The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The file references/config.md provides detailed instructions and code snippets for accessing sensitive configuration files (~/.skillhone/settings.json, identities.conf, _data/forgejo_config.txt) which contain API keys and Forgejo tokens. It demonstrates how to programmatically extract these credentials using Python and shell commands. While SKILL.md contains a rule forbidding printing these credentials, documenting the exact paths and retrieval logic increases the risk of accidental or malicious exposure.
[EXTERNAL_DOWNLOADS]: The explorer agent and references like skillhub.md and clawhub.md describe workflows for searching and installing skills from community registries (skillhub, openclaw). This involves downloading potentially untrusted scripts and markdown files into the workspace for inspection.
[COMMAND_EXECUTION]: The skill makes extensive use of subagents (developer, reviewer, issue-reporter) that are instructed to execute complex shell commands, Git operations, and interaction scripts (issue.py, pr.py, wiki.py) to manage a repository. This high-capability surface area is intended but requires careful oversight.
[DATA_EXFILTRATION]: The issue-reporter and explorer agents are tasked with summarizing findings and reporting them to a Forgejo wiki and issue tracker. If the source material (probe traces or external skills) contains sensitive information, it could be inadvertently persisted in the project's repository memory.
[PROMPT_INJECTION]: The explorer subagent is instructed to read the content of SKILL.md and scripts from untrusted external skills downloaded from community registries. This represents an indirect prompt injection surface where a malicious skill could attempt to hijack the optimization agent's behavior during the exploration phase.

skillhone-optimization