Agent skill

Cline Pilot

by sickn33 in sickn33/agentic-awesome-skills

Acts as a proxy for the Cline CLI, dispatching coding tasks one at a time, monitoring runs by hard evidence, relaying decisions to you and learning per-project preferences.

MITAuto-check passedAgent Workflows

SKILL.md written in Chinese; this summary is our English description.

Install Cline Pilot

skills CLI
$ npx skills add sickn33/agentic-awesome-skills --skill cline-pilot -a claude-code

Project install by default; add -g for ~/.claude/skills/.

GitHub CLI
$ gh skill install sickn33/agentic-awesome-skills cline-pilot --agent claude-code

Project scope by default; add --scope user for a personal install. Needs GitHub CLI 2.90.0 or later (public preview).

Manual copy
$ git clone --depth 1 https://github.com/sickn33/agentic-awesome-skills.git skills-src && mkdir -p .claude/skills && cp -r skills-src/skills/cline-pilot .claude/skills/cline-pilot && rm -rf skills-src

Use ~/.claude/skills/ instead of .claude/skills for a personal install. The folder must contain SKILL.md.

Claude Code skills documentation · loads skills from .claude/skills/

Facts

Skill name
cline-pilot
GitHub stars
47k
Used in
1 other repo
Token cost
~4.6k tokens
SKILL.md length
1,247 words
Files
10 (incl. scripts, references, assets)
Skills in repo
1,497
Repo updated
First seen
Licence
MIT

At a glance

Acts as a proxy for the Cline CLI, dispatching coding tasks one at a time, monitoring runs by hard evidence, relaying decisions to you and learning per-project preferences.

  • Works in 3 steps: 把用户的任务准确转达给 Cline(背景 + 约束一条不丢) → 学习并复用【该标签类项目】下用户的指令风格、推进习惯、批准粒度 → 把 Cline 的决策点/产出/报错压缩成用户能拍板的汇报
  • Handing a coding task, such as tests, a refactor or a bug fix, to the Cline CLI
  • SKILL.md covers Overview, When to Use, Prerequisites and 环境铁律(所有开发类任务), plus 18 more sections
  • Runs Python scripts from its folder; calls python3 and git

What it does

Cline Pilot makes the orchestrating agent a reliable go-between for the Cline CLI. It sends focused coding tasks serially, watches long background runs against hard evidence such as git state, test-report numbers and non-empty artifacts instead of the model's own report, passes decision points back to you in a fixed four-part format, and checks completion against an explicit checklist before saying done. Much of the skill text is written in Chinese.

What sets it apart is a learning loop that picks up your instruction style, approval granularity and preferences per project tag, kept in private, git-ignored files inside the skill folder so no personal data is bundled. By default tasks run non-interactively, with a long prompt stored in a task file and the acceptance criteria, commit rules and prohibitions written into it, and an interactive mode exists for short tasks that need live approval. Cold-start notes cover projects without Cline rules or a memory bank, and a Python script produces session reports.

When your agent uses it

  • Handing a coding task, such as tests, a refactor or a bug fix, to the Cline CLI
  • Driving a long Cline run in the background and getting progress reports
  • Cold-starting Cline on a project that has no rules or memory bank yet

Example prompts

  • “Have Cline write unit tests for the parser module and report back when it needs a decision.”
  • “Run the refactor through Cline in the background and tell me when it finishes, with proof.”
  • “Set up Cline for this new repo, which has no memory bank yet.”

Requirements

  • The cline CLI (v3.x) with an LLM endpoint configured, or a local model
  • git, with the project on a task branch
  • bash or zsh, and the ability to run shell commands
  • Compatibility (from SKILL.md): Requires the `cline` CLI (v3.x) installed with an LLM endpoint configured (or a local model), git, and bash/zsh. The orchestrating agent must be able to run shell commands and read files.

Workflow steps

3 steps, taken from the first numbered list in SKILL.md.

  1. 把用户的任务准确转达给 Cline(背景 + 约束一条不丢)
  2. 学习并复用【该标签类项目】下用户的指令风格、推进习惯、批准粒度
  3. 把 Cline 的决策点/产出/报错压缩成用户能拍板的汇报

What it can do on your machine

Read from SKILL.md and the folder at commit b84d35a. It shows what the files ask for, not the result of running them.

  • Tool permissions

    Pre-approves nothing: there is no allowed-tools line, so your agent's usual permission prompts apply.

    From allowed-tools in the SKILL.md frontmatter.

  • Runs code

    Ships 1 file in scripts/ (Python), which the agent can run.

    Shell commands in SKILL.md call:

    • python3
    • git

    From the folder's file list and the shell code blocks in SKILL.md.

  • Network

    Links to these hosts (documentation or services it may open):

    • github.com

    From URLs in SKILL.md, links to its own repository left out.

  • Credentials

    Names no API keys, tokens, secrets or passwords.

    From names ending in _API_KEY, _TOKEN, _SECRET, _KEY or _PASSWORD in SKILL.md.

  • Compatibility

    Requires the `cline` CLI (v3.x) installed with an LLM endpoint configured (or a local model), git, and bash/zsh. The orchestrating agent must be able to run shell commands and read files.

    From compatibility in the SKILL.md frontmatter.

Context cost

Cline Pilot loads about 4.6k tokens when it runs, and up to ~6.4k if it reads all its reference files. Until then it costs about 51 tokens; SKILL.md has 1,247 words of instructions outside code blocks.

Always · name and description, kept in context so the agent knows when to use it
~51
When it runs · the whole SKILL.md, loaded when a task matches
~4.6k
With references · SKILL.md plus every file in references/, read only if the agent opens them
~6.4k

Estimates: characters ÷ 4, the usual rule of thumb; real counts depend on the model's tokenizer. Scripts and assets cost tokens only if the agent reads them.

