diff --git a/.agent-tasks/2026-03-20--8--flush-tracker-stale-consumer/discarded-threads.md b/.agent-tasks/2026-03-20--8--flush-tracker-stale-consumer/discarded-threads.md
new file mode 100644
index 0000000000..f468f9b495
--- /dev/null
+++ b/.agent-tasks/2026-03-20--8--flush-tracker-stale-consumer/discarded-threads.md
@@ -0,0 +1,3 @@
+# Discarded threads
+
+(none yet)
diff --git a/.agent-tasks/2026-03-20--8--flush-tracker-stale-consumer/open-questions.md b/.agent-tasks/2026-03-20--8--flush-tracker-stale-consumer/open-questions.md
new file mode 100644
index 0000000000..266f6451f8
--- /dev/null
+++ b/.agent-tasks/2026-03-20--8--flush-tracker-stale-consumer/open-questions.md
@@ -0,0 +1,3 @@
+# Open questions
+
+(none yet)
diff --git a/.agent-tasks/2026-03-20--8--flush-tracker-stale-consumer/progress.md b/.agent-tasks/2026-03-20--8--flush-tracker-stale-consumer/progress.md
new file mode 100644
index 0000000000..c002ab7ee2
--- /dev/null
+++ b/.agent-tasks/2026-03-20--8--flush-tracker-stale-consumer/progress.md
@@ -0,0 +1,31 @@
+# Progress log
+
+## 2026-03-20
+
+### Initial analysis
+- Read FlushTracker, ShapeLogCollector, Consumer, ConsumerRegistry, ShapeCleaner source
+- Read existing test files for both FlushTracker and ShapeLogCollector
+- Identified the exact code path: Consumer.handle_materializer_down → {:stop, :shutdown, state} → terminate → handle_writer_termination clause 3 (:ok, no cleanup)
+- Existing tests cover: consumer crash during broadcast (detected by ConsumerRegistry), multi-fragment crash between fragments
+- Missing: consumer dying out-of-band AFTER successful broadcast delivery, with no subsequent transactions to that shape
+
+### Implementation
+- Wrote 3 FlushTracker unit tests showing stale entries blocking advancement indefinitely
+- Wrote 3 SLC integration tests:
+  - Main bug test: kill consumer with :kill (skips terminate/remove_shape), send txns only to other shapes, verify flush stuck
+  - Contrast test: graceful termination (runs terminate → remove_shape) allows flush to advance
+  - Recovery test: when a txn finally touches the dead shape's table, undeliverable detection cleans up
+- Key design decisions:
+  - Used two separate tables (table_a, table_b) so transactions can selectively target one shape
+  - Used :kill to simulate the end state of handle_materializer_down path (dead process, no cleanup)
+  - The @describetag is needed (not @tag) to propagate inspector config to setup functions
+
+### Operational issues
+- Initially edited files in ~/code/electric-sql/electric instead of ~/agents/github/erik/repos/electric
+  - Had to create a patch and git apply it to the correct repo
+- Used @tag instead of @describetag, which doesn't propagate to setup functions
+- First version of tests used `refute_receive` for the initial flush notification, but FlushTracker does partially advance when one shape catches up (to one below the stuck shape's offset). Fixed to `assert_receive` the partial advance, then `refute_receive` further advances.
+
+### PR
+- Created PR #4035: https://github.com/electric-sql/electric/pull/4035
+- Added "claude" label for automated review
diff --git a/.agent-tasks/2026-03-20--8--flush-tracker-stale-consumer/prompt.md b/.agent-tasks/2026-03-20--8--flush-tracker-stale-consumer/prompt.md
new file mode 100644
index 0000000000..6b86366c08
--- /dev/null
+++ b/.agent-tasks/2026-03-20--8--flush-tracker-stale-consumer/prompt.md
@@ -0,0 +1,11 @@
+# Task prompt
+
+Issue #8 from electric-sql/alco-agent-tasks: "FlushTracker stalling when tracked consumer dies out-of-band"
+
+Write unit tests that exercise the edge case where a consumer process is tracked by FlushTracker but then dies independently (via handle_materializer_down with :shutdown reason), leaving a stale shape entry that permanently blocks flush advancement.
