diff --git a/AGENTS.md b/AGENTS.md
new file mode 100644
index 00000000..15e161f5
--- /dev/null
+++ b/AGENTS.md
@@ -0,0 +1,34 @@
+# Paddler — Stateful Load Balancer for llama.cpp
+
+## Overview
+
+Load balancer + reverse proxy for llama.cpp servers. Distributes requests by available slots, buffers when all busy.
+
+## Key Files
+
+| File | Purpose |
+|------|---------|
+| `src/main.rs` | CLI (balancer / agent subcommands) |
+| `src/balancer/proxy_service.rs` | Proxy logic, model detection (`--check-model`), upstream peer creation |
+| `src/balancer/upstream_peer_pool.rs` | Peer registry, `use_best_peer()`, slot tracking |
+| `src/balancer/upstream_peer.rs` | Peer struct, `Ord` sort (usable → most idle → least processing) |
+| `src/balancer/request_context.rs` | Per-request context, `take_slot()` / `release_slot()` |
+| `src/agent/monitoring_service.rs` | Periodic llama.cpp `/slots` polling |
+| `src/agent/reporting_service.rs` | Agent → balancer WebSocket keepalive |
+| `src/llamacpp/llamacpp_client.rs` | llama.cpp API client |
+
+## Slot Tracking
+
+- Per-request: `take_slot()` / `release_slot()` update counts immediately on request/response
+- Heartbeat: agent polls llama.cpp every N ms (`--monitoring-interval`, default 10000), reconciles drift
+- Peer sort: usable first, then most idle slots DESC, least processing ASC
+
+## Test Infrastructure
+
+- `tests/integration_tests/` — Rust integration tests (cucumber-style)
+- `tests/scripts/test-model-detection.sh` — bash test suite, 19 scenarios
+  - `PADDLER_PROXY_HOST=<host> PADDLER_MODEL=<model> bash tests/scripts/test-model-detection.sh`
+
+## Makefile Targets
+
+`build` · `test` · `integration_tests` · `shell_tests` · `fmt` · `clean`
diff --git a/Cargo.lock b/Cargo.lock
index 2628020d..be47fb6f 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -2571,6 +2571,7 @@ dependencies = [
  "mime_guess",
  "pingora",
  "ratatui",
+ "regex",
  "reqwest",
  "reqwest-eventsource",
  "rust-embed",
diff --git a/Cargo.toml b/Cargo.toml
index 042e5f5a..5ad3d89c 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -9,7 +9,7 @@ documentation = "https://github.com/distantmagic/paddler/blob/main/README.md"
 version = "1.2.1"
 
 [workspace]
-members = ["integration_tests"]
+members = ["tests/integration_tests"]
 
 [dependencies]
 actix = "0.13.5"
@@ -32,6 +32,7 @@ serde_json = "1.0.140"
 tokio = { version = "1.45.1", features = ["full"] }
 tokio-stream = { version = "0.1.17", features = ["sync"] }
 url = { version = "2.5.4", features = ["serde"] }
+regex = "1.11.1"
 
 chrono = { version = "0.4.41", optional = true }
 crossterm = { version = "0.28.1", features = ["event-stream"], optional = true }
diff --git a/Makefile b/Makefile
index c36066ac..1609223e 100644
--- a/Makefile
+++ b/Makefile
@@ -31,12 +31,16 @@ clean:
 .PHONY: fmt
 fmt: node_modules
 	./jarmuz-fmt.mjs
-	$(MAKE) -C integration_tests fmt
+	$(MAKE) -C tests/integration_tests fmt
 
 .PHONY: integration_tests
 integration_tests:
 	cargo build
-	$(MAKE) -C integration_tests test
+	$(MAKE) -C tests/integration_tests test
+
+.PHONY: shell_tests
+shell_tests:
+	bash tests/scripts/test-model-detection.sh
 
 .PHONY: test
 test: integration_tests
diff --git a/jarmuz/worker-prettier.mjs b/jarmuz/worker-prettier.mjs
index 5e16d4fe..3d13f9b1 100644
--- a/jarmuz/worker-prettier.mjs
+++ b/jarmuz/worker-prettier.mjs
@@ -4,7 +4,7 @@ command(`
   npm exec prettier --
     --plugin=prettier-plugin-organize-imports
     --write
-    integration_tests/tests/fixtures
+    tests/integration_tests/tests/fixtures
     jarmuz
     resources
     *.mjs
diff --git a/resources/ts/components/AgentsList.tsx b/resources/ts/components/AgentsList.tsx
index 8b41aee9..1c0cbcfb 100644
--- a/resources/ts/components/AgentsList.tsx
+++ b/resources/ts/components/AgentsList.tsx
@@ -1,5 +1,5 @@
 import clsx from "clsx";
-import React, { CSSProperties } from "react";
+import React, { CSSProperties, useState } from "react";
 
 import { type Agent } from "../schemas/Agent";
 
@@ -9,27 +9,145 @@ import {
   agentUsage,
   agentUsage__progress,
   agentsTable,
+  sortIndicator,
+  sortIndicatorAsc,
+  sortIndicatorDesc,
 } from "./Dashboard.module.css";
 
 function formatTimestamp(timestamp: number): string {
   return new Date(timestamp * 1000).toLocaleString();
 }
 
+type SortColumn =
+  | "name"
+  | "model"
+  | "issues"
+  | "llamacppAddr"
+  | "lastUpdate"
+  | "idleSlots"
+  | "processingSlots";
+
+function getSortIndicator(
+  sortConfig: { key: SortColumn; direction: "ascending" | "descending" },
+  currentKey: SortColumn
+): React.ReactNode {
+  if (sortConfig.key !== currentKey) {
+    return null;
+  }
+  const className = clsx(sortIndicator, sortConfig.direction === "ascending" ? sortIndicatorAsc : sortIndicatorDesc);
+  return (
+    <span className={className}>
+      {sortConfig.direction === "ascending" ? "↑" : "↓"}
+    </span>
+  );
+}
+
 export function AgentsList({ agents }: { agents: Array<Agent> }) {
+  const [sortConfig, setSortConfig] = useState<{
+    key: SortColumn;
+    direction: "ascending" | "descending";
+  }>({ key: "name", direction: "ascending" });
+
+  function sortAgents(agents: Array<Agent>): Array<Agent> {
+    const sortableAgents = [...agents];
+    sortableAgents.sort(function (a, b) {
+      const { key, direction } = sortConfig;
+
+      // Helper function to get comparison value based on column type
+      function getValue(agent: Agent, key: SortColumn): string | number {
+        switch (key) {
+          case "name":
+            return agent.status.agent_name || "";
+          case "model":
+            return agent.status.model || "";
+          case "llamacppAddr":
+            return agent.status.external_llamacpp_addr;
+          case "lastUpdate":
+            return agent.last_update.secs_since_epoch;
+          case "idleSlots":
+            return agent.status.slots_idle;
+          case "processingSlots":
+            return agent.status.slots_processing;
+          default:
+            return "";
+        }
+      }
+
+      // Special handling for issues column
+      if (key === "issues") {
+        const hasIssuesA = a.status.error !== null;
+        const hasIssuesB = b.status.error !== null;
+        if (hasIssuesA !== hasIssuesB) {
+          return direction === "ascending" ? (hasIssuesA ? 1 : -1) : (hasIssuesA ? -1 : 1);
+        }
+        const errorA = a.status.error || "";
+        const errorB = b.status.error || "";
+        if (errorA < errorB) return direction === "ascending" ? -1 : 1;
+        if (errorA > errorB) return direction === "ascending" ? 1 : -1;
+        return 0;
+      }
+
+      const valueA = getValue(a, key);
+      const valueB = getValue(b, key);
+
+      // Handle string comparison
+      if (typeof valueA === "string" && typeof valueB === "string") {
+        if (valueA < valueB) return direction === "ascending" ? -1 : 1;
+        if (valueA > valueB) return direction === "ascending" ? 1 : -1;
+        return 0;
+      }
+
+      // Handle numeric comparison
+      if (typeof valueA === "number" && typeof valueB === "number") {
+        if (valueA < valueB) return direction === "ascending" ? -1 : 1;
+        if (valueA > valueB) return direction === "ascending" ? 1 : -1;
+        return 0;
+      }
+
+      return 0;
+    });
+    return sortableAgents;
+  }
+
+  function requestSort(key: SortColumn) {
+    let direction: "ascending" | "descending" = "ascending";
+    if (sortConfig.key === key && sortConfig.direction === "ascending") {
+      direction = "descending";
+    }
+    setSortConfig({ key, direction });
+  }
+
+  const sortedAgents = sortAgents(agents);
+
   return (
     <table className={agentsTable}>
       <thead>
         <tr>
-          <th>Name</th>
-          <th>Issues</th>
-          <th>Llama.cpp address</th>
-          <th>Last update</th>
-          <th>Idle slots</th>
-          <th>Processing slots</th>
+          <th onClick={function () { requestSort("name"); }}>
+            Name{getSortIndicator(sortConfig, "name")}
+          </th>
+          <th onClick={function () { requestSort("model"); }}>
+            Model{getSortIndicator(sortConfig, "model")}
+          </th>
+          <th onClick={function () { requestSort("issues"); }}>
+            Issues{getSortIndicator(sortConfig, "issues")}
+          </th>
+          <th onClick={function () { requestSort("llamacppAddr"); }}>
+            Llama.cpp address{getSortIndicator(sortConfig, "llamacppAddr")}
+          </th>
+          <th onClick={function () { requestSort("lastUpdate"); }}>
+            Last update{getSortIndicator(sortConfig, "lastUpdate")}
+          </th>
+          <th onClick={function () { requestSort("idleSlots"); }}>
+            Idle slots{getSortIndicator(sortConfig, "idleSlots")}
+          </th>
+          <th onClick={function () { requestSort("processingSlots"); }}>
+            Processing slots{getSortIndicator(sortConfig, "processingSlots")}
+          </th>
         </tr>
       </thead>
       <tbody>
-        {agents.map(function ({
+        {sortedAgents.map(function ({
           agent_id,
           last_update,
           quarantined_until,
@@ -47,13 +165,12 @@ export function AgentsList({ agents }: { agents: Array<Agent> }) {
             quarantined_until;
 
