菜单
Cross-model adversarial review + decision review. You orchestrate: translate user intent to panel flags, resolve doc names, pick lineages from models.json. Read-only JSON verdicts from scripts/panel.mjs + scripts/runner.mjs.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Use when a request needs product/design exploration, unclear requirements, multiple viable approaches, UI/UX choices, or new behavior whose intent is not yet bounded.
Use when 2+ independent tasks, failures, research lanes, exploration lanes, or path-scoped investigations can run without shared state, shared files, or sequential dependency.
Use when executing a written implementation plan inline in the current session with checkpoints and verification.
Use when adding or changing a capability surface such as a public function, API, command, prompt, workflow, schema/helper, policy, or reusable instruction.
Use when implementation is complete, verification passes, and the remaining decision is how to integrate, merge, PR, clean up, or finish the branch.
Use when creating or revising docs, plans, records, PR text, handoffs, or other prose that should be sparse, direct, and low-context.
| name | redteam |
| description | Cross-model adversarial review + decision review. You orchestrate: translate user intent to panel flags, resolve doc names, pick lineages from models.json. Read-only JSON verdicts from scripts/panel.mjs + scripts/runner.mjs. |
| user-invocable | true |
| triggers | ["redteam","adversarial review","review the diff","hack the code","implementation plan review","plan review","decision review","ADR review"] |
| allowed-tools | Bash, Read, Glob, Grep, AskUserQuestion |
The user speaks plainly. You run slash commands, resolve docs, translate intent to flags, and pick lineages from canonical files. The harness does not parse natural language.
| File | Purpose |
|---|---|
/home/mhugo/.claude/redteam/docs/orchestration.md | When to use which command; user language → flags |
/home/mhugo/.claude/redteam/docs/lineages.md | When to pick which lineage; speed/reliability hints |
/home/mhugo/.claude/redteam/docs/execution.md | Foreground vs background; do not poll |
/home/mhugo/.claude/redteam/models.json | Lineage policy — provider order and narrow model params |
/home/mhugo/.claude/redteam/model-bench.json | Generated lane evidence when present |
When the user names a lineage ("use qwen") or you need --models, read models.json for lineage names. Concrete provider/model ids come from configured Kimi aliases plus live provider catalogs. Default panel routing may use model-bench.json, but only for configured, live, metadata-backed models with a distinct failover.
| Command | Use |
|---|---|
/redteam:review | Review code/diff before shipping (default lean: 2 lineages) |
/redteam:architect | Review architecture, ADRs, system design, and route choices through the architect lane |
/redteam:plan | Review an implementation plan for grounding, sequencing, and falsifier gaps |
/redteam:decision | Review a non-architecture decision (doc or pasted prose) |
/redteam:hack | Security-only review |
/redteam:bughunt | Whole-codebase bug hunt |
/redteam:harvest | Positive pattern harvest |
/redteam:verify | Refute reported bugs or validate a fix |
/redteam:ultrareview | Deep multi-lens pre-merge sweep |
/redteam:bench | Discover/benchmark lane candidates and write routing evidence |
Other skills that explicitly call out redteam as a quality gate:
verification-before-completion: For high-stakes or contested claims, run /redteam:verify before declaring done.finishing-a-development-branch: Before the 4-option menu for non-trivial branches, optionally run /redteam:ultrareview.writing-skills: Before declaring a skill done, run /redteam:review with the security (hack) lane.quality-contracts: For contested or high-stakes contracts, run /redteam:review.purpose-first-tdd: For falsifier validation, run /redteam:verify.writing-plans: Before dispatching implementers, run /redteam:plan.systematic-debugging: Phase 1 — if similar bugs suspected, run /redteam:bughunt.Full implementation (commands, scripts, prompts, schemas) lives at /home/mhugo/.claude/redteam/. The kimi.plugin.json manifest there defines the plugin contract. This dotfile entry is the discoverable skill pointer that the skill tool loads.