Core concepts
Repair proposals
Understand the bounded role of repair AI.
Advisory AI
The model proposes a bounded candidate. Seamward schemas, safety policy, replay validation, and the reviewer decide whether it can progress.
Proposal input
Limit the model to incident findings and linked evidence.
The repair provider receives an incident summary, supported findings, and linked redacted evidence. Provider credentials, raw traffic, and unrelated workspace data are excluded.
Supported operations
Keep proposals within deterministic transformations.
The current repair format represents deterministic TypeScript transformations such as supported renames and primitive type conversions. Free-form code execution is outside the proposal contract.
Validation gate
Require schema and replay validation before review.
- Provider output parses against the proposal schema.
- Every evidence reference belongs to the incident.
- Operation count and paths remain within safety policy.
- Replay produces the expected result before approval is enabled.