feat: make context length configurable instead of hardcoded

cli/src/commands/autopilot/mod.rs

@@ -1648,6 +1648,9 @@ impl stakpak_gateway::dispatcher::RunOverrideResolver for ProfileRunOverrideReso
                 .map(stakpak_gateway::client::AutoApproveOverride::AllowList),
             system_prompt: resolved.system_prompt,
             max_turns: resolved.max_turns,
+            context_window: resolved.context_window,
+            context_budget_threshold: resolved.context_budget_threshold,
+            keep_last_n_assistant_messages: resolved.keep_last_n_assistant_messages,
         };
 
         if overrides.is_empty() {

cli/src/commands/watch/commands/run.rs

@@ -901,6 +901,9 @@ fn resolve_schedule_profile_overrides(
         auto_approve: normalized_auto_approve,
         system_prompt: resolved.system_prompt,
         max_turns: resolved.max_turns,
+        context_window: resolved.context_window,
+        context_budget_threshold: resolved.context_budget_threshold,
+        keep_last_n_assistant_messages: resolved.keep_last_n_assistant_messages,
     };
 
     let overrides = if overrides.is_empty() {

cli/src/config/app.rs

@@ -55,6 +55,12 @@ pub struct AppConfig {
     pub system_prompt: Option<String>,
     /// Optional max turn override for sessions using this profile.
     pub max_turns: Option<usize>,
+    /// Optional context window override for sessions using this profile.
+    pub context_window: Option<u64>,
+    /// Optional context budget threshold for sessions using this profile.
+    pub context_budget_threshold: Option<f32>,
+    /// Optional keep-last-N-assistant-messages for sessions using this profile.
+    pub keep_last_n_assistant_messages: Option<usize>,
     /// Unique ID for anonymous telemetry
     pub anonymous_id: Option<String>,
     /// Whether to collect telemetry data
@@ -170,6 +176,9 @@ impl AppConfig {
             subagent: profile_config.subagent,
             system_prompt: profile_config.system_prompt,
             max_turns: profile_config.max_turns,
+            context_window: profile_config.context_window,
+            context_budget_threshold: profile_config.context_budget_threshold,
+            keep_last_n_assistant_messages: profile_config.keep_last_n_assistant_messages,
             anonymous_id: settings.anonymous_id,
             collect_telemetry: settings.collect_telemetry,
             editor: settings.editor,

cli/src/config/profile.rs

@@ -99,6 +99,23 @@ pub struct ProfileConfig {
     #[serde(skip_serializing_if = "Option::is_none")]
     pub max_turns: Option<usize>,
 
+    /// Override the model's context window size (in tokens).
+    /// When set, this value replaces the model's default context window
+    /// for budget and trimming calculations.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub context_window: Option<u64>,
+
+    /// Fraction of the context window at which context trimming triggers.
+    /// Range: 0.1–1.0 (e.g. 0.8 = start trimming at 80% of context window).
+    /// Default: 0.8 (80%).
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub context_budget_threshold: Option<f32>,
+
+    /// Number of recent assistant messages to keep untrimmed during context
+    /// trimming. Default: 5.
+    #[serde(skip_serializing_if = "Option::is_none")]
+    pub keep_last_n_assistant_messages: Option<usize>,
+
     // =========================================================================
     // Legacy model fields - kept for backward compatibility during migration
     // These are read but deprecated (will migrate to 'model' field)
@@ -140,6 +157,9 @@ impl ProfileConfig {
                 recent_models: default.recent_models.clone(),
                 system_prompt: default.system_prompt.clone(),
                 max_turns: default.max_turns,
+                context_window: default.context_window,
+                context_budget_threshold: default.context_budget_threshold,
+                keep_last_n_assistant_messages: default.keep_last_n_assistant_messages,
                 // Enable warden for readonly sandboxed execution
                 warden: Some(WardenConfig::readonly_profile()),
                 // Don't copy allowed_tools/auto_approve - readonly has its own restrictions
@@ -437,6 +457,15 @@ impl ProfileConfig {
             max_turns: self
                 .max_turns
                 .or_else(|| other.and_then(|config| config.max_turns)),
+            context_window: self
+                .context_window
+                .or_else(|| other.and_then(|config| config.context_window)),
+            context_budget_threshold: self
+                .context_budget_threshold
+                .or_else(|| other.and_then(|config| config.context_budget_threshold)),
+            keep_last_n_assistant_messages: self
+                .keep_last_n_assistant_messages
+                .or_else(|| other.and_then(|config| config.keep_last_n_assistant_messages)),
             // Legacy fields - kept for reading only, not merged
             eco_model: None,
             smart_model: None,

