@tangle-network/agent-eval 0.100.1 → 0.101.0

package/CHANGELOG.md

@@ -4,6 +4,25 @@ All notable changes to `@tangle-network/agent-eval` and its sibling `agent-eval-
 
 ---
 
+## [0.100.3] — 2026-07-01 — product benchmark contract + eval fixture UX
+
+### Added
+
+- Published the `@tangle-network/agent-eval/product-benchmark` subpath so product agents can share one strict product-benchmark manifest, record, artifact, and integrity validator instead of copying Agent Lab or product-local schema code.
+- **Vercel-style eval fixture loading in `/campaign`.** `discoverEvalFixtures`, `loadEvalFixture`, `loadEvalFixtureScenarios`, and `planEvalFixtureRun` let agents use the simple `evals/<name>/PROMPT.md + EVAL.ts + package.json` shape while still executing through the existing `runCampaign` primitive.
+- **Dry-run planning for campaigns.** `planCampaignRun` reports `totalCells`, `cellsCached`, `cellsToRun`, per-cell cache paths, and miss reasons before any agent work starts. This is the cheap proof before spending tokens.
+- **`dispatchRef` on `runCampaign`.** Callers can include the model/tool/prompt/runtime identity in the manifest when the same dispatch function name can run different behavior.
+
+### Fixed
+
+- **Campaign resumability now validates `manifestHash` before reusing a cached cell.** Reusing the same `runDir` after changing scenario payloads, judges, seed/reps, or `dispatchRef` no longer serves stale cells that only match by `cellId`.
+
+### Docs
+
+- Added `docs/eval-fixtures.md` and `examples/eval-fixtures-quickstart/` so agents can add fixture-backed evals without rediscovering the campaign plumbing.
+
+---
+
 ## [0.96.4] — 2026-06-22 — multishot fatal tool errors
 
 ### Added

package/dist/adapters/http.d.ts

@@ -1,4 +1,4 @@
-import { S as Scenario, D as DispatchFn, b as DispatchContext } from '../types-Cv1bo4_a.js';
+import { S as Scenario, D as DispatchFn, b as DispatchContext } from '../types-fWqEJm7h.js';
 import '../run-record-DEwidcqn.js';
 import '@tangle-network/agent-interface';
 import '../errors-CzMUYo7b.js';

package/dist/adapters/langchain.d.ts

@@ -1,4 +1,4 @@
-import { S as Scenario, J as JudgeScore, D as DispatchFn, a as JudgeConfig } from '../types-Cv1bo4_a.js';
+import { S as Scenario, J as JudgeScore, D as DispatchFn, a as JudgeConfig } from '../types-fWqEJm7h.js';
 import '../run-record-DEwidcqn.js';
 import '@tangle-network/agent-interface';
 import '../errors-CzMUYo7b.js';

package/dist/adapters/otel.d.ts

@@ -1,5 +1,5 @@
 import { TraceSpanEvent, HostedClient } from '../hosted/index.js';
-import '../types-Cv1bo4_a.js';
+import '../types-fWqEJm7h.js';
 import '../run-record-DEwidcqn.js';
 import '@tangle-network/agent-interface';
 import '../errors-CzMUYo7b.js';

