feat(inference): register burst-v1 tier; SuperContext scout uses it (#4260)

gitbooks/features/model-routing/README.md

@@ -39,6 +39,7 @@ The router wraps several pre-created providers (Anthropic, OpenAI, Google, Groq,
 | `hint:vision` | A vision-capable model | Screenshots, image attachments, OCR |
 | `hint:summarize` | A model good at compression | Memory tree summary builders |
 | `hint:code` | A code-tuned model | Native coder turns |
+| `hint:burst` | A high-throughput, low-cost model | Cheap, latency-tolerant pre-flight sweeps — e.g. the SuperContext scout |
 
 The exact mappings are configurable; the defaults ship sensible per-provider routes.
 

gitbooks/features/super-context.md

@@ -44,6 +44,8 @@ New thread, first message
 
 Because the scout is **read-only**, it can never take an action on a fresh thread — it only reads and summarizes. And because it runs in the harness rather than as an optional tool, the redundant `agent_prepare_context` tool is suppressed for that turn, so the agent doesn't do the work twice.
 
+The scout runs on the **`burst` tier** (`hint = "burst"` → `burst-v1` on the managed backend) — a cheap, high-throughput, non-reasoning model. The sweep is a latency-tolerant pre-flight pass, so raw throughput on a fast model is a better fit than the pricier agentic/reasoning tiers. See [Automatic Model Routing](model-routing/README.md).
+
 ***
 
 ## Safety and robustness

src/openhuman/agent/cost.rs

@@ -101,6 +101,15 @@ pub const PRICING_TABLE: &[ModelPricing] = &[
         cached_input_per_mtok_usd: 0.003625,
         output_per_mtok_usd: 0.87,
     },
+    // Burst tier — high-throughput, low-cost model; flat rate both directions,
+    // no prompt cache (so cached rate mirrors the input rate). Used by the
+    // SuperContext scout.
+    ModelPricing {
+        model: "burst-v1",
+        input_per_mtok_usd: 0.208,
+        cached_input_per_mtok_usd: 0.208,
+        output_per_mtok_usd: 0.208,
+    },
     // Vision tier — multimodal; estimate only. The backend's echoed
     // `charged_amount_usd` is authoritative when present.
     ModelPricing {
@@ -252,6 +261,17 @@ mod tests {
         assert_eq!(p.output_per_mtok_usd, 15.0);
     }
 
+    #[test]
+    fn lookup_pricing_has_a_burst_row() {
+        // The burst tier (SuperContext scout) must price from its own row —
+        // NOT via the $3/$15 fallback, which would inflate first-turn scout cost
+        // and could trip budget gates.
+        let p = lookup_pricing("burst-v1");
+        assert_eq!(p.model, "burst-v1");
+        assert_eq!(p.input_per_mtok_usd, 0.208);
+        assert_eq!(p.output_per_mtok_usd, 0.208);
+    }
+
     #[test]
     fn lookup_pricing_falls_back_for_unknown_model() {
         let p = lookup_pricing("totally-unknown-model");

src/openhuman/agent_registry/agents/context_scout/agent.toml

@@ -32,9 +32,11 @@ omit_profile = false
 omit_memory_md = false
 
 [model]
-# Multi-step gathering loop (recall → maybe fetch → assess), so the agentic
-# tier fits — not the slow reasoning tier (the scout is meant to be cheap).
-hint = "agentic"
+# Multi-step gathering loop (recall → maybe fetch → assess). Rides the
+# high-throughput `burst` tier: the scout is a cheap, latency-tolerant,
+# non-reasoning pre-flight pass, so raw throughput on a fast model beats the
+# pricier agentic/reasoning tiers. Resolves to `burst-v1` on the managed backend.
+hint = "burst"
 
 [tools]
 # Curated read-only context-gathering surface. No writes, no shell, no

src/openhuman/agent_registry/agents/loader.rs

@@ -988,6 +988,14 @@ mod tests {
         let def = find("context_scout");
         assert_eq!(def.agent_tier, AgentTier::Worker);
         assert_eq!(def.sandbox_mode, SandboxMode::ReadOnly);
+        // Super-context scout rides the cheap, high-throughput `burst` tier
+        // (resolves to `burst-v1` on the managed backend) — not the pricier
+        // agentic/reasoning tiers.
+        assert!(
+            matches!(&def.model, ModelSpec::Hint(h) if h == "burst"),
+            "context_scout must spawn on the burst tier, got {:?}",
+            def.model
+        );
         // Bundle cap — load-bearing for the parent's context budget. Leaves
         // room for the `recommended_skills` block alongside summary + plan.
         assert_eq!(def.max_result_chars, Some(5000));

