requesting-code-reviewlisted
Install: claude install-skill felipemelendez/llm-orchestrator
# Requesting code review
Two required reviews plus one conditional, in order. Each returns an `Issues:` block.
## Two paths
This review can run two ways. The markdown stages below are the **canonical** path. When the
Workflow tool is available, prefer the accelerated path; routing — and what each path does that
the other does not — follows `using-workflows`.
One difference to know when comparing results: the markdown path stops after Stage 1 only on a
**Critical** finding, while the workflow early-exits on **critical or important** — the workflow
gates more strictly.
**This flow cannot delegate to the native review skills.** As of Claude Code v2.1.215,
*"Claude no longer runs the `/verify` and `/code-review` skills on its own"* — both carry
`disable-model-invocation: true` and cannot be preloaded into a subagent. They are excellent
**manual** passes (`/code-review xhigh` on a real diff is worth running by hand), but a skill
cannot invoke them, so the plugin's own reviewer agents are the only automatable substrate.
What stays this skill's contract either way: Stage 1 spec compliance (native review checks
correctness, never whether the diff implements the approved spec), the spec-gates-quality
order, and the evidence rules below.
### Preferred path (Workflow tool present)
1. Compute the security boolean from the single source of truth — source
`scripts/lib/orch-signals.sh` and test the diff against `$ORCH_SIG_SECURITY_DIFF` (the same
grep shown in Stage 3 below). D