The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill coordinates workflows with external models and CLIs, creating an indirect prompt injection surface where untrusted downstream outputs could potentially contain instructions aimed at manipulating the primary controller.
Ingestion points: External output text and execution metadata are ingested into the workspace via recursive-router-invoke.py and recursive-router-invoke.ps1 scripts.
Boundary markers: The documentation states that routed output is untrusted and requires verification, but it does not specify explicit structural boundary markers or escaping logic for the dispatch payloads.
Capability inventory: The skill invokes local sub-processes via Python and PowerShell to manage configuration, model discovery, and execution paths, and dispatches commands directly to external binaries (codex, kimi, opencode-cli.exe).
Sanitization: Relying on the primary orchestrator to perform verification against repository state, the skill does not detail programmatic sanitization or filtering of text returned by the external tools.
[COMMAND_EXECUTION]: The skill requires executing local automation scripts and shell commands (python and pwsh) to probe environmental configurations and launch external CLI utilities.

recursive-router