AION: Next-Generation Tasks and Practical Harness for Time Series

Tianxiang Zhan1, Xiaobao Song2,1, Tong Guan3,1, Shirui Pan1, Ming Jin1
1Griffith University    2Shenzhen University    3Zhejiang University
arXiv preprint, 2026
Paper Code arXiv Docs

Quick Start

1. Install OpenCode

curl -fsSL https://opencode.ai/install | bash

2. Install AION Plugin (one command)

curl -fsSL https://raw.githubusercontent.com/ztxtech/aion/dev/scripts/install.sh | bash

This installs the aion-ts CLI globally. Then install AION into any project:

cd your-project
aion-ts init .

3. Run

opencode

OpenCode auto-discovers the plugin. On the first turn, AION asks about language and interactive mode, then dispatches agents.

AION Workflow

Finite-state execution with iterative refinement and dual-review gates

%%{init: {'theme': 'base', 'flowchart': { 'useMaxWidth': true, 'htmlLabels': false, 'curve': 'basis', 'nodeSpacing': 80, 'rankSpacing': 90, 'padding': 20 }, 'themeVariables': { 'primaryColor': '#ffffff', 'primaryTextColor': '#1e293b', 'primaryBorderColor': '#7C3AED', 'lineColor': '#94a3b8', 'secondaryColor': '#f8fafc', 'tertiaryColor': '#f1f5f9', 'fontFamily': 'Inter, sans-serif', 'fontSize': '15px' }}}%%
flowchart LR
    classDef default fill:#ffffff,stroke:#7C3AED,stroke-width:2px,color:#1e293b,rx:6px,ry:6px
    classDef task fill:#ffffff,stroke:#7C3AED,stroke-width:3px,color:#1e293b
    classDef review fill:#fffbeb,stroke:#f59e0b,stroke-width:3px,color:#1e293b
    classDef pass fill:#ecfdf5,stroke:#10b981,stroke-width:2px,color:#047857
    classDef fail fill:#fef2f2,stroke:#ef4444,stroke-width:2px,color:#dc2626
    classDef rebuttal fill:#fff7ed,stroke:#f97316,stroke-width:2px,color:#c2410c
    classDef loop fill:#f8fafc,stroke:#8B5CF6,stroke-width:2px,color:#7C3AED,stroke-dasharray: 5 5

    A["01 Task"] --> B["02 Info"]
    B --> C["03 Run"]
    C --> D{"04 Review"}
    
    D -->|PASS| E["Done"]
    D -->|FAIL| F["Fix"]
    
    F --> G["Loop"]
    G -.-> B
    G -.-> C
    
    class A,B,C task
    class D review
    class E pass
    class F fail
    class G loop
    
    style A width:90px
    style B width:90px
    style C width:90px
    style D width:120px
    style E width:80px
    style F width:80px
    style G width:80px

The AION workflow operates as a finite-state machine with mandatory dual-review gates. After Task Analysis and Information collection, the coder runs experiments. Every output must pass both ts-critic (time-series expert + Pareto governance) and c-critic (final minimal-context cold-start critique). On Pass, the system finalizes with memory sync and trace logging. On Fail, the rebuttal protocol triggers, and the system loops back to gather more evidence or refine the experiment. No output leaves the system without passing layered critics.

Four-Layer Architecture

Task Layer
requirements-analyst Task specification & requirement extraction
Workspace Layer
information-collector 17 skills Evidence collection, tool orchestration
Execution Layer
agent coder 8 protocols, governance hierarchy
Review Layer
ts-critic c-critic Validity, temporal & completeness checks

Each layer constrains the operations of the one below it. The Task layer defines what to solve and extracts requirements. The Workspace layer manages evidence and tools available for the task. The Execution layer governs how the system acts under constraints. The Review layer validates if progress and outputs are acceptable before any output leaves the system.

Key Components

6 Specialized Agents

Main orchestrator, requirements analyst, information collector, coder, ts-critic, and c-critic with strict governance hierarchy.

17 Reusable Skills

From time-series analysis and forecasting to experiment tracking, report writing, and GitHub evidence search.

8 Runtime Protocols

Dispatch, reportback, rebuttal, stop-go, lifecycle, memory-sync, runtime-events, and compaction protocols.

5 Evaluation Contracts

Test suites, graders, scorecards, regression matrices, and release gates for quality assurance.

Memory & Trace

Per-task execution traces and cross-task persistent memory with context snapshots and decision logs.

Built on OpenCode

Extends the OpenCode AI coding agent with time-series-specific governance, skills, and safety gates.

34 Tools (20 AION + 14 Team)

Experiment, governance, memory, safety, plan, session, Hugging Face, and team-coordination primitives — all auto-traced, all return JSON.

Hugging Face Integration

Zero-dependency HF Hub REST tools: search, info, ingest, suggest. 24h cache. CLI mirror (aion-ts datasets) for offline prep.

Ablation & Statistical Rigor

HARD GATES: ablation is the SOLE arbiter of "best method". Beyond p-value: SHAP, residual diagnosis, drift analysis, and sensitivity analysis are all mandatory.

Governance Hierarchy

c-critic Final minimal-context cold-start critique — ultimate authority
ts-critic Time-series expert + Pareto governance — highest gate
Main Agent Orchestration and dispatch — no closeout authority above critics
Other Subagents Requirements analyst, information collector, coder

The main agent owns dispatch and execution, but does not own closeout authority above the critics. In blocker judgment, rebuttal verdicts, route rollback, stop-go, and final-delivery decisions: c-critic > ts-critic > main agent > other subagents.

BibTeX

@misc{zhan2026aion,
  title={AION: Next-Generation Tasks and Practical Harness for Time Series},
  author={Tianxiang Zhan and Xiaobao Song and Tong Guan and Shirui Pan and Ming Jin},
  year={2026},
  eprint={2605.25045},
  archivePrefix={arXiv},
  url={https://arxiv.org/abs/2605.25045}
}