minhtran3124/agent-harness/correctness-review

Adversarial Correctness Review

Review runtime behavior, not style or plan compliance. Determine the intended BASE..HEAD range (merge-base for a branch, working-tree diff for uncommitted work, explicit range when supplied). In SDD, use the commit before task 1 through current HEAD.

When SDD supplies REVIEW_PACKAGE_PATH, pass it into the FIND prompts, which own how to read it. It is runtime evidence, not an oracle: retain this skill's plan-blindness and do a named focused read outside it only for a concrete runtime risk. Standalone calls without a package keep the existing range-construction behavior.

Pipeline

1. Read review-config.json, then render six independent FIND prompts from correctness-reviewer-prompt.md: enclosing-function,

removed-behavior, call-site-impact, stack-defects, guard-completeness, and prior-art. Each candidate names a concrete trigger and wrong result.

1. Deduplicate by (file, line). Angles are provenance only—never evidence and never scorer input.
2. Dispatch one independent SCORE prompt per remaining location using

correctness-scorer-prompt.md. The scorer reads the code/diff, not finder reasoning.

1. Route scores at the configured threshold to classification; record lower scores as advisory in

SUMMARY.md (or inline standalone). Never lower the threshold below its configured floor.

1. Before classifying, read `rules/auto-correct-scope.md`. Rule 1–3 findings may be

fixed; Rule 4 is STOP and goes to ESCALATIONS.md (or directly to the user standalone).

1. Re-review each Rule 1–3 fix. Cap each finding at three rounds; if open blocking findings do not

decrease and the reviewed diff hash is unchanged, escalate immediately.

The scorer owns anchor scores, unreadable-code caps, JSON schema, and threshold rationale. The finder owns the six-angle methods and output shape. Compose isolated child prompts with scripts/render_skill_prompt.py; never assume they inherited controller context.

Completion gate

Before reporting success or handing to finishing-a-development-branch, every finding is either fixed with a commit SHA or durably recorded: advisory/carry-over in SUMMARY.md, or a STOP/capped blocker in ESCALATIONS.md. Anything else blocks completion.

/code-review is a sibling cleanup review, not a replacement for this runtime-bug oracle.