Blast radius measures how far a failure or harmful change can propagate through systems, data, or operations. A change that is correct inside one package can still break its consumers when the dependency graph is incomplete or ignored. section 10.2.1
The shared-utils failure
An agent changed shared-utils, passed that package's tests, and submitted a correct local diff. Nobody tested notification-service or the five other importing packages, so a load-related failure surfaced during the 11 PM batch run. The local change was small; its operational blast radius was not. section 10.2.1
Reduce blast radius by making relationships and risk explicit:
- Keep dependency graphs complete and enforce strict dependencies so agents can navigate relationships and identify affected consumers. section 10.2.6
- Test beyond the edited package when downstream consumers may be affected; isolated green tests do not establish contextual safety. section 10.2.1
- Use sandboxing and staged controls for agents, accepting friction to limit the consequences of mistakes. section 5.8
- Classify changes by module risk, ownership, downstream reach, and sensitive surfaces before allowing automatic merging. section 10.5.2
Listing 10.4 Policy-based ownership with triage rules
libs/core: owner: platform-team green: "coverage > 95%, diff < 50 lines, no new dependencies" yellow: "platform-team-oncall reviews" red: "platform-team-lead + security-review" services/checkout: owner: payments-team green: "tests pass, no schema changes, no auth modifications" yellow: "payments-team-member reviews" red: "payments-team-lead reviews" sla: "99.99% - minimum Yellow review for all changes" services/notification: owner: platform-team green: "tests pass, no external API changes, diff < 100 lines" yellow: "any platform-team member reviews" red: "platform-team-lead reviews"
Use policy-based ownership rules to distinguish low-risk changes from changes requiring team or security review. section 10.5.2







