agentleFS
relorureloru/new-relo/.claude/skills/kv/SKILL.md

kv

Skill1 starsChanged todayMITView on GitHub

Inspect and (carefully) edit the production WEATHER KV namespace — the cache behind crosbynews.com. Always uses `--remote` so it reads real production state, not local miniflare. Knows the nine content keys, `weather` + `calendar` + `water` + `fishing` + `tropics` + `pollen` + `traffic` + `burnban` (cron-owned) and `news` (routine-owned), plus `cron_status`, the Web Push state, `news_blocklist`, the MCP usage counters (`mcp_metrics` + `mcpm:*`), and the cron's bookkeeping for river flood products (`river_products`), and the orphaned `indexnow_state`. The long-term record is in D1, not KV. Use to check cache freshness or debug /news, /calendar, /water, /fishing, /tropics, /pollen, /traffic, /burn-ban, and the weather pages.

  • Deletes or force-pushes

The file

We split this skill into the rules it gives an agent

3 rules. About 0 are common, 0 are distinctive, and 3 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈0

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 3

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Read and manage the WEATHER KV namespace.

    line 10
  2. Very specific:

    ONLY out-of-band by scripts/fetch-news.mjs (a Claude routine).

    line 105
  3. Very specific:

    Read it through /api/mcp-usage rather than raw KV when you just want the numbers — it merges the unfolded shards, which a bare get does not.

    line 162

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

WEATHER KV

Read and manage the WEATHER KV namespace. Run from the repo root so wrangler resolves --binding WEATHER from wrangler.jsonc (namespace id da96de7daed84b69b32778058b374d5f is the fallback via --namespace-id).

The one rule: always --remote

wrangler kv key ... defaults to local miniflare state. Without --remote, a get reports "Value not found" even though production has the key. Every command below passes --remote — keep it.

And local state PERSISTS across runs. It lives in .wrangler/state/v3/kv/ (gitignored) and survives wrangler dev restarts, reboots, and days of elapsed time — so "a fresh local dev server" is not a cold cache. A local run cold-warms KV once and every later run reads that same snapshot back, however old it is.

This is not academic: a wrangler dev session on 2026-08-05 was served a weather value written on 2026-08-01, whose 48 hourly periods had all elapsed two days earlier — and it was briefly misread as an empty cache, because "local dev must be cold" is the intuitive assumption and it is wrong. To get a genuinely cold local cache, delete the state:

rm -rf .wrangler/state          # next `wrangler dev` starts truly empty

The nine content keys (different owners, different risk)

