Hymenium: An AI Coding Agent Workflow Orchestration Engine Written in Rust

Hymenium is an AI coding agent workflow orchestration engine written in Rust, created to solve the problem of process control relying on an "honor system" in multi-agent collaboration. It automates the implementer/reviewer pattern, phase gates, and retry mechanisms, converting textual agreements into workflow rules enforced by the engine. It is an important component of the Basidiocarp ecosystem.

As the capabilities of AI coding agents improve, multi-agent collaboration to complete complex tasks has become a trend. However, current collaboration relies on document-driven instructions like CLAUDE.md or AGENTS.md, where agents easily misinterpret or skip steps. Phase transitions depend on an "honor system", manual orchestration is hard to scale to more than two concurrent handovers, and human intervention is needed when tasks get stuck. The core problem lies in the lack of a real orchestration engine to enforce workflow rules.

Hymenium is positioned as a coordination layer (not an underlying execution environment) and is part of the Basidiocarp ecosystem. Its core capabilities include: 1. Handover document parsing: Reads structured Markdown and rejects invalid documents at the entry point; 2. Automatic decomposition: Splits large handovers into focused sub-handovers; 3. Workflow templates: Declarative templates that allow adding new types without modifying the engine; 4. Phase gates: Enforces prerequisites for transitions (e.g., reviews require code diffs and verification results from implementers); 5. Progress monitoring: Polls status and escalates to operators on timeout; 6. Retry and recovery: Closes stuck agents on heartbeat timeout and restarts with a reduced scope.

Hymenium has a built-in implementer/reviewer template flow: Handover document → Parsing and decomposition → Create implementer task → Poll for completion (including diff and verification) → Gate check → Create reviewer task → Poll for approval → Close task; if stuck, it automatically closes the agent and restarts with a reduced scope. This flow ensures reviews do not start early, implementers produce valid results, and the system recovers automatically without manual monitoring.

Hymenium belongs to the Basidiocarp ecosystem, whose components include annulus (cross-ecosystem operations), canopy (multi-agent coordination ledger), cap (Web dashboard), etc., each with clear responsibilities. The advantages of choosing Rust as the technology: Memory safety (eliminates errors like null pointers), performance (zero-cost abstractions with low overhead), concurrency (safely handles multiple workflows), reliability (catches issues at compile time).

Applicable scenarios: Automated code review (assign agents for review in CI/CD), complex feature development (decompose into subtasks for parallel execution), test generation (implementers write code + reviewers generate tests), document generation (multi-agent collaboration to maintain technical documents). CLI commands: dispatch (schedule handover), status (check workflow status), cancel (cancel workflow), reconcile (coordinate phases), decompose (decompose handover for implementation).

Hymenium promotes AI coding agent collaboration from relying on agent self-discipline via textual agreements to deterministic workflows enforced by the engine. Through mechanisms like phase gates and automatic recovery, it turns multi-agent collaboration from an experimental technology into a production-ready solution. As the Basidiocarp ecosystem improves, more innovative applications will emerge, pushing the boundaries of AI-assisted development. Project address: https://github.com/basidiocarp/hymenium