rsc-universal 0.1.1

package/skills/redis/evals/cases.yaml

@@ -0,0 +1,43 @@
+skill: redis
+
+should_trigger:
+  - prompt: "Our product cache key expires and the database gets hammered by a flood of concurrent requests until it's rebuilt."
+    why: Cache stampede / thundering herd — the core caching correctness contract this skill owns.
+  - prompt: "Two cron workers across our instances sometimes run the same nightly job. I need only one to actually do it."
+    why: Distributed lock / mutual exclusion — SET NX PX + Lua release + the de-dup decision.
+  - prompt: "My rate limiter occasionally lets bursts through and some users get throttled forever. I do INCR then EXPIRE in the app."
+    why: Non-obvious — the lost-EXPIRE race that leaves a TTL-less immortal counter, the canonical rate-limit foot-gun.
+  - prompt: "BullMQ jobs keep getting marked stalled and reprocessed. How do the job locks and lockRenewTime work?"
+    why: Redis queue reliability and stalled-job lock renewal — at-least-once + idempotency territory.
+  - prompt: "A reviewer flagged `KEYS user:*` in our request hot path. Is that actually a problem?"
+    why: Non-obvious ops foot-gun — KEYS is O(N) and blocks the single thread; route to SCAN. Review trigger.
+  - prompt: "Should we use Redlock or is a single SET NX PX enough for our locks? Do we need fencing tokens?"
+    why: Lock correctness decision — single-instance vs Redlock vs fencing for hard mutual exclusion.
+  - prompt: "Our Redis cache started returning OOM errors and refusing writes when memory filled up."
+    why: Eviction policy — a cache left on the default noeviction with no TTLs; route to allkeys-lru + TTL.
+
+should_not_trigger:
+  - prompt: "Build a durable background job queue in Postgres using SELECT ... FOR UPDATE SKIP LOCKED."
+    route_to: postgresdb
+    why: The durable relational store IS the queue — that's the postgresdb boundary, not a Redis queue.
+  - prompt: "Add semantic similarity search over our document embeddings with ANN indexing."
+    route_to: vector-db
+    why: Vector similarity as the primary job; Redis vector sets are only noted here, not owned.
+  - prompt: "Provision and size an ElastiCache / fly.io Redis cluster, set the node type and replicas."
+    route_to: fly-io
+    why: Provisioning and console/cluster sizing, not the client-side correctness contract.
+  - prompt: "Tune Cache-Control headers and ISR revalidation for our Next.js pages."
+    route_to: nextjs
+    why: HTTP/CDN/edge caching, not the Redis data-layer contract.
+  - prompt: "We need a high-throughput analytical store for clickstream events with columnar aggregation."
+    route_to: clickhouse-analytics
+    why: Analytical/event store, not a hot-path cache, lock, rate limiter, or queue.
+
+capability:
+  - scenario: "We have a hot product-detail cache that stampedes the DB on expiry, a cron that must run exactly once across 3 instances, and a per-API-key rate limiter built on INCR then EXPIRE. Fix all three correctly."
+    must_include:
+      - "Stampede: TTL with jitter plus a recompute-lock (SET NX) or probabilistic early expiry (XFetch); lock TTL must exceed worst-case rebuild."
+      - "Lock: SET <random-token> NX PX <ttl> released only via a Lua compare-and-delete (GET == my token then DEL); mentions TTL-vs-work-duration and fencing tokens for hard mutual exclusion."
+      - "Rate limiter: collapse INCR + EXPIRE into one atomic EVAL (or set EXPIRE only when n == 1); names the lost-EXPIRE race that creates an immortal counter."
+      - "Notes at-least-once / idempotency wherever double-execution is possible (the cron, the rate-limit edge)."
+      - "Examples stay client-agnostic (Lua / redis-cli), not tied to a single SDK."

package/skills/redis/references/caching.md

@@ -0,0 +1,116 @@
+# Caching — depth
+
+Read this when the cache is more than `SET k v EX`: read-through/write-through/write-behind,
+multi-level caches, probabilistic early expiry, negative caching, and hot/big-key mitigation.
