feat: add xAI Grok integration with its current features

[askiiRobotics]

Hey folks! Just a heads up - I mainly tested this with xAI/Grok since
that's what I wanted to try out, and it felt solid enough to use as a
foundation. Haven't thoroughly tested with other providers yet, so if you
notice any issues or regressions with OpenAI/Anthropic/etc., please feel
free to jump in and fix this PR! 🙏

• Complete xAI provider with the current Grok models (2/3/4 series)
• Deferred completion support for long-running requests
• Concurrent request handling for improved performance
• Web search tool integration for real-time information
• Image generation and vision capabilities
• Advanced reasoning support with model-specific validation
• Test coverage for (almost) all new xAI capabilities

Configuration enhancements:

• Runtime options support via CLI flags (--deferred, --deferred-timeout)
• Provider-level deferred completion configuration
• Auto-enable rules for smart feature activation
• Model-specific reasoning effort validation (xAI: low/high only)
• Enhanced agent configuration with override capabilities

Documentation and schema updates:

• README updates with xAI model details
• Updated configuration schema with new provider options
• Updated CLI help text and examples

[Tycale]

Thanks for this PR, I wanted to try Grok4 with OpenCode but got this message:

failed to process events: POST "https://api.x.ai/v1/chat/completions": 400 Bad Request "Invalid request content: Invalid function schema."

[askiiRobotics]

Thanks for this PR, I wanted to try Grok4 with OpenCode but got this message:

failed to process events: POST "https://api.x.ai/v1/chat/completions": 400 Bad Request "Invalid request content: Invalid function schema."

Dear Tycale,
Yes, you were right — that issue did exist. The code definitely needed some improvements, and the paths for models with thinking capabilities but no reasoning (e.g., the Grok-4 case) were incorrect.

This has now been fixed — thanks for pointing it out!

Heads-up: Still includes many untested features and plenty of room for improvement beyond basic chat. Core functionality is there, but it needs proper integration with UI and other components.

[Tycale]

Thanks, I will ASAP !

[Tycale]

hmm, it's kind of working, I got some responses but I have a lot of

failed to process events: reasoning request failed: failed to send request: Post "https://api.x.ai/v1/chat/completions": context deadline exceeded (Client.Timeout exceeded while awaiting headers)

(Grok4)

[Tycale]

Just noticed that switching back to o3 will produce the following error :

│                                                                                                                                                                                                            │
│Message:                                                                                                                                                                                                    │
│  POST "https://api.openai.com/v1/chat/completions": 400 Bad Request {                                                                                                                                      │
│      "message": "Invalid schema for function 'web_search': ['query'] is not of type 'object', 'boolean'.",                                                                                                 │
│      "type": "invalid_request_error",                                                                                                                                                                      │
│      "param": "tools[11].function.parameters",                                                                                                                                                             │
│      "code": "invalid_function_parameters"                                                                                                                                                                 │
│    }

[askiiRobotics]

hmm, it's kind of working, I got some responses but I have a lot of

failed to process events: reasoning request failed: failed to send request: Post "https://api.x.ai/v1/chat/completions": context deadline exceeded (Client.Timeout exceeded while awaiting headers)

(Grok4)

Just noticed that switching back to o3 will produce the following error :

│                                                                                                                                                                                                            │
│Message:                                                                                                                                                                                                    │
│  POST "https://api.openai.com/v1/chat/completions": 400 Bad Request {                                                                                                                                      │
│      "message": "Invalid schema for function 'web_search': ['query'] is not of type 'object', 'boolean'.",                                                                                                 │
│      "type": "invalid_request_error",                                                                                                                                                                      │
│      "param": "tools[11].function.parameters",                                                                                                                                                             │
│      "code": "invalid_function_parameters"                                                                                                                                                                 │
│    }

Thanks for pointing this out — you're right on both counts.

• The Grok-4 issue (context deadline exceeded) was related to how reasoning was being handled for models with internal-only reasoning capabilities. This has now been addressed.
• The OpenAI error stemmed from tool schema incompatibility — particularly with the web_search tool being assigned to models that don’t support it.

I’ve just pushed a commit that should resolve both: feat: Add provider-based tool filtering for model compatibility

This should prevent the mismatches you were seeing and improve fallback behavior.

One caveat: I still have concerns about pricing accuracy for xAI — this will need another pass to verify.

Appreciate the testing and feedback — feel free to flag anything else that comes up!

[Tycale]

➜  opencode git:(askiiRobotics/main) go build -o opencode                                                                                                                                       
# github.com/opencode-ai/opencode/internal/llm/tools
internal/llm/tools/web_search.go:147:3: unknown field Providers in struct literal of type ToolInfo

Can be fixed with

➜  opencode git:(askiiRobotics/main) ✗ git diff                                                                                                                                                 
diff --git a/internal/llm/tools/tools.go b/internal/llm/tools/tools.go
index bf0f8df..542fc90 100644
--- a/internal/llm/tools/tools.go
+++ b/internal/llm/tools/tools.go
@@ -10,6 +10,7 @@ type ToolInfo struct {
        Description string
        Parameters  map[string]any
        Required    []string
+       Providers   []string
 }

 type toolResponseType string

I will be able to play a bit more now :)

Thanks for this PR, I really appreciate it ! 🙏🏻

[Tycale]

Grok 4, 32k token-per-minute cap is somehow annoying… C’mon, Elon, let me spend my money ! 😢

[kujtimiihoxha]

@askiiRobotics we have added you to the early access for what will become the next version of opencode would love yo get your feedback (you should be invited on discord).

@Tycale make sure to submit https://charm.sh/early-access/ if you want to try out the new version.

[Tycale]

@kujtimiihoxha I’d be keen to also try it out and share my feedback. I’m a heavy user of Claude Code and OpenCode, and I’ve also relied on aider.chat quite a bit in the past.

[ruslanvasylev]

Just merge it! YOLO!