@mutagent/evaluator 0.1.0-alpha.3 → 0.1.0-alpha.5

package/.claude/skills/mutagent-evaluator/scripts/run-deterministic.ts

@@ -32,6 +32,7 @@ import {
   type EvalMatrix,
   type RowResultValue,
   type ScorecardCriterion,
+  CheckMethod,
   RowResult,
   Track,
   trackForCheckMethod,
@@ -84,14 +85,96 @@ function isNonEmpty(value: unknown): boolean {
   return true;
 }
 
+/** A STRUCTURED, non-empty value: a non-empty object or array. A present-but-
+ *  scalar/empty artifact is NOT structured — the typebox-schema re-exec fails it
+ *  (a hollow presence-pass under the old proxy). PURE. */
+function isStructuredNonEmpty(value: unknown): boolean {
+  if (Array.isArray(value)) return value.length > 0;
+  if (value !== null && typeof value === "object") return Object.keys(value).length > 0;
+  return false;
+}
+
+/** Scan a produced artifact for an EXPLICIT gate-failure marker: a falsy pass
+ *  flag (`pass`/`ok`/`passed` === false) or a failing `status` stamp
+ *  (`fail`/`failed`/`error`). Conservative + shallow so it stays deterministic
+ *  and does not false-fail on unrelated substrings. PURE. */
+function hasFailureMarker(value: unknown): boolean {
+  if (value === null || typeof value !== "object") return false;
+  const obj = value as Record<string, unknown>;
+  for (const flag of ["pass", "ok", "passed"] as const) {
+    if (obj[flag] === false) return true;
+  }
+  const status = obj["status"];
+  if (typeof status === "string") {
+    const s = status.toLowerCase();
+    if (s === "fail" || s === "failed" || s === "error") return true;
+  }
+  return false;
+}
+
+/**
+ * Item #4 — the REAL live re-execution of a deterministic row. Replaces the
+ * presence-proxy (`the artifact exists → mark it covered`) with an ACTUAL
+ * checkMethod-keyed predicate over the artifact's LIVE content:
+ *   - typebox-schema     : the produced artifact must be a STRUCTURED, non-empty
+ *                          object/array — a present-but-malformed scalar now FAILS
+ *                          instead of hollow-passing (real shape conformance).
+ *   - gate               : the evidence must be present AND carry NO explicit
+ *                          failure marker (fail-loud gate — `hasFailureMarker`).
+ *   - deterministic-script: the produced artifact must be present + non-empty
+ *                          (the script actually emitted its output).
+ * Returns pass/fail when the criterion maps to a live bundle artifact; returns
+ * `null` when NO artifact answers it offline — the caller then records the
+ * documented seam-skip. A genuinely un-runnable row would require executing the
+ * SUBJECT'S OWN script, which the judge-only cell never does (EV-051) and which
+ * dispatches NO sub-agents (PR-ORCH-01). PURE + deterministic (no clock/random/
+ * network) ⇒ C-PIN byte-identical reruns.
+ */
+export function reexecDeterministicCheck(
+  criterion: Criterion,
+  bundle: RunBundle,
+): RowResultValue | null {
+  const artifactKey = evidenceArtifactFor(criterion, bundle);
+  if (artifactKey == null) return null; // no live artifact → not re-executable here
+  const value = bundle.data[artifactKey];
+  // HONESTY on the depth of each re-exec (do NOT mistake these for deep checks):
+  //   - typebox-schema : a real SHAPE upgrade over presence (rejects a present-but-
+  //                      scalar), but it is NOT a full TypeBox `Value.Check` against
+  //                      the produced artifact's declared schema — it asserts
+  //                      "structured + non-empty", not field-level conformance.
+  //   - gate           : `hasFailureMarker` is TOP-LEVEL-ONLY (scans the object's own
+  //                      pass/ok/passed/status keys) — it does NOT recurse into nested
+  //                      gate evidence, so a failure buried in a child object is missed.
+  //   - deterministic-script : semantically ≡ the OLD presence proxy (present +
+  //                      non-empty). It is re-run against LIVE data (so it is a real
+  //                      re-execution, not a cached verdict), but carries no deeper
+  //                      predicate than presence. Deepening any of these to a true
+  //                      re-run of the subject's own script is the documented seam
+  //                      (out of scope for a judge-only / no-sub-agent cell).
+  switch (criterion.checkMethod) {
+    case CheckMethod.TypeboxSchema:
+      return isStructuredNonEmpty(value) ? RowResult.Pass : RowResult.Fail;
+    case CheckMethod.Gate:
+      return isNonEmpty(value) && !hasFailureMarker(value) ? RowResult.Pass : RowResult.Fail;
+    case CheckMethod.DeterministicScript:
+    default:
+      return isNonEmpty(value) ? RowResult.Pass : RowResult.Fail;
+  }
+}
+
 export interface EvaluateOptions {
   /**
-   * When false (default), criteria that need a LIVE re-exec of the subject's own
-   * script (beyond bundle presence/shape) return `skip` with a documented seam
-   * note instead of false-passing. A future integration may set this true and
-   * provide a re-exec callback.
+   * Item #4 — the LIVE RE-EXEC executor. When supplied it RUNS the real
+   * deterministic check against live bundle data (see `reexecDeterministicCheck`,
+   * the default) and its pass/fail return is authoritative — replacing the
+   * presence-proxy that marked a row covered from mere artifact existence. A
+   * `null` return means "this executor cannot answer this row" ⇒ the legacy
+   * (byte-stable) path runs. ABSENT ⇒ no re-exec wired: fully back-compatible,
+   * byte-identical to before (the presence proxy + documented seam-skip). The
+   * executor is PURE + dispatches NO sub-agents (PR-ORCH-01) + never fixes the
+   * subject (EV-051).
    */
-  liveReexec?: (criterion: Criterion) => RowResultValue | null;
+  liveReexec?: (criterion: Criterion, bundle: RunBundle) => RowResultValue | null;
 }
 
