@tangle-network/agent-eval 0.94.0 → 0.95.1

package/dist/{chunk-EGPMSBEZ.js → chunk-7QTQKIDD.js}

@@ -1,7 +1,7 @@
 import {
   runCampaign,
   summarizeBackendIntegrity
-} from "./chunk-2K6UUZ7P.js";
+} from "./chunk-2T4EZACH.js";
 import {
   callLlm
 } from "./chunk-CWNP4DV4.js";
@@ -868,173 +868,6 @@ function parseReflectionResponse(raw, maxProposals) {
   return out;
 }
 
-// src/campaign/drivers/gepa.ts
-var REFLECTION_SYSTEM = 'You are an expert prompt engineer performing GEPA-style reflective mutation. You are given a prompt surface, its top trials (preserve what works) and its bottom trials (the evidence to fix). For each proposal, reason in this order before writing the payload: (1) LOCALIZE \u2014 point to the exact span of the current surface responsible for a bottom-trial failure; (2) DIAGNOSE the root cause (a missing rule, an ambiguous instruction, an over-broad directive), not just the symptom; (3) propose the MINIMAL, GENERALIZABLE edit that fixes the whole failure class \u2014 state it as a rule the agent should follow, never a patch memorized to the shown trials (that is overfitting and will not transfer to the held-out set); (4) PRESERVE every instruction the top trials depend on \u2014 do not delete or weaken working guidance. Put this localize\u2192diagnose\u2192fix reasoning in each proposal\'s `rationale`. Output ONLY a JSON object of shape {"proposals":[{"label":string,"rationale":string,"payload":string}]} where each `payload` is the FULL improved surface text. No prose outside the JSON.';
-var COMBINE_SYSTEM = 'You are an expert prompt engineer performing a GEPA "combine complementary lessons" merge. You are given several non-dominated versions of one surface; each is uniquely best on different scenarios. Produce ONE new version that keeps what makes each version strong on its winning scenarios and resolves conflicts in favor of the more general rule. Output ONLY a JSON object of shape {"proposals":[{"label":string,"rationale":string,"payload":string}]} with exactly one proposal whose `payload` is the FULL merged surface text. No prose outside the JSON.';
-function gepaDriver(opts) {
-  const evidenceK = opts.evidenceK ?? 3;
-  const combineParents = opts.combineParents ?? true;
-  const combineMaxParents = opts.combineMaxParents ?? 4;
-  if (combineParents && combineMaxParents < 1) {
-    throw new Error("gepaDriver: combineMaxParents must be >= 1 when combineParents is enabled");
-  }
-  return {
-    kind: "gepa",
-    async propose(ctx) {
-      const parent = typeof ctx.currentSurface === "string" ? ctx.currentSurface : JSON.stringify(ctx.currentSurface);
-      const constraints = opts.constraints;
-      const preserveSections = constraints?.preserveSections !== void 0 ? constraints.preserveSections.length === 0 ? extractH2Sections(parent) : constraints.preserveSections : null;
-      const maxEdits = constraints?.maxSentenceEdits;
-      const out = [];
-      const seen = /* @__PURE__ */ new Set();
-      const accept = (payload, label, rationale) => {
-        const text = typeof payload === "string" ? payload.trim() : "";
-        if (!text || text === parent || seen.has(text)) return;
-        if (preserveSections && !validatePreservedSections(text, preserveSections)) return;
-        if (maxEdits !== void 0 && countSentenceEdits(parent, text) > maxEdits * 2) return;
-        seen.add(text);
-        out.push({ surface: text, label, rationale });
-      };
-      const stringParents = (combineParents ? ctx.paretoParents ?? [] : []).filter((p) => typeof p.surface === "string").sort((a, b) => b.composite - a.composite).slice(0, combineMaxParents);
-      if (stringParents.length > 1) {
-        const combinePrompt = buildCombinePrompt({
-          target: opts.target,
-          parents: stringParents,
-          evidenceK
-        });
-        const combineResult = await callLlm(
-          {
-            model: opts.model,
-            messages: [
-              { role: "system", content: COMBINE_SYSTEM },
-              { role: "user", content: combinePrompt }
-            ],
-            jsonMode: true,
-            temperature: opts.temperature ?? 0.7,
-            maxTokens: opts.maxTokens ?? 6e3
-          },
-          opts.llm
-        );
