@dreb/coding-agent 2.65.0 → 2.66.1

package/dist/core/agent-session.js

@@ -15,7 +15,7 @@
 import { copyFileSync, existsSync, mkdirSync, realpathSync, writeFileSync } from "node:fs";
 import { tmpdir } from "node:os";
 import { basename, dirname, join, resolve } from "node:path";
-import { isContextOverflow, modelsAreEqual, resetApiProviders, supportsMax, supportsXhigh } from "@dreb/ai";
+import { isContextOverflow, isOutputLimitExhaustion, modelsAreEqual, resetApiProviders, supportsMax, supportsXhigh, } from "@dreb/ai";
 import { getDocsPath } from "../config.js";
 import { theme } from "../modes/interactive/theme/theme.js";
 import { sleep } from "../utils/sleep.js";
@@ -1036,8 +1036,7 @@ export class AgentSession {
                 type: "length_retry",
                 attempt: event.attempt,
                 maxAttempts: event.maxAttempts,
-                previousMaxTokens: event.previousMaxTokens,
-                nextMaxTokens: event.nextMaxTokens,
+                maxTokens: event.maxTokens,
                 discardedPartial: event.discardedPartial,
             };
             await this._extensionRunner.emit(extensionEvent);
@@ -2209,8 +2208,14 @@ export class AgentSession {
         if (assistantIsFromBeforeCompaction) {
             return;
         }
-        // Case 1: Overflow - LLM returned context overflow error
-        if (sameModel && isContextOverflow(assistantMessage, contextWindow)) {
+        // Case 1: Overflow - LLM returned a context error, or exhausted its
+        // configured output limit when that output could not fit beside the request.
+        const estimatedInputTokens = estimateContextTokens(this.agent.state.messages.slice(0, -1)).tokens;
+        if (sameModel &&
+            isContextOverflow(assistantMessage, contextWindow, {
+                maxOutputTokens: this.model?.maxTokens,
+                estimatedInputTokens,
+            })) {
             // K3 auto context tier: an overflow in the 256k tier upgrades to the
             // 1M tier instead of compacting — the Kimi backend grows the prompt
             // cache seamlessly. This runs even when compaction is disabled since
@@ -2247,6 +2252,10 @@ export class AgentSession {
             await this._trackAutoCompaction("overflow", true, requestWillFollow);
             return;
         }
+        // A genuine output-limit exhaustion cannot be fixed by discarding conversation
+        // history. Only the context-constrained case above enters compact-and-retry.
+        if (isOutputLimitExhaustion(assistantMessage))
+            return;
         // Case 2: Threshold - context is getting large
         // For error messages (no usage data), estimate from last successful response.
         // This ensures sessions that hit persistent API errors (e.g. 529) can still compact.
@@ -2822,6 +2831,7 @@ export class AgentSession {
                     parentModel: () => this.model?.id,
                     parentSessionFile: () => this.sessionFile,
                     parentSessionId: () => this.sessionId,
+                    singleModelMode: () => this.settingsManager.getSingleModelMode(),
                     modelRegistry: this._modelRegistry,
                     getAgentModelsForAgent: (name) => this.settingsManager?.getAgentModelsForAgent(name),
                     defaultThinkingLevel: () => this.settingsManager.getDefaultThinkingLevel() ?? DEFAULT_THINKING_LEVEL,