 /**
@@ -104,21 +187,31 @@ export function evaluateDeterministic(
   opts: EvaluateOptions = {},
 ): ScorecardCriterion {
   const track = Track.Deterministic;
+
+  // Item #4 — LIVE RE-EXEC FIRST. When an executor is wired it RUNS the real
+  // deterministic check against live data; a pass/fail is authoritative and
+  // REPLACES the presence-proxy (no more "an artifact exists → covered"). A null
+  // return means the executor could not answer this row ⇒ fall through to the
+  // (byte-stable) legacy path below.
+  const live = opts.liveReexec?.(criterion, bundle) ?? null;
+  if (live === RowResult.Pass || live === RowResult.Fail || live === RowResult.Incomplete) {
+    return {
+      dimension: criterion.dimension,
+      severity: criterion.severity,
+      checkMethod: criterion.checkMethod,
+      track,
+      result: live,
+      detail: `live re-exec (${criterion.checkMethod}): real deterministic check over the produced artifact → ${live}`,
+    };
+  }
+
   const artifactKey = evidenceArtifactFor(criterion, bundle);
 
   if (artifactKey == null) {
-    // No bundle artifact answers this row offline -> live-reexec seam.
-    const live = opts.liveReexec?.(criterion) ?? null;
-    if (live) {
-      return {
-        dimension: criterion.dimension,
-        severity: criterion.severity,
-        checkMethod: criterion.checkMethod,
-        track,
-        result: live,
-        detail: "evaluated via live-reexec callback",
-      };
-    }
+    // No bundle artifact answers this row offline -> documented seam-skip. NOT
+    // silent: the coverage-honesty warning (computeCoverage) surfaces it, and
+    // fully executing it would require running the subject's OWN script — which
+    // the judge-only cell never does (EV-051) and which dispatches no sub-agents.
     return {
       dimension: criterion.dimension,
       severity: criterion.severity,
@@ -130,6 +223,9 @@ export function evaluateDeterministic(
     };
   }
 
