@cursor/july 0.1.92 → 0.1.94

package/dist/docs/assets/{deployment.md.DX_hc3ze.js → deployment.md.D9msOFOW.js}

@@ -1,4 +1,4 @@
-import{_ as e,c as a,o as t,ag as i}from"./chunks/framework.BCISBCiQ.js";const c=JSON.parse('{"title":"Deployment","description":"Deploy the Agent SDK with Cursor-managed hosting or on infrastructure you control.","frontmatter":{"title":"Deployment","description":"Deploy the Agent SDK with Cursor-managed hosting or on infrastructure you control."},"headers":[],"relativePath":"deployment.md","filePath":"deployment.md"}'),n={name:"deployment.md"};function h(o,s,l,r,d,p){return t(),a("div",null,[...s[0]||(s[0]=[i(`<h1 id="deploy-the-agent-sdk" tabindex="-1">Deploy the Agent SDK <a class="header-anchor" href="#deploy-the-agent-sdk" aria-label="Permalink to &quot;Deploy the Agent SDK&quot;">​</a></h1><p>Both options run the same agent project and HTTP API. Channel delivery paths differ. Cursor-managed hosting is preferred for most agents.</p><table tabindex="0"><thead><tr><th>Option</th><th>Use it when</th><th>You manage</th></tr></thead><tbody><tr><td>Cursor-managed hosting (preferred)</td><td>You want the shortest path from a Git repo to a running agent</td><td>Agent code, external storage, declared egress, and deployment secrets</td></tr><tr><td>Self-hosting</td><td>You need your own network, proxy, persistent filesystem, or process controls</td><td>Agent code, Node process, TLS, auth, secrets, durable state, monitoring, and upgrades</td></tr></tbody></table><h2 id="cursor-managed-hosting" tabindex="-1">Cursor-managed hosting <a class="header-anchor" href="#cursor-managed-hosting" aria-label="Permalink to &quot;Cursor-managed hosting&quot;">​</a></h2><p>Cursor builds the selected Git ref into a deployment. The deployment exposes a stable URL while Cursor manages its runtime lifecycle.</p><h3 id="before-you-deploy" tabindex="-1">Before you deploy <a class="header-anchor" href="#before-you-deploy" aria-label="Permalink to &quot;Before you deploy&quot;">​</a></h3><ul><li>Confirm managed hosting is enabled for the account and team.</li><li>Sign in with an account holding team-admin deployment permission.</li><li>Add <code>@cursor/july</code> to the agent project.</li></ul><p>For a GitHub source, install the Cursor GitHub App on the repository owner and grant it access to the repository. Cursor builds through its repository integration, not your local Git credentials. Commit and push the Git ref before deploying it.</p><h3 id="declare-hosting-needs" tabindex="-1">Declare hosting needs <a class="header-anchor" href="#declare-hosting-needs" aria-label="Permalink to &quot;Declare hosting needs&quot;">​</a></h3><p>If the agent needs extra egress or deployment secrets, add a <code>hosting</code> block to <code>agent/agent.ts</code>:</p><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { defineAgent } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;@cursor/july&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
+import{_ as s,c as a,o as t,ag as i}from"./chunks/framework.BCISBCiQ.js";const c=JSON.parse('{"title":"Deployment","description":"Deploy the Agent SDK with Cursor-managed hosting or on infrastructure you control.","frontmatter":{"title":"Deployment","description":"Deploy the Agent SDK with Cursor-managed hosting or on infrastructure you control."},"headers":[],"relativePath":"deployment.md","filePath":"deployment.md"}'),n={name:"deployment.md"};function o(h,e,l,r,d,p){return t(),a("div",null,[...e[0]||(e[0]=[i(`<h1 id="deploy-the-agent-sdk" tabindex="-1">Deploy the Agent SDK <a class="header-anchor" href="#deploy-the-agent-sdk" aria-label="Permalink to &quot;Deploy the Agent SDK&quot;">​</a></h1><p>Both options run the same agent project and HTTP API. Channel delivery paths differ. Cursor-managed hosting is preferred for most agents.</p><table tabindex="0"><thead><tr><th>Option</th><th>Use it when</th><th>You manage</th></tr></thead><tbody><tr><td>Cursor-managed hosting (preferred)</td><td>You want the shortest path from a Git repo to a running agent</td><td>Agent code, external storage, declared egress, and deployment secrets</td></tr><tr><td>Self-hosting</td><td>You need your own network, proxy, persistent filesystem, or process controls</td><td>Agent code, Node process, TLS, auth, secrets, durable state, monitoring, and upgrades</td></tr></tbody></table><h2 id="cursor-managed-hosting" tabindex="-1">Cursor-managed hosting <a class="header-anchor" href="#cursor-managed-hosting" aria-label="Permalink to &quot;Cursor-managed hosting&quot;">​</a></h2><p>Cursor builds the selected Git ref into a deployment. The deployment exposes a stable URL while Cursor manages its runtime lifecycle.</p><h3 id="before-you-deploy" tabindex="-1">Before you deploy <a class="header-anchor" href="#before-you-deploy" aria-label="Permalink to &quot;Before you deploy&quot;">​</a></h3><ul><li>Confirm managed hosting is enabled for the account and team.