Low Impact
Examples: Typo fix, log message, formatting change.
Behavior: Direct action, no checkpoints. Skips validation against spec, comprehension checkpoint, reasoning log, and pre-PR review.
Every change is classified by impact before the agent acts. This determines the level of ceremony, verification, and approval required.
Low Impact
Examples: Typo fix, log message, formatting change.
Behavior: Direct action, no checkpoints. Skips validation against spec, comprehension checkpoint, reasoning log, and pre-PR review.
Medium Impact
Examples: New function, local refactoring, adding a validation.
Behavior: Plan + confirmation + automated review. The agent presents the approach and waits for the developer to confirm understanding.
High Impact
Examples: New service, schema change, public API modification.
Behavior: Mandatory ADR + explicit approval + full review + all checkpoints. Not eligible for streamlined mode.
flowchart TD
start["Change requested"] --> classify{"Classify impact"}
classify -->|"Typo, log, format"| low["Low Impact"]
classify -->|"New function, refactor"| medium["Medium Impact"]
classify -->|"Service, API, schema"| high["High Impact"]
low --> action["Direct action"]
medium --> plan["Present plan"] --> confirm["Developer confirms"] --> implement["Implement + auto review"]
high --> adr["Create/reference ADR"] --> approval["Explicit approval"] --> checkpoint["Comprehension checkpoint"] --> fullreview["Implement + full review"]
If a task classified as low turns out to be more complex during implementation, it is reclassified upward to medium. The agent pauses and applies the appropriate ceremony.
| Checkpoint | Low | Medium | High |
|---|---|---|---|
| Validation against spec | Skip | Apply | Apply |
| Comprehension checkpoint | Skip | Apply | Apply |
| Reasoning log entry | Skip | Optional | Required |
| Pre-PR review | Skip | Automated | Full review |
| ADR required | No | No | Yes |
| Trade-off explanation | No | Yes | Yes |
| Developer approval | No | Yes | Yes + explicit |