+  // LEGACY presence proxy — reached ONLY when no live-reexec executor is wired
+  // (kept for byte-identity with pre-Item-#4 callers). With the default executor
+  // wired (audit-run.ts) this branch is never taken for a mapped row.
   const present = isNonEmpty(bundle.data[artifactKey]);
   return {
     dimension: criterion.dimension,
@@ -138,7 +234,7 @@ export function evaluateDeterministic(
     track,
     result: present ? RowResult.Pass : RowResult.Fail,
     detail: present
-      ? `evidence artifact '${artifactKey}' present + non-empty`
+      ? `evidence artifact '${artifactKey}' present + non-empty (presence proxy — no live re-exec wired)`
       : `evidence artifact '${artifactKey}' MISSING/empty in bundle (fail-loud)`,
   };
 }

package/.claude/skills/mutagent-evaluator/scripts/run-evaluate.ts

@@ -542,7 +542,7 @@ export function readRunVerdictFiles(input: EvaluateRunInput): MatrixVerdictFile[
  * REPORT — build the rich `EvalReportInput` from the REAL run outputs (the folded
  * scorecard + per-criterion verdicts + the per-trajectory Judge-Agent verdict
  * files + the matrix criteria + the EV-051 handover) and render the 5-tab HTML
- * eval-report into `<reportDir>/report.html`. This is the wiring D-2/D-3 had been
+ * eval-report into `<reportDir>/evaluation-report.html`. This is the wiring D-2/D-3 had been
  * MISSING: `*evaluate` produced a scorecard but emitted no report. The report
  * consumes the §9.4 judge-walk fields when present (side-by-side) and degrades to
  * the per-trajectory scorecard otherwise. Returns the written report path.
@@ -555,6 +555,9 @@ export function writeRunReport(
   result: EvaluateRunResult,
   runId: string,
   cwd?: string,
+  // PRD R-4 — render audience (resolved by the caller: explicit flag → config.default_audience →
+  // "client"). Omitted ⇒ the renderer defaults to "client" (leak-safe). "internal" shows §5.
+  audience?: "client" | "internal",
 ): string {
   const matrixVerdictFiles = readRunVerdictFiles(input);
   const reportInput = buildEvalReportInput({
@@ -586,9 +589,10 @@ export function writeRunReport(
       cPin: true,
     },
   });
+  if (audience) reportInput.audience = audience; // else renderEvalReport defaults to "client"
   const dir = reportDir(runId, cwd);
   mkdirSync(dir, { recursive: true });
-  const outPath = join(dir, "report.html");
+  const outPath = join(dir, "evaluation-report.html");
   writeFileSync(outPath, renderEvalReport(reportInput));
   return outPath;
 }

