@tea-agent/loop-agent 0.39.0-next.2 → 0.39.0-next.21

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

@@ -3,7 +3,8 @@ import path from "node:path";
 import { generateTaskDagUseCase } from "../../application/dag/generate-task-dag.js";
 import { writeJsonAtomic } from "../../infrastructure/harness/atomic-write.js";
 import { DAG_RUNS_DIR, isTerminalDagRunStatus, locateDagRun, readDagRunSpec, readDagRunState, } from "./lifecycle.js";
-import { parseDagSpec } from "./types.js";
+import { deriveDagRerunFeedback } from "./rerun-feedback.js";
+import { parseDagSpec, } from "./types.js";
 export const RERUN_TASK_LINEAGE_REL_PATH = ".runtime/rerun-task-lineage.json";
 export function parseDagRerunTaskArgs(args) {
     if (args.length === 0) {
@@ -218,6 +219,7 @@ async function writeRerunTaskLineage(cwd, newRunId, input) {
         parentRunId: input.parentRunId,
         reason: input.reason,
         requestId: input.requestId,
+        feedback: input.feedback,
         createdAt: new Date().toISOString(),
     }, { repoRoot: cwd, allowCompletedFactsWrite: true });
 }
@@ -228,12 +230,30 @@ function blockedResult(input) {
         reason: input.reason,
         requestId: input.requestId,
         mode: "blocked",
+        feedback: input.feedback ?? {
+            inherited: false,
+            reason: "no-unresolved-verify-review-feedback",
+        },
         blockedReason: input.blockedReason,
         ...(input.note ? { note: input.note } : {}),
         ...(input.taskId ? { taskId: input.taskId } : {}),
         ...(input.conflictPaths ? { conflictPaths: input.conflictPaths } : {}),
     };
 }
+async function readGeneratedRerunFeedback(outputPath) {
+    const raw = JSON.parse(await readFile(outputPath, "utf8"));
+    return parseDagSpec(raw).rerunFeedback;
+}
+function summarizeFeedback(feedback, fallbackReason) {
+    return feedback
+        ? {
+            inherited: true,
+            sourceNodeId: feedback.sourceNodeId,
+            digest: feedback.digest,
+            bindingStatus: feedback.binding.status,
+        }
+        : { inherited: false, reason: fallbackReason };
+}
 function extractRunSummary(generateResult) {
     if (generateResult.mode === "generate+basic-validate")
         return undefined;
@@ -277,7 +297,26 @@ export async function rerunStandaloneTask(input) {
             note: eligibility.note,
         });
     }
+    let parentFeedback;
+    let missingFeedbackReason = "no-unresolved-verify-review-feedback";
+    try {
+        parentFeedback = await deriveDagRerunFeedback({
+            parentRunId: input.parentRunId,
+            taskId: taskResolution.taskId,
+            runDir: located.runDir,
+            spec,
+            state,
+        });
+    }
+    catch {
+        // Feedback is additive recovery context. A corrupt/missing historical
+        // artifact must be reported honestly but must not block an otherwise
+        // eligible full rerun.
+        missingFeedbackReason = "feedback-unavailable";
+    }
     const execute = input.execute ?? true;
+    let effectiveFeedback;
+    let expectedRerunFeedbackDigest;
     // AC-001 generated-output-conflict preflight: when a real execution is
     // requested, first generate the DAG draft (execute:false) so we can inspect
     // the regenerated spec's requireChangedFiles writer write sets against
@@ -296,15 +335,23 @@ export async function rerunStandaloneTask(input) {
             dryRun: false,
             profile: input.profile ?? "auto",
             profileExplicit: input.profileExplicit ?? false,
+            rerunFeedback: parentFeedback,
         });
         if (preflight.mode === "generate+basic-validate") {
             let conflictPaths = [];
             try {
                 const raw = await readFile(preflight.outputPath, "utf8");
                 const preflightSpec = parseDagSpec(JSON.parse(raw));
+                effectiveFeedback = preflightSpec.rerunFeedback;
+                expectedRerunFeedbackDigest = effectiveFeedback?.digest;
+                if (parentFeedback && !effectiveFeedback) {
+                    throw new Error("rerun feedback missing from generated preflight DAG");
+                }
                 conflictPaths = await assessGeneratedOutputConflict(preflightSpec, input.cwd);
             }
-            catch {
+            catch (error) {
+                if (parentFeedback)
+                    throw error;
                 // If the draft cannot be parsed, fall through to the real execution
                 // path: execution's own validation will surface the structural defect.
