vc-predict

Run vc-predict with a description of the proposed change. Five personas covering architecture, security, performance, UX, and hidden assumptions analyze it independently, then debate conflicts and return a GO, CAUTION, or STOP verdict with mitigations.

Simple mode reasons from the proposal description alone. Deep mode first spawns a research subagent that scans git log, prior phase reports, and test failure history, so personas can reference concrete past attempts and known landmines.

Use deep mode when the approach touches surfaces with failure history such as auth, billing, container lifecycle, or streaming, when a similar pattern was attempted before, or when the plan is complex. It also triggers via the --deep flag.

Skip it for trivial or low-risk changes, already-approved work with no open design questions, and pure dependency upgrades without API changes. Use a debugger skill for bugs and a planning skill for already-decided tasks.

A STOP verdict occurs when security finds an unmitigable auth bypass or data exposure, the architect finds a fundamental design incompatibility, performance finds unacceptable latency with no workaround, or a false assumption invalidates the entire approach.