@agent-native/core 0.90.9 → 0.90.11

package/corpus/core/docs/content/locales/pt-BR/agent-native-toolkit.mdx

@@ -23,7 +23,7 @@ being rebuilt in slightly different ways every time.
   <Screen
     surface="desktop"
     html={
-      "<div style='min-height:560px;box-sizing:border-box;padding:24px;display:grid;grid-template-columns:200px 1fr 280px;gap:16px'><aside class='wf-card' style='display:flex;flex-direction:column;gap:8px;padding:14px'><div style='display:flex;align-items:center;gap:8px'><div style='width:26px;height:26px;border-radius:8px;background:var(--wf-accent-soft);display:flex;align-items:center;justify-content:center;font-size:12px;flex:none'>A</div><strong style='flex:1'>Acme</strong><span data-icon='chevronDown'></span></div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px;border-radius:8px;background:var(--wf-accent-soft)'><span data-icon='search'></span> Home</div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px'><span data-icon='edit'></span> Documents</div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px'><span data-icon='user'></span> Team</div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px'><span data-icon='settings'></span> Settings</div><div style='flex:1'></div><div style='display:flex;align-items:center;gap:8px;padding-top:10px;border-top:1.4px solid var(--wf-line)'><div style='width:24px;height:24px;border-radius:999px;background:var(--wf-accent-soft);flex:none'></div><small style='flex:1'>Steve</small><span data-icon='dots'></span></div></aside><main class='wf-card' style='display:flex;flex-direction:column;gap:14px'><div style='display:flex;align-items:center;gap:10px'><h2 style='margin:0'>Q3 Forecast</h2><span class='wf-pill'>Org</span><div style='flex:1'></div><div style='display:flex'><div style='width:26px;height:26px;border-radius:999px;background:var(--wf-accent-soft);border:2px solid var(--wf-card);flex:none'></div><div style='width:26px;height:26px;border-radius:999px;background:var(--wf-accent-soft);border:2px solid var(--wf-card);margin-left:-8px;flex:none'></div><div style='width:26px;height:26px;border-radius:999px;background:var(--wf-accent);border:2px solid var(--wf-card);margin-left:-8px;flex:none'></div></div><button class='primary'><span data-icon='user'></span> Share</button></div><div class='wf-box' style='display:flex;flex-direction:column;gap:10px'><small class='wf-muted'>Revenue by month</small><div style='display:flex;align-items:flex-end;gap:10px;height:120px'><div style='flex:1;height:45%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:65%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:55%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:80%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:95%;background:var(--wf-accent);border-radius:4px'></div></div></div><div style='display:flex;gap:10px'><div class='wf-box' style='flex:1'><small class='wf-muted'>ARR</small><br/><strong>$4.2M</strong></div><div class='wf-box' style='flex:1'><small class='wf-muted'>Win rate</small><br/><strong>28%</strong></div><div class='wf-box' style='flex:1'><small class='wf-muted'>Pipeline</small><br/><strong>3.2x</strong></div></div></main><aside class='wf-card' style='display:flex;flex-direction:column;gap:10px'><strong>Agent</strong><div class='wf-box'>Summarize Q3 versus Q2 and flag risks.</div><div class='wf-box' style='background:var(--wf-accent-soft)'>Revenue up 12%. Two pipeline risks flagged.</div><div style='flex:1'></div><div class='wf-card' style='padding:10px;display:flex;flex-direction:column;gap:8px'><span class='wf-muted'>Ask about this dashboard...</span><div style='display:flex;align-items:center;gap:6px'><span class='wf-pill'>Q3 Forecast</span><span class='wf-pill'>GPT-5</span><div style='flex:1'></div><button class='primary'><span data-icon='send'></span></button></div></div></aside></div>"
+      "<div style='min-height:560px;box-sizing:border-box;display:grid;grid-template-columns:200px 1fr 280px;gap:16px'><aside class='wf-card' style='display:flex;flex-direction:column;gap:8px;padding:14px'><div style='display:flex;align-items:center;gap:8px'><div style='width:26px;height:26px;border-radius:8px;background:var(--wf-accent-soft);display:flex;align-items:center;justify-content:center;font-size:12px;flex:none'>A</div><strong style='flex:1'>Acme</strong><span data-icon='chevronDown'></span></div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px;border-radius:8px;background:var(--wf-accent-soft)'><span data-icon='search'></span> Home</div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px'><span data-icon='edit'></span> Documents</div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px'><span data-icon='user'></span> Team</div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px'><span data-icon='settings'></span> Settings</div><div style='flex:1'></div><div style='display:flex;align-items:center;gap:8px;padding-top:10px;border-top:1.4px solid var(--wf-line)'><div style='width:24px;height:24px;border-radius:999px;background:var(--wf-accent-soft);flex:none'></div><small style='flex:1'>Steve</small><span data-icon='dots'></span></div></aside><main class='wf-card' style='display:flex;flex-direction:column;gap:14px'><div style='display:flex;align-items:center;gap:10px'><h2 style='margin:0'>Q3 Forecast</h2><div style='flex:1'></div><div style='display:flex'><div style='width:26px;height:26px;border-radius:999px;background:var(--wf-accent-soft);border:2px solid var(--wf-card);flex:none'></div><div style='width:26px;height:26px;border-radius:999px;background:var(--wf-accent-soft);border:2px solid var(--wf-card);margin-left:-8px;flex:none'></div><div style='width:26px;height:26px;border-radius:999px;background:var(--wf-accent);border:2px solid var(--wf-card);margin-left:-8px;flex:none'></div></div><button class='primary'><span data-icon='user'></span> Share</button></div><div class='wf-box' style='display:flex;flex-direction:column;gap:10px'><small class='wf-muted'>Revenue by month</small><div style='display:flex;align-items:flex-end;gap:10px;height:120px'><div style='flex:1;height:45%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:65%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:55%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:80%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:95%;background:var(--wf-accent);border-radius:4px'></div></div></div><div style='display:flex;gap:10px'><div class='wf-box' style='flex:1'><small class='wf-muted'>ARR</small><br/><strong>$4.2M</strong></div><div class='wf-box' style='flex:1'><small class='wf-muted'>Win rate</small><br/><strong>28%</strong></div><div class='wf-box' style='flex:1'><small class='wf-muted'>Pipeline</small><br/><strong>3.2x</strong></div></div></main><aside class='wf-card' style='display:flex;flex-direction:column;gap:10px'><strong>Agent</strong><div class='wf-box'>Summarize Q3 versus Q2 and flag risks.</div><div class='wf-box' style='background:var(--wf-accent-soft)'>Revenue up 12%. Two pipeline risks flagged.</div><div style='flex:1'></div><div class='wf-card' style='padding:10px;display:flex;flex-direction:column;gap:8px'><span class='wf-muted'>Ask about this dashboard...</span><div style='display:flex;align-items:center;gap:6px'><span class='wf-pill'>Q3 Forecast</span><span class='wf-pill'>GPT-5</span><div style='flex:1'></div><button class='primary'><span data-icon='send'></span></button></div></div></aside></div>"
     }
   />
 </WireframeBlock>

