Code and project review

Most review attention goes to new diffs; most risk lives in the code nobody has looked at for a year. This skill covers the retrospective forms: reviewing the EXISTING codebase with findings ranked and then implemented, and reviewing the PROJECT on a cadence - after milestones, before releases and submissions. Division of labor: rseng-pair-programming owns diff-time pre-review of new work; rseng-version-control-review owns the review process and rules; this skill owns review as an undertaking on what is already there. Review-then-implement is a two-phase contract: findings first, agreement second, changes third - never silent rewriting under the review's flag.

The codebase review

Scope before depth - review the code that matters most:

1. Map risk: recently churned complex modules (rseng-software-metrics' hotspot map), result-bearing code paths (what feeds published numbers), security surfaces (rseng-security), and anything the tests do not reach (rseng-testing's coverage). Review THOSE deeply; skim the rest.
2. Sweep with the pack's lenses, each producing findings with file:line evidence: correctness and silent-failure modes (rseng-defensive-coding), numerical practice (rseng-numerical-accuracy), design and coupling (rseng-software-design), performance red flags on hot paths (rseng-performance-profiling), hygiene and dead code (rseng-code-quality), reproducibility hazards - unseeded randomness, unpinned environments, hidden state (rseng-reproducibility, rseng-notebooks).
3. Rank honestly: must-fix (wrong results possible, data loss, security), should-fix (maintenance debt with a price), and nice (style beyond the linter's remit). Tier calibrates the bar (rseng-quality-framework): analysis code is not reviewed to infrastructure standards.
4. Deliver findings BEFORE changing anything: the ranked report with evidence is the review; the user picks what gets implemented (defaults offered: all must-fix, should-fix by effort).

Implementing the agreed improvements

• One finding, one commit: small reviewable changes referencing the finding (rseng-version-control-review), tests first or alongside for behavior-touching fixes - and characterization tests BEFORE touching anything load-bearing and untested (rseng-legacy-code's safety net; review of legacy code IS the legacy workflow).
• Behavior-preserving by default: refactors keep the tests green; anything that changes results is flagged loudly and decided by the human (the science is theirs - rseng-pair-programming's boundary).
• Re-run the full verification after the batch: tests, linters, metrics delta (rseng-software-metrics shows the improvement measurably - report before/after numbers).
• Route the residue: declined findings recorded with reasons in the decision log (rseng-project-tracking), recurring patterns into lint rules or checklists, and the transferable insights into LESSONS.md (rseng-lessons-learned).

The recurring project review

After a major task, milestone or before a release/submission, review the PROJECT, not just the code - a lightweight standing agenda:

• Code: the codebase review above, scoped to what changed since last time plus one rotating deep-dive area.
• Quality posture: indicator walk (rseng-quality-framework), FAIR/security assessments re-run and deltas reported (rseng-fairguard, rseng-security).
• Records current: README and docs truthful (rseng-documentation), citation and contributor records match reality (rseng-citation-metadata), aidecl.yaml up to date (rseng-ai-declaration), tracker and milestones honest (rseng-project-tracking).
• Reproducibility spot-check: the one-command path still works from clean (rseng-reproducibility).
• Close the loop: findings become tracked tasks with owners; the review note (date, scope, findings, actions) goes into the project record - and the next review starts by checking the last one's actions landed.

Offer this review PROACTIVELY at natural moments - a milestone just closed, a release is near - as a short menu ("want the post-milestone review? code, records, repro - about 20 minutes"), respecting a decline (rseng-trainer's consensual-offer rule).

Review craft

The comment discipline from rseng-pair-programming applies unchanged: evidence with every claim, reasons with every request, blocking separated from preference, praise where it is earned - and for research code, "what does this number correspond to physically" remains a first-class review question (rseng-research-integrity). Reviews are of code, never of authors; findings in inherited code especially are archaeology, not blame (rseng-lessons-learned's blame-free wording).

Working with this skill

This skill is source-independent: it encodes structured code-review practice applied retrospectively to research software. It composes the pack's lenses into one undertaking and feeds rseng-project-tracking and rseng-lessons-learned.

Learn more (verified):

• https://google.github.io/eng-practices/review/ - Google engineering review practices
• https://bssw.io/items?topic=peer-code-review - BSSW peer code review resources
• https://conventionalcomments.org - conventional comments

<!-- related-skills:begin -->

Related skills

Check whether any of these applies before moving on:

• rseng-fairguard - assessment reruns during milestone reviews
• rseng-legacy-code - characterization tests before implementing fixes
• rseng-lessons-learned - findings become recorded lessons
• rseng-project-tracking - findings become tracked tasks
• rseng-quality-framework - tier calibrates the review bar
• rseng-software-metrics - hotspot map scopes the review

<!-- related-skills:end -->