The Agent Skills Directory

[COMMAND_EXECUTION]: Shell script templates in 'references/hook-templates.md', 'references/visibility-patterns.md', and 'references/lifecycle-reference.md' utilize 'eval echo "$FILE_PATH"' to process path variables extracted from tool execution payloads. This pattern allows for arbitrary command injection if the file path contains shell metacharacters. Evidence: 'ABSOLUTE_PATH=$(eval echo "$FILE_PATH")' found in multiple templates.
[PROMPT_INJECTION]: The skill provides logic for hooks that process untrusted data, creating an indirect prompt injection surface. 1. Ingestion points: Data enters via standard input using 'PAYLOAD=$(cat)' in 'SKILL.md' and 'references/hook-templates.md'. 2. Boundary markers: None are defined in the templates to isolate untrusted tool inputs. 3. Capability inventory: The skill demonstrates capabilities for shell execution (eval), file searching (find, grep), and tool manipulation (jq). 4. Sanitization: The templates do not perform validation or escaping on variables extracted from the tool payload before they are evaluated in shell contexts.
[EXTERNAL_DOWNLOADS]: The documentation references external requirements such as 'jq' and the 'Bun' runtime. It also provides links to well-known GitHub repositories under the 'anthropics' organization and the author's development repository.

hooks-development