loopx-agent · Nov 8, 2026
@@ -6,8 +6,8 @@
 <meta property="og:type" content="article"><meta property="og:site_name" content="LoopX"><meta property="og:locale" content="en_US"><meta property="og:title" content="An Agent needs a way to look things up, too"><meta property="og:description" content="Follow one voice task through progressive reads, JSON Pointers, exact sharing, compaction-cycle deltas, and complete recovery."><meta property="og:url" content="https://loopx-project.github.io/loopx/blog/context-delivery/"><meta name="twitter:card" content="summary"><link rel="icon" href="data:,"><link rel="stylesheet" href="../blog.css">
 </head><body><a class="skip-link" href="#main">Skip to content</a>
 <header class="site-header"><a class="brand" href="../../"><span class="product-mark" aria-hidden="true"><i></i><i></i></span>LoopX</a><nav aria-label="Main navigation"><a href="../">Blog</a><a href="../../docs/book/">Docs</a></nav><div class="header-actions"><a class="language" href="../zh/context-delivery/" lang="zh-CN" hreflang="zh-CN">中文</a><a class="github" href="https://github.com/loopx-project/loopx">GitHub ↗</a></div></header>
-<main id="main"><div class="article-heading"><a class="back" href="../">← All articles</a><p class="eyebrow">Context delivery</p><h1>An Agent needs a way to look things up, too</h1><p class="lead">Follow one voice task through progressive reads, JSON Pointers, exact sharing, compaction-cycle deltas, and complete recovery.</p><div class="byline"><a href="https://github.com/loopx-project/loopx">LoopX</a><span>·</span><time datetime="2026-11-01">November 1, 2026</time><span>·</span><span>Updated November 3</span></div></div>
-<div class="article-layout"><aside class="toc"><details><summary>On this page</summary><nav aria-label="On this page"><a href="#example">Follow one task: give an assistant voice input</a><a href="#known-paths">When you know the locations, read them together</a><a href="#discover">When the structure is unfamiliar, index the relevant subtree</a><a href="#requirements">See requirements first, then open their evidence</a><a href="#sharing">Share exact records within one delivery</a><a href="#cycle">On the next continuation, deliver what changed</a><a href="#native-path">When enabled, the default output is a reading entrypoint</a><a href="#recovery">Recover missing context while preserving the original work</a><a href="#evidence">What the measurements establish</a><a href="#ablation-evidence">Multi-task and layered ablation</a><a href="#reproduce">Read the same material yourself</a></nav></details></aside><article class="prose"><p>Long-running work accumulates material: the original goal, later constraints, attempted routes, failed experiments and current progress. Sending everything again on each continuation repeats more and more content; keeping only a summary can make the original basis of a decision difficult to recover.</p>
+<main id="main"><div class="article-heading"><a class="back" href="../">← All articles</a><p class="eyebrow">Context delivery</p><h1>An Agent needs a way to look things up, too</h1><p class="lead">Follow one voice task through progressive reads, JSON Pointers, exact sharing, compaction-cycle deltas, and complete recovery.</p><div class="byline"><a href="https://github.com/loopx-project/loopx">LoopX</a><span>·</span><time datetime="2026-11-01">November 1, 2026</time><span>·</span><span>Updated November 9</span></div></div>
+<div class="article-layout"><aside class="toc"><details><summary>On this page</summary><nav aria-label="On this page"><a href="#example">Follow one task: give an assistant voice input</a><a href="#known-paths">When you know the locations, read them together</a><a href="#discover">When the structure is unfamiliar, index the relevant subtree</a><a href="#requirements">See requirements first, then open their evidence</a><a href="#sharing">Share exact records within one delivery</a><a href="#cycle">On the next continuation, deliver what changed</a><a href="#native-path">When enabled, the default output is a reading entrypoint</a><a href="#recovery">Recover missing context while preserving the original work</a><a href="#evidence">What the measurements establish</a><a href="#ablation-evidence">Multi-task and layered ablation</a><a href="#multitask-context-byte-ratios">8-task Four-Arm byte ratios</a><a href="#reproduce">Read the same material yourself</a></nav></details></aside><article class="prose"><p>Long-running work accumulates material: the original goal, later constraints, attempted routes, failed experiments and current progress. Sending everything again on each continuation repeats more and more content; keeping only a summary can make the original basis of a decision difficult to recover.</p>
 <p>LoopX's context delivery separates these jobs: <strong>preserve complete sources, read the facts needed for the current decision, expand supporting material around questions, and reuse unchanged structures exactly within a valid compaction cycle.</strong></p>
 <p>This article describes the native Codex path in released <a href="https://github.com/loopx-project/loopx/tree/v2.2.0">LoopX 2.2.0</a>. Supported POSIX Codex App/SSH sessions and new benchmark heartbeat trials default to context delivery; saved off remains effective and explicit off is available. When enabled, complete content is saved first and the Agent receives a reading entrypoint for current facts. See the <a href="https://github.com/loopx-project/loopx/releases/tag/v2.2.0">2.2.0 release notes</a> for publication and validation scope. The synthetic measurements below retain their original source and experimental conditions.</p>
 <p>Think of a capture as a complete book. An index locates its chapters; a JSON Pointer addresses a place inside it. Progressive reading determines how deeply to open the material now. Differential delivery determines what can be reused on the next continuation.</p>
@@ -266,6 +266,96 @@ <h3 id="ablation-evidence">Multi-task and layered ablation: benefit chain and bo
 </tbody>
 </table></div>
 <p>This dual-track audit shows why <strong>control-packet byte reduction (−74.25% / −38.54%) must not be equated with whole-episode model token savings (−12.07% uncached, −7.46% total), nor used to hide early-window indexing overhead (4.05× in the first 2 hours) or cross-task counterexamples.</strong> Context Delivery’s practical value lies in slowing context bloat across 10–18 hour runs so that, together with Stride, preserved attention converts into effective late-stage experiments (see <a href="../edgebench-feedback-and-memory/#later-common-windows">EdgeBench late-stage breakthroughs and five-arm attribution</a>).</p>
+<h3 id="multitask-context-byte-ratios">Multi-task control-plane byte ratios across 8 latest Four-Arm Context ON 18h trajectories (Progressive vs. Delta)</h3>
+<p>To separate the independent byte reductions of <strong>progressive projection (<code>required-context.json</code>)</strong> and <strong>in-cycle differential delivery (<code>delivery.json</code>)</strong> across real long-horizon runs, we audited all <strong>8 <code>Context ON</code> 18h trajectories from the latest Four-Arm core suite (<code>N = 213</code> model-facing wakes: <code>101</code> new-session/post-compaction baseline reads and <code>112</code> in-cycle <code>Delta</code> continuations)</strong>:</p>
+<figure class="diagram"><p class="diagram-label">8-task Four-Arm Context ON byte ratios (N = 213 wakes, 112 in-cycle Deltas)</p><a href="../images/context-delivery/multitask-byte-ratios-en.svg" target="_blank" rel="noopener" aria-label="Open the 8-task Progressive vs. Delta control-plane byte ratio chart at full size"><img src="../images/context-delivery/multitask-byte-ratios-en.svg" alt="8-Task Four-Arm Context ON 18h Control-Plane Byte Share: Progressive 33.4% (-66.6%), In-Cycle Delta 35.8% (-64.2%), Prog+Delta First-Screen 3.0% / 5.0%" width="1220" height="520" loading="lazy" style="width:100%;height:auto;border-radius:10px;margin:8px 0 12px 0"></a><figcaption>Left (A): Per-task share of Progressive projection bytes (blue), in-cycle Delta payload bytes (teal) and Prog+Delta first-screen body bytes (purple) relative to same-wake Full packets. Right (B): Cumulative control-plane byte reduction across all 213 wakes from 22.41 MB to 7.48 MB (Progressive), 4.49 MB (in-cycle Delta), and 0.37 MB / 0.63 MB (first-screen body / CLI response). Open image for full size.</figcaption></figure>
+<div class="table-scroll" role="region" tabindex="0" aria-label="8-task Four-Arm Context ON control-plane byte ratio table"><table>
+<thead>
+<tr>
+<th>Task (Four-Arm Context ON 18h)</th>
+<th style="text-align:right">Wakes (base / Δ)</th>
+<th style="text-align:right">Full packets</th>
+<th style="text-align:right">Progressive projection (share)</th>
+<th style="text-align:right">In-cycle Delta payload (vs same-wake Full)</th>
+<th style="text-align:right">Prog+Delta first-screen (body / CLI)</th>
+</tr>
+</thead>
+<tbody>
+<tr>
+<td>Grid Robot (<code>on_stride</code>)</td>
+<td style="text-align:right">29 (15 / 14)</td>
+<td style="text-align:right">4,405,651 B</td>
+<td style="text-align:right">1,166,467 B (<strong>26.5%</strong>, −73.5%)</td>
+<td style="text-align:right">474,016 B (<strong>20.7%</strong>, −79.3%)</td>
+<td style="text-align:right">41,878 / 76,650 B (<strong>1.8% / 3.3%</strong>)</td>
+</tr>
+<tr>
+<td>VLIW Scheduling (<code>on_stride</code>)</td>
+<td style="text-align:right">14 (5 / 9)</td>
+<td style="text-align:right">1,297,500 B</td>
+<td style="text-align:right">505,826 B (<strong>39.0%</strong>, −61.0%)</td>
+<td style="text-align:right">382,664 B (<strong>45.2%</strong>, −54.8%)</td>
+<td style="text-align:right">30,860 / 50,152 B (<strong>3.6% / 5.9%</strong>)</td>
+</tr>
+<tr>
+<td>FFmpeg Encoding (<code>on_stride</code>)</td>
+<td style="text-align:right">26 (12 / 14)</td>
+<td style="text-align:right">2,924,522 B</td>
+<td style="text-align:right">874,104 B (<strong>29.9%</strong>, −70.1%)</td>
+<td style="text-align:right">759,787 B (<strong>47.1%</strong>, −52.9%)</td>
+<td style="text-align:right">43,807 / 75,305 B (<strong>2.7% / 4.7%</strong>)</td>
+</tr>
+<tr>
+<td>OpenTTD Transport (<code>on_stride</code>)</td>
+<td style="text-align:right">31 (19 / 12)</td>
+<td style="text-align:right">3,288,042 B</td>
+<td style="text-align:right">1,050,623 B (<strong>32.0%</strong>, −68.0%)</td>
+<td style="text-align:right">436,053 B (<strong>30.8%</strong>, −69.2%)</td>
+<td style="text-align:right">36,753 / 64,619 B (<strong>2.6% / 4.6%</strong>)</td>
+</tr>
+<tr>
+<td>Ad Placement (<code>on_stride</code>)</td>
+<td style="text-align:right">29 (12 / 17)</td>
+<td style="text-align:right">3,115,820 B</td>
+<td style="text-align:right">1,019,704 B (<strong>32.7%</strong>, −67.3%)</td>
+<td style="text-align:right">708,828 B (<strong>39.4%</strong>, −60.6%)</td>
+<td style="text-align:right">64,174 / 102,437 B (<strong>3.6% / 5.7%</strong>)</td>
+</tr>
+<tr>
+<td>Borden Inversion (<code>on_stride</code>)</td>
+<td style="text-align:right">28 (11 / 17)</td>
+<td style="text-align:right">2,855,664 B</td>
+<td style="text-align:right">1,001,254 B (<strong>35.1%</strong>, −64.9%)</td>
+<td style="text-align:right">625,709 B (<strong>34.4%</strong>, −65.6%)</td>
+<td style="text-align:right">50,602 / 90,024 B (<strong>2.8% / 4.9%</strong>)</td>
+</tr>
+<tr>
+<td>Rust Workspace (<code>on_stride</code>)</td>
+<td style="text-align:right">15 (4 / 11)</td>
+<td style="text-align:right">1,173,832 B</td>
+<td style="text-align:right">439,814 B (<strong>37.5%</strong>, −62.5%)</td>
+<td style="text-align:right">363,310 B (<strong>42.1%</strong>, −57.9%)</td>
+<td style="text-align:right">40,956 / 64,251 B (<strong>4.7% / 7.4%</strong>)</td>
+</tr>
+<tr>
+<td>ANN Vector Search (<code>on_stride</code>)</td>
+<td style="text-align:right">41 (23 / 18)</td>
+<td style="text-align:right">3,347,659 B</td>
+<td style="text-align:right">1,228,159 B (<strong>36.7%</strong>, −63.3%)</td>
+<td style="text-align:right">737,968 B (<strong>38.6%</strong>, −61.4%)</td>
+<td style="text-align:right">65,419 / 107,112 B (<strong>3.4% / 5.6%</strong>)</td>
+</tr>
+<tr>
+<td><strong>8-Run Weighted Total</strong></td>
+<td style="text-align:right"><strong>213 (101 / 112)</strong></td>
+<td style="text-align:right"><strong>22,408,690 B (100%)</strong></td>
+<td style="text-align:right"><strong>7,479,469 B (33.4%, −66.6%)</strong></td>
+<td style="text-align:right"><strong>4,488,335 B (35.8%, −64.2%)</strong></td>
+<td style="text-align:right"><strong>374,449 / 630,550 B (3.0% / 5.0%)</strong></td>
+</tr>
+</tbody>
+</table></div>
+<p>Two orthogonal byte-reduction stages appear consistently across all 8 tasks: (1) <strong>Progressive projection (<code>required-context.json</code>)</strong> reduces full decision packets across all 213 wakes from <code>22.41 MB</code> to <code>7.48 MB</code> (<strong><code>33.4%</code> mean share, <code>−66.6%</code></strong>, per-task range <code>26.5%–39.0%</code>) by folding unselected candidates and historical long-tail detail behind on-demand JSON Pointers; (2) <strong>in-cycle Delta delivery (<code>delivery.json</code>)</strong> reduces the 112 in-cycle wakes from <code>12.53 MB</code> to <code>4.49 MB</code> (<strong><code>35.8%</code> mean share, <code>−64.2%</code></strong>, per-task range <code>20.7%–47.1%</code>); and (3) <strong>their combined <code>--initial-read</code> first-screen entrypoint</strong> carries only <strong><code>3.0%</code> of raw bytes (<code>5.0%</code> including the CLI JSON reader envelope, i.e., <code>−97.0% / −95.0%</code>)</strong>. In addition, the recently landed <strong>C1 unified controller/executor projection (<a href="https://github.com/loopx-project/loopx/pull/8266">#8266</a> / <a href="https://github.com/loopx-project/loopx/pull/8304">#8304</a>)</strong> further reduces steady-state per-turn control-plane material by <strong><code>26.7%</code></strong> (<code>69.9 KB → 51.3 KB</code>) and cuts post-writeback reprojection latency from <code>57.1ms</code> to <code>3.3ms</code>.</p>
 <p>2.2.0 ships complete sources, file entrypoints, batch selection, index depth, applicable acceptance layering, exact sharing, cycle deltas and recovery. Broader decision-stage candidate, preference and Explore delivery and consistency across all entrypoints remain unfinished. Disable in Settings → This machine → Capabilities → Differential context → Codex; saved off survives upgrade. A command or new benchmark trial can use <code>--context-delivery off</code>. Trials do not inherit the machine switch, and existing frozen trials are not rewritten. Under the default policy, ordinary CLI, unsupported hosts and non-heartbeat benchmarks retain their full paths; an explicit unsupported on request is rejected. Unavailable hooks use full delivery. Saving configuration, actual installation and model adoption remain separate checks.</p>
 <h2 id="reproduce">Read the same material yourself</h2>
 <p><a href="example.py">example.py</a> generates this complete synthetic voice task and invokes the existing reader, progressive projection and delta owners. It checks selected values, directory addresses, all ten requirements, shared references and complete recovery. It registers no Goal, admits or settles no Turn, and calls no model or speech service.</p>