Limitations

leech is deliberately narrow. Knowing what it does not do will save you time.

LeetCode side

Unofficial API. All requests go to POST /graphql, which LeetCode does not document or guarantee. It can change without notice. (The old GET /api/submissions/ endpoint no longer exists — leech doesn't use it.)
Session cookies expire. LEETCODE_SESSION is a browser session cookie. When it expires, the action fails with a "session cookie is likely invalid or expired" error. You must refresh both secrets from DevTools. There is no long-lived token.
Locked/premium problems are dropped when question filters are active. Difficulty and tag filters need problem metadata; if a problem's metadata can't be fetched (locked, premium, or deleted), it's dropped rather than guessed at. With only submission-level filters, it never needs the metadata.
Rate limits are real. leech is single-threaded and sleeps client.delayMs between calls (default 250 ms). Large first-time syncs take a while. Raise delayMs if you're throttled.
Chinese site is a host switch, not a different schema. site: leetcode.cn changes the host, URLs, and asset source; cookies and schema are the same.

One commit per submission. This is the design, but it does mean a first-time sync of years of history produces a large number of commits, all in a single push. There is no squash option.
Re-syncs overwrite files with the same path. An improved solution replaces the old file. To preserve every attempt, include submission.id (or another unique key) in the filename.
Changing commit.prefix resets the watermark and re-syncs everything. Pick a prefix and keep it, or point it at your previous tool's commit header to inherit history.
Requires permissions: contents: write. Without it the ref update fails.
force: false. If the branch moves between the watermark scan and the push, the run is rejected rather than clobbering the new commits. Use a concurrency group (and cancel-in-progress: false) to serialize runs.
@main doesn't work as a ref. dist/ is built and attached to release tags, not committed to the branch. Reference a tag (@<ActionTag />).

No Windows. leech targets ubuntu-latest; bash is the default shell.
No repository checkout by default. Hooks run in the action's environment. If a hook needs the repo's files, add actions/checkout to the workflow first.
No inherited secrets. The child environment is scrubbed of INPUT_*, GITHUB_TOKEN/GH_TOKEN, cookies, and ACTIONS_* tokens. Pass what a hook needs explicitly via env.
A failing hook can abort the whole run. With onError: fail (the default), one bad command means nothing is pushed. Use warn or skip where that's too strict.

config must be one inline YAML string. GitHub Actions only accepts flat key-value inputs, so a structured YAML mapping is silently treated as unknowns. Keep the | and the indentation.
No nested {{ }} inside filter arguments. Filter arguments are plain expressions: use codeBlock(submission.lang), not codeBlock({{ submission.lang }}).
Undefined variables render empty by default. Set render.throwOnUndefined: true while developing templates so typos fail loudly.
Root-absolute asset paths are a markdown/SSG concept. In Typst, #image("/images/...") is a filesystem-absolute path, so asset rewriting is of limited use there.

Node ≥ 20 for the CLI; the action runs on the runner's Node 24.
Nix users: when pnpm-lock.yaml changes, nix build reports the new dependency hash to paste into flake.nix's fetchPnpmDeps.
Not affiliated with LeetCode. Use at your own risk, and mind LeetCode's terms of service.