The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill provides commands in references/ci-regression.md that use sudo to modify system-level CPU and kernel settings, such as the scaling governor, Turbo Boost, and SMT (Hyper-Threading). While these are intended for stabilizing self-hosted CI runners and include warnings, they involve elevated privilege operations.
[EXTERNAL_DOWNLOADS]: The skill fetches the benchstat tool from Go's official repository. It also instructs the agent to install several other third-party utilities from external sources, including:
filippo.io/mostly-harmless/benchdiff@latest
github.com/knqyf263/cob@latest
go.bobheadxi.dev/gobenchdata@latest
github.com/felixge/fgprof (via github.com/felixge/fgprof)
[INDIRECT_PROMPT_INJECTION]: The skill identifies a surface for indirect prompt injection when processing benchmark data.
Ingestion points: Benchmark results are read from files such as old.txt, new.txt, and bench.json in SKILL.md, references/benchstat.md, and references/ci-regression.md.
Boundary markers: The instructions do not define clear boundary markers or instructions to ignore embedded content when the agent processes the output of benchstat or gobenchdata.
Capability inventory: The skill has access to Bash for executing arbitrary Go and system commands, WebFetch for network access, and broad file system permissions (Read, Edit, Write).
Sanitization: No sanitization or validation of the benchmark result file content is mentioned before the agent performs deep analysis on the performance indicators.
[COMMAND_EXECUTION]: The skill frequently uses the Bash tool to execute go test, benchstat, pprof, trace, and other diagnostic utilities. It also passes environment variables (e.g., GODEBUG, GOSSAFUNC) directly to the shell.

golang-benchmark