openclacky 0.9.3 → 0.9.4

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: d842684d3cae23106509a9e47be986d51e518f8b953e66011783b32e1a121a6e
-  data.tar.gz: 34a8ef45f736724fd7e8356e032a4cd1102bd1a28b47f412419895bd8575caf7
+  metadata.gz: a9c12ba4b5e244e6dbbec7d723dc6d0f9c7e4e4523a82bafbb96f668e7b1a0f4
+  data.tar.gz: 5d6e7d71829aa2bafbe68dc825b6ec0e7bd4c1ad004804d43ef9121a46fa6671
 SHA512:
-  metadata.gz: cc9f6ee3d0b8ebf01346261dfd9dbcf6a77654bc40632b5730432fbe5ce23c604f30135739507c1802e837a4baf6097609ba26397c9eea2ec4555b115cd73dc9
-  data.tar.gz: 20ad77eb3191fadd1d4340ce7a92b16489be68e83dc4d33fa1d2e802a821074d1964f9a1e7afadcd7cbf9846b81dabebd5cbd918441bb3b3528bfd6212fb5f56
+  metadata.gz: 1dd7ede95260c506b7dfb89cd5596ea4ba9ac4e6949e17067b84ec6a9a2686892a27fafd1e25bf80a6c0ade211f398c3f6b8ab87d3e7167ea808fe88398f46be
+  data.tar.gz: 38771322fb7651494b4595d82e315a3a6ddf902ae3bf4141d4ea01a63a65ffd1c9adb1dcf91da43d29ead77f673025e1559d0ad688fafbfa7673738dcf393052

data/CHANGELOG.md

@@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [0.9.4] - 2026-03-16
+
+### Fixed
+- **Prompt cache strategy reverted to simple last-message anchoring**: the experimental assistant-message-anchored cache strategy introduced in v0.9.3 was causing regressions; caching is now restored to a simpler, proven approach where the last message is used as the cache breakpoint
+
 ## [0.9.3] - 2026-03-16
 
 ### Added

data/lib/clacky/client.rb

@@ -143,38 +143,16 @@ module Clacky
     # ── Prompt caching helpers ────────────────────────────────────────────────
 
     # Add cache_control marker to the appropriate message in the array.
-    #
-    # Strategy: mark the SECOND-TO-LAST non-injected assistant message.
-    #
-    # Rationale: Anthropic prompt caching is prefix-based. If we mark the last
-    # message (typically the latest user turn or tool_result), the cached prefix
-    # changes every request because tool results and user inputs vary each turn.
-    # This causes alternating miss/hit patterns observed in production.
-    #
-    # By placing the breakpoint on the most recent ASSISTANT message that precedes
-    # the current user turn, we cache everything up to (and including) the last LLM
-    # reply — a stable prefix. The new user turn + any tool results live AFTER the
-    # breakpoint and are not cached (they change every request anyway).
-    #
-    # Special cases:
-    #   - Compression instruction as last message: skip it, find the assistant before it.
-    #   - Fewer than 2 messages: fall back to marking the last message.
+    # Strategy: mark the last message, unless that message is a compression
+    # instruction (system_injected: true) — in that case mark the one before it.
     def apply_message_caching(messages)
       return messages if messages.empty?
 
-      # Walk backwards to find the last assistant message that is not system_injected.
-      # That is the stable "end of history" to anchor the cache breakpoint on.
-      cache_index = nil
-      messages.each_with_index.reverse_each do |msg, idx|
-        next if msg[:system_injected]
-        if msg[:role] == "assistant"
-          cache_index = idx
-          break
-        end
-      end
-
-      # Fallback: if no assistant message found, mark the last message
-      cache_index ||= messages.length - 1
+      cache_index = if is_compression_instruction?(messages.last)
+                      [messages.length - 2, 0].max
+                    else
+                      messages.length - 1
+                    end
 
       messages.map.with_index do |msg, idx|
         idx == cache_index ? add_cache_control_to_message(msg) : msg
@@ -199,18 +177,8 @@ module Clacky
       msg.merge(content: content_array)
     end
 
-    # Only true for the compression-instruction user message inserted by MessageCompressor.
-    # Skill shim messages (also system_injected: true) must NOT be treated as compression
-    # instructions — doing so shifts the cache breakpoint into the volatile skill content
-    # block, causing a full cache miss on every slash command turn.
     def is_compression_instruction?(message)
-      return false unless message.is_a?(Hash)
-      return false unless message[:system_injected] == true
-      return false unless message[:role] == "user"
-
-      content = message[:content].to_s
-      content.include?("CRITICAL: TASK CHANGE - MEMORY COMPRESSION MODE") ||
-        content.include?("MEMORY COMPRESSION MODE")
+      message.is_a?(Hash) && message[:system_injected] == true
     end
 
     # ── HTTP connections ──────────────────────────────────────────────────────

data/lib/clacky/version.rb

@@ -1,5 +1,5 @@
 # frozen_string_literal: true
 
 module Clacky
-  VERSION = "0.9.3"
+  VERSION = "0.9.4"
 end

metadata

@@ -1,7 +1,7 @@
 --- !ruby/object:Gem::Specification
 name: openclacky
 version: !ruby/object:Gem::Version
-  version: 0.9.3
+  version: 0.9.4
 platform: ruby
 authors:
 - windy