@@ -11,7 +11,11 @@ import {
   TRANSFER_MEASURED_MCP_RESULT_BYTES,
   TRANSFER_MEASURED_TOOLS,
 } from "./fixtures/transferBudget.ts";
-import { LONG_TURN_WORK_ROWS } from "./TransferBudgetV2Fixture.integration.ts";
+import {
+  LONG_TURN_WORK_ROWS,
+  SHELL_IDLE_SUBAGENTS,
+  SHELL_SETTLED_THREADS,
+} from "./TransferBudgetV2Fixture.integration.ts";
 
 export interface WebSocketCatchUpMeasurement extends WebSocketTransferTotals {
   readonly mode: "replay" | "snapshot";
@@ -36,6 +40,12 @@ export interface TransferBudgetRun {
   /** The compact bounded snapshot web and mobile load when the long-turn thread opens. */
   readonly longTurnColdOpen: TimelineHttpMeasurement;
   readonly longTurnHistoryPage: TimelineHttpMeasurement;
+  /** The same reads as an older client that only sends `compactTurnItems`, to show the savings. */
+  readonly longTurnColdOpenLegacy: TimelineHttpMeasurement;
+  readonly longTurnHistoryPageLegacy: TimelineHttpMeasurement;
+  /** The cold-connect shell snapshot after the settled-thread fleet is written. */
+  readonly fleetShell: TimelineHttpMeasurement;
+  readonly fleetShellLegacy: TimelineHttpMeasurement;
 }
 
 interface ProviderTransferBudget {
@@ -46,11 +56,12 @@ interface ProviderTransferBudget {
   readonly measuredTurnWebSocketMessages: number;
 }
 
+// The measured turn decodes to about 22 KB; without live body dedupe its repeated answer makes it 26 KB.
 const TRANSFER_BUDGET = {
   totalWireBytes: 7_000,
   threadSnapshotWireBytes: 5_000,
   measuredTurnWebSocketWireBytes: 2_000,
-  measuredTurnWebSocketDecodedBytes: 30_000,
+  measuredTurnWebSocketDecodedBytes: 25_000,
   measuredTurnWebSocketMessages: 8,
 } satisfies ProviderTransferBudget;
 
@@ -59,15 +70,20 @@ const TRANSFER_BUDGETS: Readonly<Record<string, ProviderTransferBudget>> = {
   claudeAgent: TRANSFER_BUDGET,
 };
 
-// Decoded bytes, not wire bytes: phones parse and keep the decoded JSON. About
-// 30% above the observed 92 KB; without row budgets inside user turns the cold
-// open carried all 920 rows (1.07 MB). Kept apart from TRANSFER_BUDGET because
-// the PR comment validates that object's keys.
+// Decoded bytes, not wire bytes: phones parse and keep the decoded JSON. About 7% above the
+// observed 103 KB, so a cold open without body dedupe (117 KB), the ID table, or row budgets
+// (all 920 rows) fails. Kept apart from TRANSFER_BUDGET because the PR comment validates its keys.
 const LONG_TURN_BUDGET = {
-  decodedBytes: 120_000,
+  decodedBytes: 110_000,
   rows: 75,
 };
 
+// About 30% above the observed 42 KB. Without the shell opt-ins it carries all 142 rows (217 KB).
+const FLEET_SHELL_BUDGET = {
+  decodedBytes: 55_000,
+  rows: 27,
+};
+
 function totalWireBytes(run: TransferBudgetRun): number {
   return run.threadSnapshot.wireBytes + run.measuredTurnWebSocket.wireBytes;
 }
@@ -160,21 +176,28 @@ function webSocketRows(
   ];
 }
 
-function longTurnRows(
+function timelineRows(
   provider: ProviderDriverKind,
   phase: string,
   measured: TimelineHttpMeasurement,
+  budget: { readonly decodedBytes: number; readonly rows: number },
 ): string[] {
   return [
     infoRow(provider, phase, "HTTP wire", measured.http.wireBytes),
-    row(
-      provider,
-      phase,
-      "HTTP decoded",
-      measured.http.decodedBodyBytes,
-      LONG_TURN_BUDGET.decodedBytes,
-    ),
-    row(provider, phase, "rows", measured.rows, LONG_TURN_BUDGET.rows, String),
+    row(provider, phase, "HTTP decoded", measured.http.decodedBodyBytes, budget.decodedBytes),
+    row(provider, phase, "rows", measured.rows, budget.rows, String),
+  ];
+}
+
+function legacyRows(
+  provider: ProviderDriverKind,
+  phase: string,
+  measured: TimelineHttpMeasurement,
+): string[] {
+  return [
+    infoRow(provider, `${phase}, older client`, "HTTP wire", measured.http.wireBytes),
+    infoRow(provider, `${phase}, older client`, "HTTP decoded", measured.http.decodedBodyBytes),
+    infoRow(provider, `${phase}, older client`, "rows", measured.rows, String),
   ];
 }
 
