@tea-agent/loop-agent 0.43.0-next.18 → 0.43.0-next.19

package/dist/workflows/dag/frontend-shape.js

@@ -9,6 +9,7 @@
  */
 import { createHash } from "node:crypto";
 import { z } from "zod";
+import { collectFrontendSharedSignals, isNarrowProductSurface } from "./frontend-risk.js";
 /** Fixture-only node counts. They are NOT the authoritative 15/13/11
  * topology: split-required has no count. */
 export const FRONTEND_SHAPE_NODE_COUNTS = {
@@ -17,16 +18,6 @@ export const FRONTEND_SHAPE_NODE_COUNTS = {
     standard: 11,
     high: 11,
 };
-const SHAPE_ESCALATION_PATTERNS = [
-    { id: "auth-permission", re: /\bauth\b|\bpermission\b|\brbac\b|登录|权限|\bjwt\b|\bsession\b/i },
-    { id: "sensitive-action", re: /\bpayment\b|delete\s+account|支付|删除用户|\bdestructive\b/i },
-    { id: "ssr-boundary", re: /\bssr\b|\brsc\b|server\s+component|\bhydration\b|server-client/i },
-    { id: "new-route", re: /new\s+route|新增路由|createBrowserRouter|add\s+route/i },
-    { id: "shared-public-api", re: /public\s+api|shared\s+component|公共组件|破坏性|breaking\s+change/i },
-    { id: "schema-api-change", re: /\bopenapi\b|graphql\s+schema|api\s+contract|接口变更|schema\s+change/i },
-    { id: "new-dependency", re: /new\s+dependency|add\s+dependency|npm\s+install|新增依赖/i },
-    { id: "cross-domain-paths", re: /多个领域|cross-module|multiple\s+packages/i },
-];
 const SHAPE_SMALL_PATTERNS = [
     { id: "single-component", re: /single\s+component|单一组件|局部样式|style\s+tweak|css\s+only|文案|copy\s+change/i },
     { id: "no-remote", re: /no\s+api|无\s*api|无\s*接口|本地状态|local\s+state\s+only|pure\s+ui/i },
@@ -59,48 +50,13 @@ function pathSpreadScore(paths) {
     return tops.size;
 }
 function collectEscalationSignals(blob, supervised) {
-    const signals = [];
-    for (const pattern of SHAPE_ESCALATION_PATTERNS) {
-        if (hasAffirmativeShapeSignal(blob, pattern))
-            signals.push(pattern.id);
-    }
-    if (supervised)
-        signals.push("supervised-profile");
+    const signals = collectFrontendSharedSignals(blob, supervised);
     if (SHAPE_SPLIT_REQUIRED_PATTERN.test(blob) &&
         !SHAPE_SPLIT_REQUIRED_NEGATION.test(blob)) {
         signals.push("large-requirement-split");
     }
     return [...new Set(signals)];
 }
-function isExplicitShapeExclusion(clause, pattern) {
-    const match = new RegExp(pattern.re.source, "i").exec(clause);
-    if (!match)
-        return false;
-    const prefix = clause.slice(0, match.index).trim();
-    const suffix = clause.slice(match.index + match[0].length).trim();
-    const actionInSignal = ["new-route", "new-dependency"].includes(pattern.id);
-    if (suffix && !(actionInSignal && /^no$/i.test(prefix) && /^is\s+(?:required|needed)$/i.test(suffix)))
-        return false;
-    if (/^(?:不涉及|不(?:修改|变更|调整|改变|新增|添加)|无需(?:修改|变更|调整|改变|新增|添加)|no\s+changes?\s+to|(?:do\s+not|don't)\s+(?:change|modify))$/i.test(prefix))
-        return true;
-    return actionInSignal && /^(?:不|禁止|无需|无须|不需要|no|without|do\s+not|don't)$/i.test(prefix);
-}
-/** Only a complete, independent scope exclusion can suppress a risk signal.
- * Nested negation, conditions and trailing consequences remain risk evidence;
- * an excluded occurrence never cancels another occurrence elsewhere. */
-function hasAffirmativeShapeSignal(blob, pattern) {
-    for (const sentence of blob.split(/[\n。.;；!?！？]/)) {
-        if (!pattern.re.test(sentence))
-            continue;
-        // Commas can attach a condition to an exclusion. A comma-separated list
-        // is independent only when every item is itself an explicit exclusion.
-        const clauses = sentence.split(/[，,、]/).map(part => part.trim().replace(/^(?:[-*+]|\d+\))\s+/, ""));
-        if (clauses.every(clause => SHAPE_ESCALATION_PATTERNS.some(candidate => isExplicitShapeExclusion(clause, candidate))))
-            continue;
-        return true;
-    }
-    return false;
-}
 /** Negation applies to the API mention itself, never to unrelated Mock policy. */
 function hasFrontendRemoteSignal(blob) {
     const remaining = blob
@@ -126,7 +82,7 @@ function resolveBaselineFrontendTaskShape(input) {
     }
     const paths = input.allowedPaths ?? [];
     const spread = pathSpreadScore(paths);
-    const concentrated = spread <= 1 && paths.length <= 4;
+    const concentrated = paths.length > 0 && spread <= 1 && paths.length <= 4;
     const hasApiSignal = hasFrontendRemoteSignal(blob);
     const complexity = (input.complexity ?? "").toLowerCase();
     // split-required wins over everything: it is a termination/escalation
@@ -172,17 +128,17 @@ function resolveBaselineFrontendTaskShape(input) {
             reason: "micro scope: trivial single-file change without remote API",
         };
     }
-    // small: no escalation, no supervised, small keywords, concentrated paths,
-    // no remote API signal.
-    if (smallHits.length > 0 &&
-        concentrated &&
-        !hasApiSignal &&
-        (complexity === "small" || complexity === "low")) {
+    // small: no escalation, no supervised, narrow single-domain delivery
+    // surface, no remote API signal, and not an explicitly large task. Keyword
+    // and explicit "small/low" complexity hints are no longer required — they
+    // were a fragile proxy for "this is genuinely a small change" and blocked
+    // genuinely-small tasks whose authors used different wording.
+    if (isNarrowProductSurface(paths) && !hasApiSignal && complexity !== "large") {
         return {
             shape: "small",
-            signals: smallHits,
+            signals: smallHits.length > 0 ? smallHits : ["structured-small"],
             rejectedSignals: [],
-            reason: `small signals without escalation: ${smallHits.join(", ")}`,
+            reason: "narrow single-domain delivery surface without remote API or escalation; small topology",
         };
     }
     const rejected = smallHits.map((id) => `small:${id}:blocked-by-${hasApiSignal ? "api-signal" : "path-spread"}`);

