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

fix(translator): zai thinkingFormat sends reasoning.effort object

Browse Source 
Z.ai / GLM-5.2+ require a top-level reasoning_effort (low/high/max)
alongside thinking:{type:"enabled"} to control reasoning depth; the zai
branch previously only set thinking and dropped reasoning_effort, so every
GLM-5.x request ran at the model default (max). Gate the field behind
GLM-5.2+ (thinkingEffortSupported in capabilities.js) since older GLM
(4.x, 5.0, 5.1, 5-turbo, 5v-turbo) do not read it, and map client levels
to the exact low/high/max values z.ai accepts.

extractThinking now checks reasoning_effort/reasoning.effort before the
thinking object so a client-supplied effort is not overwritten by
thinking:{type:"enabled"} mapping to mode:auto.

Fixes #2721
This commit is contained in:
qingyong
2026-08-28 12:32:26 +07:00
committed by decolua
parent cadef6c4ff
commit 56a40765e9
3 changed files with 59 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
open-sse/translator/concerns/thinkingUnified.js
View File

@@ -58,6 +58,15 @@ export function extractThinking(body) {
return { mode: "level", level: e };
}
// OpenAI chat / Responses shape — check effort first (zai sends both thinking object and reasoning.effort)
const effort = body.reasoning_effort ?? (typeof body.reasoning === "object" ? body.reasoning?.effort : null);
if (typeof effort === "string" && effort) {
const e = effort.toLowerCase();
if (e === "none" || e === "off") return { mode: "none" };
if (e === "auto") return { mode: "auto" };
return { mode: "level", level: e };
}
// Claude shape
const t = body.thinking;
if (t && typeof t === "object") {
@@ -69,15 +78,6 @@ export function extractThinking(body) {
}
}
// OpenAI chat / Responses shape
const effort = body.reasoning_effort ?? (typeof body.reasoning === "object" ? body.reasoning?.effort : null);
if (typeof effort === "string" && effort) {
const e = effort.toLowerCase();
if (e === "none" || e === "off") return { mode: "none" };
if (e === "auto") return { mode: "auto" };
return { mode: "level", level: e };
}
// Gemini shape (top-level, generationConfig, or request envelope)
const tc = body.thinkingConfig || body.generationConfig?.thinkingConfig || body.request?.generationConfig?.thinkingConfig;
if (tc && typeof tc === "object") {
@@ -270,6 +270,18 @@ function applyFormat(fmt, body, cfg, caps, supportedLevels) {
// Z.ai ignores thinking.disabled → must use enable_thinking:false to turn off.
if (none && canDisable) { body.enable_thinking = false; delete body.thinking; break; }
body.thinking = { type: "enabled" };
// reasoning_effort is only read by z.ai from GLM-5.2 onward — older GLM ignores it
// (see thinkingEffortSupported in capabilities.js). Skip on unsupported models so we
// don't send a field the API doesn't recognize.
if (caps.thinkingEffortSupported) {
const zaiLvl = toLevel(eff);
// GLM-5.3 only accepts exactly low|high|max (anything else errors); GLM-5.2 accepts
// a wider set but z.ai maps low/medium->high and xhigh->max server-side anyway, so
// this 3-value mapping matches both.
body.reasoning_effort = (zaiLvl === "low" || zaiLvl === "minimal") ? "low"
: (zaiLvl === "high" || zaiLvl === "medium") ? "high"
: "max";
}
break;
}
case "qwen": {