Tidal Threads

Compared to v0.39.0, Neva v0.40.0 redesigns streams as an explicit tagged-union protocol, expands immutable list operations, and makes common scalar collections materially faster without changing their public Neva types.

Highlights

• stream<T> is now the tagged union Open | Data T | Close, rather than a struct carrying an implicit index and terminal flag. Stream boundaries and payloads are explicit, and an index is created only when a program asks for one with streams.Enumerate<T>.
• streams.Just<T> adapts one value to a stream, while streams.Enumerate<T> adds explicit zero-based indexes when they are needed.
• lists.Append<T> is the renamed lists.Push<T> operation. Together with new Prepend<T> and Concat<T>, it provides immutable list updates.
• Scalar list<T> and dict<T> values now use native typed runtime storage where possible. Scalar-list traversal is 75–87% faster; constructing a 512-int list is 82% faster, uses 64% less memory, and makes 255 fewer allocations.
• neva tool <name> provides a consistent CLI entry point for an installed neva-* developer tool, including neva tool lsp.
• Compiler-version mismatch errors now name the workspace module, point to its neva.yml or neva.yaml, and recommend neva upgrade.

Streams: a model with explicit states

Previously, every item in a stream<T> carried control metadata even when a program only needed the value:

pub type stream<T> struct {
    data T
    idx int
    last bool
}

Now a stream is a tagged union with exactly three possible states:

pub type stream<T> union {
    Open
    Data T
    Close
}

Open marks the beginning of a stream, Data carries a value, and Close marks completion. Sources such as streams.Range, streams.FromList, and streams.Just emit this protocol; ordinary user code normally passes streams to standard combinators rather than handling the three cases itself.

This is Neva's return to the canonical FBP stream design: bracket information packets delimit a sequence of ordinary data packets. It follows the model described by FBP pioneer J. Paul Morrison, with whom Neva's author corresponded while developing the language. The earlier indexed structure was an intermediate design; experience showed that the explicit FBP protocol is the more idiomatic and robust model.

streams.Map, Filter, ForEach, and other combinators use Switch<stream<T>> internally: they forward Open, transform or handle Data, and forward Close. The three paths are intentionally shown only as a schematic here — the actual Map graph also has a FanIn and ordering/backpressure wiring that must not be omitted:

Open  -> ... -> Open
Data  -> handler -> Data
Close -> ... -> Close

The model makes the protocol visible in the type system, removes a stored index from the default representation, and keeps indexing as an opt-in operation.

streams.Just<T>

pub def Just<T>(data T) (res stream<T>)

Just adapts one value to an API that expects a stream. Given 42, it emits Open, Data(42), then Close.

This matters when a higher-level API is stream-oriented but the caller has one value. For example, the PNG example creates one image.Pixel, turns it into a one-item stream with Just, and passes that stream to image.New:

new image.New
just streams.Just<image.Pixel>
---
newPixel -> just -> new

streams.Enumerate<T>

pub type Enumerated<T> struct {
    idx int
    item T
}

pub def Enumerate<T>(data stream<T>) (res stream<Enumerated<T>>)

Use Enumerate when an index is part of the program's logic. It attaches 0, 1, 2, … to Data items instead of making every stream item retain an index by default.

std/lists

This release adds three list operations. As with the rest of Neva's value operations, each returns a new list; the input list is not mutated.

lists.Append<T> — renamed from Push<T>

pub def Append<T>(lst list<T>, data T) (res list<T>)

The dataflow below appends 3 to [1, 2] and produces [1, 2, 3].

append lists.Append<int>
---
[1, 2] -> append:lst
3 -> append:data
append:res -> :res // produces [1, 2, 3]

lists.Prepend<T>

pub def Prepend<T>(lst list<T>, data T) (res list<T>)

Prepending 1 to [2, 3] produces [1, 2, 3].

lists.Concat<T>

pub def Concat<T>(left list<T>, right list<T>) (res list<T>)

Joining [1, 2] and [3, 4] produces [1, 2, 3, 4].

concat lists.Concat<int>
---
[1, 2] -> concat:left
[3, 4] -> concat:right
concat:res -> :res // produces [1, 2, 3, 4]

Faster scalar lists and dictionaries

This release includes a runtime optimization for common scalar collections. On the measured hot path, scalar-list traversal is 75–87% faster; constructing a 512-int list is 82% faster and uses 64% less memory. At the language level, list<T> and dict<T> are unchanged.

Internally, homogeneous scalar values no longer have to be stored as a generic runtime-message box per element when a typed representation is available.

Conceptually, the runtime can now keep an integer list or dictionary in native typed storage:

before: []Msg{Int(1), Int(2), Int(3)}
after:  []int64{1, 2, 3}

before: map[string]Msg{"answer": Int(42)}
after:  map[string]int64{"answer": 42}

Heterogeneous and non-scalar values retain the existing generic representation. The runtime boxes values only at boundaries that actually require generic messages.

On an Apple M1 Pro with Go 1.26.5, the benchmarked implementation delivers:

• scalar-list traversal: 75–87% faster for 8–1024 integer items;
• construction of a 512-int list: 82% faster, 64% fewer bytes, and 255 fewer allocations;
• construction of a 512-int dictionary: 10.6% faster, 33.7% fewer bytes, and 255 fewer allocations;
• full list_at and list_slice paths: 3–4% faster with one to two fewer allocations;
• dictionary lookup: neutral in the full port path; no regression observed.