           return (
-            <tr
-              className={clsx(agentRow, {
-                [agentRowError]: hasIssues,
-              })}
-              key={agent_id}
-            >
+              <tr
+                className={clsx(agentRow, hasIssues ? agentRowError : undefined)}
+                key={agent_id}
+              >
               <td>{status.agent_name}</td>
+              <td>{status.model}</td>
               <td>
                 {status.error && (
                   <>
diff --git a/resources/ts/components/Dashboard.module.css b/resources/ts/components/Dashboard.module.css
index 86f49b91..85458e4e 100644
--- a/resources/ts/components/Dashboard.module.css
+++ b/resources/ts/components/Dashboard.module.css
@@ -7,11 +7,29 @@
   th {
     border: 1px solid var(--color-border);
     padding: var(--spacing-base);
+    cursor: pointer;
 
     p + p {
       margin-top: var(--spacing-half);
     }
   }
+
+  th:hover {
+    background-color: var(--color-hover);
+  }
+}
+
+.sortIndicator {
+  margin-left: var(--spacing-half);
+  font-size: 0.8em;
+}
+
+.sortIndicatorAsc {
+  color: var(--color-success);
+}
+
+.sortIndicatorDesc {
+  color: var(--color-error);
 }
 
 .agentRow.agentRowError {
diff --git a/resources/ts/schemas/Agent.ts b/resources/ts/schemas/Agent.ts
index 678bd124..c2eea84d 100644
--- a/resources/ts/schemas/Agent.ts
+++ b/resources/ts/schemas/Agent.ts
@@ -5,6 +5,7 @@ import { StatusUpdateSchema } from "./StatusUpdate";
 export const AgentSchema = z
   .object({
     agent_id: z.string(),
+    model: z.string().nullable(),
     last_update: z.object({
       nanos_since_epoch: z.number(),
       secs_since_epoch: z.number(),
diff --git a/resources/ts/schemas/StatusUpdate.ts b/resources/ts/schemas/StatusUpdate.ts
index 7543d925..52747e5b 100644
--- a/resources/ts/schemas/StatusUpdate.ts
+++ b/resources/ts/schemas/StatusUpdate.ts
@@ -14,6 +14,7 @@ export const StatusUpdateSchema = z
     is_unexpected_response_status: z.boolean().nullable(),
     slots_idle: z.number(),
     slots_processing: z.number(),
+    model: z.string().nullable(),
   })
   .strict();
 
diff --git a/src/agent/monitoring_service.rs b/src/agent/monitoring_service.rs
index 4837493f..1eb14658 100644
--- a/src/agent/monitoring_service.rs
+++ b/src/agent/monitoring_service.rs
@@ -23,6 +23,7 @@ pub struct MonitoringService {
     monitoring_interval: Duration,
     name: Option<String>,
     status_update_tx: Sender<Bytes>,
+    check_model: bool, // Store the check_model flag
 }
 
 impl MonitoringService {
@@ -32,6 +33,7 @@ impl MonitoringService {
         monitoring_interval: Duration,
         name: Option<String>,
         status_update_tx: Sender<Bytes>,
+        check_model: bool, // Include the check_model flag
     ) -> Result<Self> {
         Ok(MonitoringService {
             external_llamacpp_addr,
@@ -39,6 +41,7 @@ impl MonitoringService {
             monitoring_interval,
             name,
             status_update_tx,
+            check_model,
         })
     }
 
@@ -50,6 +53,15 @@ impl MonitoringService {
             .filter(|slot| slot.is_processing)
             .count();
 
+        let model: Option<String> = if self.check_model {
+            match self.llamacpp_client.get_model().await {
+                Ok(model) => model,
+                Err(_) => None,
+            }
+        } else {
+            Some("".to_string())
+        };
+
         StatusUpdate {
             agent_name: self.name.to_owned(),
             error: slots_response.error,
@@ -63,6 +75,7 @@ impl MonitoringService {
             is_unexpected_response_status: slots_response.is_unexpected_response_status,
             slots_idle: slots_response.slots.len() - slots_processing,
             slots_processing,
+            model,
         }
     }
 
@@ -109,4 +122,4 @@ impl Service for MonitoringService {
     fn threads(&self) -> Option<usize> {
         Some(1)
     }
-}
+}
\ No newline at end of file
diff --git a/src/balancer/proxy_service.rs b/src/balancer/proxy_service.rs
index cfab2dfd..62b243f5 100644
--- a/src/balancer/proxy_service.rs
+++ b/src/balancer/proxy_service.rs
@@ -1,11 +1,13 @@
 use std::sync::atomic::AtomicUsize;
 use std::sync::atomic::Ordering;
 use std::sync::Arc;
+use std::sync::LazyLock;
 use std::time::Duration;
 
 use async_trait::async_trait;
 use bytes::Bytes;
 use log::error;
+use log::info;
 use pingora::http::RequestHeader;
 use pingora::proxy::ProxyHttp;
 use pingora::proxy::Session;
@@ -17,6 +19,51 @@ use pingora::Result;
 use crate::balancer::request_context::RequestContext;
 use crate::balancer::upstream_peer_pool::UpstreamPeerPool;
 
+static MODEL_REGEX: LazyLock<regex::Regex> = LazyLock::new(|| {
+    regex::Regex::new(r#""model"\s*:\s*"([^"]*)""#).expect("model regex is valid")
+});
+
+/// Extract the "model" field from a JSON request body.
+///
+/// Strategy:
+/// 1. If the body contains invalid UTF-8, truncate to the last valid boundary.
+/// 2. Try proper JSON parsing first (most reliable, no false positives).
+/// 3. If JSON parsing fails, try regex extraction as a fallback.
+fn extract_model_from_body(body_bytes: &Bytes) -> Option<String> {
+    // Handle invalid UTF-8 by truncating to the last valid boundary
+    let effective_bytes = match std::str::from_utf8(body_bytes) {
+        Ok(_) => body_bytes.clone(),
+        Err(e) => {
+            let valid_up_to = e.valid_up_to();
+            info!("Invalid UTF-8 in request body. Truncating from {} bytes to {} bytes for model extraction.", body_bytes.len(), valid_up_to);
+            body_bytes.slice(0..valid_up_to)
+        }
+    };
+
+    // Try proper JSON parsing first
+    if let Ok(json_value) = serde_json::from_slice::<serde_json::Value>(&effective_bytes) {
+        if let Some(model) = json_value.get("model").and_then(|v| v.as_str()) {
+            let model = model.to_string();
+            info!("Model in request: {:?}", model);
+            return Some(model);
+        }
+    }
+
+    // Fallback: regex extraction on the raw text
+    info!("Failed to parse JSON payload, trying regex extraction");
+    let body_str = String::from_utf8_lossy(&effective_bytes);
+    if let Some(caps) = MODEL_REGEX.captures(&body_str) {
+        if let Some(model_match) = caps.get(1) {
+            let model = model_match.as_str().to_string();
+            info!("Model via regex: {:?}", model);
+            return Some(model);
+        }
+    }
+
+    info!("Failed to extract model from request body");
+    None
+}
+
 struct RequestBufferGuard<'a>(&'a AtomicUsize);
 
 impl<'a> RequestBufferGuard<'a> {
@@ -41,6 +88,7 @@ pub struct ProxyService {
     buffered_request_timeout: Duration,
     max_buffered_requests: usize,
     rewrite_host_header: bool,
+    check_model: bool,
     slots_endpoint_enable: bool,
     upstream_peer_pool: Arc<UpstreamPeerPool>,
 }
@@ -48,6 +96,7 @@ pub struct ProxyService {
 impl ProxyService {
     pub fn new(
         rewrite_host_header: bool,
+        check_model: bool,
         slots_endpoint_enable: bool,
         upstream_peer_pool: Arc<UpstreamPeerPool>,
         buffered_request_timeout: Duration,
@@ -55,6 +104,7 @@ impl ProxyService {
     ) -> Self {
         Self {
             rewrite_host_header,
+            check_model,
             slots_endpoint_enable,
             upstream_peer_pool,
             buffered_request_timeout,
@@ -73,6 +123,7 @@ impl ProxyHttp for ProxyService {
             slot_taken: false,
             upstream_peer_pool: self.upstream_peer_pool.clone(),
             uses_slots: false,
+            requested_model: None,
         }
     }
 
