feat(framework): score committed solutions without an agent run - #206
Conversation
`pnpm eval:solution -- --eval <id>` boots the eval's environment, copies a directory from the eval's `solutions/` over the workspace, and runs the scorer. Deterministic, so a verdict that moves is the scorer changing rather than the agent. The overlay copy lands between `localDir` and `supabase start`, so a solution's migrations apply on boot with no per-eval apply step. Eval discovery and run setup move to `eval-discovery.ts` so the agent and solution entry points can't drift on what an eval is. Local-stack only. A tools eval's starting state is a seeded hosted project, not a workspace to copy files into. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes AI-1080
Stacked on #205. Review that one first, or read this diff against its branch.
Problem
Checking whether a scorer's checks behave takes an agent run. That run is nondeterministic and costs a model call, so you can't separate a scorer defect from agent variance, and a check that passes broken work is hard to catch before the eval ships. #205 tells contributors to score example solutions. Nothing scored them.
Solution
pnpm eval:solution -- --eval <id>, which boots the eval's environment, copies a directory from itssolutions/over the workspace, and runs the scorer with no agent.--solution <name>scores one.localDirandsupabase start, so a solution's migrations apply on boot with no per-eval apply step. That is the whole mechanism.eval-discovery.ts, so the agent and solution entry points can't drift on what an eval is.results/. Exported workspaces land in.solution-runs/, gitignored.build-cli-002-declarative-schemaand the expected-failure list that goes with them.README.mdand in the guidance from docs: cover example solutions and check design for scorers #205.Local-stack evals only. A tools eval's starting state is a seeded hosted project, not a workspace to copy files into, so it needs its own apply path.
Manual testing
Stop any local Supabase stack, since the run needs ports 54321-54329.
Score both solutions:
Output matches
solutions/README.md.greenfails only the tool-call check,migration-onlyalso fails the schema-file check:Run
git status results/. It is clean, and.solution-runs/build-cli-002-declarative-schema/holds both exported workspaces.