GHAgentCI: A Dataset of Coding-Agent Invocations in CI Workflow Histories

GHAgentCI: A Dataset of Coding-Agent Invocations in CI Workflow Histories Daniel Ogenrwot

, Jorge Gonzalo Delgado Cervantes

, Woojoong Kim

, John Businge

arXiv:2611.15522v1 [cs.SE] 11 Nov 2026

University of Nevada Las Vegas Las Vegas, USA {ogenrwot, delgaj10, kimw19}@unlv.nevada.edu, john.businge@unlv.edu

Abstract—Coding agents are increasingly integrated into GitHub Actions to review pull requests, triage issues and modify code. Identifying these integrations is challenging because agentic workflows often resemble ordinary CI files, with evidence appearing in action references, shell commands and workflow conventions. We present GHAgentCI, a dataset of 15K+ workflow histories from 6K+ of 196K+ mined GitHub repositories, comprising 159K+ revisions and 205K+ agent-evidence records linked to workflow triggers and declared execution constraints. Our open-source extractor, cai-miner, detects agent-related evidence using an extensible catalogue. Manual validation of 576 stratified records confirmed 96.1% of dedicated-action matches and 98.8% of GitHub Agentic Workflow matches, with lower rates for other evidence kinds. Exploratory analyses reveal a shift toward command-line and GitHub Agentic Workflow integrations, differences in declared safeguards, and frequent removal of detected agentic workflows. GHAgentCI supports longitudinal studies of coding-agent integration into CI. Index Terms—coding agents, continuous integration, GitHub Actions, agentic workflows, software supply chain, mining software repositories

containing 159K+ revisions and 205K+ agent-evidence records. The dataset links step- and file-level evidence to workflow triggers, permissions and actor conditions (Section VI).To demonstrate its research potential, we characterise how agent integration mechanisms, declared execution constraints and workflow retention have changed over time (Section VII). II. R ELATED W ORK

Prior research examines coding-agent outputs and configurations. AIDev collects agent-authored pull requests [4], while Robbes et al. study adoption through commits and pull requests [3], [12]. Other studies investigate agent context and configuration files [13]–[16], with GHAgentFiles providing longitudinal histories [10]. Ghaleb [17] studies agentauthored pull requests modifying CI configurations. These works examine agent instructions and outputs rather than their invocation from CI workflows. Research on GitHub Actions examines adoption and use [18], [19], migration from other CI services [20], and effects on pull requests [21]. Cardoen I. I NTRODUCTION et al. [11] provide general GitHub Actions workflow histories Coding agents are increasingly integrated into repository without identifying coding-agent integrations. GHAW-H [22] automation. GitHub Actions can invoke them on pull re- tracks GitHub Agentic Workflows in 262 repositories, linking quests, issue comments and schedules to review changes, Markdown specifications to compiled lock files, but covers only triage issues and modify code. Dedicated actions such as gh-aw. Neither dataset systematically detects integrations across claude-code-action [1], command-line tools and GitHub dedicated actions, shell commands and other mechanisms. Agentic Workflows (gh-aw) [2] support this automation. Unlike GHAgentCI complements these datasets with longitudinal interactive agent use, which has received research attention [3]– evidence of coding-agent integrations, workflow triggers and [9], these workflows can process external contributions under declared safeguards. Repository and file-level metadata support CI-defined permissions and conditions. Studying this practice linkage with existing datasets. requires examining invocation mechanisms, triggers, declared III. D ETECTING AGENT I NVOCATIONS constraints and their evolution. Identifying these integrations at scale is challenging. Agentic We developed cai-miner, an open-source tool for detecting workflows often lack distinctive filenames under .github/- agent-related evidence in GitHub Actions workflow histories. workflows/. Evidence may appear in action references, shell Unlike agent configuration files, agentic workflows often have commands or gh-aw compiler metadata. Copilot’s reserved no distinctive filenames under .github/workflows/ and must be setup workflow configures agents without invoking them, while identified by their contents. Based on vendor documentation and credentials such as ANTHROPIC_API_KEY may appear existing workflows, we identified four integration mechanisms. without an agent invocation. Existing datasets cover agent- First, dedicated actions, such as anthropics/claude-codeauthored changes, agent configuration files or general workflow action, invoke agents through workflow steps. Second, shell histories [4], [10], [11], but do not systematically link agent- steps may execute agent command-line interfaces (CLIs), related evidence to workflow triggers, declared constraints and such as claude -p or codex exec. Third, Copilot’s reserved their evolution. workflow file, copilot-setup steps.yml, configures its execution To address this gap, we present GHAgentCI, a dataset of environment. Fourth, GitHub Agentic Workflows (gh-aw) com15K+ workflow-file histories from 6K+ GitHub repositories, piles Markdown specifications into .lock.yml workflows whose

