claude-overnight 1.55.1 → 1.55.2

package/dist/bin/evolve.js

@@ -34,6 +34,7 @@ Options:
   --population <n>        Max population size (default: 8)
   --plateau <n>           Stop early if no improvement for N generations (default: 3)
   --reps <n>              Repetitions per (variant, case, model) for noise floor (default: 1)
+  --concurrency <n>       Max in-flight eval calls (default: 8; bump for slow endpoints)
   --judge                 Use llm-judge for content scoring (costs extra API calls)
   --judge-model <model>   Model to use for the judge (default: same as eval-model)
   --judge-top-n <n>       Judge only the top-N variants per generation (default: 4)
@@ -112,6 +113,10 @@ function parseArgs() {
                 opts.reps = parseInt(v, 10);
                 i++;
                 break;
+            case "--concurrency":
+                opts.concurrency = parseInt(v, 10);
+                i++;
+                break;
             case "--judge":
                 opts.useJudge = true;
                 break;
@@ -215,6 +220,7 @@ async function main() {
         populationCap: opts.population,
         plateauGenerations: opts.plateau,
         repetitions: opts.reps > 1 ? opts.reps : undefined,
+        concurrency: opts.concurrency,
         judge: opts.useJudge
             ? {
                 model: opts.judgeModel ?? opts.evalModel,

package/dist/core/_version.d.ts

@@ -1 +1 @@
-export declare const VERSION = "1.55.1";
+export declare const VERSION = "1.55.2";

package/dist/core/_version.js

@@ -1,2 +1,2 @@
 // Auto-generated by build — do not edit manually.
-export const VERSION = "1.55.1";
+export const VERSION = "1.55.2";

package/dist/prompt-evolution/evaluator.js

@@ -22,7 +22,7 @@ import { defaultCallModel, attemptJsonParse, } from "./transport.js";
 export async function buildMatrix(variants, cases, opts) {
     const models = opts.models && opts.models.length > 0 ? opts.models : [opts.model];
     const reps = Math.max(1, opts.repetitions ?? 1);
-    const concurrency = opts.concurrency ?? 4;
+    const concurrency = opts.concurrency ?? 8;
     const transport = opts.callModel ?? defaultCallModel;
     // Build the full job list: (variant × case × model × rep).
     const jobs = [];
@@ -35,13 +35,19 @@ export async function buildMatrix(variants, cases, opts) {
             }
         }
     }
-    // Raw results, keyed by variant:case:model, each an array of per-rep results.
+    // Work-stealing pool: keep `concurrency` jobs in flight at all times so a
+    // slow call (Kimi at 4 min/call is typical) doesn't block the others in its
+    // slice. Previous batch-and-wait loop serialized the slowest job in every
+    // window of `concurrency`.
     const rawByKey = new Map();
     let done = 0;
-    for (let i = 0; i < jobs.length; i += concurrency) {
-        const batch = jobs.slice(i, i + concurrency);
-        const batchResults = await Promise.all(batch.map((job) => runSingle(job, opts, transport)));
-        for (const r of batchResults) {
+    let next = 0;
+    const worker = async () => {
+        while (true) {
+            const i = next++;
+            if (i >= jobs.length)
+                return;
+            const r = await runSingle(jobs[i], opts, transport);
             const key = `${r.variantId}:${r.caseHash}:${r.model ?? ""}`;
             const arr = rawByKey.get(key) ?? [];
             arr.push(r);
@@ -49,7 +55,8 @@ export async function buildMatrix(variants, cases, opts) {
             done++;
             opts.onProgress?.(done, jobs.length, r.caseName, r.variantId);
         }
-    }
+    };
+    await Promise.all(Array.from({ length: Math.min(concurrency, jobs.length) }, worker));
     // Collapse reps: one aggregated EvaluationResult per (variant, case, model).
     const aggregated = new Map();
     for (const [key, runs] of rawByKey) {
@@ -198,11 +205,19 @@ async function runJudge(variants, cases, models, aggregated, judge) {
             }
         }
     }
-    // Run judge calls with modest concurrency to stay under provider rate limits.
-    const concurrency = 3;
-    for (let i = 0; i < jobs.length; i += concurrency) {
-        await Promise.all(jobs.slice(i, i + concurrency).map((fn) => fn()));
-    }
+    // Work-stealing pool for judge calls — modest concurrency to stay under
+    // provider rate limits, but no slice-blocking.
+    const judgeConcurrency = 3;
+    let nextJob = 0;
+    const judgeWorker = async () => {
+        while (true) {
+            const i = nextJob++;
+            if (i >= jobs.length)
+                return;
+            await jobs[i]();
+        }
+    };
+    await Promise.all(Array.from({ length: Math.min(judgeConcurrency, jobs.length) }, judgeWorker));
 }
 function averageDimensions(scores) {
     if (scores.length === 0)

package/dist/prompt-evolution/index.d.ts

@@ -52,6 +52,8 @@ export interface EvolveOpts {
     evalModels?: string[];
     /** Repetitions per (variant, case, model). Default 1. Recommended ≥3 for noise floor. */
     repetitions?: number;
+    /** Max in-flight eval calls. Default 8. Raise for slow endpoints, lower for strict rate limits. */
+    concurrency?: number;
     /** Optional llm-judge — replaces the heuristic content score for top-N variants each gen. */
     judge?: JudgeOpts & {
         topN?: number;

package/dist/prompt-evolution/index.js

@@ -63,7 +63,7 @@ export async function evolvePrompt(opts) {
             models: opts.evalModels,
             baseUrl: opts.baseUrl,
             authToken: opts.authToken,
-            concurrency: 4,
+            concurrency: opts.concurrency ?? 8,
             repetitions: opts.repetitions,
             judge: opts.judge,
             onProgress: (done, total, caseName, variantId) => {
@@ -178,7 +178,7 @@ export async function evolvePrompt(opts) {
         models: opts.evalModels,
         baseUrl: opts.baseUrl,
         authToken: opts.authToken,
-        concurrency: 4,
+        concurrency: opts.concurrency ?? 8,
         repetitions: opts.repetitions,
         judge: opts.judge,
     });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-overnight",
-  "version": "1.55.1",
+  "version": "1.55.2",
   "description": "Overnight parallel coding agents in git worktrees, with a self-curating skill memory that improves while the run is going. Mix Claude Opus as planner, Kimi 2.6 or Cursor composer-2 as cheap fast worker, Gemini or Qwen for bulk implementation. Multi-wave autonomous loop that plans, executes, reviews, and steers itself until the objective is met. Crash-safe resume, rate-limit aware, usage cap preserves headroom for your interactive Claude Code.",
   "type": "module",
   "bin": {

package/plugins/claude-overnight/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-overnight",
-  "version": "1.55.1",
+  "version": "1.55.2",
   "description": "Claude Code skill for understanding, installing, and inspecting claude-overnight runs: overnight parallel coding agents in git worktrees with a self-curating skill memory, multi-wave steering, three-layer review, and crash-safe resume. Mix Opus planner with Kimi 2.6, Cursor composer-2, Gemini, Qwen, or any Anthropic-compatible worker.",
   "author": {
     "name": "Francesco Fornace"