The Agent Skills Directory

Indirect Prompt Injection: The skill is designed to ingest and summarize conversation history from local database files, which constitutes processing untrusted data.
Ingestion points: The scripts/inspect_sessions.py script reads from checkpoints and writes tables in a local SQLite database.
Boundary markers: The script produces structured JSON output, and the agent is instructed to synthesize this data rather than repeating it verbatim, which helps maintain clear boundaries between system instructions and processed data.
Capability inventory: The script performs read-only database operations and outputs data to stdout; it does not include file-write or network capabilities.
Sanitization: Content is automatically truncated to a configurable character limit (default 4,000) to mitigate issues with oversized inputs.
Dynamic Execution: The inspection script uses dynamic process management to ensure it runs in a compatible Python environment.
Pattern: The script utilizes os.execve() to replace the current process with a different Python interpreter if the current environment lacks necessary dependencies.
Context: This behavior is used for environment discovery, specifically looking for the dcode or deepagents-code runtime on the system to ensure it can correctly deserialize message blobs using the LangGraph library.
Command Execution: The script executes shell commands to verify local environment configuration.
Pattern: subprocess.run() is used to probe discovered Python binaries by attempting to import required modules.
Context: These commands are used to validate that a candidate interpreter found on the system PATH is suitable for running the inspection logic.

deepagents-thread-inspector