package/dist/workflows/dag/frontend-test-execution-evidence.js

@@ -1,21 +1,18 @@
 import path from "node:path";
 import { readFile } from "node:fs/promises";
 import { z } from "zod";
-/** Observe supported frozen Vitest commands without changing their selection. */
-export async function withFrontendTestReporter(input) {
-    // Compound commands/wrappers need their own adapter: never append flags to
-    // a different shell command, change selection, or override a user reporter.
-    if (/[;&|<>`\n\r$]/.test(input.command) || /--(?:reporter|outputFile)/.test(input.command))
-        return undefined;
-    const direct = /^(?:(?:npx|pnpm exec|yarn exec)\s+)?vitest\s+run(?:\s|$)/.test(input.command.trim());
-    const npm = /^npm\s+(?:run\s+([\w:-]+)|(test))(?=\s|$)(.*)$/.exec(input.command.trim());
+/** Decide whether a bare vitest/npm-test command is observable and, if so,
+ * return the command with the separator (" --") a reporter would need. */
+async function adaptPlainTestCommand(command, cwd) {
+    const direct = /^(?:(?:npx|pnpm exec|yarn exec)\s+)?vitest\s+run(?:\s|$)/.test(command.trim());
+    const npm = /^npm\s+(?:run\s+([\w:-]+)|(test))(?=\s|$)(.*)$/.exec(command.trim());
     let separator = "";
     if (!direct) {
         if (!npm)
             return undefined;
         let manifest;
         try {
-            manifest = JSON.parse(await readFile(path.join(input.cwd, "package.json"), "utf8"));
+            manifest = JSON.parse(await readFile(path.join(cwd, "package.json"), "utf8"));
         }
         catch {
             return undefined;
@@ -27,8 +24,36 @@ export async function withFrontendTestReporter(input) {
             return undefined;
         separator = npm[3]?.trim() ? "" : " --";
     }
+    return `${command}${separator}`;
+}
+/** Observe supported frozen Vitest commands without changing their selection. */
+export async function withFrontendTestReporter(input) {
+    // Compound commands/wrappers need their own adapter: never append flags to
+    // a different shell command, change selection, or override a user reporter.
+    if (/[;&|<>`\n\r$]/.test(input.command) || /--(?:reporter|outputFile)/.test(input.command))
+        return undefined;
+    // Frozen verify commands are env-prefixed bash wrappers —
+    // `HARNESS_ALLOW_ACTIVE_DAG_RUNS=1 'bash' '-lc' 'npm test'` — so the JSON
+    // reporter must be injected INSIDE the quoted payload; appending it to the
+    // wrapper would target bash's own CLI and produce no test report at all
+    // (smoke r54: every behavior verification target then failed the trace
+    // gate with "test execution report unavailable").
+    const bashWrapped = /^((?:[A-Za-z_][A-Za-z0-9_]*=(?:'[^']*'|[^\s'"]+)\s+)*)'bash'\s+'-lc'\s+'([^']*)'\s*$/.exec(input.command);
+    if (bashWrapped) {
+        const inner = bashWrapped[2];
+        if (/[;&|<>`\n\r$]/.test(inner) || /--(?:reporter|outputFile)/.test(inner))
+            return undefined;
+        const adaptedInner = await adaptPlainTestCommand(inner, input.cwd);
+        if (!adaptedInner)
+            return undefined;
+        const quote = (value) => `'${value.replace(/'/g, `'"'"'`)}'`;
+        return input.command.replace(`'${inner}'`, `'${adaptedInner} --reporter=json --outputFile=${quote(input.reportPath)}'`);
+    }
+    const adapted = await adaptPlainTestCommand(input.command, input.cwd);
+    if (!adapted)
+        return undefined;
     const quote = (value) => `'${value.replace(/'/g, `'"'"'`)}'`;
-    return `${input.command}${separator} --reporter=json --outputFile=${quote(input.reportPath)}`;
+    return `${adapted} --reporter=json --outputFile=${quote(input.reportPath)}`;
 }
 const resultSchema = z.object({
     testResults: z.array(z.object({

