Struct ChatRequest 

pub struct ChatRequest {

    pub model_id: String,
    pub messages: Vec<ChatMessage>,
    pub system_prompt: String,
    pub instructions: Option<String>,
    pub reasoning: ReasoningLevel,
    pub temperature: f32,
    pub max_tokens: usize,
    pub tools: Vec<ToolDefinition>,
    pub ollama_num_ctx: Option<u32>,
    pub ollama_allow_ram_offload: Option<bool>,
    pub resolved_context_window: Option<usize>,
    pub resolved_max_output: Option<usize>,
    pub output_schema: Option<Value>,
    pub suppress_auto_compact: bool,
}

Inputs a model needs to generate a turn. Built by the reducer from Session + Settings + current MERMAID.md context. Pure data — no provider-specific knowledge here (that’s in providers::model::*::chat).

Fields

model_id: String

messages: Vec<ChatMessage>

system_prompt: String

instructions: Option<String>

MERMAID.md content to suffix onto the system prompt. None if no file was loaded for this project.

reasoning: ReasoningLevel

temperature: f32

max_tokens: usize

tools: Vec<ToolDefinition>

Tool definitions advertised to the model. Combination of the built-in tool set + any advertised MCP tools from McpState.

ollama_num_ctx: Option<u32>

Per-model Ollama num_ctx override (/context <n>/max), or None to auto-fit. The one provider-specific knob that rides on the request, the same way reasoning does — so a live /context change applies on the next turn without rebuilding the cached provider. Ignored by non-Ollama providers.

ollama_allow_ram_offload: Option<bool>

Live Ollama RAM-offload toggle (/context offload on|off). Rides on the request like ollama_num_ctx so a toggle applies on the next turn without rebuilding the cached provider (whose config is frozen at startup). None falls back to the persisted [ollama] allow_ram_offload. Ignored by non-Ollama providers.

resolved_context_window: Option<usize>

The model’s real context window, filled by the effect layer from cache-first live discovery (resolve_context_window) just before dispatch. The reducer always initializes it None; None at the adapter means “unknown” (no window clamp).

resolved_max_output: Option<usize>

The model’s real per-response output ceiling, filled alongside resolved_context_window. None means unknown — adapters that require a concrete max_tokens (Anthropic) fall back to a floor.

output_schema: Option<Value>

mermaid run --output-schema: the JSON Schema the FORMATTING turn must conform to. Set only on that dedicated turn (never during the agentic loop — Gemini rejects tools+schema, Ollama’s format would degrade tool calling). When Some, the request carries no tools: the reducer sends none and the effect runner skips the built-ins.

suppress_auto_compact: bool

Pause automatic threshold compaction for this turn. Set from RuntimeState::auto_compact_suppressed after an auto-compaction failure, so a chronically failing summarizer doesn’t burn a model call every turn. Rides on the request like ollama_num_ctx because the effect preflight sees only the request, never RuntimeState. Cleared by a successful compaction, a manual /compact, or a conversation switch.

Trait Implementations

impl Clone for ChatRequest

fn clone(&self) -> ChatRequest

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for ChatRequest

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

Formats the value using the given formatter.