gate
Gate — harness-portable quality gate
!IMPORTANT: Follow this process exactly. Do not skip steps.
This skill is a gate, not a fixer. It returns a verdict; it does not modify code.
It orchestrates with a single primitive every agent harness has: spawn a subagent, read its result. No Workflow tool, no agent teams — so it runs unchanged on Claude Code and Codex. The companion gate-wf skill runs the same review logic on the Claude Code Workflow engine (deterministic static script, --resume caching); prefer it when Workflows are enabled.
Skill version: 8. Cache entries are keyed on this — bumping invalidates all caches at once. v8: CLAUDE.local.md at the repo root joins the context bundle alongside CLAUDE.md, so a personal, git-ignored rule file reaches context-checker and its MUST/SHOULD clauses synthesize findings. The CLAUDE.md freshness fold now hashes file contents instead of git log — a git-ignored rule file has no commit, so the old fold silently no-opped and a rule edit served a stale cached verdict. Also: adr/ joins ADR_ROOT_CANDIDATES and every ADR root is now walked recursively — a repo keeping its ADRs at adr/, or its rules in .claude/rules/<domain>/, had them read by nothing. Companions that @-reference another candidate are deduped. v7: ponytail-reviewer greps the repo for an existing equivalent of every export the diff adds (ponytail-exists) — duplication of code the repo already has was in no reviewer's scope. v6: simplify-reviewer and slop are un-merged into two reviewers (one rule set each), plus a new ponytail-reviewer on the over-engineering axis (/ponytail-review) — 6 base reviewers instead of 4. Rule ids are unchanged, so existing dismissals survive.
Prerequisites
- A harness with a subagent primitive (Claude Code
Agent/Tasktool, Codex subagents). Reviewer subagent types (bug-reviewer,solid-reviewer, …skeptic,context-checker) ship with this plugin at the plugin rootagents/*.md— they are the single source of the review logic, shared withgate-wf. - The reviewer skill dependencies of Step 0.