chore: version up to 2.80.40 and package with ADR-0008

src/agent.ts

@@ -551,14 +551,13 @@ export class AgentExecutor {
             //   (2) 대화 기록을 남은 예산에 맞게 압축하고 (UI 표시용 chatHistory 는 건드리지 않음)
             //   (3) 동적으로 출력 상한(maxOutputTokens)을 계산한다.
             // ──────────────────────────────────────────────────────────────────
-            // Genuinely tiny models (≤3B) sometimes fail on prompts that fit their *nominal* context
-            // but exceed their *effective* capability (server log shows truncated=0 yet eval time≈0ms —
-            // the model emitted EOS as the first token). When detected, budget against a smaller
-            // effective window so the system prompt / RAG / history get shrunk proactively.
-            // Note: 4B+ models (e.g. gemma-4-e4b with a 100k+ window) are competent — DON'T cap them,
-            // or the output budget gets squeezed to the minimum and answers come out truncated.
+            // Optional opt-in guard (g1nation.smallModelContextCap, OFF/0 by default): some very small
+            // models (≤3B) emit EOS as the first token when the prompt is near their context window
+            // even though it nominally fits. If the user opted in, budget ≤3B models against that
+            // smaller effective window. Never applied to 4B+ models, and never when the setting is 0 —
+            // capping squeezes the output-token budget, so it's a knob, not a default.
             const modelParamB = estimateModelParamsB(actualModel);
-            const smallModelCap = config.smallModelContextCap; // 0 disables this guard
+            const smallModelCap = config.smallModelContextCap; // 0 = disabled (default)
             const cappedForSmallModel = smallModelCap > 0
                 && modelParamB !== null && modelParamB <= 3
                 && config.contextLength > smallModelCap;

src/config.ts

@@ -122,7 +122,7 @@ export function getConfig(): IAgentConfig {
             return v === 'truncateMiddle' || v === 'rollingWindow' ? v : 'stopAtLimit';
         })(),
         autoCompactHistory: cfg.get<boolean>('autoCompactHistory', true),
-        smallModelContextCap: Math.max(0, cfg.get<number>('smallModelContextCap', 16384)),
+        smallModelContextCap: Math.max(0, cfg.get<number>('smallModelContextCap', 0)),
         autoContinueOnOutputLimit: cfg.get<boolean>('autoContinueOnOutputLimit', true),
         maxAutoContinuations: Math.max(0, Math.min(10, cfg.get<number>('maxAutoContinuations', 4))),
         finalOnlyRetryOnThoughtLeak: cfg.get<boolean>('finalOnlyRetryOnThoughtLeak', true)