pubrun 0.2.0

A lightweight Python library for capturing execution context.

README | Architecture | Functional Spec | API | CLI | Configuration | Manifest

pubrun

Let your code monitor itself and write its own Methods section while you go to the pub.

pubrun is a stupidly simple, zero-dependency¹ Python library that eliminates the boilerplate of documenting methodology, tracking versions, recording inputs, and monitoring resources — making it dramatically easier to publish, share, and reproduce your models and research. If you're feeling formal, you can think of "publication-ready runner" as the meaning of the name.

Installation

pip install pubrun

On Python 3.8–3.10, this also installs tomli (a backport of the standard-library tomllib). On Python 3.11+, there are zero runtime dependencies.

Quick Start

import pubrun  # That's it 90% of the time!

or

pubrun -h  # Lots of info here.

That's it. No frameworks, no heavy integrations, no syntax hijacking. When the script exits, pubrun silently generates a structured, lightweight footprint in your local ./runs/ directory.

[!NOTE] Console capture: By default, pubrun tees stdout and stderr to log files in the run directory. Your terminal output is unchanged, but a copy is saved alongside the manifest. If your script produces very high output volume, you can disable this with capture_mode = "off" in .pubrun.toml or via pubrun.start(console={"capture_mode": "off"}). See Configuration for details.

See CLI Reference and API Reference for full details.

Features

• Automatic Execution Tracing — Captures environment variables, hardware specs, and dependency graphs without manual configuration.
• Publication-Ready Output — Generates LaTeX/Markdown methodology blocks ready for academic papers.
• Semantic Diffing — Compares execution footprints to identify subtle but critical differences between runs.
• Secret Redaction — Automatically detects and redacts passwords, tokens, and API keys in environment variables and CLI arguments.
• Codebase Drift Detection — Compares current code state against the execution snapshot to highlight changes.
• Cross-Platform Reproducibility — Extracts initialization commands for seamless environment replication.
• HPC Optimized — Supports global parent-child manifest hydration to minimize overhead on massive clusters.

The Problem

Modern scientific workflows rely on implicit state. When it's time to publish a paper or ship a model, researchers are forced to retroactively piece together their methodology — PyTorch versions, OS constraints, hardware parameters — from memory.

The Solution

pubrun permanently ends this friction.

With a single import pubrun, the library quietly traces your script execution, hashes your environment dependencies, detects codebase drift, and compiles publication-ready Computational Methodology LaTeX/Markdown blocks so your run is instantly citable.

Lazy Initialization (Explicit Tracking)

By default, simply importing pubrun spins up an invisible tracer. If you want to import pubrun without instantly generating a footprint until you explicitly call pubrun.start(), set this in your .pubrun.toml:

[core]
auto_start = false

Or set the environment variable before import:

import os
os.environ["PUBRUN_AUTO_START"] = "false"

import pubrun
# No directory is generated until you say so.

pubrun.start(output_dir="./custom_storage", profile="deep")

Now extract your method paragraph for your paper:

pubrun methods --format latex

Sample Output

Computational experiments were executed on a machine running Linux (5.15.0-91-generic) equipped with an Intel(R) Core(TM) i7-12700H and 32.0 GB of RAM. The execution environment relied on Python 3.10.12 (CPython). Key dependencies tracked include torch (v2.0.1) and numpy (v1.24.3). To guarantee computational reproducibility, the exact state of the source code was anchored at Git commit a1b2c3d4. Environment and execution provenance were tracked using the pubrun library [1].

[!NOTE] Windows support: pubrun works on Windows, but some capture engines have reduced functionality. Process uid/gid fields are not available, and os.system interception uses shell-string parsing rather than structured argument lists. All other features work identically.

---

CLI Reference

The pubrun CLI provides eight commands and diagnostic flags, all designed to work equally well on a developer laptop or across a Slurm array of thousands of HPC jobs.

pubrun cite

Generates the bibliographic citation for crediting this library in your paper.

pubrun cite --style bibtex

pubrun methods

Translates raw JSON diagnostic payloads into publication-ready methodology paragraphs.

pubrun methods [RUN_DIR] --format markdown|latex

pubrun report

A diagnostic viewer that surfaces execution timing, hardware, dependencies, and codebase drift. Accepts multiple run directories for sequential evaluation.

pubrun report ./runs/pubrun-A ./runs/pubrun-B --deep

pubrun rerun

Extracts the exact shell command needed to reproduce a run.

pubrun rerun ./runs/pubrun-A

pubrun diff

Generates a semantic side-by-side comparison between two execution traces, filtering volatile noise (timestamps, PIDs) by default.

pubrun diff ./runs/pubrun-A ./runs/pubrun-B --same --basic --wrap

pubrun meta

Generates a standalone environment snapshot for HPC parent-child hydration.

pubrun meta --out ./runs/meta.json --deep

pubrun status

Lists all runs with their current status (completed, failed, interrupted, running, crashed, ghost), or inspects a specific run in detail. Detects active processes via cross-platform PID liveness checks.

pubrun status              # Compact table of all runs
pubrun status -v           # Verbose listing with PID, RSS, CPU, events
pubrun status a3f9         # Inspect a specific run by ID prefix
pubrun status --dir /path  # Scan a non-default output directory

pubrun clean

Interactively delete old run directories. Lists candidates with age and size, then prompts for confirmation.

pubrun clean                        # Interactive: list and confirm
pubrun clean --older-than 7d --yes  # Non-interactive: delete all completed runs older than 7 days
pubrun clean --status crashed --yes # Delete all crashed runs
pubrun clean --dry-run              # Preview what would be deleted

Diagnostic Flags

Flag	Description
--version	Print the installed pubrun version and exit
--create-config	Bootstrap a fully commented .pubrun.toml file
--show-config	Print the default configuration to the terminal
--info	Display system capabilities and pubrun version
--run-tests	Execute the built-in self-test suite

See CLI Reference for full details and examples.

