Consolidate the blog tag vocabulary onto the tier-2 category terms

[hotlong]

Why

46 posts have accumulated free-form tags (Ontology, Palantir, AI Governance, Open Protocol, Perspective, and their translated variants). Tag and cluster pages are the natural home for mainstream category keywords: they can rank for mid-tail terms without putting those terms in the homepage or the title template, which is exactly the tier-2 placement proposed in the keyword-architecture issue in this batch.

Today they cannot do that job, because the vocabulary is inconsistent across posts and across locales.

Depends on

The keyword-architecture issue in this batch — it fixes the tier-2 term list. Do not start this one until that list is settled, or the consolidation will have to be redone against a different vocabulary.

Scope

1. Inventory every tag currently in use across all 46 posts and all locales.
2. Map them onto the agreed tier-2 vocabulary; collapse synonyms.
3. Apply the consolidated set. Translated locales keep tags translated (AGENTS.md), so the mapping is per-locale — the English canonical term drives which tag applies, and each locale carries its own rendering of it.
4. Confirm the resulting tag/cluster pages emit correct canonical and hreflang, same as any other page.

Watch for

• zh-Hant is generated. Change zh-Hans tags and regenerate with pnpm gen:zh-hant; never hand-edit the Traditional files.
• Do not delete a tag that is the only route to a post. Check the cluster pages render non-empty after the change.
• Frontmatter fields other than title/description/tags stay untouched across locales — topic, audience, date, status must remain identical to the source.

Acceptance

• One consolidated vocabulary, documented on this issue
• pnpm content:lint, pnpm check, pnpm build, pnpm seo:smoke all pass
• Spot-check at least two cluster pages in the in-app browser, in English and one other locale

[hotlong]

Dependency cleared: the tier list is settled on #77 (maintainer 「同意」, 2026-08-26; tier table in the #77 report; carve-out in the #77 ACCEPT). Every wave-1 body edit has landed (#130, #103, #116, #86 on main at e5a0f00), so the frontmatter surface is quiet.

Seat rulings for this card: metadata driven development is never a tag — a tag cannot carry the AI-author qualifier the ruling requires, so drop it rather than ship it bare. Tier-2 terms are the tag vocabulary; tier-3 comparison terms (Palantir, Retool, …) stay only where the post is about that vendor. Only tags changes — title, description, topic, audience, date, status and bodies are untouched in every locale.

Claim: PM loop round 4 (wave 2)
Session: session_01FeA1nwBz1ohH65dvffUGKr
Branch: claude/issue-78-tag-vocabulary
Worktree: www.objectos.ai-issue-78
Domain: repo:www.objectos.ai (single-lane execution seat)
File surface: frontmatter tags of every content/blog/*/index*.mdx (en + authored locales), regenerated index.zh-Hant.mdx; src/lib/clusters.ts:43 adjective reading noted on #77 (decide, do not widen).
Container & model: M, mode:subagent, model: opus
Clause-②: no
Serial constraints cleared: #92's new post directory and the four open how-to PRs (#147 #151 #152 #153) carry their own tags on their own branches — out of this card's reach; the seat reconciles them against the documented vocabulary at their slot-day merge. #68/#69/#70 are held until this lands.

---

Generated by Claude Code

[hotlong]

Dev claim (implementation seat)

Session: session_01FeA1nwBz1ohH65dvffUGKr
Branch: claude/issue-78-tag-vocabulary (pushed empty as the write-route probe before the first edit)
Worktree: www.objectos.ai-issue-78
Repo: objectstack-ai/www.objectos.ai
File surface: frontmatter tags: of content/blog/*/index*.mdx only (en + authored locales), plus the regenerated index.zh-Hant.mdx.
Not touching: scripts/, astro.config.mjs, src/layouts/, src/content.config.ts (#104 / #79 / #150 own them this wave), and no body / title / description / topic / audience / date / status edits anywhere.

Working from the settled vocabulary on #77: the maintainer's 2026-08-26 「同意」 tier-1 ruling, the tier table in the #77 os-dev-report, and the title carve-out in the #77 ACCEPT. Per the seat ruling on this card, metadata driven development is never minted as a tag.

