v1.26.0 — Agentic Workflows and LLM Core Microservice

Released 2026-04-08. GitHub release.

Agentic Workflows

v1.26.0 introduces a first-class workflow engine for orchestrating multi-step agentic processes across microservices.

workflow Package

The new workflow package provides the primitives for defining and executing workflows:

• Graph — A directed graph of tasks and transitions. Tasks are added by URL, transitions connect them sequentially or conditionally. Conditional transitions use state expressions (e.g. toolsRequested==true). Fan-out over arrays is supported via AddTransitionForEach. Subgraph invocation enables workflow composition. Graphs self-validate and can render as Mermaid diagrams.
• Flow — The state carrier passed to each task. Provides typed getters/setters (GetString, SetInt, etc.) for reading and writing shared state. Tasks can issue control signals: Goto (jump to a task), Retry (with exponential backoff), Sleep (delay execution), Interrupt (pause for external input), and Subgraph (delegate to another workflow).
• Reducers — Define how concurrent state modifications from parallel fan-out tasks are merged during fan-in: replace (last write wins, the default), append (concatenate arrays), add (sum numeric values), and union (merge arrays with deduplication).

Foreman Core Microservice

The foreman.core microservice orchestrates workflow execution. It manages a sharded SQL-backed persistence layer and a configurable worker pool for concurrent step processing.

Key endpoints include: Create, Start / StartNotify, Run, Snapshot, History, Resume, Fork, Continue, Retry, Cancel, List, Await, BreakBefore, and CreateTask.

The Foreman emits an OnFlowStopped event when a flow reaches a terminal status (completed, failed, cancelled, or interrupted). It tracks metrics for flows started, flows terminated, steps executed, queue depth, and steps recovered after lease expiry. A configurable retention policy purges old flows automatically.

Task and Workflow Features for Microservices

Any microservice can now define task endpoints and workflow graphs as first-class features alongside functions, web handlers, and events.

• Tasks are functional endpoints on port :428 that receive a *workflow.Flow and operate on shared state. Typed input arguments are automatically bound from flow state, and output arguments are written back.
• Workflows are graph-definition endpoints on port :428 that return a *workflow.Graph describing the task topology, transitions, and reducers.

New agent skills microbus/add-task and microbus/add-workflow scaffold these features.

LLM Core Microservice

The llm.core microservice bridges LLM tool-calling protocols with Microbus endpoints. It provides a Chat endpoint that manages multi-turn conversations with automatic tool resolution and execution.

The LLM service uses a pluggable provider architecture. The active provider is selected via the ProviderHostname configuration property, defaulting to claude.llm.core. Three provider microservices ship out of the box:

Each provider translates the common Turn interface to its vendor-specific API via the HTTP egress proxy. Tools are resolved from OpenAPI specs of Microbus endpoints, allowing LLMs to call any microservice function as a tool.

Internally, the LLM service is itself implemented as an agentic workflow (ChatLoop) with tasks for initialization, LLM calls, response processing, and tool execution — demonstrating the workflow engine in a real use case.

Restructured Agent Coding Rules

The agent coding rules in .claude/rules/ have been restructured to compress their size within LLM context windows. Topic-specific guidance such as workflows.txt and sequel.txt is separated into dedicated files, loaded on demand only when relevant to the microservice being modified.