various fixes, added new anthropic models, smarter caching

Author: orual

Cargo.lock

@@ -26,7 +26,7 @@ summarization_prompt = """You are the subconscious of Entropy, a digital entity,
 
 [model]
 provider = "Anthropic"
-model = "claude-opus-4"
+model = "claude-opus-4-1"
 temperature = 0.8
 
 

bsky_agent/agents/entropy.toml

@@ -105,7 +105,7 @@ shared = true
 
 [model]
 provider = "Anthropic"
-model = "claude-sonnet-4"
+model = "claude-sonnet-4-5"
 temperature = 0.8         # Higher for creative public interactions
 
 [[agent.tool_rules]]
@@ -543,6 +543,7 @@ default_filter = { nsids = [
     "did:plc:ouylf4wfd75rdqupddfmeerk", # nutty.land (kris nuttycombe)
     "did:plc:dzvxvsiy3maw4iarpvizsj67", # dollspace.gay
     "did:plc:bnqkww7bjxaacajzvu5gswdf", # shreyan
+    "did:plc:hu35oubkccqrxl4ldgczpgw7", # siobhan
 ], allow_any_mentions = true, keywords = [ # Keywords to include
 ], languages = [ # Languages to include
 ], require_agent_participation = true, exclude_keywords = [
@@ -561,6 +562,7 @@ default_filter = { nsids = [
     "did:plc:boia3kqcyo3qnjw5fmqknib4", # jo wynter cannot stop pushing buttons
     "did:plc:vszw3ess46odfhnzdsy4huae", # se gyges got got one too many times
     "did:plc:lw7hallqsgrhwumgudkph44j",
+    "did:plc:djpcowss7aokaydcdraoaovz", # misha, siobhan's partner who wouldn't appreciate pattern replying
 ] }
 jetstream_endpoint = "ws://jetstream.hose:6008/subscribe"
 auto_connect_firehose = true # Control when to connect

bsky_agent/constellation.toml

@@ -1110,6 +1110,17 @@ impl MessageEndpoint for BlueskyEndpoint {
             tags.push(agent_name);
         }
 
+        if rich_text.grapheme_len() > 300 {
+            return Err(crate::CoreError::ToolExecutionFailed {
+                tool_name: "bluesky_endpoint".to_string(),
+                cause: format!(
+                    "Post text is too long ({} graphemes, max is 300)",
+                    rich_text.grapheme_len()
+                ),
+                parameters: serde_json::json!({ "text": rich_text.text }),
+            });
+        }
+
         // Create the post
         let agent = &self.agent;
         let text_copy = text.clone();

crates/pattern_core/src/context/message_router.rs

@@ -10,7 +10,7 @@ use crate::{
     Result,
     id::AgentId,
     memory::MemoryBlock,
-    message::{CacheControl, Message, MessageBatch},
+    message::{CacheControl, Message, MessageBatch, MessageOptions},
     tool::{DynamicTool, ToolRegistry},
 };
 use regex::Regex;
@@ -98,10 +98,11 @@ impl MemoryContext {
             // TODO: Make head/tail counts configurable via ContextConfig
             // For now, show first 1 and last 2 summary blocks.
             let clipped = clip_archive_summary(summary, 2, 4);
-            messages.push(crate::message::Message::system(format!(
+            let mut message = crate::message::Message::system(format!(
                 "Previous conversation summary:\n{}",
                 clipped
-            )));
+            ));
+            messages.push(message);
         }
 
         // Then add all the regular messages from batches
@@ -740,20 +741,18 @@ You MUST follow these workflow rules exactly (they will be enforced by the syste
         let mut result_batches = Vec::new();
         let total_batches = final_batches.len();
         let mut cache_points_used = 0;
-        const MAX_CACHE_POINTS: usize = 4;
+        const MAX_CACHE_POINTS: usize = 2; // one in system, one in tools, two left
 
