The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes tools-manifest.json which contains schemas captured from an MCP server. This data ingestion creates a surface area where malicious tool definitions or descriptions could attempt to influence the agent's domain mapping and code generation logic.
Ingestion points: The skill reads tools-manifest.json as a primary input for Stage 2 of the pipeline.
Boundary markers: There are no explicit delimiters or instructions provided to the agent to disregard potential instructions embedded within the tool schemas.
Capability inventory: The skill utilizes local script execution capabilities via bun and npx to perform generation and end-to-end testing.
Sanitization: No sanitization or filtering of the schema content is described; the validation focuses on structural Zod compliance rather than content safety.
[COMMAND_EXECUTION]: The skill instructs the agent to execute a series of local scripts to validate the generated domain map and the resulting SDK.
Evidence: The Validation section includes commands such as bun scripts/generate-sdk.ts, npx vitest run, bun scripts/e2e-test.ts, and bun scripts/validate-generated.ts.
Context: These scripts are part of the internal development and testing workflow for the Stitch SDK.
[DYNAMIC_EXECUTION]: The skill is a component of a code generation pipeline where the output produced by the agent (domain-map.json) is used by local tooling to generate executable SDK code.
Evidence: The skill is explicitly described as Stage 2 of a generation pipeline that drives codegen, using scripts/generate-sdk.ts to transform the intermediate representation into final code.

stitch-sdk-domain-design