Changelog

The page was carrying too much UI furniture for a loader set. This keeps the gallery easier to scan, makes the preview area feel closer to the reference, and gives users the missing “how do I actually use this?” path without sending them elsewhere.

The docs need to feel like a first-party Agent Mag product surface, not a lightly themed framework docs install. This gives builders clearer navigation, better copy/export affordances for agent context, and a larger product reference map while keeping planned APIs visibly labeled as planned/spec material.

CopyDesign is meant to be a useful SEO and ecosystem wedge: paste a public website and get a design-system starter pack for agents, builders, and product teams instead of a brittle site clone. The launch surface supports the free one-page demo while leaving room for keyed API and CLI workflows.

The original docs implementation PR and the CI follow-up PR were already merged. This update gives the follow-up branch a mergeable GitHub PR target against main without reintroducing stale local changes.

Dot Matrix should feel like an Agent Mag-owned open-source tool, not a full-width borrowed gallery. The public subdomain now behaves like a focused standalone website, while the main Agent Mag domain gets the standard tool-detail/dashboard surface at /tools/dotmatrix.

Dot Matrix gives Agent Mag a focused tool microsite that is useful on its own, supports the CLI ecosystem story, and creates a clean install path for reusable UI assets without copying unlicensed source from the UX reference repo.

The merged Dot Matrix CLI changes triggered the Publish CLI packages workflow on main. The GitHub Packages mirror step builds a temporary package containing only package.json, README.md, and dist, then npm publish runs prepublishOnly: tsc. That temp package has no tsconfig.json, so tsc prints help and exits with code 1.

Docs are becoming first-party agent context, not a skill bundle. This makes the docs page useful for humans, copy-paste workflows, CLI-driven context installs, and a public generated mirror while keeping content/docs in the private site repo as the source of truth.

The docs context PR was merged, but the follow-up main push checks exposed two CI-only issues: the CLI publish job compiles against the current tool registry union, and the mirror workflow failed immediately because the optional mirror token was empty.

Agent Mag needs a repeatable visual language for imagery across editorial, newsletter, social, skills, CLI, and sponsor surfaces. This documents the concrete prompt workflow, visual DNA, subject system, export specs, naming rules, and QA checklist so future assets stay consistent.

GitHub Actions should stay focused on CI/CD, deploy orchestration, and manual emergency runs. These recurring jobs are product operations: they write Cosmos data, call Firecrawl/Azure OpenAI, score sponsor leads, expire paid placements, and generate business rollups. Moving them into Azure makes the operating system visible inside agentmag-rg, improves secret posture with Key Vault + managed identity, gives each job its own Azure schedule/retry/timeout/log history, and creates a scalable foundation for future consumer-facing automation like image generation and screenshots.

The first merged Deploy Agent Mag Actions run failed before creating any Container Apps resources because the workflow identity does not have Microsoft.App/register/action at subscription scope. The providers are already registered in the subscription, so the deploy script should not call az provider register on every run.

The Container Apps rollout broke the main deploy workflow and partially created scheduled Azure jobs before the deployment completed. This rollback returns the repo to the exact pre-rollout code state immediately after PR 63, merge commit 7a53a24 / tag v0.1.11, so the Azure migration can be redone one workflow at a time.

Azure AI Search should ground OpenClaw, not dump raw search results into the chat before the answer. This keeps retrieval useful while preventing unrelated jobs, skills, events, or article chunks from appearing as mindless cards.

The docs should still rely on Fumadocs for the hard parts: routing, MDX loading, search, TOC, page metadata, and mobile behavior. This update only replaces the visual layer so /docs feels like Agent Mag instead of the default Fumadocs/Geist theme.

Agent Mag needs a repeatable visual language for imagery across editorial, newsletter, social, skills, CLI, and sponsor surfaces. This documents the concrete prompt workflow, visual DNA, subject system, export specs, naming rules, and QA checklist so future assets stay consistent.

The footer utility links should read as Privacy, Terms, RSS, Sitemap, then Documentation, matching the requested order while keeping the docs link available in the footer.

The changelog should be readable without zooming. Screenshots are the evidence for what shipped, so they need room to breathe and captions that explain the image instead of just naming the update number.

