| name | post-milestone-review |
| description | Read-only conformance audit of a completed milestone against CLAUDE.md acceptance criteria/invariants and the DESIGN.md contracts. Use after finishing a milestone, before starting the next. |
| allowed-tools | Read, Grep, Glob, Bash |
| argument-hint | [milestone-number] |
- Confirm $ARGUMENTS is recorded as done: it has a detailed
M$ARGUMENTS (done) entry in
MILESTONES.md and a matching one-line entry in CLAUDE.md's "## Completed milestones" index.
Do NOT rely on "## Current focus" — a finalized milestone should have been moved out of it by
/implement-milestone's finalize step. If either record is missing, stop and flag the actual
state rather than guessing:
- Still listed as active/next in "## Current focus" with implementation commits on the branch
or
main → it was implemented but never finalized (the finalize bookkeeping is missing). Say so
and point the user to finalize it first — this is the /implement-milestone finalize step,
not a re-implementation.
- Absent everywhere and no implementation commits → not implemented yet; run
/implement-milestone first.
- Present but numbered differently than the work actually shipped → the milestone number is wrong.
Post-Milestone Review: Milestone $ARGUMENTS
This is a read-only audit. Do not edit, create, or fix any files — report only.
- Re-read
CLAUDE.md and DESIGN.md fresh; don't rely on conversation memory of what they say.
- Run the test suite and
devtools::check() (use --as-cran for any release-track milestone, e.g.
the M3 CRAN-readiness work) and report actual results, not assumptions.
- For each acceptance criterion for Milestone $ARGUMENTS in
CLAUDE.md: state Met / Partially Met
/ Not Met, with the specific file/test as evidence.
- For each of the eight invariants in
CLAUDE.md: state whether the implementation upholds it,
citing the file/location. Pay special attention to:
- Invariant 1 — no
*_cpp function exposed as user API; all validation in the R wrapper.
- Invariant 2 — the surrogate null statistic exactly matches the observed statistic.
- Invariant 5 — any C++ change carries a numerical-regression test; confirm it actually
compares against a reference and would catch a divergence.
Flag silent violations even if no test currently catches them.
- Compiled-code hygiene specific to this package:
- No build artifacts (
src/*.o, src/*.so, src/*.dll, Rplots.pdf, .DS_Store) are tracked
(git ls-files).
RcppExports.cpp/RcppExports.R are clean against a fresh Rcpp::compileAttributes() (no
uncommitted reordering churn).src/Makevars* does not link OpenMP unless #pragma omp is actually used (the M2 decision).
- List untested edge cases implied by
DESIGN.md/CLAUDE.md that the current suite does not
exercise (e.g., short-series abort path, na.rm = FALSE, realized window length, NA-laden
surrogate windows, coupled-vs-independent surrogate p-value sanity, Granger degrees-of-freedom
guard).
- Dependency hygiene: anything newly in
Imports that should be Suggests (or vice versa); any
new dependency added without a flag; confirm heavy compute stayed in the C++ cores.
- Roxygen check: does documentation explain why defaults were chosen (per CLAUDE.md's
documentation standard), not just what — including the
suggest_wcc_params() heuristic and the
surrogate-null semantics.
- Milestone-log hygiene: confirm
MILESTONES.md is the only detailed log and is internally
consistent — entries run in numeric order with no gaps (M1..M$ARGUMENTS all present), the
CLAUDE.md "## Completed milestones" index has exactly one matching line per entry, and the
narrative is not duplicated into DESIGN.md §15 (which must remain the forward roadmap + a pointer).
Flag any missing, out-of-order, or duplicated entry — this is the class of drift that motivated
the single-log structure.
- Output a single triaged list: Blocking (must fix before the next milestone) / Should-fix /
Nice-to-have. End with one line: READY or NOT READY for the next milestone, and why.
Do not rewrite, fix, or refactor anything during this review — that's a separate step.