---

Generated by Claude Code

---

Generated by Claude Code

[hotlong]

os-dev-report

{
"issue": 78,
"status": "done",
"branch": "claude/issue-78-tag-vocabulary",
"pr": "#160",
"premise_still_valid": true,
"summary": "Replaced the free-form blog tag vocabulary with a closed one built from the #77 tier-2 ruling: 996 distinct tag strings across the eight locales become 172 (en 148 to 25). Seven tier-2 canonical terms, one authored rendering per locale, plus tier-3 proper nouns kept only on the article that compares against them; 'metadata driven development' deliberately not minted as a tag, per the seat ruling. Four commits so review can read them separately (en / zh-Hans / ja+de+es+fr+ko / zh-Hant regeneration). Two of the card's premises need correcting, neither fatal. (1) The corpus is 52 English posts, not 46. (2) There are NO tag pages: src/lib/posts.ts routes term hubs off topic, audience, solutions and industries only, so tags reach exactly three consumers (unlinked chips in ArticleCard and BlogPost, JSON-LD keywords, RSS categories). That makes the card's 'do not delete a tag that is the only route to a post' guard trivially satisfied, and means the consolidation buys consistency and a reconcilable vocabulary rather than rankings. The decisive evidence: the ai-native-app-platform cluster page still links all five of its articleSlugs, and four of those five are posts whose tags are now empty. Consequence the seat should see: 24 of 52 English posts (162 files across locales) now carry an empty tags list, because they contain no tier-2 term and compare against no vendor. That emptiness is a measurement, not a hole papered over, and is filed as #158; inventing tags would have hidden it. Cards keep their topic/audience/solutions/industries chips regardless, so nothing renders bare. The full vocabulary table (canonical English, per-locale rendering, collapsed former tags, post count) is in the PR body and reproduced below.",
"tests": "Gate union run through the shared lock AFTER the final commit, at head 2279f1d. Wrapper verdict: 'os-verify-lock: VERDICT command-exit 0 - held the lock 74s - waited 65s'. Exit status captured before any pipe (redirect to file, then read). Each gate's own verdict line: content:lint 'content lint passed (334 files, 44 glossary terms checked)'; check 'Result (133 files): - 0 errors'; build '[build] 866 page(s) built in 58.76s' plus '[build] Complete!'; seo:smoke 'SEO smoke test passed (865 HTML pages checked)'. git status --porcelain after build was empty, which proves the committed zh-Hant is byte-identical to gen-zh-hant output. Frontmatter invariance proved mechanically and re-run after each of the four commits: git diff -U0 origin/main filtered to added/removed lines that are not diff headers and not tags-block lines returned 0 lines, so title, description, topic, audience, date, status and every body byte are unchanged in all 334 files. Browser spot-check with Chromium at 1440x900 and 390x844, screenshots reviewed: /en/blog/topics/ai-agents/ 200 with 12 cards and consolidated chips, /de/blog/topics/ai-agents/ 200 with 7 cards showing Ontologie, MCP, KI-Agenten, Palantir, /en/ai-native-app-platform/ 200 linking 5 articles, the ontology-vs post rendering its three tags, and ai-expense-audit rendering an empty tag list without breakage. Canonical plus the full eight-locale hreflang set and x-default verified present in dist on all three hub pages; they are structurally independent of tags since Base.astro builds them from its canonical/alternates props. No changeset owed: this repo has no .changeset directory.",
"open_questions": [
{
"question": "Should the closed vocabulary stay at tier-2 plus tier-3 proper nouns only, or should a bounded third bucket of retained non-tier-2 terms be added back so that the 24 category-free posts still carry keywords?",
"options": [
"A. Ship as-is: vocabulary is exactly the seven tier-2 canonicals plus per-article proper nouns. 24 English posts carry an empty tags list. Fully mechanical to reconcile: every tag on a future PR must be one of the seven, or a vendor the post compares against.",
"B. Add a documented retained set (for example Self-hosted, Vibe coding, Low-code, Technical debt, Data sovereignty, Audit) so no post ends empty. Keeps more keyword signal, but mints vocabulary the maintainer never ruled on and reopens the failure mode this card exists to close, since the next author has no rule for what may join it.",
"C. Ship A now, and treat the empty posts as a content gap fixed by editing their descriptions and bodies to use the category vocabulary, which is #158."
],
"recommendation": "A, with C as the follow-up. B fails the very test that made this card necessary: the seat has to reconcile the four open how-to PRs and #92 against this record, and only a closed list makes that a mechanical check rather than a judgement call each time. A also keeps me from minting vocabulary outside the ruling. The cost of A is real but small and measured: tags are not routes here and never ranked anything, so what the 24 posts lose is JSON-LD keywords and RSS categories, not discoverability. If the seat prefers B, it is one commit per locale group to add terms back, and the vocabulary module in the PR is written to make that a data-only change."
}
],
"out_of_scope_findings": [
"filed as #158: 24 of 52 English posts contain no tier-2 category term in slug, title, description or tags; the retrieval gap #77 predicted, now counted",
"filed as #159: Japanese renders 'semantic layer' as the literal calque 意味層 rather than セマンティックレイヤー; kept in this pass because the rule was to preserve authored renderings, flagged for a native reviewer",
"NOT refiled, already open as #144: OpenCC s2twp renders 权限 as 許可權 in every generated zh-Hant page; pre-existing generator behaviour visible in the regenerated tags, not introduced here",
"repaired inside this card rather than filed, since it is a tags-only edit in the declared surface: the Japanese file for enterprise-ontology-race-open-vs-closed carried '微软 Fabric' (Simplified-Chinese characters for Microsoft) and the Korean carried '마이크로소프트 Fabric'; tier-3 proper nouns are now the vendor's own name in every locale",
"self-reported for the seat to clean up: filing #158 with a 'finding' label silently minted that label in this repo, which does not use it (its vocabulary is seo / pm:dispatched / glossary / content:backfill-en). I removed the label from #158, so it now exists with zero issues and can be deleted; no tool available to me deletes a repository label"
]
}

