The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local CLI tools like git, gh, and npm to capture repository state and verify project figures during session transitions. These operations are conducted using standard, well-known binaries already present in developer environments.\n- [PERSISTENCE]: The context-pressure-install.sh script modifies the agent's configuration file (e.g., ~/.claude/settings.json) to register a UserPromptSubmit hook and wrap the statusLine command. These changes enable the skill to persistently monitor context window usage and alert the agent when a handoff is recommended due to context pressure.\n- [DYNAMIC_EXECUTION]: The handoff-verify.mjs script parses and executes commands extracted from handoff files. This behavior is secured by a comprehensive, per-binary allow-list (WHITELIST) that restricts execution to a safe, read-only subset of tools and flags, effectively mitigating the risk of arbitrary command execution.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests data from markdown-based handoff files which may contain embedded instructions. It mitigates this attack surface by implementing a 'Read' mode that re-measures and confirms all project figures against the live environment through restricted tools, ensuring the agent does not act on manipulated or stale data.

session-handoff