- 撤销上轮 inject 行过滤:裁历史中间行从首个被删行起前缀缓存全断,等于全价重读 (memory_note 坐会话开头附近,伤害≈整段) - 一并修同类更隐蔽病灶:withProfile=false 会把第 1 条画像快照抽掉,缓存从 index 1 就断——画像快照与占位裁剪全部照常(与常规请求同一拼法),压缩只换尾巴 - 摘要看不到距尾≥2轮的已省原料可接受:那些多已被后续 assistant 正文复述/已落盘 - 教训入码:文件内写明「压缩要同构、标题可狠筛,两套逻辑不可串用」+冒烟改断言 字节同构
68 lines
4.8 KiB
JavaScript
68 lines
4.8 KiB
JavaScript
// 临时冒烟:重负载占位裁剪(跑完即删)。
|
||
const assert = require('assert');
|
||
const llm = require('../server/llm');
|
||
|
||
const big = n => 'x'.repeat(n);
|
||
const mk = (id, role, extra) => Object.assign({ id, role, type: null, content: '' }, extra);
|
||
const call = (id, name, args) => ({ id, type: 'function', function: { name, arguments: JSON.stringify(args) } });
|
||
|
||
const rows = [
|
||
mk(1, 'user', { content: '抓X存notes再写summary' }),
|
||
mk(2, 'assistant', { tool_calls: JSON.stringify([call('c1', 'fetch', { url: 'https://e.com/x' })]) }),
|
||
mk(3, 'tool', { tool_call_id: 'c1', name: 'fetch', content: 'HTTP 200\n' + big(8000) }),
|
||
mk(4, 'assistant', { tool_calls: JSON.stringify([call('c2', 'write_file', { path: 'notes.md', content: big(12000) })]) }),
|
||
mk(5, 'tool', { tool_call_id: 'c2', name: 'write_file', content: '已创建文件 notes.md(12003 字节)' }),
|
||
mk(6, 'assistant', { tool_calls: JSON.stringify([call('c3', 'write_file', { path: 'summary.md', content: '小内容' })]) }),
|
||
mk(7, 'tool', { tool_call_id: 'c3', name: 'write_file', content: '已创建文件 summary.md(9 字节)' }),
|
||
mk(8, 'assistant', { content: '两篇已写好' }),
|
||
];
|
||
const session = () => ({ history: rows.slice(), profile_snapshot: '', blocksCache: new Map() });
|
||
|
||
// A) 全历史:距尾够远 → fetch 回执与 write_file(大) 参数被占位;write_file(小)/回执/收尾不动
|
||
let msgs = llm.buildMessages(session(), 'Simplified Chinese', null, null);
|
||
const a = msgs.filter(m => m.role === 'assistant' && m.tool_calls);
|
||
const argsNotes = JSON.parse(a.find(m => m.tool_calls[0].id === 'c2').tool_calls[0].function.arguments);
|
||
assert(argsNotes.content === '[OMITTED, size=12000]', 'write_file 大参数应占位: ' + argsNotes.content);
|
||
const argsSum = JSON.parse(a.find(m => m.tool_calls[0].id === 'c3').tool_calls[0].function.arguments);
|
||
assert.strictEqual(argsSum.content, '小内容', '低于 TRIM_MIN 不裁');
|
||
const t = msgs.find(m => m.role === 'tool' && m.tool_call_id === 'c1');
|
||
assert(t.content === '[OMITTED, http_status=200, size=8009]', 'fetch 回执应占位: ' + t.content);
|
||
assert(msgs.find(m => m.role === 'tool' && m.tool_call_id === 'c2').content.includes('已创建文件'), 'write 回执不动');
|
||
|
||
// B) 截到第 5 行(write 刚落盘的当次调用):都太新鲜,不裁
|
||
msgs = llm.buildMessages({ history: rows.slice(0, 5), profile_snapshot: '', blocksCache: new Map() }, 'Simplified Chinese', null, null);
|
||
assert(msgs.find(m => m.role === 'tool' && m.tool_call_id === 'c1').content.length > 7000, 'fetch 距尾 1 轮应保留');
|
||
assert.strictEqual(JSON.parse(msgs.filter(m => m.tool_calls).pop().tool_calls[0].function.arguments).content.length, 12000, 'write 参数距尾 0 轮应保留');
|
||
|
||
// C) 压缩回合与常规拼装字节同构,仅尾部多一条指令(前缀缓存铁律)
|
||
const normal = llm.buildMessages(session(), 'Simplified Chinese', null, null);
|
||
const cm = llm.compressMessages(session(), 'Simplified Chinese', null, null);
|
||
assert(cm.length === normal.length + 1, '压缩=常规+尾指令');
|
||
assert(JSON.stringify(cm.slice(0, -1)) === JSON.stringify(normal), '前缀字节同构');
|
||
assert(/compact|concise/.test(cm.at(-1).content), '尾指令在位');
|
||
|
||
// D) read_file 回执同款
|
||
const r3orig = rows[2];
|
||
rows[2] = mk(3, 'tool', { tool_call_id: 'c1', name: 'read_file', content: 'HTTP 200\n' + big(8000) });
|
||
rows[1] = mk(2, 'assistant', { tool_calls: JSON.stringify([call('c1', 'read_file', { path: 'a.md' })]) });
|
||
msgs = llm.buildMessages(session(), 'Simplified Chinese', null, null);
|
||
assert(msgs.find(m => m.tool_call_id === 'c1' && m.role === 'tool').content === '[OMITTED, size=8009]', 'read_file 应占位');
|
||
rows[2] = r3orig;
|
||
rows[1] = mk(2, 'assistant', { tool_calls: JSON.stringify([call('c1', 'fetch', { url: 'https://e.com/x' })]) });
|
||
|
||
// E) 开关:0=关
|
||
llm.setPayloadTrim({ trim_write_rounds: 0, trim_result_rounds: 0 });
|
||
msgs = llm.buildMessages(session(), 'Simplified Chinese', null, null);
|
||
assert(msgs.find(m => m.role === 'tool' && m.tool_call_id === 'c1').content.length > 7000, '0=关');
|
||
llm.setPayloadTrim({}); // 回落默认 1/2
|
||
|
||
// F) 非法/缺失参数不炸(arguments 非 JSON、无 url 的 fetch;需距尾够远)
|
||
rows.push(mk(9, 'assistant', { tool_calls: JSON.stringify([call('c9', 'fetch', {})]) }));
|
||
rows.push(mk(10, 'tool', { tool_call_id: 'c9', name: 'fetch', content: 'HTTP 404\n' + big(700) }));
|
||
rows.push(mk(11, 'assistant', { content: '页面不存在,换个 url?' }));
|
||
rows.push(mk(12, 'assistant', { tool_calls: JSON.stringify([call('c12', 'read_file', { path: 'notes.md' })]) }));
|
||
msgs = llm.buildMessages(session(), 'Simplified Chinese', null, null);
|
||
assert(msgs.find(m => m.tool_call_id === 'c9').content === '[OMITTED, http_status=404, size=709]', '缺 url 仍占位: ' + msgs.find(m => m.tool_call_id === 'c9').content);
|
||
|
||
console.log('TRIM-SMOKE-OK');
|