Agent Mag AI should stay OpenClaw under the hood, but it needs a real knowledge layer instead of relying on static/local heuristics. Azure AI Search gives us hybrid retrieval over articles, skills, jobs, tools, resources, docs, CLI facts, and OpenClaw brain context so chat, site search, related content, and later sponsor matching can share the same product spine.

The changelog should feel like a proper public product history: detailed enough for readers, partners, and future diligence, but still safe for a private repo. This keeps GitHub as internal release infrastructure while the website presents polished public release notes.

Agent Mag needs a real documentation foundation, not a one-off page. This gives the product a standard docs architecture that can scale across APIs, tools, skills, credits, dashboard workflows, and open-source projects while keeping the brand under one Agent Mag ecosystem.

The repo sidebar currently says no packages are published because that panel only reflects GitHub Packages, not the public npm package. This keeps the public install story clean as npx agentmag add while filling GitHub Packages as a quiet mirror for repository credibility.

This keeps the documentation surface cleaner and makes /docs feel like a first-class Agent Mag page instead of a detached documentation island.

The website /changelog should be the public narrative hub for the whole Agent Mag ecosystem: app, tools, skills, CLI-adjacent work, and release screenshots. The app repo's own Release notes should stay app-specific, but the public website changelog should include the external Tools and Skills release history too.

The changelog data was backfilled, but newer screenshot-heavy releases initially failed during media upload because the workflow fell back to article image storage while still targeting a missing release-media container. This update fixes that fallback path so release/changelog sync can complete with PR screenshots.

This makes Feed feel more like a daily briefing and Read feel more like a source-backed article archive without turning either page into a cluttered marketing surface. The newsletter card remains exactly where it was on /read.

The changelog should feel like a proper public product history: detailed enough for readers, partners, and future diligence, but still safe for a private repo. This keeps GitHub as internal release infrastructure while the website presents polished public release notes.

The editorial automation needed to be dependable after merge, not just functional in a happy path. This update makes failures easier to diagnose, reduces false-green workflow runs, verifies generated image assets after publish, and gives users visible confidence that Read/Feed articles are source-backed.

Release notes were too thin to act as a product history, and there was no public changelog for readers, partners, or acquirers to inspect. This makes Release notes the operational source of truth while publishing a richer, public, feed-friendly changelog from the same structured release document.

The feed hero/card content was forcing the shell wider than the viewport. When Agent Mag AI opened in the right rail, the rail was technically 380px wide, but it started too far to the right and spilled off-screen. Other pages did not show it as clearly because their main content was easier to shrink.

This completes the Agent Mag distribution path for the new public security review tool: tools repo source, website listing, detail page, npm CLI installer, and the private Agent Mag workflow that uses GPT-5.4 on updates.

The previous catalog freshness path was not production-grade: long sequential Firecrawl runs could time out before upserting, Ollama was fetched during page render instead of synced durably, and malformed Firecrawl records could leak through /api/models. This update makes the catalog refresh faster, safer, and easier to verify.

We just hit a repeatable workflow problem while updating PR 45: screenshots can be captured correctly but still render broken in GitHub if the PR body uses private-repo raw links. This skill turns that learning into a durable workflow so future updates can be created cleanly in new chats without rediscovering the same edge cases.

The editorial automation needed to be dependable after merge, not just functional in a happy path. This update makes failures easier to diagnose, reduces false-green workflow runs, verifies generated image assets after publish, and gives users visible confidence that Read/Feed articles are source-backed.

The site already had feed UI, Firecrawl configuration, and DB-aware article detail code, but there was no crawler/writer pipeline that could discover sources, generate original Agent Mag coverage, record ingestion metadata, and publish validated articles. /read and the public machine-readable surfaces were still mostly tied to MDX, so generated articles would not consistently show up where readers, search engines, or AI crawlers expect them.

The workflow also needed a clear operator notification loop. Without a status email, you have to manually inspect GitHub Actions or database documents to know whether scraping, writing, and publishing actually worked.

The user-facing issue was that clicks could appear unreliable because Next client navigation was waiting on RSC/session payloads while the page also emitted console errors. The MetaMask/SES messages are browser-extension noise, but /api/sidebar returning 404 and render-time browser globals were real app-side issues that could contribute to noisy hydration and navigation behavior.

This keeps the first client render stable with the server-rendered HTML, then reads browser-only data after hydration through React's external-store API.

