GitCode (Issue / PR)

Project-local skill for sciencediscovery on gitcode.com (not under skills/).

New issues and pull requests for this repository are not filed here. Use gh against openJiuwen-ai/sciencediscovery, and write the title and body in English. This skill is for an existing GitCode mirror item that still has to be read or commented on.

Use the command table below, gitcode schema, or gitcode <cmd> --help to resolve flags. Only if a flag is still unclear, run gitcode api repos/gitcode-cli/cli/contents/docs, select the download_url for COMMANDS.md, AUTH.md, or AI-GUIDE.md, and read that raw.gitcode.com content. Do not scrape GitCode blob-page HTML.

Rules

1. Run gitcode, never bare gc (often a git commit alias).
2. Install/helpers via uv; never commit CLI binaries or tokens.
3. Never print tokens (auth token / auth status --show-token).
4. Prefer --json; if create/edit JSON is thin (missing html_url or with the wrong number), confirm with view/list.
5. For a complete Issue inventory or count, never rely on issue list's default --limit 30. Pass an explicit limit larger than the expected result set, check the returned JSON array length, and continue with --page 2, --page 3, … whenever a page length equals the requested limit. Stop only after a shorter page, then deduplicate by Issue number before counting.
6. Destructive ops (close/reopen/merge/delete…) need user intent + --yes.
7. Scope: only repos the user named (or cwd remote when clearly that clone).
8. Transports: metadata/API → gitcode CLI; code/MR head → SSH git. HTTPS git easily fails (hangs without a credential helper; web/HTTPS fetches commonly return 403), so always use SSH and do not scrape gitcode.com HTML.
9. Follow Cross-references and issue association for visible links and unverified auto-close behavior.
10. sciencediscovery merge requests: each operator uses their own GitCode fork of openJiuwen/sciencediscovery (do not hard-code a login). Resolve <gitcode-login> from gitcode auth status --json. Push the source branch to that fork (common remote name gitcode-fork) and create with --head <gitcode-login>:<branch>. Do not git push origin <task-branch> onto openJiuwen/sciencediscovery. If the fork is missing: gitcode repo fork openJiuwen/sciencediscovery --json.

Hosts

gitcode.com and atomgit.com are hostnames for the same website/platform: repositories and their issue/PR numbers are identical. Do not treat AtomGit as a separate product. For authored or reported repo, issue, and PR URLs, prefer https://gitcode.com/...; rewrite AtomGit hosts and CLI html_url paths such as /merge_requests/N to these canonical forms:

• Issue: https://gitcode.com/openJiuwen/sciencediscovery/issues/9
• PR list: https://gitcode.com/openJiuwen/sciencediscovery/pulls
• PR detail: https://gitcode.com/owner/repo/pull/N
• SSH: git@gitcode.com:owner/repo.git and git@atomgit.com:owner/repo.git are equivalent; prefer writing/configuring the gitcode.com form, even when an existing remote uses AtomGit.

Install / auth (once)

uv tool install gitcode-cli
export PATH="$HOME/.local/bin:$PATH"
gitcode version
gitcode auth status --json   # need logged_in + token_valid

Token priority: GC_TOKEN > GITCODE_TOKEN > ~/.config/gc/auth.json. No browser on box → user creates classic token at https://gitcode.com/setting/token-classic and runs gitcode auth login on a private TTY (do not ask them to paste token into chat). Fallback install: uv venv .local/gitcode-venv && uv pip install --python .local/gitcode-venv gitcode-cli. Binary resolve: command -v gitcode → $HOME/.local/bin/gitcode → .local/gitcode-venv/bin/gitcode. PyPI stale: uv tool install --from 'https://gitcode.com/gitcode-cli/cli/releases/download/<ver>/gitcode_cli-<ver>-py3-none-any.whl' gitcode-cli.

Targeting

Always pass -R owner/repo when not on cwd origin. Forms: owner/repo, HTTPS URL, SSH URL (CLI API only — not git transport).

