The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill generates shell scripts at runtime to facilitate validation tasks. In references/validation-phase.md, the orchestrator writes logic to $RESEARCH_TMPDIR/cursor-validation-launch.sh and subsequently executes it via bash.
[PRIVILEGE_ESCALATION]: The skill explicitly modifies file permissions using chmod +x on the dynamically created cursor-validation-launch.sh script to enable its execution.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted data from multiple sources, creating a vulnerability surface where malicious content could influence agent behavior.
Ingestion points: Untrusted data enters the agent context from the local repository (via Read, Grep, Glob) and from the web (via WebSearch and WebFetch in the External Comparisons lane).
Boundary markers: The skill attempts to mitigate injection by wrapping untrusted data in namespaced XML tags (e.g., <reviewer_research_findings>, <lane_N_output_path>) and providing explicit instructions to treat the content as data rather than instructions.
Capability inventory: The skill environment is highly capable, allowing for Bash execution, spawning new subagents with the Agent tool, and performing network operations.
Sanitization: The validate-citations utility described in references/citation-validation-phase.md includes robust SSRF protections, such as private IP range rejection, DNS resolution checks, and connection pinning.
[COMMAND_EXECUTION]: The skill relies heavily on Bash to execute the larch.sh driver script for various tasks, including session setup, token telemetry, and background job management (bgjob).
[EXTERNAL_DOWNLOADS]: The skill uses WebSearch and WebFetch tools to retrieve information from the internet for research purposes and to validate cited URLs and DOIs.

research