cli/src/config/profile_resolver.rs

@@ -4,13 +4,16 @@ use stakpak_shared::utils::normalize_optional_string;
 
 use super::AppConfig;
 
-#[derive(Debug, Clone, Default, PartialEq, Eq)]
+#[derive(Debug, Clone, Default, PartialEq)]
 pub(crate) struct ResolvedProfileOverrides {
     pub model: Option<String>,
     pub auto_approve: Option<Vec<String>>,
     pub allowed_tools: Option<Vec<String>>,
     pub system_prompt: Option<String>,
     pub max_turns: Option<usize>,
+    pub context_window: Option<u64>,
+    pub context_budget_threshold: Option<f32>,
+    pub keep_last_n_assistant_messages: Option<usize>,
 }
 
 pub(crate) fn resolve_profile_run_overrides(
@@ -24,12 +27,18 @@ pub(crate) fn resolve_profile_run_overrides(
     let allowed_tools = normalize_tool_list(config.allowed_tools);
     let system_prompt = normalize_optional_string(config.system_prompt);
     let max_turns = config.max_turns;
+    let context_window = config.context_window;
+    let context_budget_threshold = config.context_budget_threshold;
+    let keep_last_n_assistant_messages = config.keep_last_n_assistant_messages;
 
     if model.is_none()
         && auto_approve.is_none()
         && allowed_tools.is_none()
         && system_prompt.is_none()
         && max_turns.is_none()
+        && context_window.is_none()
+        && context_budget_threshold.is_none()
+        && keep_last_n_assistant_messages.is_none()
     {
         return None;
     }
@@ -40,6 +49,9 @@ pub(crate) fn resolve_profile_run_overrides(
         allowed_tools,
         system_prompt,
         max_turns,
+        context_window,
+        context_budget_threshold,
+        keep_last_n_assistant_messages,
     })
 }
 

libs/api/src/client/mod.rs

@@ -61,6 +61,17 @@ pub struct AgentClientConfig {
     pub store_path: Option<String>,
     /// Hook registry for lifecycle events
     pub hook_registry: Option<HookRegistry<AgentState>>,
+    /// How many recent assistant messages to keep untrimmed when context
+    /// trimming is triggered (default: 5).
+    pub keep_last_n_assistant_messages: Option<usize>,
+    /// Fraction of the context window at which trimming triggers
+    /// (e.g. 0.8 = 80%, default: 0.8).
+    pub context_budget_threshold: Option<f32>,
+    /// Override the model's context window size (in tokens).
+    /// When set, replaces the model's built-in `context_window` for budget
+    /// calculations. Useful for local/custom models where the window may
+    /// not be auto-detected correctly.
+    pub context_window: Option<u64>,
 }
 
 impl AgentClientConfig {
@@ -94,6 +105,24 @@ impl AgentClientConfig {
         self.hook_registry = Some(registry);
         self
     }
+
+    /// Set context trimming: number of recent assistant messages to preserve
+    pub fn with_keep_last_n_assistant_messages(mut self, n: usize) -> Self {
+        self.keep_last_n_assistant_messages = Some(n);
+        self
+    }
+
+    /// Set context trimming: budget threshold (0.0–1.0)
+    pub fn with_context_budget_threshold(mut self, threshold: f32) -> Self {
+        self.context_budget_threshold = Some(threshold);
+        self
+    }
+
+    /// Set context window override (in tokens)
+    pub fn with_context_window(mut self, window: u64) -> Self {
+        self.context_window = Some(window);
+        self
+    }
 }
 
