The Agent Skills Directory

[PROMPT_INJECTION]: The skill ingests untrusted data from external TAPD work items (descriptions and comments), which creates a potential surface for indirect prompt injection.
Ingestion points: Found in references/intake-gate.md where descriptions and comments are retrieved via MCP.
Boundary markers: The skill implements specific delimiters (【FLOW-SPEC-BEGIN:<spec_id>】) to distinguish managed specification blocks.
Capability inventory: The skill utilizes various MCP tools for creating and updating requirements, tasks, and comments.
Sanitization: This risk is effectively mitigated by the skill's mandatory safety policy (references/safety-policy.md), which requires dry-runs and explicit human confirmation for all remote modifications.
[COMMAND_EXECUTION]: Local branch binding and environment setup involve executing shell commands.
The bundled Node.js CLI tool (scripts/tapd-context/dist/cli.js) executes standard Git commands to manage local branch state and bindings.
A diagnostic script scripts/mcp_probe.mjs spawns child processes to verify connectivity and tool availability of the TAPD MCP server.
Both utilities are well-documented, operate on local state, and include mechanisms for redacting sensitive environment variables from their output.
[EXTERNAL_DOWNLOADS]: The skill's setup instructions involve downloading and running dependencies from well-known sources.
It utilizes the mcp-server-tapd package from the official Python package registry (PyPI) to handle communication with the Tencent TAPD API.
User information is retrieved from the official Tencent TAPD API domain (api.tapd.cn) via standard authenticated requests in scripts/get_current_user.py.

tapd