diff --git a/.gitignore b/.gitignore index e43b0f9..edf0661 100644 --- a/.gitignore +++ b/.gitignore @@ -1 +1,3 @@ .DS_Store +__pycache__/ +.worktrees/ diff --git a/evolvex/index.html b/evolvex/index.html index 37cce4f..c83f8c6 100644 --- a/evolvex/index.html +++ b/evolvex/index.html @@ -1,784 +1,15 @@ - + - - - EvolveX | Self-Evolving Agent Framework by Simple Agent Lab - - - - - - - - - - - - - - - - - - - - - - - - - - + RSIHub has moved | Simple Agent Lab + + + + - - -
- - - Simple Agent Lab - - -
- -
-
-

- Framework - 开源框架 -

-

EvolveX

-

- Build agents that improve — and keep the evidence. - 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。 -

-
- - GitHub - - - - Documentation - 文档 - - - - Quick start - 快速开始 - - -
- -
- 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 EvolveX does - EvolveX 做什么 -

-

- EvolveX gives an agent a controlled way to improve itself. It runs candidates against a fixed evaluator, keeps the evidence for every generation, and carries verified improvements forward without letting candidate code rewrite the rules that score it. - EvolveX 为智能体提供一种受控的自我改进方式：候选在固定的评估器上运行，每一代都保留证据，经验证的改进被传递下去——同时不允许候选代码改写为其打分的规则。 -

-
- - - - - - - - - - - - - - - -
- For agent builders - 面向智能体开发者 - - For researchers - 面向研究者 - - Evidence built in - 内建证据 -
- Improve prompts, skills, harnesses, and agent code in a reusable experiment workspace. - 在可复用的实验工作区中改进提示词、技能、harness 与智能体代码。 - - Compare evolution strategies under fixed evaluation and mutation boundaries. - 在固定的评估与变异边界下比较不同进化策略。 - - Connect every candidate to scores, artifacts, archive records, and Git lineage. - 将每个候选与得分、产物、归档记录和 Git 谱系关联起来。 -
-
- -

- How EvolveX works - EvolveX 如何工作 -

-

- Every recipe composes the same loop: - 每个 recipe 都由同一个循环组成： -

-

select → evaluate → analyze → mutate → gate → record

-
- EvolveX architecture: five built-in strategies and custom recipes compose a loop of select, rollout and evaluation, analyze, mutate, gate, and record. The target and selected operators occupy a declared mutable surface. The evaluator, runtime, surface check, and stamped evidence remain protected from candidate changes. -
-

- 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。 -

- -

- What can evolve - 什么可以进化 -

-
- - - - - - - - - - - - - - - - - - - - - - - - - -
- Surface - 可进化面 - - Examples - 示例 - - Best fit - 适用场景 -
- prompts and skills - 提示词与技能 - - system prompts, task skills, reusable instructions - 系统提示词、任务技能、可复用指令 - - policy and behavior improvement - 策略与行为改进 -
- harnesses and target code - harness 与目标代码 - - tools, orchestration, agent implementation - 工具、编排、智能体实现 - - agent engineering - 智能体工程 -
- selected evolution operators - 被选定的进化操作符 - - analysis or mutation policy chosen by a recipe - recipe 选定的分析或变异策略 - - controlled co-evolution - 受控的共同进化 -
-
-

- Each recipe declares its mutable paths. Evaluators, archive stamps, and the vendored framework mechanism stay outside that surface. - 每个 recipe 声明自己的可变路径。评估器、归档盖章与内置的框架机制始终位于可变面之外。 -

- -

Recipes

-
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
- Choose this when you want to… - 当你想要… - Recipe - Mutable surface - 可变面 -
- improve one candidate from its current best parent - 从当前最优父代持续改进单个候选 - hill_climbtarget
- evolve prompts and reusable agent skills - 进化提示词与可复用的智能体技能 - aevolveprompt and target skills
- engineer the agent harness against evaluator feedback - 依据评估器反馈改进 agent harness - ahetarget
- balance multiple objectives with minibatch validation - 通过 minibatch 验证平衡多个目标 - gepaprompt and task skill
- co-evolve the target and selected evolution policy - 让目标与选定的进化策略共同进化 - hyperagentstarget and selected operators
-
-

