@desplega.ai/agent-swarm 2.7.0 → 2.7.1

package/src/commands/onboard/compose-generator.ts

@@ -1,3 +1,4 @@
+import { AGENT_FS_VERSION } from "./agent-fs-version.ts";
 import { composeCommandText, engineForCommands } from "./container-engine.ts";
 import { expandServices } from "./service-names.ts";
 import type { OnboardState } from "./types.ts";
@@ -71,6 +72,91 @@ const ENV_SENTRY_AUTH_TOKEN = "      - SENTRY_AUTH_TOKEN=${SENTRY_AUTH_TOKEN}";
 // biome-ignore lint/suspicious/noTemplateCurlyInString: Docker Compose env var syntax
 const ENV_SENTRY_ORG = "      - SENTRY_ORG=${SENTRY_ORG}";
 
+// ── Bundled agent-fs + TEI (`onboard --yes`, flag `bundleAgentFs`) ──
+const APP_URL = "https://app.agent-swarm.dev";
+const AGENT_FS_INTERNAL_URL = "http://agent-fs:7433";
+const TEI_MODEL = "nomic-ai/nomic-embed-text-v1.5";
+// Compose `${PUBLIC_MCP_BASE_URL:-http://localhost:<port>}`; the caller appends `<port>}`.
+const ENV_AGENT_FS_PUBLIC_URL_PREFIX =
+  "      - AGENT_FS_PUBLIC_URL=${PUBLIC_MCP_BASE_URL:-http://localhost:";
+const TEI_IMAGE_EXPRESSION =
+  // biome-ignore lint/suspicious/noTemplateCurlyInString: Docker Compose env var syntax
+  "${TEI_IMAGE:-ghcr.io/huggingface/text-embeddings-inference:cpu-1.9.4}";
+
+/** The API side of the bundle: agent-fs proxy, Comb, dashboard links, TEI memory. */
+function appendBundleApiEnvironment(lines: string[]) {
+  lines.push(`      - AGENT_FS_API_URL=${AGENT_FS_INTERNAL_URL}`);
+  lines.push("      - AGENT_FS_PROXY=true");
+  lines.push("      - COMB_ENABLED=true");
+  lines.push("      - AGENT_FS_LINK_TARGET=dashboard");
+  lines.push("      - EMBEDDING_API_BASE_URL=http://tei:80/v1");
+  lines.push(`      - EMBEDDING_MODEL=${TEI_MODEL}`);
+  lines.push("      - EMBEDDING_API_KEY=local-tei");
+  // Quoted: unquoted, `KEY=search_query: ` parses as a mapping and drops the trailing space.
+  lines.push('      - "EMBEDDING_QUERY_PREFIX=search_query: "');
+  lines.push('      - "EMBEDDING_DOCUMENT_PREFIX=search_document: "');
+}
+
+function appendBundleServices(lines: string[], state: OnboardState, apiPort: number) {
+  lines.push("");
+  lines.push("  agent-fs:");
+  lines.push(`    image: "ghcr.io/desplega-ai/agent-fs:${AGENT_FS_VERSION}"`);
+  lines.push("    container_name: swarm-agent-fs");
+  lines.push(`    pull_policy: ${state.pullPolicy}`);
+  lines.push("");
+  lines.push("    environment:");
+  lines.push("      - AGENT_FS_HOME=/data");
+  lines.push("      - AGENT_FS_STORAGE_PROVIDER=local");
+  lines.push("      - AGENT_FS_LOCAL_ROOT=/data/storage");
+  lines.push("      - HOST=0.0.0.0");
+  lines.push("      - SERVER_HOST=0.0.0.0");
+  lines.push("      - PORT=7433");
+  lines.push("      - SERVER_PORT=7433");
+  lines.push("      - EMBEDDING_PROVIDER=local");
+  // The browser reaches agent-fs through the API proxy, on the tunnel URL when there is one.
+  lines.push(`${ENV_AGENT_FS_PUBLIC_URL_PREFIX}${apiPort}}/agentfs`);
+  lines.push(`      - AGENT_FS_APP_URL=${APP_URL}`);
+  lines.push("");
+  lines.push("    volumes:");
+  lines.push("      - agent_fs_data:/data");
+  lines.push("");
+  lines.push("    healthcheck:");
+  lines.push("      # The agent-fs image has no curl or wget.");
+  lines.push(
+    `      test: ["CMD-SHELL", "node -e \\"fetch('http://127.0.0.1:7433/health').then((r) => process.exit(r.ok ? 0 : 1)).catch(() => process.exit(1))\\""]`,
+  );
+  lines.push("      interval: 10s");
+  lines.push("      timeout: 5s");
+  lines.push("      retries: 30");
+  lines.push("      start_period: 10s");
+  lines.push("");
+  lines.push("    restart: unless-stopped");
+
+  lines.push("");
+  lines.push("  tei:");
+  lines.push(`    image: "${TEI_IMAGE_EXPRESSION}"`);
+  lines.push("    container_name: swarm-tei");
+  lines.push(`    pull_policy: ${state.pullPolicy}`);
+  // --max-batch-tokens 2048: the default warmup passed 5.4 GB and was OOM-killed on an 8 GB box.
+  // --max-client-batch-size 256: the default 32 rejects documents with more than 32 chunks.
+  lines.push(
+    `    command: ["--model-id", "${TEI_MODEL}", "--max-batch-tokens", "2048", "--max-client-batch-size", "256"]`,
+  );
+  lines.push("    mem_limit: 2g");
+  lines.push("");
+  lines.push("    volumes:");
+  lines.push("      - tei_data:/data");
+  lines.push("");
+  lines.push("    healthcheck:");
+  lines.push('      test: ["CMD-SHELL", "curl -sf http://127.0.0.1:80/health || exit 1"]');
+  lines.push("      interval: 10s");
+  lines.push("      timeout: 5s");
+  lines.push("      retries: 30");
+  lines.push("      start_period: 180s");
+  lines.push("");
+  lines.push("    restart: unless-stopped");
+}
+
 function appendProviderEnvironment(lines: string[], state: OnboardState, includeHarness: boolean) {
   if (includeHarness) lines.push(ENV_HARNESS_PROVIDER);
   switch (state.provider) {
@@ -155,7 +241,7 @@ export function generateCompose(state: OnboardState): string {
   lines.push("      - MEMORY_DEMOTION_FLOOR=${MEMORY_DEMOTION_FLOOR-1.0}");
   appendProviderEnvironment(lines, state, false);
   lines.push(`      - MCP_BASE_URL=http://localhost:${port}`);
-  lines.push("      - APP_URL=https://app.agent-swarm.dev");
+  lines.push(`      - APP_URL=${APP_URL}`);
 
