Make Codex and Claude Code subscriptions go further: a local model router that gives routine agent work to economical models and keeps your strongest model for the hard parts.
Rust ai-agentsai-gatewayclaude-codecodexcoding-agentcost-optimizationdeepseekllm-gatewayllm-routerlocal-firstmodel-routingrusttask-routingtauri
docs: add bilingual HiRoute open-source introduction (#54)
澄潭b3721b043d37
diff --git a/apps/website/README.md b/apps/website/README.mdindex 359ee36d..b7915d60 100644--- a/apps/website/README.md+++ b/apps/website/README.md@@ -39,6 +39,8 @@ The repository-root `news/index.json` indexes bilingual articles in `news/`.Both GitHub and the website use those same Markdown sources. Add an entry with aunique slug, date, title, description, and file for each language; the article lists,language links and newest-article homepage entry follow the index automatically.+For an older introduction with no recorded publication date, omit `date`; it+appears after dated articles and leaves the latest-article homepage entry intact.An optional localized `shortTitle` keeps the homepage link concise; without one,it uses the full article title. Public article URLs remain under `/news/`.Article images live in `news/assets/` and are copied to ignored `public/news-assets/`diff --git a/apps/website/scripts/validate-build.mjs b/apps/website/scripts/validate-build.mjsindex 4bc11616..8f280a2d 100644--- a/apps/website/scripts/validate-build.mjs+++ b/apps/website/scripts/validate-build.mjs@@ -6,6 +6,7 @@ const dist = path.join(root, 'dist');const required = ['news/index.html', 'en/news/index.html','news/astra-qwen-smart-routing/index.html', 'en/news/astra-qwen-smart-routing/index.html',+ 'news/hiroute-open-source/index.html', 'en/news/hiroute-open-source/index.html','index.html', 'en/index.html', 'download/index.html', 'en/download/index.html','changelog/index.html', 'en/changelog/index.html','docs/index.html', 'en/docs/index.html',diff --git a/apps/website/src/components/NewsPage.astro b/apps/website/src/components/NewsPage.astroindex 1a6420f6..7dad9ada 100644--- a/apps/website/src/components/NewsPage.astro+++ b/apps/website/src/components/NewsPage.astro@@ -14,7 +14,7 @@ const body = post ? await renderNews(slug, language) : '';{post ? <><a class="text-link" href={newsRoute('', language)}>← {en ? 'All articles' : '全部文章'}</a><article class="news-article">- <header><time datetime={post.date}>{post.date}</time><h1>{title}</h1></header>+ <header>{post.date && <time datetime={post.date}>{post.date}</time>}<h1>{title}</h1></header><div class="news-prose" set:html={body} /><a class="text-link" href={`https://github.com/higress-group/HiRoute/blob/main/news/${post.files[language]}`}>{en ? 'Read the source on GitHub' : '在 GitHub 阅读原文'} ↗</a></article>@@ -22,7 +22,7 @@ const body = post ? await renderNews(slug, language) : '';<p class="eyebrow">{en ? 'From the HiRoute team' : '来自 HiRoute 团队'}</p><h1>{en ? 'Articles' : '文章'}</h1><p>{description}</p>- <div class="news-list">{posts.map(p => <article class="news-card"><time datetime={p.date}>{p.date}</time><h2><a href={newsRoute(p.slug, language)}>{p.title[language]}</a></h2><p>{p.description[language]}</p><a class="text-link" href={newsRoute(p.slug, language)}>{en ? 'Read the article' : '阅读全文'} →</a></article>)}</div>+ <div class="news-list">{posts.map(p => <article class="news-card">{p.date && <time datetime={p.date}>{p.date}</time>}<h2><a href={newsRoute(p.slug, language)}>{p.title[language]}</a></h2><p>{p.description[language]}</p><a class="text-link" href={newsRoute(p.slug, language)}>{en ? 'Read the article' : '阅读全文'} →</a></article>)}</div></>}</div></SiteLayout>diff --git a/apps/website/src/lib/news.mjs b/apps/website/src/lib/news.mjsindex ef19941d..f6ae6d18 100644--- a/apps/website/src/lib/news.mjs+++ b/apps/website/src/lib/news.mjs@@ -9,12 +9,13 @@ const seen = new Set();for (const post of posts) {if (!/^[a-z0-9]+(?:-[a-z0-9]+)*$/.test(post.slug) || seen.has(post.slug)) throw Error('invalid or duplicate news slug');seen.add(post.slug);- if (!/^\d{4}-\d{2}-\d{2}$/.test(post.date)) throw Error('invalid news date');+ if (post.date !== undefined && !/^\d{4}-\d{2}-\d{2}$/.test(post.date)) throw Error('invalid news date');for (const language of ['zh', 'en']) {if (!post.title[language] || !post.description[language] || !/^[A-Za-z0-9.-]+\.md$/.test(post.files[language])) throw Error('incomplete news translation');}}-posts.sort((a, b) => b.date.localeCompare(a.date) || a.slug.localeCompare(b.slug));+// Undated backfills follow dated news and do not replace the latest homepage article.+posts.sort((a, b) => (b.date ?? '').localeCompare(a.date ?? '') || a.slug.localeCompare(b.slug));export const newsRoute = (slug, language) => `${language === 'en' ? '/en' : ''}/news/${slug ? slug + '/' : ''}`;export function articleLink(href) {diff --git a/apps/website/src/styles/site.css b/apps/website/src/styles/site.cssindex 2639fba1..cb41ffa8 100644--- a/apps/website/src/styles/site.css+++ b/apps/website/src/styles/site.css@@ -484,10 +484,10 @@ html:lang(en) .section-title>p{flex:0 1 335px}.hero-article:hover{border-color:#aeb8e8;background:#eef0fb}.hero-article-label{flex:none;padding:3px 7px;border-radius:4px;background:#e8ecfc;color:#4255b8;font-size:11px;font-weight:600}.hero-article-copy{min-width:0;text-wrap:pretty}.hero-article-arrow{flex:none;margin-left:auto}-.news-shell{max-width:1024px;padding-top:52px;padding-bottom:80px}.news-shell h1{font-size:clamp(28px,4vw,44px);line-height:1.3;letter-spacing:-.025em}+.news-shell{max-width:1024px;padding-top:52px;padding-bottom:80px}.news-shell h1{font-size:clamp(28px,4vw,44px);line-height:1.3;letter-spacing:-.025em;text-wrap:balance}.news-article{max-width:860px;margin:24px auto}.news-article header{margin-bottom:32px}.news-shell time{font-size:14px;color:#606a7d}.news-prose{font-size:17px;line-height:1.85;overflow-wrap:anywhere}.news-prose h2{font-size:27px;line-height:1.4;margin:48px 0 18px}.news-prose p,.news-prose ul,.news-prose ol{margin:20px 0}.news-prose ul,.news-prose ol{padding-left:26px}.news-prose li{margin:8px 0}-.news-prose img{display:block;max-width:100%;height:auto;border-radius:14px;margin:30px auto}.news-prose a{color:#4055ce;text-decoration:underline;text-underline-offset:3px}+.news-prose img{display:block;width:100%;max-width:100%;height:auto;border-radius:14px;margin:30px auto}.news-prose a{color:#4055ce;text-decoration:underline;text-underline-offset:3px}.news-prose table{display:block;overflow-x:auto;border-collapse:collapse;width:100%;font-size:14px;line-height:1.6;margin:28px 0}.news-prose th,.news-prose td{border:1px solid #dce1eb;padding:12px;min-width:100px;text-align:left}.news-prose th{background:#eef1f8}.news-prose pre{overflow-x:auto;background:#172034;color:#e6edf8;padding:20px;border-radius:12px;font-size:14px;line-height:1.65}.news-prose code{font-size:.9em}.news-list{display:grid;gap:24px;margin:36px 0}.news-card{border:1px solid #dce1eb;border-radius:18px;padding:28px;background:white}.news-card h2{font-size:25px;line-height:1.4}.news-card p{line-height:1.75;color:#535d72}@media(max-width:700px){.news-shell{padding-top:32px}.news-prose{font-size:16px}.news-card{padding:20px}.news-prose h2{font-size:23px}.news-prose th,.news-prose td{padding:8px}}diff --git a/apps/website/tests/news.test.mjs b/apps/website/tests/news.test.mjsindex 7b092094..4f00e1c3 100644--- a/apps/website/tests/news.test.mjs+++ b/apps/website/tests/news.test.mjs@@ -28,3 +28,11 @@ test('routing case keeps numerical evidence and the public reproduction link', aassert.match(html, /github\.com\/higress-group\/HiRoute\/blob\/main\/experiments\//);}});++test('the backfilled open-source introduction leaves dated news ahead of it', () => {+ const introduction = posts.findIndex(post => post.slug === 'hiroute-open-source');+ const caseStudy = posts.findIndex(post => post.slug === 'astra-qwen-smart-routing');+ assert.ok(caseStudy >= 0 && introduction > caseStudy);+ assert.equal(posts[introduction].date, undefined);+ assert.ok(posts[0].date, 'the homepage latest-article entry must remain dated news');+});diff --git a/news/ILLUSTRATIONS.md b/news/ILLUSTRATIONS.mdindex bc862cc2..ffbc7d2b 100644--- a/news/ILLUSTRATIONS.md+++ b/news/ILLUSTRATIONS.md@@ -147,3 +147,55 @@ those responsibilities rather than prescribing internal process or crate boundarThe article describes the Higress engineering lineage separately from HiRoute's Rustimplementation. The existing product overview remains in the introduction; thehistorical prompt file also retains an earlier, unexecuted redraw brief.++## Open-source introduction++The [Chinese](hiroute-open-source.zh-CN.md) and [English](hiroute-open-source.en.md)+introduction preserve the reviewed launch article's six chapters. It is an undated+backfill, rather than a new launch announcement replacing the homepage case study.++- The new [Chinese routing overview](assets/launch-product-routing-zh.png) and+ [local architecture](assets/launch-local-architecture-zh.png) are explanatory+ image-generation illustrations from the reviewed launch draft. The matching+ [English overview](assets/launch-product-routing-en.svg) and+ [architecture](assets/launch-local-architecture-en.svg) are editable SVGs with+ the same independent routing layers, local boundary and call directions.+ Decision services are external; CPA connects subscriptions. Neither diagram+ depicts an embedded Higress/Envoy dependency.+- The [ecosystem figure](assets/launch-agent-ecosystem.png) comes from the+ user-supplied AGNTCon + MCPCon China keynote. Its original labels are English,+ so both article versions use the same image. The ecosystem is background+ context, not a claim that HiRoute integrates every depicted project.+- Context continuity has [Chinese](assets/launch-context-continuity-zh.svg) and+ [English](assets/launch-context-continuity-en.svg) vector versions. A rebuilt+ complete prefix differs even if the same model continues; an unchanged common+ prefix may remain reusable. New user requests are reassessed, and different+ models do not share KV Cache. These are mechanism illustrations.+- The new native stage-performance captures+ ([Chinese](assets/launch-stage-competence-zh.png),+ [English](assets/launch-stage-competence-en.png)) reuse the current homepage's+ exact localized captures. They show constructed documentation data for Qwen+ Flash and Qwen Max, not new benchmark measurements; see the+ [original capture provenance](../decision-extensions/assets/README.md).+- Research cost, unattended handoffs and historical HTTPX observation reuse the+ existing case study's localized assets and original evidence. Their numerical+ values, stage counts and unrated final stage remain unchanged.++The three setup steps use actual native macOS product captures, with matching+language for each article:++| Step | Chinese | English |+| --- | --- | --- |+| Add models | [Capture](assets/launch-model-connections-zh.png) | [Capture](assets/launch-model-connections-en.png) |+| Smart saving | [Capture](assets/launch-smart-saving-zh.png) | [Capture](assets/launch-smart-saving-en.png) |+| Agent connection | [Capture](assets/launch-agent-connections-zh.png) | [Capture](assets/launch-agent-connections-en.png) |++The Chinese smart-saving capture was made on 2026-10-08; the other setup captures+were made on 2026-10-09. Both smart-saving views show an unsaved Qwen 3.8 Flash+economy group and Qwen 3.8 Max primary group. No plan was saved, enabled or published+for the captures. Agent views show the connection controls, without changing agent+configuration. The English interface was selected through the product's language+setting, then restored to its original setting. Native JPEG captures were encoded+as PNG with the decoded pixels and dimensions verified unchanged. Product text+was not translated by retouching screenshots. These setup pictures illustrate the+real configuration flow; they are not inference or acceptance evidence.diff --git a/news/README.md b/news/README.mdindex 344ef455..8e4c7d11 100644--- a/news/README.md+++ b/news/README.md@@ -3,5 +3,8 @@| Date | English | 简体中文 || --- | --- | --- || 2026-10-04 | [Smart routing: Astra × Qwen](2026-10-04-astra-qwen.en.md) | [智能路由省钱实战](2026-10-04-astra-qwen.zh-CN.md) |+| — | [HiRoute is open source](hiroute-open-source.en.md) | [HiRoute 正式开源](hiroute-open-source.zh-CN.md) |Articles here are the canonical source for the bilingual website Articles pages. Add a dated Markdown pair and an entry in `index.json` for future posts. A localized `shortTitle` is optional for the compact homepage link.++Omit `date` for a backfilled introduction whose original publication date is not recorded. It follows dated news in the article list and does not replace the latest dated article on the homepage. The open-source introduction is such a backfill; the existing Astra × Qwen case study remains the homepage entry.diff --git a/news/assets/launch-agent-connections-en.png b/news/assets/launch-agent-connections-en.pngnew file mode 100644index 00000000..8c8d996bBinary files /dev/null and b/news/assets/launch-agent-connections-en.png differdiff --git a/news/assets/launch-agent-connections-zh.png b/news/assets/launch-agent-connections-zh.pngnew file mode 100644index 00000000..943da490Binary files /dev/null and b/news/assets/launch-agent-connections-zh.png differdiff --git a/news/assets/launch-agent-ecosystem.png b/news/assets/launch-agent-ecosystem.pngnew file mode 100644index 00000000..b824739aBinary files /dev/null and b/news/assets/launch-agent-ecosystem.png differdiff --git a/news/assets/launch-context-continuity-en.svg b/news/assets/launch-context-continuity-en.svgnew file mode 100644index 00000000..e4005252--- /dev/null+++ b/news/assets/launch-context-continuity-en.svg@@ -0,0 +1 @@+<svg xmlns="http://www.w3.org/2000/svg" width="1600" height="1080" viewBox="0 0 1600 1080" role="img"><title>Context continuity and KV Cache reuse</title><defs><marker id="arrow" markerWidth="8" markerHeight="8" refX="7" refY="4" orient="auto"><path d="M 0 0 L 8 4 L 0 8 Z" fill="#AAA1BE" /></marker></defs><g font-family="Arial,Helvetica,sans-serif"><rect x="0" y="0" width="1600" height="1080" rx="0" fill="#FFFFFF" stroke="none" stroke-width="2" /><rect x="56" y="38" width="505" height="47" rx="23" fill="#F1ECFF" stroke="none" stroke-width="2" /><text x="79" y="70" font-size="25" fill="#6942D8" font-weight="600">HiRoute · Context continuity</text><text x="56" y="148" font-size="44" fill="#15203A" font-weight="700">Context continuity: smarter routing, high KV Cache hit rates</text><text x="56" y="195" font-size="28" fill="#647086" font-weight="400">Keep one model while context stays continuous; reassess when history is rebuilt.</text><rect x="56" y="241" width="469" height="429" rx="22" fill="#F8F6FE" stroke="#E2D9F8" stroke-width="2" /><text x="81" y="290" font-size="30" fill="#5834BB" font-weight="700">01 Before rebuild · Model A</text><text x="81" y="332" font-size="25" fill="#69738A" font-weight="400">Tool results extend this turn</text><rect x="82" y="374" width="165" height="53" rx="9" fill="#7754D7" stroke="none" stroke-width="2" /><text x="91" y="409" font-size="21" fill="#FFFFFF" font-weight="500">Prior context</text><rect x="253" y="374" width="65" height="53" rx="9" fill="#E8DEFF" stroke="none" stroke-width="2" /><text x="262" y="409" font-size="21" fill="#5F40AB" font-weight="500">Input</text><rect x="82" y="453" width="165" height="53" rx="9" fill="#7754D7" stroke="none" stroke-width="2" /><text x="91" y="488" font-size="21" fill="#FFFFFF" font-weight="500">Prior context</text><rect x="253" y="453" width="65" height="53" rx="9" fill="#7754D7" stroke="none" stroke-width="2" /><text x="262" y="488" font-size="21" fill="#FFFFFF" font-weight="500">Input</text><rect x="324" y="453" width="80" height="53" rx="9" fill="#E8DEFF" stroke="none" stroke-width="2" /><text x="333" y="488" font-size="21" fill="#5F40AB" font-weight="500">Tool 1</text><rect x="82" y="532" width="165" height="53" rx="9" fill="#7754D7" stroke="none" stroke-width="2" /><text x="91" y="567" font-size="21" fill="#FFFFFF" font-weight="500">Prior context</text><rect x="253" y="532" width="65" height="53" rx="9" fill="#7754D7" stroke="none" stroke-width="2" /><text x="262" y="567" font-size="21" fill="#FFFFFF" font-weight="500">Input</text><rect x="324" y="532" width="80" height="53" rx="9" fill="#7754D7" stroke="none" stroke-width="2" /><text x="333" y="567" font-size="21" fill="#FFFFFF" font-weight="500">Tool 1</text><rect x="410" y="532" width="80" height="53" rx="9" fill="#E8DEFF" stroke="none" stroke-width="2" /><text x="419" y="567" font-size="21" fill="#5F40AB" font-weight="500">Tool 2</text><text x="82" y="626" font-size="24" fill="#5834BB" font-weight="600">Stable model, growing cache reuse</text><path d="M 535 451 L 571 451" fill="none" stroke="#AAA1BE" stroke-width="3" marker-end="url(#arrow)" /><rect x="590" y="241" width="418" height="429" rx="22" fill="#FAFBFD" stroke="#DFE4EE" stroke-width="2" /><text x="615" y="290" font-size="30" fill="#25334D" font-weight="700">02 Context rebuilt</text><text x="615" y="332" font-size="22" fill="#69738A" font-weight="400">Compacted history breaks continuity</text><rect x="615" y="369" width="367" height="58" rx="10" fill="#ECEEF4" stroke="none" stroke-width="2" /><text x="637" y="407" font-size="27" fill="#58647A" font-weight="600">Old prefix → New prefix</text><text x="615" y="472" font-size="25" fill="#37435C" font-weight="600">The full old prefix breaks</text><text x="615" y="514" font-size="25" fill="#69738A" font-weight="400">Even if model A continues</text><rect x="615" y="552" width="367" height="80" rx="14" fill="#F0EAFE" stroke="none" stroke-width="2" /><text x="636" y="585" font-size="27" fill="#6542C3" font-weight="600">Reassess which model</text><text x="636" y="619" font-size="27" fill="#6542C3" font-weight="600">should take the next stage</text><path d="M 1018 451 L 1054 451" fill="none" stroke="#AAA1BE" stroke-width="3" marker-end="url(#arrow)" /><rect x="1073" y="241" width="471" height="429" rx="22" fill="#F0FAF7" stroke="#CDE9DF" stroke-width="2" /><text x="1098" y="290" font-size="26" fill="#217D6B" font-weight="700">03 After rebuild · New prefix</text><rect x="1098" y="320" width="178" height="49" rx="22" fill="#E1F2EB" stroke="none" stroke-width="2" /><text x="1128" y="354" font-size="27" fill="#217D6B" font-weight="600">Continue A</text><text x="1291" y="354" font-size="24" fill="#7C8E87" font-weight="400">or</text><rect x="1331" y="320" width="181" height="49" rx="22" fill="#E1F2EB" stroke="none" stroke-width="2" /><text x="1361" y="354" font-size="27" fill="#217D6B" font-weight="600">Choose B</text><text x="1098" y="418" font-size="23" fill="#3C6F62" font-weight="600">Build cache hits on the new prefix</text><rect x="1098" y="454" width="255" height="53" rx="9" fill="#298F7A" stroke="none" stroke-width="2" /><text x="1118" y="489" font-size="24" fill="#FFFFFF" font-weight="500">New context</text><rect x="1098" y="530" width="255" height="53" rx="9" fill="#298F7A" stroke="none" stroke-width="2" /><text x="1118" y="565" font-size="24" fill="#FFFFFF" font-weight="500">New context</text><rect x="1360" y="530" width="153" height="53" rx="9" fill="#D5EEE6" stroke="none" stroke-width="2" /><text x="1374" y="565" font-size="24" fill="#217D6B" font-weight="500">New input</text><text x="1098" y="627" font-size="23" fill="#217D6B" font-weight="600">Keep this model; reuse its cache</text><rect x="56" y="707" width="1488" height="87" rx="18" fill="#F5F2FD" stroke="none" stroke-width="2" /><text x="86" y="763" font-size="31" fill="#5735B7" font-weight="600">Continuous prefixes keep cache hits high; rebuilt prefixes allow a new choice.</text><rect x="56" y="828" width="1488" height="125" rx="18" fill="#FAFBFD" stroke="#E5E9F0" stroke-width="2" /><text x="82" y="873" font-size="28" fill="#33425A" font-weight="600">Another decision point: a new user request</text><text x="82" y="918" font-size="23" fill="#69738A" font-weight="400">Each new user request is reassessed. Appending a message can still preserve the old prefix on the same model.</text><text x="56" y="1001" font-size="22" fill="#7B8495" font-weight="400">An unchanged common prefix may still be reused on the same model; this does not imply a total cache reset.</text><text x="56" y="1040" font-size="22" fill="#7B8495" font-weight="400">Different models do not share KV Cache. Cache hits also depend on the connected model service.</text></g></svg>diff --git a/news/assets/launch-context-continuity-zh.svg b/news/assets/launch-context-continuity-zh.svgnew file mode 100644index 00000000..3b1a84d0--- /dev/null+++ b/news/assets/launch-context-continuity-zh.svg@@ -0,0 +1 @@+<svg xmlns="http://www.w3.org/2000/svg" width="1600" height="1080" viewBox="0 0 1600 1080"><defs><marker id="arrow" markerWidth="8" markerHeight="8" refX="7" refY="4" orient="auto"><path d="M 0 0 L 8 4 L 0 8 Z" fill="#AAA1BE"/></marker></defs><g font-family="PingFang SC,Microsoft YaHei,Arial,sans-serif"><rect x="0" y="0" width="1600" height="1080" rx="0" fill="#FFFFFF" stroke="none" stroke-width="2"/><rect x="56" y="38" width="505" height="47" rx="23" fill="#F1ECFF" stroke="none" stroke-width="2"/><text x="79" y="70" font-size="25" fill="#6942D8" font-weight="600">HiRoute · 上下文保持 ContextHold</text><text x="56" y="148" font-size="46" fill="#15203A" font-weight="700">上下文保持:兼顾智能换模与 KV Cache 高命中</text><text x="56" y="195" font-size="28" fill="#647086" font-weight="400">同一段连续上下文内稳住模型;上下文重建后,再按需选择。</text><rect x="56" y="241" width="469" height="429" rx="22" fill="#F8F6FE" stroke="#E2D9F8" stroke-width="2"/><text x="81" y="290" font-size="30" fill="#5834BB" font-weight="700">01 重建前 · 模型 A</text><text x="81" y="332" font-size="25" fill="#69738A" font-weight="400">同轮工具调用,历史不断追加</text><rect x="82" y="374" width="165" height="53" rx="9" fill="#7754D7" stroke="none" stroke-width="2"/><text x="91" y="409" font-size="21" fill="#FFFFFF" font-weight="500">原有上下文</text><rect x="253" y="374" width="65" height="53" rx="9" fill="#E8DEFF" stroke="none" stroke-width="2"/><text x="262" y="409" font-size="21" fill="#5F40AB" font-weight="500">输入</text><rect x="82" y="453" width="165" height="53" rx="9" fill="#7754D7" stroke="none" stroke-width="2"/><text x="91" y="488" font-size="21" fill="#FFFFFF" font-weight="500">原有上下文</text><rect x="253" y="453" width="65" height="53" rx="9" fill="#7754D7" stroke="none" stroke-width="2"/><text x="262" y="488" font-size="21" fill="#FFFFFF" font-weight="500">输入</text><rect x="324" y="453" width="80" height="53" rx="9" fill="#E8DEFF" stroke="none" stroke-width="2"/><text x="333" y="488" font-size="21" fill="#5F40AB" font-weight="500">工具①</text><rect x="82" y="532" width="165" height="53" rx="9" fill="#7754D7" stroke="none" stroke-width="2"/><text x="91" y="567" font-size="21" fill="#FFFFFF" font-weight="500">原有上下文</text><rect x="253" y="532" width="65" height="53" rx="9" fill="#7754D7" stroke="none" stroke-width="2"/><text x="262" y="567" font-size="21" fill="#FFFFFF" font-weight="500">输入</text><rect x="324" y="532" width="80" height="53" rx="9" fill="#7754D7" stroke="none" stroke-width="2"/><text x="333" y="567" font-size="21" fill="#FFFFFF" font-weight="500">工具①</text><rect x="410" y="532" width="80" height="53" rx="9" fill="#E8DEFF" stroke="none" stroke-width="2"/><text x="419" y="567" font-size="21" fill="#5F40AB" font-weight="500">工具②</text><text x="82" y="626" font-size="24" fill="#5834BB" font-weight="600">稳定模型,持续复用 KV Cache</text><path d="M 535 451 L 571 451" fill="none" stroke="#AAA1BE" stroke-width="3" marker-end="url(#arrow)"/><rect x="590" y="241" width="418" height="429" rx="22" fill="#FAFBFD" stroke="#DFE4EE" stroke-width="2"/><text x="615" y="290" font-size="30" fill="#25334D" font-weight="700">02 上下文重建边界</text><text x="615" y="332" font-size="25" fill="#69738A" font-weight="400">压缩或重建历史后无法续接</text><rect x="615" y="369" width="367" height="58" rx="10" fill="#ECEEF4" stroke="none" stroke-width="2"/><text x="637" y="407" font-size="27" fill="#58647A" font-weight="600">旧前缀 → 新前缀</text><text x="615" y="472" font-size="25" fill="#37435C" font-weight="600">旧的完整前缀无法继续复用</text><text x="615" y="514" font-size="25" fill="#69738A" font-weight="400">即使继续同一模型也一样</text><rect x="615" y="552" width="367" height="80" rx="14" fill="#F0EAFE" stroke="none" stroke-width="2"/><text x="636" y="585" font-size="27" fill="#6542C3" font-weight="600">此时,重新判断</text><text x="636" y="619" font-size="27" fill="#6542C3" font-weight="600">接下来适合的模型</text><path d="M 1018 451 L 1054 451" fill="none" stroke="#AAA1BE" stroke-width="3" marker-end="url(#arrow)"/><rect x="1073" y="241" width="471" height="429" rx="22" fill="#F0FAF7" stroke="#CDE9DF" stroke-width="2"/><text x="1098" y="290" font-size="29" fill="#217D6B" font-weight="700">03 重建后 · 新的前缀</text><rect x="1098" y="320" width="178" height="49" rx="22" fill="#E1F2EB" stroke="none" stroke-width="2"/><text x="1128" y="354" font-size="27" fill="#217D6B" font-weight="600">继续 A</text><text x="1291" y="354" font-size="24" fill="#7C8E87" font-weight="400">或</text><rect x="1331" y="320" width="181" height="49" rx="22" fill="#E1F2EB" stroke="none" stroke-width="2"/><text x="1361" y="354" font-size="27" fill="#217D6B" font-weight="600">选择 B</text><text x="1098" y="418" font-size="23" fill="#3C6F62" font-weight="600">沿着新前缀,继续积累缓存命中</text><rect x="1098" y="454" width="255" height="53" rx="9" fill="#298F7A" stroke="none" stroke-width="2"/><text x="1118" y="489" font-size="24" fill="#FFFFFF" font-weight="500">新的上下文</text><rect x="1098" y="530" width="255" height="53" rx="9" fill="#298F7A" stroke="none" stroke-width="2"/><text x="1118" y="565" font-size="24" fill="#FFFFFF" font-weight="500">新的上下文</text><rect x="1360" y="530" width="153" height="53" rx="9" fill="#D5EEE6" stroke="none" stroke-width="2"/><text x="1374" y="565" font-size="24" fill="#217D6B" font-weight="500">新增内容</text><text x="1098" y="627" font-size="23" fill="#217D6B" font-weight="600">随后保持模型,持续复用缓存</text><rect x="56" y="707" width="1488" height="87" rx="18" fill="#F5F2FD" stroke="none" stroke-width="2"/><text x="86" y="763" font-size="34" fill="#5735B7" font-weight="600">前缀连续时保持高命中,前缀重建时按需换模型。</text><rect x="56" y="828" width="1488" height="125" rx="18" fill="#FAFBFD" stroke="#E5E9F0" stroke-width="2"/><text x="82" y="873" font-size="28" fill="#33425A" font-weight="600">另一类决策点:新用户请求</text><text x="82" y="918" font-size="25" fill="#69738A" font-weight="400">新用户请求也会重新决策;仅追加新消息时,继续原模型仍可能复用旧前缀。</text><text x="56" y="1001" font-size="24" fill="#7B8495" font-weight="400">同模型仍可能复用未变化的公共前缀,图示不代表所有缓存都失效。</text><text x="56" y="1040" font-size="24" fill="#7B8495" font-weight="400">不同模型不共享 KV Cache,实际命中取决于模型服务与配置。</text></g></svg>\ No newline at end of filediff --git a/news/assets/launch-local-architecture-en.svg b/news/assets/launch-local-architecture-en.svgnew file mode 100644index 00000000..b7769d48--- /dev/null+++ b/news/assets/launch-local-architecture-en.svg@@ -0,0 +1 @@+<svg xmlns="http://www.w3.org/2000/svg" width="1600" height="1000" viewBox="0 0 1600 1000" role="img"><defs><marker id="arrow" markerWidth="8" markerHeight="8" refX="7" refY="4" orient="auto-start-reverse"><path d="M 0 0 L 8 4 L 0 8 Z" fill="#7650D4"/></marker></defs><g font-family="Arial,Helvetica,sans-serif"><title>HiRoute: a local routing and collaboration engine</title><desc>Desktop and CLI configure the local service and Rust gateway. Agents call the gateway, which connects to decision services, model APIs, and subscriptions via CPA; execution records stay local.</desc><rect x="0" y="0" width="1600" height="1000" rx="0" fill="#fff" stroke="none" stroke-width="2"/><text x="800" y="95" font-size="46" fill="#142341" font-weight="700" text-anchor="middle">HiRoute: a local routing and collaboration engine</text><rect x="50" y="172" width="1090" height="670" rx="26" fill="#F7F4FE" stroke="#D9CFF1" stroke-width="2"/><text x="85" y="219" font-size="24" fill="#7350C5" font-weight="700" text-anchor="start">YOUR COMPUTER</text><path d="M 440 300 H 530" fill="none" stroke="#7650D4" stroke-width="3" stroke-linejoin="round" marker-end="url(#arrow)"/><path d="M 710 355 V 435" fill="none" stroke="#7650D4" stroke-width="3" stroke-linejoin="round" marker-end="url(#arrow)"/><text x="727" y="402" font-size="20" fill="#7B8294" font-weight="400" text-anchor="start">Configuration</text><path d="M 410 513 H 500" fill="none" stroke="#7650D4" stroke-width="3" stroke-linejoin="round" marker-start="url(#arrow)" marker-end="url(#arrow)"/><text x="285" y="420" font-size="22" fill="#7350C5" font-weight="400" text-anchor="start">Requests / responses</text><path d="M 965 513 H 1190" fill="none" stroke="#7650D4" stroke-width="3" stroke-linejoin="round" marker-start="url(#arrow)" marker-end="url(#arrow)"/><path d="M 915 435 V 394 H 1167 V 300 H 1190" fill="none" stroke="#7650D4" stroke-width="3" stroke-linejoin="round" marker-start="url(#arrow)" marker-end="url(#arrow)"/><path d="M 870 610 V 641 H 1037 V 670" fill="none" stroke="#7650D4" stroke-width="3" stroke-linejoin="round" marker-end="url(#arrow)"/><path d="M 1125 736 H 1190" fill="none" stroke="#7650D4" stroke-width="3" stroke-linejoin="round" marker-end="url(#arrow)"/><path d="M 710 610 V 702" fill="none" stroke="#7650D4" stroke-width="3" stroke-linejoin="round" marker-end="url(#arrow)"/><rect x="90" y="245" width="350" height="110" rx="18" fill="#fff" stroke="#DED6F1" stroke-width="2"/><text x="265.0" y="290" font-size="28" fill="#142341" font-weight="700" text-anchor="middle">Desktop / CLI</text><text x="265.0" y="323" font-size="21" fill="#64718A" font-weight="400" text-anchor="middle">Connect sources · Publish plans</text><rect x="530" y="245" width="360" height="110" rx="18" fill="#fff" stroke="#DED6F1" stroke-width="2"/><text x="710.0" y="290" font-size="28" fill="#142341" font-weight="700" text-anchor="middle">Local service</text><text x="710.0" y="323" font-size="23" fill="#64718A" font-weight="400" text-anchor="middle">Plans · Permissions · Tasks</text><rect x="90" y="442" width="320" height="145" rx="18" fill="#fff" stroke="#DED6F1" stroke-width="2"/><text x="250.0" y="487" font-size="28" fill="#142341" font-weight="700" text-anchor="middle">Existing agent</text><text x="250.0" y="520" font-size="23" fill="#64718A" font-weight="400" text-anchor="middle">Executes your work</text><rect x="500" y="435" width="465" height="175" rx="18" fill="#7650D4" stroke="none" stroke-width="2"/><text x="732.5" y="480" font-size="28" fill="#fff" font-weight="700" text-anchor="middle">HiRoute Gateway · Rust</text><text x="732.5" y="513" font-size="21" fill="#F0EAFD" font-weight="400" text-anchor="middle">Decision adapters · Capability checks</text><text x="732.5" y="544" font-size="23" fill="#F0EAFD" font-weight="400" text-anchor="middle">Ordered candidate failover</text><rect x="500" y="702" width="360" height="112" rx="18" fill="#fff" stroke="#DED6F1" stroke-width="2"/><text x="680.0" y="747" font-size="28" fill="#142341" font-weight="700" text-anchor="middle">Local execution records</text><text x="680.0" y="780" font-size="21" fill="#64718A" font-weight="400" text-anchor="middle">Sessions · Usage · Stage scores</text><rect x="950" y="670" width="175" height="132" rx="18" fill="#fff" stroke="#DED6F1" stroke-width="2"/><text x="1037.5" y="715" font-size="30" fill="#142341" font-weight="700" text-anchor="middle">CPA</text><text x="1037.5" y="748" font-size="23" fill="#64718A" font-weight="400" text-anchor="middle">Subscription</text><text x="1037.5" y="779" font-size="23" fill="#64718A" font-weight="400" text-anchor="middle">access</text><rect x="1190" y="232" width="360" height="136" rx="18" fill="#fff" stroke="#DED6F1" stroke-width="2"/><text x="1370.0" y="277" font-size="28" fill="#142341" font-weight="700" text-anchor="middle">Decision models &</text><text x="1370.0" y="310" font-size="28" fill="#64718A" font-weight="700" text-anchor="middle">extensions</text><text x="1370.0" y="341" font-size="21" fill="#64718A" font-weight="400" text-anchor="middle">Jev · Bailian · Custom services</text><rect x="1190" y="442" width="360" height="145" rx="18" fill="#fff" stroke="#DED6F1" stroke-width="2"/><text x="1370.0" y="523.5" font-size="28" fill="#142341" font-weight="700" text-anchor="middle">Model APIs</text><rect x="1190" y="667" width="360" height="145" rx="18" fill="#fff" stroke="#DED6F1" stroke-width="2"/><text x="1370.0" y="748.5" font-size="28" fill="#142341" font-weight="700" text-anchor="middle">Subscription models</text><text x="800" y="916" font-size="32" fill="#7650D4" font-weight="600" text-anchor="middle">Configuration and records stay local; connected services perform inference.</text><text x="800" y="963" font-size="25" fill="#64718A" font-weight="400" text-anchor="middle">Local model services can also be connected.</text></g></svg>diff --git a/news/assets/launch-local-architecture-zh.png b/news/assets/launch-local-architecture-zh.pngnew file mode 100644index 00000000..d5df8400Binary files /dev/null and b/news/assets/launch-local-architecture-zh.png differdiff --git a/news/assets/launch-model-connections-en.png b/news/assets/launch-model-connections-en.pngnew file mode 100644index 00000000..9f81dd06Binary files /dev/null and b/news/assets/launch-model-connections-en.png differdiff --git a/news/assets/launch-model-connections-zh.png b/news/assets/launch-model-connections-zh.pngnew file mode 100644index 00000000..49a7b0e4Binary files /dev/null and b/news/assets/launch-model-connections-zh.png differdiff --git a/news/assets/launch-product-routing-en.svg b/news/assets/launch-product-routing-en.svgnew file mode 100644index 00000000..7db22b59--- /dev/null+++ b/news/assets/launch-product-routing-en.svg@@ -0,0 +1 @@+<svg xmlns="http://www.w3.org/2000/svg" width="1600" height="960" viewBox="0 0 1600 960" role="img"><defs><marker id="arrow" markerWidth="8" markerHeight="8" refX="7" refY="4" orient="auto-start-reverse"><path d="M 0 0 L 8 4 L 0 8 Z" fill="#7650D4"/></marker></defs><g font-family="Arial,Helvetica,sans-serif"><title>Choose the agent for the task, the model for each stage</title><desc>Independent task routing and model routing, using existing subscriptions, APIs, free quotas, and local models.</desc><rect x="0" y="0" width="1600" height="960" rx="0" fill="#fff" stroke="none" stroke-width="2"/><text x="800" y="91" font-size="46" fill="#142341" font-weight="700" text-anchor="middle">Choose the agent for the task, the model for each stage</text><path d="M 415 225 V 197 Q 415 182 430 182 H 1170 Q 1185 182 1185 197 V 225" fill="none" stroke="#7650D4" stroke-width="3" stroke-linejoin="round"/><rect x="575" y="145" width="450" height="72" rx="24" fill="#7650D4" stroke="none" stroke-width="2"/><text x="800" y="192" font-size="35" fill="#fff" font-weight="700" text-anchor="middle">HiRoute routing plans</text><rect x="50" y="252" width="730" height="431" rx="22" fill="#FBFAFE" stroke="#DAD0F7" stroke-width="2"/><text x="415" y="323" font-size="38" fill="#142341" font-weight="700" text-anchor="middle">Who does the work?</text><rect x="820" y="252" width="730" height="431" rx="22" fill="#FBFAFE" stroke="#DAD0F7" stroke-width="2"/><text x="1185" y="323" font-size="38" fill="#142341" font-weight="700" text-anchor="middle">What does this stage need?</text><rect x="95" y="383" width="190" height="184" rx="16" fill="#F1ECFF" stroke="#E0D5FC" stroke-width="2"/><rect x="166" y="401" width="48" height="48" rx="24" fill="#7650D4" stroke="none" stroke-width="2"/><text x="190" y="435" font-size="29" fill="#fff" font-weight="700" text-anchor="middle">1</text><text x="190" y="484" font-size="27" fill="#142341" font-weight="700" text-anchor="middle">Main agent</text><text x="190" y="518" font-size="21" fill="#56647D" font-weight="400" text-anchor="middle">Receives the</text><text x="190" y="546" font-size="21" fill="#56647D" font-weight="400" text-anchor="middle">user’s task</text><rect x="330" y="383" width="190" height="184" rx="16" fill="#F1ECFF" stroke="#E0D5FC" stroke-width="2"/><rect x="401" y="401" width="48" height="48" rx="24" fill="#7650D4" stroke="none" stroke-width="2"/><text x="425" y="435" font-size="29" fill="#fff" font-weight="700" text-anchor="middle">2</text><text x="425" y="484" font-size="27" fill="#142341" font-weight="700" text-anchor="middle">Task routing</text><text x="425" y="518" font-size="21" fill="#56647D" font-weight="400" text-anchor="middle">Selects an</text><text x="425" y="546" font-size="21" fill="#56647D" font-weight="400" text-anchor="middle">open worker plan</text><rect x="565" y="383" width="190" height="184" rx="16" fill="#F1ECFF" stroke="#E0D5FC" stroke-width="2"/><rect x="636" y="401" width="48" height="48" rx="24" fill="#7650D4" stroke="none" stroke-width="2"/><text x="660" y="435" font-size="29" fill="#fff" font-weight="700" text-anchor="middle">3</text><text x="660" y="484" font-size="27" fill="#142341" font-weight="700" text-anchor="middle">Worker agent</text><text x="660" y="518" font-size="21" fill="#56647D" font-weight="400" text-anchor="middle">Completes the</text><text x="660" y="546" font-size="21" fill="#56647D" font-weight="400" text-anchor="middle">delegated work</text><rect x="865" y="383" width="190" height="184" rx="16" fill="#F1ECFF" stroke="#E0D5FC" stroke-width="2"/><rect x="936" y="401" width="48" height="48" rx="24" fill="#7650D4" stroke="none" stroke-width="2"/><text x="960" y="435" font-size="29" fill="#fff" font-weight="700" text-anchor="middle">1</text><text x="960" y="484" font-size="27" fill="#142341" font-weight="700" text-anchor="middle">Existing agent</text><text x="960" y="518" font-size="21" fill="#56647D" font-weight="400" text-anchor="middle">Continues the</text><text x="960" y="546" font-size="21" fill="#56647D" font-weight="400" text-anchor="middle">current work</text><rect x="1100" y="383" width="190" height="184" rx="16" fill="#F1ECFF" stroke="#E0D5FC" stroke-width="2"/><rect x="1171" y="401" width="48" height="48" rx="24" fill="#7650D4" stroke="none" stroke-width="2"/><text x="1195" y="435" font-size="29" fill="#fff" font-weight="700" text-anchor="middle">2</text><text x="1195" y="484" font-size="27" fill="#142341" font-weight="700" text-anchor="middle">Model routing</text><text x="1195" y="518" font-size="21" fill="#56647D" font-weight="400" text-anchor="middle">Chooses for</text><text x="1195" y="546" font-size="21" fill="#56647D" font-weight="400" text-anchor="middle">this stage</text><rect x="1335" y="383" width="190" height="184" rx="16" fill="#F1ECFF" stroke="#E0D5FC" stroke-width="2"/><rect x="1406" y="401" width="48" height="48" rx="24" fill="#7650D4" stroke="none" stroke-width="2"/><text x="1430" y="435" font-size="29" fill="#fff" font-weight="700" text-anchor="middle">3</text><text x="1430" y="484" font-size="27" fill="#142341" font-weight="700" text-anchor="middle">Model &</text><text x="1430" y="518" font-size="21" fill="#56647D" font-weight="400" text-anchor="middle">reasoning settings</text><path d="M 293 476 H 320" fill="none" stroke="#7650D4" stroke-width="3" stroke-linejoin="round" marker-end="url(#arrow)"/><path d="M 528 476 H 555" fill="none" stroke="#7650D4" stroke-width="3" stroke-linejoin="round" marker-end="url(#arrow)"/><path d="M 1063 476 H 1090" fill="none" stroke="#7650D4" stroke-width="3" stroke-linejoin="round" marker-end="url(#arrow)"/><path d="M 1298 476 H 1325" fill="none" stroke="#7650D4" stroke-width="3" stroke-linejoin="round" marker-end="url(#arrow)"/><text x="1185" y="614" font-size="26" fill="#344B72" font-weight="400" text-anchor="middle">Economy / primary models</text><text x="1185" y="652" font-size="25" fill="#344B72" font-weight="400" text-anchor="middle">or custom writing / review branches</text><text x="800" y="761" font-size="37" fill="#7650D4" font-weight="700" text-anchor="middle">Enable task routing and model routing independently</text><rect x="50" y="810" width="1500" height="107" rx="20" fill="#F1ECFF" stroke="none" stroke-width="2"/><rect x="78" y="834" width="280" height="60" rx="14" fill="#7650D4" stroke="none" stroke-width="2"/><text x="218" y="873" font-size="24" fill="#fff" font-weight="700" text-anchor="middle">Use existing compute</text><text x="520" y="875" font-size="29" fill="#142341" font-weight="700" text-anchor="middle">Subscriptions</text><text x="811" y="875" font-size="29" fill="#142341" font-weight="700" text-anchor="middle">Model APIs</text><text x="1090" y="875" font-size="29" fill="#142341" font-weight="700" text-anchor="middle">Free quotas</text><text x="1370" y="875" font-size="29" fill="#142341" font-weight="700" text-anchor="middle">Local models</text></g></svg>diff --git a/news/assets/launch-product-routing-zh.png b/news/assets/launch-product-routing-zh.pngnew file mode 100644index 00000000..fd9d992fBinary files /dev/null and b/news/assets/launch-product-routing-zh.png differdiff --git a/news/assets/launch-smart-saving-en.png b/news/assets/launch-smart-saving-en.pngnew file mode 100644index 00000000..6273d76dBinary files /dev/null and b/news/assets/launch-smart-saving-en.png differdiff --git a/news/assets/launch-smart-saving-zh.png b/news/assets/launch-smart-saving-zh.pngnew file mode 100644index 00000000..e5956a78Binary files /dev/null and b/news/assets/launch-smart-saving-zh.png differdiff --git a/news/assets/launch-stage-competence-en.png b/news/assets/launch-stage-competence-en.pngnew file mode 100644index 00000000..6ae09e2bBinary files /dev/null and b/news/assets/launch-stage-competence-en.png differdiff --git a/news/assets/launch-stage-competence-zh.png b/news/assets/launch-stage-competence-zh.pngnew file mode 100644index 00000000..6d8dfe13Binary files /dev/null and b/news/assets/launch-stage-competence-zh.png differdiff --git a/news/hiroute-open-source.en.md b/news/hiroute-open-source.en.mdnew file mode 100644index 00000000..3300281e--- /dev/null+++ b/news/hiroute-open-source.en.md@@ -0,0 +1,201 @@+# HiRoute Is Open Source: Help Your Agents Do More and Your Subscriptions Go Further++An agent task can move from research and code reading to design and debugging. Each step calls for different model capabilities. Use your strongest model throughout, and routine work consumes valuable subscription capacity. Switch to an economy model, and you risk getting stuck on a critical decision or a difficult bug. With several agents, subscriptions, and providers' APIs available, working out how to combine them becomes another task in itself.++The Higress team built HiRoute to make that combination easier. **HiRoute is a local engine for intelligent agent routing and collaboration.** Keep using the agent you know, let economy models handle routine work, and bring in a primary model for demanding stages. When a task benefits from a separate worker, delegate it to another agent. The aim is to make your existing tools and compute work together, while reducing unnecessary spending and manual handoffs.++We previewed HiRoute's open-source plans in the keynote at AGNTCon + MCPCon China and formally open sourced it at the Apsara Conference. You can now download it, use it, and help build it.++[](assets/launch-product-routing-en.svg)++*Figure 1. Task routing and model routing can be enabled independently or used together.*++## 1. Why open source: connect scattered agents, models, and compute++HiRoute grew out of three kinds of fragmentation we encounter every day.++Models have different strengths. Summarizing research, reasoning through a difficult problem, writing code, and handling long context all make different demands on capability and cost.