+All patterns are client-agnostic (`redis-cli` / Lua / pseudo-client).
+
+## Cache topologies
+
+| Pattern | Who reads/writes the DB | Failure mode | Use when |
+| --- | --- | --- | --- |
+| **Cache-aside** (lazy) | the app, on a miss | stale until TTL; first read is slow | the default; read-heavy, tolerant of brief staleness |
+| **Read-through** | the cache layer/lib | same as aside, hidden behind the lib | you want the cache to own the fetch |
+| **Write-through** | app writes cache + DB synchronously | write latency = DB + cache; double write to keep consistent | reads must never see stale after a write |
+| **Write-behind** (write-back) | app writes cache, a worker flushes to DB later | data loss window if Redis dies before flush | write-heavy, can tolerate a flush-lag loss window |
+
+Default to **cache-aside with a TTL**. Reach for write-through only when a read immediately after a
+write must reflect it; reach for write-behind only when you can afford to lose the un-flushed window
+(and you have AOF or a durable journal to bound it).
+
+## TTL + jitter
+
+A fixed TTL applied to keys warmed together synchronizes their expiry — they all die in the same
+second and stampede the origin simultaneously. Always jitter.
+
+```text
+ttl = base + random(0, base * 0.2)   # e.g. 300s base, up to +60s jitter
+SET product:42 <json> EX ttl
+```
+
+*Why:* spreading expirations turns one giant stampede into a trickle of independent misses.
+
+## Negative caching
+
+Caching only hits means every request for a non-existent id falls through to the origin — an easy
+DoS via random ids. Cache the miss too, with a *short* TTL so a row that appears soon isn't hidden long.
+
+```text
+val = GET product:42
+if val == "__MISS__": return null          # cached absence, no DB call
+if val == nil:
+    row = db.fetch(42)
+    if row == null: SET product:42 "__MISS__" EX 30   # short negative TTL
+    else:           SET product:42 row EX 300
+```
+
+Keep the negative TTL much shorter than the positive one (seconds, not minutes) so freshly created
+rows surface quickly.
+
+## Stampede prevention
+
+When a hot key expires, N concurrent readers all miss and hit the origin at once. Three approaches,
+in increasing sophistication:
+
+### 1. Recompute lock (request coalescing)
+
+One reader claims the rebuild; the rest serve stale or briefly wait.
+
+```text
+val = GET product:42
+if val != nil: return val
+if SET product:42:lock 1 NX PX 5000 == OK:   # 5000ms MUST exceed worst-case rebuild
+    val = db.fetch(42); SET product:42 val EX (300 + jitter); DEL product:42:lock
+    return val
+else:
+    return GET product:42:stale  or  sleep(50ms) and retry GET
+```
+
+*Why the lock TTL matters:* if it is shorter than the rebuild, a second reader claims the lock while
+the first is still rebuilding and you stampede anyway. Size it above the slowest origin call.
+
+### 2. Probabilistic early expiration (XFetch)
+
+Refresh *before* the key actually expires, with a probability that rises as expiry approaches, so one
+reader renews early while everyone else still hits a warm key. The classic XFetch test:
+
+```text
+# delta  = how long the last recompute took (seconds)
+# beta   = tuning knob, ~1.0 (higher = refresh earlier)
+# expiry = absolute unix time the key expires
+recompute if  now - delta * beta * ln(random(0,1))  >= expiry
+```
+
+Store `delta` and `expiry` alongside the value (a small hash works). Because the random term only
+crosses the threshold for an unlucky few before TTL, exactly one or two readers refresh early and the
+stampede never forms. No lock, no stale window.
+
+### 3. Serve-stale-while-revalidate
+
+Keep two TTLs: a soft "fresh until" and a hard "evict at". Past the soft TTL, serve the stale value
+*and* kick off one async refresh. Readers never block; the origin sees one refresh per key per window.
