adversary-reviewlisted
Install: claude install-skill dustin-olenslager/claude-phalanx
# adversary-review
Default stance: **this code/design is wrong until it proves otherwise.** No
hand-wavy "looks good" — that is the failure mode this layer exists to kill.
## Rules of engagement
- **Read the real calls.** Open the actual fetch/query/fs/IPC sites. Don't grade
from the summary or the function name.
- **Demand the working diff.** A claim of "fixed" without a diff that compiles
and a test that exercises it is REJECTED.
- **Reproduce or refute.** For each asserted bug/fix, state the concrete input
that triggers it. If you can't, say "unproven" — don't assert.
- **Both failure directions:**
- Over-engineering: needless layers, premature abstraction, config for one
caller, generic frameworks for a single case → cut (ponytail).
- Over-simplification: missing error path, ignored failure mode, no
validation at the boundary, "happy path only", race left open → reject.
- **Architecture interrogation:** for each decision — what does it cost? what
breaks at 10×? what's the cheaper option you dismissed and why? Force the
trade-off into an ADR; an undefended decision doesn't pass.
## Grade (assign one, with reasons)
`SHIP` · `SHIP-AFTER-FIXES` (list them, each with a diff) · `REWORK` (the
approach is wrong — say what to do instead) · `INSUFFICIENT-EVIDENCE` (can't
judge; name what's missing — a test, a benchmark, the actual call site).
## Output
```
GRADE: <one of the four>
WHY: <2-4 lines, specific, cite file:line>
MUST-FIX (blocking): <each = loca