chore(llm): remove redundant schema feature gate from zeph-llm

CHANGELOG.md

@@ -23,6 +23,7 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 ### Changed
 
+- refactor(zeph-llm): remove redundant `schema` feature gate — `schemars` is now a mandatory dependency of `zeph-llm`; all `#[cfg(feature = "schema")]` / `#[cfg_attr(feature = "schema", ...)]` annotations removed; `chat_typed`, `chat_typed_erased`, structured output types, and the `extractor` module are always compiled (#2100)
 - Promote `scheduler` and `guardrail` features to the default feature set; users with `default-features = false` are unaffected
 
 ### Fixed

crates/zeph-llm/Cargo.toml

@@ -14,8 +14,6 @@ description = "LLM provider abstraction with Ollama, Claude, OpenAI, and Candle
 readme = "README.md"
 
 [features]
-default = ["schema"]
-schema = ["dep:schemars"]
 stt = ["reqwest/multipart"]
 candle = ["dep:audioadapter-buffers", "dep:candle-core", "dep:candle-nn", "dep:candle-transformers", "dep:hf-hub", "dep:tokenizers", "dep:symphonia", "dep:rubato"]
 cuda = ["candle", "candle-core/cuda", "candle-nn/cuda", "candle-transformers/cuda"]
@@ -37,7 +35,7 @@ rand.workspace = true
 rand_distr.workspace = true
 reqwest = { workspace = true, features = ["json", "rustls", "stream"] }
 rubato = { workspace = true, optional = true }
-schemars = { workspace = true, optional = true }
+schemars = { workspace = true }
 serde = { workspace = true, features = ["derive"] }
 serde_json.workspace = true
 symphonia = { workspace = true, optional = true, features = ["mp3", "ogg", "wav", "flac", "pcm"] }

crates/zeph-llm/src/any.rs

@@ -10,9 +10,7 @@ use crate::mock::MockProvider;
 use crate::ollama::OllamaProvider;
 use crate::openai::OpenAiProvider;
 use crate::orchestrator::ModelOrchestrator;
-#[cfg(feature = "schema")]
 use schemars::JsonSchema;
-#[cfg(feature = "schema")]
 use serde::de::DeserializeOwned;
 
 use crate::provider::{
@@ -67,7 +65,6 @@ impl AnyProvider {
     /// # Errors
     ///
     /// Returns an error if the provider fails or the response cannot be parsed.
-    #[cfg(feature = "schema")]
     pub async fn chat_typed_erased<T>(&self, messages: &[Message]) -> Result<T, crate::LlmError>
     where
         T: DeserializeOwned + JsonSchema + 'static,
@@ -610,7 +607,6 @@ mod tests {
         assert!(debug.contains("OpenAi"));
     }
 
-    #[cfg(feature = "schema")]
     #[tokio::test]
     async fn chat_typed_erased_dispatches_to_mock() {
         #[derive(Debug, serde::Deserialize, schemars::JsonSchema, PartialEq)]

crates/zeph-llm/src/claude/mod.rs

@@ -818,7 +818,6 @@ impl LlmProvider for ClaudeProvider {
         true
     }
 
-    #[cfg(feature = "schema")]
     async fn chat_typed<T>(&self, messages: &[Message]) -> Result<T, LlmError>
     where
         T: serde::de::DeserializeOwned + schemars::JsonSchema + 'static,

crates/zeph-llm/src/claude/types.rs

@@ -569,7 +569,6 @@ pub(super) struct AnthropicTool<'a> {
     pub input_schema: &'a serde_json::Value,
 }
 
-#[cfg(feature = "schema")]
 #[derive(Serialize)]
 pub(super) struct TypedToolRequestBody<'a> {
     pub model: &'a str,
@@ -589,7 +588,6 @@ pub(super) struct TypedToolRequestBody<'a> {
     pub context_management: Option<ContextManagement>,
 }
 
-#[cfg(feature = "schema")]
 #[derive(Serialize)]
 pub(super) struct ToolChoice<'a> {
     pub r#type: &'a str,

crates/zeph-llm/src/compatible.rs

@@ -113,7 +113,6 @@ impl LlmProvider for CompatibleProvider {
         self.inner.supports_structured_output()
     }
 