package/dist/campaign/index.d.ts

@@ -1,11 +1,11 @@
-import { S as SignedManifest, B as BackendIntegrityReport, C as CompletionRequirement, R as RuntimeEventLike, a as CompletionVerdict, P as ProducedState, b as CorrectnessChecker } from '../pre-registration-DB8oDqZJ.js';
-export { L as LlmJudgeDimension, c as LlmJudgeOptions, l as llmJudge } from '../pre-registration-DB8oDqZJ.js';
+import { S as SignedManifest, B as BackendIntegrityReport, C as CompletionRequirement, R as RuntimeEventLike, a as CompletionVerdict, P as ProducedState, b as CorrectnessChecker } from '../pre-registration-BjGZf9YA.js';
+export { L as LlmJudgeDimension, c as LlmJudgeOptions, l as llmJudge } from '../pre-registration-BjGZf9YA.js';
 import { A as AnalyzeTracesOptions, a as AnalyzeTracesInput, b as AnalyzeTracesResult } from '../analyst-C8HHvfJp.js';
-import { S as Scenario, M as MutableSurface, b as DispatchContext, a as JudgeConfig, e as GenerationRecord, g as Gate, J as JudgeScore, L as LabeledScenarioStore, s as LabeledScenarioWrite, t as LabeledScenarioSampleArgs, u as LabeledScenarioRecord, v as LabelTrust, f as SurfaceProposer, w as ProposedCandidate, x as ProposeContext, y as LabeledScenarioSource, C as CampaignResult, o as CodeSurface } from '../types-Cv1bo4_a.js';
-export { k as CampaignAggregates, l as CampaignArtifactWriter, m as CampaignCellResult, n as CampaignCostMeter, z as CampaignTokenUsage, d as CampaignTraceWriter, D as DispatchFn, h as GateContext, j as GateDecision, G as GateResult, p as GenerationCandidate, A as JudgeAggregate, c as JudgeDimension, i as Mutator, O as OptimizationProposer, q as OptimizerConfig, P as ParetoParent, R as RedactionStatus, B as ScenarioAggregate, r as SessionScript, T as TraceSpan, E as isProposedCandidate, F as labelTrustRank } from '../types-Cv1bo4_a.js';
-import { R as RunCampaignOptions, c as RunImprovementLoopOptions, C as CampaignStorage } from '../gepa-BRgNnmGZ.js';
-export { e as GepaProposerConstraints, G as GepaProposerOptions, O as OpenAutoPrOptions, h as OpenAutoPrResult, b as RunImprovementLoopResult, a as RunOptimizationOptions, j as RunOptimizationResult, k as countSentenceEdits, l as defaultRenderDiff, m as extractH2Sections, f as fsCampaignStorage, g as gepaProposer, i as inMemoryCampaignStorage, o as openAutoPr, r as runCampaign, d as runImprovementLoop, n as runOptimization, s as surfaceHash } from '../gepa-BRgNnmGZ.js';
-export { A as AxisEvidence, a as AxisVerdict, B as BuildEvidenceVectorOptions, k as BuildLoopProvenanceArgs, D as DefaultProductionGateOptions, l as EmitLoopProvenanceArgs, m as EmitLoopProvenanceResult, E as EvidenceVector, b as EvolutionaryProposerOptions, H as HeldOutGateOptions, n as LoopProvenanceBackend, o as LoopProvenanceCandidate, L as LoopProvenanceRecord, O as ObjectiveSource, P as ParetoSignificanceGateOptions, c as PromotionObjective, d as PromotionPolicy, R as RunEvalOptions, e as buildEvidenceVector, q as buildLoopProvenanceRecord, f as composeGate, g as defaultProductionGate, s as emitLoopProvenance, h as evolutionaryProposer, i as heldOutGate, t as loopProvenanceSpans, p as paretoPolicy, j as paretoSignificanceGate, u as provenanceRecordPath, v as provenanceSpansPath, r as runEval, w as surfaceContentHash } from '../provenance-B0SZw1z2.js';
+import { S as Scenario, M as MutableSurface, b as DispatchContext, a as JudgeConfig, e as GenerationRecord, g as Gate, J as JudgeScore, L as LabeledScenarioStore, s as LabeledScenarioWrite, t as LabeledScenarioSampleArgs, u as LabeledScenarioRecord, v as LabelTrust, f as SurfaceProposer, w as ProposedCandidate, x as ProposeContext, y as LabeledScenarioSource, C as CampaignResult, o as CodeSurface } from '../types-fWqEJm7h.js';
+export { k as CampaignAggregates, l as CampaignArtifactWriter, m as CampaignCellResult, n as CampaignCostMeter, z as CampaignTokenUsage, d as CampaignTraceWriter, D as DispatchFn, h as GateContext, j as GateDecision, G as GateResult, p as GenerationCandidate, A as JudgeAggregate, c as JudgeDimension, i as Mutator, O as OptimizationProposer, q as OptimizerConfig, P as ParetoParent, R as RedactionStatus, B as ScenarioAggregate, r as SessionScript, T as TraceSpan, E as isProposedCandidate, F as labelTrustRank } from '../types-fWqEJm7h.js';
+import { e as CampaignRunPlan, P as PlanCampaignRunOptions, C as CampaignStorage, R as RunCampaignOptions, c as RunImprovementLoopOptions } from '../gepa-CEy1AIWp.js';
+export { h as CampaignRunPlanCell, j as GepaProposerConstraints, G as GepaProposerOptions, O as OpenAutoPrOptions, k as OpenAutoPrResult, b as RunImprovementLoopResult, a as RunOptimizationOptions, l as RunOptimizationResult, m as countSentenceEdits, n as defaultRenderDiff, o as extractH2Sections, f as fsCampaignStorage, g as gepaProposer, i as inMemoryCampaignStorage, p as openAutoPr, q as planCampaignRun, r as runCampaign, d as runImprovementLoop, s as runOptimization, t as surfaceHash } from '../gepa-CEy1AIWp.js';
+export { A as AxisEvidence, a as AxisVerdict, B as BuildEvidenceVectorOptions, k as BuildLoopProvenanceArgs, D as DefaultProductionGateOptions, l as EmitLoopProvenanceArgs, m as EmitLoopProvenanceResult, E as EvidenceVector, b as EvolutionaryProposerOptions, H as HeldOutGateOptions, n as LoopProvenanceBackend, o as LoopProvenanceCandidate, L as LoopProvenanceRecord, O as ObjectiveSource, P as ParetoSignificanceGateOptions, c as PromotionObjective, d as PromotionPolicy, R as RunEvalOptions, e as buildEvidenceVector, q as buildLoopProvenanceRecord, f as composeGate, g as defaultProductionGate, s as emitLoopProvenance, h as evolutionaryProposer, i as heldOutGate, t as loopProvenanceSpans, p as paretoPolicy, j as paretoSignificanceGate, u as provenanceRecordPath, v as provenanceSpansPath, r as runEval, w as surfaceContentHash } from '../provenance-DdfmVfqR.js';
 import { E as EProcessState, a as PairedBootstrapResult } from '../statistics-xP-cWc5k.js';
 import { L as LlmClientOptions } from '../llm-client-Bj7g0rqu.js';
 import { AgentProfile } from '@tangle-network/agent-interface';
@@ -139,6 +139,56 @@ interface FailureModeRecallJudgeOptions {
  */
 declare function failureModeRecallJudge(opts?: FailureModeRecallJudgeOptions): JudgeConfig<AnalystArtifact, AnalystScenario>;
 
