@exulu/backend 3.7.3 → 3.7.4

package/dist/index.d.cts

@@ -2252,11 +2252,6 @@ interface PythonSetupResult {
     /** Full output from setup script */
     output?: string;
 }
-/**
- * Check if Python environment is already set up
- * Note: This only checks if the venv exists, not if packages are installed.
- * Use validatePythonEnvironment() for a more thorough check.
- */
 declare function isPythonEnvironmentSetup(packageRoot?: string): boolean;
 /**
  * Set up the Python environment by running the setup script

package/dist/index.d.ts

@@ -2252,11 +2252,6 @@ interface PythonSetupResult {
     /** Full output from setup script */
     output?: string;
 }
-/**
- * Check if Python environment is already set up
- * Note: This only checks if the venv exists, not if packages are installed.
- * Use validatePythonEnvironment() for a more thorough check.
- */
 declare function isPythonEnvironmentSetup(packageRoot?: string): boolean;
 /**
  * Set up the Python environment by running the setup script

package/dist/index.js

@@ -5,7 +5,7 @@ import {
   isPythonEnvironmentSetup,
   setupPythonEnvironment,
   validatePythonEnvironment
-} from "./chunk-T6JVFT7L.js";
+} from "./chunk-27K2CO47.js";
 import {
   COMPACTION_INSUFFICIENT,
   ContextCompactionRequiredError,
@@ -88,7 +88,7 @@ import {
   verifyCredentialNonce,
   waitForLiteLLMReady,
   withRetry
-} from "./chunk-BNTL6LYY.js";
+} from "./chunk-AWMU6QXB.js";
 import {
   LiteLLMAdminError,
   findLiteLLMModel,
@@ -8285,10 +8285,12 @@ var generateSync = async ({
   }
   let project;
   let sessionItems;
+  let sessionOwnerId;
   if (session) {
     const sessionData = await getSession({ sessionID: session });
     sessionItems = sessionData.session_items;
     project = sessionData.project;
+    sessionOwnerId = sessionData.user ?? void 0;
   }
   const model = languageModel;
   console.log("[EXULU] Model created for generating sync.");
@@ -8392,7 +8394,8 @@ var generateSync = async ({
     agent,
     memoryItems,
     contextWindow,
-    disabledTools
+    disabledTools,
+    sessionOwnerId
   );
   const agenticEntry = currentTools?.find((t) => t.id === "agentic_context_search");
   const agenticToolKey = agenticEntry ? sanitizeToolName(agenticEntry.name) : void 0;
@@ -8627,10 +8630,12 @@ var generateStream = async ({
   let previousMessagesContent = previousMessages || [];
   let project;
   let sessionItems;
+  let sessionOwnerId;
   if (session) {
     const sessionData = await getSession({ sessionID: session });
     project = sessionData.project;
     sessionItems = sessionData.session_items;
+    sessionOwnerId = sessionData.user ?? void 0;
     console.log("[EXULU] loading previous messages from session: " + session);
     const previousMessages2 = await getAgentMessages({
       session,
@@ -8810,7 +8815,8 @@ When a tool execution is not approved by the user, do not retry it unless explic
     agent,
     memoryItems,
     contextWindow,
-    disabledTools
+    disabledTools,
+    sessionOwnerId
   );
   console.log("[EXULU] Converted tools", Object.keys(tools));
   const includesContextSearchTool = currentTools?.some(

package/dist/python-setup-JZGHWQCG.js

@@ -0,0 +1,17 @@
+import "dotenv/config";
+import {
+  getPackageRoot,
+  getPythonSetupInstructions,
+  getPythonVenvPath,
+  isPythonEnvironmentSetup,
+  setupPythonEnvironment,
+  validatePythonEnvironment
+} from "./chunk-27K2CO47.js";
+export {
+  getPackageRoot,
+  getPythonSetupInstructions,
+  getPythonVenvPath,
+  isPythonEnvironmentSetup,
+  setupPythonEnvironment,
+  validatePythonEnvironment
+};

