@tea-agent/loop-agent 0.42.0-next.8 → 0.42.0

package/dist/executors/dag-pi-executor.js

@@ -12,25 +12,255 @@ import { createPiReadBudgetCustomTools, } from "./pi-read-budget-policy.js";
 import { cleanupPlaywrightCliDefaultSession, createPlaywrightCliTool, PI_COMMAND_CAPABILITY_REGISTRY, resolveCaseIdFromWriteSet, resolveEvidenceDirFromWriteSet, } from "./pi-playwright-cli-tool.js";
 import { dagCommandPolicyAllows, resolveDagCommandPolicy, } from "../workflows/dag/types.js";
 import { parseLedgerJson } from "../task/source-prepare/ledger.js";
+import { splitDocumentIntoFragments, sha256Text, } from "../task/source-prepare/fragment-inventory.js";
 import { redactPromptForLog, truncateOutput, } from "../shared/output-truncation.js";
 import { GitStatusUnavailableError, pathsChangedDuringRun, readGitStatusPorcelain, recoverRootNulArtifact, snapshotGitStatusPathFingerprints, snapshotGitStatusPorcelain, validateShellWriteGuard, } from "./shell-write-guard.js";
 import { captureWorkspaceWriteSnapshot, diffWorkspaceWriteSnapshots, } from "./workspace-write-snapshot.js";
 import { pathMatchesPattern } from "../shared/git-progress.js";
-import { isTargetTemplateTransientRetryNode, isWriterEmptyDiffRetryCandidate, isWriterTransportRetryCandidate, INCOMPLETE_WRITE_SET_RETRY_CATEGORY, WRITER_CLEAN_TIMEOUT_RETRY_CATEGORY, WRITER_EMPTY_DIFF_RETRY_CATEGORY, } from "../workflows/dag/retry-policy.js";
+import { readTypedEventStoreFromJsonl, typedEventPayloadSha256, } from "../workflows/dag/frontend-typed-event-store.js";
+import { collectCanonicalStateFlowNames, frontendEvidenceExpectationSchema, resolveFrontendContractRequirements, validateFrontendRequiredDeliverables, } from "../workflows/dag/frontend-contract-facts.js";
+import { isTargetTemplateTransientRetryNode, isWriterEmptyDiffRetryCandidate, isWriterTransportRetryCandidate, INCOMPLETE_WRITE_SET_RETRY_CATEGORY, OUTPUT_LIMIT_RETRY_CATEGORY, STRUCTURED_OUTPUT_RETRY_CATEGORY, WRITER_CLEAN_TIMEOUT_RETRY_CATEGORY, WRITER_EMPTY_DIFF_RETRY_CATEGORY, } from "../workflows/dag/retry-policy.js";
 import { assessBackendTestMdPlanCompleteness, assessBackendTestMdWriterCompleteness, assessBackendTestPytestPlanCompleteness, assessBackendTestPytestWriterCompleteness, assessBackendTestShardChildCompleteness, backendTestWriterProgressRoleForTask, classifyBackendTestWriterCompletenessFailure, isBackendTestCompletenessRetryCandidate, isBackendTestMdPlanTask, isBackendTestPytestCollectionRepairOutcomeRecoveryCandidate, isBackendTestPytestPlanTask, isBackendTestShardChildTask, writeBackendTestWriterProgressArtifacts, } from "../workflows/dag/backend-test-writer-completeness.js";
 import { resolveBackendTestLayout } from "../workflows/dag/backend-test-layout.js";
+import { assessBackendTestPlanProtocol } from "../workflows/dag/backend-test-plan-protocol.js";
 import { frontendTestLayoutFromSpec } from "../workflows/dag/frontend-test-layout.js";
 import { redactSecrets, truncateUtf8Preview } from "../shared/preview.js";
 import { writeEffectiveContextReceipt } from "../workflows/dag/context-receipt.js";
 /**
- * Writer classification for a length-stopped thinking-only attempt: the model
- * exhausted its output budget thinking but never issued a write/edit tool call
- * and produced zero attributed diff. This is a terminal, non-retryable
- * diagnosis (the same prompt + model will hit the same budget wall); the
- * recommendation is a model switch plus a fresh run. It must NOT mask a
- * recoverable partial-write-set (incomplete-write-set) upgrade.
+ * Legacy diagnostic label retained for artifact compatibility. New executions
+ * classify this signal as output-limit so the node can retry incrementally.
  */
 export const WRITER_THINKING_EXHAUSTED_CATEGORY = "writer-thinking-exhausted";
+/**
+ * Legacy diagnostic label retained for artifact compatibility. New executions
+ * classify this signal as output-limit and preserve committed typed facts.
+ */
+export const PLANNER_THINKING_EXHAUSTED_CATEGORY = "planner-thinking-exhausted";
+/**
+ * Parallel coverage shards namespace their verification target ids with
+ * `VT-SHARD-<shard>-` so the reducer can detect cross-shard conflicts. Once
+ * every shard's facts are known the namespace must be restored: the canonical
+ * contract has to carry the exact ids the task source froze (a surviving
+ * `VT-SHARD-N-…` id is a contract-requirement gap that design review
+ * rejects). Stripping happens per shard before adoption — after the strip,
+ * the existing identity-conflict check catches genuine cross-shard
+ * duplicate base ids and fails the merge closed.
+ */
+/**
+ * Deterministic pre-reduce for parallel coverage shard facts. Shards partition
+ * requirements, but a frozen verification target can legitimately be derived
+ * by several shards (one VT covers multiple ACs across shard boundaries), so
+ * same-id verification targets are MERGED: requirementIds and uiStates union,
+ * while divergent file/commandId is a real conflict. Everything else passes
+ * through unchanged.
+ */
+export function reduceParallelCoverageShardRecords(records) {
+    // Resolve replacements inside each source shard before comparing shards.
+    // A correction is an event-log operation, not a second independent target;
+    // retaining both records makes a valid same-shard correction look like a
+    // cross-shard conflict during the later reduction.
+    const byShard = new Map();
+    for (const record of records) {
+        const shard = byShard.get(record.attemptId) ?? [];
+        shard.push(record);
+        byShard.set(record.attemptId, shard);
+    }
+    const effectiveRecords = [];
+    for (const shardRecords of byShard.values()) {
+        const effective = [];
+        for (const record of shardRecords) {
+            const fact = record.fact;
+            const entry = fact?.entry;
+            const kind = typeof fact?.kind === "string" ? fact.kind : "";
+            const identity = (kind === "plan-requirement" || kind === "plan-verification-target") &&
+                typeof entry?.id === "string"
+                ? `${kind}:${entry.id}`
+                : undefined;
+            if (!identity) {
+                effective.push(record);
+                continue;
+            }
+            const replaces = typeof fact?.replaces === "string" ? `${kind}:${fact.replaces}` : undefined;
+            if (replaces) {
+                for (let index = effective.length - 1; index >= 0; index -= 1) {
+                    const prior = effective[index];
+                    const priorFact = prior.fact;
+                    const priorEntry = priorFact?.entry;
+                    const priorIdentity = typeof priorFact?.kind === "string" &&
+                        typeof priorEntry?.id === "string"
+                        ? `${priorFact.kind}:${priorEntry.id}`
+                        : undefined;
+                    if (priorIdentity === replaces)
+                        effective.splice(index, 1);
+                }
+            }
+            // Keep the source record immutable. The replacement is represented by
+            // the latest event and its own payload hash.
+            effective.push(record);
+        }
+        effectiveRecords.push(...effective);
+    }
+    const verificationTargets = new Map();
+    const merged = [];
+    for (const record of effectiveRecords) {
+        const fact = record.fact;
+        if (!fact || typeof fact !== "object")
+            continue;
+        if (fact.kind !== "plan-verification-target") {
+            merged.push(record);
+            continue;
+        }
+        const entry = fact.entry;
+        if (!entry || typeof entry.id !== "string")
+            continue;
+        const existing = verificationTargets.get(entry.id);
+        if (!existing) {
+            verificationTargets.set(entry.id, {
+                record, entry: { ...entry },
+                sources: [`${record.attemptId}:${record.eventId}:${record.payloadSha256}`],
+            });
+            continue;
+        }
+        // Never mutate the entry held by the source record. The reducer creates
+        // a new merged entry and recomputes the payload hash for that derived
+        // fact, leaving source-event replay integrity intact.
+        const baseEntry = { ...existing.entry };
+        for (const field of ["commandId", "commandLabel", "file", "scope"]) {
+            if (entry[field] !== baseEntry[field]) {
+                throw new Error(`frontend plan coverage shard reduce conflict: verification target ${entry.id} has divergent ${field} "${String(entry[field])}" vs "${String(baseEntry[field])}"`);
+            }
+        }
+        const unionSorted = (a, b) => {
+            const left = Array.isArray(a) ? a : [];
+            const right = Array.isArray(b) ? b : [];
+            return [...new Set([...left, ...right])].sort();
+        };
+        baseEntry.requirementIds = unionSorted(baseEntry.requirementIds, entry.requirementIds);
+        baseEntry.uiStates = unionSorted(baseEntry.uiStates, entry.uiStates);
+        const mergedFact = { ...fact, entry: { ...baseEntry } };
+        existing.entry = baseEntry;
+        existing.sources.push(`${record.attemptId}:${record.eventId}:${record.payloadSha256}`);
+        existing.record = {
+            ...existing.record,
+            fact: mergedFact,
+            payloadSha256: typedEventPayloadSha256(mergedFact),
+        };
+    }
+    merged.push(...[...verificationTargets.values()].map((item) => {
+        // The reducer owns every VT revision, including a one-shard partial
+        // result. Content-addressed revisions replay idempotently, while an
+        // expanded/replaced source set appends an explicit replacement rather
+        // than masquerading as a changed source event or deleting audit facts.
+        const fact = {
+            ...item.record.fact,
+            replaces: item.entry.id,
+            entry: {
+                ...item.entry,
+                requirementIds: [...new Set(planFactStringList(item.entry.requirementIds))].sort(),
+                uiStates: [...new Set(planFactStringList(item.entry.uiStates))].sort(),
+            },
+        };
+        const payloadSha256 = typedEventPayloadSha256(fact);
+        const revisionId = createHash("sha256")
+            .update(JSON.stringify({ sources: [...new Set(item.sources)].sort(), payloadSha256 }))
+            .digest("hex");
+        return {
+            ...item.record,
+            attemptId: "frontend-plan-coverage-reducer",
+            eventId: `coverage-reduced:${revisionId}`,
+            requestId: `coverage-reduced:${revisionId}`,
+            fact, payloadSha256,
+        };
+    }));
+    return merged;
+}
+export function normalizeParallelCoverageShardRecords(records, shardNumber) {
+    const prefix = `VT-SHARD-${shardNumber}-`;
+    // Models sometimes drop the `VT-` stem when applying the shard namespace
+    // (observed: frozen `VT-X` became `VT-SHARD-2-X`), so restoring the
+    // canonical id requires re-adding the stem after the strip.
+    const stripId = (id) => {
+        if (typeof id !== "string" || !id.startsWith(prefix))
+            return id;
+        const base = id.slice(prefix.length);
+        return base.startsWith("VT-") ? base : `VT-${base}`;
+    };
+    const stripIdList = (ids) => Array.isArray(ids) ? ids.map((id) => stripId(id)) : ids;
+    return records.map((record) => {
+        const fact = record.fact;
+        if (!fact || typeof fact !== "object")
+            return record;
+        const kind = fact.kind;
+        const entry = fact.entry;
+        if (!entry || typeof entry !== "object")
+            return record;
+        let rewritten;
+        const replaces = kind === "plan-verification-target" ? stripId(fact.replaces) : fact.replaces;
+        if (kind === "plan-verification-target") {
+            const id = stripId(entry.id);
+            if (id !== entry.id || replaces !== fact.replaces)
+                rewritten = { ...entry, id };
+        }
+        else if (kind === "plan-requirement") {
+            const verificationTargetIds = stripIdList(entry.verificationTargetIds);
+            if (verificationTargetIds !== entry.verificationTargetIds) {
+                rewritten = { ...entry, verificationTargetIds };
+            }
+        }
+        else if (kind === "state-flow") {
+            const rewriteBoundTargets = (item) => {
+                if (!item || typeof item !== "object")
+                    return item;
+                return {
+                    ...item,
+                    verificationTargetIds: stripIdList(item.verificationTargetIds),
+                };
+            };
+            const uiStates = Array.isArray(entry.uiStates)
+                ? entry.uiStates.map(rewriteBoundTargets)
+                : entry.uiStates;
+            const interactions = Array.isArray(entry.interactions)
+                ? entry.interactions.map(rewriteBoundTargets)
+                : entry.interactions;
+            if (uiStates !== entry.uiStates || interactions !== entry.interactions) {
+                rewritten = { ...entry, uiStates, interactions };
+            }
+        }
+        return rewritten
+            ? {
+                ...record,
+                fact: { ...fact, ...(replaces !== undefined ? { replaces } : {}), entry: rewritten },
+                // Keep the integrity hash consistent with the rewritten
+                // payload, or the adoption replay check reports the
+                // normalized record as a tampered source event.
+                payloadSha256: typedEventPayloadSha256({
+                    ...fact,
+                    ...(replaces !== undefined ? { replaces } : {}),
+                    entry: rewritten,
+                }),
+            }
+            : record;
+    });
+}
+export function isPlannerThinkingExhausted(result, committedAnyFacts) {
+    if (result.ok)
+        return false;
+    const evidence = readWriterThinkingExhaustionEvidence(result);
+    if (evidence.stopReason !== "length")
+        return false;
+    if (evidence.thinkingObserved !== true)
+        return false;
+    if (committedAnyFacts)
+        return false;
+    if ((result.assistantText ?? "").trim())
+        return false;
+    // Gateways sometimes relabel a length-stopped stream as `network` or
+    // `nonzero-exit`; provider evidence outweighs the transport label.
+    if (result.failureCategory &&
+        !["empty-output", "network", "nonzero-exit", "unknown"].includes(result.failureCategory))
+        return false;
+    return true;
+}
 /**
  * The writer session burned an excessive token budget (a read-edit-test loop
  * that never converged) and still failed. Distinct from empty-output so the
@@ -69,8 +299,6 @@ function readWriterThinkingExhaustionEvidence(result) {
 export function isWriterThinkingExhausted(result, mapped, changeManifestChangedFiles) {
     if (mapped.ok)
         return false;
-    if (mapped.failureCategory !== "empty-output")
-        return false;
     const evidence = readWriterThinkingExhaustionEvidence(result);
     if (evidence.stopReason !== "length")
         return false;
@@ -78,6 +306,13 @@ export function isWriterThinkingExhausted(result, mapped, changeManifestChangedF
         return false;
     if ((evidence.writeToolCallCount ?? 0) !== 0)
         return false;
+    // Gateways sometimes classify a length-stopped stream as `network` or
+    // `nonzero-exit` because the terminal event is carried in stderr. The
+    // provider evidence is stronger than that transport label when no write
+    // tool was called and the run produced no diff.
+    if (mapped.failureCategory &&
+        !["empty-output", "network", "nonzero-exit", "unknown"].includes(mapped.failureCategory))
+        return false;
     if (changeManifestChangedFiles === undefined)
         return false;
     if (changeManifestChangedFiles.length !== 0)
@@ -217,6 +452,8 @@ export const FRONTEND_CONTRACT_RECORD_TOOL_NAMES = [
     "record_handoff_intent",
     "record_open_question",
     "record_split_proposal",
+    "record_ui_state",
+    "record_required_deliverables",
 ];
 export const FRONTEND_CONTRACT_TERMINAL_TOOL_NAMES = [
     "finalize_contract",
@@ -230,6 +467,7 @@ export const FRONTEND_SCOUT_EVIDENCE_TOOL_NAMES = [
 export const FRONTEND_PLAN_RECORD_TOOL_NAMES = [
     "record_route_selection",
     "record_component_choice",
+    "record_state_registry",
     "record_state_flow",
     "record_data_flow",
     "record_mock_api",
@@ -873,8 +1111,111 @@ async function resolveFrontendPlanNewComponentSourceReferences(input) {
     }
 }
 /**
- * A+B: `frontend-plan-pi` now records its decision ledger through seven
- * incremental `record_*` tools (origin=plan) and closes with exactly one
+ * Frozen canonical requirements keyed by id, resolved from the source-fidelity
+ * ledger before the contract node starts. record_requirement commits these
+ * runtime-owned values so the model can never rewrite authoritative requirement
+ * text or stringify the fragment bindings.
