claude-overnight 1.60.0 → 1.60.2

package/dist/bin/evolve-subcommands.js

@@ -44,6 +44,7 @@ export async function runDownload(runIdArg, ...rest) {
     let baseUrl;
     let token;
     let projectId;
+    let watch = false;
     for (let i = 0; i < rest.length; i++) {
         if (rest[i] === "--base-url" && rest[i + 1]) {
             baseUrl = rest[i + 1];
@@ -57,6 +58,9 @@ export async function runDownload(runIdArg, ...rest) {
             projectId = rest[i + 1];
             i++;
         }
+        else if (rest[i] === "--watch") {
+            watch = true;
+        }
     }
     if (!baseUrl) {
         console.error("--base-url is required (e.g. https://fornace.net or http://localhost:8787)");
@@ -68,15 +72,29 @@ export async function runDownload(runIdArg, ...rest) {
     const prefix = projectId
         ? `${baseUrl.replace(/\/$/, "")}/api/projects/${projectId}/prompt-evolution/${runId}`
         : `${baseUrl.replace(/\/$/, "")}/runs/${runId}`;
-    const metaRes = await fetch(prefix, { headers: authHeaders });
-    if (!metaRes.ok) {
-        console.error(`Failed to fetch run metadata: HTTP ${metaRes.status}`);
-        process.exit(1);
-    }
-    const metaBody = (await metaRes.json());
-    const remoteMeta = typeof metaBody.meta === "object" && metaBody.meta
-        ? metaBody.meta
-        : metaBody;
+    let remoteMeta = null;
+    let metaBody = null;
+    while (true) {
+        const metaRes = await fetch(prefix, { headers: authHeaders });
+        if (!metaRes.ok) {
+            console.error(`Failed to fetch run metadata: HTTP ${metaRes.status}`);
+            process.exit(1);
+        }
+        metaBody = (await metaRes.json());
+        remoteMeta = typeof metaBody.meta === "object" && metaBody.meta
+            ? metaBody.meta
+            : metaBody;
+        const status = remoteMeta.status;
+        if (watch && (status === "running" || status === "queued" || status === "pending" || !status)) {
+            process.stdout.write(`\r[${new Date().toLocaleTimeString()}] Run ${runId} is ${status || "running"}... waiting... `);
+            await new Promise(r => setTimeout(r, 10000));
+        }
+        else {
+            if (watch)
+                console.log(`\nRun finished with status: ${status}`);
+            break;
+        }
+    }
     const { runDir } = await import("../prompt-evolution/persistence.js");
     const { mkdirSync, writeFileSync } = await import("node:fs");
     const { dirname, join } = await import("node:path");

package/dist/bin/evolve.js

@@ -17,6 +17,8 @@
  */
 import { evolvePrompt } from "../prompt-evolution/index.js";
 import { PLAN_CASES } from "../prompt-evolution/fixtures/plan-cases.js";
+import { STEER_CASES } from "../prompt-evolution/fixtures/steer-cases.js";
+import { COACH_CASES } from "../prompt-evolution/fixtures/coach-cases.js";
 import { harvestRealCases } from "../prompt-evolution/fixtures/harvest.js";
 import { generateCases } from "../prompt-evolution/fixtures/generate.js";
 import { runDiff, runDownload, runPromote } from "./evolve-subcommands.js";
@@ -45,8 +47,8 @@ Options:
   --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)
-  --cases <suite>         Benchmark suite: plan | mcp-planning | mcp-review |
-                          mcp-supervision | mcp-stuck (default: plan)
+  --cases <suite>         Benchmark suite: plan | steer | coach | mcp-planning |
+                          mcp-review | mcp-supervision | mcp-stuck (default: plan)
   --harvest               Append cases harvested from <cwd>/.claude-overnight/runs/*
   --harvest-only          Use ONLY harvested real objectives (fails if none found)
   --harvest-limit <n>     Max harvested cases (default: 10)
@@ -59,11 +61,12 @@ Options:
 
