The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze traces from external LLM applications, which could potentially contain adversarial instructions or malicious content.
Ingestion Point: Phase 1 fetches traces into error-analysis/traces.jsonl from the Langfuse API or local JSONL files.
Boundary Markers: The methodology strictly enforces a human-in-the-loop "Open Coding" phase where every failure note must be reviewed and confirmed by a human user via AskUserQuestion before being clustered into a taxonomy.
Capability Inventory: The agent uses Bash for network requests and can invoke subagents for evaluation execution.
Sanitization: The process relies on manual human verification to gate the interpretation of all content before it becomes part of the persistent taxonomy or eval recommendations.
[COMMAND_EXECUTION]: The skill provides Bash snippets for interacting with the Langfuse API using curl and jq for trace normalization.
The implementation includes sophisticated defensive logic (in SKILL.md and references/langfuse-traces.md) that restricts the use of insecure HTTP to local endpoints only.
It proactively checks for and refuses URLs containing userinfo components to prevent common credential-stealing patterns.
The credential usage is restricted to standard environment variable references (e.g., -u "$LANGFUSE_PUBLIC_KEY:$LANGFUSE_SECRET_KEY"), preventing secret values from being logged or echoed during execution.

error-analysis