Skip to content

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

refactor(session_host): unify session model resolution logic

Steven Enamakel5b6b5cec292d
diff --git a/crates/openhuman-core/src/agent/session_host/builder/factory.rs b/crates/openhuman-core/src/agent/session_host/builder/factory.rs
index 77ac3d79d7..c84c1744fd 100644
--- a/crates/openhuman-core/src/agent/session_host/builder/factory.rs
+++ b/crates/openhuman-core/src/agent/session_host/builder/factory.rs
@@ -227,86 +227,14 @@ impl OpenHumanSessionHost {
crate::tools::filter_tools_by_user_preference(&mut tools, &enabled_tools);
}
- // Route the main agent's chat through the unified per-workload
- // factory so the user's "Reasoning" routing in the AI settings
- // panel (e.g. `reasoning_provider = "anthropic:claude-..."`)
- // actually takes effect. The factory returns a (Provider, model)
- // tuple — the resolved model wins over the legacy `default_model`
- // fallback so explicit picks like `anthropic:claude-sonnet-4-5`
- // actually use claude-sonnet-4-5 end to end (sending the abstract
- // "hint:reasoning" tier name to Anthropic would 404).
- //
- // When `reasoning_provider` is unset or `"cloud"`, the factory
- // resolves to the primary cloud (OpenHuman by default), so the
- // baseline behaviour is identical to the legacy
- // `create_intelligent_routing_provider` path.
- //
- // The ReliableProvider retry/backoff + model-fallback wrapper is
- // re-layered on top of the factory's resolved backend below (issue
- // #4249, 1c). `model_routes` translation and intelligent local/cloud
- // task hinting now live in the unified routing layer (router.rs) rather
- // than a per-session wrapper, so they are not re-wrapped here.
- // The Master OpenHumanSessionHost's definition selects the coding workload so the
- // default user-facing turn has a model suitable for the direct
- // inspect → edit → verify loop. Legacy/no-definition callers retain
- // the configured chat default. Other specialised roles still select
- // their model in the sub-agent runner.
- // Only the explicit `hint:<role>` form routes to a specialised
- // workload — legacy tier literals like `hint:reasoning` (which the
- // bootstrap historically pinned as `default_model` for everyone)
- // fall through to `chat`. This preserves `config.chat_provider` for
- // legacy callers. A built-in Master OpenHumanSessionHost has an explicit
- // `hint:coding`, while existing pre-registry callers continue using
- // `config.default_model` unchanged.
- let provider_role =
- provider_role_for_definition(config.default_model.as_deref(), target_def);
- // Retry/backoff is now owned by the crate `RetryPolicy` at the harness
- // model call (issue #4249, Phase 3a) — see `tinyagents::run_policy_for`.
- // The turn path therefore no longer wraps the resolved provider in
- // `ReliableProvider`; wrapping it here plus the crate retry would
- // double-retry every transient error. Cross-route fallback is likewise
- // the crate registry `FallbackPolicy`, so `config.reliability.*` no longer
- // layers on the turn path (it still governs the non-seam provider paths).
- let (mut model_name, supports_native) = match &turn_model_source {
- Some((source, model_name)) => {
- (model_name.clone(), source.supports_native_tool_calling())
- }
- None => {
- let (resolved_chat_model, model_name) =
- crate::inference::provider::create_chat_model_with_model_id_in(
- ctx,
- provider_role,
- config,
- config.default_temperature,
- )?;
- let supports_native = resolved_chat_model
- .profile()
- .is_none_or(|profile| profile.tool_calling);
- (model_name, supports_native)
- }
- };
- log::info!(
- "[session-builder] agent_id={} provider_role={} resolved_model={} supports_native_tools={}",
- agent_id,
- provider_role,
- model_name,
- supports_native
- );
- let target_agent_id = target_def
- .map(|def| def.id.as_str())
- .unwrap_or("orchestrator");
- let target_is_lead = target_def
- .map(|def| !def.subagents.is_empty())
- .unwrap_or(true);
- if let Some(pinned_model) = config.configured_agent_model(target_agent_id, target_is_lead) {
- log::debug!(
- "[session-builder] agent_id={} using config-level model pin model={}",
- target_agent_id,
- pinned_model
- );
- model_name = pinned_model.to_string();
- }
-
+ let (provider_role, model_name, supports_native) =
+ super::factory_resolution::resolve_session_model(
+ ctx,
+ config,
+ agent_id,
+ target_def,
+ turn_model_source.as_ref(),
+ )?;
// Resolve the user-configured vision flag for the (now-final) model while
// the full `Config` / `model_registry` is in scope — the turn engine only
// sees `MultimodalConfig`. Stored on the session and surfaced to the image
diff --git a/crates/openhuman-core/src/agent/session_host/builder/factory_resolution.rs b/crates/openhuman-core/src/agent/session_host/builder/factory_resolution.rs
index 16d479f83e..5fd8fcee43 100644
--- a/crates/openhuman-core/src/agent/session_host/builder/factory_resolution.rs
+++ b/crates/openhuman-core/src/agent/session_host/builder/factory_resolution.rs
@@ -44,3 +44,59 @@ pub(crate) fn provider_role_for_definition(
#[cfg(test)]
#[path = "factory_resolution_tests.rs"]
mod tests;
+
+/// Resolve the model source, workload role, and native-tool capability for a session.
+///
+/// Session construction and triage direct turns share the same resolved source when one
+/// is already available; otherwise the ordinary provider factory owns construction.
+pub(crate) fn resolve_session_model(
+ ctx: &crate::core::runtime::DispatchCtx,
+ config: &crate::config::Config,
+ agent_id: &str,
+ target_def: Option<&crate::agent_registry::host_definition::AgentDefinition>,
+ turn_model_source: Option<&(crate::agent::tinyagents::TurnModelSource, String)>,
+) -> anyhow::Result<(&'static str, String, bool)> {
+ // The Master OpenHumanSessionHost definition selects the coding workload so the default
+ // user-facing turn has a model suitable for the direct inspect → edit → verify loop.
+ // Legacy/no-definition callers retain the configured chat default.
+ let provider_role = provider_role_for_definition(config.default_model.as_deref(), target_def);
+ // Retry/backoff is owned by the crate RetryPolicy at the harness model call. The turn path
+ // must not wrap the provider again or it will double-retry. Cross-route fallback likewise
+ // belongs to the crate registry FallbackPolicy.
+ let (mut model_name, supports_native) = match turn_model_source {
+ Some((source, model_name)) => (model_name.clone(), source.supports_native_tool_calling()),
+ None => {
+ let (model, model_name) =
+ crate::inference::provider::create_chat_model_with_model_id_in(
+ ctx,
+ provider_role,
+ config,
+ config.default_temperature,
+ )?;
+ let supports_native = model.profile().is_none_or(|profile| profile.tool_calling);
+ (model_name, supports_native)
+ }
+ };
+ log::info!(
+ "[session-builder] agent_id={} provider_role={} resolved_model={} supports_native_tools={}",
+ agent_id,
+ provider_role,
+ model_name,
+ supports_native
+ );
+ let target_agent_id = target_def
+ .map(|def| def.id.as_str())
+ .unwrap_or("orchestrator");
+ let target_is_lead = target_def
+ .map(|def| !def.subagents.is_empty())
+ .unwrap_or(true);
+ if let Some(pinned_model) = config.configured_agent_model(target_agent_id, target_is_lead) {
+ log::debug!(
+ "[session-builder] agent_id={} using config-level model pin model={}",
+ target_agent_id,
+ pinned_model
+ );
+ model_name = pinned_model.to_string();
+ }
+ Ok((provider_role, model_name, supports_native))
+}