The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes shell scripts and local development tools to perform its functions. It explicitly calls bash .agents/skills/perf-optimization/scripts/go-pprof.sh and various language-specific benchmark commands like go test -race. While these are expected for a performance tool, they represent a significant capability surface.
[INDIRECT_PROMPT_INJECTION]: The skill accepts user-provided profiling data and URLs as input for its analysis phase. This untrusted data is then processed and passed to subagents without explicit boundary markers or sanitization, potentially allowing malicious content within a profile to influence the analysis or subsequent code generation.
Ingestion points: User-provided profile files or URLs in Phase 1.
Boundary markers: None identified in the instructions for wrapping or delimiting profiling data.
Capability inventory: Execution of shell scripts, writing and running benchmarks, modifying source code, and creating git commits.
Sanitization: No evidence of input validation or sanitization for the profiling data or URLs before they are processed by analysis scripts.
[DYNAMIC_EXECUTION]: In Phase 4, the skill is designed to write new code (fixes) and benchmark tests, which are then executed. This dynamic creation and execution of code is based on findings derived from the analysis of potentially untrusted profiling data.
[PROMPT_INJECTION]: The skill includes an 'Entry-point guard' directive that attempts to control the agent's behavior based on its role (aborting if it is a subagent). This is a form of instruction that overrides default agent autonomy or platform-level dispatching logic.

perf-optimize