         // Calculate which batches should get cache points
         // Strategy: cache early context and recent context
         let cache_positions: Vec<usize> = if total_batches <= MAX_CACHE_POINTS {
             // If we have 4 or fewer batches, cache the first message of each
-            (0..total_batches).collect()
+            vec![0, total_batches - 1]
         } else {
             // Otherwise, distribute cache points strategically:
-            // 1. First batch (early context)
-            // 2. 25% through
-            // 3. 50% through
-            // 4. Most recent batch
-            vec![0, total_batches / 4, total_batches / 2, total_batches - 1]
+            // 1. 50% through
+            // 2. Most recent batch
+            vec![total_batches / 2, total_batches - 1]
         };
 
         for (batch_idx, mut batch) in final_batches.into_iter().enumerate() {

crates/pattern_core/src/context/mod.rs

@@ -1604,6 +1604,7 @@ impl AgentContext {
                     // Estimate current token usage
                     let memory_blocks = self.handle.memory.get_all_non_recall();
                     let system_tokens = self.estimate_system_prompt_tokens(&memory_blocks);
+                    // TODO: need to not count prev turn thinking tokens in this value i think.
                     let message_tokens: usize = history
                         .batches
                         .iter()

crates/pattern_core/src/context/state.rs

@@ -905,23 +905,24 @@ impl BlueskyPost {
                         .images
                         .iter()
                         .map(|img| {
-                            // Try fullsize first, fallback to thumb with appropriate CDN format
-                            if img.fullsize.is_empty() {
-                                // Fallback to thumbnail - extract CID from blob JSON and use thumbnail CDN format
-                                if img.thumb.starts_with("http") {
-                                    img.thumb.clone()
-                                } else if let Some(cid) = extract_cid_from_blob(&img.thumb) {
-                                    if cid.starts_with("http") {
-                                        cid
-                                    } else {
-                                        format!("https://cdn.bsky.app/img/feed_thumbnail/plain/{}/{}@jpeg", did, cid)
-                                    }
+                            // Use thumbnail - extract CID from blob JSON and use thumbnail CDN format
+                            if img.thumb.starts_with("http") {
+                                img.thumb.clone()
+                            } else if let Some(cid) = extract_cid_from_blob(&img.thumb) {
+                                if cid.starts_with("http") {
+                                    cid
                                 } else {
-                                    // Final fallback: treat as plain CID
-                                    format!("https://cdn.bsky.app/img/feed_thumbnail/plain/{}/{}@jpeg", did, img.thumb)
+                                    format!(
+                                        "https://cdn.bsky.app/img/feed_thumbnail/plain/{}/{}@jpeg",
+                                        did, cid
+                                    )
                                 }
                             } else {
-                                convert_blob_to_url(&img.fullsize, did)
+                                // Final fallback: treat as plain CID
+                                format!(
+                                    "https://cdn.bsky.app/img/feed_thumbnail/plain/{}/{}@jpeg",
+                                    did, img.thumb
+                                )
                             }
                         })
                         .collect(),
@@ -1012,18 +1013,24 @@ impl BlueskyPost {
                             count: images.images.len(),
                             alt_texts: images.images.iter().map(|img| img.alt.clone()).collect(),
                             urls: images.images.iter().map(|img| {
-                                if img.fullsize.starts_with("http") {
-                                    img.fullsize.clone()
-                                } else if let Some(cid) = extract_cid_from_blob(&img.fullsize) {
+                                // Use thumbnail - extract CID from blob JSON and use thumbnail CDN format
+                                if img.thumb.starts_with("http") {
+                                    img.thumb.clone()
+                                } else if let Some(cid) = extract_cid_from_blob(&img.thumb) {
                                     if cid.starts_with("http") {
                                         cid
                                     } else {
-                                        // Use the parent post's DID (not the quoted post's DID) - thumbnail for LLM
-                                        format!("https://cdn.bsky.app/img/feed_thumbnail/plain/{}/{}@jpeg", did, cid)
+                                        format!(
+                                            "https://cdn.bsky.app/img/feed_thumbnail/plain/{}/{}@jpeg",
+                                            did, cid
+                                        )
                                     }
                                 } else {
-                                    // Use the parent post's DID for the fallback case too
-                                    format!("https://cdn.bsky.app/img/feed_thumbnail/plain/{}/{}@jpeg", did, img.thumb)
+                                    // Final fallback: treat as plain CID
+                                    format!(
+                                        "https://cdn.bsky.app/img/feed_thumbnail/plain/{}/{}@jpeg",
+                                        did, img.thumb
+                                    )
                                 }
                             }).collect(),
                         })
@@ -3240,7 +3247,7 @@ impl BlueskyFirehoseSource {
                 "\n💭 Reply option: @{} ({})\n",
                 post.handle, post.uri
             ));
-            message.push_str("If you choose to reply (by using send_message with target_type bluesky and the target_id set to the uri), your response must contain under 300 characters or it will be truncated.\n");
+            message.push_str("If you choose to reply (by using send_message with target_type bluesky and the target_id set to the uri), your response must contain under 300 graphemes or it won't go through. You can thread a longer reply by requesting a heartbeat when making a reply, posting the first part of the the reply, and then replying to the uri you get back from the first successful reply with the second part\n");
         }
 