- See the recipe guide for each strategy’s workflow and configuration. - 每个策略的工作流与配置见 recipe 指南。 -

- -

- Skill evolution showcase - 技能进化 Showcase -

-

- EvolveX can improve a Skill as a complete package: instructions, references, and validation scripts evolve together while a frozen evaluator keeps the comparison honest. In this local Paper2Poster run, the same Codex model and paper prompt produced both LoRA posters below. - EvolveX 可以把一个 Skill 作为完整的包来改进：指令、参考资料与验证脚本共同进化，同时由冻结的评估器保证比较的公平。在这次本地 Paper2Poster 运行中，同一个 Codex 模型和同一条论文提示词生成了下面两张 LoRA 海报。 -

-
-
- Generation zero LoRA research poster with a generic dashboard-style layout -
- - Gen 0 · minimal 12-line Skill - Gen 0 · 12 行的最小 Skill - - Deterministic geometry gate failed: 14 text elements overflowed the SVG viewBox. - 未通过确定性几何门控：14 个文本元素溢出了 SVG viewBox。 -
-
-
- Generation two LoRA research poster with a paper-specific editorial layout and low-rank matrix visualization -
- - Gen 2 · evolved editorial Skill - Gen 2 · 进化后的编辑版 Skill - - Passed deterministic renderability and geometry gates; paper fidelity remained advisory reviewer feedback. - 通过了确定性可渲染与几何门控；论文忠实度作为建议性评审反馈保留。 -
-
-
-

- 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。 -

- -

- Benchmark results - 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。 -

-
- 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。 -

-
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
- Target agent - 目标智能体 - - Method - 方法 - TrainGateSealedOverall
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%)
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%)
GEPA58.0% → 68.0%
(+10.0%)		57.9% → 68.4%
(+10.5%)		70.0% → 75.0%
(+5.0%)		60.7% → 69.7%
(+9.0%)
CodexAHE58.0% → 74.0%
(+16.0%)		52.6% → 47.4%
(−5.2%)		65.0% → 70.0%
(+5.0%)		58.4% → 67.4%
(+9.0%)
Hyperagents58.0% → 72.0%
(+14.0%)		52.6% → 57.9%
(+5.3%)		65.0% → 75.0%
(+10.0%)		58.4% → 69.7%
(+11.3%)
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%)
GEPA58.0% → 58.0%
(+0.0%)		52.6% → 52.6%
(+0.0%)		65.0% → 65.0%
(+0.0%)		58.4% → 58.4%
(+0.0%)
-
- -

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 - TrainGateSealedOverall
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%)
Hyperagents30.0% → 38.0%
(+8.0%)		35.0% → 45.0%
(+10.0%)		18.5% → 37.0%
(+18.5%)		27.8% → 39.2%
(+11.4%)
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%)
GEPA30.0% → 32.0%
(+2.0%)		35.0% → 45.0%
(+10.0%)		18.5% → 25.9%
(+7.4%)		27.8% → 33.0%
(+5.2%)
CodexAHE30.0% → 36.0%
(+6.0%)		30.0% → 45.0%
(+15.0%)		7.4% → 14.8%
(+7.4%)		23.7% → 32.0%
(+8.3%)
Hyperagents30.0% → 36.0%
(+6.0%)		30.0% → 50.0%
(+20.0%)		7.4% → 48.1%
(+40.7%)		23.7% → 42.3%
(+18.6%)
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%)
GEPA30.0% → 36.0%
(+6.0%)		30.0% → 35.0%
(+5.0%)		7.4% → 14.8%
(+7.4%)		23.7% → 29.9%
(+6.2%)
-
- -

- Trustworthy by construction - 从构造上可信 -

-

- EvolveX separates evolvable policy from the mechanism that judges it: - EvolveX 将可进化的策略与评判它的机制分离： -

