diff --git a/Cargo.lock b/Cargo.lock
index 303118e8..73ba5baf 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -3792,7 +3792,7 @@ dependencies = [
 
 [[package]]
 name = "openfang-api"
-version = "0.3.44"
+version = "0.3.46"
 dependencies = [
  "async-trait",
  "axum",
@@ -3829,7 +3829,7 @@ dependencies = [
 
 [[package]]
 name = "openfang-channels"
-version = "0.3.44"
+version = "0.3.46"
 dependencies = [
  "async-trait",
  "axum",
@@ -3861,7 +3861,7 @@ dependencies = [
 
 [[package]]
 name = "openfang-cli"
-version = "0.3.44"
+version = "0.3.46"
 dependencies = [
  "clap",
  "clap_complete",
@@ -3888,7 +3888,7 @@ dependencies = [
 
 [[package]]
 name = "openfang-desktop"
-version = "0.3.44"
+version = "0.3.46"
 dependencies = [
  "axum",
  "open",
@@ -3914,7 +3914,7 @@ dependencies = [
 
 [[package]]
 name = "openfang-extensions"
-version = "0.3.44"
+version = "0.3.46"
 dependencies = [
  "aes-gcm",
  "argon2",
@@ -3942,7 +3942,7 @@ dependencies = [
 
 [[package]]
 name = "openfang-hands"
-version = "0.3.44"
+version = "0.3.46"
 dependencies = [
  "chrono",
  "dashmap",
@@ -3959,7 +3959,7 @@ dependencies = [
 
 [[package]]
 name = "openfang-kernel"
-version = "0.3.44"
+version = "0.3.46"
 dependencies = [
  "async-trait",
  "chrono",
@@ -3996,7 +3996,7 @@ dependencies = [
 
 [[package]]
 name = "openfang-memory"
-version = "0.3.44"
+version = "0.3.46"
 dependencies = [
  "async-trait",
  "chrono",
@@ -4015,7 +4015,7 @@ dependencies = [
 
 [[package]]
 name = "openfang-migrate"
-version = "0.3.44"
+version = "0.3.46"
 dependencies = [
  "chrono",
  "dirs 6.0.0",
@@ -4034,7 +4034,7 @@ dependencies = [
 
 [[package]]
 name = "openfang-runtime"
-version = "0.3.44"
+version = "0.3.46"
 dependencies = [
  "anyhow",
  "async-trait",
@@ -4068,7 +4068,7 @@ dependencies = [
 
 [[package]]
 name = "openfang-skills"
-version = "0.3.44"
+version = "0.3.46"
 dependencies = [
  "chrono",
  "hex",
@@ -4091,7 +4091,7 @@ dependencies = [
 
 [[package]]
 name = "openfang-types"
-version = "0.3.44"
+version = "0.3.46"
 dependencies = [
  "async-trait",
  "chrono",
@@ -4110,7 +4110,7 @@ dependencies = [
 
 [[package]]
 name = "openfang-wire"
-version = "0.3.44"
+version = "0.3.46"
 dependencies = [
  "async-trait",
  "chrono",
@@ -8773,7 +8773,7 @@ checksum = "b9cc00251562a284751c9973bace760d86c0276c471b4be569fe6b068ee97a56"
 
 [[package]]
 name = "xtask"
-version = "0.3.44"
+version = "0.3.46"
 
 [[package]]
 name = "yoke"
diff --git a/Cargo.toml b/Cargo.toml
index b84e92b1..e20dedf1 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -18,7 +18,7 @@ members = [
 ]
 
 [workspace.package]
-version = "0.3.45"
+version = "0.3.46"
 edition = "2021"
 license = "Apache-2.0 OR MIT"
 repository = "https://github.com/RightNow-AI/openfang"
diff --git a/crates/openfang-api/src/channel_bridge.rs b/crates/openfang-api/src/channel_bridge.rs
index bf95f398..42b8587c 100644
--- a/crates/openfang-api/src/channel_bridge.rs
+++ b/crates/openfang-api/src/channel_bridge.rs
@@ -56,6 +56,8 @@ use std::sync::Arc;
 use std::time::{Duration, Instant};
 use tracing::{error, info, warn};
 
+use openfang_runtime::str_utils::safe_truncate_str;
+
 /// Wraps `OpenFangKernel` to implement `ChannelBridgeHandle`.
 pub struct KernelBridgeAdapter {
     kernel: Arc<OpenFangKernel>,
@@ -378,7 +380,8 @@ impl ChannelBridgeHandle for KernelBridgeAdapter {
                 .map(|e| e.name.clone())
                 .unwrap_or_else(|| t.agent_id.to_string());
             let status = if t.enabled { "on" } else { "off" };
-            let id_short = &t.id.0.to_string()[..8];
+            let id_str = t.id.0.to_string();
+            let id_short = safe_truncate_str(&id_str, 8);
             msg.push_str(&format!(
                 "  [{}] {} -> {} ({:?}) fires:{} [{}]\n",
                 id_short,
@@ -417,7 +420,8 @@ impl ChannelBridgeHandle for KernelBridgeAdapter {
             .kernel
             .triggers
             .register(agent.id, pattern, prompt.to_string(), 0);
-        let id_short = &trigger_id.0.to_string()[..8];
+        let id_str = trigger_id.0.to_string();
+        let id_short = safe_truncate_str(&id_str, 8);
         format!("Trigger created [{id_short}] for agent '{agent_name}'.")
     }
 
@@ -432,7 +436,8 @@ impl ChannelBridgeHandle for KernelBridgeAdapter {
             1 => {
                 let t = matched[0];
                 if self.kernel.triggers.remove(t.id) {
-                    format!("Trigger [{}] removed.", &t.id.0.to_string()[..8])
+                    let id_str = t.id.0.to_string();
+                    format!("Trigger [{}] removed.", safe_truncate_str(&id_str, 8))
                 } else {
                     "Failed to remove trigger.".to_string()
                 }
@@ -455,7 +460,8 @@ impl ChannelBridgeHandle for KernelBridgeAdapter {
                 .map(|e| e.name.clone())
                 .unwrap_or_else(|| job.agent_id.to_string());
             let status = if job.enabled { "on" } else { "off" };
-            let id_short = &job.id.0.to_string()[..8];
+            let id_str = job.id.0.to_string();
+            let id_short = safe_truncate_str(&id_str, 8);
             let sched = match &job.schedule {
                 openfang_types::scheduler::CronSchedule::Cron { expr, .. } => expr.clone(),
                 openfang_types::scheduler::CronSchedule::Every { every_secs } => {
@@ -515,7 +521,8 @@ impl ChannelBridgeHandle for KernelBridgeAdapter {
 