         // Collect memory blocks
@@ -4169,8 +4176,6 @@ impl LexiconIngestor for PostIngestor {
                     if let Some(buffer) = &self.buffer {
                         let mut buffer_guard = buffer.lock();
                         if buffer_guard.queue_for_processing(event.clone()) {
-                            // Successfully queued for later processing
-                            tracing::debug!("Queued post for rate-limited processing");
                         } else {
                             // Queue full, drop the event
                             tracing::warn!(
@@ -4182,9 +4187,6 @@ impl LexiconIngestor for PostIngestor {
                         buffer_guard.push(event);
                     }
                 } else {
-                    // Can send immediately
-                    tracing::debug!("📤 Sending post to channel: @{}", post_to_filter.handle);
-
                     // Always mark activity for watchdog regardless of send success - we're processing messages
                     let mut act = self.last_activity_time.lock().await;
                     *act = std::time::Instant::now();
@@ -4206,9 +4208,6 @@ impl LexiconIngestor for PostIngestor {
                             if let Some(buffer) = &self.buffer {
                                 let mut buffer_guard = buffer.lock();
                                 if buffer_guard.queue_for_processing(event.clone()) {
-                                    tracing::debug!(
-                                        "Successfully queued post for rate-limited processing"
-                                    );
                                 } else {
                                     tracing::warn!(
                                         "Backpressure: processing queue full; dropping post from {}",

crates/pattern_core/src/data_source/bluesky.rs

@@ -542,7 +542,7 @@ pub fn strip_dt(s: &str) -> &str {
 
 /// Convert SurrealDB's Datetime type to a chrono DateTime
 pub fn from_surreal_datetime(dt: surrealdb::Datetime) -> chrono::DateTime<chrono::Utc> {
-    let datetime = chrono::NaiveDateTime::parse_from_str(&dt.to_string(), "d'%FT%T%.6fZ'")
+    let datetime = chrono::NaiveDateTime::parse_from_str(&dt.to_string(), "d'%FT%T%.fZ'")
         .expect("should be valid ISO-8601");
 
     datetime.and_utc()

crates/pattern_core/src/db/mod.rs

@@ -1019,7 +1019,9 @@ impl Request {
         for msg in &mut self.messages {
             if msg.role == ChatRole::User || msg.role == ChatRole::System {
                 if let MessageContent::Text(text) = &msg.content {
-                    let timestamp = msg.created_at.naive_local();
+                    use chrono::TimeZone;
+                    let time_zone = chrono::Local::now().timezone();
+                    let timestamp = time_zone.from_utc_datetime(&msg.created_at.naive_utc());
                     // injecting created time in to make agents less likely to be confused by artifacts and more temporally aware.
                     msg.content = MessageContent::Text(format!(
                         "<time_sync>created: {}</time_sync>\n{}",

crates/pattern_core/src/message.rs

@@ -124,10 +124,10 @@ impl ResponseOptions {
                 "anthropic-beta".to_string(),
                 "computer-use-2025-01-24".to_string(),
             ));
-            headers.push((
-                "anthropic-beta".to_string(),
-                "context-1m-2025-08-07".to_string(),
-            ));
+            // headers.push((
+            //     "anthropic-beta".to_string(),
+            //     "context-1m-2025-08-07".to_string(),
+            // ));
             headers.push((
                 "anthropic-beta".to_string(),
                 "code-execution-2025-05-22".to_string(),
@@ -426,7 +426,7 @@ impl GenAiClient {
                         if let ImageSource::Url(url) = source {
                             // Quick HEAD request to check if URL is accessible
                             let client = reqwest::Client::builder()
-                                .timeout(std::time::Duration::from_secs(5))
+                                .timeout(std::time::Duration::from_secs(2))
                                 .build()
                                 .ok();