Existing proposals
Problem
Agent Zero should not depend exclusively on reviewer feedback to discover defects. After the v0.1 repair loop is reliable, it needs a proactive review mode that inspects pull-request changes, verifies suspected defects, and decides whether a finding is safe enough to fix automatically.
Parent roadmap: #5
Desired outcome
Deliver the v0.2 proactive review and autofix milestone.
Acceptance goals:
- inspect pull-request diffs proactively
- produce findings with severity, confidence, and evidence
- verify suspected defects before proposing changes
- support confidence-gated autofix policies
- distinguish safe mechanical fixes from changes requiring approval
- add configurable repository-level autofix policy
- improve multi-file reasoning and regression detection
Likely area
Agent runtime
Safety and compatibility
High impact because proactive fixes can mutate code without an explicit review comment requesting the change. Autofix must be opt-in, policy-controlled, bounded by confidence and repair budgets, and verified using repository-native checks. Low-confidence or high-impact changes must require human approval.
Alternatives considered
Always-autofix was rejected because confidence alone cannot make every code change safe. Review-only proactive findings are safer but do not meet the roadmap goal of bounded autonomous repair, so the intended design combines evidence-based findings with explicit policy gates.
Existing proposals
Problem
Agent Zero should not depend exclusively on reviewer feedback to discover defects. After the v0.1 repair loop is reliable, it needs a proactive review mode that inspects pull-request changes, verifies suspected defects, and decides whether a finding is safe enough to fix automatically.
Parent roadmap: #5
Desired outcome
Deliver the v0.2 proactive review and autofix milestone.
Acceptance goals:
Likely area
Agent runtime
Safety and compatibility
High impact because proactive fixes can mutate code without an explicit review comment requesting the change. Autofix must be opt-in, policy-controlled, bounded by confidence and repair budgets, and verified using repository-native checks. Low-confidence or high-impact changes must require human approval.
Alternatives considered
Always-autofix was rejected because confidence alone cannot make every code change safe. Review-only proactive findings are safer but do not meet the roadmap goal of bounded autonomous repair, so the intended design combines evidence-based findings with explicit policy gates.