@tea-agent/loop-agent 0.39.0-beta.13 → 0.39.0-beta.15

package/dist/workflows/dag/frontend-review-context.js

@@ -287,9 +287,20 @@ export async function runFrontendReviewContextGate(input) {
     if (!parsedContract.success) {
         throw new FrontendReviewContextFailure("review-context-invalid-contract", "frontend review context invalid implementation contract");
     }
+    // Authorized diff surface = deliverable files + every verification-target
+    // file the contract references. The writer legitimately writes VT test
+    // artifacts (behavior verification), and the reviewer must audit their
+    // diffs; scoping this to targets.files alone failed the baseline-overlap
+    // check on app.test.js (r19 extreme smoke).
+    const authorizedChangedPaths = [
+        ...new Set([
+            ...parsedContract.data.targets.files,
+            ...parsedContract.data.verificationTargets.map((target) => target.file),
+        ]),
+    ];
     const diff = await runFrontendWorktreeDiffGate({
         ...input,
-        authorizedChangedPaths: parsedContract.data.targets.files,
+        authorizedChangedPaths,
     });
     const verificationTrace = await readRequiredJson(input.runDir, "contracts/frontend-verification-trace.json");
     const repairAssessment = await readOptionalRepairAssessment(input.runDir);

package/dist/workflows/dag/frontend-shadow-dual-write.js

@@ -260,7 +260,12 @@ export function restorePlanPatchFromCommittedFacts(records) {
         if (isRecord(fact.patch))
             return fact.patch;
     }
-    return undefined;
+    // No finalize-published snapshot: the receipt fix loop can end an attempt
+    // before any finalize_plan succeeds while dozens of record_* facts are
+    // already committed. Assemble the patch from those record facts instead of
+    // declaring the ledger missing (r19: "ledger missing" discarded a ledger
+    // with 92 committed facts).
+    return assemblePlanPatchFromCommittedFacts(records);
 }
 /**
  * A+B (AC-004): reverse of `planLedgerFactsFromPatch` for the incremental
@@ -280,12 +285,33 @@ export function assemblePlanPatchFromCommittedFacts(records, contractRequirement
         PLAN_LEDGER_FACT_KINDS.includes(fact.kind));
     if (facts.length === 0)
         return undefined;
-    for (const fact of facts) {
-        if (isRecord(fact.patch))
-            return fact.patch;
-    }
+    // Singleton record facts: the LAST committed fact wins. The model corrects
+    // a rejected plan by re-recording the fact (r18: 39 finalize retries never
+    // converged because the first state-flow fact kept shadowing the
+    // corrections). Requirements / verification targets / evidence gaps stay
+    // additive (aggregated below); component choices accumulate; singletons
+    // supersede.
+    const lastByKind = (kind) => {
+        let found;
+        for (const fact of facts) {
+            if (fact.kind === kind)
+                found = fact;
+        }
+        return found;
+    };
+    // Patch snapshots (published by earlier finalize calls) carry no routes;
+    // exclude them so a snapshot cannot shadow the model's route selection.
+    const routeSurface = (() => {
+        let found;
+        for (const fact of facts) {
+            if (fact.kind !== "target-surface" || isRecord(fact.patch))
+                continue;
+            found = fact;
+        }
+        return found;
+    })();
     const patch = {};
-    const targetSurface = facts.find((fact) => fact.kind === "target-surface");
+    const targetSurface = routeSurface;
     if (targetSurface) {
         const routes = canonicalStringSet(asStringArray(targetSurface.routes));
         if (routes.length > 0)
@@ -295,14 +321,34 @@ export function assemblePlanPatchFromCommittedFacts(records, contractRequirement
     const collectedChoices = componentChoices.flatMap((fact) => Array.isArray(fact.uiComponentChoices)
         ? fact.uiComponentChoices.filter(isRecord)
         : []);
+    // Later corrections supersede earlier submissions per purpose: without
+    // this, a re-recorded choice (fixing a wrong specReference) would appear
+    // twice and the reviewer would still audit the stale entry (r19).
+    const choicesByPurpose = new Map();
+    const dedupedChoices = [];
+    for (const choice of collectedChoices) {
+        const purpose = asString(choice.purpose);
+        if (!purpose) {
+            dedupedChoices.push(choice);
+            continue;
+        }
+        if (choicesByPurpose.has(purpose)) {
+            const at = dedupedChoices.findIndex((existing) => asString(existing.purpose) === purpose);
+            dedupedChoices[at] = choice;
+        }
+        else {
+            choicesByPurpose.set(purpose, choice);
+            dedupedChoices.push(choice);
+        }
+    }
     const stylingStrategy = componentChoices
         .map((fact) => asString(fact.stylingStrategy))
         .find((value) => value.length > 0);
-    if (collectedChoices.length > 0)
-        patch.uiComponentChoices = collectedChoices;
+    if (dedupedChoices.length > 0)
+        patch.uiComponentChoices = dedupedChoices;
     if (stylingStrategy)
         patch.stylingStrategy = stylingStrategy;
-    const stateFlow = facts.find((fact) => fact.kind === "state-flow");
+    const stateFlow = lastByKind("state-flow");
     if (stateFlow) {
         if (Array.isArray(stateFlow.uiStates)) {
             patch.uiStates = stateFlow.uiStates.filter(isRecord);
@@ -311,17 +357,17 @@ export function assemblePlanPatchFromCommittedFacts(records, contractRequirement
             patch.interactions = stateFlow.interactions.filter(isRecord);
         }
     }
-    const mockApi = facts.find((fact) => fact.kind === "mock-api");
+    const mockApi = lastByKind("mock-api");
     if (mockApi && isRecord(mockApi.mockApi)) {
         patch.mockApi = mockApi.mockApi;
     }
-    const deviation = facts.find((fact) => fact.kind === "design-deviation");
+    const deviation = lastByKind("design-deviation");
     if (deviation) {
         const conflicts = canonicalStringSet(asStringArray(deviation.conflicts));
         if (conflicts.length > 0)
             patch.designEvidence = { conflicts };
     }
-    const dependency = facts.find((fact) => fact.kind === "dependency");
+    const dependency = lastByKind("dependency");
     const dependencyPolicy = asString(dependency?.policy);
     if (dependencyPolicy)
         patch.dependencyPolicy = dependencyPolicy;

