# Cy Review Round > Performs a comprehensive code review of a spec implementation and generates a review round directory with issue files compatible with cy-fix-reviews. Use when reviewing implemented spec tasks, creating a manual review round without an external provider, or performing a quality audit of code changes. Do not use for fetching reviews from external providers, fixing existing review issues, executing spec tasks, or editing source code. - Skill: `compozy/cy-review-round` (Agent Skill, multi-file: 3 files) - Install (CLI): `npx skillmds@latest add compozy/cy-review-round` - Raw SKILL.md: https://api.skillmd.com/api/skills/compozy/cy-review-round/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Security, Product & Planning - Author: compozy (https://skillmd.com/u/compozy) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/compozy/cy-review-round --- # Review Round Perform a structured code review of a spec implementation and produce a review round directory that the `cy-fix-reviews` workflow can process. ## Required Inputs - Feature name identifying the `.compozy/tasks//` directory. - Optional: specific files or directories to scope the review. ## Workflow 1. Determine the review round directory. - Derive the spec directory from the feature name: `.compozy/tasks//`. - Verify the spec directory exists. If it does not, stop and report the missing directory. - List existing `reviews-NNN/` subdirectories to determine the next round number. If none exist, use round 1. - If prior review rounds exist, read their issue files to build a list of already-known issues. The current round must only contain NEW issues not already tracked in prior rounds. Do not re-flag issues that are pending, valid, or resolved in earlier rounds. - Determine the review round directory path: `.compozy/tasks//reviews-NNN/` with the round number zero-padded to 3 digits. Do NOT create it yet — wait until step 4 confirms there are issues to write. This avoids leaving empty directories when the review finds no issues. 2. Identify the review scope. - Read `_spec.md` and `_tasks.md` from the spec directory to understand what was implemented and why, plus the contract catalogs `_user_stories.md` and `_tests.md` when present. - Read ADRs from `.compozy/tasks//adrs/` for architectural decision context. - If `_spec.md` is missing, warn that the review will lack requirements context but proceed with a code-quality-only review. - If the user provided specific files or directories, scope the review to those paths. - If no explicit scope was provided, resolve the actual PR or repository base and inspect its merge-base-to-head diff, including owned working-tree changes when requested. Reuse an established scope; ask only if the intended range remains ambiguous. - Map the relevant changed paths and consumers locally; use bounded read-only agents only for substantial independent review slices. Reuse current implementation research. 3. Perform the code review. - Read `references/review-criteria.md` for severity definitions and evaluation areas. - **Prioritize the review scope.** For a large scope, review slice by slice instead of silently sampling: partition the diff by the task boundaries that produced it (task checkpoint commits, `_tasks.md` slices) and review each partition against its task's contract, core implementation files first. When the scope cannot be partitioned and still exceeds what a complete read can honestly cover, say so in the summary and recommend splitting the delivery — a review that silently degrades to sampling at scale certifies nothing. - Inspect every changed behavior in the prioritized scope, including enclosing contracts and affected consumers. Read full files when needed to understand their invariants; unchanged boilerplate does not require repeated reading. - **Requirements validation**: If `_spec.md` was available in step 2, cross-check the implementation against every stated requirement, acceptance criterion, and architectural decision — including every acceptance criterion and edge case in `_user_stories.md` when it exists. Flag any requirement that is missing, partially implemented, or implemented differently than specified. These are correctness issues — assign severity based on the gap's impact (critical if a core feature is missing, high if behavior deviates from spec, medium if an edge case from the spec is unhandled). - **Mission fit**: when `_spec.md` states a Motivating Problem, verify the delivered work solves it and name the slice that does — checking the implementation against the spec alone cannot catch a spec that drifted from its own mission. An ADR that narrowed or deferred the Motivating Problem without the user's recorded sign-off is itself a finding, at the severity of the gap it created. - **Test-contract parity**: If `_tests.md` exists, verify that every test ID assigned in completed tasks' `## Tests` sections is implemented in the suite and asserts the behavior the contract specifies. A missing case, or a hollow one that exists without asserting the contracted behavior, is an issue — assign severity based on the impact of the behavior left unverified. - Use the relevant evaluation areas for the changed behavior: Security, Correctness, Concurrency, Performance and Scalability, Error Handling, Code Quality and Maintainability, Testing, Architecture, and Operations. - Identify issues in severity order: critical first, then high, medium, and low. - For each issue record: the file path relative to the repository root, the approximate line number, the severity level, a concise title (max 72 characters), and a detailed review comment describing the problem and a suggested fix. - **Deduplicate before writing.