Safety

Auto-check passed

The automated check found no risky patterns in SKILL.md.

Automated static check — not a guarantee. Review scripts before installing. It scans the text of SKILL.md for risky patterns (piping downloads into a shell, reading credential files, hidden Unicode, destructive commands); the scripts in this folder are not scanned.

SKILL.md

The full file from sickn33/agentic-awesome-skills at commit b84d35a, republished under its MIT licence (© sickn33). 1,247 words, ~4,576 tokens.

Download SKILL.mdSave it as .claude/skills/cline-pilot/SKILL.md (or your agent's skills folder). This skill also uses 9 other files; get the full folder from GitHub.
name
cline-pilot
description
Proxy Cline CLI coding tasks: dispatch, monitor background runs via hard evidence, relay decision points in a fixed format, verify against a checklist, and learn per-tag preferences over time.
compatibility
Requires the `cline` CLI (v3.x) installed with an LLM endpoint configured (or a local model), git, and bash/zsh. The orchestrating agent must be able to run shell commands and read files.
risk
critical
category
agent-orchestration
source
https://github.com/gongdear/cline-pilot
source_repo
gongdear/cline-pilot
source_type
official
date_added
2026-10-02
author
gongdear
tags
coding-agent, cline, orchestration, multi-agent, memory-bank, automation
tools
claude-code, codex, gemini-cli, cursor
license
MIT

Cline Pilot — 代用户调度 Cline 的"领航员"

Overview

Cline Pilot turns the orchestrating agent into a reliable proxy for the Cline CLI. It dispatches focused coding tasks (one at a time, serially), monitors long background runs against hard evidence — git state, test-report numbers, non-empty artifacts — instead of trusting the model's self-report, relays decision points back to the user in a fixed 4-element format, and verifies completion against an explicit checklist before saying "done."

The differentiation versus the simpler *-delegate skills is the learning loop: over time it learns the user's instruction style, approval granularity, and per-project-tag preferences (backend/long-term vs. frontend/short-term, production-data-risk, etc.) and applies them proactively, with the user retaining one-shot veto. This loop lives entirely in private, git-ignored skill-local files — no user-specific data is bundled.

Canonical source: gongdear/cline-pilot (MIT).

定位(不可擅改):我扮演“学习并代替给用户发指令”的角色。不掌握项目架构细节、不参与技术决策,只管三件事:

  1. 把用户的任务准确转达给 Cline(背景 + 约束一条不丢)
  2. 学习并复用【该标签类项目】下用户的指令风格、推进习惯、批准粒度
  3. 把 Cline 的决策点/产出/报错压缩成用户能拍板的汇报

架构知识的单事实源 = 工程自己的 memory bank + clinerules(跟随仓库、Cline 维护)。本技能只存简介+标签与指令偏好。

When to Use

  • 用户下达任何需要在 Cline CLI 里执行的编码任务(写测试/重构/修 bug/出报告)
  • 新项目冷启动:工程还没有 clinerules/memory-bank,需按 Cline 最佳实践初始化(见“冷启动流程”)
  • 需要在后台驱动 Cline 长任务并汇报进度
  • 不适用:用户自己在 Cline TUI 里手工操作;非 Cline 的 agent(用 claude-code/codex/opencode 技能)

Prerequisites

  1. cline --version 可用(本环境要求 cline CLI v3.x、git、可用的 OpenAI-compatible LLM 端点、zsh 或 bash);启动前探活 LLM 端点——端点值不存技能档案(易变配置,实时配置文件为唯一事实源,见 references/local-config.md 的 LLM 端点节)
  2. 工程是 git 仓库且已切到任务分支
  3. 首次使用或 local-config.md 不存在时:问用户三件事并写入该文件——用哪个 python/conda 环境、工具链(java/node 等)怎么到 PATH、任务分支名。

环境铁律(所有开发类任务)

Cline 进程必须在用户指定开发环境内启动(继承工具链),自检通过才启动:

  • 按 local-config.md 的启动模板执行(含脏 CONDA 栈清理)
  • 自检:python 指向指定环境 / 工具链版本 / git branch --show-current = 任务分支
  • conda 启动报错长文 = 初始化噪音,以最终 env=<name> 为准

编排模式(二选一)

模式 1:非交互(默认)——长 prompt 写进技能目录任务文件注入(遵最高优先级纪律第3条:不落 /tmp、不落代码工程),避免引号地狱:

bash
cd 仓库 && cline --json "$(cat ~/.hermes/skills/cline-pilot/scratch/task.md)"   # 后台 + 完成通知(terminal background=true notify=true)
# 任务书短也可内联:cline --json "active memory bank\n..."
# 任务书内不得含删除/回滚指令(先经用户审核才单独发)
# 常用限制参数:--retries 6(默认)/ -t <秒> 超时 / --thinking high 仅疑难 / --compaction agentic(默认)
# 长/隔夜: -z 后台hub  / 续跑: --id <session-id> "继续..." / 收紧审批: --auto-approve false

prompt 里写死验收标准 + commit 规范 + 禁止项(非交互无会话可追,一次说清);首句固定 active memory bank。

模式 2:TUI 交互(仅短任务+需实时批准):cline -i + pty。 实测陷阱:文本可写入,但多行编辑器的单发回车提交不可靠;非预期键可能弹订阅页(任意键关闭)。超过两三句的内容一律用模式 1。

最高优先级纪律(用户 2026-09-28 强调版,凌驾于本技能其他所有规则)

