The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill instructions reference the sensitive local file path ~/.openhands/agent-canvas/api-key.txt used by the vendor's tools to persist session API keys.
[COMMAND_EXECUTION]: The skill provides numerous shell command examples using curl and a provided Python client to interact with remote API endpoints and manage agent behavior.
[EXTERNAL_DOWNLOADS]: The skill fetches data and downloads conversation trajectories from the official OpenHands Cloud API at https://app.all-hands.dev.
[DYNAMIC_EXECUTION]: Shell examples in SKILL.md use Python here-docs to dynamically generate and execute scripts for processing and filtering API response data.
[INDIRECT_PROMPT_INJECTION]: The skill ingests external data from the OpenHands API and extracts contents from downloaded zip files, presenting an attack surface where malicious data could influence agent behavior.
Ingestion points: Conversation event search results and trajectory zip files processed in scripts/openhands_api.py.
Boundary markers: No explicit delimiters or instructions are used to separate untrusted event data from agent instructions during processing.
Capability inventory: The skill includes capabilities to execute bash commands on remote agent servers and perform local file system writes and extraction.
Sanitization: External event data and extracted file contents are processed without specific sanitization against instruction injection patterns.

openhands-api