メニュー
Use when scanning a PR diff for test-runtime policy violations (network egress, subprocess spawning, sensitive-env access, writes outside the working tree, cache tampering, toolchain modification, or combined indicators).
デフォルトでは、最初にソースを確認する Prompt が選択されています。直接コマンドに切り替えるか、ローカルコピーをダウンロードすることもできます。
インストールを決める前に、SKILL.md と SkillsMP に表示されている付属ファイルをお読みください。
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。
直接コマンドでは確認用 Prompt が省略されます。実行前にソースを確認してください。
npx skills add https://github.com/nasa/fprime --skill ci-test-runtime-policyコマンドは1行のまま表示されます。コピー前に横へスクロールして全体を確認してください。
ローカルで確認しますか？SkillsMP が現在取得できるファイルをダウンロードできます。
Work with F Prime's custom CMake layer: the `register_fprime_*` API, build targets, toolchains and platform files, implementation selection, configuration modules, sub-builds, and `settings.ini`. Use when adding or wiring a module into a build, adding a custom build target, porting to a new platform or toolchain, choosing or defining an Os implementation, or diagnosing generate/build failures that are not compiler errors.
Use when an F Prime development task should be driven to convergence through repeated local test-and-review iterations — running unit, integration, and manual GDS tests plus the full multi-agent review locally, and fixing findings until tests pass and findings trend to zero. Requires explicit user approval before use. Keywords: F Prime, iterative development, iteration loop, local review, convergence.
Use when posting inline review comments, the per-agent summary review, or interacting with GitHub review threads (resolve, un-resolve, reply) for an F Prime PR review agent.
SOC 職業分類に基づく
SKILL.md を表示中
| name | ci-test-runtime-policy |
| description | Use when scanning a PR diff for test-runtime policy violations (network egress, subprocess spawning, sensitive-env access, writes outside the working tree, cache tampering, toolchain modification, or combined indicators). |
This skill defines the test-runtime policy the security agent enforces on PRs. Tests, fixtures, conftests, CI helpers, and workflow steps must not initiate outbound network connections, spawn processes outside the test harness, read sensitive environment variables, write outside the working tree, tamper with cache / persistence mechanisms, or modify the toolchain in ways that affect other runs.
The recipes below are starter signals, not the contract. The contract is the agent's judgement that the change conforms to the policy. A motivated author can craft code that evades any specific regex; the agent is therefore required to read PR-added lines in test and CI paths with attention to intent — anything that smells off in the runtime path is in scope, regardless of whether it matches the patterns below.
This skill is consumed by security-review.agent.md (the runtime-policy
category).
For every recipe:
ctypes, os.urandom to seed
a name resolver), code in nonstandard file types loaded by tests.The agent does NOT need a regex match to raise a finding; the policy violation is sufficient grounds.
git diff <base>...<head> -- \
'*test*' '*tests*' '*Test*' '*Tests*' \
'*fixture*' '*Fixture*' '*setup*' \
'.github/' 'ci/' '*conftest*' | \
grep -nE '^\+' | \
grep -nE 'socket\(|connect\(|getaddrinfo|gethostbyname|requests\.|urllib|urlopen|httpx|aiohttp|fetch\(|curl |wget |nc |netcat |/dev/tcp/|os\.popen|subprocess\.(Popen|run|check_call|check_output|call)|popen|system\(|exec\(|fork\(|http://|https://|ws://|wss://|ftp://'
Policy: test code must not initiate outbound network connections.
The agent inspects each hit and decides:
**must fix**.pypi.org,
github.com, project mirrors) consistent with the existing build
pattern → not a finding.**suggestion**, with a
maintainer ping if the legitimacy is unclear.git diff <base>...<head> -- '*test*' '*tests*' '*Test*' '*Tests*' \
'*setup*' '.github/' 'ci/' | \
grep -nE '^\+' | \
grep -nE 'subprocess\.|Popen|os\.system|os\.popen|os\.exec|posix_spawn|posixpath\.|shell=True|fork\(|execv|execp|execl|execve|/bin/sh|/bin/bash|sh -c|bash -c'
Policy: test code must not spawn processes outside the test harness.
