test(cli): cover provider-reported context output cap

Add regression coverage for capOutputTokens preferring the provider-reported
context size (image/vision input priced by the provider) and falling back to
the media-normalized floor when reported usage is smaller or absent.

.changeset/preserve-image-output-capacity.md

@@ -2,4 +2,4 @@
 "@kilocode/cli": patch
 ---
 
-Preserve model output capacity when requests contain encoded images.
+Preserve model output capacity when requests contain encoded images. The output token cap now uses the provider-reported context size from the previous turn, so image and vision input is measured by the provider instead of by encoded payload size.

packages/opencode/src/session/prompt.ts

@@ -1685,10 +1685,11 @@ export const layer = Layer.effect(
             tools,
             model,
             toolChoice: format.type === "json_schema" ? "required" : undefined,
-            // kilocode_change - feed the provider-reported context size from the last finished
+            // kilocode_change start - feed the provider-reported context size from the last finished
             // turn into the output-token cap, so image/vision input is measured by the provider
             // rather than by encoded payload bytes (see KiloLLM.capOutputTokens)
             reportedContextTokens: lastFinished ? KiloSessionOverflow.count(lastFinished.tokens) : undefined,
+            // kilocode_change end
           })
 
           if (structured !== undefined) {

packages/opencode/test/kilocode/session-overflow.test.ts

@@ -157,6 +157,42 @@ describe("Kilo request estimation", () => {
     expect(cap).toBeGreaterThanOrEqual(1_024)
     expect(cap).toBeLessThan(32_000)
   })
+
+  test("prefers provider-reported context over the client estimate for images", () => {
+    // The client cannot price encoded image bytes, but the provider reported a
+    // large vision-token cost for the last turn.
+    const mdl = model({ context: 300_000, output: 32_000 })
+    const messages = [
+      {
+        role: "user",
+        content: [{ type: "image", image: `data:image/png;base64,${"x".repeat(600_000)}` }],
+      },
+    ] satisfies ModelMessage[]
+
+    // Without reported usage the media-normalized estimate leaves output untouched.
+    expect(KiloLLM.capOutputTokens({ model: mdl, messages, tools: {}, configured: 32_000 })).toBe(32_000)
+
+    // With the provider-reported context size, output is capped to fit real usage.
+    expect(KiloLLM.capOutputTokens({ model: mdl, messages, tools: {}, configured: 32_000, reported: 280_000 })).toBe(
+      17_952,
+    )
+  })
+
+  test("uses the media-normalized floor when reported usage is smaller", () => {
+    const mdl = model({ context: 200_000, output: 32_000 })
+    const messages = [{ role: "user" as const, content: "x".repeat(600_000) }]
+
+    const withoutReported = KiloLLM.capOutputTokens({ model: mdl, messages, tools: {}, configured: 32_000 })
+    const withStaleReported = KiloLLM.capOutputTokens({
+      model: mdl,
+      messages,
+      tools: {},
+      configured: 32_000,
+      reported: 1_000,
+    })
+    expect(withStaleReported).toBe(withoutReported)
+    expect(withStaleReported).toBeLessThan(32_000)
+  })
 })
 
 describe("Kilo preflight compaction", () => {