The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes scripts/control_kernel_init.py, a bootstrap script that initializes the control metalayer by writing configuration files and schemas to a target repository. It also provides scripts/conversation-bridge-hook.sh, designed to run as a git hook to automate the generation of conversation history.
[DATA_EXPOSURE]: The script scripts/conversation-history.py accesses agent transcripts stored in the user's home directory (~/.claude/projects/) to generate project documentation in docs/conversations/. This creates episodic memory by exposing previous session contents within the repository documentation.
[INDIRECT_PROMPT_INJECTION]: The 'Consciousness Stack' feature ingests historical conversation logs and plant observations into the agent's context. This represents an indirect prompt injection surface where untrusted data from previous sessions or external plant states could influence agent behavior.
Ingestion points: Previous session logs are parsed and written to docs/conversations/, which the agent is instructed to read at session start.
Boundary markers: The generated documentation uses Obsidian callouts (e.g., > [!quote]) to delimit user messages, though it lacks explicit 'ignore embedded instructions' warnings.
Capability inventory: The framework configures the Arcan runtime with broad capabilities, including recursive filesystem access and arbitrary command execution (exec:*).
Sanitization: The _callout_safe function in scripts/conversation-history.py provides basic sanitization by stripping XML/HTML tags and converting markdown headers to bold text to prevent callout breakage.

agentic-control-kernel