Struct PawanConfig 

pub struct PawanConfig {

    pub config_version: u32,
    pub provider: LlmProvider,
    pub model: String,
    pub base_url: Option<String>,
    pub dry_run: bool,
    pub auto_backup: bool,
    pub require_git_clean: bool,
    pub bash_timeout_secs: u64,
    pub tool_call_idle_timeout_secs: u64,
    pub max_file_size_kb: usize,
    pub max_tool_iterations: usize,
    pub max_context_tokens: usize,
    pub system_prompt: Option<String>,
    pub temperature: f32,
    pub top_p: f32,
    pub max_tokens: usize,
    pub thinking_budget: usize,
    pub max_retries: usize,
    pub fallback_models: Vec<String>,
    pub max_result_chars: usize,
    pub reasoning_mode: bool,
    pub healing: HealingConfig,
    pub targets: HashMap<String, TargetConfig>,
    pub tui: TuiConfig,
    pub mcp: HashMap<String, McpServerEntry>,
    pub permissions: HashMap<String, ToolPermission>,
    pub cloud: Option<CloudConfig>,
    pub models: ModelRouting,
    pub eruka: ErukaConfig,
    pub use_ares_backend: bool,
    pub use_coordinator: bool,
    pub skills_repo: Option<PathBuf>,
    pub local_first: bool,
    pub local_endpoint: Option<String>,
}

Main configuration for Pawan

Fields

config_version: u32

Config version for migration tracking (default: 1)

provider: LlmProvider

LLM provider to use

model: String

LLM model to use for coding tasks

base_url: Option<String>

Override the API base URL (e.g. “http://localhost:8080/v1” for llama.cpp). Takes priority over OPENAI_API_URL / NVIDIA_API_URL env vars.

dry_run: bool

Enable dry-run mode (show changes without applying)

auto_backup: bool

Create backups before editing files

require_git_clean: bool

Require clean git working directory

bash_timeout_secs: u64

Timeout for bash commands (seconds)

tool_call_idle_timeout_secs: u64

Timeout for tool calls that remain idle (seconds) Default: 300 (5 minutes)

max_file_size_kb: usize

Maximum file size to read (KB)

max_tool_iterations: usize

Maximum tool iterations per request

max_context_tokens: usize

Maximum context tokens before pruning

system_prompt: Option<String>

System prompt override

temperature: f32

Temperature for LLM responses

top_p: f32

Top-p sampling parameter

max_tokens: usize

Maximum tokens in response

thinking_budget: usize

Maximum tokens allowed for reasoning/thinking (0 = unlimited). When set, pawan tracks thinking vs action token usage per call. If thinking exceeds this budget, a warning is logged.

max_retries: usize

Maximum retries for LLM API calls (429 or 5xx errors)

fallback_models: Vec<String>

Fallback models to try when primary model fails

max_result_chars: usize

Maximum characters in tool result before truncation

reasoning_mode: bool

Enable reasoning/thinking mode (for DeepSeek/Nemotron models)

healing: HealingConfig

Healing configuration

targets: HashMap<String, TargetConfig>

Target projects

tui: TuiConfig

TUI configuration

mcp: HashMap<String, McpServerEntry>

MCP server configurations

permissions: HashMap<String, ToolPermission>

Tool permission overrides (tool_name -> permission)

cloud: Option<CloudConfig>

Cloud fallback: when primary model fails, fall back to cloud provider. Enables hybrid local+cloud routing.

models: ModelRouting

Task-type model routing: use different models for different task categories. If not set, all tasks use the primary model.

eruka: ErukaConfig

Eruka context engine integration (3-tier memory injection)

use_ares_backend: bool

Use ares-server’s LLMClient + ToolCoordinator primitives instead of pawan’s built-in OpenAI-compatible backend. Requires the ares feature flag to be enabled when building pawan-core. When true, pawan delegates LLM generation to ares which provides connection pooling, loop detection, and unified multi-provider support. Default: false (backwards compatible).

use_coordinator: bool

Use the ToolCoordinator for tool-calling loops instead of pawan’s built-in implementation. When true, delegates tool execution to the coordinator which provides parallel execution, timeouts, and consistent error handling. Default: false (backwards compatible).

skills_repo: Option<PathBuf>

Optional path to a skills repository (directory of SKILL.md files).

Mirrors the dstack pattern: public repo + private skills linked by config. When set, pawan discovers all SKILL.md files under this path at runtime via thulp-skill-files SkillLoader. Useful for linking private skill libraries without embedding them in the public repo.

Resolution order:

1. PAWAN_SKILLS_REPO environment variable
2. skills_repo field in pawan.toml
3. ~/.config/pawan/skills if it exists
4. None (no skill discovery beyond the project SKILL.md)

local_first: bool

Prefer local inference over cloud when a local model server is reachable. Before each session pawan probes local_endpoint (or the Ollama default http://localhost:11434/v1) with a 100 ms TCP timeout. If the server responds, it is used instead of the configured cloud provider. If the server is unreachable the configured provider is used as normal. Default: false (always use configured provider).

local_endpoint: Option<String>

Local inference endpoint URL for the local_first probe. Must be an OpenAI-compatible /v1 endpoint (Ollama, llama.cpp, LM Studio, …). Defaults to http://localhost:11434/v1 (Ollama) when not set.

Implementations

impl PawanConfig

pub fn load(path: Option<&PathBuf>) -> Result<Self>

Load configuration from file

pub fn apply_env_overrides(&mut self)

Apply environment variable overrides (PAWAN_MODEL, PAWAN_PROVIDER, etc.)

pub fn get_target(&self, name: &str) -> Option<&TargetConfig>

Get target by name

pub fn get_system_prompt(&self) -> String

Get the system prompt, with optional project context injection. Loads from PAWAN.md, AGENTS.md, CLAUDE.md, or .pawan/context.md.

pub fn resolve_skills_repo(&self) -> Option<PathBuf>

Resolve the effective skills repository path using the dstack pattern: env var > config field > default ~/.config/pawan/skills > None.

Returns Some(path) only if the resolved path exists as a directory. This allows public pawan repos to link to private skill libraries without embedding them — the path is configured per-machine.

pub fn auto_discover_mcp_servers(&mut self) -> Vec<String>

Auto-discover MCP server binaries in PATH and register any that aren’t already configured. Returns the names of newly-discovered servers.

Supported auto-discovery targets:

• eruka-mcp — Eruka context memory (anti-hallucination knowledge store)
• daedra — web search across 9 backends
• deagle-mcp — deagle code intelligence graph

Existing entries in the mcp HashMap are never overwritten. This makes the auto-discovery idempotent and safe to call at every agent startup.

pub fn discover_skills_from_repo(&self) -> Vec<(String, String, PathBuf)>

Discover all SKILL.md files in the configured skills repository using thulp-skill-files SkillLoader.

Returns a list of (skill_name, description, file_path) tuples. The caller is responsible for deciding which skills to inject into the system prompt or present to the user.

The skills repository is never compiled into the pawan binary — this enables the “public repo links to private skills via config” pattern used by dstack for dirmacs/skills.

pub fn use_thinking_mode(&self) -> bool

Check if thinking mode should be enabled. Applicable to DeepSeek, Gemma-4, GLM, Qwen, and Mistral Small 4+ models on NIM.

Trait Implementations

impl Clone for PawanConfig

fn clone(&self) -> PawanConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for PawanConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for PawanConfig

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for PawanConfig

where PawanConfig: Default,

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for PawanConfig

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.