-    #[cfg(feature = "schema")]
     async fn chat_typed<T>(&self, messages: &[Message]) -> Result<T, LlmError>
     where
         T: serde::de::DeserializeOwned + schemars::JsonSchema + 'static,

crates/zeph-llm/src/lib.rs

@@ -12,7 +12,6 @@ pub mod claude;
 pub mod compatible;
 pub mod ema;
 pub mod error;
-#[cfg(feature = "schema")]
 pub mod extractor;
 pub mod gemini;
 pub mod http;
@@ -35,7 +34,6 @@ pub mod whisper;
 
 pub use claude::{ThinkingConfig, ThinkingEffort};
 pub use error::LlmError;
-#[cfg(feature = "schema")]
 pub use extractor::Extractor;
 pub use gemini::ThinkingLevel as GeminiThinkingLevel;
 pub use provider::{ChatStream, LlmProvider, StreamChunk, ThinkingBlock};

crates/zeph-llm/src/openai/mod.rs

@@ -552,7 +552,6 @@ impl LlmProvider for OpenAiProvider {
         true
     }
 
-    #[cfg(feature = "schema")]
     async fn chat_typed<T>(&self, messages: &[Message]) -> Result<T, LlmError>
     where
         T: serde::de::DeserializeOwned + schemars::JsonSchema + 'static,
@@ -1150,7 +1149,6 @@ struct EmbeddingData {
     embedding: Vec<f32>,
 }
 
-#[cfg(feature = "schema")]
 #[derive(Serialize)]
 struct TypedChatRequest<'a> {
     model: &'a str,
@@ -1179,14 +1177,12 @@ impl CompletionTokens {
     }
 }
 
-#[cfg(feature = "schema")]
 #[derive(Serialize)]
 struct ResponseFormat<'a> {
     r#type: &'a str,
     json_schema: JsonSchemaFormat<'a>,
 }
 
-#[cfg(feature = "schema")]
 #[derive(Serialize)]
 struct JsonSchemaFormat<'a> {
     name: &'a str,

crates/zeph-llm/src/openai/tests.rs

@@ -238,7 +238,6 @@ fn vision_chat_request_serialization_uses_gpt5_completion_tokens() {
     assert!(!json.contains("\"max_tokens\":55"));
 }
 
-#[cfg(feature = "schema")]
 #[test]
 fn typed_chat_request_serialization_uses_gpt5_completion_tokens() {
     let msgs = [ApiMessage {

crates/zeph-llm/src/orchestrator/classifier.rs

@@ -3,8 +3,7 @@
 
 use crate::provider::{Message, Role};
 
-#[cfg_attr(feature = "schema", derive(schemars::JsonSchema))]
-#[derive(Debug, Clone, serde::Deserialize)]
+#[derive(Debug, Clone, serde::Deserialize, schemars::JsonSchema)]
 pub struct ModelSelection {
     pub model: String,
     pub reason: String,

crates/zeph-llm/src/orchestrator/mod.rs

@@ -182,7 +182,6 @@ impl ModelOrchestrator {
             .expect("default provider must exist")
     }
 
-    #[cfg(feature = "schema")]
     async fn try_llm_routing(&self, messages: &[Message]) -> Option<String> {
         if !self.llm_routing {
             return None;
@@ -225,12 +224,6 @@ impl ModelOrchestrator {
         }
     }
 
-    #[cfg(not(feature = "schema"))]
-    #[allow(clippy::unused_async)]
-    async fn try_llm_routing(&self, _messages: &[Message]) -> Option<String> {
-        None
-    }
-
     async fn chat_with_fallback(&self, messages: &[Message]) -> Result<String, LlmError> {
         if let Some(selected) = self.try_llm_routing(messages).await
             && let Some(provider) = self.providers.get(&selected)

crates/zeph-llm/src/provider.rs

@@ -3,7 +3,6 @@
 
 use std::future::Future;
 use std::pin::Pin;
-#[cfg(feature = "schema")]
 use std::{
     any::TypeId,
     collections::HashMap,
@@ -15,7 +14,6 @@ use serde::{Deserialize, Serialize};
 
 use crate::error::LlmError;
 
-#[cfg(feature = "schema")]
 static SCHEMA_CACHE: LazyLock<Mutex<HashMap<TypeId, (serde_json::Value, String)>>> =
     LazyLock::new(|| Mutex::new(HashMap::new()));
 
