The Agent Skills Directory

[OBFUSCATION]: The skill repository includes hidden executable code. The file tests/test_bounded_ct_ml_route.py reads and concatenates content from six Base64-encoded files (tests/fixtures/bounded-ct-ml-route/code-fixture.b64.part00 to part05) to reconstruct a ZIP archive containing research-related implementation code.
[DYNAMIC_EXECUTION]: The test suite performs dynamic loading of Python scripts. Specifically, tests/test_bounded_ct_ml_route.py extracts a Python file (_bounded_ct_ml_route_impl.py) from an archive and executes its contents using importlib.util.spec_from_file_location and spec.loader.exec_module.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data sources, creating a potential surface for indirect prompt injection attacks.
Ingestion points: Research literature, Zotero records, and novelty search results are ingested into the agent context in the literature-review/, zotero/, and novelty-review/ stages.
Boundary markers: The system utilizes 'frozen' objectives and verifier-backed transitions, as documented in SKILL.md and the references/ contracts, which serve as delimiters for external data.
Capability inventory: The framework includes scripts like scripts/harness_runtime.py that can write files and manage research task execution based on state replayed from event logs.
Sanitization: The skill implements schema validation (e.g., references/research-commitment.schema.json) for its internal configuration, although natural language content in research papers is not sanitized for instructions.
[COMMAND_EXECUTION]: Several integration test files, such as tests/test_harness_runtime.py and tests/test_research_pack.py, utilize the subprocess module to execute command-line instructions, specifically running the skill's own utility scripts to verify their behavior.

research-pipeline-planner