@tangle-network/agent-runtime 0.85.0 → 0.86.0

package/dist/agent.js

@@ -1,7 +1,7 @@
-import "./chunk-DLAEEF26.js";
+import "./chunk-3TZOXS7B.js";
 import {
   createSandboxForSpec
-} from "./chunk-XN5TIJGP.js";
+} from "./chunk-F7OO2SKB.js";
 import "./chunk-UD4BHQMI.js";
 import "./chunk-BZF3KQ6G.js";
 import {

package/dist/{chunk-DLAEEF26.js → chunk-3TZOXS7B.js}

@@ -15,7 +15,7 @@ import {
   settledToIteration,
   supervise,
   withDriverExecutor
-} from "./chunk-XN5TIJGP.js";
+} from "./chunk-F7OO2SKB.js";
 import {
   addTokenUsage,
   isAbortError,
@@ -977,49 +977,68 @@ function defineLeaderboard(spec) {
       return p;
     };
     let shotNonce = 0;
-    const dispatch = spec.dispatch ?? loopDispatch({
-      sandboxClient,
-      toLoopOptions: (scenario, profile) => {
-        const axis = harnessAxisOf(profile);
-        const modelId = bareModel(axis?.model ?? models[0] ?? "");
-        return {
-          // naiveDriver = the no-signal retry floor: re-run the same case as
-          // an independent attempt until one scores (>0) or the shot cap.
-          driver: naiveDriver({
-            continuation: "",
-            applyContinuation: (task) => task,
-            maxIterations: shots
-          }),
-          agentRun: {
-            profile,
-            taskToPrompt: (s) => `${promptOf(s)}
+    const dispatch = spec.dispatch ?? ((profile, scenario, dispatchCtx) => {
+      const cellDispatch = loopDispatch({
+        sandboxClient,
+        toLoopOptions: (cellScenario, cellProfile) => {
+          const axis = harnessAxisOf(cellProfile);
+          const modelId = bareModel(axis?.model ?? models[0] ?? "");
+          return {
+            // naiveDriver = the no-signal retry floor: re-run the same case as
+            // an independent attempt until one scores (>0) or the shot cap.
+            driver: naiveDriver({
+              continuation: "",
+              applyContinuation: (task) => task,
+              maxIterations: shots
+            }),
+            agentRun: {
+              profile: cellProfile,
+              taskToPrompt: (s) => `${promptOf(s)}
 
 <!-- independent-attempt:${shotNonce++} -->`,
-            ...axis ? {
-              sandboxOverrides: {
-                backend: {
-                  type: axis.harness,
-                  model: { ...spec.modelBackend, model: modelId }
+              ...axis ? {
+                sandboxOverrides: {
+                  backend: {
+                    type: axis.harness,
+                    model: { ...spec.modelBackend, model: modelId }
+                  }
                 }
+              } : {}
+            },
+            output: {
+              parse: (events) => spec.parseOutput ? spec.parseOutput(events, cellScenario.case) : (
+                // The default decode produces string — the TArtifact
+                // default. A structured-TArtifact spec supplies parseOutput
+                // (documented on the field), so this cast never lies.
+                collectAgentResponseText(events) ?? ""
+              )
+            },
+            validator: {
+              validate: async (output) => {
+                const s = normalizeScore(spec.score(output, cellScenario.case));
+                return { valid: s.composite > 0, score: s.composite };
               }
-            } : {}
-          },
-          output: {
-            parse: (events) => {
-              spec.onCellEvents?.(events, scenario.case);
-              return spec.parseOutput ? spec.parseOutput(events, scenario.case) : collectAgentResponseText(events) ?? "";
-            }
-          },
-          validator: {
-            validate: async (output) => {
-              const s = normalizeScore(spec.score(output, scenario.case));
-              return { valid: s.composite > 0, score: s.composite };
+            },
+            task: cellScenario,
+            maxIterations: shots
+          };
+        },
+        toArtifact: (result) => {
+          for (const iter of result.iterations) {
+            spec.onCellEvents?.(iter.events, scenario.case, {
+              index: iter.index,
+              ...iter.error ? { error: iter.error.message } : {},
+              ...iter.verdict ? { verdict: { score: iter.verdict.score } } : {}
+            });
+            if (spec.resolveModel) {
+              const served = spec.resolveModel(iter.events);
+              if (served !== void 0) dispatchCtx.cost.observeModel?.(served);
             }
-          },
-          task: scenario,
-          maxIterations: shots
-        };
-      }
+          }
+          return result.winner?.output;
+        }
+      });
+      return cellDispatch(profile, scenario, dispatchCtx);
     });
     await spec.setup?.(ctx);
     try {
@@ -5851,4 +5870,4 @@ export {
   computeFindingId,
   makeFinding2 as makeFinding
 };
-//# sourceMappingURL=chunk-DLAEEF26.js.map
+//# sourceMappingURL=chunk-3TZOXS7B.js.map