@reventlessdev/reventless-seed-aws 1.0.0-alpha.3 → 1.0.0-alpha.30

package/lib/bs/src/ReventlessSeedAws_Reset.res

@@ -0,0 +1,1103 @@
+// AWS "reset" for the seed harness: truncate a deployed stack's durable stores
+// — every DynamoDB table (EventLog, DcbEventLog, QueryDb) and every S3 bucket
+// (tasks, served images) the framework created for the stack — leaving the
+// infrastructure in place so the stack reads empty and is re-seedable. It is the
+// exact inverse of seeding: `Seed.Runner.assertStoreEmpty` refuses to seed a
+// non-empty store; this makes a non-empty store empty again.
+//
+// A wipe is irreversible, so every step is fail-closed and production must be
+// unreachable at several at once (the separate-AWS-account guarantee is the one
+// deferred layer — see docs/plans/done/seed-reset-and-fresh-store-guard.md):
+//
+//   1. name allowlist    — the stack name must match ^(alpha|dev|pr-.+)$
+//   2. wipeable flag      — the target's own `Pulumi.<stack>.yaml` must declare
+//                           `reventless:wipeable: true` (read via `pulumi config`)
+//   3. tag-scoped discovery — targets are found ONLY through the framework's
+//                           `reventless:platform=<project>` + `environment=<stack>`
+//                           tags (Resource Groups Tagging API). Scoping on BOTH
+//                           matters: `environment` is the stack NAME alone, so two
+//                           projects sharing a stack name (e.g. `alpha`) in one
+//                           account collide on it; `platform` (the Pulumi project)
+//                           keeps a wipe inside the one project. No name globbing.
+//   4. per-resource tag re-check — every discovered resource is re-verified to
+//                           carry both tags before a single delete is issued
+//   5. dry-run default + typed confirm — lists what it would empty and stops.
+//                           Interactively, re-typing the exact stack name confirms
+//                           a real wipe (no env var). Without a TTY (CI),
+//                           REVENTLESS_WIPE_CONFIRM=<stack> is the equivalent.
+//
+// Gates 1 and 2 are AND-ed: a stray `wipeable: true` on an unlisted name is
+// still refused, and a listed name without the flag is still refused. Production
+// stacks satisfy neither.
+//
+// A deployment is usually several Pulumi projects sharing a stack name — the
+// platform project plus one per domain plugin, each its own `reventless:platform`.
+// The caller passes those projects as `targets`; the operator picks a scope
+// (`domain` = all plugins, a single plugin, `platform`, or `everything`), and
+// gate 2 + discovery run per selected project. Wiping domain data alone leaves
+// the platform's plugin registry intact, so the store stays re-seedable.
+//
+// Stack resolution reuses `ReventlessSeedAws` (same `pulumi` subprocess, same
+// backend pinning). AWS credentials come from the ambient chain (env / profile /
+// SSO); only the region is resolved explicitly and pinned on the environment, so
+// every SDK client targets the same account/region the tags were read from.
+
+open ReventlessSeed
+
+module Ddb = AwsSdk.DynamoDb
+module Rgt = AwsSdk.ResourceGroupsTaggingApi
+module S3 = AwsSdk.S3
+module Lambda = AwsSdk.Lambda
+module Quiesce = ReventlessSeedAws_Quiesce
+
+// The Pulumi project name, which the framework stamps as `reventless:platform` on
+// every resource (`Plugin.res`: `platformName = getProjectName()`). Discovery
+// MUST scope on this as well as the stack: `reventless:environment` carries only
+// the stack *name*, so two different Pulumi projects deployed with the same stack
+// name (e.g. `alpha`) in one account/region share that tag. Filtering by platform
+// too keeps a wipe inside the one project the operator is standing in. Read from
+// the target project's Pulumi.yaml — the same dir the pulumi subprocess uses.
+let projectName = (~projectDir: string): string => {
+  let path = `${projectDir}/Pulumi.yaml`
+  let raw = try NodeFs.readFileSync(path) catch {
+  | _ => throw(Seed.Failed(`could not read ${path} to scope the wipe to this Pulumi project.`))
+  }
+  switch raw
+  ->String.split("\n")
+  ->Array.find(line => line->String.trim->String.startsWith("name:")) {
+  | Some(line) =>
+    let trimmed = line->String.trim
+    trimmed
+    ->String.slice(~start=String.length("name:"), ~end=String.length(trimmed))
+    ->String.trim
+    ->String.replaceRegExp(/^[\"']|[\"']$/g, "")
+  | None => throw(Seed.Failed(`could not find a \`name:\` field in ${path}.`))
+  }
+}
+
+// A read of a JSON object field — items come back from the document client as
+// already-unmarshalled JSON values, so a delete key is just the key attributes
+// picked back out.
+let field = (json: JSON.t, key: string): option<JSON.t> =>
+  switch json {
+  | Object(obj) => obj->Dict.get(key)
+  | _ => None
+  }
+
+let asString = (json: JSON.t): option<string> =>
+  switch json {
+  | String(s) => Some(s)
+  | _ => None
+  }
+
+let chunk = (arr: array<'a>, size: int): array<array<'a>> => {
+  let out = []
+  let i = ref(0)
+  while i.contents < arr->Array.length {
+    out->Array.push(arr->Array.slice(~start=i.contents, ~end=i.contents + size))
+    i := i.contents + size
+  }
+  out
+}
+
+// ── Gates ───────────────────────────────────────────────────────────────────
+
+// Fail-closed name allowlist. A denylist ("everything except prod") fails open
+// the day a new prod-like stack is added and forgotten; this fails closed.
+let nameAllowlist = /^(alpha|dev|pr-.+)$/
+
+// The stack's fully-resolved Pulumi config, read once. Reading it whole (rather
+// than `pulumi config get <key>`, which exits non-zero for BOTH a missing key
+// and a missing/unreachable stack) lets a refusal say *which* it was — an
+// unreadable config is a pulumi/backend problem; an absent flag is a missing
+// opt-in. `pulumi config --json` maps each key to `{value, secret}`.