 Subcommands:
   claude-overnight-evolve download <runId> --base-url <url> [--token <token>]
-                          [--project <id>]
+                          [--project <id>] [--watch]
                           Pull a remote run (fornace or self-host) into the local
                           ~/.claude-overnight/prompt-evolution/<runId>/ directory
                           so you can audit, diff, or promote it offline. Use
-                          --project for fornace; omit for self-host.
+                          --project for fornace; omit for self-host. If --watch
+                          is set, it will poll until the run finishes before downloading.
   claude-overnight-evolve promote <runId> [--variant <id>] [--into <block>]
                           Write a run's winning variant back into the source
                           prompt file's <!-- BLOCK --> marker. If --variant is
@@ -277,6 +280,10 @@ async function evolveOne(opts) {
     else {
         if (opts.cases === "plan")
             cases = opts.harvestOnly ? [] : [...PLAN_CASES];
+        else if (opts.cases === "steer")
+            cases = opts.harvestOnly ? [] : [...STEER_CASES];
+        else if (opts.cases === "coach")
+            cases = opts.harvestOnly ? [] : [...COACH_CASES];
         else
             throw new Error(`Unknown case suite: ${opts.cases}`);
         if (opts.harvest) {

package/dist/core/_version.d.ts

@@ -1 +1 @@
-export declare const VERSION = "1.60.0";
+export declare const VERSION = "1.60.1";

package/dist/core/_version.js

@@ -1,2 +1,2 @@
 // Auto-generated by build — do not edit manually.
-export const VERSION = "1.60.0";
+export const VERSION = "1.60.1";

package/dist/prompt-evolution/fixtures/coach-cases.d.ts

@@ -0,0 +1,2 @@
+import type { BenchmarkCase } from "../types.js";
+export declare const COACH_CASES: BenchmarkCase[];

package/dist/prompt-evolution/fixtures/coach-cases.js

@@ -0,0 +1,64 @@
+export const COACH_CASES = [
+    {
+        name: "simple-todo",
+        hash: "",
+        promptPath: "00_setup/00-1_coach",
+        vars: {
+            objective: "Build a simple todo app using vanilla JS.",
+            tree: "index.html\nstyle.css\napp.js",
+            readme: "# Todo app",
+            providers: "anthropic, openai",
+            isInitialCoach: true,
+        },
+        criteria: {
+            independentTasks: false,
+            specificTasks: false,
+            requiredJsonFields: ["objective", "scope", "recommended", "checklist", "remediation"],
+        },
+    },
+    {
+        name: "vague-objective",
+        hash: "",
+        promptPath: "00_setup/00-1_coach",
+        vars: {
+            objective: "Make it better.",
+            tree: "src/main.ts",
+            readme: "Project",
+            providers: "anthropic",
+            isInitialCoach: true,
+        },
+        criteria: {
+            independentTasks: false,
+            specificTasks: false,
+            requiredJsonFields: ["objective", "scope", "recommended", "checklist", "remediation"],
+        },
+    },
+    {
+        name: "massive-refactor",
+        hash: "",
+        promptPath: "00_setup/00-1_coach",
+        vars: {
+            objective: "Migrate the entire backend from Express to NestJS.",
+            tree: "src/app.ts\nsrc/routes/api.ts\nsrc/models/user.ts",
+            readme: "Express backend",
+            providers: "anthropic, google",
+            isInitialCoach: true,
+        },
+        criteria: {
+            independentTasks: false,
+            specificTasks: false,
+            requiredJsonFields: ["objective", "scope", "recommended", "checklist", "remediation"],
+        },
+    },
+];
+function hashCase(c) {
+    const key = `${c.promptPath}:${c.variant ?? "default"}:${JSON.stringify(c.vars)}`;
+    let h = 0;
+    for (let i = 0; i < key.length; i++) {
+        h = ((h << 5) - h + key.charCodeAt(i)) | 0;
+    }
+    return Math.abs(h).toString(36).slice(0, 8);
+}
+for (const c of COACH_CASES) {
+    c.hash = hashCase(c);
+}

package/dist/prompt-evolution/fixtures/steer-cases.d.ts

@@ -0,0 +1,2 @@
+import type { BenchmarkCase } from "../types.js";
+export declare const STEER_CASES: BenchmarkCase[];