   if (state.integrations.slack) {
     lines.push("      - SLACK_DISABLE=false");
@@ -168,6 +254,19 @@ export function generateCompose(state: OnboardState): string {
     lines.push("      - GITHUB_DISABLE=false");
   }
 
+  if (state.bundleAgentFs) {
+    appendBundleApiEnvironment(lines);
+    lines.push("");
+    // The boot seeder provisions agent-fs, so it must be up first. TEI only needs to
+    // start: a first boot downloads 1.1 GB of weights, and memory falls back to
+    // keyword search until it is ready.
+    lines.push("    depends_on:");
+    lines.push("      agent-fs:");
+    lines.push("        condition: service_healthy");
+    lines.push("      tei:");
+    lines.push("        condition: service_started");
+  }
+
   lines.push("");
   lines.push("    ports:");
   lines.push(`      - "${port}:3013"`);
@@ -238,6 +337,12 @@ export function generateCompose(state: OnboardState): string {
       lines.push(ENV_SENTRY_ORG);
     }
 
+    if (state.bundleAgentFs) {
+      lines.push(`      - AGENT_FS_API_URL=${AGENT_FS_INTERNAL_URL}`);
+      lines.push(`      - APP_URL=${APP_URL}`);
+      lines.push("      - AGENT_FS_LINK_TARGET=dashboard");
+    }
+
     lines.push("");
     lines.push("    ports:");
     lines.push(`      - "${port}:3000"`);
@@ -255,6 +360,8 @@ export function generateCompose(state: OnboardState): string {
     lines.push("    restart: unless-stopped");
   }
 