package/corpus/core/docs/content/locales/pt-BR/tracking.mdx

@@ -200,7 +200,7 @@ configureTracking({
   endpoint: "https://analytics.example.com/api/analytics/track",
   sessionReplay: {
     enabled: true,
-    sampleRate: 0.1,
+    sampleRate: 1,
   },
   getDefaultProps: (_event, props) => ({
     ...props,
@@ -218,14 +218,14 @@ As raízes dos modelos do Agent Native já chamam `configureTracking()`. As impl
 VITE_AGENT_NATIVE_ANALYTICS_PUBLIC_KEY=anpk_...
 VITE_AGENT_NATIVE_ANALYTICS_ENDPOINT=https://analytics.example.com/api/analytics/track
 VITE_AGENT_NATIVE_SESSION_REPLAY_ENABLED=true
-VITE_AGENT_NATIVE_SESSION_REPLAY_SAMPLE_RATE=0.1
+VITE_AGENT_NATIVE_SESSION_REPLAY_SAMPLE_RATE=1
 ```
 
 O auxiliar do navegador também realiza uma leitura best-effort e não bloqueante da sessão de autenticação atual do Agent Native. O replay é, por padrão, exclusivo para usuários conectados: quando `sessionReplay` está habilitado, a gravação não começa a menos que a sessão seja resolvida para um endereço de e-mail de usuário. Replays de usuários conectados incluem `userId`/`userEmail` baseados em e-mail, além de `orgId`.
 
 Defina `sessionReplay.requireSignedInUser: false` ou `VITE_AGENT_NATIVE_SESSION_REPLAY_REQUIRE_AUTH=false` apenas para uma implantação de replay anônimo intencional. Quando o controle de autenticação está desabilitado, as gravações anônimas permanecem consultáveis por visitante anônimo, sessão, aplicativo/modelo, hostname e caminho.
 
-O replay de sessão é amostrado de forma determinística por sessão do navegador. Um `sampleRate` de `0.1` registra cerca de 10% das sessões elegíveis; use `1` quando a população elegível for intencionalmente pequena, como em dogfooding somente para usuários conectados.
+O replay de sessão é amostrado de forma determinística por sessão do navegador. Um `sampleRate` de `1` registra todas as sessões elegíveis; use valores menores, como `0.1`, quando a população elegível for intencionalmente ampla.
 
 Os padrões de privacidade são intencionalmente conservadores, mas ainda úteis para a reprodução:
 

package/corpus/core/docs/content/locales/zh-CN/agent-native-toolkit.mdx

@@ -23,7 +23,7 @@ being rebuilt in slightly different ways every time.
   <Screen
     surface="desktop"
     html={
-      "<div style='min-height:560px;box-sizing:border-box;padding:24px;display:grid;grid-template-columns:200px 1fr 280px;gap:16px'><aside class='wf-card' style='display:flex;flex-direction:column;gap:8px;padding:14px'><div style='display:flex;align-items:center;gap:8px'><div style='width:26px;height:26px;border-radius:8px;background:var(--wf-accent-soft);display:flex;align-items:center;justify-content:center;font-size:12px;flex:none'>A</div><strong style='flex:1'>Acme</strong><span data-icon='chevronDown'></span></div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px;border-radius:8px;background:var(--wf-accent-soft)'><span data-icon='search'></span> Home</div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px'><span data-icon='edit'></span> Documents</div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px'><span data-icon='user'></span> Team</div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px'><span data-icon='settings'></span> Settings</div><div style='flex:1'></div><div style='display:flex;align-items:center;gap:8px;padding-top:10px;border-top:1.4px solid var(--wf-line)'><div style='width:24px;height:24px;border-radius:999px;background:var(--wf-accent-soft);flex:none'></div><small style='flex:1'>Steve</small><span data-icon='dots'></span></div></aside><main class='wf-card' style='display:flex;flex-direction:column;gap:14px'><div style='display:flex;align-items:center;gap:10px'><h2 style='margin:0'>Q3 Forecast</h2><span class='wf-pill'>Org</span><div style='flex:1'></div><div style='display:flex'><div style='width:26px;height:26px;border-radius:999px;background:var(--wf-accent-soft);border:2px solid var(--wf-card);flex:none'></div><div style='width:26px;height:26px;border-radius:999px;background:var(--wf-accent-soft);border:2px solid var(--wf-card);margin-left:-8px;flex:none'></div><div style='width:26px;height:26px;border-radius:999px;background:var(--wf-accent);border:2px solid var(--wf-card);margin-left:-8px;flex:none'></div></div><button class='primary'><span data-icon='user'></span> Share</button></div><div class='wf-box' style='display:flex;flex-direction:column;gap:10px'><small class='wf-muted'>Revenue by month</small><div style='display:flex;align-items:flex-end;gap:10px;height:120px'><div style='flex:1;height:45%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:65%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:55%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:80%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:95%;background:var(--wf-accent);border-radius:4px'></div></div></div><div style='display:flex;gap:10px'><div class='wf-box' style='flex:1'><small class='wf-muted'>ARR</small><br/><strong>$4.2M</strong></div><div class='wf-box' style='flex:1'><small class='wf-muted'>Win rate</small><br/><strong>28%</strong></div><div class='wf-box' style='flex:1'><small class='wf-muted'>Pipeline</small><br/><strong>3.2x</strong></div></div></main><aside class='wf-card' style='display:flex;flex-direction:column;gap:10px'><strong>Agent</strong><div class='wf-box'>Summarize Q3 versus Q2 and flag risks.