feat: token compression at 80 % of max_context_tokens

README.md

@@ -29,6 +29,50 @@ the `shai` binary will be installed in `$HOME/.local/bin`
 
 ## Configure a provider and Run!
 
+### Configuration files
+
+Shai can be configured via **configuration files** written in JSON. By default, the configuration file is `auth.config` located in `~/.config/shai/`. The file defines the list of LLM providers, the selected provider, model, and tool call method.
+
+#### Example `.shai.config`
+```json
+{
+  "providers": [
+    {
+      "provider": "ovhcloud",
+      "env_vars": {
+        "OVH_BASE_URL": "https://gpt-oss-120b.endpoints.kepler.ai.cloud.ovh.net/api/openai_compat/v1"
+      },
+      "model": "gpt-oss-120b",
+      "tool_method": "FunctionCall",
+      "max_context_tokens": 8192
+    }
+  ],
+  "selected_provider": 0
+}
+```
+
+- **providers**: an array of provider definitions. Each provider can specify environment variables (`env_vars`), the model name, the tool call method (`FunctionCall` or `Chat`), and optionally `max_context_tokens` to limit the context size.
+- **selected_provider**: the index of the provider to use (starting at `0`).
+- **max_context_tokens** (optional, per provider): maximum number of tokens that can be sent in the context to the LLM. If omitted, the default for the model is used.
+
+You can create multiple configuration files for different agents (see the *Custom Agent* section). To use a specific configuration, place the file in `~/.config/shai/agents/` and run the agent by its filename (without the `.config` extension):
+```
+shai my_custom_agent
+```
+
+Shai will automatically load the configuration, set the required environment variables, and use the selected provider for all subsequent interactions.
+
+### Using the configuration
+
+- **Automatic loading**: If a `.shai.config` file is present in the current directory, Shai will load it automatically.
+- **Explicit loading**: Use the `--config <path>` flag to specify a custom configuration file:
+```
+shai --config ~/.config/shai/agents/example.config
+```
+
+The configuration system allows you to switch providers, models, or tool call methods without recompiling the binary.
+
+
 By default `shai` uses OVHcloud as an anonymous user meaning you will be rate limited! If you want to sign in with your account or select another provider, run:
 
 ```

shai-cli/src/headless/app.rs

@@ -96,11 +96,15 @@ impl AppHeadless {
             .run().await;
 
         match result {
-            Ok(AgentResult { success, message, trace: agent_trace }) => {
+            Ok(AgentResult { success, message, trace: full_trace, compressed_trace }) => {
                 if trace {
-                    println!("{}", serde_json::to_string_pretty(&agent_trace)?);
+                    #[cfg(debug_assertions)]
+                    {
+                        println!("{}", serde_json::to_string_pretty(&compressed_trace)?);
+                    }
+                    println!("{}", serde_json::to_string_pretty(&full_trace)?);
                 } else {
-                    if let Some(message) = agent_trace.last() {
+                    if let Some(message) = full_trace.last() {
                         match message {
                             ChatMessage::Assistant { content: Some(ChatMessageContent::Text(content)), .. } => {
                                 println!("{}",content);

shai-cli/src/headless/tools.rs

@@ -39,7 +39,7 @@ impl ToolName {
             ToolName::Bash => "bash",
             ToolName::Edit => "edit",
             ToolName::Fetch => "fetch",
-            ToolName::Find => "find",
+            ToolName::Find => "search",
             ToolName::Ls => "ls",
             ToolName::MultiEdit => "multiedit",
             ToolName::Read => "read",
@@ -54,7 +54,7 @@ impl ToolName {
             "bash" => Some(ToolName::Bash),
             "edit" => Some(ToolName::Edit),
             "fetch" => Some(ToolName::Fetch),
-            "find" => Some(ToolName::Find),
+            "search" => Some(ToolName::Find),
             "ls" => Some(ToolName::Ls),
             "multiedit" => Some(ToolName::MultiEdit),
             "read" => Some(ToolName::Read),

shai-cli/src/tui/app.rs

@@ -22,6 +22,7 @@ use shai_core::logging::LoggingConfig;
 use shai_core::runners::coder::coder::coder;
 use shai_core::tools::{ToolCall, ToolResult};
 use shai_llm::{LlmClient, ToolCallMethod};
+use shai_llm::tool::max_context::get_max_context;
 use ratatui::{
     layout::{Constraint, Direction, Layout},
     style::{Color, Style},
@@ -72,6 +73,8 @@ pub struct App<'a> {
 
     pub(crate) total_input_tokens: u32,
     pub(crate) total_output_tokens: u32,
+    pub(crate) current_tokens: usize,
+    pub(crate) max_context: usize,
 }
 
 
@@ -83,6 +86,7 @@ impl App<'_> {
             let config = AgentConfig::load(agent_name)?;
 
             println!("\x1b[2m░ agent {} - {} on {}\x1b[0m", agent_name, config.llm_provider.model, config.llm_provider.provider);
