The Agent Skills Directory

[Local Credential Management]: The skill requires access to Buildkite and GitHub tokens stored in the user's home directory (~/.buildkite_token and ~/.github_vllm_token). It demonstrates good security posture by explicitly instructing the use of umask 077 and chmod 600 to ensure these files are only readable by the owner. These credentials are used solely for their intended purposes with well-known service APIs.
[Trusted API Interactions]: The workflow involves automated communication with api.buildkite.com and api.github.com. These operations are standard for the skill's purpose of triaging builds and filing repository issues. All network traffic is directed toward well-known, legitimate service providers.
[Indirect Prompt Injection Surface]: The skill processes Buildkite logs, which are external data sources (Category 8).
Ingestion points: Buildkite logs are fetched and cleaned in Steps 4 and 5.
Boundary markers: The skill relies on human oversight as a primary boundary; Step 7 mandates that the agent must draft the issue body and wait for explicit user approval before posting.
Capability inventory: The skill has the ability to create and update GitHub issues via authenticated curl commands.
Sanitization: The skill employs regex-based cleaning to strip ANSI escape codes and platform-specific timestamp markers from the logs before processing.
[Human-in-the-Loop Validation]: A key security feature is the explicit requirement in Step 7 to draft and confirm all public posts. This prevents the agent from autonomously filing issues or modifying repository state without direct user supervision, effectively mitigating risks associated with processing untrusted log data.

vllm-pytorch-ci-triage