claude-overnight 1.57.4 → 1.58.0

package/dist/bin/evolve.d.ts

@@ -8,7 +8,7 @@
  *
  * Examples:
  *   claude-overnight-evolve --prompt 10_planning/10-3_plan --eval-model claude-haiku-4-5 --generations 3
- *   claude-overnight-evolve --target mcp-browser --prompt-kind plan-supervision --eval-model kimi-k2-6
+ *   claude-overnight-evolve --target mcp-browser --prompt-kind plan-supervision --eval-model kimi-for-coding
  *
  * Requires ANTHROPIC_API_KEY (or ANTHROPIC_AUTH_TOKEN) in env. When `--target
  * mcp-browser` is used the cwd must be the MCP-browser repo root (so

package/dist/bin/evolve.js

@@ -8,7 +8,7 @@
  *
  * Examples:
  *   claude-overnight-evolve --prompt 10_planning/10-3_plan --eval-model claude-haiku-4-5 --generations 3
- *   claude-overnight-evolve --target mcp-browser --prompt-kind plan-supervision --eval-model kimi-k2-6
+ *   claude-overnight-evolve --target mcp-browser --prompt-kind plan-supervision --eval-model kimi-for-coding
  *
  * Requires ANTHROPIC_API_KEY (or ANTHROPIC_AUTH_TOKEN) in env. When `--target
  * mcp-browser` is used the cwd must be the MCP-browser repo root (so
@@ -28,7 +28,10 @@ Options:
   --prompt <path>         Prompt file path (claude-overnight)
   --prompt-kind <kind>    MCP-browser prompt kind: planning | review | evolution |
                           goal-refinement | plan-supervision | simple-supervision | stuck-analysis
-  --eval-model <model>    Fast model for evaluation (default: claude-haiku-4-5)
+  --eval-model <model>    Fast model for evaluation (default: claude-haiku-4-5).
+                          For Kimi endpoints use "kimi-for-coding" (stable alias,
+                          auto-upgrades as flagship revs). For Moonshot platform
+                          API use "kimi-k2.6" (dot, not dash).
   --eval-models <list>    Comma-separated list to run cross-model (overrides --eval-model)
   --mutate-model <model>  Smarter model for mutation (defaults to eval-model)
   --generations <n>       Number of evolution generations (default: 10)
@@ -37,6 +40,11 @@ Options:
   --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)
   --batch                 Use provider batch API (50% cheaper, slower wall-clock)
+  --batch-base-url <url>  Override base URL for batch only (e.g. api.moonshot.ai/v1
+                          when online uses api.kimi.com/coding)
+  --batch-auth-token <t>  Override auth token for batch only
+  --batch-model <model>   Override model for batch only (e.g. "kimi-k2.6" for
+                          Moonshot platform when online uses "kimi-for-coding")
   --adaptive-cap <n>      Adaptive sampling: extend reps up to N when σ > threshold (default: off)
   --adaptive-threshold <x> σ threshold that triggers an extra rep (default: 0.1)
   --judge                 Use llm-judge for content scoring (costs extra API calls)
@@ -138,6 +146,18 @@ function parseArgs() {
             case "--batch":
                 opts.batch = true;
                 break;
+            case "--batch-base-url":
+                opts.batchBaseUrl = v;
+                i++;
+                break;
+            case "--batch-auth-token":
+                opts.batchAuthToken = v;
+                i++;
+                break;
+            case "--batch-model":
+                opts.batchModel = v;
+                i++;
+                break;
             case "--adaptive-cap":
                 opts.adaptiveCap = parseInt(v, 10);
                 i++;
@@ -328,6 +348,9 @@ async function evolveOne(opts) {
         repetitions: opts.reps > 1 ? opts.reps : undefined,
         concurrency: opts.concurrency,
         batch: opts.batch || undefined,
+        batchBaseUrl: opts.batchBaseUrl,
+        batchAuthToken: opts.batchAuthToken,
+        batchModel: opts.batchModel,
         adaptiveReps: opts.adaptiveCap
             ? { cap: opts.adaptiveCap, threshold: opts.adaptiveThreshold }
             : undefined,

package/dist/core/_version.d.ts

@@ -1 +1 @@
-export declare const VERSION = "1.57.4";
+export declare const VERSION = "1.58.0";

package/dist/core/_version.js

@@ -1,2 +1,2 @@
 // Auto-generated by build — do not edit manually.
-export const VERSION = "1.57.4";
+export const VERSION = "1.58.0";