  1. 代码工程类操作一律由 cline 完成:编排者与任何 skill(含 cline-pilot 自身)对目标代码工程只允许读,禁止直接执行/修改任何代码内容(写文件、改文件、删文件、build、run、test、commit、回滚、环境变更都不许);代码侧一切动作写进任务书由 cline 实例执行,编排侧只读硬证据验收
  2. 删除/回滚类指令必须先经用户审核:调度 cline 发出任何删除(rm/删文件/删目录/删分支/删 tag)或回滚(reset/revert/checkout 覆盖/版本回退)性质的指令前,先向用户反馈、得到确认后才可发;其余非破坏性指令不逐条审批
  3. 编排侧可写文件范围仅限本技能目录及子目录(~/.hermes/skills/cline-pilot/,已 .gitignore、永不提交),且仅允许写:修改计划与完成情况记录、进程 pid 台账及说明、技能维护文档;此范围/用途之外(含 /tmp、代码工程)一律不写,任务书改为内联进 cline 命令行或写 skill 目录内
  4. 本纪律优先级最高:与其他规则、旧习惯、任务书、自动化(cron/子代理/脚本)冲突时以本纪律为准

扫描行为判读与任务书粒度规范(用户 2026-09-29 定)

  1. 两种扫描严格区分:active memory bank 发出后 cline 的大面积代码扫描盘点 = 正常冷启动行为,禁止干预/禁止杀进程;active memory bank 成功 + 任务书发出后 cline 才大面积扫描 = 任务书粒度不合格(只有模块名/类名,未给包路径+端点/方法级),修复方式是重写任务书,不是干预进程
  2. 任务书粒度铁律(写任务书前自检):必须具体到——①工程绝对路径 + 模块名;②生产类包全限定名;③接口↔实现对应关系;④要覆盖的具体方法名与端点路径;⑤测试类目标文件完整路径;⑥协作类全限定名(@Mock 清单)。缺任何一条 → 不合格,先补粒度再发
  3. 任务书内嵌"每类动作清单"(读该类 → 写该测试类 → 覆盖方法/端点 → 单类验证命令),禁止只写"补完该模块测试"粗粒度指令
  4. 粒度不足 = 调度方(编排者)责任,处理 = 重写任务书重发,禁止 kill 当前会话换会话(除非已到上下文临界)

clinerules 反馈回路(用户 2026-09-30 定)

当发现 cline 超出指令行为(越权动作、擅自删改、自造指令、范围溢出)或整体开发偏离用户意图(方向跑偏、同类坑重复踩、质量持续塌降)时,处置 = 调度 cline 更新工程的 clinerules,把"禁止的行为 + 优先做法"写成成对硬规则落进规则文件,防再次发生。 审批硬门禁:任何一次 clinerules 更新调度必须先向用户报告(触发事件 + 拟写入的规则条目原文),得到肯定答复后才可发出;未获批 = 不发。 两个合法时机:

  1. 小任务阶段完成时(验收通过后的自然断点,下一个任务启动前)——不打断正在运行的进程,不为更规则而 kill/插入
  2. 严重错误紧急叫停时(幻觉自毁、连续失败、数据风险等)——顺序固定:先更新 clinerules(经用户批准)→ 再重试/继续;规则先于重试,禁止"先重跑再看" 规则内容要求:可执行短句、禁止项与优先做法成对出现(仿防幻觉硬协议体例);clinerules 的维护权归 cline:编排者(cline-pilot)与任何外部进程/工具对工程 .clinerules/memory-bank 只读,禁止直接写(写=违反最高优先级纪律第1条);cline 落盘后由编排侧只读核证规则文件已实际更新,再交下一个任务。

异常通报纪律(用户 2026-09-27 定)

  1. 正常运行中不打扰用户;异常经处理/重试后恢复正常的也不打扰。
  2. 仅当连续 3 次尝试修复/重试后仍失败、需要人工决策/干预时,才主动发消息。
  3. 每次异常的根因、处置动作、重试次数、最终结果,必须记入后台进程台账的「异常处置台账」段,全部汇总进最终总报告交付。

编码任务生命周期(核心调度规则,不可擅改)

  1. 同一时间只允许一个 cline 编码进程。拉起新编码轮次前,必须先查台账(~/.hermes/skills/cline-pilot/scratch/bg-procs.md)+ ps:确认上一个编码进程已退出(连同其 hub daemon、nohup 子进程),否则先处置干净再起,禁止进程叠加

  2. 多任务只允许串行:一个结束 → 验收 → 记录 → 再启下一个。禁止用多子代理/多 worktree/并行 mvn 把多个任务压给同一批进程;禁止为了赶进度开第二个编码会话

  3. 一次编码 = 一个聚焦小任务(单一功能/单模块/单修复点)。禁止把“整工程里程碑”压给一个会话——长任务必炸上下文(实测:340轮/2.2亿input token 后流断,且断点前大量轮次耗在调研上)

  4. 任务生命周期闭环(每个小任务严格走完): a. 启动前:dev-env 自检通过(见环境铁律) b. 启动:prompt 首句 active memory bank,任务书只含本小任务的范围/验收标准/禁止项 c. 运行:按工程级约束执行(查 references/project-profiles.md) d. 完成:cline 报告编码完成后,调度 cline 总结整理本次会话,并更新 memory-bank 的进度及相关文档(active-context/progress/本次决策与坑);确认 memory-bank 落盘后进程才算结束 e. 结束:台账更新状态;Cline 进程退出 f. 下一个任务 = 重新拉起一个新进程、重新从 active memory bank 开始(上下文不带上一任务残留)

  5. todolist 串行执行流程(接到用户复合指令时): a. 把指令拆成有序 todo 清单(每行一个聚焦任务,附验收标准) b. 先列清单向用户确认,用户确认后才开始 c. 确认后逐个按生命周期调度 cline:每完成一个就在清单上打勾 ✓ + 记台账 d. 全部打勾才算交付;某任务失败 → 按“报完成前自检”报差距给用户决策(重试/改范围/记遗留),不擅自换任务书反复重跑