+
+Key constraints:
+- Use different approaches for each test
+- Avoid mocking too many components
+- Avoid inventing calls or messages inside the test body
+- Try setting conditions so the app hits the edge case naturally
diff --git a/.agent-tasks/2026-03-20--8--flush-tracker-stale-consumer/task.md b/.agent-tasks/2026-03-20--8--flush-tracker-stale-consumer/task.md
new file mode 100644
index 0000000000..fe311e8a0f
--- /dev/null
+++ b/.agent-tasks/2026-03-20--8--flush-tracker-stale-consumer/task.md
@@ -0,0 +1,20 @@
+# Task: FlushTracker stale consumer edge case tests
+
+## Problem
+
+When a consumer dies via `handle_materializer_down` with `:shutdown` reason:
+1. Consumer stops with `{:stop, :shutdown, state}` (skipping `stop_and_clean`)
+2. `terminate/2` calls `handle_writer_termination` clause 3 which returns `:ok` without cleanup
+3. ConsumerRegistry ETS entry is NOT removed (unregister_name is a no-op)
+4. ShapeLogCollector is NOT notified to remove the shape from FlushTracker
+5. If no future transactions affect that shape, the stale FlushTracker entry persists indefinitely
+6. This blocks `last_global_flushed_offset` advancement → unbounded WAL growth
+
+This only affects `allow_subqueries` stacks since `handle_materializer_down` requires materializers.
+
+## Goal
+
+Write tests demonstrating this edge case at different levels:
+1. FlushTracker unit level: a shape tracked but never flushed/removed blocks advancement
+2. ShapeLogCollector integration level: a consumer that dies out-of-band after receiving a transaction leaves FlushTracker stuck
+3. Full integration: demonstrate the materializer-death path that triggers handle_materializer_down → consumer death → stale FlushTracker
diff --git a/packages/sync-service/test/electric/replication/shape_log_collector/flush_tracker_test.exs b/packages/sync-service/test/electric/replication/shape_log_collector/flush_tracker_test.exs
index 6fabe24447..af6aa1df51 100644
--- a/packages/sync-service/test/electric/replication/shape_log_collector/flush_tracker_test.exs
+++ b/packages/sync-service/test/electric/replication/shape_log_collector/flush_tracker_test.exs
@@ -457,6 +457,83 @@ defmodule Electric.Replication.ShapeLogCollector.FlushTrackerTest do
     end
   end
 
+  describe "out-of-band consumer death leaves stale FlushTracker entry" do
+    test "shape tracked across many transactions never advances global offset", %{
+      tracker: tracker
+    } do
+      # Simulate: two shapes receive txn at lsn 10. One shape's consumer dies
+      # out-of-band (no handle_shape_removed ever called). Subsequent txns only
+      # affect the alive shape. The dead shape's entry pins the global offset.
+      tracker = handle_txn(tracker, batch(lsn: 10, last_offset: 10), ["alive", "dead"])
+
+      # alive flushes lsn 10
+      tracker = FlushTracker.handle_flush_notification(tracker, "alive", LogOffset.new(10, 10))
+      # Global offset is pinned one below dead's initial position (tx_offset=9)
+      assert_receive {:flush_confirmed, 8}
+
+      # Many subsequent transactions only affect the alive shape
+      tracker =
+        Enum.reduce(11..20, tracker, fn lsn, tracker ->
+          tracker
+          |> handle_txn(batch(lsn: lsn, last_offset: 10), ["alive"])
+          |> FlushTracker.handle_flush_notification("alive", LogOffset.new(lsn, 10))
+        end)
+
+      # Despite 10 more transactions fully flushed by the alive shape,
+      # global offset has not moved past the dead shape's stuck position
+      refute_receive {:flush_confirmed, 10}
+      refute FlushTracker.empty?(tracker)
+    end
+
+    test "removing the stale shape after prolonged stall unblocks to latest seen offset", %{
+      tracker: tracker
+    } do
+      # Same setup: dead shape pins the offset
+      tracker = handle_txn(tracker, batch(lsn: 10, last_offset: 10), ["alive", "dead"])
+
+      tracker = FlushTracker.handle_flush_notification(tracker, "alive", LogOffset.new(10, 10))
+      assert_receive {:flush_confirmed, 8}
+
+      # More txns flow through alive shape only
+      tracker =