 // =============================================================================
@@ -221,8 +250,9 @@ impl AgentClient {
         hook_registry.register(
             LifecycleEvent::BeforeInference,
             Box::new(TaskBoardContextHook::new(TaskBoardContextHookOptions {
-                keep_last_n_assistant_messages: Some(5), // Keep the last 5 assistant messages in context
-                context_budget_threshold: Some(0.8),     // defaults to 0.8 (80%)
+                keep_last_n_assistant_messages: config.keep_last_n_assistant_messages.or(Some(5)),
+                context_budget_threshold: config.context_budget_threshold.or(Some(0.8)),
+                context_window: config.context_window,
             })),
         );
         let hook_registry = Arc::new(hook_registry);

libs/api/src/local/context_managers/task_board_context_manager.rs

@@ -8,6 +8,7 @@ use stakpak_shared::models::{
 pub struct TaskBoardContextManager {
     keep_last_n_assistant_messages: usize,
     context_budget_threshold: f32,
+    context_window: Option<u64>,
 }
 
 impl super::ContextManager for TaskBoardContextManager {
@@ -466,6 +467,7 @@ mod tests {
         TaskBoardContextManager::new(TaskBoardContextManagerOptions {
             keep_last_n_assistant_messages: 2, // Only keep last 2 assistant messages untrimmed
             context_budget_threshold: 0.8,
+            context_window: None,
         })
     }
 
@@ -1306,6 +1308,7 @@ mod tests {
         let cm = TaskBoardContextManager::new(TaskBoardContextManagerOptions {
             keep_last_n_assistant_messages: 2,
             context_budget_threshold: 0.8,
+            context_window: None,
         });
 
         // Build: user, assistant, user, assistant, user, user, assistant
@@ -1410,6 +1413,7 @@ mod tests {
         let cm = TaskBoardContextManager::new(TaskBoardContextManagerOptions {
             keep_last_n_assistant_messages: 2,
             context_budget_threshold: 0.8,
+            context_window: None,
         });
 
         // Build 10 turns of user/assistant
@@ -1514,6 +1518,7 @@ mod tests {
         let cm = TaskBoardContextManager::new(TaskBoardContextManagerOptions {
             keep_last_n_assistant_messages: 1,
             context_budget_threshold: 0.8,
+            context_window: None,
         });
 
         // Realistic agent flow:
@@ -1618,6 +1623,7 @@ mod tests {
         let cm = TaskBoardContextManager::new(TaskBoardContextManagerOptions {
             keep_last_n_assistant_messages: 10,
             context_budget_threshold: 0.8,
+            context_window: None,
         });
 
         // Only 3 assistant messages but keep_last_n = 10
@@ -1682,6 +1688,7 @@ mod tests {
         let cm = TaskBoardContextManager::new(TaskBoardContextManagerOptions {
             keep_last_n_assistant_messages: 0,
             context_budget_threshold: 0.8,
+            context_window: None,
         });
 
         let messages = vec![
@@ -1735,6 +1742,7 @@ mod tests {
         let cm = TaskBoardContextManager::new(TaskBoardContextManagerOptions {
             keep_last_n_assistant_messages: 1,
             context_budget_threshold: 0.8,
+            context_window: None,
         });
 
         // Use large assistant messages and small user messages so that trimming
@@ -1812,6 +1820,7 @@ mod tests {
         let cm = TaskBoardContextManager::new(TaskBoardContextManagerOptions {
             keep_last_n_assistant_messages: 2,
             context_budget_threshold: 0.8,
+            context_window: None,
         });
 
