当前迁移状态： 本仓库正收敛 Volto、Swrm 和连接层。源码导入与构建检查不代表正式部署或真实任务验收完成；最新状态见 HANDOFF.md。
volto/：Volto Hub 和界面。swrm/：执行器、路由与跨主机并发。evalid/：运行验证与证据封存。integrations/：Hub 与执行网络的连接入口。- Issue、PR 和 Discussions 以本仓库为入口；当前执行看板为 phanes RSI 异构 Agent 网络（SSH fleet + GitHub issue bus）。既有 Phanes / Agent Task Project 保留历史与跨项目视图。
组件保留各自许可证：Volto 使用 PolyForm Noncommercial 1.0.0，Swrm 使用 Apache 2.0。其他文件适用其自身声明。
A unified agent network: one control plane, distributed heterogeneous execution, unified resource management, and evidence-based self-improvement.
Phanes brings three existing systems together into a single network for running fleets of AI coding agents:
- Volto is the control plane — user entry point, agent identity, goals and tasks, authorization, and the final accept/reject decision.
- swrm is the execution plane — resource registry, scheduling across machines and model backends, run/attempt state, cancellation and recovery, and run evidence.
- RSI (agent-knowledge) is the learning layer — versioned experience, reusable patterns, evaluation rules, and improvement proposals distilled from completed work.
GitHub sits alongside them as the evidence surface for code work: issues, pull requests, reviews, and releases that reference run records.
Status: The design has passed a multi-model adversarial review and is being built in stages (see Roadmap). This repository contains the component sources and integration entry points as the consolidation proceeds; deployment and end-to-end acceptance remain in progress.
Everything in Phanes flows through one versioned contract:
Goal → Task → Run / Attempt → Evidence
- A Goal states intent and acceptance criteria. A Task is a unit of work derived from it.
- Each execution is a Run made of Attempts, addressed by a stable launch key end to end. A transport retry reuses the same key; the same key with different content is rejected.
- Every attempt produces Evidence: outputs, verification results, and review records bound to the exact candidate (commit SHA or artifact digest).
- Final acceptance belongs to the control plane and must cite independent verification. A worker's self-report is never an acceptance basis.
Resources (hosts, backends, models) are described separately from agents: capabilities, repository mappings, permissions, concurrency limits, and health. Credentials are stored by reference only.
- One source of truth per concern. Run facts live in the execution plane; acceptance decisions live in the control plane. No component keeps a second copy of another's truth.
- Authority is enforced by real credentials and isolation, not by prompts. Workers run under dedicated identities in isolated workspaces, with no access to production, deployment, or scheduling credentials.
- No silent recovery. A run in an unknown state keeps its slot, raises an alert, and waits for reconciliation or a human decision — it is never auto-released or blindly retried.
- Deduplication survives reconnects. Dispatch intent is persisted before a start command is sent; a lost receipt triggers a query, not a duplicate launch. Stale attempts must be confirmed terminated before a retry or reassignment.
- Self-improvement is evidence-driven and gated. RSI proposes experience and policy changes; weighted routing stays record-only until controlled comparisons on the same task type and environment justify enabling it, and every change can roll back.
| Stage | Focus |
|---|---|
| A | Stable foundation: hub upgrade and acceptance, contract freeze, integration of audited fixes in risk order |
| B | Static unified resource pool: single-writer scheduler with a transactional dispatch log, quota grouped by credential reference |
| C | Reliable dispatch and a real closed loop: atomic launch registration, persistent termination marks, independent verification before acceptance |
| D | Networking and A2A migration: mesh connectivity pilot, capability-by-capability migration with a rollback path for each |
| E | Evidence-driven self-evolution: regression cards, complete outcome/cost records, gated policy changes with small-scope trials |
Formal release of each stage requires passing negative tests (wrong-host receipts, forbidden data access, duplicate and out-of-order results) and a full trace from every accepted result back to its task, base, candidate, verification, review, and policy version.