  • weather — NWS forecast + active alerts, shape { updated, place, periods, hourly, alerts } (hourly is the array loadWeather() checks to decide the cache is fresh). Written by the cron (*/15 * * * *) and warmed by loadWeather() on a cold cache. Inspecting is safe; a bad or deleted value self-heals within 15 min (or on the next request). Low risk.
  • calendar — Crosby ISD school calendar, shape { updated, events: [{ summary, location, start, allDay, end }] } (events is the array loadCalendar() checks for freshness). Written by the same cron (throttled to ~24h — Crosby ISD's calendar changes rarely) and warmed by loadCalendar() on a cold cache. Like weather, a bad/deleted value self-heals on the next cron or request. Low risk.
  • water — river/bayou gauges, shape { updated, gauges: [...] } (gauges is what loadWater() checks). Written by the same cron every tick; cold-warms on read. Self-heals like weather. Low risk.
  • fishing — USGS real-time water conditions for the waters people fish, shape { updated, stations: [{ id, params:{tempC,do,ph,turb,gageFt}, observed }] } (stations is what loadFishing() checks). Written by the same cron every tick; cold-warms on read. Self-heals like water. Low risk.
  • tropics — Atlantic tropical outlook from two NHC products, shape { updated, storms: [...], disturbances: [...] | null, outlookIssued }. storms is NHC CurrentStorms.json (named cyclones only); disturbances is the Tropical Weather Outlook (areas being watched, with formation chances). loadTropics() checks BOTH are arrays and cold-warms otherwise, which is what migrates pre-2026-08 entries that carry no disturbances. An empty storms is the normal quiet-basin state, NOT an error — but on its own it is not an all-clear, which is the bug this shape exists to prevent. disturbances: null means the outlook could not be read and is NOT the same as [] (nothing being watched); hand-editing one into the other makes the page assert a quiet basin nobody confirmed. Written by the same cron, throttled ~1h June-November (Atlantic hurricane season, Central time) and ~24h the rest of the year; cold-warms on read. Self-heals. Low risk.
  • pollen — the Houston Health Department's measured daily pollen & mold count, shape { updated, countDate, url, groups: {tree,weed,grass,mold}, species } (loadPollen() checks groups + countDate; countDate is the CT calendar day the count is for — weekends carry Friday's, that's normal). Written by the same cron, skipped entirely on Sat/Sun (Central time) and on weekdays only when the cached countDate isn't today's date yet — HHD publishes one count per weekday morning, so this is "have we got today's count" rather than a flat age throttle — and at most hourly, since that condition never settles on a day HHD publishes nothing (City of Houston holidays) and would otherwise retry all 96 ticks; cold-warms on read. Self-heals. Low risk. A frozen countDate with a fresh updated is the failure to look for here — it means selection is missing published counts, not that the fetch is broken (four occurrences Aug-Sep 2026, all silent; see docs/pages/pollen.md). .github/workflows/pollen-watch.yml now files a pollen-watch issue when HHD's index advertises a count newer than /api/pollen serves.
  • burnban — Harris County outdoor-burning ban status from the Texas A&M Forest Service, shape { updated, status, startDate, neighbors, statusSince, trackingSince } (loadBurnBan() checks status is "Yes" or "No"; startDate is ISO or null; neighbors is [{ county, status }] for the seven bordering counties, missing a county whose TFS row was unusable — an entry without the field is refreshed on the next cron tick). Don't hand-write statusSince / trackingSince — trackingSince is the first observation ever and never moves, statusSince moves only on a real status flip, and when they are EQUAL the page says "unchanged since we began tracking" rather than claiming a transition. Faking them makes /burn-ban assert a ban ended on a day it did not. Deleting the key is safe: the next cron reseeds both to now, which reads honestly as "no observed history yet". Written by the same cron, throttled ~4h (TFS's feed updates roughly daily, on a county judge's order rather than a schedule); cold-warms on read. Self-heals. Low risk.
  • traffic — Crosby-corridor road incidents + lane closures from Houston TranStar's public RSS feeds, shape { updated, incidents, closures } (loadTraffic() checks that incidents is PRESENT — either side is null when that feed was unreachable at the last refresh, vs [] = quiet roads, the normal state). Written by the same cron every tick; cold-warms on read. Self-heals. Low risk.
  • news — local news, shape { updated, items: [...], source }. Written ONLY out-of-band by scripts/fetch-news.mjs (a Claude routine); the Worker just renders it. Overwriting or deleting news loses the snapshot until the next routine run (up to ~a day). Treat writes/deletes here as destructive.

Also present (Web Push, don't hand-edit): push_notified (cron dedupe list of already-pushed alert IDs — first created when a severe warning actually pushes, so it's absent until then; that's the normal quiet state, not a bug) and one entry per subscriber under the push: prefix (anonymous push subscriptions). Deleting a push: entry just unsubscribes that device; deleting push_notified re-notifies every active severe warning next tick. list shows these alongside the content keys.

Also present (health, cron-owned): cron_status — {at, feeds:{<name>: {at, ok, error?, changedAt, hash}}}, written by the cron at the END of every tick and read only by /api/health. Two independent things live in each entry: at/ok/error are the last refresh attempt (a throttled tick records nothing and its previous attempt is carried forward), while changedAt/hash track when the cached CONTENT last actually changed — the cron fingerprints every feed key each tick, including routine-owned news, so a feed that keeps refreshing successfully while serving frozen data is visible. Deleting the key is harmless: every field reports null until the next tick rewrites it and the change stamps re-seed. Never hand-edit it — a fabricated ok:true would mask a genuinely failing upstream.

Also present (editorial, worker-owned): news_blocklist — {articleLink: blockedAtMs} of news articles the owner hid via the /news?admin=<ADMIN_KEY> nuke (written by /api/news/delete + /api/news/restore, read by loadNews() and the news routine). Deleting it just un-hides every article; it self-prunes past 60 days. Safe to inspect; deleting is low-risk (only un-hides).