                 match self.kernel.cron_scheduler.add_job(job, false) {
                     Ok(id) => {
-                        let id_short = &id.0.to_string()[..8];
+                        let id_str = id.0.to_string();
+                        let id_short = safe_truncate_str(&id_str, 8);
                         format!("Job [{id_short}] created: '{cron_expr}' -> {agent_name}: \"{message}\"")
                     }
                     Err(e) => format!("Failed to create job: {e}"),
@@ -537,7 +544,8 @@ impl ChannelBridgeHandle for KernelBridgeAdapter {
                         let j = matched[0];
                         match self.kernel.cron_scheduler.remove_job(j.id) {
                             Ok(_) => {
-                                format!("Job [{}] '{}' removed.", &j.id.0.to_string()[..8], j.name)
+                                let id_str = j.id.0.to_string();
+                                format!("Job [{}] '{}' removed.", safe_truncate_str(&id_str, 8), j.name)
                             }
                             Err(e) => format!("Failed to remove job: {e}"),
                         }
@@ -569,7 +577,8 @@ impl ChannelBridgeHandle for KernelBridgeAdapter {
                         };
                         match self.kernel.send_message(j.agent_id, &message).await {
                             Ok(result) => {
-                                let id_short = &j.id.0.to_string()[..8];
+                                let id_str = j.id.0.to_string();
+                                let id_short = safe_truncate_str(&id_str, 8);
                                 format!("Job [{id_short}] ran:\n{}", result.response)
                             }
                             Err(e) => format!("Failed to run job: {e}"),
@@ -589,7 +598,8 @@ impl ChannelBridgeHandle for KernelBridgeAdapter {
         }
         let mut msg = format!("Pending approvals ({}):\n", pending.len());
         for req in &pending {
-            let id_short = &req.id.to_string()[..8];
+            let id_str = req.id.to_string();
+            let id_short = safe_truncate_str(&id_str, 8);
             let age_secs = (chrono::Utc::now() - req.requested_at).num_seconds();
             let age = if age_secs >= 60 {
                 format!("{}m", age_secs / 60)
@@ -630,10 +640,11 @@ impl ChannelBridgeHandle for KernelBridgeAdapter {
                 ) {
                     Ok(_) => {
                         let verb = if approve { "Approved" } else { "Rejected" };
+                        let id_str = req.id.to_string();
                         format!(
                             "{} [{}] {} — {}",
                             verb,
-                            &req.id.to_string()[..8],
+                            safe_truncate_str(&id_str, 8),
                             req.tool_name,
                             req.agent_id
                         )
diff --git a/crates/openfang-api/static/index_body.html b/crates/openfang-api/static/index_body.html
index a57e8e87..e2d6fb90 100644
--- a/crates/openfang-api/static/index_body.html
+++ b/crates/openfang-api/static/index_body.html
@@ -736,7 +736,7 @@
                   <span class="text-xs" style="color:var(--danger)" x-text="formatRecordingTime()"></span>
                 </div>
                 <textarea id="msg-input" rows="1" :placeholder="recording ? 'Recording... release to send' : 'Message OpenFang... (/ for commands)'"
-                          @keydown.enter.prevent="if(!$event.isComposing && !$event.shiftKey){if(showModelPicker && filteredModelPicker.length){pickModel(filteredModelPicker[modelPickerIdx].id)}else if(showSlashMenu && filteredSlashCommands.length){executeSlashCommand(filteredSlashCommands[slashIdx].cmd)}else{sendMessage()}}"
+                          @keydown.enter.prevent="if(!$event.isComposing && $event.keyCode !== 229 && !$event.shiftKey){if(showModelPicker && filteredModelPicker.length){pickModel(filteredModelPicker[modelPickerIdx].id)}else if(showSlashMenu && filteredSlashCommands.length){executeSlashCommand(filteredSlashCommands[slashIdx].cmd)}else{sendMessage()}}"
                           @keydown.escape="showSlashMenu = false; showModelPicker = false"
                           @keydown.arrow-up.prevent="if(showModelPicker){modelPickerIdx = Math.max(0, modelPickerIdx - 1)}else if(showSlashMenu){slashIdx = Math.max(0, slashIdx - 1)}"
                           @keydown.arrow-down.prevent="if(showModelPicker){modelPickerIdx = Math.min(filteredModelPicker.length - 1, modelPickerIdx + 1)}else if(showSlashMenu){slashIdx = Math.min(filteredSlashCommands.length - 1, slashIdx + 1)}"
@@ -769,7 +769,7 @@
                     <div class="model-switcher-dropdown" x-show="showModelSwitcher" x-transition:enter="transition ease-out duration-150" x-transition:enter-start="opacity-0 transform translate-y-1" x-transition:enter-end="opacity-100 transform translate-y-0" x-transition:leave="transition ease-in duration-100" x-transition:leave-start="opacity-100" x-transition:leave-end="opacity-0">
                       <div class="model-switcher-search">
                         <svg width="14" height="14" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" style="flex-shrink:0;opacity:0.5"><circle cx="11" cy="11" r="8"/><line x1="21" y1="21" x2="16.65" y2="16.65"/></svg>
-                        <input id="model-switcher-search" type="text" x-model="modelSwitcherFilter" placeholder="Search models..." @keydown.escape.stop="showModelSwitcher = false" @keydown.arrow-down.prevent="modelSwitcherIdx = Math.min(modelSwitcherIdx + 1, filteredSwitcherModels.length - 1)" @keydown.arrow-up.prevent="modelSwitcherIdx = Math.max(modelSwitcherIdx - 1, 0)" @keydown.enter.prevent="filteredSwitcherModels[modelSwitcherIdx] && switchModel(filteredSwitcherModels[modelSwitcherIdx])">
+                        <input id="model-switcher-search" type="text" x-model="modelSwitcherFilter" placeholder="Search models..." @keydown.escape.stop="showModelSwitcher = false" @keydown.arrow-down.prevent="modelSwitcherIdx = Math.min(modelSwitcherIdx + 1, filteredSwitcherModels.length - 1)" @keydown.arrow-up.prevent="modelSwitcherIdx = Math.max(modelSwitcherIdx - 1, 0)" @keydown.enter.prevent="!$event.isComposing && $event.keyCode !== 229 && filteredSwitcherModels[modelSwitcherIdx] && switchModel(filteredSwitcherModels[modelSwitcherIdx])">
                         <select x-model="modelSwitcherProviderFilter" style="background:var(--surface2);border:1px solid var(--border);border-radius:6px;color:var(--text-dim);font-size:11px;padding:2px 6px;cursor:pointer;font-family:var(--font-mono);flex-shrink:0">
                           <option value="">All</option>
                           <template x-for="pn in switcherProviders" :key="pn">
@@ -1091,7 +1091,7 @@
                   <div x-show="spawnStep === 1">
                     <div class="form-group">
                       <label>Agent Name</label>
-                      <input class="form-input" x-model="spawnForm.name" placeholder="my-agent" @keydown.enter="nextStep()">
+                      <input class="form-input" x-model="spawnForm.name" placeholder="my-agent" @keydown.enter="if(!$event.isComposing && $event.keyCode !== 229) nextStep()">
                     </div>
                     <div class="form-group">
                       <label>Emoji</label>
@@ -2172,7 +2172,7 @@
             <!-- Search bar with live search and clear button -->
             <div class="search-input mb-4" style="position:relative">
               <span style="color:var(--text-muted)"><svg viewBox="0 0 24 24" width="14" height="14" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><circle cx="11" cy="11" r="8"/><path d="m21 21-4.35-4.35"/></svg></span>
-              <input placeholder="Search ClawHub skills... (type to search)" x-model="clawhubSearch" @input="onSearchInput()" @keydown.enter="searchClawHub()" @keydown.escape="clearSearch()" x-ref="clawhubSearchInput">
+              <input placeholder="Search ClawHub skills... (type to search)" x-model="clawhubSearch" @input="onSearchInput()" @keydown.enter="if(!$event.isComposing && $event.keyCode !== 229) searchClawHub()" @keydown.escape="clearSearch()" x-ref="clawhubSearchInput">
               <button x-show="clawhubSearch" @click="clearSearch()" class="search-clear-btn" title="Clear search (Esc)">&times;</button>
             </div>
 
@@ -4576,7 +4576,7 @@ args = ["-y", "@modelcontextprotocol/server-filesystem", "/path"]</pre>
               <div class="card" style="border-left:3px solid var(--accent)">
                 <div class="form-group" style="margin-bottom:8px">
                   <label>Agent Name</label>
-                  <input class="form-input" type="text" x-model="agentName" placeholder="my-assistant" style="max-width:320px" @keydown.enter="createAgent()">
+                  <input class="form-input" type="text" x-model="agentName" placeholder="my-assistant" style="max-width:320px" @keydown.enter="if(!$event.isComposing && $event.keyCode !== 229) createAgent()">
                 </div>
                 <div class="text-xs text-dim" x-text="'Will use ' + templates[selectedTemplate].provider + ' / ' + templates[selectedTemplate].model + ' with ' + profileInfo(templates[selectedTemplate].profile).label + ' profile'"></div>
                 <div class="mt-2">
@@ -4622,7 +4622,7 @@ args = ["-y", "@modelcontextprotocol/server-filesystem", "/path"]</pre>
               <!-- Input -->
               <div style="display:flex;gap:8px;margin-top:12px">
                 <input class="form-input" type="text" x-model="tryItInput" placeholder="Type a message..."
-                       @keydown.enter="sendTryItMessage(tryItInput)" :disabled="tryItSending" style="flex:1">
+                       @keydown.enter="if(!$event.isComposing && $event.keyCode !== 229) sendTryItMessage(tryItInput)" :disabled="tryItSending" style="flex:1">
                 <button class="btn btn-primary btn-sm" @click="sendTryItMessage(tryItInput)" :disabled="tryItSending || !tryItInput.trim()">Send</button>
               </div>
             </div>
diff --git a/crates/openfang-api/static/js/pages/agents.js b/crates/openfang-api/static/js/pages/agents.js
index 041b13a4..21f506f1 100644
--- a/crates/openfang-api/static/js/pages/agents.js
+++ b/crates/openfang-api/static/js/pages/agents.js
@@ -1,6 +1,21 @@
 // OpenFang Agents Page — Multi-step spawn wizard, detail view with tabs, file editor, personality presets
 'use strict';
 
