The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes findings from external AI sidecars (e.g., Codex, Gemini), which can be manipulated to include malicious instructions.
Ingestion points: The skill reads output from sidecar processes stored in files like out.txt.
Boundary markers: It requires all accepted findings to be source-groundable (referencing specific file lines or quoting text) to prevent phantom findings.
Capability inventory: The skill has access to Read, Bash, Grep, and Glob tools for processing these outputs.
Sanitization: A residency closure scan (scripts/residency_closure_scan.py) is performed on all payloads before they are sent to external models to prevent data leakage.
[COMMAND_EXECUTION]: The skill uses shell commands and local scripts to orchestrate sidecar discovery and verification.
Evidence: Executes scripts/sidecar_wait.sh, scripts/target_freeze.sh, and scripts/residency_closure_scan.py.
Evidence: Invokes external command-line interfaces including codex, agy, gemini, and gh copilot.
[EXTERNAL_DOWNLOADS]: The skill utilizes well-known service tools that may perform background downloads during initialization.
Evidence: References gh copilot for verification, noting that the launcher may fetch the core CLI component upon its first invocation.

auto-decorrelation