The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill implements a delegation mechanism in tools/_delegate.py that uses importlib.util and exec_module to dynamically load and run Python scripts from a sibling directory (../trtc/tools/). This allows for the execution of arbitrary code based on computed file system paths.
[EXTERNAL_DOWNLOADS]: During the project scaffolding phase described in references/02-path-a-scaffold-template.md, the skill uses npx degit to pull remote project templates and npm install to download numerous third-party dependencies from the NPM registry. While these sources (like Vite) are generally well-known, the skill performs these installations automatically at runtime.
[COMMAND_EXECUTION]: The skill frequently invokes shell commands via Bash and subprocess.run() (e.g., in tests/test_chat_bundle_contract.py and various flows) to perform file system operations, package management, and session state updates.
[CREDENTIALS_UNSAFE]: The skill's interactive information gathering (references/02-path-a-questions.md) explicitly requests the user's Tencent Cloud IM SecretKey. Although instructions exist to clear the key from temporary storage after writing it to a debug file, the agent is directed to process and store this raw secret initially.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user input and stores it in YAML session files (.docs-query.yaml, .trtc-session.yaml) as lastPrompt. It also pipes user input into shell commands using printf '%s' '{"text":"<用户原文>"}' | python3 tools/reporting.py prompt --input-stdin (as seen in references/13-reporting.md), creating a surface for potential prompt injection through processed data.

trtc-chat