The Agent Skills Directory

[SAFE]: The skill implements a robust 'probe ledger' system that baselines the repository state using git stash and file snapshots before any probes are conducted. It mandates a reverse-patch cleanup process to ensure all temporary modifications are reverted without overwriting the user's original uncommitted work.
[SAFE]: Explicit operational boundaries prevent the skill from applying fixes, strictly limiting its output to diagnosis and failing reproductions. This 'no-fix' policy ensures human review is mandatory before any permanent changes are made to the codebase.
[SAFE]: Tool usage for network operations (WebSearch, WebFetch) is explicitly scoped to decoding observed signals like error messages or library documentation, with no patterns suggesting data exfiltration or unauthorized credential access.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a potential attack surface for indirect prompt injection.
Ingestion points: The skill ingests untrusted data via WebFetch (for error decoding) and Read (for local file analysis in SKILL.md).
Boundary markers: The instructions do not specify the use of delimiters or 'ignore' instructions when processing external web content or file data.
Capability inventory: The skill is granted access to high-privilege tools including Bash, Edit, and Write (specified in the YAML frontmatter).
Sanitization: The skill mitigates the risk of malicious instructions in data through its mandatory 'ritual' which requires a 'Proven cause' to pass a three-step on/off toggle test. This empirical requirement serves as a logical filter against data-driven prompt injections that might attempt to steer the agent towards incorrect or malicious actions.

debugkit