Documentation
A complete reference for AetherGit — architecture, formal definitions, caching topology, and validation
Published research. AetherGit is at v1.0.0. The manuscript is published on Zenodo. No empirical work has been performed.
Table of Contents
Overview
AetherGit is a high-throughput, always-online version control framework designed specifically for agentic software engineering. By decoupling incremental changesets from recursively computed derived metadata and enforcing strict client-server asymmetry, AetherGit eliminates repository packfiles, lock-heavy merge queues, and full history fetching.
| Field | Value |
|---|---|
| Full title | AetherGit: Scalable Source Control for Autonomous AI Agents |
| Version | 1.0.0 |
| Stage | Formalization · Research Paper |
| License | MIT (code) · CC-BY-4.0 (docs, manuscript) |
| Domain | Computer Science / Software Engineering |
| DOI | 10.5281/zenodo.23966896 |
"Adapting version control infrastructure to the reality of autonomous AI agents requires a fundamental paradigm shift away from decentralized, offline-first architectures."
Formal Definitions
Definition 1 — Repository State Space. Let the repository history be represented as a Directed Acyclic Graph (DAG) G = (V, E), where V = {c₀, c₁, c₂, …} corresponds to the set of immutable changesets, and E ⊂ V × V defines explicit parent-child derivation dependencies.
Unlike traditional VCS where a node encapsulates a full pointer structure to an entire file tree, a changeset in AetherGit is strictly defined as a delta-tuple over content-addressed storage blobs:
cᵢ = ⟨P(cᵢ), Δᵢ, tᵢ, sigᵢ⟩
Where:
P(cᵢ) = {p₁, p₂, …, pₖ} ⊂ V= ordered set of parent changesetstᵢ= monotonic logical timestampsigᵢ= agent cryptographically signed identityΔᵢ = {(path, hash(B)) | path ∈ Ω, B ∈ Blobs}= localized modification delta set
Definition 2 — Recursive Derived State. For any derived data domain D, the state D(cᵢ) associated with changeset cᵢ is uniquely specified by the recursive composition function Φ_D over its immediate parent set:
D(cᵢ) = Φ_D({D_p | p ∈ P(cᵢ)}, Δᵢ)
Definition 3 — Cache Hit Probability. Given an always-online client topology featuring a tiered hierarchy of regional edge caches (C₁), node daemons (C₂), and authoritative storage (S), the probability P(Hit) of resolving tree objects without querying the authoritative storage tier is bounded by:
P(Hit) = 1 − ∏_{k=1}^{N}(1 − h_k) ≥ 1 − ε
Because content addressing enforces strict immutability:
∂ CacheState / ∂ Time = 0 (for all materialized objects)
Architectural Pillars
| Pillar | Description | Status |
|---|---|---|
| Client-Server Asymmetry | Server performs all history calculations; client is bounded cache. | Specified |
| Sparse & Virtualized Checkouts | On-demand file materialization; zero disk I/O for unread directories. | Specified |
| Lock-Free Changesets | Distributed WAL + CAS primitives; no global reference locks. | Specified |
| Tiered Cache Topology | Node daemons + regional caches + authoritative storage. | Specified |
Caching Topology
AetherGit employs a three-tier hierarchy:
| Tier | Component | Role |
|---|---|---|
| C₁ | Node-Local Daemon | Local cache & disk layout |
| C₂ | Regional Cache | Tiered in-memory cache |
| S | Authoritative Server | Spanner DB / Distributed Blob Store |
The probability of resolving tree objects without querying the authoritative storage tier is bounded by P(Hit) = 1 − ∏(1 − h_k) ≥ 1 − ε. Because content addressing enforces strict immutability, cache invalidation overhead is strictly zero.
Validation
| Check | Result |
|---|---|
| Formal Definitions | 3 (all defined) |
| Architectural Pillars | 4 (all specified) |
| Performance Theorems | 2 (O(|Δ|), P(Hit) ≥ 1−ε) |
| Open Research Questions | 2 (OQ1, OQ2) |
| Empirical Evaluation | Not executed |
| Causal Claims | None (formal analysis only) |
Scope & Limits
What AetherGit supports:
- Formal repository state space (DAG topology).
- Recursive derived state calculus.
- Probabilistic cache hit bounds.
- Four architectural pillars.
- Open manuscript (CC-BY-4.0) and open code (MIT).
What AetherGit does not support:
- Empirical benchmarking against hyper-scale monorepos (not executed).
- Causal claims (formal analysis only).
- Human evaluation.
Design philosophy. AetherGit is a research paper, not a predictive model or a policy tool. Its purpose is to specify an architecture precise enough to be tested against actual agentic workloads — not to make forecasts or normative claims.
Citation
APA: Baladi, S. (2026). AetherGit: Scalable Source Control for Autonomous AI Agents (Version 1.0.0). Zenodo. https://doi.org/10.5281/zenodo.23966896
@techreport{baladi2026aethergit,
author = {Baladi, Samir},
title = {AetherGit: Scalable Source Control for Autonomous AI Agents},
institution = {Ronin Institute / Rite of Renaissance},
type = {Research Paper},
year = {2026},
month = {November},
version = {1.0.0},
doi = {10.5281/zenodo.23966896},
url = {https://doi.org/10.5281/zenodo.23966896}
}