                 conflictPaths = [];
@@ -332,7 +379,17 @@ export async function rerunStandaloneTask(input) {
         dryRun: false,
         profile: input.profile ?? "auto",
         profileExplicit: input.profileExplicit ?? false,
+        rerunFeedback: parentFeedback,
+        expectedRerunFeedbackDigest,
     });
+    if (parentFeedback && !effectiveFeedback) {
+        effectiveFeedback = await readGeneratedRerunFeedback(generateResult.outputPath);
+    }
+    if (expectedRerunFeedbackDigest &&
+        effectiveFeedback?.digest !== expectedRerunFeedbackDigest) {
+        throw new Error(`rerun feedback drift after generation: expected ${expectedRerunFeedbackDigest}, got ${effectiveFeedback?.digest ?? "missing"}`);
+    }
+    const feedbackSummary = summarizeFeedback(effectiveFeedback, missingFeedbackReason);
     const newRunId = extractNewRunId(generateResult);
     const runSummary = extractRunSummary(generateResult);
     const completedSuccessfully = !execute ||
@@ -342,6 +399,7 @@ export async function rerunStandaloneTask(input) {
             parentRunId: input.parentRunId,
             reason: input.reason,
             requestId: input.requestId,
+            feedback: feedbackSummary,
         });
     }
     const result = {
@@ -350,6 +408,7 @@ export async function rerunStandaloneTask(input) {
         reason: input.reason,
         requestId: input.requestId,
         mode: "generate+validate+execute",
+        feedback: feedbackSummary,
         taskId: taskResolution.taskId,
         ...(newRunId ? { newRunId } : {}),
         ...(runSummary && !completedSuccessfully && "status" in runSummary

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

@@ -21,6 +21,8 @@ export const STRUCTURED_OUTPUT_RETRY_CATEGORY = "output-too-large";
 export const PROTOCOL_INVALID_RETRY_CATEGORY = "protocol-invalid";
 /** Recoverable model artifact/schema formatting failure on read-only structured nodes. */
 export const STRUCTURED_ARTIFACT_INVALID_RETRY_CATEGORY = "invalid-output";
+/** Provider stopReason=length truncated the response before the JSON contract completed. */
+export const STRUCTURED_OUTPUT_TRUNCATED_RETRY_CATEGORY = "structured-output-truncated";
 /** Retry only a proven no-op from an explicitly opt-in bounded Pi writer. */
 export const WRITER_EMPTY_DIFF_RETRY_CATEGORY = "writer-empty-diff";
 /** Retry when a backend-test writer finished but Completeness Gate found missing/broken targets. */
@@ -36,6 +38,7 @@ export const STRUCTURED_REQUIRED_DAG_RETRY_CATEGORIES = [
     ...DEFAULT_DAG_RETRY_CATEGORIES,
     STRUCTURED_OUTPUT_RETRY_CATEGORY,
     STRUCTURED_ARTIFACT_INVALID_RETRY_CATEGORY,
+    STRUCTURED_OUTPUT_TRUNCATED_RETRY_CATEGORY,
 ];
 /** Categories allowed on nodes that declare a machine-readable outputProtocol. */
 export const PROTOCOL_AWARE_DAG_RETRY_CATEGORIES = [
@@ -48,6 +51,7 @@ export const ALL_DAG_RETRY_CATEGORIES = [
     STRUCTURED_OUTPUT_RETRY_CATEGORY,
     PROTOCOL_INVALID_RETRY_CATEGORY,
     STRUCTURED_ARTIFACT_INVALID_RETRY_CATEGORY,
+    STRUCTURED_OUTPUT_TRUNCATED_RETRY_CATEGORY,
     WRITER_EMPTY_DIFF_RETRY_CATEGORY,
     INCOMPLETE_WRITE_SET_RETRY_CATEGORY,
     WRITER_CLEAN_TIMEOUT_RETRY_CATEGORY,
@@ -197,6 +201,24 @@ function isExclusivePiImplementer(task) {
         !task.dynamicCondition &&
         !task.dynamicLoopUntil);
 }
+/**
+ * Provider/transport failure categories that can crash a writer executor
+ * after some workspace writes already landed on disk. A continuation may
+ * restart that writer itself; it must never import or normalize the ERROR fact.