+ */
+async function resolveFrontendCanonicalRequirements(input) {
+    const binding = input.sourceBinding;
+    if (!binding || binding.schemaVersion !== 2 || !binding.ledgerPath) {
+        return new Map();
+    }
+    const absolutePath = path.resolve(input.cwd, binding.ledgerPath);
+    const workspaceRoot = path.resolve(input.cwd);
+    if (absolutePath !== workspaceRoot &&
+        !absolutePath.startsWith(`${workspaceRoot}${path.sep}`)) {
+        throw new Error("canonical requirement ledger escapes workspace");
+    }
+    try {
+        const raw = await readFile(absolutePath, "utf8");
+        if (sha256Text(raw) !== binding.ledgerSha256) {
+            throw new Error("source ledger is stale");
+        }
+        const ledger = parseLedgerJson(raw);
+        const sourceFragments = new Map();
+        const boundFragments = new Map(ledger.fragments.map((fragment) => [fragment.id, fragment]));
+        for (const sourcePath of new Set(ledger.fragments.map((fragment) => fragment.path))) {
+            const sourceAbsolute = path.resolve(workspaceRoot, sourcePath);
+            if (!sourceAbsolute.startsWith(`${workspaceRoot}${path.sep}`)) {
+                throw new Error("source fragment escapes workspace");
+            }
+            const content = await readFile(sourceAbsolute, "utf8");
+            for (const fragment of splitDocumentIntoFragments({ path: sourcePath, content })) {
+                const frozen = boundFragments.get(fragment.id);
+                if (frozen && frozen.sha256 === sha256Text(fragment.text)) {
+                    sourceFragments.set(fragment.id, fragment.text);
+                }
+            }
+        }
+        return new Map(ledger.canonicalRequirements.map((requirement) => [
+            requirement.id,
+            {
+                text: requirement.text,
+                sourceFragmentIds: [...(requirement.sourceFragmentIds ?? [])],
+                sourceFragments,
+            },
+        ]));
+    }
+    catch (error) {
+        throw new Error(`canonical requirement sources unavailable: ${error instanceof Error ? error.message : String(error)}`);
+    }
+}
+/**
+ * Authoritative UI state ids declared by the contract node
+ * (ui-state-declaration facts). Empty when the source declares no UI-state
+ * table, in which case the plan's state vocabulary is registry-only.
+ */
+async function resolveFrontendDeclaredUiStateIds(input) {
+    try {
+        const records = await readTypedEventStoreFromJsonl(path.join(input.runDir, "frontend-contract-pi", "contract-typed-facts.jsonl"));
+        return records.flatMap((record) => {
+            const fact = record.fact;
+            return record.phase === "committed" &&
+                fact.kind === "ui-state-declaration" &&
+                typeof fact.id === "string"
+                ? [fact.id]
+                : [];
+        });
+    }
+    catch {
+        // No contract facts (or no declared states): registry-only vocabulary.
+        return [];
+    }
+}
+/**
+ * Resolve the frozen canonical behavior verification-target ids the PRD
+ * declares. The requirement text is the same authority the design review reads
+ * when it rejects `VT-SHARD-N-…` / variant ids as a contract-requirement gap,
+ * so extracting `VT-…` tokens from the frozen contract requirement facts makes
+ * that authority deterministic instead of prose-only. An empty result (no PRD
+ * declared any behavior target id) disables the canonical check and keeps the
+ * historical free-form path.
+ */
+export async function resolveFrontendCanonicalVerificationTargetIds(input) {
+    try {
+        const records = await readTypedEventStoreFromJsonl(path.join(input.runDir, "frontend-contract-pi", "contract-typed-facts.jsonl"));
+        const ids = new Set();
+        for (const record of records) {
+            const fact = record.fact;
+            if (record.phase !== "committed" || fact.kind !== "requirement") {
+                continue;
+            }
+            const text = typeof fact.text === "string" ? fact.text : "";
+            for (const match of text.matchAll(/\bVT-[A-Z0-9][A-Z0-9_-]*\b/g)) {
+                ids.add(match[0]);
+            }
+        }
+        return [...ids].sort();
+    }
+    catch {
+        // No contract facts (or unreadable): fall back to no canonical set.
+        return [];
+    }
+}
+/**
+ * A+B: `frontend-plan-pi` records its decision ledger through incremental
+ * `record_*` tools (origin=plan) and closes with exactly one
  * `finalize_plan` terminal. A later attempt may explicitly adopt a quarantined
  * fact via `adopt_staged_fact`. Flush writes `plan-typed-facts.jsonl` for the
  * node validator / compile authority.
@@ -965,23 +1306,24 @@ export async function createFrontendPlanLedgerTools(input) {
         paths: stringArray,
         conflicts: stringArray,
     }, { additionalProperties: false });
-    // Enum fields use literal unions, not advisory strings: a soft Type.String
-    // lets the model commit values like type="behavior" that pass the tool
-    // boundary, flush into the ledger, and only fail the compile-time zod enum
-    // — a deterministic attempt failure the model could have fixed in-node.
-    const verificationTargetTypeSchema = Type.Union([
-        Type.Literal("static"),
+    // Verification mode is runtime-owned (contract v2): the plan submits a
+    // commandId referencing the frozen command directory, never a type. A
+    // soft Type.String would let the model commit values that only fail at
+    // compile time — keep the reference a required string and validate it
+    // against the directory at the tool boundary below.
+    const verificationTargetScopeSchema = Type.Union([
         Type.Literal("unit"),
         Type.Literal("component"),
         Type.Literal("integration"),
-        Type.Literal("mock"),
     ]);
     const verificationTargetSchema = Type.Object({
         id: Type.String({}),
-        type: verificationTargetTypeSchema,
-        commandLabel: Type.String({}),
+        commandId: Type.String({
+            description: "Frozen command directory key (e.g. verify-npm-run-build); the runtime resolves mode and label from it.",
+        }),
         file: Type.String({}),
         requirementIds: stringArray,
+        scope: Type.Optional(verificationTargetScopeSchema),
         uiStates: Type.Optional(Type.Array(Type.String({}), {
             description: "Optional only at this tool boundary. An omitted value is deterministically recorded as []. Pass an explicit array for new calls.",
         })),
@@ -1007,6 +1349,12 @@ export async function createFrontendPlanLedgerTools(input) {
             line: Type.Optional(Type.Number({})),
         }, { additionalProperties: false })),
         rationale: Type.String({}),
+        covers: Type.Optional(Type.Array(Type.String({}), {
+            description: "UI state and/or interaction names this single component choice covers (one choice may cover many ids).",
+        })),
+        evidencePath: Type.Optional(Type.String({
+            description: "REQUIRED for decision=reuse-existing: repo-relative path whose existing file is the reuse evidence. Greenfield paths must use decision=new.",
+        })),
     }, { additionalProperties: false });
     const stringList = (value) => Array.isArray(value)
         ? value.filter((item) => typeof item === "string")
@@ -1085,7 +1433,7 @@ export async function createFrontendPlanLedgerTools(input) {
     const recordComponentChoiceTool = defineTool({
         name: "record_component_choice",
         label: "record_component_choice",
-        description: "Record ONE component choice (origin=plan component-choice fact). Declare every UI purpose's component selection. For decision=new, pass sourceRequirementIds containing the frozen requirement ID(s) that mandate the component and sourceFragmentId selecting one frozen citation listed in the plan checklist; the runtime validates the relation and derives the exact PRD specReference. Do not read the PRD or invent a path/line. decision=reuse-existing is only for components that already exist in the repo (e.g. reusing ActiveRunBadge's styling convention). Omit rationale for reuse-existing; it is optional. Call up to 5 component choices per assistant message (batching reduces API round trips and rate-limit risk); never more than 5 per message. Optionally include stylingStrategy (set it once, on the first call). Example: {\"choice\": {\"purpose\": \"<interaction or UI state name>\", \"component\": \"<component name>\", \"decision\": \"new\"}, \"sourceRequirementIds\": [\"<AC-XXX mandating this component>\"], \"sourceFragmentId\": \"<REQ-SRC-...>\"}",
+        description: "Record ONE component choice (origin=plan component-choice fact). One choice may cover MULTIPLE UI states/interactions via covers: [\"<state-or-interaction name>\", ...]; do not emit one row per interaction. For decision=new, pass sourceRequirementIds containing the frozen requirement ID(s) that mandate the component and sourceFragmentId selecting one frozen citation listed in the plan checklist; the runtime validates the relation and derives the exact PRD specReference. Do not read the PRD or invent a path/line. decision=reuse-existing is only for components that already exist in the repo and REQUIRES evidencePath: a repo-relative path to the existing file that proves the reuse — the runtime verifies the file exists (fresh evidence); a path with no existing file is greenfield and must use decision=new instead. Call up to 5 component choices per assistant message; never more than 5 per message. Optionally include stylingStrategy (set it once, on the first call). Example: {\"choice\": {\"purpose\": \"<interaction or UI state name>\", \"component\": \"<component name>\", \"decision\": \"new\", \"covers\": [\"<other interaction names this component also serves>\"]}, \"sourceRequirementIds\": [\"<AC-XXX mandating this component>\"], \"sourceFragmentId\": \"<REQ-SRC-...>\"} or {\"choice\": {\"purpose\": \"FocusQueuePanel\", \"component\": \"FocusQueuePanel\", \"decision\": \"reuse-existing\", \"evidencePath\": \"src/journal.ts\"}}",
         promptSnippet: "Record 1-5 component choices (up to 5 per message).",
         parameters: Type.Object({
             choice: uiComponentChoiceSchema,
@@ -1136,6 +1484,46 @@ export async function createFrontendPlanLedgerTools(input) {
                     ...(citation.line !== undefined ? { line: citation.line } : {}),
                 };
             }
+            if (choice.decision === "reuse-existing") {
+                // reuse-existing must cite a file that actually exists NOW (fresh
+                // existence evidence, same semantics as Scout pathEvidence.fresh).
+                // A path with no file on disk is greenfield: decision=new is the
+                // only honest choice (dogfood run dag-1788504923861-0f7b17a9
+                // claimed 18 reuse-existing conventions in a not-yet-written
+                // src/planner.ts and every gate let it through).
+                const evidencePath = typeof choice.evidencePath === "string"
+                    ? choice.evidencePath.trim()
+                    : "";
+                if (!evidencePath) {
+                    return planToolReceipt({
+                        ok: false,
+                        kind: "component-choice",
+                        error: "decision=reuse-existing requires evidencePath naming the existing repo file that proves the reuse; if the file does not exist yet, use decision=new",
+                    });
+                }
+                if (input.workspaceRoot) {
+                    const absolute = path.resolve(input.workspaceRoot, evidencePath);
+                    const workspaceRoot = path.resolve(input.workspaceRoot);
+                    const contained = absolute === workspaceRoot ||
+                        absolute.startsWith(`${workspaceRoot}${path.sep}`);
+                    let exists = false;
+                    if (contained) {
+                        try {
+                            exists = (await stat(absolute)).isFile();
+                        }
+                        catch {
+                            exists = false;
+                        }
+                    }
+                    if (!exists) {
+                        return planToolReceipt({
+                            ok: false,
+                            kind: "component-choice",
+                            error: `decision=reuse-existing evidencePath "${evidencePath}" has no fresh existence evidence (file not found in the workspace); reuse requires an existing file — use decision=new for greenfield paths`,
+                        });
+                    }
+                }
+            }
             const components = typeof choice.component === "string" ? [choice.component] : [];
             const result = await adoptPlanFact("component-choice", `${attemptId}:record_component_choice:${randomUUID()}`, {
                 kind: "component-choice",
@@ -1159,10 +1547,86 @@ export async function createFrontendPlanLedgerTools(input) {
             return planToolReceipt(echo);
         },
     });
+    // Global UX vocabulary: one compact registry committed BEFORE any state
+    // flow details. Coverage may slice by AC; UX must not — the registry is
+    // the anti-duplication anchor that keeps every later slice on the same
+    // named concepts instead of re-inventing them per AC chunk.
+    const recordStateRegistryTool = defineTool({
+        name: "record_state_registry",
+        label: "record_state_registry",
+        description: "Commit the GLOBAL UX vocabulary (origin=plan state-registry fact) BEFORE any record_state_flow call: uiStateNames (use the contract's declared authoritative state ids when provided) and interactionNames (stable behavior-domain kebab-case names, e.g. planner-task-edit / focus-queue-move — one name per behavior domain, never one per AC). Empty arrays explicitly mean the request has no UI state or interaction vocabulary. Re-record the full vocabulary to correct it; the latest commit wins, but it may not remove names still referenced by committed state-flow facts. Example: {\"uiStateNames\": [\"planner-empty\"], \"interactionNames\": [\"planner-task-create\", \"focus-queue-move\"]}",
+        promptSnippet: "Record the global UI-state/interaction vocabulary once, before any state flow.",
+        parameters: Type.Object({
+            uiStateNames: stringArray,
+            interactionNames: stringArray,
+        }, { additionalProperties: false }),
+        async execute(_toolCallId, params) {
+            const uiStateNames = [...new Set(stringList(params?.uiStateNames).map((name) => name.trim()))];
+            const interactionNames = [
+                ...new Set(stringList(params?.interactionNames).map((name) => name.trim())),
+            ];
+            if (uiStateNames.some((name) => name.length === 0) || interactionNames.some((name) => name.length === 0)) {
+                return planToolReceipt({
+                    ok: false,
+                    kind: "state-registry",
+                    error: "record_state_registry names must be non-empty strings",
+                });
+            }
+            const invalidInteractionNames = interactionNames.filter((name) => !/^[a-z0-9]+(?:-[a-z0-9]+)*$/.test(name));
+            if (invalidInteractionNames.length > 0) {
+                return planToolReceipt({
+                    ok: false,
+                    kind: "state-registry",
+                    error: `record_state_registry interactionNames must be stable kebab-case behavior-domain names: ${invalidInteractionNames.join(", ")}`,
+                });
+            }
+            if (input.declaredUiStateIds &&
+                input.declaredUiStateIds.length > 0) {
+                const declared = new Set(input.declaredUiStateIds);
+                const undeclared = uiStateNames.filter((name) => !declared.has(name));
+                if (undeclared.length > 0) {
+                    return planToolReceipt({
+                        ok: false,
+                        kind: "state-registry",
+                        error: `record_state_registry uiStateNames are not declared by the contract's authoritative UI-state table: ${undeclared.join(", ")}; use the declared ids (${input.declaredUiStateIds.join(", ")})`,
+                    });
+                }
+                const missing = input.declaredUiStateIds.filter((name) => !uiStateNames.includes(name));
+                if (missing.length > 0) {
+                    return planToolReceipt({
+                        ok: false,
+                        kind: "state-registry",
+                        error: `record_state_registry must include every state from the contract's authoritative UI-state table; missing: ${missing.join(", ")}`,
+                    });
+                }
+            }
+            // A correction is deterministic only when the new last-wins registry
+            // still contains every live name already committed by state-flow facts.
+            // This permits adding a missed concept, while preventing a registry edit
+            // from retroactively orphaning earlier slices.