-        const merged = parseReflectionResponse(combineResult.content, 1)[0];
-        if (merged) {
-          accept(
-            merged.payload,
-            merged.label || "pareto-combine",
-            merged.rationale || `combined ${stringParents.length} non-dominated parents (gen ${stringParents.map((p) => p.generation).join(",")})`
-          );
-        }
-      }
-      const reflectCount = Math.max(0, ctx.populationSize - out.length);
-      if (reflectCount > 0) {
-        const { top, bottom, target } = buildEvidence(ctx, evidenceK, opts.target);
-        const userPrompt = buildReflectionPrompt({
-          target,
-          parentPayload: parent,
-          topTrials: top,
-          bottomTrials: bottom,
-          childCount: reflectCount,
-          mutationPrimitives: opts.mutationPrimitives
-        });
-        const analyst = renderAnalystEvidence(ctx.findings, ctx.report);
-        const finalPrompt = analyst ? `${userPrompt}
-
-${analyst}` : userPrompt;
-        const result = await callLlm(
-          {
-            model: opts.model,
-            messages: [
-              { role: "system", content: REFLECTION_SYSTEM },
-              { role: "user", content: finalPrompt }
-            ],
-            jsonMode: true,
-            temperature: opts.temperature ?? 0.7,
-            maxTokens: opts.maxTokens ?? 6e3
-          },
-          opts.llm
-        );
-        for (const proposal of parseReflectionResponse(result.content, reflectCount)) {
-          accept(proposal.payload, proposal.label, proposal.rationale);
-        }
-      }
-      return out.slice(0, ctx.populationSize);
-    }
-  };
-}
-function buildCombinePrompt(args) {
-  const lines = [
-    `You are merging ${args.parents.length} versions of: ${args.target}.`,
-    "",
-    "Each version is on the Pareto frontier \u2014 none dominates the others; each",
-    "wins on different scenarios. Combine their complementary strengths into",
-    "ONE version. Below, each version lists the scenarios it scores highest on.",
-    ""
-  ];
-  args.parents.forEach((p, i) => {
-    const tag = String.fromCharCode(65 + i);
-    const best = Object.entries(p.objectives).sort((a, b) => b[1] - a[1]).slice(0, args.evidenceK).map(([id, score]) => `${id} (${score.toFixed(2)})`);
-    lines.push(
-      `### Version ${tag} (mean ${p.composite.toFixed(2)}; strongest on: ${best.join(", ") || "n/a"})`,
-      "```",
-      p.surface,
-      "```",
-      ""
-    );
-  });
-  lines.push(
-    "Return ONE merged version that would score well on the union of every",
-    "version's winning scenarios. Keep each version's specific winning rule;",
-    "where two rules conflict, prefer the more general one and note the choice",
-    "in your rationale."
-  );
-  return lines.join("\n");
-}
-function extractH2Sections(text) {
-  const out = [];
-  for (const line of text.split("\n")) {
-    const match = /^##\s+(.+?)\s*$/.exec(line);
-    if (match) out.push(match[1]);
-  }
-  return out;
-}
-function countSentenceEdits(baseline, candidate) {
-  const norm = (s) => s.split(/(?<=[.!?])\s+|\n/g).map((p) => p.trim()).filter((p) => p.length > 0);
-  const a = new Set(norm(baseline));
-  const b = new Set(norm(candidate));
-  let edits = 0;
-  for (const s of a) if (!b.has(s)) edits++;
-  for (const s of b) if (!a.has(s)) edits++;
-  return edits;
-}
-function validatePreservedSections(candidate, required) {
-  if (required.length === 0) return true;
-  const have = new Set(extractH2Sections(candidate));
-  for (const section of required) {
-    if (!have.has(section)) return false;
-  }
-  return true;
-}
-function buildEvidence(ctx, evidenceK, baseTarget) {
-  const last = ctx.history.at(-1);
-  if (!last || last.candidates.length === 0) {
-    return { top: [], bottom: [], target: baseTarget };
-  }
-  const best = [...last.candidates].sort((a, b) => b.composite - a.composite)[0];
-  if (!best) return { top: [], bottom: [], target: baseTarget };
-  const byScore = [...best.scenarios].sort((a, b) => b.composite - a.composite);
-  const toTrace = (s) => ({
-    id: s.scenarioId,
-    score: s.composite,
-    // The judge's "why it scored low" — grounds the reflection on real failure
-    // patterns instead of blind rephrasing. Generalizable by the judge contract.