package/dist/workflows/dag/frontend-typed-event-store.js

@@ -31,18 +31,11 @@ export const TYPED_EVENT_FACT_KINDS = [
     "provider-ended-without-terminal",
     "read-only-fact-channel",
 ];
-/**
- * M2 shadow-only fact catalog (AC-001 shadow dual write). These facts are
- * audit-only observations derived from the authoritative text/JSON outputs of
- * the frontend-contract-pi / frontend-scout-pi / frontend-plan-pi nodes. They
- * are never branch-control signals and never participate in writer admission.
- *
- * Kept deliberately separate from `TYPED_EVENT_FACT_KINDS`: widening the shadow
- * catalog must never widen the provider capability matrix (which stays frozen
- * at its seven kinds). `target-surface` is shared by the scout and plan
- * origins, so the union of the two catalogs is 12 unique kinds.
- */
-export const SHADOW_FACT_KINDS = [
+/** Frontend DAG decision facts. These are committed runtime facts, not a
+ * secondary text channel; Policy, Admission and Plan compilation read them
+ * directly. Kept separate from provider capability facts so the provider
+ * matrix remains frozen. */
+export const FRONTEND_DAG_FACT_KINDS = [
     "requirement",
     "constraint",
     "open-question",
@@ -56,11 +49,26 @@ export const SHADOW_FACT_KINDS = [
     "design-deviation",
     "dependency",
 ];
+/** Experimental decision Plan fact catalog. The decision tools emit semantic
+ * decisions (module placement, reuse, state ownership, Mock boundary, behavior
+ * verification focus); `data-flow` and `dependency` reuse the DAG catalog above.
+ * Kept as a dedicated catalog so widening the decision surface never widens the
+ * frozen provider / event / review / design catalogs. */
+export const FRONTEND_DECISION_FACT_KINDS = [
+    "module-placement",
+    "reuse-decision",
+    "state-ownership",
+    "api-mock-boundary",
+    "verification-focus",
+];
+/** Terminal fact for the experimental decision Plan path. Kept separate from the
+ * decision fact catalog so it never gets collected as a decision entry. */
+export const DECISION_TERMINAL_FACT_KINDS = ["finalize_decision"];
 /**
  * M5 review terminal fact catalog (AC-001). These two committed typed tools are
  * the authoritative terminal verdict of `frontend-review-pi`. They are kept in
  * a dedicated catalog so widening the review terminal surface never widens the
- * frozen 7/7/12 provider/shadow/event catalogs.
+ * frozen provider/event catalogs.
  */
 export const REVIEW_TERMINAL_FACT_KINDS = [
     "approve_review",
@@ -72,7 +80,7 @@ export const REVIEW_TERMINAL_FACT_KINDS = [
  * `approve_design` / `request_design_changes` instead of the legacy first-line
  * `VERDICT: pass|request-revision` text protocol. Kept in a dedicated catalog
  * so widening the design terminal surface never widens the frozen provider /
- * shadow / event / review catalogs.
+ * event / review catalogs.
  */
 export const DESIGN_TERMINAL_FACT_KINDS = [
     "approve_design",
@@ -84,7 +92,7 @@ export const DESIGN_TERMINAL_FACT_KINDS = [
  * open-question/split-proposal facts plus the `contract-finalized` terminal
  * fact carrying the terminal disposition (`ready | ready-with-assumptions |