+            const liveNames = collectCanonicalStateFlowNames(readCommittedEvents(store, attemptId));
+            const orphanedUiStates = [...liveNames.uiStateNames].filter((name) => !uiStateNames.includes(name));
+            const orphanedInteractions = [...liveNames.interactionNames].filter((name) => !interactionNames.includes(name));
+            if (orphanedUiStates.length > 0 || orphanedInteractions.length > 0) {
+                return planToolReceipt({
+                    ok: false,
+                    kind: "state-registry",
+                    error: `record_state_registry cannot remove names still referenced by committed state-flow facts (uiStates: ${orphanedUiStates.join(", ") || "none"}; interactions: ${orphanedInteractions.join(", ") || "none"}); first correct/remove those state-flow entries, then re-record the full registry`,
+                });
+            }
+            const result = await adoptPlanFact("state-registry", `${attemptId}:record_state_registry:${randomUUID()}`, {
+                kind: "state-registry",
+                origin: "plan",
+                uiStateNames,
+                interactionNames,
+            });
+            return planToolReceipt(result);
+        },
+    });
     const recordStateFlowTool = defineTool({
         name: "record_state_flow",
         label: "record_state_flow",
-        description: "Record UI states and interactions as an origin=plan state-flow fact. To correct stale named entries from an earlier UX batch, include removeUiStateNames and/or removeInteractionNames. Example: {\"uiStates\": [{\"name\": \"<state>\", \"applicable\": true, \"expectedBehavior\": \"<behavior>\", \"implementationTargets\": [\"<file>\"], \"verificationTargetIds\": [\"<VT-XXX>\"]}], \"interactions\": [{\"name\": \"<interaction>\", \"trigger\": \"<user event>\", \"expectedBehavior\": \"<behavior>\", \"implementationTargets\": [\"<file>\"], \"verificationTargetIds\": [\"<VT-XXX>\"]}]}",
+        description: "Record UI states and interactions as an origin=plan state-flow fact. REQUIRES a committed record_state_registry vocabulary first, and every name here must be in that registry; uiState names must also be contract-declared authoritative ids when the contract declares them. To correct stale named entries from an earlier UX batch, include removeUiStateNames and/or removeInteractionNames. Example: {\"uiStates\": [{\"name\": \"<state>\", \"applicable\": true, \"expectedBehavior\": \"<behavior>\", \"implementationTargets\": [\"<file>\"], \"verificationTargetIds\": [\"<VT-XXX>\"]}], \"interactions\": [{\"name\": \"<interaction>\", \"trigger\": \"<user event>\", \"expectedBehavior\": \"<behavior>\", \"implementationTargets\": [\"<file>\"], \"verificationTargetIds\": [\"<VT-XXX>\"]}]}",
         promptSnippet: "Record the plan state-flow fact.",
         parameters: Type.Object({
             uiStates: Type.Array(uiStateSchema),
@@ -1266,9 +1730,63 @@ export async function createFrontendPlanLedgerTools(input) {
                 }
                 interactions.push({ ...interaction, name: resolvedName });
             }
+            // UX vocabulary gate: every recorded state/interaction name must be
+            // declared in the committed global registry first. This keeps UX out
+            // of AC-number slicing — the model commits one compact vocabulary
+            // (record_state_registry), then fills behavior-domain details, and a
+            // later slice cannot silently rename an earlier concept (dogfood
+            // dag-1788504923861-0f7b17a9: 5 renames of "prioritize" + 7 of
+            // "focus queue" across AC chunks).
             const states = uiStates
                 .map((state) => (typeof state?.name === "string" ? state.name : ""))
                 .filter(Boolean);
+            const committedRegistry = readCommittedEvents(store, attemptId)
+                .map((event) => event.fact)
+                .filter((fact) => isRecordObject(fact) && fact.kind === "state-registry")
+                .at(-1);
+            if (!committedRegistry) {
+                return planToolReceipt({
+                    ok: false,
+                    kind: "state-flow",
+                    error: "record_state_flow requires a committed UX registry first: call record_state_registry with the full uiStateNames/interactionNames vocabulary, then record state flows against it",
+                });
+            }
+            const registryUiStateNames = new Set(stringList(committedRegistry.uiStateNames));
+            const registryInteractionNames = new Set(stringList(committedRegistry.interactionNames));
+            const unknownStates = states.filter((name) => !registryUiStateNames.has(name));
+            if (unknownStates.length > 0) {
+                return planToolReceipt({
+                    ok: false,
+                    kind: "state-flow",
+                    error: `record_state_flow uiState names not in the committed registry: ${unknownStates.join(", ")}; re-record record_state_registry with the complete vocabulary first (registered: ${[...registryUiStateNames].join(", ") || "(none)"})`,
+                });
+            }
+            const unknownInteractions = interactions
+                .map((interaction) => typeof interaction?.name === "string" ? interaction.name : "")
+                .filter((name) => name && !registryInteractionNames.has(name));
+            if (unknownInteractions.length > 0) {
+                return planToolReceipt({
+                    ok: false,
+                    kind: "state-flow",
+                    error: `record_state_flow interaction names not in the committed registry: ${unknownInteractions.join(", ")}; re-record record_state_registry with the complete vocabulary first (registered: ${[...registryInteractionNames].join(", ") || "(none)"})`,
+                });
+            }
+            // Authoritative UI states: when the contract node declared the
+            // source's UI-state table, planner states must bind those ids — no
+            // invented variants (planner-empty/create-invalid/no-results/
+            // editing/focus-full/storage-unavailable, not "planner-list").
+            if (input.declaredUiStateIds &&
+                input.declaredUiStateIds.length > 0) {
+                const declared = new Set(input.declaredUiStateIds);
+                const undeclaredStates = states.filter((name) => !declared.has(name));
+                if (undeclaredStates.length > 0) {
+                    return planToolReceipt({
+                        ok: false,
+                        kind: "state-flow",
+                        error: `record_state_flow uiState names are not declared by the contract's authoritative UI-state table: ${undeclaredStates.join(", ")}; use the declared ids (${input.declaredUiStateIds.join(", ")}), or record a design deviation if the source table is genuinely incomplete`,
+                    });
+                }
+            }
             const result = await adoptPlanFact("state-flow", `${attemptId}:record_state_flow:${randomUUID()}`, {
                 kind: "state-flow",
                 origin: "plan",
@@ -1427,6 +1945,15 @@ export async function createFrontendPlanLedgerTools(input) {
                     });
                 }
             }
+            if (id &&
+                activeRequirementScope.length > 0 &&
+                !activeRequirementScope.includes(id)) {
+                return planToolReceipt({
+                    ok: false,
+                    kind: "plan-requirement",
+                    error: `record_plan_requirement id "${id}" is outside this session's requirement scope [${activeRequirementScope.join(", ")}]`,
+                });
+            }
             const result = await adoptPlanFact("plan-requirement", `${attemptId}:record_plan_requirement:${randomUUID()}`, {
                 kind: "plan-requirement",
                 origin: "plan",
@@ -1439,7 +1966,11 @@ export async function createFrontendPlanLedgerTools(input) {
     const recordPlanVerificationTargetTool = defineTool({
         name: "record_plan_verification_target",
         label: "record_plan_verification_target",
-        description: "Commit one plan verification target entry (origin=plan plan-verification-target fact). For non-static targets, call once per distinct behavior, not mechanically once per requirement: one target may cover multiple related requirementIds. For a correction, re-submit the same id with replace=true; the ledger compiles the latest replacement. A non-static target id is the stable trace token that implementation must place in a real describe/it/test title; static targets are traced by file and command only. Entry carries id, type, commandLabel, file, requirementIds, and uiStates; free-form symbol text is not accepted. IMPORTANT: batch up to 4 record_* calls per assistant message; never batch more than 4 — a larger single message risks output truncation under a small output window. Example: {\"entry\": {\"id\": \"VT-DASHBOARD-SHELL\", \"type\": \"unit\", \"commandLabel\": \"<frozen command label>\", \"file\": \"<test file>\", \"requirementIds\": [\"AC-001\", \"AC-002\"], \"uiStates\": []}}",
+        description: "Commit one plan verification target entry (origin=plan plan-verification-target fact). Reference a frozen verification command by commandId (see the frozen command directory in your prompt: static commands are project-wide checks traced by file and command only; behavior commands need a test file whose describe/it/test title contains the target id). For behavior targets, call once per distinct behavior, not mechanically once per requirement: one target may cover multiple related requirementIds. For a correction, re-submit the same id with replace=true; the ledger compiles the latest replacement. A behavior target id is the stable trace token that implementation must place in a real describe/it/test title. Entry carries id, commandId, file, requirementIds, and uiStates; optional scope (unit | component | integration) is display-only. Free-form symbol text is not accepted. IMPORTANT: batch up to 4 record_* calls per assistant message; never batch more than 4 — a larger single message risks output truncation under a small output window. Example: {\"entry\": {\"id\": \"VT-DASHBOARD-SHELL\", \"commandId\": \"<frozen behavior command id>\", \"file\": \"<test file>\", \"requirementIds\": [\"AC-001\", \"AC-002\"], \"uiStates\": []}}" +
+            (input.canonicalVerificationTargetIds &&
+                input.canonicalVerificationTargetIds.length > 0
+                ? ` Frozen canonical behavior target ids (use exactly for behavior targets): ${input.canonicalVerificationTargetIds.join(", ")}.`
+                : ""),
         promptSnippet: "Commit 1-4 plan verification target entries (up to 4 per message).",
         parameters: Type.Object({
             entry: verificationTargetSchema,
@@ -1454,19 +1985,61 @@ export async function createFrontendPlanLedgerTools(input) {
                     error: "record_plan_verification_target requires a non-empty entry object",
                 });
             }
-            // Belt-and-braces for providers that do not strictly enforce the
-            // tool-schema enum: reject an invalid type here with the allowed
-            // values so the model can re-record in-node instead of the whole
-            // attempt dying at compile time on the strict zod enum.
-            const verificationTargetType = rawEntry.type;
-            if (typeof verificationTargetType !== "string" ||
-                !["static", "unit", "component", "integration", "mock"].includes(verificationTargetType)) {
+            // Contract v2: the plan references a frozen command by commandId;
+            // mode and label are resolved by the runtime from the frozen
+            // command directory (run.json frontend-verify-shell bundle lanes).
+            // Reject unknown ids and behavior commands bound to non-test
+            // files here so the model fixes them in-node instead of the
+            // attempt dying at materialization or at the writer focused-check.
+            const verificationCommandId = typeof rawEntry.commandId === "string"
+                ? rawEntry.commandId.trim()
+                : "";
+            if (!verificationCommandId) {
                 return planToolReceipt({
                     ok: false,
                     kind: "plan-verification-target",
-                    error: `record_plan_verification_target entry.type must be one of static | unit | component | integration | mock (received ${JSON.stringify(verificationTargetType ?? null)}); for a runtime behavior check use type=mock or type=integration`,
+                    error: `record_plan_verification_target entry.commandId is required (received ${JSON.stringify(rawEntry.commandId ?? null)}); pick one id from the frozen command directory in your prompt`,
                 });
             }
+            const { deriveFrontendVerifyCommandDirectoryFromRun, isFrontendTestFilePath, } = await import("../workflows/dag/frontend-implementation-contract.js");
+            const verifyDirectory = await deriveFrontendVerifyCommandDirectoryFromRun(input.runDir);
+            if (verifyDirectory.length > 0) {
+                const directoryEntry = verifyDirectory.find((entry) => entry.commandId === verificationCommandId);
+                if (!directoryEntry) {
+                    return planToolReceipt({
+                        ok: false,
+                        kind: "plan-verification-target",
+                        error: `record_plan_verification_target verification-target-unknown-command: unknown commandId "${verificationCommandId}"; available frozen commands: [${verifyDirectory.map((entry) => `${entry.commandId} (${entry.mode}: ${entry.label})`).join(", ")}]`,
+                    });
+                }
+                if (directoryEntry.mode === "behavior" &&
+                    typeof rawEntry.file === "string" &&
+                    !isFrontendTestFilePath(rawEntry.file)) {
+                    return planToolReceipt({
+                        ok: false,
+                        kind: "plan-verification-target",
+                        error: `record_plan_verification_target verification-target-phase-mismatch: behavior command "${directoryEntry.label}" (${directoryEntry.commandId}) must bind a test file (__tests__/, tests?/, e2e/, cypress/, *.test.*, *.spec.*, *.cy.*); received file "${rawEntry.file}"`,
+                    });
+                }
+                // Canonical-identity check for behavior targets: the PRD freezes
+                // the exact behavior verification-target ids (e.g.
+                // VT-SMOKE-COUNTER-BEHAVIOR). A committed non-canonical id is
+                // immutable and design review rejects it as a
+                // contract-requirement gap, so reject invented ids here.
+                if (directoryEntry.mode === "behavior" &&
+                    input.canonicalVerificationTargetIds &&
+                    input.canonicalVerificationTargetIds.length > 0) {
+                    const canonicalTargetId = typeof rawEntry.id === "string" ? rawEntry.id.trim() : "";
+                    if (canonicalTargetId &&
+                        !input.canonicalVerificationTargetIds.includes(canonicalTargetId)) {
+                        return planToolReceipt({
+                            ok: false,
+                            kind: "plan-verification-target",
+                            error: `record_plan_verification_target id "${canonicalTargetId}" is not a frozen canonical behavior verification target; canonical ids are: ${input.canonicalVerificationTargetIds.join(", ")}`,
+                        });
+                    }
+                }
+            }
             // Duplicate-id rejection: committed typed facts are immutable, so
             // re-recording the same VT id would deadlock the compile by default.
             // replace=true is the explicit in-node correction path.
@@ -1535,6 +2108,15 @@ export async function createFrontendPlanLedgerTools(input) {
                     : [];
             }));
             const unknownRequirementIds = stringList(rawEntry.requirementIds).filter((id) => !declaredRequirementIds.has(id));
+            const outOfScopeRequirementIds = stringList(rawEntry.requirementIds).filter((id) => activeRequirementScope.length > 0 &&
+                !activeRequirementScope.includes(id));
+            if (outOfScopeRequirementIds.length > 0) {
+                return planToolReceipt({
+                    ok: false,
+                    kind: "plan-verification-target",
+                    error: `record_plan_verification_target references requirements outside this session's scope [${activeRequirementScope.join(", ")}]: ${outOfScopeRequirementIds.join(", ")}`,
+                });
+            }
             if (unknownRequirementIds.length > 0) {
                 return planToolReceipt({
                     ok: false,
@@ -1581,6 +2163,18 @@ export async function createFrontendPlanLedgerTools(input) {
                     error: "record_plan_evidence_gap requires a non-empty description describing the gap",
                 });
             }
+            const requirementId = typeof entry.requirementId === "string"
+                ? entry.requirementId
+                : undefined;
+            if (requirementId &&
+                activeRequirementScope.length > 0 &&
+                !activeRequirementScope.includes(requirementId)) {
+                return planToolReceipt({
+                    ok: false,
+                    kind: "plan-evidence-gap",
+                    error: `record_plan_evidence_gap requirementId "${requirementId}" is outside this session's requirement scope [${activeRequirementScope.join(", ")}]`,
+                });
+            }
             const result = await adoptPlanFact("plan-evidence-gap", `${attemptId}:record_plan_evidence_gap:${randomUUID()}`, { kind: "plan-evidence-gap", origin: "plan", entry });
             return planToolReceipt(result);
         },
@@ -1613,6 +2207,29 @@ export async function createFrontendPlanLedgerTools(input) {
                         ? { realIntegrationGap: params.realIntegrationGap }
                         : {}),
                 };
+                const latestRegistry = [...committed]
+                    .reverse()
+                    .map((record) => record.fact)
+                    .find((fact) => isRecordObject(fact) && fact.kind === "state-registry");
+                if (latestRegistry) {
+                    const registryUiStates = new Set(stringList(latestRegistry.uiStateNames));
+                    const registryInteractions = new Set(stringList(latestRegistry.interactionNames));
+                    const live = collectCanonicalStateFlowNames(committed);
+                    const missingUiStates = [...registryUiStates].filter((name) => !live.uiStateNames.has(name));
+                    const missingInteractions = [...registryInteractions].filter((name) => !live.interactionNames.has(name));
+                    const undeclaredUiStates = [...live.uiStateNames].filter((name) => !registryUiStates.has(name));
+                    const undeclaredInteractions = [...live.interactionNames].filter((name) => !registryInteractions.has(name));
+                    if (missingUiStates.length > 0 ||
+                        missingInteractions.length > 0 ||
+                        undeclaredUiStates.length > 0 ||
+                        undeclaredInteractions.length > 0) {
+                        return planToolReceipt({
+                            ok: false,
+                            kind: "finalize_plan",
+                            error: `finalize_plan UX registry mismatch: every registered name must have one live state-flow entry and every live entry must be registered (missing uiStates: ${missingUiStates.join(", ") || "none"}; missing interactions: ${missingInteractions.join(", ") || "none"}; undeclared uiStates: ${undeclaredUiStates.join(", ") || "none"}; undeclared interactions: ${undeclaredInteractions.join(", ") || "none"})`,
+                        });
+                    }
+                }
                 // Front-load the node's compile + policy gates into the finalize
                 // receipt (same pipeline the design-policy shell and the node
                 // self-check run: merge the patch onto the runtime skeleton,
@@ -1639,9 +2256,10 @@ export async function createFrontendPlanLedgerTools(input) {
                     catch (error) {
                         if (error instanceof PlanPolicyPrecheckFailure) {
                             // Template the fix: every uncovered interaction / state
-                            // maps to a ready-to-submit record_component_choice
-                            // call. One reuse-existing choice covers all
-                            // behavioural interactions.
+                            // maps to a record_component_choice skeleton. Reuse is
+                            // never implied: the operator/model must fill an existing
+                            // evidencePath or switch the choice to decision=new with
+                            // the required frozen source citation.
