README_ZENODO.md

Zenodo publication instructions

This repository is prepared for an open Publication / Preprint deposit. Machine-readable record metadata is in .zenodo.json; citation metadata is in CITATION.cff.

Record metadata

• Title: Dissecting Repository-Scale Code-Agent Harnesses: Retrieval, Context, and Action Interfaces Under Model-in-the-Loop Evaluation
• Creator: Mandeep Sidhu (Independent Researcher)
• Resource type: Publication / Preprint
• Access: Open
• Paper and derived-data license: CC BY 4.0
• Software license: MIT
• Version: 1.0.0-preprint
• Reserved DOI: 10.5281/zenodo.21781711
• Keywords: code agents; repository-scale code editing; retrieval-augmented generation; LLM evaluation; software engineering; agent harnesses; reproducibility

Suggested description:

A five-study controlled evaluation of repository-navigation and editing harnesses for local LLM coding agents. The release reports 5,453 audited experimental cells across three public repositories and three local models. Study 5 contributes 2,826 model-in-the-loop cells covering lexical, syntax, and dense retrieval components; retrieval-by-action interactions; graph, query, tool, and packing ablations; and a 17-task held-out validation. No universal harness winner is claimed: quality ranks transfer weakly, token-cost ranks transfer strongly, and only one held-out cell resolves.

Files to upload

Primary record:

1. output/pdf/dissecting_repository_scale_code_agent_harnesses.pdf
2. output/releases/agent-harness-publication-package.tar.gz
3. output/releases/agent-harness-publication-package.tar.gz.sha256

Upload raw trajectories as a separate compressed file or a linked versioned deposit. Include an archive-level SHA-256 manifest. Do not upload local model weights, repository checkouts, embedding caches, or unrelated machine data.

The compact archive includes source, frozen configurations and task manifests, derived cell-level evidence, preregistrations, preflight records, manuscript source/PDF, licenses, citation metadata, and deterministic checksums. It deliberately excludes raw trajectories and heavyweight local dependencies.

DOI and publication sequence

The reserved DOI is 10.5281/zenodo.21781711. It is embedded in the manuscript, repository landing page, and citation metadata. Zenodo registers the DOI when the deposit is published, so the DOI URL may not resolve while the record is still a draft.

1. Confirm that the reserved DOI shown in the Zenodo draft is exactly 10.5281/zenodo.21781711.
2. Commit the DOI-bearing metadata and create a signed or annotated Git tag such as v1.0.0-preprint.
3. Rebuild the PDF and deterministic publication archive from the tagged revision; update every checksum.
4. Upload the three primary files and, if ready, the separately checksummed raw archive.
5. Review Zenodo's rendered PDF, creator order, title, abstract, licenses, keywords, file checksums, and DOI before clicking Publish.

Once published, treat changed files or conclusions as a new Zenodo version, not a silent replacement of the archived snapshot.

Final release gate

• The repository contains LICENSE, LICENSE-DATA, CITATION.cff, .zenodo.json, REPRODUCING.md, and THIRD_PARTY_NOTICES.md.
• The author is no longer anonymous in the manuscript or PDF metadata.
• Exactly one canonical paper PDF is present.
• The full test suite and both paper-consistency checks pass.
• LaTeX has no undefined citation/reference or overfull-box warnings.
• Every rendered PDF page has been visually inspected.
• The compact archive builds twice to the same SHA-256 from the release commit.
• Secret scanning and executable-source portability checks pass.