         // 10 turns → 20 messages, small window → establishes a trim index
@@ -1848,6 +1857,7 @@ mod tests {
         let cm_generous = TaskBoardContextManager::new(TaskBoardContextManagerOptions {
             keep_last_n_assistant_messages: 8,
             context_budget_threshold: 0.8,
+            context_window: None,
         });
 
         let (_, metadata2) = cm_generous.reduce_context_with_budget(messages, 100, metadata1, None);
@@ -1870,6 +1880,7 @@ mod tests {
         let cm = TaskBoardContextManager::new(TaskBoardContextManagerOptions {
             keep_last_n_assistant_messages: 2,
             context_budget_threshold: 0.8,
+            context_window: None,
         });
 
         // Build a conversation that's just under threshold without tools
@@ -1974,6 +1985,7 @@ mod tests {
         let cm = TaskBoardContextManager::new(TaskBoardContextManagerOptions {
             keep_last_n_assistant_messages: 3,
             context_budget_threshold: 0.8,
+            context_window: None,
         });
 
         // 5 turns of: user → assistant(tool_call) → tool(result) → assistant(follow-up)
@@ -2080,6 +2092,7 @@ mod tests {
         let cm = TaskBoardContextManager::new(TaskBoardContextManagerOptions {
             keep_last_n_assistant_messages: 50,
             context_budget_threshold: 0.8,
+            context_window: None,
         });
 
         // Simulate a 200k context window model (like Claude)
@@ -2362,6 +2375,7 @@ mod tests {
         let cm = TaskBoardContextManager::new(TaskBoardContextManagerOptions {
             keep_last_n_assistant_messages: 50,
             context_budget_threshold: 0.3,
+            context_window: None,
         });
 
         // Simulate a session with 10 turns (20 messages) — well under 50
@@ -2435,6 +2449,7 @@ mod tests {
         let cm = TaskBoardContextManager::new(TaskBoardContextManagerOptions {
             keep_last_n_assistant_messages: 3,
             context_budget_threshold: 0.8,
+            context_window: None,
         });
 
         // 6 turns: user + assistant with large content.
@@ -2519,6 +2534,7 @@ mod tests {
         let cm = TaskBoardContextManager::new(TaskBoardContextManagerOptions {
             keep_last_n_assistant_messages: 20,
             context_budget_threshold: 0.8,
+            context_window: None,
         });
 
         // Simulate a session with 30 turns of tool-heavy interaction.
@@ -2651,6 +2667,7 @@ mod tests {
         let cm = TaskBoardContextManager::new(TaskBoardContextManagerOptions {
             keep_last_n_assistant_messages: 2,
             context_budget_threshold: 0.8,
+            context_window: None,
         });
 
         // Build conversation that exceeds threshold
@@ -2742,6 +2759,7 @@ mod tests {
         let cm = TaskBoardContextManager::new(TaskBoardContextManagerOptions {
             keep_last_n_assistant_messages: 10,
             context_budget_threshold: 0.8,
+            context_window: None,
         });
 
         let messages = vec![
@@ -2790,6 +2808,7 @@ mod tests {
         let cm = TaskBoardContextManager::new(TaskBoardContextManagerOptions {
             keep_last_n_assistant_messages: 50, // high keep_last_n, like production
             context_budget_threshold: 0.3,
+            context_window: None,
         });
 
         // 5 turns with large assistant responses
@@ -2882,13 +2901,19 @@ pub struct TaskBoardContextManagerOptions {
     pub keep_last_n_assistant_messages: usize,
     /// Fraction of context window at which trimming triggers (e.g., 0.8 = 80%)
     pub context_budget_threshold: f32,
+    /// Override the model's context window size (in tokens).
+    /// When set, replaces the model's built-in `context_window` for budget
+    /// calculations. Useful for local/custom models where the window may
+    /// not be auto-detected correctly.
+    pub context_window: Option<u64>,
 }
 
 impl TaskBoardContextManager {
     pub fn new(options: TaskBoardContextManagerOptions) -> Self {
         Self {
             keep_last_n_assistant_messages: options.keep_last_n_assistant_messages,
             context_budget_threshold: options.context_budget_threshold,
+            context_window: options.context_window,
         }
     }
 }