-
    -
  1. - The evaluator is frozen. Candidates cannot change the scoring contract. - 评估器被冻结。候选无法更改打分契约。 -
    2. -
  2. - Mutation is bounded. Each recipe declares which target and operator paths may change. - 变异有边界。每个 recipe 声明哪些目标与操作符路径可以改变。 -
    3. -
  3. - Evaluation is canonical. New generations are scored from clean candidate snapshots. - 评估是规范的。新的世代从干净的候选快照中打分。 -
    4. -
  4. - Evidence is durable. Reports recompute results from stamped archive.jsonl records and Git generation tags. - 证据是持久的。报告从盖章的 archive.jsonl 记录与 Git 世代标签中重新计算结果。 -
    5. -
-

- Operators run as subprocesses rather than being imported into the framework process. See the design guide for the complete ownership model and invariants. - 操作符以子进程方式运行，而非被导入框架进程。完整的所有权模型与不变量见设计指南。 -

- -

- Project status & roadmap - 项目状态与路线图 -

-

- EvolveX is an active prototype for research and controlled experimentation. The current focus is reliable experiment mechanics, local-first workflows, and composable strategies for different agent-evolution scenarios. - EvolveX 是一个面向研究与受控实验的活跃原型。当前重点是可靠的实验机制、本地优先的工作流，以及面向不同智能体进化场景的可组合策略。 -

- - -

- Documentation - 文档 -

-
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
- Document - 文档 - - Purpose - 用途 -
- Documentation site - 文档站 - - Installation, operation, concepts, guides, and reference. - 安装、操作、概念、指南与参考。 -
Quick start - Recipe launcher setup and configuration. - Recipe 启动器的安装与配置。 -
Design - System model, ownership boundaries, and invariants. - 系统模型、所有权边界与不变量。 -
Architecture - Enforced source-module map and line budgets. - 强制执行的源码模块映射与行数预算。 -
Recipes - Supported evolution strategies. - 支持的进化策略。 -
Contributing - Development setup and repository conventions. - 开发环境搭建与仓库规范。 -
-
- -

- License - 许可协议 -

-

- EvolveX is licensed under Apache-2.0. See NOTICE for required attributions. - EvolveX 以 Apache-2.0 协议开源，所需的署名信息见 NOTICE。 -

- - - - Back to Simple Agent Lab - 返回 Simple Agent Lab - -
-
- - +

RSIHub has moved to its new project page.

diff --git a/index.html b/index.html index e9046c7..7b5f7d3 100644 --- a/index.html +++ b/index.html @@ -9,19 +9,19 @@ content="D94IoHgseJNY5YrF_3tcZM9igQlPpZqWRlPSXJP3W_g" /> - Simple Agent Lab | EvolveX, Self-Evolving Agents, AI4AI & RSI + Simple Agent Lab | RSIHub, Self-Evolving Agents, AI4AI & RSI - + - +
New 最新 - EvolveX + RSIHub Build agents that improve and keep the evidence. A simple framework for agent evolution. 让智能体持续改进，并保留证据。一个简单的智能体进化框架。 - + Code - + Docs 文档 diff --git a/evolvex/assets/architecture.svg b/rsihub/assets/architecture.svg similarity index 99% rename from evolvex/assets/architecture.svg rename to rsihub/assets/architecture.svg index ada59bd..17c02b4 100644 --- a/evolvex/assets/architecture.svg +++ b/rsihub/assets/architecture.svg @@ -1,5 +1,5 @@ - EvolveX architecture + 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. + + + + + RSIHub | Self-Evolving Agent Framework by Simple Agent Lab + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+ + + Simple Agent Lab + + +
+ +
+
+

+ Framework + 开源框架 +

+

RSIHub

+

+ Build agents that improve — and keep the evidence. + 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。 +

+
+ + GitHub + + + + Documentation + 文档 + + + + Quick start + 快速开始 + + +
+ +
+ 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 +

+

+ RSIHub gives an agent a controlled way to improve itself. It runs candidates against a fixed evaluator, keeps the evidence for every generation, and carries verified improvements forward without letting candidate code rewrite the rules that score it. + RSIHub 为智能体提供一种受控的自我改进方式：候选在固定的评估器上运行，每一代都保留证据，经验证的改进被传递下去——同时不允许候选代码改写为其打分的规则。 +