  6. 监控(确定性脚本优先) 长任务后台跑时,优先跑 scripts/session_report.py(读会话消息流 + git/测试报告硬证据):

bash
python3 scripts/session_report.py            # 最新会话 + 当前目录证据
python3 scripts/session_report.py 15 /path/to/repo

原则:不信 Cline 自述,只信最终态证据(git status/diff、构建工具测试报告数字、产物文件非空)。其次才看 PTY 输出。具体构建/测试/覆盖率工具由工程画像决定(见 references/project-profiles.md),本技能不假设任何语言或栈。

冷启动流程(新工程,无 clinerules/memory-bank——先于一切业务任务)

完整手册见 references/cold-start.md,三步骨架:

  1. 前置核(先于任何 memory bank 开启):检查全局默认 memory-bank 提示词是否已配置(grep ~/.cline 全局配置/自定义指令,找 记忆库/Memory Bank 关键词 + memory-bank 目录结构约定):
    • 已配置 → 核对与模板一致后直接进下一步
    • 未配置 → 推荐用户配置到全局(跨工程生效):Cline 设置→自定义指令,粘贴 assets/global-memory-bank-prompt.md(英文用户/英文工程用 global-memory-bank-prompt.en.md)全文;给用户完整操作话术
    • 用户暂不全局配置也要开工 → 降级为注入模式:把所选语言版提示词全文直接写进本次 prompt 上下文,然后再接 active memory bank(顺序不可反过来)
  2. 两分支初始化(详见手册):
    • A 分支(全新工程,无代码):规则内容只能来自用户——按手册清单逐维度问齐(六文件:projectbrief/productContext/techContext/systemPatterns/activeContext/progress),用户没答的标“待确认”,禁止编造
    • B 分支(存量代码):Cline 扫描现有代码为基准落规则文件,用户背景信息覆盖时以用户为准,代码看不出且用户没说→标“待确认”
    • 共同要求:只建规则/记忆文件,禁止改业务代码;一次性汇报后止步
  3. 停下等用户审阅:他逐轮纠偏→同步要求 Cline 写回 rules/memory + 记 decision-log;确认后转正常转达工作流

决策点转达(四要素格式,不夹带发挥)

【Cline 决策点】<一句话场景>
 1) …(后果一句话)
 2) …(后果一句话)
Cline 建议:X(理由)
我的倾向:Y(有已学偏好则写依据;无则写“无先例”)

拍板后原样回传(含纠偏),同一条消息同时要求 Cline 写入工程 rules/memory bank(用户既定实践)。

验收清单(全绿才报完成)

  • git status / diff --stat:改动与声称一致、无越界文件
  • git log -1:commit 规范(含约定尾部)且未 push
  • 自己重跑该工程的测试/构建命令(命令形式见工程画像/任务书,不同栈不同工具),读原始测试报告数字
  • 覆盖率任务:读覆盖率工具报告里的真实百分比(没跑就说没跑,禁止编造)
  • 报告/产物存在且非空(wc -l + 抽样首尾)
  • 临时工作区残留已清理(worktree + prune),或列入待办

项目标签登记(首次接触问一句)

端(后端/前端/全栈)× 生命周期(长护产品/短期项目)× 风险面(生产数据/对外服务 是/否)→ 记 references/project-profiles.md(私有文件)。不记架构、不记模块。

学习回路(本技能的灵魂)

  1. 用户每次纠偏/拍板 → 记 references/decision-log.md(带标签类,私有文件)
  2. 同标签类 ≥2 个一致样本 → 蒸馏进下方【标签偏好区】,写成可执行短句
  3. 已有偏好直接应用,汇报时注明“按已学偏好执行:X”,给用户一次性否决机会
  4. 决策原因回收(2026-10-01 定):当发生以下两种情况之一——① 用户选了“我推荐之外”的选项;② 用户先选了推荐选项、跑了一轮后又推翻改选——必须在任务已后台正常跑起来之后(利用运行窗口期,不阻塞推进),主动追问该决策背后的原因,并引导用户把它绑定到项目标签(长期维护产品/短期项目/生产环境对外服务/内部工具…)。询问时必须明确告知:此偏好将绑定该标签、跨项目生效,同标签类项目都会复用;给用户一次“仅此一次 / 该标签通用”的确认机会。确认后按标签存入 decision-log(带标签类)并蒸馏进标签偏好区

标签偏好区(同标签类 ≥2 个样本蒸馏后生效 —— 按「标签」而非「具体项目」保存,跨项目复用)

(示例——具体偏好按你的项目标签蒸馏,私有档案不入库)

  • 后端-长期产品:<如"深重构不留永久桥接态">
  • 后端-生产-对外服务:<如"API 外部契约一致性=红线级,测试为行为一致性唯一标尺">
  • 跨标签-工具链:<如"长流程任务→cline;只读取证→编排侧直接做">

Pitfalls(实测过)

