test(e2e): recompute #275 arithmetic fixtures; isolate e2e schema preload (release gate)

Author: ualtinok

packages/e2e-tests/bunfig.toml

@@ -0,0 +1,5 @@
+# Keep direct plugin imports in e2e tests away from the user's shared database.
+# This is especially important for Pi subagent tests, which open storage without
+# first booting the isolated Pi harness process.
+[test]
+preload = ["../plugin/test-preload.ts"]

packages/e2e-tests/tests/context-limits.test.ts

@@ -50,6 +50,9 @@ describe("context-limit resolution", () => {
         const sessionId = await h.createSession();
         await h.sendPrompt(sessionId, "probe turn for context-limit resolution.");
 
+        // Mock Anthropic uses a shared context window: resolveLimit reserves
+        // min(8,192, 25% of 50,000) = 8,192 output tokens, leaving 41,808
+        // usable input tokens. The exact pressure is 20,000 / 41,808 * 100.
         // Give the event handler a moment to persist last_context_percentage.
         await Bun.sleep(300);
 
@@ -59,12 +62,12 @@ describe("context-limit resolution", () => {
             .get(sessionId) as { last_context_percentage: number } | null;
 
         const pct = row?.last_context_percentage ?? 0;
-        console.log(`[TEST] last_context_percentage = ${pct} (expected exactly 40 for 20K/50K)`);
+        console.log(
+            `[TEST] last_context_percentage = ${pct} (expected 47.83773440489858 for 20K/41,808)`,
+        );
 
-        // 20_000 input_tokens / 50_000 context_limit = exactly 40%. Mock returns
-        // exact usage and plugin rounds to 1 decimal; this must be 40.0.
-        // Tightened from the old 35-45 range (Finding #5 — the range was masking
-        // a real regression surface because ANY value in [35,45] passed).
-        expect(pct).toBe(40);
+        // Keep the exact value: a broad range would hide changes to the
+        // shared-window output reservation arithmetic.
+        expect(pct).toBe(47.83773440489858);
     }, 60_000);
 });

packages/e2e-tests/tests/pi-context-limits.test.ts

@@ -60,6 +60,8 @@ describe("pi context-limit resolution", () => {
             { timeoutMs: 5_000, label: "pi last_context_percentage persisted" },
         );
 
-        expect(pct).toBe(40);
+        // Pi's shared-window model reserves min(8,192, 25% of 50,000) =
+        // 8,192 output tokens, so 20,000 / 41,808 * 100 is the exact pressure.
+        expect(pct).toBe(47.83773440489858);
     }, 60_000);
 });

packages/e2e-tests/tests/pi-overflow-recovery.test.ts

@@ -118,7 +118,10 @@ describe("pi context overflow detection", () => {
         );
 
         expect(afterNext.detected_context_limit).toBe(120_000);
-        expect(afterNext.last_context_percentage).toBe(50);
+        // The detected combined window is narrowed before reservation:
+        // 120,000 - min(8,192, 25% of 120,000) = 111,808 usable tokens.
+        // The next pass therefore reports 60,000 / 111,808 * 100 exactly.
+        expect(afterNext.last_context_percentage).toBe(53.66342301087579);
     }, 120_000);
 
     it("does not persist detected_context_limit for non-overflow rate-limit errors", async () => {