+let readStackConfig = (~projectDir, ~backend, ~stack): result<dict<JSON.t>, string> =>
+  switch try Some(
+    ReventlessSeedAws.pulumi(~projectDir, ~backend, ["config", "--json", "--stack", stack]),
+  ) catch {
+  | _ => None
+  } {
+  | None =>
+    Error(
+      `could not read the Pulumi config for stack "${stack}" — is pulumi logged in to the right backend, and is the stack deployed?`,
+    )
+  | Some(raw) =>
+    switch try Some(JSON.parseOrThrow(raw)) catch {
+    | _ => None
+    } {
+    | Some(Object(obj)) => Ok(obj)
+    | _ => Error(`could not parse \`pulumi config --json\` output for stack "${stack}".`)
+    }
+  }
+
+let configValue = (cfg: dict<JSON.t>, key: string): option<string> =>
+  cfg->Dict.get(key)->Option.flatMap(v => v->field("value"))->Option.flatMap(asString)
+
+// ── Discovery ─────────────────────────────────────────────────────────────────
+
+type resource =
+  | Table(string)
+  | Bucket(string)
+  | Function(string)
+  | Other
+
+// arn:aws:dynamodb:<region>:<acct>:table/<name>  |  arn:aws:s3:::<bucket>
+//   |  arn:aws:lambda:<region>:<acct>:function:<name>
+//
+// A Lambda ARN can carry a trailing `:<version|alias>` qualifier. The tagging
+// API returns the unqualified form, but taking the segment after `:function:`
+// up to the next colon rather than to the end of the string means a qualified
+// one still yields the function name instead of a name no API accepts.
+let classify = (arn: string): resource =>
+  if arn->String.startsWith("arn:aws:dynamodb:") {
+    switch arn->String.split(":table/") {
+    | [_, name] => Table(name)
+    | _ => Other
+    }
+  } else if arn->String.startsWith("arn:aws:s3:::") {
+    Bucket(arn->String.slice(~start=String.length("arn:aws:s3:::"), ~end=String.length(arn)))
+  } else if arn->String.startsWith("arn:aws:lambda:") {
+    switch arn->String.split(":function:") {
+    | [_, rest] =>
+      switch rest->String.split(":") {
+      | [name] => Function(name)
+      | [name, _qualifier] => Function(name)
+      | _ => Other
+      }
+    | _ => Other
+    }
+  } else {
+    Other
+  }
+
+let tagValue = (tags: array<Rgt.GetResourcesCommand.tag>, key: string): option<string> =>
+  tags->Array.findMap(t => t.key == key ? Some(t.value) : None)
+
+/** What one project's tags resolve to: the stores to empty, and the runtimes
+    that have to be held still while that happens. */
+type discovered = {
+  tables: array<string>,
+  buckets: array<string>,
+  functions: array<string>,
+}
+
+// Discovery is scoped by BOTH tags (platform AND environment) — the Tagging API
+// ANDs multiple TagFilters — so it only ever sees the target project's own stack,
+// never a same-named stack from another project. Each returned resource is then
+// re-checked to carry both tags (a mismatch aborts the whole run rather than
+// being skipped) as a second, per-resource guard.
+//
+// Lambda functions are discovered alongside the stores, under the same tags and
+// the same per-resource re-check. They are never deleted from — they are what
+// the wipe has to hold still, and the tag scope that decides what may be emptied
+// is exactly the scope that decides what may be held.
+let discover = async (~region, ~stack, ~platform): discovered => {
+  let client = Rgt.client(~region, ())
+  let tables = []
+  let buckets = []
+  let functions = []
+  let rec loop = async (token: option<string>): unit => {
+    let out = await Rgt.GetResourcesCommand.send(
+      client,
+      Rgt.GetResourcesCommand.make({
+        tagFilters: [
+          {key: "reventless:platform", values: [platform]},
+          {key: "reventless:environment", values: [stack]},
+        ],
+        resourceTypeFilters: ["dynamodb:table", "s3", "lambda:function"],
+        resourcesPerPage: 100,
+        paginationToken: ?token,
+      }),
+    )
+    out.resourceTagMappingList
+    ->Option.getOr([])
+    ->Array.forEach(m => {
+      let tags = m.tags->Option.getOr([])
+      let assertTag = (key, expected) =>
+        switch tagValue(tags, key) {
+        | Some(v) if v == expected => ()
+        | other =>
+          throw(
+            Seed.Failed(
+              `refusing: discovered resource ${m.resourceARN} carries ${key}=${other->Option.getOr(
+                  "<none>",
+                )}, not "${expected}".`,
+            ),
+          )
+        }
+      assertTag("reventless:platform", platform)
+      assertTag("reventless:environment", stack)
+      switch classify(m.resourceARN) {
+      | Table(name) => tables->Array.push(name)
+      | Bucket(name) => buckets->Array.push(name)
+      | Function(name) => functions->Array.push(name)
+      | Other => ()
+      }
+    })
+    switch out.paginationToken {
+    | Some(t) if t != "" => await loop(Some(t))
+    | _ => ()
+    }
+  }
+  await loop(None)
+  {tables, buckets, functions}
+}
+
+// ── Declared object stores ────────────────────────────────────────────────────
+//
+// A store declared by a `@storageRef` field is provisioned by the PLATFORM
+// deploy — the serving CDN and the presign services live there, and a
+// shared-layout bucket holds several plugins' stores, so no plugin stack can own
+// it. Its bucket therefore carries the platform project's `reventless:platform`
+// tag, which makes it invisible to a domain-scoped tag discovery and, worse,
+// wholesale-emptiable by a platform-scoped one.