+/** Escape a string for use inside TOML triple-quoted strings ("""\n...\n""").
+ *  Backslashes are escaped, and runs of 3+ consecutive double-quotes are
+ *  broken up so the TOML parser never sees an unintended closing delimiter.
+ */
+function tomlMultilineEscape(s) {
+  return s.replace(/\\/g, '\\\\').replace(/"""/g, '""\\"');
+}
+
+/** Escape a string for use inside a TOML basic (single-line) string ("...").
+ *  Backslashes, double-quotes, and common control chars are escaped.
+ */
+function tomlBasicEscape(s) {
+  return s.replace(/\\/g, '\\\\').replace(/"/g, '\\"').replace(/\n/g, '\\n').replace(/\r/g, '\\r').replace(/\t/g, '\\t');
+}
+
 function agentsPage() {
   return {
     tab: 'agents',
@@ -411,7 +426,7 @@ function agentsPage() {
       var f = this.spawnForm;
       var si = this.spawnIdentity;
       var lines = [
-        'name = "' + f.name.replace(/\\/g, '\\\\').replace(/"/g, '\\"') + '"',
+        'name = "' + tomlBasicEscape(f.name) + '"',
         'module = "builtin:chat"'
       ];
       if (f.profile && f.profile !== 'custom') {
@@ -420,7 +435,7 @@ function agentsPage() {
       lines.push('', '[model]');
       lines.push('provider = "' + f.provider + '"');
       lines.push('model = "' + f.model + '"');
-      lines.push('system_prompt = """\n' + f.systemPrompt.replace(/\\/g, '\\\\').replace(/"""/g, '""\\"') + '\n"""');
+      lines.push('system_prompt = """\n' + tomlMultilineEscape(f.systemPrompt) + '\n"""');
       if (f.profile === 'custom') {
         lines.push('', '[capabilities]');
         if (f.caps.memory_read) lines.push('memory_read = ["*"]');
@@ -698,12 +713,12 @@ function agentsPage() {
     },
 
     async spawnBuiltin(t) {
-      var toml = 'name = "' + t.name + '"\n';
-      toml += 'description = "' + t.description.replace(/"/g, '\\"') + '"\n';
+      var toml = 'name = "' + tomlBasicEscape(t.name) + '"\n';
+      toml += 'description = "' + tomlBasicEscape(t.description) + '"\n';
       toml += 'module = "builtin:chat"\n';
       toml += 'profile = "' + t.profile + '"\n\n';
       toml += '[model]\nprovider = "' + t.provider + '"\nmodel = "' + t.model + '"\n';
-      toml += 'system_prompt = """\n' + t.system_prompt + '\n"""\n';
+      toml += 'system_prompt = """\n' + tomlMultilineEscape(t.system_prompt) + '\n"""\n';
 
       try {
         var res = await OpenFangAPI.post('/api/agents', { manifest_toml: toml });
diff --git a/crates/openfang-api/static/js/pages/wizard.js b/crates/openfang-api/static/js/pages/wizard.js
index a8b3e2f8..3f8924bc 100644
--- a/crates/openfang-api/static/js/pages/wizard.js
+++ b/crates/openfang-api/static/js/pages/wizard.js
@@ -1,6 +1,16 @@
 // OpenFang Setup Wizard — First-run guided setup (Provider + Agent + Channel)
 'use strict';
 
+/** Escape a string for use inside TOML triple-quoted strings ("""\n...\n"""). */
+function wizardTomlMultilineEscape(s) {
+  return s.replace(/\\/g, '\\\\').replace(/"""/g, '""\\"');
+}
+
+/** Escape a string for use inside a TOML basic (single-line) string ("..."). */
+function wizardTomlBasicEscape(s) {
+  return s.replace(/\\/g, '\\\\').replace(/"/g, '\\"').replace(/\n/g, '\\n').replace(/\r/g, '\\r').replace(/\t/g, '\\t');
+}
+
 function wizardPage() {
   return {
     step: 1,
@@ -462,12 +472,12 @@ function wizardPage() {
       }
 
       var toml = '[agent]\n';
-      toml += 'name = "' + name.replace(/"/g, '\\"') + '"\n';
-      toml += 'description = "' + tpl.description.replace(/"/g, '\\"') + '"\n';
+      toml += 'name = "' + wizardTomlBasicEscape(name) + '"\n';
+      toml += 'description = "' + wizardTomlBasicEscape(tpl.description) + '"\n';
       toml += 'profile = "' + tpl.profile + '"\n\n';
       toml += '[model]\nprovider = "' + provider + '"\n';
       toml += 'model = "' + model + '"\n';
-      toml += 'system_prompt = """\n' + tpl.system_prompt + '\n"""\n';
+      toml += 'system_prompt = """\n' + wizardTomlMultilineEscape(tpl.system_prompt) + '\n"""\n';
 
       this.creatingAgent = true;
       try {
diff --git a/crates/openfang-desktop/src/server.rs b/crates/openfang-desktop/src/server.rs
index 276baa09..984f2316 100644
--- a/crates/openfang-desktop/src/server.rs
+++ b/crates/openfang-desktop/src/server.rs
@@ -6,6 +6,7 @@
 use openfang_api::server::build_router;
 use openfang_kernel::OpenFangKernel;
 use std::net::{SocketAddr, TcpListener};
+use std::sync::atomic::{AtomicBool, Ordering};
 use std::sync::Arc;
 use tokio::sync::watch;
 use tracing::{error, info};
@@ -20,24 +21,40 @@ pub struct ServerHandle {
     shutdown_tx: watch::Sender<bool>,
     /// Join handle for the background server thread.
     server_thread: Option<std::thread::JoinHandle<()>>,
+    /// Track whether shutdown has already been initiated to prevent double shutdown.
+    shutdown_initiated: Arc<AtomicBool>,
 }
 
 impl ServerHandle {
     /// Signal the server to shut down and wait for the background thread.
     pub fn shutdown(mut self) {
-        let _ = self.shutdown_tx.send(true);
-        if let Some(handle) = self.server_thread.take() {
-            let _ = handle.join();
+        // Only proceed if shutdown hasn't been initiated yet
+        if self
+            .shutdown_initiated
+            .compare_exchange(false, true, Ordering::SeqCst, Ordering::Relaxed)
+            .is_ok()
+        {
+            let _ = self.shutdown_tx.send(true);
+            if let Some(handle) = self.server_thread.take() {
+                let _ = handle.join();
+            }
+            self.kernel.shutdown();
+            info!("OpenFang embedded server stopped");
         }
-        self.kernel.shutdown();
-        info!("OpenFang embedded server stopped");
     }
 }
 
 impl Drop for ServerHandle {
     fn drop(&mut self) {
-        let _ = self.shutdown_tx.send(true);
-        // Best-effort: don't block in drop, the thread will exit on its own.
+        // Only send shutdown signal if it hasn't been initiated yet
+        if self
+            .shutdown_initiated
+            .compare_exchange(false, true, Ordering::SeqCst, Ordering::Relaxed)
+            .is_ok()
+        {
+            let _ = self.shutdown_tx.send(true);
+            // Best-effort: don't block in drop, the thread will exit on its own.
+        }
     }
 }
 
