⚙️ FEATURE: Graduated autonomy levels replacing binary dangerous=True/False

[FernandoCelmer]

Problem

pycodeloop tools use a binary dangerous: bool flag. This forces an all-or-nothing choice: either all dangerous tools require human confirmation, or none do. There is no way to allow low-risk mutations autonomously while still requiring approval for high-risk ones.

Expected behavior

codeloop = CodeLoop(config=Config(
    autonomy="safe_execute",  # allow reads + low-risk writes, require approval for high-risk
))

Policy table:

Level	read	execute_low_risk	execute_high_risk
manual	allow	require_approval	deny
safe_execute	allow	allow	require_approval
full_project_loop	allow	allow	allow

Suggested implementation

• Add AutonomyLevel enum in pycodeloop/core/autonomy.py
• Add operation: str class attribute to tools: "read", "execute_low_risk", "execute_high_risk"
• Replace tool.dangerous → confirm check in Agent._execute() with gate(level, tool.operation) -> GateDecision
• Add autonomy: str = "safe_execute" to Config
• Record gate decisions in JSONL trace

References

• "Code as Agent Harness" — arXiv 2605.18747, Section 3.1 (Harness interface)