@@ -216,6 +239,12 @@ export function transferBudgetViolations(runs: ReadonlyArray<TransferBudgetRun>)
         LONG_TURN_BUDGET.decodedBytes,
       ],
       ["long-turn history page rows", run.longTurnHistoryPage.rows, LONG_TURN_BUDGET.rows],
+      [
+        "fleet shell snapshot decoded bytes",
+        run.fleetShell.http.decodedBodyBytes,
+        FLEET_SHELL_BUDGET.decodedBytes,
+      ],
+      ["fleet shell snapshot rows", run.fleetShell.rows, FLEET_SHELL_BUDGET.rows],
     ] as const;
     for (const [metric, observed, maximum] of checks) {
       if (observed > maximum) {
@@ -234,6 +263,7 @@ export function formatTransferBudgetReport(runs: ReadonlyArray<TransferBudgetRun
     "The measured turn is observed on three sockets at once: one with only the thread subscription (the capped rows), one with only the shell subscription, and a second client holding both. Server egress is the sum of the three. After the turn the second client disconnects and new sockets resubscribe from the cursor it held before the turn, which is the cursor a backgrounded phone would hold. SQL statements are the SQL client's per-statement spans, counted across V2 persistence and the HTTP/WebSocket handlers.",
     `Scenario: ${TRANSFER_HISTORY_TURN_COUNT} historical turns with ${TRANSFER_HISTORY_TOOLS_PER_TURN} command tools and one retained ${formatBytes(TRANSFER_HISTORY_MCP_RESULT_BYTES)} MCP result each, followed by one measured turn with ${TRANSFER_MEASURED_TOOLS} command tools and a retained ${formatBytes(TRANSFER_MEASURED_MCP_RESULT_BYTES)} MCP result. Synthetic V2 domain events exercise persistence, wire projection, and production HTTP/subscription handlers; this does not measure provider adapter ingestion. Payloads contain no user data.`,
     `The long-turn rows open a second thread through the compact bounded snapshot web and mobile use, then load one older page. Its last user turn holds ${LONG_TURN_WORK_ROWS} rows of reasoning, commands, file changes, MCP calls, progress messages, checkpoints, and subagents.`,
+    `The fleet shell rows read the cold-connect shell after ${SHELL_SETTLED_THREADS} settled threads and ${SHELL_IDLE_SUBAGENTS} idle subagent children are added. Budgeted reads send the opt-ins current clients send; "older client" rows show the same read without the opt-ins added since row budgets.`,
     "",
     "| Provider | Total thread wire | Budget | Result |",
     "| --- | ---: | ---: | --- |",
@@ -311,8 +341,17 @@ export function formatTransferBudgetReport(runs: ReadonlyArray<TransferBudgetRun
         run.reconnectShell,
       ),
       infoRow(run.provider, "reconnect", "SQL statements", run.reconnectSqlStatements, String),
-      ...longTurnRows(run.provider, "long-turn cold open", run.longTurnColdOpen),
-      ...longTurnRows(run.provider, "long-turn history page", run.longTurnHistoryPage),
+      ...timelineRows(run.provider, "long-turn cold open", run.longTurnColdOpen, LONG_TURN_BUDGET),
+      ...legacyRows(run.provider, "long-turn cold open", run.longTurnColdOpenLegacy),
+      ...timelineRows(
+        run.provider,
+        "long-turn history page",
+        run.longTurnHistoryPage,
+        LONG_TURN_BUDGET,
+      ),
+      ...legacyRows(run.provider, "long-turn history page", run.longTurnHistoryPageLegacy),
+      ...timelineRows(run.provider, "fleet shell snapshot", run.fleetShell, FLEET_SHELL_BUDGET),
+      ...legacyRows(run.provider, "fleet shell snapshot", run.fleetShellLegacy),
     );
   }