@songsid/agend 2.1.5-beta.11 → 2.1.5-beta.13

package/dist/fleet-context.d.ts

@@ -153,6 +153,13 @@ export interface FleetContext {
     cancelInstallSession?(): Promise<string>;
     /** Human-readable effective model for an instance (resolves inherited defaults). */
     modelDisplayForInstance?(name: string): string;
+    /** How this instance's backend takes a reasoning-effort setting, if at all. */
+    effortStrategyFor?(name: string): "runtime" | "restart" | "unsupported";
+    /** Configured effort for an instance: per-instance, else fleet default, else none. */
+    resolveInstanceEffort?(name: string): {
+        effort: string | null;
+        source: "instance" | "fleet-default" | "unset";
+    };
     /**
      * Show a model-selection inline keyboard for the given instance in a TG topic.
      * Returns a fallback text message if no model list is available (caller should send it).

package/dist/fleet-manager.d.ts

@@ -88,6 +88,15 @@ export interface DeliveryOptions {
  * The wait is announced before it starts, so it reads as progress, not a stall.
  */
 export declare const CLI_ENV_PROBE_DEADLINE_MS = 16000;
+/**
+ * How many CLIs may cold-start at once, from BOTH memory and cores.
+ *
+ * Memory alone said 10 on any host with roughly 3GB free, so a three-core box
+ * started ten CLIs together, saturated the CPU, and healthy starts then missed
+ * their startup budget — which used to cost the user their conversation. Cores
+ * bound how many can actually make progress; memory bounds how many fit.
+ */
+export declare function deriveSpawnConcurrency(freeMemMB: number, cores: number): number;
 export declare class FleetManager implements FleetContext, LifecycleContext, ArchiverContext, StatuslineWatcherContext, OutboundContext, AgentEndpointContext {
     dataDir: string;
     private static signalTarget;

package/dist/fleet-manager.js

@@ -1,7 +1,7 @@
 import { existsSync, readFileSync, mkdirSync, writeFileSync, unlinkSync, rmSync, readdirSync, renameSync, copyFileSync, chmodSync, statSync } from "node:fs";
 import { randomBytes } from "node:crypto";
 import { spawnSync } from "node:child_process";
-import { freemem, totalmem } from "node:os";
+import { freemem, totalmem, cpus } from "node:os";
 import { createServer } from "node:http";
 import { join, dirname, basename } from "node:path";
 import { fileURLToPath } from "node:url";
@@ -262,6 +262,18 @@ const CLI_ENV_FRESH_MS = 60 * 60 * 1000;
  * The wait is announced before it starts, so it reads as progress, not a stall.
  */
 export const CLI_ENV_PROBE_DEADLINE_MS = 16_000;
+/**
+ * How many CLIs may cold-start at once, from BOTH memory and cores.
+ *
+ * Memory alone said 10 on any host with roughly 3GB free, so a three-core box
+ * started ten CLIs together, saturated the CPU, and healthy starts then missed
+ * their startup budget — which used to cost the user their conversation. Cores
+ * bound how many can actually make progress; memory bounds how many fit.
+ */
+export function deriveSpawnConcurrency(freeMemMB, cores) {
+    const byMemory = Math.floor(freeMemMB / 300);
+    return Math.max(2, Math.min(10, byMemory, Math.max(1, cores)));
+}
 export class FleetManager {
     dataDir;
     static signalTarget = null;
@@ -452,8 +464,7 @@ export class FleetManager {
         const explicit = this.fleetConfig?.defaults?.startup?.concurrency;
         if (explicit != null)
             return Math.max(1, Math.min(20, explicit));
-        const freeMemMB = Math.round(freemem() / (1024 * 1024));
-        return Math.max(2, Math.min(10, Math.floor(freeMemMB / 300)));
+        return deriveSpawnConcurrency(Math.round(freemem() / (1024 * 1024)), cpus().length);
     }
     /** Wire the one fleet-wide storm into notification and recovery surfaces. */
     bindStormWindowEvents() {
@@ -8096,7 +8107,7 @@ You CAN write code snippets, explain code, and answer technical questions direct
 - **Delegate to 1 instance**: scoped to one project/repo, needs file access or execution.
 - **Coordinate multiple**: spans repos, outputs feed each other, or parallel helps (max 3 per task).
 
-Instance discovery order: list_teams() → list_instances() → describe_instance() → create_instance() only if nothing fits. Prefer reuse; never duplicate a running instance.
+Instance discovery: start with list_instances(); follow the guidance in its response. Prefer reuse; never duplicate a running instance.
 
 ## Reply Contract