The detailed benchmark method and raw comparison are in PR #1127.

Developer tools

Run an installed language server through the Neva CLI:

neva tool lsp

neva tool is a dispatcher for installed executables named neva-<name>; it does not install or update tools itself.

Smaller developer-experience improvements

• Compiler-version mismatch errors now identify the affected workspace module, show where its declared version lives, and suggest neva upgrade.
• Tagged-union documentation now matches the compiler: declare tagged union { ... } types rather than the removed untagged-union syntax.

Included PRs

• #1042 Redesign stream as union and add streams.Enumerate
• #1047 Redesign stream item helpers and enumerate support
• #1127 Typed scalar list/dict containers and runtime fast paths
• #1144 Validate union literal payloads
• #1151, #1153, #1154, #1155, #1157 Runtime message and collection work
• #1158 CLI developer-tool manager
• #1161 Tagged-union documentation correction
• #1162 Compiler-version mismatch guidance

Documentation, CI, benchmark, and engineering-harness improvements in the same release window are included as maintenance work.

Full Changelog

v0.39.0...v0.40.0

Operating Handles

Compared to v0.38.0, this release expands the standard library with practical OS and file-handle capabilities, improves parser error handling, and tightens release/version maintenance.

Summary

• Added explicit std/io file handles: File, Open, Create, Close, ReadAllFile, and WriteAllFile.
• Existing filename-based std/io.ReadAll and std/io.WriteAll remain available as convenience APIs for whole-file reads and writes.
• Expanded std/os with typed environment, process, filesystem, and temp-path primitives.
• Fixed std/os.Args runtime list construction.
• Parser recovery now returns compiler errors for missing connection/sender/receiver sides instead of panicking.
• Release prep now bumps repository version references consistently across manifests, docs, generated headers, and tests.
• Repository AI harness assets were consolidated under .codex/.

Standard library: file handles

std/io now supports explicit open/create/read/write/close flows for programs that need to keep file access state visible in the graph.

import { bytes, io }

def WriteExample(start any) (stop any, err error) {
	create io.Create?
	from_string bytes.FromString
	write io.WriteAllFile?
	close io.Close?
	---
	:start -> [
		'out.txt' -> create:filename,
		'hello' -> from_string:data
	]
	from_string:res -> write:data
	create:res -> write:file
	write:res -> close:file
	close:res -> :stop
}

Standard library: OS primitives

std/os now covers common environment, process, filesystem, and temporary-path operations through typed, dataflow-friendly components. Query-style components are signal-triggered so programs stay explicit about when OS data is read.

import { os }

def ListCwd(start any) (res list<os.DirEntry>, err error) {
	getwd os.Getwd?
	read_dir os.ReadDir?
	---
	:start -> getwd:sig
	getwd:res -> read_dir:path
	read_dir:res -> :res
}

Parser and runtime reliability

• Malformed connection shapes that previously could panic during parser recovery now produce compiler.Error.
• File-handle runtime paths include lifecycle and registry coverage.
• New std/os runtime helpers are covered by targeted unit and e2e tests.

Internal and tooling notes

• Version references were synchronized to 0.39.0 across checked-in manifests, examples, docs, generated headers, and test expectations.
• The release-neva skill now documents repo-wide version bump requirements.
• Legacy .agent, .claude/rules, and .opencode harness roots were consolidated under .codex/.

Included PRs

• #1139 fix(parser): return compiler errors for missing conn sides
• #1141 [codex] consolidate AI harness
• #1040 stdlib/io: add explicit Open/Create/Close file-handle API
• #1041 Implement std/os package for env/process/fs primitives
• #1143 release: bump version to v0.39.0

Full Changelog

v0.38.0...v0.39.0

Typed Waters

Compared to v0.37.1, this release makes generic node usage stricter and more explicit, improves runtime safety around edge cases, and continues internal tooling groundwork.

Summary

• Generic node instantiations now require explicit type arguments (implicit any fallback removed).
• Added analyzer + e2e regression coverage so this rule cannot silently regress.
• Runtime safety improved for list_at out-of-bounds flow and ArrayInport.Select behavior.
• pkg/view now exposes port order and DI args for downstream visual/tooling work.
• Runtime microbench suite expanded for scalar list/dict message paths.
• Docs now clarify any semantics and container specialization policy.

Language change: explicit generic args for nodes

This change is specifically about generic nodes at initialization/callsites.
If a node is generic, pass <T> explicitly when you instantiate it.

Before (now invalid, missing type arg):

import {
    fmt
    runtime
}

def Main(start any) (stop any) {
    println fmt.Println
    panic runtime.Panic
    ---
    :start -> "hello" -> println:data
    println:res -> :stop
    println:err -> panic
}

After (valid):

import {
    fmt
    runtime
}

def Main(start any) (stop any) {
    println fmt.Println<string>
    panic runtime.Panic
    ---
    :start -> "hello" -> println:data
    println:res -> :stop
    println:err -> panic
}

Internal and tooling notes

• Deterministic projection/ordering guardrails in view tests were strengthened.
• Version references were synchronized to 0.38.0 across release-related files.
• Release process docs/harness were tightened for draft-first delivery and explicit publish approval.

