ai-safe2-framework

AI SAFE² CLI: Post-1.0 Capability Priorities

What 1.0 includes, what it deliberately leaves out, and what should be built next

AI SAFE² CLI

Framework Home CLI Guide Operator Workflow Stability Policy Examples

The 1.0 value proposition

CLI 1.0 is a local-first evidence and decision-support layer for agent systems. It gives agents a stable machine interface and gives people readable evidence without granting the CLI deployment, exception, or risk-acceptance authority.

Included in 1.0

Deliberately not included

These exclusions are security and evidence boundaries, not missing marketing claims. A post-1.0 feature should narrow a boundary only with explicit consent, least privilege, negative tests, and truthful coverage reporting.

Priority order

Each priority is a capability-sized pull request. Progression is based on user value and prerequisite relationships, not a calendar.

Priority 1 — Guided agent onboarding and orchestration

Status: Implemented in CLI 1.1.0 through safe2 start; retained here as the acceptance record for the capability.

Problem: 1.0 has the necessary primitives, but a first-time user or agent must still assemble several commands and source artifacts.

Build: Add a safe2 start or equivalent guided workflow that selects an outcome, checks installation, initializes safely, discovers the environment, runs the bounded assessment, and writes a next-step card. It must preview the plan, require explicit consent for content/config inspection, refuse overwrite, and stop before remote access or remediation.

Done when: A clean-install evaluator can move from “point my agent at this repository” to a reviewable local assessment without reading the full command reference, and every skipped or unauthorized evidence domain remains visible.

Delivered boundary: The command previews without writing, requires --execute plus interactive confirmation or explicit --yes, preserves content/config inspection as separate consent, reuses but never replaces an existing project configuration, publishes into a new evidence directory, and performs no remote access or remediation.

Priority 2 — Harness admission-hook kit

Problem: Polling and manual skill gates detect change, but cannot reliably block a skill, tool, or configuration before a harness activates it.

Build: A small, versioned hook protocol and reference admission service for pre-install, pre-load, pre-tool, and post-task events. Keep the native gate synchronous and bounded; allow attributed third-party scanners to add evidence without becoming the decision authority.

Done when: At least two materially different harnesses can call the same protocol, hostile controls are blocked before activation, unavailable optional providers degrade honestly, and no hook gains broader authority than the host harness already granted.

Priority 3 — Maintained native adapter pack

Problem: Provider-neutral intake works, but users still write translation glue and can misstate what a harness export proves.

Build: Supported exporters or translators for the highest-adoption harnesses, starting with documented, stable export or hook surfaces. Candidate coverage includes Codex, Claude Code, Hermes, OpenClaw, Antigravity, and common agent/CI telemetry. MiroFish, TinyFish, bots, Muse/Dot-style systems, and other runtimes enter through the same conformance kit when a stable authorized interface exists.

Done when: Each adapter records provider and version, native source type, collection scope, missing domains, redaction behavior, and exact-source binding; passes shared conformance fixtures; and never claims access it did not have.

Priority 4 — Explicit remote and cloud inventory connectors

Problem: A workstation assessment does not provide a complete picture when agents also run in containers, remote hosts, CI, or cloud services.

Build: Opt-in connectors for named targets and least-privilege control-plane exports. Start with containers and CI artifacts, then cloud inventories. Never scan a network or discover accounts ambiently.

Done when: Local, WSL, container, CI, remote-host, and cloud evidence can be joined into one system identity while unreachable, unauthorized, stale, and not-applicable coverage remain distinct.

Priority 5 — Technology Contribution Profile automation

Problem: The repository has a strong human method for deciding what a tool contributes, but no CLI schema or command yet produces the reusable Technology Card.

Build: Versioned TCP source and result schemas, conservative evidence mapping, card rendering, comparison, and manifest integration. Preserve native scores and keep technology contribution separate from organizational AISM maturity and Challenge Lab C0-C5.

Done when: A user can compare alternatives by control fit, enforcement location, evidence strength, failure behavior, retained authority, integration cost, and residual risk without converting a vendor claim into conformance.

Priority 6 — Challenge Lab adoption and independent replication kits

Problem: Challenge 001 is executable, but independent participants still need easier packaging, submission validation, and result comparison.

Build: One-command participant bundles, adapter templates, offline verification, reproducible reports, and public result metadata that preserves raw divergence. Add new challenges only after the participation workflow is proven.

Done when: A third party can run or translate its own harness result, publish a verifiable evidence bundle, and compare it without changing the frozen challenge or claiming that translated agreement proves independence.

What should not be the next build

Release sequencing rule

Ship a capability only when its user journey, contract, negative cases, security boundary, recovery path, documentation, and clean-install acceptance are complete. Integration breadth is valuable after the common evidence layer is stable; the moat is comparable, reproducible evaluation across systems—not the number of logos in a README.