</li><li>Sign in with an account holding team-admin deployment permission.</li><li>Add <code>@cursor/july</code> to the agent project.</li></ul><p>For a GitHub source, install the Cursor GitHub App on the repository owner and grant it access to the repository. Cursor builds through its repository integration, not your local Git credentials. Commit and push the Git ref before deploying it.</p><h3 id="declare-hosting-needs" tabindex="-1">Declare hosting needs <a class="header-anchor" href="#declare-hosting-needs" aria-label="Permalink to &quot;Declare hosting needs&quot;">​</a></h3><p>If the agent needs extra egress or deployment secrets, add a <code>hosting</code> block to <code>agent/agent.ts</code>:</p><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { defineAgent } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;@cursor/july&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
 <span class="line"></span>
 <span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">export</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> default</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> defineAgent</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">({</span></span>
 <span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  hosting: {</span></span>
@@ -17,7 +17,7 @@ import{_ as e,c as a,o as t,ag as i}from"./chunks/framework.BCISBCiQ.js";const c
 <span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> deployment</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> weather-agent</span></span>
 <span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> logs</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --prod</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --slug</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> weather-agent</span></span></code></pre></div><h3 id="set-deployment-secrets" tabindex="-1">Set deployment secrets <a class="header-anchor" href="#set-deployment-secrets" aria-label="Permalink to &quot;Set deployment secrets&quot;">​</a></h3><p>A deployment must exist before you can set its secrets. Pass names only. Enter values at the hidden prompt, or pipe one line per name:</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> secrets</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> set</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> weather-agent</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> WEATHER_API_KEY</span></span>
 <span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> secrets</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> list</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> weather-agent</span></span>
-<span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> deploy</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --slug</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> weather-agent</span></span></code></pre></div><p><code>secrets list</code> returns names and creation times, never values. The engine reads secret changes on its next deploy.</p><p>Do not pass <code>NAME=VALUE</code> on the command line. That form lands in shell history and in agent-captured terminals. The CLI refuses it unless you add <code>--from-argv</code>. For non-interactive input without argv, redirect a file or pipe stdin.</p><h3 id="credential-output-in-terminals" tabindex="-1">Credential output in terminals <a class="header-anchor" href="#credential-output-in-terminals" aria-label="Permalink to &quot;Credential output in terminals&quot;">​</a></h3><p>Treat any command that prints a secret as a credential event. Keep it out of agent-captured terminals when you can.</p><table tabindex="0"><thead><tr><th>Command</th><th>What prints</th></tr></thead><tbody><tr><td><code>secrets set</code> / <code>secrets list</code></td><td>Names only. Values never print.</td></tr><tr><td><code>rotate-pod-credential</code></td><td>Masked key only.</td></tr><tr><td>First <code>deploy</code> / <code>rotate-token</code></td><td>Full alias token once. Save it outside the agent transcript; it cannot be retrieved later.</td></tr><tr><td><code>deployment --json</code></td><td>May include short-lived <code>engineAccess.headers</code>. Treat JSON as a credential.</td></tr></tbody></table><p>Do not verify secrets with <code>echo &quot;$SECRET&quot;</code>, <code>printenv</code>, or by pasting values into chat. Use <code>secrets list</code> for names, then redeploy and exercise the feature that needs the secret.</p><h3 id="choose-durable-storage" tabindex="-1">Choose durable storage <a class="header-anchor" href="#choose-durable-storage" aria-label="Permalink to &quot;Choose durable storage&quot;">​</a></h3><p>Hosted filesystem state can reset during a deploy or runtime replacement. Prefer <a href="./storage.html"><code>cursorHostedStorage</code></a> (<code>@cursor/july/storage/cursor-hosted</code>) so durable records land in Cursor&#39;s Bugbot <code>agent_serve_*</code> tables through a control-plane HTTP proxy (pod credential auth — no database URL in the engine). Do not put <code>BUGBOTDB_URL</code> or <code>AGENT_SERVE_DEPLOYMENT_ID</code> in <code>hosting.secretNames</code>. Self-host with your own <code>defineStorage</code> backend or a persistent <code>--state-root</code> when the complete filesystem must survive.