile: _tf, frontendTrace: _ft, ...rest }) => attachReuseOutcome(normalizeRepository(rest), stateData.reuseOutcomes)));
+        const sanitizedNewTests = await Promise.all(dedupedNewTests.map(({ scenarioFile: _sf, traceFile: _tf, frontendTrace: _ft, ...rest }) => attachReuseOutcome(normalizeRepository(rest), stateData.reuseOutcomes, stateData.reuseHandOffs)));
         const report = {
             businessCaseAnalysis: params.businessCaseAnalysis,
             newTestsCreated: sanitizedNewTests,
@@ -773,9 +1154,10 @@ export function registerSubmitReportTool(server) {
                 const { testFilePath: _tfp, ...wire } = attachVideoPath(normalizeRepository(row), stateData.executionVideos);
                 return wire;
             }),
-            issuesFound: params.issuesFound.map(normalizeRepository),
+            issuesFound,
             nextSteps: params.nextSteps ?? [],
-            commitMessage: (params.commitMessage ?? "").replace(/[\r\n]+/g, " ").trim() || DEFAULT_COMMIT_MESSAGE,
+            commitMessage: (params.commitMessage ?? "").replace(/[\r\n]+/g, " ").trim() ||
+                DEFAULT_COMMIT_MESSAGE,
         };
         const reportJson = JSON.stringify(report, null, 2);
         // Beside the state file, which was read successfully above — so this directory is
@@ -815,6 +1197,7 @@ export function registerSubmitReportTool(server) {
                 testResultCount: String(params.testResults.length),
                 payloadBytes: String(reportJson.length),
                 ...computeReportMetrics({ ...params, testMaintenance }),
+                ...languageTelemetry,
             }).catch(() => { });
         }
     });