diff --git a/crates/core/src/migrate/store.rs b/crates/core/src/migrate/store.rs
index a343e16..ec68759 100644
--- a/crates/core/src/migrate/store.rs
+++ b/crates/core/src/migrate/store.rs
@@ -13,10 +13,7 @@ use fjall::{
     CompressionType, Database, Keyspace, KeyspaceCreateOptions, KvSeparationOptions,
 };
 use mail_parser::MessageParser;
-use tantivy::{
-    indexer::{LogMergePolicy, NoMergePolicy},
-    Index, IndexWriter, TantivyDocument,
-};
+use tantivy::{indexer::NoMergePolicy, Index, IndexWriter, TantivyDocument};
 use uuid::Uuid;
 
 use crate::{
diff --git a/crates/core/src/store/tantivy/attachment.rs b/crates/core/src/store/tantivy/attachment.rs
index 5129061..1b14009 100644
--- a/crates/core/src/store/tantivy/attachment.rs
+++ b/crates/core/src/store/tantivy/attachment.rs
@@ -78,6 +78,10 @@ pub struct IndexManager {
 }
 
 impl IndexManager {
+    pub(crate) fn index_writer(&self) -> &Arc<Mutex<IndexWriter>> {
+        &self.index_writer
+    }
+
     pub async fn shutdown(&self) {
         let mut guard = self.handle.lock().await;
         if let Some(handle) = guard.take() {
diff --git a/crates/core/src/store/tantivy/dedup.rs b/crates/core/src/store/tantivy/dedup.rs
new file mode 100644
index 0000000..177e512
--- /dev/null
+++ b/crates/core/src/store/tantivy/dedup.rs
@@ -0,0 +1,711 @@
+use std::collections::HashMap;
+
+use tantivy::schema::Term;
+use tantivy::{IndexReader, IndexWriter};
+
+use crate::common::periodic::{PeriodicTask, TaskHandle};
+use crate::context::BichonTask;
+use crate::error::code::ErrorCode;
+use crate::error::BichonResult;
+use crate::raise_error;
+use crate::store::tantivy::attachment::ATTACHMENT_MANAGER;
+use crate::store::tantivy::envelope::ENVELOPE_MANAGER;
+use crate::store::tantivy::fields::{
+    F_ACCOUNT_ID, F_CONTENT_HASH, F_ID, F_INGEST_AT, F_MAILBOX_ID,
+};
+use crate::store::tantivy::schema::SchemaTools;
+
+// ─── Types ────────────────────────────────────────────────────────────────────
+
+/// A single document candidate for deduplication.
+/// Holds just enough information to compare and delete duplicates.
+struct DedupEntry {
+    /// Unix timestamp (seconds) when this document was ingested.
+    /// Used to determine which copy to keep: we always keep the latest,
+    /// so that a post-uidvalidity-reset uid is preferred over a stale one.
+    ingest_at: i64,
+    /// The email's f_id value, used to delete the duplicate email (via term
+    /// query on f_id) and to cascade-delete attachments whose f_envelope_id
+    /// matches this id.
+    email_id: String,
+}
+
+/// Dedup map for one account.
+/// Key   = (mailbox_id, content_hash)  — stable identity across uidvalidity resets
+/// Value = all documents sharing that key, to be reduced to exactly one.
+type DedupMap = HashMap<(u64, String), Vec<DedupEntry>>;
+
+// ─── Public entry point ───────────────────────────────────────────────────────
+
+/// Background deduplication task.
+///
+/// Iterates over every account found in the index and removes duplicate emails
+/// within each (mailbox_id, content_hash) group, keeping the most recently
+/// ingested copy.
+///
+/// For each duplicate email removed, all attachments in the attachment index
+/// whose f_envelope_id matches the removed email's f_id are also deleted.
+///
+/// Why keep the *latest* ingest_at?
+///   After a uidvalidity reset the server reassigns UIDs. If we kept an old
+///   copy (lower ingest_at) its uid would be stale, and uid-based incremental
+///   sync would re-download emails that are already present.
+///
+/// Processing is done account-by-account so that peak memory is bounded by
+/// the largest single account rather than the entire index.