@@ -180,10 +231,83 @@ impl ProxyHttp for ProxyService {
             }
             "/chat/completions" => true,
             "/completion" => true,
+            "/v1/completions" => true,
             "/v1/chat/completions" => true,
             _ => false,
         };
 
+        info!("upstream_peer - {:?} request | rewrite_host_header? {} check_model? {}", session.req_header().method, self.rewrite_host_header, self.check_model);
+
+        // Check if the request method is POST and the content type is JSON
+        if self.check_model && ctx.uses_slots {
+            info!("Checking model...");
+            if session.req_header().method == "POST" {
+                // Check if the content type is application/json
+                if let Some(content_type) = session.get_header("Content-Type") {
+                    if let Ok(content_type_str) = content_type.to_str() {
+                        if content_type_str.contains("application/json") {
+                            // Enable retry buffering to preserve the request body, reference: https://github.com/cloudflare/pingora/issues/349#issuecomment-2377277028
+                            session.enable_retry_buffering();
+
+                            // Read one chunk from the body. The model field is typically
+                            // the first field in the JSON, so one read is sufficient.
+                            // The retry buffer (64KB) captures what was read.
+                            let read_result = session.read_body_or_idle(false).await;
+                            match read_result {
+                                Ok(Some(_)) => {
+                                    // Check if the retry buffer was truncated (body > 64KB).
+                                    // If truncated, get_retry_buffer() returns None, so we
+                                    // cannot extract the model and must reject the request.
+                                    if session.retry_buffer_truncated() {
+                                        error!("Request body exceeds 64KB retry buffer limit, cannot determine model");
+                                        session
+                                            .respond_error(pingora::http::StatusCode::BAD_REQUEST.as_u16())
+                                            .await?;
+                                        return Err(Error::new_down(pingora::ErrorType::ConnectRefused));
+                                    }
+
+                                    if let Some(body_bytes) = session.get_retry_buffer() {
+                                        ctx.requested_model = extract_model_from_body(&body_bytes);
+                                    } else {
+                                        info!("Retry buffer is empty after reading body chunk");
+                                    }
+                                }
+                                Ok(None) => {
+                                    info!("Request body is empty");
+                                }
+                                Err(e) => {
+                                    error!("Failed to read request body: {e}");
+                                    session
+                                        .respond_error(pingora::http::StatusCode::BAD_REQUEST.as_u16())
+                                        .await?;
+                                    return Err(Error::new_down(pingora::ErrorType::ConnectRefused));
+                                }
+                            }
+                        }
+                    }
+                }
+            }
+
+            // abort if model has not been set
+            if ctx.requested_model.is_none() {
+                info!("Model missing in request");
+                session
+                    .respond_error(pingora::http::StatusCode::BAD_REQUEST.as_u16())
+                    .await?;
+
+                return Err(Error::new_down(pingora::ErrorType::ConnectRefused));
+            } else if !ctx.has_peer_supporting_model() {
+                info!("Model {:?} not supported by upstream", ctx.requested_model);
+                session
+                    .respond_error(pingora::http::StatusCode::NOT_FOUND.as_u16())
+                    .await?;
+
+                return Err(Error::new_down(pingora::ErrorType::ConnectRefused));
+            } else {
+                info!("Model {:?}", ctx.requested_model);
+            }
+        }
+
         let peer = tokio::select! {
             result = async {
                 loop {
@@ -232,7 +356,11 @@ impl ProxyHttp for ProxyService {
             }
         };
 
-        Ok(HttpPeer::new(peer.status.external_llamacpp_addr, false, "".into()).into())
+        let mut http_peer = HttpPeer::new(peer.status.external_llamacpp_addr, false, "".into());
+        // Expire pooled upstream connections after 30s of inactivity to prevent
+        // stale connections from accumulating (idle_timeout was None by default).
+        http_peer.options.idle_timeout = Some(Duration::from_secs(30));
+        Ok(http_peer.into())
     }
 
     async fn upstream_request_filter(
diff --git a/src/balancer/request_context.rs b/src/balancer/request_context.rs
index eb8b1c11..da308717 100644
--- a/src/balancer/request_context.rs
+++ b/src/balancer/request_context.rs
@@ -2,6 +2,7 @@ use std::sync::Arc;
 
 use anyhow::anyhow;
 use log::error;
+use log::info;
 use pingora::Error;
 use pingora::Result;
 
@@ -13,6 +14,7 @@ pub struct RequestContext {
     pub selected_peer: Option<UpstreamPeer>,
     pub upstream_peer_pool: Arc<UpstreamPeerPool>,
     pub uses_slots: bool,
+    pub requested_model: Option<String>,
 }
 
 impl RequestContext {
@@ -30,16 +32,19 @@ impl RequestContext {
         }
     }
 
-    pub fn use_best_peer_and_take_slot(&mut self) -> anyhow::Result<Option<UpstreamPeer>> {
+    pub fn use_best_peer_and_take_slot(&mut self, model: Option<String>) -> anyhow::Result<Option<UpstreamPeer>> {
         if let Some(peer) = self.upstream_peer_pool.with_agents_write(|agents| {
+            let model_str = model.as_deref().unwrap_or("");
             for peer in agents.iter_mut() {
-                if peer.is_usable() {
-                    peer.take_slot()?;
+                let is_usable = peer.is_usable();
+                let is_usable_for_model = peer.is_usable_for_model(model_str);
 
+                if is_usable && (model.is_none() || is_usable_for_model) {
+                    info!("Peer {} is usable: {}, usable for model '{}': {}", peer.agent_id, is_usable, model_str, is_usable_for_model);
+                    peer.take_slot()?;
                     return Ok(Some(peer.clone()));
                 }
             }
-
             Ok(None)
         })? {
             self.upstream_peer_pool.restore_integrity()?;
@@ -52,11 +57,26 @@ impl RequestContext {
         }
     }
 
+    pub fn has_peer_supporting_model(&self) -> bool {
+        let model_str = self.requested_model.as_deref().unwrap_or("");
+        match self.upstream_peer_pool.with_agents_read(|agents| {
+            for peer in agents.iter() {
+                if peer.supports_model(model_str) {
+                    return Ok(true);
+                }
+            }
+            Ok(false)
+        }) {
+            Ok(result) => result,
+            Err(_) => false, // or handle the error as needed
+        }
+    }
+
     pub fn select_upstream_peer(&mut self) -> Result<()> {
         let result_option_peer = if self.uses_slots && !self.slot_taken {
-            self.use_best_peer_and_take_slot()
+            self.use_best_peer_and_take_slot(self.requested_model.clone())
         } else {
-            self.upstream_peer_pool.use_best_peer()
+            self.upstream_peer_pool.use_best_peer(self.requested_model.clone())
         };
 
         self.selected_peer = match result_option_peer {
@@ -95,6 +115,7 @@ mod tests {
             selected_peer: None,
             upstream_peer_pool,
             uses_slots: true,
+            requested_model: Some("llama3".to_string()),
         }
     }
 
@@ -105,7 +126,7 @@ mod tests {
 
         pool.register_status_update("test_agent", mock_status_update("test_agent", 0, 0))?;
 
-        assert!(ctx.use_best_peer_and_take_slot().unwrap().is_none());
+        assert!(ctx.use_best_peer_and_take_slot(ctx.requested_model.clone()).unwrap().is_none());
 
         assert!(!ctx.slot_taken);
         assert_eq!(ctx.selected_peer, None);
diff --git a/src/balancer/status_update.rs b/src/balancer/status_update.rs
index 6f2c5409..5e811640 100644
--- a/src/balancer/status_update.rs
+++ b/src/balancer/status_update.rs
@@ -20,6 +20,7 @@ pub struct StatusUpdate {
     pub is_unexpected_response_status: Option<bool>,
     pub slots_idle: usize,
     pub slots_processing: usize,
+    pub model: Option<String>,
 }
 
 impl StatusUpdate {
diff --git a/src/balancer/test/mock_status_update.rs b/src/balancer/test/mock_status_update.rs
index 07d12a1f..8cc9da41 100644
--- a/src/balancer/test/mock_status_update.rs
+++ b/src/balancer/test/mock_status_update.rs
@@ -22,5 +22,6 @@ pub fn mock_status_update(
         is_unexpected_response_status: Some(false),
         slots_idle,
         slots_processing,
+        model: Some("llama3".to_string()),
     }
 }
diff --git a/src/balancer/upstream_peer.rs b/src/balancer/upstream_peer.rs
index a8eaff95..24b52d62 100644
--- a/src/balancer/upstream_peer.rs
+++ b/src/balancer/upstream_peer.rs
@@ -13,6 +13,7 @@ use crate::balancer::status_update::StatusUpdate;
 #[derive(Clone, Debug, Eq, Serialize, Deserialize)]
 pub struct UpstreamPeer {
     pub agent_id: String,
+    pub model: Option<String>,
     pub last_update: SystemTime,
     pub quarantined_until: Option<SystemTime>,
     pub slots_taken: usize,
@@ -24,6 +25,7 @@ impl UpstreamPeer {
     pub fn new_from_status_update(agent_id: String, status: StatusUpdate) -> Self {
         Self {
             agent_id,
+            model: status.model.clone(),
             last_update: SystemTime::now(),
             quarantined_until: None,
             slots_taken: 0,
@@ -36,6 +38,14 @@ impl UpstreamPeer {
         !self.status.has_issues() && self.status.slots_idle > 0 && self.quarantined_until.is_none()
     }
 
