The Agent Skills Directory

[PERSISTENCE]: The skill registers multiple lifecycle hooks in agent configuration files (e.g., settings.json for Claude Code, hooks.json for Cursor) to manage project state, session capture, and auto-actualization.
Modifies ~/.claude/settings.json, ~/.cursor/hooks.json, and other agent-specific configuration paths.
This persistence is essential for the skill's documented purpose of maintaining long-term memory across sessions.
[COMMAND_EXECUTION]: The tool performs extensive local shell command execution for developer workflows.
Scripts like mb-test-run.sh and mb-rules-check.sh invoke local testing and linting tools.
The mb-metrics.sh script detects project stacks and executes language-specific metric collectors.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted repository content while having access to powerful tools, creating a potential attack surface.
Ingestion points: Analyzes source code and markdown files via mb-context.sh and mb-codegraph.py (file paths specified in SKILL.md).
Boundary markers: Employs specific markdown markers (e.g., ) to identify its own data blocks.
Capability inventory: Possesses full Bash execution, Write/Edit file access, and network research capabilities through subagents.
Sanitization: Implements sc_redact_secrets and sc_strip_private to prevent leakage of credentials or marked private data into session logs or LLM summaries.
[EXTERNAL_DOWNLOADS]: Fetches metadata from established registries to support maintenance and research functions.
mb-version-check.sh queries GitHub and PyPI APIs for update information.
The mb-research subagent dispatches queries to external documentation and web sources as part of its research role.
[DYNAMIC_EXECUTION]: Implements dynamic module loading for language grammars.
memory_bank_skill/codegraph_treesitter.py uses __import__ to load language-specific tree-sitter modules dynamically based on the detected project stack.

memory-bank