Repository navigation
Releases
v0.13.0
0.13.0
API request planning can propose actual OAS-typed values for all pending calls in one invocation. Core validates each call against its unchanged contract, retains independently accepted calls, and attributes literal values to the original messages in code. The model no longer selects character positions, original-message IDs or lexical source IDs in this opt-in protocol. Declared enum, boolean and null interpretation can be included in the same proposal and is recorded separately from verbatim source evidence.
generateWorkflowApiValues provides shared value planning; generateApiValues provides the same protocol for a scoped call. analyzeApiValueAnswers handles one actual follow-up reply. selectWorkflowApiBindingSources selects producers jointly while callIds can restrict work to pending calls without removing other source candidates. Field binding, cycle validation, confirmation, checkpoints and visible retry policy retain their existing owners.
Identity checks no longer construct complete value/model contracts. JSON-only source consumers avoid lexical catalog allocation. Anonymous standalone validators share a bounded content-and-dialect cache; exported embedded schema IDs remain source-scoped so equal field schemas do not collide in a parent schema. Explicit OAS resource IDs remain unchanged.
Compatibility
These are additive, opt-in APIs. Existing source-selection APIs and legacy fingerprint bytes remain compatible. New value receipts use a distinct contract and must not be interpreted as previous source-catalog receipts. Hosts must migrate their state, independently replay current receipts and start a fresh planning session across this migration. No automatic host migration, hidden model retry, default insertion, type coercion or local package substitution is included.
Value proposals and reply extraction use function calling with strict: false: partial OAS objects must be allowed to omit unanswered fields, while provider strict mode requires every declared object property. Actual field types and enums remain in the supplied schema, and Core performs the unchanged OAS checks per call. This is not a claim of provider-enforced strict enum output or proof of semantic correctness. The joint producer decision uses a closed strict schema.
Verification and limitations
See the verification record. Source tests, an isolated host-consumer patch, private live-model input planning and local n8n execution are separate verification scopes. The existing Ajv runtime failure for a declared __proto__ property remains unresolved. Previously disclosed transitive dependency limitations are unchanged; no new security-audit claim is made.
Product installation is tracked separately in the verification record. Full product generation, remote n8n registration/execution and service deployment are not verified by these source checks. Private original messages, OAS documents, host source, credentials and local artifacts are excluded from public archives.
Packages
v0.12.0
0.12.0
Workflow input planning can now use one original-source catalog for all pending API calls. Core derives values, types and message ownership from source references. The model chooses OAS target templates and compatible original references together, without separately generating values, quotations, message IDs, fingerprints or shared-input indices.
Core compiles each call against its original OAS, explicit bindings and confirmed inputs. Independently checked calls remain available when another call is rejected. Serialized receipts are replayed against the current originals and call contracts before reuse. Missing values remain input questions; there is no type coercion, source reassignment or invented default.
Compatibility
Core adds the shared catalog, target description, compilation, fingerprint and receipt-validation APIs through its public export. LangChain adds generateWorkflowApiArguments. Existing single-call APIs retain their current protocol. The new planner requires strict function-calling structured output and makes one literal planning call per invocation; optional semantic enum interpretation is separate and may add model calls. Hosts own visible bounded replanning and retention of previously returned, validated batches.
No implicit retry or host migration is included in the library. A whole-proposal syntax error can reject the current batch, and an interrupted invocation has no accepted receipt. Previous validated batches can still be retained by the host. Valid source ownership and JSON types do not prove business intent or remote service availability.
Verification and limitations
The verification record separates source tests, host-consumer checks and a private live-model input-planning probe from product installation and execution. The existing Ajv prototype-key runtime regression remains unresolved. This release does not resolve the separately observed remote model timeout or prove full workflow generation, remote n8n registration or business API execution.
Private original messages, OAS documents, host source, credentials and local artifacts are excluded from the public packages. Product integration and service deployment are separate steps.
Packages
v0.11.0
0.11.0
Tracked API arguments now propose their OAS destinations and original sources together as one construction plan. Core checks each destination/source pair, assembles original values through explicit item/property mappings, and validates the resulting request and source receipts. A destination-only choice can no longer freeze the next model call into a source contract with no possible answer.