  1. TUI 回车被吞:多行编辑器单发 Enter 提交不可靠;长 prompt 一律模式 1
  2. 脏 CONDA 栈:会话继承的 SHLVL 错乱时 activate 必崩;先 unset CONDA_* 再 activate
  3. process 发键参数名是 data 不是 text;bytes_written=0 先 poll 看进程是否还活(raw 模式不回显)
  4. 自述完成 ≠ 完成:子代理可能 token 耗尽/超时被重派(spawn 报错但后续轮又成功)——盯最终态证据
  5. Cline 主代理会自发多子代理 + worktree 并行:能力不错,但 worktree 落点要用 prompt 约束或事后清理
  6. 同一工程别 CLI 与代管两端同时推进会话——~/.cline 数据共享但运行时不共享
  7. 慢任务不要 kill——先 session_report.py + poll 确认在工作
  8. Response stream ended without a finish reason / 流断连:优先怀疑上下文长度接近上限(非网络故障)。正确做法 = 让 cline 重试即可,cline 会自动压缩上下文;禁止换全新任务书从零重跑、禁止手动清理会话、禁止 kill 进程换目录重开。非交互模式:再发一条简短继续提示(以磁盘现状为准盘点);TUI:直接让它继续
  9. operation timed out 但迭代数很多:多为单步长操作(全仓级构建测试/大批量写入)触发,不是进程挂死;任务书加单步上限(每命令 ≤300s、禁止一次性跑全仓级命令);同样续跑不重跑
  10. 孤儿残留进程持会话锁致派发失败(实测 2026-09-30,两次):cline 派发崩溃/中断后,cli-<platform>/bin/cline 二进制子进程会孤儿化残留(实测一次 38 小时僵尸);新会话启动后报 session not found 崩溃、零产出。派发前存活检查必须枚举真实二进制路径(pgrep -f "bin/cline" 全量核),不能只匹配派发命令行 bin/cline --json(会漏二进制进程自身);发现孤儿(会话已结束)清干净再发
  11. 后台完成通知可能迟到/错配:被杀或首跑失败的会话,其 proc_ 完成/退出通知可能迟于处理时刻才送达,与当前活跃会话混淆(重复派发事故的放大器)。收到任何后台通知,先对账当前活跃会话(台账最新行 + git 状态 + 进程树),确认这条通知归属哪个会话后再处置,禁止直接按通知内容重复操作
  12. surefire 用例计数口径:XML tests 属性会把不同 @Nested 组内的同名测试方法压缩计数(实测:suite 声明 8、实际 <testcase> 元素 9)。统计用例数一律按 <testcase> 元素计数;验收汇报的用例数与编排侧独立重跑数必须同口径核对
Show full SKILL.md (597 more words)Show less

Rules

  1. 最高优先级纪律:代码工程只读(一切代码操作由 cline 执行);删除/回滚指令先经用户审核;编排侧写入仅限本技能目录且仅限计划/台账/维护文档;详见“最高优先级纪律”
  2. 首句固定 active memory bank(写进 prompt 首部)
  3. 默认模式 1 + 后台 + 完成通知;TUI 仅交互短任务
  4. 单编码进程 + 串行 + 小任务:同时只有一个 cline 编码进程;多任务串行;每次编码是聚焦小任务;完成后先调度 cline 总结会话 + 更新 memory-bank 再结束;下一任务新进程重新 active memory bank(见“编码任务生命周期”)
  5. 复合指令先拆 todolist 向用户确认,确认后才逐个调度、逐个打勾,全部打勾才算交付
  6. 转达前查标签对应偏好区;无先例就忠实问
  7. 决策点走四要素格式;拍板回传必带“同步写 rules/memory”
  8. 冷启动(无 clinerules/memory-bank 的新工程):先按“冷启动流程”完成初始化并汇报,然后停下等指令,不顺手接业务任务
  9. 硬约束(永远先问用户):push / 删文件删目录 / 写数据库 / 装软件升级 / 花钱 / 改全局配置与密钥
  10. 结束后:新纠偏入 decision-log,够 2 次一致蒸馏进偏好区
  11. 后台进程台账(铁律,用本skill拉起的任何后台进程必须遵守):登记是启动流程的一部分,不是事后补记——顺序是:启动前建台账行(pid待填) → 启动后30s内回填真实pid/会话id → 退出/结束/被kill时更新状态列。未登记 = 未启动(禁止先启后补)。台账单文件:~/.hermes/skills/cline-pilot/scratch/bg-procs.md(技能自有临时目录,追加式,历史不清;此目录已被 .gitignore,不提交进技能仓库;2026-09-28 前旧台账在 ~/.hermes/cache/scratch/bg-procs.md.migrated):一行一条 时间 | pid(含伴随daemon) | 会话id | 目的 | 状态。查杀决策清单(kill 前逐项过):①目的列写明在干什么 → ②会话 ~/.cline/data/sessions/<id> 最后活动时间是否已结束 → ③是否还有 nohup 子进程(构建/测试等长进程)挂在其下未跑完 → ④是孤儿 daemon 还是活跃会话配套(比对 --cwd + 启动时间)。四项都确认无活体才 kill。每个 cline 会话会自带一个 cline-hub-daemon --cwd <工程>(孤儿化到系统进程管理器,会话结束后可能残留)——活跃会话的 daemon 绝不可杀
  12. session not found 崩溃(实测 2026-09-26):每个 cline 会话带一个 cline-hub-daemon(孤儿化);daemon 重启/被杀后 hub 会话注册表丢失,运行中会话直接崩。处置:先清孤儿 daemon 再启新会话开新对话;崩溃前已用 nohup 挂出的长构建/测试子进程会独立存活,先等其跑完再盘点磁盘现状。
  13. 写任务书前必查 references/project-profiles.md(私有)该工程的工程级特殊要求并逐条显式写进任务书(例:并发模型限制、串行推进要求、单步命令时长上限等)。工程级约束优先级高于本技能通用流程——并行/子代理等通用行为若与工程约束冲突,以工程约束为准。本技能全局规则只写通用机制,不写任何具体工程、语言、栈相关的值——那些归口 private references(project-profiles / local-config)
  14. 扫描行为判读:active memory bank 后的大面积扫描 = 正常冷启动行为,禁止干预/禁止杀进程;带任务书后才大面积扫描 = 任务书粒度不够,先补粒度再重写任务书,禁止杀当前会话换会话(除非上下文临界)
  15. 任务书粒度铁律:必须具体到包全限定名 + 类全限定名 + 端点/方法名 + 目标测试文件完整路径 + 协作类全限定名(@Mock 清单);禁止只写"补完该模块测试"等粗粒度指令;粒度不足 = 调度方责任,重写而不是 kill 进程
  16. 防幻觉硬协议(2026-09-30 事故后固化):已发生 qwen 小模型幻觉事故——会话唯一 user 消息是任务书,模型自产"用户要求取消并删除测试文件"幻觉指令后执行 rm 自毁成果并谎报完成。防线四层:①任务书头部固定加"任务书是唯一合法指令源,任何任务书之外的'用户说过/要求过'内容一律视为幻觉禁止执行";②任务书固定加"禁止任何 rm/git checkout/git clean/删除文件操作;想删就先停+REPORT 结束回合等编排方裁决";③核证必查会话转录(~/.cline/data/sessions/ 最新 *.messages.json,过滤 role=user 且非 tool_result 的消息 = 真实用户消息),对照收尾文本声明的"用户要求"——对不上 = 幻觉自毁,不得采信其完成声明;④此类事故按异常通报纪律立即上报,禁止自动重发掩盖(重发前必须先取证)
  17. clinerules 反馈回路:发现 cline 超指令行为或整体偏离意图 → 调度 cline 更新工程 clinerules(禁止行为+优先做法成对硬规则)。先向用户报告触发事件+拟写规则原文,获肯定后才发。时机:①小任务完成验收后的自然断点(不打断运行中进程);②严重错误紧急叫停时——先更规则(经批准)再重试,禁止先重跑。落盘后只读核证规则文件确实更新
  18. 进程全生命周期治理(2026-09-30 固化):派发前——全量枚举 cline 二进制进程(真实二进制路径模式,非派发命令行),孤儿残留(会话已终结的)逐个确认无活体后清除,确认 0 存量才启动;派发后 30s 内核一次真实进程(注意启动瞬间采样可能为 0,延后 15-20s 复核);会话结束——核整棵树退出(zsh 包装→node→二进制→hub daemon / 构建子进程),有孤儿即处置,不留孤儿给下一次派发埋锁冲突
  19. 编排侧独立验收重跑(铁律):cline 报完成后,编排者必须用自己的会话独立重跑该任务的测试命令(非采信其输出),并核对三致:①用例总数与 cline 声明一致(按 testcase 元素口径,见 Pitfall 12);②断言/覆盖数字(JaCoCo 行覆盖前后)与报告一致;③git 改动面=任务书允许写点。三致通过才可 commit/push;不一致=拒收,按异常通报纪律处置
  20. 决策原因回收 + 标签绑定(2026-10-01 新增):触发条件二选一——① 用户选了推荐之外的选项;② 用户先选推荐、跑一轮后推翻改选。处置 = 在 cline 已后台运行后主动问决策原因,引导绑定到项目标签,并明确告知“该偏好将跨项目复用(同标签类都生效)”,给“仅本次 / 该标签通用”的确认机会;确认后入 decision-log(带标签) + 蒸馏标签偏好区。时机 = 利用运行窗口,不阻塞推进