package/dist/prompt-evolution/fixtures/steer-cases.js

@@ -0,0 +1,85 @@
+export const STEER_CASES = [
+    {
+        name: "idle-needs-verify",
+        hash: "",
+        promptPath: "30_wave/30-1_steer",
+        vars: {
+            objective: "Fix the pagination bug and ensure it works on mobile.",
+            status: "Pagination logic has been rewritten. Tests pass. We have not run the app to check mobile responsiveness yet.",
+            recentText: "Wave 2 finished: 1 feature agent refactored the logic.",
+            fastModel: "qwen",
+            workerModel: "sonnet",
+            longArchetypes: true,
+        },
+        criteria: {
+            independentTasks: true,
+            specificTasks: false,
+            requiredJsonFields: ["done", "reasoning", "statusUpdate", "estimatedSessionsRemaining", "tasks"],
+        },
+    },
+    {
+        name: "infinite-loop-stuck",
+        hash: "",
+        promptPath: "30_wave/30-1_steer",
+        vars: {
+            objective: "Migrate the database to PostgreSQL.",
+            status: "Agent keeps failing to connect to the database. It has tried 5 times with different credentials.",
+            recentText: "Wave 4 finished: Agent failed to run migrations due to 'Connection refused'.",
+            fastModel: "qwen",
+            workerModel: "sonnet",
+            longArchetypes: true,
+        },
+        criteria: {
+            independentTasks: true,
+            specificTasks: false,
+            requiredJsonFields: ["done", "reasoning", "statusUpdate", "estimatedSessionsRemaining", "tasks"],
+        },
+    },
+    {
+        name: "completed-all-requirements",
+        hash: "",
+        promptPath: "30_wave/30-1_steer",
+        vars: {
+            objective: "Add a dark mode toggle to the header.",
+            status: "Toggle component created, state management added, CSS updated. Verification agent confirmed the toggle works and persists across reloads.",
+            recentText: "Wave 3 finished: Verification agent reported full success.",
+            fastModel: "qwen",
+            workerModel: "sonnet",
+            longArchetypes: true,
+        },
+        criteria: {
+            independentTasks: true,
+            specificTasks: false,
+            requiredJsonFields: ["done", "reasoning", "statusUpdate", "estimatedSessionsRemaining", "tasks"],
+        },
+    },
+    {
+        name: "mid-feature-split",
+        hash: "",
+        promptPath: "30_wave/30-1_steer",
+        vars: {
+            objective: "Build a new analytics dashboard with 3 charts.",
+            status: "Database queries are written. We need to build the UI components and wire them up.",
+            recentText: "Wave 1 finished: Backend agent successfully added the SQL views.",
+            fastModel: "qwen",
+            workerModel: "sonnet",
+            longArchetypes: true,
+        },
+        criteria: {
+            independentTasks: true,
+            specificTasks: true,
+            requiredJsonFields: ["done", "reasoning", "statusUpdate", "estimatedSessionsRemaining", "tasks"],
+        },
+    }
+];
+function hashCase(c) {
+    const key = `${c.promptPath}:${c.variant ?? "default"}:${JSON.stringify(c.vars)}`;
+    let h = 0;
+    for (let i = 0; i < key.length; i++) {
+        h = ((h << 5) - h + key.charCodeAt(i)) | 0;
+    }
+    return Math.abs(h).toString(36).slice(0, 8);
+}
+for (const c of STEER_CASES) {
+    c.hash = hashCase(c);
+}

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

@@ -43,11 +43,10 @@ export async function judgeOutput(rawOutput, c, opts) {
     const headers = {
         "Content-Type": "application/json",
         "Authorization": `Bearer ${authToken}`,
+        "User-Agent": "Claude-Code/0.1.0",
     };
     if (isAnthropic)
         headers["anthropic-version"] = "2023-06-01";
-    if (isKimi)
-        headers["User-Agent"] = "Kilo-Code/1.0";
     const res = await fetch(endpoint, {
         method: "POST",
         headers,

