The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local shell scripts (scripts/compile.sh) that coordinate the compilation process. These scripts utilize system utilities like find, grep, md5sum, and python3 for logic and JSON processing. The script also provides support for the claude CLI binary when using the claude-cli runtime configuration.
[DATA_EXFILTRATION]: The core functionality of the skill involves reading local artifacts from the .agents/ directory and transmitting their contents to external LLM services (including Anthropic and OpenAI) or a user-defined Ollama host. This data flow is documented and intended for the purpose of synthesizing knowledge across files.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it aggregates raw, potentially untrusted knowledge artifacts (research, learnings) into a large prompt sent to an LLM. Malicious content within these artifacts could attempt to override the compiler's instructions.
Ingestion points: The script scripts/compile.sh reads markdown files from .agents/learnings, .agents/patterns, .agents/research, .agents/retros, .agents/forge, and .agents/knowledge.
Boundary markers: The compiler prompt uses --- FILE: <file_path> --- as a delimiter to separate different input artifacts.
Capability inventory: The skill performs network API requests and has the capability to write multiple files to the .agents/compiled/ directory.
Sanitization: No explicit sanitization or filtering of the artifact content is performed before interpolation into the LLM prompt.
[EXTERNAL_DOWNLOADS]: The skill makes network requests via curl to well-known LLM provider APIs (api.anthropic.com, api.openai.com) and Ollama instances. These endpoints are standard for the technologies utilized by the skill.

compile