+//
+// Tag discovery cannot fix that: the tag says which project *built* the bucket,
+// and the answer needed here is which plugin's data is *in* it — at prefix
+// granularity, since one bucket holds several plugins' stores. The platform
+// already publishes exactly that mapping as its `objectStores` stack output, so
+// the reset reads ownership from the declaration rather than inferring it.
+
+/** One provisioned store: which plugin owns it, and where its objects live. */
+type objectStore = {
+  qualified: string,
+  plugin: string,
+  store: string,
+  bucketName: string,
+  keyPrefix: string,
+}
+
+// A store's key is `{plugin}.{store}`, split at the FIRST dot: a registered
+// plugin name cannot contain one, a store name could in principle.
+let splitQualified = (key: string): option<(string, string)> =>
+  key
+  ->String.indexOfOpt(".")
+  ->Option.map(i => (
+    key->String.slice(~start=0, ~end=i),
+    key->String.slice(~start=i + 1, ~end=key->String.length),
+  ))
+
+/** Parse the platform stack's `objectStores` output. Absent is normal — a
+    platform may declare no stores — so only a malformed entry is an error, and
+    it is an error rather than a skip because a store the reset cannot read is a
+    store it would silently leave behind. */
+let parseObjectStores = (json: option<JSON.t>): result<array<objectStore>, string> =>
+  switch json {
+  | None => Ok([])
+  | Some(Object(entries)) =>
+    entries
+    ->Dict.toArray
+    ->Array.reduce(Ok([]), (acc, (qualified, entry)) =>
+      switch acc {
+      | Error(_) as failed => failed
+      | Ok(stores) =>
+        switch (
+          splitQualified(qualified),
+          entry->field("bucketName")->Option.flatMap(asString),
+          entry->field("keyPrefix")->Option.flatMap(asString),
+        ) {
+        | (Some((plugin, store)), Some(bucketName), Some(keyPrefix)) =>
+          Ok(Array.concat(stores, [{qualified, plugin, store, bucketName, keyPrefix}]))
+        | _ =>
+          Error(
+            `the platform stack's \`objectStores\` output has a malformed entry for "${qualified}" — ` ++ `expected a {plugin}.{store} key carrying bucketName and keyPrefix.`,
+          )
+        }
+      }
+    )
+  | Some(_) => Error("the platform stack's `objectStores` output is not an object.")
+  }
+
+/** Refuse any store set a prefix-scoped wipe cannot separate.
+
+    Equality is the cross-plugin collision: two plugins declaring one store name
+    land on one prefix inside a shared bucket, and nothing distinguishes their
+    objects. Containment is the same problem one level up — the delete prefix is
+    `{keyPrefix}/`, so a store rooted at `a` encloses one rooted at `a/b`.
+    Comparing containment rather than equality is what keeps this correct once a
+    prefix carries path structure.
+
+    Fail-closed, and before anything is counted: this is the tool that destroys
+    data, so it refuses rather than skipping, and it does not assume an upstream
+    deploy-time check ran. */
+let validateStores = (stores: array<objectStore>): result<unit, string> =>
+  switch stores->Array.find(s => s.keyPrefix == "" || s.store->String.includes("/")) {
+  | Some(s) =>
+    Error(
+      `store "${s.qualified}" has an unusable key prefix ("${s.keyPrefix}") — ` ++ `a store name may not be empty or contain "/".`,
+    )
+  | None =>
+    switch stores->Array.findMap(a =>
+      stores->Array.findMap(b =>
+        if a.qualified == b.qualified || a.bucketName != b.bucketName {
+          None
+        } else if a.keyPrefix == b.keyPrefix {
+          Some(
+            `stores "${a.qualified}" and "${b.qualified}" both live at ` ++
+            `${a.bucketName}/${a.keyPrefix}/ — a prefix-scoped wipe cannot tell their objects ` ++
+            `apart. Rename one store, or qualify the \`@storageRef\` annotation if they were ` ++ `meant to be one shared store.`,
+          )
+        } else if b.keyPrefix->String.startsWith(a.keyPrefix ++ "/") {
+          Some(
+            `store "${a.qualified}" (${a.bucketName}/${a.keyPrefix}/) encloses "${b.qualified}" ` ++
+            `(${b.keyPrefix}/) — wiping the first would delete the second's objects. Rename one.`,
+          )
+        } else {
+          None
+        }
+      )
+    ) {
+    | Some(message) => Error(message)
+    | None => Ok()
+    }
+  }
+
+// ── Counting (dry-run) ─────────────────────────────────────────────────────────
+
+let countTable = async (table: string): int => {
+  let rec loop = async (start: option<dict<JSON.t>>, acc: int): int => {
+    let out = await Ddb.DocumentClient.ScanCommand.send(
+      Ddb.DocumentClient.ScanCommand.make({
+        tableName: table,
+        select: #COUNT,
+        exclusiveStartKey: ?start,
+      }),
+    )
+    let acc = acc + out.count->Option.getOr(0)
+    switch out.lastEvaluatedKey {
+    | Some(k) => await loop(Some(k), acc)
+    | None => acc
+    }
+  }
+  await loop(None, 0)
+}
+
+// `~prefix` narrows the count to one declared store inside a shared bucket;
+// omitted, it counts the whole bucket.
+let countBucket = async (bucket: string, ~prefix: option<string>=?): int => {
+  let rec loop = async (keyMarker, versionMarker, acc): int => {
+    let out = await S3.ListObjectVersionsCommand.send(
+      S3.ListObjectVersionsCommand.make({
+        bucket,
+        ?prefix,
+        ?keyMarker,
+        versionIdMarker: ?versionMarker,
+      }),
+    )
+    let n =
+      out.versions->Option.getOr([])->Array.length +
+        out.deleteMarkers->Option.getOr([])->Array.length
+    if out.isTruncated->Option.getOr(false) {
+      await loop(out.nextKeyMarker, out.nextVersionIdMarker, acc + n)
+    } else {
+      acc + n
+    }
+  }
+  await loop(None, None, 0)
+}
+
+// ── Wiping ──────────────────────────────────────────────────────────────────
+
+// BatchWrite takes ≤ 25 requests and may return some UnprocessedItems under
+// throttling; resend those. Capped so a persistently-failing table surfaces as
+// an error rather than looping forever.