package/dist/prompt-evolution/evaluator-judge.js

@@ -82,8 +82,12 @@ async function runJudgeBatch(cells, judge, opts) {
         : null;
     const transport = opts.batchCallModel ?? batchCallModel;
     const results = await transport(batchJobs, {
-        baseUrl: judge.baseUrl ?? opts.baseUrl,
-        authToken: judge.authToken ?? opts.authToken,
+        // Judge batch follows the same override hierarchy as eval batch: if a
+        // dedicated batch endpoint is set on EvalOpts, use it; else fall back
+        // to the judge's own endpoint or the main one.
+        baseUrl: opts.batchBaseUrl ?? judge.baseUrl ?? opts.baseUrl,
+        authToken: opts.batchAuthToken ?? judge.authToken ?? opts.authToken,
+        modelOverride: opts.batchModel,
         maxTokens: judge.maxTokens ?? 2048,
         resumeBatchId: existing?.batchId,
         onSubmitted: (batchId, p) => {

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

@@ -53,6 +53,17 @@ export interface EvalOpts {
     callModel?: CallModel;
     /** Use provider batch API instead of online calls (50% cheaper, slower wall-clock). */
     batch?: boolean;
+    /**
+     * Override base URL for batch submissions only — lets batch hit a
+     * different endpoint than online. Key use-case: Kimi users whose online
+     * traffic runs through api.kimi.com/coding (which has no batch) but
+     * whose batch traffic should go to api.moonshot.ai/v1.
+     */
+    batchBaseUrl?: string;
+    /** Override auth token for batch when batchBaseUrl needs a different key. */
+    batchAuthToken?: string;
+    /** Override model for batch submissions (e.g., kimi-k2.6 when online uses kimi-for-coding). */
+    batchModel?: string;
     /** Run id — required when batch=true so state is crash-resumable. */
     runId?: string;
     /** Current generation number — used to key batch state. */

package/dist/prompt-evolution/evaluator.js

@@ -268,8 +268,9 @@ async function runBatchPath(jobs, opts, rawByKey) {
         : null;
     const transport = opts.batchCallModel ?? batchCallModel;
     const results = await transport(batchJobs, {
-        baseUrl: opts.baseUrl,
-        authToken: opts.authToken,
+        baseUrl: opts.batchBaseUrl ?? opts.baseUrl,
+        authToken: opts.batchAuthToken ?? opts.authToken,
+        modelOverride: opts.batchModel,
         maxTokens: opts.maxTokens,
         resumeBatchId: existing?.batchId,
         onSubmitted: (batchId, p) => {

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

@@ -56,6 +56,12 @@ export interface EvolveOpts {
     concurrency?: number;
     /** Use provider batch API instead of online calls. 50% cheaper, slower wall-clock. */
     batch?: boolean;
+    /** Override base URL for batch submissions only. */
+    batchBaseUrl?: string;
+    /** Override auth token for batch submissions only. */
+    batchAuthToken?: string;
+    /** Override model for batch submissions (e.g. kimi-k2.6 when online uses kimi-for-coding). */
+    batchModel?: string;
     /** Adaptive sampling cap (opt-in). Keeps adding reps to noisy cells up to this count. */
     adaptiveReps?: {
         cap: number;

package/dist/prompt-evolution/index.js

@@ -74,6 +74,9 @@ export async function evolvePrompt(opts) {
             repetitions: opts.repetitions,
             judge: opts.judge,
             batch: opts.batch,
+            batchBaseUrl: opts.batchBaseUrl,
+            batchAuthToken: opts.batchAuthToken,
+            batchModel: opts.batchModel,
             adaptiveReps: opts.adaptiveReps,
             runId,
             generation: gen,
@@ -220,6 +223,9 @@ export async function evolvePrompt(opts) {
             concurrency: opts.concurrency ?? 8,
             repetitions: opts.repetitions,
             batch: opts.batch,
+            batchBaseUrl: opts.batchBaseUrl,
+            batchAuthToken: opts.batchAuthToken,
+            batchModel: opts.batchModel,
             adaptiveReps: opts.adaptiveReps,
             runId,
             generation: generations + 1,