                             const suggestions = error.findings
                                 .filter((finding) => finding.code === "ui-design-coverage-missing" &&
                                 finding.path)
@@ -1652,6 +2270,7 @@ export async function createFrontendPlanLedgerTools(input) {
                                         purpose: finding.path,
                                         component: "<name the existing or new component>",
                                         decision: "reuse-existing",
+                                        evidencePath: "<existing repo file that proves this reuse>",
                                     },
                                 },
                             }));
@@ -1737,6 +2356,7 @@ export async function createFrontendPlanLedgerTools(input) {
         customTools: [
             recordRouteSelectionTool,
             recordComponentChoiceTool,
+            recordStateRegistryTool,
             recordStateFlowTool,
             recordDataFlowTool,
             recordMockApiTool,
@@ -1748,6 +2368,66 @@ export async function createFrontendPlanLedgerTools(input) {
             adoptStagedFactTool,
             finalizePlanTool,
         ],
+        adoptCommittedFacts: async (records) => {
+            for (const record of records) {
+                if (record.phase !== "committed")
+                    continue;
+                const fact = record.fact;
+                if (!fact || typeof fact !== "object" || Array.isArray(fact))
+                    continue;
+                const kind = typeof fact.kind === "string"
+                    ? fact.kind
+                    : "plan-fact";
+                const entry = fact.entry;
+                const identity = (kind === "plan-requirement" || kind === "plan-verification-target") &&
+                    typeof entry?.id === "string"
+                    ? `${kind}:${entry.id}`
+                    : undefined;
+                const sourceShardPrefix = `${attemptId}:parallel-merge:${record.attemptId}:`;
+                const requestId = `${sourceShardPrefix}${record.eventId}`;
+                const committed = readCommittedEvents(store, attemptId);
+                const replay = committed.find((candidate) => candidate.requestId === requestId);
+                if (replay) {
+                    if (replay.payloadSha256 !== record.payloadSha256) {
+                        throw new Error(`frontend plan shard fact merge replay conflict: source event ${record.eventId} changed payload`);
+                    }
+                    continue;
+                }
+                const explicitReplacement = typeof fact.replaces === "string" &&
+                    fact.replaces === entry?.id;
+                const conflicting = committed.find((candidate) => {
+                    const candidateFact = candidate.fact;
+                    const candidateEntry = candidateFact.entry;
+                    const sameSourceShard = candidate.requestId.startsWith(sourceShardPrefix);
+                    return (candidateFact.kind === kind &&
+                        typeof candidateEntry?.id === "string" &&
+                        `${kind}:${candidateEntry.id}` === identity &&
+                        !(sameSourceShard && explicitReplacement));
+                });
+                if (conflicting) {
+                    // Two shards may honestly emit the same fact (e.g. both
+                    // derive the same frozen verification target). Identical
+                    // payloads are duplicates to skip; divergent payloads are
+                    // a real conflict the ladder must resolve.
+                    if (conflicting.payloadSha256 === record.payloadSha256) {
+                        continue;
+                    }
+                    const fromSameSourceShard = conflicting.requestId.startsWith(sourceShardPrefix);
+                    if (fromSameSourceShard) {
+                        throw new Error(`frontend plan shard fact merge conflict: ${identity} was already committed by another shard with a different payload`);
+                    }
+                    // Divergent payload from a different source attempt means a
+                    // ladder retry re-derived the coverage phase: the fresh
+                    // derivation supersedes the stored record.
+                    const storeWithRecords = store;
+                    storeWithRecords.records = storeWithRecords.records.filter((candidate) => candidate.eventId !== conflicting.eventId);
+                }
+                const result = await adoptPlanFact(kind, requestId, fact);
+                if (!result.ok) {
+                    throw new Error(`frontend plan shard fact merge failed: ${result.error}`);
+                }
+            }
+        },
         setActiveRequirementScope: (requirementIds) => {
             activeRequirementScope = [
                 ...new Set(requirementIds.filter((id) => id.trim().length > 0)),
@@ -1786,6 +2466,10 @@ function contractFactSourceSpan(fact) {
     if (Array.isArray(fact.sourceRefs) && fact.sourceRefs.length > 0) {
         return fact.sourceRefs;
     }
+    if (Array.isArray(fact.sourceFragmentIds) &&
+        fact.sourceFragmentIds.some((value) => typeof value === "string" && value.trim().length > 0)) {
+        return fact.sourceFragmentIds;
+    }
     if (typeof fact.source === "string" && fact.source.trim().length > 0) {
         return fact.source;
     }
@@ -1835,6 +2519,7 @@ export function buildFrontendTaskContractVNext(records) {
         requirements,
         constraints: byKind("constraint"),
         evidenceExpectations: byKind("evidence-expectation"),
+        deliverableDeclarations: byKind("required-deliverables").at(-1)?.items ?? [],
         handoffIntents: byKind("handoff-intent"),
         openQuestions: byKind("open-question"),
         splitProposals: byKind("split-proposal"),
@@ -1842,7 +2527,7 @@ export function buildFrontendTaskContractVNext(records) {
     };
 }
 /**
- * A+B: `frontend-contract-pi` incremental contract tools. Six `record_*` tools
+ * A+B: `frontend-contract-pi` incremental contract tools. The `record_*` tools
  * commit origin=contract facts and `finalize_contract` commits the terminal
  * disposition (ready | ready-with-assumptions | blocked + blockingOwner).
  */
@@ -1894,9 +2579,7 @@ export async function createFrontendContractTools(input) {
         }
     }
     const recordKinds = {
-        record_requirement: "requirement",
         record_constraint: "constraint",
-        record_evidence_expectation: "evidence-expectation",
         record_handoff_intent: "handoff-intent",
         record_open_question: "open-question",
         record_split_proposal: "split-proposal",
@@ -1916,42 +2599,177 @@ export async function createFrontendContractTools(input) {
             return receipt(result);
         },
     }));
-    // OpenSpec selection committed as individual typed facts (one path per
-    // call) so a large candidate set never exceeds a single model output
-    // budget: each tool call carries exactly one {path, disposition,
-    // rationale} row and the ledger accumulates them across calls. Only
-    // positive classifications (required | relevant) are legal; unmentioned
-    // candidates default to irrelevant at the prewrite gate.
-    const recordOpenspecSelectionTool = defineTool({
-        name: "record_openspec_selection",
-        label: "record_openspec_selection",
-        description: "Commit one OpenSpec candidate classification (origin=contract openspec-selection fact). Call once per path you actually use or consult: required (must be read and cited) or relevant (informs planning). Never call it for irrelevant candidates — unmentioned candidates default to irrelevant. You may call it many times; one row per call.",
-        promptSnippet: "Commit one OpenSpec candidate classification (required | relevant); one path per call; skip irrelevant candidates.",
-        parameters: Type.Object({
-            path: Type.String({
-                description: "Repo-relative candidate spec path, e.g. openspec/project-specs/ui/ucp-components-md/AdvancedSearch.md",
-            }),
-            disposition: Type.Enum({
-                required: "required",
-                relevant: "relevant",
-            }),
-            rationale: Type.String({}),
-        }, { additionalProperties: false }),
+    // record_requirement is runtime-owned by design: the requirement TEXT and
+    // sourceFragmentIds come from the frozen source-fidelity ledger, never from
+    // model-authored prose. The model only names the canonical id it confirms.
+    // This closes the free-shape hole (additionalProperties:true accepted
+    // `statement` rewrites and JSON-stringified `sourceFragmentIds` arrays,
+    // which then reached the planner as empty text + dead fragment bindings).
+    const recordRequirementTool = defineTool({
+        name: "record_requirement",
+        label: "record_requirement",
+        description: 'Confirm one canonical ledger requirement (origin=contract requirement fact). Pass ONLY the canonical id listed in the <frontend_contract_input> inventory, e.g. {"id": "AC-001"} — the runtime commits the authoritative text and sourceFragmentIds from the frozen ledger. Never pass text/statement/sourceFragmentIds yourself: free-form rewrites and JSON-stringified fragment arrays are rejected. IMPORTANT: batch up to 5 record_* calls per message, starting from the FIRST message.',
+        promptSnippet: "Confirm 1-5 canonical requirements by id (up to 5 per message).",
+        parameters: Type.Object({ id: Type.String({ description: "Canonical ledger requirement id (e.g. AC-001)" }) }, { additionalProperties: false }),
         async execute(_toolCallId, params) {
-            const path = typeof params?.path === "string" ? params.path : "";
-            const disposition = params?.disposition;
-            const rationale = typeof params?.rationale === "string" ? params.rationale : "";
-            if (!path || !disposition || !rationale.trim()) {
+            const id = typeof params?.id === "string" ? params.id.trim() : "";
+            if (!id) {
                 return receipt({
                     ok: false,
-                    kind: "openspec-selection",
-                    error: "record_openspec_selection requires non-empty path, disposition (required|relevant), and rationale",
+                    kind: "requirement",
+                    error: "record_requirement requires the canonical requirement id",
                 });
             }
-            const result = await adoptContractFact("openspec-selection", {
-                kind: "openspec-selection",
+            const canonical = input.canonicalRequirements?.get(id);
+            if (!canonical) {
+                const known = [...(input.canonicalRequirements?.keys() ?? [])];
+                return receipt({
+                    ok: false,
+                    kind: "requirement",
+                    error: `record_requirement id "${id}" is not a canonical ledger requirement; canonical ids are: ${known.join(", ") || "(none)"}`,
+                });
+            }
+            const result = await adoptContractFact("requirement", {
+                kind: "requirement",
                 origin: "contract",
-                path,
+                disposition: "explicit",
+                id,
+                text: canonical.text,
+                sourceFragmentIds: canonical.sourceFragmentIds,
+            });
+            return receipt(result);
+        },
+    });
+    // Authoritative UI state declarations: the contract node extracts the
+    // PRD/reference UI-state table into structured facts so the planner binds
+    // uiStates to declared ids instead of inventing names (dogfood
+    // dag-1788504923861-0f7b17a9: 10 invented list-visibility variants, 0 of
+    // the 7 PRD states).
+    const recordUiStateTool = defineTool({
+        name: "record_ui_state",
+        label: "record_ui_state",
+        description: 'Declare ONE authoritative UI state extracted from the task source\'s UI-state table (origin=contract ui-state-declaration fact). Call once per declared state, exactly as the source names it: {"id": "<state id from the source table>", "trigger": "<when this state applies>", "observableOutcome": "<what the user can observe>"}. The planner must bind these ids later; do not rename or invent states.',
+        promptSnippet: "Declare 1-5 authoritative UI states (up to 5 per message).",
+        parameters: Type.Object({
+            id: Type.String({ description: "UI state id exactly as the source table declares it" }),
+            trigger: Type.String({ description: "When this state applies" }),
+            observableOutcome: Type.String({ description: "Observable result for the user" }),
+        }, { additionalProperties: false }),
+        async execute(_toolCallId, params) {
+            const id = typeof params?.id === "string" ? params.id.trim() : "";
+            const trigger = typeof params?.trigger === "string" ? params.trigger.trim() : "";
+            const observableOutcome = typeof params?.observableOutcome === "string"
+                ? params.observableOutcome.trim()
+                : "";
+            if (!id || !trigger || !observableOutcome) {
+                return receipt({
+                    ok: false,
+                    kind: "ui-state-declaration",
+                    error: "record_ui_state requires non-empty id, trigger, and observableOutcome",
+                });
+            }
+            const result = await adoptContractFact("ui-state-declaration", {
+                kind: "ui-state-declaration",
+                origin: "contract",
+                id,
+                trigger,
+                observableOutcome,
+            });
+            return receipt(result);
+        },
+    });
+    const evidenceStatus = Type.Enum({
+        required: "required", optional: "optional", "not-applicable": "not-applicable",
+    });
+    const recordEvidenceExpectationTool = defineTool({
+        name: "record_evidence_expectation",
+        label: "record_evidence_expectation",
+        description: 'Record evidence requirements as {requirementId:"<canonical id>",evidence:{static:"required|optional|not-applicable",behavior:"required|optional|not-applicable",mock:"required|optional|not-applicable","real-integration":"required|optional|not-applicable"}}. Declare at least one lane; later calls replace only the lanes they name.',
+        parameters: Type.Object({
+            requirementId: Type.String(),
+            evidence: Type.Object({
+                static: Type.Optional(evidenceStatus),
+                behavior: Type.Optional(evidenceStatus),
+                mock: Type.Optional(evidenceStatus),
+                "real-integration": Type.Optional(evidenceStatus),
+            }, { additionalProperties: false }),
+        }, { additionalProperties: false }),
+        async execute(_toolCallId, params) {
+            try {
+                const fact = frontendEvidenceExpectationSchema.parse(params);
+                if (!input.canonicalRequirements?.has(fact.requirementId)) {
+                    throw new Error(`unknown canonical requirement ${fact.requirementId}`);
+                }
+                return receipt(await adoptContractFact("evidence-expectation", {
+                    kind: "evidence-expectation", origin: "contract", ...fact,
+                }));
+            }
+            catch (error) {
+                return receipt({
+                    ok: false, kind: "evidence-expectation",
+                    error: error instanceof Error ? error.message : String(error),
+                });
+            }
+        },
+    });
+    const recordRequiredDeliverablesTool = defineTool({
+        name: "record_required_deliverables",
+        label: "record_required_deliverables",
+        description: 'Declare the complete source-required file deliverables as {items:[{path,requirementId,sourceFragmentId}]}. Interpret obligations from the original source, including lists/tables: permissions (allowedPaths/only allowed to modify), prohibitions, examples and read-only references are NOT delivery obligations. Each path must appear exactly in its frozen requirement-bound source fragment. Submit {items:[]} explicitly if no files are mandatory. A correction replaces the whole inventory. Required before finalize_contract ready.',
+        parameters: Type.Object({ items: Type.Array(Type.Object({
+                path: Type.String(), requirementId: Type.String(), sourceFragmentId: Type.String(),
+            }, { additionalProperties: false })) }, { additionalProperties: false }),
+        async execute(_toolCallId, params) {
+            try {
+                const declaration = validateFrontendRequiredDeliverables(params, input.canonicalRequirements ?? new Map());
+                return receipt(await adoptContractFact("required-deliverables", {
+                    kind: "required-deliverables", origin: "contract", ...declaration,
+                }));
+            }
+            catch (error) {
+                return receipt({
+                    ok: false, kind: "required-deliverables",
+                    error: error instanceof Error ? error.message : String(error),
+                });
+            }
+        },
+    });
+    // OpenSpec selection committed as individual typed facts (one path per
+    // call) so a large candidate set never exceeds a single model output
+    // budget: each tool call carries exactly one {path, disposition,
+    // rationale} row and the ledger accumulates them across calls. Only
+    // positive classifications (required | relevant) are legal; unmentioned
+    // candidates default to irrelevant at the prewrite gate.