</div><div class='wf-box' style='background:var(--wf-accent-soft)'>Revenue up 12%. Two pipeline risks flagged.</div><div style='flex:1'></div><div class='wf-card' style='padding:10px;display:flex;flex-direction:column;gap:8px'><span class='wf-muted'>Ask about this dashboard...</span><div style='display:flex;align-items:center;gap:6px'><span class='wf-pill'>Q3 Forecast</span><span class='wf-pill'>GPT-5</span><div style='flex:1'></div><button class='primary'><span data-icon='send'></span></button></div></div></aside></div>"
+      "<div style='min-height:560px;box-sizing:border-box;display:grid;grid-template-columns:200px 1fr 280px;gap:16px'><aside class='wf-card' style='display:flex;flex-direction:column;gap:8px;padding:14px'><div style='display:flex;align-items:center;gap:8px'><div style='width:26px;height:26px;border-radius:8px;background:var(--wf-accent-soft);display:flex;align-items:center;justify-content:center;font-size:12px;flex:none'>A</div><strong style='flex:1'>Acme</strong><span data-icon='chevronDown'></span></div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px;border-radius:8px;background:var(--wf-accent-soft)'><span data-icon='search'></span> Home</div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px'><span data-icon='edit'></span> Documents</div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px'><span data-icon='user'></span> Team</div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px'><span data-icon='settings'></span> Settings</div><div style='flex:1'></div><div style='display:flex;align-items:center;gap:8px;padding-top:10px;border-top:1.4px solid var(--wf-line)'><div style='width:24px;height:24px;border-radius:999px;background:var(--wf-accent-soft);flex:none'></div><small style='flex:1'>Steve</small><span data-icon='dots'></span></div></aside><main class='wf-card' style='display:flex;flex-direction:column;gap:14px'><div style='display:flex;align-items:center;gap:10px'><h2 style='margin:0'>Q3 Forecast</h2><div style='flex:1'></div><div style='display:flex'><div style='width:26px;height:26px;border-radius:999px;background:var(--wf-accent-soft);border:2px solid var(--wf-card);flex:none'></div><div style='width:26px;height:26px;border-radius:999px;background:var(--wf-accent-soft);border:2px solid var(--wf-card);margin-left:-8px;flex:none'></div><div style='width:26px;height:26px;border-radius:999px;background:var(--wf-accent);border:2px solid var(--wf-card);margin-left:-8px;flex:none'></div></div><button class='primary'><span data-icon='user'></span> Share</button></div><div class='wf-box' style='display:flex;flex-direction:column;gap:10px'><small class='wf-muted'>Revenue by month</small><div style='display:flex;align-items:flex-end;gap:10px;height:120px'><div style='flex:1;height:45%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:65%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:55%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:80%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:95%;background:var(--wf-accent);border-radius:4px'></div></div></div><div style='display:flex;gap:10px'><div class='wf-box' style='flex:1'><small class='wf-muted'>ARR</small><br/><strong>$4.2M</strong></div><div class='wf-box' style='flex:1'><small class='wf-muted'>Win rate</small><br/><strong>28%</strong></div><div class='wf-box' style='flex:1'><small class='wf-muted'>Pipeline</small><br/><strong>3.2x</strong></div></div></main><aside class='wf-card' style='display:flex;flex-direction:column;gap:10px'><strong>Agent</strong><div class='wf-box'>Summarize Q3 versus Q2 and flag risks.</div><div class='wf-box' style='background:var(--wf-accent-soft)'>Revenue up 12%. Two pipeline risks flagged.</div><div style='flex:1'></div><div class='wf-card' style='padding:10px;display:flex;flex-direction:column;gap:8px'><span class='wf-muted'>Ask about this dashboard...</span><div style='display:flex;align-items:center;gap:6px'><span class='wf-pill'>Q3 Forecast</span><span class='wf-pill'>GPT-5</span><div style='flex:1'></div><button class='primary'><span data-icon='send'></span></button></div></div></aside></div>"
     }
   />
 </WireframeBlock>

package/corpus/core/docs/content/locales/zh-CN/tracking.mdx

@@ -200,7 +200,7 @@ configureTracking({
   endpoint: "https://analytics.example.com/api/analytics/track",
   sessionReplay: {
     enabled: true,
-    sampleRate: 0.1,
+    sampleRate: 1,
   },
   getDefaultProps: (_event, props) => ({
     ...props,
@@ -218,14 +218,14 @@ Agent Native 模板根已经调用了 `configureTracking()`。托管模板部署
 VITE_AGENT_NATIVE_ANALYTICS_PUBLIC_KEY=anpk_...
 VITE_AGENT_NATIVE_ANALYTICS_ENDPOINT=https://analytics.example.com/api/analytics/track
 VITE_AGENT_NATIVE_SESSION_REPLAY_ENABLED=true
-VITE_AGENT_NATIVE_SESSION_REPLAY_SAMPLE_RATE=0.1
+VITE_AGENT_NATIVE_SESSION_REPLAY_SAMPLE_RATE=1
 ```
 
