Forges of Sight and Speed

Three Explore agents sent forth in parallel — one to audit, one to map, one to research. The audit returned with fifteen manual events scattered across the codebase, plus Umami's automatic pageviews flowing from a self-hosted instance at analytics.fenrirledger.com. The map counted fourteen marketing routes, nine app routes, thirty chronicles — most untracked beyond mere pageviews.

The research delivered the verdict Odin needed: Umami is not Heap. No DOM autocapture. No retroactive event definition. Beyond pageviews, every click must be tagged by hand. A second agent dispatched to compare GA4, PostHog, Microsoft Clarity, FullStory, Hotjar, LogRocket, Mixpanel, Amplitude — confirmed PostHog as the modern Heap killer: 1M events/mo free, autocapture, replay, retroactive Actions, flags, all in one snippet.

The plan crystallised — a vendor-neutral abstraction layer with adapters fanning out to every upstream tracker. One analytics.track() call, three vendors notified. Freya summoned to interview Odin: env-gating decided, PII policy decided (strip nothing — card numbers are never collected), scroll depth at 25/50/75/100%, priority split between Wave 0–1 (high) and the rest (normal).

The /plan-w-team skill filed eleven child issues sequentially, capturing numbers for dependencies. Epic tracker #2289 birthed with a four-column dependency table. All eleven linked as GitHub sub-issues via the addSubIssue GraphQL mutation. The auto-close workflow already stood watch — when every child closes, the tracker closes itself.

The dashboard ran but every issue showed READY — the parser confused by a stray Priority column in the dependency table. The tracker body rewritten with the canonical four columns; the dashboard re-ran clean. Only #2278 ready, the rest blocked downstream as designed.

FiremanDecko dispatched to GKE — claude-sonnet-4-6, on-demand node, branch enhancement/issue-2278-analytics-abstraction. The job manifest written, prompt assembled, kubectl applied. Session issue-2278-step1-firemandecko-c50bd59c began its work: refactor track.ts into a shim, build router.ts, write the Umami adapter as the reference implementation. Auto-resume chain set — Loki would awaken on green verify.

Wave 0 closed cleanly. Three Wave 1 issues unblocked, zero file overlap between them: #2279 AnalyticsProvider for single-source-of-truth pageviews via usePathname + useSearchParams; #2283 TrackedLink + TrackedButton wrappers around next/link and shadcn Button with required event prop; #2285 Blog cluster-filter events on chip click with prevCluster funnel signal.

Three K8s Jobs fanned out simultaneously to the fenrir-agents namespace — well under the 32-pod quota. Each FiremanDecko spawned its own branch, its own todos, its own chain continuation back to Loki. The forge ran hot but stable.

Grep led to the root: StatusIcon.tsx:118-143 rendered a 📌 SVG when job.status === "cached", while CardPinButton already conveyed pin state through its gold fill. Two pins, one card. Issue #2291 filed, FiremanDecko dispatched.

Then Odin's second blade fell — the unpin confirmation text "✕ unpin?" spilled past the button bounds, breaking the card layout. The agent's job e1f304bd killed mid-flight with kubectl delete job. The issue body rewritten in place per the team norm (no comments with corrections — always rewrite). A fresh agent 4d08d49e spawned with the expanded scope: two bugs, one PR, status-icon hidden for cached jobs and confirm state stripped to a single red ✕ glyph.

The infrastructure confessed: infrastructure/helm/fenrir-app/values-prod.yaml:29-30 set backendConfig.cdn.enabled: true. Google Cloud CDN ran in front of www.fenrirledger.com, and after every deploy its edge cache lay cold — first-visit users in each POP paid full origin latency until traffic warmed the stone.

The honest answer delivered: Google Cloud CDN has no native prewarm API. Three patterns exist — multi-region GET storm via Cloud Run Jobs (the standard), Cloud Monitoring uptime checks (zero-code accidental warming), or switch to a CDN that supports prewarm (Fastly, Cloudflare). Odin chose the storm.

POPs explained — Points of Presence, the physical edge caches Google operates across ~187 locations. US examples named (Ashburn, Atlanta, Chicago, Dallas, Denver, LA, Miami, NYC, San Jose, Seattle). Then the brutal truth surfaced: GCP cannot target a specific POP. Anycast routing decides. Cloud Run egresses at the GCP region level — us-west1 means Oregon, not SF Bay Area. To target POPs requires third-party probe vendors (Catchpoint, ThousandEyes) at $50–$200/mo.

Eight rounds of grilling answered: region-only model for V1, post-deploy GitHub Actions trigger on main only, same-origin assets recursive (no third-party), cap at discovered_count * 1.25, registrable-domain match (*.fenrirledger.com catches www, apex, subdomains), refactor crawl-fenrir.mjs into a shared async-iterator module in the same PR, bounded worker pool with 10 concurrent default (50 max), 15s AbortController per request.

The plan rewritten with full CI/CD scope. Three child issues filed: #2296 extract scripts/lib/crawl-core.mjs as an async-iterator BFS crawler, refactor crawl-fenrir.mjs to consume it, build scripts/prime-cdn.mjs locally. #2297 Dockerize and deploy via Terraform to eight US Cloud Run Jobs (us-central1, us-east1, us-east4, us-east5, us-west1, us-west2, us-west3, us-west4). #2298 wire GitHub Actions matrix workflow post-deploy with PR-comment aggregator.

Epic tracker #2299 filed with a four-column dependency table, three children linked as sub-issues. Wave 0 dispatched — FiremanDecko began session b3cc27cb with an acceptance test baked in: round-2 cache hit rate must exceed 0.5 against the live site to prove priming worked.

Wave 1 closed in unison. Four Wave 2 issues unblocked, zero file overlap: #2280 PostHog adapter with autocapture and session replay (maskAllInputs: true, env-gated, ships dark); #2281 Heap adapter with the official Heap snippet (disableHeapPageView: true so the provider owns pageviews); #2284 Chronicle MDX wrapper auto-emitting chronicle.read on mount and chronicle.scroll at 25/50/75/100% depth — every current and future MDX post tracked without per-post code; #2286 ESLint rule warning on raw <button>/<a>/<Link>/<Button> without tracking signal.

Four K8s Jobs spawned in a single message. Pod count climbed to six — still 26 of headroom. Each adapter ships dark behind its env-var gate; no beacons fly until #2282 pushes the secrets via /manage-secrets.

The CDN epic dashboard ran. #2296 still RUNNING — FiremanDecko session b3cc27cb deep in the work of extracting the crawler core and building prime-cdn.mjs. Wave 1 (#2297 Dockerize + Terraform) and Wave 2 (#2298 GitHub Actions matrix) both blocked downstream as designed.

Nothing to dispatch. The forges run, the watch continues. Five agents across two epics still in flight on GKE Autopilot — four analytics adapters and the CDN crawler — each on its own branch, each chained back to Loki on green verify, each chained to merge on Loki PASS. When #2296 closes, the cache forge advances to Wave 1; when #2280 and #2281 close, the analytics secrets unlock the lights-on moment for PostHog and Heap.