The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied variables such as <CANN_PATH>, <ATB_REPO_PATH>, and <DOCKER_NAME> which are directly interpolated into shell commands (e.g., docker exec, source, cd).
Ingestion points: Path and container name variables provided by the user in response to skill prompts.
Boundary markers: None present; the skill assumes valid input.
Capability inventory: Full shell command execution via bash and docker exec in the SKILL.md instructions.
Sanitization: No explicit sanitization or validation of the user-provided paths is performed before execution.
[EXTERNAL_DOWNLOADS]: The skill facilitates the download of third-party dependencies (nlohmann/json, googletest, pybind11, doxygen) and vendor-specific repositories from GitHub and GitCode mirrors.
Evidence: Instructions use git clone and sed to point to https://gitcode.com and https://github.com for library dependencies.
[COMMAND_EXECUTION]: The skill executes local build scripts and setup commands within a Docker environment.
Evidence: Commands like bash scripts/build.sh testframework and source <CANN_PATH>/set_env.sh are core to the skill's functionality.
[DYNAMIC_CONTEXT_INJECTION]: The skill uses PreToolUse and PostToolUse hooks to execute shell commands for environment validation and limiting compilation attempts.
Evidence: The hooks section in SKILL.md contains shell scripts that run to verify the existence of set_env.sh and the repository directory.

atb-testframework-build