src/openhuman/config/mod.rs

@@ -48,10 +48,10 @@ pub use schema::{
     StorageProviderSection, StreamMode, TeamModelConfig, TelegramConfig, TokenjuiceConfig,
     UpdateConfig, UpdateRestartStrategy, VoiceActivationMode, VoiceServerConfig, WebSearchConfig,
     WebhookConfig, DEFAULT_CLOUD_LLM_MODEL, DEFAULT_MEMORY_SYNC_INTERVAL_SECS, DEFAULT_MODEL,
-    MEMORY_SYNC_INTERVAL_PRESETS_SECS, MODEL_AGENTIC_V1, MODEL_CHAT_V1, MODEL_CODING_V1,
-    MODEL_REASONING_QUICK_V1, MODEL_REASONING_V1, MODEL_SUMMARIZATION_V1, MODEL_VISION_V1,
-    SEARCH_ENGINE_BRAVE, SEARCH_ENGINE_DISABLED, SEARCH_ENGINE_MANAGED, SEARCH_ENGINE_PARALLEL,
-    SEARCH_ENGINE_QUERIT,
+    MEMORY_SYNC_INTERVAL_PRESETS_SECS, MODEL_AGENTIC_V1, MODEL_BURST_V1, MODEL_CHAT_V1,
+    MODEL_CODING_V1, MODEL_REASONING_QUICK_V1, MODEL_REASONING_V1, MODEL_SUMMARIZATION_V1,
+    MODEL_VISION_V1, SEARCH_ENGINE_BRAVE, SEARCH_ENGINE_DISABLED, SEARCH_ENGINE_MANAGED,
+    SEARCH_ENGINE_PARALLEL, SEARCH_ENGINE_QUERIT,
 };
 pub use schemas::{
     all_controller_schemas as all_config_controller_schemas,

src/openhuman/config/schema/identity_cost.rs

@@ -150,7 +150,7 @@ impl Default for CostConfig {
 /// Default pricing for popular models (USD per 1M tokens)
 fn get_default_pricing() -> HashMap<String, ModelPricing> {
     use super::types::{
-        MODEL_AGENTIC_V1, MODEL_CHAT_V1, MODEL_CODING_V1, MODEL_REASONING_QUICK_V1,
+        MODEL_AGENTIC_V1, MODEL_BURST_V1, MODEL_CHAT_V1, MODEL_CODING_V1, MODEL_REASONING_QUICK_V1,
         MODEL_REASONING_V1,
     };
 
@@ -192,6 +192,14 @@ fn get_default_pricing() -> HashMap<String, ModelPricing> {
             output: 3.30,
         },
     );
+    // Burst tier — high-throughput, low-cost model; flat rate both directions.
+    prices.insert(
+        MODEL_BURST_V1.into(),
+        ModelPricing {
+            input: 0.208,
+            output: 0.208,
+        },
+    );
 
     prices
 }

src/openhuman/config/schema/types.rs

@@ -14,6 +14,10 @@ pub const MODEL_CHAT_V1: &str = "chat-v1";
 /// Legacy low-latency chat tier slug retained for older persisted configs.
 pub const MODEL_REASONING_QUICK_V1: &str = "reasoning-quick-v1";
 pub const MODEL_CODING_V1: &str = "coding-v1";
+/// High-throughput "burst" tier served by the managed backend. Cheap, fast,
+/// non-reasoning, text-only, 128k context, no prompt cache; used by the
+/// super-context scout. Managed-backend only (no BYOK knob).
+pub const MODEL_BURST_V1: &str = "burst-v1";
 pub const MODEL_SUMMARIZATION_V1: &str = "summarization-v1";
 /// Multimodal (image-input) tier. Managed backend serves this with the vision
 /// flag enabled; the vision sub-agent rides this tier via `hint:vision`.

src/openhuman/inference/model_context.rs

@@ -6,7 +6,8 @@
 //! metadata is not yet available.
 
 use crate::openhuman::config::{
-    MODEL_AGENTIC_V1, MODEL_CHAT_V1, MODEL_CODING_V1, MODEL_REASONING_QUICK_V1, MODEL_REASONING_V1,
+    MODEL_AGENTIC_V1, MODEL_BURST_V1, MODEL_CHAT_V1, MODEL_CODING_V1, MODEL_REASONING_QUICK_V1,
+    MODEL_REASONING_V1,
 };
 