+
+ + + + + + + + + + + + + + + +
+ For agent builders + 面向智能体开发者 + + For researchers + 面向研究者 + + Evidence built in + 内建证据 +
+ Improve prompts, skills, harnesses, and agent code in a reusable experiment workspace. + 在可复用的实验工作区中改进提示词、技能、harness 与智能体代码。 + + Compare evolution strategies under fixed evaluation and mutation boundaries. + 在固定的评估与变异边界下比较不同进化策略。 + + Connect every candidate to scores, artifacts, archive records, and Git lineage. + 将每个候选与得分、产物、归档记录和 Git 谱系关联起来。 +
+
+ +

+ How RSIHub works + RSIHub 如何工作 +

+

+ Every recipe composes the same loop: + 每个 recipe 都由同一个循环组成： +

+

select → evaluate → analyze → mutate → gate → record

+
+ RSIHub architecture: five built-in strategies and custom recipes compose a loop of select, rollout and evaluation, analyze, mutate, gate, and record. The target and selected operators occupy a declared mutable surface. The evaluator, runtime, surface check, and stamped evidence remain protected from candidate changes. +
+

+ 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。 +

+ +

+ What can evolve + 什么可以进化 +

+
+ + + + + + + + + + + + + + + + + + + + + + + + + +
+ Surface + 可进化面 + + Examples + 示例 + + Best fit + 适用场景 +
+ prompts and skills + 提示词与技能 + + system prompts, task skills, reusable instructions + 系统提示词、任务技能、可复用指令 + + policy and behavior improvement + 策略与行为改进 +
+ harnesses and target code + harness 与目标代码 + + tools, orchestration, agent implementation + 工具、编排、智能体实现 + + agent engineering + 智能体工程 +
+ selected evolution operators + 被选定的进化操作符 + + analysis or mutation policy chosen by a recipe + recipe 选定的分析或变异策略 + + controlled co-evolution + 受控的共同进化 +
+
+

+ Each recipe declares its mutable paths. Evaluators, archive stamps, and the vendored framework mechanism stay outside that surface. + 每个 recipe 声明自己的可变路径。评估器、归档盖章与内置的框架机制始终位于可变面之外。 +

+ +

Recipes

+
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+ Choose this when you want to… + 当你想要… + Recipe + Mutable surface + 可变面 +
+ improve one candidate from its current best parent + 从当前最优父代持续改进单个候选 + hill_climbtarget
+ evolve prompts and reusable agent skills + 进化提示词与可复用的智能体技能 + aevolveprompt and target skills
+ engineer the agent harness against evaluator feedback + 依据评估器反馈改进 agent harness + ahetarget
+ balance multiple objectives with minibatch validation + 通过 minibatch 验证平衡多个目标 + gepaprompt and task skill
+ co-evolve the target and selected evolution policy + 让目标与选定的进化策略共同进化 + hyperagentstarget and selected operators
+
+

+ See the recipe guide for each strategy’s workflow and configuration. + 每个策略的工作流与配置见 recipe 指南。 +

+ +

+ Skill evolution showcase + 技能进化 Showcase +

+

+ RSIHub can improve a Skill as a complete package: instructions, references, and validation scripts evolve together while a frozen evaluator keeps the comparison honest. In this local Paper2Poster run, the same Codex model and paper prompt produced both LoRA posters below. + RSIHub 可以把一个 Skill 作为完整的包来改进：指令、参考资料与验证脚本共同进化，同时由冻结的评估器保证比较的公平。在这次本地 Paper2Poster 运行中，同一个 Codex 模型和同一条论文提示词生成了下面两张 LoRA 海报。 +

