The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Bash to execute standard file system operations. Evidence includes calls to git rev-parse --show-toplevel to identify the project root, mkdir -p for session directory creation, and cp for consolidating task artifacts. These operations are limited to local environment management and are consistent with the tool's stated purpose.
[DATA_EXFILTRATION]: The workflow involves searching and reading codebase files (e.g., README, documentation, and existing code) using tools like Grep, Glob, and Read to gather context for planning. All gathered information is processed locally and stored in the project's .workflow/.planning/ directory. No external network calls or data transmission patterns were found.
[INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection as it processes untrusted data from the codebase and user input to generate plans.
Ingestion points: Uses mcp__ace-tool__search_context, Grep, and Read to ingest file content during the 'Understanding' and 'Planning' phases.
Boundary markers: Uses markdown headers (e.g., ## 任务池) to separate domain sections, though explicit 'ignore' instructions for embedded data are absent.
Capability inventory: Includes capabilities for writing files, executing shell commands for directory management, and calling subsequent workflows.
Sanitization: No explicit sanitization or filtering of codebase content is mentioned before it is interpolated into the planning templates.

collaborative-plan-with-file