 /// Conservative default for OpenHuman abstract tier models (tokens).
@@ -126,6 +127,9 @@ fn tier_context_window(model: &str) -> Option<u64> {
         MODEL_REASONING_V1 => Some(TIER_REASONING_CONTEXT),
         MODEL_AGENTIC_V1 | MODEL_CODING_V1 => Some(TIER_LARGE_CONTEXT),
         "summarization-v1" => Some(TIER_SUMMARIZATION_CONTEXT),
+        // Burst tier advertises a 128k window on the managed backend. Matched on
+        // the `burst-v1` alias before any substring fallbacks below.
+        MODEL_BURST_V1 => Some(TIER_STANDARD_CONTEXT),
         MODEL_CHAT_V1 | MODEL_REASONING_QUICK_V1 | "chat" => Some(TIER_STANDARD_CONTEXT),
         m if m.starts_with("gemma") || m.contains(":1b") || m.contains("270m") => {
             Some(TIER_LOCAL_CONTEXT)
@@ -277,6 +281,9 @@ mod tests {
         assert_eq!(context_window_for_model("reasoning-v1"), Some(1_000_000));
         assert_eq!(context_window_for_model("agentic-v1"), Some(200_000));
         assert_eq!(context_window_for_model("chat-v1"), Some(128_000));
+        // Burst tier — 128k on the managed backend. Matched on the alias, not
+        // the local-gemma 8k substring arm.
+        assert_eq!(context_window_for_model("burst-v1"), Some(128_000));
         assert_eq!(
             context_window_for_model("reasoning-quick-v1"),
             Some(128_000)

src/openhuman/inference/provider/factory.rs

@@ -77,7 +77,7 @@ pub(crate) const NO_MODEL_CONFIGURED_ANCHOR: &str = "resolved to an empty model
 
