From e243abb869d158643c1d471fd952f16fd777d60c Mon Sep 17 00:00:00 2001 From: Octane Date: Fri, 14 Aug 2026 12:01:30 +0800 Subject: [PATCH 01/20] Refresh the RSIHub project page against the latest repository - Replace the benchmark section with the v2 figure and the combined Train / Full Benchmark table, matching the current README numbers. - Add the four-part composition model (stage, operator, recipe, evaluate) and the operator CLI workflow to How RSIHub works. - Link the showcase paragraph to the result snapshot, frozen rubric, and minimal seed Skill. - Sync the subtitle and descriptions to the new framing (evaluator- driven learning, controllable modification). - Add the RSIHub ring mark above the page title and page styles for the mark and multi-line code blocks. - Bump the sitemap lastmod and update the expected asset list in the migration tests. Co-Authored-By: Claude Fable 5 --- assets/styles.css | 20 ++ rsihub/assets/benchmark-results-rsihub-v2.svg | 135 ++++++++++ rsihub/assets/benchmark-results.svg | 133 ---------- rsihub/assets/rsihub-mark.svg | 8 + rsihub/index.html | 243 ++++++++---------- sitemap.xml | 2 +- tests/test_rsihub_migration.py | 3 +- 7 files changed, 278 insertions(+), 266 deletions(-) create mode 100644 rsihub/assets/benchmark-results-rsihub-v2.svg delete mode 100644 rsihub/assets/benchmark-results.svg create mode 100644 rsihub/assets/rsihub-mark.svg diff --git a/assets/styles.css b/assets/styles.css index 8421710..d9010d5 100644 --- a/assets/styles.css +++ b/assets/styles.css @@ -247,6 +247,12 @@ footer a:hover { color: var(--fg); } .project-page { max-width: 760px; } .project-page h1 { margin-bottom: 1rem; } +.project-page .page-mark { + display: block; + width: 56px; + height: 56px; + margin-bottom: 18px; +} .project-page .tagline { margin-bottom: .5rem; font: 650 19px/1.5 var(--sans); @@ -318,6 +324,20 @@ footer a:hover { color: var(--fg); } background: var(--rule); font: 500 .82em ui-monospace, SFMono-Regular, Menlo, Consolas, monospace; } +.project-page pre { + margin: 1.3rem 0 1.7rem; + padding: 14px 18px; + border-radius: 10px; + background: var(--rule); + overflow-x: auto; + font: 500 13px/1.7 ui-monospace, SFMono-Regular, Menlo, Consolas, monospace; +} +.project-page pre code { + padding: 0; + border-radius: 0; + background: none; + font: inherit; +} .project-page .doc { min-width: 0; } .toc-rail { display: none; } .project-page .abstract { diff --git a/rsihub/assets/benchmark-results-rsihub-v2.svg b/rsihub/assets/benchmark-results-rsihub-v2.svg new file mode 100644 index 0000000..6dab105 --- /dev/null +++ b/rsihub/assets/benchmark-results-rsihub-v2.svg @@ -0,0 +1,135 @@ + +RSIHub benchmark results +Terminal Bench 2 and Tau cubed Banking results for four evolution methods with MiniSWE and Codex target agents. MiniSWE is grouped left and Codex right in each panel. The vertical axes use truncated ranges (50–80% and 5–45%). Dark bar sections show seed scores above each axis baseline; light sections show improvement to the best score. + + +Terminal Bench 2 + +50% + +55% + +60% + +65% + +70% + +75% + +80% + + +MiniSWE + + +55.1 +56.2(+1.1) +AHE + + +55.1 +68.5(+13.4) +Hyperagents + + +55.1 +65.2(+10.1) +A-Evolve + + +55.1 +59.6(+4.5) +GEPA +Codex + + +60.7 +66.3(+5.6) +AHE + + +60.7 +70.8(+10.1) +Hyperagents + + +60.7 +64.0(+3.3) +A-Evolve + + +60.7 +60.7(0.0) +GEPA +Tau³ Banking + +5% + +10% + +15% + +20% + +25% + +30% + +35% + +40% + +45% + + +MiniSWE + + +12.4 +22.7(+10.3) +AHE + + +12.4 +28.9(+16.5) +Hyperagents + + +12.4 +24.7(+12.3) +A-Evolve + + +12.4 +22.7(+10.3) +GEPA +Codex + + +11.3 +26.8(+15.5) +AHE + + +11.3 +39.2(+27.9) +Hyperagents + + +11.3 +13.4(+2.1) +A-Evolve + + +11.3 +15.5(+4.2) +GEPA + + +MiniSWE + + +Codex +dark = seed · light = improvement · labels show best (change) · target: GPT-5.4-high · meta: GPT-5.4-xhigh + diff --git a/rsihub/assets/benchmark-results.svg b/rsihub/assets/benchmark-results.svg deleted file mode 100644 index 49a0c4b..0000000 --- a/rsihub/assets/benchmark-results.svg +++ /dev/null @@ -1,133 +0,0 @@ - -RSIHub benchmark results -Terminal Bench 2 and Tau cubed Banking results for four evolution methods with MiniSWE and Codex target agents. MiniSWE is grouped left and Codex right in each panel. Dark bar sections show seed scores; light sections show improvement to the best score. - - -Terminal Bench 2 - -0% - -10% - -20% - -30% - -40% - -50% - -60% - -70% - -80% - - -MiniSWE Agent - - -60.7 -71.9 (+11.2) -AHE - - -60.7 -69.7 (+9.0) -Hyperagents - - -60.7 -69.7 (+9.0) -A Evolve - - -60.7 -69.7 (+9.0) -GEPA -Codex - - -58.4 -67.4 (+9.0) -AHE - - -58.4 -69.7 (+11.3) -Hyperagents - - -58.4 -58.4 (+0.0) -A Evolve - - -58.4 -58.4 (+0.0) -GEPA -Tau³ Banking - -0% - -10% - -20% - -30% - -40% - -50% - - -MiniSWE Agent - - -27.8 -33.0 (+5.2) -AHE - - -27.8 -39.2 (+11.4) -Hyperagents - - -27.8 -35.1 (+7.3) -A Evolve - - -27.8 -33.0 (+5.2) -GEPA -Codex - - -23.7 -32.0 (+8.3) -AHE - - -23.7 -42.3 (+18.6) -Hyperagents - - -23.7 -34.0 (+10.3) -A Evolve - - -23.7 -29.9 (+6.2) -GEPA - - -MiniSWE Agent - - -Codex -dark = seed · light = improvement · labels show best (+gain) · target: GPT-5.4-high · meta: GPT-5.4-xhigh - diff --git a/rsihub/assets/rsihub-mark.svg b/rsihub/assets/rsihub-mark.svg new file mode 100644 index 0000000..ca93ef8 --- /dev/null +++ b/rsihub/assets/rsihub-mark.svg @@ -0,0 +1,8 @@ + + RSIHub ring mark + Four separated arc segments, one in each ramp color, set a quarter turn apart around an open center. + + + + + diff --git a/rsihub/index.html b/rsihub/index.html index 7db0b4f..e15e94a 100644 --- a/rsihub/index.html +++ b/rsihub/index.html @@ -8,7 +8,7 @@ RSIHub | Self-Evolving Agent Framework by Simple Agent Lab
+