+    pub fn supports_model(&self, requested_model: &str) -> bool {
+        requested_model.is_empty() || self.model.as_deref() == Some(requested_model)
+    }
+
+    pub fn is_usable_for_model(&self, requested_model: &str) -> bool {
+        self.is_usable() && (requested_model.is_empty() || self.model.as_deref() == Some(requested_model))
+    }
+
     pub fn release_slot(&mut self) -> Result<()> {
         if self.slots_taken < 1 {
             return Err(anyhow!(
@@ -59,6 +69,7 @@ impl UpstreamPeer {
         self.last_update = SystemTime::now();
         self.quarantined_until = None;
         self.slots_taken_since_last_status_update = 0;
+        self.model = status_update.model.clone();
         self.status = status_update;
     }
 
@@ -110,6 +121,7 @@ mod tests {
     fn create_test_peer() -> UpstreamPeer {
         UpstreamPeer {
             agent_id: "test_agent".to_string(),
+            model: Some("llama3".to_string()),
             last_update: SystemTime::now(),
             quarantined_until: None,
             slots_taken: 0,
@@ -130,6 +142,7 @@ mod tests {
                 is_unexpected_response_status: None,
                 slots_idle: 5,
                 slots_processing: 0,
+                model: Some("llama3".to_string()),
             },
         }
     }
@@ -177,7 +190,6 @@ mod tests {
     #[test]
     fn test_update_status() {
         let mut peer = create_test_peer();
-
         let slots: Vec<Slot> = vec![];
         let slots_idle = slots.iter().filter(|slot| !slot.is_processing).count();
 
@@ -194,6 +206,7 @@ mod tests {
             is_unexpected_response_status: None,
             slots_idle,
             slots_processing: slots.len() - slots_idle,
+            model: Some("llama3".to_string())
         };
 
         peer.update_status(status_update);
diff --git a/src/balancer/upstream_peer_pool.rs b/src/balancer/upstream_peer_pool.rs
index 2e6924a9..4593bab9 100644
--- a/src/balancer/upstream_peer_pool.rs
+++ b/src/balancer/upstream_peer_pool.rs
@@ -2,6 +2,7 @@ use std::sync::atomic::AtomicUsize;
 use std::sync::RwLock;
 use std::time::Duration;
 use std::time::SystemTime;
+use log::info;
 
 use anyhow::anyhow;
 use anyhow::Result;
@@ -159,10 +160,15 @@ impl UpstreamPeerPool {
         })
     }
 
-    pub fn use_best_peer(&self) -> Result<Option<UpstreamPeer>> {
-        self.with_agents_read(|agents| {
+    pub fn use_best_peer(&self, model: Option<String>) -> Result<Option<UpstreamPeer>> {
+        self.with_agents_write(|agents| {
             for peer in agents.iter() {
-                if peer.is_usable() {
+                let model_str = model.as_deref().unwrap_or("");
+                let is_usable = peer.is_usable();
+                let is_usable_for_model = peer.is_usable_for_model(model_str);
+
+                if is_usable && (model.is_none() || is_usable_for_model) {
+                    info!("Peer {} is usable: {}, usable for model '{}': {}", peer.agent_id, is_usable, model_str, is_usable_for_model);
                     return Ok(Some(peer.clone()));
                 }
             }
@@ -263,7 +269,7 @@ mod tests {
         pool.register_status_update("test2", mock_status_update("test2", 3, 0))?;
         pool.register_status_update("test3", mock_status_update("test3", 0, 0))?;
 
-        let best_peer = pool.use_best_peer()?.unwrap();
+        let best_peer = pool.use_best_peer(None)?.unwrap();
 
         assert_eq!(best_peer.agent_id, "test1");
         assert_eq!(best_peer.status.slots_idle, 5);
diff --git a/src/cmd/agent.rs b/src/cmd/agent.rs
index aef1b711..578f5dbe 100644
--- a/src/cmd/agent.rs
+++ b/src/cmd/agent.rs
@@ -18,6 +18,7 @@ pub fn handle(
     management_addr: SocketAddr,
     monitoring_interval: Duration,
     name: Option<String>,
+    check_model: bool, // Include the check_model flag
 ) -> Result<()> {
     let (status_update_tx, _status_update_rx) = channel::<Bytes>(1);
 
@@ -29,6 +30,7 @@ pub fn handle(
         monitoring_interval,
         name,
         status_update_tx.clone(),
+        check_model, // Pass the check_model flag
     )?;
 
     let reporting_service = ReportingService::new(management_addr, status_update_tx)?;
@@ -45,4 +47,4 @@ pub fn handle(
     pingora_server.add_service(monitoring_service);
     pingora_server.add_service(reporting_service);
     pingora_server.run_forever();
-}
+}
\ No newline at end of file
diff --git a/src/cmd/balancer.rs b/src/cmd/balancer.rs
index dbca8bc4..f5ac3d4d 100644
--- a/src/cmd/balancer.rs
+++ b/src/cmd/balancer.rs
@@ -2,6 +2,7 @@ use std::net::SocketAddr;
 use std::sync::Arc;
 #[cfg(feature = "statsd_reporter")]
 use std::time::Duration;
+use log::info;
 
 use anyhow::Result;
 use pingora::proxy::http_proxy_service;
@@ -24,6 +25,7 @@ pub fn handle(
     metrics_endpoint_enable: bool,
     reverseproxy_addr: &SocketAddr,
     rewrite_host_header: bool,
+    check_model: bool,
     slots_endpoint_enable: bool,
     #[cfg(feature = "statsd_reporter")] statsd_addr: Option<SocketAddr>,
     #[cfg(feature = "statsd_reporter")] statsd_prefix: String,
@@ -45,6 +47,7 @@ pub fn handle(
         &pingora_server.configuration,
         ProxyService::new(
             rewrite_host_header,
+            check_model,
             slots_endpoint_enable,
             upstream_peer_pool.clone(),
             buffered_request_timeout,
@@ -76,5 +79,7 @@ pub fn handle(
         pingora_server.add_service(statsd_service);
     }
 
+    info!("rewrite_host_header? {} check_model? {} slots_endpoint_enable? {}", rewrite_host_header, check_model, slots_endpoint_enable);
+
     pingora_server.run_forever();
 }
diff --git a/src/llamacpp/llamacpp_client.rs b/src/llamacpp/llamacpp_client.rs
index 742025f4..4e802164 100644
--- a/src/llamacpp/llamacpp_client.rs
+++ b/src/llamacpp/llamacpp_client.rs
@@ -1,16 +1,19 @@
 use std::net::SocketAddr;
 use std::time::Duration;
 
+use anyhow::anyhow;
 use anyhow::Result;
 use reqwest::header;
 use url::Url;
 
 use crate::llamacpp::slot::Slot;
 use crate::llamacpp::slots_response::SlotsResponse;
+use crate::llamacpp::models_response::ModelsResponse;
 
 pub struct LlamacppClient {
     client: reqwest::Client,
     slots_endpoint_url: String,
+    models_endpoint_url: String,
 }
 
 impl LlamacppClient {
@@ -36,6 +39,7 @@ impl LlamacppClient {
         Ok(Self {
             client: builder.build()?,
             slots_endpoint_url: Url::parse(&format!("http://{addr}/slots"))?.to_string(),
+            models_endpoint_url: Url::parse(&format!("http://{addr}/v1/models"))?.to_string(),
         })
     }
 
@@ -115,4 +119,40 @@ impl LlamacppClient {
             },
         }
     }
+
+    pub async fn get_model(&self) -> Result<Option<String>> {
+        let url = self.models_endpoint_url.to_owned();
+
+        let response = match self.client.get(url.clone()).send().await {
+            Ok(resp) => resp,
+            Err(err) => {
+                return Err(anyhow!(
+                    "Request to '{}' failed: '{}'; connect issue: {}; decode issue: {}; request issue: {}; status issue: {}; status: {:?}",
+                    url,
+                    err,
+                    err.is_connect(),
+                    err.is_decode(),
+                    err.is_request(),
+                    err.is_status(),
+                    err.status()
+                ));
+            }
+        };
+
+        match response.status() {
+            reqwest::StatusCode::OK => {
+                let models_response: ModelsResponse = response.json().await?;
+                if let Some(models) = models_response.models {
+                    if models.is_empty() {
+                        Ok(None)
+                    } else {
+                        Ok(models.first().and_then(|m| Some(m.model.clone())))
+                    }
+                } else {
+                    Ok(None)
+                }
+            },
+            _ => Err(anyhow!("Unexpected response status")),
+        }
+    }
 }
diff --git a/src/llamacpp/mod.rs b/src/llamacpp/mod.rs
index 2b1e1d24..68f5cae8 100644
--- a/src/llamacpp/mod.rs
+++ b/src/llamacpp/mod.rs
@@ -1,3 +1,4 @@
 pub mod llamacpp_client;
 pub mod slot;
 pub mod slots_response;
+pub mod models_response;
diff --git a/src/llamacpp/models_response.rs b/src/llamacpp/models_response.rs
new file mode 100644
index 00000000..d7d04b50
--- /dev/null
+++ b/src/llamacpp/models_response.rs
@@ -0,0 +1,13 @@
+// paddler/src/llamacpp/models_response.rs
+use serde::Deserialize;
+
+#[derive(Debug, Deserialize)]
+pub struct ModelsResponse {
+    pub models: Option<Vec<Model>>,
+}
+
+#[derive(Debug, Deserialize)]
+pub struct Model {
+    pub model: String,
+    // Add other fields as needed
+}
diff --git a/src/main.rs b/src/main.rs
index 007721c3..baa24430 100644
--- a/src/main.rs
+++ b/src/main.rs
@@ -87,6 +87,10 @@ enum Commands {
         #[arg(long)]
         /// Name of the agent (optional)
         name: Option<String>,
+
+        #[arg(long)]
+        /// Flag whether to check the model served by llama.cpp and reject requests for other models
+        check_model: bool,
     },
     /// Balances incoming requests to llama.cpp instances and optionally provides a web dashboard
     Balancer {
@@ -134,6 +138,10 @@ enum Commands {
         /// Enable the slots endpoint (not recommended)
         slots_endpoint_enable: bool,
 
