Agent Skills
› brycewang-stanford/Awesome-Journal-Skills
› issta-artifact-evaluation
issta-artifact-evaluation
GitHub指导为ISSTA会议打包工具、基准和结果以参与工件评估。涵盖ACM徽章（可用、功能、可复用、复现）的达标要求，包括Zenodo归档、容器化、匿名审查副本及可运行入口点，确保满足评估者对快速验证和结果复现的核心需求。
Trigger Scenarios
准备ISSTA会议的工件评估提交材料
配置ACM工件徽章所需的归档和容器环境
生成用于匿名评审的工件副本
编写确保结果可复现的脚本和文档
Install
npx skills add brycewang-stanford/Awesome-Journal-Skills --skill issta-artifact-evaluation -g -y
SKILL.md
Frontmatter
{
"name": "issta-artifact-evaluation",
"description": "Use when packaging an ISSTA tool, benchmark, and results for the artifact-evaluation track, covering the ACM badges (Artifacts Available via Zenodo, Evaluated Functional and Reusable, Results Reproduced), the anonymous review-time copy, containerization, a runnable entry point, and what ISSTA artifact evaluators actually try first."
}
ISSTA Artifact Evaluation
Use this to package evidence for ISSTA's artifact-evaluation track. ISSTA has a genuinely strong artifact culture: a runnable tool and a shared benchmark are treated as normal, and the badges carry weight. Reopen the current artifact call before packaging — the badge set and the archival requirement are cycle-specific.
The ACM badge targets
| Badge | What it certifies | What you must ship |
|---|---|---|
| Artifacts Available | The artifact is publicly, permanently retrievable | A DOI-issuing deposit (Zenodo) — a GitHub link alone does not qualify |
| Artifacts Evaluated — Functional | The artifact runs and does what the paper says | A working entry point, dependencies, and a documented expected output |
| Artifacts Evaluated — Reusable | Others can inspect, adapt, and build on it | Clean structure, real documentation, and configurability beyond the paper's runs |
| Results Reproduced | The paper's key results regenerate from the artifact | Scripts that produce the paper's tables/figures within stated tolerance |
Available is about archival; Functional and Reusable are about engineering quality; Results Reproduced is about matching the paper. They are earned independently — decide which you are going for before you package.
Artifact plan
- Decide the badge set, then package to its bar. A Reusable badge needs documentation and structure a Functional-only artifact can skip.
- Ship the tool with a single documented entry point and a container or environment file, so an evaluator reaches a result without reconstructing your machine.
- Pin subjects and benchmark versions: the exact Defects4J revision, subject-program commit SHAs, and any fuzzing seed corpora, archived rather than referenced by name.
- Provide a results-regeneration path: a script that takes logged runs to the paper's tables, so "Results Reproduced" is a command, not an argument.
- Anonymize the review-time copy — repository owners, commit authors, container labels — because ISSTA artifact review is double-anonymous alongside the paper.
- After acceptance, make the Zenodo deposit public and citable, and record its DOI for the camera-ready badge display.
What ISSTA evaluators try first
- The README's quick-start, then the single command that produces one headline number. If that fails on a clean machine, no amount of internal quality is visible.
- A small, fast subset that finishes in minutes, before any full multi-day fuzzing or symbolic-execution campaign. Ship a "smoke" configuration explicitly.
- The mapping from a paper claim to the artifact output that supports it; an artifact whose outputs cannot be tied back to Table N reads as unverifiable.
Handling long-running and non-deterministic tools
Testing and analysis artifacts often run for hours and vary between runs. Package for that reality:
artifact/
README.md # quick-start, smoke config, full config, expected outputs, runtime
Dockerfile # pinned toolchain and dependencies
subjects/ # pinned subject SHAs or a fetch script that pins them
run_smoke.sh # minutes: reproduces one representative row
run_full.sh # hours/days: reproduces all tables
results/ # logged raw outputs from the authors' runs
scripts/tables.py # regenerates paper tables from results/
State the run count and expected variance for non-deterministic results, and have the regeneration script accept the evaluator's fresh runs as well as the shipped logs, so a partial reproduction still lands on the paper's numbers within tolerance.
Calibration anchors
- Evaluators are time-boxed; assume they run the smoke config and skim the full one. Design so the smoke path alone justifies Functional.
- Badge names, the Zenodo requirement, and any single-blind vs. double-blind detail vary by cycle; verify against the current artifact call rather than a past year.
Output format
[Badge target] Available / Functional / Reusable / Results Reproduced
[Entry point] <command + smoke runtime>
[Pinned subjects] <benchmark version / SHAs archived?>
[Reproduction level] turnkey / scripted / descriptive / weak
[Anonymity risks] <owners/authors/labels/paths>
[Fixes before upload] <ordered list>
Version History
- 9f86f09 Current 2026-07-19 16:06
