Skip to content
/ aux Public

[AUx: Agentic Unix] A very particular set of skills.

License

Notifications You must be signed in to change notification settings

JordanGunn/aux

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

22 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

AUx: Agentic Unix

ASI Aligned

If you are looking for ransom, I can tell you I don't have money. But what I do have is a very particular set of skills, skills I have acquired over a very long career.


Don't want to read this?

Use your Agent!

The repository provides an llms.txt file (see spec) for agent fetching.

If you don't want to read the docs, ask your agent to do it for you:

Please fetch the AUx skills repository and summarize the contents from llms.txt.


aux is a collection of atomic agent skills that wrap classic Unix read-only commands (find, grep, diff, ls) with a thin agentic layer.

The tools already exist.
The power comes from letting agents do what they're good at—parameter selection—while keeping execution fully deterministic, inspectable, and replayable.

Note: All skills in this repository are designed in accordance with ASI (Agentic Skill Interface) — a governance framework for deterministic, auditable agent skills.


What problem this solves

Classic Unix tools are:

  • extremely powerful
  • extremely reliable
  • extremely unfriendly at scale

Humans are bad at:

  • choosing exhaustive search terms
  • anticipating edge cases
  • writing correct regexes under time pressure
  • scaling these operations across large trees

Agents are excellent at those things, but unreliable if given too much authority.

aux bridges that gap.


What aux is

  • A facade layer over existing Unix commands
  • Each skill maps 1:1 to a familiar verb:
    • /findfd → file enumeration
    • /greprg → text search
  • Regex is expressed explicitly as pattern semantics inside /find and /grep plans
  • Natural language is used only to express intent
  • Agents generate parameters, not actions
  • Scripts perform all execution

What aux is not

By design, AUx skill are not:

  • A monolithic "search tool"
  • A prompt framework
  • RAG
  • Embeddings
  • Magic
  • Mutation (by default)

If a task can be deterministically scripted, the agent is not allowed to do it.


Core principles (ASI-aligned)

  • Deterministic maximalism

    • Scripts own execution
    • argv arrays, not shell strings
    • Bounded scope, time, and output
  • Subjective minimalism

    • Agents select parameters only where humans fail
    • No interpretation unless explicitly requested
  • Reason-after-reduction

    • Shrink the surface first
    • Reason on results, not the filesystem
  • Replayability

    • Every run produces artifacts
    • No silent overwrites
    • Explicit lifecycle controls

Skills

Skill Description Wraps Schema
grep Agent-assisted text search rg (ripgrep) aux grep --schema
find Agent-assisted file enumeration fd / fdfind aux find --schema
diff Deterministic git diff inspection git diff aux diff --schema
ls Deterministic directory state inspection filesystem + git status aux ls --schema

Schema is source of truth. Run aux <skill> --schema to get the current JSON schema for any skill.


Skill model

Each skill is fully independent and self-contained:

aux/
  skills/         # All skills live here
    grep/
    find/
    diff/
    ls/
  cli/            # Unified execution backend
  docs/           # Documentation
  scripts/        # Installation scripts

Each skill owns:

  • its schemas
  • its scripts
  • its artifacts
  • its lifecycle

They share conventions, not shared code.


Results

In controlled experiments comparing agent performance with and without aux skills:

Metric Without Skills With Skills Improvement
Context consumed 47% (182k tokens) 25% (98k tokens) -47% tokens
Files read 35 28 -20%
Search operations ~200-300 file scans 6 targeted passes -98%
Output quality Excellent Excellent Same

Key finding: Skills reduce token consumption by nearly half while maintaining output quality. The "search-before-read" discipline prevents agents from over-reading files they don't need.


Quick Start

# Install the CLI
cd cli && pip install -e .

# Verify dependencies
aux doctor

# Get schema for any skill
aux grep --schema
aux find --schema

# Run a search
aux grep "TODO" --root ./src --glob "*.py"

TLDR

aux makes classic Unix commands agent-operable
without sacrificing determinism, safety, or trust.

About

[AUx: Agentic Unix] A very particular set of skills.

Resources

License

Contributing

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published