Struct SessionConfig 

pub struct SessionConfig {

    pub max_context_tokens: usize,
    pub compaction_threshold: usize,
    pub compaction_buffer_percent: Option<u8>,
    pub store_path: Option<PathBuf>,
    pub auto_title: bool,
    pub title_model: Option<String>,
    pub tool_profile: String,
    pub prompt_cache: String,
    pub cache_warm_minutes: u64,
    pub undo_depth: usize,
    pub model_fast: Option<String>,
    pub goal_max_turns: u32,
    pub goal_model: Option<String>,
    pub model_smol: Option<String>,
    pub model_plan: Option<String>,
    pub prewalk: Option<bool>,
    pub model_race: String,
    pub race_after_ms: u64,
    pub fallback_models: Vec<String>,
    pub stream_retry_secs: u64,
    pub response_cache: String,
    pub preconnect: bool,
    pub intent_guidance: String,
    pub compaction_llm: String,
    pub compaction_method: String,
    pub reasoning_effort: Option<String>,
    pub magic_keywords: MagicKeywordsConfig,
    pub stream_rules: Vec<StreamRuleConfig>,
    pub headless_ask: Option<HeadlessAskMode>,
}

Fields

max_context_tokens: usize

compaction_threshold: usize

compaction_buffer_percent: Option<u8>

Share of the model’s context window to keep free. Auto-compaction fires at the lower of compaction_threshold and window × (100 − p) / 100. Unset uses compaction_threshold alone.

store_path: Option<PathBuf>

auto_title: bool

Auto-name sessions: project placeholder → first-message heuristic → small-model refine after the first turn. Default on.

title_model: Option<String>

Optional title model: provider/model or bare model id (same provider). Empty = pick a known small/fast sibling (haiku / mini / flash).

tool_profile: String

Tools advertised to the model: core (default, ~12 tools, faster TTFT) or full (every built-in including github/web/lsp).

prompt_cache: String

Anthropic-style prompt cache policy: auto (default) or none.

cache_warm_minutes: u64

Keep the prompt cache warm for this many idle minutes after a turn (TUI). Each keep-alive is a paid request that re-reads the cached prefix; 0 (default) turns it off.

undo_depth: usize

Turns /undo can step back. The oldest entries drop first past the cap; 0 = unlimited (memory grows with every turn).

model_fast: Option<String>

Fast model for trivial chat (provider/model or bare model id). Empty = auto-pick small sibling of the session model (haiku/mini/flash).

goal_max_turns: u32

Turns a /goal may take before it pauses (counted from when it was set or last resumed); 0 = no cap.

goal_model: Option<String>

Model that checks /goal (provider/model or bare id). Empty = the session model.

model_smol: Option<String>

Cheap model for task / swarm fan-out (provider/model or bare id). Empty = auto-pick the small sibling of the session model.

model_plan: Option<String>

Model used while the plan agent is active (provider/model or bare id).

prewalk: Option<bool>

Plan on the strong model, then hand the rest of the turn to the cheap one (model_smol, else the small sibling) after the first edit. Unset = off.

model_race: String

First-token race partner: off (default), auto (small sibling), or provider/model. When set, a slow primary TTFT starts the partner.

race_after_ms: u64

How long to wait for the primary first token before opening model_race.

fallback_models: Vec<String>

provider/model refs tried in order when the active model fails to open a stream (rate limit after named credentials, 5xx, network, timeout). Each needs a key: same provider, config api_key, or <PROVIDER>_API_KEY.

stream_retry_secs: u64

Re-send a step whose stream drops mid-response (network, timeout, 5xx, truncated body) for up to this many seconds after the first drop. The partial draft is discarded. Default 60; 0 turns it off.

response_cache: String

Process-local text-only response cache: auto (default) or off.

preconnect: bool

Open the provider connection while the prompt is being typed, so the first request skips DNS/TCP/TLS setup. Sends no credentials or prompt.

intent_guidance: String

Heuristic intent posture for build turns: auto (default), off, or always.

When enabled, high-confidence question/plan signals inject an ephemeral <whycodes_intent> block into the LLM request (not session history) so the model answers or plans instead of over-eager edits. Hard modes (ask / plan) still enforce tool denylists regardless of this flag.

compaction_llm: String

Full-replace compact (manual /compact and auto over threshold): "auto" (default) asks the session model for a structured summary; "off" is the local stub only.

compaction_method: String

Which compaction runs first: "auto" uses the provider’s own compaction when the model has it (Anthropic on-demand compaction, beta) and falls back to full-replace; "local" (default) is full-replace only.

reasoning_effort: Option<String>

OpenAI-compat / xAI reasoning_effort: low | medium | high | xhigh. Empty = family default (medium when the model supports effort). xhigh (Max) is grok-4.6+ and GLM-5+ only; a missing level clamps to high (GLM-5+ has no medium and defaults to high).

magic_keywords: MagicKeywordsConfig

Standalone prose keywords (ultrathink, orchestrate) that inject a hidden per-turn instruction. Default on.

stream_rules: Vec<StreamRuleConfig>

Regexes matched against streamed assistant text. A hit aborts the current LLM request and injects hint before the next step.

headless_ask: Option<HeadlessAskMode>

How --format json / stream-json treat permission ask.

Unset = deny (fail-closed). allow restores the old auto-approve behaviour; --approve-tools on the CLI also forces allow. TUI and --plain ignore this — they still prompt (or follow WHYCODES_AUTO_*).

Trait Implementations

impl Clone for SessionConfig

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 SessionConfig

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

Formats the value using the given formatter.

impl Default for SessionConfig

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for SessionConfig

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

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for SessionConfig

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

where __S: Serializer,

Serialize this value into the given Serde serializer.