Included PRs

• #1130 view: deterministic overload ordering guard + plan sync
• #1131 runtime: expand Msg scalar list/dict microbench coverage
• #1132 analyzer: remove implicit any default for generic node instantiations
• #1133 analyzer: add regression tests for explicit generic node args
• #1134 docs: clarify any and container specialization policy
• #1135 runtime: add safe Select/list_at coverage and fix bounds panic
• #1136 feat(view): expose port order and DI args
• #1137 release: bump version to v0.38.0

Full Changelog

v0.37.1...v0.38.0

Trace Output Polish

Patch release after v0.37.0 focused on trace output consistency and JSONL cleanup.

What changed

• Panic trace rendering aligned with the new pretty tree view (receiver <- sender, branch lines).
• Runtime new now propagates signal causes, so causal chains can reach :start.
• JSONL trace events omit port.Index when it is null (kept only for array ports).
• Added focused runtime tests for cause propagation and JSONL encoding behavior.

Included PRs

• #1128
• #1129

The Last Mile Begins

Compared to v0.36.1, this release starts the final language-completion track: runtime tracing first, then visual tooling, then debugger stack.

Summary

• Added runtime tracing for panic termination paths.
• Standardized real-time trace file output to JSONL.
• Added pkg/view foundation primitives for upcoming visual tooling.

Why this matters

In control-flow languages, production failures are explained by stack traces.
In Neva, execution is dataflow, so failure context is a causality graph of messages. This release makes that graph observable in runtime output.

Runtime Tracing

When panic happens, Neva now prints a panic cause dataflow trace to stderr.

Example A — minimal panic flow

import {
	fmt
	runtime
}

def Main(start any) (stop any) {
	panic runtime.Panic
	printf fmt.Printf
	---
	:start -> [
		'value=$1' -> printf:tpl,
		10 -> printf:args[0]
	]
	printf:err -> panic
	printf:sig -> :stop
}

Pretty trace view (target format):

panic:data <- printf:err
├─ printf:args[0] <- __newv2__2
│  └─ __newv2__2 <- :start
└─ printf:tpl <- __newv2__1
   └─ __newv2__1 <- :start

Example B — branched causes converging into one panic

import {
	fmt
	runtime
	strconv
}

def Main(start any) (stop any) {
	panic runtime.Panic
	printf fmt.Printf
	left_atoi strconv.Atoi
	right_atoi strconv.Atoi
	---
	:start -> [
		'left=$0 right=$1 extra=$2' -> printf:tpl,
		'10' -> left_atoi,
		'20' -> right_atoi
	]
	[left_atoi:err, right_atoi:err, printf:err] -> panic
	left_atoi:res -> printf:args[0]
	right_atoi:res -> printf:args[1]
	printf:sig -> :stop
}

Pretty trace view (target format):

panic:data <- __fan_in__2
└─ __fan_in__2:data[2] <- printf:err
   ├─ printf:tpl <- __newv2__7
   │  └─ __newv2__7 <- :start
   ├─ printf:args[0] <- left_atoi:res
   │  └─ left_atoi:data <- __newv2__8
   │     └─ __newv2__8 <- :start
   └─ printf:args[1] <- right_atoi:res
      └─ right_atoi:data <- __newv2__9
         └─ __newv2__9 <- :start

Note:
Current runtime output still exposes internal synthetic node names (for example __newv2__*) and may stop short of :start for some runtime funcs. The next tracing patch will align runtime output with the target view above.

JSONL Trace Files

Trace events written to file are standardized as JSONL (v=2, one event per line).

Real sample from trace.log:

{"message":{},"port":{"Path":"","Port":"start"},"event":"sent","causeIndexes":null,"v":2,"index":1}
{"message":"value=$1","port":{"Path":"__newv2__1","Port":"res"},"event":"sent","causeIndexes":null,"v":2,"index":4}
{"message":"value=$1","port":{"Path":"printf","Port":"tpl"},"event":"recv","v":2,"index":4}
{"message":10,"port":{"Path":"printf","Port":"args","Index":0},"event":"recv","v":2,"index":5}

Visual Tooling Foundation

This release also includes initial pkg/view projection primitives used by the read-only visual tooling track.

Related work

• Tracing + visual tooling tracker: #1118
• Visual editor track: #1050
• View foundation PR: #1123

Support the project

• Open Collective: https://opencollective.com/nevalang
• Star the repo: https://github.com/nevalang/neva

Maintenance Throughput

The previous "Engineered Stability" release v0.36.0 tightened delivery quality gates.

v0.36.1 is a focused patch release that continues this maintenance track: wider runtime operator coverage, benchmark surface expansion, lint debt reduction, and CI/release automation cleanup.

🌟 Summary

• Added missing runtime builtin coverage for unary negate and slice operator paths.
• Fixed delayed-echo example wait-group accounting to prevent hanging behavior.
• Expanded runtime benchmark coverage (foundation, support-wired, simple operators, and atomic builtin batches).
• Continued strict golangci-lint rollout with targeted suppression cleanup via refactors.
• Tightened OpenCode/GitHub review and release-marketing workflows.
• Updated AGENTS/tooling guidance and stdlib maintenance documentation.

