fix(agent): reserve the requested max output when deciding to compact

src/background/agent/agent.ts

@@ -138,7 +138,7 @@ const MAX_EMPTY_LLM_ATTEMPTS = 2;
 // ---------------------------------------------------------------------------
 
 const DEFAULT_CONTEXT_WINDOW = 262000;
-const RESERVE_TOKENS = 16000;      // Reserved for the response
+const RESERVE_TOKENS = 16000;      // Minimum reserved for the response
 const KEEP_RECENT_TOKENS = 40000;  // Recent tokens retained during compaction
 const TOOL_RESULT_RECENT_ROUNDS = 2;
 const IMAGE_MESSAGE_RECENT_LIMIT = 8;
@@ -1479,7 +1479,7 @@ export class Agent {
         messageTokens,
         inputTokens: tokens,
         contextWindow: this._contextWindow(),
-        reserveTokens: RESERVE_TOKENS,
+        reserveTokens: this._reserveTokens(),
         maxKeepTokens: KEEP_RECENT_TOKENS,
         force,
         retainLatestRound: trigger === "manual" || trigger === "slash_command",
@@ -1489,7 +1489,7 @@ export class Agent {
     this._traceContextEstimate("compact_check", contextEstimate, {
       compactThreshold: plan.threshold,
       contextWindow: this._contextWindow(),
-      reserveTokens: RESERVE_TOKENS,
+      reserveTokens: this._reserveTokens(),
       toolSchemaCount: Array.isArray(toolSchemas) ? toolSchemas.length : 0,
       storedMessageCount: this.messages.length,
       sendMessageCount: contextEstimate.messageCount,
@@ -1584,7 +1584,7 @@ export class Agent {
       this._traceContextEstimate("compact_after", updatedEstimate, {
         compactThreshold: this._compactThreshold(),
         contextWindow: this._contextWindow(),
-        reserveTokens: RESERVE_TOKENS,
+        reserveTokens: this._reserveTokens(),
         toolSchemaCount: Array.isArray(toolSchemas) ? toolSchemas.length : 0,
         storedMessageCount: this.messages.length,
         boundaryId: this._currentContextBoundaryId(),
@@ -2190,7 +2190,7 @@ export class Agent {
       snapshot = {
         contextUsageId: this._makeSpanId("ctx", ++this._contextUsageSeq),
         contextWindow,
-        reserveTokens: RESERVE_TOKENS,
+        reserveTokens: this._reserveTokens(),
         threshold: this._compactThreshold(),
         inputTokensEstimate: Math.ceil(estimate),
         messageCount: Math.max(0, Number(extra.messageCount) || 0),
@@ -2257,7 +2257,13 @@ export class Agent {
   }
 
   _compactThreshold(runProfile = this._runProfile) {
-    return compactThreshold(this._contextWindow(runProfile), RESERVE_TOKENS);
+    return compactThreshold(this._contextWindow(runProfile), this._reserveTokens(runProfile));
+  }
+
+  /** Room left for the response: at least the max_tokens each request asks for. */
+  _reserveTokens(runProfile = this._runProfile) {
+    const maxOutputTokens = Number(runProfile?.model?.maxOutputTokens) || 0;
+    return Math.max(RESERVE_TOKENS, maxOutputTokens);
   }
 
   _makeSpanId(kind: string, id: unknown) {

tests/unit/test-compaction-reserve.mjs

@@ -0,0 +1,29 @@
+import assert from "node:assert/strict";
+
+globalThis.chrome ||= {
+  storage: { local: { get: async () => ({}), set: async () => {} }, onChanged: { addListener() {} } },
+  runtime: { id: "test", getManifest: () => ({ version: "0" }) },
+};
+
+const { Agent } = await import("../../src/background/agent/agent.ts");
+
+function agentWith(model) {
+  return Object.assign(Object.create(Agent.prototype), { _runProfile: { model }, _runtimeConfig: {} });
+}
+
+// Compaction must leave room for the max_tokens the request will ask for, or a strict
+// OpenAI-compatible server rejects input + max_tokens > context window.
+for (const model of [
+  { contextWindow: 131072, maxOutputTokens: 32768 }, // custom model: default output limit
+  { contextWindow: 200000, maxOutputTokens: 64000 },
+  { contextWindow: 128000, maxOutputTokens: 8192 },
+]) {
+  const threshold = agentWith(model)._compactThreshold();
+  assert.ok(threshold + model.maxOutputTokens <= model.contextWindow,
+    `compaction threshold ${threshold} + max output ${model.maxOutputTokens} exceeds window ${model.contextWindow}`);
+}
+
+// Small output limits keep the existing 16,000-token floor.
+assert.equal(agentWith({ contextWindow: 128000, maxOutputTokens: 8192 })._compactThreshold(), 128000 - 16000);
+
+console.log("compaction reserve tests passed");