CLAUDE.md — Zhaoyang Chu’s Academic Homepage

Inherits the global workflow in ~/.claude/CLAUDE.md (Git-as-universal-workflow, branch → PR → merge, tag conventions). This file adds repo-specific context so any future session can edit without re-asking.

What this repo is

My personal academic homepage, served at https://zhaoyang-chu.github.io via GitHub Pages. Built on the AcadHomepage Jekyll template (RayeRen/acad-homepage.github.io). It is a long-lived platform for my academic brand: I update papers, news, and bio here regularly.

The whole site is effectively a single page — almost all content lives in one file.

Where things live

Path	What it is	Touch frequency
`_pages/about.md`	The only content page — the homepage. News, Publications, Honors, Educations all live here.	⭐ Primary
`_config.yml`	Site config: title, avatar, location, social links (Google Scholar, GitHub, LinkedIn, ORCID, Twitter, email).	Occasional
`_data/navigation.yml`	Top nav-bar anchors. Add an entry when adding a new top-level `#` section to about.md.	When adding sections
`images/`	Per-paper thumbnails + `profile.jpg` avatar.	When adding a paper
`assets/`	`ZhaoyangChu_CV.pdf`, paper PDFs, css/js/fonts.	When updating CV
`_includes/`, `_layouts/`, `_sass/`	Template HTML + styles. Rarely touched.	Almost never
`google_scholar_crawler/` + `.github/workflows/google_scholar_crawler.yaml`	Auto-updates citation counts.	Automated, but can silently stall — see Citation crawler below

Identity (keep consistent everywhere)

Second-year PhD student (started Sep 2025 — bump the year in about.md and README.md each September), Dept. of Computer Science, UCL. Co-supervised by Dr. He Ye and Prof. Federica Sarro (always in this order).
MSc from HUST (advisor Prof. Yao Wan); collaborations with Prof. Lingming Zhang (UIUC) and Prof. Hongyu Zhang (Chongqing Univ.).
Research: intersection of SE and AI — coding agents: how to build, evaluate, and make them reliable on real-world tasks (same wording as the CV).
Primary email: zhaoyang.chu.25@ucl.ac.uk (also zychu418@gmail.com). This is the email in _config.yml and about.md — keep them in sync.

Most frequent task: adding a publication

Drop the thumbnail into images/ (jpg/png; rendered at width="100%" inside the box).
In _pages/about.md, copy an existing <div class='paper-box'> block into the Publications section. Publications are ordered Preprints first, then by venue/recency (newest/most prestigious near the top). Exception: the representative work (currently TerminalWorld) stays pinned at the very top even after it is accepted.
The block template:

<div class='paper-box'><div class='paper-box-image'><div><div class="badge">VENUE or "Preprint"</div><img src='images/THUMB.jpg' alt="sym" width="100%"></div></div>
<div class='paper-box-text' markdown="1">

**Paper Title**.<br>
Author A, [**Zhaoyang Chu**](https://zhaoyang-chu.github.io/), Author C†.<br>
[**VENUE YEAR**](venue-url). *Full venue name*.<br>
\[ [Paper](url) \] \[ [Code](url) \] \[ [Homepage](url) \] \[ [Dataset](url) \]

