v0.5.0 — M5: the learned control plane (AI manages AI)

M5 is complete. THALIOX now governs itself with a learned, falsifiable, fully in-system control plane (RFC-0007) — "AI manages AI" as a running loop, not a slogan.

The four stages

• M5a — the closed loop (runtime::control): observe the cluster as a fixed-width state vector → a swappable Policy (transparent heuristic baseline) → actuate only through M1–M4's invariant-guarded mechanisms (self_heal / migrate / budget grants). The plane invents no new way to touch an agent.
• M5b — the governor is an agent (Governor): it thinks (spends budget, INV-1 — starved ⇒ governs nothing), acts under capability (INV-2 — out-of-scope ⇒ denied), is audited (INV-4), with Shadow / Canary / Act modes gated in-system. No human floor, no master key (INV-5 self-sovereignty).
• M5c — the learned policy π_θ + falsification gate E5 (runtime::learn): trained in a deterministic cluster simulator seeded from replayed audit traces (the INV-4 ledger is the dataset); invariants are action-space masks, never reward terms; reward = budget-efficiency under a hard survival floor. π_θ may not actuate until it strictly beats the heuristic baseline on a held-out suite with zero violations — then it is promoted Shadow → Canary → Act and auto-demoted on any regression, no human on any rung.
• M5d — self-optimization: the refill becomes a learned, graded adaptive-compute knob (50/100/200 tokens, priced by per-actuation overhead — the first concrete F10 step), and the self-update verdict (promote or roll back a staged candidate generation, real update.rs mechanism underneath) is decided from observed post-update yield instead of a hand-set threshold.

The gate, passed in CI

On the held-out suite π_θ scores 0.719 vs the heuristic's 0.631 mean budget-efficiency, zero invariant violations, full survival; on the update scenario it concludes all 8 hidden-quality candidates (4 promoted, 4 rolled back) while the verdict-less baseline leaves them dangling (0.772 vs 0.612).

All four gates green: fmt · clippy -D warnings · test (127) · doc -D warnings.

Next: M6 — H2 begins, pushing the stack down.

v0.4.0 — M4 cluster + multi-platform

THALIOX is now an operating system for a distributed society of agents. This release completes the H1 software arc through M4.

Highlights since v0.1.0

• M2 microVM-ization — one-command deploy + snapshot/restore + self-update rollback; the agent runs inside a real Firecracker microVM (vsock deploy, VM snapshot/restore), validated on KVM bare-metal. (RFC-0004)
• M3 multi-instance HA — per-field CRDT merge, Node + migrate, and a Supervisor (heartbeat → self-heal → reconcile). (RFC-0005)
• M4 cluster + multi-platform (RFC-0006):
  • fabric — VectorMessage transport between agents (in-process) and across nodes (TCP), INV-2/INV-3 enforced at the boundary.
  • cross-host live migration validated on two KVM machines — at both process-level and microVM-level, covering the full {VM, host-process} × {VM, host-process} migration matrix.
  • teams in four paradigms — Pipeline / Hierarchy / Market / Swarm — with INV-2 on every cross-member hop and INV-3 on every agent's input.
  • cluster front door — the api gateway generalized: one capability-based authorization model (INV-2 at the door), an SSE streaming surface, and peer routing for a fleet across nodes.

INV-2 and INV-3 hold between agents and at the cluster door — the team/cluster boundary is not a hole in the invariants.

All four CI gates green: fmt · clippy -D warnings · test (103) · doc -D warnings.

Next: M5 learned control plane (RL scheduling + a supervisor agent + self-optimization — "AI manages AI").

Full roadmap: docs/MASTER_PLAN.md · Cluster design: docs/rfcs/0006-cluster-multiplatform.md

v0.1.0 — M1 Single-Node MVP

THALIOX 的编程模型成立。 M1 把 RFC-0001 TAM 抽象机 从纸面契约变成可运行、可证伪的代码:一个单机 agent,在 TAM 五不变量约束下,能自主完成任务。

交付

• 认知 — 统一 LlmProvider:Anthropic Messages / OpenAI Chat Completions(及任意兼容网关)+ 离线 mock。
• 记忆 — SemanticSpace 向量记忆。
• 工具 + 自主闭环 — Agent::run(goal):模型自己决定调哪个工具(web_search / fetch)、执行、结果喂回、再思考。
• 注意力预算 — 预留→真实 token 对账(INV-1),失败退款。
• 能力门控 — 签名 + 过期 + scope 前置(INV-2),全程审计(INV-4)。
• API 网关 — axum HTTP:生命周期 + think / remember / recall / invoke + 审计。

实测(glm-5.1 + Tavily)

模型自主调 web_search → 真实搜索 → 一句话总结;审计 Think → ToolInvoke → Think,预算逐笔对账。

质量

四门全绿:fmt · clippy -D warnings · test(30) · doc -D warnings。8 crate · 6 example。

详见 docs/M1-MILESTONE.md。下一站 M2 microVM 化。