-    ...s.notes ? { failureNote: s.notes } : {}
-  });
-  const top = byScore.slice(0, evidenceK).map(toTrace);
-  const bottom = byScore.slice(-evidenceK).reverse().map(toTrace);
-  const weakest = Object.entries(best.dimensions).sort((a, b) => a[1] - b[1]).slice(0, 3).map(([dim, value]) => `${dim} (${value.toFixed(2)})`);
-  const target = weakest.length > 0 ? `${baseTarget} \u2014 weakest dimensions: ${weakest.join(", ")}` : baseTarget;
-  return { top, bottom, target };
-}
-
 // src/campaign/gates/heldout-gate.ts
 function heldOutGate(options) {
   const deltaThreshold = options.deltaThreshold ?? 0.5;
@@ -1251,6 +1084,7 @@ function labelTrustRank(trust) {
 // src/campaign/presets/run-optimization.ts
 import { createHash } from "crypto";
 async function runOptimization(opts) {
+  const { proposer } = opts;
   const promoteTopK = opts.promoteTopK ?? 2;
   if (typeof opts.runDir !== "string" || opts.runDir.trim().length === 0) {
     throw new Error("runOptimization: runDir is required and must be a non-empty string");
@@ -1273,9 +1107,9 @@ async function runOptimization(opts) {
     toParetoParent(opts.baselineSurface, winnerSurfaceHash, baselineCampaign, -1)
   ];
   for (let gen = 0; gen < opts.maxGenerations; gen++) {
-    if (opts.driver.decide?.({ history }).stop) break;
+    if (proposer.decide?.({ history }).stop) break;
     const paretoParents = computeParetoFrontier(scored);
-    const proposed = await opts.driver.propose({
+    const proposed = await proposer.propose({
       currentSurface: currentSurfaces[0] ?? opts.baselineSurface,
       history,
       findings: currentFindings,
@@ -1423,9 +1257,9 @@ async function runImprovementLoop(opts) {
       "runImprovementLoop: autoOnPromote='config' is deferred to Pass B (requires shadow deploy + rollback + ensemble judges). Use 'pr' or 'none' in v0.40."
     );
   }
-  if (opts.tracing === "off" && opts.driver) {
+  if (opts.tracing === "off" && opts.proposer) {
     throw new Error(
-      "runImprovementLoop: tracing='off' is forbidden when a driver is wired. The improvement loop without traces is unattributable; candidate surfaces cannot be cited back to spans and the optimization dataset goes unfed."
+      "runImprovementLoop: tracing='off' is forbidden when a proposer is wired. The improvement loop without traces is unattributable; candidate surfaces cannot be cited back to spans and the optimization dataset goes unfed."
     );
   }
   if (opts.autoOnPromote === "pr" && (!opts.ghOwner || !opts.ghRepo)) {
@@ -1443,7 +1277,7 @@ async function runImprovementLoop(opts) {
   const dispatchTimeoutMs = opts.dispatchTimeoutMs ?? DEFAULT_DISPATCH_TIMEOUT_MS;
   const optimization = await runOptimization({ ...opts, dispatchTimeoutMs });
   const winnerIsBaseline = optimization.winnerSurfaceHash === surfaceHash(opts.baselineSurface);
-  const { runCampaign: runCampaign2 } = await import("./run-campaign-7WNXMDSN.js");
+  const { runCampaign: runCampaign2 } = await import("./run-campaign-WXY7KI67.js");
   const baselineOnHoldout = await runCampaign2({
     ...opts,
     dispatchTimeoutMs,
@@ -1538,6 +1372,173 @@ ${fmt(winnerSurface)}`;
   return lines.join("\n");
 }
 
+// src/campaign/proposers/gepa.ts
+var REFLECTION_SYSTEM = 'You are an expert prompt engineer performing GEPA-style reflective mutation. You are given a prompt surface, its top trials (preserve what works) and its bottom trials (the evidence to fix). For each proposal, reason in this order before writing the payload: (1) LOCALIZE \u2014 point to the exact span of the current surface responsible for a bottom-trial failure; (2) DIAGNOSE the root cause (a missing rule, an ambiguous instruction, an over-broad directive), not just the symptom; (3) propose the MINIMAL, GENERALIZABLE edit that fixes the whole failure class \u2014 state it as a rule the agent should follow, never a patch memorized to the shown trials (that is overfitting and will not transfer to the held-out set); (4) PRESERVE every instruction the top trials depend on \u2014 do not delete or weaken working guidance. Put this localize\u2192diagnose\u2192fix reasoning in each proposal\'s `rationale`. Output ONLY a JSON object of shape {"proposals":[{"label":string,"rationale":string,"payload":string}]} where each `payload` is the FULL improved surface text. No prose outside the JSON.';
+var COMBINE_SYSTEM = 'You are an expert prompt engineer performing a GEPA "combine complementary lessons" merge. You are given several non-dominated versions of one surface; each is uniquely best on different scenarios. Produce ONE new version that keeps what makes each version strong on its winning scenarios and resolves conflicts in favor of the more general rule. Output ONLY a JSON object of shape {"proposals":[{"label":string,"rationale":string,"payload":string}]} with exactly one proposal whose `payload` is the FULL merged surface text. No prose outside the JSON.';