package/.claude/skills/mutagent-evaluator/scripts/sync-eval-criteria.ts

@@ -0,0 +1,244 @@
+/**
+ * scripts/sync-eval-criteria.ts — the EVAL-LEG reconcile CONTRACT (Wave-2 W2I5 · KP-003).
+ * ---------------------------------------------------------------------------
+ * The THIRD leg of the def → impl → eval triad. Today `#sync-spec` reconciles two
+ * legs — spec ↔ impl. When an implementation amends (the ⑤ IMPROVE loop's ai-engineer,
+ * or a brownfield drift), the eval criteria that GROUND the subject's evaluation can go
+ * stale w.r.t. the changed impl. This module is the EVALUATOR-SIDE CONTRACT that the
+ * reconcile (ai-architect REASONING, session-dispatched — Model-B) READS and WRITES:
+ *
+ *   1. WHICH eval leg applies — the subject-kind resolver. The eval leg's shape differs
+ *      by subject kind (this is why W2I5 was gated behind W2I1's code-quality leg):
+ *        - agent / skill / composite subject → the EVAL-SUITE criteria (the evaluator's
+ *          discovered / maintained criteria for that subject).
+ *        - code subject → the CODE-QUALITY criteria (W2I1's `DEFAULT_CODE_QUALITY_CRITERIA`
+ *          / `#mode-judge-code-quality`).
+ *   2. A deterministic STALENESS FLAG predicate — the eval-leg half of the drift trigger
+ *      (mirrors the builder freshness probe's spec/code compare; same 3-value semantics).
+ *   3. The monotonic APPLY primitive — the ai-architect reasons a criteria delta; this
+ *      applies it as criteria MAINTENANCE: upsert-by-id, append the novel, and NEVER drop
+ *      an existing criterion (EV-053, guarded by `assertMonotonicGrowth`).
+ *
+ * JUDGE-ONLY PRESERVED (EV-051): this file NEVER scores a subject and NEVER decides a
+ * subject pass/fail. It FLAGS freshness (pure epoch compare) and GROWS a criteria set
+ * (pure set algebra). The reconcile REASONING — which criteria a changed impl makes stale,
+ * what new criterion it needs — is the ai-architect's, not this code's (Model-B: code =
+ * drift predicates + deterministic maintenance only, never agent dispatch).
+ *
+ * PURE + deterministic (C-PIN): no clock / random / network; same inputs → byte-identical.
+ */
+import { type Static, Type } from "@sinclair/typebox";
+import { Value } from "@sinclair/typebox/value";
+
+import { assertMonotonicGrowth } from "./living-suite.ts";
+
+// ── subject kind → eval leg ──────────────────────────────────────────────────
+
+/** The subject kinds the triad reconciles. `agent|skill|composite` come from the
+ *  agentspec `definition.identity.kind`; `code` is the ⑤ IMPROVE code-target subject. */
+export const EvalSubjectKind = {
+  Agent: "agent",
+  Skill: "skill",
+  Composite: "composite",
+  Code: "code",
+} as const;
+export type EvalSubjectKindValue = (typeof EvalSubjectKind)[keyof typeof EvalSubjectKind];
+
+/** Which eval leg grounds a subject's evaluation — the W2I5 core rule. */
+export const EvalLegKind = {
+  /** agent / skill / composite → the evaluator's discovered/maintained eval-suite criteria. */
+  EvalSuite: "eval-suite",
+  /** code → W2I1's code-quality criteria (`DEFAULT_CODE_QUALITY_CRITERIA`). */
+  CodeQuality: "code-quality",
+} as const;
+export type EvalLegKindValue = (typeof EvalLegKind)[keyof typeof EvalLegKind];
+
+/**
+ * Resolve WHICH eval leg a subject kind reconciles against. Code subjects use the
+ * code-quality criteria (a deterministic-test-gate PLUS a quality judge); every other
+ * kind uses the discovered eval-suite criteria. This single mapping is the shape
+ * differentiator the whole triad turns on.
+ */
+export function evalLegForSubjectKind(kind: EvalSubjectKindValue): EvalLegKindValue {
+  return kind === EvalSubjectKind.Code ? EvalLegKind.CodeQuality : EvalLegKind.EvalSuite;
+}
+
+// ── the staleness flag predicate (eval-leg half of the drift trigger) ────────
+
+/** The eval-leg freshness verdict — mirrors the builder freshness probe's 3-value spec/code semantics. */
+export const EvalLegDrift = {
+  /** no eval-criteria artifact yet — the eval leg must be CONSTRUCTED (cold). */
+  MissingEval: "missing-eval",
+  /** the eval criteria are as fresh as the impl — no reconcile write needed. */
+  InSync: "in-sync",
+  /** the impl is newer than the eval criteria — the eval leg is flagged for RECONCILE. */
+  NeedsSync: "needs-sync",
+} as const;
+export type EvalLegDriftValue = (typeof EvalLegDrift)[keyof typeof EvalLegDrift];
+
+/** Effective freshness epochs (seconds) the flag predicate compares. `null` = unknown/absent. */
+export interface EvalLegFreshnessInput {
+  /** effective freshness epoch of the eval-criteria artifact; `null` = artifact absent. */
+  evalCriteriaEpoch: number | null;
+  /** effective freshness epoch of the implementation; `null` = unknown. */
+  implEpoch: number | null;
+}
+
+/**
+ * FLAG whether the eval leg is stale w.r.t. the impl. Deterministic — a pure epoch
+ * compare, identical in spirit to the builder `check-sync-spec` spec/code compare so the
+ * two legs agree on "impl is newer → reconcile". This is a FLAG only: it decides nothing
+ * about the subject and proposes no criteria — the reconcile reasoning is the agent's.
+ */
+export function flagEvalLegDrift(input: EvalLegFreshnessInput): EvalLegDriftValue {
+  if (input.evalCriteriaEpoch == null) return EvalLegDrift.MissingEval;
+  if (input.implEpoch != null && input.implEpoch > input.evalCriteriaEpoch) {
+    return EvalLegDrift.NeedsSync;
+  }
+  return EvalLegDrift.InSync;
+}
+
+/** The resolved eval leg for a subject: which criteria kind + the current staleness flag. */
+export interface EvalLegAssessment {
+  leg: EvalLegKindValue;
+  drift: EvalLegDriftValue;
+}
+
+/**
+ * Convenience: resolve the eval leg for a subject kind AND flag its staleness in one