+        Enum.reduce(11..15, tracker, fn lsn, tracker ->
+          tracker
+          |> handle_txn(batch(lsn: lsn, last_offset: 10), ["alive"])
+          |> FlushTracker.handle_flush_notification("alive", LogOffset.new(lsn, 10))
+        end)
+
+      # Now simulate detection of the dead consumer (e.g. a sweep or monitor)
+      tracker = FlushTracker.handle_shape_removed(tracker, "dead")
+
+      # Global offset jumps all the way to the latest seen offset (lsn 15)
+      assert_receive {:flush_confirmed, 15}
+      assert FlushTracker.empty?(tracker)
+    end
+
+    test "stale shape blocks advancement even when its tracked offset is much older than current",
+         %{tracker: tracker} do
+      # Dead shape gets tracked at a very early offset
+      tracker = handle_txn(tracker, batch(lsn: 1, last_offset: 10), ["dead"])
+
+      # Alive shape joins later and processes many transactions
+      tracker =
+        Enum.reduce(2..50, tracker, fn lsn, tracker ->
+          tracker
+          |> handle_txn(batch(lsn: lsn, last_offset: 10), ["alive"])
+          |> FlushTracker.handle_flush_notification("alive", LogOffset.new(lsn, 10))
+        end)
+
+      # The global offset is still stuck at lsn 1's prev_log_offset position
+      # because the dead shape was never removed
+      refute_receive {:flush_confirmed, 2}
+      refute FlushTracker.empty?(tracker)
+
+      # Clean it up
+      _tracker = FlushTracker.handle_shape_removed(tracker, "dead")
+      assert_receive {:flush_confirmed, 50}
+    end
+  end
+
   # Helper: calls handle_txn_fragment with shapes_with_changes defaulting to
   # all affected shapes (the common case for single-fragment transactions).
   defp handle_txn(tracker, fragment, affected_shapes) do
diff --git a/packages/sync-service/test/electric/replication/shape_log_collector_test.exs b/packages/sync-service/test/electric/replication/shape_log_collector_test.exs
index a90f4b57c5..54492dfd9f 100644
--- a/packages/sync-service/test/electric/replication/shape_log_collector_test.exs
+++ b/packages/sync-service/test/electric/replication/shape_log_collector_test.exs
@@ -1280,6 +1280,235 @@ defmodule Electric.Replication.ShapeLogCollectorTest do
     end
   end
 
+  describe "FlushTracker stalling when tracked consumer dies out-of-band" do
+    @two_table_inspector Support.StubInspector.new(%{
+                           {1234, {"public", "table_a"}} => [
+                             %{name: "id", type: "int8", pk_position: 0}
+                           ],
+                           {5678, {"public", "table_b"}} => [
+                             %{name: "id", type: "int8", pk_position: 0}
+                           ]
+                         })
+
+    @shape_a Shape.new!("table_a", inspector: @two_table_inspector)
+    @shape_b Shape.new!("table_b", inspector: @two_table_inspector)
+
+    @describetag inspector: @two_table_inspector
+    setup :setup_log_collector
+
+    setup ctx do
+      parent = self()
+
+      consumers = [
+        {:alive,
+         start_supervised!(
+           {Support.TransactionConsumer,
+            id: :alive,
+            stack_id: ctx.stack_id,
+            parent: parent,
+            shape: @shape_a,
+            shape_handle: "shape-alive"},
+           id: {:consumer, :alive}
+         )},
+        {:doomed,
+         start_supervised!(
+           {Support.TransactionConsumer,
+            id: :doomed,
+            stack_id: ctx.stack_id,
+            parent: parent,
+            shape: @shape_b,
+            shape_handle: "shape-doomed"},
+           id: {:consumer, :doomed}
+         )}
+      ]
+
+      %{consumers: consumers}
+    end
+
+    # This test documents a known bug (electric-sql/electric#3980):
+    # when a consumer dies without calling remove_shape (e.g. via handle_materializer_down
+    # with :shutdown reason), and no subsequent transactions touch that shape's table,
+    # FlushTracker stays stuck indefinitely. When a fix is implemented (e.g. PID monitoring
+    # or periodic liveness sweep), this test should be updated to assert the fix instead.
