F
felipe-diniz/keel
Claude Code plugin: explore, implement, test, research and review agents, a single pre-commit gate for…
★ 4·更新 2026-11-16
在 GitHub 查看 ↗链接已于 2026-11-17 核查 · HTTP 200 · 打开的是厂商自己的官网
- ★ 数取值
- 4 ★ · felipe-diniz/keel · 取自 GitHub 搜索 API · (2026-11-19 20:00:48 核对)
- 来源明细
- 从该仓库的 GitHub 页面直接读取 · 条目类型:开源项目
- 收录与活跃
- 收录于 2026-11-16 · 最近见到 2026-11-20 · 共见到 5 次 · 被 2 个来源列出
- 评分构成
- 综合分 203 · 4 ★ · 2 个来源 · repo
常见问题(答案都来自我们的记录)
它适合谁 / 用来做什么?
我们把它归在这些标签下: Agent 技能。标签来自条目自身,不是我们强加的分类。
谁维护它、还在更新吗?
仓库
felipe-diniz/keel;我们收录于 2026-11-16;最近见到 2026-11-20。这里没有"官方认证"或"质量评级"——我们只记录来源与时间。README 原文(英文,未翻译) · 在 GitHub 读完整 README →
# keel
以下摘录自该仓库的 README(片段)
A Claude Code plugin that ships a ready-made working method: five agents, a single pre-commit review gate, a skill for running several fronts in parallel as background sessions, web skills (public sites, product screens, the modern web platform, SEO), an LGPD skill that teaches the law and hunts personal-data leaks in code, and per-developer versioned project memory. The main session thinks and decides. Wide reading, isolated writing and web research run outside it, in separate contexts.
Install it and start working. There is no setup step.
## Why use it
Quality where it matters
There is one review gate, keel:critical-review, before the commit. It runs only when the session changed a critical-risk invariant (money, auth, tenant isolation and the like).
It runs once, at the end, with tests green, at max effort. It reviews the critical hunks and their siblings (callers, consumers, the tests on that path).
Changes with no critical risk ship with tests and lint. Docs, i18n and the rest pay no review tax.
The implement agent does not touch critical domains. That work stays in the main session.
Lower cost and faster sessions
Wide reading goes to keel:explore (haiku, read-only), which returns only a synthesis with file:line references. Isolated writing goes to keel:implement (sonnet), and tests to keel:test (sonnet), reviewed by the main session. Web research goes to keel:research (sonnet). The main context stays small, so each turn is cheaper and compaction comes later.
The lean profile used for background sessions cut the per-call base from 72k to 40k tokens. The skill listing in background sessions went from 22,662 to 4,779 characters.
Each rule lives in one place, and a skill keeps a short core with the detail in references read on demand. In 3.4.0 the coordinator's core went from 23,395 to 19,693 bytes, under the 20,000 characters that come back after compaction, and what every session carries stayed at about 13.8 KB with a new skill on board. Test-pinned caps keep it there (CHANGELOG).
The 400k compaction window cost the same as 300k in measurement and compacts less often. Each compaction loses detail.
The rules say what compaction must preserve: the request in the user's own words, literal approvals, files, commands with their results, and the gate state.
Long autonomous runs
Parallel fronts run as real background sessions (claude --bg) for 12 to 20 hours, one session per front, with a queue, a resource window and a commit mode.
Questions from every front are asked up front, before the run starts.
With nobody there to type /clear, a relay replaces it: the coordinator opens its successor, checks that it is alive and exits.
Shared, versioned memory per developer
Each developer writes to their own folder in .claude/memory/, so merges never conflict on the index. The team still reads each other's lessons, and git gives history and review. Details in Versioned memory.
Guardrails enforced by hooks, not by hope
The gate hook refuses a call without a briefing, and asks for confirmation from the second pass in a session.
Inside the reviewer, a Bash guard blocks anything that changes state: git that mutates the tree, writes outside the temp directory, full test suites, builds, servers, migrations, package installs and external network.
Install it and start working. There is no setup step.
## Why use it
Quality where it matters
There is one review gate, keel:critical-review, before the commit. It runs only when the session changed a critical-risk invariant (money, auth, tenant isolation and the like).
It runs once, at the end, with tests green, at max effort. It reviews the critical hunks and their siblings (callers, consumers, the tests on that path).
Changes with no critical risk ship with tests and lint. Docs, i18n and the rest pay no review tax.
The implement agent does not touch critical domains. That work stays in the main session.
Lower cost and faster sessions
Wide reading goes to keel:explore (haiku, read-only), which returns only a synthesis with file:line references. Isolated writing goes to keel:implement (sonnet), and tests to keel:test (sonnet), reviewed by the main session. Web research goes to keel:research (sonnet). The main context stays small, so each turn is cheaper and compaction comes later.
The lean profile used for background sessions cut the per-call base from 72k to 40k tokens. The skill listing in background sessions went from 22,662 to 4,779 characters.
Each rule lives in one place, and a skill keeps a short core with the detail in references read on demand. In 3.4.0 the coordinator's core went from 23,395 to 19,693 bytes, under the 20,000 characters that come back after compaction, and what every session carries stayed at about 13.8 KB with a new skill on board. Test-pinned caps keep it there (CHANGELOG).
The 400k compaction window cost the same as 300k in measurement and compacts less often. Each compaction loses detail.
The rules say what compaction must preserve: the request in the user's own words, literal approvals, files, commands with their results, and the gate state.
Long autonomous runs
Parallel fronts run as real background sessions (claude --bg) for 12 to 20 hours, one session per front, with a queue, a resource window and a commit mode.
Questions from every front are asked up front, before the run starts.
With nobody there to type /clear, a relay replaces it: the coordinator opens its successor, checks that it is alive and exits.
Shared, versioned memory per developer
Each developer writes to their own folder in .claude/memory/, so merges never conflict on the index. The team still reads each other's lessons, and git gives history and review. Details in Versioned memory.
Guardrails enforced by hooks, not by hope
The gate hook refuses a call without a briefing, and asks for confirmation from the second pass in a session.
Inside the reviewer, a Bash guard blocks anything that changes state: git that mutates the tree, writes outside the temp directory, full test suites, builds, servers, migrations, package installs and external network.
收录徽章(可选)
想把收录标记放进 README 就复制这行,可选,不影响收录。
[](https://anclum.com/zh/s/felipe-diniz-keel/)
- 来源
- github · https://github.com/felipe-diniz/keel
- 按领域浏览
- Agent 技能
- Traction
- 4 ★
- 收录来源
- 2 个独立来源
- 收录于
- 今天
更多条目
所属领域与标签
每个标签都能点进去看同类的全部条目。数字是当前库里的真实条数。