openova/openova
1
0
Fork 0
Code Issues Pull Requests Actions 4 Packages Projects Releases Wiki Activity

fix(axon): cap assistant messages and total payload to prevent WAF rejection on long conversations

Browse Source 
WAF anomaly scoring accumulates across the entire request body. After 2-3 turns,
assistant responses containing infrastructure terms (security, scanning, etc.)
push the total past the threshold. Added per-assistant trim (1500 chars) and a
12000-char sliding window that drops oldest messages.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
e3mrah 2026-04-26 08:44:33 +02:00
parent 40c4abe4f6
commit 00ddc1437c
1 changed files with 27 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

35
products/axon/src/providers/vllm.ts
View File

@ -33,6 +33,8 @@ export class VllmProvider {
}
private static readonly SYSTEM_MSG_MAX_CHARS = 6000;
private static readonly ASSISTANT_MSG_MAX_CHARS = 1500;
private static readonly TOTAL_MSG_MAX_CHARS = 12000;
private sanitizeMessages(messages: ChatMessage[]): ChatMessage[] {
let seenSystem = false;
@ -45,17 +47,34 @@ export class VllmProvider {
deduped.push(msg);
}
return deduped.map((msg) => {
if (msg.role !== "system" || !msg.content || msg.content.length <= VllmProvider.SYSTEM_MSG_MAX_CHARS) {
return msg;
}
const limit = VllmProvider.SYSTEM_MSG_MAX_CHARS;
const trimmed = deduped.map((msg) => {
if (!msg.content) return msg;
let limit: number;
if (msg.role === "system") limit = VllmProvider.SYSTEM_MSG_MAX_CHARS;
else if (msg.role === "assistant") limit = VllmProvider.ASSISTANT_MSG_MAX_CHARS;
else return msg;
if (msg.content.length <= limit) return msg;
const headSize = Math.floor(limit * 0.7);
const tailSize = limit - headSize;
const head = msg.content.slice(0, headSize);
const tail = msg.content.slice(-tailSize);
return { ...msg, content: `${head}\n\n[...condensed...]\n\n${tail}` };
return { ...msg, content: `${msg.content.slice(0, headSize)}\n\n[...condensed...]\n\n${msg.content.slice(-tailSize)}` };
});
const totalLimit = VllmProvider.TOTAL_MSG_MAX_CHARS;
const totalChars = trimmed.reduce((sum, m) => sum + (m.content?.length ?? 0), 0);
if (totalChars <= totalLimit) return trimmed;
const systemMsgs = trimmed.filter((m) => m.role === "system");
const nonSystemMsgs = trimmed.filter((m) => m.role !== "system");
const sysChars = systemMsgs.reduce((sum, m) => sum + (m.content?.length ?? 0), 0);
let budget = totalLimit - sysChars;
const kept: ChatMessage[] = [];
for (let i = nonSystemMsgs.length - 1; i >= 0; i--) {
const len = nonSystemMsgs[i].content?.length ?? 0;
if (budget - len < 0 && kept.length > 0) break;
kept.unshift(nonSystemMsgs[i]);
budget -= len;
}
return [...systemMsgs, ...kept];
}
private cleanPayload(body: ChatCompletionRequest, stream: boolean): Record<string, unknown> {