The chat was hallucinating install commands, losing job context across follow-ups, and sometimes returning cards without useful conversational text. This update gives the AI stronger product/data rails while keeping GPT-4o/OpenClaw as the conversational layer.

The site already had feed UI, Firecrawl configuration, and DB-aware article detail code, but there was no crawler/writer pipeline that could discover sources, generate original Agent Mag coverage, record ingestion metadata, and publish validated articles. /read and the public machine-readable surfaces were still mostly tied to MDX, so generated articles would not consistently show up where readers, search engines, or AI crawlers expect them.

The workflow also needed a clear operator notification loop. Without a status email, you have to manually inspect GitHub Actions or database documents to know whether scraping, writing, and publishing actually worked.

PostHog was partially installed, but the app did not have a complete Product OS layer. The missing pieces were reliable early client initialization, stable event naming, server-side route capture, CLI attribution, feature flag/dashboard setup, and proper enrichment support for country/IP analysis. This update turns PostHog into the analytics source for Agent Mag growth, skills/CLI adoption, newsletter conversion, chat usage, revenue intent, web vitals, errors, and replay QA.

The PostHog integration PR was merged after an SEO auto-fix made src/app/skills/[slug]/page.tsx a Client Component while still exporting generateMetadata. Next.js disallows that and CI failed during pnpm build.

Sponsors do not need an admin console. They need a clean workbench for the surfaces they buy and manage: editorial, newsletter, public listings, and boosts. This keeps the sponsor product simple while still using the same Agent Mag app, auth, Cosmos data layer, and Stripe checkout flow.

Agent Mag AI was treating first-party CLI commands like agentmag add as unknown and over-triggering model cards on install-command follow-ups. This gives it concise public Agent Mag product knowledge without dumping raw source code or private implementation internals into OpenClaw.

The /models page was built with shadcn components (rounded buttons, Badge pills, Input with ring focus) while the rest of the site uses a raw editorial style (sharp borders, mono text, inverted black/white chips). This made /models visually inconsistent with /skills, /jobs, and /read.

Agent Mag AI needed to become an opinionated product guide that can actually help users pick models from the Agent Mag catalog, not just answer generically or show hardcoded cards. The follow-up keeps that intelligence but makes the first visible result much faster by relying on deterministic recommendation cards first and GPT-4o for the chat explanation. This also prevents malformed scraped model docs from appearing in recommendations and makes local development less fragile when OpenClaw gateway settings change.

PostHog was partially installed, but the app did not have a complete Product OS layer. The missing pieces were reliable early client initialization, stable event naming, server-side route capture, CLI attribution, feature flag/dashboard setup, and proper enrichment support for country/IP analysis. This update turns PostHog into the analytics source for Agent Mag growth, skills/CLI adoption, newsletter conversion, chat usage, revenue intent, web vitals, errors, and replay QA.

Individual model pages could crash because the page was passing title and items into RelatedContent, but that component expects a links object. That broke direct detail URLs like /models/ollama--kimi-k2.6 for users browsing from the models directory.

The row click target also became fragmented after the copy command UI was added, so the list cards needed a cleaner link overlay that still lets the copy button work.

The app currently presents authentication like a public product, while the actual intention is a private alpha restricted to approved emails. This update makes the alpha posture explicit in the product copy and keeps local development access controlled without changing the broader production direction.

This gives Agent Mag first-party local model install commands from the models directory and makes the npm package useful for local Ollama model adoption.

AgentMag skills need to be canonical, SEO/GEO-ready, and safe to grow. This keeps published skills anchored to GitHub while creating a reviewed path for scaling external skills.sh imports without exposing low-quality or unsafe pages.

Individual model pages could crash because the page was passing title and items into RelatedContent, but that component expects a links object. That broke direct detail URLs like /models/ollama--kimi-k2.6 for users browsing from the models directory.

The skills registry is one of Agent Mag's strongest ecosystem assets, so registry releases belong in the public website changelog.

Including skills releases helps users and buyers see the growth of the installable skill ecosystem alongside the app itself.

Agent Mag Tools turns the publication into a practical utility layer for AI agent builders, not just a place to read about releases.

Including tools releases in the public changelog makes the website the single narrative surface for product, CLI, skills, and open-source utility work.