+
+## Multi-level caching
+
+In-process LRU (per app instance) in front of Redis cuts Redis round-trips for the hottest keys. The
+cost is coherence: a write must invalidate *both* levels, and different instances can hold different
+local copies for up to the local TTL. Keep the local TTL very short (1-5s) so divergence is bounded,
+and treat Redis as the source of truth the local cache merely fronts.
+
+## Invalidation
+
+- **On write, delete the key** (cache-aside) — the next read rebuilds. Simple and correct as long as
+  every write path deletes.
+- **Never** rely on manual invalidation *instead of* a TTL: you will miss a code path and serve stale
+  forever. TTL is the backstop; explicit invalidation is the latency optimization on top.
+- **Tag/group invalidation** (drop everything for a tenant): keep a set of member keys per tag and
+  `DEL` them together, or namespace by a version counter you bump (`v:tenant:7 -> 12`, keys read
+  `tenant:7:v12:*`) so a bump orphans the whole old generation to be evicted by LRU.
+
+## Hot keys and big keys
+
+- **Hot key** (one key, enormous read rate) can saturate a single shard/CPU. Mitigate with a
+  per-instance local cache in front, or shard the value across N suffixed keys read at random.
+- **Big key** (one huge value or a multi-million-element collection) blocks the single thread on
+  every access and on eviction. Find them with `redis-cli --bigkeys` and `MEMORY USAGE <key>`; split
+  big hashes/lists into smaller keyed chunks; never `DEL` a giant key on the hot path — `UNLINK` it
+  (frees memory in a background thread).

package/skills/redis/references/locks-and-rate-limiting.md

@@ -0,0 +1,140 @@
+# Locks and rate limiting — depth
+
+The full Redlock algorithm and its critique, lease renewal, fencing tokens, and all four
+rate-limit algorithms with complete Lua. Client-agnostic.
+
+## Single-instance lock (the baseline)
+
+```text
+SET resource:lock <random-token> NX PX <ttl-ms>   # acquire: only if absent, auto-expires
+# ... protected work ...
+# release via the compare-and-delete Lua below — NEVER a bare DEL
+```
+
+The token must be unguessable and unique per acquisition (a UUID or `<host>:<pid>:<rand>`), because
+release is "delete only if the value is still mine":
+
+```lua
+-- KEYS[1] = lock key, ARGV[1] = my token. Returns 1 if I held and released, else 0.
+if redis.call('GET', KEYS[1]) == ARGV[1] then
+  return redis.call('DEL', KEYS[1])
+else
+  return 0
+end
+```
+
+## Lease renewal (watchdog)
+
+If protected work can outlast the TTL, the lock expires mid-task and a second worker starts. Renew
+the TTL periodically — but only if you still own it (same token guard), or you will extend a lock
+someone else now holds:
+
+```lua
+-- KEYS[1] = lock key, ARGV[1] = my token, ARGV[2] = new ttl ms.
+if redis.call('GET', KEYS[1]) == ARGV[1] then
+  return redis.call('PEXPIRE', KEYS[1], ARGV[2])
+else
+  return 0
+end
+```
+
+Run this from a watchdog at ~`ttl/3`. If a renewal ever returns 0, you lost the lock — stop the work,
+do not assume you still hold it.
+
+## Redlock (multi-instance) and why it is contested
+
+Redlock acquires the same lock on a majority of N independent masters (e.g. 3 of 5), each with
+`SET NX PX`, and considers the lock held only if the majority succeeded *within* a small fraction of
+the TTL (so the remaining validity time is still positive). Release deletes on all N.
+
+The critique (Kleppmann; acknowledged by antirez): Redlock's safety rests on **bounded clock drift**
+and **no long process pauses**. A GC pause, a hypervisor stop-the-world, or an NTP jump can let the
+TTL expire on the holder's wall clock while the holder still believes it is inside the lease; a second
+client then acquires, and both act. Majority-acquisition does not fix this — it is a *timing* problem,
+not a *quorum* problem.
