fix(graders): guard judge against unverifiable SDK-shape claims#126
fix(graders): guard judge against unverifiable SDK-shape claims#126frederikprijck wants to merge 1 commit into
Conversation
The LLM judge was producing false negatives by asserting, from stale
model memory, that valid code was broken — e.g. claiming
'from fastapi_plugin import Auth0FastAPI' would raise ImportError (it is
the correct public import for auth0-fastapi-api, verified against 1.0.0b7),
and calling useAuth0/getSession/{ accessToken } 'invented' when all three
are real in @auth0/auth0-nuxt 1.1.0. These sank otherwise-correct runs.
Instruct the judge to grade only what the code shows and not to fail code
over import-path, module-name, or method-signature claims it cannot verify
from the code in front of it, since a package's public import surface often
differs from its internal module name.
default.md feeds prompts.generated.ts (gitignored, rebuilt on build).
|
Warning Review limit reached
Next review available in: 59 minutes Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (1)
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
1 participant
What
Hardens the shared LLM-judge system prompt (
packages/evals-core/src/graders/prompts/default.md) so the judge stops failing correct code over SDK-shape claims it cannot verify from the code in front of it.Why
The judge produced false negatives by asserting, from stale/incorrect model memory, that valid code was broken. Two confirmed cases from the latest matrix:
from fastapi_plugin import Auth0FastAPI"would cause an ImportError." Verified againstauth0-fastapi-api==1.0.0b7:fastapi_pluginis the correct public import;auth0_fastapi_api(what the judge said was right) does not exist. The agent's own runtimeimport mainsucceeded in the trace.useAuth0,getSession, and{ accessToken }destructuring "invented." Verified against@auth0/auth0-nuxt@1.1.0: all three are real (useAuth0(event): Auth0Client,getSession(): Promise<SessionData | undefined>,getAccessToken(): Promise<TokenSet>whereTokenSet.accessToken: string).These verdicts sank otherwise-correct runs (fastapi 84.2 was the only sub-85 run in the matrix, driven partly by these two judge failures).
Change
Adds guidance to
default.mdinstructing the judge to grade only what the code shows and not to assert an import path is wrong, that a package exposes a symbol under a different module name, or that a signature is invalid unless the code itself contradicts the question — because a package's public import surface often differs from its internal module name.default.mdis the source forprompts.generated.ts(gitignored, regenerated on build viascripts/generate-prompts.mjs).Verification
npm run build✓ (5/5 packages)npm test✓ (all tasks; 484 tests in evals-core)npm run lint✓ /npm run format✓ (unchanged)Notes
Companion to
fix/quickstart-env-prompt-wording(#125), which addresses the separate hardcoded-secret failures. Independent — either can merge first.