codebuddy-help-me

在主 Agent 里调用本机腾讯 CodeBuddy CLI（codebuddy，别名 cbc）执行任务。

用

发给你的 Agent 一句话：

安装这个技能：https://github.com/mhgd3250905/codebuddy-help-me

手动：

git clone https://github.com/mhgd3250905/codebuddy-help-me.git ~/.agents/skills/codebuddy-help-me

示例

技能名放在任务前或后都行：

• 默认模型（跟随 codebuddy config list，当前 hy4-preview）：

  $codebuddy-help-me 审查登录流程的输入校验
  审查登录流程的输入校验 $codebuddy-help-me
  

• 显式指定模型：

  $codebuddy-help-me 用 glm-5.3 审查登录流程的输入校验
  审查登录流程的输入校验，使用模型 glm-5.3 $codebuddy-help-me
  

宿主 CLI 调用

宿主 Agent 可向 doctor/run/batch 传递 --model（省略时跟随 codebuddy config list 的默认模型，当前 hy4-preview；--effort 独立传参，默认 high，与模型解耦、无后缀推导）：

# 检查环境并验证模型可用性
python3 scripts/cb_helper.py doctor --json
python3 scripts/cb_helper.py doctor --json --model glm-5.3

# 单任务派发（显式模型）
python3 scripts/cb_helper.py run --preset review-local --model glm-5.3 --request-stdin

# 独立多任务派发（单批次统一模型，支持 1..3 路并发；这里使用默认模型）
python3 scripts/cb_helper.py batch --request-stdin

请求是单行 UTF-8 JSON；格式化多行请求请改用 --request-file。最小四字段模板、preset 额外字段和错误说明见 SKILL.md 与 references/fast-path.md。 run/batch 的五个 preset 默认传入完全免打扰三件套（-y --permission-mode bypassPermissions，并由 helper 仅向子进程注入 CODEBUDDY_IS_SANDBOX=1），任务授权和结构化验收不变；doctor 只做能力探测。

单任务与多任务 Batch

• 单任务（run）：单个明确任务（如审查、修改或验证），由 helper 运行单 lane CodeBuddy 并返回结构化结果。支持 --model 与独立的 --effort（默认 high）。
• 独立多任务（batch）：多个互不依赖的独立任务，由宿主 Agent 在单次调用中提交 batch（支持 1..3 路并发，默认 3），内部统一调度，无需手动开启多个终端；--run-timeout 是每个 lane 的独立 deadline。
• 并发与写保护：只读任务或不重叠工作区任务可并发执行；同一或重叠工作区的写任务（change/repair）自动互斥，避免重叠并发写入。

前提

codebuddy 已安装并登录（codebuddy --version 能跑通，实测 2.143.0）。本机有可用的 python3。仅支持 POSIX（macOS/Linux），v1 不支持 Windows。技能不负责安装和登录。

致谢：本技能是 mhgd3250905/antigravity-help-me 的 MIT 衍生项目。

---

English

Call your local Tencent CodeBuddy CLI (codebuddy, alias cbc) to execute tasks from inside your main Agent.

Usage — send your Agent one line:

Install this skill: https://github.com/mhgd3250905/codebuddy-help-me

Manually:

git clone https://github.com/mhgd3250905/codebuddy-help-me.git ~/.agents/skills/codebuddy-help-me

Example — skill name before or after the task, either works:

• Default model (follows codebuddy config list, currently hy4-preview):

  $codebuddy-help-me review input validation in the login flow
  review input validation in the login flow $codebuddy-help-me
  

• Explicit model:

  $codebuddy-help-me use glm-5.3 to review input validation in the login flow
  review input validation in the login flow with model glm-5.3 $codebuddy-help-me
  

Host CLI Invocations

The host agent can pass --model to doctor, run, or batch (omitting it follows the default model from codebuddy config list, currently hy4-preview; --effort is passed independently, defaults to high, and is decoupled from the model with no suffix derivation):

# Probe environment and verify model availability
python3 scripts/cb_helper.py doctor --json
python3 scripts/cb_helper.py doctor --json --model glm-5.3

# Single-task dispatch with an explicit model
python3 scripts/cb_helper.py run --preset review-local --model glm-5.3 --request-stdin

# Independent multi-task dispatch (one model per batch, 1..3 parallel lanes; default model shown)
python3 scripts/cb_helper.py batch --request-stdin

Requests are a single UTF-8 JSON line; use --request-file for formatted multi-line JSON. See SKILL.md and references/fast-path.md for the minimal four-field template, preset requirements, and validation errors. All five run/batch presets pass the full hands-free trio by default (-y --permission-mode bypassPermissions, plus CODEBUDDY_IS_SANDBOX=1 injected into the subprocess only), while task authorization and structured acceptance remain unchanged; doctor only probes capabilities.

Single Task vs. Batch

• Single task (run): Use for a single scoped task (review, edit, or verification). The helper manages one CodeBuddy lane and returns structured results. Supports --model and an independent --effort (default high).
• Independent multi-task (batch): When dispatching multiple independent jobs, the host agent submits a batch in a single invocation with bounded concurrency (1..3, default 3), without requiring manually opened multiple terminals; --run-timeout acts as an independent per-lane deadline.
• Concurrency & write safety: Read-only tasks and non-overlapping workspaces run concurrently; write tasks (change/repair) on overlapping workspaces are automatically serialized to avoid parallel overlapping writes.

Prerequisite — codebuddy is installed and signed in (codebuddy --version works; tested against 2.143.0). A working python3 is available. POSIX only (macOS/Linux); Windows is not supported in v1. The skill does not install or sign in for you.

Acknowledgment: this skill is an MIT-licensed derivative of mhgd3250905/antigravity-help-me.