** If the same pattern (e.g., missing nil check, missing error wrap) appears in multiple files, create one issue for the most representative instance and list the other affected files in its Review Comment. Do not create N identical issues for N files exhibiting the same root cause. One issue per distinct problem, not per occurrence. - **Verify before flagging.** Before creating an issue, check whether the pattern is intentional: look for adjacent comments explaining the choice, ADR references, or test coverage that validates the behavior. If code looks suspicious but has a clear justification (e.g., `// nolint: intentionally ignoring close error on read-only file`), do not create an issue. Only flag patterns that are genuinely problematic, not merely unconventional. - Skip issues already owned by linters or formatters. Reuse current scoped gate evidence; run the owning lane only when its result is needed, not an unconditional full lint pass. - **Focus on signal, not volume.** Aim for fewer, higher-quality issues rather than an exhaustive list. Deduplicate by root cause and retain actionable findings supported by evidence. Issue counts are not quality targets or caps. - Also note well-implemented aspects of the code. These observations inform the summary but do not produce issue files. - If no issues are found after a thorough review, report that the implementation looks clean and skip steps 4 through 6. Do not create the review round directory. 4. Generate issue files. - Create the review round directory determined in step 1. - Read `references/issue-template.md` for the canonical format. - For each issue identified in step 3, create an `issue_NNN.md` file in the review round directory. - Issue numbering starts at `001` and increments sequentially. - Each file must use this exact structure: ``` --- provider: manual pr: round: round_created_at: status: pending file: path/to/file.go line: 42 severity: high author: claude-code provider_ref: --- # Issue NNN: ## Review Comment <detailed review body> ## Triage - Decision: `UNREVIEWED` - Notes: ``` - The `<author>` field must be `claude-code`. - The `provider_ref` field must be empty. - The `provider` field must be `manual`. - The `pr` field is empty for manual reviews. If the user provides a PR number, include it. - The `round` field must match the directory number as an integer (not zero-padded). - The `round_created_at` field must use the same current UTC RFC3339 timestamp in every issue in this round. - The `severity` field must be exactly one of: `critical`, `high`, `medium`, `low`. 5. Summarize and present the review. - Print a summary listing: - **Merge recommendation**: If any critical or high issues exist, state "Needs fixes before merge" with the blocking issues. If only medium/low issues exist, state "Safe to merge with follow-ups." If no issues, state "Clean — ready to merge." - Total issues found, broken down by severity (critical, high, medium, low). - The review round directory path. - The full list of generated issue file names. - Well-implemented aspects observed during the review. - Suggest running `compozy loop run --workspace <ref> --name review-and-fix --input task_name=<name>` to process the review round. 6. Verify before completion. - Use installed `cy-final-verify` before claiming the review round is complete. - Read back each generated issue file and verify the frontmatter parses correctly. - Verify every issue file in the round has matching `provider`, `pr`, `round`, and `round_created_at` values. - Confirm the review round directory follows the `reviews-NNN` naming convention. ## Critical Rules - Do not fix the issues found. This skill only identifies and documents issues. The `cy-fix-reviews` workflow handles remediation. - Do not create issue files for problems that linters or formatters already catch. - Every issue file must have valid YAML frontmatter parseable by `prompt.ParseReviewContext()`. - Do not create or maintain review `_meta.md`; round metadata lives in each issue file frontmatter. - Do not create empty review rounds. If no issues are found, report a clean review and do not create the round directory. - Do not modify any source code files. This is a review-only skill. - Do not call provider-specific scripts or `gh` mutations. ## Error Handling - If the spec directory does not exist, stop and report the missing directory. - If no files can be identified for review and the user did not provide explicit paths, ask the user to specify files. - If `_spec.md` is missing, warn about the lack of requirements context but proceed with code-quality-only review. - If the review round directory cannot be created, stop and report the filesystem error. - If writing an issue file fails, stop and report which file could not be written. - If the relevant lint gate cannot run (build errors, missing tools), note the failure in the summary and proceed with the review. Do not skip the review because linting failed — just acknowledge that linter-overlap filtering could not be applied.