 fn is_abstract_tier_model(model: &str) -> bool {
     use crate::openhuman::config::{
-        MODEL_AGENTIC_V1, MODEL_CHAT_V1, MODEL_CODING_V1, MODEL_REASONING_QUICK_V1,
+        MODEL_AGENTIC_V1, MODEL_BURST_V1, MODEL_CHAT_V1, MODEL_CODING_V1, MODEL_REASONING_QUICK_V1,
         MODEL_REASONING_V1, MODEL_SUMMARIZATION_V1, MODEL_VISION_V1,
     };
     let trimmed = model.trim();
@@ -86,6 +86,7 @@ fn is_abstract_tier_model(model: &str) -> bool {
         || trimmed == MODEL_CHAT_V1
         || trimmed == MODEL_AGENTIC_V1
         || trimmed == MODEL_CODING_V1
+        || trimmed == MODEL_BURST_V1
         || trimmed == MODEL_VISION_V1
         || trimmed == MODEL_SUMMARIZATION_V1
 }
@@ -109,6 +110,7 @@ pub fn resolve_model_for_hint(hint_or_tier: &str, config: &Config) -> String {
         ("chat", crate::openhuman::config::MODEL_CHAT_V1),
         ("agentic", crate::openhuman::config::MODEL_AGENTIC_V1),
         ("coding", crate::openhuman::config::MODEL_CODING_V1),
+        ("burst", crate::openhuman::config::MODEL_BURST_V1),
         ("vision", crate::openhuman::config::MODEL_VISION_V1),
         (
             "summarization",
@@ -125,6 +127,7 @@ pub fn resolve_model_for_hint(hint_or_tier: &str, config: &Config) -> String {
         (crate::openhuman::config::MODEL_REASONING_QUICK_V1, "chat"),
         (crate::openhuman::config::MODEL_AGENTIC_V1, "agentic"),
         (crate::openhuman::config::MODEL_CODING_V1, "coding"),
+        (crate::openhuman::config::MODEL_BURST_V1, "burst"),
         (crate::openhuman::config::MODEL_VISION_V1, "vision"),
         (
             crate::openhuman::config::MODEL_SUMMARIZATION_V1,
@@ -184,7 +187,7 @@ pub fn resolve_model_for_hint(hint_or_tier: &str, config: &Config) -> String {
 /// to the backend.
 pub(crate) fn is_known_openhuman_tier(model: &str) -> bool {
     use crate::openhuman::config::{
-        MODEL_AGENTIC_V1, MODEL_CHAT_V1, MODEL_CODING_V1, MODEL_REASONING_QUICK_V1,
+        MODEL_AGENTIC_V1, MODEL_BURST_V1, MODEL_CHAT_V1, MODEL_CODING_V1, MODEL_REASONING_QUICK_V1,
         MODEL_REASONING_V1, MODEL_SUMMARIZATION_V1, MODEL_VISION_V1,
     };
     matches!(
@@ -193,6 +196,7 @@ pub(crate) fn is_known_openhuman_tier(model: &str) -> bool {
             | MODEL_CHAT_V1
             | MODEL_AGENTIC_V1
             | MODEL_CODING_V1
+            | MODEL_BURST_V1
             | MODEL_REASONING_QUICK_V1
             | MODEL_SUMMARIZATION_V1
             | MODEL_VISION_V1
@@ -200,6 +204,7 @@ pub(crate) fn is_known_openhuman_tier(model: &str) -> bool {
             | "hint:chat"
             | "hint:agentic"
             | "hint:coding"
+            | "hint:burst"
             | "hint:summarization"
             | "hint:vision"
     )
@@ -219,7 +224,7 @@ pub(crate) fn is_known_openhuman_tier(model: &str) -> bool {
 /// ([`crate::openhuman::inference::model_context::model_vision_enabled`]).
 pub(crate) fn oh_tier_supports_vision(model: &str) -> bool {
     use crate::openhuman::config::{
-        MODEL_AGENTIC_V1, MODEL_CHAT_V1, MODEL_CODING_V1, MODEL_REASONING_QUICK_V1,
+        MODEL_AGENTIC_V1, MODEL_BURST_V1, MODEL_CHAT_V1, MODEL_CODING_V1, MODEL_REASONING_QUICK_V1,
         MODEL_REASONING_V1, MODEL_SUMMARIZATION_V1, MODEL_VISION_V1,
     };
     match model {
@@ -231,6 +236,8 @@ pub(crate) fn oh_tier_supports_vision(model: &str) -> bool {
         MODEL_REASONING_QUICK_V1 => false,
         MODEL_AGENTIC_V1 | "hint:agentic" => false,
         MODEL_CODING_V1 | "hint:coding" => false,
+        // Burst is a text-only tier.
+        MODEL_BURST_V1 | "hint:burst" => false,
         MODEL_SUMMARIZATION_V1 | "hint:summarization" => false,
         _ => false,
     }
@@ -259,6 +266,11 @@ pub fn provider_for_role(role: &str, config: &Config) -> String {
         "reasoning" => config.reasoning_provider.as_deref(),
         "agentic" => config.agentic_provider.as_deref(),
         "coding" => config.coding_provider.as_deref(),
+        // Burst is managed-backend only (no per-workload provider knob): it rides
+        // the hosted high-throughput tier. Unset → falls through to
+        // `primary_cloud` (→ managed `burst-v1`) like the other tier-specific
+        // background workloads.
+        "burst" => None,
         // Tier-specific multimodal model; like `agentic` it is NOT part of the
         // chat-tier BYOK inheritance below — when unset it falls through to
         // `primary_cloud` (→ managed `vision-v1`).
@@ -299,13 +311,19 @@ pub fn provider_for_role(role: &str, config: &Config) -> String {
         if !matches!(role, "chat" | "reasoning" | "coding") {
             if let Some(chat) = config.chat_provider.as_deref() {
                 if crate::openhuman::inference::local::profile::is_local_provider_string(chat) {
+                    // burst is managed-backend only — there is no `burst_provider`
+                    // knob, so don't suggest setting one.
+                    let override_hint = if role == "burst" {
+                        "managed-backend only; no per-workload override".to_string()
+                    } else {
+                        format!("set {role}_provider explicitly to override")
+                    };
                     log::info!(
                         "[providers][local-fallback] role={} using managed backend (chat is \
-                         local '{}' but background workloads require cloud — set \
-                         {}_provider explicitly to override)",
+                         local '{}' but background workloads require cloud — {})",
                         role,
                         chat,
-                        role
+                        override_hint
                     );
                 }
             }
@@ -871,12 +889,18 @@ pub(crate) fn create_local_chat_provider_from_string(
 /// every attached image — leaving the managed vision sub-agent blind.
 fn managed_tier_for_role(role: &str) -> Option<&'static str> {
     use crate::openhuman::config::{
-        MODEL_AGENTIC_V1, MODEL_CHAT_V1, MODEL_CODING_V1, MODEL_REASONING_V1, MODEL_VISION_V1,
+        MODEL_AGENTIC_V1, MODEL_BURST_V1, MODEL_CHAT_V1, MODEL_CODING_V1, MODEL_REASONING_V1,
+        MODEL_VISION_V1,
     };
     match role {
         "reasoning" => Some(MODEL_REASONING_V1),
         "agentic" => Some(MODEL_AGENTIC_V1),
         "coding" => Some(MODEL_CODING_V1),
+        // Burst rides the managed backend's high-throughput tier. Pinned here
+        // (rather than collapsing to `default_model`) so the `hint = "burst"`
+        // sub-agent — the super-context scout — actually reaches `burst-v1`.
+        // There is no `burst_provider` knob: burst is managed-only.
+        "burst" => Some(MODEL_BURST_V1),
         "vision" => Some(MODEL_VISION_V1),
         // Background subconscious tick/triage: pinned to the lightweight chat
         // tier (see the doc above for why it is pinned despite being background).
@@ -977,6 +1001,7 @@ fn make_openhuman_backend(
         Some("chat") => crate::openhuman::config::MODEL_CHAT_V1.to_string(),
         Some("agentic") => crate::openhuman::config::MODEL_AGENTIC_V1.to_string(),
         Some("coding") => crate::openhuman::config::MODEL_CODING_V1.to_string(),
+        Some("burst") => crate::openhuman::config::MODEL_BURST_V1.to_string(),
         Some("summarization") => crate::openhuman::config::MODEL_SUMMARIZATION_V1.to_string(),
         Some("vision") => crate::openhuman::config::MODEL_VISION_V1.to_string(),
         Some(_) => {

