The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted user input (task descriptions and specifications) and interpolates this data into prompts for agents with high capabilities.
Ingestion Points: User input enters the system via the initial_idea and spec_path arguments in SKILL.md and workflow.py.
Capability Inventory: The skill possesses extensive capabilities, including writing files (write_artifact), spawning subagents capable of shell execution (spawn_subagent), and delegating to child workflows (TeamWorkflow) that modify source code.
Boundary Markers: While the system utilizes structured output markers (STRUCTURED_OUTPUT_START) for internal agent communication, there are no strict sanitization routines or explicit boundary markers to isolate the user-provided 'idea' from the instructions passed to the specification and implementation subagents.
Sanitization: No explicit sanitization or filtering of the user-provided content is performed before it is used to drive the autonomous development pipeline.
[COMMAND_EXECUTION]: The skill is designed to generate and execute shell commands for the purpose of task verification and QA.
Evidence: The SKILL.md and workflow.py files describe a process where subagents generate verification_command scripts that are executed to prove completion of tasks. Phase 4 (QA) specifically includes an empirical mutation-test lane and code verification routines.

autopilot