+
+### Fencing tokens — the only real fix
+
+Issue a **monotonically increasing token** with every successful acquisition (e.g. `INCR
+resource:fence`). The protected resource (DB, file store, queue) records the highest token it has
+accepted and **rejects any write carrying a lower token**. Now even a paused-then-resumed holder is
+harmless: its token is stale, so its write is refused. No pure-Redis lock provides fencing; you must
+add the counter and enforce it at the resource.
+
+### Decision
+
+- **Idempotent / best-effort work** (cron de-dup, cache warm): single-instance `SET NX PX` is fine —
+  an occasional double-run is harmless.
+- **Hard mutual exclusion** (money, non-idempotent side effects): a Redis lock alone is *not* a safety
+  guarantee. Fence the resource, or use a real lease/consensus system (ZooKeeper, etcd, a DB advisory
+  lock). Redlock buys availability across instances, not correctness under pauses.
+
+## Rate limiting — four algorithms
+
+Every one must be a **single atomic `EVAL`**. The classic bug is splitting count and TTL: if the
+process dies between `INCR` and `EXPIRE`, the key never expires and the user is throttled forever.
+
+### Fixed window
+
+```lua
+-- KEYS[1]=counter, ARGV[1]=limit, ARGV[2]=window-seconds. 1=allow, 0=deny.
+local n = redis.call('INCR', KEYS[1])
+if n == 1 then redis.call('EXPIRE', KEYS[1], ARGV[2]) end
+if n > tonumber(ARGV[1]) then return 0 end
+return 1
+```
+
+Cheapest (one integer per key). Flaw: allows up to ~2× the limit across a window boundary (a burst at
+the end of one window plus a burst at the start of the next).
+
+### Sliding-window log
+
+Exact: store one timestamp per request in a sorted set, prune the old, count what remains.
+
+```lua
+-- KEYS[1]=zset, ARGV[1]=now-ms, ARGV[2]=window-ms, ARGV[3]=limit, ARGV[4]=member.
+redis.call('ZREMRANGEBYSCORE', KEYS[1], 0, ARGV[1] - ARGV[2])   -- drop entries older than window
+local count = redis.call('ZCARD', KEYS[1])
+if count >= tonumber(ARGV[3]) then return 0 end
+redis.call('ZADD', KEYS[1], ARGV[1], ARGV[4])                   -- record this request
+redis.call('PEXPIRE', KEYS[1], ARGV[2])                         -- self-clean idle keys
+return 1
+```
+
+Exact, but O(requests) memory per key — expensive at high limits.
+
+### Sliding-window counter
+
+Approximates the log with two fixed-window counters (current + previous), weighting the previous one
+by how far into the current window you are. Near-exact at a fraction of the memory of the log. Keep
+both counters and compute `prev * (1 - elapsed_fraction) + curr` in Lua; deny if that exceeds the limit.
+
+### Token bucket
+
+Allows controlled bursts: a bucket of capacity C refills at R tokens/sec; each request costs one token.
+
+```lua
+-- KEYS[1]=hash{tokens,ts}, ARGV[1]=capacity, ARGV[2]=refill/sec, ARGV[3]=now-sec, ARGV[4]=cost.
+local s = redis.call('HMGET', KEYS[1], 'tokens', 'ts')
+local tokens = tonumber(s[1]) or tonumber(ARGV[1])
+local ts     = tonumber(s[2]) or tonumber(ARGV[3])
+tokens = math.min(tonumber(ARGV[1]), tokens + (ARGV[3] - ts) * ARGV[2])   -- refill by elapsed time
+if tokens < tonumber(ARGV[4]) then
+  redis.call('HSET', KEYS[1], 'tokens', tokens, 'ts', ARGV[3]); return 0
+end
+tokens = tokens - tonumber(ARGV[4])
+redis.call('HSET', KEYS[1], 'tokens', tokens, 'ts', ARGV[3])
+redis.call('PEXPIRE', KEYS[1], math.ceil(tonumber(ARGV[1]) / tonumber(ARGV[2]) * 1000))
+return 1
+```
+
+Smallest memory (one small hash per key), smooth average rate, configurable burst. The usual default.
