fix: use max output size for completion budget

.changeset/max-output-size-completion-budget.md

@@ -0,0 +1,6 @@
+---
+"@moonshot-ai/agent-core": patch
+"@moonshot-ai/kimi-code": patch
+---
+
+Use configured model output limits as completion token caps.

packages/agent-core/src/agent/compaction/full.ts

@@ -243,6 +243,7 @@ export class FullCompaction {
       const provider = applyCompletionBudget({
         provider: this.agent.config.provider,
         budget: resolveCompletionBudget({
+          maxOutputSize: this.agent.config.maxOutputSize,
           reservedContextSize: this.agent.kimiConfig?.loopControl?.reservedContextSize,
         }),
         capability: this.agent.config.modelCapabilities,

packages/agent-core/src/agent/config/index.ts

@@ -127,6 +127,10 @@ export class ConfigState {
     return this.tryResolvedProviderConfig()?.modelCapabilities ?? UNKNOWN_CAPABILITY;
   }
 
+  get maxOutputSize(): number | undefined {
+    return this.tryResolvedProviderConfig()?.maxOutputSize;
+  }
+
   private get resolvedProviderConfig(): ResolvedRuntimeProvider | undefined {
     if (this._modelAlias === undefined) return undefined;
     return this.agent.modelProvider?.resolveProviderConfig(this._modelAlias);

packages/agent-core/src/agent/index.ts

@@ -201,6 +201,7 @@ export class Agent {
     const provider = this.config.provider.withThinking(this.config.thinkingLevel);
     const loopControl = this.kimiConfig?.loopControl;
     const completionBudgetConfig = resolveCompletionBudget({
+      maxOutputSize: this.config.maxOutputSize,
       reservedContextSize: loopControl?.reservedContextSize,
     });
     return new KosongLLM({

packages/agent-core/src/session/provider-manager.ts

@@ -17,6 +17,7 @@ export interface ResolvedRuntimeProvider {
   readonly providerName: string;
   readonly provider: KosongProviderConfig;
   readonly modelCapabilities: ModelCapability;
+  readonly maxOutputSize?: number | undefined;
 }
 
 interface ProviderManagerOptions {
@@ -115,6 +116,7 @@ export class ProviderManager implements ModelProvider {
       providerName,
       provider,
       modelCapabilities: resolveModelCapabilities(alias, provider),
+      maxOutputSize: alias.maxOutputSize,
     };
   }
 

packages/agent-core/src/utils/completion-budget.ts

@@ -16,6 +16,7 @@ const DEFAULT_UNKNOWN_CONTEXT_FALLBACK = 32000;
  * non-positive env values disable clamping.
  */
 export function resolveCompletionBudget(args: {
+  readonly maxOutputSize?: number | undefined;
   readonly reservedContextSize?: number;
   readonly env?: NodeJS.ProcessEnv;
 }): CompletionBudgetConfig | undefined {
@@ -28,6 +29,9 @@ export function resolveCompletionBudget(args: {
   if (fromLegacy !== 'absent') {
     return fromLegacy === 'disabled' ? undefined : { hardCap: fromLegacy };
   }
+  if (args.maxOutputSize !== undefined) {
+    return { hardCap: args.maxOutputSize };
+  }
   if (args.reservedContextSize !== undefined && args.reservedContextSize > 0) {
     return { fallback: args.reservedContextSize };
   }

packages/agent-core/test/harness/runtime-provider.test.ts

@@ -88,6 +88,7 @@ describe('resolveRuntimeProvider model metadata', () => {
             provider: 'openai',
             model: 'gpt-runtime',
             maxContextSize: 200000,
+            maxOutputSize: 24000,
             capabilities: ['tool_use'],
           },
         },
@@ -106,6 +107,7 @@ describe('resolveRuntimeProvider model metadata', () => {
       tool_use: true,
       max_context_tokens: 200000,
     });
+    expect(resolved.maxOutputSize).toBe(24000);
   });
 
   it('uses config Kimi capabilities without requiring an api key during OAuth setup', () => {

packages/agent-core/test/utils/completion-budget.test.ts

@@ -172,6 +172,24 @@ describe('resolveCompletionBudget', () => {
     expect(budget?.hardCap).toBe(2048);
   });
 
+  it('uses maxOutputSize as the hard cap when env vars are unset', () => {
+    const budget = resolveCompletionBudget({
+      maxOutputSize: 2048,
+      reservedContextSize: 1000,
+      env: {},
+    });
+    expect(budget?.hardCap).toBe(2048);
+    expect(budget?.fallback).toBeUndefined();
+  });
+
+  it('lets env vars override maxOutputSize', () => {
+    const budget = resolveCompletionBudget({
+      maxOutputSize: 2048,
+      env: { KIMI_MODEL_MAX_COMPLETION_TOKENS: '4096' },
+    });
+    expect(budget?.hardCap).toBe(4096);
+  });
+
   it('uses reservedContextSize as the unknown-context fallback when no env var is set', () => {
     const budget = resolveCompletionBudget({
       reservedContextSize: 12345,