  * blocked` + optional `blockingOwner`). Kept in a dedicated catalog so
- * widening the contract surface never widens the frozen provider / shadow /
+ * widening the contract surface never widens the frozen provider /
  * event / review / design catalogs.
  */
 export const CONTRACT_FACT_KINDS = [
@@ -142,8 +150,8 @@ export const FRONTEND_ISSUE_CATEGORY_VALUES = [
 ];
 export const frontendIssueCategorySchema = z.enum(FRONTEND_ISSUE_CATEGORY_VALUES);
 /** Runtime control facts committed at frontend shape boundaries. Kept outside
- * the frozen provider/shadow catalogs so adding control authority cannot widen
- * either legacy capability surface. */
+ * the frozen provider catalog so adding control authority cannot widen that
+ * capability surface. */
 export const FRONTEND_SHAPE_FACT_KINDS = ["frontend-shape"];
 export const typedEventPhaseSchema = z.enum([
     "staged",
@@ -153,7 +161,9 @@ export const typedEventPhaseSchema = z.enum([
 const ALL_TYPED_EVENT_FACT_KIND_VALUES = [
     ...new Set([
         ...TYPED_EVENT_FACT_KINDS,
-        ...SHADOW_FACT_KINDS,
+        ...FRONTEND_DAG_FACT_KINDS,
+        ...FRONTEND_DECISION_FACT_KINDS,
+        ...DECISION_TERMINAL_FACT_KINDS,
         ...REVIEW_TERMINAL_FACT_KINDS,
         ...DESIGN_TERMINAL_FACT_KINDS,
         ...FRONTEND_SHAPE_FACT_KINDS,
@@ -166,7 +176,7 @@ const ALL_TYPED_EVENT_FACT_KIND_VALUES = [
 export const typedEventFactKindSchema = z.enum(ALL_TYPED_EVENT_FACT_KIND_VALUES);
 /** A typed event fact is a `kind`-discriminated record. The schema stays
  * permissive beyond `kind` (arbitrary JSON-safe payload) so provider/model
- * shadow facts can carry their own fields, but `kind` is always validated. */
+ * provider facts can carry their own fields, but `kind` is always validated. */
 export const typedEventFactSchema = z
     .record(z.unknown())
     .superRefine((value, ctx) => {
@@ -178,8 +188,7 @@ export const typedEventFactSchema = z
         });
     }
 });
-/** A single review finding shared by both typed terminal verdicts. Mirrors the
- * legacy JSON review finding shape so typed↔JSON equivalence stays total. */
+/** A single review finding shared by both typed terminal verdicts. */
 export const reviewFindingSchema = z
     .object({
     severity: z.enum(["Critical", "Important", "Minor", "Info"]),
@@ -190,8 +199,7 @@ export const reviewFindingSchema = z
 })
     .strict();
 /** `approve_review` terminal fact payload. Approval may carry informational
- * findings (Minor/Info) but never Critical/Important — the same rule the legacy
- * JSON verdict applies to `verdict: "pass"`. */
+ * findings (Minor/Info) but never Critical/Important. */
 export const approveReviewFactSchema = z
     .object({
     kind: z.literal("approve_review"),
@@ -221,8 +229,7 @@ export const requestReviewChangesFactSchema = z
 })
     .strict();
 /** `approve_design` terminal fact payload. Approval may carry informational
- * findings (Minor/Info) but never Critical/Important — the same rule the legacy
- * text protocol applies to `VERDICT: pass`. */
+ * findings (Minor/Info) but never Critical/Important. */
 export const approveDesignFactSchema = z
     .object({
     kind: z.literal("approve_design"),