package/ee/invoke-skills/artifact-filter.test.ts

@@ -0,0 +1,49 @@
+import { isIgnoredArtifactPath, capArtifacts, needsDownload } from "./artifact-filter";
+
+describe("isIgnoredArtifactPath — dependency and cache trees are not session artifacts", () => {
+  it("ignores virtualenvs, node_modules, caches and VCS metadata at any depth", () => {
+    for (const p of [
+      "venv/pyvenv.cfg",
+      "venv/lib/python3.11/site-packages/docx/api.py",
+      ".venv/bin/python",
+      "node_modules/lodash/index.js",
+      "scripts/__pycache__/create.cpython-311.pyc",
+      ".cache/pip/http/x",
+      ".git/objects/ab/cd",
+      "project/env/lib/python3.11/site-packages/x.py",
+    ]) {
+      expect(isIgnoredArtifactPath(p)).toBe(true);
+    }
+  });
+
+  it("keeps the files the agent actually produces", () => {
+    for (const p of ["Besprechungsprotokoll.docx", "create_minutes.py", "out/report.pdf", "Angebot/Angebot_107426.docx"]) {
+      expect(isIgnoredArtifactPath(p)).toBe(false);
+    }
+  });
+});
+
+describe("capArtifacts — a bash command that touches thousands of files must not flood the tool result", () => {
+  const many = Array.from({ length: 250 }, (_, i) => ({ relativePath: `out/file_${i}.txt`, url: `https://s3/${i}` }));
+
+  it("returns everything below the cap untouched", () => {
+    const { kept, omitted } = capArtifacts(many.slice(0, 10), 50);
+    expect(kept).toHaveLength(10);
+    expect(omitted).toBe(0);
+  });
+
+  it("keeps only the first N and reports how many were omitted", () => {
+    const { kept, omitted } = capArtifacts(many, 50);
+    expect(kept).toHaveLength(50);
+    expect(kept[0]!.relativePath).toBe("out/file_0.txt");
+    expect(omitted).toBe(200);
+  });
+});
+
+describe("needsDownload — resyncing S3 session files into a live sandbox", () => {
+  it("downloads when the local file is missing or differs in size, skips identical files", () => {
+    expect(needsDownload(undefined, 100)).toBe(true);
+    expect(needsDownload(99, 100)).toBe(true);
+    expect(needsDownload(100, 100)).toBe(false);
+  });
+});

package/ee/invoke-skills/artifact-filter.ts

@@ -0,0 +1,38 @@
+/**
+ * The sandbox mirrors every file a shell command creates under the session directory to
+ * S3 and lists each one (with a presigned URL) in the tool result. A single
+ * `python -m venv venv && pip install python-docx` creates thousands of files, which
+ * turned one tool result into ~1.8M characters and killed the session's context window.
+ */
+
+const IGNORED_SEGMENTS = new Set([
+  "venv",
+  ".venv",
+  "env",
+  "node_modules",
+  "__pycache__",
+  "site-packages",
+  "dist-packages",
+  ".cache",
+  ".git",
+  ".pytest_cache",
+  ".mypy_cache",
+  ".ipynb_checkpoints",
+]);
+
+/** True when any path segment is a dependency, cache or VCS directory. */
+export function isIgnoredArtifactPath(relativePath: string): boolean {
+  return relativePath.split(/[\\/]+/).some((segment) => IGNORED_SEGMENTS.has(segment));
+}
+
+export const DEFAULT_ARTIFACT_CAP = 50;
+
+export function capArtifacts<T>(artifacts: T[], max: number = DEFAULT_ARTIFACT_CAP): { kept: T[]; omitted: number } {
+  if (artifacts.length <= max) return { kept: artifacts, omitted: 0 };
+  return { kept: artifacts.slice(0, max), omitted: artifacts.length - max };
+}
+
+/** Whether an S3 session file must be (re)downloaded into the live sandbox directory. */
+export function needsDownload(localSize: number | undefined, remoteSize: number): boolean {
+  return localSize === undefined || localSize !== remoteSize;
+}