@@ -61,6 +78,7 @@ pub fn start_server() -> Result<ServerHandle, Box<dyn std::error::Error>> {
 
     let (shutdown_tx, shutdown_rx) = watch::channel(false);
     let kernel_clone = kernel.clone();
+    let shutdown_initiated = Arc::new(AtomicBool::new(false));
 
     let server_thread = std::thread::Builder::new()
         .name("openfang-server".into())
@@ -83,6 +101,7 @@ pub fn start_server() -> Result<ServerHandle, Box<dyn std::error::Error>> {
         kernel,
         shutdown_tx,
         server_thread: Some(server_thread),
+        shutdown_initiated,
     })
 }
 
diff --git a/crates/openfang-kernel/src/kernel.rs b/crates/openfang-kernel/src/kernel.rs
index 1bea626a..8f6dbc6b 100644
--- a/crates/openfang-kernel/src/kernel.rs
+++ b/crates/openfang-kernel/src/kernel.rs
@@ -806,12 +806,20 @@ impl OpenFangKernel {
             use openfang_runtime::embedding::create_embedding_driver;
             let configured_model = &config.memory.embedding_model;
             if let Some(ref provider) = config.memory.embedding_provider {
-                // Explicit config takes priority — use the configured embedding model
+                // Explicit config takes priority — use the configured embedding model.
+                // If the user left embedding_model at the default ("all-MiniLM-L6-v2"),
+                // pick a sensible default for the chosen provider so we don't send a
+                // local model name to a cloud API.
+                let model = if configured_model == "all-MiniLM-L6-v2" {
+                    default_embedding_model_for_provider(provider)
+                } else {
+                    configured_model.as_str()
+                };
                 let api_key_env = config.memory.embedding_api_key_env.as_deref().unwrap_or("");
                 let custom_url = config.provider_urls.get(provider.as_str()).map(|s| s.as_str());
-                match create_embedding_driver(provider, configured_model, api_key_env, custom_url) {
+                match create_embedding_driver(provider, model, api_key_env, custom_url) {
                     Ok(d) => {
-                        info!(provider = %provider, model = %configured_model, "Embedding driver configured from memory config");
+                        info!(provider = %provider, model = %model, "Embedding driver configured from memory config");
                         Some(Arc::from(d))
                     }
                     Err(e) => {
@@ -821,14 +829,14 @@ impl OpenFangKernel {
                 }
             } else if std::env::var("OPENAI_API_KEY").is_ok() {
                 let model = if configured_model == "all-MiniLM-L6-v2" {
-                    "text-embedding-3-small"
+                    default_embedding_model_for_provider("openai")
                 } else {
                     configured_model.as_str()
                 };
                 let openai_url = config.provider_urls.get("openai").map(|s| s.as_str());
                 match create_embedding_driver("openai", model, "OPENAI_API_KEY", openai_url) {
                     Ok(d) => {
-                        info!("Embedding driver auto-detected: OpenAI");
+                        info!(model = %model, "Embedding driver auto-detected: OpenAI");
                         Some(Arc::from(d))
                     }
                     Err(e) => {
@@ -839,14 +847,14 @@ impl OpenFangKernel {
             } else {
                 // Try Ollama (local, no key needed)
                 let model = if configured_model == "all-MiniLM-L6-v2" {
-                    "nomic-embed-text"
+                    default_embedding_model_for_provider("ollama")
                 } else {
                     configured_model.as_str()
                 };
                 let ollama_url = config.provider_urls.get("ollama").map(|s| s.as_str());
                 match create_embedding_driver("ollama", model, "", ollama_url) {
                     Ok(d) => {
-                        info!("Embedding driver auto-detected: Ollama (local)");
+                        info!(model = %model, "Embedding driver auto-detected: Ollama (local)");
                         Some(Arc::from(d))
                     }
                     Err(e) => {
@@ -4919,6 +4927,21 @@ fn apply_budget_defaults(
     }
 }
 
+/// Pick a sensible default embedding model for a given provider when the user
+/// configured an explicit `embedding_provider` but left `embedding_model` at the
+/// default value (which is a local model name that cloud APIs wouldn't recognise).
+fn default_embedding_model_for_provider(provider: &str) -> &'static str {
+    match provider {
+        "openai" => "text-embedding-3-small",
+        "mistral" => "mistral-embed",
+        "cohere" => "embed-english-v3.0",
+        // Local providers use nomic-embed-text as a good default
+        "ollama" | "vllm" | "lmstudio" => "nomic-embed-text",
+        // Other OpenAI-compatible APIs typically support the OpenAI model names
+        _ => "text-embedding-3-small",
+    }
+}
+
 /// Infer provider from a model name when catalog lookup fails.
 ///
 /// Uses well-known model name prefixes to map to the correct provider.
diff --git a/crates/openfang-runtime/src/agent_loop.rs b/crates/openfang-runtime/src/agent_loop.rs
index 36de035e..1e73f7f9 100644
--- a/crates/openfang-runtime/src/agent_loop.rs
+++ b/crates/openfang-runtime/src/agent_loop.rs
@@ -917,7 +917,11 @@ async fn call_with_retry(
             Err(e) => {
                 // Use classifier for smarter error handling
                 let raw_error = e.to_string();
-                let classified = llm_errors::classify_error(&raw_error, None);
+                let status = match &e {
+                    LlmError::Api { status, .. } => Some(*status),
+                    _ => None,
+                };
+                let classified = llm_errors::classify_error(&raw_error, status);
                 warn!(
                     category = ?classified.category,
                     retryable = classified.is_retryable,
@@ -1027,7 +1031,11 @@ async fn stream_with_retry(
             }
             Err(e) => {
                 let raw_error = e.to_string();
-                let classified = llm_errors::classify_error(&raw_error, None);
+                let status = match &e {
+                    LlmError::Api { status, .. } => Some(*status),
+                    _ => None,
+                };
+                let classified = llm_errors::classify_error(&raw_error, status);
                 warn!(
                     category = ?classified.category,
                     retryable = classified.is_retryable,
@@ -1813,6 +1821,7 @@ pub async fn run_agent_loop_streaming(
 /// 6. `[TOOL_CALL]...[/TOOL_CALL]` blocks (JSON or arrow syntax) — issue #354
 /// 7. `<tool_call>{"name":"tool","arguments":{...}}</tool_call>` — Qwen3, issue #332
 /// 8. Bare JSON `{"name":"tool","arguments":{...}}` objects (last resort, only if no tags found)
+/// 9. `<function name="tool" parameters="{...}" />` — XML attribute style (Groq/Llama)
 ///
 /// Validates tool names against available tools and returns synthetic `ToolCall` entries.
 fn recover_text_tool_calls(text: &str, available_tools: &[ToolDefinition]) -> Vec<ToolCall> {
@@ -2148,6 +2157,53 @@ fn recover_text_tool_calls(text: &str, available_tools: &[ToolDefinition]) -> Ve
         }
     }
 
+    // Pattern 9: <function name="tool" parameters="{...}" /> — XML attribute style
+    // Groq/Llama sometimes emit self-closing XML with name/parameters attributes.
+    // The parameters value is HTML-entity-escaped JSON (&quot; etc.).
+    {
+        use regex_lite::Regex;
+        // Match both self-closing <function ... /> and <function ...></function>
+        let re = Regex::new(
+            r#"<function\s+name="([^"]+)"\s+parameters="([^"]*)"[^/]*/?>"#
+        ).unwrap();
+        for caps in re.captures_iter(text) {
+            let tool_name = caps.get(1).unwrap().as_str();
+            let raw_params = caps.get(2).unwrap().as_str();
+
+            if !tool_names.contains(&tool_name) {
+                warn!(tool = tool_name, "XML-attribute tool call for unknown tool — skipping");
+                continue;
+            }
+
+            // Unescape HTML entities (&quot; &amp; &lt; &gt; &apos;)
+            let unescaped = raw_params
+                .replace("&quot;", "\"")
+                .replace("&amp;", "&")
+                .replace("&lt;", "<")
+                .replace("&gt;", ">")
+                .replace("&apos;", "'");
+
+            let input: serde_json::Value = match serde_json::from_str(&unescaped) {
+                Ok(v) => v,
+                Err(e) => {
+                    warn!(tool = tool_name, error = %e, "Failed to parse XML-attribute tool call params — skipping");
+                    continue;
+                }
+            };
+
+            if calls.iter().any(|c| c.name == tool_name && c.input == input) {
+                continue;
+            }
+
+            info!(tool = tool_name, "Recovered XML-attribute tool call → synthetic ToolUse");
+            calls.push(ToolCall {
+                id: format!("recovered_{}", uuid::Uuid::new_v4()),
+                name: tool_name.to_string(),
+                input,
+            });
+        }
+    }
+
     // Pattern 8: Bare JSON tool call objects in text (common Ollama fallback)
     // Matches: {"name":"tool_name","arguments":{"key":"value"}} not already inside tags
     // Only try this if no calls were found by tag-based patterns, to avoid false positives.
@@ -3571,6 +3627,47 @@ mod tests {
         assert_eq!(calls[0].input["command"], "ls");
     }
 