+    test "consumer killed out-of-band after delivery permanently blocks flush advancement", ctx do
+      register_as_replication_client(ctx.stack_id)
+
+      lsn = Lsn.from_integer(10)
+
+      # Transaction that inserts into both tables → both shapes get tracked in FlushTracker
+      txn =
+        transaction(100, lsn, [
+          %Changes.NewRecord{
+            relation: {"public", "table_a"},
+            record: %{"id" => "1"},
+            log_offset: LogOffset.new(lsn, 0)
+          },
+          %Changes.NewRecord{
+            relation: {"public", "table_b"},
+            record: %{"id" => "1"},
+            log_offset: LogOffset.new(lsn, 1)
+          }
+        ])
+
+      assert :ok = ShapeLogCollector.handle_event(txn, ctx.stack_id)
+
+      # Both consumers receive the transaction
+      assert_receive {Support.TransactionConsumer, {:alive, _}, [_]}
+      assert_receive {Support.TransactionConsumer, {:doomed, _}, [_]}
+
+      # Kill the doomed consumer with :kill to prevent terminate from running.
+      # This simulates a consumer dying via handle_materializer_down → {:stop, :shutdown}
+      # where handle_writer_termination clause 3 returns :ok without cleanup.
+      # Using :kill here achieves the same end state: dead process, no remove_shape call.
+      {_, doomed_pid} = List.keyfind!(ctx.consumers, :doomed, 0)
+      kill_consumer(doomed_pid, :kill)
+
+      # The alive consumer flushes its data
+      ShapeLogCollector.notify_flushed(ctx.stack_id, "shape-alive", LogOffset.new(lsn, 1))
+
+      # Flush boundary partially advances: the alive shape caught up, but the dead
+      # shape pins the global offset at its prev_log_offset (tx_offset=9) minus 1 = 8.
+      assert_receive {:flush_boundary_updated, 8}, 100
+
+      # Send more transactions that ONLY affect table_a (the alive shape).
+      # Since no transaction touches table_b, SLC never discovers the dead consumer.
+      for i <- 11..15 do
+        lsn_i = Lsn.from_integer(i)
+
+        txn_i =
+          transaction(100 + i, lsn_i, [
+            %Changes.NewRecord{
+              relation: {"public", "table_a"},
+              record: %{"id" => "#{i}"},
+              log_offset: LogOffset.new(lsn_i, 0)
+            }
+          ])
+
+        assert :ok = ShapeLogCollector.handle_event(txn_i, ctx.stack_id)
+        assert_receive {Support.TransactionConsumer, {:alive, _}, [_]}
+        ShapeLogCollector.notify_flushed(ctx.stack_id, "shape-alive", LogOffset.new(lsn_i, 0))
+      end
+
+      # Despite 5 more fully-flushed transactions through lsn 15, the flush boundary
+      # is permanently stuck at 8. It never reaches the current lsn (15).
+      # This is the bug: the dead consumer's stale FlushTracker entry blocks WAL flush.
+      refute_receive {:flush_boundary_updated, _}, 100
+    end
+
+    # NOTE: This test uses Support.TransactionConsumer whose terminate/2 always
+    # calls ShapeLogCollector.remove_shape. The real Consumer does NOT do this
+    # when dying with :shutdown (handle_writer_termination clause 3 returns :ok).
+    # This test demonstrates that the cleanup mechanism works when invoked,
+    # serving as a contrast to the "killed out-of-band" test above where it isn't.