Hits in test code that did not previously invoke subprocesses are at
least **could fix**; if the subprocess is invoked with shell=True
and the command line is composed from PR-controlled or external
inputs, it is **must fix**.
git diff <base>...<head> -- '*test*' '*tests*' '*Test*' '*Tests*' \
'*setup*' '.github/' 'ci/' | \
grep -nE '^\+' | \
grep -nE 'os\.environ|getenv\(|ENV\[|process\.env|System\.getenv|\$\{?(GITHUB_TOKEN|GH_TOKEN|NPM_TOKEN|PYPI|AWS_|GCP_|AZURE_|SECRET|TOKEN|PASSWORD|API_KEY|PRIVATE_KEY|SSH_)'
Policy: test code must not read sensitive environment variables (tokens, secrets, credentials).
Reading GITHUB_TOKEN in a path that also has outbound network
activity is a strong signal of unauthorized secret access →
**must fix**. Reading other sensitive env vars without obvious
legitimate purpose → at least **suggestion** with a maintainer
ping.
git diff <base>...<head> -- '*test*' '*tests*' '*Test*' '*Tests*' \
'*setup*' '.github/' 'ci/' | \
grep -nE '^\+' | \
grep -nE 'open\([^,]*[\"'"'"']/(etc|home|root|tmp|var|usr|opt|boot|sys|dev|proc)/|>/etc/|>/home/|>>/etc/|>>/home/|chown |chmod 7|/usr/bin/install |mv /|cp /|symlink |os\.rename\([^,]*[\"'"'"']/'
Policy: test code must not write outside the working tree.
Tests that write to /etc, /usr, /home, or modify ownership /
permissions outside the working tree are at least **suggestion**;
writing to runner-shared paths (/tmp/<known cache>, the GHA cache
directory, the artifact directory) requires verification and is at
least **could fix**.
git diff <base>...<head> -- '.github/' 'ci/' | \
grep -nE '^\+' | \
grep -nE 'actions/cache|setup-buildx-action|cache-from|cache-to|GITHUB_TOKEN|secrets\.|env:|GITHUB_PATH|GITHUB_ENV|::set-(env|output)|::add-path|core\.exportVariable|core\.addPath'
Policy: test code must not tamper with cache, artifact, or cross-run persistence mechanisms in ways influenced by PR content.
New uses of cache actions or persistence mechanisms in combination
with PR-controllable code → **must fix** if the cache key is
influenced by PR content; otherwise at least **could fix**.
::set-env, ::set-output, ::add-path are deprecated workflow
commands and indicate untrusted command injection; their appearance
in a PR-added workflow is **must fix**.
git diff <base>...<head> -- 'requirements*.txt' 'pyproject.toml' \
'setup.py' 'setup.cfg' 'package.json' 'package-lock.json' \
'pnpm-lock.yaml' 'yarn.lock' '.github/workflows/' \
'.github/actions/' '.github/scripts/' \
'Dockerfile*' 'docker-compose*' | \
grep -nE '^\+'
Policy: toolchain modifications that touch secrets or network egress fall under this policy in addition to supply-chain review.
Any + line here is supply-chain-reviewer territory primarily; the
security agent additionally reviews for runtime-policy contributions:
curl … | sh, wget … | bash → **must fix** unless the source
is hash-pinned and provably trustworthy.**must fix** if a
package install path also touches secrets or network egress in
combination.If two or more of §1, §3, §5 are true on PR-added lines, the
combination is a strong indicator of an attempted runner compromise.
The agent emits a runtime-policy **must fix** with body:
[Security] **must fix** Multiple runtime-policy indicators introduced into the
CI path: <list the matches>. The combination resembles an attempted
runner compromise (network egress + sensitive-env read + persistence).
The agent recommends not running CI on this PR until manually
reviewed.
cc @<maintainers> — multi-indicator CI risk; please confirm.
This is the case that most strongly forces CI safety: No-Go AND
the aggregator's Recommend: Close heuristic (review-summary
agent §5e trigger 2).
The recipes match the common shapes. An author who wants to slip past them has many low-effort options:
importlib, __import__,
getattr chain).utils.sanity_check() that resolves a hostname).ctypes to call socket(2) directly,
reading /proc/self/environ instead of os.environ)..cfg, .json consumed by a test that evals it).The agent therefore reads the PR-added lines in test and CI paths for intent, not pattern. If a function added in a test path constructs and executes a command, opens a network primitive, or reads from a path outside the working tree — regardless of how it phrases the call — it is in scope. When in doubt the agent posts at the appropriate severity with a maintainer ping per the low-confidence rubric in the review contract.
Scan PR-added lines in test / CI / setup paths for runtime-policy violations (outbound connections, subprocess spawning, sensitive-env reads, writes outside the working tree, cache or toolchain tampering) and combined indicators. Greps are starter signals; the policy is the contract. Tag at severity; pile-up forces must-fix + maintainer ping.