fix(shared): restrict cached Anthropic 1M context to GA 4.6 models only

2026-03-25 14:29:59 +09:00
parent 719a58270b
commit bf804b0626
3 changed files with 137 additions and 48 deletions
@@ -135,9 +135,96 @@ describe("context-window-monitor modelContextLimitsCache", () => {
    })
  })

-  describe("#given Anthropic provider with cached context limit and 1M mode disabled", () => {
+  describe("#given Anthropic 4.6 provider with cached context limit and 1M mode disabled", () => {
    describe("#when cached usage is below threshold of cached limit", () => {
      it("#then should respect the cached limit and skip the reminder", async () => {
+        // given
+        const modelContextLimitsCache = new Map<string, number>()
+        modelContextLimitsCache.set("anthropic/claude-sonnet-4-6", 500000)
+
+        const hook = createContextWindowMonitorHook({} as never, {
+          anthropicContext1MEnabled: false,
+          modelContextLimitsCache,
+        })
+        const sessionID = "ses_anthropic_cached_limit_respected"
+
+        await hook.event({
+          event: {
+            type: "message.updated",
+            properties: {
+              info: {
+                role: "assistant",
+                sessionID,
+                providerID: "anthropic",
+                modelID: "claude-sonnet-4-6",
+                finish: true,
+                tokens: {
+                  input: 150000,
+                  output: 0,
+                  reasoning: 0,
+                  cache: { read: 10000, write: 0 },
+                },
+              },
+            },
+          },
+        })
+
+        // when
+        const output = createOutput()
+        await hook["tool.execute.after"]({ tool: "bash", sessionID, callID: "call_1" }, output)
+
+        // then — 160K/500K = 32%, well below 70% threshold
+        expect(output.output).toBe("original")
+      })
+    })
+
+    describe("#when cached usage exceeds threshold of cached limit", () => {
+      it("#then should use the cached limit for the reminder", async () => {
+        // given
+        const modelContextLimitsCache = new Map<string, number>()
+        modelContextLimitsCache.set("anthropic/claude-sonnet-4-6", 500000)
+
+        const hook = createContextWindowMonitorHook({} as never, {
+          anthropicContext1MEnabled: false,
+          modelContextLimitsCache,
+        })
+        const sessionID = "ses_anthropic_cached_limit_exceeded"
+
+        await hook.event({
+          event: {
+            type: "message.updated",
+            properties: {
+              info: {
+                role: "assistant",
+                sessionID,
+                providerID: "anthropic",
+                modelID: "claude-sonnet-4-6",
+                finish: true,
+                tokens: {
+                  input: 350000,
+                  output: 0,
+                  reasoning: 0,
+                  cache: { read: 10000, write: 0 },
+                },
+              },
+            },
+          },
+        })
+
+        // when
+        const output = createOutput()
+        await hook["tool.execute.after"]({ tool: "bash", sessionID, callID: "call_1" }, output)
+
+        // then — 360K/500K = 72%, above 70% threshold, uses cached 500K limit
+        expect(output.output).toContain("context remaining")
+        expect(output.output).toContain("500,000-token context window")
+      })
+    })
+  })
+
+  describe("#given older Anthropic provider with cached context limit and 1M mode disabled", () => {
+    describe("#when cached usage would only exceed the incorrect cached limit", () => {
+      it("#then should ignore the cached limit and use the 200K default", async () => {
        // given
        const modelContextLimitsCache = new Map<string, number>()
        modelContextLimitsCache.set("anthropic/claude-sonnet-4-5", 500000)
@@ -146,7 +233,7 @@ describe("context-window-monitor modelContextLimitsCache", () => {
          anthropicContext1MEnabled: false,
          modelContextLimitsCache,
        })
-        const sessionID = "ses_anthropic_cached_limit_respected"
+        const sessionID = "ses_anthropic_older_model_ignores_cached_limit"

        await hook.event({
          event: {
@@ -173,51 +260,9 @@ describe("context-window-monitor modelContextLimitsCache", () => {
        const output = createOutput()
        await hook["tool.execute.after"]({ tool: "bash", sessionID, callID: "call_1" }, output)

-        // then — 160K/500K = 32%, well below 70% threshold
-        expect(output.output).toBe("original")
-      })
-    })
-
-    describe("#when cached usage exceeds threshold of cached limit", () => {
-      it("#then should use the cached limit for the reminder", async () => {
-        // given
-        const modelContextLimitsCache = new Map<string, number>()
-        modelContextLimitsCache.set("anthropic/claude-sonnet-4-5", 500000)
-
-        const hook = createContextWindowMonitorHook({} as never, {
-          anthropicContext1MEnabled: false,
-          modelContextLimitsCache,
-        })
-        const sessionID = "ses_anthropic_cached_limit_exceeded"
-
-        await hook.event({
-          event: {
-            type: "message.updated",
-            properties: {
-              info: {
-                role: "assistant",
-                sessionID,
-                providerID: "anthropic",
-                modelID: "claude-sonnet-4-5",
-                finish: true,
-                tokens: {
-                  input: 350000,
-                  output: 0,
-                  reasoning: 0,
-                  cache: { read: 10000, write: 0 },
-                },
-              },
-            },
-          },
-        })
-
-        // when
-        const output = createOutput()
-        await hook["tool.execute.after"]({ tool: "bash", sessionID, callID: "call_1" }, output)
-
-        // then — 360K/500K = 72%, above 70% threshold, uses cached 500K limit
+        // then
        expect(output.output).toContain("context remaining")
-        expect(output.output).toContain("500,000-token context window")
+        expect(output.output).toContain("200,000-token context window")
      })
    })
  })