Skill v1.0.0
currentLLM-judged scan90/100name: vool-code-review description: "Review a diff before it lands: correctness at the owning seams, test coverage of the new behaviour, and honesty of the claims. Use when the user asks to review a diff, a patch, or a change before push." version: 1.0.0 license: MIT author: sls_0x status: p1 type: capability_skill triggers: [review, diff, patch, before push, look over my change] allowed-tools: [workspace.git_diff, workspace.read_file, workspace.search_text, workspace.run_lint] capabilities: [workspace.read, workspace.git, workspace.validate] permissions: [read_files] effects: [read_only, validation_command] inputs: [workspace_root, diff_ref optional] outputs: [review_findings, verdict] stop-conditions: [review delivered with findings and a verdict, no diff exists to review - say so, findings list exhausted - do not pad with style nits] recovery: [if the diff is empty, review the named files directly and label it a file review not a diff review, if lint cannot run, review without it and mark that limit in the verdict] cumulative-test-law: [tests/native_skills must pass before this package changes, selection phrase "review my diff before I push" must keep selecting this package, review findings must cite the diff hunks they came from] id: code-review risk-class: read_only task-families: [workspace_audit] capability-families: [workspace] tool-intents: [workspace.git_diff, workspace.read_file, workspace.search_text, workspace.run_lint] permitted-tools: [workspace.git_diff, workspace.read_file, workspace.search_text, workspace.run_lint] prerequisites: [] expected-outputs: [review_findings, verdict] verification: [evidence_cited] stopping-conditions: ["stop when every finding carries its file:line and the verdict names the decision"] incompatible-with: [] priority: 30
Code Review
You review the CHANGE, not the world. Findings cite hunks; the verdict states land / fix-first / reject and why.
Scope
IN: the diff — correctness at its seams, coverage of new behaviour, honesty of claims (tests that pin the change, no silent widening), lint signal. OUT: rewrites of untouched code, style bikeshedding beyond the linter's own rules, approving by vibes.
Required capabilities (declaring is not granting)
workspace.read, workspace.git, workspace.validate. Declaring is not permission; reads stay read-class and gated.
Procedure
workspace.git_diffthe change; if empty, ask which files or fall back to
the named files, labelled as a file review.
workspace.read_fileeach hunk's owning file with context.workspace.search_textfor the callers the change can break.workspace.run_lintfor mechanical signal.- Deliver findings (each citing its hunk) + verdict: land / fix-first /
reject.
Stop conditions
Stop when findings are delivered. No padding: a missing finding is honest, a manufactured one is noise.
Recovery
Lint unavailable: say the verdict carries no lint signal. Diff unreadable: report the exact failure, do not review from imagination.
Cumulative test law
tests/native_skills is the pack; review-behaviour edits re-run it.
No direct execution bypass
Diff and lint go through the runtime tool door (execute_runtime_tool). No direct git or linter invocation around it.