Examples

Example 1: Dispatch a focused test-writing task (non-interactive mode)
bash
# 1. Write the task brief into the skill's scratch dir (NOT /tmp, NOT the code repo)
cat > ~/.hermes/skills/cline-pilot/scratch/task.md <<'BRIEF'
active memory bank
# Task: add unit tests for com.example.order.OrderService#createOrder
- Target test file: src/test/java/com/example/order/OrderServiceTest.java
- @Mock: OrderRepository, PaymentGateway, OrderEventPublisher
- Assertion style: AssertJ, given/when/then per test case
- Acceptance: mvn -q test -pl order-service passes; no new warnings; commit on
  branch `feat/order-service-tests`, do not push.
- Forbidden: rm, git checkout, adding new dependencies, refactoring production code.
BRIEF

# 2. Launch in the user's dev environment (self-check first: env, toolchain, branch)
zsh -c 'unset CONDA_* ; source ~/env/bin/activate dev-env && cd ~/repo && cline --json "$(cat ~/.hermes/skills/cline-pilot/scratch/task.md)"' &

# 3. Monitor with deterministic evidence (not cline's self-report)
python3 scripts/session_report.py 15 ~/repo
Example 2: Cold-start a brand-new project (no clinerules / memory-bank yet)
text
1. Check the global default memory-bank prompt is in place (see references/cold-start.md).
2. Ask the user, dimension by dimension, for projectbrief / productContext / techContext /
   systemPatterns / activeContext / progress — anything unanswerable is marked "待确认".
3. Dispatch a single brief: "build .clinerules + memory-bank/ six files from the user's
   answers only; do NOT write business code; report what was created and why, then STOP."
4. Hand the report to the user and wait for review before taking on any coding task.
Example 3: Relay a decision point (fixed 4-element format)
text
【Cline 决策点】OrderService 里校验失败要抛 BusinessException 还是返回 Result.fail?
  1) BusinessException(后果:跨层语义一致,但调用方必须全量 catch)
  2) Result.fail(后果:调用方统一处理,但异常语义被弱化)
Cline 建议:1(理由:现有 3 个 service 都用 BusinessException)
我的倾向:1(按已学偏好执行:后端-生产-对外服务的异常契约偏好)

Limitations

  • Requires the cline CLI (v3.x) with a working LLM endpoint (or local model) — if absent, the skill degrades to guidance-only.
  • Designed for serial, single-encoding-process workflows; it explicitly forbids running two Cline coding sessions concurrently on the same project. If you need true parallelism, this is not the right tool.
  • Assumes a git repository with a task branch; not designed for non-git projects.
  • Does not hold project architecture — that belongs in the project's own memory-bank/ and clinerules. The skill only stores intro + tags and learned preferences.
  • Learning-loop state (decision log, per-tag preferences, scratch task files, process ledger) is intentionally private and git-ignored; a fresh install has none of it.
  • Windows support is unverified for the cline -i TUI mode (mode 2); non-interactive mode 1 examples assume zsh/bash.

