ai-safe2-framework

Guided Start

safe2 start turns the first local AI SAFE² assessment into one previewable, consent-bound workflow. It is designed for a person delegating to an agent and for an agent that needs a clear stopping point before it changes anything.

Quick start

Preview the plan. This validates the target’s filesystem metadata and, when present, reads and validates the bounded local AI SAFE² configuration to show the effective inspection scope. It writes no files:

safe2 start .

Execute after reviewing the target, inspection consent, output, and boundaries:

safe2 start . --execute

The command asks for confirmation. An authorized non-interactive caller can use:

safe2 start . --execute --yes

--yes means only that the displayed local plan was authorized. It does not authorize remediation, network access, deployment, publication, or acceptance of risk.

Choose inspection scope

Option Local behavior Default
no inspection flags Inventory recognized local metadata and security assets enabled
--scan-content Read bounded project content for static analysis disabled
--inspect-config Emit allowlisted, redacted configuration structure disabled; requires --scan-content
--wsl Inventory WSL distribution names when available disabled

No option in this command enables remote or cloud access. If an existing validated project configuration already enables content collection, the preview shows the effective read as enabled and identifies its source as configuration. Execution is bound to that configuration digest and refuses to proceed if it changes after preview.

Outputs

The default destination is .safe2/onboarding. Select another new directory with --output-dir. Existing destinations are rejected.

Artifact Audience Meaning
onboarding-plan.json agent/reviewer Exact approved scope, consent, steps, and stops
onboarding-result.json agent/reviewer Run outcome, assessment summary, authorization boundary, and integrity digest
next-step-card.md human Short action path and explicit boundary
configuration-snapshot.toml reviewer Exact bounded configuration approved for this run
assessment/assessment.json agent/reviewer Canonical assessment, coverage, limitations, and next actions
assessment/decision-card.md human Detailed readable assessment
assessment/environment.json agent/reviewer Local discovery and posture evidence
assessment/project-scan.json agent/reviewer Consented scan or explicit not_requested state
assessment/manifest.json verifier Hash-bound assessment artifact inventory

completed means the approved workflow finished and wrote valid artifacts. It does not mean the project is safe, compliant, conformant, or ready to deploy.

Recovery and repeat runs

The project configuration is created first because the assessment consumes it. If later assessment work fails, the secure configuration may remain and will be reused on the next run. Partial evidence is staged and removed rather than published. A successful bundle is never overwritten; choose a new output path for another run.

Agent handoff

Run `safe2 start .` and show me the plan. Distinguish facts, requested reads,
missing coverage, and actions the command will not take. Do not add `--execute`,
`--yes`, `--scan-content`, `--inspect-config`, or `--wsl` without my explicit
approval. After execution, show me both Decision Cards and the canonical JSON
paths. Stop before remediation, remote access, deployment, publication, policy
change, exception, or risk acceptance.