+    test "consumer that calls remove_shape on termination allows flush to advance", ctx do
+      register_as_replication_client(ctx.stack_id)
+
+      lsn = Lsn.from_integer(10)
+
+      # Transaction affecting both tables
+      txn =
+        transaction(100, lsn, [
+          %Changes.NewRecord{
+            relation: {"public", "table_a"},
+            record: %{"id" => "1"},
+            log_offset: LogOffset.new(lsn, 0)
+          },
+          %Changes.NewRecord{
+            relation: {"public", "table_b"},
+            record: %{"id" => "1"},
+            log_offset: LogOffset.new(lsn, 1)
+          }
+        ])
+
+      assert :ok = ShapeLogCollector.handle_event(txn, ctx.stack_id)
+      assert_receive {Support.TransactionConsumer, {:alive, _}, [_]}
+      assert_receive {Support.TransactionConsumer, {:doomed, _}, [_]}
+
+      # Alive consumer flushes — partial advance to 8 (doomed still pending)
+      ShapeLogCollector.notify_flushed(ctx.stack_id, "shape-alive", LogOffset.new(lsn, 1))
+      assert_receive {:flush_boundary_updated, 8}, 100
+
+      # Stop the doomed consumer gracefully — terminate runs, calls remove_shape,
+      # which triggers FlushTracker.handle_shape_removed. This is the CORRECT path.
+      {_, doomed_pid} = List.keyfind!(ctx.consumers, :doomed, 0)
+      Process.unlink(doomed_pid)
+      stop_supervised!({:consumer, :doomed})
+
+      # Now the flush boundary advances all the way to lsn because the stale shape
+      # was cleaned up by the graceful termination path
+      expected_lsn = Lsn.to_integer(lsn)
+      assert_receive {:flush_boundary_updated, ^expected_lsn}, 200
+    end
+
+    test "transaction to dead shape's table eventually unblocks flush via undeliverable detection",
+         ctx do
+      register_as_replication_client(ctx.stack_id)
+
+      lsn = Lsn.from_integer(10)
+
+      txn =
+        transaction(100, lsn, [
+          %Changes.NewRecord{
+            relation: {"public", "table_a"},
+            record: %{"id" => "1"},
+            log_offset: LogOffset.new(lsn, 0)
+          },
+          %Changes.NewRecord{
+            relation: {"public", "table_b"},
+            record: %{"id" => "1"},
+            log_offset: LogOffset.new(lsn, 1)
+          }
+        ])
+
+      assert :ok = ShapeLogCollector.handle_event(txn, ctx.stack_id)
+      assert_receive {Support.TransactionConsumer, {:alive, _}, [_]}
+      assert_receive {Support.TransactionConsumer, {:doomed, _}, [_]}
+
+      # Kill the doomed consumer out-of-band
+      {_, doomed_pid} = List.keyfind!(ctx.consumers, :doomed, 0)
+      kill_consumer(doomed_pid, :kill)
+
+      # Alive flushes — partial advance to 8
+      ShapeLogCollector.notify_flushed(ctx.stack_id, "shape-alive", LogOffset.new(lsn, 1))
+      assert_receive {:flush_boundary_updated, 8}, 100
+
+      # Now a NEW transaction comes that affects BOTH tables.
+      # ConsumerRegistry.broadcast detects the dead doomed consumer → undeliverable.
+      # SLC calls FlushTracker.handle_shape_removed for undeliverable shapes.
+      lsn2 = Lsn.from_integer(20)
+
+      txn2 =
+        transaction(200, lsn2, [
+          %Changes.NewRecord{
+            relation: {"public", "table_a"},
+            record: %{"id" => "2"},
+            log_offset: LogOffset.new(lsn2, 0)
+          },
+          %Changes.NewRecord{
+            relation: {"public", "table_b"},
+            record: %{"id" => "2"},
+            log_offset: LogOffset.new(lsn2, 1)
+          }
+        ])
+
+      log =
+        ExUnit.CaptureLog.capture_log(fn ->
+          assert :ok = ShapeLogCollector.handle_event(txn2, ctx.stack_id)
+        end)
+
+      assert log =~ "Consumer processes crashed or missing during broadcast"
+
+      assert_receive {Support.TransactionConsumer, {:alive, _}, [_]}
+
+      # The stale entry was removed by the undeliverable detection in publish/1.
+      # The alive consumer flushes the second transaction.
+      ShapeLogCollector.notify_flushed(ctx.stack_id, "shape-alive", LogOffset.new(lsn2, 1))
+
+      expected_lsn = Lsn.to_integer(lsn2)
+      assert_receive {:flush_boundary_updated, ^expected_lsn}, 100
+    end
+  end
+
   defp transaction(xid, lsn, changes) do
     last_log_offset =
       case Enum.reverse(changes) do