# aiZee

> Sovereign AI engineering control plane — governance enforced on the action path, not advised around it.

## Docs

- [README.md](README.md): install, quickstart, command surface
- [spec.md](spec.md): architecture + contract
- [SKILL.md](SKILL.md): agent-facing usage guide
- [AGENTS.md](AGENTS.md): contributor + worker-agent contract
- [aizee_mcp/API.md](aizee_mcp/API.md): MCP tool reference

## Key paths

- `aizee_cli.py`: CLI entry point (~40 commands)
- `runtime/kernel.py`: gate facade (Probity → Guardian → Policy → Loop → Budget → Audit)
- `runtime/doctor.py`: structured health verdicts + next steps
- `runtime/gateway_proxy.py`: governed LLM proxy with provider failover
- `runtime/verification.py`: verification-depth dial (off/final/checkpoints/strict)
- `runtime/find.py`: NL discovery over skills/workflows/tech-stack
- `runtime/spec/changes.py`: OpenSpec-style change folders
- `runtime/tool_envelope.py`: uniform MCP response envelope + error codes
- `runtime/api_session.py`: governed API sessions replayed from curl/HAR captures (kernel-gated, same-host)
- `runtime/secret_scrub.py` + `runtime/sensitive.py`: capture-time secret scrubbing + placeholder injection
- `runtime/skill_lint.py` + `runtime/skill_scan.py`: skill-library lint/collision + prompt-injection scanner
- `runtime/context_composition.py`: per-call context-window composition telemetry (segment shares, utilization)
- `runtime/inventory_note.py`: plain-fact install inventory rendered for prompt injection (`aizee inventory`)
- `runtime/plugin_transactions.py`: lockfile-serialized plugins.yaml mutations (`aizee plugin enable|disable`)
- `aizee_mcp/tools/`: MCP tool modules (auto-discovered `*_tools.py`)

## Conventions

- `AIZEE_ROOT` env var points at the OS root; `config.discover_root()` resolves it.
- New runtime modules: `runtime/<name>.py`, wire through `kernel.py`, tests in `tests/`.
- MCP tools return the envelope `{schema_version:"aizee.tool.v1", ok, result, error:{code}, meta}`.