++Agents have different workflows. Some people prefer Codex's execution flow, others know Claude Code best, and many use additional tools for research or specialized engineering work.++Compute is spread across subscriptions, API keys from different providers, and local services. Capabilities you have already paid for or configured can be difficult to use together.++[](assets/launch-agent-ecosystem.png)++*Figure 2. Agent applications, frameworks, and runtime infrastructure keep diversifying. Users need a simpler way to combine them.*++Long tasks add a time dimension to these choices. A job may start with research, move through design and implementation, and end with validation and difficult repairs. A model that works well for gathering information may not be enough for an architectural decision. Once the hard part is solved, the remaining steps may no longer need the most expensive capabilities.++Users therefore care about the cost of the whole task: whether the strongest model is used where it matters, whether subscription capacity is reserved for difficult work, whether failures allow a controlled handoff, and what evidence can inform the next choice.++HiRoute turns those choices into capabilities you can configure, reuse, and inspect. Open source lets developers review routing behavior, connect their own workflows, and improve the system through real tasks.++Configuration, routing, and observation run locally. Connected model services perform inference. Individual developers and teams building their own agents can start with tools they already use.++## 2. Product value: more capable agents, more useful subscriptions++### Combine models and subscriptions; reserve primary capacity for critical work++The most immediate benefit is keeping your strongest capabilities available for work that needs them.++If you have a supported subscription, such as Codex, alongside a lower-cost API, HiRoute can connect that subscription through CLIProxyAPI (CPA) and include both sources in one routing plan.++An economy API can handle research summaries and routine edits, while a primary model from your subscription handles complex analysis and critical repairs. That can reduce expensive calls and make subscription capacity go further. For example, place Qwen 3.8 Flash in the economy group and choose a stronger model suited to your tasks for the primary group.++### Connect decision models such as Jev; choose capabilities for the task++As agent work becomes longer, developers are taking an interest in System One models that specialize in fast decisions. [Jev](https://openrouter.ai/typesafe/jev-1.13/) is one example: its categories, probabilities, and scores are useful for deciding how work should proceed. **HiRoute includes Jev in its built-in decision-model system.**++Under **Models → Decision models**, connect a decision model from Bailian (Alibaba Cloud Model Studio) or TypeSafe's Jev through OpenRouter. You do not need to deploy your own decision service.++A decision model judges the current task's category and difficulty and can assess the preceding execution stage. HiRoute applies your thresholds, model groups, and candidate order to make the selection. Simple tasks use the economy group; complex tasks, or work following a stage that falls below your competence standard, use the primary group. You can edit the criteria and assessment prompts so that “simple,” “complex,” and “good enough” reflect your own work.++This also applies beyond coding. For an article, define separate **writing** and **review** branches. Writing can emphasize structure, expression, and use of source material; review can focus on evidence, factual accuracy, and omissions. Each branch can have regular and primary models. You define the categories and the standards.++A writing-stage score belongs to that actual execution stage. It is not automatically applied to a review task.++### Preserve context continuity; maintain high KV Cache hit rates++An agent usually completes work through repeated tool calls, responses, and further reasoning. Context grows throughout that process. High KV Cache hit rates help reduce input costs and response latency.++HiRoute uses **context continuity**, implemented by ContextHold, to recognize calls whose history continues within the same user turn. It retains the routing decision and prefers the eligible current model, allowing the growing context prefix to keep benefiting from cache reuse.++When context compaction or history rebuilding breaks the old complete prefix and the previous decision can no longer be continued, HiRoute reassesses the model at that boundary. Whether it keeps the same model or chooses another, execution then continues along the new prefix and builds cache hits again.++A new user request also triggers a fresh decision. Simply appending a message does not necessarily invalidate the previous prefix: continuing with the same model can still reuse existing cache.++[](assets/launch-context-continuity-en.svg)++*Figure 3. Keep the model stable while context is continuous. When the old complete prefix breaks, reassess the model and continue reusing KV Cache along the new prefix.*++### Give agents distinct roles; keep your familiar workflow++Choose agents around the work as well. Keep your usual main agent, delegate independent research, implementation, or test tasks to configured workers, then bring their results together. Task routing answers “who should do this?” Model routing answers “what capabilities does this stage need?” You can use either layer on its own or combine them.++### Record execution and handoffs; make model choices with evidence++When a candidate is rate limited or unavailable, HiRoute tries the next eligible candidate within the plan's boundaries. It stops explicitly when candidates are exhausted. Sessions and Model performance let you inspect which model actually ran each stage, why a selection occurred, how much it used, and whether the stage received an assessment. Your next routing plan can draw on your own execution records as well as general benchmarks.++[](assets/launch-stage-competence-en.png)++*Figure 4. Model performance is organized by execution stage. Open a stage to inspect the assessment evidence and execution history.*++## 3. Proof in practice: lower costs and less manual intervention++### Lower costs: reserve strong reasoning for critical decisions, at over 90% lower cost++Our first case was a deliberately structured engineering research task: investigate six messaging systems, produce 60 source-verification cards for each, and deliver a decision memo on Kafka and SQL consistency.++The research required careful checking against official sources. The final memo asked harder questions: after failures and retries, could the design prevent duplicate business processing or missed operations? Finding that a component “supports transactions” was only the start; the agent had to decide whether that guarantee covered the whole business workflow. A large volume of research supported a few critical decisions—a familiar structure in technology selection and system migration.++We ran the same task in three modes: all GPT-6 Astra, HiRoute mixed routing, and all Qwen 3.8 Flash. In the mixed mode, Qwen handled the six research packages and Astra handled the critical decision memo.++In the two paired runs where both mixed routing and all Astra passed the original whole-task acceptance criteria, mixed routing reduced API-equivalent cost by **92.01% and 91.39%**, respectively.++Across those accepted runs, average source-verification accuracy was 99.03% for mixed routing and 99.72% for all Astra. Both passed the critical-decision review.++Three reviews of the critical memo show why the division matters: mixed routing and all Astra each passed 3/3 without major errors; all Qwen Flash passed 1/3. An economy model can do substantial research work, while the decisive reasoning still needs sufficient capability. Matching capabilities to the right stages makes it possible to maintain the delivery standard while spending less.++[](assets/research-cost-en.svg)++*Figure 5. The chart compares average costs across the two accepted paired runs; the percentages in the text describe each run separately. Costs include all attempts and routing assessments, using frozen API prices. They do not represent subscription-bill savings. The [full HiRoute case study](https://hiroute.ai/en/news/astra-qwen-smart-routing/) links to the process and experiment records.*++### Less intervention: automatic model handoffs complete a long task from one instruction++Our second case focused on sustained delivery. The task was to add synchronous and asynchronous streaming JSON iteration to HTTPX: support different stream formats, handle encoding and consumption state, and yield parsed values before later input arrived. We gave native Codex one instruction and let it work.