package/dist/prompt-evolution/transport-batch.d.ts

@@ -12,7 +12,7 @@
  *   - openrouter → NO batch support; throws (caller must fall back to online)
  *
  * Custom IDs route results back to the right (variant, case, model, rep)
- * cell. The evaluator builds ids like `v0:h_abc:kimi-k2-6:r0`.
+ * cell. The evaluator builds ids like `v0:h_abc:kimi-for-coding:r0`.
  *
  * Poll state is persisted via `persistBatchState` so a crashed or
  * restarted run can resume without resubmitting.
@@ -27,6 +27,14 @@ export interface BatchJob {
 export interface BatchOpts {
     baseUrl?: string;
     authToken?: string;
+    /**
+     * Override model for the batch submission. Moonshot's batch API only
+     * accepts kimi-k2.5 or kimi-k2.6 — NOT the kimi-for-coding alias that the
+     * coding endpoint uses. When batch is enabled against a Kimi stack, set
+     * this so online eval keeps using kimi-for-coding while batch uses the
+     * concrete version.
+     */
+    modelOverride?: string;
     maxTokens?: number;
     /** Poll interval starts here and doubles to `pollMaxMs`. Defaults 30s → 5min. */
     pollStartMs?: number;

package/dist/prompt-evolution/transport-batch.js

@@ -12,19 +12,30 @@
  *   - openrouter → NO batch support; throws (caller must fall back to online)
  *
  * Custom IDs route results back to the right (variant, case, model, rep)
- * cell. The evaluator builds ids like `v0:h_abc:kimi-k2-6:r0`.
+ * cell. The evaluator builds ids like `v0:h_abc:kimi-for-coding:r0`.
  *
  * Poll state is persisted via `persistBatchState` so a crashed or
  * restarted run can resume without resubmitting.
  */
+import { VERSION } from "../core/_version.js";
+const USER_AGENT = `claude-overnight-evolve/${VERSION}`;
 export function detectBatchProvider(baseUrl) {
     const url = (baseUrl ?? "https://api.anthropic.com").toLowerCase();
     if (/(^|\/\/)(api\.)?anthropic\.com/.test(url))
         return "anthropic";
+    // Providers with no batch support — caller auto-falls back to online.
+    // - OpenRouter: no batch API at all.
+    // - api.kimi.com/coding: Moonshot's coding-specific endpoint; synchronous
+    //   only (30 concurrent, 300-1200 req/5hr) with no /v1/files upload flow.
+    //   Moonshot's generic platform.moonshot.ai might have batch; this one
+    //   doesn't.
     if (/openrouter/.test(url))
         return "unsupported";
-    // Everything else that speaks /v1/chat/completions — OpenAI, Kimi, Moonshot,
-    // DeepSeek — exposes an OpenAI-compatible batch endpoint.
+    if (/(api\.)?kimi\.com\/coding/.test(url))
+        return "unsupported";
+    // Everything else that speaks /v1/chat/completions — OpenAI, DeepSeek,
+    // DashScope in OpenAI-compat mode — exposes an OpenAI-compatible batch
+    // endpoint we can ride.
     return "openai-compatible";
 }
 export async function batchCallModel(jobs, opts) {
@@ -32,7 +43,11 @@ export async function batchCallModel(jobs, opts) {
         return new Map();
     const provider = detectBatchProvider(opts.baseUrl);
     if (provider === "unsupported") {
-        throw new Error(`Batch API not supported for baseUrl=${opts.baseUrl}; use online transport`);
+        throw new Error(`Batch API not supported for baseUrl=${opts.baseUrl}. ` +
+            `Options: (1) omit --batch and use online transport, or (2) point ` +
+            `the batch call at a provider with batch support (e.g. set --batch-base-url ` +
+            `https://api.moonshot.ai/v1 --batch-model kimi-k2.6 for Kimi users whose ` +
+            `online endpoint is api.kimi.com/coding).`);
     }
     if (provider === "anthropic")
         return runAnthropicBatch(jobs, opts);
@@ -45,6 +60,7 @@ async function runAnthropicBatch(jobs, opts) {
     const headers = {
         "Content-Type": "application/json",
         "Authorization": `Bearer ${authToken}`,
+        "User-Agent": USER_AGENT,
         "anthropic-version": "2023-06-01",
         "anthropic-beta": "message-batches-2024-09-24",
     };
@@ -53,7 +69,7 @@ async function runAnthropicBatch(jobs, opts) {
         const body = JSON.stringify({
             requests: jobs.map((j) => {
                 const params = {
-                    model: j.model,
+                    model: opts.modelOverride ?? j.model,
                     max_tokens: opts.maxTokens ?? 4096,
                     messages: [{ role: "user", content: j.userText }],
                 };
@@ -115,7 +131,10 @@ async function runAnthropicBatch(jobs, opts) {
 async function runOpenAIBatch(jobs, opts) {
     const baseUrl = (opts.baseUrl ?? "https://api.openai.com").replace(/\/$/, "");
     const authToken = opts.authToken ?? process.env.ANTHROPIC_AUTH_TOKEN ?? process.env.ANTHROPIC_API_KEY ?? "";
-    const authHeaders = { "Authorization": `Bearer ${authToken}` };
+    const authHeaders = {
+        "Authorization": `Bearer ${authToken}`,
+        "User-Agent": USER_AGENT,
+    };
     let batchId = opts.resumeBatchId;
     let outputFileId;
     if (!batchId) {
@@ -129,7 +148,7 @@ async function runOpenAIBatch(jobs, opts) {
                 custom_id: j.customId,
                 method: "POST",
                 url: "/v1/chat/completions",
-                body: { model: j.model, max_tokens: opts.maxTokens ?? 4096, messages },
+                body: { model: opts.modelOverride ?? j.model, max_tokens: opts.maxTokens ?? 4096, max_completion_tokens: opts.maxTokens ?? 4096, messages },
             });
         }).join("\n");
         const form = new FormData();