+let rec sendBatch = async (
+  table: string,
+  requests: array<Ddb.DocumentClient.BatchWriteCommand.writeRequest>,
+  ~attempt: int,
+): unit =>
+  if requests->Array.length > 0 {
+    if attempt > 8 {
+      throw(
+        Seed.Failed(
+          `table ${table}: ${requests
+            ->Array.length
+            ->Int.toString} item(s) still unprocessed after 8 retries.`,
+        ),
+      )
+    }
+    let out = await Ddb.DocumentClient.BatchWriteCommand.send(
+      Ddb.DocumentClient.BatchWriteCommand.make({
+        requestItems: Dict.fromArray([(table, requests)]),
+      }),
+    )
+    let unprocessed =
+      out.unprocessedItems->Option.flatMap(d => d->Dict.get(table))->Option.getOr([])
+    if unprocessed->Array.length > 0 {
+      await sendBatch(table, unprocessed, ~attempt=attempt + 1)
+    }
+  }
+
+let truncateTable = async (table: string): unit => {
+  let desc = await Ddb.DynamoDb.DescribeTableCommand.send(
+    Ddb.DynamoDb.DescribeTableCommand.make({tableName: table}),
+  )
+  let keyAttrs =
+    desc.table->Option.flatMap(t => t.keySchema)->Option.getOr([])->Array.map(k => k.attributeName)
+  if keyAttrs->Array.length == 0 {
+    throw(Seed.Failed(`could not read a key schema for table ${table}.`))
+  }
+  // Project only the key attributes (aliased to dodge reserved words) so the
+  // scan carries just what a delete needs.
+  let names = keyAttrs->Array.mapWithIndex((name, i) => (`#k${i->Int.toString}`, name))
+  let projection = names->Array.map(((alias, _)) => alias)->Array.join(", ")
+  let rec loop = async (start: option<dict<JSON.t>>): unit => {
+    let out = await Ddb.DocumentClient.ScanCommand.send(
+      Ddb.DocumentClient.ScanCommand.make({
+        tableName: table,
+        exclusiveStartKey: ?start,
+        projectionExpression: projection,
+        expressionAttributeNames: names->Dict.fromArray,
+      }),
+    )
+    let requests =
+      out.items
+      ->Option.getOr([])
+      ->Array.map(item => {
+        let key =
+          keyAttrs
+          ->Array.filterMap(attr => item->field(attr)->Option.map(v => (attr, v)))
+          ->Dict.fromArray
+        ({deleteRequest: {key: key}}: Ddb.DocumentClient.BatchWriteCommand.writeRequest)
+      })
+    let batches = chunk(requests, Ddb.DocumentClient.BatchWriteCommand.maxBatchSize)
+    for i in 0 to batches->Array.length - 1 {
+      switch batches->Array.get(i) {
+      | Some(b) => await sendBatch(table, b, ~attempt=1)
+      | None => ()
+      }
+    }
+    switch out.lastEvaluatedKey {
+    | Some(k) => await loop(Some(k))
+    | None => ()
+    }
+  }
+  await loop(None)
+}
+
+// One ListObjectVersions page returns ≤ 1000 entries (versions + delete
+// markers), and DeleteObjects takes ≤ 1000, so one list page maps to one delete.
+let emptyBucket = async (bucket: string, ~prefix: option<string>=?): unit => {
+  let rec loop = async (keyMarker, versionMarker): unit => {
+    let out = await S3.ListObjectVersionsCommand.send(
+      S3.ListObjectVersionsCommand.make({
+        bucket,
+        ?prefix,
+        ?keyMarker,
+        versionIdMarker: ?versionMarker,
+      }),
+    )
+    let ids =
+      Array.concat(out.versions->Option.getOr([]), out.deleteMarkers->Option.getOr([]))->Array.map((
+        v
+      ): S3.DeleteObjectsCommand.objectIdentifier => {key: v.key, versionId: v.versionId})
+    if ids->Array.length > 0 {
+      let res = await S3.DeleteObjectsCommand.send(
+        S3.DeleteObjectsCommand.make({
+          bucket,
+          delete: {objects: ids, quiet: true},
+        }),
+      )
+      switch res.errors {
+      | Some(errs) if errs->Array.length > 0 =>
+        throw(
+          Seed.Failed(
+            `failed to delete ${errs->Array.length->Int.toString} object(s) from ${bucket}: ${errs
+              ->Array.get(0)
+              ->Option.flatMap(e => e.message)
+              ->Option.getOr("unknown")}`,
+          ),
+        )
+      | _ => ()
+      }
+    }
+    if out.isTruncated->Option.getOr(false) {
+      await loop(out.nextKeyMarker, out.nextVersionIdMarker)
+    }
+  }
+  await loop(None, None)
+}
+
+// ── Targets & scope ───────────────────────────────────────────────────────────
+
+// A deployment is several Pulumi projects sharing a stack name — the platform
+// project plus one per domain plugin — each a separate `reventless:platform`. A
+// target names one such project by the directory its `Pulumi.<stack>.yaml` lives
+// in (relative to the seed cwd), plus a menu label and whether it holds domain
+// data or platform bookkeeping. The caller declares them; the reset never guesses
+// the topology.
+type group =
+  | Domain
+  | Platform
+
+type target = {
+  projectDir: string,
+  label: string,
+  group: group,
+  // The name this project's plugin REGISTERS, when it differs from the
+  // operator-facing `label` (`Catalog` against `catalog`). It is what the
+  // platform's `objectStores` keys are qualified by, so it is how a declared
+  // store is attributed to a target. Declared rather than case-folded from the
+  // label: the caller states the topology, the reset never guesses it.