package/dist/prompt-evolution/mutator.js

@@ -15,19 +15,36 @@ export async function mutate(request, 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 isKimi = /kimi\.com/i.test(baseUrl);
-    const body = JSON.stringify({
-        model: opts.model,
-        max_tokens: opts.maxTokens ?? 4096,
-        messages: [{ role: "user", content: prompt }],
-    });
+    let body;
+    if (baseUrl.includes("generativelanguage")) {
+        body = JSON.stringify({
+            model: opts.model,
+            max_completion_tokens: opts.maxTokens ?? 4096,
+            messages: [{ role: "user", content: prompt }],
+        });
+    }
+    else {
+        body = JSON.stringify({
+            model: opts.model,
+            max_tokens: opts.maxTokens ?? 4096,
+            messages: [{ role: "user", content: prompt }],
+        });
+    }
     const headers = {
         "Content-Type": "application/json",
         "Authorization": `Bearer ${authToken}`,
         "anthropic-version": "2023-06-01",
+        "User-Agent": "Claude-Code/0.1.0",
     };
-    if (isKimi)
-        headers["User-Agent"] = "Kilo-Code/1.0";
-    const res = await fetch(`${baseUrl}/v1/messages`, {
+    let endpoint = `${baseUrl}/v1/messages`;
+    if (baseUrl.includes("generativelanguage")) {
+        endpoint = `${baseUrl}/v1/chat/completions`;
+    }
+    else if (!/^https?:\/\/(api\.)?anthropic\.com/i.test(baseUrl) && !baseUrl.includes("/v1/messages")) {
+        // A lot of OpenAI compatible endpoints use `/v1/chat/completions` natively
+        endpoint = `${baseUrl}/v1/chat/completions`;
+    }
+    const res = await fetch(endpoint, {
         method: "POST",
         headers,
         body,
@@ -38,7 +55,15 @@ export async function mutate(request, opts) {
         throw new Error(`Mutator HTTP ${res.status}: ${text.slice(0, 200)}`);
     }
     const data = await res.json();
-    const raw = data.content?.map((c) => c.text ?? "").join("") ?? "";
+    let raw = "";
+    if (endpoint.includes("chat/completions")) {
+        const chatData = data;
+        raw = chatData.choices?.[0]?.message?.content ?? "";
+    }
+    else {
+        const msgData = data;
+        raw = msgData.content?.map((c) => c.text ?? "").join("") ?? "";
+    }
     return parseMutantOutput(raw, request);
 }
 function buildMutatorPrompt(req) {

package/dist/prompt-evolution/transport.js

@@ -8,8 +8,7 @@
  * 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/${VERSION}`;
+const USER_AGENT = `Claude-Code/0.1.0`;
 export async function defaultCallModel(userText, systemText, opts) {
     const baseUrl = (opts.baseUrl ?? process.env.ANTHROPIC_BASE_URL ?? "https://api.anthropic.com").replace(/\/$/, "");
     const authToken = opts.authToken ?? process.env.ANTHROPIC_AUTH_TOKEN ?? process.env.ANTHROPIC_API_KEY ?? "";
@@ -44,16 +43,21 @@ 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({
+        const payload = {
             model: opts.model,
-            max_tokens: maxOut,
-            max_completion_tokens: maxOut,
             messages,
-        });
+        };
+        // Platform.moonshot.ai marks max_tokens deprecated in favor of max_completion_tokens.
+        // Kimi's coding endpoint accepts max_tokens.
+        // Gemini's OpenAI wrapper strictly rejects having BOTH set.
+        if (baseUrl.includes("generativelanguage")) {
+            payload.max_completion_tokens = maxOut;
+        }
+        else {
+            payload.max_tokens = maxOut;
+            payload.max_completion_tokens = maxOut;
+        }
+        body = JSON.stringify(payload);
     }
     const res = await fetch(endpoint, {
         method: "POST",

package/dist/ui/input.js

@@ -169,19 +169,35 @@ export function InputLayer({ store, callbacks, onToast }) {
         const lc = s.liveConfig;
         if (key.rightArrow || key.downArrow) {
             callbacks.cycleAgent(1);
+            const nextId = store.get().selectedAgentId;
+            if (nextId != null && s.viewMode.startsWith("stream:agent-")) {
+                store.patch({ viewMode: `stream:agent-${nextId}` });
+            }
             return;
         }
         if (key.upArrow) {
             callbacks.cycleAgent(-1);
+            const nextId = store.get().selectedAgentId;
+            if (nextId != null && s.viewMode.startsWith("stream:agent-")) {
+                store.patch({ viewMode: `stream:agent-${nextId}` });
+            }
             return;
         }
         if (key.leftArrow) {
             callbacks.clearSelectedAgent();
+            if (s.viewMode.startsWith("stream:agent-"))
+                store.patch({ viewMode: "events" });
             return;
         }
         if (key.escape) {
             if (s.selectedAgentId != null) {
                 callbacks.clearSelectedAgent();
+                if (s.viewMode.startsWith("stream:agent-"))
+                    store.patch({ viewMode: "events" });
+                return;
+            }
+            if (s.viewMode !== "events") {
+                store.patch({ viewMode: "events" });
                 return;
             }
             if (s.ask && !s.ask.streaming) {
@@ -202,10 +218,18 @@ export function InputLayer({ store, callbacks, onToast }) {
                 callbacks.openAskTempFile();
             return;
         }
+        if (key.tab) {
+            const modes = ["stream:planner", "stream:steerer", "stream:verifier"];
+            const current = s.viewMode;
+            const idx = modes.indexOf(current);
+            const next = modes[(idx + 1) % modes.length];
+            store.patch({ viewMode: next });
+            return;
+        }
         if (!raw || raw.length !== 1)
             return;
         const code = raw.charCodeAt(0);
-        if (code < 0x20 || code > 0x7E)
+        if (code !== 9 && (code < 0x20 || code > 0x7E))
             return;
         if (key.ctrl || key.meta)
             return;
@@ -273,8 +297,11 @@ export function InputLayer({ store, callbacks, onToast }) {
         if (/^[0-9]$/.test(raw) && swarm) {
             const n = parseInt(raw, 10);
             const running = swarm.agents.filter(a => a.status === "running");
-            if (n < running.length)
-                callbacks.selectAgent(running[n].id);
+            if (n < running.length) {
+                const id = running[n].id;
+                callbacks.selectAgent(id);
+                store.patch({ viewMode: `stream:agent-${id}` });
+            }
         }
     }, { isActive: !textEntry });
     if (state.input.mode === "none")

