Compaction overhaul (#5186)

Co-authored-by: Douwe Osinga <douwe@squareup.com>
Co-authored-by: David Katz <dkatz@squareup.com>
Co-authored-by: Alex Hancock <alexhancock@block.xyz>

crates/goose/src/providers/formats/anthropic.rs

@@ -90,10 +90,7 @@ pub fn format_messages(messages: &[Message]) -> Vec<Value> {
                 MessageContent::ToolConfirmationRequest(_tool_confirmation_request) => {
                     // Skip tool confirmation requests
                 }
-                MessageContent::ContextLengthExceeded(_) => {
-                    // Skip
-                }
-                MessageContent::SummarizationRequested(_) => {
+                MessageContent::ConversationCompacted(_) => {
                     // Skip
                 }
                 MessageContent::Thinking(thinking) => {

crates/goose/src/providers/formats/bedrock.rs

@@ -48,10 +48,7 @@ pub fn to_bedrock_message_content(content: &MessageContent) -> Result<bedrock::C
             // Redacted thinking blocks are not supported in Bedrock - skip
             bedrock::ContentBlock::Text("".to_string())
         }
-        MessageContent::ContextLengthExceeded(_) => {
-            bail!("ContextLengthExceeded should not get passed to the provider")
-        }
-        MessageContent::SummarizationRequested(_) => {
+        MessageContent::ConversationCompacted(_) => {
             bail!("SummarizationRequested should not get passed to the provider")
         }
         MessageContent::ToolRequest(tool_req) => {

crates/goose/src/providers/formats/databricks.rs

@@ -127,10 +127,7 @@ fn format_messages(messages: &[Message], image_format: &ImageFormat) -> Vec<Data
                         }
                     }
                 }
-                MessageContent::ContextLengthExceeded(_) => {
-                    continue;
-                }
-                MessageContent::SummarizationRequested(_) => {
+                MessageContent::ConversationCompacted(_) => {
                     continue;
                 }
                 MessageContent::ToolResponse(response) => {

crates/goose/src/providers/formats/openai.rs

@@ -95,10 +95,7 @@ pub fn format_messages(messages: &[Message], image_format: &ImageFormat) -> Vec<
                     // Redacted thinking blocks are not directly used in OpenAI format
                     continue;
                 }
-                MessageContent::ContextLengthExceeded(_) => {
-                    continue;
-                }
-                MessageContent::SummarizationRequested(_) => {
+                MessageContent::ConversationCompacted(_) => {
                     continue;
                 }
                 MessageContent::ToolRequest(request) => match &request.tool_call {

crates/goose/src/providers/formats/snowflake.rs

@@ -53,10 +53,7 @@ pub fn format_messages(messages: &[Message]) -> Vec<Value> {
                 MessageContent::ToolConfirmationRequest(_) => {
                     // Skip tool confirmation requests
                 }
-                MessageContent::ContextLengthExceeded(_) => {
-                    // Skip
-                }
-                MessageContent::SummarizationRequested(_) => {
+                MessageContent::ConversationCompacted(_) => {
                     // Skip
                 }
                 MessageContent::Thinking(_thinking) => {

crates/goose/src/providers/usage_estimator.rs

@@ -1,6 +1,6 @@
 use crate::conversation::message::Message;
 use crate::providers::base::ProviderUsage;
-use crate::token_counter::create_async_token_counter;
+use crate::token_counter::create_token_counter;
 use anyhow::Result;
 use rmcp::model::Tool;
 
@@ -17,7 +17,7 @@ pub async fn ensure_usage_tokens(
         return Ok(());
     }
 
-    let token_counter = create_async_token_counter()
+    let token_counter = create_token_counter()
         .await
         .map_err(|e| anyhow::anyhow!("Failed to create token counter: {}", e))?;
 

crates/goose/src/providers/utils.rs

@@ -1,6 +1,8 @@
 use super::base::Usage;
 use super::errors::GoogleErrorCode;
+use crate::config::paths::Paths;
 use crate::model::ModelConfig;
+use crate::providers::errors::{OpenAIError, ProviderError};
 use anyhow::Result;
 use base64::Engine;
 use regex::Regex;
@@ -12,8 +14,6 @@ use std::io::Read;
 use std::path::Path;
 use std::time::Duration;
 
-use crate::providers::errors::{OpenAIError, ProviderError};
-
 #[derive(serde::Deserialize)]
 struct OpenAIErrorResponse {
     error: OpenAIError,
@@ -461,14 +461,32 @@ pub fn emit_debug_trace<T1, T2>(
     T1: ?Sized + Serialize,
     T2: ?Sized + Serialize,
 {
-    tracing::debug!(
-        model_config = %serde_json::to_string_pretty(model_config).unwrap_or_default(),
-        input = %serde_json::to_string_pretty(payload).unwrap_or_default(),
-        output = %serde_json::to_string_pretty(response).unwrap_or_default(),
-        input_tokens = ?usage.input_tokens.unwrap_or_default(),
-        output_tokens = ?usage.output_tokens.unwrap_or_default(),
-        total_tokens = ?usage.total_tokens.unwrap_or_default(),
-    );
+    let logs_dir = Paths::in_state_dir("logs");
+
+    if let Err(e) = std::fs::create_dir_all(&logs_dir) {
+        tracing::warn!("Failed to create logs directory: {}", e);
+        return;
+    }
+
+    let log_path = |i| logs_dir.join(format!("llm_request.{}.json", i));
+
+    for i in (0..4).rev() {
+        let _ = std::fs::rename(log_path(i), log_path(i + 1));
+    }
+
+    let data = serde_json::json!({
+        "model_config": model_config,
+        "input": payload,
+        "output": response,
+        "usage": usage,
+    });
+
+    if let Err(e) = std::fs::write(
+        log_path(0),
+        serde_json::to_string_pretty(&data).unwrap_or_default(),
+    ) {
+        tracing::warn!("Failed to write log file: {}", e);
+    }
 }
 
 /// Safely parse a JSON string that may contain doubly-encoded or malformed JSON.