The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied Rust code and development queries as primary inputs. It mitigates potential injection risks by enforcing a rigid output structure consisting of specific versioned contracts (e.g., rust_change_contract/v1) and mandatory safety checklists.
Ingestion points: User prompts containing Rust code snippets or development requirements defined in SKILL.md ('Use when' section).
Boundary markers: The skill includes specific 'Safety rules' and a 'Completion Checklist' that force the agent to categorize and validate inputs before generating plans.
Capability inventory: The skill is restricted to generating text-based planning artifacts. It explicitly forbids the agent from running toolchains, compilers, or sanitizers ('Do not run cargo, Miri, sanitizers, or any toolchain from OMH core').
Sanitization: Output is constrained to predefined schemas, and any code involving unsafe or FFI primitives is automatically escalated to a higher-rigor safety checklist.
[COMMAND_EXECUTION]: While the skill mentions various shell commands (e.g., cargo build, cargo clippy, miri), it identifies these as commands to be performed by an external executor rather than the agent itself. The skill explicitly states 'OMH runs no toolchain' in both references/rust-discipline.md and references/ub-escalation.md, maintaining a safe execution boundary.
[SAFE]: The skill uses clear instructional language and provides 'Recovery Notes' to handle edge cases conservatively, such as escalating safety reviews when a crate cannot be fully inspected.

omh-rust