+    // --- Pattern 9: XML-attribute style <function name="..." parameters="..." /> ---
+
+    #[test]
+    fn test_recover_xml_attribute_basic() {
+        let tools = vec![ToolDefinition {
+            name: "web_search".into(),
+            description: "Search".into(),
+            input_schema: serde_json::json!({}),
+        }];
+        let text = r#"<function name="web_search" parameters="{&quot;query&quot;: &quot;best crypto 2024&quot;}" />"#;
+        let calls = recover_text_tool_calls(text, &tools);
+        assert_eq!(calls.len(), 1);
+        assert_eq!(calls[0].name, "web_search");
+        assert_eq!(calls[0].input["query"], "best crypto 2024");
+    }
+
+    #[test]
+    fn test_recover_xml_attribute_unknown_tool() {
+        let tools = vec![ToolDefinition {
+            name: "web_search".into(),
+            description: "Search".into(),
+            input_schema: serde_json::json!({}),
+        }];
+        let text = r#"<function name="unknown_tool" parameters="{&quot;x&quot;: 1}" />"#;
+        let calls = recover_text_tool_calls(text, &tools);
+        assert!(calls.is_empty());
+    }
+
+    #[test]
+    fn test_recover_xml_attribute_non_selfclosing() {
+        let tools = vec![ToolDefinition {
+            name: "shell_exec".into(),
+            description: "Execute".into(),
+            input_schema: serde_json::json!({}),
+        }];
+        let text = r#"<function name="shell_exec" parameters="{&quot;command&quot;: &quot;ls&quot;}"></function>"#;
+        let calls = recover_text_tool_calls(text, &tools);
+        assert_eq!(calls.len(), 1);
+        assert_eq!(calls[0].name, "shell_exec");
+    }
+
     // --- Helper function tests ---
 
     #[test]
diff --git a/crates/openfang-runtime/src/drivers/anthropic.rs b/crates/openfang-runtime/src/drivers/anthropic.rs
index 23862bc7..6f1179ac 100644
--- a/crates/openfang-runtime/src/drivers/anthropic.rs
+++ b/crates/openfang-runtime/src/drivers/anthropic.rs
@@ -27,7 +27,10 @@ impl AnthropicDriver {
         Self {
             api_key: Zeroizing::new(api_key),
             base_url,
-            client: reqwest::Client::new(),
+            client: reqwest::Client::builder()
+                .user_agent(crate::USER_AGENT)
+                .build()
+                .unwrap_or_default(),
         }
     }
 }
diff --git a/crates/openfang-runtime/src/drivers/gemini.rs b/crates/openfang-runtime/src/drivers/gemini.rs
index c7adb3f0..03585513 100644
--- a/crates/openfang-runtime/src/drivers/gemini.rs
+++ b/crates/openfang-runtime/src/drivers/gemini.rs
@@ -32,7 +32,10 @@ impl GeminiDriver {
         Self {
             api_key: Zeroizing::new(api_key),
             base_url,
-            client: reqwest::Client::new(),
+            client: reqwest::Client::builder()
+                .user_agent(crate::USER_AGENT)
+                .build()
+                .unwrap_or_default(),
         }
     }
 }
diff --git a/crates/openfang-runtime/src/drivers/mod.rs b/crates/openfang-runtime/src/drivers/mod.rs
index 00be0e17..fdd62aaf 100644
--- a/crates/openfang-runtime/src/drivers/mod.rs
+++ b/crates/openfang-runtime/src/drivers/mod.rs
@@ -15,11 +15,12 @@ use crate::llm_driver::{DriverConfig, LlmDriver, LlmError};
 use openfang_types::model_catalog::{
     AI21_BASE_URL, ANTHROPIC_BASE_URL, CEREBRAS_BASE_URL, CHUTES_BASE_URL, COHERE_BASE_URL,
     DEEPSEEK_BASE_URL, FIREWORKS_BASE_URL, GEMINI_BASE_URL, GROQ_BASE_URL, HUGGINGFACE_BASE_URL,
-    LEMONADE_BASE_URL, LMSTUDIO_BASE_URL, MINIMAX_BASE_URL, MISTRAL_BASE_URL, MOONSHOT_BASE_URL,
-    OLLAMA_BASE_URL, OPENAI_BASE_URL, OPENROUTER_BASE_URL, PERPLEXITY_BASE_URL,
-    QIANFAN_BASE_URL, QWEN_BASE_URL, REPLICATE_BASE_URL, SAMBANOVA_BASE_URL, TOGETHER_BASE_URL,
-    VENICE_BASE_URL, VLLM_BASE_URL, VOLCENGINE_BASE_URL, VOLCENGINE_CODING_BASE_URL,
-    XAI_BASE_URL, ZAI_BASE_URL, ZAI_CODING_BASE_URL, ZHIPU_BASE_URL, ZHIPU_CODING_BASE_URL,
+    KIMI_CODING_BASE_URL, LEMONADE_BASE_URL, LMSTUDIO_BASE_URL, MINIMAX_BASE_URL,
+    MISTRAL_BASE_URL, MOONSHOT_BASE_URL, OLLAMA_BASE_URL, OPENAI_BASE_URL,
+    OPENROUTER_BASE_URL, PERPLEXITY_BASE_URL, QIANFAN_BASE_URL, QWEN_BASE_URL,
+    REPLICATE_BASE_URL, SAMBANOVA_BASE_URL, TOGETHER_BASE_URL, VENICE_BASE_URL, VLLM_BASE_URL,
+    VOLCENGINE_BASE_URL, VOLCENGINE_CODING_BASE_URL, XAI_BASE_URL, ZAI_BASE_URL,
+    ZAI_CODING_BASE_URL, ZHIPU_BASE_URL, ZHIPU_CODING_BASE_URL,
 };
 use std::sync::Arc;
 
@@ -149,11 +150,16 @@ fn provider_defaults(provider: &str) -> Option<ProviderDefaults> {
             api_key_env: "",
             key_required: false,
         }),
-        "moonshot" | "kimi" => Some(ProviderDefaults {
+        "moonshot" | "kimi" | "kimi2" => Some(ProviderDefaults {
             base_url: MOONSHOT_BASE_URL,
             api_key_env: "MOONSHOT_API_KEY",
             key_required: true,
         }),
+        "kimi_coding" => Some(ProviderDefaults {
+            base_url: KIMI_CODING_BASE_URL,
+            api_key_env: "KIMI_API_KEY",
+            key_required: true,
+        }),
         "qwen" | "dashscope" | "model_studio" => Some(ProviderDefaults {
             base_url: QWEN_BASE_URL,
             api_key_env: "DASHSCOPE_API_KEY",
@@ -174,7 +180,7 @@ fn provider_defaults(provider: &str) -> Option<ProviderDefaults> {
             api_key_env: "ZHIPU_API_KEY",
             key_required: true,
         }),
-        "zai" => Some(ProviderDefaults {
+        "zai" | "z.ai" => Some(ProviderDefaults {
             base_url: ZAI_BASE_URL,
             api_key_env: "ZHIPU_API_KEY",
             key_required: true,
@@ -323,6 +329,22 @@ pub fn create_driver(config: &DriverConfig) -> Result<Arc<dyn LlmDriver>, LlmErr
         )));
     }
 
+    // Kimi for Code — Anthropic-compatible endpoint
+    if provider == "kimi_coding" {
+        let api_key = config
+            .api_key
+            .clone()
+            .or_else(|| std::env::var("KIMI_API_KEY").ok())
+            .ok_or_else(|| {
+                LlmError::MissingApiKey("Set KIMI_API_KEY environment variable".to_string())
+            })?;
+        let base_url = config
+            .base_url
+            .clone()
+            .unwrap_or_else(|| KIMI_CODING_BASE_URL.to_string());
+        return Ok(Arc::new(anthropic::AnthropicDriver::new(api_key, base_url)));
+    }
+
     // All other providers use OpenAI-compatible format
     if let Some(defaults) = provider_defaults(provider) {
         let api_key = config
@@ -453,6 +475,8 @@ pub fn known_providers() -> &'static [&'static str] {
         "minimax",
         "zhipu",
         "zhipu_coding",
+        "zai",
+        "kimi_coding",
         "qianfan",
         "volcengine",
         "chutes",
@@ -551,12 +575,14 @@ mod tests {
         assert!(providers.contains(&"minimax"));
         assert!(providers.contains(&"zhipu"));
         assert!(providers.contains(&"zhipu_coding"));
+        assert!(providers.contains(&"zai"));
+        assert!(providers.contains(&"kimi_coding"));
         assert!(providers.contains(&"qianfan"));
         assert!(providers.contains(&"volcengine"));
         assert!(providers.contains(&"chutes"));
         assert!(providers.contains(&"codex"));
         assert!(providers.contains(&"claude-code"));
-        assert_eq!(providers.len(), 32);
+        assert_eq!(providers.len(), 34);
     }
 