</p><h3 id="use-the-hosted-agent" tabindex="-1">Use the hosted agent <a class="header-anchor" href="#use-the-hosted-agent" aria-label="Permalink to &quot;Use the hosted agent&quot;">​</a></h3><p>The CLI handles authentication for <code>--prod</code> commands. External HTTP clients that hit the stable alias URL must send <code>X-Agent-Alias-Token</code> on every request. Authored channel auth still runs after that gate.</p><p>Use <code>publicEndpoint()</code> on a custom channel that verifies its own provider signature. Cursor then serves that channel path without the alias token. The built-in session API still requires the token.</p><p>Use a relay or self-host when the channel cannot authenticate requests itself.</p><p>Use <code>--prod</code> with the normal client commands:</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> playground</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --prod</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --slug</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> weather-agent</span></span>
+<span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> deploy</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --slug</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> weather-agent</span></span></code></pre></div><p><code>secrets list</code> returns names and creation times, never values. The engine reads secret changes on its next deploy.</p><p>Do not pass <code>NAME=VALUE</code> on the command line. That form lands in shell history and in agent-captured terminals. The CLI refuses it unless you add <code>--from-argv</code>. For non-interactive input without argv, redirect a file or pipe stdin.</p><h3 id="credential-output-in-terminals" tabindex="-1">Credential output in terminals <a class="header-anchor" href="#credential-output-in-terminals" aria-label="Permalink to &quot;Credential output in terminals&quot;">​</a></h3><p>Treat any command that prints a secret as a credential event. Keep it out of agent-captured terminals when you can.</p><table tabindex="0"><thead><tr><th>Command</th><th>What prints</th></tr></thead><tbody><tr><td><code>secrets set</code> / <code>secrets list</code></td><td>Names only. Values never print.</td></tr><tr><td><code>rotate-pod-credential</code></td><td>Masked key only.</td></tr><tr><td>First <code>deploy</code> / <code>rotate-token</code></td><td>Full alias token once. Save it outside the agent transcript; it cannot be retrieved later.</td></tr><tr><td><code>deployment --json</code></td><td>May include short-lived <code>engineAccess.headers</code>. Treat JSON as a credential.</td></tr></tbody></table><p>Do not verify secrets with <code>echo &quot;$SECRET&quot;</code>, <code>printenv</code>, or by pasting values into chat. Use <code>secrets list</code> for names, then redeploy and exercise the feature that needs the secret.</p><h3 id="choose-durable-storage" tabindex="-1">Choose durable storage <a class="header-anchor" href="#choose-durable-storage" aria-label="Permalink to &quot;Choose durable storage&quot;">​</a></h3><p>Hosted filesystem state can reset during a deploy or runtime replacement. Prefer <a href="./storage.html"><code>cursorHostedStorage</code></a> so records survive replace. Do not put platform storage or deployment-identity names in <code>hosting.secretNames</code>. Self-host with your own <code>defineStorage</code> backend or a persistent <code>--state-root</code> when the complete filesystem must survive.</p><h3 id="use-the-hosted-agent" tabindex="-1">Use the hosted agent <a class="header-anchor" href="#use-the-hosted-agent" aria-label="Permalink to &quot;Use the hosted agent&quot;">​</a></h3><p>The CLI handles authentication for <code>--prod</code> commands. External HTTP clients that hit the stable alias URL must send <code>X-Agent-Alias-Token</code> on every request. Authored channel auth still runs after that gate.</p><p>Use <code>publicEndpoint()</code> on a custom channel that verifies its own provider signature. Cursor then serves that channel path without the alias token. The built-in session API still requires the token.</p><p>Use a relay or self-host when the channel cannot authenticate requests itself.</p><p>Use <code>--prod</code> with the normal client commands:</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> playground</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --prod</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --slug</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> weather-agent</span></span>
 <span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> chat</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --prod</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --slug</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> weather-agent</span></span>