Listing 1: Excerpt of the frozen catalogue: the files to extract, and the patterns that identify Claude and Copilot, by evidence kind. [paths] github_actions

= [".github/workflows/*.yml", ".github/workflows/*.yaml"] agentic_workflows = [".github/workflows/*.md"] [agents.claude] uses = ["anthropics/claude-code-action", "anthropics/claude-code-base-action"] run = ["claude -p", "claude --print", "@anthropic-ai/claude-code", ...] url = ["api.anthropic.com"] env = ["ANTHROPIC_API_KEY", "CLAUDE_CODE_OAUTH_TOKEN"] [agents.copilot] path = ["copilot-setup-steps.yml", "copilot-setup-steps.yaml"] job = ["copilot-setup-steps"] ...

Because not every match indicates an invocation, we distinguish strong evidence from weaker signals, such as credentials and inference endpoints, which may also occur in ordinary LLM-backed applications. Weak evidence alone does not classify a revision as agentic. Records retain their declaring scope: a workflow-level ANTHROPIC_API_KEY produces one file-level record, not one per step. Multiple matches within a step are combined into one record, retaining the strongest evidence kind. Reserved Copilot setup files remain configuration evidence, not direct invocations. Each record preserves its evidence kind and scope, distinguishing invocation declarations from indirect configuration evidence. IV. B UILDING THE DATASET

TABLE I: Evidence kinds, example patterns, evidence tiers, and the number of agent-evidence records associated with each kind. Only strong evidence classifies a workflow revision as agentic. kind

matched in

example

tier

records

uses run path ghaw job url env

step uses: step run: file name lock-file header job identifier step content env:, any scope

anthropics/claude-code-action claude -p copilot-setup-steps.yml engine declared by gh-aw copilot-setup-steps api.anthropic.com ANTHROPIC_API_KEY

strong strong strong strong weak weak weak

69,104 64,822 8,805 36,657 –* 1,969 24,963

* Reserved job identifiers occur only together with the reserved file name, which decides.

To study coding-agent adoption in CI, we selected repositories independently of agent usage. We queried SEART [23] on 3 October 2026 for non-fork GitHub repositories with at least 300 commits and a last-commit date within the preceding year, as recorded in the SEART export. This excluded 143,943 inactive repositories, leaving 196,938 candidates. Unlike GHAgentFiles [10], we did not require repositories to be at least one year old, which would exclude newer projects potentially adopting coding agents. Instead, we flagged the 172,607 repositories meeting this criterion for comparisons with GHAgentFiles. We cloned repositories between 3 October 2026 and 5 October 2026 using partial clones excluding blobs larger than 64 KB. To avoid truncated workflow histories, cai-miner separately fetched workflow blobs exceeding this limit and recorded blob request and read counts to check extraction completeness. Repositories exceeding the initial 15-minute mining budget were retried with a one-hour budget. After retries, 681 repositories remained unmined, mostly because they were deleted or private. These and four deliberately excluded gh-aw development repositories remain in the dataset with their status. Finally, we examined matches predating LLM-based coding agents. Some Markdown frontmatter keys matched the gh-aw detector spuriously. Following GHAgentFiles [10], we used ChatGPT’s release on 30 November 2022 as a conservative cutoff, excluding 3 evidence records from 3 histories. The earliest retained agentic workflows include PR-Agent and CodeRabbit, used for pull request reviews in 2023.

