Double-fork daemonization to fully detach the process.

Compute site_packages_path and multiversion_base directly from a Python executable path, without relying on ConfigManager lookup. IMPORT

Detect binary executables by reading file headers/magic numbers.

Detect if we're in a circular import scenario.

Detect corruption patterns in conda command stderr output. Args: stderr_output: Standard error from failed conda command

🚀 ULTRA-FAST Conflict Detection (Redis Pipelining). Replaces 2.3ms of disk I/O with ~0.1ms of cache lookup.

Detect the Python implementation (cpython, pypy, graalpy, etc.). Returns: str: Normalized implementation name ('cpython'

Discover CUDA library paths for this package spec. Dynamically detects CUDA version requirement (cu11 vs cu12).

ULTRA-FAST targeted discovery for known packages. Only searches expected locations instead of entire filesystem. Args:

Auto-discover all site-packages directories for known Python interpreters by reading the SHM registry JSON written by WorkerPoolDaemon.

Query omnipkg for all torch versions installed in the current Python's bubbles, then return up to `want` specs that each have a *distinct* CU

Kill a list of process dicts; return (killed, total).

Download and extract source code.

Parse ELF64 section headers from raw bytes. Returns {section_name: (sh_type, vma, file_offset, size)}. Silently returns {} on

Emergency on-the-spot metadata healing when corruption is detected during KB scan. Returns True if healing succeeded, False otherwise

Ensure critical omnipkg runtime deps are installed in a newly adopted interpreter.

Creates .omnipkg_config.json next to the interpreter if it doesn't exist. Called from the dispatcher at the earliest moment we know a path is

Self-healing shim check for the native/primary Python. The adoption path short-circuits with "already available" for the native interpre

Pre-flight check: warn if no numpy bubble dir exists for a torch/tensorflow spec. KB-guided range lookup requires pre_core which is n

Creates a shim directory where 'python', 'python3', and 'pip' point back to the omnipkg dispatcher.

Populate _uv_exe_cached / _uv_cache_dir / _uv_ffi_run on core if not yet done.

Write a minimal .omnipkg_config.json next to the worker's Python executable so ConfigManager._get_our_config_path() finds it and doesn't fall

Activate multiple packages sequentially, with dependency-aware ordering. Ordering rule: if package A's bubble directory contains pa

Estimate directory size with early termination for safety.

Evict the longest-idle non-dirty worker. Call with self._lock held.

CPU-only mode: Run computation on CPU without any GPU transfers. Uses zero-copy SHM like Test 17. This is faster than Hybrid

Execute code with CUDA IPC tensors.

Hybrid mode: Copy to CPU SHM, worker copies to GPU. NOTE: Benchmarks show this is the SLOWEST mode (14ms vs 1.5ms Universal).

PyTorch 1.x native IPC (framework-managed).

Execute a Python script (existing functionality)

Universal CUDA IPC - client pre-allocates both buffers, worker just opens handles.

Execute command with bubbled dependencies

Pull every failing package name out of pip's stderr.

Extract the import name by finding the most common top-level directory in the RECORD file (excluding common non-package directories).

Extract the problematic package name from error output.

Parses all Project-URL metadata headers into a structured dict and extracts the most useful individual links (PyPI, GitHub, Homepage)

enclave Extract version from various pip download filename formats across pip eras. Handles package name normalization (dashes become

O(n) scan for missing METADATA files across main env + all bubbles. Pure is_dir/exists calls — no parsing, no pip, no Redis.

Fetch package metadata and compatibility data from GitHub

Lightweight PyPI fetch - just gets the latest version number. Used for background cache refreshes.

Finds the latest version for each dependency before a given date.

Find newest package version with wheels for target Python version

FIXED: Authoritatively finds files belonging to the specific distribution. Also collects native extension files (.so / .pyd / .dll) f

Find the Python executable in a directory, checking common locations.

CRITICAL HEALING LOGIC: Fixes import failures by copying missing modules. This is what was missing from the streamlined version.

Force a module to complete its initialization.

Format uptime in human-readable format.

Generate shell-specific activation script

Generate smart import candidates based on package name patterns.

Gets the version of a package actively installed in the current Python environment using pip show.

Read top_level.txt to get actual import names. This fixes issues where package name != import name (e.g. tomli -> tomli, autocommand

Gets a list of standard binary paths to search for executables.

Ask the target interpreter to run _get_core_dependencies for its Python version. Returns a list of PEP 508 dep strings like "aiohttp>=3.13.3"

Helper to get current timestamp for knowledge base entries.

Compute env_id the same way core.py does — canonical resolved path hash.

(ROBUST VERSION) Runs a subprocess to get the authoritative file list. The script is now hardened to handle corrupted metadata and pr

Constructs the deterministic slot name for this package in the global cache. Format: ``<canonical_name>-<version>-<record_hash[:16]>

Creates a temp venv USING THE CORRECT PYTHON CONTEXT to discover dependencies.

Uses PyPI JSON API to get the release date for a specific version.

Get the authoritative list of import candidates for a package. Uses multiple strategies in order of reliability.

(Authoritative) Finds import candidates by reading top_level.txt from the package's .dist-info directory. Handles PyPI's inconsistent

Gets or creates a file lock for a specific package INSTALLATION. This prevents race conditions when multiple threads try to install

Builds or loads a FAST hash index using multiple strategies: 1. Isolated subprocess for authoritative file lists (preferred)

FIXED: Returns per-Python config path, not global config PRIORITY ORDER: 1. Environment variable override (for testing)

Helper to find all modules related to a package in sys.modules.

Get current platform information for wheel matching.

Generates the Redis key prefix for a specific Python version string.

Build a ``{canonical_name: version}`` snapshot of everything installed alongside (or under) this package. For active/main-en

FIXED: Parses the security report from `safety`, correctly handling both the legacy object format ({'pkg': [...]}) and the modern lis

Returns a human-readable description of swap speed.

Helper to get a valid system language code.

Get or create the daemon client and resolve py3.11 path once.

Returns the exact set of folder/file names that must be symlinked from the global cache slot into a bubble so Python's import machine

"Compile-as-ghost" architecture (v3): For each (_, target) in swaps, compile dispatcher.c to a unique temp path (e.g. %TEMP%\\omnipkg\\_

Canonical path for a given content SHA.

Handle circular imports by pre-loading dependencies.

Handle C++/Python boundary imports with lazy initialization.

Logic for the parent process on Unix after forking.

Handle partial initialization.