Struct CodingAgentConfig 

pub struct CodingAgentConfig {

    pub model: String,
    pub system: Option<String>,
    pub base_system: Option<String>,
    pub workspace: PathBuf,
    pub skills: Option<Vec<Skill>>,
    pub disabled_skills: BTreeSet<String>,
    pub max_turn_steps: u32,
    pub thinking: ThinkingEffort,
    pub tools: Option<Arc<ShellTools>>,
    pub extra_resolver: Option<Arc<dyn ToolResolver>>,
    pub transcript: Option<TranscriptSettings>,
    pub summary_worker: Option<Arc<dyn Summarizer>>,
    pub shell_env: BTreeMap<String, String>,
    pub steering: Option<Arc<dyn Steering>>,
    pub mcp_servers: Vec<McpServerLine>,
    pub disabled_mcp_servers: BTreeSet<String>,
    pub environment: Option<String>,
    pub user_preferences: Option<String>,
    pub vision: Option<String>,
}

Everything a CodingAgent needs besides the provider.

Fields

model: String

Model identifier, passed through to ChatParams.

system: Option<String>

System prompt stamped into every request. None uses the base prompt — base_system when set, the built-in default otherwise; setting it replaces everything, it does not append.

base_system: Option<String>

The base prompt from the user’s SYSTEM.md, a Jinja template rendered by render_system — a template error fails the agent’s construction. None uses the built-in default_system.

workspace: PathBuf

Root all tools are scoped to; paths that escape it are rejected in-band to the model.

skills: Option<Vec<Skill>>

Skills already collected elsewhere, for the prompt’s skills list. None discovers from the filesystem — the global config directory and the workspace’s .rhema/skills. A frontend with no filesystem to discover from (the cloud sandbox) seeds the skill files itself and passes the list; the model reads bodies through the shell, so each path must resolve inside the workspace.

disabled_skills: BTreeSet<String>

Skill folder names to drop from whichever list the prompt carries — the persisted off switch a frontend reads out of its config. The filter runs after collection and after discovery alike, so both supply paths honor it; an empty set (the default) changes nothing.

max_turn_steps: u32

Cap on steps per turn; a turn still requesting tools past it fails with AgentError::MaxIterations.

thinking: ThinkingEffort

How much reasoning the model should do; providers without a knob ignore it, and ThinkingEffort::Default sends nothing.

tools: Option<Arc<ShellTools>>

The shell resolver the agent’s tools run on. None builds one over workspace — the default. A frontend that shares its shell with an extension (the subagents extension reads the turn’s approval hook off it) brings its own.

extra_resolver: Option<Arc<dyn ToolResolver>>

An extension resolver merged over the shell’s tools through Overlay: its specs join the prompt’s tool list and its tools resolve by name, shadowing a built-in of the same name. None runs the built-ins alone.

transcript: Option<TranscriptSettings>

The transcript: a bounded request view over the unbounded session, plus the zoom tool to read back through it. On by default and below the budget a pass-through, so small sessions are byte-identical to a transcript-less run. None is the off switch: requests carry the whole history. See specs/rhema-transcript.md.

summary_worker: Option<Arc<dyn Summarizer>>

The compaction’s model fill: a cheap worker the transcript’s stepped compaction runs at flushes, rewriting each new block section’s extractive stand-in into a dense markdown block. The worker never blocks or fails a request — a failed call leaves the stand-in in place. None keeps the compaction extractive: still stepped, still cache-stable, only less dense. See specs/rhema-transcript.md.

shell_env: BTreeMap<String, String>

Extra environment the shell tool’s commands run with, layered over the inherited one — how a frontend provisions its agent’s client tools (a daemon URL, a scoped token, a PATH addition) without touching the process environment.

steering: Option<Arc<dyn Steering>>

A steering source the turn loop drains: one User-role message per drain lands in the history at a step boundary, so work that finished outside the turn (a background subagent) is announced to the model. None — the default — runs the turn exactly as before: the loop is the provided run_turn with no injection.

mcp_servers: Vec<McpServerLine>

The wired MCP servers, as the prompt’s mcp section carries them — the lines a frontend’s wire step produced (spec: specs/rhema-mcp.md §10). Empty renders no section: a frontend that wires nothing advertises nothing.

disabled_mcp_servers: BTreeSet<String>

MCP server names to drop from the prompt’s section — the persisted off switch a frontend reads out of its stored settings, the disabled_skills mechanism. Applied at stamp time over mcp_servers by config name; the wire step filters the same set before anything spawns.

environment: Option<String>

The environment section rendered into the system prompt: machine facts probed where commands run (see the environment module). None renders the section empty. A verbatim system prompt ignores it, like every template section.

user_preferences: Option<String>

The user-preferences section rendered into the system prompt — the contents of the user’s global USER.md, verbatim, a frontend that has one reads it and passes it here. None renders nothing.

vision: Option<String>

The prompt section documenting the wired vision macro — the see bash function’s documentation, produced by the frontend’s provisioning and passed verbatim. None renders no section: a frontend that wired nothing advertises nothing, the same rule the mcp_servers section follows.

Implementations

impl CodingAgentConfig

pub const DEFAULT_MAX_TURN_STEPS: u32 = 2048

Default for max_turn_steps. Long autonomous sessions — builds, rebases, multi-package sweeps — routinely run hundreds of steps on cheap models under a watching eye; the cap exists to end a runaway loop, not to ration the model. --max-steps is the tighter leash for unattended turns on expensive models.

pub fn new(model: impl Into<String>, workspace: impl Into<PathBuf>) -> Self

Config naming model, rooted at workspace, defaults elsewhere.

Trait Implementations

impl Clone for CodingAgentConfig

fn clone(&self) -> Self

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for CodingAgentConfig

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

Formats the value using the given formatter.