Need	Use
Title/body/state/comments/create/edit/review	gitcode …
Patch summary	gitcode pr diff N (text preferred; binaries have no preview)
Full tree / deep review	SSH remote git@gitcode.com:owner/repo.git; fetch refs/merge-requests/N/head
Scratch clone	.tmp/<repo-name>/
Long bodies	.tmp/*.md + --body-file

Issue / PR commands

# issues
gitcode issue create -R owner/repo --title "…" --body-file body.md --json
# Complete inventory: 30 is the CLI default and is commonly truncated.
gitcode issue list -R owner/repo --state open --limit 100 --page 1 --json
gitcode issue view N -R owner/repo --json
gitcode issue view N -R owner/repo --comments --json
gitcode issue edit N -R owner/repo --body-file body.md --json
gitcode issue comment N -R owner/repo --body "…" --json
gitcode issue comments N -R owner/repo --json
gitcode issue comment edit <comment_id> -R owner/repo --body "…"   # may not support --json
gitcode issue close N -R owner/repo --yes --json
gitcode issue reopen N -R owner/repo --yes --json

# PRs — sciencediscovery: source branch on the operator's own fork, never on origin
gitcode pr create -R openJiuwen/sciencediscovery --head <gitcode-login>:branch --base main --title "…" --body-file pr.md --json
gitcode pr create -R upstream/repo --head myfork:branch --title "…" --body-file pr.md --json
gitcode pr create -R owner/repo --head branch --fill --json
gitcode pr create -R owner/repo --head branch --title "WIP" --draft --json
gitcode pr list -R owner/repo --state open --json
gitcode pr view N -R owner/repo --json
gitcode pr view N -R owner/repo --comments --json
gitcode pr diff N -R owner/repo
gitcode pr edit N -R owner/repo --body-file pr.md --json
gitcode pr comment N -R owner/repo --body "…" --json
gitcode pr comment N -R owner/repo --body "…" --path path/to/file.py --position 12 --json
gitcode pr comments N -R owner/repo --json
gitcode pr reply N -R owner/repo --discussion <id> --body "…"
gitcode pr review N -R owner/repo --comment "…" --json   # or --comment-file
gitcode pr merge N -R owner/repo --yes --json            # optional --method squash|rebase
gitcode pr close|reopen|ready N -R owner/repo --yes --json
gitcode pr checkout N -R owner/repo
gitcode repo fork owner/repo --json
gitcode repo view owner/repo --json
gitcode schema
gitcode schema "issue create"

PR view JSON shapes

gitcode pr view changes its top-level JSON shape when comments are requested:

• gitcode pr view N -R owner/repo --json returns the PR object directly.
• gitcode pr view N -R owner/repo --comments --json returns { "pull_request": <PR object>, "comments": [...] }.

Do not read .number, .head, or .labels at the top level of the comments form; those fields are below .pull_request. Optional arrays such as labels may be null, so normalize them with // [] before iterating.

# PR metadata only
gitcode pr view N -R owner/repo --json |
  jq '{number, title, state, head_sha: .head.sha,
       labels: [((.labels // [])[]) | .name]}'

# PR metadata and comments
gitcode pr view N -R owner/repo --comments --json |
  jq '{pr: (.pull_request | {number, title, state, head_sha: .head.sha}),
       comments: [(.comments // [])[] |
         {author: (.user.login // .author.login // .author_name),
          body: (.body // .note // ""), created_at}]}'

Use gitcode pr comments N -R owner/repo --json when only the comment list is needed.

For complete Issue inventories, treat length == --limit as “possibly truncated,” not as a final count. Fetch subsequent pages with the same filters until one returns fewer rows than the limit; combine the pages and deduplicate by .number. If the first page returns fewer than the explicit limit, it is complete for those filters at that retrieval time.