+  if (state.bundleAgentFs) appendBundleServices(lines, state, port);
+
   // ── Volumes section ──
   lines.push("");
   lines.push("volumes:");
@@ -264,6 +371,10 @@ export function generateCompose(state: OnboardState): string {
   for (const svc of expanded) {
     lines.push(`  swarm_${svc.sanitizedName}:`);
   }
+  if (state.bundleAgentFs) {
+    lines.push("  agent_fs_data:");
+    lines.push("  tei_data:");
+  }
 
   return `${lines.join("\n")}\n`;
 }

package/src/commands/onboard/env-generator.ts

@@ -43,6 +43,15 @@ export function generateEnv(state: OnboardState): string {
   lines.push("MEMORY_RATERS=implicit-citation,explicit-self");
   lines.push("MEMORY_DEMOTION_FLOOR=1.0");
 
+  if (state.bundleAgentFs) {
+    lines.push("");
+    lines.push("# Bundled agent-fs (local disk) and TEI (memory embeddings); settings are in");
+    lines.push(
+      "# docker-compose.yml. Generate without them: agent-swarm onboard --yes --no-agent-fs",
+    );
+    if (state.teiImage) lines.push(`TEI_IMAGE=${state.teiImage}`);
+  }
+
   // ── Authentication ──
   lines.push("");
   lines.push("# === Authentication ===");

package/src/commands/onboard/tei-image.ts

@@ -0,0 +1,12 @@
+/** TEI image for the bundled embeddings server on arm64 hosts. */
+export const TEI_ARM64_IMAGE = "ghcr.io/huggingface/text-embeddings-inference:cpu-arm64-1.9.4";
+
+/**
+ * TEI publishes one tag per CPU architecture, so Compose `platform:` cannot
+ * pick it. `install.sh` runs onboard on the Docker host, so the host arch is
+ * the engine arch. Returns `null` for amd64 and anything else: the compose
+ * default (`cpu-1.9.4`, amd64) applies.
+ */
+export function teiImageForArch(arch: string): string | null {
+  return arch === "arm64" || arch === "aarch64" ? TEI_ARM64_IMAGE : null;
+}

package/src/commands/onboard/types.ts

@@ -95,6 +95,10 @@ export interface OnboardState {
   containerEngine: ContainerEnginePreference;
   outputDir: string;
   nonInteractive: boolean;
+  /** Bundle agent-fs (local disk) and TEI. Only `--yes` without `--no-agent-fs` sets it. */
+  bundleAgentFs: boolean;
+  /** TEI image override for the host arch; `null` keeps the compose default (amd64). */
+  teiImage: string | null;
   error: string | null;
   logs: LogLine[];
 }
@@ -114,6 +118,7 @@ export interface OnboardProps {
   maxConcurrentTasks?: string;
   pullPolicy?: string;
   containerEngine?: string;
+  noAgentFs?: boolean;
 }
 
 export type PullPolicy = "always" | "missing" | "never";
@@ -212,6 +217,8 @@ export const INITIAL_STATE: OnboardState = {
   containerEngine: "auto",
   outputDir: process.cwd(),
   nonInteractive: false,
+  bundleAgentFs: false,
+  teiImage: null,
   error: null,
   logs: [],
 };

package/src/commands/onboard.tsx

@@ -1,4 +1,5 @@
 import crypto from "node:crypto";
+import os from "node:os";
 import { Select } from "@inkjs/ui";
 import { Box, Text, useApp, useInput } from "ink";
 import { useCallback, useEffect, useRef, useState } from "react";
@@ -35,6 +36,7 @@ import { PostTaskStep } from "./onboard/steps/post-task.tsx";
 import { PrereqCheckStep } from "./onboard/steps/prereq-check.tsx";
 import { ReviewStep } from "./onboard/steps/review.tsx";
 import { StartStep } from "./onboard/steps/start.tsx";
+import { teiImageForArch } from "./onboard/tei-image.ts";
 import {
   getStepProgress,
   INITIAL_STATE,
@@ -61,6 +63,7 @@ export function Onboard({
   maxConcurrentTasks,
   pullPolicy,
   containerEngine,
+  noAgentFs = false,
 }: OnboardProps) {
   const { exit } = useApp();
   const [state, setState] = useState<OnboardState>(() => {
@@ -229,8 +232,15 @@ export function Onboard({
       addLog(`Integrations: ${enabled.join(", ")}`);
     }
 
