From 9997f5c0454eaf68905a861fad15e669a57e815e Mon Sep 17 00:00:00 2001 From: huangruiteng Date: Tue, 10 Nov 2026 15:09:56 +0800 Subject: [PATCH] docs(blog): lead context delivery with progressive and delta evidence Signed-off-by: huangruiteng --- .../public/blog/context-delivery/index.html | 322 +++++++++--------- .../blog/zh/context-delivery/index.html | 320 ++++++++--------- 2 files changed, 329 insertions(+), 313 deletions(-) diff --git a/apps/presentation/site/public/blog/context-delivery/index.html b/apps/presentation/site/public/blog/context-delivery/index.html index 9c177029b6..79650dd58d 100644 --- a/apps/presentation/site/public/blog/context-delivery/index.html +++ b/apps/presentation/site/public/blog/context-delivery/index.html @@ -6,12 +6,173 @@
LoopX
中文GitHub ↗
-
← All articles

Context delivery

An Agent needs a way to look things up, too

One voice task, from choosing facts to fitting the first read, preserving exact values, and recovering without replaying work.

LoopX··Updated November 8
-

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.

+
← All articles

Context delivery

An Agent needs a way to look things up, too

One voice task, from choosing facts to fitting the first read, preserving exact values, and recovering without replaying work.

LoopX··Updated November 10
+

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.

LoopX's context delivery separates these jobs: 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.

+

Results first: progressive reading, deltas and task ablation

+

Eight tasks: progressive reading and in-cycle deltas

+

This section recalculates the published eight-task Four-Arm Context ON 18h aggregates to distinguish Progressive and in-cycle Delta transport costs. The source reports 213 model-facing wakes: 101 new-session or post-compaction baseline reads and 112 in-cycle continuations. Per-task cells retain the published values; totals are recomputed from those cells. Raw per-wake logs are not included here.

+
    +
  • Progressive reading: across all 213 wakes, full packets fall from 22.41 MB to 7.29 MB (−67.5%). The projection retains 32.5% of original bytes, with a per-task range of 26.5%–39.0%; unselected candidates and historical detail remain available on demand.
    • +
  • In-cycle deltas: across 112 continuations, corresponding full packets fall from approximately 12.53 MB to 4.49 MB (−64.2%). Deltas retain 35.8% of same-wake full bytes, with a per-task range of 20.7%–47.1%.
    • +
  • Progressive + delta first read: for those same 112 continuations, the --initial-read body carries 3.0% of original bytes, or 5.0% including the CLI JSON envelope (−97.0% / −95.0%). First-read size is not whole-episode model input.
    • +
+

8-task Four-Arm Context ON byte ratios (N = 213 wakes, 112 in-cycle Deltas)

8-Task Four-Arm Context ON 18h Control-Plane Byte Share: Progressive 32.5% (-67.5%), In-Cycle Delta 35.8% (-64.2%), Prog+Delta First-Screen 3.0% / 5.0%
A: Per-task shares, with Progressive over all wakes and Delta/first-read over in-cycle wakes only. B: Separate cohort totals: Progressive 7.29 / 22.41 MB over 213 wakes; Delta 4.49 / approximately 12.53 MB over 112 wakes. First-read body / CLI totals are 0.374 / 0.631 MB for the same 112 wakes. The two denominators are not interchangeable. Open image for full size.
+
Expand the eight-task data table +
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
Task (Four-Arm Context ON 18h)Wakes (base / Δ)Full packetsProgressive projection (share)In-cycle Delta payload (vs same-wake Full)Prog+Delta first-screen (body / CLI)
Grid Robot (on_stride)29 (15 / 14)4,405,651 B1,166,467 B (26.5%, −73.5%)474,016 B (20.7%, −79.3%)41,878 / 76,650 B (1.8% / 3.3%)
VLIW Scheduling (on_stride)14 (5 / 9)1,297,500 B505,826 B (39.0%, −61.0%)382,664 B (45.2%, −54.8%)30,860 / 50,152 B (3.6% / 5.9%)
FFmpeg Encoding (on_stride)26 (12 / 14)2,924,522 B874,104 B (29.9%, −70.1%)759,787 B (47.1%, −52.9%)43,807 / 75,305 B (2.7% / 4.7%)
OpenTTD Transport (on_stride)31 (19 / 12)3,288,042 B1,050,623 B (32.0%, −68.0%)436,053 B (30.8%, −69.2%)36,753 / 64,619 B (2.6% / 4.6%)
Ad Placement (on_stride)29 (12 / 17)3,115,820 B1,019,704 B (32.7%, −67.3%)708,828 B (39.4%, −60.6%)64,174 / 102,437 B (3.6% / 5.7%)
Borden Inversion (on_stride)28 (11 / 17)2,855,664 B1,001,254 B (35.1%, −64.9%)625,709 B (34.4%, −65.6%)50,602 / 90,024 B (2.8% / 4.9%)
Rust Workspace (on_stride)15 (4 / 11)1,173,832 B439,814 B (37.5%, −62.5%)363,310 B (42.1%, −57.9%)40,956 / 64,251 B (4.7% / 7.4%)
ANN Vector Search (on_stride)41 (23 / 18)3,347,659 B1,228,159 B (36.7%, −63.3%)737,968 B (38.6%, −61.4%)65,419 / 107,112 B (3.4% / 5.6%)
8-Run Weighted Total213 (101 / 112)22,408,690 B (100%)7,285,951 B (32.5%, −67.5%)4,488,335 B (35.8%, −64.2%)374,449 / 630,550 B (3.0% / 5.0%)
+
+ +

Basis and source: the published eight task rows sum to 7,285,951 B / 22,408,690 B = 32.5% for Progressive across all 213 wakes. Delta and first-screen shares of approximately 35.8% and 3.0% / 5.0% cover only the 112 in-cycle wakes. Their Full denominator was published rounded as 12.53 MB; it is not the 213-wake total. This is an aggregate recalculation, not a new experiment or an independent audit of per-wake logs. Public aggregate data and provenance.

Implementation status: #8266 remains an open C1 proposal. The final merged scope of #8304 preserves precision while reusing ordering and a verified digest. These cohort bytes do not establish a shipped unified Controller/Worker gain.

+

Multi-task and layered ablation: benefit chain and boundary counterexamples

+

In EdgeBench ANN vector search optimization (gpt-6.1-sol / xhigh, 18h, subject to Recall@10 ≥ 0.95) and cross-task ablations, we separate Context Delivery (progressive selection and cycle deltas) from Stride (evidence-conditioned experiment sizing), auditing selected control-packet bytes, main-session model input tokens, and local/early-window counterexamples side by side (reconciled from archived trajectories and receipts by @shangzh0):

+

Multi-task ablation and layered accounting

Progressive Context Delivery and Stride Ablation: ANN 18h Peak QPS Progression and Layered Token/Byte Savings vs Boundary Counterexamples
Top (A): ANN 18h six-arm peak valid throughput QPS (Recall@10 ≥ 0.95). Bottom (B): Layered control-packet byte and main-session input token savings vs Context off (left, negative) alongside disclosed local/early-window overhead and cross-task counterexamples (right, positive). Open image for full size.
+

This dual-track audit shows why 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. The reported ANN runs contain both lower input totals and later useful experiments. That observation motivates testing whether context delivery helps sustain long research, but does not by itself attribute the outcome to saved attention, separate every interaction with Stride, or establish a gain on another task. The November source changes described later were not treatments in these archived runs. See EdgeBench late-stage breakthroughs and five-arm attribution for the original research context.

+
Expand the ablation results and counterexamples +
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
Dimension or experimental armObserved figureScope and interpretation
ANN 18h: Official / Single / Native-Goal5,753 / 5,706 / 4,372 QPSOfficial peaked at 1.05h and remained on a plateau afterward
ANN 18h: Explore · Context off5,766 QPS (+0.2%)Peaked at 11.37h; input accounting is reported separately below
ANN 18h: Explore · Context on6,020 QPS (+4.6%)Peaked at 14.43h; this run's later peak does not isolate the cause
ANN 18h: Explore · Context on + Stride7,159 QPS (+24.4%)Breakthrough at 9.19h (QT_8bit_direct HNSW-SQ8, Recall@10 = 0.9585)
Control-packet bytes: 24 delta events / 46 paired events−74.25% / −38.54%Paired selected bytes dropped from 3,089,841 B to 1,898,889 B (−25.75% across 34 Stride-arm events)
Main-session input tokens: full-18h uncached / total−12.07% / −7.46%Uncached input 8.57M → 7.53M (−7.91% on the 16h common window); total input 272.03M → 251.74M
Counterexample 1: project+read process stdout+57% to +83% (1.57×–1.83×)CLI JSON framing and pagination metadata make process stdout larger than raw selected file bytes
Counterexample 2: First-2h early-window uncached input+304.7% (0.25M → 1.00M)Baseline establishment and staged discovery add upfront work; later amortization is not guaranteed by this early window
Counterexample 3: Historical Grid (lower is better)Official 1967 < Off 2054 < On+Stride 2088 < On 2242Cross-task counterexample under differing durations and resources; retained without causal overclaim
+

The core native Codex path ships in LoopX 2.2.0. 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 2.2.0 release notes for publication and validation scope.

Think of a capture as a complete saved snapshot, like a book whose edition is fixed. 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.

-

This November 8 revision also follows the first-read, staged-scope, transport and numeric-fidelity changes merged after 2.2.0, checked at source 611019243e. Those sections describe that source snapshot, not an upgrade already installed everywhere. The original teaching measurements and archived research results below retain their earlier versions and conditions.

+

This November 8 revision also follows the first-read, staged-scope, transport and numeric-fidelity changes merged after 2.2.0, checked at source 611019243e. Those sections describe that source snapshot, not an upgrade already installed everywhere. The teaching measurements and opening archived research results retain their earlier versions and conditions.

Follow one task: give an assistant voice input

Suppose an assistant needs to understand a voice note and reply in its original chat. Its complete material contains the goal, provider constraints, privacy requirements, acceptance and a long discussion of alternatives. This is a teaching example; the example program below generates the complete data in an isolated directory.

capture.json
@@ -239,161 +400,8 @@ 
A shorter view must preserve the original values

 
What the measurements establish

 
The example separates three mechanisms: avoid repeatedly moving history not currently needed, combine reads that belong together, and avoid retransmitting unchanged verified structures. Complete sources, requirements and recovery connect them.

 
Reader figures include actual CLI page metadata; local path length affects a small amount of framing. Projection figures compare capture bodies; the cycle table compares codec output. The reproduction script checks these distinct measurements and exact reconstruction.

-
Real long-running tasks must also count necessary later reads, recovery, cache hits, model input/output, tool calls, waiting and human attention. Effectiveness depends on constraints understood, decisions made and accepted outcomes. The synthetic voice example above measures the mechanical path; below we examine multi-task and layered ablation observations on 18-hour trajectories.

-
Multi-task and layered ablation: benefit chain and boundary counterexamples

-
In EdgeBench ANN vector search optimization (gpt-6.1-sol / xhigh, 18h, subject to Recall@10 ≥ 0.95) and cross-task ablations, we separate Context Delivery (progressive selection and cycle deltas) from Stride (evidence-conditioned experiment sizing), auditing selected control-packet bytes, main-session model input tokens, and local/early-window counterexamples side by side (reconciled from archived trajectories and receipts by @shangzh0):

-
Multi-task ablation and layered accounting
Progressive Context Delivery and Stride Ablation: ANN 18h Peak QPS Progression and Layered Token/Byte Savings vs Boundary Counterexamples
Top (A): ANN 18h six-arm peak valid throughput QPS (Recall@10 ≥ 0.95). Bottom (B): Layered control-packet byte and main-session input token savings vs Context off (left, negative) alongside disclosed local/early-window overhead and cross-task counterexamples (right, positive). Open image for full size.

-

-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
Dimension or experimental armObserved figureScope and interpretation
ANN 18h: Official / Single / Native-Goal5,753 / 5,706 / 4,372 QPSOfficial peaked at 1.05h and remained on a plateau afterward
ANN 18h: Explore · Context off5,766 QPS (+0.2%)Peaked at 11.37h; input accounting is reported separately below
ANN 18h: Explore · Context on6,020 QPS (+4.6%)Peaked at 14.43h; this run's later peak does not isolate the cause
ANN 18h: Explore · Context on + Stride7,159 QPS (+24.4%)Breakthrough at 9.19h (QT_8bit_direct HNSW-SQ8, Recall@10 = 0.9585)
Control-packet bytes: 24 delta events / 46 paired events−74.25% / −38.54%Paired selected bytes dropped from 3,089,841 B to 1,898,889 B (−25.75% across 34 Stride-arm events)
Main-session input tokens: full-18h uncached / total−12.07% / −7.46%Uncached input 8.57M → 7.53M (−7.91% on the 16h common window); total input 272.03M → 251.74M
Counterexample 1: project+read process stdout+57% to +83% (1.57×–1.83×)CLI JSON framing and pagination metadata make process stdout larger than raw selected file bytes
Counterexample 2: First-2h early-window uncached input+304.7% (0.25M → 1.00M)Baseline establishment and staged discovery add upfront work; later amortization is not guaranteed by this early window
Counterexample 3: Historical Grid (lower is better)Official 1967 < Off 2054 < On+Stride 2088 < On 2242Cross-task counterexample under differing durations and resources; retained without causal overclaim

-
This dual-track audit shows why 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. The reported ANN runs contain both lower input totals and later useful experiments. That observation motivates testing whether context delivery helps sustain long research, but does not by itself attribute the outcome to saved attention, separate every interaction with Stride, or establish a gain on another task. The November source changes above were not treatments in these archived runs. See EdgeBench late-stage breakthroughs and five-arm attribution for the original research context.

-
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 --context-delivery off. 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.

-
Multi-task control-plane byte ratios across 8 latest Four-Arm Context ON 18h trajectories (Progressive vs. Delta)

-
This section recalculates the published eight-task Four-Arm Context ON 18h aggregates to distinguish Progressive and in-cycle Delta transport costs. The source reports 213 model-facing wakes: 101 new-session or post-compaction baseline reads and 112 in-cycle continuations. Per-task cells retain the published values; totals are recomputed from those cells. Raw per-wake logs are not included here.

-
8-task Four-Arm Context ON byte ratios (N = 213 wakes, 112 in-cycle Deltas)
8-Task Four-Arm Context ON 18h Control-Plane Byte Share: Progressive 32.5% (-67.5%), In-Cycle Delta 35.8% (-64.2%), Prog+Delta First-Screen 3.0% / 5.0%
A: Per-task shares, with Progressive over all wakes and Delta/first-read over in-cycle wakes only. B: Separate cohort totals: Progressive 7.29 / 22.41 MB over 213 wakes; Delta 4.49 / approximately 12.53 MB over 112 wakes. First-read body / CLI totals are 0.374 / 0.631 MB for the same 112 wakes. The two denominators are not interchangeable. Open image for full size.

-

-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
Task (Four-Arm Context ON 18h)Wakes (base / Δ)Full packetsProgressive projection (share)In-cycle Delta payload (vs same-wake Full)Prog+Delta first-screen (body / CLI)
Grid Robot (on_stride)29 (15 / 14)4,405,651 B1,166,467 B (26.5%, −73.5%)474,016 B (20.7%, −79.3%)41,878 / 76,650 B (1.8% / 3.3%)
VLIW Scheduling (on_stride)14 (5 / 9)1,297,500 B505,826 B (39.0%, −61.0%)382,664 B (45.2%, −54.8%)30,860 / 50,152 B (3.6% / 5.9%)
FFmpeg Encoding (on_stride)26 (12 / 14)2,924,522 B874,104 B (29.9%, −70.1%)759,787 B (47.1%, −52.9%)43,807 / 75,305 B (2.7% / 4.7%)
OpenTTD Transport (on_stride)31 (19 / 12)3,288,042 B1,050,623 B (32.0%, −68.0%)436,053 B (30.8%, −69.2%)36,753 / 64,619 B (2.6% / 4.6%)
Ad Placement (on_stride)29 (12 / 17)3,115,820 B1,019,704 B (32.7%, −67.3%)708,828 B (39.4%, −60.6%)64,174 / 102,437 B (3.6% / 5.7%)
Borden Inversion (on_stride)28 (11 / 17)2,855,664 B1,001,254 B (35.1%, −64.9%)625,709 B (34.4%, −65.6%)50,602 / 90,024 B (2.8% / 4.9%)
Rust Workspace (on_stride)15 (4 / 11)1,173,832 B439,814 B (37.5%, −62.5%)363,310 B (42.1%, −57.9%)40,956 / 64,251 B (4.7% / 7.4%)
ANN Vector Search (on_stride)41 (23 / 18)3,347,659 B1,228,159 B (36.7%, −63.3%)737,968 B (38.6%, −61.4%)65,419 / 107,112 B (3.4% / 5.6%)
8-Run Weighted Total213 (101 / 112)22,408,690 B (100%)7,285,951 B (32.5%, −67.5%)4,488,335 B (35.8%, −64.2%)374,449 / 630,550 B (3.0% / 5.0%)

-
Two orthogonal byte-reduction stages appear consistently across all 8 tasks: (1) Progressive projection (required-context.json) reduces full decision packets across all 213 wakes from 22.41 MB to 7.29 MB (32.5% weighted byte share, −67.5%, per-task range 26.5%–39.0%) by folding unselected candidates and historical long-tail detail behind on-demand JSON Pointers; (2) in-cycle Delta delivery (delivery.json) reduces the 112 in-cycle wakes from 12.53 MB to 4.49 MB (35.8% weighted byte share, −64.2%, per-task range 20.7%–47.1%); and (3) their combined --initial-read first-screen entrypoint carries only 3.0% of raw bytes (5.0% including the CLI JSON reader envelope, i.e., −97.0% / −95.0%).

+
Real long-running tasks must also count necessary later reads, recovery, cache hits, model input/output, tool calls, waiting and human attention. Effectiveness depends on constraints understood, decisions made and accepted outcomes. The synthetic voice example above measures the mechanical path; the opening eight-task comparison and 18-hour ablations separately examine transport cost, model input and task outcomes in long-running research.

 
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 --context-delivery off. 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.

-
-
Basis and source: the published eight task rows sum to 7,285,951 B / 22,408,690 B = 32.5% for Progressive across all 213 wakes. Delta and first-screen shares of approximately 35.8% and 3.0% / 5.0% cover only the 112 in-cycle wakes. Their Full denominator was published rounded as 12.53 MB; it is not the 213-wake total. This is an aggregate recalculation, not a new experiment or an independent audit of per-wake logs. Public aggregate data and provenance.
Implementation status: #8266 remains an open C1 proposal. The final merged scope of #8304 preserves precision while reusing ordering and a verified digest. These cohort bytes do not establish a shipped unified Controller/Worker gain.

 
Read the same material yourself

 
example.py 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.

 
Prepare the original measurement checkout, a4a0fad889d91cba0d945ce658788bdb76ba351e, and its compatible dependencies. Run from that checkout’s root, substituting the downloaded program path. The output directory must be empty.

diff --git a/apps/presentation/site/public/blog/zh/context-delivery/index.html b/apps/presentation/site/public/blog/zh/context-delivery/index.html
index c006bc0ef7..85af793ab2 100644
--- a/apps/presentation/site/public/blog/zh/context-delivery/index.html
+++ b/apps/presentation/site/public/blog/zh/context-delivery/index.html
@@ -7,11 +7,172 @@
 
 
LoopX
ENGitHub ↗

 
← 全部文章
Context delivery
Agent 读上下文，也需要一套“查书”的方法
沿一个语音任务，讲清选读、完整首读、按阶段补读、精确复用与恢复。
LoopX··11 月 10 日更新

-
Agent 做长任务，资料会越来越多：原目标、后来追加的约束、试过的方案、失败的实验、当前做到哪里。每次继续都把整份材料再塞一遍，重复会不断累积；只留一段摘要，又可能找不回决定所依赖的原话。

+
Agent 做长任务，资料会越来越多：原目标、后来追加的约束、试过的方案、失败的实验、当前做到哪里。每次继续都把整份材料再塞一遍，重复会不断累积；只留一段摘要，又可能找不回决定所依赖的原话。

 
LoopX 的 context delivery 把这件事拆开做：完整资料保存下来，当前决定需要的事实先读，支持材料沿问题展开；同一个有效 compaction 周期内，再精确复用没有变化的结构。

+
先看数据：渐进读取、差分交付与任务消融

+
八题对照：渐进读取与周期差分

+
本节复算已公开的八题四臂 Context ON 18h 聚合表，以区分 Progressive 与周期内 Delta 的传输开销。原表报告 213 次模型侧唤醒：101 次新会话或压缩后的基线读取，以及 112 次周期内续跑。逐题单元格保留来源原值，合计按表重算；这里不附原始逐 wake 日志。

+
    
+
  • 渐进读取：全部 213 次唤醒，完整包 22.41 MB → 7.29 MB（−67.5%）。投影保留原包的 32.5%，各题为 26.5%–39.0%；未选中候选与历史长尾详情按需展开。
    • 
+
  • 周期内差分：112 次同周期续跑，对应完整包 约 12.53 MB → 4.49 MB（−64.2%）。差分保留同组原包的 35.8%，各题为 20.7%–47.1%。
    • 
+
  • 渐进＋差分首读：在同一组 112 次续跑中，--initial-read 正文占原包 3.0%，含 CLI JSON 封装为 5.0%（净减 −97.0% / −95.0%）。首读入口大小不代表整轮模型输入。
    • 
+

+
八题四臂 Context ON 控制面字节比例（N = 213 唤醒，112 周期内 Delta）
八题四臂 Context ON 18h 轨迹控制面字节比例：Progressive 占 32.5% (-67.5%)，同周期 Delta 占 35.8% (-64.2%)，Prog+Delta 首屏占 3.0% / 5.0%
A：逐题占比，Progressive 覆盖全部唤醒，Delta／首读仅覆盖周期内唤醒。B：分开汇总：213 次唤醒的 Progressive 为 7.29 / 22.41 MB；112 次周期内唤醒的 Delta 为 4.49 / 约 12.53 MB。同一 112 次的首读正文／CLI 为 0.374 / 0.631 MB。两个分母不可混用。可打开原图查看。

+
展开八题逐项数据
+

+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
题目（四臂 Context ON 18h）唤醒数（基线 / Δ）Full 全量包Progressive 必要视图（占比）同周期 Delta 载荷（占同轮 Full）Prog+Delta 首屏（正文 / 含封装）
Grid 路径优化（on_stride）29（15 / 14）4,405,651 B1,166,467 B（26.5%, −73.5%）474,016 B（20.7%, −79.3%）41,878 / 76,650 B（1.8% / 3.3%）
VLIW 调度优化（on_stride）14（5 / 9）1,297,500 B505,826 B（39.0%, −61.0%）382,664 B（45.2%, −54.8%）30,860 / 50,152 B（3.6% / 5.9%）
FFmpeg 编码加速（on_stride）26（12 / 14）2,924,522 B874,104 B（29.9%, −70.1%）759,787 B（47.1%, −52.9%）43,807 / 75,305 B（2.7% / 4.7%）
OpenTTD 路网规划（on_stride）31（19 / 12）3,288,042 B1,050,623 B（32.0%, −68.0%）436,053 B（30.8%, −69.2%）36,753 / 64,619 B（2.6% / 4.6%）
Ad Placement 广告拍卖（on_stride）29（12 / 17）3,115,820 B1,019,704 B（32.7%, −67.3%）708,828 B（39.4%, −60.6%）64,174 / 102,437 B（3.6% / 5.7%）
Borden 震源反演（on_stride）28（11 / 17）2,855,664 B1,001,254 B（35.1%, −64.9%）625,709 B（34.4%, −65.6%）50,602 / 90,024 B（2.8% / 4.9%）
Rust 工作区重构（on_stride）15（4 / 11）1,173,832 B439,814 B（37.5%, −62.5%）363,310 B（42.1%, −57.9%）40,956 / 64,251 B（4.7% / 7.4%）
ANN 向量检索（on_stride）41（23 / 18）3,347,659 B1,228,159 B（36.7%, −63.3%）737,968 B（38.6%, −61.4%）65,419 / 107,112 B（3.4% / 5.6%）
8 题合计（加权总体）213（101 / 112）22,408,690 B（100%）7,285,951 B（32.5%, −67.5%）4,488,335 B（35.8%, −64.2%）374,449 / 630,550 B（3.0% / 5.0%）

+

+
+
口径与来源：上表逐题数据复算自已公开的八题表，Progressive 合计为 7,285,951 B / 22,408,690 B = 32.5%，覆盖全部 213 次唤醒。Delta 与首屏的 35.8%、3.0% / 5.0% 仅对应 112 次周期内唤醒；其 Full 分母原文仅报告约 12.53 MB，不能与 213 次总量混用。这里没有新增实验，也未独立核对逐 wake 原始记录。公开聚合数据与来源。
实现状态：#8266 仍是开放的 C1 提案；#8304 的最终合入范围是保留精度的排序与摘要复用。上述字节数据不归因为统一 Controller/Worker 已上线的收益。

+
真实模型与多题消融实证：收益链路与边界反例

+
在 EdgeBench ANN 向量检索优化（gpt-6.1-sol / xhigh，18h，约束 Recall@10 ≥ 0.95）与跨题消融中，我们将 Context Delivery（渐进读取与周期差分）与 Stride（基于证据差分的步长控制）拆开对照，同时统计控制包选中字节、主会话模型输入 Token 与局部/早窗反例（基于 @shangzh0 归档轨迹与控制面回执复算）：

+
多题消融与分层计量审计
Context 渐进式交付与 Stride 步幅控制消融：ANN 18h 峰值吞吐 QPS 对比与分层 Token/字节节省及边界反例
上图（A）：ANN 18h 六臂峰值有效吞吐 QPS（Recall@10 ≥ 0.95）。下图（B）：相对 Context off 的分层字节与主会话输入 Token 节省（左，负值）及诚实披露的局部/早窗开销与跨题反例（右，正值）。点击查看原图。

+
这组双轨实证说明：控制包裁剪率（−74.25% / −38.54%）不能直接等同于整轮模型 Token 节省（未缓存 −12.07%、总输入 −7.46%），更不能掩盖前 2 小时早窗的前置建索引开销（4.05×）和跨题反例。 报告中的 ANN 运行同时出现了较低输入总量和后程有效实验，值得据此继续检验上下文交付能否支撑长程研究。但这些观察本身不能将成果归因于省下的注意力、拆清它与 Stride 的全部交互，也不能证明换一题仍有收益。后文介绍的 11 月源码修复并不是这些归档运行的实验干预。原研究条件见 EdgeBench 后程突破与五臂归因。

+
展开逐项消融结果与反例
+

+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
观测维度或实验臂对照数值口径与解释
ANN 18h：Official / Single / Native-Goal5,753 / 5,706 / 4,372 QPSOfficial 在 1.05h 触顶后进入长平台期
ANN 18h：Explore · Context off5,766 QPS（+0.2%）11.37h 触顶；输入成本在下方单列统计
ANN 18h：Explore · Context on6,020 QPS（+4.6%）14.43h 触顶；该次运行的后程峰值不能单独隔离原因
ANN 18h：Explore · Context on + Stride7,159 QPS（+24.4%）9.19h 突破（QT_8bit_direct HNSW-SQ8，Recall@10 = 0.9585）
控制包字节：24 个 Delta 事件 / 46 个配对事件−74.25% / −38.54%配对总选中字节由 3,089,841 B 降至 1,898,889 B（Stride 臂 34 事件为 −25.75%）
主会话输入 Token：全 18h 未缓存 / 总输入−12.07% / −7.46%未缓存输入 8.57M → 7.53M（16h 公共窗口为 −7.91%）；总输入 272.03M → 251.74M
边界反例 1：project+read 进程 stdout+57% ~ +83%（1.57×–1.83×）CLI JSON 封装与分页导航使进程 stdout 高于被选中的原始文件字节
边界反例 2：前 2h 早窗未缓存输入 Token+304.7%（0.25M → 1.00M）建立基线与分步选读增加前置工作；这个早窗不能保证后续摊薄
边界反例 3：跨题 Historical Grid（越低越好）Official 1967 < Off 2054 < On+Stride 2088 < On 2242不同时长与资源条件下的历史跨题反例；保留负向观察，不作跨题因果外推

+

 
核心 Codex 原生路径已随 LoopX 2.2.0 发布。受支持的 POSIX Codex App／SSH 会话与新的 benchmark heartbeat 默认开启 context delivery；已有关闭设置继续生效，也可以显式关闭。开启后完整内容先保存，Agent 收到当前读取入口。具体发布与验证范围见 2.2.0 版本说明。

 
可以把它想成查一本厚书：capture 是完整保存的快照，像一本版本固定的书；目录告诉我们章节在哪，JSON Pointer 是书内地址。渐进式读取决定这一刻翻到多深；差分决定下一轮哪些内容可以沿用。

-
11 月 8 日这次修订，还接入了 2.2.0 之后合入的首读、阶段化范围读取、进程传输和数字精度修复，核对源码为 611019243e。对应章节描述的是这份源码，不代表所有安装都已升级。下文原有的教学测量与归档研究结果仍保留各自较早的版本和条件。

+
11 月 8 日这次修订，还接入了 2.2.0 之后合入的首读、阶段化范围读取、进程传输和数字精度修复，核对源码为 611019243e。对应章节描述的是这份源码，不代表所有安装都已升级。文中的教学测量与开篇归档研究结果仍保留各自较早的版本和条件。

 
用“管家读语音”串起一次工作

 
假设要让管家理解语音消息，并在原对话返回结果。完整资料里保留了目标、服务约束、隐私要求、验收，以及很长的方案讨论。下面是教学示意，后文的示例程序会生成完整数据，可在隔离目录复现。

 
capture.json
@@ -239,161 +400,8 @@ 
视图变短，原值也要保真

 
这些数据说明了什么

 
这组例子把三个收益来源分开了：少重复搬当前用不上的历史，少发本可合并的读取请求，少重复交付已经核验的不变结构。 完整来源、要求与恢复路径贯穿其中。

 
上面的 reader 数字包含实际 CLI 响应中的分页元数据；本机路径长度会影响少量封装字节。原生投影数字比较捕获正文，周期表比较编码层，三种口径分别列出。示例脚本可复算这些值与完整还原结果。

-
放回真实长程任务，还要把必要补读、恢复、缓存命中、模型输入输出、工具调用、等待与人工介入一起算进去；效果则看是否漏掉约束、做错决定，最后交付了什么。前面的语音合成例子测清了机械路径，下面补入真实 18 小时长程任务上的多题与分层消融观察。

-
真实模型与多题消融实证：收益链路与边界反例

-
在 EdgeBench ANN 向量检索优化（gpt-6.1-sol / xhigh，18h，约束 Recall@10 ≥ 0.95）与跨题消融中，我们将 Context Delivery（渐进读取与周期差分）与 Stride（基于证据差分的步长控制）拆开对照，同时统计控制包选中字节、主会话模型输入 Token 与局部/早窗反例（基于 @shangzh0 归档轨迹与控制面回执复算）：

-
多题消融与分层计量审计
Context 渐进式交付与 Stride 步幅控制消融：ANN 18h 峰值吞吐 QPS 对比与分层 Token/字节节省及边界反例
上图（A）：ANN 18h 六臂峰值有效吞吐 QPS（Recall@10 ≥ 0.95）。下图（B）：相对 Context off 的分层字节与主会话输入 Token 节省（左，负值）及诚实披露的局部/早窗开销与跨题反例（右，正值）。点击查看原图。

-

-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
观测维度或实验臂对照数值口径与解释
ANN 18h：Official / Single / Native-Goal5,753 / 5,706 / 4,372 QPSOfficial 在 1.05h 触顶后进入长平台期
ANN 18h：Explore · Context off5,766 QPS（+0.2%）11.37h 触顶；输入成本在下方单列统计
ANN 18h：Explore · Context on6,020 QPS（+4.6%）14.43h 触顶；该次运行的后程峰值不能单独隔离原因
ANN 18h：Explore · Context on + Stride7,159 QPS（+24.4%）9.19h 突破（QT_8bit_direct HNSW-SQ8，Recall@10 = 0.9585）
控制包字节：24 个 Delta 事件 / 46 个配对事件−74.25% / −38.54%配对总选中字节由 3,089,841 B 降至 1,898,889 B（Stride 臂 34 事件为 −25.75%）
主会话输入 Token：全 18h 未缓存 / 总输入−12.07% / −7.46%未缓存输入 8.57M → 7.53M（16h 公共窗口为 −7.91%）；总输入 272.03M → 251.74M
边界反例 1：project+read 进程 stdout+57% ~ +83%（1.57×–1.83×）CLI JSON 封装与分页导航使进程 stdout 高于被选中的原始文件字节
边界反例 2：前 2h 早窗未缓存输入 Token+304.7%（0.25M → 1.00M）建立基线与分步选读增加前置工作；这个早窗不能保证后续摊薄
边界反例 3：跨题 Historical Grid（越低越好）Official 1967 < Off 2054 < On+Stride 2088 < On 2242不同时长与资源条件下的历史跨题反例；保留负向观察，不作跨题因果外推

-
这组双轨实证说明：控制包裁剪率（−74.25% / −38.54%）不能直接等同于整轮模型 Token 节省（未缓存 −12.07%、总输入 −7.46%），更不能掩盖前 2 小时早窗的前置建索引开销（4.05×）和跨题反例。 报告中的 ANN 运行同时出现了较低输入总量和后程有效实验，值得据此继续检验上下文交付能否支撑长程研究。但这些观察本身不能将成果归因于省下的注意力、拆清它与 Stride 的全部交互，也不能证明换一题仍有收益。上文 11 月的源码修复并不是这些归档运行的实验干预。原研究条件见 EdgeBench 后程突破与五臂归因。

-
2.2.0 已交付完整来源、文件入口、批量选择、目录深度、适用验收分层、精确共享、周期差分与恢复。更广的候选、偏好、Explore 按决策阶段统一交付及全入口一致性仍有未完成部分。在「设置 → 本机 → 能力 → 差分上下文 → Codex」可以关闭，已有 off 会保留；单次命令或新 benchmark trial 可用 --context-delivery off。Trial 不继承本机开关，已有冻结 trial 不被重写。默认策略下，普通 CLI、不支持的宿主和非 heartbeat benchmark 保留原完整路径；显式请求不支持的 on 会明确拒绝，hooks 不可用时使用完整交付。保存配置、实际安装与模型采用分别核验。

-
八题四臂 Context ON：按唤醒组比较上下文字节

-
本节复算已公开的八题四臂 Context ON 18h 聚合表，以区分 Progressive 与周期内 Delta 的传输开销。原表报告 213 次模型侧唤醒：101 次新会话或压缩后的基线读取，以及 112 次周期内续跑。逐题单元格保留来源原值，合计按表重算；这里不附原始逐 wake 日志。

-
八题四臂 Context ON 控制面字节比例（N = 213 唤醒，112 周期内 Delta）
八题四臂 Context ON 18h 轨迹控制面字节比例：Progressive 占 32.5% (-67.5%)，同周期 Delta 占 35.8% (-64.2%)，Prog+Delta 首屏占 3.0% / 5.0%
A：逐题占比，Progressive 覆盖全部唤醒，Delta／首读仅覆盖周期内唤醒。B：分开汇总：213 次唤醒的 Progressive 为 7.29 / 22.41 MB；112 次周期内唤醒的 Delta 为 4.49 / 约 12.53 MB。同一 112 次的首读正文／CLI 为 0.374 / 0.631 MB。两个分母不可混用。可打开原图查看。

-

-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
题目（四臂 Context ON 18h）唤醒数（基线 / Δ）Full 全量包Progressive 必要视图（占比）同周期 Delta 载荷（占同轮 Full）Prog+Delta 首屏（正文 / 含封装）
Grid 路径优化（on_stride）29（15 / 14）4,405,651 B1,166,467 B（26.5%, −73.5%）474,016 B（20.7%, −79.3%）41,878 / 76,650 B（1.8% / 3.3%）
VLIW 调度优化（on_stride）14（5 / 9）1,297,500 B505,826 B（39.0%, −61.0%）382,664 B（45.2%, −54.8%）30,860 / 50,152 B（3.6% / 5.9%）
FFmpeg 编码加速（on_stride）26（12 / 14）2,924,522 B874,104 B（29.9%, −70.1%）759,787 B（47.1%, −52.9%）43,807 / 75,305 B（2.7% / 4.7%）
OpenTTD 路网规划（on_stride）31（19 / 12）3,288,042 B1,050,623 B（32.0%, −68.0%）436,053 B（30.8%, −69.2%）36,753 / 64,619 B（2.6% / 4.6%）
Ad Placement 广告拍卖（on_stride）29（12 / 17）3,115,820 B1,019,704 B（32.7%, −67.3%）708,828 B（39.4%, −60.6%）64,174 / 102,437 B（3.6% / 5.7%）
Borden 震源反演（on_stride）28（11 / 17）2,855,664 B1,001,254 B（35.1%, −64.9%）625,709 B（34.4%, −65.6%）50,602 / 90,024 B（2.8% / 4.9%）
Rust 工作区重构（on_stride）15（4 / 11）1,173,832 B439,814 B（37.5%, −62.5%）363,310 B（42.1%, −57.9%）40,956 / 64,251 B（4.7% / 7.4%）
ANN 向量检索（on_stride）41（23 / 18）3,347,659 B1,228,159 B（36.7%, −63.3%）737,968 B（38.6%, −61.4%）65,419 / 107,112 B（3.4% / 5.6%）
8 题合计（加权总体）213（101 / 112）22,408,690 B（100%）7,285,951 B（32.5%, −67.5%）4,488,335 B（35.8%, −64.2%）374,449 / 630,550 B（3.0% / 5.0%）

-
可以看到两层正交且稳定的字节压缩效应：（1）Progressive 渐进式投影（required-context.json）在全部 213 次唤醒上将完整决策包从 22.41 MB 压至 7.29 MB（均值 32.5%，净减 −67.5%，各题稳定在 26.5%–39.0%），折叠掉未选中候选与历史长尾详情；（2）Delta 周期内差分（delivery.json）在 112 次同周期续跑唤醒上，将对应的 12.53 MB 全量包压至 4.49 MB（均值 35.8%，净减 −64.2%，各题 20.7%–47.1%）；（3）两者叠加的 --initial-read 首屏入口仅占原包的 3.0%（含 CLI JSON 封装为 5.0%，即净减 −97.0% / −95.0%）。

+
放回真实长程任务，还要把必要补读、恢复、缓存命中、模型输入输出、工具调用、等待与人工介入一起算进去；效果则看是否漏掉约束、做错决定，最后交付了什么。前面的语音合成例子测清了机械路径；开篇的八题字节对照与 18 小时消融观察，则分别检查真实长程任务中的传输开销、模型输入和任务结果。

 
2.2.0 已交付完整来源、文件入口、批量选择、目录深度、适用验收分层、精确共享、周期差分与恢复。更广的候选、偏好、Explore 按决策阶段统一交付及全入口一致性仍有未完成部分。在「设置 → 本机 → 能力 → 差分上下文 → Codex」可以关闭，已有 off 会保留；单次命令或新 benchmark trial 可用 --context-delivery off。Trial 不继承本机开关，已有冻结 trial 不被重写。默认策略下，普通 CLI、不支持的宿主和非 heartbeat benchmark 保留原完整路径；显式请求不支持的 on 会明确拒绝，hooks 不可用时使用完整交付。保存配置、实际安装与模型采用分别核验。

-
-
口径与来源：上表逐题数据复算自已公开的八题表，Progressive 合计为 7,285,951 B / 22,408,690 B = 32.5%，覆盖全部 213 次唤醒。Delta 与首屏的 35.8%、3.0% / 5.0% 仅对应 112 次周期内唤醒；其 Full 分母原文仅报告约 12.53 MB，不能与 213 次总量混用。这里没有新增实验，也未独立核对逐 wake 原始记录。公开聚合数据与来源。
实现状态：#8266 仍是开放的 C1 提案；#8304 的最终合入范围是保留精度的排序与摘要复用。上述字节数据不归因为统一 Controller/Worker 已上线的收益。

 
用同一份材料动手读一遍

 
example.py 完整生成这个合成语音任务，调用现有 reader、渐进投影与差分 owner，逐项核对选择值、目录地址、全部 10 条要求、共享引用与完整恢复。程序不注册 Goal，不做准入或结算，也不调用模型或语音服务。

 
先准备原测量的源码 a4a0fad889d91cba0d945ce658788bdb76ba351e 及其兼容依赖，在该 checkout 根目录运行。将下载的程序路径填入下方；输出目录必须为空。