 <span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> run</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --prod</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --slug</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> weather-agent</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --message</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;Forecast for Paris&quot;</span></span>
 <span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> sessions</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --prod</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --slug</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> weather-agent</span></span>
@@ -25,32 +25,27 @@ import{_ as e,c as a,o as t,ag as i}from"./chunks/framework.BCISBCiQ.js";const c
 <span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  -H</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;X-Agent-Alias-Token: </span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">$AGENT_ALIAS_TOKEN</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;</span></span></code></pre></div><p><code>--prod</code> commands don&#39;t use the alias token. If it is lost or exposed, rotate it. The old token stops working immediately:</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> rotate-token</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> weather-agent</span></span></code></pre></div><h3 id="connect-github" tabindex="-1">Connect GitHub <a class="header-anchor" href="#connect-github" aria-label="Permalink to &quot;Connect GitHub&quot;">​</a></h3><p>Let the hosted engine pull Cursor SCM events. Repeat <code>--cursor-events-repo</code> for each repository whose events should wake the agent:</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> deploy</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> \\</span></span>
 <span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  --slug</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> pr-approver</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> \\</span></span>
 <span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  --cursor-events-repo</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> acme/checkout</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> \\</span></span>
-<span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  --cursor-events-repo</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> acme/payments</span></span></code></pre></div><p>The agent still needs a <code>githubChannel()</code> declaration for the events it handles. This delivery path needs no public GitHub webhook URL. The flag requires Cursor SCM-event access for the deployment credential.</p><p>For outbound GitHub calls, use <code>githubChannel({ cursorAccount: true })</code> and grant the team&#39;s Cursor GitHub App access to each repository. Alternatively, add dedicated GitHub credentials as deployment secrets. See the <a href="./guides/github.html">GitHub guide</a>.</p><h3 id="connect-slack" tabindex="-1">Connect Slack <a class="header-anchor" href="#connect-slack" aria-label="Permalink to &quot;Connect Slack&quot;">​</a></h3><p>Hosted Slack supports the team&#39;s Cursor Slack app or a dedicated Socket Mode app.</p><p>Use the Cursor Slack app when mentions and direct messages are enough:</p><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { slackChannel } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;@cursor/july/channels/slack&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
-<span class="line"></span>
-<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">export</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> default</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> slackChannel</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">({</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  cursorAccount: </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">true</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  agentName: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;PrApprover&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">});</span></span></code></pre></div><p>The team must have the Cursor Slack app installed and Slack event relay access enabled. This mode needs no Slack token secrets. It doesn&#39;t support channel-post watches, tool approvals, or interactivity.</p><p>Use a dedicated Socket Mode app for those features or a separate bot identity:</p><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { slackChannel } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;@cursor/july/channels/slack&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