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

@@ -3210,6 +3210,7 @@ async function buildFrontendHybridDagFromTask(sources) {
                     ...(requiresOpenspecClassification ? ["When a component choice uses an OpenSpec selection, cite that selection; otherwise do not classify unrelated candidates."] : []),
                     "Call finalize_plan exactly once after the necessary typed facts. Return no Markdown narrative.",
                     "TOOL-ONLY PLAN: Do not read Contract/Scout stdout, task sources, or Scout-confirmed target files. Contract and Scout already own evidence discovery; use the injected upstream facts, record a genuine evidence gap when those facts are insufficient, and start committing record_* facts immediately. For decision=new, pass sourceRequirementIds to record_component_choice; runtime derives the exact PRD citation from the frozen ledger.",
+                    "Output budget protocol (hard, max output <=16K per turn): never enumerate-reason the whole requirement list before your first record_* call — that reasoning burns the entire output budget and the attempt dies with zero committed facts. Process requirements in order: think about ONE requirement briefly, immediately emit its record calls (up to 5 per message), then move to the next. If your budget runs low, stop recording and call finalize_plan with what is committed — the retry ladder continues the remainder in a fresh session.",
                     fixedVerificationContext,
                     scopedOpenspecContext,
                     mockContextBlock,

package/dist/workflows/dag/types.js

@@ -315,9 +315,11 @@ export const dagFrontendVerificationBundleSchema = z
     behaviorEvidence: dagShellVerifyEvidenceSchema,
     lintBaselineNodeId: dagFrontendNodeIdSchema.optional(),
     writerNodeIds: z.array(dagFrontendNodeIdSchema).optional(),
-    /** M8: verify-shell only ever runs in initial mode; a failure is terminal
-     * (routes to recovery) and never selects a same-run repair branch. */
-    mode: z.literal("initial"),
+    /** "initial" = first-pass assess bundle; "repair" = the convergence
+     * loop's post-repair reverify, which re-runs the same frozen commands
+     * and fails on any failure (still no same-run repair branch inside the
+     * node itself). */
+    mode: z.enum(["initial", "repair"]),
 })
     .superRefine((bundle, context) => {
     const groups = [

package/harness.json

@@ -83,8 +83,8 @@
     "pi": {
       "description": "Pi 负责规划、评审、诊断；当 DAG toolProfile=write 时也可做有界写入。模型按复杂度三档配置，支持 provider/model 字符串或带 thinking 的对象；斜杠前为 Pi provider，后为 modelId，勿只写裸 modelId。",
       "LOW": "wizard-local/minimax-m3",
-      "MED": "wizard-local/gpt-5.6-sol",
-      "HIGH": "wizard-local/gpt-5.6-sol"
+      "MED": "wizard-local/gpt-5.5",
+      "HIGH": "wizard-local/gpt-5.5"
     }
   }
-}
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tea-agent/loop-agent",
-  "version": "0.39.0-beta.13",
+  "version": "0.39.0-beta.15",
   "type": "module",
   "bin": {
     "loop-agent": "bin/loop-agent.js",
@@ -106,4 +106,4 @@
     "vitest": "^3.2.4"
   },
   "packageManager": "pnpm@10.33.0+sha512.10568bb4a6afb58c9eb3630da90cc9516417abebd3fabbe6739f0ae795728da1491e9db5a544c76ad8eb7570f5c4bb3d6c637b2cb41bfdcdb47fa823c8649319"
-}
+}