메뉴
Deep code review of a diff. Launches the code-review workflow — finder agents scan each review angle, an independent verifier judges every candidate, then a ranked, capped findings report. Read-only; it reviews, it does not edit.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
| name | code-review |
| description | Deep code review of a diff. Launches the code-review workflow — finder agents scan each review angle, an independent verifier judges every candidate, then a ranked, capped findings report. Read-only; it reviews, it does not edit. |
| whenToUse | When the user runs /code-review to review a PR, branch, ref range, path, or the current working changes before merge. Trigger phrases — "code review", "review this PR", "review my changes", "/code-review". |
| argumentHint | [pr-number | branch | ref-range | path | instructions] |
| userInvocable | true |
| modelInvocable | false |
| context | inline |
The user invoked /code-review. Any text in the <command-args> block is the
optional review target — a PR number, a branch, a ref range, a path, or a
free-form instruction (e.g. "only review src/foo.ts", "focus on error
handling"). If empty, the workflow reviews the current branch diff.
Your job is not to review inline. Launch the bundled code-review workflow —
it fans out one finder agent per review angle (correctness + cleanup), streams
each candidate to an independent verifier, optionally sweeps for gaps, then ranks
and caps the findings. You orchestrate it; the agents do the work.
Call the Workflow tool exactly once:
Workflow({ name: "code-review", args: "<level> <target>" })
<target> is the user's text from <command-args>, passed verbatim (omit to
review the current branch diff).<level> is the effort, chosen by you:
high, xhigh, or max, treat that as
their explicit choice and pass it through unchanged.Say at most one line before launching (the level you picked and why), then make the call. Do not read the diff or review anything yourself first — that is the workflow's job.
The workflow returns a summary, ranked findings (each with file, line, and a
failure scenario), the refuted candidates, and stats. Render the findings for
the user as a clean, readable review — never dump raw JSON. Lead with the
summary, then the findings most-severe first (correctness before cleanup), each
as file:line — one-line issue followed by the failure scenario and the fix.
Note how many candidates were refuted during verification.
This is a review, not an edit. After presenting it, offer to fix the findings — but do not change code until the user approves.
Run a prompt or slash command on a recurring interval (e.g. /loop 5m /foo). Omit the interval to let the model self-pace.
Deep whole-repo security audit. Launches the deep-security-review workflow — a prep agent maps the attack surface, specialized auditors scan each vulnerability class, candidates are debated for-and-against reachability, survivors are proven and chained, then a strict report is rendered and independently re-verified. OWASP Top 10, CWE Top 25, NIST SP 800-53/SSDF; target repo is read-only while report artifacts are written under the configured reports directory.
Reflective memory consolidation — review recent activity, synthesize learnings into the memory directory, and prune stale entries. Use when the user says "dream", "learn", "consolidate memories", "organize your memories", or asks Otherside to reflect on what it has learned.
Interview the user relentlessly about a plan or design until reaching shared understanding, resolving each branch of the decision tree. Use when user wants to stress-test a plan, get grilled on their design, or mentions "grill me".
Initialize new OTHERSIDE.md file(s) and optional skills/hooks with codebase documentation
Code review a pull request — summarize changes, flag issues, suggest improvements.