relai regression select examines a proposed change and recommends a small set
of existing RELAI tests, custom adapter tasks, or individual benchmark samples.
It does not run them or change code.
--base defaults to HEAD. The selector reads the working tree diff and Git
status relative to that revision; use --change to supply intent, especially
when the change is only proposed or lives in an untracked file. Use
--agent-target in a multi-target project; it is required when multiple
targets are registered. --max-targets defaults to 3. The selector can read
the current project source and EvalsWiki context when available. Tests
without a declared agent target remain eligible for a named target; tests
declared for another target are excluded. A missing Wiki is reported as
unavailable and does not prevent selection.
For large benchmarks, RELAI considers at most ten rows per benchmark and
reports that limit as a coverage gap.
--result-json writes a versioned result with selected checks, reasons for
each choice, suggested simulation commands, and any coverage gaps. An empty
selection is valid when RELAI explains why no existing check is useful. A
coverage gap means the current test catalog may miss an affected behavior; it
is not a passing result.
With --agent-interaction, stdout carries one
relai.agent_interaction.v2 completion document pointing to the saved result.
RELAI writes that result automatically when
--result-json is omitted. The command never asks for approval or runs tests.
Selection keeps no session: Ctrl-C cancels it (exit status 130, with a
cancelled document under --agent-interaction), and when the
invocation timeout is reached it exits with status 124;
re-run the command to select again.
Run the suggested relai simulate command for each selected check. For a behavior
change, run the most relevant case before and after the fix when feasible;
after a direct fix, run the selected cases against the changed code. Inspect
simulation scores and trajectories before claiming the change is safe. If no
existing case measures the risk, consider adding grounded coverage. For a
change that needs repeated candidate generation and evaluation, use
relai optimize.