     #[test]
@@ -642,6 +668,14 @@ mod tests {
         std::env::remove_var("NVIDIA_API_KEY");
     }
 
+    #[test]
+    fn test_provider_defaults_kimi_coding() {
+        let d = provider_defaults("kimi_coding").unwrap();
+        assert_eq!(d.base_url, "https://api.kimi.com/coding");
+        assert_eq!(d.api_key_env, "KIMI_API_KEY");
+        assert!(d.key_required);
+    }
+
     #[test]
     fn test_custom_provider_explicit_key_with_url() {
         // When api_key is explicitly passed, it should be used regardless of env var.
diff --git a/crates/openfang-runtime/src/drivers/openai.rs b/crates/openfang-runtime/src/drivers/openai.rs
index 93448d2f..2bc74b6f 100644
--- a/crates/openfang-runtime/src/drivers/openai.rs
+++ b/crates/openfang-runtime/src/drivers/openai.rs
@@ -26,11 +26,19 @@ impl OpenAIDriver {
         Self {
             api_key: Zeroizing::new(api_key),
             base_url,
-            client: reqwest::Client::new(),
+            client: reqwest::Client::builder()
+                .user_agent(crate::USER_AGENT)
+                .build()
+                .unwrap_or_default(),
             extra_headers: Vec::new(),
         }
     }
 
+    /// True if this provider is Moonshot/Kimi and requires reasoning_content on assistant messages with tool_calls.
+    fn kimi_needs_reasoning_content(&self, model: &str) -> bool {
+        self.base_url.contains("moonshot") || model.to_lowercase().contains("kimi")
+    }
+
     /// Create a driver with additional HTTP headers (e.g. for Copilot IDE auth).
     pub fn with_extra_headers(mut self, headers: Vec<(String, String)>) -> Self {
         self.extra_headers = headers;
@@ -59,6 +67,9 @@ struct OaiRequest {
     /// Request usage stats in streaming responses (OpenAI extension, supported by Groq et al).
     #[serde(skip_serializing_if = "Option::is_none")]
     stream_options: Option<serde_json::Value>,
+    /// Moonshot Kimi K2.5: disable thinking so multi-turn with tool_calls works without preserving reasoning_content.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    thinking: Option<serde_json::Value>,
 }
 
 /// Returns true if a model uses `max_completion_tokens` instead of `max_tokens`.
@@ -89,6 +100,12 @@ fn rejects_temperature(model: &str) -> bool {
         || m.contains("-reasoning")
 }
 
+/// Returns true if a model only accepts temperature = 1 (e.g. Moonshot Kimi K2/K2.5).
+fn temperature_must_be_one(model: &str) -> bool {
+    let m = model.to_lowercase();
+    m.starts_with("kimi-k2") || m == "kimi-k2.5" || m == "kimi-k2.5-0711"
+}
+
 #[derive(Debug, Serialize)]
 struct OaiMessage {
     role: String,
@@ -98,6 +115,9 @@ struct OaiMessage {
     tool_calls: Option<Vec<OaiToolCall>>,
     #[serde(skip_serializing_if = "Option::is_none")]
     tool_call_id: Option<String>,
+    /// Moonshot Kimi: sent as empty string on assistant messages with tool_calls when using Kimi (thinking is disabled for multi-turn compatibility).
+    #[serde(skip_serializing_if = "Option::is_none")]
+    reasoning_content: Option<String>,
 }
 
 /// Content can be a plain string or an array of content parts (for images).
@@ -190,6 +210,7 @@ impl LlmDriver for OpenAIDriver {
                 content: Some(OaiMessageContent::Text(system.clone())),
                 tool_calls: None,
                 tool_call_id: None,
+                reasoning_content: None,
             });
         }
 
@@ -203,6 +224,7 @@ impl LlmDriver for OpenAIDriver {
                             content: Some(OaiMessageContent::Text(text.clone())),
                             tool_calls: None,
                             tool_call_id: None,
+                            reasoning_content: None,
                         });
                     }
                 }
@@ -212,6 +234,7 @@ impl LlmDriver for OpenAIDriver {
                         content: Some(OaiMessageContent::Text(text.clone())),
                         tool_calls: None,
                         tool_call_id: None,
+                        reasoning_content: None,
                     });
                 }
                 (Role::Assistant, MessageContent::Text(text)) => {
@@ -220,6 +243,7 @@ impl LlmDriver for OpenAIDriver {
                         content: Some(OaiMessageContent::Text(text.clone())),
                         tool_calls: None,
                         tool_call_id: None,
+                        reasoning_content: None,
                     });
                 }
                 (Role::User, MessageContent::Blocks(blocks)) => {
@@ -241,6 +265,7 @@ impl LlmDriver for OpenAIDriver {
                                     )),
                                     tool_calls: None,
                                     tool_call_id: Some(tool_use_id.clone()),
+                                    reasoning_content: None,
                                 });
                             }
                             ContentBlock::Text { text } => {
@@ -263,6 +288,7 @@ impl LlmDriver for OpenAIDriver {
                             content: Some(OaiMessageContent::Parts(parts)),
                             tool_calls: None,
                             tool_call_id: None,
+                            reasoning_content: None,
                         });
                     }
                 }
@@ -308,6 +334,11 @@ impl LlmDriver for OpenAIDriver {
                             Some(tool_calls)
                         },
                         tool_call_id: None,
+                        reasoning_content: if has_tool_calls && self.kimi_needs_reasoning_content(&request.model) {
+                            Some(String::new())
+                        } else {
+                            None
+                        },
                     });
                 }
                 _ => {}
@@ -346,11 +377,25 @@ impl LlmDriver for OpenAIDriver {
             messages: oai_messages,
             max_tokens: mt,
             max_completion_tokens: mct,
-            temperature: if rejects_temperature(&request.model) { None } else { Some(request.temperature) },
+            temperature: if self.kimi_needs_reasoning_content(&request.model) {
+                // Kimi with thinking disabled uses fixed 0.6 for multi-turn compatibility.
+                Some(0.6)
+            } else if temperature_must_be_one(&request.model) {
+                Some(1.0)
+            } else if rejects_temperature(&request.model) {
+                None
+            } else {
+                Some(request.temperature)
+            },
             tools: oai_tools,
             tool_choice,
             stream: false,
             stream_options: None,
+            thinking: if self.kimi_needs_reasoning_content(&request.model) {
+                Some(serde_json::json!({"type": "disabled"}))
+            } else {
+                None
+            },
         };
 
         let max_retries = 3;
@@ -615,6 +660,7 @@ impl LlmDriver for OpenAIDriver {
                 content: Some(OaiMessageContent::Text(system.clone())),
                 tool_calls: None,
                 tool_call_id: None,
+                reasoning_content: None,
             });
         }
 
@@ -627,6 +673,7 @@ impl LlmDriver for OpenAIDriver {
                             content: Some(OaiMessageContent::Text(text.clone())),
                             tool_calls: None,
                             tool_call_id: None,
+                            reasoning_content: None,
                         });
                     }
                 }
@@ -636,6 +683,7 @@ impl LlmDriver for OpenAIDriver {
                         content: Some(OaiMessageContent::Text(text.clone())),
                         tool_calls: None,
                         tool_call_id: None,
+                        reasoning_content: None,
                     });
                 }
                 (Role::Assistant, MessageContent::Text(text)) => {
@@ -644,6 +692,7 @@ impl LlmDriver for OpenAIDriver {
                         content: Some(OaiMessageContent::Text(text.clone())),
                         tool_calls: None,
                         tool_call_id: None,
+                        reasoning_content: None,
                     });
                 }
                 (Role::User, MessageContent::Blocks(blocks)) => {
@@ -661,6 +710,7 @@ impl LlmDriver for OpenAIDriver {
                                 )),
                                 tool_calls: None,
                                 tool_call_id: Some(tool_use_id.clone()),
+                                reasoning_content: None,
                             });
                         }
                     }
@@ -703,6 +753,11 @@ impl LlmDriver for OpenAIDriver {
                             Some(tool_calls_out)
                         },
                         tool_call_id: None,
+                        reasoning_content: if has_tool_calls && self.kimi_needs_reasoning_content(&request.model) {
+                            Some(String::new())
+                        } else {
+                            None
+                        },
                     });
                 }
                 _ => {}
@@ -741,11 +796,24 @@ impl LlmDriver for OpenAIDriver {
             messages: oai_messages,
             max_tokens: mt,
             max_completion_tokens: mct,
-            temperature: if rejects_temperature(&request.model) { None } else { Some(request.temperature) },
+            temperature: if self.kimi_needs_reasoning_content(&request.model) {
+                Some(0.6)
+            } else if temperature_must_be_one(&request.model) {
+                Some(1.0)
+            } else if rejects_temperature(&request.model) {
+                None
+            } else {
+                Some(request.temperature)
+            },
             tools: oai_tools,
             tool_choice,
             stream: true,
             stream_options: Some(serde_json::json!({"include_usage": true})),
+            thinking: if self.kimi_needs_reasoning_content(&request.model) {
+                Some(serde_json::json!({"type": "disabled"}))
+            } else {
+                None
+            },
         };
 
