Skip to content

feat(judge): support crush as a judge CLI [7/9] - #31

Open
LarsArtmann wants to merge 7 commits into
cosmtrek:masterfrom
LarsArtmann:judge-crush-cli
Open

LarsArtmann wants to merge 7 commits into
cosmtrek:masterfrom
LarsArtmann:judge-crush-cli

Conversation

@LarsArtmann

Copy link
Copy Markdown

What

The evaluation judge can now run through the crush CLI, alongside claude and codex:

  • SupportedCLIs gains crush; auto-detection and the UI's judge picker pick it up
  • crush run --quiet --verbose reads the prompt from stdin (large evidence documents never hit argv limits); --verbose stderr names the model that answered, parsed like the codex preamble
  • the run is pinned to the judge workdir; the Crush adapter (feat(crush): session adapter for Crush [3/3] #27) marks sessions recorded there auxiliary via the existing judge.IsWorkDir belt
  • judge/adapter home-dir resolution unified through adapter.HomePath (one MINDWALK_HOME contract)
  • Cache.Store errors now name the session key, and its temp-file cleanup is errcheck-clean

Sealed-subprocess stance is unchanged: no tools, no user config, verdicts still derived mechanically from findings.

Why

Users who only have Crush installed currently cannot evaluate sessions at all; the judge is otherwise CLI-agnostic.

Series

Part 7/9; stacks on #30 (and #25–#29 before it). "Files changed" shows the cumulative stack — this PR's own delta is just the seventh commit, feat(judge): support crush as a judge CLI.

Validation

  • go test -count=1 ./... green at the tip
  • new test: crushModel reads the verbose log line and stays empty without one

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant