fix(code-review): require explicit user invocation

[ZaunEkko]

Summary

• mark /code-review as manual-only with disable-model-invocation: true
• preserve explicit user invocation of /code-review and /code-review --comment
• prevent models and subagents from programmatically re-entering the full multi-agent review workflow

Motivation

The code-review command launches an initial gate agent, a guidance-discovery agent, a change-summary agent, four parallel reviewers, and additional validators for reported findings. Programmatic invocation from one of those descendants can recursively re-enter the complete workflow and multiply the agent tree.

disable-model-invocation is the documented frontmatter control for commands that require explicit user judgment: it keeps the slash command available to users while removing it from the model's programmatic command invocation surface.

This is a plugin-level mitigation related to #77414 and #68110. It cuts off recursive model-initiated re-entry into this specific multi-agent workflow, but it does not replace host-level descendant budgets, breadth limits, resource backpressure, or tree-wide cancellation.

Behavior change

Invocation path	Before	After
User types /code-review	Allowed	Allowed
User types /code-review --comment	Allowed	Allowed
Model or subagent invokes code-review programmatically	Allowed	Blocked
Subagent directly invokes Agent	Allowed	Unchanged

Test plan

• Validate plugins/code-review with claude plugin validate
• Validate plugins/code-review with claude plugin validate --strict
• Parse the command frontmatter as YAML and confirm disable-model-invocation is boolean true
• Run git diff --check
• Confirm only plugins/code-review/commands/code-review.md changed
• Confirm the branch is based on the latest upstream/main

No Agent integration test was run because /code-review intentionally launches multiple agents; this PR is limited to static frontmatter and plugin validation on an unrestricted workstation.

Generated with Claude Code