 浏览器端辅助函数还会尽力地、非阻塞地读取当前 Agent Native 认证会话。默认情况下，重放仅限已登录用户：启用 `sessionReplay` 后，除非会话解析出用户邮箱地址，否则不会开始录制。已登录用户的重放会包含基于邮箱的 `userId`/`userEmail` 以及 `orgId`。
 
 仅在有意进行匿名重放部署时，才设置 `sessionReplay.requireSignedInUser: false` 或 `VITE_AGENT_NATIVE_SESSION_REPLAY_REQUIRE_AUTH=false`。禁用认证门控后，匿名录制仍可按匿名访客、会话、应用/模板、主机名和路径进行查询。
 
-会话重放按浏览器会话确定性采样。`sampleRate` 为 `0.1` 时，会记录约 10% 的符合条件的会话；当符合条件的用户群体本就很小时（例如仅限登录用户的内部测试），可使用 `1`。
+会话重放按浏览器会话确定性采样。`sampleRate` 为 `1` 时，会记录所有符合条件的会话；当符合条件的用户群体有意较大时，可使用较低值，例如 `0.1`。
 
 隐私默认设置刻意保持保守，但仍能满足回放的可用性：
 

package/corpus/core/docs/content/locales/zh-TW/agent-native-toolkit.mdx

@@ -23,7 +23,7 @@ being rebuilt in slightly different ways every time.
   <Screen
     surface="desktop"
     html={
-      "<div style='min-height:560px;box-sizing:border-box;padding:24px;display:grid;grid-template-columns:200px 1fr 280px;gap:16px'><aside class='wf-card' style='display:flex;flex-direction:column;gap:8px;padding:14px'><div style='display:flex;align-items:center;gap:8px'><div style='width:26px;height:26px;border-radius:8px;background:var(--wf-accent-soft);display:flex;align-items:center;justify-content:center;font-size:12px;flex:none'>A</div><strong style='flex:1'>Acme</strong><span data-icon='chevronDown'></span></div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px;border-radius:8px;background:var(--wf-accent-soft)'><span data-icon='search'></span> Home</div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px'><span data-icon='edit'></span> Documents</div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px'><span data-icon='user'></span> Team</div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px'><span data-icon='settings'></span> Settings</div><div style='flex:1'></div><div style='display:flex;align-items:center;gap:8px;padding-top:10px;border-top:1.4px solid var(--wf-line)'><div style='width:24px;height:24px;border-radius:999px;background:var(--wf-accent-soft);flex:none'></div><small style='flex:1'>Steve</small><span data-icon='dots'></span></div></aside><main class='wf-card' style='display:flex;flex-direction:column;gap:14px'><div style='display:flex;align-items:center;gap:10px'><h2 style='margin:0'>Q3 Forecast</h2><span class='wf-pill'>Org</span><div style='flex:1'></div><div style='display:flex'><div style='width:26px;height:26px;border-radius:999px;background:var(--wf-accent-soft);border:2px solid var(--wf-card);flex:none'></div><div style='width:26px;height:26px;border-radius:999px;background:var(--wf-accent-soft);border:2px solid var(--wf-card);margin-left:-8px;flex:none'></div><div style='width:26px;height:26px;border-radius:999px;background:var(--wf-accent);border:2px solid var(--wf-card);margin-left:-8px;flex:none'></div></div><button class='primary'><span data-icon='user'></span> Share</button></div><div class='wf-box' style='display:flex;flex-direction:column;gap:10px'><small class='wf-muted'>Revenue by month</small><div style='display:flex;align-items:flex-end;gap:10px;height:120px'><div style='flex:1;height:45%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:65%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:55%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:80%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:95%;background:var(--wf-accent);border-radius:4px'></div></div></div><div style='display:flex;gap:10px'><div class='wf-box' style='flex:1'><small class='wf-muted'>ARR</small><br/><strong>$4.2M</strong></div><div class='wf-box' style='flex:1'><small class='wf-muted'>Win rate</small><br/><strong>28%</strong></div><div class='wf-box' style='flex:1'><small class='wf-muted'>Pipeline</small><br/><strong>3.2x</strong></div></div></main><aside class='wf-card' style='display:flex;flex-direction:column;gap:10px'><strong>Agent</strong><div class='wf-box'>Summarize Q3 versus Q2 and flag risks.</div><div class='wf-box' style='background:var(--wf-accent-soft)'>Revenue up 12%. Two pipeline risks flagged.</div><div style='flex:1'></div><div class='wf-card' style='padding:10px;display:flex;flex-direction:column;gap:8px'><span class='wf-muted'>Ask about this dashboard...</span><div style='display:flex;align-items:center;gap:6px'><span class='wf-pill'>Q3 Forecast</span><span class='wf-pill'>GPT-5</span><div style='flex:1'></div><button class='primary'><span data-icon='send'></span></button></div></div></aside></div>"
+      "<div style='min-height:560px;box-sizing:border-box;display:grid;grid-template-columns:200px 1fr 280px;gap:16px'><aside class='wf-card' style='display:flex;flex-direction:column;gap:8px;padding:14px'><div style='display:flex;align-items:center;gap:8px'><div style='width:26px;height:26px;border-radius:8px;background:var(--wf-accent-soft);display:flex;align-items:center;justify-content:center;font-size:12px;flex:none'>A</div><strong style='flex:1'>Acme</strong><span data-icon='chevronDown'></span></div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px;border-radius:8px;background:var(--wf-accent-soft)'><span data-icon='search'></span> Home</div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px'><span data-icon='edit'></span> Documents</div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px'><span data-icon='user'></span> Team</div><div style='display:flex;align-items:center;gap:8px;padding:7px 9px'><span data-icon='settings'></span> Settings</div><div style='flex:1'></div><div style='display:flex;align-items:center;gap:8px;padding-top:10px;border-top:1.4px solid var(--wf-line)'><div style='width:24px;height:24px;border-radius:999px;background:var(--wf-accent-soft);flex:none'></div><small style='flex:1'>Steve</small><span data-icon='dots'></span></div></aside><main class='wf-card' style='display:flex;flex-direction:column;gap:14px'><div style='display:flex;align-items:center;gap:10px'><h2 style='margin:0'>Q3 Forecast</h2><div style='flex:1'></div><div style='display:flex'><div style='width:26px;height:26px;border-radius:999px;background:var(--wf-accent-soft);border:2px solid var(--wf-card);flex:none'></div><div style='width:26px;height:26px;border-radius:999px;background:var(--wf-accent-soft);border:2px solid var(--wf-card);margin-left:-8px;flex:none'></div><div style='width:26px;height:26px;border-radius:999px;background:var(--wf-accent);border:2px solid var(--wf-card);margin-left:-8px;flex:none'></div></div><button class='primary'><span data-icon='user'></span> Share</button></div><div class='wf-box' style='display:flex;flex-direction:column;gap:10px'><small class='wf-muted'>Revenue by month</small><div style='display:flex;align-items:flex-end;gap:10px;height:120px'><div style='flex:1;height:45%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:65%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:55%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:80%;background:var(--wf-accent-soft);border-radius:4px'></div><div style='flex:1;height:95%;background:var(--wf-accent);border-radius:4px'></div></div></div><div style='display:flex;gap:10px'><div class='wf-box' style='flex:1'><small class='wf-muted'>ARR</small><br/><strong>$4.2M</strong></div><div class='wf-box' style='flex:1'><small class='wf-muted'>Win rate</small><br/><strong>28%</strong></div><div class='wf-box' style='flex:1'><small class='wf-muted'>Pipeline</small><br/><strong>3.2x</strong></div></div></main><aside class='wf-card' style='display:flex;flex-direction:column;gap:10px'><strong>Agent</strong><div class='wf-box'>Summarize Q3 versus Q2 and flag risks.</div><div class='wf-box' style='background:var(--wf-accent-soft)'>Revenue up 12%. Two pipeline risks flagged.</div><div style='flex:1'></div><div class='wf-card' style='padding:10px;display:flex;flex-direction:column;gap:8px'><span class='wf-muted'>Ask about this dashboard...</span><div style='display:flex;align-items:center;gap:6px'><span class='wf-pill'>Q3 Forecast</span><span class='wf-pill'>GPT-5</span><div style='flex:1'></div><button class='primary'><span data-icon='send'></span></button></div></div></aside></div>"
     }
   />
 </WireframeBlock>