headers declare the agent engine. Detecting only dedicated actions would therefore miss other integration mechanisms. We encode these traces in a declarative TOML catalogue (Listing 1). The [paths] section specifies files to extract, including gh-aw Markdown sources, while agent sections define detection patterns for the evidence kinds in Table I. We constructed the catalogue from documentation for nine agents: Claude, Codex, Copilot, Gemini, Cursor, Aider, Devin, PRAgent and CodeRabbit. We also included 1,122 action repositories listed in GitHub Marketplace’s ai-assisted category on 8 September 2026. Actions associated with known vendors added Goose, OpenCode and OpenHands; the remaining actions form a generic_ai group, which may include general LLM utilities. The catalogue covers 12 named agents and was frozen before mining. We retain generic_ai matches in the dataset but V. VALIDATION exclude them from the reported adoption counts and analyses. We sampled 576 records across strata defined by the deciding Given a local repository clone, cai-miner traverses the default branch’s first-parent history, following cofee [10]. It evidence kind. The 384 named-agent records were allocated records workflow revisions with their git metadata and change proportionally across uses, run, path, gh-aw and parse-failed types. A logical file retains its identifier across modifications strata, with at least 30 each, targeting a ±5% margin at 95% and renames, but receives a new one if deleted and recreated. confidence. Weak-only and generic_ai evidence received 96 Each distinct YAML revision is parsed and matched against records each. Within each stratum, we used systematic sampling the catalogue at the job and step levels. For gh-aw, cai- with a random start, ordered by agent, and selected at most miner parses compiler headers and Markdown frontmatter two records per repository. across compiler versions, extracting the declared engine, tools, For each record, two authors independently inspected the permitted outputs, network policy and sandbox. Unparseable workflow at its commit, without seeing the detector’s clasrevisions fall back to text matching and are flagged accordingly. sification. A yes label required evidence of a coding-agent

TABLE II: Manual validation by evidence kind. Strict rates treat maybe as incorrect. Intervals are 95% Wilson confidence intervals. Kind

n Yes Maybe

No Confirmed (%) 95% CI

gh-aw 80 79 uses 128 123 run 116 95 path 30 26 parse-failed 30 14 weak-only 96 41 generic AI 96 49

0 3 4 0 1 19 12

Total

39 110

576 427

1 2 17 4 15 36 35

98.8 96.1 81.9 86.7 46.7 42.7 51.0

[93.3,99.8] [91.2,98.3] [73.9,87.8] [70.3,94.7] [30.2,63.9] [33.3,52.7] [41.2,60.8]

– –

TABLE III: Content of the dataset file

description

revisions.csv.gz

Compressed CSV file with one row per touch of a workflow file in an agentic history. Compressed CSV file with one row per agentevidence record per revision. Compressed CSV file with one row per revision of a gh-aw artifact: engine, compiler and schema versions, model, tools, MCP tools, safe outputs, secrets, sandboxing, network policy and permissions. Compressed CSV file with one row per mined repository: commit count, workflow and agentic file counts, agents used, first and last agentic dates, blob accounting and status. Compressed archive of every distinct workflow revision.

