claude-overnight 1.50.3 → 1.50.6

package/dist/core/_version.d.ts

@@ -1 +1 @@
-export declare const VERSION = "1.50.3";
+export declare const VERSION = "1.50.6";

package/dist/core/_version.js

@@ -1,2 +1,2 @@
 // Auto-generated by build — do not edit manually.
-export const VERSION = "1.50.3";
+export const VERSION = "1.50.6";

package/dist/core/types.d.ts

@@ -310,6 +310,10 @@ export interface RunState extends RunConfigBase {
     coachedObjective?: string;
     /** Unix timestamp (ms) when the coach produced the accepted rewrite. */
     coachedAt?: number;
+    /** Run-branch created when mergeStrategy="branch" — must survive resume so subsequent waves accumulate into it instead of orphan branches. */
+    runBranch?: string;
+    /** Branch (or detached HEAD sha) the user was on before the run-branch was created — restored on run completion. */
+    originalRef?: string;
 }
 /** Function that returns a rate-limit snapshot with optional context token info. */
 export type RLGetter = () => {

package/dist/planner/query.js

@@ -21,10 +21,16 @@ let _envResolver;
 export function setPlannerEnvResolver(fn) {
     _envResolver = fn;
 }
-// ── Cursor proxy: direct HTTP bypass ──
-// SDK spawns 4+ subprocesses (~15s each) for the proxy; one direct POST is 4-10x faster.
-function isCursorProxyEnv(env) {
-    return !!env?.CURSOR_API_KEY && !env?.ANTHROPIC_API_KEY;
+// ── Direct HTTP bypass for non-Anthropic endpoints ──
+// The Claude Code CLI subprocess spawned by @anthropic-ai/claude-agent-sdk validates
+// model names against its built-in Anthropic list and rejects custom ids (qwen3.6-plus,
+// composer-2, etc.) pre-flight, even when ANTHROPIC_BASE_URL points at an Anthropic-
+// compatible proxy. Bypass the SDK with a direct POST for any non-anthropic.com base.
+function shouldUseDirectFetch(env) {
+    const base = env?.ANTHROPIC_BASE_URL;
+    if (!base)
+        return false;
+    return !/^https?:\/\/(api\.)?anthropic\.com/i.test(base);
 }
 async function runViaDirectFetch(prompt, opts, onLog) {
     const env = opts.env ?? _envResolver?.(opts.model);
@@ -39,30 +45,34 @@ async function runViaDirectFetch(prompt, opts, onLog) {
         await apiEndpointLimiter.waitIfNeeded();
         const waited = await rl.waitIfNeeded();
         if (waited > 0)
-            onLog(`Cursor proxy rate gate — waited ${Math.round(waited / 1000)}s`, "event");
+            onLog(`Planner proxy rate gate — waited ${Math.round(waited / 1000)}s`, "event");
         const res = await fetch(`${baseUrl}/v1/messages`, {
             method: "POST",
-            headers: { "Content-Type": "application/json", "Authorization": `Bearer ${authToken}` },
+            headers: {
+                "Content-Type": "application/json",
+                "Authorization": `Bearer ${authToken}`,
+                "anthropic-version": "2023-06-01",
+            },
             body: JSON.stringify({ model: opts.model, max_tokens: 8192, messages: [{ role: "user", content: prompt }] }),
         });
         if (res.status === 429 && attempt < MAX_RETRIES) {
             const waitMs = BACKOFF[attempt];
-            onLog(`Cursor proxy rate limited — waiting ${Math.round(waitMs / 1000)}s`, "event");
+            onLog(`Planner proxy rate limited — waiting ${Math.round(waitMs / 1000)}s`, "event");
             await sleep(waitMs);
             continue;
         }
         if (!res.ok)
-            throw new Error(`Cursor proxy ${res.status}: ${(await res.text().catch(() => ""))}`);
+            throw new Error(`Planner proxy ${res.status}: ${(await res.text().catch(() => ""))}`);
         rl.record();
         apiEndpointLimiter.record();
         const data = await res.json();
         return data.content?.[0]?.text ?? "";
     }
-    throw new Error("Cursor proxy direct fetch failed after retries");
+    throw new Error("Planner proxy direct fetch failed after retries");
 }
 export async function runPlannerQuery(prompt, opts, onLog) {
     const env = opts.env ?? _envResolver?.(opts.model);
-    if (isCursorProxyEnv(env))
+    if (shouldUseDirectFetch(env))
         return runViaDirectFetch(prompt, opts, onLog);
     const MAX_RETRIES = 3;
     const BACKOFF = [30_000, 60_000, 120_000];

package/dist/planner/steering.js

@@ -27,7 +27,11 @@ export const STEER_SCHEMA = {
         required: ["done", "tasks", "reasoning", "statusUpdate", "estimatedSessionsRemaining"],
     },
 };
-const PROMPT_BUDGET = 6000;
+// The base 30-1_steer template alone is ~7 KB, so any budget below that is
+// unreachable no matter how aggressively we trim variables. 20 KB leaves room
+// for the template + moderate run memory while still being a tiny fraction of
+// any planner's context window.
+const PROMPT_BUDGET = 20_000;
 const DEFAULT_CAPS = {
     milestones: 2000, designs: 1500, reflections: 1000,
     verifications: 1000, previousRuns: 800, userGuidance: 4000,

