Struct LearningConfig 

pub struct LearningConfig {

    pub enabled: bool,
    pub auto_activate: bool,
    pub min_failures: u32,
    pub improve_threshold: f64,
    pub rollback_threshold: f64,
    pub min_evaluations: u32,
    pub max_versions: u32,
    pub cooldown_minutes: u64,
    pub correction_detection: bool,
    pub correction_confidence_threshold: f32,
    pub detector_mode: DetectorMode,
    pub judge_model: String,
    pub feedback_provider: ProviderName,
    pub judge_adaptive_low: f32,
    pub judge_adaptive_high: f32,
    pub correction_recall_limit: u32,
    pub correction_min_similarity: f32,
    pub auto_promote_min_uses: u32,
    pub auto_promote_threshold: f64,
    pub auto_demote_min_uses: u32,
    pub auto_demote_threshold: f64,
    pub cross_session_rollout: bool,
    pub min_sessions_before_promote: u32,
    pub min_sessions_before_demote: u32,
    pub max_auto_sections: u32,
    pub domain_success_gate: bool,
    pub arise_enabled: bool,
    pub arise_min_tool_calls: u32,
    pub arise_trace_provider: ProviderName,
    pub stem_enabled: bool,
    pub stem_min_occurrences: u32,
    pub stem_min_success_rate: f64,
    pub stem_provider: ProviderName,
    pub stem_retention_days: u32,
    pub stem_pattern_window_days: u32,
    pub erl_enabled: bool,
    pub erl_extract_provider: ProviderName,
    pub erl_max_heuristics_per_skill: u32,
    pub erl_dedup_threshold: f32,
    pub erl_min_confidence: f64,
    pub d2skill_enabled: bool,
    pub d2skill_max_corrections: u32,
    pub d2skill_provider: ProviderName,
}

Self-learning and skill evolution configuration, nested under [skills.learning] in TOML.

When enabled = true, Zeph tracks skill performance and can automatically improve or roll back skill definitions based on usage outcomes (ARISE, STEM, D2Skill pipelines).

Example (TOML)

[skills.learning]
enabled = true
auto_activate = false
min_failures = 3

Fields

enabled: bool

Enable self-learning pipelines. Default: false.

auto_activate: bool

Automatically activate improved skill versions without user confirmation. Default: false.

min_failures: u32

improve_threshold: f64

rollback_threshold: f64

min_evaluations: u32

max_versions: u32

cooldown_minutes: u64

correction_detection: bool

correction_confidence_threshold: f32

detector_mode: DetectorMode

Detector strategy: “regex” (default) or “judge”.

judge_model: String

Model for the judge detector (e.g. “claude-sonnet-4-6”). Empty = use primary provider.

feedback_provider: ProviderName

Provider name from [[llm.providers]] for detector_mode = "model" (LlmClassifier).

Empty = use the primary provider. Named but not found in registry = log warning, degrade to regex-only. Never fails startup.

judge_adaptive_low: f32

Regex confidence below this value is treated as “not a correction” — judge not invoked.

judge_adaptive_high: f32

Regex confidence at or above this value is accepted without judge confirmation.

correction_recall_limit: u32

correction_min_similarity: f32

auto_promote_min_uses: u32

auto_promote_threshold: f64

auto_demote_min_uses: u32

auto_demote_threshold: f64

cross_session_rollout: bool

When true, auto-promote and auto-demote decisions require the skill to have been used across at least min_sessions_before_promote (for promotion) or min_sessions_before_demote (for demotion) distinct conversation sessions. Prevents trust transitions from a single long session.

min_sessions_before_promote: u32

Minimum number of distinct conversation_id values in skill_outcomes before auto-promotion is eligible. Only checked when cross_session_rollout = true.

min_sessions_before_demote: u32

Minimum distinct sessions before auto-demotion when cross_session_rollout = true.

Default 1 (demotion can happen after a single bad session by default). Separate from min_sessions_before_promote because demotion should be fast (low threshold) while promotion benefits from conservative validation (higher threshold).

max_auto_sections: u32

Maximum number of top-level content sections (markdown H2 headers) allowed in auto-generated skill bodies. Bodies exceeding this limit are rejected by validate_body_sections().

domain_success_gate: bool

When true, auto-generated skill versions must pass a domain-conditioned evaluation before promotion. If the improved body drifts from the original skill’s domain, activation is skipped (the version is still saved for manual review).

arise_enabled: bool

Enable ARISE trace-based skill improvement (disabled by default).

arise_min_tool_calls: u32

Minimum tool calls in a turn to trigger ARISE trace improvement.

arise_trace_provider: ProviderName

Provider name from [[llm.providers]] for ARISE trace summarization. Empty = fall back to primary provider.

stem_enabled: bool

Enable STEM automatic tool pattern detection and skill generation (disabled by default).

stem_min_occurrences: u32

Minimum occurrences of a tool sequence before generating a skill candidate.

stem_min_success_rate: f64

Minimum success rate of the pattern before generating a skill candidate.

stem_provider: ProviderName

Provider name from [[llm.providers]] for STEM skill generation. Empty = fall back to primary provider.

stem_retention_days: u32

Days to retain rows in skill_usage_log before pruning.

stem_pattern_window_days: u32

Window in days for pattern detection queries (limits scan cost on large tables).

erl_enabled: bool

Enable ERL post-task heuristic extraction (disabled by default).

erl_extract_provider: ProviderName

Provider name from [[llm.providers]] for ERL heuristic extraction. Empty = fall back to primary provider.

erl_max_heuristics_per_skill: u32

Maximum heuristics prepended per skill at match time.

erl_dedup_threshold: f32

Text similarity threshold (Jaccard) for heuristic deduplication. When exact text match exceeds this, increment use_count instead of inserting.

erl_min_confidence: f64

Minimum confidence to include a heuristic at match time.

d2skill_enabled: bool

Enable D2Skill step-level error correction (disabled by default).

Requires arise_enabled = true to populate corrections from ARISE traces. If d2skill_enabled = true and arise_enabled = false, existing corrections are still applied but no new ones are generated via ARISE.

d2skill_max_corrections: u32

Maximum corrections to inject per failure event.

d2skill_provider: ProviderName

Provider name from [[llm.providers]] for correction extraction from ARISE traces. Empty = fall back to primary provider.

Trait Implementations

impl Clone for LearningConfig

fn clone(&self) -> LearningConfig

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for LearningConfig

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

Formats the value using the given formatter.

impl Default for LearningConfig

fn default() -> Self

Returns the “default value” for a type.

impl<'de> Deserialize<'de> for LearningConfig

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

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for LearningConfig

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

where __S: Serializer,

Serialize this value into the given Serde serializer.