The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs local command execution for its installation and runtime lifecycle hooks.
The scripts/install.sh script uses standard utilities like cp, chmod, and jq to configure the environment and merge lifecycle hooks into the project settings.
The eval/run_eval.py script utilizes subprocess.run to execute the skill's internal installer and harvester scripts during automated testing.
Runtime hooks (stop.sh, precompact.sh, session_start.sh) execute the skill's Python scripts using fixed local paths.
[INDIRECT_PROMPT_INJECTION]: The skill processes agent transcript data, which constitutes a data ingestion surface.
Ingestion points: assets/harvest.py reads transcript history from local JSONL files provided by the platform.
Boundary markers: The skill employs a 'two-channel' trust model, harvesting data from trusted message roles (user and assistant) while explicitly excluding tool_result and tool_use blocks. It also requires specific markers to appear at the start of a line to prevent smuggling.
Capability inventory: The skill manages a local cache in the .context/ directory and is triggered by platform-defined hooks.
Sanitization: assets/context_ledger.py implements a _neutralize_fences function that escapes <data> tokens in content originating from untrusted channels, preventing potential fence breakout in the final context digest.
[PRIVILEGE_ESCALATION]: The installation script modifies file permissions to enable execution.
scripts/install.sh uses chmod +x on the skill's provided scripts and hooks.
This operation is standard for the installation of executable lifecycle hooks and is performed only on the skill's own assets.

context-hygiene-kit