+ */
+export const WRITER_TRANSPORT_FAILURE_CATEGORIES = new Set([
+    "unavailable",
+    "network",
+    "rate-limit",
+    "timeout",
+]);
+export function isTransportCrashedExclusiveWriter(task, failureCategory) {
+    if (!task)
+        return false;
+    return (isExclusivePiImplementer(task) &&
+        failureCategory !== undefined &&
+        WRITER_TRANSPORT_FAILURE_CATEGORIES.has(failureCategory));
+}
 /**
  * Static eligibility for the requireChangedFiles writer-empty-diff class.
  * The Pi executor assigns writer-empty-diff only after post-write-guard

package/dist/workflows/dag/types.js

@@ -193,12 +193,38 @@ export const dagFrontendPrewriteGateSchema = z.object({
     requireSourceFreshness: z.literal(true),
     implementationWriteSet: z.array(z.string().min(1)).min(1).optional(),
     openspecCandidatePaths: z.array(openspecCandidatePathSchema).default([]),
+    /**
+     * Generation-frozen component/theme specification bucket. Derived from
+     * `designEvidence.classified.component ∪ theme ∪ rule.components` at DAG
+     * generation. When non-empty and the canonical contract has UI-visible
+     * work, the prewrite gate requires a non-empty `uiComponentChoices` and
+     * cross-checks each `specReference.path` against this set + the effective
+     * plan openspec-citations block. Legacy DAGs without this field keep the
+     * field-absent semantics (component-choices-missing is not enforced).
+     */
+    componentSpecCandidatePaths: z.array(openspecCandidatePathSchema).optional(),
     /**
      * OpenSpec gate policy. Omitted on legacy DAGs: the runtime treats
      * undefined as scan-strict so pre-existing full-read semantics are
      * preserved. Generation always writes it explicitly for new DAGs.
      */
     openspecPolicy: z.enum(["cited", "scan-strict"]).optional(),
+    /**
+     * When true, the effective plan node (planFromNodeId when it is selected
+     * over the fallback) emits an RFC 7386 merge-patch delta instead of a full
+     * contract. The gate applies it on the original contract node
+     * (planFallbackFromNodeIds[0]) before analysis/materialization. Legacy DAGs
+     * without this field keep full-contract semantics.
+     */
+    revisionPatch: z.boolean().optional(),
+    /**
+     * Deterministically rendered plan.md artifact name (same outputDir as the
+     * contract). Defaults to frontend-plan.md at runtime.
+     */
+    planMdArtifactName: z
+        .string()
+        .regex(/^[a-z0-9][a-z0-9._-]*\.md$/)
+        .optional(),
     /**
      * Generation-frozen candidate source metadata (diagnostic). cited mode
      * candidates are declared ∪ taskSourceCited; scan-strict candidates are
@@ -397,6 +423,8 @@ export const dagBackendTestPipelineSchema = z.enum([
     "markdown-scenario-param",
     "markdown-execute-html",
     "markdown-manifest",
+    "markdown-report",
+    "ingest-backend-test-gap",
 ]);
 export const dagFrontendBrowserToolPreflightSchema = z.object({}).strict();
 export const dagFrontendTestEvidenceValidationSchema = z.object({}).strict();
@@ -574,6 +602,13 @@ export const dagDynamicExpansionSchema = z.object({
         maxTotalTokens: z.number().int().positive().optional(),
     })
         .optional(),
+    /**
+     * When true, any child blocked by the aggregate token budget fails the map
+     * barrier instead of being retained as a case-level blocked outcome.
+     * Backend-test enables this because partial Markdown/pytest module sets are
+     * invalid generation assets; frontend-test keeps the default false.
+     */
+    failOnTokenBudgetExhaustion: z.boolean().optional(),
     /**
      * When true, map child ERROR/auth/timeout is recorded as case-level
      * failed/blocked evidence and the map barrier still succeeds (frontend-test).
@@ -678,6 +713,8 @@ export const dagConvergenceSpecSchema = z
  */
 export const structuredContractOutputSchemaIds = [
     "frontend-implementation-contract-v1",
+    "frontend-implementation-contract-plan-patch-v1",
+    "frontend-implementation-contract-revision-patch-v1",
 ];
 export const dagTaskSchema = z.object({ id: z.string().regex(/^[a-z][a-z0-9-]*$/, "task id must be kebab-case"),
     depends_on: z.array(z.string()).default([]),
@@ -739,8 +776,25 @@ export const dagTaskSchema = z.object({ id: z.string().regex(/^[a-z][a-z0-9-]*$/
         .object({
         schemaId: z.enum(structuredContractOutputSchemaIds),
         retryOnInvalid: z.boolean().default(true),
+        /**
+         * Optional deterministic base object for patch-producing structured
+         * nodes. The frontend plan patch validator applies the model delta to
+         * this run-owned skeleton before strict contract validation.
