From 09a4d3f15a40453c56ba1a1d37967ce4d2513430 Mon Sep 17 00:00:00 2001
From: Patrick Patel <patrickpatel96@gmail.com>
Date: Tue, 24 Feb 2026 22:21:05 -0600
Subject: [PATCH] fix: Resolve Claude-compatible provider for backlog plan when
 client sends model (#809)

When the Plan dialog sends a model (e.g. MiniMax-M2.1 from phase
settings), the server now:

- Calls getProviderByModelId() so the correct provider config
  (baseUrl, credentials) is used for backlog plan generation.
- Falls back to getPhaseModelWithOverrides('backlogPlanningModel')
  when model lookup finds no provider, so the phase's provider is
  used when the model matches.
- Uses a plain system prompt instead of the claude_code preset when
  a Claude-compatible provider is set; the preset is for native
  Claude CLI and can break requests to MiniMax/GLM APIs.

Previously the request was sent to the default Anthropic endpoint
and/or used the preset, causing plan generation to fail for
MiniMax/GLM users.

Co-authored-by: Cursor <cursoragent@cursor.com>
---
 .../src/routes/backlog-plan/generate-plan.ts  | 36 ++++++++++++++++++-
 1 file changed, 35 insertions(+), 1 deletion(-)

diff --git a/apps/server/src/routes/backlog-plan/generate-plan.ts b/apps/server/src/routes/backlog-plan/generate-plan.ts
index 2f47d293..4a1dc95a 100644
--- a/apps/server/src/routes/backlog-plan/generate-plan.ts
+++ b/apps/server/src/routes/backlog-plan/generate-plan.ts
@@ -35,6 +35,7 @@ import {
   getUseClaudeCodeSystemPromptSetting,
   getPromptCustomization,
   getPhaseModelWithOverrides,
+  getProviderByModelId,
 } from '../../lib/settings-helpers.js';
 
 /** Maximum number of retry attempts for transient CLI failures */
@@ -231,6 +232,35 @@ export async function generateBacklogPlan(
       effectiveModel = resolved.model;
       thinkingLevel = resolved.thinkingLevel;
       credentials = await settingsService?.getCredentials();
+      // Resolve Claude-compatible provider when client sends a model (e.g. MiniMax, GLM)
+      if (settingsService) {
+        const providerResult = await getProviderByModelId(
+          effectiveModel,
+          settingsService,
+          '[BacklogPlan]'
+        );
+        if (providerResult.provider) {
+          claudeCompatibleProvider = providerResult.provider;
+          if (providerResult.credentials) {
+            credentials = providerResult.credentials;
+          }
+        }
+        // Fallback: use phase settings provider if model lookup found nothing (e.g. model
+        // string format differs from provider's model id, but backlog planning phase has providerId).
+        if (!claudeCompatibleProvider) {
+          const phaseResult = await getPhaseModelWithOverrides(
+            'backlogPlanningModel',
+            settingsService,
+            projectPath,
+            '[BacklogPlan]'
+          );
+          const phaseResolved = resolvePhaseModel(phaseResult.phaseModel);
+          if (phaseResult.provider && phaseResolved.model === effectiveModel) {
+            claudeCompatibleProvider = phaseResult.provider;
+            credentials = phaseResult.credentials ?? credentials;
+          }
+        }
+      }
     } else if (settingsService) {
       // Use settings-based model with provider info
       const phaseResult = await getPhaseModelWithOverrides(
@@ -291,8 +321,12 @@ CRITICAL INSTRUCTIONS:
 
 ${userPrompt}`;
       finalSystemPrompt = undefined; // System prompt is now embedded in the user prompt
+    } else if (claudeCompatibleProvider) {
+      // Claude-compatible providers (MiniMax, GLM, etc.) use a plain API; do not use
+      // the claude_code preset (which is for Claude CLI/subprocess and can break the request).
+      finalSystemPrompt = systemPrompt;
     } else if (useClaudeCodeSystemPrompt) {
-      // Use claude_code preset for Claude models so the SDK subprocess
+      // Use claude_code preset for native Claude so the SDK subprocess
       // authenticates via CLI OAuth or API key the same way all other SDK calls do.
       finalSystemPrompt = {
         type: 'preset',