Blueprint: code-review ecology

Objective

Review code changes through multiple independent lenses—correctness, security, performance, maintainability, and contract compatibility—while preserving structured findings and preventing reviewers from modifying acceptance policy.

Components

• deterministic linters and static analyzers;
• language-specific review specialists;
• security specialist;
• performance and concurrency specialist;
• test-generation model in a sandbox;
• finding deduplicator and severity calibrator;
• acceptance gate tied to repository policy;
• persistent code beads for findings and remediation.

Review flow

FUNCTION review_change(change, repository_policy)
    deterministic <- RUN_STATIC_AND_TEST_SUITES(change)
    model_findings <- RUN_INDEPENDENTLY(
        specialists = SELECT_REVIEW_SPECIALISTS(change),
        input = SANITIZED_CHANGE_CONTEXT(change)
    )

    normalized <- NORMALIZE_AND_DEDUPLICATE(deterministic, model_findings)
    verified <- VERIFY_FINDINGS_WITH_TESTS_OR_EVIDENCE(normalized)
    beads <- CREATE_ACTIONABLE_BEADS(verified)

    verdict <- APPLY_REPOSITORY_ACCEPTANCE_POLICY(
        deterministic,
        verified,
        repository_policy
    )

    RETURN {verdict, beads, trace}
END FUNCTION

Breeding loop

Use confirmed missed defects, false positives, review latency, and language coverage as evidence. Train or distill narrow specialists offline. Code-breeding proposals that alter analyzers or policy follow a separate software pipeline.

Safety boundaries

Review models have read-only repository snapshots, no production credentials, no ability to merge code, and no authority to mark their own findings resolved. Generated tests run in disposable environments with network denied.

Metrics

Confirmed-defect recall, false-positive rate, severity calibration, time to verified finding, duplicated findings, test-generation validity, review cost, and developer override reasons.

Code beads

Each accepted finding becomes a structured bead with file, line, severity, evidence, dependency, acceptance criteria, and discovered-from links. This supports handoff across short-lived agent sessions without relying on conversation memory.