luulam/9router
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(translator): drop replayed reasoning fields for Groq/Mistral/Cerebras (#4220)

Browse Source 
Strict OpenAI-compatible validators reject unknown assistant-message
fields: Groq 400 ("property 'reasoning_content' is unsupported"),
Mistral 422 ("extra_forbidden"), Cerebras 400 ("wrong_api_format").

Clients driving reasoning models (Hermes Agent, and anything following
the DeepSeek/Kimi convention) echo the previous turn's reasoning_content
on every assistant message, so from the second turn on every request to
these providers fails and a fallback combo silently skips them.

Add a dropMessageFields rule to paramSupport.js that strips
reasoning_content / reasoning / reasoning_details from assistant turns
for groq, mistral, and cerebras.
This commit is contained in:
Anantachoke
2026-09-21 19:58:34 +07:00
parent 253199f16f
commit 7c2b1fe3e1
2 changed files with 62 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
open-sse/translator/concerns/paramSupport.js
View File

@@ -24,6 +24,15 @@ const STRIP_RULES = [
// "integer above maximum value, expected <= 32768". Pin an explicit endpoint cap;
// min() with the model ceiling still applies if a variant's own limit is lower.
{ provider: "volcengine-ark", match: /kimi/i, maxOutputCap: 32768, clampToModelMaxOutput: true },
// Strict OpenAI-compatible validators reject unknown assistant-message fields.
// Clients that talk to reasoning models (e.g. Hermes) echo the prior turn's
// reasoning back on every assistant message; Groq answers 400 and Mistral 422
// ("extra_forbidden") on it, which knocks these providers out of every
// multi-turn combo. Providers that *require* the field (DeepSeek, Kimi) are
// handled by reasoningContentInjector and are not listed here.
{ provider: "groq", dropMessageFields: ["reasoning_content", "reasoning", "reasoning_details"] },
{ provider: "mistral", dropMessageFields: ["reasoning_content", "reasoning", "reasoning_details"] },
{ provider: "cerebras", dropMessageFields: ["reasoning_content", "reasoning", "reasoning_details"] },
];
// Test a rule's match (regex or predicate) against the model id.
@@ -47,6 +56,15 @@ export function stripUnsupportedParams(provider, model, body) {
for (const key of rule.drop || []) {
if (body[key] !== undefined) delete body[key];
}
// Per-message field drop (assistant turns only — that is where clients replay reasoning).
if (Array.isArray(rule.dropMessageFields) && Array.isArray(body.messages)) {
for (const msg of body.messages) {
if (!msg || msg.role !== "assistant") continue;
for (const key of rule.dropMessageFields) {
if (msg[key] !== undefined) delete msg[key];
}
}
}
// CF Workers AI oneOf root schema only accepts content as plain string (#1926)
if (rule.flattenContent && Array.isArray(body.messages)) {
for (const msg of body.messages) {