steps.csv.gz ghaw_manifests.csv.gz

repositories.csv.gz

invocation or, for reserved Copilot setup files, valid agentcontent.tar.gz specific configuration. Merely naming or installing an agent was insufficient. Weak-only matches required additional evidence of agent execution, while ambiguous cases were labeled maybe. TABLE IV: Characteristics of the dataset. The % indicates the proportion of elements in the last snapshot w.r.t. all snapshots. Only yes labels counted as confirmed. The authors agreed on 97.7% of records (κ = 0.96) and resolved disagreements characteristic all snapshots last snapshot through discussion. Table II reports the adjudicated labels and # repositories 6,345 4,598 (72.5%) confirmation rates by evidence kind. with gh-aw workflows 506 377 (74.5%) # histories 15,928 9,942 (62.4%) Dedicated actions and gh-aw yielded the highest confirmation # commits 104,545 6,923 (6.6%) rates (96.1% and 98.8%), followed by shell commands (81.9%), # workflow revisions 159,878 9,942 (6.2%) some of which matched agent names or installation commands agentic 125,266 9,265 (7.4%) # agent records 205,832 12,944 (6.3%) without execution. Parse-failed records reached 46.7%, weakvia uses 69,104 5,050 (7.3%) only evidence 42.7%, and generic AI actions 51.0%. The via run 64,822 3,421 (5.3%) latter two groups are retained separately rather than counted as via path 8,805 992 (11.3%) via ghaw 36,657 2,153 (5.9%) confirmed invocations. To examine missed invocations, we remined 772 randomly selected repositories containing workflows and inspected 200 unflagged workflows selected using broad keywords (e.g., llm, agent). Eleven (5.5%; 95% Wilson CI: kept separate. The dataset also includes 36,660 gh-aw artifact 3.1–9.6%) contained an agent invocation. This is the missed- revisions from 506 repositories. invocation rate within the keyword-selected sample, not overall VII. E XPLORATORY E MPIRICAL A NALYSIS recall. We examine three questions using named agents with strong evidence only. VI. T HE GHAGENT CI DATASET RQ1: How have coding-agent integration mechanisms in CI GHAgentCI comprises four compressed CSV tables (Ta- evolved? ble III) and an archive of distinct workflow revisions stored Fig. 1 shows how newly detected agentic workflows are by SHA-256 content hash. The tables are keyed by repository, integrated into CI. In 2025 Q2, 73.9% used dedicated vendor file identifier and commit hash. The repository table includes actions, compared with 42.1% in 2026 Q3. Over the same agentic and non-agentic repositories, providing an adoption period, gh-aw increased from 0.4% to 26.6%, and CLI invocadenominator. The revision table tracks each change to a file tions from 7.3% to 27.6%, while reserved Copilot setup files that is agentic at some point, from creation to deletion, and declined from 18.4% to 3.7%. At the latest repository snapshot, retains the git metadata of GHAgentFiles [10] to support the most common mechanism varies by agent: dedicated actions linking the datasets. Each steps-table row records agent-related account for 77% of Claude workflows, CLI calls for 51% of evidence at the step level (job and step) or file level (reserved Codex workflows, and gh-aw for 48% of Copilot workflows. filename, workflow-level credential or gh-aw header). A field- These findings show that identifying agentic workflows solely level schema accompanies the dataset, which is released with through dedicated action references misses a growing share of cai-miner. them. Table IV summarises GHAgentCI. Histories span 16 July RQ2: How do workflow triggers relate to execution con2023 to 5 October 2026 and include revisions before agent straints? introduction and after removal. 26,932 records contain weak Fig. 2 examines declared safeguards across 9,265 agentic evidence only. The share of repositories with detected agentic workflows. Inferred tasks are mainly review (38.4%) and workflow histories is 7.9% among repositories younger than one triage (34.4%), rather than implementation (5.7%). Among year, compared with 2.6% among older repositories. Another issue- or comment-triggered workflows, 85% restrict triggering 417 repositories contain only generic_ai actions and are actors, but only 19% explicitly declare token permissions and

CLI in a shell step

reserved file name

521

1,919

3,844

1,910

4,772

100

gh-aw

2,782

% still agentic

% of new agentic workflows

100

dedicated action

gh-aw

75 CLI

50 25 0

80 60 Copilot (1,577 repositories) Codex (355 repositories) Claude (4,084 repositories) other (630 repositories)

40 20

action

0 2025 Q2

2025 Q3

2025 Q4

2026 Q1

2026 Q2

0

2026 Q3

Fig. 1: Integration mechanisms detected in newly agentic workflows by quarter; numbers above the bars indicate workflow counts. Quarters that ended after cloning started are omitted. permissions declared

actor restricted

3

6

9

12

15

months since the repository first invoked an agent in CI

Fig. 3: Repository retention of detected agentic workflows by first adopted agent (Kaplan–Meier; 95% confidence bands). Simultaneous adopters count in multiple groups; curves stop when fewer than 50 repositories remain under observation.

