Skip to content

SysOps Agent

Local-first, model-assisted system operations with deterministic execution controls, verification predicates, recovery loops, and a redacted audit trail.

Status: public alpha (0.1.0a1). SysOps Agent executes commands on the host. It is not production-ready, does not provide a security sandbox, and must not be run unattended or against untrusted objectives. Use an isolated disposable machine or VM.

CI Security License

What it is

SysOps Agent is an experimental Python control plane around two locally hosted GGUF models:

Role Default artifact Responsibility
Navigator Qwen3-4B-Q5_K_M.gguf planning, reasoning, verification, recovery
Coder Qwen2.5-CODER-3B-Q6_k.gguf bounded command and terminal-action authoring

Models run in separate llama-server processes. An optional escalation oracle can be reached over HTTP; it is disabled with SISTEMISTA_ORACLE=0. Remote oracle use sends redacted prompt context outside the host and therefore changes the privacy boundary.

The runtime wraps model output with deterministic controls: workspace confinement, command classification, typed success predicates, bounded recovery, credential redaction, audit events, and distinct process exit codes.

Current limitations

  • Only goals classified SAFE reach planning; review-required, malformed, or unavailable classification fails closed before any command runs.
  • The current policy is read-anywhere/write-workspace. Host package, service, registry, ACL, account, power, and process mutations are refused; there is no unconfined environment flag.
  • Windows is the actively exercised platform. Unix paths exist but are not yet certified.
  • Hardware capacity and throughput depend on the chosen quantization and llama.cpp backend; no minimum-VRAM or performance guarantee is made.
  • Model weights and llama-server binaries are external artifacts and are not bundled.
  • A successful run means declared verification predicates passed; it is not proof that every operational side effect was intended or complete.

Quick start

Prerequisites:

git clone https://github.com/PatrickDev-it/sysops-agent.git
cd sysops-agent
py -3.12 -m venv .venv
.\.venv\Scripts\python -m pip install --upgrade pip
.\.venv\Scripts\python -m pip install -e .
$env:SISTEMISTA_LLAMA_SERVER_BIN = "C:\path\to\llama-server.exe"
$env:SISTEMISTA_ORACLE = "0"
.\.venv\Scripts\sysops-agent "inspect disk pressure" --workspace C:\safe\disposable-workspace

On Unix-like hosts, activate .venv/bin/activate and use POSIX paths. This path is available for development, not yet part of the certified support statement.

The process exits with 0 only for COMPLETE; 1 means incomplete, 2 refused, 3 runtime error, and 130 operator interrupt.

Configuration

The important environment variables are:

Variable Default Purpose
SISTEMISTA_LLAMA_SERVER_BIN repo binary or PATH explicit llama-server executable
SISTEMISTA_VAR_DIR workspace/var relocatable runtime state root
SISTEMISTA_NAV_PORT 8090 navigator server port
SISTEMISTA_CODER_PORT 8091 coder server port
SISTEMISTA_ORACLE 1 enable optional escalation oracle
SISTEMISTA_ORACLE_URL http://127.0.0.1:8081 oracle endpoint
SISTEMISTA_ORACLE_TOKEN unset bearer token required for a remote oracle

Non-loopback oracle endpoints must use HTTPS and authentication. Runtime state, weights, and binaries are ignored by Git.

Architecture

objective
   |
   v
preflight -> safety classification -> environment discovery
   |                                  |
   v                                  v
planner/navigation model -> typed plan -> deterministic command gates
                                              |
                                              v
                                      PTY/subprocess runtime
                                              |
                                              v
                                 predicates -> verify -> recover
                                              |
                                              v
                                  redacted trace + telemetry

Source is under workspace/src; hermetic regression tests are under workspace/tests; benchmarks and their evidence are under workspace/benchmarks and workspace/validation. Architecture decisions and publication gates are recorded in .sinapsi/rfc.

Development

py -3.12 -m venv .venv
.\.venv\Scripts\python -m pip install -e ".[dev]"
.\.venv\Scripts\ruff check .
.\.venv\Scripts\ruff format --check .
.\.venv\Scripts\mypy
.\.venv\Scripts\pytest -q
.\.venv\Scripts\python -m build

CI runs lint, formatting, static typing, tests on Windows and Ubuntu, package builds, dependency audit, and a full-history Gitleaks scan. See CONTRIBUTING.md before opening a change.

Branch and release policy

  • development: integration branch and current public alpha.
  • validation: candidates that have passed the secure-execution gate.
  • production: release-only branch after every publication RFC is closed.

The repository being public does not imply production readiness. There is no stable release until the validation evidence and release supply chain meet the promotion gate.

Security and privacy

Read SECURITY.md before deployment or vulnerability reporting. SysOps Agent is local-first, but an enabled remote oracle receives redacted operational context. Redaction is a defense-in-depth control, not a guarantee that arbitrary sensitive content cannot leave the host. Confinement is an application policy, not kernel isolation.

License

Apache License 2.0. External model weights and llama.cpp retain their own licenses and are not distributed by this repository; see NOTICE.

About

Local-first AI agent for bounded system operations, verification, recovery, and auditable execution.

Topics

Resources

Code of conduct

Contributing

Security policy

Stars

Watchers

Forks

Releases

Packages

Used by

Contributors

Languages