The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data such as application logs and tool outputs, creating an attack surface for indirect prompt injection.
Ingestion points: Untrusted data enters the agent context through file reads, log ingestion, and tool output capture, as described in log-debugging-hygiene.md and reversible-compression.md.
Boundary markers: The findings-format.md module defines a structured template for agent findings using YAML frontmatter to separate control metadata from task content.
Capability inventory: The skill utilizes Bash, Read, and Grep tools, and coordinates complex workflows via subagents through the Task tool.
Sanitization: The skill provides extensive logic for filtering logs using rg, jq, and awk, and summarizing tool outputs via the Reversible Compression (CCR) pattern before they are stored in the conversation context.
[EXTERNAL_DOWNLOADS]: The log-debugging-hygiene.md module includes a command snippet that suggests using the uv tool to fetch the tiktoken package from the Python Package Index (PyPI) for token measurement. PyPI is a well-known service and tiktoken is a standard utility.
[DYNAMIC_EXECUTION]: Several modules reference helper scripts located in the plugin's scripts/ directory, such as coordination_workspace.py, context_retrieve.py, and agent_memory.py. These scripts are executed via python3 to manage coordination workspaces, triage memory tiers, and retrieve archived context handles.

context-optimization