+
+
+ Generation zero LoRA research poster with a generic dashboard-style layout +
+ + Gen 0 · minimal 12-line Skill + Gen 0 · 12 行的最小 Skill + + Deterministic geometry gate failed: 14 text elements overflowed the SVG viewBox. + 未通过确定性几何门控：14 个文本元素溢出了 SVG viewBox。 +
+
+
+ Generation two LoRA research poster with a paper-specific editorial layout and low-rank matrix visualization +
+ + Gen 2 · evolved editorial Skill + Gen 2 · 进化后的编辑版 Skill + + Passed deterministic renderability and geometry gates; paper fidelity remained advisory reviewer feedback. + 通过了确定性可渲染与几何门控；论文忠实度作为建议性评审反馈保留。 +
+
+
+

+ 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。 +

+ +

+ Benchmark results + 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。 +

+
+ 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。 +

+
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+ Target agent + 目标智能体 + + Method + 方法 + TrainGateSealedOverall
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%)
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%)
GEPA58.0% → 68.0%
(+10.0%)		57.9% → 68.4%
(+10.5%)		70.0% → 75.0%
(+5.0%)		60.7% → 69.7%
(+9.0%)
CodexAHE58.0% → 74.0%
(+16.0%)		52.6% → 47.4%
(−5.2%)		65.0% → 70.0%
(+5.0%)		58.4% → 67.4%
(+9.0%)
Hyperagents58.0% → 72.0%
(+14.0%)		52.6% → 57.9%
(+5.3%)		65.0% → 75.0%
(+10.0%)		58.4% → 69.7%
(+11.3%)
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%)
GEPA58.0% → 58.0%
(+0.0%)		52.6% → 52.6%
(+0.0%)		65.0% → 65.0%
(+0.0%)		58.4% → 58.4%
(+0.0%)
+
+ +

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 + TrainGateSealedOverall
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%)
Hyperagents30.0% → 38.0%
(+8.0%)		35.0% → 45.0%
(+10.0%)		18.5% → 37.0%
(+18.5%)		27.8% → 39.2%
(+11.4%)
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%)
GEPA30.0% → 32.0%
(+2.0%)		35.0% → 45.0%
(+10.0%)		18.5% → 25.9%
(+7.4%)		27.8% → 33.0%
(+5.2%)
CodexAHE30.0% → 36.0%
(+6.0%)		30.0% → 45.0%
(+15.0%)		7.4% → 14.8%
(+7.4%)		23.7% → 32.0%
(+8.3%)
Hyperagents30.0% → 36.0%
(+6.0%)		30.0% → 50.0%
(+20.0%)		7.4% → 48.1%
(+40.7%)		23.7% → 42.3%
(+18.6%)
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%)
GEPA30.0% → 36.0%
(+6.0%)		30.0% → 35.0%
(+5.0%)		7.4% → 14.8%
(+7.4%)		23.7% → 29.9%
(+6.2%)
+
+ +

+ Trustworthy by construction + 从构造上可信 +

+

+ RSIHub separates evolvable policy from the mechanism that judges it: + RSIHub 将可进化的策略与评判它的机制分离： +

+
    +
  1. + The evaluator is frozen. Candidates cannot change the scoring contract. + 评估器被冻结。候选无法更改打分契约。 +
    2. +
  2. + Mutation is bounded. Each recipe declares which target and operator paths may change. + 变异有边界。每个 recipe 声明哪些目标与操作符路径可以改变。 +
    3. +
  3. + Evaluation is canonical. New generations are scored from clean candidate snapshots. + 评估是规范的。新的世代从干净的候选快照中打分。 +
    4. +
  4. + Evidence is durable. Reports recompute results from stamped archive.jsonl records and Git generation tags. + 证据是持久的。报告从盖章的 archive.jsonl 记录与 Git 世代标签中重新计算结果。 +
    5. +
+

+ Operators run as subprocesses rather than being imported into the framework process. See the design guide for the complete ownership model and invariants. + 操作符以子进程方式运行，而非被导入框架进程。完整的所有权模型与不变量见设计指南。 +

+ +

+ Project status & roadmap + 项目状态与路线图 +

+

+ RSIHub is an active prototype for research and controlled experimentation. The current focus is reliable experiment mechanics, local-first workflows, and composable strategies for different agent-evolution scenarios. + RSIHub 是一个面向研究与受控实验的活跃原型。当前重点是可靠的实验机制、本地优先的工作流，以及面向不同智能体进化场景的可组合策略。 +

