Skip to content

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

refactor(inference): improve cloud slug model resolution logic

Steven Enamakel07124a531123
diff --git a/crates/openhuman-core/src/inference/provider/factory/cloud_slug.rs b/crates/openhuman-core/src/inference/provider/factory/cloud_slug.rs
index 94ab5b7e64..0a7030119a 100644
--- a/crates/openhuman-core/src/inference/provider/factory/cloud_slug.rs
+++ b/crates/openhuman-core/src/inference/provider/factory/cloud_slug.rs
@@ -9,6 +9,8 @@ use crate::inference::provider::factory::access_gates::verify_backend_session_ac
use crate::inference::provider::factory::access_gates::verify_session_active;
use crate::inference::provider::fallback_diagnostics;
use tinyinference_llm::providers::anthropic::endpoint_is_anthropic_messages;
+use tinyinference_llm::providers::openai::codex::resolve_codex_routing;
+use tinyinference_llm::providers::registry::resolve_slug_model;
/// Look up a `cloud_providers` entry by slug and build the provider.
/// The shared resolution for a `<slug>:<model>` cloud provider — the cloud
@@ -47,87 +49,74 @@ pub(super) fn resolve_cloud_slug_for<'a>(
)
})?;
- // Resolve effective model: use provided model if non-empty, else fall back
- // to the entry's legacy default_model (if any), else empty → error.
- let mut effective_model = if model.trim().is_empty() {
- entry.default_model.clone().unwrap_or_default()
- } else {
- model.to_string()
- };
-
- // Guard: if effective_model is still empty after fallback, bail with an
- // actionable error. Sending an empty model string to providers like
- // nvidia-nim causes a 400 "model field is required" — a confusing error
- // that obscures the real cause (missing model in the provider string or
- // unset default_model on the config entry).
- // See https://github.com/tinyhumansai/openhuman/issues/2784.
- //
- // OpenhumanJwt entries are exempt: they always delegate to
- // make_openhuman_backend which derives the model from config.default_model,
- // ignoring whatever effective_model we computed here.
let hosted = crate::config::schema::cloud_providers::is_hosted_auth_style_for(
extensions,
entry.auth_style,
);
- if !hosted && effective_model.trim().is_empty() {
- log::warn!(
- "[providers][chat-factory] role={} slug={} resolved to empty model — \
- provider string must include a model id (e.g. '{}:<model-id>') or \
- set default_model on the cloud_providers entry",
- role,
- slug,
- slug,
- );
- anyhow::bail!(
- "[chat-factory] no model configured: role '{}' resolved to an empty model id for slug '{}'. \
- Include a model in the provider string (e.g. '{slug}:<model-id>') or \
- set default_model on the cloud_providers entry for slug '{slug}'.",
+ // Hosted routes derive their model from the managed provider and retain the
+ // old permissive behavior. BYOK model fallback and hint classification are
+ // shared with the provider owner.
+ let effective_model = if hosted {
+ if model.trim().is_empty() {
+ entry.default_model.clone().unwrap_or_default()
+ } else {
+ model.to_string()
+ }
+ } else {
+ match resolve_slug_model(model, entry.default_model.as_deref(), |candidate| {
+ crate::inference::provider_model::is_abstract_model_for(
+ extensions,
+ PROVIDER_OPENHUMAN,
+ candidate,
+ )
+ }) {
+ Ok(model) => model,
+ Err(tinyinference_llm::providers::registry::SlugModelError::MissingModel) => {
+ log::warn!(
+ "[providers][chat-factory] role={} slug={} resolved to empty model — \
+ provider string must include a model id (e.g. '{}:<model-id>') or \