+        #[arg(long)]
+        /// Flag to check the model served by llama.cpp and reject requests for other models
+        check_model: bool,
+
         #[cfg(feature = "statsd_reporter")]
         #[arg(long, value_parser = parse_socket_addr)]
         /// Address of the statsd server to report metrics to
@@ -171,6 +179,7 @@ fn main() -> Result<()> {
             management_addr,
             monitoring_interval,
             name,
+            check_model,
         }) => cmd::agent::handle(
             match external_llamacpp_addr {
                 Some(addr) => addr.to_owned(),
@@ -181,6 +190,7 @@ fn main() -> Result<()> {
             management_addr.to_owned(),
             monitoring_interval.to_owned(),
             name.to_owned(),
+            *check_model
         ),
         Some(Commands::Balancer {
             buffered_request_timeout,
@@ -192,6 +202,7 @@ fn main() -> Result<()> {
             metrics_endpoint_enable,
             reverseproxy_addr,
             rewrite_host_header,
+            check_model,
             slots_endpoint_enable,
             #[cfg(feature = "statsd_reporter")]
             statsd_addr,
@@ -213,6 +224,7 @@ fn main() -> Result<()> {
                 metrics_endpoint_enable.to_owned(),
                 reverseproxy_addr,
                 rewrite_host_header.to_owned(),
+                *check_model,
                 slots_endpoint_enable.to_owned(),
                 #[cfg(feature = "statsd_reporter")]
                 statsd_addr.to_owned(),
diff --git a/integration_tests/.gitignore b/tests/integration_tests/.gitignore
similarity index 100%
rename from integration_tests/.gitignore
rename to tests/integration_tests/.gitignore
diff --git a/integration_tests/Cargo.lock b/tests/integration_tests/Cargo.lock
similarity index 100%
rename from integration_tests/Cargo.lock
rename to tests/integration_tests/Cargo.lock
diff --git a/integration_tests/Cargo.toml b/tests/integration_tests/Cargo.toml
similarity index 100%
rename from integration_tests/Cargo.toml
rename to tests/integration_tests/Cargo.toml
diff --git a/integration_tests/Makefile b/tests/integration_tests/Makefile
similarity index 100%
rename from integration_tests/Makefile
rename to tests/integration_tests/Makefile
diff --git a/integration_tests/rust-toolchain b/tests/integration_tests/rust-toolchain
similarity index 100%
rename from integration_tests/rust-toolchain
rename to tests/integration_tests/rust-toolchain
diff --git a/integration_tests/rustfmt.toml b/tests/integration_tests/rustfmt.toml
similarity index 100%
rename from integration_tests/rustfmt.toml
rename to tests/integration_tests/rustfmt.toml
diff --git a/integration_tests/tests/features/agent/agent_monitors.feature b/tests/integration_tests/tests/features/agent/agent_monitors.feature
similarity index 100%
rename from integration_tests/tests/features/agent/agent_monitors.feature
rename to tests/integration_tests/tests/features/agent/agent_monitors.feature
diff --git a/integration_tests/tests/features/balancer/balance_requests.feature b/tests/integration_tests/tests/features/balancer/balance_requests.feature
similarity index 100%
rename from integration_tests/tests/features/balancer/balance_requests.feature
rename to tests/integration_tests/tests/features/balancer/balance_requests.feature
diff --git a/integration_tests/tests/features/balancer/cors_headers.feature b/tests/integration_tests/tests/features/balancer/cors_headers.feature
similarity index 100%
rename from integration_tests/tests/features/balancer/cors_headers.feature
rename to tests/integration_tests/tests/features/balancer/cors_headers.feature
diff --git a/integration_tests/tests/features/balancer/export_metrics.feature b/tests/integration_tests/tests/features/balancer/export_metrics.feature
similarity index 100%
rename from integration_tests/tests/features/balancer/export_metrics.feature
rename to tests/integration_tests/tests/features/balancer/export_metrics.feature
diff --git a/integration_tests/tests/features/balancer/report_metrics.feature b/tests/integration_tests/tests/features/balancer/report_metrics.feature
similarity index 100%
rename from integration_tests/tests/features/balancer/report_metrics.feature
rename to tests/integration_tests/tests/features/balancer/report_metrics.feature
diff --git a/integration_tests/tests/fixtures/llamacpp-server-mock.mjs b/tests/integration_tests/tests/fixtures/llamacpp-server-mock.mjs
similarity index 100%
rename from integration_tests/tests/fixtures/llamacpp-server-mock.mjs
rename to tests/integration_tests/tests/fixtures/llamacpp-server-mock.mjs
diff --git a/integration_tests/tests/fixtures/statsd-server-mock.mjs b/tests/integration_tests/tests/fixtures/statsd-server-mock.mjs
similarity index 100%
rename from integration_tests/tests/fixtures/statsd-server-mock.mjs
rename to tests/integration_tests/tests/fixtures/statsd-server-mock.mjs
diff --git a/integration_tests/tests/paddler/agent_response.rs b/tests/integration_tests/tests/paddler/agent_response.rs
similarity index 100%
rename from integration_tests/tests/paddler/agent_response.rs
rename to tests/integration_tests/tests/paddler/agent_response.rs
diff --git a/integration_tests/tests/paddler/agents_collection.rs b/tests/integration_tests/tests/paddler/agents_collection.rs
similarity index 100%
rename from integration_tests/tests/paddler/agents_collection.rs
rename to tests/integration_tests/tests/paddler/agents_collection.rs
diff --git a/integration_tests/tests/paddler/assert_balancer_table.rs b/tests/integration_tests/tests/paddler/assert_balancer_table.rs
similarity index 100%
rename from integration_tests/tests/paddler/assert_balancer_table.rs
rename to tests/integration_tests/tests/paddler/assert_balancer_table.rs
diff --git a/integration_tests/tests/paddler/balancer_instance.rs b/tests/integration_tests/tests/paddler/balancer_instance.rs
similarity index 100%
rename from integration_tests/tests/paddler/balancer_instance.rs
rename to tests/integration_tests/tests/paddler/balancer_instance.rs
diff --git a/integration_tests/tests/paddler/balancer_management_client.rs b/tests/integration_tests/tests/paddler/balancer_management_client.rs
similarity index 100%
rename from integration_tests/tests/paddler/balancer_management_client.rs
rename to tests/integration_tests/tests/paddler/balancer_management_client.rs
diff --git a/integration_tests/tests/paddler/expression/given_agent_is_registered.rs b/tests/integration_tests/tests/paddler/expression/given_agent_is_registered.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/given_agent_is_registered.rs
rename to tests/integration_tests/tests/paddler/expression/given_agent_is_registered.rs
diff --git a/integration_tests/tests/paddler/expression/given_agent_is_running.rs b/tests/integration_tests/tests/paddler/expression/given_agent_is_running.rs
similarity index 95%
rename from integration_tests/tests/paddler/expression/given_agent_is_running.rs
rename to tests/integration_tests/tests/paddler/expression/given_agent_is_running.rs
index 6fa35a7a..813fdc2e 100644
--- a/integration_tests/tests/paddler/expression/given_agent_is_running.rs
+++ b/tests/integration_tests/tests/paddler/expression/given_agent_is_running.rs
@@ -21,7 +21,7 @@ pub async fn given_agent_is_attached(
 
     world.agents.instances.insert(
         agent_name.clone(),
-        Command::new("../target/debug/paddler")
+        Command::new("../../target/debug/paddler")
             .arg("agent")
             .arg(format!("--name={agent_name}"))
             .arg(format!(
diff --git a/integration_tests/tests/paddler/expression/given_agent_monitors_llamacpp_every_milliseconds.rs b/tests/integration_tests/tests/paddler/expression/given_agent_monitors_llamacpp_every_milliseconds.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/given_agent_monitors_llamacpp_every_milliseconds.rs
rename to tests/integration_tests/tests/paddler/expression/given_agent_monitors_llamacpp_every_milliseconds.rs
diff --git a/integration_tests/tests/paddler/expression/given_balancer_allows_cors_host.rs b/tests/integration_tests/tests/paddler/expression/given_balancer_allows_cors_host.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/given_balancer_allows_cors_host.rs
rename to tests/integration_tests/tests/paddler/expression/given_balancer_allows_cors_host.rs
diff --git a/integration_tests/tests/paddler/expression/given_balancer_is_running.rs b/tests/integration_tests/tests/paddler/expression/given_balancer_is_running.rs
similarity index 95%
rename from integration_tests/tests/paddler/expression/given_balancer_is_running.rs
rename to tests/integration_tests/tests/paddler/expression/given_balancer_is_running.rs
index 965bbf12..bf44dfe0 100644
--- a/integration_tests/tests/paddler/expression/given_balancer_is_running.rs
+++ b/tests/integration_tests/tests/paddler/expression/given_balancer_is_running.rs
@@ -12,7 +12,7 @@ pub async fn given_balancer_is_running(world: &mut PaddlerWorld) -> Result<()> {
         return Err(anyhow::anyhow!("Balancer is already running"));
     }
 