+function gepaProposer(opts) {
+  const evidenceK = opts.evidenceK ?? 3;
+  const combineParents = opts.combineParents ?? true;
+  const combineMaxParents = opts.combineMaxParents ?? 4;
+  if (combineParents && combineMaxParents < 1) {
+    throw new Error("gepaProposer: combineMaxParents must be >= 1 when combineParents is enabled");
+  }
+  return {
+    kind: "gepa",
+    async propose(ctx) {
+      const parent = typeof ctx.currentSurface === "string" ? ctx.currentSurface : JSON.stringify(ctx.currentSurface);
+      const constraints = opts.constraints;
+      const preserveSections = constraints?.preserveSections !== void 0 ? constraints.preserveSections.length === 0 ? extractH2Sections(parent) : constraints.preserveSections : null;
+      const maxEdits = constraints?.maxSentenceEdits;
+      const out = [];
+      const seen = /* @__PURE__ */ new Set();
+      const accept = (payload, label, rationale) => {
+        const text = typeof payload === "string" ? payload.trim() : "";
+        if (!text || text === parent || seen.has(text)) return;
+        if (preserveSections && !validatePreservedSections(text, preserveSections)) return;
+        if (maxEdits !== void 0 && countSentenceEdits(parent, text) > maxEdits * 2) return;
+        seen.add(text);
+        out.push({ surface: text, label, rationale });
+      };
+      const stringParents = (combineParents ? ctx.paretoParents ?? [] : []).filter((p) => typeof p.surface === "string").sort((a, b) => b.composite - a.composite).slice(0, combineMaxParents);
+      if (stringParents.length > 1) {
+        const combinePrompt = buildCombinePrompt({
+          target: opts.target,
+          parents: stringParents,
+          evidenceK
+        });
+        const combineResult = await callLlm(
+          {
+            model: opts.model,
+            messages: [
+              { role: "system", content: COMBINE_SYSTEM },
+              { role: "user", content: combinePrompt }
+            ],
+            jsonMode: true,
+            temperature: opts.temperature ?? 0.7,
+            maxTokens: opts.maxTokens ?? 6e3
+          },
+          opts.llm
+        );
+        const merged = parseReflectionResponse(combineResult.content, 1)[0];
+        if (merged) {
+          accept(
+            merged.payload,
+            merged.label || "pareto-combine",
+            merged.rationale || `combined ${stringParents.length} non-dominated parents (gen ${stringParents.map((p) => p.generation).join(",")})`
+          );
+        }
+      }
+      const reflectCount = Math.max(0, ctx.populationSize - out.length);
+      if (reflectCount > 0) {
+        const { top, bottom, target } = buildEvidence(ctx, evidenceK, opts.target);
+        const userPrompt = buildReflectionPrompt({
+          target,
+          parentPayload: parent,
+          topTrials: top,
+          bottomTrials: bottom,
+          childCount: reflectCount,
+          mutationPrimitives: opts.mutationPrimitives
+        });
+        const analyst = renderAnalystEvidence(ctx.findings, ctx.report);
+        const finalPrompt = analyst ? `${userPrompt}
+
+${analyst}` : userPrompt;
+        const result = await callLlm(
+          {
+            model: opts.model,
+            messages: [
+              { role: "system", content: REFLECTION_SYSTEM },
+              { role: "user", content: finalPrompt }
+            ],
+            jsonMode: true,
+            temperature: opts.temperature ?? 0.7,
+            maxTokens: opts.maxTokens ?? 6e3
+          },
+          opts.llm
+        );
+        for (const proposal of parseReflectionResponse(result.content, reflectCount)) {
+          accept(proposal.payload, proposal.label, proposal.rationale);
+        }
+      }
+      return out.slice(0, ctx.populationSize);
+    }
+  };
+}
+function buildCombinePrompt(args) {
+  const lines = [
+    `You are merging ${args.parents.length} versions of: ${args.target}.`,
+    "",
+    "Each version is on the Pareto frontier \u2014 none dominates the others; each",
+    "wins on different scenarios. Combine their complementary strengths into",
+    "ONE version. Below, each version lists the scenarios it scores highest on.",
+    ""
+  ];
+  args.parents.forEach((p, i) => {
+    const tag = String.fromCharCode(65 + i);
+    const best = Object.entries(p.objectives).sort((a, b) => b[1] - a[1]).slice(0, args.evidenceK).map(([id, score]) => `${id} (${score.toFixed(2)})`);
+    lines.push(
+      `### Version ${tag} (mean ${p.composite.toFixed(2)}; strongest on: ${best.join(", ") || "n/a"})`,
+      "```",
+      p.surface,
+      "```",
+      ""
+    );
+  });
+  lines.push(
+    "Return ONE merged version that would score well on the union of every",
+    "version's winning scenarios. Keep each version's specific winning rule;",
+    "where two rules conflict, prefer the more general one and note the choice",
+    "in your rationale."