package/dist/prompt-evolution/transport.js

@@ -8,17 +8,22 @@
  * Supports both Anthropic-native and OpenAI-compatible endpoints so we can
  * run the same eval against Haiku, Kimi, and OpenRouter without a rewrite.
  */
+import { VERSION } from "../core/_version.js";
+const USER_AGENT = `claude-overnight-evolve/${VERSION}`;
 export async function defaultCallModel(userText, systemText, opts) {
     const baseUrl = (opts.baseUrl ?? process.env.ANTHROPIC_BASE_URL ?? "https://api.anthropic.com").replace(/\/$/, "");
     const authToken = opts.authToken ?? process.env.ANTHROPIC_AUTH_TOKEN ?? process.env.ANTHROPIC_API_KEY ?? "";
     const isAnthropic = /^https?:\/\/(api\.)?anthropic\.com/i.test(baseUrl);
-    const isKimi = /kimi\.com/i.test(baseUrl);
+    // Identify ourselves honestly. Kimi's coding-endpoint docs explicitly say
+    // "Tampering with the client identifier (User-Agent) is considered a
+    // violation." The previous "Kilo-Code/1.0" was impersonating a third-party
+    // tool; we now send our real binary name + version.
     const headers = {
         "Content-Type": "application/json",
         "Authorization": `Bearer ${authToken}`,
+        "User-Agent": USER_AGENT,
     };
-    if (isKimi)
-        headers["User-Agent"] = "Kilo-Code/1.0";
+    const maxOut = opts.maxTokens ?? 4096;
     let endpoint;
     let body;
     if (isAnthropic) {
@@ -26,7 +31,7 @@ export async function defaultCallModel(userText, systemText, opts) {
         headers["anthropic-version"] = "2023-06-01";
         const payload = {
             model: opts.model,
-            max_tokens: opts.maxTokens ?? 4096,
+            max_tokens: maxOut, // Anthropic uses max_tokens, not max_completion_tokens.
             messages: [{ role: "user", content: userText }],
         };
         if (systemText)
@@ -39,9 +44,14 @@ export async function defaultCallModel(userText, systemText, opts) {
         if (systemText)
             messages.push({ role: "system", content: systemText });
         messages.push({ role: "user", content: userText });
+        // Platform.moonshot.ai marks max_tokens deprecated in favor of
+        // max_completion_tokens. Kimi's coding endpoint still accepts max_tokens.
+        // Sending both is safe — OpenAI, Moonshot, DeepSeek, and Kimi all tolerate
+        // the extra field, and we're future-proof against the deprecation.
         body = JSON.stringify({
             model: opts.model,
-            max_tokens: opts.maxTokens ?? 4096,
+            max_tokens: maxOut,
+            max_completion_tokens: maxOut,
             messages,
         });
     }

package/docs/prompt-evolution-research.md

@@ -183,7 +183,7 @@ Your laptop can be off the whole time.
 npm run evolve -- --prompt 10_planning/10-3_plan --eval-model claude-haiku-4-5 --generations 10
 
 # Evolve an MCP-browser supervision prompt
-npm run evolve -- --target mcp-browser --prompt-kind plan-supervision --eval-model kimi-k2-6 --generations 10
+npm run evolve -- --target mcp-browser --prompt-kind plan-supervision --eval-model kimi-for-coding --generations 10
 ```
 
 ### Via Platform API (runs on server)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-overnight",
-  "version": "1.57.4",
+  "version": "1.58.0",
   "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": "1.57.4",
+  "version": "1.58.0",
   "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"