Runtime & Correctness

Runtime builtin breadth completed for benchmark matrix

Runtime/operator work filled missing paths and removed stale suppression-heavy patterns:

• missing builtin neg/slice function coverage added and benchmarked
• operator creator/refactor passes reduced nolint usage in runtime funcs
• printf formatter internals were split and covered with focused tests

Example reliability fix

A delayed-echo example synchronization bug was fixed by correcting wait-group count wiring.

Benchmarks & Measurement

This release significantly expands reproducible benchmark slices for runtime behavior:

• foundation runtime e2e baseline slice
• support-wired atomic fan in/out coverage
• startup noop and atomic builtin operator batches
• scalar cast, routing/type, else/error/gate, and simple operator baselines

These changes improve visibility into runtime performance trends without introducing new language syntax.

CI, Lint, and Release Operations

• lint workflow naming/config alignment and stricter phase-1 profile enforcement
• contextual error propagation improvements in tooling paths
• release automation updates for Telegram and Discord announcement flows
• OpenCode PR-review policy tightened to focus on required changes

Docs & Harness Guidance

• hierarchical AGENTS harness and cross-tool rule shims added
• stdlib component coverage policy documented
• legacy .agent layout cleanup to keep harness guidance consistent

Language Surface

No language-level syntax or type-system semantic changes were introduced in this patch window.

📑 Related PRs

• #1112 typesystem: remove nolint suppressions via refactors
• #1111 runtime/funcs: remove nolint suppressions from operator creators
• #1110 chore(golangci): add phase-1 runtime funcs exclusions
• #1109 refactor: add contextual errors in tooling paths
• #1107 ci(lint): rename golangci check and set explicit config
• #1106 refactor(runtime): split printf formatter into focused helpers
• #1105 test(runtime): add coverage for printf template formatting
• #1103 chore: strict golangci profile with phase-1 suppression baseline
• #1102 opencode review: remove positive validation from AI PR comments
• #1095 chore(agent): clean up legacy .agent layout
• #1092 OpenCode PR review: only required changes, no praise
• #1091 Add Discord release news workflow
• #1090 benchmarks: add simple operator baseline slice
• #1089 examples: fix delayed echo wait group count
• #1087 ci: auto-run OpenCode review-pull-request skill on PR open
• #1086 runtime: add missing builtin neg/slice funcs and complete atomic builtin breadth
• #1085 benchmarks: add atomic builtin else/error and gate batch
• #1084 benchmarks: add atomic builtin routing/type benchmark batch
• #1083 benchmarks: add atomic scalar-cast builtin benchmarks
• #1082 docs: record stdlib coverage policy
• #1081 benchmarks: add atomic builtin float/string operator batch
• #1080 benchmarks: add startup noop and atomic builtin operator batch
• #1079 docs: add cross-tool rule shims
• #1078 benchmarks: add support-wired atomic fan in/out
• #1077 docs: adopt hierarchical AGENTS harness
• #1076 docs: tighten telegram release post skill format
• #1075 benchmarks: add atomic one-shot builtin batch
• #1074 ci: fix release telegram workflow for release events
• #1073 benchmarks: foundation runtime e2e baseline slice

Full Changelog: v0.36.0...v0.36.1

Post-release asset refresh (May 9, 2026)

Release binaries were re-uploaded from main commit f9dbb8e6 to fix CLI-reported compiler version metadata (neva version now reports 0.36.1).

Engineered Stability

The previous "Pragmatic Power" release v0.35 focused on language and stdlib leverage.

v0.36 is about making that leverage safer and more repeatable in day-to-day work: stricter release builds, stronger quality gates, cleaner CI lanes, security/toolchain refresh, and reduced maintenance noise.

🌟 Summary

• Security remediation pass completed across Go toolchain and dependencies.
• Repo toolchain moved to go1.26.1 (including CI).
• Release build arguments were centralized and normalized for deterministic artifacts.
• CI pipeline now runs in explicit stages (lint -> unit -> e2e) and uses shared e2e binary caching.
• Local/CI quality gates are now explicit (lint, test-unit, vulncheck, quality-gate, quality-gate-ci).
• e2e/examples test lane returned to default Go package parallelism after hardening.
• NilAway cleanup landed across internal compiler packages.
• Release workflow guidance gained a compact repo-local skill with snippet quality checks.
• Published releases now have an automated Telegram announcement path.
• Stale lint suppressions were removed and fieldalignment cleanup was applied.
• Agent guidance docs were consolidated into a compact router and updated with modern Go defaults.

🔒 Security & Toolchain

Vulnerability remediation

A dedicated sweep addressed reachable vulnerability findings and upgraded affected modules:

• github.com/go-git/go-git/v5 v5.11.0 -> v5.16.5
• github.com/cloudflare/circl v1.3.7 -> v1.6.3
• golang.org/x/crypto v0.21.0 -> v0.45.0
• golang.org/x/net v0.23.0 -> v0.47.0

go mod tidy was re-run to normalize the module graph after the update.

Go toolchain update

Neva now targets:

• go 1.26
• toolchain go1.26.1

CI workflows were aligned to the same version to keep local/CI behavior consistent.

📦 Release Build Discipline

Release build flags are now standardized in shared backend helpers and aligned with Make targets:

• -trimpath
• -buildvcs=false
• -ldflags="-s -w"

This keeps end-user binaries compact and more reproducible across release paths (install, native backend, wasm backend, and release builds).