+type EvalFixtureValidationMode = 'vitest' | 'none';
+interface EvalFixtureFile {
+    path: string;
+    sha256: string;
+    bytes: number;
+}
+interface EvalFixture {
+    name: string;
+    path: string;
+    promptPath: string;
+    evalPath?: string;
+    packageJsonPath?: string;
+    prompt: string;
+    files: EvalFixtureFile[];
+    fingerprint: string;
+}
+interface EvalFixtureScenario extends Scenario {
+    kind: 'eval-fixture';
+    fixtureName: string;
+    fixturePath: string;
+    promptPath: string;
+    evalPath?: string;
+    packageJsonPath?: string;
+    prompt: string;
+    fingerprint: string;
+}
+interface EvalFixtureLoadOptions {
+    /** `vitest` requires EVAL.ts/EVAL.tsx and package.json type=module. `none` only requires PROMPT.md. */
+    validation?: EvalFixtureValidationMode;
+    /** Extra caller-owned knobs that affect fixture behavior, folded into the fingerprint. */
+    fingerprintConfig?: unknown;
+}
+interface LoadEvalFixtureScenariosOptions extends EvalFixtureLoadOptions {
+    names?: string[];
+}
+interface PlanEvalFixtureRunOptions<TArtifact = unknown> extends Pick<PlanCampaignRunOptions<EvalFixtureScenario, TArtifact>, 'dispatchRef' | 'judges' | 'seed' | 'reps' | 'resumable' | 'runDir'> {
+    evalsDir: string;
+    validation?: EvalFixtureValidationMode;
+    fingerprintConfig?: unknown;
+    names?: string[];
+    storage?: CampaignStorage;
+}
+type EvalFixtureRunPlan = CampaignRunPlan & {
+    fixtures: Array<Pick<EvalFixtureScenario, 'fixtureName' | 'fixturePath' | 'fingerprint'>>;
+};
+declare function discoverEvalFixtures(evalsDir: string): string[];
+declare function loadEvalFixture(evalsDir: string, name: string, options?: EvalFixtureLoadOptions): EvalFixture;
+declare function loadEvalFixtureScenarios(evalsDir: string, options?: LoadEvalFixtureScenariosOptions): EvalFixtureScenario[];
+declare function planEvalFixtureRun<TArtifact = unknown>(options: PlanEvalFixtureRunOptions<TArtifact>): EvalFixtureRunPlan;
+
 /**
  * Anytime-valid sequential promotion gate — an e-process (betting
  * test-martingale, see `eProcess` in `statistics.ts`) over paired
@@ -1475,4 +1525,4 @@ declare function gitWorktreeAdapter(opts: GitWorktreeAdapterOptions): WorktreeAd
  *  as a ref under the adapter's worktree dir. */
 declare function resolveWorktreePath(surface: CodeSurface, worktreeDir?: string): string;
 