         // Retry loop for the initial HTTP request
diff --git a/crates/openfang-runtime/src/lib.rs b/crates/openfang-runtime/src/lib.rs
index 22019ae0..5722d148 100644
--- a/crates/openfang-runtime/src/lib.rs
+++ b/crates/openfang-runtime/src/lib.rs
@@ -3,6 +3,10 @@
 //! Manages the agent execution loop, LLM driver abstraction,
 //! tool execution, and WASM sandboxing for untrusted skill/plugin code.
 
+/// Default User-Agent header sent with all outgoing HTTP requests.
+/// Some LLM providers (e.g. Moonshot, Qwen) reject requests without one.
+pub const USER_AGENT: &str = "openfang/0.3.45";
+
 pub mod a2a;
 pub mod agent_loop;
 pub mod apply_patch;
diff --git a/crates/openfang-runtime/src/llm_errors.rs b/crates/openfang-runtime/src/llm_errors.rs
index 5599b0f9..0b76c0d9 100644
--- a/crates/openfang-runtime/src/llm_errors.rs
+++ b/crates/openfang-runtime/src/llm_errors.rs
@@ -281,6 +281,7 @@ pub fn classify_error(message: &str, status: Option<u16>) -> ClassifiedError {
                     return build(LlmErrorCategory::Auth);
                 }
             }
+            404 => return build(LlmErrorCategory::ModelNotFound),
             _ => {}
         }
     }
diff --git a/crates/openfang-runtime/src/model_catalog.rs b/crates/openfang-runtime/src/model_catalog.rs
index 2ad54866..8822c460 100644
--- a/crates/openfang-runtime/src/model_catalog.rs
+++ b/crates/openfang-runtime/src/model_catalog.rs
@@ -7,10 +7,10 @@ use openfang_types::model_catalog::{
     AuthStatus, ModelCatalogEntry, ModelTier, ProviderInfo, AI21_BASE_URL, ANTHROPIC_BASE_URL,
     BEDROCK_BASE_URL, CEREBRAS_BASE_URL, CHUTES_BASE_URL, COHERE_BASE_URL, DEEPSEEK_BASE_URL,
     FIREWORKS_BASE_URL, GEMINI_BASE_URL, GITHUB_COPILOT_BASE_URL, GROQ_BASE_URL,
-    HUGGINGFACE_BASE_URL, LMSTUDIO_BASE_URL, MINIMAX_BASE_URL, MISTRAL_BASE_URL,
-    MOONSHOT_BASE_URL, OLLAMA_BASE_URL, LEMONADE_BASE_URL, OPENAI_BASE_URL, OPENROUTER_BASE_URL,
-    PERPLEXITY_BASE_URL, QIANFAN_BASE_URL, QWEN_BASE_URL, REPLICATE_BASE_URL,
-    SAMBANOVA_BASE_URL, TOGETHER_BASE_URL, VENICE_BASE_URL, VLLM_BASE_URL,
+    HUGGINGFACE_BASE_URL, KIMI_CODING_BASE_URL, LEMONADE_BASE_URL, LMSTUDIO_BASE_URL,
+    MINIMAX_BASE_URL, MISTRAL_BASE_URL, MOONSHOT_BASE_URL, OLLAMA_BASE_URL, OPENAI_BASE_URL,
+    OPENROUTER_BASE_URL, PERPLEXITY_BASE_URL, QIANFAN_BASE_URL, QWEN_BASE_URL,
+    REPLICATE_BASE_URL, SAMBANOVA_BASE_URL, TOGETHER_BASE_URL, VENICE_BASE_URL, VLLM_BASE_URL,
     VOLCENGINE_BASE_URL, VOLCENGINE_CODING_BASE_URL, XAI_BASE_URL, ZAI_BASE_URL,
     ZAI_CODING_BASE_URL, ZHIPU_BASE_URL, ZHIPU_CODING_BASE_URL,
 };
@@ -688,6 +688,15 @@ fn builtin_providers() -> Vec<ProviderInfo> {
             auth_status: AuthStatus::Missing,
             model_count: 0,
         },
+        ProviderInfo {
+            id: "kimi_coding".into(),
+            display_name: "Kimi for Code".into(),
+            api_key_env: "KIMI_API_KEY".into(),
+            base_url: KIMI_CODING_BASE_URL.into(),
+            key_required: true,
+            auth_status: AuthStatus::Missing,
+            model_count: 0,
+        },
         ProviderInfo {
             id: "qianfan".into(),
             display_name: "Baidu Qianfan".into(),
@@ -2991,6 +3000,37 @@ fn builtin_models() -> Vec<ModelCatalogEntry> {
             aliases: vec!["codegeex".into()],
         },
         // ══════════════════════════════════════════════════════════════
+        // Z.AI Coding / GLM Coding Models (2)
+        // ══════════════════════════════════════════════════════════════
+        ModelCatalogEntry {
+            id: "glm-5-coding".into(),
+            display_name: "GLM-5 Coding".into(),
+            provider: "zai_coding".into(),
+            tier: ModelTier::Frontier,
+            context_window: 131_072,
+            max_output_tokens: 16_384,
+            input_cost_per_m: 2.00,
+            output_cost_per_m: 8.00,
+            supports_tools: true,
+            supports_vision: false,
+            supports_streaming: true,
+            aliases: vec!["glm-5-code".into(), "glm-coding".into()],
+        },
+        ModelCatalogEntry {
+            id: "glm-4.7-coding".into(),
+            display_name: "GLM-4.7 Coding".into(),
+            provider: "zai_coding".into(),
+            tier: ModelTier::Smart,
+            context_window: 131_072,
+            max_output_tokens: 16_384,
+            input_cost_per_m: 1.50,
+            output_cost_per_m: 5.00,
+            supports_tools: true,
+            supports_vision: false,
+            supports_streaming: true,
+            aliases: vec!["glm-4.7-code".into()],
+        },
+        // ══════════════════════════════════════════════════════════════
         // Moonshot / Kimi (5)
         // ══════════════════════════════════════════════════════════════
         ModelCatalogEntry {
@@ -3061,6 +3101,23 @@ fn builtin_models() -> Vec<ModelCatalogEntry> {
             supports_tools: true,
             supports_vision: true,
             supports_streaming: true,
+            aliases: vec!["kimi-k2.5-0711".into()],
+        },
+        // ══════════════════════════════════════════════════════════════
+        // Kimi for Code (1)
+        // ══════════════════════════════════════════════════════════════
+        ModelCatalogEntry {
+            id: "kimi-for-coding".into(),
+            display_name: "Kimi For Coding".into(),
+            provider: "kimi_coding".into(),
+            tier: ModelTier::Frontier,
+            context_window: 262_144,
+            max_output_tokens: 32_768,
+            input_cost_per_m: 0.0,
+            output_cost_per_m: 0.0,
+            supports_tools: true,
+            supports_vision: true,
+            supports_streaming: true,
             aliases: vec![],
         },
         // ══════════════════════════════════════════════════════════════
@@ -3492,7 +3549,7 @@ mod tests {
     #[test]
     fn test_catalog_has_providers() {
         let catalog = ModelCatalog::new();
-        assert_eq!(catalog.list_providers().len(), 37);
+        assert_eq!(catalog.list_providers().len(), 38);
     }
 
