The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill captures and packages runtime execution data from LangChain/LangGraph, including tool inputs, model messages, and event streams. This creates a surface for indirect prompt injection if the resulting bundle is later analyzed by an AI agent.
Ingestion points: The skill ingests data via graph.astream_events in SKILL.md (Step 2) and captures input_str and output_snippet through a DebugCallbackHandler described in references/callback-propagation.md.
Boundary markers: The skill does not explicitly apply boundary delimiters or "ignore embedded instructions" warnings to the logged content within the generated events.jsonl or callbacks.txt files.
Capability inventory: The skill possesses the capability to write to the file system (tarfile, Path.write_text) and execute local commands for metadata collection.
Sanitization: Significant sanitization measures are implemented in SKILL.md (Step 5) and references/sanitization-checklist.md, utilizing regex patterns to redact various API keys (OpenAI, Anthropic, Google, LangSmith), bearer tokens, and PII before the final archive is created.
[COMMAND_EXECUTION]: The skill performs command execution to collect environment metadata during the manifest assembly process.
Evidence: In SKILL.md (Step 1) and references/env-manifest-template.md, the code uses subprocess.check_output([sys.executable, "-m", "pip", "show", name]) to retrieve package versions.
Context: The execution is restricted to a hardcoded list of RELEVANT Python packages (e.g., langchain-core, langgraph, langsmith) and is used solely for diagnostic reporting.

langchain-debug-bundle