+  plugin?: string,
+}
+
+let pluginOf = (t: target): string => t.plugin->Option.getOr(t.label)
+
+// A target resolved to its discovered, counted stores, ready to report and wipe.
+type resolved = {
+  target: target,
+  platform: string,
+  tables: array<string>,
+  tableCounts: array<int>,
+  bucketCounts: array<(string, int)>,
+  storeCounts: array<(objectStore, int)>,
+  /** The project's Lambda functions — not wiped, held still while it is. */
+  functions: array<string>,
+}
+
+// Picks which targets to wipe. `domain` (every domain plugin) leads and is the
+// default; each single domain plugin follows so one plugin's data can be wiped
+// alone; then `platform`; then `everything`. `SEED_RESET_SCOPE` (domain |
+// platform | everything | a plugin label) selects non-interactively.
+let chooseScope = async (~targets: array<target>): array<target> => {
+  let domain = targets->Array.filter(t => t.group == Domain)
+  let platform = targets->Array.filter(t => t.group == Platform)
+  let labelsOf = ts => ts->Array.map(t => t.label)->Array.join(", ")
+  switch Seed.Prompt.envValue("SEED_RESET_SCOPE")->Option.map(String.toLowerCase) {
+  | Some("domain") => domain
+  | Some("platform") => platform
+  | Some("all") | Some("everything") | Some("both") => targets
+  | Some(other) =>
+    switch targets->Array.find(t => t.label->String.toLowerCase == other) {
+    | Some(t) => [t]
+    | None =>
+      throw(
+        Seed.Failed(
+          `SEED_RESET_SCOPE="${other}" is not a scope — use domain, platform, everything, or a plugin label (${labelsOf(
+              domain,
+            )}).`,
+        ),
+      )
+    }
+  | None =>
+    let options = []
+    if domain->Array.length > 0 {
+      options->Array.push((`domain — ${labelsOf(domain)}`, domain))
+
+      // Single-plugin entries only when there is more than one, else they just
+      // duplicate the `domain` entry.
+      if domain->Array.length > 1 {
+        domain->Array.forEach(t => options->Array.push((t.label, [t])))
+      }
+    }
+    if platform->Array.length > 0 {
+      options->Array.push((`platform — ${labelsOf(platform)}`, platform))
+    }
+    if domain->Array.length > 0 && platform->Array.length > 0 {
+      options->Array.push((`everything — ${labelsOf(targets)}`, targets))
+    }
+    await Seed.Prompt.select(~title="Reset scope:", ~options)
+  }
+}
+
+// ── Orchestration ─────────────────────────────────────────────────────────────
+
+// Gate 2 per target: its own `Pulumi.<stack>.yaml` must declare wipeable, and
+// must resolve a region. Reasons are distinct (config unreadable vs opt-in
+// missing vs region missing), each saying what to fix. Returns the target's
+// region so the caller can insist every selected target shares one.
+let gateTarget = (~target: target, ~backend, ~stack): string => {
+  let cfg = switch readStackConfig(~projectDir=target.projectDir, ~backend, ~stack) {
+  | Ok(c) => c
+  | Error(message) => throw(Seed.Failed(message))
+  }
+  switch configValue(cfg, "reventless:wipeable")->Option.map(v =>
+    v->String.trim->String.toLowerCase
+  ) {
+  | Some("true") => ()
+  | _ =>
+    throw(
+      Seed.Failed(
+        `${target.label}: stack "${stack}" does not declare \`reventless:wipeable: "true"\` in its Pulumi.${stack}.yaml — refusing. Add that line only on disposable dev stacks (see the alpha example configs).`,
+      ),
+    )
+  }
+  switch Seed.Prompt.envValue("AWS_REGION") {
+  | Some(r) if r != "" => r
+  | _ =>
+    switch configValue(cfg, "aws:region") {
+    | Some(r) if r != "" => r
+    | _ =>
+      throw(
+        Seed.Failed(
+          `${target.label}: could not resolve the AWS region — set AWS_REGION or \`aws:region\` in the stack config.`,
+        ),
+      )
+    }
+  }
+}
+
+/** Opt out of holding the platform still for the wipe. The hold needs Lambda
+    permissions the reset did not previously ask for, so credentials that lack
+    them have a way through — at the cost of the emptiness guarantee, which the
+    note printed alongside it says plainly. */
+let noQuiesce = (): bool =>
+  switch Seed.Prompt.envValue("SEED_RESET_NO_QUIESCE")->Option.map(v =>
+    v->String.trim->String.toLowerCase
+  ) {
+  | Some("1") | Some("true") | Some("yes") => true
+  | _ => false
+  }
+
+/** What to say when a store is not empty after the wipe.
+
+    The old message reported only a total and advised re-running. That advice was
+    wrong in exactly the case that occurs: a runtime holding its state in memory
+    rewrites the store on every invocation, so every re-run reproduces the same
+    failure and the message sends the operator round the loop again.