-export { type AcceptedEdit, type AceProposerOptions, type AnalystArtifact, type AnalystScenario, type ApplySkillPatchResult, type BuildAnalystSurfaceDispatchOptions, type CampaignBreakdown, CampaignResult, CampaignStorage, CodeSurface, type CompareProposersOptions, type DimensionRegression, DispatchContext, type FailureModeRecallJudgeOptions, type FapoAttributionSignals, type FapoEntryConfig, type FapoFailureCluster, type FapoOptimizationLevel, type FapoProposerOptions, type FapoReviewInput, type FapoReviewIssue, type FapoReviewResult, type FapoScopeContract, FsLabeledScenarioStore, type FsLabeledScenarioStoreOptions, Gate, GenerationRecord, type GitWorktreeAdapterOptions, type HaloProposerOptions, type HeldoutSignificance, type HeldoutSignificanceOptions, type JsonPrimitive, type JsonValue, JudgeConfig, JudgeScore, LabelTrust, LabeledScenarioRecord, LabeledScenarioSampleArgs, LabeledScenarioSource, LabeledScenarioStore, LabeledScenarioStoreError, LabeledScenarioWrite, type MemoryCurationProposerOptions, MutableSurface, type OptimizerEntryConfig, type PairedHoldout, type ParameterCandidate, type ParameterChange, type ParameterSweepProposerOptions, type PlaybackContext, type PlaybackDriver, type PlaybackStep, type PolicyEditProposerOptions, type ProfileDispatchFn, ProfileMatrixError, type ProfileSummary, ProposeContext, type ProposePatchesArgs, ProposedCandidate, type ProposerComparison, type ProposerEntry, type ProposerPairwise, type ProposerScore, type RejectedEdit, RunCampaignOptions, RunImprovementLoopOptions, type RunProfileMatrixOptions, type RunProfileMatrixResult, type RunSkillOptOptions, type RunSkillOptResult, Scenario, type ScenarioRollup, type ScoreboardRenderOptions, type ScoreboardRow, type ScoreboardSummary, type SequentialDecideFn, type SequentialDecideOptions, type SequentialDecision, type SequentialObservation, type SequentialPairedGate, type SequentialPairedGateOptions, type SkillOptEpochRecord, type SkillOptEvidence, type SkillOptProposer, type SkillOptProposerOptions, type SkillPatch, type SkillPatchOp, SkillPatchParseError, type SkillPatchRejection, SurfaceProposer, type TraceAnalystProposerOptions, type UserStory, type UserStoryVerdict, type Worktree, type WorktreeAdapter, WorktreeAdapterError, aceProposer, applySkillPatch, buildAnalystSurfaceDispatch, campaignBreakdown, campaignMeanComposite, compareProposers, detectScale, dimensionRegressions, extractFapoAttributionSignals, failureModeRecallJudge, fapoEscalationEntry, fapoProposer, gepaParetoEntry, gepaReflectionEntry, gitWorktreeAdapter, haloProposer, heldoutSignificance, makePlaybackDispatch, memoryCurationProposer, pairHoldout, parameterSweepProposer, parseSkillPatchResponse, patchEditCount, policyEditProposer, renderScoreboardMarkdown, resolveWorktreePath, runProfileMatrix, runSkillOpt, scoreUserStory, scoreboardSummary, sequentialDecide, sequentialPairedGate, skillOptEntry, skillOptProposer, traceAnalystProposer, userStoryScoreboard };
+export { type AcceptedEdit, type AceProposerOptions, type AnalystArtifact, type AnalystScenario, type ApplySkillPatchResult, type BuildAnalystSurfaceDispatchOptions, type CampaignBreakdown, CampaignResult, CampaignRunPlan, CampaignStorage, CodeSurface, type CompareProposersOptions, type DimensionRegression, DispatchContext, type EvalFixture, type EvalFixtureFile, type EvalFixtureLoadOptions, type EvalFixtureRunPlan, type EvalFixtureScenario, type EvalFixtureValidationMode, type FailureModeRecallJudgeOptions, type FapoAttributionSignals, type FapoEntryConfig, type FapoFailureCluster, type FapoOptimizationLevel, type FapoProposerOptions, type FapoReviewInput, type FapoReviewIssue, type FapoReviewResult, type FapoScopeContract, FsLabeledScenarioStore, type FsLabeledScenarioStoreOptions, Gate, GenerationRecord, type GitWorktreeAdapterOptions, type HaloProposerOptions, type HeldoutSignificance, type HeldoutSignificanceOptions, type JsonPrimitive, type JsonValue, JudgeConfig, JudgeScore, LabelTrust, LabeledScenarioRecord, LabeledScenarioSampleArgs, LabeledScenarioSource, LabeledScenarioStore, LabeledScenarioStoreError, LabeledScenarioWrite, type LoadEvalFixtureScenariosOptions, type MemoryCurationProposerOptions, MutableSurface, type OptimizerEntryConfig, type PairedHoldout, type ParameterCandidate, type ParameterChange, type ParameterSweepProposerOptions, PlanCampaignRunOptions, type PlanEvalFixtureRunOptions, type PlaybackContext, type PlaybackDriver, type PlaybackStep, type PolicyEditProposerOptions, type ProfileDispatchFn, ProfileMatrixError, type ProfileSummary, ProposeContext, type ProposePatchesArgs, ProposedCandidate, type ProposerComparison, type ProposerEntry, type ProposerPairwise, type ProposerScore, type RejectedEdit, RunCampaignOptions, RunImprovementLoopOptions, type RunProfileMatrixOptions, type RunProfileMatrixResult, type RunSkillOptOptions, type RunSkillOptResult, Scenario, type ScenarioRollup, type ScoreboardRenderOptions, type ScoreboardRow, type ScoreboardSummary, type SequentialDecideFn, type SequentialDecideOptions, type SequentialDecision, type SequentialObservation, type SequentialPairedGate, type SequentialPairedGateOptions, type SkillOptEpochRecord, type SkillOptEvidence, type SkillOptProposer, type SkillOptProposerOptions, type SkillPatch, type SkillPatchOp, SkillPatchParseError, type SkillPatchRejection, SurfaceProposer, type TraceAnalystProposerOptions, type UserStory, type UserStoryVerdict, type Worktree, type WorktreeAdapter, WorktreeAdapterError, aceProposer, applySkillPatch, buildAnalystSurfaceDispatch, campaignBreakdown, campaignMeanComposite, compareProposers, detectScale, dimensionRegressions, discoverEvalFixtures, extractFapoAttributionSignals, failureModeRecallJudge, fapoEscalationEntry, fapoProposer, gepaParetoEntry, gepaReflectionEntry, gitWorktreeAdapter, haloProposer, heldoutSignificance, loadEvalFixture, loadEvalFixtureScenarios, makePlaybackDispatch, memoryCurationProposer, pairHoldout, parameterSweepProposer, parseSkillPatchResponse, patchEditCount, planEvalFixtureRun, policyEditProposer, renderScoreboardMarkdown, resolveWorktreePath, runProfileMatrix, runSkillOpt, scoreUserStory, scoreboardSummary, sequentialDecide, sequentialPairedGate, skillOptEntry, skillOptProposer, traceAnalystProposer, userStoryScoreboard };

