The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The agent is designed to maintain a persistent 'sanctum' by recording session logs and distilling them into long-term memory. This process involves processing potentially untrusted data from conversation history without explicit sanitization or boundary markers, which could allow adversarial input to influence future agent logic.
Ingestion points: Raw session logs stored in sessions/ and curated into MEMORY.md during autonomous Pulse sessions.
Boundary markers: None; the agent relies on its own distillation logic and character persona to manage memory without specific delimiters for user-provided data.
Capability inventory: The agent can execute shell commands via uv run and has write access to its sanctum directory to store memory and new scripts.
Sanitization: No automated sanitization of log content prior to curation is detailed in the templates.
[DYNAMIC_EXECUTION]: The 'Capability Authoring' guide (references/capability-authoring.md) instructs the agent to expand its functionality by writing Python or bash scripts to the capabilities/ directory. These agent-generated scripts are then executed to perform deterministic tasks, facilitating dynamic code generation and execution within the local environment.
[COMMAND_EXECUTION]: The skill requires the execution of initialization and runtime scripts (scripts/init-sanctum.py and scripts/wake.py) to manage the agent's identity and state transitions. These commands are executed locally using the uv environment during skill activation.

bmad-agent-sentinel