---

Monitoring Runs

pubrun tracks the lifecycle of every run from start to finish, enabling real-time and post-hoc inspection of execution state.

Lock Files and Liveness Detection

When a run starts, pubrun writes a .pubrun.lock file to the run directory containing the PID, start timestamp, hostname, and git commit. This file is removed when the run finalizes normally.

If a process is killed (SIGKILL, OOM, power loss), the lock file persists. pubrun status detects these orphaned runs by checking whether the recorded PID is still alive and whether its start time matches (to handle PID recycling). Runs are classified as:

Status	Meaning
completed	Manifest exists, outcome is "completed"
failed	Manifest exists, outcome is "failed"
interrupted	Run received SIGINT, SIGTERM, or SIGHUP (e.g., Ctrl+C)
running	Lock file present, process is alive
crashed	Lock file present, process is dead
ghost	Run entered ghost mode (filesystem write failure at init)

Signal and Exit Code Capture

pubrun installs non-intrusive signal handlers that record OS signals (SIGINT, SIGTERM, SIGHUP, etc.) received during execution. These handlers chain to any pre-existing handlers — if the importing script has its own SIGINT handler, it is called normally after pubrun records the signal.

The process exit code is also captured at finalization. All signal and exit data appears in the "signals" section of the manifest:

{
  "signals_received": [
    {"signal": 2, "signal_name": "SIGINT", "timestamp_utc": 1780250544.068}
  ],
  "exit_code": 0,
  "exit_exception": null
}

Signal capture is configurable via [capture.signals].enabled in .pubrun.toml.

Live Process Inspection

For running processes, pubrun status <run-id> shows live resource usage (RSS memory and CPU percent) queried cross-platform:

• Linux: reads from /proc/<pid>/status and /proc/<pid>/stat
• macOS: queries via ps
• Windows: uses ctypes (kernel32/psapi) and wmic

No external dependencies are required.

---

Advanced HPC Ecosystems (Global Hydration)

If you run thousands of array jobs across a cluster, you don't want each child run wasting time and disk logging identical dependency graphs. pubrun supports parent-child manifest hydration.

Step 1: Snap the Parent Cluster

On the head node, snapshot the global environment:

pubrun meta --out ./runs/meta.json --deep

This generates a deep metadata map of hardware, environment variables, and the full Python package tree.

Step 2: Hydrate Children

In your Slurm script, reference the parent snapshot:

export PUBRUN_META_REF=meta.json
python minimal_script.py

Child scripts automatically skip heavy footprint tracking. When you run pubrun report or pubrun methods, the orchestrator detects the PUBRUN_META_REF, pulls in the parent meta.json context, and stitches the complete hardware and dependency picture back together. It also compares script timestamps against the parent snapshot and warns you if environmental drift has been detected.

---

Configuration

pubrun supports a hierarchical configuration system (highest to lowest precedence):

1. API overrides — pubrun.start(profile="deep")
2. Environment variables — PUBRUN_AUTO_START=false
3. Local project config — .pubrun.toml or .config/pubrun/config.toml
4. User home config — ~/.config/pubrun/config.toml
5. Built-in defaults — default.toml (shipped with the library)

Generate a Configuration File

pubrun --create-config

See Configuration Reference for all settings and examples.

---

Security & Redaction

pubrun automatically detects and redacts sensitive values (passwords, tokens, API keys, credentials) in both environment variables and CLI arguments before writing them to the manifest. Redaction is destructive by default — raw values are replaced with {"representation": "redacted"}, and no hashes are generated, to prevent brute-force attacks.

Both environment variable and argv redaction are independently configurable:

[redaction]
env_enabled = true    # Redact matching environment variable values
argv_enabled = true   # Redact matching CLI argument values

See Configuration Reference for the full redaction policy and regex pattern.

---

Roadmap

Future

1. Sphinx / MkDocs integration — Generate hosted API documentation from docstrings.
2. GitHub Actions CI — Automated test matrix on push/PR.
3. Plugin / extension model — Formal extension points for custom capture engines.
4. Artifact registration API — register_artifact() for tracking user-produced output files.
5. Custom metadata API — register_metadata() for injecting structured data into the manifest.
6. Timestamped console capture — standard mode prepends timestamps to log lines, enabling pubrun combined (below).
7. pubrun combined command — Interleaves stdout and stderr from one or more runs using log timestamps. Requires timestamped capture (item 6).

---

Acknowledgements

pubrun was redesigned and rewritten from pre-existing custom libraries, code fragments, scripts, and ideas spanning almost two decades, with the assistance of Google Antigravity for its official release.

License

Released under the BSD 3-Clause License. Copyright (c) 2007-2026 Gabriele Fariello. See the LICENSE file for full terms.

---

README | Architecture | Functional Spec | API | CLI | Configuration | Manifest

1. On Python 3.11+, pubrun uses only the standard library. On Python 3.8–3.10, the sole runtime dependency is tomli (a backport of the standard-library tomllib). ↩