# Free-Tier Survival Plan — leaksnipe.win estate
Verified against the live account + Cloudflare docs, 2026-10-24 (Truffle).
Replaces the Grok roadmap draft (its zone-multiplication premise is false — see §1).

## 1. The caps that actually bind (all ACCOUNT-level, Cloudflare docs "Account plan limits")
- Workers: 100,000 requests/day · 10 ms CPU/invocation · 50 subrequests/request · 100 workers
- Cron triggers: **5 per account total**
- KV: 100k reads/day · 1,000 writes/day · 1 GB
- D1: 5M reads/day · 100k writes/day · 5 GB
- R2: 10 GB free, egress free on every plan
- Durable Objects (free): 100k requests/day · 13,000 GB-s/day · SQLite classes only
- Zones do NOT multiply any of this. Subdomain zones are Enterprise-only (API refuses
  subdomain zone creation: error 1116, verified live 2026-10-24).
- Zone plan (Pro Website on leaksnipe.win) is separate from the Workers Paid
  subscription. The Oct 26 deadline is the Workers subscription ($84.77 invoice).

## 2. Estate state (measured tonight)
- 64 workers · ~150,700 requests/day (7-day GraphQL analytics) before tonight's fix
- kalshi-ladder-worker alone: ~105,800/day (70% of all traffic)
- All 7 DO namespaces: `use_sqlite: true` → free-safe, no migrations owed
- No Images / Browser Rendering / Vectorize / Queue bindings anywhere
- Containers: containers-template + leaksnipe-beta would stop on free (paid-only);
  both non-production (~4 req/week)
- Cron triggers: 9 → 8 after cutting dead vvv-price-alert (VVV route ended Oct 19/20)

## 3. Done tonight
- **Edge-cache rule for ladder.leaksnipe.win GETs, 60s TTL** (zone cache ruleset, last
  position so it wins for this host). Why it's safe: the worker serves a *cached
  window state* refreshed on its 3-min cron and self-reports `fetched_at` — the body
  cannot change between refreshes, so 60s of edge TTL loses no freshness that exists.
  Effect (analytics, 20-min window after deploy): ladder invocations 18 vs ~1,460
  baseline expectation. Projected estate total: **~47k/day — under the cap.**
  Rollback: delete the rule described "free-tier survival 2026-10-24" from the
  http_request_cache_settings entrypoint.
- vvv-price-alert cron deleted (worker untouched).
- UVIX router de-gated + dex-direct.leaksnipe.win created and registered
  (uvix:dex, uvix:dex-direct, manifest v1.3.2) — all agent surfaces answer unauthenticated.

## 4. Staged, not fired (only bind if the plan actually drops)
- **Cron cut 8 → 5**: script `apply-cron-cut-when-downgraded.py` in this directory.
  Keep: kalshi-ladder-worker, scalper-market-data, wecryp-price-proxy,
  kodak-market-archiver, kodak-collector.
  Cut: muse-mcp (warmer only — bridge still answers live), mem-gateway (daily
  maintenance), leaksnipe-files (hourly janitor). NOT fired tonight: on the paid plan
  these crons do real work; cutting them now buys nothing.
- **DO-alarm substitution**: the three cut jobs can move into Durable Object alarms
  (alarms are not cron triggers) on the existing SQLite DOs. Code change per worker;
  alarms burn the DO free budget (100k req/day) — precision tool, spec only.
- **dex-rugcheck subrequest fix (Grok-side code)**: ~136 RPC calls/check vs the
  50-subrequest free cap. JSON-RPC batching collapses it to a handful of fetches.
  Without this, the checker breaks on free no matter what else is done.
- **Static-assets offload**: Workers Static Assets requests are unmetered; only
  code-routed paths count. Candidate: maurofanellijr-home (~1.3k/day). Requires a
  homepage-worker redeploy — do not touch while v8 awaits Mauro's verdict.

## 5. The structural lever (his build, with Grok)
High-frequency internal polling is the load. Tailscale Serve (svc:grokdex pattern)
moves agent-to-agent traffic off Cloudflare entirely — unmetered, and immune to the
billing state of the account. Every internal read that moves to the tailnet is a
request the free cap never sees.

## 6. Bottom line
With the ladder cache rule in place, the estate fits under the free cap as measured
(~47k/day projected vs 100k). The remaining free-tier breaks, in order: dex-rugcheck
subrequests (code, Grok), cron 8>5 (staged script), Containers templates (accepted
loss), 10 ms CPU on the heavy fan-out workers (unmeasured risk). Paying the $84.77
by Oct 26 still erases the entire question.