package/ee/invoke-skills/create-sandbox.ts

@@ -8,6 +8,7 @@ import { join, dirname, resolve, relative, posix } from 'node:path'
 import { exec, spawn } from 'node:child_process'
 import { promisify } from 'node:util'
 import { listS3ObjectsByPrefix, getS3ObjectBytes, uploadFile, getPresignedUrl, type S3FileObject } from '@SRC/uppy/index.ts'
+import { isIgnoredArtifactPath, capArtifacts, needsDownload } from './artifact-filter'
 import { getNpmGlobalRoot } from '@SRC/exulu/system-dependencies.ts'
 import type { ExuluConfig } from '@SRC/exulu/app/index.ts'
 import { createBashTool, type Sandbox } from "bash-tool";
@@ -270,6 +271,7 @@ async function restoreArtifactsFromS3(
     sessionId: string,
     userId: number | string,
     config: ExuluConfig,
+    opts: { onlyMissing?: boolean } = {},
 ): Promise<void> {
     const userPrefix = `user_${userId}/sessions/${sessionId}/`
     let objects: S3FileObject[]
@@ -297,8 +299,18 @@ async function restoreArtifactsFromS3(
         const idx = obj.key.indexOf(userPrefix)
         const relativePath = idx >= 0 ? obj.key.slice(idx + userPrefix.length) : ''
         if (!relativePath) continue // directory marker or unexpected key shape
+        if (isIgnoredArtifactPath(relativePath)) continue // never pull dependency trees back
 
         const localPath = join(sessionDir, relativePath)
+        if (opts.onlyMissing) {
+            let localSize: number | undefined
+            try {
+                localSize = (await stat(localPath)).size
+            } catch {
+                localSize = undefined
+            }
+            if (!needsDownload(localSize, obj.size)) continue
+        }
         try {
             // Use binary-safe fetch — session artifacts now include PDFs, .docx
             // and other binary formats (from user uploads as well as agent
@@ -380,6 +392,20 @@ export async function downloadKeyIntoSandbox(opts: {
  * cache AND no session directory on disk), previously persisted artifacts for
  * the session are restored from S3 into the fresh session directory.
  */
+/**
+ * Loaded lazily: python-setup resolves the package root via import.meta.url, which
+ * must not be evaluated when this module is merely imported (jest runs CJS).
+ */
+async function resolvePythonVenvPath(): Promise<string | undefined> {
+    try {
+        const { getPythonVenvPath } = await import('@SRC/utils/python-setup.ts')
+        return getPythonVenvPath()
+    } catch (err) {
+        console.warn('[SKILLS] Could not resolve the Python venv for the session sandbox; skill scripts fall back to the system python.', err)
+        return undefined
+    }
+}
+
 export async function createSessionSandbox(
     sessionId: string,
     skills: SkillRef[],
@@ -404,6 +430,16 @@ export async function createSessionSandbox(
             cached.installedSkills.set(skill.id, skill.current_version)
         }
 
+        // Files uploaded through the side panel (or by an email routine) after the
+        // sandbox was created live only in S3 until re-synced; without this the agent's
+        // `ls` never sees them although the system prompt promises it will.
+        if (userId && config.fileUploads) {
+            try {
+                await restoreArtifactsFromS3(cached.handle.sessionDir, sessionId, userId, config, { onlyMissing: true })
+            } catch (err) {
+                console.error(`[SKILLS] Failed to re-sync S3 session files for session ${sessionId}; continuing.`, err)
+            }
+        }
         return cached.handle
     }
 
@@ -439,8 +475,9 @@ export async function createSessionSandbox(
     // Restore artifacts from S3 only on a true cold start. If the dir already
     // existed, the local files are at least as new as S3 and may contain
     // unsaved-to-S3 writes from a prior process.
-    if (userId && config.fileUploads && !dirExisted) {
-        await restoreArtifactsFromS3(sessionDir, sessionId, userId, config)
+    if (userId && config.fileUploads) {
+        // Cold start restores everything; a surviving on-disk dir only fetches what changed.
+        await restoreArtifactsFromS3(sessionDir, sessionId, userId, config, { onlyMissing: dirExisted })
     }
 
