The Agent Skills Directory

[COMMAND_EXECUTION]: The skill frequently utilizes the Bash tool across multiple roles. The coordinator uses it for session directory setup, while the profiler and benchmarker roles use it to execute project-specific diagnostic and benchmarking tools (e.g., cargo bench, go test -bench, or manual timed executions). The optimizer role additionally employs a CLI tool via Bash to implement code changes.
[PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection because it ingests and processes untrusted data from the local project environment.
Ingestion points: Roles such as the profiler and optimizer read source code, configuration files (e.g., package.json, Cargo.toml), and performance reports as specified in role-specs/profiler.md and role-specs/optimizer.md.
Boundary markers: The roles/coordinator/commands/dispatch.md file defines structured Markdown templates that use headers (PURPOSE, TASK, CONTEXT) and --- delimiters to separate instructions from external data.
Capability inventory: The agents possess capabilities to execute shell commands (Bash), modify the filesystem (Write, Edit), and spawn further sub-agents (Agent), which could be exploited if an injection occurs.
Sanitization: No explicit logic is provided to sanitize or validate the content of files read from the project before they are included in the prompts for the worker agents.

team-perf-opt