Predict the blast radius
before the merge.
Most AI coding tools see snippets. Cascade reads the entire repository, traces every cross-service call, follows env-vars into Docker, K8s, CI — and tells you exactly which lines will burn when your diff merges. Live analysis runs on **IBM Bob Shell** (non-interactive, full CASCADE-ARCHITECT prompt on the server).
What snippet AIs miss
Cross-service breaks
A signature change in `services/auth` silently rewires four callers in `services/billing`. The compiler shrugs. Cascade catches it.
Silent runtime failures
`SMTP_SERVER` → `MAIL_HOST` in code, but .env / docker-compose / K8s manifest disagree. No type checker sees this. Cascade does.
No false positives
Refactor `calculateTax` with identical contract? Cascade flags it MEDIUM, not CRITICAL. Trust depends on restraint.
Reference scenarios
Pipeline
- 01parseunified diff → changed-symbol candidates
- 02contextload cross-service source files
- 03analyzeinvoke IBM Bob Shell with cascade-architect
- 04classifyCRITICAL · HIGH · MEDIUM · LOW
- 05emitJSON · Mermaid · Jest tests · PR comment
Watch the agent think — live.
Run a preset and watch Bob Shell return the reasoning trace and strict JSON impact report. Engine · IBM Bob Shell.