🤖 fix: omit providerMetadata from compaction summary to fix context display (coder#762)

After compaction, context usage was inflated because the summary
message's `providerMetadata` contained `cacheCreationInputTokens` from
the pre-compaction context.

This fix omits `providerMetadata` from the compaction summary message.
Context usage now correctly shows just the compaction output tokens
until the next API request.

Cost tracking remains accurate via the `historicalUsage` field.

_Generated with `mux`_

Author: ethanndickson

src/node/services/compactionHandler.test.ts

@@ -255,7 +255,7 @@ describe("CompactionHandler", () => {
         model: "claude-3-5-sonnet-20241022",
         usage,
         duration: 2000,
-        providerMetadata: { foo: "bar" },
+        providerMetadata: { anthropic: { cacheCreationInputTokens: 50000 } },
         systemMessageTokens: 100,
       });
       await handler.handleCompletion(event);
@@ -266,14 +266,15 @@ describe("CompactionHandler", () => {
       });
       expect(summaryEvent).toBeDefined();
       const sevt = summaryEvent?.data.message as MuxMessage;
+      // providerMetadata is omitted to avoid inflating context with pre-compaction cacheCreationInputTokens
       expect(sevt.metadata).toMatchObject({
         model: "claude-3-5-sonnet-20241022",
         usage,
         duration: 2000,
-        providerMetadata: { foo: "bar" },
         systemMessageTokens: 100,
         compacted: true,
       });
+      expect(sevt.metadata?.providerMetadata).toBeUndefined();
     });
 
     it("should emit stream-end event to frontend", async () => {

src/node/services/compactionHandler.ts

@@ -112,7 +112,10 @@ export class CompactionHandler {
     }
     const deletedSequences = clearResult.data;
 
-    // Create summary message with metadata
+    // Create summary message with metadata.
+    // We omit providerMetadata because it contains cacheCreationInputTokens from the
+    // pre-compaction context, which inflates context usage display. The historicalUsage
+    // field preserves full cost accounting from pre-compaction messages.
     const summaryMessage = createMuxMessage(
       `summary-${Date.now()}-${Math.random().toString(36).substring(2, 11)}`,
       "assistant",
@@ -123,7 +126,6 @@ export class CompactionHandler {
         model: metadata.model,
         usage: metadata.usage,
         historicalUsage,
-        providerMetadata: metadata.providerMetadata,
         duration: metadata.duration,
         systemMessageTokens: metadata.systemMessageTokens,
         muxMetadata: { type: "normal" },