+    const bundleAgentFs = !noAgentFs;
+    if (bundleAgentFs) {
+      addLog("Bundling agent-fs and TEI (local files and memory; --no-agent-fs to skip)");
+    }
+
     goToStep("generate", {
       nonInteractive: true,
+      bundleAgentFs,
+      teiImage: bundleAgentFs ? teiImageForArch(os.arch()) : null,
       deployType: "local",
       presetId: selectedPreset.id,
       services: selectedPreset.services,
@@ -250,7 +260,7 @@ export function Onboard({
       sentryToken: process.env.SENTRY_AUTH_TOKEN || "",
       sentryOrg: process.env.SENTRY_ORG || "",
     });
-  }, [yes, preset, state.step, addLog, goToStep, goToError]);
+  }, [yes, preset, noAgentFs, state.step, addLog, goToStep, goToError]);
 
   // Interactive welcome: auto-advance after detecting existing setup
   useEffect(() => {

package/src/commands/runner.ts

@@ -1222,6 +1222,7 @@ export async function provisionAgentFsAfterRegistration(opts: {
  * - TEMPLATE_REGISTRY_URL — read per registry fetch.
  * - SLACK_DISABLE — read by the prompt builder to gate the Slack tool section.
  * - SWARM_ORG_NAME — read per telemetry event for org identity.
+ * - AGENT_FS_LINK_TARGET — read by the prompt builder for the file-link base.
  *
  * CLAUDE_TRANSPORT stays in the per-session environment. Caching a scoped
  * value here would prevent deletion from restoring the configured default.
@@ -1241,6 +1242,7 @@ export const RELOADABLE_ENV_KEYS: ReadonlySet<string> = new Set([
   "TEMPLATE_REGISTRY_URL",
   "SLACK_DISABLE",
   "SWARM_ORG_NAME",
+  "AGENT_FS_LINK_TARGET",
   // Gate the public-ranking `ranking.usage` emit, so clearing the website or
   // turning the setting off stops it on running workers.
   "SWARM_ORG_WEBSITE",
@@ -2714,6 +2716,8 @@ export interface RunningTask {
   assistantText?: { value?: string };
   /** Last progress value, retained until the final API writes settle. */
   finalProgress?: { value?: string };
+  /** Latest cost snapshot captured from `result` events before completion settles. */
+  lastCost?: { value?: CostData };
   /** Last time `reconcileActiveTasks` read this task's server-side status. */
   lastReconcileAt?: number;
   /** When the runner asked the session to abort (cancel or reconcile). */
@@ -4553,6 +4557,7 @@ async function spawnProviderProcess(
   // session promise resolves (see `trackAssistantText`).
   const assistantTextHolder: { value?: string } = {};
   const finalProgressHolder: { value?: string } = {};
+  const lastCostHolder: { value?: CostData } = {};
 
   // Context usage throttle: max 1 snapshot per 30 seconds
   let lastContextPostTime = 0;
@@ -4761,6 +4766,11 @@ async function spawnProviderProcess(
           break;
         }
         case "result":
+          lastCostHolder.value = {
+            ...event.cost,
+            taskId: realTaskId,
+            sessionId: opts.runnerSessionId,
+          };
           {
             // Mid-session snapshot — fires on each streamed "result" event,
             // necessarily before `ProviderResult.appliedReasoningEffort` exists
@@ -4968,6 +4978,7 @@ async function spawnProviderProcess(
         }
 
         // Retain cost data for the concurrent final API writes.
+        result.cost ??= lastCostHolder.value;
         if (result.cost) {
           result.cost = { ...result.cost, taskId: realTaskId, sessionId: opts.runnerSessionId };
           sessionSpan.setAttributes({
@@ -5130,6 +5141,7 @@ async function spawnProviderProcess(
     model: model || undefined,
     assistantText: assistantTextHolder,
     finalProgress: finalProgressHolder,
+    lastCost: lastCostHolder,
   };
   runningTaskForSessionInit = runningTask;
   if (pendingHarnessVariant !== undefined) {
@@ -5146,7 +5158,7 @@ async function spawnProviderProcess(
       runningTask.result ??= r;
     })
     .catch(() => {
-      runningTask.result ??= { exitCode: 1, isError: true };
+      runningTask.result ??= { exitCode: 1, isError: true, cost: lastCostHolder.value };
     });
 
   return runningTask;
@@ -5316,6 +5328,7 @@ export async function reconcileActiveTasks(
         exitCode: 1,
         isError: true,
         sessionId: task.session.sessionId,
+        cost: task.lastCost?.value,
         failureReason: "runner exited without result: provider session did not settle after abort",
       };
     }