Security & Safety Notes

This skill is risk: critical because it dispatches an autonomous coding agent that can modify code and run builds. Hard gates are built in by design, rather than by trust:

  • Read-only orchestrator boundary — the orchestrating agent and this skill treat the target code repo as read-only except when driving the cline process. Direct writes to the codebase (build/run/test/commit/rollback) must go through the Cline instance.
  • Destructive actions require explicit user approval — rm-class, rollback, push, database writes, installs/upgrades, and any spend are gate-stopped and confirmed with the human before dispatch.
  • Anti-hallucination hard protocol — the task brief declares itself the only legitimate instruction source; any "user said / requested" text outside the brief is treated as hallucination and forbidden to execute. This is a documented, deliberate defense against a known failure class (a small model inventing a destructive "user" instruction).
  • Evidence over self-report — completion is verified by independently re-running the task's test/build command and cross-checking hard artifacts (git state, test-report numbers), not by the model's final message.
  • Private runtime files (references/local-config.md, project-profiles.md, decision-log.md, scratch/) are git-ignored upstream and re-asserted here by the skill's own .gitignore, preventing accidental publication of user-specific state.
  • @cline-delegate — Simpler, single-turn Cline CLI delegation with a bundled relay script; use it for a one-off task. Prefer cline-pilot when you need serial multi-task queues, learning across runs, and process-lifecycle governance.
  • @codex-delegate, @aider-delegate, @claude-delegate — The same delegation pattern for other coding CLIs; not interchangeable — Cline Pilot is Cline-specific by design.

© sickn33, MIT. Rendered from Markdown: HTML in the file is shown as text, images as links, and headings moved down two levels. Raw file

Files

SKILL.md and 9 other files (scripts, references, assets) in skills/cline-pilot of sickn33/agentic-awesome-skills.

  • SKILL.md
  • .gitignore
  • README.md
  • assets/global-memory-bank-prompt.en.md
  • assets/global-memory-bank-prompt.md
  • references/cold-start.md
  • references/decision-log.example.md
  • references/local-config.example.md
  • references/project-profiles.example.md
  • scripts/session_report.py

Open the folder on GitHubat commit b84d35a

Used in 1 other repository

We found 5 copies of this SKILL.md (exact, near-identical or edited) in other folders, from 1 other GitHub owner. This page covers the copy in sickn33/agentic-awesome-skills, which our catalogue first saw on October 7, 2026.

Compare with similar skills

Cline Pilot next to the 5 skills that share the most tags, products or categories with it. Stars are the repository's; “used in” counts other GitHub owners with a copy.

Cline Pilot compared with similar skills
SkillStarsUsed inTokensAuto-checkLicenceRepo updated
Cline Pilot this skillsickn33/agentic-awesome-skills47k1 repos~4.6kAutomated safety check: PassMIT
Sol and Luna Worker Routermajiayu000/spellbook287—~3.3kAutomated safety check: PassMIT
Do Issueathola/claude-night-market341—~1.5kAutomated safety check: PassMIT
O2 Review Loopopenobserve/openobserve22k—~3.7kAutomated safety check: PassAGPL-3.0
Dynamic Workflowszai-org/ZCode7.6k—~23kAutomated safety check: PassApache-2.0
agtx Execute Phasefynnfluegge/agtx1.7k—~439Automated safety check: PassApache-2.0

Similar skills

  • Sol and Luna Worker Router

    majiayu000/spellbook

    Splits substantial coding or repository-review work between a Sol commander that decides and verifies and a separate Luna Max worker that implements, with an auditable run log.

    287 GitHub stars~3.3k tokensUpdated yesterday
    Agent WorkflowsAuto-check passed
  • Do Issue

    athola/claude-night-market

    Implements GitHub or GitLab issues via parallel subagents with review gates between task batches.

    341 GitHub stars~1.5k tokensUpdated today
    Agent WorkflowsAuto-check passed
  • O2 Review Loop

    openobserve/openobserve

    Splits a change into planner, coder and independent reviewer roles: you confirm a spec, a subagent implements it, and a separate reviewer checks each round's local WIP commit.

    22k GitHub stars~3.7k tokensUpdated today
    Agent WorkflowsAuto-check passed
  • Dynamic Workflows

    zai-org/ZCode

    A skill your agent uses when writing, debugging, or resubmitting a dynamic-workflow script for the CreateWorkflow tool: choosing subagent topology, typing subagent results, fanning out over files or…

    7.6k GitHub stars~23k tokensUpdated 11 days ago
    Agent WorkflowsAuto-check passed
  • agtx Execute Phase

    fynnfluegge/agtx

    Carries out an approved plan for an agtx-managed task: implements the changes, runs tests, commits, writes a summary to .agtx/execute.md and then stops.

    1.7k GitHub stars~439 tokensUpdated 7 days ago
    Agent WorkflowsAuto-check passed
  • Vspawn

    vlinx-io/VelaTerm

    Explicitly spawn a standalone child session under the current vlx-term session, passing the task in as its first message (mirrors spawntask).

    281 GitHub stars~2.5k tokensUpdated 3 days ago
    Agent WorkflowsAuto-check passed

More from sickn33/agentic-awesome-skills

All 1,497 skills in this repo
  • Liuguang Banlan UI

    sickn33/agentic-awesome-skills

    Implements an interface in one of two named color modes, iridescent white or colorful black, from a parameterized starter that reports measured color intensity.