+    const recordOpenspecSelectionTool = defineTool({
+        name: "record_openspec_selection",
+        label: "record_openspec_selection",
+        description: "Commit one OpenSpec candidate classification (origin=contract openspec-selection fact). Call once per path you actually use or consult: required (must be read and cited) or relevant (informs planning). Never call it for irrelevant candidates — unmentioned candidates default to irrelevant. You may call it many times; one row per call.",
+        promptSnippet: "Commit one OpenSpec candidate classification (required | relevant); one path per call; skip irrelevant candidates.",
+        parameters: Type.Object({
+            path: Type.String({
+                description: "Repo-relative candidate spec path, e.g. openspec/project-specs/ui/ucp-components-md/AdvancedSearch.md",
+            }),
+            disposition: Type.Enum({
+                required: "required",
+                relevant: "relevant",
+            }),
+            rationale: Type.String({}),
+        }, { additionalProperties: false }),
+        async execute(_toolCallId, params) {
+            const path = typeof params?.path === "string" ? params.path : "";
+            const disposition = params?.disposition;
+            const rationale = typeof params?.rationale === "string" ? params.rationale : "";
+            if (!path || !disposition || !rationale.trim()) {
+                return receipt({
+                    ok: false,
+                    kind: "openspec-selection",
+                    error: "record_openspec_selection requires non-empty path, disposition (required|relevant), and rationale",
+                });
+            }
+            const result = await adoptContractFact("openspec-selection", {
+                kind: "openspec-selection",
+                origin: "contract",
+                path,
                 disposition,
                 rationale,
             });
@@ -1986,6 +2804,13 @@ export async function createFrontendContractTools(input) {
                     error: "blocked disposition requires blockingOwner (blocked-human | blocked-external)",
                 });
             }
+            if (disposition !== "blocked" && input.canonicalRequirements?.size &&
+                !readCommittedEvents(store, attemptId).some((record) => record.fact.kind === "required-deliverables")) {
+                return receipt({
+                    ok: false, kind: "finalize_contract",
+                    error: "call record_required_deliverables with the complete source-bound inventory (or items:[] when none) before finalizing",
+                });
+            }
             const blockedOwner = mapContractBlockedOwner({
                 disposition: disposition ?? "",
                 blockingOwner,
@@ -2009,6 +2834,10 @@ export async function createFrontendContractTools(input) {
     return {
         customTools: [
             ...recordTools,
+            recordRequirementTool,
+            recordEvidenceExpectationTool,
+            recordUiStateTool,
+            recordRequiredDeliverablesTool,
             recordOpenspecSelectionTool,
             finalizeContractTool,
         ],
@@ -2228,6 +3057,22 @@ export async function createFrontendScoutEvidenceTools(input) {
     });
     return {
         customTools: [recordTargetSurfaceTool, recordDesignEvidenceTool],
+        committedFacts: () => readCommittedEvents(store, attemptId),
+        adoptCommittedFacts: async (records) => {
+            for (const record of records) {
+                if (record.phase !== "committed")
+                    continue;
+                const fact = record.fact;
+                if (!fact || typeof fact !== "object" || Array.isArray(fact))
+                    continue;
+                const result = await adoptScoutFact(typeof fact.kind === "string"
+                    ? fact.kind
+                    : "scout-fact", fact);
+                if (!result.ok) {
+                    throw new Error(`frontend scout shard fact merge failed: ${result.error}`);
+                }
+            }
+        },
         flush: async () => {
             const committed = readCommittedEvents(store, attemptId);
             await writeTypedEventStoreJsonl(path.join(input.runDir, input.nodeId, "scout-typed-facts.jsonl"), committed);
@@ -2620,19 +3465,30 @@ const FRONTEND_PLAN_SEGMENTS = [
         instruction: [
             "PLAN PHASE — requirement coverage only.",
             "Your ONLY job: for every frozen requirement, emit record_plan_requirement (requirement → implementation files) and record_plan_verification_target facts (verification target bound to requirement ids and files). Group related requirements under one non-static behavior target 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; never submit prose as a symbol. Do NOT record components, UI states, mock, dependency, or routes — a follow-up session owns those.",
+            "A coverage session is complete only when EVERY requirement assigned to this session (the full inventory, or the exact COVERAGE BATCH / shard list when present) has committed coverage facts: a record_plan_requirement entry plus verification targets, or a committed evidence gap. Keep committing in batches of up to 4 record_* calls per assistant message until then; do not write a concluding summary while any assigned requirement is still uncommitted — an early stop strands the remainder into a MISSING-FACT repair session and doubles the sessions needed.",
             "If a requirement genuinely cannot have a verification target, record a non-empty record_plan_evidence_gap. Do not call finalize_plan; it is not available in this phase.",
         ].join(" "),
     },
+    {
+        id: "ux-registry",
+        toolNames: new Set(["record_state_registry", "adopt_staged_fact"]),
+        instruction: [
+            "PLAN PHASE — global UX vocabulary.",
+            "Review ALL frozen requirements together and call record_state_registry exactly once with the complete UI-state and interaction vocabulary. UI states use declaredUiStates ids when present. Interaction names are stable kebab-case behavior domains; merge requirements that describe the same behavior instead of renaming it per AC slice. Empty arrays explicitly declare that no UX vocabulary applies. Do not record component choices or state-flow details in this phase.",
+            "Do not call finalize_plan; it is not available in this phase.",
+        ].join(" "),
+    },
     {
         id: "ux-local",
         toolNames: new Set([
             "record_component_choice",
             "record_state_flow",
+            "record_plan_verification_target",
             "adopt_staged_fact",
         ]),
         instruction: [
-            "PLAN PHASE — requirement-local UX decisions.",
-            "Requirements and verification targets are already committed in the ledger (do NOT re-record them; duplicates are rejected). For ONLY this requirement slice, record component choices and UI state flow. Keep these facts for this slice together. Cross-cutting data flow belongs to the global Mock/data phase. Do not record routes, Mock/API policy, dependencies, or design deviations here.",
+            "PLAN PHASE — global UX decisions.",
+            "Requirements and verification targets are already committed in the ledger. Review the complete requirement set and the committed global UX registry together, then record each component choice, UI state and interaction. Bind each applicable state to its verificationTargetIds; the runtime derives the reverse VT.uiStates relation. If a VT requires correction, record_plan_verification_target with replace:true is available after declaring its states; preserve its requirement coverage. Multiple requirements describing one behavior share one registry name and state-flow entry. Cross-cutting data flow belongs to the global Mock/data phase. Do not record routes, Mock/API policy, dependencies, or design deviations here.",
             "Do not call finalize_plan; it is not available in this phase.",
         ].join(" "),
     },
@@ -2690,40 +3546,6 @@ function planFactStringList(value) {
 function planFactScopeIntersects(fact, requirementIds) {
     return planFactStringList(fact.scopeRequirementIds).some((id) => requirementIds.has(id));
 }
-/** Apply state-flow replacement/removal semantics exactly as the plan compiler does. */
-function collectCanonicalStateFlowNames(committedFacts) {
-    const uiStateNames = new Set();
-    const interactionNames = new Set();
-    for (const value of committedFacts) {
-        const fact = committedFactFromPlanRecord(value);
-        if (!fact || fact.origin !== "plan" || fact.kind !== "state-flow")
-            continue;
-        for (const name of planFactStringList(fact.removeUiStateNames)) {
-            uiStateNames.delete(name);
-        }
-        for (const name of planFactStringList(fact.removeInteractionNames)) {
-            interactionNames.delete(name);
-        }
-        for (const state of Array.isArray(fact.uiStates) ? fact.uiStates : []) {
-            if (!state || typeof state !== "object" || Array.isArray(state))
-                continue;
-            const name = state.name;
-            if (typeof name === "string" && name.trim())
-                uiStateNames.add(name);
-        }
-        for (const interaction of Array.isArray(fact.interactions)
-            ? fact.interactions
-            : []) {
-            if (!interaction || typeof interaction !== "object" || Array.isArray(interaction)) {
-                continue;
-            }
-            const name = interaction.name;
-            if (typeof name === "string" && name.trim())
-                interactionNames.add(name);
-        }
-    }
-    return { uiStateNames, interactionNames };
-}
 /** Compute the authoritative coverage queue from the committed plan ledger. */
 export function collectFrontendPlanMissingFacts(input) {
     const requirements = new Map();
@@ -2805,6 +3627,17 @@ export function collectFrontendPlanPhaseMissingFacts(input) {
     const facts = input.committedFacts
         .map(committedFactFromPlanRecord)
         .filter((fact) => Boolean(fact && fact.origin === "plan"));
+    if (input.phase === "ux-registry") {
+        return facts.some((fact) => fact.kind === "state-registry")
+            ? []
+            : [
+                {
+                    kind: "state-registry",
+                    requirementIds: [...input.requirementIds],
+                    reason: "global UX vocabulary phase has no committed state-registry fact",
+                },
+            ];
+    }
     if (input.phase === "ux-local") {
         const needsUx = input.requirementIds.some((id) => input.behaviorRequiredRequirementIds?.includes(id));
         if (!needsUx)
@@ -2899,11 +3732,28 @@ export function batchFrontendPlanRequirements(input) {
     return batches;
 }
 const FRONTEND_PLAN_COVERAGE_MAX_RECORD_CALLS = 10;
-const FRONTEND_PLAN_UX_LOCAL_MAX_RECORD_CALLS = 6;
-// A large requirement set creates both coverage and UX-local shards. Keep a
-// safety bound, but do not let the old 32-session ceiling skip finalize for a
-// legitimate large plan.
+const FRONTEND_PLAN_COVERAGE_MAX_CONCURRENCY = 4;
+// A large requirement set creates parallel coverage shards; UX remains one
+// global decision session so behavior names and component/state facts are not
+// reinvented per AC batch. Keep a safety bound for adaptive coverage retries
+// without letting the old 32-session ceiling skip finalize.
 const FRONTEND_PLAN_BATCH_MAX_SESSIONS = 128;
+const FRONTEND_PLAN_SMALL_MAX_REQUIREMENTS = 8;
+const FRONTEND_PLAN_SMALL_MAX_ESTIMATED_CALLS = 24;
+async function mapWithConcurrency(items, limit, worker) {
+    const results = new Array(items.length);
+    let nextIndex = 0;
+    const workerCount = Math.min(Math.max(1, limit), items.length);
+    await Promise.all(Array.from({ length: workerCount }, async () => {
+        while (true) {
+            const index = nextIndex++;
+            if (index >= items.length)
+                return;
+            results[index] = await worker(items[index], index);
+        }
+    }));
+    return results;
+}
 function compactPromptString(value, maxChars) {
     if (typeof value !== "string" || value.trim().length === 0)
         return undefined;
@@ -2920,6 +3770,45 @@ function compactPromptStringArray(value, maxEntries = 12, maxChars = 180) {
         .filter((item) => item !== undefined)
         .slice(0, maxEntries);
 }
+function countFrontendPlanTargetSurfaces(basePrompt) {
+    const match = /<frontend_plan_input>[\s\S]*?<\/frontend_plan_input>/.exec(basePrompt);
+    if (!match)
+        return undefined;
+    for (const line of match[0].split(/\r?\n/)) {
+        try {
+            const payload = JSON.parse(line);
+            if (Array.isArray(payload.targetSurface))
+                return payload.targetSurface.length;
+        }
+        catch {
+            // surrounding lines are prose
+        }
+    }
+    return undefined;
+}
+/**
+ * Deterministically extract repository file paths that frozen verification
+ * commands operate on (`--config <file>`, `node --check <file>`). A frozen
+ * command whose referenced file is outside the planner's verification targets
+ * can never run: the admission writeSet derives from those targets, so the
+ * writer is not authorized to create the file and verify-shell fails ~20
+ * minutes later. Surfacing the gap as plan missing-facts lets the coverage or
+ * compact session record the missing verification target inside the same
+ * attempt instead.
+ */
+export function collectFrontendVerificationCommandFiles(basePrompt) {
+    const files = new Set();
+    const configRe = /--config\s+([\w@./-]+\.(?:js|mjs|cjs|ts|json))/g;
+    const checkRe = /node\s+--check\s+([\w@./-]+\.(?:js|mjs|cjs))/g;
+    for (const re of [configRe, checkRe]) {
+        for (const match of basePrompt.matchAll(re)) {
+            const file = match[1];
+            if (file && file.includes("/"))
+                files.add(file);
+        }
+    }
+    return [...files].sort();
+}
 /**
  * Remove the repeated full planner input from a coverage batch. The normal
  * plan prompt already contains a bounded JSON handoff, but repeating all
@@ -2990,8 +3879,8 @@ export function compactFrontendPlanPromptForRequirementSlice(basePrompt, require
                         ...(compactPromptString(target.id, 80)
                             ? { id: compactPromptString(target.id, 80) }
                             : {}),
-                        ...(compactPromptString(target.type, 40)
-                            ? { type: compactPromptString(target.type, 40) }
+                        ...(compactPromptString(target.commandId, 80)
+                            ? { commandId: compactPromptString(target.commandId, 80) }
                             : {}),
                         ...(compactPromptString(target.commandLabel, 180)
                             ? { commandLabel: compactPromptString(target.commandLabel, 180) }
@@ -3047,8 +3936,33 @@ export function compactFrontendPlanPromptForRequirementSlice(basePrompt, require
                 }];
         })
         : [];
+    const compactDeclaredUiStates = Array.isArray(payload.declaredUiStates)
+        ? payload.declaredUiStates.flatMap((value) => {
+            if (!value || typeof value !== "object" || Array.isArray(value))
+                return [];
+            const state = value;
+            const id = compactPromptString(state.id, 80);
+            if (!id)
+                return [];
+            return [{
+                    id,
+                    ...(compactPromptString(state.trigger, 180)
+                        ? { trigger: compactPromptString(state.trigger, 180) }
+                        : {}),
+                    ...(compactPromptString(state.observableOutcome, 240)
+                        ? { observableOutcome: compactPromptString(state.observableOutcome, 240) }
+                        : {}),
+                }];
+        })
+        : [];
+    const committedUx = payload.committedUx &&
+        typeof payload.committedUx === "object" &&
+        !Array.isArray(payload.committedUx)
+        ? payload.committedUx
+        : undefined;
     const compactPayload = {
         requirements: compactRequirements,
+        requiredDeliverables: Array.isArray(payload.requiredDeliverables) ? payload.requiredDeliverables : [],
         ...(compactTargetSurface.length > 0
             ? { targetSurface: compactTargetSurface }
             : {}),
@@ -3058,6 +3972,17 @@ export function compactFrontendPlanPromptForRequirementSlice(basePrompt, require
         ...(compactDesignEvidence.length > 0
             ? { designEvidence: compactDesignEvidence }
             : {}),
+        ...(compactDeclaredUiStates.length > 0
+            ? { declaredUiStates: compactDeclaredUiStates }
+            : {}),
+        ...(committedUx
+            ? {
+                committedUx: {
+                    uiStateNames: compactPromptStringArray(committedUx.uiStateNames, 40, 100),
+                    interactionNames: compactPromptStringArray(committedUx.interactionNames, 60, 120),
+                },
+            }
+            : {}),
     };
     const compactBlock = [
         "<frontend_plan_input>",
@@ -3136,8 +4061,7 @@ function compactFrontendPlanLedgerContext(input) {
                 kind: fact.kind,
                 entry: {
                     id: compactPromptString(entry.id, 80),
-                    type: compactPromptString(entry.type, 40),
-                    commandLabel: compactPromptString(entry.commandLabel, 180),
+                    commandId: compactPromptString(entry.commandId, 80),
                     file: compactPromptString(entry.file, 180),
                     requirementIds,
                     uiStates: compactPromptStringArray(entry.uiStates, 12, 100),
@@ -3155,6 +4079,8 @@ function compactFrontendPlanLedgerContext(input) {
                             purpose: compactPromptString(item.purpose, 120),
                             component: compactPromptString(item.component, 120),
                             decision: compactPromptString(item.decision, 40),
+                            covers: compactPromptStringArray(item.covers, 40, 120),
+                            evidencePath: compactPromptString(item.evidencePath, 180),
                         }];
                 }).slice(0, 24)
                 : [];
@@ -3162,6 +4088,14 @@ function compactFrontendPlanLedgerContext(input) {
                 compactFacts.push({ kind: fact.kind, uiComponentChoices: choices });
             continue;
         }
+        if (fact.kind === "state-registry") {
+            compactFacts.push({
+                kind: fact.kind,
+                uiStateNames: compactPromptStringArray(fact.uiStateNames, 40, 100),
+                interactionNames: compactPromptStringArray(fact.interactionNames, 60, 120),
+            });
+            continue;
+        }
         if (fact.kind === "state-flow") {
             compactFacts.push({
                 kind: fact.kind,
@@ -3224,8 +4158,10 @@ function compactFrontendPlanLedgerContext(input) {
 export async function runFrontendPlanSegmentedSessions(input) {
     const queue = [];
     const coverageSegment = FRONTEND_PLAN_SEGMENTS.find((segment) => segment.id === "coverage");
+    const uxRegistrySegment = FRONTEND_PLAN_SEGMENTS.find((segment) => segment.id === "ux-registry");
     const uxSegment = FRONTEND_PLAN_SEGMENTS.find((segment) => segment.id === "ux-local");
     const globalMockDataSegment = FRONTEND_PLAN_SEGMENTS.find((segment) => segment.id === "global-mock-data");
+    const finalizeSegment = FRONTEND_PLAN_SEGMENTS.find((segment) => segment.id === "finalize");
     const allRequirementIds = input.requirementIds ?? [];
     const buildPhasePrompt = (segment, missing = []) => {
         const compact = allRequirementIds.length > 0
@@ -3242,7 +4178,7 @@ export async function runFrontendPlanSegmentedSessions(input) {
                 ? ["plan-requirement", "plan-verification-target", "state-flow", "data-flow", "mock-api"]
                 : segment.id === "global-dependency-deviation"
                     ? ["dependency", "design-deviation"]
-                    : ["plan-requirement", "plan-verification-target", "component-choice", "state-flow", "data-flow", "mock-api", "design-deviation", "dependency", "target-surface"];
+                    : ["plan-requirement", "plan-verification-target", "state-registry", "component-choice", "state-flow", "data-flow", "mock-api", "design-deviation", "dependency", "target-surface"];
         const ledger = input.committedFacts
             ? compactFrontendPlanLedgerContext({
                 committedFacts: input.committedFacts(),
@@ -3265,7 +4201,7 @@ export async function runFrontendPlanSegmentedSessions(input) {
     const buildCoveragePrompt = (slice, missing = []) => [
         compactFrontendPlanPromptForRequirementSlice(input.basePrompt, slice),
         coverageSegment.instruction,
-        `COVERAGE BATCH: process ONLY these requirements in this session: ${slice.join(", ")}. Other requirements are handled by separate sessions; do not record them.`,
+        `COVERAGE BATCH: process ONLY these requirements in this session: ${slice.join(", ")}. Other requirements are handled by separate sessions; do not record them. Finish the whole list before concluding: commit every listed requirement's coverage facts (verification targets or an evidence gap), batching up to 4 record_* calls per message; an early stop re-queues the remainder as a MISSING-FACT repair session.`,
         ...(missing.length > 0
             ? [
                 "MISSING-FACT QUEUE: the previous session did not establish complete coverage. Repair ONLY these items, then re-check the slice:",
@@ -3278,14 +4214,20 @@ export async function runFrontendPlanSegmentedSessions(input) {
             ? compactFrontendPlanLedgerContext({
                 committedFacts: input.committedFacts(),
                 requirementIds: slice,
-                kinds: ["plan-requirement", "plan-verification-target", "component-choice", "state-flow"],
+                kinds: [
+                    "plan-requirement",
+                    "plan-verification-target",
+                    "state-registry",
+                    "component-choice",
+                    "state-flow",
+                ],
                 scopedKinds: ["component-choice", "state-flow"],
             })
             : "";
         return [
             compactFrontendPlanPromptForRequirementSlice(input.basePrompt, slice),
             uxSegment.instruction,
-            `UX LOCAL BATCH: process ONLY these requirements: ${slice.join(", ")}.`,
+            `GLOBAL UX SCOPE: process the complete requirement set together: ${slice.join(", ")}. Do not split or rename one behavior by requirement id.`,
             ledger,
             ...(missing.length > 0
                 ? [
@@ -3295,6 +4237,74 @@ export async function runFrontendPlanSegmentedSessions(input) {
                 : []),
         ].filter(Boolean).join("\n\n");
     };
+    const buildUxRegistryPrompt = (missing = []) => {
+        const ledger = input.committedFacts
+            ? compactFrontendPlanLedgerContext({
+                committedFacts: input.committedFacts(),
+                requirementIds: allRequirementIds,
+                kinds: [
+                    "plan-requirement",
+                    "plan-verification-target",
+                    "state-registry",
+                ],
+            })
+            : "";
+        const compact = allRequirementIds.length > 0
+            ? compactFrontendPlanPromptForRequirementSlice(input.basePrompt, allRequirementIds, {
+                includeVerificationTargets: false,
+                includeDesignEvidence: false,
+                includeChecklist: false,
+            })
+            : input.basePrompt;
+        return [
+            compact,
+            uxRegistrySegment.instruction,
+            ledger,
+            ...(missing.length > 0
+                ? [
+                    "MISSING-FACT QUEUE: commit the global UX registry, then re-check this phase:",
+                    ...missing.map((item) => `- ${item.reason}`),
+                ]
+                : []),
+        ].filter(Boolean).join("\n\n");
+    };
+    const buildCompactLocalPrompt = (missing = []) => {
+        const ledger = input.committedFacts
+            ? compactFrontendPlanLedgerContext({
+                committedFacts: input.committedFacts(),
+                requirementIds: allRequirementIds,
+                kinds: [
+                    "plan-requirement",
+                    "plan-verification-target",
+                    "state-registry",
+                    "component-choice",
+                    "state-flow",
+                ],
+            })
+            : "";
+        return [
+            compactFrontendPlanPromptForRequirementSlice(input.basePrompt, allRequirementIds),
+            "PLAN PHASE — compact local planning for a small frontend request.",
+            "Review all listed requirements together and record_state_registry first with one global UX vocabulary. Then record every plan-requirement and verification-target fact, followed by the component-choice and state-flow facts needed by the observable UX. Do not read the repository or task source; use only the committed input above. Do not call finalize_plan in this session.",
+            "TOOL-FIRST: your first assistant actions must be record_* tool calls, at most 2-3 facts per message. Do not draft the whole analysis before recording; if a fact is uncertain, record it with an evidence gap instead of reasoning longer.",
+            ledger,
+            ...(missing.length > 0
+                ? [
+                    "MISSING-FACT QUEUE: repair ONLY these items, then re-check the compact local phase:",
+                    ...missing.map((item) => `- ${item.kind}${item.id ? ` ${item.id}` : ""} for ${item.requirementIds.join(", ")}: ${item.reason}`),
+                ]
+                : []),
+        ].filter(Boolean).join("\n\n");
+    };
+    const compactFinalizeInstruction = "This is a small-request compact pass. Reconcile the committed local facts with route, data-flow, Mock/API, dependency and deviation policy, then call finalize_plan exactly once.";
+    const buildCompactFinalizePrompt = (missing = []) => [buildPhasePrompt(finalizeSegment, missing), compactFinalizeInstruction].join("\n\n");
+    const mapPlannerExhaustion = (r, committedAnyFacts) => isPlannerThinkingExhausted(r, committedAnyFacts)
+        ? {
+            ...r,
+            failureCategory: OUTPUT_LIMIT_RETRY_CATEGORY,
+            stderr: `${r.stderr}\n${OUTPUT_LIMIT_RETRY_CATEGORY}: stopReason=length ended the turn before the next typed fact; preserve committed facts and retry only the unfinished phase`.trim(),
+        }
+        : r;
     // An empty list means "ledger unreadable / unknown" and falls back to one
     // unscoped coverage session; a non-empty list enables deterministic sharding.
