The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent on how to build handlers for external data (wire-protocol JSON) that can trigger high-capability actions, such as private simulator API calls and subprocess execution. This represents a potential indirect prompt injection surface if inputs are not strictly validated.
Ingestion points: The skill defines the baguette serve WebSocket and baguette input stdin as the primary entry points for external JSON envelopes (documented in Phase 0 of SKILL.md).
Boundary markers: Guidelines recommend the use of structured parsing utilities (Field.required*) and the Gesture.parse pattern to validate data before execution (described in SKILL.md and references/domain-models.md).
Capability inventory: The skill guides implementation of code that interacts with private SimulatorKit / IndigoHID symbols and spawns host processes using a Subprocess collaborator wrapping Foundation.Process (documented in SKILL.md and references/domain-models.md).
Sanitization: Instructions mandate failing loudly on parse errors and moving all logic into rich domain models to ensure input is handled safely (described in references/domain-models.md).
[DYNAMIC_EXECUTION]: The skill documents patterns for runtime resolution of private framework symbols, which is a core requirement for the tool's intended purpose of simulator interaction.
Evidence: Phase 2 of SKILL.md describes the resolution of private SimulatorKit symbols within infrastructure adapters like IndigoHIDInput using dynamic loading techniques.

app-implement-feature