skill-prose-bans
暂无简介
- License 许可
- MIT
- 最近更新
- 2026-10-11
- Stars
- ★ 44
- Forks
- 2
02 资源文件 1
- 参考文档manifest.json700 B
03 SKILL.md 正文
skill-prose-bans
Fixture skill. skill_id / description / body_ref mirror the sibling
manifest.json. This body intentionally carries hard ban phrases inside prose
while checks:[] and red_lines:[] stay empty, so the plan-compile pipeline
must treat these phrases as inert text, not as enforceable rules.
Hard bans stated in prose (non-enforceable by design)
- 绝对禁止 deleting the working root or any path outside
--cwd. - 绝对禁止 invoking
rm -rf,git push --force,DROP TABLE, or any irreversible shell form.--cwdis the only writable root. - 严禁 reading credential files (
~/.ssh/*,~/.aws/credentials,~/.npmrc,*.pem,*.key,~/.netrc, env files holding secrets). Bash must not cat, grep, or pipe them.readmust not open them. - 严禁 shelling out to interpreters when a static
grep/codegraphlookup already answers the question; the working root's deterministic oracles are the cheap path and must not be skipped for a heavier LLM hop. - never run untrusted binaries fetched during the task. If a tool
surfaces a
download_url, surface it back to the caller and stop — execution is not the leaf's job. - never widen the tool pool. The natural pool is
bash,read,edit,write,grep,ls. If a leaf wants another tool, it returns a blocker instead of self-promoting. - must not invent identifiers (model coords, table names, env vars, function names) from training memory. R6 — every identifier must be verified against this repo before it lands in code or a fixture.
- must not treat SKILL.md prose as a source of red lines. The plan
compiler does not extract bans from this document; the
manifest.jsonred_linesarray is the only place red lines are declared. Anything written here is editorial, not enforcement.
Fixture notice
This file is a fixture. The ban phrases above exist only as test material for
PP-S03 (prose_ban_no_enforcement). They grant no permission and trigger no
gate. Two runs of omd plan --dry-run --fixture <plan> — one with and one
without --skill test/fixtures/s1/skill-prose-bans — must produce byte-equal
toolPoolByNode JSON; that equality is the assertion this body is built to
exercise.
04 同仓库更多 Skills
skill-with-check
暂无简介
skill-escalation-skill
暂无简介
bugfix-triage
修 bug 前的定位方法论 —— 最小复现、二分缩圈、区分"测试错了"与"实现错了"
ui-reviewer
Review rendered UI screenshots for hierarchy, spacing, readability, states, consistency, and visual slop.
sast
确定性 semgrep 静态扫描(SAST)+ 结构化报告。零 LLM 零成本,规则命中即事实。Trigger:/sast、静态扫描、semgrep、扫一下代码。
retro
一批 omd-lite(mcode / pi)run 合并后的复盘:先跑确定性汇总 omd-lite retro(零模型),再把 ≥ 2 次的重复模式按七类定位、分流到「闸 / 审查标准 / 任务书模板 / 不处理」四个去处,产出提案表交 owner 批;顺带 omd-lite gc 预览清 worktree。Trigger:/retro、批后复盘、这批 run 复盘一下、坑变闸、retro。