package/dist/workflows/dag/frontend-verification-trace.js

@@ -260,6 +260,8 @@ export async function evaluateFrontendWriterTracePreflight(input) {
             file: target.file,
             ...(target.symbol ? { symbol: target.symbol } : {}),
             status: fileTrace.issues.length ? "failed" : "ok",
+            evidence: fileTrace.issues.length ? "unavailable" : "file-exists",
+            behaviorCoverage: "unconfirmed",
             matchedNodeIds: [],
             issues: fileTrace.issues,
         });
@@ -296,7 +298,10 @@ export async function evaluateVerificationTargets(input) {
         });
         issues.push(...fileTrace.issues);
         const executed = input.executedTests?.get(target.commandLabel);
-        if (target.mode !== "static" && executed !== undefined) {
+        if (target.mode !== "static" && executed === undefined) {
+            issues.push(`test execution report unavailable for behavior target: ${target.id}`);
+        }
+        else if (target.mode !== "static" && executed !== undefined) {
             // safePath permits "./" prefixes; reporter files are cwd-relative.
             const targetFile = path.posix.normalize(target.file.replace(/\\/g, "/"));
             const matching = executed.filter((test) => test.file === targetFile);
@@ -314,6 +319,10 @@ export async function evaluateVerificationTargets(input) {
             file: target.file,
             symbol: target.symbol,
             status,
+            evidence: issues.length
+                ? "unavailable"
+                : target.mode === "static" ? "command-execution" : "test-file-execution",
+            behaviorCoverage: "unconfirmed",
             matchedNodeIds: owners ?? [],
             issues,
         });
@@ -520,7 +529,7 @@ export function formatFrontendVerificationTraceStdout(result) {
         `Browser: ${result.browserStatus}`,
         `Visual: ${result.visualStatus}`,
         `Targets: ${result.targets.length}`,
-        ...result.targets.map((target) => `- ${target.id}: ${target.status} label=${JSON.stringify(target.commandLabel)} file=${target.file}`),
+        ...result.targets.map((target) => `- ${target.id}: execution=${target.status} evidence=${target.evidence} behavior-coverage=${target.behaviorCoverage} label=${JSON.stringify(target.commandLabel)} file=${target.file}`),
     ];
     if (result.artifactPath) {
         lines.push(`Artifact: ${result.artifactPath}`);

