The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests model implementation code from an external "upstream" repository which is subsequently executed within the local environment.
Ingestion points: Code is cloned from the vLLM upstream repository (Step 2 in references/procedure.md) into a temporary folder.
Capability inventory: The agent has access to Bash, python3, and ssh, enabling it to run the migrated code during verification steps.
Boundary markers: No specific delimiters or warnings are used when the agent processes the external code.
Sanitization: While scripts/validate_migration.py performs static analysis for imports and API compatibility, it does not sanitize the logical content of the migrated model files.
[COMMAND_EXECUTION]: The skill performs automated management of remote servers using SSH and Docker commands.
Evidence: The scripts/e2e_remote_serve.sh script executes commands on a remote host using SSH with the -o StrictHostKeyChecking=no flag, which bypasses host identity verification.
[EXTERNAL_DOWNLOADS]: The skill downloads source code from an external repository to facilitate the migration process.
Evidence: references/procedure.md Step 2 instructs the agent to clone the vllm-project/vllm repository from GitHub.
[REMOTE_CODE_EXECUTION]: The skill facilitates the execution of code downloaded from external sources.
Evidence: scripts/serve.sh uses the --trust-remote-code flag when launching the vLLM server, allowing the execution of arbitrary code from HuggingFace model checkpoints.
Evidence: Migrated code from the upstream repository is directly executed during the import test in references/procedure.md Step 4.2.C.

model-migrate-flagos