+
    +
  • + Scenario-oriented recipes: compose the current operator library into opinionated recipes for different agent-evolution use cases. + 面向场景的 recipes：将现有操作符库组合成针对不同智能体进化用例的成熟配方。 +
    • +
  • + Local-first workflows: make lightweight, Docker-free iteration a first-class path for trusted local agents, prompts, skills, and small features. + 本地优先的工作流：让轻量、无需 Docker 的迭代成为受信任的本地智能体、提示词、技能与小功能的一等路径。 +
    • +
  • + More method integrations: add evolution and search methods while preserving the shared evaluator, lineage, and evidence contracts. + 更多方法集成：在保持共享的评估器、谱系与证据契约的前提下，加入更多进化与搜索方法。 +
    • +
+ +

+ Documentation + 文档 +

+
+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
+ Document + 文档 + + Purpose + 用途 +
+ Documentation site + 文档站 + + Installation, operation, concepts, guides, and reference. + 安装、操作、概念、指南与参考。 +
Quick start + Recipe launcher setup and configuration. + Recipe 启动器的安装与配置。 +
Design + System model, ownership boundaries, and invariants. + 系统模型、所有权边界与不变量。 +
Architecture + Enforced source-module map and line budgets. + 强制执行的源码模块映射与行数预算。 +
Recipes + Supported evolution strategies. + 支持的进化策略。 +
Contributing + Development setup and repository conventions. + 开发环境搭建与仓库规范。 +
+
+ +

+ License + 许可协议 +

+

+ RSIHub is licensed under Apache-2.0. See NOTICE for required attributions. + RSIHub 以 Apache-2.0 协议开源，所需的署名信息见 NOTICE。 +