Finite declared input roots retain OAS-compatible source choices. Generic descendant construction is available only under roots whose original OAS permits objects or arrays. Scalar-only calls do not receive an unrelated global JSON candidate catalog. Wrong message attribution reports the original message identities containing the exact complete literal as rejection facts; code never reassigns a source, splits a comma list, coerces a number to text or invents missing input.
Compatibility
Public argument input/result shapes are unchanged. Core adds listApiLiteralContainerTargets through its public export. Custom models and test doubles must return the complete existing literalInputs/literalSources proposal in generate_api_arguments; the former select_api_literal_targets call is removed. Each attempt uses at most one tracked literal call, while optional enum interpretation may add one call. Fully bound/confirmed inputs remain code-only. Legacy untracked value generation keeps its existing protocol.
The library adds no hidden retry. Hosts own explicit bounded replanning, progress, source identity and budgets. The example graph retains rejection accounting while forwarding only the immediately preceding rejection for the unchanged source. Transport, authentication, configuration and enum errors remain separate failures.
Verification and limitations
The verification record distinguishes source checks, independent package consumption, isolated n8n execution and private live-model checks. The existing Ajv prototype-key runtime regression remains unresolved. Model semantics can still be wrong, and a bounded host may exhaust its attempts; validated provenance and types do not prove business intent.
Private messages, OAS documents, credentials and host application code are excluded. Public installation, full hosted workflow generation, remote registration, business API execution and deployment are separate checks.
Packages
v0.10.1
0.10.1
Repeated schema projection now reuses validated preparation for the exact JSON declaration, including declarations cloned during material rebinding or checkpoint restoration. Preparation retains at most 32 resource contexts per process. Changing any source content requires a different preparation, even when its $id stays the same. Invalid JSON values and unresolved references remain errors; failed preparation is never retained.
Prepared resources are immutable internally. Public projected schemas remain independent mutable JSON values, so changing a returned schema cannot change another consumer's source contract. Reference enumeration for native output composition validates the selected source and pointer without constructing output schemas. Output schema generation and compilation retain their original contract checks.
Compatibility
Public function signatures, model tool schemas, source receipts and generated workflow behavior are unchanged. This is not a retry policy, a model/review budget increase, a reference-removal workaround or a host runtime upgrade. A bounded preparation cache can evict an entry; a later use then validates and compiles it again. Package consumers continue to own their Node runtime and deployment.
Verification and limitations
Source type checks, lint, formatting, focused recursive-resource and native-output tests, graph-document checks and local n8n 2.37.10 integration runs were exercised. Details and final counts are recorded in the verification record.
A private connection-step replay produced byte-identical model messages and output schema before and after this change. Its measured preparation time fell from approximately 25 seconds to approximately 3 seconds; separate instrumentation reduced compiling Ajv instances from 3,677 to 3 and compile calls from 7,354 to 6. These are measurements of one captured input, not a general latency guarantee or proof that an external connection failure is fixed. Private source documents, messages, responses and credentials are excluded from this repository.
The existing Ajv prototype-key runtime regression remains unresolved. Local execution tests include expected failures for invalid data and missing values; passing means those cases stopped at the expected boundary, not that invalid data was accepted. Full hosted-product generation, remote workflow registration and service deployment are separate checks.
Packages
Source synchronization, the matching annotated tag, public npm archive verification and the published GitHub Release are separate release steps. Product installation and deployment are reported separately.
v0.10.0
0.10.0
Tracked API argument generation now selects concrete request destinations before selecting original value sources. Core resolves each selected path against its remaining OpenAPI value contract, including finite paths through recursive schemas and dynamic properties. JSON source candidates must satisfy that destination's schema. A number candidate is no longer offered for a string array item. Text remains text, and the complete request and original source receipts are still checked without coercion.
Core exports describeApiLiteralTargets and ApiLiteralTargetContract. OAS validation issues retain the concrete failing property or array-item path. The reference host reserves both literal-planning calls before starting dependent argument work.
Compatibility
This changes the tracked model tool protocol and invocation count. Custom models and test doubles must handle select_api_literal_targets, followed by generate_api_arguments when literals are supplied. The public generateApiArguments input and result shapes remain unchanged. Supplied literals use two explicit calls, an explicit no-literal decision uses one, and wholly bound or confirmed inputs use none. Optional enum interpretation can add one call. Update host reservations and use new planning sessions rather than replaying old outputs as destination plans. Untracked legacy value generation keeps its single-call protocol. There is no automatic checkpoint migration, hidden retry or value conversion.