+  );
+  return lines.join("\n");
+}
+function extractH2Sections(text) {
+  const out = [];
+  for (const line of text.split("\n")) {
+    const match = /^##\s+(.+?)\s*$/.exec(line);
+    if (match) out.push(match[1]);
+  }
+  return out;
+}
+function countSentenceEdits(baseline, candidate) {
+  const norm = (s) => s.split(/(?<=[.!?])\s+|\n/g).map((p) => p.trim()).filter((p) => p.length > 0);
+  const a = new Set(norm(baseline));
+  const b = new Set(norm(candidate));
+  let edits = 0;
+  for (const s of a) if (!b.has(s)) edits++;
+  for (const s of b) if (!a.has(s)) edits++;
+  return edits;
+}
+function validatePreservedSections(candidate, required) {
+  if (required.length === 0) return true;
+  const have = new Set(extractH2Sections(candidate));
+  for (const section of required) {
+    if (!have.has(section)) return false;
+  }
+  return true;
+}
+function buildEvidence(ctx, evidenceK, baseTarget) {
+  const last = ctx.history.at(-1);
+  if (!last || last.candidates.length === 0) {
+    return { top: [], bottom: [], target: baseTarget };
+  }
+  const best = [...last.candidates].sort((a, b) => b.composite - a.composite)[0];
+  if (!best) return { top: [], bottom: [], target: baseTarget };
+  const byScore = [...best.scenarios].sort((a, b) => b.composite - a.composite);
+  const toTrace = (s) => ({
+    id: s.scenarioId,
+    score: s.composite,
+    // The judge's "why it scored low" — grounds the reflection on real failure
+    // patterns instead of blind rephrasing. Generalizable by the judge contract.
+    ...s.notes ? { failureNote: s.notes } : {}
+  });
+  const top = byScore.slice(0, evidenceK).map(toTrace);
+  const bottom = byScore.slice(-evidenceK).reverse().map(toTrace);
+  const weakest = Object.entries(best.dimensions).sort((a, b) => a[1] - b[1]).slice(0, 3).map(([dim, value]) => `${dim} (${value.toFixed(2)})`);
+  const target = weakest.length > 0 ? `${baseTarget} \u2014 weakest dimensions: ${weakest.join(", ")}` : baseTarget;
+  return { top, bottom, target };
+}
+
 // src/campaign/provenance.ts
 import { createHash as createHash2 } from "crypto";
 import { join as join2 } from "path";
@@ -1824,9 +1825,6 @@ export {
   buildReflectionPrompt,
   renderAnalystEvidence,
   parseReflectionResponse,
-  gepaDriver,
-  extractH2Sections,
-  countSentenceEdits,
   heldOutGate,
   openAutoPr,
   campaignMeanComposite,
@@ -1837,6 +1835,9 @@ export {
   surfaceHash,
   runImprovementLoop,
   defaultRenderDiff,
+  gepaProposer,
+  extractH2Sections,
+  countSentenceEdits,
   surfaceContentHash,
   buildLoopProvenanceRecord,
   loopProvenanceSpans,
@@ -1844,4 +1845,4 @@ export {
   provenanceSpansPath,
   emitLoopProvenance
 };
-//# sourceMappingURL=chunk-EGPMSBEZ.js.map
+//# sourceMappingURL=chunk-7QTQKIDD.js.map