action pinned to a SHA

issue or comment (3,070)

VIII. T HREATS TO VALIDITY

pull request (4,879) pull_request_target (327) push (1,436) schedule (1,594) manual dispatch (4,336) called by a workflow (449) 0

25

50

75

100

% of agentic workflows started by the event

Fig. 2: Declared safeguards by trigger at the latest commit. Counts are workflows; those with multiple triggers appear in multiple groups. SHA pinning applies to dedicated actions.

24% of dedicated actions are SHA-pinned. Scheduled and manually triggered workflows show the opposite pattern. Of 327 workflows using pull_request_target, only 55% restrict triggering actors, despite the potential exposure of repository secrets to fork-triggered workflows. Among 1,120 gh-aw sources, 50.8% declare custom network policies and 2.7% explicit sandboxing. RQ3: How long do repositories retain agentic workflows? A repository remains agentic while any detected agentic workflow is present, regardless of workflow replacement. Fig. 3 shows Kaplan–Meier retention estimates: 10% lose detected agentic workflows within one month, 20% within three months, and 44% within a year. Another 174 repositories removed workflows on the adoption day. Copilot shows the longest retention; Codex’s curve ends after approximately 8 months, with only 21 repositories observed for a full year. Removals cluster across related repositories: 33 Pulumi repositories removed Copilot setup workflows together, while 32 R-package repositories (e.g., r-dbi, r-lib, tidyverse) stopped on 13 September 2026.

SEART-based sampling excludes small and inactive repositories, which may adopt agents differently. Rewritten git histories may affect commit dates, while default-branch first-parent traversal misses workflows confined to other branches [24]. Like GHAgentFiles [10], our detection catalogue relies on vendor documentation that may be incomplete or change over time, potentially biasing historical trends. cai-miner may miss invocations through unrecognised wrappers, composite actions or external reusable workflows. Some run patterns match agent names or installation commands without execution, while unparseable revisions use less precise text matching. Validation identifies false positives and some missed cases but cannot establish overall recall (Section V). Detected agent evidence reflects workflow declarations, not observed executions. Moving agents to external reusable workflows may appear as abandonment, while organisation-wide changes create dependencies between repository histories. Inferred tasks labels and actor-restriction flags support exploratory analysis, not ground truth or definitive security assessments. IX. C ONCLUSION We presented GHAgentCI, a dataset of 15,928 agentic workflow histories spanning 159,878 revisions and 205,832 agentevidence records across 6,345 GitHub repositories. We also provide cai-miner, an open-source extractor with an extensible catalogue for detecting agent integrations in CI. Unlike datasets focused on agent-produced artifacts or configuration files, GHAgentCI captures how coding agents are integrated into CI, under which declared constraints, and how these integrations evolve. Together, the dataset and extractor support longitudinal studies of coding-agent adoption in software delivery pipelines. Future releases could extend the detection catalogue and resolve external reusable workflows, improving coverage of currently unrecognized integrations.

Research opportunities: The preliminary analysis presented in this section only scratches the surface. GHAgentCI supports studies of coding-agent governance and adoption over time. DATA AVAILABILITY Linking it with GHAgentFiles [10] could compare context-file restrictions with CI safeguards, while AIDev [4] could support GHAgentCI and its replication package are archived on Zencomparisons of pull request review outcomes across CI-invoked odo at https://doi.org/10.5281/zenodo.24088832. The project and interactively used agents, where invocation origins can be website at https://unlv-evol.github.io/cai-miner/ provides an established. interactive explorer, dataset documentation and executable

