メニュー
Run shell/terminal commands in the Agentium Docker stack safely — pytest, alembic, ruff/black/mypy, docker compose, make, and Postgres/Redis/Chroma inspection. Covers the ShellTool list-vs-shell rule and safe-bash discipline.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Use the 'vector_db' agent tool to read from and write to the shared ChromaDB vector store that powers Agentium's collective RAG memory. Covers query/get/add actions, the writable-collection allow-list, metadata conventions, and how to surface this knowledge to other agents through ChromaDB.
Core operating knowledge every Agentium agent should consult before acting: which tool or skill to use for a given intent (fetch a URL, search the web, run a shell command, read/write collective memory, access host files), host-system access patterns, and general best practices. Use it when you are unsure which capability applies, or before touching the host filesystem or running commands. Skill file at backend/.agentium/skills/operating_knowledge/SKILL.md.
Crawl a website beyond a single page using the web_crawler tool: depth-limited, polite link traversal that follows links and returns each page as clean Markdown. Use it when you need multi-page content, sitelinks, or structured site reading rather than one URL (web_fetch) or search results (web_search). Covers robots.txt respect, rate limiting, and depth limits. Skill file at backend/.agentium/skills/web_crawling/SKILL.md.
Run code in an isolated Docker sandbox via the code_execution tool. Raw data never leaves the sandbox; you receive only a structured summary. Use for safe computation, data transforms, or running untrusted snippets. Restricted to 0xxxx/1xxxx/2xxxx tiers. Skill file at backend/.agentium/skills/code_execution/SKILL.md.
Discover registered Agentium tools by describing what you need, using the tool_search tool. Returns ranked tool names and descriptions scoped to your tier, so you can find the right tool without seeing the entire tool list. Use the get action to fetch one tool's full descriptor. Skill file at backend/.agentium/skills/tool_search/SKILL.md.
Fetch a URL and return its page content as clean Markdown with a token budget, so an agent can read a web page without burning context. Use the web_fetch tool when you have a specific URL and need its content (not search results). Skill file at backend/.agentium/skills/web_fetch/SKILL.md.
| name | bash |
| description | Run shell/terminal commands in the Agentium Docker stack safely — pytest, alembic, ruff/black/mypy, docker compose, make, and Postgres/Redis/Chroma inspection. Covers the ShellTool list-vs-shell rule and safe-bash discipline. |
| skill_type | automation |
| domain | devops |
| complexity | intermediate |
| tags | ["bash","shell","docker","terminal","devops","testing","migrations"] |
| creator_tier | head |
Agentium runs entirely in Docker. You execute shell through ShellTool.execute(command: List[str]),
which calls subprocess.run without a shell.
Pipes, redirects, &&, ||, $(), and globs do NOT work unless you wrap the whole
command in bash -lc "..." and quote the inner string:
bash -lc "docker compose exec -T backend pytest tests/foo.py -q 2>&1 | tail -40".
make + docker compose (make up, make down,
make restart, make test-integration, make audit, make benchmark,
make docker-scout).docker compose exec -T backend bash -lc '...'.Follow tools/host_path.py: ~ → /host_home, other absolute paths → /host,
/tmp and relative → container-local. Never write Sovereign/user files into the
container filesystem.
See safety.md. One-liner: set -euo pipefail, quote "${var}", [[ ]] not
[ ], no eval, check return codes, idempotent, explicit ./ for globs, preview
destructive ops. ShellTool already blocks rm -rf /, mkfs, dd if=/dev/zero,
shutdown, reboot.
See commands.md for copy-paste recipes (tests, DB/alembic, Redis/Chroma, lint/
format/type, logs/health, lifecycle, git).
This skill ships runnable helpers. The __SKILL_DIR__ token resolves to this
skill's absolute path inside the container
(/app/backend/.agentium/skills/bash). Invoke a helper through bash -lc:
bash -lc '__SKILL_DIR__/scripts/agent-health.sh'
Datasets (if any) live at __SKILL_DIR__/datasets/.
Unhealthy services, migration conflicts, port clashes, offline model downloads,
secret-scan failures. Agent foot-guns: guessing commands, unquoted vars, ignoring
non-zero exits, unvetted rm -rf.