The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from conversation history files (MEMORY.md and CONTEXT.md) to extract topics, entities, and decisions for scoring.
Ingestion points: Reads full conversation history from MEMORY.md and current context from CONTEXT.md in score.py.
Boundary markers: None. The script processes raw text using regular expressions and keyword counters.
Capability inventory: The script performs read operations on history files and write operations to its own state file (state.yaml). It does not execute subprocesses, network requests, or arbitrary code.
Sanitization: Extracted text fragments for 'blind spots' are truncated to 80-120 characters before being stored or displayed. Topic extraction uses keyword frequency counting, which naturally filters complex instruction sequences.
[REMOTE_CODE_EXECUTION]: No remote code execution patterns were found. The Python script uses standard libraries and does not include any mechanisms to download and execute external scripts or packages.
[DATA_EXFILTRATION]: No network operations or data exfiltration attempts were detected. The skill performs all processing locally and does not interact with external domains.
[COMMAND_EXECUTION]: The script uses standard file I/O to manage its state and read workspace logs. It does not spawn shell commands or utilize user input to build system calls.

context-assembly-scorer