Compaction: ignore tool result details in oversized checks (#24057)

* Compaction: ignore tool result details in oversized checks

* Tests/Compaction: type estimateTokens message callback

src/agents/compaction.tool-result-details.test.ts

@@ -3,7 +3,7 @@ import { beforeEach, describe, expect, it, vi } from "vitest";
 
 const piCodingAgentMocks = vi.hoisted(() => ({
   generateSummary: vi.fn(async () => "summary"),
-  estimateTokens: vi.fn(() => 1),
+  estimateTokens: vi.fn((_message: unknown) => 1),
 }));
 
 vi.mock("@mariozechner/pi-coding-agent", async () => {
@@ -17,7 +17,7 @@ vi.mock("@mariozechner/pi-coding-agent", async () => {
   };
 });
 
-import { summarizeWithFallback } from "./compaction.js";
+import { isOversizedForSummary, summarizeWithFallback } from "./compaction.js";
 
 describe("compaction toolResult details stripping", () => {
   beforeEach(() => {
@@ -64,4 +64,23 @@ describe("compaction toolResult details stripping", () => {
     expect(serialized).not.toContain("Ignore previous instructions");
     expect(serialized).not.toContain('"details"');
   });
+
+  it("ignores toolResult.details when evaluating oversized messages", () => {
+    piCodingAgentMocks.estimateTokens.mockImplementation((message: unknown) => {
+      const record = message as { details?: unknown };
+      return record.details ? 10_000 : 10;
+    });
+
+    const toolResult = {
+      role: "toolResult",
+      toolCallId: "call_1",
+      toolName: "browser",
+      isError: false,
+      content: [{ type: "text", text: "ok" }],
+      details: { raw: "x".repeat(100_000) },
+      timestamp: 2,
+    } as unknown as AgentMessage;
+
+    expect(isOversizedForSummary(toolResult, 1_000)).toBe(false);
+  });
 });

src/agents/compaction.ts

@@ -152,7 +152,7 @@ export function computeAdaptiveChunkRatio(messages: AgentMessage[], contextWindo
  * If single message > 50% of context, it can't be summarized safely.
  */
 export function isOversizedForSummary(msg: AgentMessage, contextWindow: number): boolean {
-  const tokens = estimateTokens(msg) * SAFETY_MARGIN;
+  const tokens = estimateCompactionMessageTokens(msg) * SAFETY_MARGIN;
   return tokens > contextWindow * 0.5;
 }
 
@@ -240,7 +240,7 @@ export async function summarizeWithFallback(params: {
   for (const msg of messages) {
     if (isOversizedForSummary(msg, contextWindow)) {
       const role = (msg as { role?: string }).role ?? "message";
-      const tokens = estimateTokens(msg);
+      const tokens = estimateCompactionMessageTokens(msg);
       oversizedNotes.push(
         `[Large ${role} (~${Math.round(tokens / 1000)}K tokens) omitted from summary]`,
       );