The Agent Skills Directory

[COMMAND_EXECUTION]: Multiple scripts within the skill, including date_provenance.py, project_binding.py, sync_project.py, session_discovery.py, and lint_wiki.py, invoke the git command-line utility via subprocess.run. This is used to resolve project metadata, verify file tracking, and analyze repository history for the purpose of building a temporal axis.
[COMMAND_EXECUTION]: The ingest_docs.py script executes a sibling Python tool, ticket-autopilot.py, using sys.executable. This execution is localized to the agent's environment and is required to parse structured ticket data from the project's documentation files.
[INDIRECT_PROMPT_INJECTION]: The skill facilitates the ingestion of external source materials (in raw/sources/), repository documentation, and agent transcripts from .claude or .codex directories. While this data is synthesized by the agent to create wiki pages, the process presents an attack surface for indirect prompt injection if those source files contain malicious instructions. The skill addresses this risk through strict layout enforcement and structural linting, though it ultimately relies on the platform's safety guardrails during content synthesis.
[DATA_EXPOSURE]: session_discovery.py and session_ingest.py access local agent history stored in ~/.claude/projects/ and ~/.codex/sessions/. This access is a core feature enabling the skill to 'discover' and digest past agent sessions into the wiki. The skill minimizes exposure by generating pointers and summaries (digests) rather than copying full transcript content into the wiki tree.

llm-wiki