Skip to content

The fastest, cheapest, most efficient open-source agent harness. Run more than 500 agents on a $10 VPS.

refactor(inference): drop managed reasoning wire helpers

Steven Enamakelbeac74c55939
diff --git a/crates/openhuman-tinyhumans/src/inference/managed_reasoning.rs b/crates/openhuman-tinyhumans/src/inference/managed_reasoning.rs
index b21ae8bd10..2e21a0078e 100644
--- a/crates/openhuman-tinyhumans/src/inference/managed_reasoning.rs
+++ b/crates/openhuman-tinyhumans/src/inference/managed_reasoning.rs
@@ -1,5 +1,4 @@
-//! Managed-backend wire extensions: the `reasoning` object and the `thread_id`
-//! field, derived from the provider-neutral request.
+//! OpenHuman's provider-neutral hint for a managed call that disables reasoning.
use serde_json::Value;
use tinyinference_llm::model::ModelRequest;
@@ -16,81 +15,6 @@ pub(crate) fn without_reasoning(mut request: ModelRequest) -> ModelRequest {
request
}
-/// Translate the request's reasoning choice into the managed backend's wire
-/// field: the OpenRouter-style `reasoning` object, which the backend forwards
-/// upstream.
-///
-/// Sources, highest first: a `reasoning` object the caller already put in
-/// provider options (left alone); the `without_reasoning` hint (core, `REASONING_OFF_METADATA_KEY`), which marks
-/// one helper call specifically; the request's provider-neutral
-/// `ModelRequest::reasoning` (the user's thinking level, from the harness
-/// `RunPolicy::default_reasoning`). The
-/// neutral field is consumed here, so the OpenAI-compatible transport does not
-/// also emit a top-level `reasoning_effort` for the same choice.
-pub(super) fn apply_reasoning_hint(mut request: ModelRequest) -> ModelRequest {
- if request.provider_options.get("reasoning").is_some() {
- return request;
- }
- let wants_off = request
- .metadata
- .get(REASONING_OFF_METADATA_KEY)
- .and_then(Value::as_bool)
- .unwrap_or(false);
- let neutral = request.reasoning.take();
- let wire = if wants_off {
- Some(serde_json::json!({ "enabled": false }))
- } else {
- neutral.as_ref().and_then(managed_reasoning_wire)
- };
- let Some(wire) = wire else {
- return request;
- };
- let mut options = request.provider_options.clone();
- if !options.is_object() {
- options = Value::Object(serde_json::Map::new());
- }
- if let Some(map) = options.as_object_mut() {
- map.insert("reasoning".to_string(), wire.clone());
- }
- log::debug!("[inference][managed] reasoning for this call: {wire}");
- request.with_provider_options(options)
-}
-
-/// The OpenRouter `reasoning` object for a provider-neutral config: `none`
-/// disables reasoning, an explicit budget becomes `max_tokens`, and any other
-/// effort is sent by name. An empty config sends nothing.
-fn managed_reasoning_wire(reasoning: &tinyinference_llm::model::ReasoningConfig) -> Option<Value> {
- use tinyinference_llm::model::ReasoningEffort;
- if reasoning.effort == Some(ReasoningEffort::None) {
- return Some(serde_json::json!({ "enabled": false }));
- }
- if let Some(budget) = reasoning.budget_tokens {
- return Some(serde_json::json!({ "max_tokens": budget }));
- }
- reasoning
- .effort
- .map(|effort| serde_json::json!({ "effort": effort.as_str() }))
-}
-
-/// Inject this managed model's explicitly owned thread into provider options.
-/// This backend-only wire extension is never inferred from ambient state.
-pub(super) fn with_thread_id(request: ModelRequest, thread_id: Option<&str>) -> ModelRequest {
- let Some(thread_id) = thread_id else {
- return request;
- };
- let mut options = request.provider_options.clone();
- if !options.is_object() {
- options = Value::Object(serde_json::Map::new());
- }
- if let Some(map) = options.as_object_mut() {
- map.insert(
- "thread_id".to_string(),
- Value::String(thread_id.to_string()),
- );
- }
- request.with_provider_options(options)
-}
-
#[cfg(test)]
#[path = "managed_reasoning_tests.rs"]
mod tests;
diff --git a/crates/openhuman-tinyhumans/src/inference/managed_reasoning_tests.rs b/crates/openhuman-tinyhumans/src/inference/managed_reasoning_tests.rs
index bfb8a428c4..08a68c846d 100644
--- a/crates/openhuman-tinyhumans/src/inference/managed_reasoning_tests.rs
+++ b/crates/openhuman-tinyhumans/src/inference/managed_reasoning_tests.rs
@@ -2,22 +2,15 @@ use super::*;
use tinyinference_llm::message::Message;
#[test]
-fn with_thread_id_injects_explicit_thread() {
- let request = ModelRequest::new(vec![Message::user("hi")]);
- let updated = with_thread_id(request, Some("thread-42"));
- assert_eq!(
- updated.provider_options["thread_id"],
- serde_json::json!("thread-42")
- );
-}
+fn without_reasoning_marks_the_request_without_rewriting_other_options() {
+ let options = serde_json::json!({ "transport": { "trace": true } });
+ let mut request = ModelRequest::new(vec![Message::user("hi")])
+ .with_provider_options(options.clone());
+ request.metadata = serde_json::json!({ "caller": "kept" });
-#[test]
-fn an_explicit_reasoning_option_is_left_alone() {
- let request = ModelRequest::new(vec![Message::user("hi")])
- .with_provider_options(serde_json::json!({ "reasoning": { "effort": "high" } }));
- let updated = apply_reasoning_hint(request);
- assert_eq!(
- updated.provider_options["reasoning"],
- serde_json::json!({ "effort": "high" })
- );
+ let updated = without_reasoning(request);
+
+ assert_eq!(updated.metadata["caller"], "kept");
+ assert_eq!(updated.metadata[REASONING_OFF_METADATA_KEY], true);
+ assert_eq!(updated.provider_options, options);
}