Fix the agent, command, and hook defects the skill-focused review missed

Context

The story behind this plan — what prompted the work and why it matters now.

A review on 2026-07-16 covered all 59 skills across the 13 marketplace plugins and produced three plans, but examined none of the 18 commands, 22 agents, or 6 hooks. This plan covers that gap. The findings are ranked below by blast radius, and the first is a live defect, not a style issue.

The agents key is tools:, not allowed-tools:. All seven plan-agent/agents/plan-reviewer-*.md files declare allowed-tools: Read, Glob, Grep, Bash, which is the skills key. On an agent it is not recognised, so the declaration is silently ignored and the agent inherits everything. The contrast is visible inside this very marketplace: product-plans/agents/product-reviewer-*.md declare tools: Read, Glob, Grep, Bash(git *) and are correctly scoped. Confirmed empirically from a live session's agent registry, where plan-agent:plan-reviewer-architecture lists as "(Tools: All tools)" while product-plans:product-reviewer-pm lists as "(Tools: Read, Glob, Grep, Bash(git *))". Both clusters intended the same restriction; only one got it. The consequence is seven agents whose whole job is to read a plan and report findings holding Write, Edit, and Bash against the repo.

A trap worth naming explicitly: the maintainer's memory records that allowed-tools is valid in skill files and that an IDE linter flagging it is wrong. That note is correct — for skills. It does not transfer to agents, and it is the single most likely reason this finding gets waved off. The key is valid in one component type and inert in the other.

Two team-defaults agents are misconfigured. css-generator.md declares MultiEdit, a tool that no longer exists in Claude Code, and has no model:. code-comments.md declares no tools: at all, so a JSDoc writer inherits Bash, WebFetch, and Agent; its description is a capability blurb with no trigger phrase, so it will not reliably activate; and its name: ts-commenter does not match its filename.

Four plan-interview commands have drifted from their own skills. deep-grill.md, plan-status.md, and documenting-plans.md do not invoke their same-named skills — they restate the whole workflow, and the copies are already 20, 153, and 383 lines apart from the skill they duplicate. The behaviours differ concretely: deep-grill's command resolves the plan from $ARGUMENTS, its skill from "a path in the user's message". plan-interview.md (397 lines) re-implements the router that also lives in its 594-line skill. Two behaviours, one name, no way for a user to know which they invoked. The repo already ships the correct shape: plan-to-html.md is frontmatter plus a single Skill(...) call.

Four hooks fire on every file edit. All four plan-agent/hooks.json PostToolUse entries match Write|Edit|MultiEdit, so every edit in every session spawns four processes to discover the file is not a plan. hooks/build-index.sh:36 compounds it: find_templates_dir() walks all of ~/.claude/plugins and the project root on each plan write, unbounded. Separately validate-plan-filename.py matches only Write|Edit, so MultiEdit bypasses the filename gate entirely — an inconsistency in the opposite direction.

Out of scope, deferred to Next Steps: consolidating plan-agent's 8 reviewer agents down to ~5 and delegating the ux/accessibility lenses to product-plans' equivalents. That removes agents users may reference and is a MAJOR-bump decision, consistent with the cleanup-first sequencing already agreed.