The Agent Skills Directory

[COMMAND_EXECUTION]: The skill installs and configures session-start hooks in .claude/settings.json and .codex/hooks.json to execute a project-local Node.js script that provides startup context.
[DATA_EXFILTRATION]: The skill manages session state and progress notes on the local filesystem. It implements safety measures by automatically updating the project's .gitignore to prevent these files from being committed or shared.
[PROMPT_INJECTION]: The skill processes content from local handover files which presents an indirect prompt injection surface. This is mitigated through explicit agent instructions and script-level sanitization.
Ingestion points: Handover data is read from handover.md and .handover/current.md during the 'boot' mode and by the handover-session-start.js hook script.
Boundary markers: SKILL.md instructs the agent to treat handover content as unverified and to summarize it while identifying stale or conflicting information.
Capability inventory: The skill is capable of reading and writing local project files (handover.md, AGENTS.md, CLAUDE.md) and modifying local configuration files for agent tools.
Sanitization: The handover-session-start.js script performs targeted extraction of markdown sections ('Goal', 'Next Action', 'Stop Conditions') and applies length limits to the emitted context.

handover