The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted financial signals and news content from external sources which are then interpolated into agent prompts for report generation and analysis.
Ingestion points: News signals are fetched via NewsNow in scripts/utils/news_tools.py and search results are gathered using Jina, DuckDuckGo, and Baidu in scripts/utils/search_tools.py.
Boundary markers: Prompts defined in references/PROMPTS.md and scripts/prompts/report_agent.py interpolate untrusted text directly into context blocks (e.g., {signals_text} or {signal_cluster_text}) without explicit instructions for the model to ignore embedded commands or markers to delineate the untrusted content.
Capability inventory: The agent has access to multiple tools in scripts/tools/toolkits.py, including network search, sentiment analysis, and stock market data retrieval.
Sanitization: No evidence of input sanitization specifically for prompt injection patterns was found; the skill primarily implements structural JSON validation and comment stripping in scripts/utils/json_utils.py.
[EXTERNAL_DOWNLOADS]: The skill makes legitimate requests to external financial data providers and news aggregators as part of its core functionality.
Evidence: scripts/utils/news_tools.py fetches news and prediction market data from newsnow.busiyi.world and gamma-api.polymarket.com.
Evidence: scripts/utils/search_tools.py and scripts/utils/content_extractor.py utilize the Jina Search and Reader APIs (s.jina.ai and r.jina.ai).
Evidence: scripts/utils/stock_tools.py uses the akshare library to fetch domestic and international stock market data.
Evidence: scripts/utils/predictor/training.py and scripts/utils/sentiment_tools.py download pre-trained models and weights from HuggingFace and other official package repositories.
[DYNAMIC_EXECUTION]: The skill utilizes standard methods for loading model weights and parsing data which involve dynamic loading.
Evidence: scripts/utils/predictor/training.py and scripts/utils/predictor/evaluation.py use torch.load to load local pre-trained model weights.
Evidence: scripts/utils/json_utils.py employs ast.literal_eval as a fallback for parsing JSON-like structures, which is a safer alternative to the built-in eval() for literal data structures.

alphaear-reporter