add conversation compaction to preserve context instead of trimming

Author: alexylon

.sofosrc

@@ -363,7 +363,7 @@ This prints:
 
 ### Anthropic API
 - Uses Claude Messages API (not legacy Completions)
-- Model: claude-sonnet-4-5 (default)
+- Model: claude-sonnet-4-6 (default)
 - Supports tool calling and server-side tools (web_search)
 - API version: 2023-06-01
 - Usage tracking: Automatic token counting and cost calculation

CHANGELOG.md

@@ -4,6 +4,15 @@ All notable changes to Sofos are documented in this file.
 
 ## [Unreleased]
 
+### Added
+- Conversation compaction: replaces naive message trimming with intelligent context preservation
+  - Two-phase approach: truncates large tool results first, then summarizes older messages via the LLM
+  - Works with both Anthropic and OpenAI providers
+  - Auto-triggers at 80% of token budget before sending the next request
+  - `/compact` command for manual compaction
+  - Shows "Compacting conversation..." animation during summarization
+  - Falls back to trimming on failure or ESC interrupt
+
 ### Changed
 - Increase `MAX_TOOL_OUTPUT_TOKENS`
 

README.md

@@ -48,6 +48,7 @@ Tested on macOS but should work on Linux and Windows as well.
 - **MCP Integration** - Connect to external tools via Model Context Protocol
 - **Visual Diffs** - Colored change display
 - **Iterative Tools** - Up to 200 tool calls per request
+- **Context Compaction** - Summarizes older messages instead of dropping them
 - **Cost Tracking** - Session token usage and cost estimates
 - **Safe Mode** - Read-only operation mode
 
@@ -95,6 +96,7 @@ sofos
 - `/resume` - Resume previous session
 - `/clear` - Clear conversation history
 - `/think [on|off]` - Toggle extended thinking (shows status if no arg)
+- `/compact` - Compress conversation history to free up context (also triggers automatically at 80% token usage)
 - `/s` - Safe mode (read-only, prompt: **`λ:`**, blinking underscore (`_`) cursor)
 - `/n` - Normal mode (all tools, prompt: **`λ>`**, default cursor)
 - `/exit`, `/quit`, `/q`, `Ctrl+D` - Exit with cost summary
@@ -133,7 +135,7 @@ Exit summary shows token usage and estimated cost (based on official API pricing
     --api-key <KEY>          Anthropic API key (overrides env var)
     --openai-api-key <KEY>   OpenAI API key (overrides env var)
     --morph-api-key <KEY>    Morph API key (overrides env var)
-    --model <MODEL>          Model to use (default: claude-sonnet-4-5)
+    --model <MODEL>          Model to use (default: claude-sonnet-4-6)
     --morph-model <MODEL>    Morph model (default: morph-v3-fast)
     --max-tokens <N>         Max response tokens (default: 8192)
 -t, --enable-thinking        Enable extended thinking (default: false)

src/api/anthropic.rs

@@ -186,7 +186,7 @@ mod tests {
     fn test_request_with_thinking() {
         let thinking = Some(Thinking::enabled(3000));
         let request = CreateMessageRequest {
-            model: "claude-sonnet-4-5".to_string(),
+            model: "claude-sonnet-4-6".to_string(),
             max_tokens: 8192,
             messages: vec![],
             system: None,

src/cli.rs

@@ -30,7 +30,7 @@ pub struct Cli {
     #[arg(long)]
     pub check_connection: bool,
 
-    #[arg(long, default_value = "claude-sonnet-4-5")]
+    #[arg(long, default_value = "claude-sonnet-4-6")]
     pub model: String,
 
     #[arg(long, default_value = "morph-v3-fast")]

src/commands/builtin.rs

@@ -47,3 +47,8 @@ pub fn normal_mode_command(repl: &mut Repl) -> Result<CommandResult> {
     repl.disable_safe_mode();
     Ok(CommandResult::Continue)
 }
+
+pub fn compact_command(repl: &mut Repl) -> Result<CommandResult> {
+    repl.handle_compact_command()?;
+    Ok(CommandResult::Continue)
+}

src/commands/mod.rs

@@ -23,6 +23,7 @@ pub enum Command {
     ThinkStatus,
     SafeMode,
     NormalMode,
+    Compact,
 }
 
 impl Command {
@@ -36,6 +37,7 @@ impl Command {
             "/think" => Some(Command::ThinkStatus),
             "/s" => Some(Command::SafeMode),
             "/n" => Some(Command::NormalMode),
+            "/compact" => Some(Command::Compact),
             _ => None,
         }
     }
@@ -50,6 +52,7 @@ impl Command {
             Command::ThinkStatus => builtin::think_status_command(repl),
             Command::SafeMode => builtin::safe_mode_command(repl),
             Command::NormalMode => builtin::normal_mode_command(repl),
+            Command::Compact => builtin::compact_command(repl),
         }
     }
 }
@@ -66,4 +69,5 @@ pub static COMMANDS: &[&str] = &[
     "/think",
     "/s",
     "/n",
+    "/compact",
 ];

src/config.rs

@@ -8,6 +8,12 @@ pub struct SofosConfig {
     pub max_file_size: usize,
     #[allow(dead_code)]
     pub max_bash_output: usize,
+    /// Auto-compact when token usage exceeds this ratio of max_context_tokens
+    pub compaction_trigger_ratio: f64,
+    /// Number of recent messages to preserve during compaction
+    pub compaction_preserve_recent: usize,
+    /// Truncate tool results longer than this (chars) during compaction
+    pub tool_result_truncate_threshold: usize,
 }
 
 impl Default for SofosConfig {
@@ -18,6 +24,9 @@ impl Default for SofosConfig {
             max_tool_iterations: 200,
             max_file_size: 10 * 1024 * 1024,
             max_bash_output: 50 * 1024 * 1024,
+            compaction_trigger_ratio: 0.80,
+            compaction_preserve_recent: 20,
+            tool_result_truncate_threshold: 2000,
         }
     }
 }
@@ -77,5 +86,8 @@ mod tests {
         assert_eq!(config.max_tool_iterations, 200);
         assert_eq!(config.max_file_size, 10 * 1024 * 1024);
         assert_eq!(config.max_bash_output, 50 * 1024 * 1024);
+        assert!((config.compaction_trigger_ratio - 0.80).abs() < f64::EPSILON);
+        assert_eq!(config.compaction_preserve_recent, 20);
+        assert_eq!(config.tool_result_truncate_threshold, 2000);
     }
 }