notebooks. The extraction tool and replication scripts are available at https://github.com/unlv-evol/cai-miner. The tool is also distributed through PyPI at https://pypi.org/project/cai-miner/ and can be installed using pip install cai-miner. R EFERENCES [1] Anthropic, “claude-code-action,” GitHub repository, 2026, accessed: Nov. 10, 2026. [Online]. Available: https://github.com/anthropics/ claude-code-action [2] GitHub, “GitHub Agentic Workflows (gh-aw),” GitHub repository, 2026, accessed: Nov. 10, 2026. [Online]. Available: https://github.com/github/ gh-aw [3] R. Robbes, T. Matricon, T. Degueule, A. Hora, and S. Zacchiroli, “Promises, perils, and (timely) heuristics for mining coding agent activity,” in Proceedings of the 23rd International Conference on Mining Software Repositories. New York, NY, USA: Association for Computing Machinery, 2026, p. 496–507. [Online]. Available: https://doi.org/10.1145/3793302.3793375 [4] H. Li, H. Zhang, and A. E. Hassan, “The rise of ai teammates in software engineering (se) 3.0: How autonomous coding agents are reshaping software engineering,” 2025. [Online]. Available: https://arxiv.org/abs/2507.15003 [5] D. Ogenrwot and J. Businge, “Agenticflict: A large-scale dataset of merge conflicts in ai coding agent pull requests on github,” in Proceedings of the 3rd ACM International Conference on AI-Powered Software, ser. AIware ’26. New York, NY, USA: Association for Computing Machinery, 2026, p. 323–331. [Online]. Available: https://doi.org/10.1145/3805760.3814923 [6] D. Ogenrwot and J. Businge, How AI Coding Agents Modify Code: A Large-Scale Study of GitHub Pull Requests. New York, NY, USA: Association for Computing Machinery, 2026, p. 924–928. [Online]. Available: https://doi.org/10.1145/3793302.3793603 [7] D. Ogenrwot and J. Businge, “Patchtrack: A comprehensive analysis of chatgpt’s influence on pull request outcomes,” Empirical Software Engineering, vol. 31, no. 5, p. 136, 2026. [Online]. Available: https://doi.org/10.1007/s10664-026-10869-5 [8] D. Ogenrwot and J. Businge, “Patchtrack: Analyzing chatgpt’s impact on software patch decision-making in pull requests,” in Proceedings of the 39th IEEE/ACM International Conference on Automated Software Engineering, ser. ASE ’24. New York, NY, USA: Association for Computing Machinery, 2024, p. 2480–2481. [Online]. Available: https://doi.org/10.1145/3691620.3695338 [9] R. Sserunjogi, D. Ogenrwot, and J. Businge, “Prompt quality and pull request outcomes: A stage-based empirical study of llm-assisted development,” 2026. [Online]. Available: https://arxiv.org/abs/2606.19644 [10] G. Cardoen, T. Mens, and A. Decan, “Ghagentfiles: A dataset of coding agent file histories in github repositories,” in Proceedings of the 42nd International Conference on Software Maintenance and Evolution, F.R.S.-FNRS - Fonds de la Recherche Scientifique. Institute of Electrical and Electronics Engineers (IEEE), 16 September 2026. [Online]. Available: https://doi.org/10.1109/ICSME67971.2026.00080 [11] G. Cardoen, T. Mens, and A. Decan, “A dataset of github actions workflow histories,” in Proceedings of the 21st International Conference on Mining Software Repositories, ser. MSR ’24. New York, NY, USA: Association for Computing Machinery, 2024, p. 677–681. [Online]. Available: https://doi.org/10.1145/3643991.3644867