package/dist/campaign/index.js

@@ -10,15 +10,16 @@ import {
   paretoPolicy,
   paretoSignificanceGate,
   runEval
-} from "../chunk-G7IB3GJ5.js";
+} from "../chunk-CHIFZIQD.js";
 import {
   agentProfileHash,
   agentProfileId,
   agentProfileModelId,
   extractProducedState,
+  harnessAxisOf,
   llmJudge,
   verifyCompletion
-} from "../chunk-VWQ6PO5O.js";
+} from "../chunk-GUII3E73.js";
 import {
   buildLoopProvenanceRecord,
   campaignBreakdown,
@@ -40,14 +41,16 @@ import {
   runOptimization,
   surfaceContentHash,
   surfaceHash
-} from "../chunk-2KTBHICD.js";
+} from "../chunk-NK77GPUH.js";
 import {
   assertRealBackend,
+  contentHash,
   fsCampaignStorage,
   inMemoryCampaignStorage,
+  planCampaignRun,
   runCampaign,
   summarizeBackendIntegrity
-} from "../chunk-HRGTA6U5.js";
+} from "../chunk-XIOQHCHU.js";
 import {
   estimateCost,
   isModelPriced
@@ -84,7 +87,9 @@ import "../chunk-PC4UYEBM.js";
 import {
   validateRunRecord
 } from "../chunk-G6S73VA7.js";
-import "../chunk-ABOIVNXL.js";
+import {
+  buildAgentProfileCell
+} from "../chunk-ABOIVNXL.js";
 import {
   canonicalize
 } from "../chunk-VSMTAMNK.js";
@@ -155,17 +160,171 @@ function failureModeRecallJudge(opts = {}) {
   };
 }
 