+pub async fn dedup_task(
+    email_reader: &IndexReader,
+    email_writer: &mut IndexWriter,
+    attachment_writer: &mut IndexWriter,
+) -> BichonResult<()> {
+    let account_ids = collect_account_ids(email_reader)?;
+    let mut total_deleted = 0u64;
+
+    for account_id in account_ids {
+        total_deleted += dedup_account(email_reader, email_writer, attachment_writer, account_id)?;
+    }
+
+    tracing::info!("dedup: finished, total removed={}", total_deleted);
+    Ok(())
+}
+
+// ─── Periodic task ──────────────────────────────────────────────────────────
+
+const DEDUP_INTERVAL: std::time::Duration = std::time::Duration::from_secs(12 * 60 * 60);
+
+/// Periodically scans the email index for duplicate (mailbox_id, content_hash)
+/// entries and removes redundant copies, keeping the most recently ingested one.
+/// Attachments belonging to removed emails are cascade-deleted from the
+/// attachment index.
+pub struct DedupTask;
+
+impl BichonTask for DedupTask {
+    fn start() -> TaskHandle {
+        let periodic_task = PeriodicTask::new("index-dedup");
+
+        let task = move |_: Option<u64>| {
+            Box::pin(async move {
+                // Acquire both writers before creating a reader. The fresh reader
+                // sees the last committed state, while the writers ensure we have
+                // exclusive access to perform deletions.
+                let mut email_writer = ENVELOPE_MANAGER.index_writer().lock().await;
+                let mut attach_writer = ATTACHMENT_MANAGER.index_writer().lock().await;
+                let email_reader = ENVELOPE_MANAGER.create_reader()?;
+
+                dedup_task(&email_reader, &mut email_writer, &mut attach_writer).await?;
+
+                // Commit any remaining changes from the dedup pass.
+                // dedup_account commits per-account, but we ensure a final commit
+                // so the attachment index is in sync.
+                crate::store::tantivy::fatal_commit(&mut attach_writer);
+
+                drop(attach_writer);
+                drop(email_writer);
+                Ok(())
+            })
+        };
+
+        periodic_task.start(task, None, DEDUP_INTERVAL, false, false)
+    }
+}
+
+// ─── Internals ────────────────────────────────────────────────────────────────
+
+/// Collect the distinct set of account_ids present in the index.
+///
+/// Scans only the account_id FAST column — no stored field reads, no I/O
+/// beyond the column file itself.
+fn collect_account_ids(reader: &IndexReader) -> BichonResult<Vec<u64>> {
+    let searcher = reader.searcher();
+    let mut ids = std::collections::HashSet::new();
+
+    for segment_reader in searcher.segment_readers() {
+        let account_col = segment_reader
+            .fast_fields()
+            .u64(F_ACCOUNT_ID)
+            .map_err(|e| raise_error!(format!("{:#?}", e), ErrorCode::InternalError))?;
+        let max_doc = segment_reader.max_doc();
+        for doc_id in 0..max_doc {
+            // Skip documents that have already been soft-deleted
+            if segment_reader.is_deleted(doc_id) {
+                continue;
+            }
+            ids.insert(account_col.values.get_val(doc_id));
+        }
+    }
+    Ok(ids.into_iter().collect())
+}
+
+/// Deduplicate all emails for a single account, and cascade-delete their attachments.
+///
+/// Strategy:
+///   1. Scan FAST columns for (mailbox_id, content_hash, ingest_at, f_id) — no heap reads.
+///   2. Group by (mailbox_id, content_hash).
+///   3. Within each group, sort descending by ingest_at and soft-delete all
+///      but the first (most recent) entry.
+///   4. For each removed email, delete all attachments in the attachment index
+///      whose f_envelope_id matches the removed email's f_id.
+///   5. Commit both writers once per account so memory is released before the
+///      next account is processed.
+///
+/// Peak memory for this function ≈ account_email_count × ~160 bytes
+/// (the extra ~80 bytes over previous version comes from storing email_id strings).