[12] R. Robbes, T. Matricon, T. Degueule, A. Hora, and S. Zacchiroli, “Agentic much? adoption of coding agents on github,” ACM Trans. Softw. Eng. Methodol., vol. 35, no. 11, Nov. 2026. [Online]. Available: https://doi.org/10.1145/3822180 [13] S. Mohsenimofidi, M. Galster, C. Treude, and S. Baltes, “Context engineering for ai agents in open-source software,” in Proceedings of the 23rd International Conference on Mining Software Repositories. New York, NY, USA: Association for Computing Machinery, 2026, p. 194–198. [Online]. Available: https://doi.org/10.1145/3793302.3793350 [14] W. Chatlatanagulchai, H. Li, Y. Kashiwa, B. Reid, K. Thonglek, P. Leelaprute, A. Rungsawang, B. Manaskasemsak, B. Adams, A. E. Hassan, and H. Iida, “Agent readmes: An empirical study of context files for agentic coding,” ACM Trans. Softw. Eng. Methodol., vol. 35, no. 11, Nov. 2026. [Online]. Available: https://doi.org/10.1145/3840295 [15] M. Galster, S. Mohsenimofidi, J. L. Lulla, M. A. Abubakar, C. Treude, and S. Baltes, “Configuring agentic ai coding tools: An exploratory study,” in Proceedings of the 3rd ACM International Conference on AI-Powered Software, ser. AIware ’26. New York, NY, USA: Association for Computing Machinery, 2026, p. 11–20. [Online]. Available: https://doi.org/10.1145/3805760.3814887 [16] M. Galster, S. Mohsenimofidi, L. Böhme, J. L. Lulla, M. A. Abubakar, C. Treude, and S. Baltes, “A dataset of agentic ai coding tool configurations,” in Proceedings of the 3rd ACM International Conference on AI-Powered Software, ser. AIware ’26. New York, NY, USA: Association for Computing Machinery, 2026, p. 314–322. [Online]. Available: https://doi.org/10.1145/3805760.3814922 [17] T. A. Ghaleb, “When ai agents touch ci/cd configurations: Frequency and success,” in Proceedings of the 23rd International Conference on Mining Software Repositories, ser. MSR ’26. New York, NY, USA: Association for Computing Machinery, 2026, p. 817–821. [Online]. Available: https://doi.org/10.1145/3793302.3793581 [18] T. Kinsman, M. Wessel, M. A. Gerosa, and C. Treude, “How do software developers use github actions to automate their workflows?” in 2021 IEEE/ACM 18th International Conference on Mining Software Repositories (MSR), 2021, pp. 420–431. [19] A. Decan, T. Mens, P. R. Mazrae, and M. Golzadeh, “On the use of github actions in software development repositories,” in 2022 IEEE International Conference on Software Maintenance and Evolution (ICSME), 2022, pp. 235–245. [20] M. Golzadeh, A. Decan, and T. Mens, “On the rise and fall of ci services in github,” in 2022 IEEE International Conference on Software Analysis, Evolution and Reengineering (SANER), 2022, pp. 662–672. [21] M. Wessel, J. Vargovich, M. A. Gerosa, and C. Treude, “Github actions: The impact on the pull request process,” Empirical Software Engineering, vol. 28, no. 6, p. 131, 2023. [Online]. Available: https://doi.org/10.1007/s10664-023-10369-w [22] P. Valenzuela-Toledo, T. Kehrer, and S. Panichella, “GHAW-H: A dataset of github agentic workflow histories,” 2026, dataset, version v0.1.2. [Online]. Available: https://doi.org/10.5281/zenodo.22084012 [23] O. Dabic, E. Aghajani, and G. Bavota, “Sampling projects in github for msr studies,” in 2021 IEEE/ACM 18th International Conference on Mining Software Repositories (MSR), 2021, pp. 560–564. [24] E. Kalliamvakou, G. Gousios, K. Blincoe, L. Singer, D. M. German, and D. Damian, “An in-depth study of the promises and perils of mining github,” Empirical Software Engineering, vol. 21, no. 5, pp. 2035–2071, 2016. [Online]. Available: https://doi.org/10.1007/s10664-015-9393-5

Related documents

• GHAgentCI: A Dataset of Coding-Agent Invocations in CI Workflow Histories#1659104
• Software Engineering in the Agent Era From Trustworthy Change to Human Agent Software Organizations#927901
• AgentPack: A Dataset of Code Changes, Co-Authored by Agents and Humans#1123950
• MTAC-IFBench: Benchmarking Instruction-Following in Multi-Turn Agentic Coding#1084892
• DynaFlow: Learning Free-Form Multi-Agent Orchestration#1471228