package/dist/providers/index.js

@@ -109,9 +109,6 @@ export function envFor(p) {
         base.ANTHROPIC_AUTH_TOKEN = key;
     }
     delete base.ANTHROPIC_API_KEY;
-    // Prevent CURSOR_API_KEY from leaking into non-proxy envs — would trip
-    // isCursorProxyEnv false-positive, silently rerouting through direct fetch
-    // which ignores outputFormat (no JSON schema enforcement).
     delete base.CURSOR_API_KEY;
     delete base.CURSOR_AUTH_TOKEN;
     return base;

package/dist/run/run.js

@@ -207,10 +207,25 @@ export async function executeRun(cfg) {
             .then(text => { display.setDebrief(text.trim().slice(0, 210), label); })
             .catch(() => { display.setDebrief(undefined); });
     };
-    // For flex + branch strategy: create one target branch
+    // For flex + branch strategy: create one target branch (or restore on resume).
+    // The run-branch + originalRef are persisted in run.json so resumes accumulate
+    // into the original branch instead of spawning orphan swarm/run-* branches.
     let runBranch;
     let originalRef;
-    if (flex && mergeStrategy === "branch" && useWorktrees && !cfg.resuming) {
+    if (cfg.resuming && cfg.resumeState) {
+        runBranch = cfg.resumeState.runBranch;
+        originalRef = cfg.resumeState.originalRef;
+        if (runBranch) {
+            try {
+                execSync(`git checkout "${runBranch}"`, { cwd, encoding: "utf-8", stdio: "pipe" });
+                console.log(chalk.dim(`  Resumed on branch: ${runBranch}\n`));
+            }
+            catch {
+                console.log(chalk.yellow(`  ⚠ Could not check out run branch ${runBranch} — wave merges may diverge\n`));
+            }
+        }
+    }
+    if (flex && mergeStrategy === "branch" && useWorktrees && !runBranch) {
         try {
             originalRef = execSync("git rev-parse --abbrev-ref HEAD", { cwd, encoding: "utf-8", stdio: "pipe" }).trim();
             if (originalRef === "HEAD")
@@ -253,6 +268,8 @@ export async function executeRun(cfg) {
         repoFingerprint,
         coachedObjective: cfg.coachedObjective,
         coachedAt: cfg.coachedAt,
+        runBranch,
+        originalRef,
     };
     const buildRunState = (varying) => composeRunState({ ...runStateBase, workerModel, plannerModel, fastModel, concurrency, usageCap }, { remaining: varying.remaining, waveNum, accCost, accCompleted, accFailed, accIn, accOut, accTools, branches }, { phase: varying.phase, currentTasks: varying.currentTasks });
     const gracefulStop = () => {

package/dist/state/run-state.d.ts

@@ -22,6 +22,8 @@ export interface RunStateBase {
     repoFingerprint: string;
     coachedObjective?: string;
     coachedAt?: number;
+    runBranch?: string;
+    originalRef?: string;
 }
 /** Live counters captured at snapshot time. */
 export interface RunStateLive {

package/dist/state/run-state.js

@@ -26,5 +26,7 @@ export function composeRunState(base, live, varying) {
         repoFingerprint: base.repoFingerprint,
         coachedObjective: base.coachedObjective,
         coachedAt: base.coachedAt,
+        runBranch: base.runBranch,
+        originalRef: base.originalRef,
     };
 }

package/dist/state/state.js

@@ -199,7 +199,12 @@ export function saveRunState(runDir, state) {
 }
 export function loadRunState(runDir) {
     try {
-        return JSON.parse(readFileSync(join(runDir, "run.json"), "utf-8"));
+        const state = JSON.parse(readFileSync(join(runDir, "run.json"), "utf-8"));
+        if (state && !Array.isArray(state.branches))
+            state.branches = [];
+        if (state && !Array.isArray(state.currentTasks))
+            state.currentTasks = [];
+        return state;
     }
     catch {
         return null;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-overnight",
-  "version": "1.50.3",
+  "version": "1.50.6",
   "description": "Parallel Claude agents in git worktrees with a usage cap that reserves headroom for your interactive Claude Code. Crash-safe resume. Provider-agnostic model catalog (Anthropic, Cursor, OpenAI, Gemini, DeepSeek, Llama, Qwen) with capability-based task scoping.",
   "type": "module",
   "bin": {

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

@@ -1,6 +1,6 @@
 {
   "name": "claude-overnight",
-  "version": "1.50.3",
+  "version": "1.50.6",
   "description": "Claude Code skill for understanding, installing, and inspecting claude-overnight runs  -- parallel Claude agents in git worktrees with thinking waves, multi-wave steering, and crash-safe resume. Supports Cursor API Proxy, Qwen, OpenRouter.",
   "author": {
     "name": "Francesco Fornace"