The Agent Skills Directory

[COMMAND_EXECUTION]: The framework sequences shell commands for repository management and external tool interaction as part of its primary design.
Evidence: templates/adws/adw_modules/git_helper.py performs core Git operations like branch management and committing via subprocess calls.
Evidence: templates/adws/adw_modules/agent_pi.py spawns the Pi coding agent subprocess to execute LLM-driven tasks.
Evidence: templates/adws/adw_modules/quality.py facilitates the execution of local development tools including linting and test runners.
Evidence: templates/adws/adw_modules/gates.py utilizes shell=True within its tests_pass factory for flexible test verification, though this is restricted to pre-defined commands in the workflow scripts.
[PROMPT_INJECTION]: The skill implements multi-agent pipelines where data is passed between agents, creating a surface for potential indirect prompt injection.
Ingestion points: Untrusted data enters the context through user-provided prompts in ADW scripts (e.g., templates/adws/adw_build.py) and is subsequently passed as JSON envelopes to downstream agents.
Boundary markers: Data is separated using Markdown headings in templates (e.g., templates/prompt_engineering/builder/user.md), but lacks hardened delimiters or explicit "ignore" instructions for the variable sections.
Capability inventory: Agents possess significant capabilities including repository modification and shell access via the underlying pi agent framework.
Sanitization: Prompt and envelope data are interpolated into Markdown templates without advanced sanitization or structural validation against injection patterns.

sssf