     const requirementIdsProvided = input.requirementIds !== undefined && input.requirementIds.length > 0;
@@ -3307,7 +4317,66 @@ export async function runFrontendPlanSegmentedSessions(input) {
         : [];
     const incompleteRequirementIds = new Set(initialMissing.flatMap((item) => item.requirementIds));
     const coverageWorkIds = (input.requirementIds ?? []).filter((id) => pending.includes(id) || incompleteRequirementIds.has(id));
-    if (!requirementIdsProvided) {
+    if (input.parallelCoverageOnly === true &&
+        requirementIdsProvided &&
+        coverageWorkIds.length === 0) {
+        // A retry may reopen a shard whose committed facts are already complete.
+        // Treat that shard as an idempotent no-op; returning the normal empty
+        // session failure would make a partially failed map impossible to resume.
+        return {
+            ok: true,
+            assistantText: "",
+            command: [],
+            durationMs: 0,
+            exitCode: 0,
+            failureCategory: "success",
+            modelDisplay: "reused-coverage-facts",
+            parsedEvents: 0,
+            stderr: "",
+            stdout: "",
+            timedOut: false,
+            attemptedModels: [],
+            fallbackUsed: false,
+            tokensUsed: 0,
+        };
+    }
+    const estimatedCalls = (input.requirementIds ?? []).reduce((total, id) => total + Math.max(1, input.requirementCosts?.get(id) ?? 2), 0);
+    const targetSurfaceCount = countFrontendPlanTargetSurfaces(input.basePrompt);
+    // Small, single-surface requests do not benefit from six isolated Pi
+    // sessions. Keep the typed ledger as the authority, but let one local
+    // session establish requirement/UX facts and one final session establish
+    // cross-cutting policy + finalize. The old sharded ladder remains available
+    // for larger plans and for the unscoped compatibility path.
+    const useCompactSmallPlan = requirementIdsProvided &&
+        input.compactSmallPlan === true &&
+        input.requirementCosts !== undefined &&
+        estimatedCalls > 12 &&
+        (input.requirementIds?.length ?? 0) <= FRONTEND_PLAN_SMALL_MAX_REQUIREMENTS &&
+        estimatedCalls <= FRONTEND_PLAN_SMALL_MAX_ESTIMATED_CALLS &&
+        targetSurfaceCount === 1;
+    if (useCompactSmallPlan) {
+        const compactLocalTools = new Set([
+            "record_plan_requirement",
+            "record_plan_verification_target",
+            "record_plan_evidence_gap",
+            "record_state_registry",
+            "record_component_choice",
+            "record_state_flow",
+            "adopt_staged_fact",
+        ]);
+        queue.push({
+            id: "compact-local",
+            toolNames: compactLocalTools,
+            requirementSlice: [...allRequirementIds],
+            prompt: buildCompactLocalPrompt(),
+        });
+        queue.push({
+            id: "finalize",
+            toolNames: null,
+            prompt: buildCompactFinalizePrompt(),
+        });
+    }
+    else if (!requirementIdsProvided) {
         queue.push({
             id: "coverage",
             toolNames: coverageSegment.toolNames,
@@ -3329,42 +4398,47 @@ export async function runFrontendPlanSegmentedSessions(input) {
             prompt: buildCoveragePrompt(slice),
         }));
     }
-    const uxCosts = new Map((input.requirementIds ?? []).map((id) => [
-        id,
-        Math.max(2, Math.min(3, input.requirementCosts?.get(id) ?? 2)),
-    ]));
-    const uxSlices = requirementIdsProvided
-        ? batchFrontendPlanRequirements({
-            requirementIds: input.requirementIds,
-            maxEstimatedRecordCalls: FRONTEND_PLAN_UX_LOCAL_MAX_RECORD_CALLS,
-            maxRequirements: 3,
-            requirementCosts: uxCosts,
-        })
-        : [[]];
-    uxSlices.forEach((slice, batchIndex) => queue.push({
-        id: `ux-local-${batchIndex + 1}`,
-        toolNames: uxSegment.toolNames,
-        ...(slice.length > 0 ? { requirementSlice: slice } : {}),
-        prompt: slice.length > 0
-            ? buildUxPrompt(slice)
-            : buildPhasePrompt(uxSegment),
-    }));
-    for (const segment of FRONTEND_PLAN_SEGMENTS) {
-        if (["coverage", "ux-local", "finalize"].includes(segment.id))
-            continue;
+    if (useCompactSmallPlan) {
+        // Compact mode already queued both sessions above.
+    }
+    else if (!input.parallelCoverageOnly) {
+        if (requirementIdsProvided) {
+            queue.push({
+                id: uxRegistrySegment.id,
+                toolNames: uxRegistrySegment.toolNames,
+                prompt: buildUxRegistryPrompt(),
+            });
+        }
+        const uxSlice = requirementIdsProvided ? [...allRequirementIds] : [];
         queue.push({
-            id: segment.id,
-            toolNames: segment.toolNames,
-            prompt: buildPhasePrompt(segment),
+            id: "ux-local-1",
+            // Compatibility path for an unreadable requirement inventory: there is
+            // no safe all-requirements registry phase, so the unscoped UX session
+            // must establish its registry before recording state flow.
+            toolNames: requirementIdsProvided
+                ? uxSegment.toolNames
+                : new Set([...(uxSegment.toolNames ?? []), "record_state_registry"]),
+            ...(uxSlice.length > 0 ? { requirementSlice: uxSlice } : {}),
+            prompt: uxSlice.length > 0
+                ? buildUxPrompt(uxSlice)
+                : buildPhasePrompt(uxSegment),
+        });
+        for (const segment of FRONTEND_PLAN_SEGMENTS) {
+            if (["coverage", "ux-registry", "ux-local", "finalize"].includes(segment.id))
+                continue;
+            queue.push({
+                id: segment.id,
+                toolNames: segment.toolNames,
+                prompt: buildPhasePrompt(segment),
+            });
+        }
+        queue.push({
+            id: finalizeSegment.id,
+            toolNames: finalizeSegment.toolNames,
+            prompt: buildPhasePrompt(finalizeSegment),
         });
     }
-    const finalizeSegment = FRONTEND_PLAN_SEGMENTS.find((segment) => segment.id === "finalize");
-    queue.push({
-        id: finalizeSegment.id,
-        toolNames: finalizeSegment.toolNames,
-        prompt: buildPhasePrompt(finalizeSegment),
-    });
-    let last;
+    let accumulated;
     let index = 0;
     let invocationCount = 0;
     while (index < queue.length) {
@@ -3389,6 +4463,12 @@ export async function runFrontendPlanSegmentedSessions(input) {
             const promptSlice = remaining.length > 0 ? remaining : session.coverageSlice;
             prompt = buildCoveragePrompt(promptSlice, session.missingFacts ?? preexistingMissing);
         }
+        else if (session.id === "compact-local") {
+            prompt = buildCompactLocalPrompt(session.missingFacts);
+        }
+        else if (session.id === "ux-registry") {
+            prompt = buildUxRegistryPrompt(session.missingFacts);
+        }
         else if (session.id.startsWith("ux-local-")) {
             // Build this at execution time: coverage facts are committed by the
             // preceding sessions and must be visible to the UX-local model.
@@ -3402,12 +4482,17 @@ export async function runFrontendPlanSegmentedSessions(input) {
             // stale queue entry from dropping facts written by earlier phases.
             const segment = FRONTEND_PLAN_SEGMENTS.find((candidate) => candidate.id === session.id);
             if (segment) {
-                prompt = buildPhasePrompt(segment, session.missingFacts);
+                prompt =
+                    session.id === "finalize" && useCompactSmallPlan
+                        ? buildCompactFinalizePrompt(session.missingFacts)
+                        : buildPhasePrompt(segment, session.missingFacts);
             }
         }
         const committedBefore = input.committedFactCount();
-        input.setActiveRequirementScope?.(session.id.startsWith("ux-local-")
-            ? session.requirementSlice ?? []
+        input.setActiveRequirementScope?.(session.coverageOnly ||
+            session.id === "compact-local" ||
+            session.id.startsWith("ux-local-")
+            ? session.coverageSlice ?? session.requirementSlice ?? []
             : []);
         if (invocationCount >= FRONTEND_PLAN_BATCH_MAX_SESSIONS)
             break;
@@ -3425,7 +4510,9 @@ export async function runFrontendPlanSegmentedSessions(input) {
                 }
                 : {}),
         });
-        last = result;
+        accumulated = accumulated
+            ? combineSequentialPiResults(accumulated, result)
+            : result;
         try {
             await input.flushLedger();
         }
@@ -3444,27 +4531,81 @@ export async function runFrontendPlanSegmentedSessions(input) {
                 committedFacts: input.committedFacts(),
             })
             : [];
+        const isCompactLocalSession = session.id === "compact-local";
+        const isUxRegistrySession = session.id === "ux-registry";
         const isUxLocalSession = session.id.startsWith("ux-local-");
-        const missingPhase = isUxLocalSession && session.requirementSlice && input.committedFacts
-            ? collectFrontendPlanPhaseMissingFacts({
-                phase: "ux-local",
-                requirementIds: session.requirementSlice,
-                committedFacts: input.committedFacts(),
-                behaviorRequiredRequirementIds: input.behaviorRequiredRequirementIds,
-            })
-            : session.id === "global-mock-data" && allRequirementIds.length > 0 && input.committedFacts
+        const missingPhase = isCompactLocalSession && input.committedFacts
+            ? [
+                ...collectFrontendPlanMissingFacts({
+                    requirementIds: allRequirementIds,
+                    committedFacts: input.committedFacts(),
+                }),
+                ...collectFrontendPlanPhaseMissingFacts({
+                    phase: "ux-registry",
+                    requirementIds: allRequirementIds,
+                    committedFacts: input.committedFacts(),
+                }),
+                ...collectFrontendPlanPhaseMissingFacts({
+                    phase: "ux-local",
+                    requirementIds: allRequirementIds,
+                    committedFacts: input.committedFacts(),
+                    behaviorRequiredRequirementIds: input.behaviorRequiredRequirementIds,
+                }),
+            ]
+            : isUxRegistrySession && input.committedFacts
                 ? collectFrontendPlanPhaseMissingFacts({
-                    phase: "global-mock-data",
+                    phase: "ux-registry",
                     requirementIds: allRequirementIds,
                     committedFacts: input.committedFacts(),
                 })
-                : [];
+                : isUxLocalSession && session.requirementSlice && input.committedFacts
+                    ? collectFrontendPlanPhaseMissingFacts({
+                        phase: "ux-local",
+                        requirementIds: session.requirementSlice,
+                        committedFacts: input.committedFacts(),
+                        behaviorRequiredRequirementIds: input.behaviorRequiredRequirementIds,
+                    })
+                    : session.id === "global-mock-data" && allRequirementIds.length > 0 && input.committedFacts
+                        ? collectFrontendPlanPhaseMissingFacts({
+                            phase: "global-mock-data",
+                            requirementIds: allRequirementIds,
+                            committedFacts: input.committedFacts(),
+                        })
+                        : [];
         const missingPhaseFacts = [...missingCoverage, ...missingPhase];
+        // Frozen verification commands must operate on files the planner has
+        // committed verification targets for; otherwise the admission writeSet
+        // cannot authorize the file and verify-shell deterministically fails.