-    let mut command = Command::new("../target/debug/paddler");
+    let mut command = Command::new("../../target/debug/paddler");
 
     command
         .arg("balancer")
diff --git a/integration_tests/tests/paddler/expression/given_balancer_reports_metrics_every_milliseconds.rs b/tests/integration_tests/tests/paddler/expression/given_balancer_reports_metrics_every_milliseconds.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/given_balancer_reports_metrics_every_milliseconds.rs
rename to tests/integration_tests/tests/paddler/expression/given_balancer_reports_metrics_every_milliseconds.rs
diff --git a/integration_tests/tests/paddler/expression/given_buffered_requests_timeout_after_milliseconds.rs b/tests/integration_tests/tests/paddler/expression/given_buffered_requests_timeout_after_milliseconds.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/given_buffered_requests_timeout_after_milliseconds.rs
rename to tests/integration_tests/tests/paddler/expression/given_buffered_requests_timeout_after_milliseconds.rs
diff --git a/integration_tests/tests/paddler/expression/given_llamacpp_completes_response_after_milliseconds.rs b/tests/integration_tests/tests/paddler/expression/given_llamacpp_completes_response_after_milliseconds.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/given_llamacpp_completes_response_after_milliseconds.rs
rename to tests/integration_tests/tests/paddler/expression/given_llamacpp_completes_response_after_milliseconds.rs
diff --git a/integration_tests/tests/paddler/expression/given_llamacpp_server_is_running.rs b/tests/integration_tests/tests/paddler/expression/given_llamacpp_server_is_running.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/given_llamacpp_server_is_running.rs
rename to tests/integration_tests/tests/paddler/expression/given_llamacpp_server_is_running.rs
diff --git a/integration_tests/tests/paddler/expression/given_request_buffering_is_disabled.rs b/tests/integration_tests/tests/paddler/expression/given_request_buffering_is_disabled.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/given_request_buffering_is_disabled.rs
rename to tests/integration_tests/tests/paddler/expression/given_request_buffering_is_disabled.rs
diff --git a/integration_tests/tests/paddler/expression/given_request_has_header_with_value.rs b/tests/integration_tests/tests/paddler/expression/given_request_has_header_with_value.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/given_request_has_header_with_value.rs
rename to tests/integration_tests/tests/paddler/expression/given_request_has_header_with_value.rs
diff --git a/integration_tests/tests/paddler/expression/given_statsd_is_running.rs b/tests/integration_tests/tests/paddler/expression/given_statsd_is_running.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/given_statsd_is_running.rs
rename to tests/integration_tests/tests/paddler/expression/given_statsd_is_running.rs
diff --git a/integration_tests/tests/paddler/expression/mod.rs b/tests/integration_tests/tests/paddler/expression/mod.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/mod.rs
rename to tests/integration_tests/tests/paddler/expression/mod.rs
diff --git a/integration_tests/tests/paddler/expression/then_average_metrics_are.rs b/tests/integration_tests/tests/paddler/expression/then_average_metrics_are.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/then_average_metrics_are.rs
rename to tests/integration_tests/tests/paddler/expression/then_average_metrics_are.rs
diff --git a/integration_tests/tests/paddler/expression/then_balancer_state_is.rs b/tests/integration_tests/tests/paddler/expression/then_balancer_state_is.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/then_balancer_state_is.rs
rename to tests/integration_tests/tests/paddler/expression/then_balancer_state_is.rs
diff --git a/integration_tests/tests/paddler/expression/then_exported_metrics_are_stored.rs b/tests/integration_tests/tests/paddler/expression/then_exported_metrics_are_stored.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/then_exported_metrics_are_stored.rs
rename to tests/integration_tests/tests/paddler/expression/then_exported_metrics_are_stored.rs
diff --git a/integration_tests/tests/paddler/expression/then_next_balancer_state_is.rs b/tests/integration_tests/tests/paddler/expression/then_next_balancer_state_is.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/then_next_balancer_state_is.rs
rename to tests/integration_tests/tests/paddler/expression/then_next_balancer_state_is.rs
diff --git a/integration_tests/tests/paddler/expression/then_reported_metrics_are_stored.rs b/tests/integration_tests/tests/paddler/expression/then_reported_metrics_are_stored.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/then_reported_metrics_are_stored.rs
rename to tests/integration_tests/tests/paddler/expression/then_reported_metrics_are_stored.rs
diff --git a/integration_tests/tests/paddler/expression/then_request_landed_in.rs b/tests/integration_tests/tests/paddler/expression/then_request_landed_in.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/then_request_landed_in.rs
rename to tests/integration_tests/tests/paddler/expression/then_request_landed_in.rs
diff --git a/integration_tests/tests/paddler/expression/then_response_code_is.rs b/tests/integration_tests/tests/paddler/expression/then_response_code_is.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/then_response_code_is.rs
rename to tests/integration_tests/tests/paddler/expression/then_response_code_is.rs
diff --git a/integration_tests/tests/paddler/expression/then_response_header_is.rs b/tests/integration_tests/tests/paddler/expression/then_response_header_is.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/then_response_header_is.rs
rename to tests/integration_tests/tests/paddler/expression/then_response_header_is.rs
diff --git a/integration_tests/tests/paddler/expression/then_response_header_is_not_present.rs b/tests/integration_tests/tests/paddler/expression/then_response_header_is_not_present.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/then_response_header_is_not_present.rs
rename to tests/integration_tests/tests/paddler/expression/then_response_header_is_not_present.rs
diff --git a/integration_tests/tests/paddler/expression/when_llamacpp_stops_running.rs b/tests/integration_tests/tests/paddler/expression/when_llamacpp_stops_running.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/when_llamacpp_stops_running.rs
rename to tests/integration_tests/tests/paddler/expression/when_llamacpp_stops_running.rs
diff --git a/integration_tests/tests/paddler/expression/when_multiple_requests_are_sent_to_path.rs b/tests/integration_tests/tests/paddler/expression/when_multiple_requests_are_sent_to_path.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/when_multiple_requests_are_sent_to_path.rs
rename to tests/integration_tests/tests/paddler/expression/when_multiple_requests_are_sent_to_path.rs
diff --git a/integration_tests/tests/paddler/expression/when_request_is_sent_to_management_endpoint.rs b/tests/integration_tests/tests/paddler/expression/when_request_is_sent_to_management_endpoint.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/when_request_is_sent_to_management_endpoint.rs
rename to tests/integration_tests/tests/paddler/expression/when_request_is_sent_to_management_endpoint.rs
diff --git a/integration_tests/tests/paddler/expression/when_request_is_sent_to_path.rs b/tests/integration_tests/tests/paddler/expression/when_request_is_sent_to_path.rs
similarity index 100%
rename from integration_tests/tests/paddler/expression/when_request_is_sent_to_path.rs
rename to tests/integration_tests/tests/paddler/expression/when_request_is_sent_to_path.rs
diff --git a/integration_tests/tests/paddler/llamacpp_instance.rs b/tests/integration_tests/tests/paddler/llamacpp_instance.rs
similarity index 100%
rename from integration_tests/tests/paddler/llamacpp_instance.rs
rename to tests/integration_tests/tests/paddler/llamacpp_instance.rs
diff --git a/integration_tests/tests/paddler/llamacpp_instance_collection.rs b/tests/integration_tests/tests/paddler/llamacpp_instance_collection.rs
similarity index 100%
rename from integration_tests/tests/paddler/llamacpp_instance_collection.rs
rename to tests/integration_tests/tests/paddler/llamacpp_instance_collection.rs