Release workflow support

Release operations were tightened further around lightweight repo-local guidance and post-publish automation:

• a compact release-neva skill now codifies the draft-first release flow and version-bump decision points
• user-facing release snippets are expected to stay concise while still being valid Nevalang
• published releases can trigger Telegram announcement generation via OpenCode and shared messaging assets

⚙️ CI, E2E & Reliability

CI flow split + faster e2e bootstrap

The test workflow was restructured into ordered jobs:

• lint
• unit_tests
• e2e_tests

pkg/e2e now supports shared binary caching with safe fallback behavior, reducing repetitive rebuild work during e2e runs.

Explicit quality gate

The repo now exposes a clearer local/CI quality path:

• make lint
• make test-unit
• make vulncheck
• make quality-gate
• make quality-gate-ci

govulncheck also runs as a dedicated workflow job, and optional Go-only hooks can mirror the same checks locally before commit.

Parallelism rollout step

After introducing safer e2e/cache mechanics, the temporary explicit -p 2 cap for e2e/examples was removed in CI. The lane now relies on default Go package parallelism.

Stdlib cache behavior clarified

Documentation/comments were tightened around stdlib cache invalidation:

• stdlib extraction remains content-based (correct for embed.FS constraints)
• metadata-based fingerprinting remains limited to on-disk repo file use-cases

🧹 Code Health & Maintenance

• Removed stale nolint suppressions across the repo.
• Applied struct field reordering where needed to satisfy govet field alignment checks.
• Fixed NilAway findings in internal compiler packages and excluded test files from make nilaway to keep the default developer path focused on production code.
• AGENTS guidance was compacted into a high-signal docs router and extended with version-aware modern Go recommendations.

🧭 Language Surface

No language-level breaking changes were introduced in this release window. The focus of v0.36 is delivery quality and operational stability.

📑 Related PRs

• #1072 ci: release marketing to Telegram via OpenCode
• #1071 Fix NilAway findings in internal code and exclude tests in Makefile
• #1070 Add explicit quality gate (lint/test/vulncheck) and remove vuln-scan skill
• #1069 skills: add compact release-neva skill and snippet quality checks
• #1068 docs: add modern Go guidance for agents
• #1066 ci(test): run e2e/examples with default Go package parallelism
• #1065 docs(cache): document stdlib fingerprinting constraints and close issue tails
• #1062 test(ci): split unit/e2e pipeline and cache e2e neva binary
• #1061 Fix vuln findings and add reusable vuln-scan skill
• #1060 docs(agents): condense AGENTS into compact router
• #1055 lint: remove stale nolint suppressions and enforce fieldalignment
• #1054 backend/golang: centralize release build args and reproducible flags

Full Changelog: v0.35.0...v0.36.0

Pragmatic Power

The previous "Back to Dataflow" release v0.34 has cleaned up the language surface and removed a lot of accidental complexity.

Current "Pragmatic Power" releasev0.35 is what comes next: making that simpler core actually pleasant and powerful in real programs. This release is more about real leverage.

🌟 Summary

• New bytes builtin data-type for effective IO operations. We no longer pretend that all bytes are strings.
• Idiomatic type convertors in stdlib (streams/dicts/strings/bytes).
• Stdlib port naming was standardized (data/res/err/sig defaults) across many components to make APIs more predictable.
• Even stricter syntax (nodes now must have explicit names in component definition).
• The maybe type is no longer special case for type-system, it's just union now.
• Removed deferred connections syntax and switched to explicit lock wiring.
• Array-bypass syntax was finalized as [*] (instead of =>), which is much clearer and less magical.
• Neva LSP was moved to separate repository for maintainability.
• Core LSP language features landed (huge step for daily DX in editors).
• Fixed old issue with zombie processes after e2e by improving pkg/e2e package for end-to-end testing.

🧠 Language & Semantics

First-class bytes

bytes is now a first-class builtin type.

Minimal snippet (nodes + network only):

read_all io.ReadAll?
bytes_to_string strings.FromBytes
println fmt.Println<any>?
---
:start -> 'bytes_roundtrip.txt' -> read_all:filename -> bytes_to_string -> println -> :stop

A concrete stdlib API example using bytes:

#extern(write_all)
pub def WriteAll(filename string, data bytes) (res any, err error)

Node (top-level) names are now required

It makes code more consistent and easier to reason about for LLMs.

// before (now invalid)
import { fmt }

def Main(start any) (stop any) {
    fmt.Println
    ...
}

// after
import { fmt }

def Main(start any) (stop any) {
    println fmt.Println
    ...
}

maybe/error modeling cleanup

maybe<T> is now regular tagged-union modeling in std/builtin.

pub type maybe<T> union {
    Some T
    None
}

pub type error struct {
    text string
    child maybe<error>
}

No special type-system path is needed for optional/error chaining.

Deferred connections removed

We finished the remaining language cleanup from v0.34 by removing deferred connection syntax.

Before, you could write sugar like a -> { b -> c }, where delivery from b to c was deferred by a. It worked, but

1. It added one more connection form i.e. made language more complex
2. It made dataflow less obvious (it's not clear that what's deferred is receiving by c and not sending by b, which is clear using explicit locks - what desugarer was doing under the hood)
3. It made 1-1 mapping from source code to visual node editor impossible, which was the most critical problem among all 3. The best way to visualize deferred connection was to "desugar it" at the level of the visual editor, which means we would desugar it two times, at the different edges of the compilation spectrum, which... Doesn't feel right, let's say.

So now this is explicit wiring via builtin.Lock:

lock Lock<string>
---
a -> lock:sig
b -> lock:data
lock -> c

Array-bypass syntax finalized as [*]

Array bypass used to use =>. Now it is explicit port-slot wildcard syntax on both sides:

// before
in:items => out:items

// after
in:items[*] -> out:items[*]

This might seem like a opinionated change but actually it's just simplification we didn't see possible before - at AST level we used to have 2 kinds of connection, a normal one and array bypass one. Now array bypass is just a special case of normal connection where array slot index is * (which is encoded as 255 - reserved uint8 value). So all connections are "normal" now. I.e. there are just "connections". Also [*] feels more consistent with [i] (e.g. [0]) rather than using different kind of arrow =>.

📦 Standard Library: Conversion Toolkit

Stdlib port naming standardization

A lot of stdlib APIs were normalized to follow port naming convention with (data, res, err, sig) with boundary exceptions only when domain naming adds real value. The convention itself was finalized in the docs/style_guide.md document.

This is not a flashy feature, but it helps to form idiomatic conventions for the language and its standard library. This particular change should make it a little bit easier to reason about the port names. We expect you to just follow the convention without asking yourself a lot about "how do I name this port?". Also should help LLMs with codegen predictability.

New / improved conversion path components

• streams.FromString(data string) (res stream<string>)
• streams.FromDict<T>(data dict<T>) (res stream<DictEntry<T>>)
• dicts.FromStream<T>(data stream<DictEntry<T>>) (res dict<T>)
• strings.FromBytes(data bytes) (res string)

This introduces idiomatic convention for data-type convertors. We have decided to continue follow "small core" philosophy and made type convertors simple components rather than language feature.

Example: dict -> stream

const dict_value dict<string> = {
    a: 'one',
    b: 'two'
}

...

dict_to_stream streams.FromDict<string>
for_each_println streams.ForEach<DictEntry<string>>{fmt.Println<any>}?
wait streams.Wait
---
:start -> $dict_value -> dict_to_stream -> for_each_println -> wait -> :stop

Example: stream -> dict (last write wins)

const dict_entries list<DictEntry<string>> = [
    { key: 'dup', value: 'one' },
    { key: 'dup', value: 'forty-two' }
]

...

list_to_stream streams.FromList<DictEntry<string>>
stream_to_dict dicts.FromStream<string>
println fmt.Println<any>?
---
:start -> $dict_entries -> list_to_stream -> stream_to_dict -> println -> :stop

Scalar conversions

Builtin scalar converters now explicitly document intent (aligned with Go):

• Int(float) -> int (truncate toward zero)
• Float(int) -> float
• String(int) -> string (Unicode code point)

This gives a sane, predictable baseline while keeping non-total parsing/formatting in stdlib (strconv style APIs).

Bytes(string) -> bytes and String(bytes) -> string in builtin are in progress.

⚙️ Tooling & Architecture

Go 1.26 migration + go fix discipline

Repository now targets:

• go 1.26
• toolchain go1.26.0

CI now enforces go fix ./... cleanliness. If you haven't read about go fix then do it. It's awesome tool that automatically rewrites legacy Go code to its modern version respecting language and stdlib changes. Now every Neva release language is going to be better and better also because of this, among with many-many other reasons.

LSP was moved + Refactoring

A lot of groundwork landed to make this split working:

• public pkg/ast, pkg/core, pkg/indexer, pkg/typesystem,
• in-repo cmd/lsp removed,
• canonical LSP implementation lives in nevalang/neva-lsp.

This is important for velocity in neva-lsp and vscode-neva: compiler core stays focused, language tooling can evolve in its own repo.

pkg/* APIs are now usable from external Go modules

Expose public APIs for external LSP extraction means you can now import Neva AST/typesystem packages directly from another Go module.

Minimal example:

package main

import (
	"fmt"

	src "github.com/nevalang/neva/pkg/ast"
	ts "github.com/nevalang/neva/pkg/typesystem"
)

func main() {
	var _ src.Component
	var _ ts.Expr
	fmt.Println("neva ast/typesystem imported successfully")
}

LSP core language features (major milestone)

Core editor features landed:

• textDocument/definition - jump to symbol definition.
• textDocument/references - find all usages.
• textDocument/rename (+ prepare rename) - safe symbol rename.
• textDocument/hover - quick symbol/type info.
• textDocument/documentSymbol - file outline navigation.
• textDocument/completion - entities, ports, and imports.
• textDocument/semanticTokens/full - syntax-aware highlighting.
• CodeLens (references, implementations) - inline code navigation counts.

This is the baseline for vscode-neva and future visual tooling over LSP transport.

🧪 Reliability & Quality

• E2E infrastructure now handles timeout/cancel more safely by cleaning process groups to avoid orphan child processes.
• Additional lint debt on main was cleaned while preparing this release baseline (staticcheck + wastedassign findings).
• Runtime JSON spacing corruption fix (#1030) was a small but important bug-fix: pretty spacing no longer mutates string payload contents.