+fn dedup_account(
+    email_reader: &IndexReader,
+    email_writer: &mut IndexWriter,
+    attachment_writer: &mut IndexWriter,
+    account_id: u64,
+) -> BichonResult<u64> {
+    let searcher = email_reader.searcher();
+    let fields = SchemaTools::email_fields();
+    eprintln!(
+        "DEBUG dedup_account: entry account={account_id} f_id_field={:?} f_content_hash_field={:?}",
+        fields.f_id, fields.f_content_hash
+    );
+    let mut map: DedupMap = HashMap::new();
+
+    // ── Phase 1: build the dedup map via FAST column scans ──────────────────
+    for segment_reader in searcher.segment_readers() {
+        let account_col = segment_reader
+            .fast_fields()
+            .u64(F_ACCOUNT_ID)
+            .map_err(|e| raise_error!(format!("{:#?}", e), ErrorCode::InternalError))?;
+        let mailbox_col = segment_reader
+            .fast_fields()
+            .u64(F_MAILBOX_ID)
+            .map_err(|e| raise_error!(format!("{:#?}", e), ErrorCode::InternalError))?;
+        let ingest_col = segment_reader
+            .fast_fields()
+            .i64(F_INGEST_AT)
+            .map_err(|e| raise_error!(format!("{:#?}", e), ErrorCode::InternalError))?;
+        // content_hash and f_id are text fields with FAST; stored as dictionary-encoded strings
+        let hash_col = segment_reader
+            .fast_fields()
+            .str(F_CONTENT_HASH)
+            .map_err(|e| raise_error!(format!("{:#?}", e), ErrorCode::InternalError))?
+            .ok_or_else(|| raise_error!(format!("FAST str column '{}' not found in segment; ensure the field is declared with FAST in the schema", F_CONTENT_HASH), ErrorCode::InternalError))?;
+        let id_col = segment_reader
+            .fast_fields()
+            .str(F_ID)
+            .map_err(|e| raise_error!(format!("{:#?}", e), ErrorCode::InternalError))?
+            .ok_or_else(|| raise_error!(format!("FAST str column '{}' not found in segment; ensure the field is declared with FAST in the schema", F_ID), ErrorCode::InternalError))?;
+
+        let max_doc = segment_reader.max_doc();
+        for doc_id in 0..max_doc {
+            if segment_reader.is_deleted(doc_id) {
+                continue;
+            }
+            // Filter to the current account without touching stored fields
+            if account_col.values.get_val(doc_id) != account_id {
+                continue;
+            }
+
+            let mailbox_id = mailbox_col.values.get_val(doc_id);
+            let ingest_at = ingest_col.values.get_val(doc_id);
+
+            // Read content_hash from the dictionary-encoded string column
+            let hash_ord = hash_col.ords().values_for_doc(doc_id as u32).next().unwrap_or(0);
+            let mut hash_buf = String::new();
+            hash_col
+                .ord_to_str(hash_ord, &mut hash_buf)
+                .map_err(|e| raise_error!(format!("{:#?}", e), ErrorCode::InternalError))?;
+            let content_hash = hash_buf;
+
+            // Read f_id from the dictionary-encoded string column
+            let id_ord = id_col.ords().values_for_doc(doc_id as u32).next().unwrap_or(0);
+            let mut id_buf = String::new();
+            id_col
+                .ord_to_str(id_ord, &mut id_buf)
+                .map_err(|e| raise_error!(format!("{:#?}", e), ErrorCode::InternalError))?;
+            let email_id = id_buf;
+
+            eprintln!(
+                "DEBUG dedup_account: account={account_id} doc_id={doc_id} mailbox={mailbox_id} hash={content_hash:?} id={email_id:?} ingest_at={ingest_at}"
+            );
+
+            map.entry((mailbox_id, content_hash))
+                .or_default()
+                .push(DedupEntry {
+                    ingest_at,
+                    email_id,
+                });
+        }
+    }
+
+    // ── Phase 2: delete duplicate emails and their attachments ───────────────
+    let attachment_fields = SchemaTools::attachment_fields();
+    let mut deleted = 0u64;
+
+    for (_key, mut entries) in map {
+        if entries.len() <= 1 {
+            // No duplicates in this group
+            continue;
+        }
+
+        // Sort descending: the most recently ingested document comes first.
+        // This ensures we keep the copy whose uid reflects the current
+        // uidvalidity, which is required for correct incremental sync.
