Neuron Redesign (vNext) — Release Notes

This release introduces a composable, layered runtime with stable protocol boundaries, execution primitives for planner/steering/streaming, and a split effect executor. It is a breaking redesign intended to make multi‑agent systems predictable, testable, and durable‑ready.

Highlights

Layer 0 stability preserved: object‑safe traits (Operator, Orchestrator, StateStore, Environment, Hook) + serde wire types
New turn engine primitives (neuron-turn-kit): ToolExecutionPlanner, ConcurrencyDecider, BatchExecutor (execution‑only), SteeringSource
New effect executor split:
- neuron-effects-core — trait + error + policy
- neuron-effects-local — in‑process effect execution (in‑order, best‑effort)
React Operator improvements (opt‑in; defaults unchanged):
- BarrierPlanner (Rho‑style Shared/Exclusive batching)
- SteeringSource (operator‑initiated mid‑loop injection + skip placeholders)
- Streaming tools (ToolDynStreaming) + HookPoint::ToolExecutionUpdate (read‑only)
- Tool‑level concurrency hints (ToolConcurrencyHint) with metadata‑based decider
Local orchestrator: minimal signal/query semantics + reference effect execution
Umbrella crate (neuron) and example custom operator included
CI: fmt, clippy ‑D warnings, tests, coverage, audit, deny, docs

New crates

neuron-turn-kit — turn engine primitives
neuron-effects-core — effect executor trait
neuron-effects-local — local effect executor impl

Backward compatibility

Old runtime/durable context is superseded by Orchestrator + Effect executor separation. Operators remain pure and durable‑agnostic.
neuron-orch-kit re‑exports the effect executor crates for back‑compat paths.
neuron-op-react now depends on neuron-turn-kit but preserves public API and default behavior.

Migration (for downstreams)

Prefer the umbrella crate with features: neuron = { version = "x.y.z", features = ["op-react", "provider-anthropic", "orch-local", "state-fs", "hooks"] }
If using reaction loop directly:
- Swap to neuron-turn-kit traits for planning/steering; defaults remain sequential/no‑steering
- Adopt ToolDynStreaming + HookPoint::ToolExecutionUpdate to observe tool chunk updates
For effect execution: use neuron-effects-local or integrate a durable executor in your orchestrator
See MIGRATION.md for full details

Deprecations

Legacy durable context patterns are deprecated; replace with Orchestrator implementations (Temporal/Restate) and a durable effect executor

Known limitations

Durable orchestrators (Temporal/Restate) are not included in this release; wiring guidance provided in docs

Acknowledgements

Thanks to the Rho integration work for motivating turn-kit and the effect split.

Assets 2

24 Feb 02:24

SecBear

neuron-v0.3.0

30097c1

neuron v0.3.0

What's New in v0.3.0

This release adds four Pydantic AI-inspired features, a new OpenTelemetry crate, and comprehensive CI improvements.

New Features

Usage Limits (UsageLimits) — enforce token, request, and tool-call budgets with configurable limits checked at three points per loop iteration, including streaming
Tool Timeout Middleware (TimeoutMiddleware) — per-tool timeout enforcement via the ToolMiddleware pipeline using tokio::time::timeout
Structured Output Validation (StructuredOutputValidator, RetryLimitedValidator) — JSON Schema validation of model output with ToolError::ModelRetry self-correction loop
OpenTelemetry Instrumentation (neuron-otel crate) — OtelHook implementing ObservabilityHook with tracing spans following GenAI semantic conventions (gen_ai.chat, gen_ai.execute_tool, etc.)

Bug Fixes

Fixed run_stream() missing usage limit enforcement entirely (3 checkpoints added)
Fixed workspace dependency versions for correct crates.io resolution

New Crate

neuron-otel — OpenTelemetry instrumentation with opt-in content capture, span hierarchy, and privacy-safe defaults

CI/CD

Automated cargo publish via release-please workflow
Manual publish workflow (workflow_dispatch) for bootstrap and fallback
Combined release PRs (no more 12 separate PRs per release)

All Published Crates

Crate	Version	crates.io
`neuron-types`	0.3.0	crates.io
`neuron-tool-macros`	0.2.1	crates.io
`neuron-tool`	0.3.0	crates.io
`neuron-context`	0.3.0	crates.io
`neuron-loop`	0.3.0	crates.io
`neuron-provider-anthropic`	0.3.0	crates.io
`neuron-provider-openai`	0.3.0	crates.io
`neuron-provider-ollama`	0.3.0	crates.io
`neuron-mcp`	0.3.0	crates.io
`neuron-runtime`	0.3.0	crates.io
`neuron-otel`	0.3.0	crates.io
`neuron`	0.3.0	crates.io