oss-docs Verify Recipe (exit gate)

Runnable checks for the sprint-contract dimensions. The scripts docs-i18n-check.sh and gen_menu.py DO NOT exist — never shell out to them; every check is inlined below. All checks are read-only; this skill never commits or pushes.

1. Build clean (build-clean, must_pass, threshold 1.0)

cd docs-site && hugo --minify --gc

• Must exit 0 AND complete warning-free (any WARN/ERROR line = FAIL).

test -f docs-site/public/sitemap.xml && echo "sitemap OK" || echo "sitemap MISSING"

2. URL blacklist (content-fidelity)

grep -rn 'docs\.moai-ai\.dev\|adk\.moai\.com\|adk\.moai\.kr' docs-site/content README*.md

• Expected: no matches. Only adk.mo.ai.kr is valid. Note: the pattern adk\.moai\.kr does not match adk.mo.ai.kr (different dot positions) — no false positive on the valid domain.

3. Mermaid direction (style-compliance)

grep -rn 'flowchart LR\|graph LR\|flowchart RL\|graph RL' docs-site/content

• Expected: no matches (TD-only rule; flowchart TD / graph TB pass).

4. 4-locale parity (locale-parity, must_pass, threshold 1.0)

File-existence parity — every ko page has en/ja/zh counterparts:

cd docs-site/content && for f in $(cd ko && find . -name '*.md'); do
  for loc in en ja zh; do
    [ -f "$loc/$f" ] || echo "MISSING: $loc/$f"
  done
done

Section-count parity per page, ratcheted against a checked-in baseline.

Comparing tree totals is not a parity check: per-page divergences in opposite directions cancel, so a page where ko leads en nets out against a page where en leads ko and the total looks healthy. Compare each page against its own three counterparts instead.

The gate is a ratchet, not an absolute check. docs-site/.locale-parity-baseline lists the pages that already diverge; the gate fails on any divergent page NOT in that list. An absolute check would fail on every baselined page from the first run, and a gate that fails on day one gets switched off — which is worse than the weak check it replaces. Ratcheting means the debt is explicit and auditable, and it can only shrink.

cd docs-site/content

# Current divergence set: pages whose ko/en/ja/zh H2-and-deeper counts disagree.
# One grep pass over the whole tree — a per-file loop over 143x4 files does not
# finish inside a 2-minute budget.
grep -rc '^#\{2,\} ' ko en ja zh --include='*.md' \
| awk -F: '
    { i=index($1,"/"); loc=substr($1,1,i-1); page=substr($1,i+1)
      n[page,loc]=$2; pages[page]=1 }
    END { for (p in pages)
            if (n[p,"en"]!=n[p,"ko"] || n[p,"ja"]!=n[p,"ko"] || n[p,"zh"]!=n[p,"ko"])
              print p }' \
| sort > /tmp/parity-now.txt

grep -v '^#' ../.locale-parity-baseline | grep -v '^[[:space:]]*$' | sort > /tmp/parity-base.txt

comm -23 /tmp/parity-now.txt /tmp/parity-base.txt   # NEW divergence  -> FAIL
comm -13 /tmp/parity-now.txt /tmp/parity-base.txt   # converged pages -> prune baseline

Failure condition (explicit): the first comm prints one or more page paths. Any output there is a FAIL — a page that was previously in parity has lost it, or a newly added page landed unbalanced. Fix the page, or (only with a deliberate decision) add it to the baseline; adding a line is admitting new debt.

The second comm is informational: those pages have converged and should be pruned from the baseline so the ratchet tightens. Not pruning is not a failure.

A missing counterpart file also surfaces here (its count reads as empty and therefore disagrees), which overlaps with the file-existence check above — that redundancy is intentional.

README 4-file heading-count parity:

grep -c '^## ' README.md README.ko.md README.ja.md README.zh.md

• Expected: identical counts across the 4 files (and identical H2 order — spot-check with grep '^## ' <file>).

5. Body-emoji scan (style-compliance)

grep -rnP '[\x{1F300}-\x{1FAFF}\x{2600}-\x{26FF}\x{2700}-\x{27BF}]' docs-site/content --include='*.md' | grep -v '{{<' | head -40

• Review each hit: body-text emoji = FAIL (use {{</* icon */>}}); preserved typographic symbols (→ ← ↓ ✓ ✗, U+2702 in handoff blocks) and branding emoji inside orchestrator-banner example code blocks are allowed — judge code-block context before flagging.

6. Version-string sync (version-sync)

Version displays must equal the release number. Extract the expected version from the SSOT and screen for stale displays:

grep -E 'version = ' docs-site/hugo.toml    # expected version, e.g. v3.1.1
grep -rn 'Release-v[0-9]' README.md README.ko.md README.ja.md README.zh.md
grep -rn '🗿 v[0-9]' docs-site/content README*.md | grep -v "$(grep -oE 'version = "v[0-9.]+"' docs-site/hugo.toml | grep -oE 'v[0-9.]+')"

• Expected: every Release-v… badge and every 🗿 v… example equals the hugo.toml version. Any stale display (badge, statusline example, update-prompt example X ⬆️ Y, version-column example) = FAIL.
• Historical citations ("introduced in vX.Y.Z", "since v3.0.0", "retired in v3.0.0") are not displays — do not flag them.
• releaseDate must be bumped together with version (hugo.toml's own two-line contract).

Scoring map (sprint contract)

A must_pass dimension below threshold blocks the harness run result (must_pass_ok: false) — fix and re-verify before handing back to the orchestrator.