📑 Related PRs

Core to this release window:

• #1051 Switch builtin maybe/error to tagged unions
• #1049 fix(e2e): clean up orphan child processes on timeout/cancel
• #1039 chore: migrate repo to Go 1.26 and add gofix CI check
• #1038 Add streams.FromString and document converter policies
• #1036 Add first-class bytes type and migrate binary APIs
• #1035 Add dict<->stream converters in std
• #1034 Add Go-parity scalar converters for builtin and strconv
• #1031 Cleanup streams API and standardize stdlib port naming
• #1030 Fix runtime JSON spacing corruption in message formatting
• #1029 feat: require explicit aliases for top-level node declarations
• #1026 chore: remove cmd/lsp after extraction to neva-lsp
• #1025 Expose public APIs for external LSP extraction
• #1024 Move LSP indexer to pkg/indexer
• #1022 refactor: move ast and core packages to pkg
• #1020 LSP: add core language features
• #1018 refactor: remove deferred connections
• #1013 Replace array-bypass => with [*]

Full Changelog: v0.34.0...v0.35.0

Back to Dataflow

🌟 Vision & Direction (Q1 2026)

This release is a turning point toward the 2026 Q1 priority: a polished, “wow‑factor” working prototype. To get there we needed to correct language design mistakes before moving forward. v0.34 finishes that simplification effort and brings Neva back to its roots: a small core and pure dataflow. We’re no longer trying to mimic conventional programming languages — we’re doubling down on what makes Neva different.

With this foundation in place, we can finally shift focus to the visual node editor. The language syntax/semantics are now kinda stable: expect small tweaks, but no major reshaping. The next frontier is stdlib APIs and patterns rather than the language itself.

This release continues the language simplification work: composite sender syntax is removed in favor of explicit stdlib components, and const/literal senders are now strictly chain-only. Alongside the language changes, you get a new neva install command, clearer compiler diagnostics, and updated stdlib APIs.

🧠 Language Changes

Why this simplification (context from #963)

Neva is a pure dataflow language: messages flow between nodes over connections. Composite senders (expressions, range syntax, ternary, switch syntax, union wrapping syntax) pull references from the network in sender position, which conflicts with chain-only semantics (constants/literals/union tags must be triggered). This created:

• Context-dependent rules (sometimes literals are allowed, sometimes not).
• Compiler↔stdlib coupling (syntax like 1..10 implies streams.Range and auto-imports).
• Visual/editor mismatch (phantom nodes/edges that don’t exist in the graph).
• Unsound or inconsistent typing for union elimination (see the Switch note below).

v0.34.0 removes these composite senders and requires explicit wiring. This makes the textual format a faithful serialization of the graph and restores a single, consistent rule: senders are nodes or chain-only literals/consts triggered by a signal.

Composite senders removed (breaking)

The following language-level constructs are removed in v0.34.0:

• Unary and binary operator expressions (e.g., -x, a + b, a && b).
• Ternary operator.
• switch language construct.
• Range syntax (1..10).
• Union senders (Type::Tag(data) as a sender).

Use explicit stdlib components instead. This keeps the text format aligned with the visual dataflow model.

Before/After: Unary & Binary Operators

// before
x + y -> out:data

// after (builtin is implicit)
add Add
---
:x -> add:left
:y -> add:right
add -> :res

Before/After: Ternary

// before
cond ? a : b -> out:data

// after (builtin is implicit)
ternary Ternary<string>
---
:cond -> ternary:if
:a -> ternary:then
:b -> ternary:else
ternary -> :res

Before/After: switch construct

// before
value -> switch {
  case1 -> handle1
  case2 -> handle2
  _ -> handle_default
}

// after (builtin is implicit)
switch Switch<string>
---
:value -> switch:data
:start -> [
    'Alice' -> switch:case[0] -> :out1,
    'Bob' -> switch:case[1] -> :out2
]
switch:else -> :out_default

Before/After: Range

// before
1..10 -> stream:data

// after
import { streams }

range streams.Range
---
:start -> [
    1 -> range:from,
    10 -> range:to
]
range:res -> :stop

Before/After: Union senders

// before
type MyUnion union { Tag int }

MyUnion::Tag(value) -> handler

// after (builtin is implicit)
type MyUnion union { Tag int }

union Union<MyUnion>
---
:data -> union:data
:start -> MyUnion::Tag -> union:tag
union -> :res

Const/literal senders must be chained (breaking)

Standalone const or literal senders are no longer allowed. They must appear in a chain triggered by a signal.

// before (invalid)
$answer -> print:data
42 -> print:data

// after
:start -> [
    $answer -> print:data,
    42 -> print:data
]

Switch<T> and the Type placeholder (type-system note)

Removing the language switch construct exposes a deeper type-system reality: union elimination is not expressible as a normal generic component. A true pattern match needs different output types per branch, but Neva’s type system assigns a single type T to Switch<T> ports. A naive Switch<T> therefore cannot statically refine types by tag.

To keep the language minimal and preserve switch-like behavior, Neva keeps Switch as a stdlib component, while the compiler applies special typing rules to its case outputs. The signature uses a placeholder type Type to indicate “compiler-determined output type” per case:

pub def Switch<T>(data T, [case] T) ([case] Type, else T)

Type is not a real runtime type. It is a documentation/typing placeholder that signals compiler participation. This is a deliberate, constrained “compiler magic” to keep the surface language small while acknowledging that union elimination is not soundly expressible as a normal generic function today (see #969 for details).