+<span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  --cursor-events-repo</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> acme/payments</span></span></code></pre></div><p>The agent still needs a <code>githubChannel()</code> declaration for the events it handles. This delivery path needs no public GitHub webhook URL. The flag requires Cursor SCM-event access for the deployment credential.</p><p>For outbound GitHub calls, use <code>githubChannel({ cursorAccount: true })</code> and grant the team&#39;s Cursor GitHub App access to each repository. Alternatively, add dedicated GitHub credentials as deployment secrets. See the <a href="./guides/github.html">GitHub guide</a>.</p><h3 id="connect-slack" tabindex="-1">Connect Slack <a class="header-anchor" href="#connect-slack" aria-label="Permalink to &quot;Connect Slack&quot;">​</a></h3><p>Hosted Slack uses a dedicated Socket Mode app:</p><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { slackChannel } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;@cursor/july/channels/slack&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
 <span class="line"></span>
 <span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">export</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> default</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> slackChannel</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">({ envPrefix: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;PR_APPROVER&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> });</span></span></code></pre></div><p>Provision the dedicated app from the dashboard. Open the deployment under <strong>Deployed Agents</strong> at <a href="https://cursor.com/dashboard" target="_blank" rel="noreferrer">cursor.com/dashboard</a>, switch to <strong>Details</strong>, and expand <strong>Slack</strong> under <strong>Integrations</strong>. <strong>Add to Slack</strong> connects the workspace with a one-time authorization, and <strong>Create Slack app</strong> creates and installs the app, then stores its tokens as deployment secrets automatically. Redeploy when prompted so the running agent picks them up. See <a href="./guides/slack.html#provision-from-the-dashboard">Provision from the dashboard</a> for the walkthrough, including workspace-admin approval.</p><p>If you created the Slack app by hand instead, set its tokens as deployment secrets yourself. The prefix selects the secret names:</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> secrets</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> set</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> pr-approver</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> \\</span></span>
 <span class="line"><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">  PR_APPROVER_SLACK_BOT_TOKEN</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> \\</span></span>
 <span class="line"><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">  PR_APPROVER_SLACK_APP_TOKEN</span></span>
-<span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> deploy</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --slug</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> pr-approver</span></span></code></pre></div><p>Without <code>envPrefix</code>, a dedicated app reads <code>SLACK_BOT_TOKEN</code> and <code>SLACK_APP_TOKEN</code>. Socket Mode needs no inbound URL. See the <a href="./guides/slack.html">Slack guide</a>.</p><h3 id="update-stop-or-delete-a-deployment" tabindex="-1">Update, stop, or delete a deployment <a class="header-anchor" href="#update-stop-or-delete-a-deployment" aria-label="Permalink to &quot;Update, stop, or delete a deployment&quot;">​</a></h3><p>Redeploy the same slug after pushing a new Git ref. The stable alias continues to point at the active generation. Follow the same source rules from <a href="#deploy-from-git">Deploy from Git</a>.</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> deploy</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> /path/to/my-agent</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --slug</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> weather-agent</span></span>
+<span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> deploy</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --slug</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> pr-approver</span></span></code></pre></div><p>Without <code>envPrefix</code>, a dedicated app reads <code>SLACK_BOT_TOKEN</code> and <code>SLACK_APP_TOKEN</code>. Socket Mode needs no inbound URL. See the <a href="./guides/slack.html">Slack guide</a>.</p><h3 id="update-stop-or-delete-a-deployment" tabindex="-1">Update, stop, or delete a deployment <a class="header-anchor" href="#update-stop-or-delete-a-deployment" aria-label="Permalink to &quot;Update, stop, or delete a deployment&quot;">​</a></h3><p>Redeploy the same slug after pushing a new Git ref. The stable alias continues to point at the latest deploy. Follow the same source rules from <a href="#deploy-from-git">Deploy from Git</a>.</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> deploy</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> /path/to/my-agent</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --slug</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> weather-agent</span></span>
