eval

Overview ¶

Package eval implements the `da eval` command tree: the operator surface over the R4 agent-evaluation harness. It mirrors the `da score` command-surface idioms (subcommands under a group, a --repo-dir root resolver, --json output) and is pure CLI wiring — every heavy stage lives behind the internal/eval seam packages (generator registry, sandbox, runner, verifier, scoring bridge, store).

• `da eval gen` synthesises one TaskSpec from the knowledge graph.
• `da eval run` drives one task end-to-end (gen → sandbox → run → verify → score → persist) and prints the R1-scored outcome.
• `da eval ls` lists the persisted eval runs under the eval root.

The gen/run/ls RunE handlers are injected by the root command rather than defined here (see NewCmd). internal/globalflagcov statically traces each handler's global-flag reads by indexing a fixed set of command packages that does not include this subpackage, so a RunE closure defined here is opaque to it ("unresolved closure"). Wiring the handlers from package commands — where they read the global --json flag through a traceable commands.Flags access — keeps global-flag coverage analysable while the run logic stays here.