The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run to interact with Git for repository inspection (e.g., status, diff) and to execute local helper scripts within the skill directory.
Evidence: Found in scripts/blueprint_runtime.py, scripts/architecture_handoff.py, and scripts/_skill_protocol_runtime.py.
Context: These calls are used for legitimate repository state assessment and protocol implementation. All calls are configured with shell=False and target known local scripts or the git binary.
[DYNAMIC_EXECUTION]: The skill employs importlib.import_module to load internal modules for handling different architecture and planning artifacts.
Evidence: scripts/blueprint_runtime.py and scripts/_skill_protocol_runtime.py use dynamic imports to load sibling scripts like architecture_handoff.py.
Context: This is a standard Python pattern for modular codebases and does not involve loading or executing code from untrusted external sources.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external request and architecture files, which could serve as a vector for indirect prompt injection if those files contain malicious instructions.
Ingestion points: scripts/blueprint_runtime.py and scripts/architecture_handoff.py parse content from repository files cited in a task request.
Boundary markers: The skill's instructions in references/delegation-protocol.md provide a strong mitigation by explicitly directing the agent to ignore any instructions found within source files.
Capability inventory: The skill's capabilities are limited to Git inspection and writes confined to the .first-four-hours/blueprints/ workspace.
Sanitization: The blueprint_runtime.py script enforces a strict grammar for generated blueprints, including filters to prevent the use of shell interpreters or mutating Git commands in the plan's verification sections.

blueprint