+ set default_model on the cloud_providers entry",
+ role,
+ slug,
+ slug,
+ );
+ anyhow::bail!(
+ "[chat-factory] no model configured: role '{}' resolved to an empty model id for slug '{}'. \
+ Include a model in the provider string (e.g. '{slug}:<model-id>') or \
+ set default_model on the cloud_providers entry for slug '{slug}'.",
+ role,
+ slug,
+ );
+ }
+ Err(tinyinference_llm::providers::registry::SlugModelError::NoConcreteDefault) => {
+ anyhow::bail!(
+ "[chat-factory] model '{}' is an abstract tier for role '{}', \
+ but cloud provider slug '{}' has no concrete default_model configured. \
+ Set cloud_providers[].default_model to a provider-native model id.",
+ model,
+ role,
+ slug
+ );
+ }
+ }
+ };
+
+ if !hosted && effective_model != model && !model.trim().is_empty() {
+ log::info!(
+ "[providers][chat-factory] role={} slug={} remapping abstract model {} -> {}",
role,
slug,
+ model,
+ effective_model
);
}
- if !hosted
- && crate::inference::provider_model::is_abstract_model_for(
- extensions,
- PROVIDER_OPENHUMAN,
- &effective_model,
- )
- {
- if let Some(default_model) = entry.default_model.as_deref().map(str::trim).filter(|m| {
- !m.is_empty()
- && !crate::inference::provider_model::is_abstract_model_for(
- extensions,
- PROVIDER_OPENHUMAN,
- m,
- )
- }) {
- log::info!(
- "[providers][chat-factory] role={} slug={} remapping abstract model {} -> {}",
- role,
- slug,
- effective_model,
- default_model
- );
- effective_model = default_model.to_string();
- } else {
- anyhow::bail!(
- "[chat-factory] model '{}' is an abstract tier for role '{}', \
- but cloud provider slug '{}' has no concrete default_model configured. \
- Set cloud_providers[].default_model to a provider-native model id (e.g. deepseek-v4-pro).",
- effective_model,
- role,
- slug
- );
- }
- }
-
log::info!(
"[providers][chat-factory] role={} slug={} model={} endpoint_host={}",
role,
slug,
effective_model,
- redact_endpoint(&entry.endpoint)
+ super::super::redact_endpoint_origin(&entry.endpoint)
);
// #5146 §2.1: a raw "failed to read API key for slug 'anthropic'" is
@@ -186,8 +175,40 @@ pub(super) fn resolve_cloud_slug_for<'a>(
anyhow::bail!("{}", missing_credentials());
}
let bearer_is_oauth = slug == "openai" && openai_bearer_is_oauth(config);
- let codex = resolve_openai_codex_routing(config, slug, &entry.endpoint, &key, bearer_is_oauth)
- .map_err(anyhow::Error::msg)?;
+ let oauth_credentials = if slug == "openai" {
+ match crate::security::credentials::openai_oauth::lookup_openai_oauth_credentials(config) {
+ Ok(credentials) => credentials,
+ Err(error) if !key.trim().is_empty() => {
+ log::warn!(
+ "[providers][openai-codex] oauth metadata unavailable; continuing with standard bearer key: {error}"
+ );
+ None
+ }
+ Err(error) => {
+ return Err(anyhow::anyhow!(
+ "[chat-factory] openai oauth lookup failed: {error}"
+ ));
+ }
+ }
+ } else {
+ None
+ };
+ let has_oauth_credentials = oauth_credentials.is_some();
+ let account_id = oauth_credentials
+ .filter(|_| bearer_is_oauth)
+ .and_then(|credentials| credentials.account_id);
+ let using_codex_oauth = bearer_is_oauth && has_oauth_credentials;
+ let mut codex = resolve_codex_routing(
+ slug,
+ &entry.endpoint,
+ using_codex_oauth,
+ account_id.as_deref(),
+ );
+ // The owner helper normalizes standard route slashes. The provider registry
+ // previously received the configured endpoint verbatim, so preserve it.
+ if !codex.using_oauth {
+ codex.endpoint.clone_from(&entry.endpoint);
+ }
Ok(CloudSlugResolution {
entry,