The AI writes the code.
Two gates decide
if it counts.

# the floor alone: deterministic, runnable
$ hexa init ./myapp --scaffold --lang rust

# the floor plus what you described,
# gated on the build AND the grade
$ hexa scaffold "A bookmark service: SQLite store, HTTP API, tag search" \
    --target ./myapp --lang rust --grade A

$ hexa do run "make add() return a + b, not a - b" \
    --file src/lib.rs \
    --evidence "cargo test --test add"

$ hexa analyze .                        # grade + rule violations
$ hexa analyze . --grade A              # fail below A
$ hexa analyze . --grade A --exit-code  # and on any violation
$ hexa graph consumers <path>           # who depends on this

# adversarial: hunt by lens, verify each finding, fix under the gate
$ hexa harden src/domain/merge.rs --gate "cargo test -p mycrate"

# cooperative: diverge, red-team, synthesize, build to the gate
$ hexa build "a bounded work queue" --target src/queue --gate "cargo test"

# either one, past any terminal or tool timeout
$ hexa harden src/ --gate "cargo test" --detach

$ hexa loop adr ADR-2609131800               # the decision
$ hexa loop gate "cargo test --test add"      # before the code
$ hexa loop evidence "cargo test -- --nocapture"
$ hexa loop stage done                       # evidence into the ADR
$ hexa loop                                  # every session's state
$ hexa spend                                 # tokens and dollars

$ hexa doctor      # non-zero when a check fails
$ hexa go          # the next right thing
$ hexa --version   # the commit it was built from