The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill uses the Agent tool and TeamCreate to spawn 'general-purpose' subagents (often using the Opus model) that are authorized to implement code, create stubs, and run test suites. This creates a delegated execution chain where the lead agent manages multiple sub-processes with command execution privileges across the pipeline.
[PROMPT_INJECTION]: The pipeline has a significant surface for indirect prompt injection. Multiple templates, such as acceptance-test-writer-prompt.md and plan-writer-prompt.md, ingest the 'FULL TEXT' of design documents and implementation plans directly into subagent prompts.
Ingestion points: Design documents (Phase 1), Implementation Plans (Phase 2), and Git diffs (Phase 3) are used as direct inputs for subagents.
Boundary markers: Explicit delimiters or 'ignore instructions' warnings are absent in the prompt templates when interpolating external text.
Capability inventory: Subagents are assigned the 'general-purpose' type, providing them with tool access for file system modification and command execution.
Sanitization: No escaping or validation of external content is performed before interpolation into subagent prompts.
[COMMAND_EXECUTION]: The skill performs extensive file system and git operations. Notably, the 'Cleanup Agent' (cleanup-prompt.md) is empowered to autonomously delete code and tests based on its own analysis of the codebase. The orchestrator also writes session metrics and decision logs to the /tmp directory.

build