← 返回 Skills 市场
Mentor
作者
Indigo Karasu
· GitHub ↗
· v2.3.0
· MIT-0
290
总下载
0
收藏
0
当前安装
3
版本数
在 OpenClaw 中安装
/install ocas-mentor
功能描述
Self-improving orchestration and evaluation engine for long-running multi-skill workflows. Analyzes journals, evaluates variants, and proposes skill improvem...
安全使用建议
What to consider before installing:
- This skill is an orchestration/evaluation engine and is designed to read all skill journals (~/openclaw/journals/) and to write proposals/decisions to the Forge intake. That behavior matches its purpose, but it also includes bundled workflow plans (notably 'contact-enrichment') that will scan your Gmail history and write facts into your Weave graph — a high-privacy operation. Only enable those plans if you understand and trust the intended data flows.
- Metadata inconsistencies to resolve first:
- The SKILL.md and plans reference a Gmail account env var ($GOG_ACCOUNT) and a 'gog' command; the skill metadata declares no required env vars or required binaries. Ask the author to declare required credentials (GOG account, Weave credentials, etc.) in the skill manifest so you can review/limit them.
- SKILL.md says it writes ExperimentRequest files to ~/openclaw/data/ocas-fellow/intake/, but skill.json does not list that write path. Confirm filesystem permissions and update the manifest to reflect all write targets.
- Operational precautions:
- Disable automated heartbeats/cron (or run them in an isolated session) until you audit the plans and confirm credential/permission mappings.
- Review bundled plans (contact-enrichment) and either remove or modify steps that access first-party signals if you do not want automated full-Gmail scans.
- If you install, consider restricting the skill's filesystem permissions (if the platform allows) or ensuring the GOG credential provided is scoped/minimal and that Weave writes are acceptable.
- If you need higher assurance, ask the publisher for an explicit list of required environment variables and exact filesystem paths the skill will read/write, and for a manifest update before enabling autonomous runs.
功能分析
Type: OpenClaw Skill
Name: ocas-mentor
Version: 2.3.0
The skill implements a high-risk self-update mechanism in SKILL.md ('mentor.update') that downloads and executes code from a remote GitHub repository, creating a potential Remote Code Execution (RCE) vector. Additionally, the 'contact-enrichment' plan (references/plans/contact-enrichment.plan.md) provides explicit instructions to scrape PII from the user's Gmail history—including relationships, life events, and employers—to build detailed profiles. While these behaviors are documented as features of the 'Mentor' orchestration role, the combination of automated code updates, broad read access to all skill journals, and deep PII extraction from private communications carries significant security and privacy risks.
能力评估
Purpose & Capability
The skill claims to be an orchestration/evaluation engine that reads journals and proposes improvements — that aligns with most instructions and the declared filesystem read of journals. However, the included workflow plans (e.g., contact-enrichment) call out first-party Gmail scanning and require a GOG account variable ($GOG_ACCOUNT) and invocation of a 'gog' tool/skill, yet the skill declares no required environment variables or required binaries. Also SKILL.md and README reference writing ExperimentRequest files to Fellow intake, but skill.json's write list omits the ocas-fellow intake path. These mismatches mean the skill's declared requirements do not fully reflect what it will attempt to do at runtime.
Instruction Scope
SKILL.md instructs Mentor to recursively read all skills' journals at ~/openclaw/journals/ (expected for evaluation). It also documents plans that will scan the user's entire Gmail history (first-party signals) and write facts into a personal graph (Weave). Those behaviors are sensitive but consistent with the 'contact-enrichment' plan. The concern is that the instructions reference environment variables, other skills, and filesystem write targets not fully declared in skill metadata (e.g., Fellow intake path), and the plan explicitly says 'Review ALL returned messages' — a high-privilege action that should be highlighted to users before enabling automated heartbeats or cron runs.
