kairn-cli 2.2.6 → 2.2.8

package/dist/cli.js

@@ -4748,6 +4748,13 @@ the agent lacks a tool it needs, or has tools that add noise without benefit.
 - Consider both additions AND removals. Remove sections that add noise without improving task performance.
 - Bloated harnesses hurt performance \u2014 trim what isn't earning its keep.
 
+## Anti-Gaming (CRITICAL)
+- Mutations must improve GENERAL-PURPOSE development quality, not target specific eval criteria.
+- You do NOT have access to scoring rubrics or expected outcomes. Diagnose problems from traces only.
+- Do NOT add over-specified rules that restate existing conventions with stronger emphasis (e.g., changing "use chalk.green for success" to "MUST use chalk.green, no exceptions"). If a convention already exists, trust it.
+- Do NOT add rules that only apply to a narrow eval scenario (e.g., write permissions for a specific directory just because one task needed it).
+- Ask: "Would this mutation help a developer working on ANY task in this project?" If not, don't propose it.
+
 Return ONLY valid JSON.`;
 var STDOUT_TRUNCATION_LIMIT = 1e3;
 var MAX_CONTEXT_CHARS = 1e5;
@@ -4806,8 +4813,6 @@ ${content}
         `### Task: ${task.id}
 - Template: ${task.template}
 - Description: ${task.description}
-- Expected outcome: ${Array.isArray(task.expected_outcome) ? task.expected_outcome.join("; ") : task.expected_outcome}
-- Scoring: ${task.scoring}
 `
       );
     }
@@ -5125,8 +5130,31 @@ async function evolve(workspacePath, tasks, kairnConfig, evolveConfig, onProgres
       break;
     }
     onProgress?.({ type: "iteration-start", iteration: iter });
-    const { results, aggregate } = await evaluateAll(
-      tasks,
+    const isFirstIter = iter === 0;
+    const isLastIter = iter === evolveConfig.maxIterations - 1;
+    const prevLog = history.length > 0 ? history[history.length - 1] : null;
+    let tasksToRun = tasks;
+    const carriedScores = {};
+    if (!isFirstIter && !isLastIter && prevLog) {
+      tasksToRun = [];
+      for (const task of tasks) {
+        const prevScore = prevLog.taskResults[task.id];
+        const prevValue = prevScore ? prevScore.score ?? (prevScore.pass ? 100 : 0) : 0;
+        if (prevValue >= 100) {
+          carriedScores[task.id] = { pass: true, score: 100 };
+          onProgress?.({
+            type: "task-skipped",
+            iteration: iter,
+            taskId: task.id,
+            message: `Skipped ${task.id} (scored 100% last iteration)`
+          });
+        } else {
+          tasksToRun.push(task);
+        }
+      }
+    }
+    const { results: evalResults, aggregate: evalAggregate } = await evaluateAll(
+      tasksToRun,
       harnessPath,
       workspacePath,
       iter,
@@ -5135,6 +5163,13 @@ async function evolve(workspacePath, tasks, kairnConfig, evolveConfig, onProgres
       evolveConfig.runsPerTask,
       evolveConfig.parallelTasks
     );
+    const results = { ...carriedScores, ...evalResults };
+    const allScores = Object.values(results);
+    const total = allScores.reduce(
+      (sum, s) => sum + (s.score ?? (s.pass ? 100 : 0)),
+      0
+    );
+    const aggregate = allScores.length > 0 ? total / allScores.length : 0;
     onProgress?.({ type: "iteration-scored", iteration: iter, score: aggregate });
     if (iter === 0) baselineScore = aggregate;
     if (iter > 0 && aggregate < bestScore) {
@@ -5841,6 +5876,9 @@ evolveCommand.command("run").description("Run tasks against the current harness"
           case "task-run":
             console.log(chalk14.dim(`      ${event.message ?? ""}`));
             break;
+          case "task-skipped":
+            console.log(chalk14.dim(`    SKIP  ${event.taskId ?? "unknown"} (100% last iteration)`));
+            break;
           case "task-scored": {
             const taskScore = event.score ?? 0;
             const taskStatus = taskScore >= 100 ? chalk14.green("PASS") : taskScore >= 60 ? chalk14.yellow("PARTIAL") : chalk14.red("FAIL");