Import blocks accept comments

Comments are now allowed inside import { ... } blocks and as trailing comments on import lines.

import {
    fmt
    strings // used for joining
}

Struct literals allow trailing commas

Multi-line struct literals can now end with a trailing comma.

const person Person = {
    name: 'Ada',
}

⚙️ Compiler & Tooling

neva install command

Build and install a Neva program to your Go bin path (GOBIN, GOPATH/bin, or ~/go/bin) using the package directory name as the binary name.

neva install path/to/pkg

Real‑world examples:

# Install a CLI from a local module
neva install ./cmd/hello

# Install the main package from a module root
neva install .

# Then run it
hello

Debug runtime validation (for language developers)

A new --debug-runtime-validation flag is available for language developers. It emits a validation helper into the generated runtime to print unconnected senders/receivers for wiring inspection, without shipping that helper in the repo runtime by default.

# Inspect runtime wiring during development
neva run --debug-runtime-validation path/to/pkg

Diagnostics improvements

• Unused inport errors now list the exact ports and use correct singular/plural grammar.
• Chained connection error locations now fall back correctly when the parser node has missing position metadata.
• Recursive unqualified component references are now a clear analyzer error with a hint to use builtin.<Name> when appropriate.

📦 Standard Library

Dotenv loaders are signal-style (breaking)

std/os/dotenv loaders no longer return dictionaries. They load into the process environment and send a completion signal.

import {
    os/dotenv
}

def Main(start any) (stop any) {
    load dotenv.Load
    ---
    :start -> load -> :stop
}

Load a specific file via LoadFrom, and use the override variants when needed:

import {
    os/dotenv
}

def Main(start any) (stop any) {
    load dotenv.LoadFrom
    ---
    :start -> './config/.env' -> load:data -> load:res -> :stop
}

import {
    os/dotenv
}

def Main(start any) (stop any) {
    load dotenv.LoadFromOverride
    ---
    :start -> './config/.env' -> load:data -> load:res -> :stop
}

Available components in std/os/dotenv:

• Load
• LoadFrom
• LoadOverride
• LoadFromOverride

New os.Environ

Retrieve the current process environment as list<string> in KEY=VALUE form. This mirrors Go’s os.Environ() API, which returns a slice of KEY=VALUE strings rather than a map, so Neva exposes the same shape for consistency with the runtime implementation.

import {
    fmt
    os
}

def Main(start any) (stop any) {
    environ os.Environ
    println fmt.Println<list<string>>
    ---
    :start -> environ -> println -> :stop
}

streams.Range is now explicit (breaking)

The .. syntax and sig-gated Range variant are removed. Use streams.Range with explicit wiring.

import { streams }

def Main(start any) (stop any) {
    range streams.Range
    ---
    :start -> [
        1 -> range:from,
        10 -> range:to
    ]
    range -> :stop
}

Union wrapping must be explicit (breaking)

Union senders are removed. Use the Union<T> component.

type MyUnion union { Tag int }

def WrapUnion(start any, data int) (res MyUnion) {
    union Union<MyUnion>
    ---
    :data -> union:data
    :start -> MyUnion::Tag -> union:tag
    union -> :res
}

🔁 Migration Guide (Quick Map)

Unary/binary expressions         -> builtin operators (e.g., Add, Sub, Eq)
ternary                          -> Ternary<T>
switch {...}                    -> Switch<T>
1..10                           -> streams.Range (from/to wiring)
Type::Tag(data) sender          -> Union<T>
Standalone literals/consts      -> Chain from a signal
Dotenv Load returning dict      -> signal-only loaders + os.Environ

📦 Stdlib Component Changes (Summary)

Added or changed components in v0.34.0:

• os.Environ (new)
• os/dotenv.Load, LoadFrom, LoadOverride, LoadFromOverride (changed to signal-style)
• streams.Range (moved to streams, single explicit variant)
• Union<T> (unchanged component; now required for union wrapping instead of union senders)

🧪 Testing & Quality

• Much stricter golangci-lint configuration with an explicit allowlist and new quality/security linters enabled. Highlights include gosec, govet with fieldalignment, depguard, gochecknoglobals, gochecknoinits, gocritic, revive, errorlint, nilerr/nilnil, copyloopvar, and tparallel. This tightened bar took focused cleanup work and materially improves correctness, security, and maintainability.
• Examples and e2e fixtures updated for chained literals/consts.
• Beyond the removals, this release adds extensive tests, fixes long‑standing bugs, and refactors complex compiler/runtime paths. It’s been nearly two months of hard work since v0.33 (Dec 8), and we’re still on track for a “Neva release every month” cadence.

🧩 Misc Fixes & Refactors

• Stdlib extraction is now checksum-based. The embedded stdlib is hashed and compared against a .checksum file under ~/neva/std. If it matches, the existing on-disk stdlib is reused; if it differs, the stdlib is fully re-extracted and the checksum updated. This prevents stale stdlib copies after upgrades while keeping fast startup when nothing changed.
• Docs updated for new grammar and language simplification.

📑 Related Issues & PRs

Issues: #963, #969, #797

Pull requests: #951, #953, #955, #956, #958, #974, #978, #981, #991, #992, #993, #994, #995, #997, #998, #1005, #1006, #1009

Full Changelog: v0.33.0...v0.34.0