kodegen_tools_browser 0.10.11

KODEGEN.ᴀɪ: Memory-efficient, Blazing-Fast, MCP tools for code generation agents.
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70

use std::sync::Arc;
use tokio::sync::Mutex;
use tracing::warn;
use kodegen_mcp_client::KodegenClient;

use crate::agent::{ActionResult, AgentError, AgentOutput, AgentResult, prompts::{AgentMessagePrompt, SystemPrompt}};
use crate::utils::AgentState;

/// Shared agent state and processing logic (can be Arc-cloned)
pub(super) struct AgentInner {
    pub(super) task: String,
    pub(super) add_infos: String,
    pub(super) mcp_client: Arc<KodegenClient>,
    pub(super) system_prompt: SystemPrompt,
    pub(super) agent_prompt: AgentMessagePrompt,
    pub(super) max_actions_per_step: usize,
    pub(super) agent_state: Arc<Mutex<AgentState>>,
    pub(super) temperature: f64,
    pub(super) max_tokens: u64,
    pub(super) vision_timeout_secs: u64,
    pub(super) llm_timeout_secs: u64,
    /// Results from the previous step's actions - fed back to LLM for learning
    pub(super) previous_action_results: Mutex<Vec<ActionResult>>,
}

/// Core processing logic
impl AgentInner {
    /// Process a single agent step internally
    pub(super) async fn process_step(&self) -> AgentResult<AgentOutput> {
        // Check if stop requested
        let agent_state = self.agent_state.lock().await;
        if agent_state.is_stop_requested() {
            return Err(AgentError::Stopped);
        }
        drop(agent_state);

        // Get previous action results for feedback to LLM
        let previous_results = {
            let guard = self.previous_action_results.lock().await;
            guard.clone()
        };

        // Get current browser state (with screenshot)
        let mut browser_state = self.get_browser_state().await?;

        // Generate agent actions using CandleFluentAi LLM (with vision analysis if screenshot available)
        // Now includes previous action results so LLM can learn from failures
        let llm_response = self.generate_actions_with_llm(&mut browser_state, &previous_results).await?;

        // Execute actions via MCP hot path
        let (action_results, errors) = self.execute_actions(llm_response.action.clone()).await?;

        // Store action results for next step's feedback loop
        {
            let mut guard = self.previous_action_results.lock().await;
            *guard = action_results;
        }

        // Log errors if any
        if !errors.is_empty() {
            warn!("Action execution errors: {:?}", errors);
        }

        // Return output with LLM-generated state (no wasteful rebuilding!)
        Ok(AgentOutput {
            current_state: llm_response.current_state,
            action: llm_response.action,
        })
    }
}