+         */
+        skeleton: z.record(z.string(), z.unknown()).optional(),
     })
         .strict()
+        .superRefine((value, ctx) => {
+        if (value.schemaId ===
+            "frontend-implementation-contract-plan-patch-v1" &&
+            !value.skeleton) {
+            ctx.addIssue({
+                code: "custom",
+                message: "frontend plan patch output requires a runtime skeleton",
+                path: ["skeleton"],
+            });
+        }
+    })
         .optional(),
     /**
      * Fail-closed outcome/diff consistency contract for bounded Pi writers.
@@ -819,6 +873,55 @@ export const dagSourceBindingSchema = z.object({
         .min(1),
     requirementIds: z.array(z.string().regex(/^(?:REQ|BR|AC)-[A-Z0-9]+(?:-[A-Z0-9]+)*$/)),
 });
+const dagRerunFeedbackEvidencePathSchema = z
+    .string()
+    .min(1)
+    .refine((value) => !/^(?:[A-Za-z]:[\\/]|[\\/]{2}|\/)/.test(value), "rerun feedback evidence paths must be repository-relative");
+/**
+ * Bounded unresolved verify/review feedback carried by a full task rerun.
+ * This is recovery context, not a mutation of the original task contract.
+ */
+export const dagRerunFeedbackSchema = z
+    .object({
+    schemaVersion: z.literal(1),
+    kind: z.literal("unresolved-terminal-feedback"),
+    parentRunId: z.string().min(1),
+    taskId: z.string().min(1),
+    sourceNodeId: z.string().min(1),
+    verdict: z.enum(["request-revision", "fail"]).optional(),
+    failureCategory: z.string().min(1).optional(),
+    feedbackText: z.string().min(1).max(12_000),
+    evidenceRefs: z
+        .array(z
+        .object({
+        nodeId: z.string().min(1),
+        preservedNodeRecordPath: dagRerunFeedbackEvidencePathSchema,
+    })
+        .strict())
+        .max(8),
+    attemptSummary: z
+        .object({
+        completedPasses: z.number().int().positive(),
+        maxPasses: z.number().int().positive(),
+        terminalReason: z.string().min(1).optional(),
+    })
+        .strict(),
+    binding: z
+        .object({
+        sourceCanonicalHash: z
+            .string()
+            .regex(/^[a-f0-9]{64}$/)
+            .optional(),
+        taskContractCanonicalHash: z
+            .string()
+            .regex(/^[a-f0-9]{64}$/)
+            .optional(),
+        status: z.enum(["exact", "changed", "unknown"]),
+    })
+        .strict(),
+    digest: z.string().regex(/^[a-f0-9]{64}$/),
+})
+    .strict();
 export const dagSpecSchema = z
     .object({
     version: dagVersionSchema,
@@ -830,6 +933,28 @@ export const dagSpecSchema = z
     sourceBinding: dagSourceBindingSchema.optional(),
     /** v4 managed Task Contract binding; only valid on version 4. */
     taskContractBinding: dagTaskContractBindingSchema.optional(),
+    /** Optional unresolved parent-run feedback for a full task rerun. */
+    rerunFeedback: dagRerunFeedbackSchema.optional(),
+    /** Frozen backend-test artifact layout (plan A); absent = historical default testcase/ layout. */
+    backendTestLayout: z
+        .object({
+        testRoot: z.string(),
+        markdownRel: z.string(),
+        scriptRel: z.string(),
+        markdownDir: z.string(),
+        scriptDir: z.string(),
+        readmePath: z.string(),
+        isDefault: z.boolean(),
+    })
+        .optional(),
+    /** Bound user shared-setup document (plan B); absent = no shared setup references expected. */
+    backendTestSharedSetup: z
+        .object({
+        path: z.string(),
+        readPath: z.string(),
+        sha256: z.string(),
+    })
+        .optional(),
     outputLanguage: dagOutputLanguageSchema.optional(),
     objective: z.string().optional(),
     successCriteria: z.array(z.string()).optional(),