@tangle-network/agent-eval 0.54.0 → 0.56.0

package/dist/contract/index.js

@@ -6,12 +6,12 @@ import {
   heldOutGate,
   runEval,
   runImprovementLoop
-} from "../chunk-YXD7GWJI.js";
+} from "../chunk-JB4UWIM6.js";
 import {
   fsCampaignStorage,
   inMemoryCampaignStorage,
   runCampaign
-} from "../chunk-J3EIOI3O.js";
+} from "../chunk-74Y2EMNH.js";
 import {
   createHostedClient
 } from "../chunk-FQK2CCIM.js";
@@ -26,14 +26,14 @@ import {
 } from "../chunk-3RF76KTD.js";
 import {
   paretoChart
-} from "../chunk-EGIPWXHL.js";
+} from "../chunk-OLIBRKRD.js";
 import {
   cohensD,
   pairedBootstrap,
   pairedMde,
   pairedTTest,
   requiredSampleSize
-} from "../chunk-WP7SY7AI.js";
+} from "../chunk-S3SDD56V.js";
 import "../chunk-GGE4NNQT.js";
 import "../chunk-47X6LRCE.js";
 import "../chunk-5BKGXME7.js";
@@ -1029,6 +1029,78 @@ function cellsToRunRecords(cells, candidateId, runId) {
   });
 }
 
+// src/contract/intake/agent-trace.ts
+function rangeLines(r) {
+  return Math.max(0, r.end_line - r.start_line + 1);
+}
+function parseAgentTrace(records) {
+  const acc = /* @__PURE__ */ new Map();
+  for (const record of records) {
+    const sha = record.vcs?.revision;
+    if (!sha) continue;
+    let a = acc.get(sha);
+    if (!a) {
+      a = {
+        models: /* @__PURE__ */ new Set(),
+        tools: /* @__PURE__ */ new Set(),
+        files: /* @__PURE__ */ new Set(),
+        conversationCount: 0,
+        lineCount: 0,
+        humanInvolved: false
+      };
+      acc.set(sha, a);
+    }
+    if (record.tool?.name) a.tools.add(record.tool.name);
+    for (const file of record.files ?? []) {
+      a.files.add(file.path);
+      for (const conv of file.conversations ?? []) {
+        a.conversationCount += 1;
+        for (const range of conv.ranges ?? []) {
+          const contributor = range.contributor ?? conv.contributor;
+          a.lineCount += rangeLines(range);
+          if (!contributor) continue;
+          if (contributor.type === "human" || contributor.type === "mixed") {
+            a.humanInvolved = true;
+          }
+          if ((contributor.type === "ai" || contributor.type === "mixed") && contributor.model_id) {
+            a.models.add(contributor.model_id);
+          }
+        }
+      }
+    }
+  }
+  const index = /* @__PURE__ */ new Map();
+  for (const [sha, a] of acc) {
+    index.set(sha, {
+      commitSha: sha,
+      aiModels: [...a.models].sort(),
+      tools: [...a.tools].sort(),
+      conversationCount: a.conversationCount,
+      fileCount: a.files.size,
+      lineCount: a.lineCount,
+      humanInvolved: a.humanInvolved
+    });
+  }
+  return index;
+}
+function partitionRunsByAuthoringModel(runs, index) {
+  const byModel = /* @__PURE__ */ new Map();
+  const unattributed = [];
+  for (const run of runs) {
+    const provenance = index.get(run.commitSha);
+    if (!provenance || provenance.aiModels.length === 0) {
+      unattributed.push(run);
+      continue;
+    }
+    for (const model of provenance.aiModels) {
+      const cohort = byModel.get(model) ?? [];
+      cohort.push(run);
+      byModel.set(model, cohort);
+    }
+  }
+  return { byModel, unattributed };
+}
+
 // src/contract/intake/feedback-table.ts
 function fromFeedbackTable(opts) {
   const { ratings, meta = [], scale, emitRaterScores = true } = opts;
@@ -1204,6 +1276,8 @@ export {
   gepaDriver,
   heldOutGate,
   inMemoryCampaignStorage,
+  parseAgentTrace,
+  partitionRunsByAuthoringModel,
   runCampaign,
   runEval,
   runImprovementLoop,