From b0d9bde66481c72405b06d34f6cfb8dcf59f2105 Mon Sep 17 00:00:00 2001 From: matevip Date: Tue, 19 May 2026 16:53:06 +0800 Subject: [PATCH] fix(agent): clamp DashScope max_tokens to the provider's 8192 ceiling --- .../mate/agent/graph/node/ReasoningNode.java | 23 ++++++++++++++++++- 1 file changed, 22 insertions(+), 1 deletion(-) diff --git a/mateclaw-server/src/main/java/vip/mate/agent/graph/node/ReasoningNode.java b/mateclaw-server/src/main/java/vip/mate/agent/graph/node/ReasoningNode.java index 468e500b..3bbbb4d0 100644 --- a/mateclaw-server/src/main/java/vip/mate/agent/graph/node/ReasoningNode.java +++ b/mateclaw-server/src/main/java/vip/mate/agent/graph/node/ReasoningNode.java @@ -71,6 +71,17 @@ public class ReasoningNode implements NodeAction { */ private static final int DEFAULT_MAX_OUTPUT_TOKENS = 16384; + /** + * DashScope's native chat API caps {@code max_tokens} at 8192 and returns a + * 400 {@code InvalidParameter} ("Range of max_tokens should be [1, 8192]") + * for anything larger. The failover layer misclassifies that 400 as + * "model not found" and silently switches to a different provider, so the + * per-call ceiling must be clamped to this value for DashScope-backed + * models — keeping {@link #DEFAULT_MAX_OUTPUT_TOKENS} for every other + * provider that does accept the larger budget. + */ + private static final int DASHSCOPE_MAX_OUTPUT_TOKENS = 8192; + /** * Hermes-agent style enforcement clause appended to every ReasoningNode * system prompt. Treats narration ("I will now …") as a protocol violation @@ -768,9 +779,19 @@ public class ReasoningNode implements NodeAction { // 始终使用 OpenAiChatOptions（而非 ToolCallingChatOptions）， // 因为 ToolCallingChatOptions 会丢失 OpenAI 特有参数（streamUsage 等）， // 导致 Kimi 等 OpenAI 兼容 API 响应异常或提前截断。 + // DashScope rejects max_tokens above its 8192 ceiling with a 400 that + // the failover layer misreads as "model not found"; clamp so a + // DashScope-backed model never overflows the provider limit. + int effectiveMaxTokens = maxOutputTokens; + if (chatModel instanceof com.alibaba.cloud.ai.dashscope.chat.DashScopeChatModel + && effectiveMaxTokens > DASHSCOPE_MAX_OUTPUT_TOKENS) { + log.debug("[ReasoningNode] Clamping max_tokens {} -> {} for DashScope-backed model", + effectiveMaxTokens, DASHSCOPE_MAX_OUTPUT_TOKENS); + effectiveMaxTokens = DASHSCOPE_MAX_OUTPUT_TOKENS; + } OpenAiChatOptions.Builder oaiBuilder = OpenAiChatOptions.builder() .toolCallbacks(toolCallbacks) - .maxTokens(maxOutputTokens); + .maxTokens(effectiveMaxTokens); if (StringUtils.hasText(effectiveReasoning)) { oaiBuilder.reasoningEffort(effectiveReasoning); }