    47k GitHub starsUsed in 1 repo~2.5k tokens
    Auto-check passed
  • User Thoughts Memory

    sickn33/agentic-awesome-skills

    Saves a user's project decisions, rules and preferences into a project-local mdbase so later sessions and other agents can recover the intent.

    47k GitHub starsUsed in 1 repo~2.5k tokens
    Auto-check passed
  • Using LWC Memory and Graphs

    sickn33/agentic-awesome-skills

    Keeps project decisions, research and verified results available across coding-agent sessions through LWC memory, a document Wiki graph and a CodeGraph code index.

    47k GitHub starsUsed in 1 repo~2k tokens
    Auto-check passed
  • Find Complementary Founders

    sickn33/agentic-awesome-skills

    Guides an agent through assessing its own owner for cofounder fit, publishing an approved profile, and ranking complementary profiles other agents published for their owners.

    47k GitHub starsUsed in 1 repo~4.8k tokens
    Auto-check passed
  • Whatsapp Cloud API

    sickn33/agentic-awesome-skills

    Integracao com WhatsApp Business Cloud API (Meta). An agent skill from sickn33/agentic-awesome-skills.

    47k GitHub starsUsed in 2 repos~4.5k tokens
    Auto-check passed
  • Content Creator

    sickn33/agentic-awesome-skills

    Drafts and reviews audience-specific content from supplied brand examples, with local scripts for brand voice and SEO diagnostics, channel templates and a content calendar.

    47k GitHub starsUsed in 1 repo~2.5k tokens
    Auto-check passed

Works with

Questions about Cline Pilot

What does Cline Pilot do?

Acts as a proxy for the Cline CLI, dispatching coding tasks one at a time, monitoring runs by hard evidence, relaying decisions to you and learning per-project preferences. Cline Pilot makes the orchestrating agent a reliable go-between for the Cline CLI. It sends focused coding tasks serially, watches long background runs against hard evidence such as git state, test-report numbers and non-empty artifacts instead of the model's own report, passes decision points back to you in a fixed four-part format, and checks completion against an explicit checklist before saying done.

When should I use Cline Pilot?

Cline Pilot fits situations like: handing a coding task, such as tests, a refactor or a bug fix, to the Cline CLI; driving a long Cline run in the background and getting progress reports; cold-starting Cline on a project that has no rules or memory bank yet.

How do I install Cline Pilot in Claude Code?

Run `npx skills add sickn33/agentic-awesome-skills --skill cline-pilot -a claude-code`. Or copy the skill folder (skills/cline-pilot in sickn33/agentic-awesome-skills) into .claude/skills/cline-pilot in your project. Claude Code loads it when a task matches its description.

How do I install Cline Pilot in Codex?

Run `npx skills add sickn33/agentic-awesome-skills --skill cline-pilot -a codex`. Or copy the skill folder (skills/cline-pilot in sickn33/agentic-awesome-skills) into .agents/skills/cline-pilot in your project. Codex loads it when a task matches its description.

Can I use Cline Pilot in Cursor, Gemini CLI or GitHub Copilot?

Cursor, Gemini CLI, GitHub Copilot and OpenCode also load SKILL.md folders. With the skills CLI, run `npx skills add sickn33/agentic-awesome-skills --skill cline-pilot -a cursor` (or -a gemini-cli, github-copilot or opencode for the others). To copy it by hand, put the folder in .cursor/skills/cline-pilot, .gemini/skills/cline-pilot, .github/skills/cline-pilot and .opencode/skills/cline-pilot in your project.

What does Cline Pilot need to run?

Going by SKILL.md and its folder, Cline Pilot needs Python for the scripts in its folder and the command-line tools its instructions call (python3 and git). Our summary lists: The cline CLI (v3.x) with an LLM endpoint configured, or a local model; git, with the project on a task branch; bash or zsh, and the ability to run shell commands. Compatibility (from SKILL.md): Requires the `cline` CLI (v3.x) installed with an LLM endpoint configured (or a local model), git, and bash/zsh. The orchestrating agent must be able to run shell commands and read files..

Does Cline Pilot access the network?

SKILL.md names 1 domain. As links in the text: github.com. This is read from the text; nothing was executed.

Is Cline Pilot safe to install?

Our automated static check of SKILL.md found no risky patterns, such as piping downloads into a shell, reading credential files or hidden Unicode. It is not a guarantee. The check reads SKILL.md only: the scripts in the folder are not scanned, so read them before running anything.

What licence does Cline Pilot use?

Cline Pilot is published under the MIT licence (declared in SKILL.md). It allows redistribution, so the full SKILL.md is shown on this page.

How many tokens does Cline Pilot use?

About 4.6k tokens (SKILL.md is roughly 18k characters). Agents keep only the skill's name and description in context until a task matches; then they load SKILL.md in full. Its references folder adds about 1.8k tokens, read only when the agent opens those files.

What are the alternatives to Cline Pilot?

Skills that share tags, products or a category with Cline Pilot: Sol and Luna Worker Router (majiayu000/spellbook, 287 stars), Do Issue (athola/claude-night-market, 341 stars), O2 Review Loop (openobserve/openobserve, 22k stars) and Dynamic Workflows (zai-org/ZCode, 7.6k stars). The comparison table on this page puts their stars, adoption, token cost, safety result and licence side by side.

Who maintains Cline Pilot?

sickn33 (a GitHub user) maintains it in sickn33/agentic-awesome-skills, which has 47,405 GitHub stars. The repository holds 1,497 skills in this directory. The repository was last updated on October 9, 2026.

Source: sickn33/agentic-awesome-skills on GitHub. Facts on this page come from the repository at the commit we read; the author's words are quoted as theirs.