Struct PawanAgent 

pub struct PawanAgent { /* private fields */ }

The main Pawan agent — handles conversation, tool calling, and self-healing.

This struct represents the core Pawan agent that handles:

• Conversation history management
• Tool calling with the LLM via pluggable backends
• Streaming responses
• Multiple LLM backends (NVIDIA API, Ollama, OpenAI)
• Context management and token counting
• Integration with Eruka for 3-tier memory injection

Implementations

impl PawanAgent

pub async fn preflight_check(&self) -> Result<()>

Pre-flight health check: verify the LLM backend is reachable before starting work. Sends a minimal “ping” message. Returns Ok(()) if the model responds.

impl PawanAgent

pub fn new(config: PawanConfig, workspace_root: PathBuf) -> Self

Create a new PawanAgent with auto-selected backend

pub fn with_tools(self, tools: ToolRegistry) -> Self

Create with a specific tool registry

pub fn tools_mut(&mut self) -> &mut ToolRegistry

Get mutable access to the tool registry (for registering MCP tools)

pub fn with_backend(self, backend: Box<dyn LlmBackend>) -> Self

Create with a custom backend

pub fn history(&self) -> &[Message]

Get the current conversation history

pub fn save_session(&self) -> Result<String>

Save current conversation as a session, returns session ID

pub fn resume_session(&mut self, session_id: &str) -> Result<()>

Resume a saved session by ID

pub async fn archive_to_eruka(&self) -> Result<()>

Archive the current conversation to Eruka’s context store. Safe to call from any async context; returns Ok even when eruka is disabled or unreachable so callers can fire-and-forget after save_session().

pub fn config(&self) -> &PawanConfig

Get the configuration

pub fn clear_history(&mut self)

Clear the conversation history

pub fn add_message(&mut self, message: Message)

Add a message to history

pub fn switch_model(&mut self, model: &str)

Switch the LLM model at runtime. Recreates the backend with the new model.

pub fn model_name(&self) -> &str

Get the current model name

pub fn get_tool_definitions(&self) -> Vec<ToolDefinition>

Get tool definitions for the LLM

pub async fn execute(&mut self, user_prompt: &str) -> Result<AgentResponse>

Execute a single prompt with tool calling support

pub async fn execute_with_callbacks( &mut self, user_prompt: &str, on_token: Option<TokenCallback>, on_tool: Option<ToolCallback>, on_tool_start: Option<ToolStartCallback>, ) -> Result<AgentResponse>

Execute with optional callbacks for streaming

pub async fn execute_with_all_callbacks( &mut self, user_prompt: &str, on_token: Option<TokenCallback>, on_tool: Option<ToolCallback>, on_tool_start: Option<ToolStartCallback>, on_permission: Option<PermissionCallback>, ) -> Result<AgentResponse>

Execute with all callbacks, including permission prompt.

pub async fn heal(&mut self) -> Result<AgentResponse>

Execute a healing task with real diagnostics

pub async fn heal_with_retries( &mut self, max_attempts: usize, ) -> Result<AgentResponse>

Execute healing with retries — calls heal(), checks for remaining errors, retries if needed.

Two-stage gate: Stage 1 — cargo check: must produce zero errors before proceeding. Stage 2 — healing.verify_cmd (optional): a user-supplied shell command (e.g. cargo test --workspace). If it exits non-zero the loop continues so the LLM can address the reported failures.

Anti-thrash guard: each Stage-1 error is fingerprinted (kind + code + message prefix). If the same fingerprint survives max_attempts consecutive rounds unchanged the loop halts rather than spinning indefinitely on an error the LLM cannot fix.

pub async fn task(&mut self, task_description: &str) -> Result<AgentResponse>

Execute a task with a specific prompt

pub async fn generate_commit_message(&mut self) -> Result<String>

Generate a commit message for current changes