Also present (MCP usage, cron-owned): mcp_metrics — aggregate counts for POST /mcp: {v, updated, rolledUpThrough, firstSeen, lifetime, months, days, rollup}, where each day holds {total, batches, methods, tools, outcomes, codes, clients, toolMs}. 90 days of daily detail; older days collapse into months; lifetime is never trimmed. Alongside it, transient mcpm:<bucket>:<shard> entries — one per isolate per 10-minute UTC bucket, written by the request path, folded by the cron and deleted, with a 6h expirationTtl as the backstop. list shows these alongside the content keys.

Never hand-edit mcp_metrics. rolledUpThrough is the high-water mark that makes the fold idempotent — KV deletes are eventually consistent, so a folded shard can still appear in list() for up to a minute, and lowering the mark re-folds it and doubles the numbers. Deleting the whole key is recoverable only in the sense that counting restarts from zero; there is no other copy. Deleting an mcpm:* shard loses at most ten minutes of one isolate's counts.

indexnow_state is orphaned: IndexNow was removed on 2026-11-13 and nothing reads or writes the key any more. It holds the last signals and a pending submission that will never be sent. Deleting it is safe, but it is a production delete, so only with the owner's go-ahead.

river_products (cron-owned) lists the ids of the NWS flood products (FLW/FLS, office HGX) the cron has already read for river flood warnings, last 300. Deleting it only makes the next tick re-read whatever products NWS still lists; every write is an upsert, so nothing is double-counted. See docs/ops/record.md.

Read it through /api/mcp-usage rather than raw KV when you just want the numbers — it merges the unfolded shards, which a bare get does not.

Read (safe)

List keys:

npx wrangler kv key list --binding WEATHER --remote

Get a key. Prefix CI=1 — wrangler prints a one-line "Cloudflare agent skills are available" banner to stdout that otherwise corrupts a JSON pipe; CI=1 suppresses it (so does ... | grep -v 'agent skills'). The weather value is large (~48 hourly periods), so pipe through a formatter (python3 -m json.tool, or head -c 800 if python3 is unavailable):

CI=1 npx wrangler kv key get weather  --binding WEATHER --remote | python3 -m json.tool | head -40
CI=1 npx wrangler kv key get calendar --binding WEATHER --remote | python3 -m json.tool | head -40
CI=1 npx wrangler kv key get water    --binding WEATHER --remote | python3 -m json.tool | head -40
CI=1 npx wrangler kv key get fishing  --binding WEATHER --remote | python3 -m json.tool | head -40
CI=1 npx wrangler kv key get tropics  --binding WEATHER --remote | python3 -m json.tool | head -40
CI=1 npx wrangler kv key get pollen   --binding WEATHER --remote | python3 -m json.tool | head -40
CI=1 npx wrangler kv key get traffic  --binding WEATHER --remote | python3 -m json.tool | head -40
CI=1 npx wrangler kv key get burnban  --binding WEATHER --remote | python3 -m json.tool | head -40
CI=1 npx wrangler kv key get news     --binding WEATHER --remote | python3 -m json.tool | head -40

For freshness, read the updated field rather than eyeballing the blob.

Write / delete (guarded — confirm first)

These are intentionally NOT pre-authorized, so they trigger a permission prompt. A value is read back with .get(key, "json"), so it must be valid JSON in the expected shape or the page that renders it breaks. State which key and why before running; for news, confirm the user accepts losing the current snapshot until the routine reruns.

npx wrangler kv key put    <key> '<json>' --binding WEATHER --remote
npx wrangler kv key delete <key>          --binding WEATHER --remote
  • Deleting weather, calendar, water, fishing, tropics, pollen, traffic, or burnban is recoverable (next request/cron re-warms it).
  • To repopulate news properly, re-run the pipeline instead of hand-writing it: CLOUDFLARE_API_TOKEN=... CLOUDFLARE_ACCOUNT_ID=... node scripts/fetch-news.mjs.

Default (no args)

List the keys, then report the updated / freshness of weather, calendar, water, fishing, tropics, pollen, traffic, burnban, and news.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.