+
+### Choosing
+
+| Need | Pick |
+| --- | --- |
+| Cheapest, OK with 2× edge bursts | fixed window |
+| Exact, low traffic | sliding-window log |
+| Near-exact, high traffic | sliding-window counter |
+| Smooth rate + intentional bursts | token bucket |

package/skills/redis/references/queues.md

@@ -0,0 +1,102 @@
+# Queues — depth
+
+Full reliable-list and Streams consumer-group implementations, stalled/pending recovery, dead-letter
+queues, idempotency, and the library comparison. Client-agnostic.
+
+## Reliable list pattern
+
+Never `RPOP` straight into a worker: a crash between the pop and the work loses the job. Move it
+atomically to a per-worker processing list, do the work idempotently, then ack by removing it. A
+separate reaper requeues anything orphaned in a processing list whose worker has died.
+
+```bash
+# 1. Atomic claim: pop from the queue, push onto THIS worker's processing list, in one command.
+job=$(redis-cli LMOVE jobs jobs:proc:w1 LEFT RIGHT)     # BLMOVE for a blocking variant
+
+# 2. Do the work — idempotently, because delivery is at-least-once (see below).
+
+# 3. Ack: remove the job from the processing list.
+redis-cli LREM jobs:proc:w1 1 "$job"
+```
+
+Reaper (runs periodically): for each `jobs:proc:*`, if its worker is gone (heartbeat key expired),
+`LMOVE` every element back to `jobs`. The trade-off: a job that was *almost* done but un-acked gets
+re-run — hence handlers must be idempotent.
+
+`LMOVE`/`BLMOVE` (Redis 6.2+) replace the older `RPOPLPUSH`/`BRPOPLPUSH`, which still work.
+
+## Streams + consumer groups (preferred for reliable jobs)
+
+Streams give per-message acks and a Pending Entries List (PEL) for free — no hand-rolled processing
+list. This is the right native primitive for reliable jobs.
+
+```bash
+# Create the group once (MKSTREAM creates the stream if absent). '$' = only new messages.
+redis-cli XGROUP CREATE jobs g1 '$' MKSTREAM
+
+# Producer: append, capping growth so the stream can't grow unbounded.
+redis-cli XADD jobs MAXLEN '~' 100000 '*' type email to user@x.com
+
+# Consumer: read new messages ('>'), each delivered message enters this consumer's PEL.
+redis-cli XREADGROUP GROUP g1 worker1 COUNT 10 BLOCK 5000 STREAMS jobs '>'
+
+# Ack a finished message — removes it from the PEL.
+redis-cli XACK jobs g1 <id>
+```
+
+### Stalled / pending recovery
+
+A worker that dies leaves its in-flight messages in the PEL, un-acked. Reclaim them with
+`XAUTOCLAIM` (Redis 6.2+), which transfers messages idle longer than a threshold to a live consumer:
+
+```bash
+# Claim messages in group g1 idle > 60000ms, reassigning them to worker2, starting at id 0.
+redis-cli XAUTOCLAIM jobs g1 worker2 60000 0
+# Inspect the backlog: XPENDING jobs g1   (summary)  /  XPENDING jobs g1 - + 10  (detail)
+```
+
+Track a delivery count per message (`XPENDING` reports it); after K deliveries, stop retrying and
+route to a dead-letter stream instead of looping forever on a poison message.
+
+### Dead-letter queue
+
+```bash
+# Poison message exceeded max deliveries: copy to a DLQ stream, then ack it off the main group.
+redis-cli XADD jobs:dead '*' orig_id <id> reason max-deliveries-exceeded payload "$payload"
+redis-cli XACK jobs g1 <id>
+```
+
+Trim both the main and DLQ streams (`MAXLEN ~`) so neither grows without bound. Redis 8.2 adds
+`XDELEX`/`XACKDEL` for ack-and-delete in one step.
+
+## Idempotency and at-least-once
+
+Every Redis queue (list or stream) is **at-least-once**: a crash after work but before ack re-runs
+the job. Make handlers idempotent:
+
+- Carry an **idempotency key** (a job id or a hash of the payload). Before the side effect, `SET
+  done:<key> 1 NX EX <ttl>`; if it returns nil, the job already ran — skip the side effect, just ack.