++The economy-first plan went through four consecutive stages: **Qwen → Astra → Qwen → Astra**. Qwen first read source code and tests. The first Astra handoff prepared a context summary; the policy in use at the time then returned execution to Qwen.++Work continued, but remained largely focused on research and investigation. At the second context handoff, the recorded competence score was 0.485, below the 0.5 threshold, and HiRoute selected Astra again.++Astra then wrote the streaming parser, ran tests, and repaired asynchronous iterator cleanup, encoding edge cases, and type-checking issues.++The task took **31 minutes 36 seconds**, with no additional human prompts and no manual code patches during execution. Independent acceptance passed all **343 checks**: 108 functional checks, 229 regression checks, and 6 incremental-streaming assertions.++This makes the value of stage feedback concrete. An economy model starts the work; at a boundary where reassessment is possible, HiRoute checks progress and lets a primary model continue implementation and validation. The execution record also distinguishes summarizing, investigating, and actually writing code, so users can see what each handoff accomplished.++[](assets/unattended-handoff-en.svg)++*Figure 6. One task crossed four model stages, with execution feedback guiding the later implementation and validation.*++[](assets/desktop-session-competence-en.png)++*Figure 7. Native observation of the historical experiment. The first Astra stage produced only a summary; the final implementation stage remains unrated. Scores from different stages are not a controlled model ranking. This run retains its historical configuration and has no all-Astra cost baseline, so it supports a handoff result rather than a cost-reduction claim. The [handoff process and acceptance records](https://hiroute.ai/en/news/astra-qwen-smart-routing/) are public.*++## 4. Getting started: three steps into your existing workflow++HiRoute provides a macOS Desktop app and a headless CLI and background service for Linux. Configure it visually on a desktop, or use the same capabilities in servers and automated workflows.++For your first setup, start with a simple **Smart saving** plan.++### Step 1: connect models and subscriptions++In **Models**, add an API, a supported subscription, or a local model service. For example, use Qwen 3.8 Flash as the economy model and Qwen 3.8 Max as the primary model. You can also combine an economy API with an existing Codex subscription. Model sources are managed together; once connected, a source can be used in multiple plans.++[](assets/launch-model-connections-en.png)++*Figure 8. Start with Add models: connect an API, scan existing subscriptions and configurations, or browse free models.*++### Step 2: create and publish a routing plan++In **Smart routing**, choose **Smart saving**, add economy and primary models, and select a decision model. Built-in connections such as Bailian and Jev through OpenRouter need connection settings, without a separately deployed extension service. Judgment thresholds and assessment criteria are collapsed by default. Begin with the two model groups, then expand the settings when you need finer control. Once enabled, the published plan becomes a reusable model entry point.++[](assets/launch-smart-saving-en.png)++*Figure 9. Choose economy and primary models for the combination. Expand judgment thresholds and advanced settings when needed.*++### Step 3: connect an agent and inspect execution++In **Agent**, enable **Model routing** and select your published plan for the agent you use. With Codex, copy the launch command after enabling routing, start a new session in your terminal, and give it tasks as usual. **Sessions** shows the actual models, usage, and costs. **Model performance** links competence scores to specific stages and execution evidence, giving you a basis for refining the next combination.++[](assets/launch-agent-connections-en.png)++*Figure 10. Enable Model routing on the Agent page to connect your usual agent to a published plan.*++## 5. Architecture: local first, built on Higress engineering experience++HiRoute follows a **local-first architecture**. Configuration, routing, and observation records run on your machine, without a HiRoute cloud relay. Your agent's model requests reach the local gateway first, then the model services you choose.++When you use a cloud execution or decision model, the requests required for inference or assessment go to that service. You can also connect local models and use your own compute.++HiRoute Gateway is built in **Rust**. It brings engineering experience from [Higress AI Gateway](https://github.com/higress-group/higress) into local agent workflows: protocol adaptation, capability and eligibility checks, bounded candidate failover, streaming delivery, and execution observation. The Desktop app uses Tauri; Desktop and CLI share the same core, so visual setup and automated calls follow the same execution rules.++[](assets/launch-local-architecture-en.svg)++*Figure 11. Configuration, routing, and execution records stay on your machine. Connected services handle model inference and decisions; local model services can also be used.*++Those engineering capabilities serve four product principles.++### Separate task and model routing; reuse the plan++Task routing chooses the agent that carries out the work; model routing chooses the capabilities used at each stage. A plan expresses model groups, candidate order, task conditions, and assessment criteria, and publication saves a definite configuration snapshot. Editing a draft does not silently change a published plan or rewrite historical records. An effective combination can be reused for later work.++### Separate judgment and execution; respect the plan's boundaries++A decision model identifies the current task category and difficulty and assesses the preceding stage when evidence is sufficient. HiRoute applies your thresholds to select a group, checks that candidates satisfy the request, calls them in order, and uses bounded failover for availability failures. It fails explicitly when candidates are exhausted. A decision cannot bypass the execution boundaries of the published plan.++### Keep continuous execution stable; reassess each new turn++ContextHold recognizes history continuation within the same turn. When authorization, the plan, and candidates remain applicable, it retains the decision and prefers the current model. Every new user message triggers reassessment; so does a history rebuild that prevents reuse of the prior decision. One upgrade does not lock subsequent turns to the upgraded model.++### Bind scores to actual stages; make evidence traceable++A valid low score from the preceding stage can trigger primary-model protection. Without sufficient evidence, the stage remains explicitly unrated. Assessments belong to the actual task and model that ran, and users can inspect the underlying execution. The next turn still receives a fresh decision.++Developers who need their own decision algorithm can connect a service through the open **HTTP decision-extension protocol**. The extension integrates its chosen decision service and implements judgment and assessment; HiRoute retains threshold policy and execution control. The project provides an [OpenAPI contract and a deployable Jev reference implementation](https://hiroute.ai/en/docs/decision-api/) to bring domain-specific expertise into the same routing flow.++## 6. What's next: enterprise governance and a broader ecosystem++HiRoute already supports **five agent integrations: Codex, Claude Code, Qoder, Pi, and DeepSeek Harness**. Its built-in catalog covers **100+ providers and 700+ model configurations**, including the GPT, Claude, Qwen, DeepSeek, Kimi, and GLM families.++It also offers **seven models with free-quota access**, which can be combined with subscriptions, APIs, and local models.++We plan to develop HiRoute along two paths.++The first is **connecting enterprise AI infrastructure**. Integration with enterprise AI gateways such as Higress can let teams centralize authorization and credentials, publish shared routing plans, and track costs. Employees continue working through their familiar local agents, while model combinations proven by the team become available to more colleagues.++The second is **broadening the open ecosystem**. We will keep connecting more agents, models, subscriptions, and decision services, with OpenCode, Hermes, and OpenClaw among the directions ahead. Developers should be able to adapt combinations as tasks and models change, while community-contributed algorithms and integrations share a common runtime.