Verification and limitations
See the source verification record. The existing Ajv prototype-key runtime regression remains unresolved. Source membership and type compatibility do not prove semantic relevance. A real-model enum proposal copied an entire answer instead of its valid enum substring; validation rejected it and one explicit, source-checked replan succeeded. The change does not guarantee that a model always follows its output schema on its first attempt.
Private live probes verified new argument generation and six independent Dev reads on real n8n 2.37.10. Their test wiring was explicit, and one test specification declared anonymous access after the actual service accepted an anonymous request. This is not full natural-language workflow regeneration, hosted-product verification, remote registration or deployment. No private specifications, credentials, business responses or company application source are included in this release.
Packages
Source synchronization, the matching annotated tag, public package/archive verification and the published GitHub Release are separate release steps. Product integration and service deployment are recorded separately.
Dependency audit reports 16 affected dependency records (10 moderate, 6 high; no critical). Dependency remediation is not included in this scoped release; a clean security audit is not claimed.
v0.9.0
0.9.0
Connection planning can express a closed API call inventory and exact material values and container sizes as typed, code-checked goals. Hosts opt in with reviewBoundary: 'declared-contract': Core and the model wire schema reject free semantic goals in this mode, while the separate initial coverage review still checks the original requirements. Omission retains the legacy semantic-review contract. Completion assessments cannot override these checks by demanding unrelated proof about hidden service implementation. Explicit unsupported requirements retain their original requirement references, affected nodes, grounds and a classified implementation, material or verification limitation. The verification target is generated configuration under supplied contracts, rather than future response values or hidden handler behavior.
Capability-owned value policies distinguish missing business inputs from generated text and internal references. Native questions identify the exact missing business fields. Undeclared ownership does not authorize a question. Missing internal material is reported as a technical limitation; it is not delegated to the user as a request to invent a schema.
New Core inspection and n8n compilation APIs preserve independently checked execution components while keeping dependent or unresolved graph regions as drafts. They retain actual configuration, prepared API inputs and bindings, diagnostics and accepted wiring. Partial artifacts do not imply that the whole request is fulfilled or that external execution succeeded.
Compatibility
This is a breaking planning-contract change. Use new connection sessions and native question receipts. needs-clarification requires fieldPointers authorized by the capability's parameterValuePolicy. Handle needs-materials in model proposal consumers, NativeParameterMaterialUnavailableError, and WorkflowConnectionUnsupportedRequirementsError. Persisted questions and checkpoints are not migrated automatically. See the contract guide.
Verification and limitations
The source verification record separates static checks, synthetic regressions, real-model probes and isolated n8n execution. Natural-language interpretation remains a model judgment. The existing Ajv prototype-key regression remains unresolved. Configured-source live Dev execution passed the positive, mismatch and missing-price conditions; its wiring was explicitly supplied for the test. Product installation, complete browser regeneration and service deployment remain separate pending steps.
Packages
The release process verifies matching source, tag, public version metadata and downloaded archive integrity before publishing these notes as the GitHub Release. Company product integration and service deployment are separate; no private specifications, credentials, company source or local diagnostic artifacts belong in these packages.
v0.8.0
0.8.0
Connection completion now reviews the actual configuration against original requirements. Current API arguments, recorded literal attribution, request/response contracts, native settings and the complete operation inventory are available through code-owned evidence. Recorded literals and their original locations are rechecked before a model can approve changed values.
After structural planning finishes, the adapter uses an independent fixed-material review prompt. Prior verdicts and rejected explanations are not supplied as semantic authority. Evidence is selected from a strict map of existing identities and paths; code supplies original quotations and causal node identities. Unreviewed, satisfied, violated and insufficient-evidence states are distinct. Negative reviews and invalid completion evidence stop explicitly instead of consuming an unchanged proposal loop.
The assertion capability exposes its actual success output contract. The optional compose-native-outputs feature builds one named result object from declared native JSON outputs. Code derives its field contracts, waits for its selected inputs, preserves original JSON values and rejects missing or ambiguous source items. Merge append retains its item-preserving behavior. These additions do not change comparison results or API response values.