</div>
</div>

My name is always bolded and linked: [**Zhaoyang Chu**](https://zhaoyang-chu.github.io/).
\* after a name = equal contribution; † = corresponding author. There’s a legend line near the top of Publications.
Only include link brackets that exist (Paper / Code / Homepage / Dataset).
Citation count: append <span class='show_paper_citations' data='HYu3DyEAAAAJ:PUB_ID'></span> after the last link bracket so the paper shows | Citations: N. Find PUB_ID (the author_pub_id) in gs_data.json on the google-scholar-stats branch — match by title. New papers not yet indexed by Google Scholar have no PUB_ID; skip the span for those (a span with a bad id throws a JS error that breaks all per-paper counts). The total-citations shields badge near the top updates automatically and needs no per-paper edit.
GitHub stars / HuggingFace downloads badges (live shields.io, appended after the citation span on the same line):
- Stars (one per GitHub repo): ![GitHub stars](https://img.shields.io/github/stars/OWNER/REPO?style=flat-square&logo=github&label=stars)
- HF downloads (one per HF dataset, shows all-time downloads): ![HF downloads](https://img.shields.io/badge/dynamic/json?style=flat-square&url=ENCODED_HF_API_URL&query=%24.downloadsAllTime&label=%F0%9F%A4%97%20downloads&color=ff9d00) where ENCODED_HF_API_URL is https://huggingface.co/api/datasets/OWNER/NAME?expand=downloadsAllTime percent-encoded. Use expand=downloadsAllTime (the plain downloads field is only the last 30 days; downloadsAllTime requires the expand). Note: use expand=..., not expand[]=... — shields rejects the bracketed form with invalid query parameter: url. 4. Add a matching News line at the top of the # 🔥 News list: - *YYYY.MM*:  🎉 [PaperName](paper-url), <short phrase>, was accepted to **VENUE**.
Bold marks only the achievement (venue, award, download count). Paper names are plain links — never bold.
The <short phrase> (~8–12 words) must come from the paper’s abstract; no invented claims, no numbers except milestone counts.
Papers without a short name: [Our paper](url) on <topic>.
Items from before the PhD start (2025.09) live inside the <details> “Show older news” fold; keep roughly the latest 8 items visible and move older ones into the fold.
1. If it’s an award or new section, also update # 🎖 Honors and Awards / _data/navigation.yml.

Conventions

Dates use *YYYY.MM* format in News and Educations.
Only papers with a public arXiv/official version go on the homepage; unpublished submissions stay off. Never write “Under Review” here — a public-but-unaccepted paper is badged Preprint.
New papers without a figure yet use images/placeholder.svg with a  comment; replace before merging.
Commented-out  paper-box / news blocks are intentionally parked (old preprints, hidden papers) — leave them unless asked.
Don’t hand-edit citation counts; the GitHub Action regenerates them on the google-scholar-stats branch weekly (Mon 08:00 UTC), on any Pages build, or via manual workflow_dispatch. A failed run never overwrites good data (the step exits before pushing).

Citation crawler (lessons from the 2026-07 → 2026-10 stall)

Three independent ways it stalls — check all three if counts look stale (git log -1 origin/google-scholar-stats):

Dependency drift. scholarly==1.5.1 imports bibtexparser.bibdatabase, removed in bibtexparser 2.0 (2026-09); requirements.txt pins bibtexparser<2. A ModuleNotFoundError in the run log means another unpinned dependency moved.
Google Scholar blocks GitHub Actions IPs — intermittently. A blocked run hangs until the 10-min timeout cancels it; the next run may succeed. Long streaks of cancelled runs are possible (9 in a row in 2026-07/09).
GitHub auto-disables the scheduled workflow after 60 days without commits (gh workflow list --all shows disabled_inactivity). Re-enable with gh workflow enable google_scholar_crawler.yaml, then gh workflow run google_scholar_crawler.yaml.

After fresh data lands, per-paper badges read it through jsDelivr, which caches the branch — purge with curl https://purge.jsdelivr.net/gh/Zhaoyang-Chu/Zhaoyang-Chu.github.io@google-scholar-stats/gs_data.json.

Running locally

./run_server.sh   # bundle exec jekyll liveserve

Environment requirements (set up once; macOS system Ruby 2.6 will not work):

Ruby 3.3.x via rbenv — pinned in .ruby-version (3.3.11), matching GitHub Pages’ own Ruby. Installed with brew install rbenv ruby-build && rbenv install 3.3.11. rbenv init is in ~/.zshrc.
UTF-8 locale — export LANG=en_US.UTF-8 LC_ALL=en_US.UTF-8 (in ~/.zshrc). Without it the sass build dies with Invalid US-ASCII character on non-ASCII chars in the styles.
Gemfile.lock was re-resolved (bundle update) off the original Ruby-2.6-era pins to versions compatible with Ruby 3.3. GitHub Pages ignores Gemfile.lock and builds with its own bundle, so this only affects local dev.

Git workflow for this repo

Per global rules: branch → PR → merge even for solo edits; keep main clean. Tag milestones, e.g. release/<event> when refreshing for a conference or job-market push.

Zhaoyang Chu (储朝阳)