     #[test]
diff --git a/crates/openfang-runtime/src/web_fetch.rs b/crates/openfang-runtime/src/web_fetch.rs
index bd916cef..22681e0e 100644
--- a/crates/openfang-runtime/src/web_fetch.rs
+++ b/crates/openfang-runtime/src/web_fetch.rs
@@ -23,6 +23,7 @@ impl WebFetchEngine {
     /// Create a new fetch engine from config with a shared cache.
     pub fn new(config: WebFetchConfig, cache: Arc<WebCache>) -> Self {
         let client = reqwest::Client::builder()
+            .user_agent(crate::USER_AGENT)
             .timeout(std::time::Duration::from_secs(config.timeout_secs))
             .gzip(true)
             .deflate(true)
@@ -71,7 +72,7 @@ impl WebFetchEngine {
             "DELETE" => self.client.delete(url),
             _ => self.client.get(url),
         };
-        req = req.header("User-Agent", "Mozilla/5.0 (compatible; OpenFangAgent/0.1)");
+        req = req.header("User-Agent", format!("Mozilla/5.0 (compatible; {})", crate::USER_AGENT));
 
         // Add custom headers
         if let Some(hdrs) = headers {
diff --git a/crates/openfang-types/src/agent.rs b/crates/openfang-types/src/agent.rs
index 3aa91416..5f002b77 100644
--- a/crates/openfang-types/src/agent.rs
+++ b/crates/openfang-types/src/agent.rs
@@ -265,7 +265,7 @@ impl Default for ResourceQuota {
             max_memory_bytes: 256 * 1024 * 1024, // 256 MB
             max_cpu_time_ms: 30_000,             // 30 seconds
             max_tool_calls_per_minute: 60,
-            max_llm_tokens_per_hour: 1_000_000,
+            max_llm_tokens_per_hour: 0, // unlimited by default
             max_network_bytes_per_hour: 100 * 1024 * 1024, // 100 MB
             max_cost_per_hour_usd: 0.0, // unlimited by default
             max_cost_per_day_usd: 0.0,   // unlimited
@@ -1169,4 +1169,131 @@ provider = "openai"
         assert_eq!(cfg.model, "gpt-4o");
         assert_eq!(cfg.provider, "openai");
     }
+
+    // ----- Multi-line system_prompt TOML tests (wizard generateToml output) -----
+
+    #[test]
+    fn test_manifest_multiline_system_prompt_toml() {
+        // This is the exact TOML format the dashboard wizard generateToml() now produces
+        let toml_str = r#"
+name = "brand-guardian"
+module = "builtin:chat"
+
+[model]
+provider = "google"
+model = "gemini-3-flash-preview"
+system_prompt = """
+You are Brand Guardian, an expert brand strategist.
+
+Your Core Mission:
+- Develop brand strategy including purpose, vision, mission, values
+- Design complete visual identity systems
+- Establish brand voice and messaging architecture
+
+Critical Rules:
+- Establish comprehensive brand foundation before tactical implementation
+- Ensure all brand elements work as a cohesive system
+"""
+"#;
+        let manifest: AgentManifest = toml::from_str(toml_str).unwrap();
+        assert_eq!(manifest.name, "brand-guardian");
+        assert_eq!(manifest.model.provider, "google");
+        assert_eq!(manifest.model.model, "gemini-3-flash-preview");
+        assert!(manifest.model.system_prompt.contains("Brand Guardian"));
+        assert!(manifest.model.system_prompt.contains("Critical Rules:"));
+        // Verify newlines are preserved
+        assert!(manifest.model.system_prompt.contains('\n'));
+    }
+
+    #[test]
+    fn test_manifest_multiline_system_prompt_with_quotes() {
+        // System prompt containing double quotes (common in persona prompts)
+        let toml_str = r#"
+name = "test-agent"
+
+[model]
+provider = "groq"
+model = "llama-3.3-70b-versatile"
+system_prompt = """
+You are a "helpful" assistant.
+When users say "hello", respond warmly.
+"""
+"#;
+        let manifest: AgentManifest = toml::from_str(toml_str).unwrap();
+        assert!(manifest.model.system_prompt.contains("\"helpful\""));
+        assert!(manifest.model.system_prompt.contains("\"hello\""));
+    }
+
+    #[test]
+    fn test_manifest_multiline_system_prompt_with_code_blocks() {
+        // System prompt containing markdown-style code blocks
+        let toml_str = r#"
+name = "coder"
+
+[model]
+provider = "deepseek"
+model = "deepseek-chat"
+system_prompt = """
+You are a coding assistant.
+
+Example output format:
+```python
+def hello():
+    print("world")
+```
+
+Always use proper indentation.
+"""
+"#;
+        let manifest: AgentManifest = toml::from_str(toml_str).unwrap();
+        assert!(manifest.model.system_prompt.contains("```python"));
+        assert!(manifest.model.system_prompt.contains("def hello()"));
+    }
+
+    #[test]
+    fn test_manifest_single_line_system_prompt_still_works() {
+        // Ensure the old single-line format still parses fine
+        let toml_str = r#"
+name = "simple"
+
+[model]
+provider = "groq"
+model = "llama-3.3-70b-versatile"
+system_prompt = "You are a helpful assistant."
+"#;
+        let manifest: AgentManifest = toml::from_str(toml_str).unwrap();
+        assert_eq!(
+            manifest.model.system_prompt,
+            "You are a helpful assistant."
+        );
+    }
+
+    #[test]
+    fn test_manifest_wizard_custom_profile_with_capabilities() {
+        // Full wizard output when profile=custom with capabilities block
+        let toml_str = r#"
+name = "brand-guardian"
+module = "builtin:chat"
+
+[model]
+provider = "google"
+model = "gemini-3-flash-preview"
+system_prompt = """
+You are Brand Guardian.
+Protect brand consistency across all touchpoints.
+"""
+
+[capabilities]
+memory_read = ["*"]
+memory_write = ["self.*"]
+"#;
+        let manifest: AgentManifest = toml::from_str(toml_str).unwrap();
+        assert_eq!(manifest.name, "brand-guardian");
+        assert!(manifest.model.system_prompt.contains("Brand Guardian"));
+        assert_eq!(manifest.capabilities.memory_read, vec!["*".to_string()]);
+        assert_eq!(
+            manifest.capabilities.memory_write,
+            vec!["self.*".to_string()]
+        );
+    }
 }
diff --git a/crates/openfang-types/src/config.rs b/crates/openfang-types/src/config.rs
index fd799292..76c3a510 100644
--- a/crates/openfang-types/src/config.rs
+++ b/crates/openfang-types/src/config.rs
@@ -1108,9 +1108,8 @@ pub struct BudgetConfig {
     pub max_monthly_usd: f64,
     /// Alert threshold as a fraction (0.0 - 1.0). Trigger warnings at this % of any limit.
     pub alert_threshold: f64,
-    /// Default per-agent hourly token limit override. When set (> 0), agents that
-    /// still have the built-in default (1,000,000) or a lower per-agent value will
-    /// be overridden to this value. Set to 0 to keep each agent's own limit.
+    /// Default per-agent hourly token limit override. When set (> 0), all agents
+    /// will be overridden to this value. Set to 0 to keep each agent's own limit.
     /// Use this to globally raise or lower the token budget for all agents.
     pub default_max_llm_tokens_per_hour: u64,
 }
diff --git a/crates/openfang-types/src/model_catalog.rs b/crates/openfang-types/src/model_catalog.rs
index 5413d26c..b38e51b6 100644
--- a/crates/openfang-types/src/model_catalog.rs
+++ b/crates/openfang-types/src/model_catalog.rs
@@ -42,7 +42,8 @@ pub const ZHIPU_CODING_BASE_URL: &str = "https://open.bigmodel.cn/api/coding/paa
 /// Z.AI domain aliases (same API, different domain).
 pub const ZAI_BASE_URL: &str = "https://api.z.ai/api/paas/v4";
 pub const ZAI_CODING_BASE_URL: &str = "https://api.z.ai/api/coding/paas/v4";
-pub const MOONSHOT_BASE_URL: &str = "https://api.moonshot.cn/v1";
+pub const MOONSHOT_BASE_URL: &str = "https://api.moonshot.ai/v1";
+pub const KIMI_CODING_BASE_URL: &str = "https://api.kimi.com/coding";
 pub const QIANFAN_BASE_URL: &str = "https://qianfan.baidubce.com/v2";
 pub const VOLCENGINE_BASE_URL: &str = "https://ark.cn-beijing.volces.com/api/v3";
 pub const VOLCENGINE_CODING_BASE_URL: &str = "https://ark.cn-beijing.volces.com/api/coding/v3";