Compatibility
This is a breaking planning-contract change. Semantic goals require scope; assessments use status and validated evidence rather than satisfied: boolean. The adapter's model selection shape differs from the Core receipt shape. Native extensions create a new source epoch and return previousSession; hosts retaining extension receipts must keep it. Existing connection checkpoints must start a new session. No automatic conversion, inferred scopes or stale receipt recovery is provided. Material-only judgments survive unrelated graph patches; workflow judgments and changed evidence are invalidated.
Hosts must preserve the original literal source messages when supplying attributed API materials and handle WorkflowConnectionCompletionIncompleteError and WorkflowConnectionCompletionEvidenceError. The reference LangGraph implements explicit termination. See the contract and migration guide.
Verification and limits
Final verification results are recorded in the source verification record. Semantic interpretations remain model assessments, not proofs of arbitrary business meaning. Exact-source verification proves the recorded token and its current value; it does not prove that a model chose the correct business role. Planning readiness is separate from external API or engine execution.
The existing Ajv prototype-key regression from the preceding release remains an acknowledged limitation. Do not interpret the focused completion tests as an entirely clean suite or a clean security audit. No company specifications, credentials or private runtime artifacts are included in this release.
Packages: Core, LangChain, n8n.
Source: d58c3e727e4962df0f59e186a07a67784014103a, annotated tag v0.8.0. All three downloaded public npm archives matched the audited local archives exactly. Product/service deployment is separate.
v0.7.2
openapi-flow 0.7.2
This release preserves recursive static OpenAPI schemas as finite JSON resources.
Resolved runtime object cycles are converted to explicit $ref links and $defs
before operation contracts cross JSON boundaries. The original catalog snapshot
is retained. No fields, nulls or validation rules are removed or coerced.
API field, array-item and object-entry projections carry their schema resources.
Deep required-input checks follow the actual value path. Host validators and n8n
standalone validators register declared nested resources before their parents,
including resources shared by several API outputs under prefixItems. Runtime
pointer lookup uses a pure closed-resource resolver without a compiler or network.
Compatibility: existing planning wire formats are unchanged. Core adds the public
listSchemaResources utility and SchemaResourceRegistration type. Static schema
recursion is supported; relocating dynamic/recursive anchor scopes during static
bundling is explicitly rejected. Non-schema structural cycles and unresolved
references fail with their source. This is not a claim of complete OAS support.
Verification passed typecheck/build, lint, formatting, 54 checks against three
private OAS documents, an independently installed strict consumer, focused
recursive regressions, and four synthetic executions
in n8n 2.37.10 with two local reads each. A separate private, explicitly assembled
six-read workflow passed actual service response guards and three comparisons;
three in-memory price mutations each failed the comparison. These checks do not
claim fresh end-to-end model planning, product browser success or deployment.
The full suite passed 1928/1929; the existing Ajv __proto__ slot-assignment
failure remains. The dependency audit still reports high 6 and moderate 11.
External dependency records are unchanged.
Packages:
v0.7.1
openapi-flow 0.7.1
This release fixes a gap between desired-region validation and the
host's bounded connection-replanning loop. Core exposes typed region validation
failures with the unadopted region and structured violations. LangChain can
replay the captured original selection against the exact current source and
create a checked rejection. The reference host counts that rejection under the
existing budget and supplies only the latest current-revision feedback. It does
not deduplicate, adopt invalid edges, fabricate a patch or retry transport errors.
Connection step results and planning errors include a private proposal-attempt
receipt, with a code-generated ID and the pre-call session/source/revision. The
receipt preserves decoded JSON before port resolution and patch derivation.
Sensitive output is explicitly withheld, and non-JSON output is unavailable.
The reference graph checkpoints the attempt before calling a separate private
sink. Its optional file-store example uses owner-only permissions, atomic record
publication and idempotent same-ID writes. Sink failure resumes without another
model invocation while the host checkpoint remains available.
Compatibility: the model's 0.7.0 desired-region wire format is unchanged. The
rejected-proposal union now also accepts canonical set-region proposals;
consumers of that union must handle this variant. Step results expose the new
receipt. Keep it out of browser events, public state, ordinary logs and prompts.
