重构: tool_registry拆分及多批改进
This commit is contained in:
@@ -458,24 +458,44 @@ impl ContextManager {
|
||||
/// 致 over_budget_trims_old 等测试失败)。若运行时日志显示连续 role 也是 1214 来源,再补合并。
|
||||
fn ensure_sequence_legal(messages: Vec<ChatMessage>) -> Vec<ChatMessage> {
|
||||
let mut skipped = 0u32;
|
||||
let result: Vec<ChatMessage> = messages
|
||||
.into_iter()
|
||||
.skip_while(|m| {
|
||||
if matches!(m.role, MessageRole::Assistant | MessageRole::Tool) {
|
||||
skipped += 1;
|
||||
true
|
||||
} else {
|
||||
false
|
||||
let mut merged = 0u32;
|
||||
let mut fixed: Vec<ChatMessage> = Vec::with_capacity(messages.len());
|
||||
for m in messages {
|
||||
// 首条必须 user:skip 开头 assistant/tool(无前置 user 的孤儿)
|
||||
if fixed.is_empty() && matches!(m.role, MessageRole::Assistant | MessageRole::Tool) {
|
||||
skipped += 1;
|
||||
continue;
|
||||
}
|
||||
// 连续同 role 合并(user content;assistant content+tool_calls;Tool 不合并——
|
||||
// 连续 tool_result 由 anthropic_compat flush_tool_results 合并成 user blocks,此处合会丢 id)
|
||||
if let Some(last) = fixed.last_mut() {
|
||||
let same_role = std::mem::discriminant(&last.role) == std::mem::discriminant(&m.role);
|
||||
if same_role && matches!(m.role, MessageRole::User | MessageRole::Assistant) {
|
||||
if !m.content.is_empty() {
|
||||
if !last.content.is_empty() {
|
||||
last.content.push('\n');
|
||||
}
|
||||
last.content.push_str(&m.content);
|
||||
}
|
||||
if matches!(m.role, MessageRole::Assistant) {
|
||||
if let Some(calls) = m.tool_calls {
|
||||
last.tool_calls.get_or_insert_with(Vec::new).extend(calls);
|
||||
}
|
||||
}
|
||||
merged += 1;
|
||||
continue;
|
||||
}
|
||||
})
|
||||
.collect();
|
||||
if skipped > 0 {
|
||||
}
|
||||
fixed.push(m);
|
||||
}
|
||||
if skipped > 0 || merged > 0 {
|
||||
tracing::warn!(
|
||||
skipped,
|
||||
"序列修复:丢弃开头的 assistant/tool 消息(Anthropic 要求首条 user,避免 1214)"
|
||||
merged,
|
||||
"序列修复:skip 开头非 user + 合并连续同 role(view-only,避免 Anthropic/GLM 1214)"
|
||||
);
|
||||
}
|
||||
result
|
||||
fixed
|
||||
}
|
||||
|
||||
/// 全量克隆(持久化 save_conversation / build_for_request 未裁剪分支,不受裁剪影响)
|
||||
@@ -952,17 +972,22 @@ mod tests {
|
||||
fn over_budget_trims_old() {
|
||||
// 小预算强制裁剪:20 条超预算,触发裁剪且保留保护区
|
||||
let mut mgr = ContextManager::new(cfg(200));
|
||||
// user/assistant 交替(真实对话序列;连续 user 会被 ensure_sequence_legal 合并,无法测条数裁剪)
|
||||
for i in 0..20 {
|
||||
mgr.push(ChatMessage::user(&format!("这是第 {} 条较长的消息用于撑爆预算", i)));
|
||||
if i % 2 == 0 {
|
||||
mgr.push(ChatMessage::user(&format!("这是第 {} 条较长的消息用于撑爆预算", i)));
|
||||
} else {
|
||||
mgr.push(ChatMessage::assistant(&format!("第 {} 条较长的回复用于撑爆预算", i)));
|
||||
}
|
||||
}
|
||||
let (msgs, trimmed) = mgr.build_for_request(0);
|
||||
assert!(trimmed, "超预算应触发裁剪");
|
||||
assert!(msgs.len() < 20, "应裁掉部分旧消息, 实际 {}", msgs.len());
|
||||
|
||||
// 保护区:最新一条必保留
|
||||
// 保护区:最新一条必保留(末条 i=19 是 assistant)
|
||||
assert_eq!(
|
||||
msgs.last().unwrap().content,
|
||||
"这是第 19 条较长的消息用于撑爆预算",
|
||||
"第 19 条较长的回复用于撑爆预算",
|
||||
"保护区最新消息被误裁"
|
||||
);
|
||||
|
||||
|
||||
Reference in New Issue
Block a user