package/corpus/core/docs/content/locales/zh-TW/tracking.mdx

@@ -200,7 +200,7 @@ configureTracking({
   endpoint: "https://analytics.example.com/api/analytics/track",
   sessionReplay: {
     enabled: true,
-    sampleRate: 0.1,
+    sampleRate: 1,
   },
   getDefaultProps: (_event, props) => ({
     ...props,
@@ -218,14 +218,14 @@ Agent Native 範本根目錄已經呼叫了 `configureTracking()`。託管範本
 VITE_AGENT_NATIVE_ANALYTICS_PUBLIC_KEY=anpk_...
 VITE_AGENT_NATIVE_ANALYTICS_ENDPOINT=https://analytics.example.com/api/analytics/track
 VITE_AGENT_NATIVE_SESSION_REPLAY_ENABLED=true
-VITE_AGENT_NATIVE_SESSION_REPLAY_SAMPLE_RATE=0.1
+VITE_AGENT_NATIVE_SESSION_REPLAY_SAMPLE_RATE=1
 ```
 
 瀏覽器輔助程式也會盡力執行、非阻塞式讀取目前的 Agent Native 驗證工作階段。重播預設僅限已登入使用者：當 `sessionReplay` 啟用時，除非工作階段解析出使用者電子郵件地址，否則不會開始錄製。已登入使用者的重播包含以電子郵件為基礎的 `userId`/`userEmail` 以及 `orgId`。
 
 僅在刻意進行匿名重播部署時，才設定 `sessionReplay.requireSignedInUser: false` 或 `VITE_AGENT_NATIVE_SESSION_REPLAY_REQUIRE_AUTH=false`。停用驗證閘控後，匿名錄製仍可依匿名訪客、工作階段、應用程式/範本、主機名稱與路徑進行查詢。
 
-工作階段重播是依瀏覽器工作階段進行確定性取樣的。`sampleRate` 為 `0.1` 表示大約記錄 10% 符合條件的工作階段；當符合條件的母體刻意較小時（例如僅限已登入使用者的內部試用），可使用 `1`。
+工作階段重播是依瀏覽器工作階段進行確定性取樣的。`sampleRate` 為 `1` 表示記錄所有符合條件的工作階段；當符合條件的母體刻意較大時，可使用較低值，例如 `0.1`。
 
 隱私預設值刻意設計得保守，但仍足以支援播放：
 

package/corpus/core/docs/content/tracking.mdx

@@ -200,7 +200,7 @@ configureTracking({
   endpoint: "https://analytics.example.com/api/analytics/track",
   sessionReplay: {
     enabled: true,
-    sampleRate: 0.1,
+    sampleRate: 1,
   },
   getDefaultProps: (_event, props) => ({
     ...props,
@@ -218,14 +218,14 @@ Agent Native template roots already call `configureTracking()`. Hosted template
 VITE_AGENT_NATIVE_ANALYTICS_PUBLIC_KEY=anpk_...
 VITE_AGENT_NATIVE_ANALYTICS_ENDPOINT=https://analytics.example.com/api/analytics/track
 VITE_AGENT_NATIVE_SESSION_REPLAY_ENABLED=true
-VITE_AGENT_NATIVE_SESSION_REPLAY_SAMPLE_RATE=0.1
+VITE_AGENT_NATIVE_SESSION_REPLAY_SAMPLE_RATE=1
 ```
 
 The browser helper also performs a best-effort, non-blocking read of the current Agent Native auth session. Replay is signed-in-only by default: when `sessionReplay` is enabled, recording does not start unless the session resolves to a user email address. Signed-in replays include email-backed `userId`/`userEmail` plus `orgId`.
 
 Set `sessionReplay.requireSignedInUser: false` or `VITE_AGENT_NATIVE_SESSION_REPLAY_REQUIRE_AUTH=false` only for an intentional anonymous replay deployment. When auth gating is disabled, anonymous recordings remain queryable by anonymous visitor, session, app/template, hostname, and path.
 
-Session replay is sampled deterministically per browser session. A `sampleRate` of `0.1` records about 10% of eligible sessions; use `1` when the eligible population is intentionally small, such as logged-in-only dogfooding.
+Session replay is sampled deterministically per browser session. A `sampleRate` of `1` records all eligible sessions; use lower values, such as `0.1`, when the eligible population is intentionally broad.
 
