v0.4.3

openapi-flow 0.4.3

Changes

• Optional OAS enum interpretation now requires an explicit enumIntentInputs
  decision in the initial request-value output. An unknown value is no longer
  automatically sent to semantic enum selection just because its OAS field is
  required or requested. An empty list skips that extra model call.
• Core validates the selected paths against actual unanswered OAS enum inputs.
  Duplicate, unknown, non-enum, already supplied, confirmed and bound targets are
  rejected. The same scope is checked again when replaying a semantic result.
• Code includes selected semantic targets in requested-input presence. The model
  need not regenerate the same path in two lists. Typed OAS values are still copied
  by code from the selected choice; display explanations do not control scheduling.

Packages

• Core
• LangChain
• n8n

Compatibility

selectOasEnums: true now supplies and requires enumIntentInputs in the model
output schema. Custom model doubles must follow that schema; old outputs are not
silently migrated. When the option is disabled, no new field or semantic call is
required. Fully bound or confirmed calls retain their code-owned no-model path.
Core's optional filter can still be omitted to discover available enum questions;
inventory discovery is not permission to interpret every unanswered input.

Verification and limitations

The release record
tracks exact checks and publication state. Focused checks cover explicit enum
selection, unknown values, literal provenance, cancellation and scoped replay.
The versioned source passed strict type/build/Lint/format checks and 1,796/1,797
enabled tests, retaining the known Ajv failure with no skips or cancellations.
Four original OAS files passed 59 checks. Eleven isolated n8n executions comprised
five successes and six deliberate invalid-input stops against local fixtures.
Controlled actual-model checks preserved an unspecified input and separately
accepted an explicitly confirmed local fixture. They are input-stage checks, not
complete workflow acceptance or business-service execution.

A fresh unchanged multi-OAS scenario reached input deferral correctly, but its
later connection plan failed: the model configured a ten-input Merge and connected
only three grouped results. This release does not claim to resolve that separate
planning error. UI verification and independent five-scenario comparisons remain
unfinished; semantic accuracy is not proved by a valid output schema.

Post-install actual-model checks exposed a remaining input-stage failure: the
model wrongly selected an unspecified input as a supplied semantic intention, so
a later enum selection still promoted an illustrative value. The explicit local
fixture control passed. The scheduling contract works for the declared decision,
but semantic classification is not solved or generally guaranteed. The failed
original result was retained without repeated attempts to obtain a passing run.

The disclosed Ajv properties.__proto__ regression remains enabled. Upstream SDK
dependency audit reports High 5 and no other severities; no clean-audit or all-tests-pass
claim is made. Host installation and service deployment are separate operations.