The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses a specialized Node.js resolver script (scripts/research_cli.mjs) to execute workspace management and research commands via child processes. It also includes a POSIX shell wrapper (scripts/agent-wrapper-posix.sh) to securely execute independent reviewer agents with minimal environment exposure.
[EXTERNAL_DOWNLOADS]: The skill utilizes npx to fetch and execute specific versions of the vendor's own CLI tools (@tiangong-ai/cli and skills). Additionally, it performs network operations through a dedicated broker to fetch research evidence from external APIs such as Brave Search and the vendor's SCI database, ensuring the AI agent does not have direct access to credentials.
[DYNAMIC_EXECUTION]: The resolver script (scripts/research_cli.mjs) dynamically loads the appropriate version of the CLI runtime based on a workspace's runtime-lock.json or setup-plan.json file. This allows for reproducible research environments while maintaining version pinning for security.
[INDIRECT_PROMPT_INJECTION]: As a tool designed to process untrusted data from the web and various document formats (PDF, XLSX, etc.), the skill possesses an inherent indirect prompt injection surface. It mitigates this via an extensive 'scientific gate' system described in its documentation (references/scientific-design.md, references/evidence-pipeline.md), which includes strict admission rules, evidence freezing, and independent reviews across different agent families.
Ingestion points: External data is ingested from web search APIs, academic databases, and user-supplied document artifacts during the discovery and acquisition stages.
Boundary markers: The system enforces 'inference gates' and 'content snapshots' to freeze evidence before analysis, ensuring that external content is categorized and verified before it reaches the reasoning stages.
Capability inventory: Subprocess calls and network fetching are strictly managed through the research CLI and a broker command.
Sanitization: All candidate leads and artifacts are deduplicated by content hash and metadata is sanitized before registration in the immutable ledger.

tiangong-auto-research