+ + + + Back to Simple Agent Lab + 返回 Simple Agent Lab + +
+
+ + + + diff --git a/sitemap.xml b/sitemap.xml index 185a6ea..9af610d 100644 --- a/sitemap.xml +++ b/sitemap.xml @@ -2,10 +2,10 @@ https://simpleagentlab.com/ - 2026-08-09 + 2026-08-12 - https://simpleagentlab.com/evolvex/ - 2026-08-09 + https://simpleagentlab.com/rsihub/ + 2026-08-12 diff --git a/tests/test_rsihub_migration.py b/tests/test_rsihub_migration.py new file mode 100644 index 0000000..84dfbab --- /dev/null +++ b/tests/test_rsihub_migration.py @@ -0,0 +1,142 @@ +from html.parser import HTMLParser +from pathlib import Path +import subprocess +import tempfile +import unittest + + +ROOT = Path(__file__).resolve().parents[1] +REPO_URL = "https://github.com/simple-agent-lab/RSIHub" +DOCS_URL = "https://simpleagentlab.com/RSIHub/" +OVERVIEW_URL = "https://simpleagentlab.com/rsihub/" + + +class LinkParser(HTMLParser): + def __init__(self): + super().__init__() + self.links = [] + self.canonicals = [] + self.refreshes = [] + + def handle_starttag(self, tag, attrs): + values = dict(attrs) + if tag == "a" and "href" in values: + self.links.append(values["href"]) + if tag == "link" and values.get("rel") == "canonical": + self.canonicals.append(values.get("href")) + if tag == "meta" and values.get("http-equiv", "").lower() == "refresh": + self.refreshes.append(values.get("content")) + + +def parse_html(relative_path): + parser = LinkParser() + parser.feed((ROOT / relative_path).read_text(encoding="utf-8")) + return parser + + +def read_tracked_text(path): + if not path.is_file(): + return None + try: + return path.read_text(encoding="utf-8") + except UnicodeDecodeError: + return None + + +class TrackedTextTests(unittest.TestCase): + def test_read_tracked_text_skips_missing_and_non_utf8_paths(self): + with tempfile.TemporaryDirectory() as directory: + root = Path(directory) + readable = root / "page.html" + binary = root / "font.woff2" + readable.write_text("RSIHub", encoding="utf-8") + binary.write_bytes(b"\xff\xfe\x00\x01") + + self.assertEqual(read_tracked_text(readable), "RSIHub") + self.assertIsNone(read_tracked_text(binary)) + self.assertIsNone(read_tracked_text(root / "deleted.html")) + self.assertIsNone(read_tracked_text(root)) + + +class RSIHubProjectPageTests(unittest.TestCase): + def test_canonical_project_page_uses_new_identity_and_urls(self): + page_path = ROOT / "rsihub/index.html" + self.assertTrue(page_path.is_file()) + text = page_path.read_text(encoding="utf-8") + parser = parse_html("rsihub/index.html") + self.assertIn("RSIHub", text) + self.assertNotIn("EvolveX", text) + self.assertNotIn("simple-agent-lab/EvolveX", text) + self.assertNotIn("simple-agent-lab.github.io/EvolveX", text) + self.assertEqual(parser.canonicals, [OVERVIEW_URL]) + self.assertIn(REPO_URL, parser.links) + self.assertIn(DOCS_URL, parser.links) + + def test_project_assets_moved_and_svg_titles_renamed(self): + assets = ROOT / "rsihub/assets" + expected = { + "architecture.svg", + "benchmark-results.svg", + "evolve-lineage.svg", + "paper-poster-lora-gen0.png", + "paper-poster-lora-gen2.png", + } + self.assertEqual({path.name for path in assets.iterdir()}, expected) + for svg in assets.glob("*.svg"): + self.assertNotIn("EvolveX", svg.read_text(encoding="utf-8")) + + +class RSIHubHomepageTests(unittest.TestCase): + def test_homepage_uses_rsihub_name_and_actions(self): + text = (ROOT / "index.html").read_text(encoding="utf-8") + parser = parse_html("index.html") + self.assertIn("RSIHub", text) + self.assertNotIn("EvolveX", text) + self.assertIn("./rsihub/", parser.links) + self.assertIn(REPO_URL, parser.links) + self.assertIn(DOCS_URL, parser.links) + + def test_homepage_structured_data_uses_canonical_rsihub_urls(self): + text = (ROOT / "index.html").read_text(encoding="utf-8") + self.assertIn(f'"@id": "{OVERVIEW_URL}#software"', text) + self.assertIn(f'"url": "{OVERVIEW_URL}"', text) + self.assertIn(f'"codeRepository": "{REPO_URL}"', text) + self.assertNotIn("simpleagentlab.com/evolvex/", text) + self.assertNotIn("simple-agent-lab/EvolveX", text) + self.assertNotIn("simple-agent-lab.github.io/EvolveX", text) + + +class RSIHubCompatibilityTests(unittest.TestCase): + def test_legacy_page_redirects_and_has_fallback(self): + parser = parse_html("evolvex/index.html") + self.assertEqual(parser.canonicals, [OVERVIEW_URL]) + self.assertIn("0; url=/rsihub/", parser.refreshes) + self.assertIn("/rsihub/", parser.links) + + def test_sitemap_contains_only_the_canonical_project_route(self): + text = (ROOT / "sitemap.xml").read_text(encoding="utf-8") + self.assertIn("https://simpleagentlab.com/rsihub/", text) + self.assertNotIn("https://simpleagentlab.com/evolvex/", text) + + def test_old_identity_is_absent_from_published_files(self): + allowed = { + Path("tests/test_rsihub_migration.py"), + } + forbidden = ("EvolveX",) + tracked = subprocess.check_output( + ["git", "ls-files", "-z"], cwd=ROOT + ).decode("utf-8").split("\0") + for tracked_path in tracked: + if not tracked_path: + continue + relative = Path(tracked_path) + path = ROOT / relative + if relative in allowed: + continue + if path.suffix.lower() in {".png", ".jpg", ".jpeg", ".gif", ".webp"}: + continue + text = read_tracked_text(path) + if text is None: + continue + for old_value in forbidden: + self.assertNotIn(old_value, text, f"{old_value!r} remains in {relative}")