The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest and process data from external DingTalk sharing links (alidocs.dingtalk.com/i/p/...) using the read_url tool in references/url-patterns.md. This represents a surface where malicious instructions embedded in a document's content or metadata could influence agent behavior.
Ingestion points: references/url-patterns.md (via the read_url tool for sharing links) and references/recipes/conventions.md (via search/list command results).
Boundary markers: The skill implements extensive node type probing using dws drive info to validate content before processing, as described in references/url-patterns.md.
Capability inventory: The agent has the capability to read/write files, manage calendars, send chat messages, and export authentication packages using the dws CLI binary.
Sanitization: The skill includes detailed instructions to normalize URLs and resolve 'dlink' (shortcut) targets to their stable original IDs before performing content-level operations to prevent redirection or identity confusion.
[COMMAND_EXECUTION]: The skill frequently executes shell commands via the dws binary to perform business logic. It provides specific instructions for constructing parallel shell execution strings (e.g., using the & operator and wait) to handle multiple resource IDs efficiently in references/recipes/conventions.md.
[DATA_EXPOSURE]: The skill defines strict handling rules for sensitive credentials. It explicitly forbids the agent from outputting or recording tokens, app secrets, or webhook tokens in SKILL.md and references/runtime-contract.md. Additionally, it provides guidance for the dws auth export command in references/global-reference.md, cautioning that authentication packages contain sensitive material and must not have their contents printed.

dingtalk-shared