package/dist/workflows/dag/frontend-writer-admission.js

@@ -22,9 +22,8 @@ export const frontendWriterAdmissionClassificationSchema = z.enum([
     "requires-human-approval",
     "blocked",
     "stale",
-    // Legacy artifacts remain parseable for diagnostics, but `admitted` is never
-    // accepted by the runtime authorization gate. Canonical producers always
-    // emit one of the four values above.
+    // `admitted` is retained only as an invalidated input value; the runtime
+    // authorization gate accepts only the four canonical classifications above.
     "admitted",
 ]);
 /** A+B (AC-006): Micro topology uses `requirement-to-target-and-evidence`
@@ -59,8 +58,7 @@ export const frontendAdmissionFailureSourceSchema = z.enum([
 /**
  * Deterministic artifact schema for `contracts/frontend-writer-admission-result.json`.
  * Scheduler admission gate and frontend recovery planning read this artifact
- * (fail-closed); it replaces the legacy frontend-prewrite-result-v1 as the
- * write authorization source.
+ * (fail-closed); it is the write authorization source for new runs.
  */
 export const frontendWriterAdmissionResultV1Schema = z
     .object({
@@ -275,45 +273,3 @@ export function deriveFrontendWriterAdmissionShell(input) {
     };
     return result;
 }
-function normalizeNewVerdict(classification) {
-    return classification === "accepted" ||
-        classification === "requires-human-approval" ||
-        classification === "admitted"
-        ? "accepted"
-        : "blocked";
-}
-function normalizeLegacyVerdict(classification) {
-    return classification === "accepted" || classification === "accepted-normalized"
-        ? "accepted"
-        : "blocked";
-}
-export function compareAdmissionToLegacyGate(input) {
-    const newVerdict = normalizeNewVerdict(input.newClassification);
-    const legacyVerdict = normalizeLegacyVerdict(input.legacyClassification);
-    return {
-        newClassification: input.newClassification,
-        legacyClassification: input.legacyClassification,
-        newVerdict,
-        legacyVerdict,
-        match: newVerdict === legacyVerdict,
-    };
-}
-/** Render a stable, deterministically ordered mismatch list for the evaluation
- * set. Input order is irrelevant: rows are sorted by a canonical key. */
-export function renderAdmissionShadowDiff(results) {
-    const indexed = results.map((result, index) => ({ ...result, index }));
-    const sorted = indexed.sort((a, b) => {
-        const keyA = `${a.legacyClassification}|${a.newClassification}`;
-        const keyB = `${b.legacyClassification}|${b.newClassification}`;
-        if (keyA !== keyB)
-            return keyA < keyB ? -1 : 1;
-        return a.index - b.index;
-    });
-    const mismatches = sorted.filter((row) => !row.match);
-    const lines = [
-        "## Admission Shadow Diff",
-        `${sorted.length} comparisons, ${mismatches.length} mismatches`,
-        ...sorted.map((row) => `- ${row.legacyClassification} -> ${row.newClassification}: legacy=${row.legacyVerdict} new=${row.newVerdict}${row.match ? " match" : " MISMATCH"}`),
-    ];
-    return lines.join("\n");
-}

package/dist/workflows/dag/frontend-writer-status.js

@@ -161,26 +161,6 @@ export function buildFrontendMicroLoopCheckpoint(input) {
         writeAttempts: input.facts.writeAttempts,
     };
 }
-/**
- * Shadow comparison between the fact-derived status and the legacy
- * `IMPLEMENTATION_OUTCOME:` first-line parse. Never used to decide ok/fail —
- * only persisted for the M4 shadow evaluation.
- */
-export function compareFactStatusToLegacyOutcome(factStatus, legacyOutcome) {
-    const factVerdict = factStatus === "changed" || factStatus === "already-satisfied"
-        ? "ok"
-        : "fail";
-    const legacyVerdict = legacyOutcome === "changed" || legacyOutcome === "already-satisfied"
-        ? "ok"
-        : "fail";
-    return {
-        factStatus,
-        legacyOutcome,
-        factVerdict,
-        legacyVerdict,
-        match: factVerdict === legacyVerdict,
-    };
-}
 /**
  * Pure session-event scanner: count write-tool events from the persisted
  * session-events.jsonl content. Fail-closed: an unparseable line never counts
@@ -249,8 +229,5 @@ export async function collectFrontendWriterFacts(input) {
         wallTimeMs: input.wallTimeMs ?? 0,
         rounds: input.rounds ?? 1,
         writeAttempts: input.writeAttempts ?? 1,
-        ...(input.firstLineText !== undefined
-            ? { firstLineText: input.firstLineText }
-            : {}),
     };
 }