The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies extensively on executing shell commands within a Docker container using docker exec. These commands are used for environment verification, tool installation, and running the primary test sequences.
Evidence: Multiple instances in Phase 0 and Phase 3, such as docker exec $CONTAINER bash -c '/workspace/.devcontainer/ensure-skillshare-linux-binary.sh && ss version'.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it reads and executes commands defined in Markdown files within the /workspace/ai_docs/tests/ directory.
Ingestion points: ai_docs/tests/<runbook_file>.md (referenced in Phase 3).
Boundary markers: Absent. The skill does not implement delimiters or safety instructions to distinguish between intended commands and potentially malicious content within the Markdown files.
Capability inventory: The skill has the capability to run arbitrary shell commands via docker exec, access the network via ss install, and manipulate the file system within the container (Phase 1, 3).
Sanitization: Absent. The mdproof tool and subsequent bash -c calls execute the content of the Markdown steps directly.
[REMOTE_CODE_EXECUTION]: The skill includes procedures to download and execute code from remote sources during the environment preparation and testing phases.
Evidence: docker exec $CONTAINER /workspace/.devcontainer/ensure-mdproof.sh downloads the mdproof tool from a GitHub release (Phase 0).
Evidence: The ss install <src> command is used to download and install skill packages from remote Git repositories (Command Policy).
[EXTERNAL_DOWNLOADS]: The skill references external URLs for obtaining tools and metadata.
Evidence: Fetches mdproof from GitHub releases.
Evidence: Metadata source field points to https://github.com/runkids/skillshare/tree/main/.skillshare/skills/cli-e2e-test.
[METADATA_POISONING]: There is a discrepancy between the declared author of the skill and the metadata provided within the skill file, which could be used to obscure the true source of the instructions.
Evidence: The skill file metadata lists the author as "Runkids," whereas the skill originates from a different context.
[DYNAMIC_EXECUTION]: The skill uses mdproof to dynamically interpret Markdown runbooks into executable bash commands at runtime. It also performs runtime compilation of Go source code within the container.
Evidence: mdproof --report json /workspace/ai_docs/tests/ (Phase 1).
Evidence: go build -o bin/skillshare ./cmd/skillshare (Test Command Policy).

skillshare-cli-e2e-test