-// src/campaign/gates/sequential.ts
+// src/campaign/fixtures.ts
 import { createHash } from "crypto";
+import { existsSync, readdirSync, readFileSync, statSync } from "fs";
+import { isAbsolute, join, relative, resolve } from "path";
+function discoverEvalFixtures(evalsDir) {
+  const root = resolve(evalsDir);
+  if (!existsSync(root)) throw new Error(`discoverEvalFixtures: evalsDir not found: ${root}`);
+  const fixtures = [];
+  const walk = (dir, base = "") => {
+    for (const entry of readdirSync(dir).sort()) {
+      if (entry.startsWith(".") || entry === "node_modules") continue;
+      const fullPath = join(dir, entry);
+      if (!statSync(fullPath).isDirectory()) continue;
+      const name = base ? `${base}/${entry}` : entry;
+      if (existsWithExactCase(fullPath, "PROMPT.md")) {
+        fixtures.push(name);
+      } else {
+        walk(fullPath, name);
+      }
+    }
+  };
+  walk(root);
+  return fixtures;
+}
+function loadEvalFixture(evalsDir, name, options = {}) {
+  const validation = options.validation ?? "vitest";
+  const root = resolve(evalsDir);
+  const fixturePath = resolve(root, name);
+  assertInside(root, fixturePath, name);
+  if (!existsSync(fixturePath) || !statSync(fixturePath).isDirectory()) {
+    throw new Error(`loadEvalFixture: fixture not found: ${name}`);
+  }
+  if (!existsWithExactCase(fixturePath, "PROMPT.md")) {
+    throw new Error(`loadEvalFixture: ${name} is missing exact-case PROMPT.md`);
+  }
+  const promptPath = join(fixturePath, "PROMPT.md");
+  const evalPath = resolveEvalPath(fixturePath);
+  const packageJsonPath = existsWithExactCase(fixturePath, "package.json") ? join(fixturePath, "package.json") : void 0;
+  if (validation !== "none") {
+    if (!evalPath)
+      throw new Error(`loadEvalFixture: ${name} is missing exact-case EVAL.ts or EVAL.tsx`);
+    if (!packageJsonPath)
+      throw new Error(`loadEvalFixture: ${name} is missing exact-case package.json`);
+    assertModulePackage(packageJsonPath, name);
+  }
+  const files = collectFixtureFiles(fixturePath);
+  return {
+    name,
+    path: fixturePath,
+    promptPath,
+    evalPath,
+    packageJsonPath,
+    prompt: readFileSync(promptPath, "utf8"),
+    files,
+    fingerprint: contentHash({
+      files,
+      config: options.fingerprintConfig ?? null
+    })
+  };
+}
+function loadEvalFixtureScenarios(evalsDir, options = {}) {
+  const names = options.names ?? discoverEvalFixtures(evalsDir);
+  return names.map((name) => {
+    const fixture = loadEvalFixture(evalsDir, name, options);
+    return {
+      id: fixture.name,
+      kind: "eval-fixture",
+      tags: ["eval-fixture"],
+      fixtureName: fixture.name,
+      fixturePath: fixture.path,
+      promptPath: fixture.promptPath,
+      evalPath: fixture.evalPath,
+      packageJsonPath: fixture.packageJsonPath,
+      prompt: fixture.prompt,
+      fingerprint: fixture.fingerprint
+    };
+  });
+}
+function planEvalFixtureRun(options) {
+  const scenarios = loadEvalFixtureScenarios(options.evalsDir, {
+    names: options.names,
+    validation: options.validation,
+    fingerprintConfig: options.fingerprintConfig
+  });
+  const plan = planCampaignRun({
+    scenarios,
+    dispatchRef: options.dispatchRef ?? "eval-fixture-dispatch",
+    judges: options.judges,
+    seed: options.seed,
+    reps: options.reps,
+    resumable: options.resumable,
+    runDir: options.runDir,
+    storage: options.storage
+  });
+  return {
+    ...plan,
+    fixtures: scenarios.map((scenario) => ({
+      fixtureName: scenario.fixtureName,
+      fixturePath: scenario.fixturePath,
+      fingerprint: scenario.fingerprint
+    }))
+  };
+}
+function existsWithExactCase(dirPath, fileName) {
+  try {
+    return readdirSync(dirPath).includes(fileName);
+  } catch {
+    return false;
+  }
+}
+function resolveEvalPath(fixturePath) {
+  if (existsWithExactCase(fixturePath, "EVAL.ts")) return join(fixturePath, "EVAL.ts");
+  if (existsWithExactCase(fixturePath, "EVAL.tsx")) return join(fixturePath, "EVAL.tsx");
+  return void 0;
+}
+function assertModulePackage(packageJsonPath, name) {
+  let parsed;
+  try {
+    parsed = JSON.parse(readFileSync(packageJsonPath, "utf8"));
+  } catch (err) {
+    throw new Error(
+      `loadEvalFixture: ${name} package.json is invalid JSON: ${err instanceof Error ? err.message : String(err)}`
+    );
+  }
+  if (typeof parsed !== "object" || parsed === null || parsed.type !== "module") {
+    throw new Error(`loadEvalFixture: ${name} package.json must set "type": "module"`);
+  }
+}
+function collectFixtureFiles(fixturePath, base = "") {
+  const files = [];
+  for (const entry of readdirSync(join(fixturePath, base)).sort()) {
+    if (entry === "node_modules" || entry === ".git") continue;
+    const relativePath = base ? `${base}/${entry}` : entry;
+    const fullPath = join(fixturePath, relativePath);
+    const stat = statSync(fullPath);
+    if (stat.isDirectory()) {
+      files.push(...collectFixtureFiles(fixturePath, relativePath));
+      continue;
+    }
+    const bytes = readFileSync(fullPath);
+    files.push({
+      path: relativePath,
+      sha256: createHash("sha256").update(bytes).digest("hex"),
+      bytes: bytes.byteLength
+    });
+  }
+  return files;
+}
+function assertInside(root, target, label) {
+  const rel = relative(root, target);
+  if (rel === "" || !rel.startsWith("..") && !isAbsolute(rel)) return;
+  throw new Error(`loadEvalFixture: fixture path escapes evalsDir: ${label}`);
+}
+
+// src/campaign/gates/sequential.ts
+import { createHash as createHash2 } from "crypto";
 function verifyManifestSync(m) {
   if (m.algo !== void 0 && m.algo !== "sha256-content") {
     throw new Error(`sequentialPairedGate: unrecognized manifest hash algo '${m.algo}'`);
   }
-  const { contentHash, algo: _algo, ...rest } = m;
+  const { contentHash: contentHash2, algo: _algo, ...rest } = m;
   void _algo;
   const bytes = JSON.stringify(canonicalize(rest));
-  const hash = createHash("sha256").update(bytes, "utf8").digest("hex");
-  return hash === contentHash;
+  const hash = createHash2("sha256").update(bytes, "utf8").digest("hex");
+  return hash === contentHash2;
 }
 function resolveConfig(opts) {
   const m = opts.preRegistration;
@@ -422,9 +581,9 @@ function sequentialDecide(options = {}) {
 }
 
