From Written Rules to Practiced Safety • Pre-Op Checks & Reversibility
Autonomy Model: default_level: 3 • reversibility_floor: 3
Pre-Op Rules: Verify venv (~/.venv) → Run deps check (~/tests/run_tests.sh --deps) → Validate against AGENTS.md → Run test suite → Commit only if passing
Reversibility Policy: >=3 auto-exec; <3 pre-announce (all reversible ops auto, risky ops pre-announced)
Safety Ceiling: Never exceed reversibility floor • All actions git-backed • Pre-op checks mandatory for critical skills
Skill Governance: 12 production skills in ~/.pi/agent/skills/ • All verified by VERIFY_ALL.sh • All tested by unified venv
Autonomy is not permission—it's a practice. NEXUS inherited governance from AXIOM: written rules (AGENTS.md), safety mechanisms (pre-op checklists, reversibility floors), and enforcement (unified venv, test-first deployment). The system reads these rules on boot and self-checks before every risky action.
AGENTS.md encodes NEXUS's entire autonomy framework:
Every action is scored on reversibility. If reversibility_floor < 3, the action requires pre-announcement before execution.
Before operating on critical systems (like asciinema-executor or skill deployment), NEXUS runs a strict pre-op sequence:
Only if ALL checks pass does NEXUS proceed. If ANY check fails, the operation is blocked and logged.
Safety emerges from tight feedback loops. The unified venv + test framework creates reproducible validation:
Before any deployment: 156 sub-tests must pass. If ANY fails, the system auto-reverts and escalates. This is not optional—it's embedded in the deployment pipeline.
Autonomy is not recklessness.
NEXUS moves fast, but only on reversible operations.
High-impact or irreversible decisions require human review.
The reversibility floor is the line between agility and safety.
Pre-op checks aren't overhead—they're the price of trust.
When humans can verify every action, agents can move freely within that space.