glmproxy 2.6.0 → 2.6.1

package/README.md

@@ -158,8 +158,8 @@ AutoClaw handles authentication automatically. When the cloud path fails, reques
 
 | ID | Name | Context | Max Output | Notes |
 |----|------|---------|------------|-------|
-| `zai_auto` | Auto | 1M | 393K | Routes to AutoClaw's optimal model |
-| `zaicoding_glm-5.3` | GLM-5.3 | 1M | 307K | Latest GLM coding model |
+| `zai_auto` | Auto | 1M | 131K | Routes to AutoClaw's optimal model (GLM-5.3-Flash today) |
+| `zaicoding_glm-5.3` | GLM-5.3 | 1M | 131K | Latest GLM coding model |
 | `zai_glm-5-turbo` | GLM-5-Turbo | 200K | 131K | Zhipu AI GLM-5 Turbo |
 | `zai_glm-5.3-flash` | GLM-5.3-Flash ("OX-alpha") | 1M | 131K | Now a regular catalog model, served straight through the cloud path |
 | `tdpsk_deepseek-v4-flash-202605` | Deepseek-V4-Flash | 1M | 393K | Fast DeepSeek model |
@@ -386,6 +386,7 @@ TRUSTED_PROXIES=127.0.0.1 glmproxy --host 0.0.0.0
 - On a 401, the proxy invalidates its cached token and you can retry immediately
 - Upstream 400 `"invalid request"` gets one retry after a 2s delay (a known upstream hiccup). Quota/plan errors are never retried
 - The cloud upstream requires AutoClaw's app system-prompt banner in every request. The proxy injects it automatically and never duplicates it. In practice it doesn't change much since your harness's own system prompt overrides it anyway
+- Max output is clamped to each model's real upstream cap (131K for every GLM model, 393K for DeepSeek). AutoClaw's runtime catalog overstates GLM-5.3's cap (307K), and asking the cloud for more than a model's real cap makes it **silently run a DeepSeek model instead and bill DeepSeek credits** — the proxy clamps so your `zai_glm-5.3` stays GLM-5.3
 - The token file is watched for changes, so AutoClaw can rotate auth mid-session without a restart
 - AutoClaw's client identity (app version, platform, channel) loads dynamically from its runtime file, same as the model catalog, so an AutoClaw app update is picked up without editing or restarting the proxy
 - The fallback model catalog lives in `lib/fallback-models.json` (override with `FALLBACK_MODELS_PATH`). The built-in list is only a last resort when AutoClaw's runtime file is unreadable

package/lib/core.js

@@ -192,7 +192,9 @@ export function readRuntimeModels(config) {
         id:            m.id,
         name:          m.name || m.id,
         contextWindow: m.contextWindow || 1_048_576,
-        maxTokens:     m.maxTokens     || 131_072,
+        // The runtime file overstates GLM-5.3's output cap (307200 vs the
+        // cloud's real 131072); never advertise more than the verified cap.
+        maxTokens:     Math.min(m.maxTokens || 131_072, OUTPUT_CAPS[m.id] ?? Infinity),
       }));
 
       if (models.length > 0) return { models, source: candidate };
@@ -227,6 +229,32 @@ export function loadModelCatalog(config) {
   return { MODELS: loadModelsFromRuntime(config) };
 }
 
+// ============================================================================
+// Real upstream output caps (probe-verified 2026-09-02)
+// ============================================================================
+// AutoClaw's runtime catalog OVERSTATES GLM-5.3's max output (307200) — the
+// cloud's real cap for every GLM model is 131072. Sending max_completion_tokens
+// above the cap makes the upstream silently substitute a deepseek model
+// (glm-5.3 → deepseek-v4-pro, flash/turbo/auto → deepseek-v4-flash) and bill
+// deepseek credits. Verified to the exact token: 131072 ok, 131073 flips.
+export const OUTPUT_CAPS = Object.freeze({
+  "zaicoding_glm-5.3":               131_072,
+  "zai_glm-5.3-flash":               131_072,
+  "zai_glm-5-turbo":                 131_072,
+  "zai_auto":                        131_072,
+  "tdpsk_deepseek-v4-flash-202605":  393_216,
+  "tdpsk_deepseek-v4-pro-202606":    393_216,
+});
+
+// Clamp a requested max output to the model's real upstream cap. Unknown
+// models (not in OUTPUT_CAPS) pass through untouched.
+export function clampMaxOutput(modelId, value) {
+  if (!Number.isFinite(value)) return value;
+  const cap = OUTPUT_CAPS[modelId];
+  if (!cap) return value;
+  return Math.min(value, cap);
+}
+
 // ============================================================================
 // Logger
 // ============================================================================
@@ -1177,8 +1205,12 @@ function buildSanitizedBody(openAIBody, upstreamModelId) {
   };
   if (typeof openAIBody.temperature === "number") sanitized.temperature = openAIBody.temperature;
   if (typeof openAIBody.top_p === "number") sanitized.top_p = openAIBody.top_p;
-  if (typeof openAIBody.max_tokens === "number") sanitized.max_tokens = openAIBody.max_tokens;
-  if (typeof openAIBody.max_completion_tokens === "number") sanitized.max_tokens = openAIBody.max_completion_tokens;
+  // Clamp max output to the model's REAL upstream cap. Exceeding it makes the
+  // cloud silently swap in a deepseek model (probe-verified 2026-09-02) and
+  // bill deepseek credits — the harness sends 393216 for everything, which
+  // every GLM model (real cap 131072) trips.
+  if (typeof openAIBody.max_tokens === "number") sanitized.max_tokens = clampMaxOutput(upstreamModelId, openAIBody.max_tokens);
+  if (typeof openAIBody.max_completion_tokens === "number") sanitized.max_tokens = clampMaxOutput(upstreamModelId, openAIBody.max_completion_tokens);
   if (openAIBody.stop !== undefined) sanitized.stop = openAIBody.stop;
   if (Array.isArray(openAIBody.tools) && openAIBody.tools.length > 0) sanitized.tools = openAIBody.tools;
   if (openAIBody.tool_choice !== undefined) sanitized.tool_choice = openAIBody.tool_choice;

package/lib/fallback-models.json

@@ -1,7 +1,7 @@
 {
   "models": [
-    { "id": "zai_auto",                       "name": "Auto",              "contextWindow": 1048576, "maxTokens": 393216 },
-    { "id": "zaicoding_glm-5.3",              "name": "GLM-5.3",           "contextWindow": 1048576, "maxTokens": 307200 },
+    { "id": "zai_auto",                       "name": "Auto",              "contextWindow": 1048576, "maxTokens": 131072 },
+    { "id": "zaicoding_glm-5.3",              "name": "GLM-5.3",           "contextWindow": 1048576, "maxTokens": 131072 },
     { "id": "zai_glm-5-turbo",                "name": "GLM-5-Turbo",       "contextWindow": 204800,   "maxTokens": 131072 },
     { "id": "zai_glm-5.3-flash",              "name": "GLM-5.3-Flash",     "contextWindow": 1048576, "maxTokens": 131072 },
     { "id": "tdpsk_deepseek-v4-flash-202605", "name": "Deepseek-V4-Flash", "contextWindow": 1048576, "maxTokens": 393216 },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "glmproxy",
-  "version": "2.6.0",
+  "version": "2.6.1",
   "description": "Local OpenAI- & Anthropic-compatible proxy for AutoClaw's Zhipu GLM-5.3 / GLM-5 / DeepSeek models",
   "main": "openai.js",
   "type": "module",