• Line comments: --position = line on the new file (right side of diff). Inline comments appear as comment_type: diff_comment.
• --approve: only with explicit user intent + approval permission; own-PR / missing role → 403 — leave a comment review instead.
• --request (request changes) unsupported; put change requests in --comment.
• Merge only with explicit user OK for that PR.
• pr create 409 / "same source branch already has an open MR !N": !N may be the PR this very call just created (observed on this repo), an earlier PR opened from the same branch, or one you should not touch. Never retry blindly and never rename the branch to dodge it. Read it back — gitcode pr view N -R owner/repo --json — compare base, head branch, head SHA, title, body, author, created time, and existing review activity against this attempt, then act on what !N actually is. Typical outcomes: this attempt landed ⇒ report !N; your own re-delivery on the same branch, where the MR already tracks the pushed head but title/body are stale ⇒ bring it up to date with gitcode pr edit N --body-file … instead of creating a second PR; head SHA still behind your push ⇒ confirm the push reached the fork before touching the PR; wrong base, another author, or review already under way ⇒ stop and report so the user can choose (edit, close, or a fresh branch). Judge from the read-back, not from the error text.

Read a PR (review prep)

1. gitcode pr view N -R owner/repo --json — title, body, SHAs, state
2. gitcode pr diff N -R owner/repo — patch overview; prefer text output, and expect no binary preview
3. gitcode api repos/owner/repo/pulls/N/files — structured file list / raw URLs
4. Full tree: SSH git fetch git@gitcode.com:owner/repo.git refs/merge-requests/N/head (GitCode MR ref; not HTTPS). Prefer this over scraping the web UI.

Cross-references and issue association

• Ordinary issue/comment reference: [#32](https://gitcode.com/openJiuwen/sciencediscovery/issues/32).
• Ordinary PR reference: [#10](https://gitcode.com/openJiuwen/sciencediscovery/pull/10). Write /pull/N in authored links even if CLI html_url says /merge_requests/N.
• To associate an issue in a PR body, retain the trigger candidate and add a readable link: Fixes #32 ([#32](https://gitcode.com/openJiuwen/sciencediscovery/issues/32)).
• Automatic closing after merge is not verified. Never promise that Fixes #N or its Markdown link will close the issue; commit messages alone do not auto-close.

openJiuwen org templates

Template source order when filing an Issue/PR:

1. This repo's own .gitcode/ Issue/PR templates, if present (openJiuwen/sciencediscovery currently has none).
2. Otherwise the openJiuwen org templates below — this is the default source, not an opt-in path.

Language: a GitHub issue or pull request is English and is created with gh — see github. This section applies only when you are still editing an item on the GitCode mirror. There, default to the Chinese templates (ISSUE_TEMPLATE.zh/ + PULL_REQUEST_TEMPLATE.md / PULL_REQUEST_TEMPLATE.zh-CN.md). Use the English set (ISSUE_TEMPLATE.en/, PULL_REQUEST_TEMPLATE.en.md) only when the person who asked explicitly wants English on that mirror item.

Upstream (do not vendor into git): openJiuwen/.gitcode @ master, tree .gitcode/ (web: https://gitcode.com/openJiuwen/.gitcode). Local cache (gitignored): {baseDir}/cache/openjiuwen-org-templates/.

FETCH="{baseDir}/scripts/fetch_openjiuwen_templates.py"
uv run --no-project "$FETCH"          # ensure cache exists
uv run --no-project "$FETCH" --force  # refresh after upstream changes

Layout is split per locale — there is no single ISSUE_TEMPLATE/ directory:

.gitcode/ISSUE_TEMPLATE.zh/*.yml        # Chinese issue forms (default)
.gitcode/ISSUE_TEMPLATE.en/*.yml        # English issue forms (+ config.yml)
.gitcode/PULL_REQUEST_TEMPLATE.md
.gitcode/PULL_REQUEST_TEMPLATE.zh-CN.md
.gitcode/PULL_REQUEST_TEMPLATE.en.md

Do not rely on any template inventory in this skill — lists go stale. At filing time:

1. Ensure cache is present (run fetch if needed).
2. List and read the actual files under …/cache/openjiuwen-org-templates/.gitcode/ (ISSUE_TEMPLATE.zh/, ISSUE_TEMPLATE.en/, PULL_REQUEST_TEMPLATE*). FETCH_META.json records what the last fetch wrote (issue_template_dirs, pr_templates), but read the files themselves before filling one in.
3. Pick the matching template for the intent from the language directory chosen above; fill required fields from that file; create with --body-file.
4. If the issue template is an issue form (YAML), convert it to Markdown in body order: write each attributes.label as a ### heading and put the answer below it; every item with validations.required: true is mandatory.
5. config.yml is issue-form chooser config, not a template — never file it as an issue body.

