| name | fix-bank |
| description | Fix typos / garbled text in a hema-2026 exam question reported through the in-app feedback button (a GitHub issue). Use when the user says "fix issue N", "fix the feedback", or points at a mcq-bank issue about a question having 錯字 / wrong text. This corrects wording only — for wrong ANSWERS use the verdict-by-oe skill instead. |
fix-bank
Repeatable process for fixing a question's text (typos, OCR garble, wording)
reported via the in-app 回報 button, which files a GitHub issue in the repo set by
GH_FEEDBACK_REPO in wrangler.toml (currently htlin222/mcq-bank).
Scope guard: this skill only rewrites stem / options_json. It does not
change answer. If the report disputes the correct answer, stop and use the
verdict-by-oe skill instead.
Key facts
- D1 database name: read it, never hardcode —
grep database_name wrangler.toml
(currently hema-2026-db). Binding is DB.
- Questions table columns:
id (e.g. 114-022), year, number, stem,
options_json, answer, group.
- options_json shape:
[{"key":"A","text":"..."},{"key":"B","text":"..."}, ...].
- Remote D1 needs a token: load it from
.env before every remote command:
set -a && . ./.env && set +a.
- The reported URL looks like
.../q/114-022 — the question id is the last segment.
- The issue owner comment often contains the corrected text — treat that as
the source of truth, then additionally fix any residual obvious typos.
Steps
1 — Read the issue
gh issue view <N> --repo htlin222/mcq-bank --json title,body,labels,comments
Extract the question id from the URL in the body (segment after /q/).
Read every comment: the owner frequently pastes the intended stem + options.
2 — Read current DB content (remote is prod)
DB=$(grep -m1 database_name wrangler.toml | sed 's/.*"\(.*\)".*/\1/')
set -a && . ./.env && set +a
wrangler d1 execute "$DB" --remote --json \
--command "SELECT id, stem, options_json, answer FROM questions WHERE id='<ID>'" 2>/dev/null
Diff the DB text against the owner's corrected text. Decide the final wording:
- Base = owner's corrected text if provided, else your best reconstruction.
- Also fix residual unambiguous errors (medical terms, 錯字). Common ones seen:
Cobolamin→Cobalamin, homocystein→homocysteine, 筆因→肇因,
ASMATS13→ADAMTS13. When a fix is a judgement call, still apply it but list
it explicitly in the close comment so the owner can object.
3 — Write an UPDATE .sql file
Write to the scratchpad (avoids shell-escaping hell). JSON has no single quotes,
so wrap values in single quotes. Use full-width Chinese parens （） as the owner does.
UPDATE questions
SET stem = '<corrected stem>',
options_json = '[{"key":"A","text":"..."},{"key":"B","text":"..."},{"key":"C","text":"..."},{"key":"D","text":"..."},{"key":"E","text":"..."}]'
WHERE id = '<ID>';
Do not touch answer (see scope guard).
4 — Apply to remote, then verify
set -a && . ./.env && set +a
wrangler d1 execute "$DB" --remote --file <scratchpad>/fix-<ID>.sql 2>/dev/null
wrangler d1 execute "$DB" --remote --json \
--command "SELECT stem, options_json FROM questions WHERE id='<ID>'" 2>/dev/null \
| python3 -c "import sys,json; r=json.load(sys.stdin)[0]['results'][0]; print(r['stem']); [print(o['key'],o['text']) for o in json.loads(r['options_json'])]"
Confirm Rows written = 2 (stem + options_json) and the printout reads correctly.
5 — Mirror to local D1
Keeps .wrangler/state in sync so local dev shows the fix.
wrangler d1 execute "$DB" --local --file <scratchpad>/fix-<ID>.sql 2>/dev/null
6 — Close the issue with a changelog
gh issue close <N> --repo htlin222/mcq-bank --comment "已修正 <ID> 的錯字。
<one line on the root cause, e.g. OCR 亂碼 / 個別錯字>
主要修正：
- \`old\` → \`new\`
...
答案維持 <X>。remote + local D1 皆已更新。"
Notes
- This edits D1 directly — there is no code change to commit.
- Re-importing questions from CSV can clobber DB edits (see the
import-clobbers-promoted-answer memory). If the fix must survive a re-import,
also update the source CSV/import for that id.
- Only touching text → answer stays. Any answer dispute → hand off to
verdict-by-oe.