 // src/campaign/labeled-store/fs-adapter.ts
-import { createHash as createHash2 } from "crypto";
-import { existsSync, mkdirSync, readFileSync, writeFileSync } from "fs";
-import { join } from "path";
+import { createHash as createHash3 } from "crypto";
+import { existsSync as existsSync2, mkdirSync, readFileSync as readFileSync2, writeFileSync } from "fs";
+import { join as join2 } from "path";
 var LabeledScenarioStoreError = class extends Error {
   constructor(code, message) {
     super(message);
@@ -436,7 +595,7 @@ var LabeledScenarioStoreError = class extends Error {
 var FsLabeledScenarioStore = class {
   constructor(options) {
     this.options = options;
-    if (!existsSync(options.root)) mkdirSync(options.root, { recursive: true });
+    if (!existsSync2(options.root)) mkdirSync(options.root, { recursive: true });
     this.now = options.now ?? Date.now;
   }
   options;
@@ -471,8 +630,8 @@ var FsLabeledScenarioStore = class {
         if (!explicit) continue;
       }
       const path = this.pathForSource(source);
-      if (!existsSync(path)) continue;
-      const lines = readFileSync(path, "utf8").split("\n").filter(Boolean);
+      if (!existsSync2(path)) continue;
+      const lines = readFileSync2(path, "utf8").split("\n").filter(Boolean);
       for (const line of lines) {
         let record;
         try {
@@ -500,11 +659,11 @@ var FsLabeledScenarioStore = class {
     let total = 0;
     for (const source of ALL_SOURCES) {
       const path = this.pathForSource(source);
-      if (!existsSync(path)) {
+      if (!existsSync2(path)) {
         bySource[source] = 0;
         continue;
       }
-      const lines = readFileSync(path, "utf8").split("\n").filter(Boolean);
+      const lines = readFileSync2(path, "utf8").split("\n").filter(Boolean);
       bySource[source] = lines.length;
       total += lines.length;
       for (const line of lines) {
@@ -584,7 +743,7 @@ var FsLabeledScenarioStore = class {
     };
   }
   pathForSource(source) {
-    return join(this.options.root, `${source}.jsonl`);
+    return join2(this.options.root, `${source}.jsonl`);
   }
 };
 var ALL_SOURCES = [
@@ -621,11 +780,11 @@ function matchesFilter(record, args, source) {
   return true;
 }
 function sha256(input) {
-  return createHash2("sha256").update(input).digest("hex").slice(0, 16);
+  return createHash3("sha256").update(input).digest("hex").slice(0, 16);
 }
 function appendLine(path, line) {
-  if (existsSync(path)) {
-    const existing = readFileSync(path, "utf8");
+  if (existsSync2(path)) {
+    const existing = readFileSync2(path, "utf8");
     writeFileSync(path, existing + line);
   } else {
     writeFileSync(path, line);
@@ -1850,8 +2009,8 @@ function renderScoreboardMarkdown(rows, opts = {}) {
 }
 
