Manage the KV cache disc usage from preferences and stats

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

src/database.rs

@@ -6,8 +6,8 @@ use std::path::Path;
 
 use crate::schema::{messages, preferences, projects, sessions};
 use crate::settings::{
-    DiagnosticPreferences, ExecutionPreferences, GenerationPreferences, ReasoningMode,
-    RuntimePreferences, SpeculativePreferences, SsdPreferences, SteeringPreferences,
+    DiagnosticPreferences, ExecutionPreferences, GenerationPreferences, KvCachePreferences,
+    ReasoningMode, RuntimePreferences, SpeculativePreferences, SsdPreferences, SteeringPreferences,
     StreamingCacheBudget,
 };
 
@@ -58,6 +58,10 @@ pub struct AppPreferences {
     /// Project the app reopens on. Cleared when that project goes away.
     pub last_project_id: Option<i32>,
     pub sidebar_width: i32,
+    pub kv_budget_gib: Option<i64>,
+    pub kv_min_tokens: Option<i32>,
+    pub kv_cold_max_tokens: Option<i32>,
+    pub kv_continued_interval_tokens: Option<i32>,
 }
 
 impl Default for AppPreferences {
@@ -103,6 +107,10 @@ impl Default for AppPreferences {
             sidebar_collapsed: false,
             last_project_id: None,
             sidebar_width: 276,
+            kv_budget_gib: None,
+            kv_min_tokens: None,
+            kv_cold_max_tokens: None,
+            kv_continued_interval_tokens: None,
         }
     }
 }
@@ -207,6 +215,28 @@ impl AppPreferences {
                     .map_err(|_| "Saved simulated memory is invalid.")?,
                 expert_profile_path: self.expert_profile_path.clone(),
             },
+            kv_cache: KvCachePreferences {
+                budget_gib: self
+                    .kv_budget_gib
+                    .map(u64::try_from)
+                    .transpose()
+                    .map_err(|_| "Saved KV cache budget is invalid.")?,
+                min_tokens: self
+                    .kv_min_tokens
+                    .map(u32::try_from)
+                    .transpose()
+                    .map_err(|_| "Saved KV cache minimum tokens is invalid.")?,
+                cold_max_tokens: self
+                    .kv_cold_max_tokens
+                    .map(u32::try_from)
+                    .transpose()
+                    .map_err(|_| "Saved KV cache cold maximum is invalid.")?,
+                continued_interval_tokens: self
+                    .kv_continued_interval_tokens
+                    .map(u32::try_from)
+                    .transpose()
+                    .map_err(|_| "Saved KV cache continued interval is invalid.")?,
+            },
         })
     }
 }
@@ -250,6 +280,10 @@ struct PreferenceChanges<'a> {
     endpoint_port: i32,
     endpoint_enabled: bool,
     endpoint_cors: bool,
+    kv_budget_gib: Option<i64>,
+    kv_min_tokens: Option<i32>,
+    kv_cold_max_tokens: Option<i32>,
+    kv_continued_interval_tokens: Option<i32>,
 }
 
 #[derive(Clone, Debug, Identifiable, Queryable, Selectable)]
@@ -534,6 +568,18 @@ impl Database {
                 endpoint_port,
                 endpoint_enabled,
                 endpoint_cors,
+                kv_budget_gib: runtime
+                    .kv_cache
+                    .budget_gib
+                    .map(i64::try_from)
+                    .transpose()
+                    .map_err(|_| "KV cache budget is too large.")?,
+                kv_min_tokens: runtime.kv_cache.min_tokens.map(|value| value as i32),
+                kv_cold_max_tokens: runtime.kv_cache.cold_max_tokens.map(|value| value as i32),
+                kv_continued_interval_tokens: runtime
+                    .kv_cache
+                    .continued_interval_tokens
+                    .map(|value| value as i32),
             })
             .returning(AppPreferences::as_returning())
             .get_result(&mut self.connection)
@@ -832,6 +878,12 @@ mod tests {
                 simulated_used_memory_gib: Some(8),
                 expert_profile_path: Some("/tmp/experts.json".into()),
             },
+            kv_cache: KvCachePreferences {
+                budget_gib: Some(16),
+                min_tokens: Some(256),
+                cold_max_tokens: Some(0),
+                continued_interval_tokens: Some(4_096),
+            },
             ..RuntimePreferences::default()
         };
         database