+        entries.sort_by_key(|e| std::cmp::Reverse(e.ingest_at));
+
+        eprintln!("DEBUG Phase2: key={_key:?} kept={} deleting={}", entries[0].email_id, entries.len() - 1);
+        // Keep entries[0], soft-delete everything else via term query on f_id
+        for entry in &entries[1..] {
+            eprintln!(
+                "DEBUG Phase2: delete_term f_id={:?} text=\"{}\"",
+                fields.f_id, &entry.email_id
+            );
+            // Remove the duplicate email from the email index
+            let email_term = Term::from_field_text(fields.f_id, &entry.email_id);
+            email_writer.delete_term(email_term);
+
+            // Cascade: remove all attachments belonging to this email.
+            // f_envelope_id in the attachment index mirrors f_id in the email index.
+            let envelope_term =
+                Term::from_field_text(attachment_fields.f_envelope_id, &entry.email_id);
+            attachment_writer.delete_term(envelope_term);
+
+            deleted += 1;
+        }
+    }
+
+    // Commit both indexes once per account so the DedupMap memory for this
+    // account can be reclaimed before the next account is processed.
+    if deleted > 0 {
+        email_writer
+            .commit()
+            .map_err(|e| raise_error!(format!("{:#?}", e), ErrorCode::InternalError))?;
+        attachment_writer
+            .commit()
+            .map_err(|e| raise_error!(format!("{:#?}", e), ErrorCode::InternalError))?;
+        tracing::info!("dedup: account={} removed={}", account_id, deleted);
+    }
+
+    Ok(deleted)
+}
+
+// ─── Tests ─────────────────────────────────────────────────────────────────────
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::store::tantivy::fields::AttachmentFields;
+    use crate::store::tantivy::fields::EmailFields;
+    use crate::store::tantivy::schema::SchemaTools;
+    use crate::store::tantivy::tokenizers::EuroTokenizer;
+    use std::collections::HashSet;
+    use std::fmt::Write;
+    use std::fs;
+    use tantivy::Index;
+    use tantivy::TantivyDocument;
+
+    fn temp_dir(prefix: &str) -> std::path::PathBuf {
+        let dir = std::env::temp_dir()
+            .join("bichon-dedup-test")
+            .join(prefix)
+            .join(uuid::Uuid::new_v4().to_string());
+        fs::create_dir_all(&dir).unwrap();
+        dir
+    }
+
+    fn make_index(dir: &std::path::Path, schema: tantivy::schema::Schema) -> Index {
+        let index = Index::create_in_dir(dir, schema).unwrap();
+        index.tokenizers().register("euro", EuroTokenizer::new());
+        index
+    }
+
+    /// Collect non-deleted f_id values from the email index.
+    fn surviving_email_ids(reader: &IndexReader) -> HashSet<String> {
+        reader
+            .reload()
+            .expect("reader reload failed");
+        let searcher = reader.searcher();
+        let mut ids = HashSet::new();
+        let segments = searcher.segment_readers();
+        eprintln!("DEBUG surviving_email_ids: segment_count={}", segments.len());
+        for (seg_idx, seg) in segments.iter().enumerate() {
+            let id_col = seg
+                .fast_fields()
+                .str(F_ID)
+                .unwrap()
+                .expect("FAST str column 'id' missing");
+            let max_doc = seg.max_doc();
+            eprintln!("DEBUG surviving_email_ids: seg={seg_idx} max_doc={max_doc}");
+            for doc_id in 0..max_doc {
+                let is_del = seg.is_deleted(doc_id);
+                let ord = id_col.ords().values_for_doc(doc_id as u32).next().unwrap_or(0);
+                let mut buf = String::new();
+                id_col.ord_to_str(ord, &mut buf).unwrap();
+                eprintln!("DEBUG surviving_email_ids: seg={seg_idx} doc_id={doc_id} is_deleted={is_del} ord={ord} buf={buf:?}");
+                if !is_del {
+                    ids.insert(buf);
+                }
+            }
+        }
+        ids
+    }
+
+    /// Collect non-deleted f_id values from the attachment index.