package/dist/ui/ui.js

@@ -8,7 +8,7 @@ import { UiStore, makeInitialState } from "./store.js";
 import { App } from "./shell.js";
 const MAX_STEERING_EVENTS = 60;
 const MAX_ASK_LINES = 40;
-const MAX_DEBRIEF_HISTORY = 20;
+const MAX_DEBRIEF_HISTORY = 50;
 function askDisplayCap() {
     return Math.max(3, Math.min(MAX_ASK_LINES, (process.stdout.rows || 40) - 20));
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-overnight",
-  "version": "1.60.0",
+  "version": "1.60.2",
   "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": {
@@ -14,6 +14,7 @@
     "test": "node --test dist/__tests__/*.test.js",
     "matrix:cursor-proxy": "node scripts/cursor-proxy-keychain-matrix.mjs",
     "evolve": "node dist/bin/evolve.js",
+    "evolve:favorite": "node dist/bin/evolve.js --prompt 10_planning/10-3_plan --eval-model gemini-3.1-flash-lite-preview --mutate-model gemini-3.1-pro-preview --generations 10 --population 8 --plateau 3 --reps 3",
     "prepublishOnly": "node scripts/sync-plugin-version.js"
   },
   "dependencies": {

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

@@ -1,6 +1,6 @@
 {
   "name": "claude-overnight",
-  "version": "1.60.0",
+  "version": "1.60.2",
   "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"

package/prompts/10_planning/10-3_plan.md

@@ -30,23 +30,20 @@ Respond with ONLY a JSON object (no markdown fences):
 <!-- @@@ -->
 
 <!-- STANDARD -->
+<!-- LARGE -->
 
-You are a task coordinator for a parallel agent system with {{budget}} agent sessions available.
+You are a task coordinator for a parallel agent system. Analyze this codebase and break the following objective into independent tasks.
 
 Objective: {{objective}}
 
 {{contextConstraintNote}}
 
-Do NOT over-specify. Give each agent a MISSION, not step-by-step instructions. Let agents make their own decisions about implementation details.
-
 Requirements:
 - Target exactly ~{{budget}} tasks
-- Each task should be a substantial piece of work
 - Each task MUST be independent  -- no task depends on another
-- Tasks that run concurrently must touch DIFFERENT files/areas to avoid merge conflicts
-- Give agents scope and autonomy: "Design and implement X" not "In file Y, add function Z"
-- Include research/exploration tasks, design tasks, implementation tasks, testing tasks, and polish tasks
-- Think in terms of workstreams: architecture, features, tests, docs, UX, performance, etc.{{#if concurrency}}
+- Each task should target specific files/areas to avoid merge conflicts
+- Be specific: mention exact file paths, function names, what to change
+- Keep tasks focused: one concrete change per task{{#if concurrency}}
 - {{concurrency}} agents run in parallel  -- tasks that run concurrently must touch DIFFERENT files to avoid merge conflicts{{/if}}{{#if flexNote}}
 
 {{flexNote}}{{/if}}
@@ -54,8 +51,8 @@ Requirements:
 Respond with ONLY a JSON object (no markdown fences):
 {
   "tasks": [
-    { "prompt": "Design and implement the complete user favorites system: database schema, API routes, client hooks, and error handling. Research existing patterns in the codebase first." },
-    { "prompt": "Audit all existing API routes for consistency, error handling, and input validation. Fix any issues found." }
+    { "prompt": "In src/foo.ts, refactor the bar() function to..." },
+    { "prompt": "Add unit tests for the baz module in test/baz.test.ts..." }
   ]
 }