open-prose

OpenProse is a reusable multi-agent orchestration framework that an AI agent embodies directly rather than shelling out to a binary. It activates when a user types 'prose ...', opens a .prose.md file with kind frontmatter, opens a .prose file, or asks for reusable multi-agent orchestration. It solves the problem of authoring, wiring, and executing durable, inspectable multi-step agent workflows: the agent reads a Markdown contract, selects a state backend, wires responsibilities, executes with host primitives, and persists run state under a selected OpenProse root. It is deliberately scoped to decline one-shot questions where a plain prompt would do.

The framework has five load-bearing pieces: Contract Markdown (the human-readable .prose.md source format), Forme (a semantic dependency-injection container that wires contracts), the Prose VM (the execution engine that runs responsibilities, functions, and pinned execution blocks), ProseScript (an imperative scripting layer for choreography such as order, loops, conditionals, parallelism, and retries), and the Responsibility Runtime (a responsibility-oriented architecture with standing goals, a Reactor, and compile/serve doctrine). A detailed 'first 90 seconds' routing table directs the agent to load only the narrowest set of reference docs for the task at hand, whether explaining OpenProse, running a responsibility, authoring new sources, writing choreography, compiling a typed-image brief, or standing up a Reactor. Execution state is persisted as inspectable files under an OpenProse root, using a published-vs-workspace discipline where canonical world-model artifacts are deterministically serialized and content-addressed, with an append-only receipt ledger enabling inspection, resumption, debugging, and auditability. Many worked examples ship with the skill.

It targets developers and power users building reusable, auditable multi-agent systems and standing automations. As an orchestration VM it executes workflows with host primitives and manages state on the local filesystem; the documented model emphasizes inspectability, deterministic serialization, and auditable receipts rather than any destructive, remote-execution, or exfiltration behavior.