Skip to content

About

Provider-neutral AI workspace framework: shared agent contracts, provider discovery, bounded runtime, and verification.

Topics

Resources

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

14 Commits

Folders and files

Repository files navigation

System Agentic AI

Provider-neutral AI workspace framework: shared operating contract, agent brain, bounded runtime, provider discovery, host skills, cross-platform bootstrap, and verification. It is a repository framework, not a GUI, SaaS product, dashboard, model bundle, account manager, or 9router distribution.

Introduction film

Watch the 54-second English introduction or the lightweight GIF preview. The film walks through a safe provider-readiness prompt, the decision-plane result, live provider inventory, READ_ONLY boundaries, and a verification receipt.

The composition is reproducible from media/, with captions, a transcript, and sanitized evidence readbacks. JSON values come from a local run pinned to public commit 7b7168a; the advisory result did not authorize or submit a task. Connector geometry and policy panels are labeled illustrative in the film. No credentials, account data, or external writes are included.

Optional paired browser extension

The browser module now supports an opt-in Chromium extension transport with explicit tab/origin pairing, native click/fill, sequential batches, navigation fencing and non-replayed uncertain receipts. GitHub, Notion and generic pages share a bounded semantic contract. Public reads and local editor fixtures were verified; authenticated private-site writes are not certified. No service, account access or model fallback is enabled by cloning this repository. See installation, permissions, evidence and limits.

First run

macOS, Linux, or Windows with Python 3.10+ and Git:

git clone https://github.com/NamAIEngineer/system-agentic-AI.git
cd system-agentic-AI
python3 scripts/setup.py --check-only
./scripts/setup.sh
./scripts/doctor.sh

Windows PowerShell:

git clone https://github.com/NamAIEngineer/system-agentic-AI.git
cd system-agentic-AI
py -3 scripts/setup.py --check-only
powershell -ExecutionPolicy Bypass -File .\scripts\setup.ps1
powershell -ExecutionPolicy Bypass -File .\scripts\doctor.ps1

Then open the cloned folder in ORCA or another compatible AI coding host and ask the host to read AGENTS.md, README.md, WORKFLOW.md, and agent-brain/operating-contract.md before initializing the workspace.

In ORCA, use this single first-run instruction:

Read AGENTS.md and fully initialize this System-Agentic-AI workspace.

ORCA auto-loading is host-specific and remains UNVERIFIED without host evidence. This sentence is the supported fallback. The host must show approval plans before network/model downloads, dependency execution, or writes outside the checkout.

setup.py and doctor.py do not install global software, modify shell profiles, change credentials, start services, or require ORCA/9router. The default setup path is local and offline. Reference repositories are optional and require an explicit command:

python3 scripts/install-references.py --check-only
python3 scripts/install-references.py --install --all --allow-network

Use --skip-references in CI. Never point the installer at an existing dirty checkout owned by another project.

The installer keeps the seven references outside Git and emits independent counts for source cloning, pinned revision verification, capability registration, and actual execution verification. A cloned or pinned reference is not automatically execution-ready. NOASSERTION_REVIEW entries require license/ownership review before running their code.

Quyet Tiny is an optional local advisory classifier:

python3 scripts/setup-quyet.py --check-only --json
python3 scripts/setup-quyet.py --verify --json
python3 scripts/setup-quyet.py --install --allow-network --allow-model-download

The install command requires separate approval for public downloads and model weights. Weights remain in user cache, outside Git. Quyet output never grants execution permission; unavailable or low-confidence results fall back to host reasoning.

What the framework provides

  • agent-brain/: provider-neutral contract, provider/module registries, decision plane, bounded adapters, tool runtime, and host skill.
  • products/agent-runtime/: optional local queue/worker source. Runtime SQLite, logs, model weights, and service state stay outside Git.
  • packages/agent-core/: reusable TypeScript capability/risk contracts.
  • product-kits/: generic scaffolds, not user projects.
  • scripts/: setup, doctor, reference installer, provider inventory, runtime CLI, browser helpers, and smoke checks.
  • docs/: architecture, host boundaries, provider status, configuration, troubleshooting, and source reports.

Host/provider support

Codex/ORCA is the only execution adapter verified on the owner machine. Claude and Gemini can consume Markdown instructions but have no included execution adapter. Local Qwen, REA, browser, ChatGPT bridge, and agent-desktop are optional integrations whose live readiness must be checked on the user's host. Run:

python3 scripts/system-agentic providers
python3 scripts/doctor.py --json

Doctor reports core_status, references_status, quyet_status, per-provider optional status, and host_execution_status. Missing optional providers do not downgrade a healthy core. A real host task is required before claiming execution readiness.

agent-desktop is optional and never bundled as a binary:

./scripts/setup-agent-desktop.sh --install
./scripts/setup-agent-desktop.sh --check

9router is an optional external integration. System Agentic AI does not require 9router. Do not copy its config, account pool, credentials, or runtime state.

Runtime task example

After a user has deliberately configured a compatible local worker:

./scripts/agentic-runtime providers
./scripts/agentic-runtime submit --workspace "$PWD" "Inspect this workspace and report evidence."

Explicit Codex processing is per-task and uses the user's own host login:

./scripts/agentic-runtime submit --backend codex --workspace "$PWD" "Run this task through Codex."

No provider is silently selected as a fallback. An unavailable optional provider returns AUTH_REQUIRED, NOT_CONFIGURED, UNAVAILABLE, UNSUPPORTED, or UNVERIFIED; it does not block the static framework.

Safety and data boundary

Keep credentials, cookies, browser profiles, account homes, model caches, queue databases, logs, private projects, plans, and generated artifacts outside Git. The framework never inherits the owner's credentials. Read SECURITY.md and docs/configuration.md before enabling a provider.

Verification

./scripts/smoke-test.sh

The smoke test is deterministic and offline. It checks Python syntax, manifest and registry shape, setup/doctor check-only paths, shell syntax, symlink safety, and CLI help. It does not claim that ORCA, a browser, 9router, a model, or a cloud account is authenticated.

License

Original System Agentic AI code in this repository is MIT-licensed where the owner has distribution rights. Reference repositories remain separate and retain their own licenses. See LICENSE and THIRD_PARTY_NOTICES.md.

About

Provider-neutral AI workspace framework: shared agent contracts, provider discovery, bounded runtime, and verification.

Topics

Resources

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages