<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>KDCube Engineering Blog</title>
    <link>https://kdcube.tech/blog.html</link>
    <atom:link href="https://kdcube.tech/blog/rss.xml" rel="self" type="application/rss+xml" />
    <description>Engineering writing from KDCube: isolated execution, timeline-first agents, cache shaping, distributed execution, workspace isolation, and ReAct v3 runtime design.</description>
    <language>en-us</language>
    <lastBuildDate>Thu, 30 Apr 2026 12:00:00 +0000</lastBuildDate>

    <item>
      <title>Isolated runtime: split executor, trusted supervisor</title>
      <link>https://kdcube.tech/blog/isolated-runtime-split-executor.html</link>
      <guid isPermaLink="true">https://kdcube.tech/blog/isolated-runtime-split-executor.html</guid>
      <pubDate>Thu, 30 Apr 2026 12:00:00 +0000</pubDate>
      <description>How KDCube isolates generated code with a split Docker executor/supervisor design: no executor network, no platform secrets, narrow mounts, supervised tool calls, and complete diagnostic logs.</description>
    </item>

    <item>
      <title>KDCube ReAct v3 Agent vs Claude Code</title>
      <link>https://kdcube.tech/blog/react-v2-vs-claude-code.html</link>
      <guid isPermaLink="true">https://kdcube.tech/blog/react-v2-vs-claude-code.html</guid>
      <pubDate>Fri, 10 Apr 2026 12:00:00 +0000</pubDate>
      <description>An architecture comparison between the KDCube ReAct v3 agent and Claude Code: runtime vs client surface, context model, tools, skills, execution, streaming, permissions, and KDCube's hosted Claude Code integration.</description>
    </item>

    <item>
      <title>Memory is not one bucket: how React remembers</title>
      <link>https://kdcube.tech/blog/how-react-remembers.html</link>
      <guid isPermaLink="true">https://kdcube.tech/blog/how-react-remembers.html</guid>
      <pubDate>Mon, 06 Apr 2026 12:00:00 +0000</pubDate>
      <description>How memory works in React v3: timeline memory, internal memory anchors, SOURCES POOL and ANNOUNCE as the attention area, namespaces, plans, compaction summaries, turn logs, workspace memory, and retrieval by path or search.</description>
    </item>

    <item>
      <title>How React manages cache without losing context</title>
      <link>https://kdcube.tech/blog/how-react-manages-cache.html</link>
      <guid isPermaLink="true">https://kdcube.tech/blog/how-react-manages-cache.html</guid>
      <pubDate>Mon, 06 Apr 2026 12:00:00 +0000</pubDate>
      <description>How React v3 manages cache across turns: previous-turn, pre-tail, and tail checkpoints; react.hide for tail cleanup; TTL pruning; compaction; and why context still stays available via path-based recovery.</description>
    </item>

    <item>
      <title>Attention Area: SOURCES POOL and ANNOUNCE at the Tail</title>
      <link>https://kdcube.tech/blog/announce-is-not-decoration.html</link>
      <guid isPermaLink="true">https://kdcube.tech/blog/announce-is-not-decoration.html</guid>
      <pubDate>Mon, 06 Apr 2026 12:00:00 +0000</pubDate>
      <description>Why React v3 keeps an attention area at the tail of the timeline: SOURCES POOL plus ANNOUNCE, with ANNOUNCE acting as the signal board for budgets, alerts, workspace state, steer, and evolving conversation-wide context.</description>
    </item>

    <item>
      <title>Why not pure tool calling?</title>
      <link>https://kdcube.tech/blog/why-not-pure-tool-calling.html</link>
      <guid isPermaLink="true">https://kdcube.tech/blog/why-not-pure-tool-calling.html</guid>
      <pubDate>Mon, 06 Apr 2026 12:00:00 +0000</pubDate>
      <description>Why React v3 does not use pure provider-native tool calling as its primary orchestration protocol, and why timeline-first context plus subscribable channeled generation is a better fit for this runtime.</description>
    </item>
  </channel>
</rss>