     // Probe whether bwrap can actually create namespaces on this host. On a
@@ -492,6 +529,9 @@ export async function createSessionSandbox(
     // fail with a clear MODULE_NOT_FOUND, matching what the user would see
     // outside the sandbox).
     const npmGlobalRoot = await getNpmGlobalRoot()
+    // The @exulu/backend Python venv carries the skill runtime deps (python-docx …).
+    // Put its bin dir first on PATH so `python3` inside the sandbox resolves to it.
+    const pythonVenvPath = await resolvePythonVenvPath()
 
     // Per-session policy. Passed to every wrapWithSandbox() invocation made
     // from within this closure. customConfig wins over the singleton's
@@ -510,6 +550,7 @@ export async function createSessionSandbox(
                 // the sandbox. Without this, `require('docx')` fails with
                 // EPERM even when NODE_PATH points the resolver here.
                 ...(npmGlobalRoot ? [npmGlobalRoot] : []),
+                ...(pythonVenvPath ? [pythonVenvPath] : []),
             ],
             allowWrite: [sessionDir],
             denyWrite: [],
@@ -563,6 +604,9 @@ export async function createSessionSandbox(
         ...configuredVariables,
         ...process.env,
         ...(npmGlobalRoot ? { NODE_PATH: npmGlobalRoot } : {}),
+        ...(pythonVenvPath
+            ? { PATH: `${join(pythonVenvPath, 'bin')}:${process.env.PATH ?? ''}`, VIRTUAL_ENV: pythonVenvPath }
+            : {}),
     }
 
     // Either wrap a command with bwrap/sandbox-exec or return it unchanged when
@@ -757,6 +801,7 @@ export async function createSessionSandbox(
             for (const entry of entries) {
                 const full = join(dir, entry.name)
                 if (full === skillsDir) continue
+                if (isIgnoredArtifactPath(relative(sessionDir, full))) continue
                 if (entry.isDirectory()) {
                     await walk(full)
                 } else if (entry.isFile()) {
@@ -917,19 +962,26 @@ export async function createSessionSandbox(
             // marker block isn't truncated. Only files with a presigned URL
             // appear here; locally-only entries would just confuse the user.
             let stdout = result?.stdout ?? ''
-            const withUrls = artifacts.filter((a) => a.url)
+            // Cap the listing: one command can touch thousands of files, and every
+            // entry carries a presigned URL. The full set is still mirrored to S3.
+            const { kept, omitted } = capArtifacts(artifacts)
+            const withUrls = kept.filter((a) => a.url)
             if (withUrls.length > 0) {
                 const lines = ['', '[exulu-artifacts]']
                 for (const a of withUrls) {
                     lines.push(`  ${a.relativePath}: ${a.url}`)
                 }
+                if (omitted > 0) {
+                    lines.push(`  … ${omitted} more file(s) were created and mirrored but are not listed here.`)
+                }
                 stdout = `${stdout}\n${lines.join('\n')}`
             }
 
             return {
                 ...result,
                 stdout,
-                artifacts,
+                artifacts: kept,
+                ...(omitted > 0 ? { artifactsOmitted: omitted } : {}),
             }
         },
     })

package/ee/python/requirements.txt

@@ -52,3 +52,8 @@ prisma==0.15.0
 # NOT included in litellm[proxy]; without it, requests to Vertex models
 # fail with "No module named 'vertexai'".
 google-cloud-aiplatform>=1.38
+
+# Skill runtime dependencies — the session sandbox puts this venv's bin dir on PATH so
+# skill scripts (e.g. the built-in docx-manipulation skill) find their libraries without
+# creating a venv per session (which mirrored thousands of files into the tool result).
+python-docx>=1.1

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@exulu/backend",
   "author": "Qventu Bv.",
-  "version": "3.7.3",
+  "version": "3.7.4",
   "main": "./dist/index.js",
   "private": false,
   "publishConfig": {