feat: add Cursor CLI agent harness and experiments

[vivek100]

Summary

• Add a Cursor CLI harness (packages/core/src/agents/cursor/) with pinned install, headless stream-json runner, transcript parser, and cursorAgent() factory wired into the registry.
• Add four experiments mirroring Claude Code: cursor-composer-2.5, cursor-composer-2.5-no-skills, cursor-grok-4.5-high, cursor-grok-4.5-high-no-skills (Supabase MCP on; skills vs none).
• Document CURSOR_API_KEY in .env.example / README and label Cursor in the results UI.

Fixes #167

How to run

# .env needs CURSOR_API_KEY; OPENAI_API_KEY for rubric/judge checks
pnpm eval -- --experiment cursor-composer-2.5 --eval resolve-dataapi-001-empty-results

Model ids (from live cursor-agent --list-models): composer-2.5, cursor-grok-4.5-high.

Verification

• Unit: pnpm exec vitest run src/agents/cursor/ (core package)
• Typecheck / format: pnpm typecheck, pnpm format:check
• Discovery: pnpm eval:dry for all four Cursor experiments
• Tools-mode smoke: cursor-composer-2.5 × resolve-dataapi-001-empty-results — agent completed with Supabase MCP (list_tables, execute_sql, apply_migration, …) and skills loaded. Scorer judge step needs OPENAI_API_KEY (eval uses judge() in addition to deterministic checks).

Test plan

• pnpm --filter @supabase-evals/core test (includes Cursor runner/parser tests)
• pnpm typecheck
• Dry-run one Cursor experiment
• Optional: add run-evals label if maintainers want full benchmark CI
• Optional: targeted eval-refresh for the new Cursor experiments

[vercel]

@vivek100 is attempting to deploy a commit to the Supabase Team on Vercel.

A member of the Team first needs to authorize it.

[mattrossman]

Hey @vivek100 , thanks for looking into this.

We're definitely interested in adding support for more harnesses including Cursor (it looks like this work overlaps with existing PR #76).

Disclaimer, the blocker here is our eval runner infra which current lives on GitHub Actions with a 256 job limit. We're working to move to a more scalable solution that will give us the headroom to support more agent variations in our benchmark matrix.

[vivek100]

Hi @mattrossman thanks for the response.

I should have looked at the open PRs first :P , PR #76 does implement the same thing.

I was working on similar evals using harbor framework and was trying to see if there are any upsides or trying the supabase eval setup.

Will mark this PR close as its not needed, 76 can be used when the infra setup is updated.