 Privacy defaults are intentionally conservative but still useful for playback:
 

package/corpus/core/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@agent-native/core",
-  "version": "0.90.9",
+  "version": "0.90.11",
   "description": "Framework for agent-native application development — where AI agents and UI share SQL state, actions, and context",
   "homepage": "https://github.com/BuilderIO/agent-native#readme",
   "bugs": {

package/corpus/core/src/agent/durable-background.ts

@@ -260,12 +260,12 @@ export function backgroundRunMarkerExpectsBackgroundRuntime(
 }
 
 export function shouldUseBackgroundFunctionTimeoutForWorker(
-  marker: unknown,
+  _marker: unknown,
 ): boolean {
-  return (
-    isInBackgroundFunctionRuntime() ||
-    backgroundRunMarkerExpectsBackgroundRuntime(marker)
-  );
+  // The dispatch marker says which URL the foreground targeted, not where the
+  // request actually landed. Only the worker runtime proof can safely lift the
+  // hosted 40s clamp to the 15-minute background-function budget.
+  return isInBackgroundFunctionRuntime();
 }
 
 export function backgroundRuntimeDiagnosticDetail(marker: unknown): string {

package/corpus/core/src/agent/engine/ai-sdk-engine.ts

@@ -19,7 +19,10 @@ import {
 } from "../../server/credential-provider.js";
 import { normalizeReasoningEffortForModel } from "../../shared/reasoning-effort.js";
 import { AI_SDK_MODEL_CONFIG, type AISDKProvider } from "../model-config.js";
-import { resolveMaxOutputTokensForEngine } from "./output-tokens.js";
+import {
+  clampThinkingBudgetTokens,
+  resolveMaxOutputTokensForEngine,
+} from "./output-tokens.js";
 import {
   engineToolsToAISDK,
   engineMessagesToAISDK,
@@ -262,16 +265,34 @@ class AISDKEngine implements AgentEngine {
       toolResultImages: this.capabilities.vision,
     });
 
+    // Resolved once so both `maxOutputTokens` (below, in the streamText call)
+    // and the thinking-budget headroom clamp agree on the same ceiling.
+    const resolvedMaxOutputTokens = resolveMaxOutputTokensForEngine(
+      this.name,
+      opts.maxOutputTokens,
+      opts.model,
+    );
+
     // Build providerOptions for Anthropic-native features when using Anthropic provider
     const providerOpts: Record<string, unknown> = {};
     if (this.provider === "anthropic" && opts.providerOptions?.anthropic) {
       const anthropicOpts = opts.providerOptions.anthropic;
       if (anthropicOpts.thinking) {
+        // Only the "enabled" config carries a numeric budgetTokens; clamp it
+        // so thinking can't consume the entire maxOutputTokens budget and
+        // leave zero room for the actual response ("adaptive" thinking has
+        // no budgetTokens field at all per @ai-sdk/anthropic's schema).
         providerOpts.anthropic = {
           ...((providerOpts.anthropic as object) ?? {}),
           thinking: {
             type: "enabled",
-            budgetTokens: anthropicOpts.thinking.budgetTokens,
+            budgetTokens:
+              typeof anthropicOpts.thinking.budgetTokens === "number"
+                ? clampThinkingBudgetTokens(
+                    anthropicOpts.thinking.budgetTokens,
+                    resolvedMaxOutputTokens,
+                  )
+                : anthropicOpts.thinking.budgetTokens,
           },
         };
       }
@@ -309,11 +330,26 @@ class AISDKEngine implements AgentEngine {
         // Gemini 3.x models reject thinkingBudget — they require thinkingLevel.
         // Gemini 2.5.x models use thinkingBudget (integer token count or -1).
         const isGemini3 = /^gemini-3/.test(opts.model);
+        const thinkingBudget = googleThinkingBudget(reasoningEffort);
         providerOpts.google = {
           ...((providerOpts.google as object) ?? {}),
           thinkingConfig: isGemini3
             ? { thinkingLevel: gemini3ThinkingLevel(reasoningEffort) }
-            : { thinkingBudget: googleThinkingBudget(reasoningEffort) },
+            : {
+                // Unlike Anthropic's adaptive thinking, Gemini 2.5's
+                // thinkingBudget IS a concrete numeric token count, so the
+                // same headroom clamp applies: at "max" effort this maps to
+                // 32000 tokens, which can equal (or exceed) a small
+                // maxOutputTokens cap and leave zero room for the actual
+                // response. Preserve Gemini's -1 "dynamic" sentinel.
+                thinkingBudget:
+                  thinkingBudget > 0
+                    ? clampThinkingBudgetTokens(
+                        thinkingBudget,
+                        resolvedMaxOutputTokens,
+                      )
+                    : thinkingBudget,
+              },
         };
       }
     }
@@ -326,11 +362,7 @@ class AISDKEngine implements AgentEngine {
         system: opts.systemPrompt,
         messages,
         tools: aiSdkTools,
-        maxOutputTokens: resolveMaxOutputTokensForEngine(
-          this.name,
-          opts.maxOutputTokens,
-          opts.model,
-        ),
+        maxOutputTokens: resolvedMaxOutputTokens,
         ...(opts.temperature !== undefined
           ? { temperature: opts.temperature }
           : {}),

package/corpus/core/src/agent/engine/anthropic-engine.ts

@@ -19,7 +19,10 @@ import {
   LLM_MISSING_CREDENTIALS_ERROR_CODE,
   LLM_MISSING_CREDENTIALS_MESSAGE,
 } from "./credential-errors.js";
-import { resolveMaxOutputTokensForEngine } from "./output-tokens.js";
+import {
+  clampThinkingBudgetTokens,
+  resolveMaxOutputTokensForEngine,
+} from "./output-tokens.js";
 import {
   engineToolsToAnthropic,
   engineMessagesToAnthropic,
@@ -67,12 +70,31 @@ class AnthropicEngine implements AgentEngine {
     const messages = engineMessagesToAnthropic(opts.messages);
     const anthropicOpts = opts.providerOptions?.anthropic;
 
+    // Resolved once so both max_tokens and the thinking-budget headroom
+    // clamp below agree on the same ceiling.
+    const resolvedMaxOutputTokens = resolveMaxOutputTokensForEngine(
+      this.name,
+      opts.maxOutputTokens,
+      opts.model,
+    );
+
     // Build extra body params for Anthropic-native features
     const extra: Record<string, unknown> = {};
     if (anthropicOpts?.thinking) {
       extra.thinking = {
         type: anthropicOpts.thinking.type,
-        budget_tokens: anthropicOpts.thinking.budgetTokens,
+        // Only the "enabled" config carries a numeric budget_tokens; clamp it
+        // so thinking can't consume the entire max_tokens budget and leave
+        // zero room for the actual response ("adaptive" thinking has no
+        // budget_tokens field at all, so it passes through unclamped).
+        budget_tokens:
+          anthropicOpts.thinking.type === "enabled" &&
+          typeof anthropicOpts.thinking.budgetTokens === "number"
+            ? clampThinkingBudgetTokens(
+                anthropicOpts.thinking.budgetTokens,
+                resolvedMaxOutputTokens,
+              )
+            : anthropicOpts.thinking.budgetTokens,
       };
     }
     if (anthropicOpts?.topK !== undefined) {
@@ -139,11 +161,7 @@ class AnthropicEngine implements AgentEngine {
 
     const requestParams: any = {
       model: opts.model,
-      max_tokens: resolveMaxOutputTokensForEngine(
-        this.name,
-        opts.maxOutputTokens,
-        opts.model,
-      ),
+      max_tokens: resolvedMaxOutputTokens,
       system: systemBlocks,
       tools: cachedTools.length > 0 ? cachedTools : undefined,
       messages: cachedMessages,

package/corpus/core/src/agent/engine/output-tokens.ts

@@ -13,6 +13,100 @@ export const DEFAULT_AI_SDK_MAX_OUTPUT_TOKENS = 4096;
 export const DEFAULT_ANTHROPIC_MAX_OUTPUT_TOKENS = 8192;
 export const DEFAULT_BUILDER_MAX_OUTPUT_TOKENS = 8192;
 
+// ---------------------------------------------------------------------------
+// Main interactive chat path
+//
+// The per-engine defaults above (4096-8192) exist for internal/eval/
+// observational-memory callers that intentionally want a small explicit cap.
+// The main interactive chat turn (the one the end user is staring at) needs
+// real headroom: on long-context or reasoning-heavy turns, a tiny completion
+// budget means extended thinking alone can consume the entire response,
+// leaving zero tokens for visible text/tool calls ("empty response" bug).
+// These helpers give the chat path a much higher floor while staying at or
+// under each model's documented ceiling — they never lower the ceiling.
+// ---------------------------------------------------------------------------
+
+/** Cap for the first attempt of an interactive chat turn. */
+export const MAIN_CHAT_MAX_OUTPUT_TOKENS_CAP = 32_000;
+/**
+ * Cap used only when retrying a turn that came back with an empty final
+ * response (see production-agent.ts's empty-final-response retry). Higher
+ * than the first-attempt cap so the retry meaningfully raises the ceiling.
+ */
+export const EMPTY_RESPONSE_RETRY_MAX_OUTPUT_TOKENS_CAP = 64_000;
+
+/**
+ * Resolve the max_output_tokens floor for the first attempt of an
+ * interactive chat turn: min(model ceiling, 32K). Always at or above the
+ * flat per-engine defaults above, regardless of whether the model is known.
+ */
+export function resolveMainChatMaxOutputTokens(modelId?: string): number {
+  return Math.min(
+    getMaxOutputTokensForModel(modelId),
+    MAIN_CHAT_MAX_OUTPUT_TOKENS_CAP,
+  );
+}
+
+/**
+ * Resolve the max_output_tokens to use when retrying a turn after an empty
+ * final response: min(model ceiling, 64K).
+ */
+export function resolveEmptyResponseRetryMaxOutputTokens(
+  modelId?: string,
+): number {
+  return Math.min(
+    getMaxOutputTokensForModel(modelId),
+    EMPTY_RESPONSE_RETRY_MAX_OUTPUT_TOKENS_CAP,
+  );
+}
+
+// ---------------------------------------------------------------------------
+// Extended-thinking budget headroom
+//
+// Anthropic's `thinking: { type: "enabled", budget_tokens }` config requires
+// budget_tokens >= 1024 and STRICTLY LESS THAN max_tokens (confirmed against
+// the installed @anthropic-ai/sdk@0.90.0 type defs — see ThinkingConfigEnabled
+// in resources/messages/messages.d.ts). budget_tokens counts toward
+// max_tokens, so an unclamped large budget can leave too little (or zero)
+// room for the actual visible completion. This clamp guarantees at least
+// max(8000, 40% of maxOutputTokens) tokens of non-thinking headroom.
+//
+// Note: this only applies to the explicit numeric-budget "enabled" config.
+// Anthropic's `type: "adaptive"` thinking config (used by the
+// reasoningEffort -> output_config.effort mapping in anthropic-engine.ts /
+// ai-sdk-engine.ts) has NO budget_tokens field at all per the SDK types, so
+// there is nothing to clamp there — those callers rely on the raised
+// maxOutputTokens ceiling above instead.
+// ---------------------------------------------------------------------------
+
+/** Anthropic's documented minimum extended-thinking budget. */
+export const ANTHROPIC_MIN_THINKING_BUDGET_TOKENS = 1024;
+
+/**
+ * Clamp a requested Anthropic thinking budget so it leaves guaranteed
+ * headroom under `maxOutputTokens` for non-thinking output, and stays within
+ * the provider's valid range (>= 1024, < maxOutputTokens).
+ */
+export function clampThinkingBudgetTokens(
+  requestedBudgetTokens: number,
+  maxOutputTokens: number,
+): number | undefined {
+  if (maxOutputTokens <= ANTHROPIC_MIN_THINKING_BUDGET_TOKENS) {
+    return undefined;
+  }
+  const headroom = Math.max(8000, Math.round(0.4 * maxOutputTokens));
+  const budgetCapForHeadroom = Math.max(
+    ANTHROPIC_MIN_THINKING_BUDGET_TOKENS,
+    maxOutputTokens - headroom,
+  );
+  // budget_tokens must stay strictly below max_tokens per the API contract.
+  const strictUpperBound = maxOutputTokens - 1;
+  return Math.max(
+    ANTHROPIC_MIN_THINKING_BUDGET_TOKENS,
+    Math.min(requestedBudgetTokens, budgetCapForHeadroom, strictUpperBound),
+  );
+}
+
 function parsePositiveInteger(value: unknown): number | null {
   if (typeof value === "string" && value.trim() === "") return null;
   const n =