+
+    So name the stores, and separate the two situations that are actually
+    different. Under a hold, nothing could start during the wipe, so the only
+    possible writer is an invocation that was already in flight — and a re-run
+    does clear that. Without a hold there is no such bound, and re-running is the
+    wrong advice. */
+let refillMessage = (~refilled: array<(string, int)>, ~heldCount: int): string => {
+  let total = refilled->Array.reduce(0, (acc, (_, n)) => acc + n)
+  let lines =
+    refilled
+    ->Array.map(((label, n)) => `\n      ${n->Int.toString->String.padStart(8, " ")}  ${label}`)
+    ->Array.join("")
+  let diagnosis = if heldCount > 0 {
+    `\n\n  The wipe deleted these and something wrote them back. All ${heldCount->Int.toString} runtime(s) ` ++
+    `in scope were held at zero concurrency for the whole wipe, so the writer was an invocation already ` ++ `in flight when the hold took effect. Re-running clears it.`
+  } else {
+    `\n\n  The wipe deleted these and a running runtime wrote them back. The stack was NOT held ` ++
+    `(SEED_RESET_NO_QUIESCE), so nothing stopped it, and re-running will not help: the writer keeps the ` ++
+    `state in memory and restores it on every invocation, identical each time. Re-run without ` ++ `SEED_RESET_NO_QUIESCE so the runtimes are held and recycled.`
+  }
+  `${total->Int.toString} item(s)/object(s) came back after the wipe:${lines}${diagnosis}`
+}
+
+let reportAll = (resolvedList: array<resolved>, ~stack, ~region): int => {
+  Seed.Runner.heading(`Reset target: stack "${stack}" in ${region}`)
+  let total = ref(0)
+  resolvedList->Array.forEach(r => {
+    Console.log("")
+    Console.log(`  ${r.target.label}  (${r.platform})`)
+    Console.log("    DynamoDB tables:")
+    r.tables->Array.forEachWithIndex((t, i) => {
+      let c = r.tableCounts->Array.get(i)->Option.getOr(0)
+      total := total.contents + c
+      Console.log(`      ${c->Int.toString->String.padStart(8, " ")}  ${t}`)
+    })
+    if r.tables->Array.length == 0 {
+      Console.log("      (none)")
+    }
+    Console.log("    S3 buckets:")
+    r.bucketCounts->Array.forEach(((b, c)) => {
+      total := total.contents + c
+      Console.log(`      ${c->Int.toString->String.padStart(8, " ")}  ${b}`)
+    })
+    if r.bucketCounts->Array.length == 0 {
+      Console.log("      (none)")
+    }
+
+    // Declared stores get their own section rather than being folded in with the
+    // plain buckets: the unit is a prefix inside a bucket that other plugins also
+    // write to, and the operator needs to see which is which before confirming.
+    if r.storeCounts->Array.length > 0 {
+      Console.log("    Object stores:")
+      r.storeCounts->Array.forEach(((s, c)) => {
+        total := total.contents + c
+        Console.log(
+          `      ${c->Int.toString->String.padStart(8, " ")}  ${s.qualified}   ` ++
+          `${s.bucketName}/${s.keyPrefix}/`,
+        )
+      })
+    }
+  })
+  total.contents
+}
+
+/**
+ * Reset a deployed stack across one or more of its Pulumi projects. Resolves the
+ * shared stack name, refuses unless it is on the name allowlist, lets the
+ * operator pick a scope (domain / a single plugin / platform / everything), then
+ * for each chosen project: refuses unless it declares itself wipeable, discovers
+ * its stores by `platform`+`environment` tag, and reports what it would empty.
+ * Only on a matching `REVENTLESS_WIPE_CONFIRM` plus a re-typed stack name does it
+ * truncate every table and empty every bucket, then verify empty.
+ *
+ * `targets` are the deployment's projects (the caller knows the topology);
+ * `stack` and `backend` mirror `ReventlessSeedAws.connect`.
+ */
+let run = (~stack=?, ~backend=?, ~targets: array<target>, ()): unit => {
+  let go = async () => {
+    try {
+      if targets->Array.length == 0 {
+        throw(Seed.Failed("no targets were declared — nothing to reset."))
+      }
+      let backend = switch Seed.Prompt.envValue("SEED_PULUMI_BACKEND") {
+      | Some(url) => Some(url)
+      | None => backend
+      }
+      // Any target's dir lists the shared stack; prefer the platform project.
+      let baseTarget =
+        targets->Array.find(t => t.group == Platform)->Option.getOr(targets->Array.getUnsafe(0))
+      let stack = await ReventlessSeedAws.resolveStack(
+        ~projectDir=baseTarget.projectDir,
+        ~backend,
+        ~stack,
+      )
+
+      if !(nameAllowlist->RegExp.test(stack)) {
+        throw(
+          Seed.Failed(
+            `stack "${stack}" is not on the wipe name-allowlist (alpha, dev, pr-*) — refusing.`,
+          ),
+        )
+      }
+
+      let selected = await chooseScope(~targets)
+
+      // Declared object stores live in a bucket the PLATFORM project owns, so
+      // resolving them reads the platform's stack output whichever scope was
+      // picked. Reading is side-effect free; the gates below decide whether
+      // anything may be deleted from it.
+      let platformTarget = targets->Array.find(t => t.group == Platform)
+      let allStores = switch platformTarget {
+      | None =>
+        // Not silent: a topology with no platform target is exactly the case
+        // where declared stores would be missed, and missing them is the bug
+        // this resolution exists to fix.
+        Console.log("")
+        Console.log(
+          "Note: no `platform` target is declared, so declared object stores could not be " ++ "resolved — any uploaded objects will be left in place.",
+        )
+        []
+      | Some(pt) =>
+        let output = ReventlessSeedAws.stackOutputs(~projectDir=pt.projectDir, ~backend, stack)
+        switch parseObjectStores(output->field("objectStores")) {
+        | Ok(stores) => stores
+        | Error(message) => throw(Seed.Failed(message))
+        }
+      }
+      switch validateStores(allStores) {
+      | Ok() => ()
+      | Error(message) => throw(Seed.Failed(`refusing: ${message}`))
+      }
+
+      // Every bucket that holds a declared store, selected or not. These are
+      // excluded from the plain per-target bucket lists below so a store is
+      // reachable ONLY through the plugin that declared it — otherwise the
+      // platform scope would empty every plugin's objects wholesale.
+      let storeBucketNames = allStores->Array.map(s => s.bucketName)
+      let selectedStores =
+        allStores->Array.filter(s => selected->Array.some(t => pluginOf(t) == s.plugin))
+
+      // Gate every selected target and collect its region; all must agree, since
+      // the DynamoDB/S3 clients read one region from the environment. When a
+      // store is in scope, the platform's own stack must also declare itself
+      // wipeable — the objects belong to a plugin, but the bucket is the
+      // platform's, and both consents are needed to delete from it.
