Rig v0.30 now available!

[joshua-mo-143]

Hey everyone! Happy to announce that Rig v0.30 is now available to try.

Encoding agentic loop behaviour into hooks

One of the issues that the current hook design has is that you can't actually "return" an early cancellation - you can only send CancelSignal::cancel() and then manually cancel. This version we have instead made a change that lets you return a HookAction, specifying whether you would either like to continue the loop or terminate early. In the case of tool calls, you can also choose to skip them.

A short snippet to illustrate the point:

pub trait PromptHook<M>: Clone + WasmCompatSend + WasmCompatSync
where
    M: CompletionModel,
{
    /// Called before the prompt is sent to the model
    fn on_completion_call(
        &self,
        _prompt: &Message,
        _history: &[Message],
    ) -> impl Future<Output = HookAction> + WasmCompatSend {
        // by returning HookAction::cont()
        // we can explicitly tell the loop to continue
        async { HookAction::cont() }
    }
    
    // .. normally more methods here, this is just a snippet
}

This is a step forward in how we think about loops internally as we want to move towards patterns that both properly encode the behaviour of the agentic loop and how to best add things like agent state/context and manually mutating the loop state without having to add things like &mut variables which make it difficult to use such things.

Rejecting tool calls in rig::agent::Agent

You can now skip tool calls! Instead of simply cancelling the loop, you can send a message back to the LLM telling it that the tool call failed with a reason, rather than whatever the result of the tool call should normally be.

impl<M: CompletionModel> StreamingPromptHook<M> for PermissionHook {
    async fn on_tool_call(
        &self,
        tool_name: &str,
        _tool_call_id: Option<String>,
        _internal_call_id: &str,
        _args: &str,
    ) -> ToolCallHookAction {
        let count = self.call_count.fetch_add(1, Ordering::SeqCst);

        if count == 0 {
            ToolCallHookAction::Skip {
                reason: format!(
                    "Tool '{}' is currently unavailable. \
                     Please use 'read_file_tail' instead to read the file.",
                    tool_name
                ),
            }
        } else {
            ToolCallHookAction::Continue
        }
    }
}

Other features

• OpenAI reasoning xhigh variant now supported
• Image tool results now supported for Gemini
• xAI integration now uses xAI Responses API
• implemented VectorStoreIndexDyn for MongoDB and Milvus (meaning that you can use them as dynamic docstores for agents)
• Added an internal tool call ID to every tool call. This can be viewed from the on_tool_call/on_tool_result hook method and makes it much more convenient to reason about tool call IDs. (edit: Additionally, some enums related to message contents may have an internal tool call ID added)

Bug fixes

• Better error handling in Azure OpenAI module
• Fixed an issue with tool server not running tools concurrently
• Fixed an issue with OpenAI Responses API requiring reasoning when using tool calls
• Fixed an issue with Deepseek's deepseek-reasoner model being broken due to reasoning not being added back to the conversation history
• Fixed an issue with the Gemini API schema shape
• Fixed an issue where tool calls in ollama streaming get ignored

Refactoring

• Refactored max_depth to max_turns().
• OpenAI Responses API now creates a system message instead of instructions, as several Responses API compliant services do not use it (xAI, LM Studio, etc...).

Miscellaneous

• Updated ollama provider module docs to match the example
• Added default type parameter T = reqwest::Client to ollama's EmbeddingModel for consistency with other providers

BREAKING

• Prompt hooks have been significantly changed. Please refer to the above notes (or check docs.rs) for how to migrate to the correct type signature.
• Max depth has now been renamed to max_turns(). PromptRequest::multi_turn() has now been renamed to PromptRequest::max_turns() and the agentbuilder equivalent is default_max_turns().

---

Changelog PR: #1288

Thanks for reading! We build rig, rig builds us.