@cloudhal/comments

First-party reader comments for Cloudhal sites (RFC-0007, ADR-0022). Guest comments are held for moderation, live in D1 outside the site graph, are server-rendered into the cached article page and purged by tag (cloudhal-comments-<nodeId>) within the Free-tier purge budget. Bots pay with a first-party proof of work; Turnstile is opt-in.

Everything here is Worker-safe: no Node-only import at any entry. The stored and returned record is the Comment contract of @cloudhal/core-contracts; no API returns an email address, an IP hash or a challenge outside the commenter accessors (invariant C1).

Settings and thread state (§4.5)

import { applySettingsInput, defaultCommentSettings, threadClosesAt, threadState } from "@cloudhal/comments";

const settings = await store.settings(siteId); // the row, or the defaults (off, rev 0)
threadState({ settings, entry: { nodeId, collection: "posts", publishedAt }, thread, hasApproved, now });
// → "open" | "closed" | "hidden" | "off"   (90 days after first publication by default)

CommentSettingsInput is validated by applySettingsInput (unknown keys refused; CommentSettingsError, code invalid_settings). collectionEnabled, closeAfterDaysFor, threadVisible and DEFAULT_RETENTION are exported too.

Text (§6.4, §8.1)

• normalizeText, normalizeName, normalizeEmail, checkText, checkName, checkEmail, codePoints, cutCodePoints, CONTRACT_FORBIDDEN — NFC, CRLF → LF, control and bidi characters removed (the contract refuses them), lengths in code points.
• evaluateContent({ name, text }, { linkLimit, watchWords, blockedWords }) → { ok, status: "pending" | "spam", labels } or { ok: false, field, problem }; compileWordList, matchesWordList, wordTokens, foldForMatch (NFKC, Turkish İ/I/ı folded together, whole words), countLinks, instructionLikeRules/isInstructionLike (INSTRUCTION_RULES_VERSION, English and Turkish), ruleLabels.
• Links: parseCommentText(text) → paragraphs → lines → {kind:"text"}|{kind:"link", href} segments; safeCommentHref, linkifyCommentLine, linkify; commentTextHtml(text) for server-rendered HTML (escaped, <p>/<br>, rel="ugc nofollow noopener noreferrer" = COMMENT_LINK_REL). Only https: URLs with a dotted host and no userinfo become links. The rules match @cloudhal/ui's interim comment-model.ts; the fixture pins them.

Proof of work (§7)

import { deriveCommentKeys, difficultyFor, issueChallenge, verifyChallenge, ReplayCache } from "@cloudhal/comments";
const keys = await deriveCommentKeys(env.CLOUDHAL_COMMENTS_KEY); // HKDF per purpose: pow, emailHmac, email, secrets
const { challenge, bits, expiresAt } = await issueChallenge(keys.pow, { siteId, target: nodeId, kind: "comment", bits: difficultyFor(today, cap), now });

v1.<payload>.<mac>, stateless, 10 minutes, ≥ 3 s fill time, 18 bits rising to 20/21 with the daily cap. solveChallenge is the reference solver (tests, the browser worker's algorithm). Single use: ReplayCache per isolate plus the unique index on challenge_id.

The pipeline (§6.2–§6.6)

import { handleCommentSubmission, handleReportSubmission, handleChallengeRequest,
         submissionResponse, challengeResponse, CommentLimiter, DailyCount, ReplayCache, memoize } from "@cloudhal/comments";
import { SqlCommentStore } from "@cloudhal/comments/sql";

// once per isolate
const limiter = new CommentLimiter(), daily = new DailyCount(), replay = new ReplayCache();
const settings = memoize(() => store.settings(env.CLOUDHAL_SITE_ID)); // 60 s

const deps: PipelineDeps = {
  siteId: env.CLOUDHAL_SITE_ID, available: true, hosts, clientAddress: request.headers.get("cf-connecting-ip")!,
  keys, now: () => new Date(), settings, resolveEntry, store, limiter, daily, replay,
  rateLimit?: (key) => …,           // optional Rate Limiting binding
  turnstile?: ({ token, remoteip, idempotencyKey }) => siteverify(…), // when the site uses Turnstile
};
return submissionResponse(await handleCommentSubmission(request, deps));
// reports: submissionResponse(await handleReportSubmission(request, deps), { pendingFragment: "comment-reported" })
// challenges: challengeResponse(await handleChallengeRequest(request, deps))

resolveEntry(nodeId) returns the published entry ({ nodeId, collection, path, publishedAt }) or null. Outcomes are { kind: "pending", stored, dropped? } (honeypot, bad proof, replay, duplicate and bans are indistinguishable from a stored comment) or { kind: "error", status, error, fields?, retryAfter? }; form posts answer 303 to #comment-pending, #comment-invalid, #comment-closed, #comment-script-required, #comment-slow-down, #comment-try-again, #comment-paused. Nothing is written before step 12; a month's first IP salt is inserted inside step 12's batch.

Store (@cloudhal/comments/sql)

new SqlCommentStore(sql, { now? }):

approve, restore, delete, replies, erasure and bans need a human actor (CommentForbiddenError, code forbidden); stale revs are CommentConflictError (conflict) for single writes and per-comment conflict results for bulk writes. audit statements are appended to the same batch, last. Statement counts per path are in the store's header and asserted by the conformance suite.

Purge planner and retention (§8.4, §9.3)

planPurge({ rows, recent, now }) → { requests, nextCheckAt, takedownWaiting }: takedowns at once (normal tags ride along), normal tags after 20 s and only while the last minute shows fewer than 3 purge requests (publish included), 100 tags a request. purgeBackoffMs, settingsPurgeTags (or "everything" above 300 entries), commentsCacheTag, PURGE_LIMITS. retentionPlan lists the bounded steps (500 rows a run).