+      let gated = switch (selectedStores->Array.length > 0, platformTarget) {
+      | (true, Some(pt)) if !(selected->Array.some(t => t.projectDir == pt.projectDir)) =>
+        Array.concat(selected, [pt])
+      | _ => selected
+      }
+      let regions = gated->Array.map(t => gateTarget(~target=t, ~backend, ~stack))
+      let region = regions->Array.getUnsafe(0)
+      if regions->Array.some(r => r != region) {
+        throw(
+          Seed.Failed(
+            `selected targets span more than one region (${regions->Array.join(
+                ", ",
+              )}) — reset them one region at a time.`,
+          ),
+        )
+      }
+      NodeProcess.env->Dict.set("AWS_REGION", region)
+
+      // A store bucket arrives by stack output, not by tag discovery, so the
+      // per-resource tag re-check has to be applied to it explicitly: confirm it
+      // carries the platform project's own `platform`+`environment` tags before
+      // anything is deleted from it. One extra tagging-API call, and only when a
+      // store is actually in scope.
+      if selectedStores->Array.length > 0 {
+        switch platformTarget {
+        | Some(pt) =>
+          let platformProject = projectName(~projectDir=pt.projectDir)
+          let platformFound = await discover(~region, ~stack, ~platform=platformProject)
+          selectedStores->Array.forEach(s =>
+            if !(platformFound.buckets->Array.includes(s.bucketName)) {
+              throw(
+                Seed.Failed(
+                  `refusing: store "${s.qualified}" names bucket ${s.bucketName}, which does not ` ++
+                  `carry reventless:platform=${platformProject} + reventless:environment=${stack}.`,
+                ),
+              )
+            }
+          )
+        | None => ()
+        }
+      }
+
+      // Discover + count each target, scoped to its own project via the platform
+      // tag so a same-named stack from another project is never touched.
+      let resolvedList = []
+      for i in 0 to selected->Array.length - 1 {
+        switch selected->Array.get(i) {
+        | Some(target) =>
+          let platform = projectName(~projectDir=target.projectDir)
+          let found = await discover(~region, ~stack, ~platform)
+          let tables = found.tables->Array.toSorted(String.compare)
+          let buckets =
+            found.buckets
+            ->Array.filter(b => !(storeBucketNames->Array.includes(b)))
+            ->Array.toSorted(String.compare)
+          let tableCounts = []
+          for j in 0 to tables->Array.length - 1 {
+            switch tables->Array.get(j) {
+            | Some(t) => tableCounts->Array.push(await countTable(t))
+            | None => ()
+            }
+          }
+          let bucketCounts = []
+          for j in 0 to buckets->Array.length - 1 {
+            switch buckets->Array.get(j) {
+            | Some(b) => bucketCounts->Array.push((b, await countBucket(b)))
+            | None => ()
+            }
+          }
+          let stores = selectedStores->Array.filter(s => s.plugin == pluginOf(target))
+          let storeCounts = []
+          for j in 0 to stores->Array.length - 1 {
+            switch stores->Array.get(j) {
+            | Some(s) =>
+              storeCounts->Array.push((
+                s,
+                await countBucket(s.bucketName, ~prefix=`${s.keyPrefix}/`),
+              ))
+            | None => ()
+            }
+          }
+          resolvedList->Array.push({
+            target,
+            platform,
+            tables,
+            tableCounts,
+            bucketCounts,
+            storeCounts,
+            functions: found.functions->Array.toSorted(String.compare),
+          })
+        | None => ()
+        }
+      }
+
+      let total = reportAll(resolvedList, ~stack, ~region)
+      if total == 0 {
+        Seed.Prompt.close()
+        Console.log("")
+        Console.log(`Nothing to reset — the selected scope already reads empty in ${region}.`)
+        NodeProcess.exit(0)
+      }
+
+      // Confirmation. Interactively, re-typing the exact stack name IS the
+      // confirmation — the validation above (allowlist + wipeable + tag checks)
+      // plus a deliberate keystroke is enough, so no env var is needed. Without a
+      // TTY (CI/scripts) there is nothing to type into, so
+      // REVENTLESS_WIPE_CONFIRM=<stack> is the equivalent opt-in. Dry-run is the
+      // default either way.
+      let interactive = NodeProcess.stdin->NodeProcess.isTTY->Option.getOr(false)
+      let confirmed = if interactive {
+        let typed = await Seed.Prompt.ask(
+          `About to permanently empty ${total->Int.toString} item(s)/object(s) across the selected scope of "${stack}". ` ++ `Type the stack name to confirm, or press Enter to keep this a dry run: `,
+        )
+        typed->String.trim == stack
+      } else {
+        Seed.Prompt.envValue("REVENTLESS_WIPE_CONFIRM") == Some(stack)
+      }
+      Seed.Prompt.close()
+      if !confirmed {
+        Console.log("")
+        Console.log(
+          interactive
+            ? "Dry run — nothing was deleted."
+            : `Dry run — nothing was deleted. Set REVENTLESS_WIPE_CONFIRM=${stack} to empty this scope non-interactively.`,
+        )
+        NodeProcess.exit(0)
+      }
+
+      // Hold the platform still before the first delete. A truncate is not
+      // durable while the runtimes that own the data are running — see
+      // ReventlessSeedAws_Quiesce for the mechanism and for why emptying the
+      // upstream stores first does not substitute for this.