Images in bodies/comments

UPLOAD="{baseDir}/scripts/upload_image.py"
uv run --no-project "$UPLOAD" -R owner/repo .tmp/image.png --json

1. List the source directory first to confirm the actual filename; do not retype a path from memory. Then copy archived/non-ASCII paths to a simple ASCII name under .tmp/ before upload.

2. Prefer --json; continue only when the command exits 0 and stdout is a single JSON object with non-empty markdown and url fields (success prints no progress). The URL must use https://raw.gitcode.com/user-images/assets/…, never bare /uploads/…. Errors/warnings go to stderr only.

3. On non-zero exit or stderr containing error:, stop. Never paste stderr/error text into an issue/PR body or replace a failed image placeholder with an empty string.

4. Build comments and long bodies in .tmp/*.md without shell expansion (use an editor/patch or a single-quoted heredoc), then pass the file with --body-file; do not use $(cat <<EOF) around Markdown.

gitcode pr comment N -R owner/repo --body-file .tmp/comment.md --json

5. After posting, read back with gitcode issue view N -R owner/repo --comments --json or gitcode pr comments N -R owner/repo --json. Confirm the comment contains the expected raw.gitcode.com embed and contains no error: file not found.

Fetching images (raw.gitcode.com)

raw.gitcode.com user-images URLs require auth — anonymous GET returns 403 no access right (verified platform behavior, not a broken link). Use the skill's download script, which reuses the same token resolution as upload_image.py and sends Authorization: Bearer <token>:

DOWNLOAD="{baseDir}/scripts/download_image.py"
# Bare URL or Markdown embed ![alt](url 'title'); saves to .tmp/<filename> by default
uv run --no-project "$DOWNLOAD" "https://raw.gitcode.com/user-images/assets/<repo_id>/<uuid>/<file>" -o .tmp/issue44_img1.png --json
uv run --no-project "$DOWNLOAD" "![alt](https://raw.gitcode.com/user-images/assets/.../x.png 'x')" -o .tmp/x.png

1. The script accepts a bare https://raw.gitcode.com/... URL or a full Markdown embed ![alt](url 'title') (e.g. copied straight from an issue body) and extracts the URL.
2. Prefer --json; continue only when the command exits 0 and stdout is a single JSON object with "success": true, status: 200, and a content_type starting with image/ (success prints no progress). The saved file path is in path.
3. Default output is .tmp/<url-filename> if .tmp/ exists, else ./<url-filename>; pass -o to override. Non-ASCII filenames (e.g. 企业微信截图_*.png) are kept as-is — copy to an ASCII name before review if needed.
4. On non-zero exit or stderr containing error:, stop. Exit 4 means auth/403 (token missing, invalid, or expired — run gitcode auth status --json); exit 1 is a transport/server error. Never treat a 403 as a broken image link.
5. To inspect images referenced in an issue/PR body, gitcode issue view N --json / gitcode pr view N --json and extract raw.gitcode.com URLs from the body field before downloading.

Wiki

No gitcode wiki subcommand. Wiki is owner/repo.wiki (SSH git@gitcode.com:owner/repo.wiki.git), default branch main, entry Home.md. Prefer SSH.

Checklist

1. gitcode version + auth status --json
2. Lock -R
3. Read before write (view/list/diff; code via SSH MR ref) — no HTML scrape
4. Follow the cross-reference rules above; do not promise unverified auto-close behavior
5. Confirm number/URL after create/edit; report /pull/N or /issues/N
6. Failures: stderr + exit class (0 ok, 1 generic, 2 usage, 3 not found, 4 auth, 5 conflict) — conflict (e.g. pr create 409) means read back the cited !N and decide from its actual state, not from the message; never invent success

Policy: skill sources under .agents/skills/gitcode/ are tracked; cache/ is not. No binaries, no vendored templates, no secrets in skill files.