+    fn surviving_attachment_ids(reader: &IndexReader) -> HashSet<String> {
+        let searcher = reader.searcher();
+        let mut ids = HashSet::new();
+        for seg in searcher.segment_readers() {
+            let id_col = seg
+                .fast_fields()
+                .str(F_ID)
+                .unwrap()
+                .expect("FAST str column 'id' missing in attachment index");
+            let max_doc = seg.max_doc();
+            for doc_id in 0..max_doc {
+                if seg.is_deleted(doc_id) {
+                    continue;
+                }
+                let ord = id_col.ords().values_for_doc(doc_id as u32).next().unwrap_or(0);
+                let mut buf = String::new();
+                id_col.ord_to_str(ord, &mut buf).unwrap();
+                ids.insert(buf);
+            }
+        }
+        ids
+    }
+
+    fn add_email(
+        f: &EmailFields,
+        w: &mut IndexWriter,
+        id: &str,
+        account: u64,
+        mailbox: u64,
+        hash: &str,
+        ingest_at: i64,
+    ) {
+        let mut doc = TantivyDocument::new();
+        doc.add_text(f.f_id, id);
+        doc.add_u64(f.f_account_id, account);
+        doc.add_u64(f.f_mailbox_id, mailbox);
+        doc.add_text(f.f_content_hash, hash);
+        doc.add_i64(f.f_ingest_at, ingest_at);
+        w.add_document(doc).unwrap();
+    }
+
+    fn add_attachment(
+        f: &AttachmentFields,
+        w: &mut IndexWriter,
+        id: &str,
+        envelope_id: &str,
+        account: u64,
+        mailbox: u64,
+    ) {
+        let mut doc = TantivyDocument::new();
+        doc.add_text(f.f_id, id);
+        doc.add_text(f.f_envelope_id, envelope_id);
+        doc.add_u64(f.f_account_id, account);
+        doc.add_u64(f.f_mailbox_id, mailbox);
+        w.add_document(doc).unwrap();
+    }
+
+    /// Prevent segment merges during dedup so delete operations are isolated
+    /// and test assertions target the exact expected document set.
+    fn apply_no_merge_policy(w: &mut IndexWriter) {
+        let mut mp = tantivy::indexer::LogMergePolicy::default();
+        mp.set_min_num_segments(500);
+        mp.set_max_docs_before_merge(1_000_000);
+        w.set_merge_policy(Box::new(mp));
+    }
+
+    struct Harness;
+
+    impl Harness {
+        async fn run<F>(
+            case: &str,
+            populate: F,
+            expected_emails: &[&str],
+            expected_attachments: &[&str],
+        ) where
+            F: FnOnce(&EmailFields, &mut IndexWriter, &AttachmentFields, &mut IndexWriter),
+        {
+            let email_schema = SchemaTools::email_schema();
+            let attach_schema = SchemaTools::attachment_schema();
+            let email_f = SchemaTools::email_fields();
+            let attach_f = SchemaTools::attachment_fields();
+
+            let email_idx = make_index(&temp_dir(case), email_schema);
+            let attach_idx = make_index(&temp_dir(case), attach_schema);
+
+            let mut email_w = email_idx.writer_with_num_threads(1, 50_000_000).unwrap();
+            let mut attach_w = attach_idx.writer_with_num_threads(1, 50_000_000).unwrap();
+            apply_no_merge_policy(&mut email_w);
+            apply_no_merge_policy(&mut attach_w);
+
+            populate(&email_f, &mut email_w, &attach_f, &mut attach_w);
+
+            email_w.commit().unwrap();
+            attach_w.commit().unwrap();
+            drop(email_w);
+            drop(attach_w);
+
+            let mut email_w2 = email_idx.writer_with_num_threads(1, 50_000_000).unwrap();
+            let mut attach_w2 = attach_idx.writer_with_num_threads(1, 50_000_000).unwrap();
+            apply_no_merge_policy(&mut email_w2);
+            apply_no_merge_policy(&mut attach_w2);
+
+            let email_r = email_idx.reader().unwrap();
+            dedup_task(&email_r, &mut email_w2, &mut attach_w2)
+                .await
+                .unwrap();
+