+        const verificationCommandFiles = collectFrontendVerificationCommandFiles(input.basePrompt);
+        if (verificationCommandFiles.length > 0 && allRequirementIds.length > 0 && input.committedFacts) {
+            const coveredVerificationFiles = new Set(input
+                .committedFacts()
+                .map(committedFactFromPlanRecord)
+                .filter((fact) => Boolean(fact && fact.origin === "plan" && fact.kind === "plan-verification-target"))
+                .map((fact) => fact.entry?.file)
+                .filter((file) => typeof file === "string"));
+            for (const file of verificationCommandFiles) {
+                if (coveredVerificationFiles.has(file))
+                    continue;
+                missingPhaseFacts.push({
+                    kind: "plan-verification-target",
+                    id: file,
+                    requirementIds: allRequirementIds.slice(0, 1),
+                    reason: `frozen verification command references ${file} but no committed verification target covers it; record a static verification target for this file so it joins the writeSet`,
+                });
+            }
+        }
         const promptForMissingPhase = (missing) => session.coverageOnly
             ? buildCoveragePrompt(session.coverageSlice ?? [], missing)
-            : isUxLocalSession
-                ? buildUxPrompt(session.requirementSlice ?? [], missing)
-                : buildPhasePrompt(globalMockDataSegment, missing);
+            : isCompactLocalSession
+                ? buildCompactLocalPrompt(missing)
+                : isUxRegistrySession
+                    ? buildUxRegistryPrompt(missing)
+                    : session.id === "finalize" && useCompactSmallPlan
+                        ? buildCompactFinalizePrompt(missing)
+                        : isUxLocalSession
+                            ? buildUxPrompt(session.requirementSlice ?? [], missing)
+                            : buildPhasePrompt(globalMockDataSegment, missing);
         if (result.ok) {
             if (missingPhaseFacts.length === 0) {
                 index += 1;
@@ -3480,9 +4621,9 @@ export async function runFrontendPlanSegmentedSessions(input) {
                 continue;
             }
             return {
-                ...result,
+                ...accumulated,
                 ok: false,
-                stderr: `${result.stderr}\nfrontend plan completeness check failed: ${missingPhaseFacts.map((item) => item.reason).join("; ")}`.trim(),
+                stderr: `${accumulated.stderr}\nfrontend plan completeness check failed: ${missingPhaseFacts.map((item) => item.reason).join("; ")}`.trim(),
                 failureCategory: "invalid-output",
             };
         }
@@ -3499,19 +4640,45 @@ export async function runFrontendPlanSegmentedSessions(input) {
             }
             if (missingPhaseFacts.length > 0) {
                 return {
-                    ...result,
+                    ...accumulated,
                     ok: false,
-                    stderr: `frontend plan completeness check failed: ${missingPhaseFacts.map((item) => item.reason).join("; ")}`,
+                    stderr: `${accumulated.stderr}\nfrontend plan completeness check failed: ${missingPhaseFacts.map((item) => item.reason).join("; ")}`.trim(),
                     failureCategory: "invalid-output",
                 };
             }
             index += 1;
             continue;
         }
+        // A length-stopped, fact-less compact session is the planner variant of
+        // writer-thinking-exhausted. Give the same scope exactly one tool-first
+        // retry before the split below re-batches the requirements, because one
+        // reinforced full-scope pass is cheaper than re-planning split halves.
+        const plannerThinkingBurn = isCompactLocalSession &&
+            committedAfter === committedBefore &&
+            !(result.assistantText ?? "").trim() &&
+            !result.stderr.trim() &&
+            !result.timedOut &&
+            readWriterThinkingExhaustionEvidence(result).stopReason === "length";
+        if (plannerThinkingBurn && (session.retryCount ?? 0) < 1) {
+            queue[index] = {
+                ...session,
+                retryCount: (session.retryCount ?? 0) + 1,
+                prompt: buildCompactLocalPrompt(),
+            };
+            continue;
+        }
         // Option 5: a multi-requirement coverage batch that failed with ZERO
         // new facts and no provider stderr is the upfront-reasoning burn —
-        // halve the slice and retry instead of failing the attempt.
-        const coverageSlice = session.coverageOnly ? session.coverageSlice : undefined;
+        // halve the slice and retry instead of failing the attempt. The compact
+        // local session participates through its requirementSlice so a
+        // thinking-burned small plan degrades into smaller batched sessions
+        // instead of replaying one full-scope prompt until the repair budget
+        // runs out.
+        const coverageSlice = session.coverageOnly
+            ? session.coverageSlice
+            : isCompactLocalSession
+                ? session.requirementSlice
+                : undefined;
         const zeroProgressBurn = coverageSlice !== undefined &&
             coverageSlice.length > 1 &&
             committedAfter === committedBefore &&
@@ -3522,6 +4689,26 @@ export async function runFrontendPlanSegmentedSessions(input) {
             const half = Math.ceil(coverageSlice.length / 2);
             const firstSlice = coverageSlice.slice(0, half);
             const secondSlice = coverageSlice.slice(half);
+            if (isCompactLocalSession) {
+                // The split halves leave compact mode: continue them as ordinary
+                // coverage sessions so every downstream ladder branch applies.
+                queue.splice(index, 1, {
+                    ...session,
+                    id: `coverage-compact-split-1`,
+                    coverageOnly: true,
+                    coverageSlice: firstSlice,
+                    requirementSlice: firstSlice,
+                    prompt: buildCoveragePrompt(firstSlice),
+                }, {
+                    ...session,
+                    id: `coverage-compact-split-2`,
+                    coverageOnly: true,
+                    coverageSlice: secondSlice,
+                    requirementSlice: secondSlice,
+                    prompt: buildCoveragePrompt(secondSlice),
+                });
+                continue;
+            }
             queue.splice(index, 1, {
                 ...session,
                 coverageSlice: firstSlice,
@@ -3533,7 +4720,7 @@ export async function runFrontendPlanSegmentedSessions(input) {
             });
             continue;
         }
-        if ((session.coverageOnly || isUxLocalSession || session.id === "global-mock-data") &&
+        if ((session.coverageOnly || isUxRegistrySession || isUxLocalSession || session.id === "global-mock-data") &&
             missingPhaseFacts.length > 0 &&
             !result.stderr.trim() &&
             !result.timedOut &&
@@ -3546,11 +4733,11 @@ export async function runFrontendPlanSegmentedSessions(input) {
             };
             continue;
         }
-        return result;
+        return mapPlannerExhaustion(accumulated, committedAfter > committedBefore);
     }
     if (index < queue.length) {
         return {
-            ...(last ?? {
+            ...(accumulated ?? {
                 ok: false,
                 stdout: "",
                 stderr: "",
@@ -3567,17 +4754,195 @@ export async function runFrontendPlanSegmentedSessions(input) {
                 tokensUsed: 0,
             }),
             ok: false,
-            stderr: `${last?.stderr ?? ""}\nfrontend plan segmentation exceeded the ${FRONTEND_PLAN_BATCH_MAX_SESSIONS}-session safety limit before finalize`.trim(),
+            stderr: `${accumulated?.stderr ?? ""}\nfrontend plan segmentation exceeded the ${FRONTEND_PLAN_BATCH_MAX_SESSIONS}-session safety limit before finalize`.trim(),
             failureCategory: "invalid-output",
         };
     }
-    return (last ?? {
+    return mapPlannerExhaustion(accumulated ?? {
         ok: false,
         stdout: "",
         stderr: "frontend plan segmentation produced no session",
         failureCategory: "empty-output",
         durationMs: 0,
-    });
+    }, false);
+}
+/**
+ * Run the two independent Scout evidence surfaces concurrently while keeping
+ * their typed-event stores isolated. The main Scout store is the only store
+ * visible to Plan; shard facts are merged in completion-order-independent
+ * order after both sessions settle. This gives discovery real parallelism
+ * without allowing sibling models to race a shared revision counter.
+ */
+function aggregateParallelPiResults(results) {
+    const first = results[0];
+    const failed = results.find((result) => !result.ok);
+    const representative = failed ?? first;
+    return {
+        ...representative,
+        ok: failed === undefined,
+        failureCategory: failed?.failureCategory ?? "success",
+        durationMs: Math.max(...results.map((result) => result.durationMs), 0),
+        exitCode: failed ? failed.exitCode : 0,
+        stderr: results.map((result) => result.stderr).filter(Boolean).join("\n"),
+        tokensUsed: results.reduce((total, result) => total + result.tokensUsed, 0),
+        parsedEvents: results.reduce((total, result) => total + result.parsedEvents, 0),
+        attemptedModels: [
+            ...new Set(results.flatMap((result) => result.attemptedModels)),
+        ],
+        fallbackUsed: results.some((result) => result.fallbackUsed),
+        timedOut: results.some((result) => result.timedOut),
+    };
+}
+function combineSequentialPiResults(first, second) {
+    return {
+        ...second,
+        durationMs: first.durationMs + second.durationMs,
+        stderr: [first.stderr, second.stderr].filter(Boolean).join("\n"),
+        tokensUsed: first.tokensUsed + second.tokensUsed,
+        parsedEvents: first.parsedEvents + second.parsedEvents,
+        attemptedModels: [
+            ...new Set([...first.attemptedModels, ...second.attemptedModels]),
+        ],
+        fallbackUsed: first.fallbackUsed || second.fallbackUsed,
+        timedOut: first.timedOut || second.timedOut,
+    };
+}
+async function runFrontendScoutParallelSessions(input) {
+    const [{ createTypedEventStore }] = await Promise.all([
+        import("../workflows/dag/frontend-typed-event-store.js"),
+    ]);
+    const shards = [
+        {
+            id: "surface",
+            toolName: "record_target_surface",
+            instruction: [
+                "PARALLEL SCOUT SHARD — target surface only.",
+                "Inspect routes, entrypoints, implementation ownership, data source, and applicable test paths.",
+                "Call record_target_surface exactly once with the complete runtime-evidenced surface. Do not call record_design_evidence.",
+            ].join(" "),
+        },
+        {
+            id: "design",
+            toolName: "record_design_evidence",
+            instruction: [
+                "PARALLEL SCOUT SHARD — design evidence only.",
+                "Inspect the frontend framework, styling/theme conventions, reusable components, and relevant design/spec files.",
+                "Call record_design_evidence for the evidence you actually read. Do not call record_target_surface.",
+            ].join(" "),
+        },
+    ];
+    const outcomes = await Promise.all(shards.map(async (shard) => {
+        let shardTools;
+        let shardResult;
+        try {
+            const store = createTypedEventStore();
+            const shardNodeId = `${input.nodeId}/parallel/${shard.id}`;
+            shardTools = await createFrontendScoutEvidenceTools({
+                attemptId: `${input.attemptId}:parallel:${shard.id}`,
+                store,
+                runDir: input.runDir,
+                nodeId: shardNodeId,
+                workspaceRoot: input.workspaceRoot,
+                sourceDeclaredPaths: input.sourceDeclaredPaths,
+            });
+            const customTools = shardTools.customTools.filter((tool) => typeof tool === "object" &&
+                tool !== null &&
+                tool.name === shard.toolName);
+            const sessionOptions = {
+                ...input.sessionOptions,
+                sessionEventsPath: path.join(input.runDir, shardNodeId, "session-events.jsonl"),
+            };
+            shardResult = await input.piStepFn({
+                ...sessionOptions,
+                prompt: `${input.basePrompt}\n\n${shard.instruction}`,
+                writerToolPolicy: {
+                    requireSdk: true,
+                    customTools: [...customTools, ...(input.readBudgetTools ?? [])],
+                },
+            });
+            await shardTools.flush();
+            return { shard, result: shardResult, tools: shardTools };
+        }
+        catch (error) {
+            const crashMessage = `frontend scout parallel shard ${shard.id} crashed: ${error instanceof Error ? error.message : String(error)}`;
+            return {
+                shard,
+                tools: shardTools,
+                result: shardResult
+                    ? {
+                        ...shardResult,
+                        ok: false,
+                        failureCategory: shardResult.ok
+                            ? "invalid-output"
+                            : shardResult.failureCategory,
+                        stderr: [shardResult.stderr, crashMessage]
+                            .filter(Boolean)
+                            .join("\n"),
+                    }
+                    : {
+                        ok: false,
+                        assistantText: "",
+                        command: [],
+                        durationMs: 0,
+                        exitCode: null,
+                        failureCategory: "tool-policy",
+                        modelDisplay: "unknown",
+                        parsedEvents: 0,
+                        stderr: crashMessage,
+                        stdout: "",
+                        timedOut: false,
+                        attemptedModels: [],
+                        fallbackUsed: false,
+                        tokensUsed: 0,
+                    },
+            };
+        }
+    }));
+    const ordered = [...outcomes].sort((left, right) => left.shard.id.localeCompare(right.shard.id));
+    try {
+        for (const outcome of ordered) {
+            if (outcome.result.ok && outcome.tools) {
+                await input.mainTools.adoptCommittedFacts(outcome.tools.committedFacts());
+            }
+        }
+    }
+    catch (error) {
+        const aggregate = aggregateParallelPiResults(ordered.map((outcome) => outcome.result));
+        return {
+            ...aggregate,
+            ok: false,
+            failureCategory: "invalid-output",
+            stderr: [
+                aggregate.stderr,
+                `frontend scout parallel fact merge failed: ${error instanceof Error ? error.message : String(error)}`,
+            ]
+                .filter(Boolean)
+                .join("\n"),
+        };
+    }
+    const failed = ordered.find((outcome) => !outcome.result.ok);
+    if (failed) {
+        const aggregate = aggregateParallelPiResults(ordered.map((outcome) => outcome.result));
+        return {
+            ...aggregate,
+            ok: false,
+            stderr: `${aggregate.stderr}\nfrontend scout parallel shard failed: ${failed.shard.id}`.trim(),
+        };
+    }
+    const committedKinds = new Set(ordered.flatMap((outcome) => (outcome.tools?.committedFacts() ?? [])
+        .map((record) => record.fact?.kind)
+        .filter((kind) => typeof kind === "string")));
+    const missing = ["target-surface", "design-evidence"].filter((kind) => !committedKinds.has(kind));
+    if (missing.length > 0) {
+        const fallback = aggregateParallelPiResults(ordered.map((outcome) => outcome.result));
+        return {
+            ...fallback,
+            ok: false,
+            failureCategory: "invalid-output",
+            stderr: `${fallback.stderr}\nfrontend scout parallel shards committed no ${missing.join(" or ")} fact`.trim(),
+        };
+    }
+    return aggregateParallelPiResults(ordered.map((outcome) => outcome.result));
 }
 export async function executeDagPiNode(input, meta, piStepFn = executePiStep, writeGuardDependencies = DEFAULT_DAG_PI_WRITE_GUARD_DEPENDENCIES) {
     const started = Date.now();
@@ -3673,6 +5038,7 @@ export async function executeDagPiNode(input, meta, piStepFn = executePiStep, wr
     let planLedgerTools;
     let contractTools;
     let scoutEvidenceTools;
+    let scoutSourceDeclaredPaths;
     let readBudgetTools;
     const commandPolicy = resolveDagCommandPolicy(input.task.commandPolicy);
     const allowsPlaywrightCli = dagCommandPolicyAllows(input.task.commandPolicy, "playwright-cli");
@@ -3816,6 +5182,10 @@ export async function executeDagPiNode(input, meta, piStepFn = executePiStep, wr
                 store,
                 runDir: meta.runDir,
                 nodeId: input.task.id,
+                canonicalRequirements: await resolveFrontendCanonicalRequirements({
+                    cwd: input.cwd,
+                    sourceBinding: meta.spec.sourceBinding,
+                }),
             });
             writerToolPolicy = {
                 requireSdk: true,
@@ -3836,16 +5206,17 @@ export async function executeDagPiNode(input, meta, piStepFn = executePiStep, wr
         try {
             const { createTypedEventStore } = await import("../workflows/dag/frontend-typed-event-store.js");
             const store = createTypedEventStore();
+            scoutSourceDeclaredPaths = await resolveFrontendScoutSourceDeclaredPaths({
+                cwd: input.cwd,
+                spec: meta.spec,
+            });
             scoutEvidenceTools = await createFrontendScoutEvidenceTools({
                 attemptId: `${meta.runId}:${input.task.id}`,
                 store,
                 runDir: meta.runDir,
                 nodeId: input.task.id,
                 workspaceRoot: input.cwd,
-                sourceDeclaredPaths: await resolveFrontendScoutSourceDeclaredPaths({
-                    cwd: input.cwd,
-                    spec: meta.spec,
-                }),
+                sourceDeclaredPaths: scoutSourceDeclaredPaths,
             });
             writerToolPolicy = {
                 requireSdk: true,
@@ -3878,6 +5249,13 @@ export async function executeDagPiNode(input, meta, piStepFn = executePiStep, wr
                     cwd: input.cwd,
                     sourceBinding: meta.spec.sourceBinding,
                 }),
+                declaredUiStateIds: await resolveFrontendDeclaredUiStateIds({
+                    runDir: meta.runDir,
+                }),
+                canonicalVerificationTargetIds: await resolveFrontendCanonicalVerificationTargetIds({
+                    runDir: meta.runDir,
+                }),
+                workspaceRoot: input.cwd,
             });
             writerToolPolicy = {
                 requireSdk: true,
@@ -4017,11 +5395,27 @@ export async function executeDagPiNode(input, meta, piStepFn = executePiStep, wr
                 }
                 : undefined,
         };
-        if (isFrontendPlanLedgerNode(input.task) && planLedgerTools) {
-            // Frontend-only split: sequential sessions with independent
-            // output budgets (coverage -> UX decisions -> finalize), mirroring the
-            // backend-test template's module sharding. Every other template keeps
-            // the single-session path below.