+      let functionNames =
+        resolvedList->Array.reduce([], (acc, r) =>
+          Array.concat(acc, r.functions->Array.filter(f => !(acc->Array.includes(f))))
+        )
+      let quiesce = !noQuiesce() && functionNames->Array.length > 0
+      let lambdaClient = Lambda.client(~region, ())
+      let held = if quiesce {
+        Seed.Runner.heading(
+          `Holding ${functionNames->Array.length->Int.toString} runtime(s) at zero concurrency …`,
+        )
+        await Quiesce.hold(~client=lambdaClient, ~functionNames)
+      } else {
+        if functionNames->Array.length > 0 {
+          Console.log("")
+          Console.log(
+            "Note: SEED_RESET_NO_QUIESCE is set — the stack's runtimes keep running through the wipe. " ++ "A runtime that holds state across invocations can write it straight back over an emptied store.",
+          )
+        }
+        []
+      }
+
+      // From here the reset owns the platform's concurrency, so every exit has
+      // to run through the release below — including the abort path. ReScript
+      // has no `finally`, so the wipe's outcome is captured as a value and
+      // re-raised after the stack has been handed back.
+      let outcome = try {
+        for i in 0 to resolvedList->Array.length - 1 {
+          switch resolvedList->Array.get(i) {
+          | Some(r) =>
+            Seed.Runner.heading(`Emptying ${r.target.label} …`)
+            for j in 0 to r.tables->Array.length - 1 {
+              switch r.tables->Array.get(j) {
+              | Some(t) =>
+                await truncateTable(t)
+                Console.log(`  truncated ${t}`)
+              | None => ()
+              }
+            }
+            for j in 0 to r.bucketCounts->Array.length - 1 {
+              switch r.bucketCounts->Array.get(j) {
+              | Some((b, _)) =>
+                await emptyBucket(b)
+                Console.log(`  emptied ${b}`)
+              | None => ()
+              }
+            }
+            // Report the count with the store, not just in the run total: "how
+            // many images went" is the question an operator is actually asking.
+            for j in 0 to r.storeCounts->Array.length - 1 {
+              switch r.storeCounts->Array.get(j) {
+              | Some((s, count)) =>
+                await emptyBucket(s.bucketName, ~prefix=`${s.keyPrefix}/`)
+                Console.log(
+                  `  emptied ${s.qualified} — ${count->Int.toString} object(s) removed from ` ++
+                  `${s.bucketName}/${s.keyPrefix}/`,
+                )
+              | None => ()
+              }
+            }
+          | None => ()
+          }
+        }
+
+        // Prove empty — the inverse of the post-seed verify. Every store that
+        // still reads non-zero is named, because "12 items remain" without
+        // saying where is the difference between a diagnosis and a mystery.
+        let refilled = []
+        for i in 0 to resolvedList->Array.length - 1 {
+          switch resolvedList->Array.get(i) {
+          | Some(r) =>
+            for j in 0 to r.tables->Array.length - 1 {
+              switch r.tables->Array.get(j) {
+              | Some(t) =>
+                let left = await countTable(t)
+                if left != 0 {
+                  refilled->Array.push((t, left))
+                }
+              | None => ()
+              }
+            }
+            for j in 0 to r.bucketCounts->Array.length - 1 {
+              switch r.bucketCounts->Array.get(j) {
+              | Some((b, _)) =>
+                let left = await countBucket(b)
+                if left != 0 {
+                  refilled->Array.push((b, left))
+                }
+              | None => ()
+              }
+            }
+            // Per-store confirmation, so "every uploaded object is gone" is an
+            // observed fact rather than an inference from a global total.
+            for j in 0 to r.storeCounts->Array.length - 1 {
+              switch r.storeCounts->Array.get(j) {
+              | Some((s, _)) =>
+                let left = await countBucket(s.bucketName, ~prefix=`${s.keyPrefix}/`)
+                if left != 0 {
+                  refilled->Array.push((`${s.qualified} (${s.bucketName}/${s.keyPrefix}/)`, left))
+                }
+                Console.log(
+                  `  verified empty: ${s.qualified} — ${left->Int.toString} object(s) remain under ` ++
+                  `${s.bucketName}/${s.keyPrefix}/`,
+                )
+              | None => ()
+              }
+            }
+          | None => ()
+          }
+        }
+        Ok(refilled)
+      } catch {
+      | Seed.Failed(message) => Error(message)
+      | exn =>
+        Error(
+          `the wipe failed: ${exn
+            ->JsExn.fromException
+            ->Option.flatMap(JsExn.message)
+            ->Option.getOr("unknown error")}`,
+        )
+      }
+
+      // Recycle while still held, so the container that eventually serves
+      // traffic is created after the stores were emptied — then hand the stack
+      // back. Both are best-effort and report rather than raise: they run on the
+      // abort path too, where throwing would bury the reason the reset failed.
+      let warnings = if held->Array.length > 0 {
+        Seed.Runner.heading(`Recycling ${held->Array.length->Int.toString} runtime(s) …`)
+        let recycled = await Quiesce.recycle(~client=lambdaClient, ~held)
+        let released = await Quiesce.release(~client=lambdaClient, ~held)
+        Array.concat(recycled, released)
+      } else {
+        []
+      }
+      if warnings->Array.length > 0 {
+        Console.error("")
+        warnings->Array.forEach(w => Console.error(`  warning: ${w}`))
+      }
+
+      switch outcome {
+      | Error(message) => throw(Seed.Failed(message))
+      | Ok(refilled) if refilled->Array.length > 0 =>
+        throw(Seed.Failed(refillMessage(~refilled, ~heldCount=held->Array.length)))
+      | Ok(_) => ()
+      }
+
+      Console.log("")
+      Console.log(
+        `Reset complete — the selected scope of "${stack}" reads empty and is re-seedable.`,
+      )
+      NodeProcess.exit(0)
+    } catch {
+    | Seed.Failed(message) =>
+      Seed.Prompt.close()
+      Console.error("")
+      Console.error(`Reset aborted — ${message}`)
+      NodeProcess.exit(1)
+    | exn =>
+      Seed.Prompt.close()
+      Console.error("")
+      Console.error("Reset aborted with an unexpected error:")
+      Console.error(
+        exn->JsExn.fromException->Option.flatMap(JsExn.message)->Option.getOr("unknown"),
+      )
+      NodeProcess.exit(1)
+    }
+  }
+  go()->ignore
+}