+            let email_r = email_idx.reader().unwrap();
+            let survivors = surviving_email_ids(&email_r);
+            let expected: HashSet<String> =
+                expected_emails.iter().map(|s| s.to_string()).collect();
+            assert_eq!(survivors, expected, "[{case}] email survivors mismatch");
+
+            let attach_r = attach_idx.reader().unwrap();
+            let att_survivors = surviving_attachment_ids(&attach_r);
+            let att_expected: HashSet<String> =
+                expected_attachments.iter().map(|s| s.to_string()).collect();
+            assert_eq!(att_survivors, att_expected, "[{case}] attachment survivors mismatch");
+        }
+    }
+
+    #[tokio::test]
+    async fn dedup_removes_duplicates_and_cascades_to_attachments() {
+        Harness::run(
+            "basic",
+            |ef, ew, af, aw| {
+                add_email(ef, ew, "dup-old", 1, 200, "hash-dup", 1000);
+                add_email(ef, ew, "dup-new", 1, 200, "hash-dup", 3000);
+                add_email(ef, ew, "unique", 1, 200, "hash-uniq", 1000);
+                add_attachment(af, aw, "att-old", "dup-old", 1, 200);
+                add_attachment(af, aw, "att-new", "dup-new", 1, 200);
+            },
+            &["dup-new", "unique"],
+            &["att-new"],
+        )
+        .await;
+    }
+
+    #[tokio::test]
+    async fn dedup_keeps_latest_among_many_duplicates() {
+        Harness::run(
+            "many-dups",
+            |ef, ew, af, aw| {
+                for (i, ts) in [50, 100, 400, 200, 300].iter().enumerate() {
+                    let id = format!("dup-{i}");
+                    add_email(ef, ew, &id, 1, 1, "H", *ts);
+                    add_attachment(af, aw, &format!("att-{i}"), &id, 1, 1);
+                }
+            },
+            &["dup-2"],      // ingest_at=400, the latest
+            &["att-2"],
+        )
+        .await;
+    }
+
+    #[tokio::test]
+    async fn dedup_no_duplicates_is_noop() {
+        Harness::run(
+            "no-dups",
+            |ef, ew, af, aw| {
+                add_email(ef, ew, "a", 1, 1, "hash-a", 100);
+                add_email(ef, ew, "b", 1, 1, "hash-b", 200);
+                add_email(ef, ew, "c", 1, 1, "hash-c", 300);
+                add_attachment(af, aw, "att-a", "a", 1, 1);
+                add_attachment(af, aw, "att-b", "b", 1, 1);
+                add_attachment(af, aw, "att-c", "c", 1, 1);
+            },
+            &["a", "b", "c"],
+            &["att-a", "att-b", "att-c"],
+        )
+        .await;
+    }
+
+    #[tokio::test]
+    async fn dedup_isolates_accounts() {
+        // Same hash, same mailbox, DIFFERENT accounts → no dedup
+        Harness::run(
+            "cross-account",
+            |ef, ew, af, aw| {
+                add_email(ef, ew, "acc1-a", 1, 1, "hash-same", 100);
+                add_email(ef, ew, "acc1-b", 1, 1, "hash-same", 200);
+                add_email(ef, ew, "acc2-a", 2, 1, "hash-same", 100);
+                add_email(ef, ew, "acc2-b", 2, 1, "hash-same", 200);
+                add_attachment(af, aw, "att-1a", "acc1-a", 1, 1);
+                add_attachment(af, aw, "att-1b", "acc1-b", 1, 1);
+                add_attachment(af, aw, "att-2a", "acc2-a", 2, 1);
+                add_attachment(af, aw, "att-2b", "acc2-b", 2, 1);
+            },
+            // Each account keeps its latest: acc1 keeps acc1-b (200>100), acc2 keeps acc2-b
+            &["acc1-b", "acc2-b"],
+            &["att-1b", "att-2b"],
+        )
+        .await;
+    }
+
+    #[tokio::test]
+    async fn dedup_isolates_mailboxes() {
+        // Same hash, same account, DIFFERENT mailboxes → no dedup
+        Harness::run(
+            "cross-mailbox",
+            |ef, ew, af, aw| {
+                add_email(ef, ew, "mb1-a", 1, 1, "hash-same", 100);
+                add_email(ef, ew, "mb2-a", 1, 2, "hash-same", 100);
+                add_email(ef, ew, "mb1-b", 1, 1, "hash-same", 200);
