The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill operates on untrusted external data during its research phases, creating an inherent surface for indirect prompt injection. This is mitigated through structured processing and verification steps.
Ingestion points: External data enters the context during the 'Dispatch + Investigate' phase (P2) via web retrieval and search result processing.
Boundary markers: The skill uses structured evidence packets (task-n.md), a citation registry (P3), and claim-evidence tables to isolate external content.
Capability inventory: The agent possesses file-writing, web-retrieval, and coordination capabilities over sub-agents.
Sanitization: All load-bearing claims must undergo a mandatory 'Counter-Review' (P6) and 'Final Verification' (P7), where findings are explicitly cross-checked against original source documents to filter out injected instructions or hallucinations.
[COMMAND_EXECUTION]: The test suite (tests/test_provider_runs.py and tests/test_research_assets.py) utilizes subprocess.run to execute local Python scripts during development testing. This usage is restricted to the validation of CLI functionality within a controlled environment and does not pose a security risk.

deep-research