+- Or make the side effect itself idempotent (upserts keyed by id, conditional writes with a fencing
+  token, external APIs that accept an idempotency key).
+
+Exactly-once delivery does not exist over a network; exactly-once *effect* is achievable only by
+making the effect idempotent.
+
+## When to use a library instead of hand-rolling
+
+| Library | Language | You get |
+| --- | --- | --- |
+| **BullMQ** | TypeScript/Node | retries, delays, repeatable jobs, priorities, a dashboard, **stalled-job lock renewal** via `lockDuration`/`lockRenewTime` |
+| **Sidekiq** | Ruby | multi-threaded workers, retries, scheduling, web UI |
+| **RQ** | Python | simple, list-based, easy to reason about |
+| **Celery** (Redis broker) | Python | mature, broad feature set, can use Redis or others |
+
+Reach for a library the moment you need retries with backoff, scheduled/delayed jobs, priorities, or a
+dashboard — do not reinvent them. BullMQ's stalled-job handling is the canonical example: a worker
+holds a lock on its active job and renews it every `lockRenewTime`; if a worker dies, the lock lapses
+after `lockDuration` and the job is moved back to be retried up to its attempt limit. Setting
+`lockDuration` below your real job duration is the usual cause of "my jobs keep getting marked
+stalled and reprocessed" — raise it above the worst-case job time (and renew), don't lower your
+expectations.
+
+Hand-roll lists/streams only for the simplest cases or when no good library exists for your runtime.

package/skills/redis/scripts/verify.sh

@@ -0,0 +1,164 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# verify.sh — redis skill gate. Run from your PROJECT root.
+#
+# What it does (read-only, idempotent, NEVER connects to a Redis server, never writes):
+#   Scans discovered Redis-touching source/config for high-confidence foot-guns. It exits
+#   non-zero ONLY on a near-certain production bug (`KEYS` / `KEYS *` in non-test source).
+#   Everything else is advisory [warn], and an absence of Redis usage is [skip] (exit 0).
+#
+# Heuristics are deliberately conservative: client SDKs and config look wildly different across
+# node-redis, ioredis, redis-py, go-redis and Lettuce, so all the soft checks are warnings you
+# eyeball, not build-breakers. The hard fail is reserved for `KEYS` because it is essentially
+# always wrong on a hot path and trivially confirmable.
+#
+# Portability: stock macOS bash 3.2 (no mapfile, no associative arrays). Arrays are initialised so
+# they expand safely under `set -u`.
+
+YELLOW=$'\033[33m'; GREEN=$'\033[32m'; RED=$'\033[31m'; NC=$'\033[0m'
+EXIT=0
+skip() { printf '%s[skip]%s %s\n' "$YELLOW" "$NC" "$*"; }
+note() { printf '%s[warn]%s %s\n' "$YELLOW" "$NC" "$*"; }
+ok()   { printf '%s[ok]%s %s\n'   "$GREEN"  "$NC" "$*"; }
+err()  { printf '%s[fail]%s %s\n' "$RED"    "$NC" "$*"; EXIT=1; }
+
+ROOT="$(pwd)"
+
+# is_test_path <path> — true for files that look like tests/specs/fixtures, where foot-guns are fine.
+is_test_path() {
+  case "$1" in
+    *test*|*Test*|*spec*|*Spec*|*__tests__*|*fixtures*|*/mocks/*) return 0 ;;
+    *) return 1 ;;
+  esac
+}
+
+# Discover candidate source files (common Redis-client languages) and redis.conf, excluding vendor dirs.