+                add_email(ef, ew, "mb2-b", 1, 2, "hash-same", 200);
+                add_attachment(af, aw, "att-1a", "mb1-a", 1, 1);
+                add_attachment(af, aw, "att-1b", "mb1-b", 1, 1);
+                add_attachment(af, aw, "att-2a", "mb2-a", 1, 2);
+                add_attachment(af, aw, "att-2b", "mb2-b", 1, 2);
+            },
+            &["mb1-b", "mb2-b"],
+            &["att-1b", "att-2b"],
+        )
+        .await;
+    }
+
+    #[tokio::test]
+    async fn dedup_multiple_attachments_per_email() {
+        // Deleting an email cascades all its attachments, not just one
+        Harness::run(
+            "multi-att",
+            |ef, ew, af, aw| {
+                add_email(ef, ew, "old", 1, 1, "H", 100);
+                add_email(ef, ew, "new", 1, 1, "H", 200);
+                // The old email has 3 attachments — all should be removed
+                add_attachment(af, aw, "att1", "old", 1, 1);
+                add_attachment(af, aw, "att2", "old", 1, 1);
+                add_attachment(af, aw, "att3", "old", 1, 1);
+                // The kept email has 2 attachments — both should survive
+                add_attachment(af, aw, "att4", "new", 1, 1);
+                add_attachment(af, aw, "att5", "new", 1, 1);
+            },
+            &["new"],
+            &["att4", "att5"],
+        )
+        .await;
+    }
+
+    /// Inspects the production email index and reports duplicate counts.
+    ///
+    /// A "duplicate" is defined as two or more emails sharing the same
+    /// (account_id, mailbox_id, content_hash) tuple.
+    ///
+    /// This test is read-only — it does not modify the index.
+    #[test]
+    fn inspect_production_duplicates() {
+        let index_path = r"E:\db\data\bichon-indices\mail_metadata";
+        let report_path = std::path::PathBuf::from(r"E:\bichon\dedup_report.txt");
+
+        let mut report = String::new();
+        let _ = writeln!(report, "opening index at {index_path}...");
+
+        let index = match Index::open_in_dir(index_path) {
+            Ok(idx) => {
+                let _ = writeln!(report, "index opened successfully");
+                idx
+            }
+            Err(e) => {
+                let _ = writeln!(report, "Failed to open index at {index_path}: {e}");
+                let _ = std::fs::write(&report_path, &report);
+                return;
+            }
+        };
+
+        let reader = match index.reader() {
+            Ok(r) => r,
+            Err(e) => {
+                let _ = writeln!(report, "Failed to create reader: {e}");
+                let _ = std::fs::write(&report_path, &report);
+                return;
+            }
+        };
+
+        reader.reload().expect("reader reload failed");
+        let searcher = reader.searcher();
+
+        let mut total_docs = 0u64;
+        let mut groups: std::collections::HashMap<u64, std::collections::HashMap<(u64, String), u64>> =
+            std::collections::HashMap::new();
+
+        for segment_reader in searcher.segment_readers() {
+            let account_col = segment_reader
+                .fast_fields()
+                .u64(F_ACCOUNT_ID)
+                .unwrap();
+            let mailbox_col = segment_reader
+                .fast_fields()
+                .u64(F_MAILBOX_ID)
+                .unwrap();
+            let hash_col = match segment_reader
+                .fast_fields()
+                .str(F_CONTENT_HASH)
+                .unwrap()
+            {
+                Some(c) => c,
+                None => {
+                    let _ = writeln!(report, "Segment has no FAST str column for content_hash, skipping");
+                    continue;
+                }
+            };
+
+            let max_doc = segment_reader.max_doc();
+            for doc_id in 0..max_doc {
+                if segment_reader.is_deleted(doc_id) {
+                    continue;
+                }
+
+                let account_id = account_col.values.get_val(doc_id);