diff --git a/integration_tests/tests/paddler/main.rs b/tests/integration_tests/tests/paddler/main.rs
similarity index 100%
rename from integration_tests/tests/paddler/main.rs
rename to tests/integration_tests/tests/paddler/main.rs
diff --git a/integration_tests/tests/paddler/metrics.rs b/tests/integration_tests/tests/paddler/metrics.rs
similarity index 100%
rename from integration_tests/tests/paddler/metrics.rs
rename to tests/integration_tests/tests/paddler/metrics.rs
diff --git a/integration_tests/tests/paddler/paddler_world.rs b/tests/integration_tests/tests/paddler/paddler_world.rs
similarity index 100%
rename from integration_tests/tests/paddler/paddler_world.rs
rename to tests/integration_tests/tests/paddler/paddler_world.rs
diff --git a/integration_tests/tests/paddler/request_builder.rs b/tests/integration_tests/tests/paddler/request_builder.rs
similarity index 100%
rename from integration_tests/tests/paddler/request_builder.rs
rename to tests/integration_tests/tests/paddler/request_builder.rs
diff --git a/integration_tests/tests/paddler/request_headers_to_be_set.rs b/tests/integration_tests/tests/paddler/request_headers_to_be_set.rs
similarity index 100%
rename from integration_tests/tests/paddler/request_headers_to_be_set.rs
rename to tests/integration_tests/tests/paddler/request_headers_to_be_set.rs
diff --git a/integration_tests/tests/paddler/statsd_instance.rs b/tests/integration_tests/tests/paddler/statsd_instance.rs
similarity index 100%
rename from integration_tests/tests/paddler/statsd_instance.rs
rename to tests/integration_tests/tests/paddler/statsd_instance.rs
diff --git a/tests/scripts/test-model-detection.sh b/tests/scripts/test-model-detection.sh
new file mode 100755
index 00000000..a29660f5
--- /dev/null
+++ b/tests/scripts/test-model-detection.sh
@@ -0,0 +1,353 @@
+#!/usr/bin/env bash
+#
+# Test script for model detection changes in proxy_service.rs
+#
+# Usage:
+#   ./test-model-detection.sh                          # uses defaults (localhost)
+#   ./test-model-detection.sh deeplearning4.stockpulse.de
+#   PADDLER_PROXY_HOST=deeplearning4.stockpulse.de ./test-model-detection.sh
+#   PADDLER_PROXY_PORT=5000 PADDLER_MGMT_PORT=8085 PADDLER_MODEL=gemma4 ./test-model-detection.sh
+#
+# Environment variables (all optional, CLI arg overrides host):
+#   PADDLER_PROXY_HOST   - reverse proxy hostname/IP   (default: localhost)
+#   PADDLER_PROXY_PORT   - reverse proxy port          (default: 5000)
+#   PADDLER_MGMT_HOST    - management API hostname/IP  (default: same as proxy host)
+#   PADDLER_MGMT_PORT    - management API port         (default: 8085)
+#   PADDLER_MODEL        - model name to test with     (default: gemma4)
+
+# ── Resolve configuration ──
+
+# Parse a host or full URL into scheme + host, defaulting to http://
+parse_url() {
+    local input="$1"
+    # Strip trailing slashes
+    input="${input%%/}"
+    input="${input%%\?}"
+    if [[ "$input" =~ ^https:// ]]; then
+        echo "https ${input#https://}"
+    elif [[ "$input" =~ ^http:// ]]; then
+        echo "http ${input#http://}"
+    else
+        echo "http $input"
+    fi
+}
+
+read -r PROXY_SCHEME PROXY_HOST <<< "$(parse_url "${PADDLER_PROXY_HOST:-${1:-localhost}}")"
+PROXY_PORT="${PADDLER_PROXY_PORT:-5000}"
+
+# Management defaults to same host and scheme as proxy if not explicitly set
+if [ -n "${PADDLER_MGMT_HOST:-}" ]; then
+    read -r MGMT_SCHEME MGMT_HOST <<< "$(parse_url "$PADDLER_MGMT_HOST")"
+else
+    MGMT_SCHEME="$PROXY_SCHEME"
+    MGMT_HOST="$PROXY_HOST"
+fi
+MGMT_PORT="${PADDLER_MGMT_PORT:-8085}"
+MODEL="${PADDLER_MODEL:-gemma4}"
+
+PROXY="${PROXY_SCHEME}://${PROXY_HOST}:${PROXY_PORT}"
+MANAGEMENT="${MGMT_SCHEME}://${MGMT_HOST}:${MGMT_PORT}"
+
+PASS=0
+FAIL=0
+
+green()  { echo -e "\033[32m  ✓ $1\033[0m"; }
+red()    { echo -e "\033[31m  ✗ $1\033[0m"; }
+yellow() { echo -e "\033[33m  ⚠ $1\033[0m"; }
+header() { echo -e "\n\033[1m── $1 ──\033[0m"; }
+
+pass() { green "$1"; PASS=$((PASS + 1)); }
+fail() { red "$1"; FAIL=$((FAIL + 1)); }
+
+check_status() {
+    local desc="$1" expected="$2" actual="$3"
+    if [ "$actual" = "$expected" ]; then
+        pass "$desc (HTTP $actual)"
+    else
+        fail "$desc (expected HTTP $expected, got HTTP $actual)"
+    fi
+}
+
+# ── Banner ──
+
+echo "======================================================"
+echo "  Paddler Model Detection Test Suite"
+echo "======================================================"
+echo "  Proxy:      $PROXY"
+echo "  Management: $MANAGEMENT"
+echo "  Model:      $MODEL"
+echo "======================================================"
+
+# ── Pre-flight: check that the management API is reachable ──
+
+header "Pre-flight checks"
+
+MGMT_STATUS=$(curl -s -o /dev/null -w "%{http_code}" --max-time 5 "$MANAGEMENT/api/v1/agents" 2>/dev/null)
+MGMT_STATUS=${MGMT_STATUS:-000}
+if [ "$MGMT_STATUS" = "200" ]; then
+    pass "Management API reachable at $MANAGEMENT"
+else
+    red "Management API unreachable (HTTP $MGMT_STATUS) — is the balancer running?"
+    echo ""
+    echo "  Configure the target host with one of:"
+    echo "    ./test-model-detection.sh <host>"
+    echo "    PADDLER_PROXY_HOST=<host> ./test-model-detection.sh"
+    echo ""
+    echo "  Available environment variables:"
+    echo "    PADDLER_PROXY_HOST   - reverse proxy host (default: localhost)"
+    echo "    PADDLER_PROXY_PORT   - reverse proxy port (default: 5000)"
+    echo "    PADDLER_MGMT_HOST    - management API host (default: same as proxy)"
+    echo "    PADDLER_MGMT_PORT    - management API port (default: 8085)"
+    echo "    PADDLER_MODEL        - model name to test (default: gemma4)"
+    echo ""
+    echo "Aborting tests."
+    exit 1
+fi
+
+# Show registered agents
+echo ""
+echo "  Registered agents:"
+curl -s "$MANAGEMENT/api/v1/agents" 2>/dev/null | python3 -m json.tool 2>/dev/null || true
+echo ""
+
+# ── Test 1: Valid request with correct model ──
+
+header "Test 1: Valid request with correct model ($MODEL)"
+
+BODY="{\"model\": \"$MODEL\", \"messages\": [{\"role\": \"user\", \"content\": \"Say hello in one word.\"}], \"max_tokens\": 10}"
+HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" \
+    -X POST "$PROXY/v1/chat/completions" \
+    -H "Content-Type: application/json" \
+    -d "$BODY" 2>/dev/null)
+check_status "POST /v1/chat/completions with model=$MODEL" "200" "$HTTP_CODE"
+
+# ── Test 2: Full functional test — actual joke response ──
+
+header "Test 2: Functional test — tell a short joke"
+
+RESPONSE=$(curl -s -X POST "$PROXY/v1/chat/completions" \
+    -H "Content-Type: application/json" \
+    -d "{\"model\": \"$MODEL\", \"messages\": [{\"role\": \"user\", \"content\": \"Tell a very short joke.\"}], \"max_tokens\": 50}" 2>/dev/null)
+
+if echo "$RESPONSE" | python3 -c "import sys,json; d=json.load(sys.stdin); assert 'choices' in d" 2>/dev/null; then
+    pass "Got valid chat completion response"
+    # Print the actual reply
+    CONTENT=$(echo "$RESPONSE" | python3 -c "import sys,json; print(json.load(sys.stdin)['choices'][0]['message']['content'].strip())" 2>/dev/null || true)
+    echo "  Response: $CONTENT"
+else
+    fail "Invalid or empty response"
+    echo "  Raw: ${RESPONSE:0:200}"
+fi
+
+# ── Test 3: Wrong model — should get 404 ──
+
+header "Test 3: Wrong model (nonexistent) — expect 404"
+
+HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" \
+    -X POST "$PROXY/v1/chat/completions" \
+    -H "Content-Type: application/json" \
+    -d '{"model": "llama-fake-model-xyz", "messages": [{"role": "user", "content": "hi"}], "max_tokens": 5}' 2>/dev/null)
+check_status "POST with wrong model" "404" "$HTTP_CODE"
+
+# ── Test 4: Missing model field — should get 400 ──
+
+header "Test 4: Missing model field — expect 400"
+
+HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" \
+    -X POST "$PROXY/v1/chat/completions" \
+    -H "Content-Type: application/json" \
+    -d '{"messages": [{"role": "user", "content": "hi"}], "max_tokens": 5}' 2>/dev/null)
+check_status "POST without model field" "400" "$HTTP_CODE"
+
+# ── Test 5: Empty body — should get 400 ──
+
+header "Test 5: Empty request body — expect 400"
+
+HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" \
+    -X POST "$PROXY/v1/chat/completions" \
+    -H "Content-Type: application/json" \
+    -d '{}' 2>/dev/null)
