The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes its internal logic and classification scripts using the Python subprocess.run function.
These executions use shell=False, which prevents traditional shell command injection by passing arguments as a list.
Script paths are resolved relative to the skill directory, and the runtime validates that these paths remain within authorized boundaries to prevent the execution of arbitrary scripts.
[PROMPT_INJECTION]: The skill accepts user-provided inputs for fields like intent and rationale, which are then included in the generated Markdown handoff.
Ingestion points: These inputs are collected via CLI flags in scripts/route_work.py and defined in the manifest skill-protocol.json.
Boundary markers: The skill does not use specific delimiters but treats these fields as descriptive text to be echoed verbatim.
Capability inventory: The skill possesses the capability to write a report file to disk and execute validation subprocesses.
Sanitization: The skill applies NFKC Unicode normalization to these inputs in scripts/route_work.py to ensure a stable presentation, though it intentionally avoids interpreting the content to maintain an authority boundary between the agent and the validator.

route-work