+SRC_FILES=()
+while IFS= read -r -d '' f; do
+  SRC_FILES+=("$f")
+done < <(
+  find "$ROOT" \
+    \( -path '*/node_modules/*' -o -path '*/.git/*' -o -path '*/vendor/*' -o -path '*/.venv/*' \
+       -o -path '*/dist/*' -o -path '*/build/*' -o -path '*/.next/*' \) -prune -o \
+    -type f \( -name '*.js' -o -name '*.ts' -o -name '*.jsx' -o -name '*.tsx' -o -name '*.mjs' \
+               -o -name '*.py' -o -name '*.go' -o -name '*.rb' -o -name '*.java' -o -name '*.kt' \
+               -o -name '*.lua' -o -name '*.php' -o -name '*.rs' \) -print0 2>/dev/null
+)
+
+CONF_FILES=()
+while IFS= read -r -d '' f; do
+  CONF_FILES+=("$f")
+done < <(
+  find "$ROOT" \
+    \( -path '*/node_modules/*' -o -path '*/.git/*' -o -path '*/vendor/*' \) -prune -o \
+    -type f \( -name 'redis.conf' -o -name 'valkey.conf' -o -name '*redis*.conf' \) -print0 2>/dev/null
+)
+
+# Decide whether this project touches Redis at all. If not, skip cleanly.
+# Use -l (newline-separated) not -Z: BSD/macOS grep ignores -Z when listing filenames, so a
+# NUL-delimited read loop would mis-parse. Filenames with newlines are not supported here (rare),
+# which is an acceptable trade for portability across BSD and GNU grep.
+REDIS_HINT=0
+HINT_FILES=()
+if [ "${#SRC_FILES[@]}" -gt 0 ]; then
+  while IFS= read -r f; do
+    [ -n "$f" ] && HINT_FILES+=("$f")
+  done < <(
+    grep -ril -E 'redis|valkey|ioredis|bullmq|sidekiq|XREADGROUP|maxmemory' "${SRC_FILES[@]}" 2>/dev/null || true
+  )
+fi
+if [ "${#HINT_FILES[@]}" -gt 0 ] || [ "${#CONF_FILES[@]}" -gt 0 ]; then
+  REDIS_HINT=1
+fi
+
+if [ "$REDIS_HINT" -eq 0 ]; then
+  skip "no Redis/Valkey usage detected in this project — nothing to check"
+  ok "verify.sh passed"
+  exit 0
+fi
+
+# Only scan the files that actually mentioned Redis (plus conf files), to cut false positives.
+SCAN_FILES=()
+for f in ${HINT_FILES[@]+"${HINT_FILES[@]}"}; do SCAN_FILES+=("$f"); done
+for f in ${CONF_FILES[@]+"${CONF_FILES[@]}"}; do SCAN_FILES+=("$f"); done
+
+# ---- 1. KEYS / KEYS * in non-test source -> HARD FAIL ----
+# Match a KEYS command (quoted, or as a client method call), not the identifier "keys" in general.
+KEYS_HIT=0
+for f in "${SCAN_FILES[@]}"; do
+  is_test_path "$f" && continue
+  if grep -Eniq "([\"'\` ]KEYS[[:space:]]+[^)]|\.keys\(|->keys\(|[\"'\`]KEYS\*)" "$f" 2>/dev/null; then
+    # Exclude obvious false friends: object key iteration like Object.keys(, dict.keys(), .Keys (Go map helpers vary; conservative).
+    if grep -Eniq "([\"'\`][[:space:]]*KEYS[[:space:]]|[\"'\`]KEYS\*|redis[^.]*\.keys\(|client\.keys\(|->keys\()" "$f" 2>/dev/null; then
+      err "$f: uses Redis KEYS (O(N), blocks the single thread). Use SCAN MATCH ... COUNT ... instead."
+      KEYS_HIT=1
+    fi
+  fi
+done
+[ "$KEYS_HIT" -eq 0 ] && ok "no Redis KEYS in non-test source"
+
+# ---- 2. FLUSHALL / FLUSHDB in non-test source -> warn ----
+for f in "${SCAN_FILES[@]}"; do
+  is_test_path "$f" && continue
+  if grep -Eniq "FLUSHALL|FLUSHDB|\.flushall\(|\.flushdb\(" "$f" 2>/dev/null; then
+    note "$f: FLUSHALL/FLUSHDB present — wipes the instance; never run on a shared/prod Redis."
