Code review on PR #26631 surfaced two blocking gaps:
1. The OpenAPI legacy shim (addLegacyErrorSchemas in public.ts) still
declared BadRequestError as the old {data, errors, success} shape,
so the SDK's typed result.error path saw fields that don't match
the actual wire body. Update the shim and regenerate the SDK so
typed consumers see {name, data:{message,kind}}.
2. The regression test only triggered kind:"Payload" (request body
parse). The user-reported failure was kind:"Body" (response encode
on a corrupt stored row). Add a second test that mirrors the OMO/
Windows scenario: stored step-finish part with NaN tokens.output
makes the messages endpoint 400 — assert the new body has the field
path in data.message.
Plus an end-to-end SDK canary in sdk-error-shape.test.ts: asserts the
v2 SDK's wrapClientError extracts data.message into Error.message and
preserves the full body in cause.body, so future regressions of either
the server shape OR wrapClientError surface immediately.
bun typecheck, all 18 tests in 4 adjacent files pass.
Kit Langton
80897e2159