feat: add Cerebras and Fireworks AI LLM providers (#2113)

* fix: improve local model provider robustness and UX

- Extract shared Docker URL rewriting and env conversion into BaseProvider
  to eliminate 4x duplicated code across Ollama and LMStudio
- Add error handling and 5s timeouts to all model-listing fetches so one
  unreachable provider doesn't block the entire model list
- Fix Ollama using createOllama() instead of mutating provider internals
- Fix LLMManager singleton ignoring env updates on subsequent requests
- Narrow cache key to only include provider-relevant env vars instead of
  the entire server environment
- Fix 'as any' casts in LMStudio and OpenAILike by using shared
  convertEnvToRecord helper
- Replace console.log/error with structured logger in OpenAILike
- Fix typo: filteredStaticModesl -> filteredStaticModels in manager
- Add connection status indicator (green/red dot) for local providers
  in the ModelSelector dropdown
- Show helpful "is X running?" message when local provider has no models

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* feat: add Cerebras LLM provider

- Add Cerebras provider with 8 models (Llama, GPT OSS, Qwen, ZAI GLM)
- Integrate @ai-sdk/cerebras@0.2.16 for compatibility
- Add CEREBRAS_API_KEY to environment configuration
- Register provider in LLMManager registry

Models included:
- llama3.1-8b, llama-3.3-70b
- gpt-oss-120b (reasoning)
- qwen-3-32b, qwen-3-235b variants
- zai-glm-4.6, zai-glm-4.7 (reasoning)

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

* feat: add Fireworks AI LLM provider

- Add Fireworks provider with 6 popular models
- Integrate @ai-sdk/fireworks@0.2.16 for compatibility
- Add FIREWORKS_API_KEY to environment configuration
- Register provider in LLMManager registry

Models included:
- Llama 3.1 variants (405B, 70B, 8B Instruct)
- DeepSeek R1 (reasoning model)
- Qwen 2.5 72B Instruct
- FireFunction V2

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

* feat: add coding-specific models to existing providers

Enhanced providers with state-of-the-art coding models:

**DeepSeek Provider:**
+ DeepSeek V3.2 (integrates thinking + tool-use)
+ DeepSeek V3.2-Speciale (high-compute variant, beats GPT-5)

**Fireworks Provider:**
+ Qwen3-Coder 480B (262K context, best for coding)
+ Qwen3-Coder 30B (fast coding specialist)

**Cerebras Provider:**
+ Qwen3-Coder 480B (2000 tokens/sec!)
- Removed deprecated models (qwen-3-32b, llama-3.3-70b)

Total new models: 4
Total coding models across all providers: 12+

Performance highlights:
- Qwen3-Coder: State-of-the-art coding performance
- DeepSeek V3.2: Integrates thinking directly into tool-use
- ZAI GLM 4.6: 73.8% SWE-bench score
- Ultra-fast inference: 2000 tok/s on Cerebras

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

* feat: add dynamic model discovery to providers

Implemented getDynamicModels() for automatic model discovery:

**DeepSeek Provider:**
- Fetches models from https://api.deepseek.com/models
- Automatically discovers new models as DeepSeek adds them
- Filters out static models to avoid duplicates

**Cerebras Provider:**
- Fetches models from https://api.cerebras.ai/v1/models
- Auto-discovers new Cerebras models
- Keeps UI up-to-date with latest offerings

**Fireworks Provider:**
- Fetches from https://api.fireworks.ai/v1/accounts/fireworks/models
- Includes context_length from API response
- Discovers new Qwen-Coder and other models automatically

**Moonshot Provider:**
- Fetches from https://api.moonshot.ai/v1/models
- OpenAI-compatible endpoint
- Auto-discovers new Kimi models

Benefits:
- ✅ No manual updates needed when providers add new models
- ✅ Users always have access to latest models
- ✅ Graceful fallback to static models if API fails
- ✅ 5-second timeout prevents hanging
- ✅ Caching system built into BaseProvider

Technical details:
- Uses BaseProvider's built-in caching system
- Cache invalidates when API keys change
- Failed API calls fallback to static models
- All endpoints have 5-second timeout protection

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

* feat: add Z.AI provider with GLM models and JWT authentication

Merged changes from PR #2069 to add Z.AI provider:
- Added GLM-4.6 (200K), GLM-4.5 (128K), and GLM-4.5 Flash models
- Implemented secure JWT token generation with HMAC-SHA256 signing
- Added dynamic model discovery from Z.AI API
- Included proper error handling and token validation
- GLM-4.6 achieves 73.8% on SWE-bench coding benchmarks

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

app/lib/modules/llm/manager.ts

@@ -9,7 +9,7 @@ export class LLMManager {
   private static _instance: LLMManager;
   private _providers: Map<string, BaseProvider> = new Map();
   private _modelList: ModelInfo[] = [];
-  private readonly _env: any = {};
+  private _env: Record<string, string> = {};
 
   private constructor(_env: Record<string, string>) {
     this._registerProvidersFromDirectory();
@@ -19,6 +19,9 @@ export class LLMManager {
   static getInstance(env: Record<string, string> = {}): LLMManager {
     if (!LLMManager._instance) {
       LLMManager._instance = new LLMManager(env);
+    } else if (Object.keys(env).length > 0) {
+      // Update env on subsequent calls so Cloudflare Workers get fresh bindings
+      LLMManager._instance._env = env;
     }
 
     return LLMManager._instance;
@@ -121,10 +124,10 @@ export class LLMManager {
     const staticModels = Array.from(this._providers.values()).flatMap((p) => p.staticModels || []);
     const dynamicModelsFlat = dynamicModels.flat();
     const dynamicModelKeys = dynamicModelsFlat.map((d) => `${d.name}-${d.provider}`);
-    const filteredStaticModesl = staticModels.filter((m) => !dynamicModelKeys.includes(`${m.name}-${m.provider}`));
+    const filteredStaticModels = staticModels.filter((m) => !dynamicModelKeys.includes(`${m.name}-${m.provider}`));
 
     // Combine static and dynamic models
-    const modelList = [...dynamicModelsFlat, ...filteredStaticModesl];
+    const modelList = [...dynamicModelsFlat, ...filteredStaticModels];
     modelList.sort((a, b) => a.name.localeCompare(b.name));
     this._modelList = modelList;