AutoFACS CLI Automation is the operational layer behind the broader AutoFACS project: a FACS (Facial Action Coding System)-grounded computer-vision and machine-learning system for facial expression recognition. The CLI automation side exists to make those workflows reproducible, auditable, and safe to run at scale.
This repository documents the Gemini CLI–oriented automation architecture used to govern that work. It shows how sensitive CV/ML tasks can be executed through bounded prompts, explicit task templates, named-file scope, visible session logging, storage-safe workflows, and human review gates.
Rather than acting like a generic "AI agent," this system is designed as a governed control plane for CV/ML operations: scraping and ingestion support, curation and triage, manifest and storage maintenance, supervised training orchestration, evaluation loops, and inference-facing operational checks.
ML workflows drift easily. Scripts accumulate hidden assumptions, storage boundaries get blurred, evaluation steps get skipped, and CLI agents can overreach if their scope is not tightly controlled.
AutoFACS CLI Automation is an attempt to solve that problem by giving Gemini CLI a disciplined operating model for sensitive CV/ML work:
- task-class prompt templates instead of one free-form agent prompt
- explicit source boundaries
- mount-safe and storage-aware execution
- required log proof
- non-destructive change handling
- human review before important state changes
Start here, then use the linked docs as a guided set instead of reading the repository file by file.
- Architecture overview — how the control plane is structured
- Workflow patterns — concrete execution patterns for audit, implementation, and ops work
- Prompt library — the public-facing design of the Gemini CLI prompt system
- Prompt evolution and governance — why the prompt library was split into task classes
- Public-facing boundaries — what this repo shares directly and what it abstracts
- Docs index — quick navigation across the documentation set
- Example configuration — a boundary-aware public config example
This is a public-facing architecture and workflow repository for the automation track. It focuses on the design patterns and operating discipline behind the system, including:
- Prompt-governed CLI execution for Gemini CLI sessions
- Task-template discipline for audit, corrective documentation, implementation, refactor, and governance checks
- Storage-safe automation around manifests, DVC boundaries, and remote resources
- Audit-first execution with visible proof of what was done
- Human-in-the-loop review for decisions that should not be silently delegated
The automation track supports the broader AutoFACS pipeline, including work such as:
- scraping and ingestion support
- quality gates and curation workflows
- storage and manifest refresh logic
- supervised training hardening
- evaluation and promotion loops
- inference and active-learning support
In other words, the CV/ML repository contains the domain logic and model-facing work; this repository documents the operational framework that helps that work run in a controlled and automated way.
The prompt library is organized as a small set of governed task templates rather than one free-form agent prompt.
-
Shared forward
Common front matter for all CLI sessions. It establishes single-mode execution, safe-halt behavior, task splitting, named-file scope, and the rule that task-local prompt contracts are allowed to be narrower than the global runtime rules. -
Audit / evidence
Used when the goal is to inspect, classify, and report facts only. No edits are allowed. -
Corrective documentation / tracker
Used when the goal is to correct overstated wording, downgrade certainty, or append historical correction notes without performing new engineering work. -
Implementation
Used when the goal is to make one narrow change after the necessary evidence has already been established. -
Refactor
Used when the goal is to improve a narrow code surface without changing external behavior. -
Governance probe
Used when the goal is to verify that the CLI is actually loading the expected governance surfaces at run time.
Each session is expected to follow a controlled lifecycle rather than free-form agent behavior:
- Protocol Gate — acknowledge runtime and safety rules
- Preflight Compliance — confirm environment state, scope, and constraints
- Execution — perform only the approved task
- Log Proof — verify physical session logs before closure
This model is meant to reduce "silent success" claims, hidden drift, and unclear state transitions.
The automation track favors Git-first history and managed archival over destructive cleanup.
- active files are versioned rather than casually replaced
- legacy artifacts are moved to archive-style locations rather than deleted
- session logs and operational evidence are preserved as part of the forensic trail
- fragile storage boundaries are treated as real operational constraints, not ignored as implementation detail
Representative principles from the AutoFACS automation approach include:
- Do not treat fragile remote storage like a normal local folder
- Prefer explicit manifests and known boundaries over recursive discovery
- Do not mark work complete without evidence
- Do not mix diagnosis and implementation in one uncontrolled session
- Do not allow agent convenience to outrank operational safety
This project is in the transition from a useful but fragile collection of scripts and prompts to a more reliable automation framework for real CV/ML operations. The current direction is toward a closed-loop system with stronger observability, safer storage interaction, disciplined training and evaluation flows, and better agent compliance.
.
├── assets/
│ └── social-preview.png
├── docs/
│ ├── README.md
│ ├── architecture.md
│ ├── example-workflows.md
│ ├── prompt-evolution.md
│ ├── prompt-library.md
│ └── public-safe-boundaries.md
├── examples/
│ └── example_config.py
├── REPO_TOPICS.txt
└── README.md
This repository documents the architecture and operating model, not every private implementation detail. It does not publish private datasets, credentials, internal infrastructure details, or non-public model artifacts.
No open source license is granted for this repository at this time.
Unless otherwise stated, all rights are reserved. Publication on GitHub does not grant permission to use, modify, or redistribute this material outside the limited rights required for viewing and platform forking under GitHub’s Terms of Service.
- AutoFACS CV/ML — model, data, training, evaluation, and inference work
- AutoFACS CLI Automation — governed Gemini CLI operating layer for those workflows