Repository navigation
test(build): add agent guidance and historical eval candidates - #2887
Merged
Merged
Conversation
mekilis
force-pushed
the
smart/agent-guidance-historical-evals
branch
from
October 21, 2026 09:47
a481139 to
d9462f7
Compare
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Adds public contributor guidance and a historical coding-agent evaluation harness. The catalog contains 60 task candidates, each pinned to a pre-fix revision with acceptance scenarios, permitted implementation paths, and regression tests from the reference fix.
The harness prepares isolated worktrees and grades implementation patches independently of candidate-authored tests. Backend CI validates task definitions and runs harness unit tests without invoking models. Usage, disposable-service prerequisites, privacy rules, and open-book benchmark limitations are documented in
evals/README.md.All 60 tasks are explicitly uncalibrated: prompt review and meaningful baseline-failure/reference-pass evidence are required before reporting agent scores.
Validation: all 60 task definitions checked; six harness unit tests passed; catalog regeneration reproduced the same contents;
actionlintand whitespace checks passed. No model runs or historical task calibration were executed.