Vocabulary record — the documented consolidation this card asks for

Tier 2 — the category backbone

Canonical (en)	zh-Hans	ja	de	es	fr	ko	Former tags collapsed in	en posts
Ontology	本体	オントロジー	Ontologie	Ontología	Ontologie	온톨로지	Ontology, Enterprise Ontology, Ontology Actions	8
Semantic layer	语义层	意味層	Semantische Schicht	Capa semántica	Couche sémantique	시맨틱 레이어	Semantic Layer	4
MCP	MCP	MCP	MCP	MCP	MCP	MCP	MCP	4
Governed tools	受治理的工具	ガバナンス付きツール	Kontrollierte Werkzeuge	Herramientas gobernadas	Outils gouvernés	거버넌스 도구	Tool Calling, Tool calling, Tool Layer, Agent Tools	4
AI agents	AI 智能体	AI エージェント	KI-Agenten	Agentes de IA	Agents IA	AI 에이전트	AI Agent, AI Agents, Agent	15
Agent permissions	智能体权限	エージェント権限	Agent-Berechtigungen	Permisos de agentes	Permissions des agents	에이전트 권한	Permissions, Permission governance, Row-Level Security, Field Masking	4
Forward deployed engineer	前沿部署工程师	フォワードデプロイドエンジニア	Forward deployed engineer	Forward deployed engineer	Forward deployed engineer	포워드 디플로이드 엔지니어	Forward-Deployed Engineer	3

zh-Hant is the s2twp conversion of the zh-Hans column and is never hand-edited. metadata driven development is absent by ruling.

Tier 3 — named products, companies and protocols, per article only, untranslated

A2A · Agentforce ×2 · Airtable ×2 · Airtable Omni · Copilot Studio · Dataverse · Fabric IQ ×2 · Lovable · Microsoft Fabric · Palantir ×6 · Palantir Foundry · Power Apps · Power Platform · Replit · Retool · ServiceNow · Snowflake · Unity Catalog

The rule that closes the list — this is what to reconcile #147 / #151 / #152 / #153 and #92 against

Every tag on a post must be either one of the seven tier-2 canonicals, in that locale's rendering, or the proper noun of a product, company or protocol the post actually compares against. Nothing else qualifies. Generic category words (Low-code, Knowledge Graph, AI-Native, Open Protocol), editorial labels (Perspective, Governance, Metadata, Audit), per-vertical descriptors, and the site's own brand names (ObjectOS, ObjectStack) are all out — 62 former English tags survive on no post.