+check_status "POST with empty JSON body" "400" "$HTTP_CODE"
+
+# ── Test 6: Invalid JSON body — should get 400 ──
+
+header "Test 6: Invalid JSON body — expect 400"
+
+HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" \
+    -X POST "$PROXY/v1/chat/completions" \
+    -H "Content-Type: application/json" \
+    -d 'not json at all' 2>/dev/null)
+check_status "POST with invalid JSON" "400" "$HTTP_CODE"
+
+# ── Test 7: Non-JSON content type — model check skipped, no model found → 400 ──
+
+header "Test 7: Non-JSON Content-Type (model check skipped, no model found)"
+
+HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" \
+    -X POST "$PROXY/v1/chat/completions" \
+    -H "Content-Type: text/plain" \
+    -d "{\"model\": \"$MODEL\"}" 2>/dev/null)
+# With --check-model, non-JSON Content-Type skips model extraction.
+# Since no model is found, the balancer rejects with 400. This is expected.
+check_status "POST with text/plain Content-Type (model not extracted)" "400" "$HTTP_CODE"
+
+# ── Test 8: GET request — no body, no model found → 400 ──
+
+header "Test 8: GET request (no body, no model found)"
+
+HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" \
+    -X GET "$PROXY/v1/chat/completions" 2>/dev/null)
+# With --check-model, GET has no body so no model is found → 400. This is expected.
+check_status "GET /v1/chat/completions (no body, no model)" "400" "$HTTP_CODE"
+
+# ── Test 9: Model in different JSON field order ──
+
+header "Test 9: Model field not first in JSON"
+
+BODY="{\"messages\": [{\"role\": \"user\", \"content\": \"Say hi.\"}], \"max_tokens\": 10, \"model\": \"$MODEL\"}"
+HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" \
+    -X POST "$PROXY/v1/chat/completions" \
+    -H "Content-Type: application/json" \
+    -d "$BODY" 2>/dev/null)
+check_status "POST with model field last in JSON" "200" "$HTTP_CODE"
+
+# ── Test 10: Model with extra whitespace in JSON ──
+
+header "Test 10: Model with extra whitespace in JSON"
+
+BODY="{  \"model\"  :  \"$MODEL\"  ,  \"messages\"  :  [  {  \"role\"  :  \"user\"  ,  \"content\"  :  \"hi\"  }  ]  ,  \"max_tokens\"  :  5  }"
+HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" \
+    -X POST "$PROXY/v1/chat/completions" \
+    -H "Content-Type: application/json" \
+    -d "$BODY" 2>/dev/null)
+check_status "POST with extra whitespace in JSON" "200" "$HTTP_CODE"
+
+# ── Test 11: Completions endpoint (not chat) ──
+
+header "Test 11: /v1/completions endpoint"
+
+HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" \
+    -X POST "$PROXY/v1/completions" \
+    -H "Content-Type: application/json" \
+    -d "{\"model\": \"$MODEL\", \"prompt\": \"Say hi\", \"max_tokens\": 5}" 2>/dev/null)
+check_status "POST /v1/completions with model=$MODEL" "200" "$HTTP_CODE"
+
+# ── Test 12: Legacy llama.cpp endpoint /completion ──
+
+header "Test 12: Legacy /completion endpoint"
+
+HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" \
+    -X POST "$PROXY/completion" \
+    -H "Content-Type: application/json" \
+    -d "{\"model\": \"$MODEL\", \"prompt\": \"Say hi\", \"n_predict\": 5}" 2>/dev/null)
+check_status "POST /completion with model=$MODEL" "200" "$HTTP_CODE"
+
+# ── Test 13: Payload >64KB — needle-in-haystack ──
+
+header "Test 13: Payload >64KB (needle-in-haystack, verifies full body forwarding)"
+
+# Generate a ~70KB payload: model field first, padding in the middle, needle at the end.
+# If the full body reaches the upstream, the model should be able to find the needle.
+NEEDLE="PADDLER_TEST_SECRET_42"
+PADDING=$(python3 -c "print('x' * 70000)")
+BIG_BODY=$(printf '{"model": "%s", "messages": [{"role": "user", "content": "%s The secret code is %s. Reply with only the secret code."}], "max_tokens": 20}' "$MODEL" "$PADDING" "$NEEDLE")
+
+BIG_RESPONSE=$(curl -s -X POST "$PROXY/v1/chat/completions" \
+    -H "Content-Type: application/json" \
+    -d "$BIG_BODY" 2>/dev/null)
+
+BIG_HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" \
+    -X POST "$PROXY/v1/chat/completions" \
+    -H "Content-Type: application/json" \
+    -d "$BIG_BODY" 2>/dev/null)
+
+if [ "$BIG_HTTP_CODE" = "200" ]; then
+    # Check if the model found the needle in the response
+    if echo "$BIG_RESPONSE" | grep -q "$NEEDLE"; then
+        pass "Got HTTP 200 and model found the needle (full body forwarded)"
+    else
+        yellow "Got HTTP 200 but model did NOT find the needle"
+        yellow "The body may have been truncated or the model ignored it"
+        echo "  Response: $(echo "$BIG_RESPONSE" | python3 -c "import sys,json; print(json.load(sys.stdin).get('choices',[{}])[0].get('message',{}).get('content','').strip())" 2>/dev/null || echo "(parse error)")"
+        fail "Needle not found in response"
+    fi
+else
+    fail "Got HTTP $BIG_HTTP_CODE (expected 200)"
+fi
+
+# ── Test 14: Image input (base64 in content, non-UTF-8 safe) ──
+
+header "Test 14: Vision request with base64 image in content"
+
+# Base64 image data (small 1x1 red PNG) — valid UTF-8 but large binary-ish payload
+BASE64_IMG="iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mNk+M9QDwADhgGAWjR9awAAAABJRU5ErkJggg=="
+VISION_BODY=$(printf '{"model": "%s", "messages": [{"role": "user", "content": [{"type": "text", "text": "Describe this image."}, {"type": "image_url", "image_url": {"url": "data:image/png;base64,%s"}}]}], "max_tokens": 10}' "$MODEL" "$BASE64_IMG")
+HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" \
+    -X POST "$PROXY/v1/chat/completions" \
+    -H "Content-Type: application/json" \
+    -d "$VISION_BODY" 2>/dev/null)
+# The model field should be extracted correctly even with base64 in the body.
+# Upstream may reject if it's not a vision model, but balancer should NOT return 400.
+check_status "POST with base64 image in content" "200" "$HTTP_CODE"
+
+# ── Test 15: Model field appears inside content string (false-positive check) ──
+
+header "Test 15: model keyword inside content string (no false positive)"
+
+# The word "model" appears in the content, but the real model field is $MODEL
+FALSE_POSITIVE_BODY=$(printf '{"model": "%s", "messages": [{"role": "user", "content": "I want to use a different model called llama3."}], "max_tokens": 10}' "$MODEL")
+HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" \
+    -X POST "$PROXY/v1/chat/completions" \
+    -H "Content-Type: application/json" \
+    -d "$FALSE_POSITIVE_BODY" 2>/dev/null)
+check_status "POST with model keyword in content (should extract $MODEL)" "200" "$HTTP_CODE"
+
+# ── Test 16: Streaming request ──
+
+header "Test 16: Streaming request (stream: true)"
+
+STREAM_BODY=$(printf '{"model": "%s", "messages": [{"role": "user", "content": "Say hi."}], "max_tokens": 10, "stream": true}' "$MODEL")
+HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" \
+    -X POST "$PROXY/v1/chat/completions" \
+    -H "Content-Type: application/json" \
+    -d "$STREAM_BODY" 2>/dev/null)
+check_status "POST with stream: true" "200" "$HTTP_CODE"
+
+# ── Test 17: GET /v1/models (listing, no body, no model check) ──
+
+header "Test 17: GET /v1/models (model listing endpoint)"
+
+HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" \
+    -X GET "$PROXY/v1/models" 2>/dev/null)
+# This is not a slots endpoint, so model check should not apply.
+# Upstream should return 200 with the model list.
+check_status "GET /v1/models" "200" "$HTTP_CODE"
+
+# ── Test 18: Model with special characters in name ──
+
+header "Test 18: Model name with special characters (hyphens, dots)"
+
+# Test with a model name that has hyphens/dots — should still be extracted
+SPECIAL_MODEL_BODY='{"model": "gemma-2-9b-it", "messages": [{"role": "user", "content": "hi"}], "max_tokens": 5}'
+HTTP_CODE=$(curl -s -o /dev/null -w "%{http_code}" \
+    -X POST "$PROXY/v1/chat/completions" \
+    -H "Content-Type: application/json" \
+    -d "$SPECIAL_MODEL_BODY" 2>/dev/null)
+# This model doesn't exist on the upstream, so we expect 404 (model not found), not 400 (extraction failed)
+check_status "POST with model name containing hyphens" "404" "$HTTP_CODE"
+
+# ── Summary ──
+
+header "Summary"
+echo "  Passed: $PASS"
+echo "  Failed: $FAIL"
+echo "  Total:  $((PASS + FAIL))"
+echo ""
+
+if [ "$FAIL" -gt 0 ]; then
+    echo "  Some tests failed. Review the output above."
+    exit 1
+else
+    echo "  All tests passed!"
+    exit 0
+fi