fix(stats): 'With context-mode' = retrieval-tool returns only, not sandbox work-output

Refines 6d667348. The earlier fold counted ALL tool_calls.bytes_returned into 'With context-mode', including ctx_execute/batch/file sandbox stdout — work-output the model would see regardless. That crushed the redirect-savings bar to a false ~43%. Now folds ONLY retrieval returns (ctx_search + ctx_fetch_and_index) = the bytes the model paid to ACCESS the kept-out content. Empirical on this session: avoided 2.5 MB vs retrieval 125 KB = 95% kept out (vs the absurd '1 B / 100%' bug and the wrong '43%'). TDD: 8.1b now seeds a work tool (excluded) + two retrieval tools (included).

src/session/analytics.ts

@@ -1275,13 +1275,14 @@ export function getRealBytesStats(opts: {
             if (snap?.bytes) snapshotBytes += Number(snap.bytes);
           } catch { /* old schema */ }
           try {
-            // MCP tool returns (ctx_execute/ctx_search stdout, redirect stubs)
-            // land in tool_calls.bytes_returned via incrementToolCall, NOT in
-            // session_events.bytes_returned (which carries snapshot-replay only).
-            // Folding them here is what the "With context-mode" bar measures;
-            // without it a tool-heavy session rendered a false "1 B / 100%".
+            // "With context-mode" = the bytes the model paid to ACCESS the
+            // kept-out content: ctx_search (query the index) + ctx_fetch_and_index
+            // (fetch + index a URL). Sandbox compute (ctx_execute/batch/file) is
+            // work-output the model would see regardless — NOT redirect savings —
+            // so it is excluded; folding it crushed the bar to a false ~43%.
             const tc = sdb.prepare(
-              "SELECT COALESCE(SUM(bytes_returned), 0) AS bytes FROM tool_calls WHERE session_id = ?",
+              `SELECT COALESCE(SUM(bytes_returned), 0) AS bytes FROM tool_calls
+               WHERE session_id = ? AND tool IN ('ctx_search', 'ctx_fetch_and_index')`,
             ).get(opts.sessionId) as { bytes: number } | undefined;
             if (tc?.bytes) bytesReturned += Number(tc.bytes);
           } catch { /* old schema: no tool_calls table */ }
@@ -1324,7 +1325,8 @@ export function getRealBytesStats(opts: {
                FROM tool_calls
                WHERE session_id IN (
                  SELECT session_id FROM session_meta WHERE project_dir = ?
-               )`,
+               )
+               AND tool IN ('ctx_search', 'ctx_fetch_and_index')`,
             ).get(opts.projectDir) as { bytes: number } | undefined;
             if (tc?.bytes) bytesReturned += Number(tc.bytes);
           } catch { /* old schema: no tool_calls table */ }
@@ -1351,7 +1353,8 @@ export function getRealBytesStats(opts: {
           } catch { /* old schema */ }
           try {
             const tc = sdb.prepare(
-              "SELECT COALESCE(SUM(bytes_returned), 0) AS bytes FROM tool_calls",
+              `SELECT COALESCE(SUM(bytes_returned), 0) AS bytes FROM tool_calls
+               WHERE tool IN ('ctx_search', 'ctx_fetch_and_index')`,
             ).get() as { bytes: number } | undefined;
             if (tc?.bytes) bytesReturned += Number(tc.bytes);
           } catch { /* old schema: no tool_calls table */ }

tests/session/real-bytes-stats.test.ts

@@ -130,12 +130,13 @@ describe("getRealBytesStats (Phase 8 renderer source-of-truth)", () => {
     expect(r.totalSavedTokens).toBeGreaterThan(9_000); // ≈ 9_500
   });
 
-  test("8.1b conversation tier: folds tool_calls.bytes_returned (MCP returns) into bytesReturned", () => {
-    // Production reality: ctx_execute / ctx_search return bytes to context via
-    // the tool_calls counter (incrementToolCall / persistToolCallCounter), NOT
-    // session_events.bytes_returned. A session with real MCP returns but no
-    // snapshot replay reported bytesReturned=0 before this fix, rendering the
-    // Section 1 bar as "With context-mode: 1 B / 100% kept out" — false.
+  test("8.1b conversation tier: 'With context-mode' folds ONLY retrieval tool returns, not sandbox work-output", () => {
+    // "With context-mode" = the bytes the model paid to ACCESS kept-out content
+    // (ctx_search / ctx_fetch_and_index), via the tool_calls counter — NOT
+    // session_events.bytes_returned (snapshot-replay only, ~0). Sandbox compute
+    // (ctx_execute) is work-output the model would see regardless, so it is
+    // EXCLUDED. Before this fix bytesReturned was 0 ("1 B / 100%"); an earlier
+    // over-broad fold counted ctx_execute too, crushing the bar to a false ~43%.
     const dir = mkSessionsDir();
     const sid = `sess-${randomUUID()}`;
     const dbPath = dbPathFor(dir, "cafebabecafebabe");
@@ -144,13 +145,18 @@ describe("getRealBytesStats (Phase 8 renderer source-of-truth)", () => {
       sid,
       [{ type: "tool_use", category: "file", data: "src/app.ts", bytesAvoided: 90_000, bytesReturned: 0 }],
       undefined,
-      [{ tool: "ctx_execute", bytesReturned: 8_000 }, { tool: "ctx_search", bytesReturned: 2_000 }],
+      [
+        { tool: "ctx_execute", bytesReturned: 800_000 },        // work-output → EXCLUDED
+        { tool: "ctx_search", bytesReturned: 2_000 },           // retrieval → INCLUDED
+        { tool: "ctx_fetch_and_index", bytesReturned: 1_500 },  // retrieval → INCLUDED
+      ],
     );
 
     const r = getRealBytesStats({ sessionId: sid, sessionsDir: dir });
 
-    // 8_000 + 2_000 from tool_calls — the bytes that actually entered context.
-    expect(r.bytesReturned).toBe(10_000);
+    // Only the two retrieval tools (2_000 + 1_500); the 800 KB ctx_execute
+    // work-output is NOT redirect savings and must not enter "With context-mode".
+    expect(r.bytesReturned).toBe(3_500);
     expect(r.bytesAvoided).toBe(90_000);
   });