++HiRoute is open source under **Apache 2.0**. Start with your first routing plan, make better use of the agents, models, and subscriptions you already have, and join us with a real workflow to improve.++[**Visit GitHub**](https://github.com/higress-group/HiRoute) · [**Download HiRoute**](https://hiroute.ai/en/download/) · [**Read the documentation**](https://hiroute.ai/en/docs/)diff --git a/news/hiroute-open-source.zh-CN.md b/news/hiroute-open-source.zh-CN.mdnew file mode 100644index 00000000..b8bc5998--- /dev/null+++ b/news/hiroute-open-source.zh-CN.md@@ -0,0 +1,201 @@+# 让 AI Agent 更能干,让模型订阅更耐用:HiRoute 正式开源++一项 Agent 任务,从查资料、读代码,到做设计、修问题,每一步对模型能力的要求并不一样。一直用主力模型,常规工作也在消耗宝贵的订阅额度;换成经济模型,又担心卡在关键判断和疑难问题上。手里有多个 Agent、几份订阅和不同厂商的 API,如何组合使用,也成了一项额外工作。++为此,Higress 团队打造了 HiRoute:一个运行在本地的 Agent 智能路由与协作引擎。继续使用你熟悉的 Agent,让经济模型承担常规工作,让主力模型处理关键阶段;需要分工时,还可以把独立任务交给其他 Agent。我们希望让你已有的工具与算力配合起来,在把工作做好的同时,减少不必要的开销和手动接力。++HiRoute 此前在 AGNTCon + MCPCon China 的 Keynote 演讲中预告开源,并在云栖大会正式开源。现在,你可以下载使用,也可以参与共建。++[](assets/launch-product-routing-zh.png)++*图 1 任务路由与模型路由可以独立启用,也可以组合使用。*++## 一、为何开源:连接分散的 Agent、模型与算力++做 HiRoute,源于我们每天都在面对的三种碎片化。++模型各有所长:资料归纳、复杂推理、编码、长上下文处理,对能力和成本的要求不同。++Agent 也各有工作方式:有人习惯 Codex 的执行流程,有人更熟悉 Claude Code,也会为研究或特定工程任务使用其他工具。++算力则分散在订阅、不同厂商的 API Key 和本地服务中。已经购买、已经配置好的能力,往往很难组合起来使用。++[](assets/launch-agent-ecosystem.png)++*图 2 Agent 应用、框架与运行基础设施持续分化,用户需要更简单的组合方式。*++长程任务又为这些选择加上了时间维度。一项工作可能从调研开始,经过设计、实现、验证,最后进入疑难修复。整理资料时合适的模型,到了架构判断时未必足够;难点解决以后,后续步骤也未必还需要持续消耗最昂贵的能力。++用户开始关心整项任务的代价:主力模型是否用在关键处,订阅额度能否留给难题,失败时能否按规则接续,以及下次选模型时有什么事实可参考。++HiRoute 希望把这些选择变成可以配置、复用和检查的日常能力。我们选择开源,也希望开发者能够审查路由行为、接入自己的工作流,并用真实任务检验和改进这套机制。++配置、执行与观测在本地运行,模型推理由所连接的服务完成;个人开发者和自建 Agent 的团队,都可以从已有工具开始。++## 二、产品价值:让 Agent 更能干,让订阅更耐用++### 统筹模型与订阅,把主力能力留给关键工作++最直接的价值,是把主力能力留给真正需要它的工作。++如果你已经购买 Codex 等受支持的订阅,又配置了价格更低的 API,HiRoute 可以通过 CLIProxyAPI(CPA)接入受支持的订阅,把这些来源放进同一份路由计划。++让经济 API 承担资料整理、常规修改,把订阅中的主力模型用于复杂分析与关键修复,就有机会减少高价调用,也让订阅额度用得更久。比如将 Qwen3.8 Flash 放进省钱组,主力组则选择适合自己任务的强模型。++### 接入 Jev 等决策模型,按任务选择合适能力++随着 Agent 工作越来越长,专门承担快速判断的 System One 模型开始进入开发者视野。[Jev](https://openrouter.ai/typesafe/jev-1.13/) 就是其中一个代表:它擅长输出类别、概率与评分,适合用来判断下一步该如何执行。**HiRoute 已将 Jev 纳入内置决策模型体系。**++在“模型 → 决策模型”中,可以连接百炼或 TypeSafe 提供的 Jev 等决策模型,无需自行部署决策服务。++决策模型判断当前任务的类别和程度,并可评价上一实际执行阶段;HiRoute 根据你设定的阈值、模型组和候选顺序执行选择。任务简单时使用省钱组,任务复杂或上一阶段未达到胜任标准时使用主力组。你可以修改判断条件和评分提示词,让“简单”“复杂”“做得好”符合自己的业务。++这也适用于编码之外的工作。写一篇文章,可以给“写稿”和“审稿”配置不同分支:写稿侧重结构、表达和素材组织,审稿侧重论证、事实与遗漏;再为每类任务设置常规与主力模型。类别由你定义,标准由你掌握。++一次写稿阶段的评分归属于那次实际执行,不会被随意套用到审稿任务上。++### 保持上下文连续,兼顾 KV Cache 高命中++Agent 完成一项工作,往往需要反复调用工具、读取结果、继续推理,前面的上下文会不断积累。保持 KV Cache 高命中,能够降低输入成本与响应等待。++HiRoute 通过“上下文保持”(ContextHold)识别同一轮中历史连续的调用,沿用已有路由决策,并尽量继续使用当前模型,让不断增长的上下文前缀持续复用缓存。++当上下文压缩或重建导致旧的完整前缀断裂、原有决策无法续用时,HiRoute 会在这一边界重新判断模型。此时,无论继续原模型还是选择新模型,后续都沿着新的上下文保持连续执行,持续积累缓存命中。++新用户请求也会触发重新决策。不过,仅追加新消息时,原有前缀未必失效;继续原模型仍可能复用已有缓存。++[](assets/launch-context-continuity-zh.svg)++*图 3 连续上下文内保持模型稳定;旧完整前缀断裂后,重新选模,再沿新前缀持续复用 KV Cache。*++### 组织多个 Agent 分工,保留熟悉的工作方式++Agent 的选择同样可以围绕工作展开。你保留惯用的主 Agent,把研究、实现或测试等独立任务委派给已配置的 Worker,再继续整合结果。任务路由回答“交给谁做”,模型路由负责“这个阶段用什么能力”;两层能力可以分别使用,也可以组合起来。++### 记录执行与接力,让模型选择有据可查++遇到限流或候选不可用,HiRoute 会在计划允许的范围内有序接力;候选耗尽时明确停止。每段工作实际用了哪个模型、发生了怎样的选择、消耗多少用量,以及是否获得阶段评分,都能在会话和模型表现中核对。这样,调整下一份路由计划时,用户可以依据自己的执行记录,而不只依赖通用榜单。++[](assets/launch-stage-competence-zh.png)++*图 4 会话中的模型表现按执行阶段呈现,可继续查看评分依据和执行过程。*++## 三、实战验证:既省成本,也省心++### 省成本:把强模型用在关键处,达标运行成本降低 90% 以上++第一项实战是专门设计的工程调研任务:研究六种消息系统,每种完成 60 张资料核验卡,最后交付一份 Kafka 与 SQL 一致性决策备忘录。++前面的工作需要仔细查证官方资料,最后则要回答更难的问题:系统故障并重试时,方案能否避免业务重复处理或遗漏?发现某个组件“支持事务”,还需要进一步判断这项保证能否覆盖整个业务流程。大量资料整理,最终服务于少数关键判断,这也是技术选型和系统迁移中常见的工作结构。++我们让三种模式处理相同任务:固定使用 GPT-6 Astra、由 HiRoute 混合路由,以及固定使用 Qwen-3.8 Flash。混合模式中,六份资料整理由 Qwen 完成,关键决策备忘录交给 Astra。++结果显示,在双方均通过原定整单验收的两次配对运行中,混合模式相较全 Astra 的 API 等价成本分别降低 **92.01% 和 91.39%**。++在这两次达标运行中,混合模式和全 Astra 的资料核验平均正确率分别为 99.03% 和 99.72%,关键决策均通过验收。++三次关键决策审查更能说明分工的价值:混合模式与全 Astra 都是 3/3 无重大错误,全 Qwen Flash 为 1/3。经济模型能够承担大量资料工作,关键判断仍需要足够的推理能力。把能力分配到合适环节,才有机会同时守住交付标准和成本。++[](assets/research-cost-zh.svg)++*图 5 成本图比较两次达标运行的平均成本;正文百分比是各次降幅。统计包含受试尝试与路由评估,按冻结价格折算 API 等价成本,不代表订阅账单节省比例。完整过程与实验记录入口见[HiRoute 官网实战详解](https://hiroute.ai/news/astra-qwen-smart-routing/)。*++### 省心:模型自动接力,一次指令完成长任务++第二项实战关注持续交付。任务要求在 HTTPX 中增加同步与异步 JSON 流式迭代,支持不同流格式、处理编码和消费状态,并在后续数据尚未到达时就产出已经解析的值。我们只向原生 Codex 下达一次任务,让它持续执行。++这次采用经济模型优先的计划,实际经过 **Qwen → Astra → Qwen → Astra** 四个连续阶段。Qwen 先阅读源码和测试;第一次切到 Astra 时,处理的是上下文摘要,随后按当时的策略回到 Qwen。++继续执行后,工作仍主要停留在研究和探查。第二次上下文交接时,记录的阶段胜任度为 0.485,低于 0.5 下限,HiRoute 再次选择 Astra。++这次接手之后,Astra 写入流式解析实现,运行测试,并继续修复异步迭代器关闭、编码边界与类型检查等问题。++整项任务耗时 31 分 36 秒,中途没有人工追加提示,也没有人工代码补丁。最终独立验收的 108 项功能、229 项回归与 6 项增量性断言,合计 **343 项全部通过**。++这项实战让阶段反馈的价值变得具体:经济模型先承担工作,在可以重新判断的交接点检查进展,让主力模型接续实现与验证。执行记录同时保留了摘要、研究和真正写入代码的区别,方便用户回看每次切换究竟带来了什么。++[](assets/unattended-handoff-zh.svg)++*图 6 一次任务经历四个模型阶段,在执行反馈后接续实现与验证。*++[](assets/desktop-session-competence-zh.png)++*图 7 历史实验的原生会话记录。首轮 Astra 只生成摘要;最终实现阶段尚未评分,不同阶段分数不构成模型排名。该任务保留当时配置,且没有全 Astra 成本对照,不作降本结论。[接力过程与验收记录](https://hiroute.ai/news/astra-qwen-smart-routing/)已公开。*++## 四、快速上手:三步接入熟悉的工作流++HiRoute 提供 macOS Desktop,以及面向 Linux 的无界面 CLI 和后台服务。你可以在桌面上完成配置,也可以把同一套能力接入服务器和自动化流程。++第一次使用,从一份简单的“智能省钱”计划开始即可。++### 第一步:接入模型与订阅++在“模型”页添加 API、受支持的订阅或本地模型服务。例如,用 Qwen3.8 Flash 作为省钱模型,Qwen3.8 Max 作为主力模型;也可以把经济型 API 与已有 Codex 订阅组合起来。模型来源统一管理,一次接入后可用于多份路由计划。++[](assets/launch-model-connections-zh.png)++*图 8 从“添加模型”开始,接入 API、扫描已有订阅与配置,或浏览免费模型。*++### 第二步:创建并发布路由计划++在“智能路由”中选择“智能省钱”,分别加入省钱和主力模型,再选择决策模型。百炼、OpenRouter Jev 等内置接入只需配置连接,无需自行部署扩展服务。判断阈值、评分标准等进阶设置默认收起,先选好两组模型就能开始;需要精细控制时,再展开调整。发布后,这份计划就成为可以重复使用的模型入口。++[](assets/launch-smart-saving-zh.png)++*图 9 选择省钱与主力模型即可配置组合,判断阈值与进阶设置按需展开。*++### 第三步:连接 Agent,查看执行效果++在“Agent”页启用模型路由,为常用 Agent 选择刚发布的计划。以 Codex 为例,启用后复制启动命令,在终端开启新会话,再像往常一样下达任务。进入“会话”,可以看到真正执行的模型、用量和成本;进入“模型表现”,还可以沿具体阶段查看胜任评分与执行证据,为下一次调整组合提供依据。++[](assets/launch-agent-connections-zh.png)++*图 10 在 Agent 页启用模型路由,将常用 Agent 接到已发布的智能路由计划。*++## 五、架构设计:本地优先,复用 Higress 工程积累++HiRoute 采用**本地优先**的架构:配置管理、路由执行和观测记录都在本机运行,无需部署 HiRoute 云端中转服务。Agent 的模型请求先进入本地网关,再前往你选择的模型服务。++选择云端执行模型或决策模型时,完成推理或判断所需的请求会发送到相应服务;选择本地模型时,也可以复用自己的本地算力。++HiRoute Gateway 使用 Rust 构建,将 [Higress AI Gateway](https://github.com/higress-group/higress) 积累的网关工程经验用于本地 Agent 工作流:统一处理协议适配、模型能力与资格检查、有边界的候选接力、流式交付和执行观测。Desktop 基于 Tauri,CLI 与桌面端共享同一套核心,让可视化操作与自动化调用遵循一致的执行规则。++[](assets/launch-local-architecture-zh.png)++*图 11 配置、路由与执行记录保留在本机;模型推理和决策由所连接的服务完成,也可以接入本地模型。*++这些工程能力最终服务于四个产品原则。++### 任务与模型分层,路由计划可以复用++任务路由安排由哪个 Agent 承担工作,模型路由安排每个阶段使用什么能力。模型组、候选顺序、任务条件和评分标准都可以在计划中明确表达;发布时保存确定的配置快照。后续修改草稿不会悄悄改变已发布计划,也不会改写历史记录,让一次有效的组合能够复用到后续任务。++### 判断与执行分离,模型选择遵循边界++决策模型判断当前任务类别、处理难度,并在证据充分时评价上一阶段。HiRoute 根据你设定的阈值选择模型组,再检查候选是否满足请求要求、按顺序调用,并在可用性故障时进行有限接力。候选耗尽就明确失败,判断结果不会绕过已发布计划的执行边界。++### 连续执行保持稳定,新的轮次重新决策++ContextHold(上下文保持)识别同轮历史续接,在授权、计划和候选仍适用时,沿用已有决策并尽量保持当前模型。每条新用户消息都会触发重新判断;历史重建导致原有决策无法复用时,也会重新判断。一次模型升级不会永久锁住后续轮次。++### 评分绑定实际阶段,执行证据可以追溯++上一阶段的有效低分可以触发主力保护,缺少证据时明确保留未评分状态。评分绑定实际执行的任务与模型,可以回到具体过程核对;新一轮仍会重新决策,一次升级不会永久锁住后续选择。++需要自己的判断算法时,开发者还可以通过开放的 **HTTP 决策扩展协议**接入服务。扩展负责对接自己的决策服务并实现判断与评分,HiRoute 保留阈值策略与执行控制。项目提供 [OpenAPI 文档和可部署的 Jev 参考实现](https://hiroute.ai/docs/decision-api/),方便把领域经验接入同一条路由流程。++## 六、未来演进:连接企业治理,拓展开放生态++HiRoute 已适配 **Codex、Claude Code、Qoder、Pi、DeepSeek Harness 五类 Agent**,内置模型目录覆盖 **100 多家服务商、700 多种模型配置**,包括 GPT、Claude、Qwen、DeepSeek、Kimi、GLM 等模型家族;++同时提供 **7 款免费额度模型接入**,可以与已有订阅、API 和本地模型组合使用。++接下来,HiRoute 将沿两条路线继续演进。++一条是**连接企业 AI 基础设施**。通过对接 Higress 等企业 AI Gateway,让团队集中管理授权与凭据、统一发布智能路由计划,并形成可追踪的成本治理。员工仍从自己熟悉的本地 Agent 发起工作,团队验证有效的模型组合则可以共享给更多成员。++另一条是**持续扩大开放生态**。继续接入更多 Agent、模型、订阅与决策服务,推进 OpenCode、Hermes、OpenClaw 等后续方向,让开发者能够随任务和模型能力的变化调整组合,也让社区贡献的决策算法与集成能力获得共同的运行基础。++HiRoute 以 **Apache 2.0 协议**开源。欢迎从自己的第一份路由计划开始,把已有 Agent、模型和订阅用得更好,也欢迎带着真实工作流参与共建。++[**访问 GitHub**](https://github.com/higress-group/HiRoute) · [**下载 HiRoute**](https://hiroute.ai/download/) · [**阅读使用文档**](https://hiroute.ai/docs/)diff --git a/news/index.json b/news/index.jsonindex 63e6d24e..8b90f284 100644--- a/news/index.json+++ b/news/index.json@@ -18,5 +18,20 @@"zh": "2026-10-04-astra-qwen.zh-CN.md","en": "2026-10-04-astra-qwen.en.md"}+ },+ {+ "slug": "hiroute-open-source",+ "title": {+ "zh": "让 AI Agent 更能干,让模型订阅更耐用:HiRoute 正式开源",+ "en": "HiRoute Is Open Source: Help Your Agents Do More and Your Subscriptions Go Further"+ },+ "description": {+ "zh": "从开源缘起、核心价值和真实实战,到三步上手与本地架构:了解 HiRoute 如何连接已有 Agent、模型与订阅,把主力能力留给关键工作。",+ "en": "Why we built HiRoute, what it delivers, and how to get started: connect your existing agents, models, and subscriptions, and reserve your strongest models for the work that needs them."+ },+ "files": {+ "zh": "hiroute-open-source.zh-CN.md",+ "en": "hiroute-open-source.en.md"+ }}]