The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted source code diffs, which creates an ingestion surface for potentially malicious instructions hidden within code logic or comments designed to manipulate the review tool's output.\n
Ingestion points: The file diff and path are ingested in SKILL.md and passed as a state object to the omh_jev_ask tool.\n
Boundary markers: The skill does not define specific delimiters or clear 'ignore instructions' markers to isolate the untrusted diff content within the tool call.\n
Capability inventory: The skill is limited to providing advisory flags and does not have the authority to perform automated actions, merges, or system-level changes based on tool outputs.\n
Sanitization: The workflow contract includes a safety rule to remove 'credential-like content' from the diff before it is transmitted to the external model.\n- [DATA_EXFILTRATION]: The skill is designed to transmit source code diffs to an external model for automated review, which is its primary intended purpose.\n
The skill implements significant safeguards for this process: it requires a mandatory warning to the user that source code 'leaves the machine' before the first ask, limits transmissions to at most 8 files per review, and mandates the removal of secrets prior to transmission to protect user credentials.

omh-jev-review-gate