Hosts must wire their own private sink for durable history. Memory checkpoints
still do not survive process replacement.
Verification includes focused rejection/correction, unchanged accepted state,
shared budgets, source-bound feedback validation, private capture, checkpoint
resume and idempotent file tests. A separate private connection-only live-model
trial captured five original proposals and completed six starts and seven edges
with zero SDK errors or warnings. That trial had no duplicate connection proposal
and did not validate a full browser run, business API execution or n8n execution.
Version-aligned verification passed typecheck/build, lint, formatting, 59 OAS
checks, an independently installed strict consumer and four synthetic n8n
executions (20 local requests). The full suite passed 1915/1916; the existing Ajv
prototype-looking-key failure remains. The audit still reports high 6 and
moderate 11. Library verification is separate from product integration,
browser testing and service deployment.
Packages:
v0.7.0
openapi-flow 0.7.0
Connection models now select the desired incoming relations and boundaries of
explicit target nodes with set-region. Core calculates additions and removals
against the accepted graph. Selecting an existing connection retains it instead
of adding a duplicate. Invalid duplicate selections, missing fields and protected
edge removal still fail explicitly; outputs are never silently repaired.
Confirmed independence is enforced before a partial graph is adopted. Necessary
connection candidates also use declared data dependencies and typed ordering,
route and join goals. These checks do not replace atomic whole-graph validation
or establish that arbitrary business semantics were understood correctly.
Root failures now identify actual incoming edges and failed preconditions. Step
feedback must match the exact current session, source and revision. The reference
host preserves the complete private rejection ledger and budget accounting while
sending only the latest current-revision rejection to the next step call. Its
ready-work scheduling policy prevents a blocked consumer bundle from repeatedly
starving necessary producer work; it creates no execution edges.
Compatibility
This minor release changes the step and connection-revision model wire contract.
Custom providers, mocks and recorded responses must replace model-generated
add/remove patches with node-local desired state. The old
WorkflowConnectionPatchSelection model type is replaced by
WorkflowConnectionRegionSelection. Core patch callers and returned decisions
retain their explicit patch API. Single ports are copied from the actual node
contract; multiple ports require explicit choices. Removed material nodes remain
explicit deletion choices. Historical feedback from another revision is rejected
before a model call. Existing checkpoints with accepted independence violations
are rejected instead of silently migrated. See the
contract and migration.
Step and revision structured-output schemas use standard local JSON references
to share endpoint choices. Custom model providers must support those references.
Packages
Verification and known limitations
Implementation checks passed 41 focused regressions, build/typecheck, lint,
formatting and six compiled-graph/document checks. Full tests passed 1,908 of
1,909; the previously documented Ajv handling of the literal __proto__
property remains a failing test and is not skipped or counted as passing.
Version-aligned checks passed 54 original-OAS regressions and independent
archive installation, strict typechecking and runtime compilation without
workspace symlinks or skipped library checks. Deterministic connection,
connection-revision and initial-material-recovery regressions passed ten isolated
n8n 2.37.10 executions and fifty local fixture requests, including baseline runs.
These synthetic runs establish engine behavior for their fixtures, not
natural-language planning accuracy or business-service behavior.
A private real-model connection-stage run used six prepared API calls and two
unchanged native configurations. It completed in 14 model calls after one goal
rejection and four connection rejections, producing six independent starts,
a six-input gather and a following comparison. Generated n8n JSON passed SDK
validation with no errors or warnings. Three earlier failed development runs
were retained and informed contract and scheduling changes. This is not a
controlled quality benchmark, a full browser generation, business API execution,
n8n engine verification or a service deployment. Latest-current feedback versus
no-feedback quality has not been compared experimentally.
Release archive and independent-consumer verification results are recorded in
the verification history. Private specifications, records, credentials,
model prompts and business responses are excluded from the public source and
packages. Dependency licenses, including n8n's conditions, apply independently.
External dependency versions are unchanged. Existing audit findings and engine
verification limits are recorded in the verification history; no clean audit or
full end-to-end success is claimed.
The current dependency audit reports 6 high and 11 moderate findings (17 total,
0 critical). These findings remain unresolved.