Install Mechanism
This is an instruction-only skill with no install spec in the registry (the SKILL.md contains a suggested one-line openclaw installer). No packages or external archives are automatically downloaded by the registry metadata — lowest install risk. Risk arises from runtime behavior (invoking other skills/tools) rather than install-time payloads.
Credentials
The skill declares no required environment variables or primary credential, but plan steps and runtime commands reference $GOG_ACCOUNT (for Gmail via 'gog'), and runtime expects other skills (weave, fellow, forge) to be available and authorized. Requesting no env creds in metadata while expecting sensitive credentials at runtime is inconsistent and could cause silent failures or unexpected prompts. The skill will perform sensitive operations (read Gmail, write to Weave) that require credentials; those should be explicitly declared and justified in the metadata.
Persistence & Privilege
always:false (good). The skill is allowed autonomous invocation (disable-model-invocation:false) — normal for skills. The skill.json lists explicit filesystem read/write paths that grant access to cross-skill journals and to the Forge intake directory; reading all journals is powerful but coherent with Mentor's role. There is a minor mismatch: SKILL.md describes writing to Fellow intake (~/openclaw/data/ocas-fellow/intake/) but skill.json's write list does not include that path — a permissions/manifest inconsistency to resolve before enabling automated heartbeats or cron jobs.
如何使用
- 确保已安装 OpenClaw（本地或 Docker 部署）
- 在对话框中输入安装命令：
/install ocas-mentor - 安装完成后，直接呼叫该 Skill 的名称或使用
/ocas-mentor触发 - 根据 Skill 的参数说明提供必要输入，即可获得结构化输出
版本历史
v2.3.0
**Added workflow plan support and expanded orchestration features:**
- Introduced workflow plans with commands to list, run, resume, and inspect plan execution (see new mentor.plan.* commands).
- Added plan templates and sample plans in the new `references/plans/` directory.
- Mentor now processes experiment results from Fellow automatically after each run.
- Expanded documentation for run completion and inter-skill handoffs.
- Updated default config and documentation, including instructions for pulling updates from source.
- Added README and detailed references for plan and workflow architecture.
v2.0.0
ocas-mentor 2.0.0
- Added new journal reference file: references/journal.md.
- SKILL.md updates: clarified Mentor’s responsibility boundaries and its cooperation with other skills.
- Expanded command documentation: added mentor.journal and detailed command outputs and side effects.
- Described how Mentor scans and ingests journals from all skills; introduced tracking of run_ids via ingestion_log.jsonl.
- Documented new paths and formats for handoffs with Forge (VariantProposal, VariantDecision) and updated OKRs and configuration samples.
- Added visibility into parallel evaluation by Elephas and explained parallel journal ingestion.
v1.2.0
- Added detailed documentation in SKILL.md explaining Mentor’s purpose, usage, commands, and internal mechanisms.
- Clarified when to use Mentor versus related skills (Sift, Forge, Dispatch, Corvus).
- Documented core orchestration and evaluation processes, including heartbeat cycles and evaluation loops.
- Described layered failure repair policies and safety invariants for reliable autonomous operation.
- Provided file and storage layout, validation rules, and references to supporting documentation.
元数据
常见问题
Mentor 是什么？
Self-improving orchestration and evaluation engine for long-running multi-skill workflows. Analyzes journals, evaluates variants, and proposes skill improvem... 它是一个面向 Claude Code / OpenClaw 的 AI Agent Skill 插件，目前累计下载 290 次。
如何安装 Mentor？
在 OpenClaw 或 Claude Code 对话框中运行命令「/install ocas-mentor」即可一键安装，无需额外配置。
Mentor 是免费的吗？
是的，Mentor 完全免费，采用 MIT-0 许可证，可自由下载、安装和使用。
Mentor 支持哪些平台？
Mentor 跨平台运行，可在任意部署了 OpenClaw / Claude Code 的环境中使用（cross-platform）。
谁开发了 Mentor？
由 Indigo Karasu（@indigokarasu）开发并维护，当前版本 v2.3.0。
推荐 Skills