+ * call — the exact pair the reconcile reads ("which leg, is it stale?").
+ */
+export function assessEvalLeg(
+  subjectKind: EvalSubjectKindValue,
+  freshness: EvalLegFreshnessInput,
+): EvalLegAssessment {
+  return { leg: evalLegForSubjectKind(subjectKind), drift: flagEvalLegDrift(freshness) };
+}
+
+// ── the criteria representation + the monotonic reconcile-apply ──────────────
+
+/**
+ * One eval criterion — subject-kind-agnostic (a stable `id` key + the binary `statement`,
+ * plus the `leg` it belongs to and an optional severity). Serves BOTH legs: a code-quality
+ * `Qn` criterion (`severity` ∈ critical|high|medium) and an eval-suite discovered criterion
+ * flow through the same shape. `additionalProperties: true` keeps a richer criterion (e.g.
+ * a code-quality `passDefinition`/`failDefinition`, an eval-suite `checkMethod`) forward-parsing.
+ */
+export const EvalCriterionSchema = Type.Object(
+  {
+    id: Type.String({ minLength: 1 }),
+    statement: Type.String({ minLength: 1 }),
+    leg: Type.Union([Type.Literal(EvalLegKind.EvalSuite), Type.Literal(EvalLegKind.CodeQuality)]),
+    severity: Type.Optional(Type.String({ minLength: 1 })),
+  },
+  { additionalProperties: true },
+);
+export type EvalCriterion = Static<typeof EvalCriterionSchema>;
+
+/**
+ * The reconcile REQUEST the ai-architect emits: the persisted criteria `existing`, plus the
+ * agent-reasoned `proposed` additions/updates. `proposed` is an UPSERT set (append the novel,
+ * revise a same-id criterion's wording for the changed impl) — it is NEVER a delete list; the
+ * monotonic-growth invariant (EV-053) forbids dropping a criterion.
+ */
+export const EvalCriteriaReconcileRequestSchema = Type.Object(
+  {
+    subjectId: Type.String({ minLength: 1 }),
+    subjectKind: Type.Union([
+      Type.Literal(EvalSubjectKind.Agent),
+      Type.Literal(EvalSubjectKind.Skill),
+      Type.Literal(EvalSubjectKind.Composite),
+      Type.Literal(EvalSubjectKind.Code),
+    ]),
+    leg: Type.Union([Type.Literal(EvalLegKind.EvalSuite), Type.Literal(EvalLegKind.CodeQuality)]),
+    existing: Type.Array(EvalCriterionSchema),
+    proposed: Type.Array(EvalCriterionSchema),
+  },
+  { additionalProperties: false },
+);
+export type EvalCriteriaReconcileRequest = Static<typeof EvalCriteriaReconcileRequestSchema>;
+
+/** Provenance for a reconcile — pure counters (NO clock, for byte-identity / C-PIN). */
+export interface EvalCriteriaReconcileProvenance {
+  /** how many genuinely-novel criteria were appended. */
+  added: number;
+  /** how many existing criteria (same id) had their content revised. */
+  updated: number;
+  /** total criteria after the reconcile. */
+  total: number;
+}
+
+/** The reconcile RESULT — the grown/maintained criteria set + provenance. */
+export interface EvalCriteriaReconcileResult {
+  subjectId: string;
+  leg: EvalLegKindValue;
+  criteria: EvalCriterion[];
+  provenance: EvalCriteriaReconcileProvenance;
+}
+
+/** Parse + narrow a reconcile request (guarded). THROWS on schema violation. PURE. */
+export function parseEvalCriteriaReconcileRequest(value: unknown): EvalCriteriaReconcileRequest {
+  if (!Value.Check(EvalCriteriaReconcileRequestSchema, value)) {
+    const first = [...Value.Errors(EvalCriteriaReconcileRequestSchema, value)][0];
+    throw new Error(
+      `parseEvalCriteriaReconcileRequest: schema violation at '${first?.path ?? "(root)"}': ` +
+        `${first?.message ?? "invalid eval-criteria reconcile request"}`,
+    );
+  }
+  return value;
+}
+
+const criterionId = (c: EvalCriterion): string => c.id;
+
+/**
+ * APPLY the reconcile: maintain the eval-criteria set for a changed impl. Criteria
+ * MAINTENANCE, not judging (EV-051): upsert each `proposed` criterion by `id` (revise a
+ * same-id criterion's content in place, keeping the id), append the genuinely-novel, and
+ * assert the id-set NEVER shrank (EV-053, via the shared `assertMonotonicGrowth`). The
+ * `leg` on every proposed criterion must match the request's `leg` — mixing a code-quality
+ * criterion into an eval-suite reconcile (or vice versa) is a hard error, not a silent
+ * accept. DETERMINISTIC + PURE: existing order preserved, novel appended in proposal order;
+ * no clock / random. Same (existing, proposed) → byte-identical result.
+ */
+export function reconcileEvalCriteria(
+  request: EvalCriteriaReconcileRequest,
+): EvalCriteriaReconcileResult {
+  const req = parseEvalCriteriaReconcileRequest(request);
+
+  for (const p of req.proposed) {
+    if (p.leg !== req.leg) {
+      throw new Error(
+        `reconcileEvalCriteria: proposed criterion '${p.id}' has leg '${p.leg}' but the ` +
+          `reconcile targets leg '${req.leg}'. An eval-suite and a code-quality criterion ` +
+          "are different legs — refusing to mix them (subject-kind integrity).",
+      );
+    }
+  }
+
+  // Insertion-ordered upsert: existing order first (content possibly revised), novel appended.
+  const byId = new Map<string, EvalCriterion>();
+  for (const c of req.existing) byId.set(c.id, c);
+
+  let added = 0;
+  let updated = 0;
+  for (const p of req.proposed) {
+    if (byId.has(p.id)) {
+      byId.set(p.id, p); // revise content, id (and thus the key/order) retained
+      updated += 1;
+    } else {
+      byId.set(p.id, p); // novel → appended after the existing keys
+      added += 1;
+    }
+  }
+
+  const criteria = [...byId.values()];
+  // EV-053: a maintained criteria set may GROW or revise, but never DROP a criterion.
+  assertMonotonicGrowth(req.existing, criteria, criterionId);
+
+  return {
+    subjectId: req.subjectId,
+    leg: req.leg,
+    criteria,
+    provenance: { added, updated, total: criteria.length },
+  };
+}