+  fi
+done
+
+# ---- 3. Lock release via DEL near a lock key without a token compare -> warn ----
+for f in "${SCAN_FILES[@]}"; do
+  is_test_path "$f" && continue
+  if grep -Eniq "lock" "$f" 2>/dev/null \
+     && grep -Eniq "(\.del\(|\.delete\(|->del\(|[\"'\`]DEL[\"'\` ])" "$f" 2>/dev/null \
+     && ! grep -Eiq "get'?\)?[[:space:]]*==[[:space:]]*ARGV|get'?[^=]*==[^=]*token|compare.?and.?del|redis\.call\('get'" "$f" 2>/dev/null; then
+    note "$f: lock + DEL but no visible token compare-and-delete — a bare DEL can release another owner's lock. Use a Lua GET==token then DEL."
+  fi
+done
+
+# ---- 4. SET ... NX for a lock with no PX/EX -> warn (lock that can't expire) ----
+for f in "${SCAN_FILES[@]}"; do
+  is_test_path "$f" && continue
+  # A SET with NX but neither PX nor EX on the same line/call is a lock that never auto-frees.
+  if grep -Eni "set[^\n]*\bNX\b" "$f" 2>/dev/null | grep -Eivq "\bP?X\b|expire|ttl" 2>/dev/null; then
+    if grep -Eni "set[^\n]*\bNX\b" "$f" 2>/dev/null | grep -Eiq "lock" 2>/dev/null; then
+      note "$f: SET ... NX on a lock with no PX/EX — a crashed owner holds it forever. Add PX <ttl-ms>."
+    fi
+  fi
+done
+
+# ---- 5. INCR not followed by an EXPIRE (counter with no TTL) -> warn ----
+for f in "${SCAN_FILES[@]}"; do
+  is_test_path "$f" && continue
+  if grep -Eniq "(\.incr\(|->incr\(|[\"'\`]INCR[\"'\` ])" "$f" 2>/dev/null \
+     && ! grep -Eniq "(expire|EXPIRE|pexpire|PEXPIRE|EX[[:space:]]|setex|SETEX)" "$f" 2>/dev/null; then
+    note "$f: INCR with no EXPIRE in the same file — a rate-limit counter with no TTL counts forever. Set EXPIRE atomically (Lua, or only when n==1)."
+  fi
+done
+
+# ---- 6/7. redis.conf eviction checks -> warn ----
+for f in ${CONF_FILES[@]+"${CONF_FILES[@]}"}; do
+  if grep -Eiq "^[[:space:]]*maxmemory[[:space:]]" "$f" 2>/dev/null \
+     && ! grep -Eiq "^[[:space:]]*maxmemory-policy" "$f" 2>/dev/null; then
+    note "$f: maxmemory set without maxmemory-policy — defaults to noeviction, which 500s a full cache. Set allkeys-lru/lfu for a cache."
+  fi
+  if grep -Eiq "^[[:space:]]*maxmemory-policy[[:space:]]+noeviction" "$f" 2>/dev/null; then
+    case "$f" in
+      *cache*|*Cache*) note "$f: maxmemory-policy noeviction on a file that looks like a cache config — a cache must evict (allkeys-lru/lfu), not fail writes." ;;
+    esac
+  fi
+done
+
+# ---- 8. while/long loop inside an inline Lua EVAL string -> warn ----
+for f in "${SCAN_FILES[@]}"; do
+  is_test_path "$f" && continue
+  if grep -Eniq "EVAL|eval" "$f" 2>/dev/null \
+     && grep -Eniq "while[[:space:]].*do|for[[:space:]].*=.*,.*do" "$f" 2>/dev/null \
+     && grep -Eniq "redis\.call" "$f" 2>/dev/null; then
+    note "$f: a loop inside Lua run via EVAL blocks ALL clients while it runs (Redis is single-threaded). Keep each EVAL a short bounded step."
+  fi
+done
+
+printf '\n'
+if [ "$EXIT" -eq 0 ]; then ok "verify.sh passed"; else err "verify.sh found failures"; fi
+exit "$EXIT"