gu-log
A bilingual AI/tech blog — great English writing, retold in Traditional Chinese (and back).
English · 繁體中文
What is gu-log?
gu-log takes the best AI / agent / tooling writing on the internet — X threads, blog posts, HN discussions, docs — and retells it in clear Traditional Chinese, with the original always linked. It also publishes original pieces and beginner tutorials. Every article ships in both zh-tw and en.
The name: gu = 菇 (mushroom) from ShroomDog. So gu-log = 菇 log = a mushroom's notebook. 🍄
Good content shouldn't be gated by language — and translating it is how we learn it too.
The cast
Article series
Every article carries a ticket ID so you can tell at a glance who picked it and why.
| Prefix | Series | Who picks | Who writes |
|---|
| GP | Gu-log Picks | ShroomDog, from published MP | Mogu, same as MP: the signature articles |
| MP | Mogu Picks | Mogu (self-selected) | Mogu writes from source material |
| SD | ShroomDog Original | ShroomDog | ShroomDog writes |
| Lv | Level-Up | — | beginner tutorials |
Quality: a two-tier bar
gu-log is a blog about AI quality, so it puts its own AI self-scores in the open — even the bad ones. Quality is gated in two layers, not one hard wall:
| Tier | Bar | Enforced by | If it doesn't pass |
|---|
| Floor (auto-gate) | real scores.vibe + required Vibe dimensions for that tribunal version + composite ≥ 3 | pre-commit hook | commit blocked — garbage never reaches main |
| PASS (editorial) | full tribunal pass bar: Vibe composite ≥ 8 with one dim ≥ 9 and no dim < 8; Fact Check, Librarian, and Fresh Eyes hard gates also pass | homepage / UI filter | still ships, but with a "refining" badge and kept off the homepage until a background pass lifts it to PASS |
Scoring runs through a 4-judge tribunal (each article, newest-first). This list is a derived view; model routing is defined in the codex-tribunal-runtime spec, so docs do not duplicate model names:
- Vibe Scorer — v9 dimensions: Persona / MoguNote / Vibe / Narrative; v8 and below also include Clarity
- Fact Checker — technical accuracy, source fidelity, logical consistency
- Librarian — glossary, cross-refs, attribution, source alignment
- Fresh Eyes — a stranger's first impression; v9 owns Clarity as a non-compensating gate
Anything sub-8 gets queued for a background rewrite (up to 3 rounds) instead of blocking the ship.
Tech stack
- Framework — Astro (Content Collections + MDX; version
declared in
package.json)
- Hosting — Vercel (auto-deploy on push to
main)
- Package manager — pnpm (the only supported one;
pnpm-lock.yaml is the source of truth)
- Fonts — Inter + Noto Sans TC
- Theme — Solarized (light) / Dracula-ish (dark), via CSS variables
Local development
pnpm install # install deps (frozen lockfile in CI)
pnpm run dev # dev server at localhost:4321
pnpm run build # production build (catches render errors)
pnpm exec astro check # TypeScript + template type checking
pnpm run validate:posts # frontmatter & content policy
pnpm run content:check # validate:posts + build, in one shot
Project structure
| Path | What lives there |
|---|
src/ | The Astro site: posts in src/content/posts/ (schema: src/content.config.ts), components, layouts, pages |
scripts/ | Content validators, git hooks, the Tribunal runner and other tooling |
tools/gp-pipeline/ | Go CLI for the source-to-post writing pipeline |
tests/ | Vitest unit tests and Playwright end-to-end specs |
openspec/ | Specs and change history |
playbooks/, docs/ | Agent playbooks, runbooks and references |
.github/workflows/ | CI, nightly and deploy workflows |
The full source layout and commands live in docs/dev-reference.md.
Quality gates & CI
CI is layered to keep PRs fast while still catching everything overnight.
Layer 1 — PR gate (.github/workflows/ci.yml, blocking):
the required check is a single ci-passed job that waits for every job in its needs: list — lint, content validation, security, unit and end-to-end tests, build and more. That list is the source of truth for what a PR must pass.
Layer 2 — nightly deep check (.github/workflows/nightly-deep.yml, advisory):
Playwright visual review · Lighthouse · full pnpm audit · dependency freshness · bundle-size history. Failures ping Telegram.
Layer 3 — post-deploy smoke test (.github/workflows/deploy-smoke-test.yml):
After each Vercel production deploy, checks the site is live, articles render, and counts match.
Two extra blocking policies worth knowing:
- Security gate (
pnpm run security:gate) — any new high/critical vulnerability fails the PR unless it has a valid, expiring entry in quality/security-allowlist.json (runtime: ≤ 14 days, dev: ≤ 45 days).
- Bundle budget (
scripts/bundle-budget-check.mjs) — global JS/CSS and single-file sizes are blocking; HTML/total/route sizes are warn-only trend monitors with growth-rate alerts.
No --no-verify. If a hook fails, the fix is to fix the code or fix the hook — never to skip it.
Contributing & docs
These are the sources of truth — read them before editing content:
