Skip to content

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

refactor(core): migrate usage and cost types to tinyinference_llm

Steven Enamakel08a24cce6f05
diff --git a/crates/openhuman-core/src/agent/agent_tests.rs b/crates/openhuman-core/src/agent/agent_tests.rs
index c8e495acd4..e1ac9f2b5d 100644
--- a/crates/openhuman-core/src/agent/agent_tests.rs
+++ b/crates/openhuman-core/src/agent/agent_tests.rs
@@ -26,7 +26,7 @@
use crate::agent::session_host::OpenHumanSessionHost;
use crate::config::AgentConfig;
-use crate::inference::provider::ChatResponse;
+use tinyinference_llm::model::ChatResponse;
use anyhow::Result;
use async_trait::async_trait;
use std::sync::{Arc, Mutex};
diff --git a/crates/openhuman-core/src/agent/cost.rs b/crates/openhuman-core/src/agent/cost.rs
index b548a38f36..108c828da9 100644
--- a/crates/openhuman-core/src/agent/cost.rs
+++ b/crates/openhuman-core/src/agent/cost.rs
@@ -16,7 +16,7 @@
//! to exist ($3 / $0.30 / $15 per MTok) and priced a glm-5.3-flash thread at
//! $4.25 when the provider billed about $0.30.
-use crate::inference::provider::BilledUsage;
+use tinyinference_llm::usage::BilledUsage;
use tinyagents_harness::cost::{CallCost, CallUsage};
/// Whether a model names an explicitly free OpenRouter variant.
diff --git a/crates/openhuman-core/src/agent/cost_tests.rs b/crates/openhuman-core/src/agent/cost_tests.rs
index 3c8078b96d..10b022da54 100644
--- a/crates/openhuman-core/src/agent/cost_tests.rs
+++ b/crates/openhuman-core/src/agent/cost_tests.rs
@@ -11,7 +11,7 @@ fn counts(input: u64, output: u64, cached: u64) -> BilledUsage {
fn record(turn: &mut tinyagents_harness::cost::TurnCost, model: &str, usage: &BilledUsage) {
use tinyagents_hooks::stop::StopHookMetrics as _;
let response = tinyinference_llm::model::response_to_model_response(
- &crate::inference::provider::ChatResponse {
+ &tinyinference_llm::model::ChatResponse {
text: Some("ok".into()),
usage: Some(*usage),
..Default::default()
diff --git a/crates/openhuman-core/src/agent/harness/graph.rs b/crates/openhuman-core/src/agent/harness/graph.rs
index ed014dc4a1..02dad590e4 100644
--- a/crates/openhuman-core/src/agent/harness/graph.rs
+++ b/crates/openhuman-core/src/agent/harness/graph.rs
@@ -179,7 +179,7 @@ pub(crate) async fn run_channel_turn_via_graph(
pause_at_cap: false,
// Bound the model's per-call output (legacy parity — channel turns ran at
// the standard per-turn budget).
- max_output_tokens: Some(crate::inference::provider::AGENT_TURN_MAX_OUTPUT_TOKENS),
+ max_output_tokens: Some(crate::agent::tinyagents::turn_policy::AGENT_TURN_MAX_OUTPUT_TOKENS),
// Context middlewares: cache-align + default tool-result byte cap (the
// channel path has no session context config to source from).
context_mw:
diff --git a/crates/openhuman-core/src/agent/session_host/driver.rs b/crates/openhuman-core/src/agent/session_host/driver.rs
index a3dd6383d7..258ef8e883 100644
--- a/crates/openhuman-core/src/agent/session_host/driver.rs
+++ b/crates/openhuman-core/src/agent/session_host/driver.rs
@@ -212,7 +212,7 @@ impl SessionDriver<OpenHumanRunContext> for OpenHumanSessionDriver {
message_run_queue: None,
early_exit_tools: &["ask_user_clarification"],
pause_at_cap: true,
- max_output_tokens: Some(crate::inference::provider::AGENT_TURN_MAX_OUTPUT_TOKENS),
+ max_output_tokens: Some(crate::agent::tinyagents::turn_policy::AGENT_TURN_MAX_OUTPUT_TOKENS),
context_mw,
tool_policy: Some(tool_policy),
deterministic_cacheable: false,
diff --git a/crates/openhuman-core/src/agent/subagent_host/ops/checkpoint.rs b/crates/openhuman-core/src/agent/subagent_host/ops/checkpoint.rs
index 0ed7441bb4..ee52de7ad5 100644
--- a/crates/openhuman-core/src/agent/subagent_host/ops/checkpoint.rs
+++ b/crates/openhuman-core/src/agent/subagent_host/ops/checkpoint.rs
@@ -1,6 +1,6 @@
//! OpenHuman billing adapter for TinyAgents' cap-hit checkpoint summary.
-use crate::inference::provider::BilledUsage;
+use tinyinference_llm::usage::BilledUsage;
use std::sync::{Arc, Mutex};
use tinyagents_harness::summarization::CapToolResult;
use tinyagents_orchestration::subagent::{summarize_at_cap, CapSummaryModel};
diff --git a/crates/openhuman-core/src/agent/subagent_host/ops/graph/dispatch.rs b/crates/openhuman-core/src/agent/subagent_host/ops/graph/dispatch.rs
index b5aa371405..c6b8dee08a 100644
--- a/crates/openhuman-core/src/agent/subagent_host/ops/graph/dispatch.rs
+++ b/crates/openhuman-core/src/agent/subagent_host/ops/graph/dispatch.rs
@@ -417,7 +417,7 @@ pub(in super::super) async fn run_subagent_via_graph(
usage.cached_input_tokens += u.cached_input_tokens();
let call_cost = crate::agent::cost::call_cost(model, &u);
usage.cost.add(call_cost);
- let billed = crate::inference::provider::BilledUsage::from_counts(
+ let billed = tinyinference_llm::usage::BilledUsage::from_counts(
u.input_tokens,
u.output_tokens,
)
diff --git a/crates/openhuman-core/src/agent/subagent_host/ops/runner.rs b/crates/openhuman-core/src/agent/subagent_host/ops/runner.rs
index b4a79fd134..f1d47c5257 100644
--- a/crates/openhuman-core/src/agent/subagent_host/ops/runner.rs
+++ b/crates/openhuman-core/src/agent/subagent_host/ops/runner.rs
@@ -22,7 +22,7 @@ use crate::agent::subagent_host::types::{
};
use crate::agent::tinyagents::budget::MAX_SPAWN_DEPTH;
use crate::agent_registry::host_definition::{AgentDefinition, AgentDefinitionRegistry};
-use crate::inference::provider::AGENT_TURN_MAX_OUTPUT_TOKENS;
+use crate::agent::tinyagents::turn_policy::AGENT_TURN_MAX_OUTPUT_TOKENS;
use tinyagents_definition::{
validate_tier_transition, AgentTier, IterationPolicy, PromptSource,
SandboxMode as AgentSandboxMode,
diff --git a/crates/openhuman-core/src/agent/tinyagents/middleware/turn_context.rs b/crates/openhuman-core/src/agent/tinyagents/middleware/turn_context.rs
index f709e30dd5..964c528059 100644
--- a/crates/openhuman-core/src/agent/tinyagents/middleware/turn_context.rs
+++ b/crates/openhuman-core/src/agent/tinyagents/middleware/turn_context.rs
@@ -275,7 +275,7 @@ impl Middleware<(), crate::agent::tinyagents::host::OpenHumanRunContext>
let host_usage =
tinyinference_llm::usage::usage_info_from_response(response)
.unwrap_or(
- crate::inference::provider::BilledUsage::from_counts(
+ tinyinference_llm::usage::BilledUsage::from_counts(
usage.input_tokens,
usage.output_tokens,
)
diff --git a/crates/openhuman-core/src/agent/tinyagents/observability/bridge_types.rs b/crates/openhuman-core/src/agent/tinyagents/observability/bridge_types.rs
index e7dfcce865..4f29b412e4 100644
--- a/crates/openhuman-core/src/agent/tinyagents/observability/bridge_types.rs
+++ b/crates/openhuman-core/src/agent/tinyagents/observability/bridge_types.rs
@@ -58,7 +58,7 @@ pub(crate) type ToolFailureMap = Arc<
>,
>;
-/// Shared FIFO carry of the per-call provider [`BilledUsage`](crate::inference::provider::BilledUsage)
+/// Shared FIFO carry of the per-call provider [`BilledUsage`](tinyinference_llm::usage::BilledUsage)
/// the model adapter observed, drained by the bridge when it records that
/// call's usage. The crate `Usage` the harness surfaces on
/// `AgentEvent::UsageRecorded` carries only token counts, so the
@@ -71,4 +71,4 @@ pub(crate) type ToolFailureMap = Arc<
/// that did not push, or an out-of-band usage event) degrades gracefully to a
/// catalogue estimate.
pub(crate) type ProviderUsageCarry =
- Arc<Mutex<std::collections::VecDeque<crate::inference::provider::BilledUsage>>>;
+ Arc<Mutex<std::collections::VecDeque<tinyinference_llm::usage::BilledUsage>>>;
diff --git a/crates/openhuman-core/src/agent/tinyagents/observability/event_bridge.rs b/crates/openhuman-core/src/agent/tinyagents/observability/event_bridge.rs
index b26d93f9c5..a44e9aef4b 100644
--- a/crates/openhuman-core/src/agent/tinyagents/observability/event_bridge.rs
+++ b/crates/openhuman-core/src/agent/tinyagents/observability/event_bridge.rs
@@ -9,7 +9,7 @@ use tokio::sync::mpsc::Sender;
use tinyinference_llm::usage::Usage;
-use crate::inference::provider::BilledUsage;
+use tinyinference_llm::usage::BilledUsage;
use crate::progress::AgentProgress;
use tinytools::humanize_tool_name;
diff --git a/crates/openhuman-core/src/agent/tinyagents/payload_summarizer.rs b/crates/openhuman-core/src/agent/tinyagents/payload_summarizer.rs
index e384feca9b..f81b02cc08 100644
--- a/crates/openhuman-core/src/agent/tinyagents/payload_summarizer.rs
+++ b/crates/openhuman-core/src/agent/tinyagents/payload_summarizer.rs
@@ -101,7 +101,7 @@ impl HarnessPayloadSummarizer<OpenHumanRunContext> for SubagentPayloadSummarizer
let thread_id = parent_ctx.data.thread_id.clone();
let max_output_tokens = definition
.max_turn_output_tokens
- .unwrap_or(crate::inference::provider::AGENT_TURN_MAX_OUTPUT_TOKENS);
+ .unwrap_or(crate::agent::tinyagents::turn_policy::AGENT_TURN_MAX_OUTPUT_TOKENS);
let child_context = unary_child_context(
parent_ctx,
&definition.id,
diff --git a/crates/openhuman-core/src/agent/tinyagents/turn_outcome.rs b/crates/openhuman-core/src/agent/tinyagents/turn_outcome.rs
index f861dbaca5..98ee963a47 100644
--- a/crates/openhuman-core/src/agent/tinyagents/turn_outcome.rs
+++ b/crates/openhuman-core/src/agent/tinyagents/turn_outcome.rs
@@ -207,7 +207,7 @@ pub(crate) fn record_unobserved_turn_usage(
?estimated_usd,
"[tinyagents] recording unobserved-turn usage into the global cost tracker"
);
- let usage = crate::inference::provider::BilledUsage::from_counts(input_tokens, output_tokens)
+ let usage = tinyinference_llm::usage::BilledUsage::from_counts(input_tokens, output_tokens)
.with_cached_input_tokens(cached_input_tokens);
// No per-call charge reached this path; record the catalog estimate when
// the model has one, else the record stays unpriced.
diff --git a/crates/openhuman-core/src/agent/tinyagents/turn_run_finalize.rs b/crates/openhuman-core/src/agent/tinyagents/turn_run_finalize.rs
index bd60e759f9..4028db9364 100644
--- a/crates/openhuman-core/src/agent/tinyagents/turn_run_finalize.rs
+++ b/crates/openhuman-core/src/agent/tinyagents/turn_run_finalize.rs
@@ -112,7 +112,7 @@ pub(super) async fn finalize_turn_outcome(
// model has a list price, else the turn's cost is unknown.
let estimate = crate::agent::cost::estimate_call_cost_usd(
model,
- &crate::inference::provider::BilledUsage::from_counts(input, output)
+ &tinyinference_llm::usage::BilledUsage::from_counts(input, output)
.with_cached_input_tokens(cached),
);
crate::agent::tinyagents::turn_outcome::record_unobserved_turn_usage(
diff --git a/crates/openhuman-core/src/flows/caps_host/prompt.rs b/crates/openhuman-core/src/flows/caps_host/prompt.rs
index a5ec797dd7..4e6c9acf43 100644
--- a/crates/openhuman-core/src/flows/caps_host/prompt.rs
+++ b/crates/openhuman-core/src/flows/caps_host/prompt.rs
@@ -18,7 +18,7 @@ pub(crate) fn build_completion_messages(messages: &[CompletionMessage]) -> Vec<T
}
/// Maps provider usage onto the established flow completion wire shape.
-pub(crate) fn usage_to_json(usage: &Option<crate::inference::provider::BilledUsage>) -> Value {
+pub(crate) fn usage_to_json(usage: &Option<tinyinference_llm::usage::BilledUsage>) -> Value {
match usage {
None => Value::Null,
Some(usage) => json!({
diff --git a/crates/openhuman-core/src/flows/caps_host/prompt_tests.rs b/crates/openhuman-core/src/flows/caps_host/prompt_tests.rs
index 78dfaba2e5..5666de8778 100644
--- a/crates/openhuman-core/src/flows/caps_host/prompt_tests.rs
+++ b/crates/openhuman-core/src/flows/caps_host/prompt_tests.rs
@@ -1,5 +1,5 @@
use super::*;
-use crate::inference::provider::BilledUsage;
+use tinyinference_llm::usage::BilledUsage;
/// The flow `agent`/`llm` node result exposes usage under these exact keys;
/// they are a contract with saved flows, so pin the literal JSON.
diff --git a/crates/openhuman-core/src/inference/reasoning_tests.rs b/crates/openhuman-core/src/inference/reasoning_tests.rs
index 7efffd319b..343dc3217d 100644
--- a/crates/openhuman-core/src/inference/reasoning_tests.rs
+++ b/crates/openhuman-core/src/inference/reasoning_tests.rs
@@ -215,7 +215,7 @@ fn root_turn_reasoning_reserves_output_room_with_a_budget() {
let reasoning = turn_reasoning_for(
&ctx,
None,
- Some(crate::inference::provider::AGENT_TURN_MAX_OUTPUT_TOKENS),
+ Some(crate::agent::tinyagents::turn_policy::AGENT_TURN_MAX_OUTPUT_TOKENS),
);
set_thread_effort(&crate::core::runtime::DispatchCtx::ambient(), thread, None);
assert_eq!(
diff --git a/crates/openhuman-core/src/platform/cost/global.rs b/crates/openhuman-core/src/platform/cost/global.rs
index 2810980444..3119c5dc0e 100644
--- a/crates/openhuman-core/src/platform/cost/global.rs
+++ b/crates/openhuman-core/src/platform/cost/global.rs
@@ -21,7 +21,7 @@ use parking_lot::RwLock;
use crate::config::CostConfig;
use crate::core::runtime::DispatchCtx;
-use crate::inference::provider::types::BilledUsage;
+use tinyinference_llm::usage::BilledUsage;
use super::tracker::CostTracker;
use super::types::{CostSource, TokenUsage, UsageScope};
diff --git a/crates/openhuman-core/src/platform/cost/report_tests.rs b/crates/openhuman-core/src/platform/cost/report_tests.rs
index aa5c6173d5..b34aa9f3fd 100644
--- a/crates/openhuman-core/src/platform/cost/report_tests.rs
+++ b/crates/openhuman-core/src/platform/cost/report_tests.rs
@@ -277,7 +277,7 @@ fn a_model_switch_in_a_thread_is_not_a_cold_call() {
#[test]
fn an_estimated_cost_lands_in_estimated_usd_not_charged() {
- use crate::inference::provider::BilledUsage;
+ use tinyinference_llm::usage::BilledUsage;
let estimated = BilledUsage::from_counts(100, 10).with_estimated_usd(0.25);
let usage = crate::platform::cost::global::build_token_usage("model/a", &estimated).unwrap();
assert_eq!(usage.cost_source, CostSource::Estimated);
diff --git a/crates/openhuman-core/src/threads/ops/usage.rs b/crates/openhuman-core/src/threads/ops/usage.rs
index 42183542f1..66f7d5c256 100644
--- a/crates/openhuman-core/src/threads/ops/usage.rs
+++ b/crates/openhuman-core/src/threads/ops/usage.rs
@@ -209,7 +209,7 @@ fn recorded_cost(
let estimate = model.and_then(|model| {
crate::agent::cost::estimate_call_cost_usd(
model,
- &crate::inference::provider::BilledUsage::from_counts(
+ &tinyinference_llm::usage::BilledUsage::from_counts(
split.unpriced_input_tokens,
split.unpriced_output_tokens,
)