package/.claude/skills/mutagent-evaluator/workflows/audit.workflow.js

@@ -47,7 +47,7 @@ phase('Deterministic spine')
 const spine = await agent(
   `Run the DETERMINISTIC spine for the master audit. Use Bash:\n` +
   `  bash ${PKG}/scripts/cli/run.sh ${PKG}/scripts/cli/audit-run.ts ${RUN_ID} --subject ${SUBJECT} --bundle-dir ${BUNDLE_DIR} --subjects-root ${PKG}/subjects --generated-at ${GEN}\n` +
-  `This writes <bundle>/audit/scorecard.json (two-track GATE+TREND) + <bundle>/audit/master-audit.html (the base report). Return the parsed JSON the script prints (subject, runId, criteria, deterministicRows, judgeRows, gateRunPass, scorecard, report).`,
+  `This writes the master-audit under the unified evaluator reports root: .mutagent/evaluator/reports/<runId>/master-audit.html + master-audit.scorecard.json (two-track GATE+TREND). Return the parsed JSON the script prints (subject, runId, criteria, deterministicRows, judgeRows, gateRunPass, scorecard, report).`,
   { label: 'spine:audit-run', phase: 'Deterministic spine' },
 )
 log(`spine: ${spine && spine.criteria} criteria, ${spine && spine.deterministicRows} deterministic + ${spine && spine.judgeRows} judge rows; gateRunPass=${spine && spine.gateRunPass}`)