ASSERT-KTH/agent-commits-raw

AI Coding-Agent Commits on GitHub

This dataset documents commits associated with four AI coding agents: Claude, OpenAI Codex, GitHub Copilot and Cursor. It contains 1,853,915 commit records across 444,055 GitHub repositories, with 220,753 identifiable GitHub user accounts recorded as commit authors. Messages and author identities are in commits; repository metadata, file changes and patch text are available in separate tables.

Dataset

Agent	Commit records	Repositories	Author accounts
Claude	401,025	208,658	115,720
OpenAI Codex	45,020	6,379	2,089
GitHub Copilot	980,957	163,026	66,263
Cursor	425,955	79,540	43,164
Unlabeled	1,343	432	194
Overall, deduplicated	1,853,915	444,055	220,753

A commit record represents a commit within a repository. Across repositories, the dataset contains 1,691,910 distinct commit SHAs. Repository counts use GitHub IDs, with recorded names as a fallback where IDs are missing.

Author accounts are distinct GitHub author IDs with account type User. The count measures identifiable accounts; one person may use multiple accounts, and an account may be shared. Author IDs are missing from 174,367 commit records.

Each commit counts once per agent, regardless of identification channel. The 385 commits with multiple agent labels appear in each relevant row. Repositories and authors can also appear under several agents; the overall row counts each identity once across the full dataset, including unlabeled records.

Collection

GitHub searches targeted daily windows between June 2025 and May 2026. Identification used agent accounts, email addresses, co-author trailers and tool names in commit messages. The matched collection used ten-minute daily windows; the expanded collection used four-hour daily windows for Codex, Copilot and Cursor. Counts describe this corpus and reflect each agent's search exposure.

For example, this query illustrates a four-hour search window:

author-email:codex@openai.com author-date:2026-01-15T00:00:00Z..2026-01-15T04:00:00Z

It finds commits whose Git author email is codex@openai.com and whose author timestamp falls between 00:00 and 04:00 UTC on January 15, 2026. The four hours describe the date range covered by the search. Other channels search agent accounts or message text, such as co-author trailers.

Search results provide the repository, commit SHA, message and URL. The repository and SHA are then used to retrieve commit details and patch text from the GitHub API. Results from different searches are merged by repository identity and SHA, retaining each agent/channel pair in identifications. See GitHub's commit search syntax.

Start with commits for messages, identities and agent labels. Load repository details, file changes and patch text as needed.

Tables

Table	Count	Size	Contents
commits	1,853,915	782.3 MiB	One commit within a repository.
repositories	441,371	21.7 MiB	One repository metadata observation.
commit_files	17,826,229	855.2 MiB	One changed file in a commit response.
patches	1,650,139	14,065.9 MiB	One patch file with its commit ID and content hash.

Repository totals count distinct IDs, or names where IDs are unavailable. The repositories table contains metadata observations; a repository may appear more than once.

Each table consists of independently readable Parquet files. Select the tables and columns needed for your analysis. The configuration name is the table name, and the split is all.

Files

data/
  commits/          Messages, identities and paired agent/channel labels
  repositories/     Recorded repository metadata
  commit_files/     Changed paths, change types and line counts
  patches/          Patch text and content hashes
raw/                Original API content
indexes/            Record locations and source-file references
manifests/          File inventories and checksums
README.md
DATA_DICTIONARY.md
PROVENANCE.md
LICENSES.md
SHA256SUMS

Read commit metadata

from itertools import islice
from datasets import load_dataset
from huggingface_hub import HfApi

REPO = "ASSERT-KTH/agent-commits-raw"
REVISION = HfApi().dataset_info(REPO).sha
commits = load_dataset(
    REPO, "commits", split="all", revision=REVISION,
    streaming=True,
    columns=["commit_id", "repo_full_name", "commit_sha",
             "commit_url", "message", "identifications"],
)
for commit in islice(commits, 3):
    print(commit)

Save REVISION with your analysis. identifications keeps each agent and channel together. message includes the full text and co-author trailers.

Find a commit and its patches

Set a recorded repository name and the full commit SHA. The small SHA-prefix index locates the metadata file and row group. Read the commit metadata first, then load its patches as needed.

import hashlib
import pyarrow.parquet as pq
from huggingface_hub import HfApi, hf_hub_download

REPO = "ASSERT-KTH/agent-commits-raw"
REVISION = HfApi().dataset_info(REPO).sha
target_repo = "Tkute0401/Urban-Realty"
target_sha = "8b0c1a43d2e80ad40929d4e949b7d49f6648cfa2"

def local(path):
    return hf_hub_download(REPO, path, repo_type="dataset", revision=REVISION)

prefix = target_sha[:2]
locations = pq.read_table(local(f"indexes/commits/{prefix}.parquet"),
    filters=[("commit_sha", "=", target_sha),
             ("repo_full_name", "=", target_repo.casefold())]).to_pylist()
if len(locations) != 1:
    raise ValueError("Select a unique repository identity from the matching records")
location = locations[0]
rows = pq.ParquetFile(local(location["commit_file"])).read_row_group(
    location["row_group"]).to_pylist()
commit = next(r for r in rows if r["commit_id"] == location["commit_id"])
print(commit["message"])

patches = pq.read_table(local(f"indexes/patches/{prefix}.parquet"),
    filters=[("commit_id", "=", commit["commit_id"])]).to_pylist()
for item in patches:
    rows = pq.ParquetFile(local(item["patch_file"])).read_row_group(
        item["row_group"]).to_pylist()
    patch = next(r for r in rows if r["patch_id"] == item["patch_id"])
    payload = patch["patch"].encode("utf-8")
    assert len(payload) == patch["bytes"]
    assert hashlib.sha256(payload).hexdigest() == patch["sha256"]
    print(patch["patch"][:500])

Join the tables

• commits.repository_observation_id joins to the same field in repositories, where owner type, owner ID, language, stars, forks and license are available.
• commit_files.commit_id and patches.commit_id join to commits.commit_id.
• For the file entries behind a commit's summary, filter commit_files.change_set_id to commits.primary_change_set_id.
• Git author/committer names and emails belong to the commit. GitHub account fields describe linked accounts.

Field definitions are in DATA_DICTIONARY.md, source-file access in PROVENANCE.md, and source rights in LICENSES.md.

Citation

Cite the dataset URL together with the fixed revision used in your analysis.