The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/orchestrate_edge_pipeline.py script uses subprocess.run to call other Python scripts within the research pipeline. The script paths (e.g., auto_detect_candidates.py, build_hints.py) are hardcoded relative to the skill's root directory, ensuring that execution is limited to intended internal components.
[INDIRECT_PROMPT_INJECTION]: The skill serves as a surface for indirect prompt injection by ingesting and processing data that may contain LLM-generated content or external research data.
Ingestion points: The orchestrator reads data from user-supplied directories (--tickets-dir), external data files (--from-ohlcv), and LLM-provided hints (--llm-ideas-file). It also ingests revision_instructions from review artifacts generated in earlier pipeline stages.
Boundary markers: The orchestrator does not define explicit boundary markers for the files it processes, relying on the schema-based parsing in downstream scripts.
Capability inventory: The skill has the capability to execute subprocesses, read from the local filesystem, and write pipeline manifests and artifacts to an output directory.
Sanitization: Path arguments are resolved using Path.resolve() to prevent traversal. The feedback loop logic in apply_revisions uses strict keyword matching (e.g., searching for "add volume filter") to modify data structures, preventing the execution of arbitrary instructions embedded in the review text.

edge-pipeline-orchestrator