From ea4b5a4dfee7bad69b56bc2d48751becb00daa7b Mon Sep 17 00:00:00 2001
From: xiaowei <xiaowei@xiaowei.local>
Date: Thu, 20 Aug 2026 00:10:51 +0800
Subject: [PATCH] =?UTF-8?q?fix(distill):=20max=5Ftokens=20600=E2=86=92800?=
 =?UTF-8?q?=20=E2=80=94=20Agnes=E6=8E=A8=E7=90=86=E6=A8=A1=E5=9E=8BJSON?=
 =?UTF-8?q?=E4=B8=8D=E5=86=8D=E6=88=AA=E6=96=AD?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

engine.go: max_tokens 600→800
consolidate.go: max_tokens 300→800
根因: Agnes 2.0-flash是推理模型，max_tokens=600里大量是推理token，
正文只剩~300→JSON截断→parse error→蒸馏退化。
800确保推理+正文都够用。
---
 go/internal/distill/consolidate.go | 2 +-
 go/internal/distill/engine.go      | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/go/internal/distill/consolidate.go b/go/internal/distill/consolidate.go
index d9f9288..d4f9ce7 100644
--- a/go/internal/distill/consolidate.go
+++ b/go/internal/distill/consolidate.go
@@ -184,7 +184,7 @@ func (e *Engine) ConsolidateMemory(target MemoryCandidate, candidates []MemoryCa
 		"model":       e.LLMModel,
 		"messages":    []map[string]string{{"role": "user", "content": prompt}},
 		"temperature": 0.1,
-		"max_tokens":  300,
+		"max_tokens":  800,
 	}
 	jsonBody, err := json.Marshal(body)
 	if err != nil {
diff --git a/go/internal/distill/engine.go b/go/internal/distill/engine.go
index 0847ef4..f5bab50 100644
--- a/go/internal/distill/engine.go
+++ b/go/internal/distill/engine.go
@@ -339,7 +339,7 @@ func (e *Engine) callLLM5D(content string) (LLMResponse, error) {
 			{"role": "user", "content": prompt},
 		},
 		"temperature": 0.2,
-		"max_tokens":  600,
+		"max_tokens":  800,
 	}
 
 	jsonBody, err := json.Marshal(body)