The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to interact with local development toolchains including pnpm, cargo, and uv to sync dependencies and boot project-scoped services. These actions are confined to the project boundary.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from a project file (.agents/environment.md) to determine its diagnostic routine. This creates an attack surface where instructions inside the manifest could attempt to influence the agent. However, the skill explicitly defines safety boundaries to mitigate this risk.
Ingestion points: .agents/environment.md (File read in SKILL.md Step 1 of Execution Protocol).
Boundary markers: The 'Operating Invariants' section explicitly instructs the agent to halt on unresolved port collisions and to forbid destructive actions.
Capability inventory: Executes toolchain syncs, service readiness probes (socket/HTTP), and test datastore migrations.
Sanitization: None explicitly mentioned for manifest content, but high-risk commands are explicitly blacklisted.
[SAFE]: The skill demonstrates security best practices by explicitly forbidding sudo usage, preventing access to production .env files, and mandating the use of synthetic mocks for secrets.

env-doctor