Skip to content

v0.3.17

Choose a tag to compare

@carefreelife98 carefreelife98 released this 25 Oct 15:58
· 49 commits to main since this release

openapi-flow 0.3.17

Keep all three packages aligned at 0.3.17. Node.js 24 or later and a compatible
host LangChain model are required.

Initial goal rejection feedback

createWorkflowConnectionGoalRejection classifies an initial connection-goal
output or contract failure against the unchanged source. Its public Zod schema
and inferred type retain the exact JSON output, validation stage/code and reason
for private host checkpoints. planWorkflowConnectionGoals accepts these records
as optional rejectedProposals, validates them before the call and makes exactly
one model invocation. No hidden loop or response repair was added to the library.

The private LangGraph reference composes automatic replanning under the existing
shared call/rejection/deadline budgets. Invalid goals never become a session or a
public graph. Changed sources, cancellation and transport failures still fail
without semantic replanning. Existing completed-goal checkpoints remain readable.

Stage-specific clarification input

Core's additive describeApiOperationCandidate returns the existing discovery
contract from an original operation. Clarification receives all supplied API
identities, methods, paths, descriptions and native feature descriptions, rather
than duplicated raw operations and configuration schemas. Later selected-operation
planning and validation still use the complete original OAS. No API filtering,
value invention, response rewriting or support restriction was introduced.

This is not a timeout fix. A recorded broad request over 176 candidates still
exceeded its unchanged 60-second model deadline after input projection. No general
latency or workflow-quality improvement is claimed.

Verification and compatibility

Strict types, lint, formatting, 56 original-OAS checks, all 21 isolated engine/
runtime suite commands and independent strict installation/compilation passed.
Final unit results were 1,374/1,375, with only the retained Ajv failure; skipped and
cancelled counts were zero. Final archive hashes match the independently installed
consumer. The first consumer attempt's incompatible OAS/fixture pair and its
corrected explicit verification are retained in the
release record.
The change is additive; single-pass APIs and complete DAG validation remain.

A separate explicit fault-injection check rejected one invalid initial goal, then
completed with two actual-model calls. Four isolated n8n 2.37.10 executions and
twenty loopback HTTP requests passed without changing the generated plan or
upstream values. This was a fixed selected-API fixture, not an end-to-end broad
query evaluation. No business-service API was executed.

A separate short clarification request over the same 176 candidates completed
in 25,071 ms with three supplied-intent statements and no questions. This is not a
full workflow run, does not replace the failed broad request and is not a benchmark.

The enabled Ajv properties.__proto__ regression and previously accepted n8n SDK
Zod 3 security limitation remain. Audit reports the same High 5 advisory/dependency
records as 0.3.16, excluding only suggested-fix metadata. The automated gate is
not a clean pass. No
external dependency version, exclusion or security override was changed. The n8n
dependency remains subject to its own Sustainable Use License. npm publication,
GitHub Release, local host installation and service deployment are separate steps.