 // src/campaign/presets/run-profile-matrix.ts
-import { createHash as createHash3 } from "crypto";
-import { join as join2 } from "path";
+import { createHash as createHash4 } from "crypto";
+import { join as join3 } from "path";
 var ProfileMatrixError = class extends AgentEvalError {
   constructor(message) {
     super("profile_matrix", message);
@@ -1861,7 +2020,7 @@ function sanitize(id) {
   return id.replace(/[^a-zA-Z0-9_-]/g, "_");
 }
 function sha(input) {
-  return createHash3("sha256").update(JSON.stringify(input)).digest("hex");
+  return createHash4("sha256").update(JSON.stringify(input)).digest("hex");
 }
 function mean2(xs) {
   return xs.length === 0 ? 0 : xs.reduce((a, b) => a + b, 0) / xs.length;
@@ -1930,6 +2089,7 @@ function buildRunRecord(args) {
     outcome,
     splitTag,
     scenarioId: cell.scenarioId,
+    ...args.agentProfileCell ? { agentProfile: args.agentProfileCell } : {},
     ...cell.error ? { failureMode: cell.error } : {}
   };
   if (args.corpusText && args.scenario) {
@@ -2008,7 +2168,14 @@ async function runProfileMatrix(opts) {
       captureSource: opts.captureSource,
       storage: opts.storage,
       now: opts.now,
-      runDir: join2(opts.runDir, sanitize(profileId))
+      runDir: join3(opts.runDir, sanitize(profileId))
+    });
+    const axis = harnessAxisOf(profile);
+    const agentProfileCell = await buildAgentProfileCell({
+      profileId,
+      sourceProfile: { kind: "agent-interface-profile", hash: profileHash },
+      model,
+      ...axis ? { harness: { id: axis.harness } } : {}
     });
     const profileRecords = [];
     for (const cell of campaign.cells) {
@@ -2021,6 +2188,7 @@ async function runProfileMatrix(opts) {
         splitTag,
         commitSha: opts.commitSha,
         matrixId,
+        agentProfileCell,
         scenario: scenarioById.get(cell.scenarioId),
         corpusText: opts.corpusText
       });
@@ -2175,7 +2343,7 @@ import { promisify } from "util";
 // src/campaign/proposers/analysis-edit.ts
 import { mkdtempSync, writeFileSync as writeFileSync2 } from "fs";
 import { tmpdir } from "os";
-import { join as join3 } from "path";
+import { join as join4 } from "path";
 var APPLY_SYSTEM = "You apply a trace-analysis report to an agent instruction prompt. Output ONLY the full revised prompt \u2014 no preamble, no commentary, no code fences. Make the minimal edits that address the report findings; preserve everything else verbatim.";
 function surfaceToPromptText(surface) {
   return typeof surface === "string" ? surface : JSON.stringify(surface);
@@ -2187,8 +2355,8 @@ function analysisEditProposer(opts) {
       const parent = surfaceToPromptText(ctx.currentSurface);
       const traces = await opts.resolveTraces(ctx) ?? "";
       if (!traces.trim()) throw new Error(opts.noTracesError);
-      const dir = mkdtempSync(join3(tmpdir(), `${opts.kind}-proposer-`));
-      const tracePath = join3(dir, "traces.jsonl");
+      const dir = mkdtempSync(join4(tmpdir(), `${opts.kind}-proposer-`));
+      const tracePath = join4(dir, "traces.jsonl");
       writeFileSync2(tracePath, traces.endsWith("\n") ? traces : `${traces}
 `);
       const report = await opts.analyze(tracePath, ctx);
@@ -2473,8 +2641,8 @@ ${report.slice(0, 800)}`,
 
 // src/campaign/worktree/index.ts
 import { execFileSync } from "child_process";
-import { existsSync as existsSync2 } from "fs";
-import { basename, isAbsolute, join as join4 } from "path";
+import { existsSync as existsSync3 } from "fs";
+import { basename, isAbsolute as isAbsolute2, join as join5 } from "path";
 var WorktreeAdapterError = class extends Error {
   constructor(message, cause) {
     super(message);
@@ -2496,13 +2664,13 @@ function slug2(label) {
 }
 function gitWorktreeAdapter(opts) {
   const git = opts.git ?? defaultGit;
-  const worktreeDir = opts.worktreeDir ?? join4(opts.repoRoot, ".worktrees");
+  const worktreeDir = opts.worktreeDir ?? join5(opts.repoRoot, ".worktrees");
   const branchPrefix = opts.branchPrefix ?? "improve";
   return {
     async create({ baseRef, label }) {
       const id = `${slug2(label)}-${Date.now().toString(36)}-${Math.random().toString(36).slice(2, 6)}`;
       const branch = `${branchPrefix}/${id}`;
-      const path = join4(worktreeDir, id);
+      const path = join5(worktreeDir, id);
       git(["worktree", "add", "-b", branch, path, baseRef], opts.repoRoot);
       return { path, branch, baseRef };
     },
@@ -2526,8 +2694,8 @@ function gitWorktreeAdapter(opts) {
   };
 }
 function resolveWorktreePath(surface, worktreeDir) {
-  if (isAbsolute(surface.worktreeRef) && existsSync2(surface.worktreeRef)) return surface.worktreeRef;
-  if (worktreeDir) return join4(worktreeDir, basename(surface.worktreeRef));
+  if (isAbsolute2(surface.worktreeRef) && existsSync3(surface.worktreeRef)) return surface.worktreeRef;
+  if (worktreeDir) return join5(worktreeDir, basename(surface.worktreeRef));
   return surface.worktreeRef;
 }
 export {
@@ -2550,6 +2718,7 @@ export {
   defaultRenderDiff,
   detectScale,
   dimensionRegressions,
+  discoverEvalFixtures,
   emitLoopProvenance,
   evolutionaryProposer,
   extractFapoAttributionSignals,
@@ -2569,6 +2738,8 @@ export {
   isProposedCandidate,
   labelTrustRank,
   llmJudge,
+  loadEvalFixture,
+  loadEvalFixtureScenarios,
   loopProvenanceSpans,
   makePlaybackDispatch,
   memoryCurationProposer,
@@ -2579,6 +2750,8 @@ export {
   paretoSignificanceGate,
   parseSkillPatchResponse,
   patchEditCount,
+  planCampaignRun,
+  planEvalFixtureRun,
   policyEditProposer,
   provenanceRecordPath,
   provenanceSpansPath,