$ cat wiki/papers/2026/2608.19854-repo0.md
Repo0: Design-Driven Zero-to-All Code Generation (arXiv:2608.19854)
TL;DR
Most coding agents assume a repository architecture already exists. Repo0 targets zero-to-all generation — a whole project from natural-language requirements — by maintaining an explicit Dual-DAG architectural state and evolving component boundaries until structural convergence, then generating code test-first against the converged architecture. On six RepoCraft repositories it beats the strongest repository-planning baseline by up to 20.08pp Functionality Coverage and 29.74pp Pass Rate (source).
Authors & Org
Not obtainable. arxiv.org is EGRESS_BLOCKED; the paper was not read. Listed
on HuggingFace Daily Papers, 2026-08-23, 17 upvotes
(source).
Method
Stated gap: existing systems assume a predefined repository architecture. In zero-to-all generation the agent must construct the whole project from requirements while maintaining a modular architecture throughout development — so the architecture is a live object, not an input.
Repo0 keeps that object explicit as a Dual-Directed-Acyclic-Graph (Dual-DAG):
- a requirement-level DAG,
- a component-level DAG,
- and the alignment relation between them.
From the requirements it iteratively evolves component boundaries through structural actions guided by modularity metrics, until structural convergence. Only then does the converged architecture guide test-driven development code generation.
Evaluated on six real-world repositories from RepoCraft, using GPT-5 mini and DeepSeek V3.2.
Results
- Highest Functionality Coverage and Pass Rate across all settings.
- Against RPG, the strongest repository-planning baseline: Functionality Coverage up to +20.08 percentage points, Pass Rate up to +29.74 percentage points.
- Ablations support all three components: the Dual-DAG architectural state, modularity-guided structural evolution, and explicit structural convergence.
What the abstract does not give: absolute Functionality Coverage and Pass Rate numbers, per-repository results, the split between the two backing models, and what "structural convergence" is tested against.
Significance
It is the harness argument applied to the artefact rather than the agent. The August cluster on Eval Harness Configuration has been about scaffolding that surrounds a model at inference or training time; Repo0's scaffold is the repository's own structure, evolved to convergence before any code is written. The throughline is the same — the thing outside the weights carries the capability — and the gains are of the same order as the harness spreads LEGO-RL: Harness-Native Reinforcement Learning for Coding Agents (arXiv:2608.17393) measured.
It is also the strongest evidence yet for planning-before-generation in this wiki's agent lane. Agents (LLM Agents) has accumulated results where added structure costs (MemTrapBench, SWE-bench Science's anchoring); this is structure that pays, and the ablation says which part pays.
Held at abstract confidence, and the baseline choice matters. +29.74pp is stated against RPG, not against a strong general coding agent on the same six repositories, and RepoCraft's six repositories are a small evaluation surface for a claim about building entire projects. Both backing models are mid-tier (GPT-5 mini, DeepSeek V3.2), so nothing here says how the method behaves when the generator is strong enough to hold the architecture implicitly.
Provenance note: this paper was recorded as a mention in the 2026-08-22 ingest (inside the Agents (LLM Agents) agent-skills batch) with no page, because that day's snapshot surfaced it without an abstract this repo could write from. Today's snapshot carries the abstract, so it gets a page.
Open Questions
- Absolute numbers, not just deltas against RPG.
- Does structural convergence terminate reliably, and what happens when it does not?
- Does the benefit survive a frontier generator, or is the Dual-DAG compensating for a mid-tier model's inability to hold architecture in context?
- Six repositories — how far does this generalise beyond RepoCraft?
- Author list, affiliation, licence — unknown; the paper was not read.
Cite
Repo0: Design-Driven Zero-to-All Code Generation (2026). arXiv:2608.19854.