package/corpus/core/src/agent/production-agent.ts

@@ -45,6 +45,7 @@ import { fireInternalDispatch } from "../server/self-dispatch.js";
 import {
   isReasoningEffort,
   normalizeReasoningEffortForModel,
+  stepDownReasoningEffort,
   type ReasoningEffort,
 } from "../shared/reasoning-effort.js";
 import { actionPreparationContinuationNote } from "./action-continuation-guidance.js";
@@ -77,7 +78,11 @@ import {
   normalizeModelForEngine,
   isResolvedEngineUsableForRequest,
 } from "./engine/index.js";
-import { resolveMaxOutputTokensForEngine } from "./engine/output-tokens.js";
+import {
+  resolveEmptyResponseRetryMaxOutputTokens,
+  resolveMainChatMaxOutputTokens,
+  resolveMaxOutputTokensForEngine,
+} from "./engine/output-tokens.js";
 import { PROVIDER_TO_ENV } from "./engine/provider-env-vars.js";
 import {
   backfillEngineMessagesToolResults,
@@ -1015,6 +1020,10 @@ const TOOL_INPUT_ACTIVITY_INTERVAL_MS = 1500;
 const ACTION_PREPARATION_NO_PROGRESS_TIMEOUT_MS = 90_000;
 const ACTION_PREPARATION_ZERO_BYTE_RESTART_LIMIT = 2;
 const MODEL_STREAM_NO_PROGRESS_TIMEOUT_MS = 90_000;
+// Raised from 1 -> 2 now that each retry actually adapts (raises the token
+// ceiling and steps reasoning effort down a tier) instead of re-issuing the
+// exact same doomed request twice.
+const EMPTY_FINAL_RESPONSE_RETRY_LIMIT = 2;
 const MAIN_CHAT_INTERNAL_CONTINUATION_LIMIT = 6;
 const RUN_BUDGET_EXHAUSTED_ERROR_CODE = "run_budget_exhausted";
 const RUN_BUDGET_EXHAUSTED_MESSAGE =
@@ -2733,7 +2742,14 @@ export async function runAgentLoop(opts: {
   }
 
   let finalGuardRetries = 0;
+  let emptyFinalResponseRetries = 0;
   let iterations = 0;
+  // Overridden (raised tokens, lowered effort) only after an empty-final-
+  // response retry below — kept separate from `opts.maxOutputTokens`/
+  // `opts.reasoningEffort` so the very first attempt is unaffected and later
+  // tool-loop turns revert to the caller's original request after a success.
+  let effectiveMaxOutputTokens = opts.maxOutputTokens;
+  let effectiveReasoningEffort = opts.reasoningEffort;
 
   // Set when an in-loop processor aborts via `abort()` / throws a `TripWire`.
   // The loop emits the `tripwire` event, surfaces the reason as a final
@@ -2843,10 +2859,10 @@ export async function runAgentLoop(opts: {
           abortSignal: signal,
           maxOutputTokens: resolveMaxOutputTokensForEngine(
             engine.name,
-            opts.maxOutputTokens,
+            effectiveMaxOutputTokens,
             model,
           ),
-          reasoningEffort: opts.reasoningEffort,
+          reasoningEffort: effectiveReasoningEffort,
           providerOptions: opts.providerOptions,
         };
 
@@ -3349,16 +3365,33 @@ export async function runAgentLoop(opts: {
       // text — typically when reasoning consumes the entire output-token
       // budget. Without a final text part the SSE stream still ends with a
       // clean `done`, which renders as a totally empty assistant bubble.
-      // Surface a plain-language error so the user knows what happened.
-      if (
+      // Retry so a reasoning-budget miss can still finish; each retry raises
+      // the token ceiling and steps reasoning effort down a tier so it's not
+      // just re-issuing the identical doomed request. If retries also come
+      // back empty, surface a plain-language error.
+      const hasEmptyFinalResponse =
         !guardEmittedFallback &&
         collectTextParts(assistantContentForHistory).trim().length === 0 &&
-        streamedAssistantText.trim().length === 0
-      ) {
+        streamedAssistantText.trim().length === 0;
+      if (hasEmptyFinalResponse) {
+        if (emptyFinalResponseRetries < EMPTY_FINAL_RESPONSE_RETRY_LIMIT) {
+          emptyFinalResponseRetries += 1;
+          effectiveMaxOutputTokens =
+            resolveEmptyResponseRetryMaxOutputTokens(model);
+          effectiveReasoningEffort = stepDownReasoningEffort(
+            effectiveReasoningEffort,
+          );
+          appendAgentLoopContinuation(messages, "max_tokens");
+          continue;
+        }
         send({
           type: "text",
           text: "The model returned an empty response. This usually means reasoning used the full output-token budget. Try again, or pick a different model from the model menu.",
         });
+      } else {
+        emptyFinalResponseRetries = 0;
+        effectiveMaxOutputTokens = opts.maxOutputTokens;
+        effectiveReasoningEffort = opts.reasoningEffort;
       }
       break;
     }
@@ -3369,6 +3402,7 @@ export async function runAgentLoop(opts: {
     // finalGuardRetries stays at 1 from a prior cycle and the guard is
     // permanently disabled for the rest of a long multi-step run.
     finalGuardRetries = 0;
+    emptyFinalResponseRetries = 0;
 
     flushUnstreamedAssistantText();
 
@@ -6784,6 +6818,11 @@ export function createProductionAgentHandler(
           orgId: getRequestOrgId() ?? null,
           attachments: requestAttachments,
           reasoningEffort,
+          // The interactive chat turn needs real completion headroom — the
+          // flat per-engine defaults (4096-8192) exist for internal/eval
+          // callers and are far below what a hard, long-context turn needs
+          // once extended thinking is in play. See output-tokens.ts.
+          maxOutputTokens: resolveMainChatMaxOutputTokens(effectiveModel),
           providerOptions: options.providerOptions,
           executionMode: requestMode,
           maxIterations: loopSettings.maxIterations,

package/corpus/core/src/agent/types.ts

@@ -180,8 +180,9 @@ export interface AgentChatRequest {
     continuationCount?: number;
     /**
      * True when the dispatcher expects the self-POST to land in a real
-     * Netlify `-background` function (15-min budget) rather than the ~60s
-     * synchronous function. See `shouldUseBackgroundFunctionTimeoutForWorker`.
+     * Netlify `-background` function rather than the ~60s synchronous function.
+     * This is diagnostic only; the 15-minute budget is unlocked by the worker's
+     * actual runtime marker.
      */
     backgroundFunctionRuntimeExpected?: boolean;
     /**