From 794db55538e01b4bf0c008c493de5d8b8bf67c5d Mon Sep 17 00:00:00 2001
From: qer <wbxl2000@outlook.com>
Date: Sat, 27 Jun 2026 23:24:57 +0800
Subject: [PATCH] fix(agent-core): cap compaction output at 128k by default
 (#1156)

---
 .changeset/compaction-default-output-cap.md   |  5 +++
 .../agent-core/src/agent/compaction/full.ts   | 25 ++++++++++++--
 .../test/agent/compaction/full.test.ts        | 33 +++++++++++++++++++
 3 files changed, 61 insertions(+), 2 deletions(-)
 create mode 100644 .changeset/compaction-default-output-cap.md

diff --git a/.changeset/compaction-default-output-cap.md b/.changeset/compaction-default-output-cap.md
new file mode 100644
index 000000000..f89cb0ed1
--- /dev/null
+++ b/.changeset/compaction-default-output-cap.md
@@ -0,0 +1,5 @@
+---
+"@moonshot-ai/kimi-code": patch
+---
+
+Cap compaction output at 128k tokens by default to avoid provider max_tokens errors.
diff --git a/packages/agent-core/src/agent/compaction/full.ts b/packages/agent-core/src/agent/compaction/full.ts
index e5f134df0..31ff5c4e4 100644
--- a/packages/agent-core/src/agent/compaction/full.ts
+++ b/packages/agent-core/src/agent/compaction/full.ts
@@ -39,6 +39,15 @@ import {
 
 export const MAX_COMPACTION_RETRY_ATTEMPTS = 5;
 
+/**
+ * Default hard cap on compaction output tokens when `maxOutputSize` is not
+ * configured on the model alias. Without this, compaction falls back to the
+ * full context window size, which exceeds the `max_tokens` ceiling enforced
+ * by many OpenAI-compatible providers. 128k matches the chat-completions
+ * ceiling applied by the OpenAI Legacy provider.
+ */
+const DEFAULT_COMPACTION_MAX_COMPLETION_TOKENS = 128 * 1024;
+
 class CompactionTruncatedError extends Error {
   constructor() {
     super('Compaction response was truncated before producing a complete summary.');
@@ -261,13 +270,25 @@ export class FullCompaction {
       await this.triggerPreCompactHook(data, tokensBefore, signal);
 
       const model = this.agent.config.model;
+      const capability = this.agent.config.modelCapabilities;
+      const maxContextTokens = capability.max_context_tokens;
+      // When the model's context window is known and the user has not set
+      // `maxOutputSize`, cap compaction output to a safe default so a large
+      // context window does not push `max_tokens` past the provider's ceiling.
+      // When the window is unknown (maxContextTokens === 0), leave
+      // `maxOutputSize` unset so `resolveCompletionBudget` falls back to the
+      // conservative unknown-context fallback.
+      const defaultCompactionCap =
+        maxContextTokens > 0
+          ? Math.min(maxContextTokens, DEFAULT_COMPACTION_MAX_COMPLETION_TOKENS)
+          : undefined;
       const provider = applyCompletionBudget({
         provider: this.agent.config.provider,
         budget: resolveCompletionBudget({
-          maxOutputSize: this.agent.config.maxOutputSize,
+          maxOutputSize: this.agent.config.maxOutputSize ?? defaultCompactionCap,
           reservedContextSize: this.agent.kimiConfig?.loopControl?.reservedContextSize,
         }),
-        capability: this.agent.config.modelCapabilities,
+        capability,
       });
 
       const delays = retryBackoffDelays(MAX_COMPACTION_RETRY_ATTEMPTS);
diff --git a/packages/agent-core/test/agent/compaction/full.test.ts b/packages/agent-core/test/agent/compaction/full.test.ts
index 4e392908a..a1a630250 100644
--- a/packages/agent-core/test/agent/compaction/full.test.ts
+++ b/packages/agent-core/test/agent/compaction/full.test.ts
@@ -1811,6 +1811,39 @@ describe('FullCompaction', () => {
     expect(compactionMaxCompletionTokens).toEqual([384000]);
   });
 
+  it('uses default 128k hardCap when maxOutputSize is not configured', async () => {
+    let callCount = 0;
+    const compactionMaxCompletionTokens: unknown[] = [];
+    const generate: GenerateFn = async (provider, _system, _tools, _history, callbacks) => {
+      callCount += 1;
+      if (callCount === 1) {
+        throw new APIContextOverflowError(400, 'Context length exceeded', 'req-default-cap');
+      }
+      if (callCount === 2) {
+        compactionMaxCompletionTokens.push(providerMaxCompletionTokens(provider));
+        return textResult('Default cap compacted summary.');
+      }
+      await callbacks?.onMessagePart?.({
+        type: 'text',
+        text: 'Recovered with default cap.',
+      });
+      return textResult('Recovered with default cap.');
+    };
+    const ctx = testAgent({ generate });
+    ctx.configure({
+      provider: CATALOGUED_PROVIDER,
+      modelCapabilities: CATALOGUED_MODEL_CAPABILITIES,
+    });
+    ctx.appendExchange(1, 'old user one', 'old assistant one', 20);
+    ctx.newEvents();
+
+    await ctx.rpc.prompt({ input: [{ type: 'text', text: 'Retry with default cap' }] });
+    await ctx.untilTurnEnd();
+
+    expect(callCount).toBe(3);
+    expect(compactionMaxCompletionTokens).toEqual([128 * 1024]);
+  });
+
   it('ignores filtered assistant placeholders when checking the retained overflow suffix', async () => {
     let callCount = 0;
     const generate: GenerateFn = async (_provider, _system, _tools, _history, callbacks) => {