Posts that end with no tags (24 en, 162 files)

ai-compliance-control · ai-content-workbench · ai-contract-risk · ai-employee-service-center · ai-expense-audit · ai-procurement-risk · ai-project-risk-assistant · ai-ticket-hub · ai-wrote-your-app-dare-to-merge · automation-cross-system-flows · automation-governed-flow-changes · automation-pause-resume-approvals · automation-trigger-model · conversational-app-iteration · eu-ai-act-runtime-audit · extend-existing-systems-with-ai · from-requirement-to-app · low-code-vs-ai-native-app-platform · manufacturing-legacy-systems-ai · metadata-not-code-generation · objectos-automation-engine · self-hosted-ai-app-platform · vibe-coding-technical-debt-2026 · why-custom-systems-die

src/lib/clusters.ts:43 — the reading #77 asked for

"Metadata-driven screens, workflows, APIs and agent tools…" was read against the tier-1 constraint and left as-is. It is an adjective describing screens, not the bare noun phrase the ruling constrains, and the line sits outside this card's tags-only surface. Decision recorded, diff not widened.

---

Generated by Claude Code

---

Generated by Claude Code

[hotlong]

ACCEPT — PR #160 (2279f1d).

Seat verification:

• Mechanical checks on the branch: 334 files, all under content/blog/; git diff -U0 origin/main... filtered to non-tags lines → 0; 24 of 52 English posts end with tags: []; resulting English vocabulary: AI agents ×15 · Ontology ×8 · Palantir ×6 · Semantic layer ×4 · MCP ×4 · Governed tools ×4 · Agent permissions ×4 · Forward deployed engineer ×3 + per-article vendor names. metadata driven development absent, per the ruling.
• Union E (main + fix(blog): make wide article tables announce that they scroll on mobile #157 + content(blog): how to configure your AI coding agent to write ObjectStack apps #151's patch + this PR, 62d3632): ✓ content lint passed (335 files, 44 glossary terms checked), astro check 0/0/0, [build] 867 page(s) built, SEO smoke test passed (866 HTML pages checked), tree clean after the build (zh-Hant is generator output). CI green on 2279f1d.
• Browser spot-check at 1440×900 and 390×844: /en/blog/topics/ai-agents/ (13 cards, chips Semantic layer · AI agents · Forward deployed engineer · Palantir · Ontology · MCP · Governed tools · …), /de/blog/topics/ai-agents/ (Ontologie · MCP · KI-Agenten · Kontrollierte Werkzeuge · …), /zh-Hans/blog/topics/ai-agents/ (本体 · MCP · 前沿部署工程师 · AI 智能体 · 受治理的工具), a tagged post (JSON-LD keywords: ["Ontology","Semantic layer","AI agents"]), an empty-tag post (ai-expense-audit, keywords [], page intact, facet chips still shown) and the ai-native-app-platform cluster (6 linked cards). Canonical and the nine-entry hreflang set present on every hub. No overflow, no page errors.

Ruling on the open question: A now, C as the follow-up (#158). The closed vocabulary is what makes reconciliation mechanical, and tags are not routes on this site (measured: term hubs key on topic/audience/solutions/industries). Option B mints tier-2 vocabulary the maintainer never ruled on, so it is not the seat's to take — the seat puts the question to the maintainer on the seat post: whether tier 2 should grow a bounded set (e.g. self-hosting, approvals/automation, modernization) so the 24 category-free posts regain keywords. Until then, every new post's tags are the seven canonicals in that locale's rendering plus the vendors it compares against.

Follow-ups: #158 (retrieval gap) and #159 (ja rendering of semantic layer) stay queued; the stray finding label the dev minted on #158 is now unused and can be deleted by the maintainer. The five pending how-to branches get a tags-only reconciliation patch next. Flipping ready and squash-merging.

---

Generated by Claude Code