+        if (isFrontendScoutEvidenceNode(input.task) &&
+            scoutEvidenceTools &&
+            input.task.complexity !== "LOW") {
+            result = await runFrontendScoutParallelSessions({
+                piStepFn,
+                sessionOptions: piSessionOptions,
+                basePrompt: input.prompt,
+                mainTools: scoutEvidenceTools,
+                runDir: meta.runDir,
+                nodeId: input.task.id,
+                attemptId: `${meta.runId}:${input.task.id}`,
+                workspaceRoot: input.cwd,
+                sourceDeclaredPaths: scoutSourceDeclaredPaths,
+                readBudgetTools: readBudgetTools?.customTools,
+            });
+        }
+        else if (isFrontendPlanLedgerNode(input.task) && planLedgerTools) {
+            // Frontend-only split: independent coverage map sessions feed a single
+            // reducer (UX decisions -> global policy -> finalize), mirroring the
+            // backend-test template's module sharding. Small plans normally have one
+            // coverage batch and retain the compact path below.
             let planRequirementIds = [];
             const planRequirementCosts = new Map();
             const behaviorRequiredRequirementIds = [];
@@ -4039,12 +5433,8 @@ export async function executeDagPiNode(input, meta, piStepFn = executePiStep, wr
                     ?.kind === "requirement")
                     .map((record) => record.fact?.id)
                     .filter((id) => typeof id === "string");
-                for (const record of contractFacts) {
-                    const fact = record.fact;
-                    if (fact?.kind !== "requirement" || typeof fact.id !== "string")
-                        continue;
-                    const evidence = fact.evidence;
-                    if (evidence?.behavior === "required")
+                for (const fact of resolveFrontendContractRequirements(contractFacts.map((record) => record.fact))) {
+                    if (fact.evidence.behavior === "required")
                         behaviorRequiredRequirementIds.push(fact.id);
                     planRequirementCosts.set(fact.id, estimateFrontendPlanRequirementRecordCalls(fact));
                 }
@@ -4052,25 +5442,189 @@ export async function executeDagPiNode(input, meta, piStepFn = executePiStep, wr
             catch {
                 // Unreadable ledger falls back to a single coverage session.
             }
-            result = await runFrontendPlanSegmentedSessions({
+            // Independent requirement-coverage batches are map workers. Each worker
+            // owns an isolated typed-event store; only after all workers settle do we
+            // merge facts into the main Plan ledger and run the single UX/global/
+            // finalize reducer. This avoids revision races while shortening the
+            // longest coverage phase for large plans.
+            const runPlanSessions = (options) => runFrontendPlanSegmentedSessions({
                 piStepFn,
-                sessionOptions: piSessionOptions,
-                basePrompt: input.prompt,
+                sessionOptions: options.sessionOptions,
+                basePrompt: options.basePrompt ?? input.prompt,
                 attempt: input.attempt ?? 1,
-                committedFactCount: () => planLedgerTools.committedFactCount(),
-                requirementIds: planRequirementIds,
+                committedFactCount: () => options.ledgerTools.committedFactCount(),
+                requirementIds: options.requirementIds ?? planRequirementIds,
                 requirementCosts: planRequirementCosts,
-                committedRequirementIds: () => planLedgerTools.committedRequirementIds(),
-                committedFacts: () => planLedgerTools.committedFacts(),
+                ...(options.parallelCoverageOnly !== undefined
+                    ? { parallelCoverageOnly: options.parallelCoverageOnly }
+                    : {}),
+                ...(options.compactSmallPlan !== undefined
+                    ? { compactSmallPlan: options.compactSmallPlan }
+                    : {}),
+                committedRequirementIds: () => options.ledgerTools.committedRequirementIds(),
+                committedFacts: () => options.ledgerTools.committedFacts(),
                 behaviorRequiredRequirementIds,
-                setActiveRequirementScope: (requirementIds) => planLedgerTools.setActiveRequirementScope(requirementIds),
+                setActiveRequirementScope: (requirementIds) => options.ledgerTools.setActiveRequirementScope(requirementIds),
                 segmentCustomTools: (toolNames) => toolNames === null
-                    ? planLedgerTools.customTools
-                    : planLedgerTools.customTools.filter((tool) => typeof tool === "object" &&
+                    ? options.ledgerTools.customTools
+                    : options.ledgerTools.customTools.filter((tool) => typeof tool === "object" &&
                         tool !== null &&
                         toolNames.has(tool.name)),
-                flushLedger: () => planLedgerTools.flush(),
+                flushLedger: () => options.ledgerTools.flush(),
             });
+            if (planRequirementIds.length > 1) {
+                const estimatedPlanCalls = planRequirementIds.reduce((total, id) => total + Math.max(1, planRequirementCosts.get(id) ?? 2), 0);
+                const compactEligibleBeforeSharding = planRequirementIds.length <= FRONTEND_PLAN_SMALL_MAX_REQUIREMENTS &&
+                    estimatedPlanCalls > 12 &&
+                    estimatedPlanCalls <= FRONTEND_PLAN_SMALL_MAX_ESTIMATED_CALLS &&
+                    countFrontendPlanTargetSurfaces(input.prompt) === 1;
+                if (compactEligibleBeforeSharding) {
+                    // Decide the small-plan topology before creating coverage shards.
+                    // Sharding first would make the compact two-session path unreachable.
+                    result = await runPlanSessions({
+                        ledgerTools: planLedgerTools,
+                        sessionOptions: piSessionOptions,
+                        compactSmallPlan: true,
+                    });
+                }
+                else {
+                    const coverageBatches = batchFrontendPlanRequirements({
+                        requirementIds: planRequirementIds,
+                        maxEstimatedRecordCalls: FRONTEND_PLAN_COVERAGE_MAX_RECORD_CALLS,
+                        maxRequirements: 4,
+                        requirementCosts: planRequirementCosts,
+                    });
+                    if (coverageBatches.length > 1) {
+                        const shardResults = await mapWithConcurrency(coverageBatches, FRONTEND_PLAN_COVERAGE_MAX_CONCURRENCY, async (slice, index) => {
+                            let shardTools;
+                            let shardResult;
+                            try {
+                                const { createTypedEventStore } = await import("../workflows/dag/frontend-typed-event-store.js");
+                                shardTools = await createFrontendPlanLedgerTools({
+                                    attemptId: `${meta.runId}:${input.task.id}:parallel:${index + 1}`,
+                                    store: createTypedEventStore(),
+                                    runDir: meta.runDir,
+                                    nodeId: `${input.task.id}/parallel/coverage-${index + 1}`,
+                                    skeleton: input.task.structuredContractOutput?.skeleton,
+                                    sourceBinding: meta.spec.sourceBinding,
+                                    requirementIds: slice,
+                                    writeSetPatterns: input.task.writeSet,
+                                    canonicalVerificationTargetIds: await resolveFrontendCanonicalVerificationTargetIds({
+                                        runDir: meta.runDir,
+                                    }),
+                                    componentNewSourceReferences: await resolveFrontendPlanNewComponentSourceReferences({
+                                        cwd: input.cwd,
+                                        sourceBinding: meta.spec.sourceBinding,
+                                    }),
+                                });
+                                const shardSessionOptions = {
+                                    ...piSessionOptions,
+                                    sessionEventsPath: path.join(meta.runDir, input.task.id, "parallel", `coverage-${index + 1}`, "session-events.jsonl"),
+                                };
+                                shardResult = await runPlanSessions({
+                                    ledgerTools: shardTools,
+                                    sessionOptions: shardSessionOptions,
+                                    requirementIds: slice,
+                                    basePrompt: `${input.prompt}\n\nPARALLEL COVERAGE SHARD ${index + 1}: use the frozen canonical behavior verification-target ids listed in the record_plan_verification_target tool description — do not prefix ids with a shard namespace or invent variant ids; identical cross-shard targets are deduped, divergent ones fail the merge.`,
+                                    parallelCoverageOnly: true,
+                                });
+                                await shardTools.flush();
+                                return { index, result: shardResult, tools: shardTools };
+                            }
+                            catch (error) {
+                                const crashMessage = `frontend plan coverage shard ${index + 1} crashed: ${error instanceof Error ? error.message : String(error)}`;
+                                return {
+                                    index,
+                                    tools: shardTools,
+                                    result: shardResult
+                                        ? {
+                                            ...shardResult,
+                                            ok: false,
+                                            failureCategory: shardResult.ok
+                                                ? "invalid-output"
+                                                : shardResult.failureCategory,
+                                            stderr: [shardResult.stderr, crashMessage]
+                                                .filter(Boolean)
+                                                .join("\n"),
+                                        }
+                                        : {
+                                            ok: false,
+                                            assistantText: "",
+                                            command: [],
+                                            durationMs: 0,
+                                            exitCode: null,
+                                            failureCategory: "tool-policy",
+                                            modelDisplay: "unknown",
+                                            parsedEvents: 0,
+                                            stderr: crashMessage,
+                                            stdout: "",
+                                            timedOut: false,
+                                            attemptedModels: [],
+                                            fallbackUsed: false,
+                                            tokensUsed: 0,
+                                        },
+                                };
+                            }
+                        });
+                        const coverageResult = aggregateParallelPiResults(shardResults.map((shard) => shard.result));
+                        let mergeFailure;
+                        try {
+                            // Flatten every shard's facts FIRST, then pre-reduce: shards
+                            // partition requirements but a frozen verification target can
+                            // span shards, so same-id targets must merge across shards
+                            // before adoption, not per shard.
+                            await planLedgerTools.adoptCommittedFacts(reduceParallelCoverageShardRecords(shardResults
+                                .filter((item) => item.result.ok && item.tools)
+                                .flatMap((shard) => normalizeParallelCoverageShardRecords(shard.tools.committedFacts(), shard.index + 1))));
+                        }
+                        catch (error) {
+                            mergeFailure = error;
+                        }
+                        const failedShard = shardResults.find((shard) => !shard.result.ok);
+                        if (mergeFailure) {
+                            result = {
+                                ...coverageResult,
+                                ok: false,
+                                failureCategory: "invalid-output",
+                                stderr: [
+                                    coverageResult.stderr,
+                                    `frontend plan coverage shard merge failed: ${mergeFailure instanceof Error ? mergeFailure.message : String(mergeFailure)}`,
+                                ]
+                                    .filter(Boolean)
+                                    .join("\n"),
+                            };
+                        }
+                        else if (failedShard) {
+                            result = {
+                                ...coverageResult,
+                                ok: false,
+                                stderr: `${coverageResult.stderr}\nfrontend plan coverage shard ${failedShard.index + 1} failed before reduce`.trim(),
+                            };
+                        }
+                        else {
+                            const reducerResult = await runPlanSessions({
+                                ledgerTools: planLedgerTools,
+                                sessionOptions: piSessionOptions,
+                            });
+                            result = combineSequentialPiResults(coverageResult, reducerResult);
+                        }
+                    }
+                    else {
+                        result = await runPlanSessions({
+                            ledgerTools: planLedgerTools,
+                            sessionOptions: piSessionOptions,
+                            compactSmallPlan: true,
+                        });
+                    }
+                }
+            }
+            else {
+                result = await runPlanSessions({
+                    ledgerTools: planLedgerTools,
+                    sessionOptions: piSessionOptions,
+                    compactSmallPlan: true,
+                });
+            }
         }
         else {
             result = await piStepFn({
@@ -4079,6 +5633,18 @@ export async function executeDagPiNode(input, meta, piStepFn = executePiStep, wr
                 ...(writerToolPolicy ? { writerToolPolicy } : {}),
             });
         }
+        try {
+            const verificationCommandFiles = collectFrontendVerificationCommandFiles(input.prompt);
+            const debugPath = path.join(meta.runDir, input.task.id, "verification-command-coverage.json");
+            await import("node:fs/promises").then(({ writeFile, mkdir }) => mkdir(path.dirname(debugPath), { recursive: true }).then(() => writeFile(debugPath, `${JSON.stringify({
+                schemaVersion: 1,
+                basePromptChars: input.prompt.length,
+                verificationCommandFiles,
+            }, null, 2)}\n`)));
+        }
+        catch {
+            // best-effort diagnostic breadcrumb
+        }
     }
     catch (error) {
         if (playwrightToolContext) {
@@ -4121,6 +5687,37 @@ export async function executeDagPiNode(input, meta, piStepFn = executePiStep, wr
         : input.task.writerOutcomePolicy
             ? WRITER_OUTCOME_PROTOCOL_LINE
             : input.task.firstProtocolLine);
+    if (input.task.id === "generate-backend-md-plan-pi") {
+        if (mapped.stopReason === "length") {
+            return {
+                ...mapped,
+                ok: false,
+                failureCategory: STRUCTURED_OUTPUT_RETRY_CATEGORY,
+                stderr: [
+                    mapped.stderr,
+                    "backend-test Markdown plan was truncated before its required protocol could be trusted: stopReason=length; retry with the final artifact first",
+                ]
+                    .filter(Boolean)
+                    .join("\n"),
+            };
+        }
+        if (mapped.ok) {
+            const protocol = assessBackendTestPlanProtocol(mapped.assistantText || mapped.stdout);
+            if (!protocol.ok) {
+                return {
+                    ...mapped,
+                    ok: false,
+                    failureCategory: "invalid-output",
+                    stderr: [
+                        mapped.stderr,
+                        `backend-test Markdown plan protocol invalid: ${protocol.issues.map((issue) => `${issue.code}: ${issue.detail}`).join("; ")}`,
+                    ]
+                        .filter(Boolean)
+                        .join("\n"),
+                };
+            }
+        }
+    }
     // A node-specific budget is enforced for every frontend reader that opts in,
     // including design/final review. Earlier code only checked contract, scout,
     // and plan nodes, leaving the two largest review sessions unbounded.
@@ -4574,7 +6171,7 @@ export async function executeDagPiNode(input, meta, piStepFn = executePiStep, wr
                 runDir: meta.runDir,
                 progress,
                 attempt: input.attempt ?? 1,
-                maxAttempts: input.task.retryPolicy?.maxAttempts ?? 3,
+                maxAttempts: input.task.retryPolicy?.maxAttempts ?? 5,
             });
             if (progress.status !== "PASS") {
                 const classified = classifyBackendTestWriterCompletenessFailure(progress);
@@ -4624,7 +6221,7 @@ export async function executeDagPiNode(input, meta, piStepFn = executePiStep, wr
         stderrParts.push(completenessFailure.detail);
     }
     if (writerThinkingExhausted) {
-        stderrParts.push(`${WRITER_THINKING_EXHAUSTED_CATEGORY}: stopReason=length, thinking observed, 0 write tool calls, 0 attributed diff; recommend a model switch and a new run`);
+        stderrParts.push(`${OUTPUT_LIMIT_RETRY_CATEGORY}: stopReason=length ended the turn before any write tool call; retry from the existing workspace and complete only unfinished targets`);
     }
     if (meta.writeGuardAttribution === "best-effort") {
         stderrParts.push("write guard note: concurrent rank writers use best-effort per-node attribution; keep same-rank writeSet entries disjoint");
@@ -4697,14 +6294,11 @@ export async function executeDagPiNode(input, meta, piStepFn = executePiStep, wr
                         rawFailureCategory === "context-overflow" &&
                             (changeManifestChangedFiles?.length ?? 0) > 0
                             ? "partial-success-with-context-overflow"
-                            : // writer-thinking-exhausted: a length-stopped thinking-only attempt with
-                                // zero write tool calls and zero attributed diff is terminal and
-                                // non-retryable; recommend a model switch + fresh run. Only applies on
-                                // the empty-output base category so provider/transport failures keep
-                                // their original category, and only when the completeness gate did not
-                                // upgrade to incomplete-write-set above.
+                            : // output-limit: a length-stopped attempt is capacity truncation, not
+                                // empty output. Preserve the raw category for diagnostics and let the
+                                // node retry from committed facts/current workspace state.
                                 writerThinkingExhausted
-                                    ? WRITER_THINKING_EXHAUSTED_CATEGORY
+                                    ? OUTPUT_LIMIT_RETRY_CATEGORY
                                     : writerCleanTimeout
                                         ? WRITER_CLEAN_TIMEOUT_RETRY_CATEGORY
                                         : // writer-budget-exhausted: the provider session consumed an
@@ -4860,6 +6454,8 @@ export function mapPiResultToDagNodeResult(result, firstProtocolLine) {
         tokensUsed: result.tokensUsed,
         parsedEvents: result.parsedEvents,
         stopReason: readWriterThinkingExhaustionEvidence(result).stopReason,
+        thinkingObserved: readWriterThinkingExhaustionEvidence(result).thinkingObserved,
+        writeToolCallCount: readWriterThinkingExhaustionEvidence(result).writeToolCallCount,
     };
 }
 function canonicalizeProtocolFirstLine(assistantText, firstProtocolLine) {