The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill captures tool inputs and outputs (via environment variables in hooks) to generate learning episodes and candidates. This ingested data is stored in local state files and used to trigger the extraction of candidate lessons based on failure modes. While lessons currently use hardcoded templates, the ingestion of untrusted data from tool outputs creates a surface for potential bias or logic manipulation if the classification logic (e.g., detecting 'search_skipped') is deceived.
Ingestion points: TOOL_INPUT and TOOL_OUTPUT environment variables are captured by hooks/pre-tool.sh and hooks/post-bash.sh and processed by runtime/learning_core.py.
Boundary markers: No explicit delimiters or 'ignore' instructions are used when storing or summarizing tool data in the observation records.
Capability inventory: The skill uses subprocess.run (in publish_event) to execute external commands and has file write capabilities to maintain local state.
Sanitization: The redact_text function uses regular expressions to remove secrets (e.g., API keys, bearer tokens, AWS credentials), and summarize_blob limits the length and structure of tool data previews.
[COMMAND_EXECUTION]: The skill executes external commands to publish events to a telemetry system.
Evidence: runtime/learning_core.py contains the publish_event function which calls subprocess.run using a command list built by build_publish_command. The command executes uv run bb publish to send JSON data to an external service.
[EXTERNAL_DOWNLOADS]: The skill relies on an external system called 'Bloodbank' and its associated CLI tool bb, which is expected to be present on the host system.
Evidence: The skill documentation and the bloodbank_root() function in runtime/learning_core.py reference a local directory ~/code/33GOD/bloodbank as the location for this external dependency.

self-improving-agent