@@ -24,7 +22,6 @@ static SCHEMA_CACHE: LazyLock<Mutex<HashMap<TypeId, (serde_json::Value, String)>
 /// # Errors
 ///
 /// Returns an error if schema serialization fails.
-#[cfg(feature = "schema")]
 pub(crate) fn cached_schema<T: schemars::JsonSchema + 'static>()
 -> Result<(serde_json::Value, String), crate::LlmError> {
     let type_id = TypeId::of::<T>();
@@ -577,7 +574,6 @@ pub trait LlmProvider: Send + Sync {
     ///
     /// Default implementation injects JSON schema into the system prompt and retries once
     /// on parse failure. Providers with native structured output should override this.
-    #[cfg(feature = "schema")]
     #[allow(async_fn_in_trait)]
     async fn chat_typed<T>(&self, messages: &[Message]) -> Result<T, LlmError>
     where
@@ -621,7 +617,6 @@ pub trait LlmProvider: Send + Sync {
 /// Strip markdown code fences from LLM output. Only handles outer fences;
 /// JSON containing trailing triple backticks in string values may be
 /// incorrectly trimmed (acceptable for MVP — see review R2).
-#[cfg(feature = "schema")]
 fn strip_json_fences(s: &str) -> &str {
     s.trim()
         .trim_start_matches("```json")
@@ -1144,37 +1139,31 @@ mod tests {
 
     // --- M27: strip_json_fences tests ---
 
-    #[cfg(feature = "schema")]
     #[test]
     fn strip_json_fences_plain_json() {
         assert_eq!(strip_json_fences(r#"{"a": 1}"#), r#"{"a": 1}"#);
     }
 
-    #[cfg(feature = "schema")]
     #[test]
     fn strip_json_fences_with_json_fence() {
         assert_eq!(strip_json_fences("```json\n{\"a\": 1}\n```"), r#"{"a": 1}"#);
     }
 
-    #[cfg(feature = "schema")]
     #[test]
     fn strip_json_fences_with_plain_fence() {
         assert_eq!(strip_json_fences("```\n{\"a\": 1}\n```"), r#"{"a": 1}"#);
     }
 
-    #[cfg(feature = "schema")]
     #[test]
     fn strip_json_fences_whitespace() {
         assert_eq!(strip_json_fences("  \n  "), "");
     }
 
-    #[cfg(feature = "schema")]
     #[test]
     fn strip_json_fences_empty() {
         assert_eq!(strip_json_fences(""), "");
     }
 
-    #[cfg(feature = "schema")]
     #[test]
     fn strip_json_fences_outer_whitespace() {
         assert_eq!(
@@ -1183,26 +1172,22 @@ mod tests {
         );
     }
 
-    #[cfg(feature = "schema")]
     #[test]
     fn strip_json_fences_only_opening_fence() {
         assert_eq!(strip_json_fences("```json\n{\"a\": 1}"), r#"{"a": 1}"#);
     }
 
     // --- M27: chat_typed tests ---
 
-    #[cfg(feature = "schema")]
     #[derive(Debug, serde::Deserialize, schemars::JsonSchema, PartialEq)]
     struct TestOutput {
         value: String,
     }
 
-    #[cfg(feature = "schema")]
     struct SequentialStub {
         responses: std::sync::Mutex<Vec<Result<String, LlmError>>>,
     }
 
-    #[cfg(feature = "schema")]
     impl SequentialStub {
         fn new(responses: Vec<Result<String, LlmError>>) -> Self {
             Self {
@@ -1211,7 +1196,6 @@ mod tests {
         }
     }
 
-    #[cfg(feature = "schema")]
     impl LlmProvider for SequentialStub {
         async fn chat(&self, _messages: &[Message]) -> Result<String, LlmError> {
             let mut responses = self.responses.lock().unwrap();
@@ -1247,7 +1231,6 @@ mod tests {
         }
     }
 
-    #[cfg(feature = "schema")]
     #[tokio::test]
     async fn chat_typed_happy_path() {
         let provider = StubProvider {
@@ -1263,7 +1246,6 @@ mod tests {
         );
     }
 
-    #[cfg(feature = "schema")]
     #[tokio::test]
     async fn chat_typed_retry_succeeds() {
         let provider = SequentialStub::new(vec![
@@ -1275,7 +1257,6 @@ mod tests {
         assert_eq!(result, TestOutput { value: "ok".into() });
     }
 
-    #[cfg(feature = "schema")]
     #[tokio::test]
     async fn chat_typed_both_fail() {
         let provider = SequentialStub::new(vec![Ok("bad json".into()), Ok("still bad".into())]);
@@ -1285,7 +1266,6 @@ mod tests {
         assert!(err.to_string().contains("parse failed after retry"));
     }
 
-    #[cfg(feature = "schema")]
     #[tokio::test]
     async fn chat_typed_chat_error_propagates() {
         let provider = SequentialStub::new(vec![Err(LlmError::Unavailable)]);
@@ -1294,7 +1274,6 @@ mod tests {
         assert!(matches!(result, Err(LlmError::Unavailable)));
     }
 
-    #[cfg(feature = "schema")]
     #[tokio::test]
     async fn chat_typed_strips_fences() {
         let provider = StubProvider {