Gandalf the Greybeard (PE)
5f3f0ab94d
Contamination class: a stale GH_CONFIG_DIR inherited from a prior session or a different agent's workspace caused generate-token.sh to write .gh-token into a foreign workspace, silently granting that agent's gh config access to the wrong token. Three hardening changes: 1. agent-setup/scripts/setup.sh — before deriving GH_CONFIG_DIR from AGENT_HOME, warn and unset any inherited value that points outside AGENT_HOME. This prevents the contaminated value from leaking into the derived path or the dotfile. 2. agent-setup/SKILL.md — correct the sourcing example from `source ~/.env` to `source "$AGENT_HOME/.env"` so the dotfile is sourced from the documented location (setup.sh writes to $AGENT_HOME/.env, not ~/ which may differ). 3. github-app-token/scripts/generate-token.sh — (a) add a hard die() guard that refuses to write the token when GH_CONFIG_DIR is outside AGENT_HOME; (b) pin GH_CONFIG_DIR="$GH_TOKEN_DIR" on the gh auth login invocation so it cannot fall back to any inherited config dir. Verified: - bash -n passes on both modified scripts - With GH_CONFIG_DIR=/tmp/someone-elses/.github AGENT_HOME=/tmp/me, setup.sh warns + overrides; generate-token.sh dies before writing. - With GH_CONFIG_DIR unset and a valid AGENT_HOME, behaviour is unchanged (token lands in $AGENT_HOME/.github). Co-Authored-By: Paperclip <noreply@paperclip.ing>
Skills
A collection of Claude Code skills — reusable tools that extend Claude Code's capabilities. Each skill lives in its own top-level directory and ships a SKILL.md (the entry point Claude Code reads when invoking the skill) plus any supporting scripts or references.
Available skills
| Skill | What it does |
|---|---|
github-app-token |
Generate a short-lived GitHub App installation access token and authenticate the gh CLI with it. |
trebuchet |
Start AI-powered penetration test scans, check scan status, and retrieve security findings reports via the Trebuchet API. |
kubernetes-reflector |
Reference for Kubernetes Reflector annotations that mirror secrets and configmaps across namespaces. |
minimax-image-generation |
Generate images from MiniMax's image-01 model via the /v1/image_generation endpoint. |
Skill layout
<skill-name>/
├── SKILL.md # Required. YAML frontmatter (name, description) + usage docs.
├── CLAUDE.md # Optional. Maintainer / implementation notes.
└── scripts/ # Optional. Bash or other implementation scripts.
Scripts use set -euo pipefail and a shared die() pattern for error handling. Scripts are invoked via bash scripts/<name>.sh (not ./scripts/<name>.sh) so that they work even when the executable bit did not survive deployment.
No build / test / lint tooling
There is no centralized build, test, or lint system. Each skill is self-contained and pulls in only standard Unix tools as declared in its SKILL.md.
Contributing
- New skills get a new top-level directory with at minimum a
SKILL.mdthat starts with YAML frontmatter on line 1. - Keep
SKILL.mdfocused on decision flow + user-facing usage. Move implementation details, API references, and rarely-needed tables intoCLAUDE.mdor areferences/subdirectory to keep per-invocation token cost low. - Add a row to the table above.