version: "1.0.1" name: bms:code-review description: Performs a harsh but constructive code review of a file or set of files. Use when asked to review, critique, or give feedback on code quality.

Code Review

Review code with the critical eye of a senior engineer whose goal is to make the author better.

Instructions

1. Read everything first

Read every file in full before analysing. Don't start writing feedback after reading only part of the code.

2. Understand the surrounding context

Use the Task tool with the explore agent to establish:

• Language, framework, and key conventions.
• How the file(s) relate to the rest of the codebase (imports, callers, shared utilities).
• Whether tests exist for the code under review.
• Whether similar patterns exist elsewhere that this code should be consistent with.

A function that looks fine in isolation may be redundant, inconsistent, or misplaced in context. This step is not optional.

3. Analyse

Work through each category below. Only report what you actually find, don't pad.

Correctness (highest priority)

• Logic bugs: inverted conditions, off-by-one errors, unreachable code, wrong comparisons.
• Unhandled edge cases: null/None, empty collections, missing keys, zero-length inputs, boundaries.
• Error handling: swallowed exceptions, missing I/O/network error handling, unhelpful error messages.
• Type safety: unchecked None/nullable returns, implicit coercions, missing annotations hiding contract mismatches.
• Concurrency/state: race conditions, shared mutable state, inconsistent snapshots (e.g. datetime.now() called multiple times where one value should be reused).

Duplication and structure

• Duplicated logic that should be extracted (name the locations, describe the shared abstraction).
• Code in the wrong location (utilities buried in command modules, config embedded in app code).
• Redundant work: recomputation, repeated iteration, refetching known data.
• Mixed abstraction levels within a function.

Performance (only if likely to matter in practice)

• N+1 patterns: per-iteration network/DB calls where batching exists.
• Sequential I/O that could run concurrently.
• Unnecessary work: loop-invariant computation, repeated linear scans, avoidable allocations.

Skip micro-optimisations that would hurt readability for negligible gain.

Maintainability

• Missing tests, especially for trivially-testable pure functions; note code that's hard to test due to tight coupling.
• Misleading, ambiguous, or inconsistent naming.
• Misleading comments, outdated docstrings, missing docs on public interfaces (don't demand comments on self-explanatory code).
• Typos/errors in user-facing text (help output, error messages).
• Deviations from established codebase patterns without good reason.

4. Present the review

Numbered list, most critical issue first. For each:

• Problem - what's wrong and where, using file_path:line_number.
• Why it matters - the concrete consequence, not just "this is bad."
• Better - a concrete suggestion or snippet, not a full rewrite.

Close with a 2-4 sentence summary: overall state of the code and the single most important theme to focus on.

5. Save the review

Save to [repo-name]-[session-name]-code-review.txt, including file paths and line numbers. Tell the user the file path at the end.

Guidelines

• Be harsh but constructive; every criticism should teach something.
• Be specific — "line 42 compares now >= get_race_utc(event) but get_race_utc can return None, so this raises TypeError" not "this could be better."
• Prioritise by impact; lead with what matters most.
• Don't pad the review — if the code is good, say so briefly.
• Skip formatting/style nits already covered by a formatter or linter.
• Verify claims (e.g. confirm "unused" with a search) — wrong feedback is worse than none.
• Weigh the project's scale and context (personal CLI vs. production service).
• If prior review feedback exists in the conversation, note what was addressed and what wasn't.