src/openhuman/inference/provider/factory_tests.rs

@@ -404,7 +404,7 @@ fn create_chat_provider_uses_role() {
 #[test]
 fn managed_backend_pins_specialised_role_to_tier() {
     use crate::openhuman::config::{
-        MODEL_AGENTIC_V1, MODEL_CODING_V1, MODEL_REASONING_V1, MODEL_VISION_V1,
+        MODEL_AGENTIC_V1, MODEL_BURST_V1, MODEL_CODING_V1, MODEL_REASONING_V1, MODEL_VISION_V1,
     };
     // default_model is chat-v1 — the value the buggy path would have leaked.
     let config = Config::default();
@@ -414,6 +414,7 @@ fn managed_backend_pins_specialised_role_to_tier() {
         ("reasoning", MODEL_REASONING_V1),
         ("agentic", MODEL_AGENTIC_V1),
         ("coding", MODEL_CODING_V1),
+        ("burst", MODEL_BURST_V1),
         ("vision", MODEL_VISION_V1),
     ] {
         let (_, model) = create_chat_provider_from_string(role, "openhuman", &config)
@@ -481,12 +482,16 @@ fn managed_backend_summarization_ignores_cloud_llm_model_override() {
 // `code_executor` agent (`hint = "coding"`) makes when it spawns.
 #[test]
 fn subagent_hint_resolves_to_tier_on_managed_backend() {
-    use crate::openhuman::config::{MODEL_AGENTIC_V1, MODEL_CODING_V1, MODEL_REASONING_V1};
+    use crate::openhuman::config::{
+        MODEL_AGENTIC_V1, MODEL_BURST_V1, MODEL_CODING_V1, MODEL_REASONING_V1,
+    };
     let config = Config::default();
     for (hint, expected_tier) in &[
         ("coding", MODEL_CODING_V1),
         ("agentic", MODEL_AGENTIC_V1),
         ("reasoning", MODEL_REASONING_V1),
+        // The super-context scout (`context_scout`) spawns with `hint = "burst"`.
+        ("burst", MODEL_BURST_V1),
     ] {
         let (_, model) =
             create_chat_provider(hint, &config).expect("create_chat_provider must succeed");
@@ -1010,10 +1015,23 @@ fn known_hints_pass() {
     assert!(is_known_openhuman_tier("hint:chat"));
     assert!(is_known_openhuman_tier("hint:agentic"));
     assert!(is_known_openhuman_tier("hint:coding"));
+    assert!(is_known_openhuman_tier("hint:burst"));
     assert!(is_known_openhuman_tier("hint:summarization"));
     assert!(is_known_openhuman_tier("hint:vision"));
 }
 
+// `hint:burst` is accepted by `is_known_openhuman_tier`, so it must also be
+// translated to `burst-v1` by the managed backend — otherwise a saved
+// `default_model = "hint:burst"` would be forwarded literally and 400.
+#[test]
+fn managed_backend_translates_hint_burst_to_burst_tier() {
+    let mut config = Config::default();
+    config.default_model = Some("hint:burst".to_string());
+    let (_, model) = create_chat_provider_from_string("chat", "openhuman", &config)
+        .expect("managed backend must build");
+    assert_eq!(model, crate::openhuman::config::MODEL_BURST_V1);
+}
+
 #[test]
 fn invalid_models_fail() {
     assert!(!is_known_openhuman_tier("deepseek-v4-pro"));