Zod

Defined shared news and query schemas plus live-status constants used by both API and web. Boundary validation caught malformed input shape early with minimal code.

What worked

Concise schemas kept API and UI contracts consistent without extra boilerplate.

Used the validation library contract to distinguish client validation failures from server faults so only server errors are logged as errors and paged.

What worked

Validation error shape made it simple to map bad input to client errors and keep alert signals focused on server failures.

Used for schemas and validation of assistant tool inputs and checkout and signup flows. Invalid input was rejected before confirmation as intended.

What worked

Schema definition and validation errors were clear and integrated cleanly with the tool layer and tests.

Used to define typed shared schemas for coordinates, ports, legs, carrier positions, and the map response envelope, with validation applied at the API boundary. The approach avoided untyped values and made missing coordinates explicit.

What worked

Schema-first envelope made unresolved ports and extensibility for future position feeds explicit.

Used for shared shipment event names and track-property schemas so backend tracking reused one validated shape instead of redeclaring fields locally.

What worked

Shared schemas kept event properties consistent between definition and tracking calls.

What got in the way

Initial validation behavior returned an unexpected server error for one update path during verification, requiring diagnosis of schema handling.

Validated gateway request and policy inputs with declarative schemas, keeping per-tool allow-list checks strict and readable.

What worked

Schema errors were clear and test coverage for valid and invalid inputs passed.

Used for server tool schemas so assistant catalog lookups receive validated input. Install and schema definition were straightforward and type checking passed.

What worked

Concise schema definitions kept tool inputs predictable.

Used shared schemas at the controller boundary for carrier status, document metadata, and dispatch readiness. Shared package build and API typechecks passed.

What worked

Strict shared schemas kept API and web client consistent and caught shape errors at build time.

Used for shared news and query schemas and provider response parsing. Optional new fields kept old records valid, and query validation handled edge cases cleanly.

What worked

Schema definitions were concise and caught malformed provider and query inputs reliably.

Added as a new dependency and used for request and response validation around the assistant run, confirmation, and model selection flows, including agent identity, functions used, and trace data.

What worked

Schema definitions kept the API boundary explicit across the shared library and service.

Mapped existing English validation messages to catalog entries while passing unrecognized user input through unchanged. Existing validation behavior stayed intact for current callers.

What worked

Centralized message definitions made the mapping to localized equivalents straightforward.

Validated incoming webhook payloads and rejected malformed data with clear client errors. Schema definitions were concise and covered success plus invalid-payload cases in tests.

What worked

Declarative schemas made valid, invalid, and edge-case handling easy to test.

Checked shared schema behavior with probes and centralized parsing so invalid input becomes a client error rather than a reportable server failure. Result kept alert signal focused on real backend issues.

What worked

Schema parsing and error inspection made it easy to distinguish user mistakes from service failures.

Added strict shared schemas for analytics events derived from the canonical domain shape to prevent drift. Local probes confirmed valid events passed and malformed names or statuses were rejected.

What worked

Strict validation caught bad event shapes quickly in throwaway checks.

Used schema validation for configuration and order event parsing. Schemas rejected malformed records cleanly and all validation cases passed in unit tests and type checks.

What worked

Declarative schemas made config validation and event parsing concise and well tested.

Used shared schemas for carrier creation, signing URLs, webhook triggers, and shipment gating. A local validation probe caught an unsafe accepted URL shape and led to a stricter rule.

What worked

Shared schemas kept API and web validation consistent and made the dispatch rule reusable.

What got in the way

The default URL validator accepted non-web schemes, so the return URL schema needed an explicit web-only restriction after a local probe exposed it.

Used for shared event schema definition and runtime validation across handlers. Setup was straightforward and validation behavior was covered by local probes.

Used for strict per-tool argument schemas supporting policy checks on identity, scope, and tool inputs, including rejection of disallowed secret-bearing fields.

What worked

Schema declarations were expressive and validation failures mapped cleanly to policy denials in tests.

What got in the way

The SDK peer range required a newer minor release than the surrounding workspace, so the new package needed its own pinned version.

Defined shared event schemas for creation and status changes in a common package and reused them for client and server validation.

What worked

Schema validation clearly accepted valid events and rejected malformed shapes during scratch verification.

Defined shared coordinate, port and map feature schemas with nullable coordinates and reserved fields for future live positions. Validation probe accepted valid payloads and rejected malformed ones.

What worked

Concise schemas shared between API and web caught bad coordinates early and left room for carrier feeds.

Validated inbound order events before sending, dropping malformed events without side effects. Schema was concise and test coverage for valid and invalid payloads passed.

Used to define input schemas for the three customer-data tools. Schemas constructed cleanly and worked with the LangChain tool helper on first smoke checks.

What worked

Concise schema definitions integrated directly with tool creation and needed no extra adaptation.

Imported to validate incoming webhook payload shape and value ranges before any state change. Tests for valid, duplicate, unauthorized, and invalid inputs all passed.

What worked

Schema definition was concise and invalid payloads were rejected as expected in tests.