Framework 开源框架 @@ -162,8 +163,8 @@

RSIHub

让智能体持续改进，并保留证据。

- A file-based framework for evaluator-driven evolution, reproducible candidate lineage, and controlled self-modification. - 一个基于文件的框架，支持评估器驱动的进化、可复现的候选谱系，以及受控的自我修改。 + A file-based evolution framework for evaluator-driven learning, reproducible candidate lineage, and controllable modification. + 一个基于文件的进化框架，支持评估器驱动的学习、可复现的候选谱系，以及可控的修改。

@@ -257,6 +258,41 @@

A recipe decides how parents are selected, how traces are analyzed, what may be edited, and which evaluations admit a new generation. The framework owns the mechanism that makes those decisions inspectable: clean candidate snapshots, protected scoring, surface enforcement, Git tags, and stamped archive records. Recipe 决定如何选择父代、如何分析轨迹、哪些内容可以被编辑、以及哪些评估可以准入新的一代。框架则负责让这些决策可检查：干净的候选快照、受保护的打分、变异面约束、Git 标签与盖章的归档记录。

+

+ The public composition model has four parts: + 公开的组合模型由四部分构成： +

+
    +
  • + a stage is a fixed lifecycle slot such as select, analyze, or mutate; + stage（阶段）是固定的生命周期槽位，例如 select、analyze、mutate； +
    • +
  • + an operator is a reusable implementation at library/<stage>/<name>.py; + operator（操作符）是位于 library/<stage>/<name>.py 的可复用实现； +
    • +
  • + a recipe is code-free selection and configuration of those operators; + recipe（配方）是对这些操作符的免代码选择与配置； +
    • +
  • + evaluate is the framework-owned trusted mechanism, never a selectable operator. + evaluate（评估）是框架持有的受信任机制，永远不是可选择的操作符。 +
    • +
+

+ Add an operator to a source checkout, validate it, and compose it without a registry edit: + 在源码检出中新增一个操作符、验证它并完成组合，全程无需修改注册表： +

+ 
uv run --frozen evolve operator new mutate my_operator
+uv run --frozen evolve operator describe mutate/my_operator
+uv run --frozen evolve operator check mutate/my_operator --config '{}'
+uv run --frozen evolve operator list mutate
+uv run --frozen evolve recipe check /path/to/my-recipe/evolve.yaml
+

+ See the operator guide for the complete workflow. + 完整工作流见 operator 指南。 +

What can evolve @@ -441,8 +477,8 @@

- Across the four-paper showcase, the deterministic completion pass rate moved from 1/4 at Gen 0 to 4/4 at Gen 2. The trials ran concurrently through Harbor’s local environment without Docker and retained ATIF trajectories plus evaluator-owned visual feedback. This is a representative evolution run rather than a broad benchmark. - 在四篇论文的 showcase 中，确定性完成通过率从 Gen 0 的 1/4 提升到 Gen 2 的 4/4。这些试验通过 Harbor 的本地环境并发运行（无需 Docker），并保留了 ATIF 轨迹与评估器持有的视觉反馈。这是一次代表性的进化运行，而非广泛的基准测试。 + Across the four-paper showcase, the deterministic completion pass rate moved from 1/4 at Gen 0 to 4/4 at Gen 2. The trials ran concurrently through Harbor’s local environment without Docker and retained ATIF trajectories plus evaluator-owned visual feedback. This is a representative evolution run rather than a broad benchmark; see the result snapshot, frozen rubric, and minimal seed Skill. + 在四篇论文的 showcase 中，确定性完成通过率从 Gen 0 的 1/4 提升到 Gen 2 的 4/4。这些试验通过 Harbor 的本地环境并发运行（无需 Docker），并保留了 ATIF 轨迹与评估器持有的视觉反馈。这是一次代表性的进化运行，而非广泛的基准测试；另见结果快照、冻结的评分标准与最小种子 Skill。

@@ -450,26 +486,25 @@

基准测试结果

- Scores are shown as seed → best, with the absolute change underneath. All runs use a GPT-5.4-high target model and a GPT-5.4-xhigh Codex meta-agent. - 分数以 seed → best 呈现，括号中为绝对变化。所有运行使用 GPT-5.4-high 目标模型与 GPT-5.4-xhigh Codex meta-agent。 + Scores are percentages shown as seed → evolved agent, with the absolute change underneath. The train score is measured on the recipe’s training split; the full benchmark score is measured across the complete benchmark. All runs use a GPT-5.4-high target model and a GPT-5.4-xhigh Codex mutate operator. + 分数为百分比，以 seed → 进化后的智能体 呈现，括号中为绝对变化。Train 分数在 recipe 的训练划分上测得；完整基准分数在整个基准上测得。所有运行使用 GPT-5.4-high 目标模型与 GPT-5.4-xhigh 的 Codex mutate 操作符。

Terminal Bench 2 and Tau cubed Banking results for AHE, Hyperagents, A Evolve, and GEPA with MiniSWE and Codex target agents. Each stacked bar labels the seed score inside the dark section and the best score plus improvement above the light section.
-

Terminal Bench 2

-

- Split: 50 train / 19 gate / 20 sealed. - 数据划分：50 train / 19 gate / 20 sealed。 -

+ - - - - + + - - - - - - - - - - - - - + + + + + - - - - - + + + - - - - - + + + - - - - - - + + + - - - - - + + + + - - - - - + + + - - - - - + + + - -
+ Benchmark + 基准 + Target agent 目标智能体 @@ -478,156 +513,102 @@

Terminal Bench 2

Method 方法 		TrainGateSealedOverall + Train score + Train 分数 + + Full benchmark score + 完整基准分数 +
MiniSWE AgentAHE58.0% → 74.0%
(+16.0%)		57.9% → 68.4%
(+10.5%)		70.0% → 70.0%
(+0.0%)		60.7% → 71.9%
(+11.2%)
Hyperagents58.0% → 68.0%
(+10.0%)		57.9% → 73.7%
(+15.8%)		70.0% → 70.0%
(+0.0%)		60.7% → 69.7%
(+9.0%)		Terminal-Bench 2
50 train / 19 gate / 20 sealed		MiniSWEAHE70.0% → 74.0%
(+4.0%)		55.1% → 56.2%
(+1.1%)
A Evolve58.0% → 68.0%
(+10.0%)		57.9% → 78.9%
(+21.0%)		70.0% → 65.0%
(−5.0%)		60.7% → 69.7%
(+9.0%)		Hyperagents58.0% → 68.0%
(+10.0%)		55.1% → 68.5%
(+13.4%)
GEPA58.0% → 68.0%
(+10.0%)		57.9% → 68.4%
(+10.5%)		70.0% → 75.0%
(+5.0%)		60.7% → 69.7%
(+9.0%)		A-Evolve66.0% → 68.0%
(+2.0%)		55.1% → 65.2%
(+10.1%)
CodexAHE58.0% → 74.0%
(+16.0%)		52.6% → 47.4%
(−5.2%)		65.0% → 70.0%
(+5.0%)		58.4% → 67.4%
(+9.0%)		GEPA58.0% → 68.0%
(+10.0%)		55.1% → 59.6%
(+4.5%)
Hyperagents58.0% → 72.0%
(+14.0%)		52.6% → 57.9%
(+5.3%)		65.0% → 75.0%
(+10.0%)		58.4% → 69.7%
(+11.3%)		CodexAHE60.0% → 74.0%
(+14.0%)		60.7% → 66.3%
(+5.6%)
A Evolve58.0% → 58.0%
(+0.0%)		52.6% → 52.6%
(+0.0%)		65.0% → 65.0%
(+0.0%)		58.4% → 58.4%
(+0.0%)		Hyperagents58.0% → 72.0%
(+14.0%)		60.7% → 70.8%
(+10.1%)
GEPA58.0% → 58.0%
(+0.0%)		52.6% → 52.6%
(+0.0%)		65.0% → 65.0%
(+0.0%)		58.4% → 58.4%
(+0.0%)		A-Evolve62.0% → 62.0%
(0.0%)		60.7% → 64.0%
(+3.3%)
-
- -

Tau³ Banking

-

- Split: 50 train / 20 gate / 27 sealed. - 数据划分：50 train / 20 gate / 27 sealed。 -

-
- - - - - - - - + + + - - - - - - - - + + + + + - - - - - + + + - - - - - + + + - - - - - + + + - - - - - - + + + + - - - - - + + + - - - - - + + + - - - - - + + +
- Target agent - 目标智能体 - - Method - 方法 - TrainGateSealedOverallGEPA64.0% → 64.0%
(0.0%)		60.7% → 60.7%
(0.0%)
MiniSWE AgentAHE30.0% → 36.0%
(+6.0%)		35.0% → 35.0%
(+0.0%)		18.5% → 25.9%
(+7.4%)		27.8% → 33.0%
(+5.2%)		Tau³ Banking
50 train / 20 gate / 27 sealed		MiniSWEAHE34.0% → 36.0%
(+2.0%)		12.4% → 22.7%
(+10.3%)
Hyperagents30.0% → 38.0%
(+8.0%)		35.0% → 45.0%
(+10.0%)		18.5% → 37.0%
(+18.5%)		27.8% → 39.2%
(+11.4%)		Hyperagents30.0% → 38.0%
(+8.0%)		12.4% → 28.9%
(+16.5%)
A Evolve30.0% → 34.0%
(+4.0%)		35.0% → 45.0%
(+10.0%)		18.5% → 29.6%
(+11.1%)		27.8% → 35.1%
(+7.3%)		A-Evolve30.0% → 34.0%
(+4.0%)		12.4% → 24.7%
(+12.3%)
GEPA30.0% → 32.0%
(+2.0%)		35.0% → 45.0%
(+10.0%)		18.5% → 25.9%
(+7.4%)		27.8% → 33.0%
(+5.2%)		GEPA30.0% → 32.0%
(+2.0%)		12.4% → 22.7%
(+10.3%)
CodexAHE30.0% → 36.0%
(+6.0%)		30.0% → 45.0%
(+15.0%)		7.4% → 14.8%
(+7.4%)		23.7% → 32.0%
(+8.3%)		CodexAHE32.0% → 36.0%
(+4.0%)		11.3% → 26.8%
(+15.5%)
Hyperagents30.0% → 36.0%
(+6.0%)		30.0% → 50.0%
(+20.0%)		7.4% → 48.1%
(+40.7%)		23.7% → 42.3%
(+18.6%)		Hyperagents34.0% → 36.0%
(+2.0%)		11.3% → 39.2%
(+27.9%)
A Evolve30.0% → 38.0%
(+8.0%)		30.0% → 45.0%
(+15.0%)		7.4% → 18.5%
(+11.1%)		23.7% → 34.0%
(+10.3%)		A-Evolve30.0% → 38.0%
(+8.0%)		11.3% → 13.4%
(+2.1%)
GEPA30.0% → 36.0%
(+6.0%)		30.0% → 35.0%
(+5.0%)		7.4% → 14.8%
(+7.4%)		23.7% → 29.9%
(+6.2%)		GEPA10.0% → 16.0%
(+6.0%)		11.3% → 15.5%
(+4.2%)
diff --git a/sitemap.xml b/sitemap.xml index 02d16b6..d20f144 100644 --- a/sitemap.xml +++ b/sitemap.xml @@ -6,7 +6,7 @@ https://simpleagentlab.com/rsihub/ - 2026-08-12 + 2026-08-14 https://simpleagentlab.com/ai4ai/ diff --git a/tests/test_rsihub_migration.py b/tests/test_rsihub_migration.py index 84dfbab..f71b22f 100644 --- a/tests/test_rsihub_migration.py +++ b/tests/test_rsihub_migration.py @@ -76,10 +76,11 @@ def test_project_assets_moved_and_svg_titles_renamed(self): assets = ROOT / "rsihub/assets" expected = { "architecture.svg", - "benchmark-results.svg", + "benchmark-results-rsihub-v2.svg", "evolve-lineage.svg", "paper-poster-lora-gen0.png", "paper-poster-lora-gen2.png", + "rsihub-mark.svg", } self.assertEqual({path.name for path in assets.iterdir()}, expected) for svg in assets.glob("*.svg"): From c44bf9ecf1358aa3947180f50ad5d4b60a284cfe Mon Sep 17 00:00:00 2001 From: Octane Date: Fri, 14 Aug 2026 13:33:55 +0800 Subject: [PATCH 02/20] Rebuild the RSIHub page header with the lockup and a rotating tagline MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit - Replace the ring mark + text title with the inline repo lockup so the wordmark follows the site's manual theme toggle (the standalone SVG only responds to prefers-color-scheme). - Add an animated hero line — evolve your on — whose slots rotate through Codex, Claude Code, a DeepSeek harness, and MiniSWE against Terminal-Bench 2, Tau³ Banking, and your own eval; plain text swaps under prefers-reduced-motion. - Remove the lineage figure and the now-unused evolve-lineage and ring-mark assets; sync the migration-test asset list. Co-Authored-By: Claude Fable 5 --- assets/styles.css | 34 ++++++++++-- rsihub/assets/evolve-lineage.svg | 22 -------- rsihub/assets/rsihub-mark.svg | 8 --- rsihub/index.html | 92 +++++++++++++++++++++++++++----- tests/test_rsihub_migration.py | 2 - 5 files changed, 109 insertions(+), 49 deletions(-) delete mode 100644 rsihub/assets/evolve-lineage.svg delete mode 100644 rsihub/assets/rsihub-mark.svg diff --git a/assets/styles.css b/assets/styles.css index d9010d5..a0f0f10 100644 --- a/assets/styles.css +++ b/assets/styles.css @@ -8,6 +8,7 @@ --selection: rgba(0, 203, 212, .18); --focus: #3c8cff; --blue: #3c8cff; + --teal: #0e7c70; --serif: "Iowan Old Style", "Songti SC", "Noto Serif CJK SC", Georgia, serif; --sans: -apple-system, BlinkMacSystemFont, "Segoe UI", "PingFang SC", "Microsoft YaHei", sans-serif; color-scheme: light dark; @@ -22,6 +23,7 @@ --header: rgba(17, 21, 19, .94); --selection: #5a3a32; --focus: #ef8068; + --teal: #00cbd4; } * { box-sizing: border-box; margin: 0; padding: 0; } @@ -247,12 +249,35 @@ footer a:hover { color: var(--fg); } .project-page { max-width: 760px; } .project-page h1 { margin-bottom: 1rem; } -.project-page .page-mark { +.project-page h1.lockup { margin: 0 0 1.4rem; } +.project-page .lockup svg { display: block; - width: 56px; - height: 56px; - margin-bottom: 18px; + width: min(300px, 74vw); + height: auto; + aspect-ratio: 174.09 / 64; +} +.lockup .rsi { fill: var(--fg); } +.lockup .ramp-third { stop-color: #00a3b0; } +.lockup .ramp-fourth { stop-color: #2fa844; } +:root[data-theme="dark"] .lockup .ramp-third { stop-color: #00cbd4; } +:root[data-theme="dark"] .lockup .ramp-fourth { stop-color: #78e85c; } +.project-page .hero-line { + margin: 0 0 1.3rem; + font: 650 clamp(20px, 3.6vw, 29px)/1.55 var(--sans); + letter-spacing: -.035em; +} +.project-page .rotor { + display: inline-block; + max-width: 100%; + padding: 0 10px 2px; + border-radius: 9px; + background: var(--rule); + color: var(--blue); + white-space: nowrap; + transition: opacity .22s ease, transform .22s ease, width .3s ease; } +.project-page .rotor-bench { color: var(--teal); } +.project-page .rotor.is-out { opacity: 0; transform: translateY(.3em); } .project-page .tagline { margin-bottom: .5rem; font: 650 19px/1.5 var(--sans); @@ -458,4 +483,5 @@ footer a:hover { color: var(--fg); } @media (prefers-reduced-motion: reduce) { html { scroll-behavior: auto; } .masthead-mark { animation: none; } + .project-page .rotor { transition: none; } } diff --git a/rsihub/assets/evolve-lineage.svg b/rsihub/assets/evolve-lineage.svg deleted file mode 100644 index 3c7839c..0000000 --- a/rsihub/assets/evolve-lineage.svg +++ /dev/null @@ -1,22 +0,0 @@ - - Verified candidate lineage - A baseline branches into evaluated candidates. The selected lineage rises through successive generations to a verified improvement while unselected candidates remain visible as evidence. - - - - - - - - - - - - - - baseline - explored candidates - selected lineage - verified generation - - diff --git a/rsihub/assets/rsihub-mark.svg b/rsihub/assets/rsihub-mark.svg deleted file mode 100644 index ca93ef8..0000000 --- a/rsihub/assets/rsihub-mark.svg +++ /dev/null @@ -1,8 +0,0 @@ - - RSIHub ring mark - Four separated arc segments, one in each ramp color, set a quarter turn apart around an open center. - - - - - diff --git a/rsihub/index.html b/rsihub/index.html index e15e94a..85cf3e8 100644 --- a/rsihub/index.html +++ b/rsihub/index.html @@ -152,12 +152,45 @@
-

Framework 开源框架

-

RSIHub

+

+ + RSIHub + The RSIHub ring mark beside the RSIHub wordmark, RSI in interface ink followed by Hub in a blue-to-green ramp. + + + + + + + + + + + + + + + + + + + + + + + + + + +

+

+ Evolve your Codex on Terminal-Bench 2. + 在 Terminal-Bench 2 上进化你的 Codex。 +

Build agents that improve — and keep the evidence. 让智能体持续改进，并保留证据。 @@ -183,17 +216,6 @@

RSIHub

-
- A baseline branches into evaluated candidates. The selected lineage rises through successive generations to a verified improvement, while unselected candidates remain visible as evidence. -
- The selected lineage rises through generations to a verified improvement; unselected candidates remain visible as evidence. - 被选中的谱系沿世代上升，抵达经验证的改进；未被选中的候选作为证据保留可见。 -
-
-

What RSIHub does RSIHub 做什么 @@ -761,5 +783,49 @@

© 2026 + + diff --git a/tests/test_rsihub_migration.py b/tests/test_rsihub_migration.py index f71b22f..299307e 100644 --- a/tests/test_rsihub_migration.py +++ b/tests/test_rsihub_migration.py @@ -77,10 +77,8 @@ def test_project_assets_moved_and_svg_titles_renamed(self): expected = { "architecture.svg", "benchmark-results-rsihub-v2.svg", - "evolve-lineage.svg", "paper-poster-lora-gen0.png", "paper-poster-lora-gen2.png", - "rsihub-mark.svg", } self.assertEqual({path.name for path in assets.iterdir()}, expected) for svg in assets.glob("*.svg"): From d9f7a63e5527a64967241f56ac94aaf14a36f335 Mon Sep 17 00:00:00 2001 From: Octane Date: Fri, 14 Aug 2026 13:51:04 +0800 Subject: [PATCH 03/20] Redesign the RSIHub page as a product landing page Replace the README-style article with a proper product homepage: - Hero: repo lockup (ring + wordmark, theme-aware inline SVG), a large two-line rotating headline (evolve your / on ) with per-language word lists, sub-copy, and CTA buttons. - A terminal card walking through the real quick-start commands with published seed/best train scores, plus an honesty note. - Stats strip (best full-benchmark gains, recipe count, license) with ramp-gradient numerals. - Sections rebuilt as landing blocks: loop pills + architecture, composition-model cards + operator CLI, audience cards, mutable- surface cards, recipe rows, showcase, benchmark chart with the full table folded into a details element, trust grid, and a closing CTA band with doc links. - All styling is page-scoped; the shared stylesheet is restored to its upstream state. Bilingual parity and the manual theme toggle are preserved throughout; reduced-motion swaps text without animation. Co-Authored-By: Claude Fable 5 --- assets/styles.css | 46 -- rsihub/index.html | 1358 ++++++++++++++++++++++++++++----------------- 2 files changed, 864 insertions(+), 540 deletions(-) diff --git a/assets/styles.css b/assets/styles.css index a0f0f10..8421710 100644 --- a/assets/styles.css +++ b/assets/styles.css @@ -8,7 +8,6 @@ --selection: rgba(0, 203, 212, .18); --focus: #3c8cff; --blue: #3c8cff; - --teal: #0e7c70; --serif: "Iowan Old Style", "Songti SC", "Noto Serif CJK SC", Georgia, serif; --sans: -apple-system, BlinkMacSystemFont, "Segoe UI", "PingFang SC", "Microsoft YaHei", sans-serif; color-scheme: light dark; @@ -23,7 +22,6 @@ --header: rgba(17, 21, 19, .94); --selection: #5a3a32; --focus: #ef8068; - --teal: #00cbd4; } * { box-sizing: border-box; margin: 0; padding: 0; } @@ -249,35 +247,6 @@ footer a:hover { color: var(--fg); } .project-page { max-width: 760px; } .project-page h1 { margin-bottom: 1rem; } -.project-page h1.lockup { margin: 0 0 1.4rem; } -.project-page .lockup svg { - display: block; - width: min(300px, 74vw); - height: auto; - aspect-ratio: 174.09 / 64; -} -.lockup .rsi { fill: var(--fg); } -.lockup .ramp-third { stop-color: #00a3b0; } -.lockup .ramp-fourth { stop-color: #2fa844; } -:root[data-theme="dark"] .lockup .ramp-third { stop-color: #00cbd4; } -:root[data-theme="dark"] .lockup .ramp-fourth { stop-color: #78e85c; } -.project-page .hero-line { - margin: 0 0 1.3rem; - font: 650 clamp(20px, 3.6vw, 29px)/1.55 var(--sans); - letter-spacing: -.035em; -} -.project-page .rotor { - display: inline-block; - max-width: 100%; - padding: 0 10px 2px; - border-radius: 9px; - background: var(--rule); - color: var(--blue); - white-space: nowrap; - transition: opacity .22s ease, transform .22s ease, width .3s ease; -} -.project-page .rotor-bench { color: var(--teal); } -.project-page .rotor.is-out { opacity: 0; transform: translateY(.3em); } .project-page .tagline { margin-bottom: .5rem; font: 650 19px/1.5 var(--sans); @@ -349,20 +318,6 @@ footer a:hover { color: var(--fg); } background: var(--rule); font: 500 .82em ui-monospace, SFMono-Regular, Menlo, Consolas, monospace; } -.project-page pre { - margin: 1.3rem 0 1.7rem; - padding: 14px 18px; - border-radius: 10px; - background: var(--rule); - overflow-x: auto; - font: 500 13px/1.7 ui-monospace, SFMono-Regular, Menlo, Consolas, monospace; -} -.project-page pre code { - padding: 0; - border-radius: 0; - background: none; - font: inherit; -} .project-page .doc { min-width: 0; } .toc-rail { display: none; } .project-page .abstract { @@ -483,5 +438,4 @@ footer a:hover { color: var(--fg); } @media (prefers-reduced-motion: reduce) { html { scroll-behavior: auto; } .masthead-mark { animation: none; } - .project-page .rotor { transition: none; } } diff --git a/rsihub/index.html b/rsihub/index.html index 85cf3e8..988f9e7 100644 --- a/rsihub/index.html +++ b/rsihub/index.html @@ -93,6 +93,325 @@ + From 296e7056b4e0ce672f5f9c7c1544f902efbd861e Mon Sep 17 00:00:00 2001 From: Octane Date: Fri, 14 Aug 2026 15:42:10 +0800 Subject: [PATCH 10/20] Rebuild the architecture figure as an animated HTML diagram The static architecture.svg (and the redundant loop-pill row above it) is replaced by an in-page diagram in the benchmark chart's visual language: blue method pills plus a dashed your-recipe slot, a blue-tinted composable-loop box whose orange dashed border marks the declared mutable surface, six stage cards (Mutate and Gate highlighted orange) with a next-generation return loop, evidence/contract connectors, and dark slate protected-mechanism cards. On section entry the pills, connector, stages, surface, and locks appear in sequence; a soft highlight ring then cycles through the six stages continuously. All motion is disabled under prefers-reduced-motion. Also: the CTA column now centers its items, fixing the git-clone chip stretching to full width, and the two now-unused SVG assets are removed with the migration-test asset list updated. Co-Authored-By: Claude Fable 5 --- rsihub/assets/architecture.svg | 126 ------------ rsihub/assets/benchmark-results-rsihub-v2.svg | 135 ------------- rsihub/index.html | 191 +++++++++++++++--- tests/test_rsihub_migration.py | 2 - 4 files changed, 164 insertions(+), 290 deletions(-) delete mode 100644 rsihub/assets/architecture.svg delete mode 100644 rsihub/assets/benchmark-results-rsihub-v2.svg diff --git a/rsihub/assets/architecture.svg b/rsihub/assets/architecture.svg deleted file mode 100644 index 17c02b4..0000000 --- a/rsihub/assets/architecture.svg +++ /dev/null @@ -1,126 +0,0 @@ - - RSIHub architecture - Evolution methods such as Hill Climb, A-Evolve, AHE, GEPA and HyperAgents plug into one loop of select, rollout, analyze, mutate, gate and record. The loop and the agent it improves sit inside a declared mutable surface. Recipes select permitted targets, operators, and stages. Only the substrate below stays frozen: the evaluator, the runtime, the surface check and the stamped evidence. - - - - EVOLUTION METHODS - Five built-in strategies, or compose your own - - Hill Climb - - A-Evolve - - AHE - - GEPA - - HyperAgents - - your recipe - - - - - - - - - - ONE COMPOSABLE LOOP - recipes choose which stages and targets may evolve - - - DECLARED MUTABLE SURFACE · target + selected operators - - - - - - - - - - - - - - - Select - choose a parent - - Rollout - run tasks + evaluator - - Analyze - inspect the traces - - Mutate - edit the candidate - - Gate - accept verified gains - - Record - tag + archive - - - - - - - - - - - - - next generation - - - - evidence - - - contracts - PROTECTED MECHANISM - outside candidate control - - Frozen evaluator - pinned scoring contract - - Locked runtime - pinned deps + digest - - Surface check - declared paths only - - Stamped evidence - archive + git lineage - diff --git a/rsihub/assets/benchmark-results-rsihub-v2.svg b/rsihub/assets/benchmark-results-rsihub-v2.svg deleted file mode 100644 index 6dab105..0000000 --- a/rsihub/assets/benchmark-results-rsihub-v2.svg +++ /dev/null @@ -1,135 +0,0 @@ - -RSIHub benchmark results -Terminal Bench 2 and Tau cubed Banking results for four evolution methods with MiniSWE and Codex target agents. MiniSWE is grouped left and Codex right in each panel. The vertical axes use truncated ranges (50–80% and 5–45%). Dark bar sections show seed scores above each axis baseline; light sections show improvement to the best score. - - -Terminal Bench 2 - -50% - -55% - -60% - -65% - -70% - -75% - -80% - - -MiniSWE - - -55.1 -56.2(+1.1) -AHE - - -55.1 -68.5(+13.4) -Hyperagents - - -55.1 -65.2(+10.1) -A-Evolve - - -55.1 -59.6(+4.5) -GEPA -Codex - - -60.7 -66.3(+5.6) -AHE - - -60.7 -70.8(+10.1) -Hyperagents - - -60.7 -64.0(+3.3) -A-Evolve - - -60.7 -60.7(0.0) -GEPA -Tau³ Banking - -5% - -10% - -15% - -20% - -25% - -30% - -35% - -40% - -45% - - -MiniSWE - - -12.4 -22.7(+10.3) -AHE - - -12.4 -28.9(+16.5) -Hyperagents - - -12.4 -24.7(+12.3) -A-Evolve - - -12.4 -22.7(+10.3) -GEPA -Codex - - -11.3 -26.8(+15.5) -AHE - - -11.3 -39.2(+27.9) -Hyperagents - - -11.3 -13.4(+2.1) -A-Evolve - - -11.3 -15.5(+4.2) -GEPA - - -MiniSWE - - -Codex -dark = seed · light = improvement · labels show best (change) · target: GPT-5.4-high · meta: GPT-5.4-xhigh - diff --git a/rsihub/index.html b/rsihub/index.html index 83c8c03..963f0c0 100644 --- a/rsihub/index.html +++ b/rsihub/index.html @@ -354,21 +354,135 @@ .block .lead { max-width: 700px; margin: 0 0 36px; } .block .after { max-width: 700px; margin: 26px 0 0; } - /* Loop pills */ - .loop { display: flex; flex-wrap: wrap; align-items: center; gap: 10px; margin: 0 0 34px; } - .loop .pill { - padding: 10px 17px; - border: 1px solid var(--rule); + /* Architecture diagram (animated, blue/orange palette) */ + .arch { border: 1px solid var(--rule); border-radius: 16px; padding: 22px 22px 18px; } + .arch-head { display: flex; justify-content: space-between; align-items: baseline; gap: 14px; flex-wrap: wrap; margin-bottom: 12px; } + .arch-cap { color: var(--soft); font: 750 11px/1.4 var(--sans); letter-spacing: .09em; text-transform: uppercase; } + .arch-note { color: var(--faint); font: 450 11.5px/1.4 var(--sans); } + .arch-pills { display: flex; flex-wrap: wrap; gap: 10px; } + .arch-m { + flex: 1; + min-width: 110px; + padding: 10px 12px; + border: 1px solid rgba(29, 78, 216, .35); + border-radius: 10px; + background: rgba(147, 197, 253, .12); + color: #1d4ed8; + font: 700 13px/1.2 var(--sans); + text-align: center; + transition: opacity .45s ease var(--d), transform .45s ease var(--d); + } + .arch-m.custom { border-style: dashed; border-color: var(--faint); background: none; color: var(--soft); } + :root[data-theme="dark"] .arch-m { color: #93c5fd; border-color: rgba(147, 197, 253, .35); background: rgba(59, 130, 246, .14); } + :root[data-theme="dark"] .arch-m.custom { border-color: var(--faint); background: none; color: var(--soft); } + .arch-down { display: block; width: 1px; height: 24px; margin: 12px auto; background: var(--faint); transform-origin: top; transition: transform .4s ease .45s; } + .arch-loop { border: 1px solid var(--rule); border-radius: 14px; padding: 16px 16px 14px; background: rgba(147, 197, 253, .07); } + :root[data-theme="dark"] .arch-loop { background: rgba(59, 130, 246, .07); } + .arch-surface { position: relative; border: 1.5px dashed rgba(194, 65, 12, .5); border-radius: 12px; padding: 22px 14px 14px; margin-top: 10px; transition: opacity .5s ease 1.3s; } + :root[data-theme="dark"] .arch-surface { border-color: rgba(253, 186, 116, .45); } + .arch-surface-label { + position: absolute; + top: 0; + left: 50%; + transform: translate(-50%, -50%); + max-width: 94%; + padding: 3px 10px; border-radius: 999px; - font: 650 14px/1 var(--sans); - color: var(--fg); + background: var(--bg); + color: #c2410c; + font: 750 10.5px/1.4 var(--sans); + letter-spacing: .06em; + text-transform: uppercase; + white-space: nowrap; + overflow: hidden; + text-overflow: ellipsis; + } + :root[data-theme="dark"] .arch-surface-label { color: #fdba74; } + .arch-stages { display: grid; grid-template-columns: repeat(6, 1fr); gap: 22px; } + .stage { + position: relative; + padding: 12px 10px 13px; + border: 1px solid var(--rule); + border-radius: 10px; + background: var(--bg); + text-align: center; + transition: opacity .45s ease var(--d), transform .45s ease var(--d); + animation: stage-pulse 7.2s ease-in-out calc(2.2s + var(--p) * 1.2s) infinite; + } + .stage b { display: block; margin-bottom: 3px; font: 750 13px/1.3 var(--sans); color: var(--fg); } + .stage small { color: var(--soft); font: 450 10.5px/1.35 var(--sans); } + .stage::after { + content: "→"; + position: absolute; + top: 50%; + right: -19px; + transform: translateY(-50%); + color: var(--faint); + font: 400 13px/1 var(--sans); + } + .arch-stages .stage:last-child::after { content: none; } + .stage.hot { border-color: rgba(194, 65, 12, .45); background: rgba(253, 186, 116, .14); } + .stage.hot b { color: #c2410c; } + :root[data-theme="dark"] .stage.hot b { color: #fdba74; } + @keyframes stage-pulse { + 0%, 12% { box-shadow: 0 0 0 3px rgba(29, 78, 216, .22); } + 18%, 100% { box-shadow: 0 0 0 0 rgba(29, 78, 216, 0); } + } + .arch-return { + position: relative; + height: 14px; + margin: 16px 6px 14px; + border: 1px solid var(--rule); + border-top: none; + border-radius: 0 0 10px 10px; + transition: opacity .5s ease 1.45s; + } + .arch-return span { + position: absolute; + top: 100%; + left: 50%; + transform: translateX(-50%); + margin-top: 3px; + color: var(--faint); + font: 550 10.5px/1.4 var(--sans); + white-space: nowrap; + } + .arch-io { display: flex; justify-content: space-between; margin: 12px 10px; color: var(--faint); font: 550 11px/1.4 var(--sans); transition: opacity .5s ease 1.4s; } + .arch-locks { display: grid; grid-template-columns: repeat(4, 1fr); gap: 12px; } + .lock { + padding: 12px 12px 13px; + border: 1px solid rgba(255, 255, 255, .08); + border-radius: 10px; + background: #0f172a; + text-align: center; + transition: opacity .45s ease var(--d), transform .45s ease var(--d); + } + :root[data-theme="dark"] .lock { background: #1e293b; } + .lock b { display: block; margin-bottom: 3px; color: #fff; font: 750 12.5px/1.3 var(--sans); } + .lock small { color: #94a3b8; font: 450 10.5px/1.35 var(--sans); } + /* Diagram stays collapsed until its section scrolls into view. */ + .js .landing .block:not(.in) .arch-m, + .js .landing .block:not(.in) .stage, + .js .landing .block:not(.in) .lock { opacity: 0; transform: translateY(10px); } + .js .landing .block:not(.in) .arch-down { transform: scaleY(0); } + .js .landing .block:not(.in) .arch-surface, + .js .landing .block:not(.in) .arch-return, + .js .landing .block:not(.in) .arch-io { opacity: 0; } + @media (max-width: 860px) { + .arch-stages { grid-template-columns: repeat(2, 1fr); gap: 14px; } + .stage::after { content: none; } + .arch-locks { grid-template-columns: repeat(2, 1fr); } + .arch-surface { padding-top: 14px; } + .arch-surface-label { + position: static; + display: block; + width: fit-content; + margin: 0 auto 12px; + transform: none; + white-space: normal; + text-align: center; + } } - .loop .pill.locked { background: var(--rule); border-color: transparent; } - .loop b { color: var(--faint); font: 400 16px/1 var(--sans); } - - /* Figures */ - .shot { margin: 0; border: 1px solid var(--rule); border-radius: 16px; overflow: hidden; } - .shot img { display: block; width: 100%; height: auto; } /* Recipe rows */ .recipes { margin-top: 6px; border-top: 1px solid var(--rule); } @@ -424,6 +538,7 @@ display: flex; flex-direction: column; justify-content: center; + align-items: center; margin-top: 32px; padding: 78px 24px 12px; border-top: 1px solid var(--rule); @@ -479,6 +594,8 @@ .terminal .t-row { animation: none; } html { scroll-snap-type: none; } .reveal, .bar .fill, .bar-top { transition: none; } + .arch-m, .stage, .lock, .arch-down, .arch-surface, .arch-return, .arch-io { transition: none; } + .stage { animation: none; } }