Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add skills/yunsii/wezdeck/exec-risknpx skills add yunsii/wezdeck --skill exec-riskgit clone --depth 1 https://github.com/yunsii/wezdeckWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00075 | $0.01849 |
| Opus 5 | $0.00037 | $0.00924 |
| Sonnet 5 | $0.00015 | $0.00370 |
| Haiku 4.5 | $0.00007 | $0.00185 |
Grade C, and why
exec-risk scanned grade C with 1 finding against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Recursive force deletehighDestructive command
rm -rf with a variable or a broad path is one typo away from removing the wrong tree.
./openclaw/scripts/claw-run.sh --dry-run 'rm -rf /tmp/x' # classify only How it starts
The opening of the file, as written. The whole thing — 170 lines — stays where its author put it; the contents beside it link to each section on GitHub.
Exec risk (layered) — option A protocol
OpenClaw exec.mode=full does not hard-block host shell. You enforce
this skill on every host command. Skipping the gate is a protocol violation.
Platform vs classifier (division of labor)
| Layer | Local setting | Role |
|---|---|---|
| OpenClaw host policy | mode=full, ask=off |
No /approve spam for normal shell |
OpenClaw strictInlineEval |
false (personal OpenClaw · Dex) |
Do not hard-block xargs / -c at platform |
This skill (claw-run / gate) |
always | Semantic risk: rules → Grok → 飞书 if danger |
strictInlineEval=false does not disable the classifier. It only removes
platform /approve on inline carriers. Risk control is claw-run + Feishu.
Prefer: rg -l 'pat' packages over find … \| xargs rg.
Classifier soft-covers high-risk inline forms (python -c, xargs rm/sh, …);
innocent xargs rg is not auto-danger — still run via claw-run.
Required entry
Prefer claw-run.sh (gate + run in one step):
# From repo root, or use absolute path under wezterm-config/openclaw/scripts/
./openclaw/scripts/claw-run.sh -- git status
./openclaw/scripts/claw-run.sh 'ls -la'
./openclaw/scripts/claw-run.sh --dry-run 'rm -rf /tmp/x' # classify only
Gate-only (inspect without running):
./openclaw/scripts/claw-exec-gate.sh '<command>'
| exit | meaning |
|---|---|
| 0 | allow (or command finished after allow) |
| 2 | human_required — do not run; ask 飞书 |
| 3 | usage / empty |
| 4 | infra fail — treat as need human (fail closed) |
Stdout/stderr:
- On deny: JSON on stdout (and stderr). Parse
human_required,layer,reason. - On allow via
claw-run: command owns stdout; gate JSON is on stderr.
Mandatory agent loop
before ANY host shell (exec / bash -c / pipelines):
1. Prefer: claw-run.sh [--] <command>
Or: claw-exec-gate.sh <command> then run only if decision=allow
2. If exit 2 or human_required=true:
- 飞书说明 layer + reason + 完整命令
- 等待用户明确同意（是 / 确认 / yes）
- 仅在同意后: claw-run.sh --force -- <same command>
3. Never invent --force. Never “先跑再报”.
4. Trivial probes still go through claw-run (rules allow instantly; no LLM).
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2d ago First seen · 170 lines · 75 tokens per session scan C 8ab3ade152f4
exec-risk is a skill published in the GitHub repository yunsii/wezdeck (5 stars, last pushed 6d ago), licensed MIT. It adds 75 tokens to every session and 1,849 once invoked, about $0.0004 per session on Opus 5. A static security scan graded it C with 1 finding (recursive force delete). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
workflow
Create, run, monitor, cancel, or resume deterministic JavaScript workflows that orchestrate multiple Codex, Claude, Cursor, or Kimi agents. Use for parallel investigation or review, staged agent pipelines, resumable long-running work, or when the user explicitly asks for a workflow, orchestration, fan-out, or several…
orchestrate
Use only when the user explicitly types /orchestrate:orchestrate to decompose a large task, spawn a tree of parallel worker/subplanner/verifier subagents, and collect structured handoffs. Do not invoke autonomously.
launch
Launch and automate VS Code (Code OSS) using agent-browser via Chrome DevTools Protocol. Use when you need to interact with the VS Code UI, automate the chat panel, test UI features, or take screenshots of VS Code. Triggers include 'automate VS Code', 'interact with chat', 'test the UI', 'take a screenshot', 'launch…
accessibility
Primary accessibility skill for VS Code. REQUIRED for new feature and contribution work, and also applies to updates of existing UI. Covers accessibility help dialogs, accessible views, verbosity settings, signals, ARIA announcements, keyboard navigation, and ARIA labels/roles.
sessions
Agent Sessions window architecture — covers the sessions-first app, layering, folder structure, chat widget, menus, contributions, entry points, and development guidelines. Use when implementing features or fixing issues in the Agent Sessions window.
component-fixtures
Use when creating or updating component fixtures for screenshot testing, or when designing UI components to be fixture-friendly. Covers fixture file structure, theming, service setup, CSS scoping, async rendering, and common pitfalls.