+                let mailbox_id = mailbox_col.values.get_val(doc_id);
+
+                let hash_ord = hash_col.ords().values_for_doc(doc_id as u32).next().unwrap_or(0);
+                let mut hash_buf = String::new();
+                hash_col.ord_to_str(hash_ord, &mut hash_buf).unwrap();
+                let content_hash = hash_buf;
+
+                total_docs += 1;
+                groups
+                    .entry(account_id)
+                    .or_default()
+                    .entry((mailbox_id, content_hash))
+                    .and_modify(|c| *c += 1)
+                    .or_insert(1);
+            }
+        }
+
+        // ── Summarize ──────────────────────────────────────────────────────────
+        let mut total_duplicate_groups = 0u64;
+        let mut total_duplicate_emails = 0u64;
+
+        for (account_id, account_groups) in &groups {
+            let mut account_dup_groups = 0u64;
+            let mut account_dup_emails = 0u64;
+            for ((_mailbox_id, _hash), count) in account_groups {
+                if *count > 1 {
+                    account_dup_groups += 1;
+                    account_dup_emails += count - 1;
+                }
+            }
+            if account_dup_groups > 0 {
+                let _ = writeln!(
+                    report,
+                    "account={account_id}: {account_dup_groups} duplicate groups, {account_dup_emails} redundant emails"
+                );
+            }
+            total_duplicate_groups += account_dup_groups;
+            total_duplicate_emails += account_dup_emails;
+        }
+
+        let _ = writeln!(
+            report,
+            "─── Summary ───\n\
+             total_docs          = {total_docs}\n\
+             accounts            = {}\n\
+             duplicate_groups    = {total_duplicate_groups}\n\
+             redundant_emails    = {total_duplicate_emails}\n\
+             unique_after_dedup  = {}",
+            groups.len(),
+            total_docs - total_duplicate_emails,
+        );
+
+        std::fs::write(&report_path, &report).unwrap();
+        println!("report written to {}", report_path.display());
+    }
+}
diff --git a/crates/core/src/store/tantivy/envelope.rs b/crates/core/src/store/tantivy/envelope.rs
index 2705da7..90d9ca9 100644
--- a/crates/core/src/store/tantivy/envelope.rs
+++ b/crates/core/src/store/tantivy/envelope.rs
@@ -86,6 +86,16 @@ pub struct IndexManager {
 }
 
 impl IndexManager {
+    pub(crate) fn index_writer(&self) -> &Arc<Mutex<IndexWriter>> {
+        &self.index_writer
+    }
+
+    pub(crate) fn create_reader(&self) -> BichonResult<IndexReader> {
+        self.index
+            .reader()
+            .map_err(|e| raise_error!(format!("{:#?}", e), ErrorCode::InternalError))
+    }
+
     pub async fn shutdown(&self) {
         let mut guard = self.handle.lock().await;
         if let Some(handle) = guard.take() {
diff --git a/crates/core/src/store/tantivy/mod.rs b/crates/core/src/store/tantivy/mod.rs
index 32a7814..0e1460f 100644
--- a/crates/core/src/store/tantivy/mod.rs
+++ b/crates/core/src/store/tantivy/mod.rs
@@ -24,6 +24,7 @@ use crate::{
 };
 
 pub mod attachment;
+pub mod dedup;
 pub mod envelope;
 pub mod fields;
 pub mod filter;
diff --git a/crates/core/src/tasks/mod.rs b/crates/core/src/tasks/mod.rs
index 2dd3f10..0f86b59 100644
--- a/crates/core/src/tasks/mod.rs
+++ b/crates/core/src/tasks/mod.rs
@@ -19,6 +19,8 @@
 use crate::common::periodic::TaskHandle;
 use crate::context::BichonTask;
 use crate::oauth2::{refresh::OAuth2RefreshTask, task::OAuth2CleanTask};
+use crate::store::tantivy::dedup::DedupTask;
+
 pub struct PeriodicTasks {
     tasks: Vec<TaskHandle>,
 }
@@ -28,6 +30,7 @@ impl PeriodicTasks {
         let mut tasks = Vec::new();
         tasks.push(OAuth2CleanTask::start());
         tasks.push(OAuth2RefreshTask::start());
+        tasks.push(DedupTask::start());
         Self { tasks }
     }