fix(context): honor compaction summary output reserve

src/context/compaction/CompactionEngine.ts

@@ -51,7 +51,7 @@ export type CompactionEngineOptions = {
   model_: string;
   /** Optional summary system prompt base; the runtime always appends its summary rubric and safety constraints. */
   systemPrompt?: string;
-  /** Max output tokens for the summary call (legacy default 20_000). */
+  /** Max output tokens for the summary call. */
   maxOutputTokens?: number;
   /** Tool names whose turns should be preserved verbatim across full compaction. */
   protectedToolNames?: Iterable<string>;
@@ -65,9 +65,10 @@ export const COMPACT_SYSTEM_PROMPT_DEFAULT =
   "You are a conversation summarizer for a coding agent. Your summary will replace " +
   "the early conversation history, so it MUST preserve all information the agent " +
   "needs to continue working without repeating past steps.";
-// Keep room for the summary input inside small configured contexts such as
-// agent.maxContextTokens: 20_000.
-export const COMPACT_MAX_OUTPUT_TOKENS = 4_000;
+// Keep room for the summary input inside small configured contexts while still
+// allowing enough output for a useful summary of a large transcript.
+export const COMPACT_MAX_OUTPUT_TOKENS = 16_000;
+const COMPACT_OUTPUT_RESERVE_CONTEXT_RATIO = 0.4;
 
 const SUMMARY_MARKDOWN_HEADINGS = [
   "Objective",
@@ -194,9 +195,19 @@ export class CompactionEngine {
     const cacheReset = input.cacheReset ?? checkpointMerged;
     const stablePrefix: CanonicalMessage[] = [];
     const planningMessages = checkpoint.liveMessages;
-    const summaryOutputReserve = positiveTokenCount(input.maxOutputTokens)
+    const requestedSummaryOutputReserve = positiveTokenCount(input.maxOutputTokens)
       ?? positiveTokenCount(this.options.maxOutputTokens)
       ?? COMPACT_MAX_OUTPUT_TOKENS;
+    const summaryContextCap = positiveTokenCount(input.effectiveContextTokens);
+    const summaryOutputReserve = summaryContextCap === undefined
+      ? requestedSummaryOutputReserve
+      : Math.max(
+          1,
+          Math.min(
+            requestedSummaryOutputReserve,
+            Math.floor(summaryContextCap * COMPACT_OUTPUT_RESERVE_CONTEXT_RATIO),
+          ),
+        );
     const targetPostTokens = positiveTokenCount(input.targetPostTokens);
     const tailRatio = clamp(input.keepTailRatio ?? DEFAULT_KEEP_TAIL_RATIO, 0, 1);
     const tailTokenBudget = targetPostTokens !== undefined
@@ -265,7 +276,7 @@ export class CompactionEngine {
             input.userInstruction,
             input.signal,
             summaryAnchors,
-            input.maxOutputTokens,
+            summaryOutputReserve,
             checkpoint.previousSummaries,
           );
           summaryMessage = wrapSummaryMessage(result.message);

tests/context/compaction-engine.spec.ts

@@ -466,6 +466,43 @@ test("summary output truncated by the provider is treated as a failed compaction
   assert.equal(result.summaryGenerated, false);
 });
 
+test("summary output reserve scales with context while preserving explicit limits", async () => {
+  const summaryRequests: CanonicalModelRequest[] = [];
+  const engine = new CompactionEngine({
+    model: {
+      async *stream(request: CanonicalModelRequest): AsyncIterable<CanonicalModelEvent> {
+        summaryRequests.push(request);
+        yield { type: "message_start", role: "assistant" };
+        yield { type: "text_delta", text: "## Objective\nContinue.\n\n## Current State\nCompacted." };
+        yield { type: "message_end", finishReason: "stop" };
+      },
+    },
+    provider: "local",
+    model_: "local-chat",
+  });
+
+  const cases = [
+    { effectiveContextTokens: 95_232, expected: 16_000 },
+    { effectiveContextTokens: 20_000, expected: 8_000 },
+    { expected: 16_000 },
+    { effectiveContextTokens: 95_232, maxOutputTokens: 1_536, expected: 1_536 },
+  ];
+  for (const testCase of cases) {
+    await engine.run({
+      trigger: "auto",
+      messages: compactFixture(),
+      keepTailRatio: 0.01,
+      effectiveContextTokens: testCase.effectiveContextTokens,
+      maxOutputTokens: testCase.maxOutputTokens,
+    });
+  }
+
+  assert.deepEqual(
+    summaryRequests.map((request) => request.maxOutputTokens),
+    cases.map((testCase) => testCase.expected),
+  );
+});
+
 test("protected turns remain in chronological order with later work", async () => {
   const engine = new CompactionEngine({
     model: {