+            self.max_context = get_max_context(&config.llm_provider.model);
 
             // Create agent from config
             let agent_builder = AgentBuilder::from_config(config).await?;
@@ -91,7 +95,7 @@ impl App<'_> {
             // Use default coder agent
             let (llm, model) = ShaiConfig::get_llm().await?;
             println!("\x1b[2m░ {} on {}\x1b[0m", model, llm.provider().name());
-
+            self.max_context = get_max_context(&model);
             Box::new(coder(Arc::new(llm), model))
         };
 
@@ -158,6 +162,13 @@ impl App<'_> {
         if let AgentEvent::TokenUsage { input_tokens, output_tokens } = &event {
             self.total_input_tokens += input_tokens;
             self.total_output_tokens += output_tokens;
+            self.current_tokens += (input_tokens + output_tokens) as usize;
+        }
+        // Update current_tokens after context compression
+        if let AgentEvent::ContextCompressed { current_tokens, .. } = &event {
+            if let Some(ct) = current_tokens {
+                self.current_tokens = *ct as usize;
+            }
         }
 
         Ok(())
@@ -182,6 +193,8 @@ impl App<'_> {
             permission_queue: VecDeque::new(),
             total_input_tokens: 0,
             total_output_tokens: 0,
+            current_tokens: 0,
+            max_context: 0,
         }
     }
 
@@ -385,7 +398,7 @@ impl App<'_> {
             AppModalState::PermissionModal { widget } => widget.height(),
         }.max(5);
         let height = modal_height
-        + 1 
+        + 2 
         + self.running_tools.len() as u16;
 
         if let Some(ref mut terminal) = self.terminal {  
@@ -395,11 +408,13 @@ impl App<'_> {
             }
 
             terminal.draw(|frame| {                    
-                let [_, inprogress, modal] = Layout::vertical([
+                let [_, inprogress, modal, ctx_line] = Layout::vertical([
                     Constraint::Length(1), // padding
                     Constraint::Length(self.running_tools.len() as u16 + 1), // running tool (if any)
-                    Constraint::Length(modal_height)])                // input or modal
-                    .areas(frame.area()); 
+                    Constraint::Length(modal_height),
+                    Constraint::Length(1)
+                ]).areas(frame.area());
+
 
                 // draw running tool
                 if !self.running_tools.is_empty() {
@@ -418,6 +433,14 @@ impl App<'_> {
                         widget.draw(frame, modal)
                     }
                 }
+                // Render context usage line
+                if self.max_context > 0 {
+                    let used = self.current_tokens;
+                    let remaining = if used > self.max_context { 0 } else { self.max_context - used };
+                    let percent = (remaining as f64 / self.max_context as f64) * 100.0;
+                    let ctx_text = format!("Context: {:.1}% remaining", percent);
+                    frame.render_widget(Span::styled(ctx_text, Style::default().fg(Color::DarkGray)), ctx_line);
+                }
             })?;
         }
         Ok(())

shai-cli/src/tui/command.rs

@@ -10,6 +10,7 @@ impl App<'_> {
             (("/auth","select a provider"), vec![]),
             (("/tc","set the tool call method: [fc | fc2 | so]"), vec!["method"]),
             (("/tokens","display token usage (input/output)"), vec![]),
+            (("/compact","trigger context compaction"), vec![]),
         ])
         .into_iter()
         .map(|((cmd,desc),args)|((cmd.to_string(),desc.to_string()),args.into_iter().map(|s|s.to_string()).collect()))
@@ -65,6 +66,12 @@ impl App<'_> {
                 );
                 self.input.alert_msg(&msg, Duration::from_secs(5));
             }
+            "/compact" => {
+                if let Some(ref agent) = self.agent {
+                    let _ = agent.controller.trigger_context_compression().await;
+                    self.input.alert_msg("Context compression triggered", Duration::from_secs(2));
+                }
+            }
             _ => {
                 self.input.alert_msg("command unknown", Duration::from_secs(1));
             }

shai-cli/src/tui/helper.rs

@@ -15,14 +15,15 @@ impl HelpArea {
             "  Available Commands:",
             "  /exit                exit from the tui",
             "  /tc <method>         set tool call method: [auto | fc | fc2 | so]",
-            "  /tokens              display token usage"
+            "  /tokens              display token usage",
+            "  /compact             trigger context compaction"
         ].join("\n").to_string()
     }
 }
 
 impl HelpArea {
     pub fn height(&self) -> u16 {
-        8 // content (3 general help lines + 1 blank + 1 header + 3 command lines)
+        9 // content (3 general help lines + 1 blank + 1 header + 4 command lines)
     }
 
     pub fn draw(&self, f: &mut Frame, area: Rect) {