CLI¶

Entry point: aire → aire.cli.main:app (Typer).

Command	Purpose
`aire init <name>`	Scaffold project (`aire.yaml`, `app.py`, `docs/`, `evals/`)
`aire run <prompt>`	One-shot generation (`--model`, `--config`)
`aire evaluate <dataset>`	Eval suite (`--model`, `--metrics`, `--output`)
`aire serve`	Serve `build_target()` from `app.py` via FastAPI/uvicorn
`aire gateway`	OpenAI-compat gateway (`-m/--model`, `-a/--alias`, `--embed-alias`, `--routing`, `--objective`, `--auth-token`, `--rate-limit`)
`aire mcp-serve`	Builtin + registered tools over MCP stdio
`aire inspect [what]`	`models` \| `tools` \| `plugins` \| `config` \| `all`
`aire plugins`	List discovered/loaded plugins
`aire doctor [--smoke] [--live] [--output PATH] [--config PATH]`	Versioned diagnostic report; offline smoke or explicit live probes
`aire scaffold <kind>`	Recipe snippet: `rag` \| `agent` \| `finetune` \| `gateway` \| `workflow`
`aire version`	Print version
`aire deploy scale`	Docker Compose + K8s scale pack
`aire foundation`	Toy architecture only (`--describe` for catalog) — not pretrained weights
`aire analytics`	Analytics report (demo counters unless process metrics / `AIRE_ANALYTICS` attached)

Examples:

aire init mybot --model mock:echo
aire doctor
aire run 'hello' --model mock:echo
aire gateway -m mock:echo
aire foundation --describe

Diagnostic evidence¶

aire doctor --smoke --output doctor.json exercises mock generation, local hashing embeddings, a real tool and cited retrieval without network or model downloads. The JSON report records version, Python, healthy, and explicit passed, failed and skipped statuses. The saved report matches stdout; required failures exit 1. Invalid configuration is reported without echoing its contents. --config requires the requested file/directory to exist.

Default doctor checks availability/configuration without executing model calls. Unrelated missing credentials and optional serving dependencies are skipped. --live adds the offline smoke, probes OpenAI/Anthropic only when a resolved configuration/environment credential exists, and probes local Ollama. Cloud health probes are bounded to ten seconds each and may incur usage; errors omit provider bodies and credentials. Client resources are closed after probing.

aire inspect config excludes provider settings and masks common secret-bearing keys recursively, including gateway bearer tokens. Add custom keys to observability.mask_fields; masking cannot detect secrets in arbitrary text. Unknown inspection selectors exit 2.