 <span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> stop</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> weather-agent</span></span>
-<span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> delete</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> weather-agent</span></span></code></pre></div><p><code>stop</code> waits for the deployment to stop unless you pass <code>--no-wait</code>. <code>delete</code> removes the deployment and waits until it is gone. See the <a href="./reference/cli.html#deploy">CLI reference</a> for the full command reference.</p><h2 id="self-host-the-agent-sdk" tabindex="-1">Self-host the Agent SDK <a class="header-anchor" href="#self-host-the-agent-sdk" aria-label="Permalink to &quot;Self-host the Agent SDK&quot;">​</a></h2><p>The Agent SDK runs as a Node HTTP server on Node 22.13 or newer. You can host it on a VM, container platform, or ECS.</p><h3 id="the-security-model-in-one-minute" tabindex="-1">The security model in one minute <a class="header-anchor" href="#the-security-model-in-one-minute" aria-label="Permalink to &quot;The security model in one minute&quot;">​</a></h3><p><code>serve</code> binds to loopback and admits direct local callers by default. Choose one of these options before exposing it:</p><ol><li>Pass <code>--bearer-token &lt;secret&gt;</code> for a shared host.</li><li>Define channel-specific auth for routes with their own credentials or signatures.</li><li>Use <code>--allow-anonymous</code> only behind an authenticating proxy.</li></ol><p>A static bearer token maps every holder to one principal. Use authored auth when callers need separate identities. See <a href="./reference/channels.html#auth-policies">Channels</a> for policy details.</p><h3 id="credentials" tabindex="-1">Credentials <a class="header-anchor" href="#credentials" aria-label="Permalink to &quot;Credentials&quot;">​</a></h3><p>A self-hosted server can read these credentials.</p><table tabindex="0"><thead><tr><th>Credential</th><th>Used for</th><th>Provide it as</th></tr></thead><tbody><tr><td>Cursor API key</td><td>model turns, cloud runtime, Cursor account MCP connections</td><td><code>agent-sdk login</code> (stores a revocable key), <code>CURSOR_API_KEY</code>, or <code>--api-key</code> / <code>serve({ apiKey })</code></td></tr><tr><td>Slack tokens</td><td>Slack channels</td><td><code>&lt;PREFIX&gt;_SLACK_BOT_TOKEN</code> + <code>&lt;PREFIX&gt;_SLACK_APP_TOKEN</code> per agent</td></tr><tr><td>GitHub webhook secret</td><td>delivery signature verification</td><td><code>GITHUB_WEBHOOK_SECRET</code>, same value on server and signer</td></tr><tr><td>GitHub API</td><td>outbound API calls</td><td>a GitHub App (<code>GITHUB_APP_ID</code> + <code>GITHUB_APP_PRIVATE_KEY</code> + installation id) or <code>GITHUB_TOKEN</code> / <code>gh auth login</code></td></tr><tr><td>MCP connection tokens</td><td>authored MCP connections</td><td>env vars your <code>mcp-connections/*.ts</code> read, or host OAuth secrets from <code>agent-sdk mcp oauth &lt;name&gt; --store</code> (<code>MCP_OAUTH_*</code>; see <a href="./guides/mcp-oauth.html">Host MCP OAuth</a>)</td></tr></tbody></table><p>Use a dedicated Cursor key per host. <code>agent-sdk whoami</code> shows the active credential. <code>logout</code> removes the stored key from the host; revoke the key in the Cursor dashboard to invalidate it. See <a href="./reference/cli.html#login-logout-whoami">CLI authentication</a> for credential resolution.</p><h3 id="state" tabindex="-1">State <a class="header-anchor" href="#state" aria-label="Permalink to &quot;State&quot;">​</a></h3><p>Place <code>--state-root</code> on a persistent volume outside the agent repository, and back it up. Sessions survive restarts only when their state does. See <a href="./storage.html">Storage</a> and <a href="./reference/sessions.html">Sessions</a> for persistence and layout details.</p><h3 id="a-single-box" tabindex="-1">A single box <a class="header-anchor" href="#a-single-box" aria-label="Permalink to &quot;A single box&quot;">​</a></h3><p>A single-host deployment needs one supervised <code>serve</code> process on a private network. Export the Cursor key and a generated bearer token in the supervisor environment:</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">export</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> CURSOR_API_KEY</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">=</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;&lt;cursor-api-key&gt;&quot;</span></span>
+<span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> delete</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> weather-agent</span></span></code></pre></div><p><code>stop</code> waits for the deployment to stop unless you pass <code>--no-wait</code>. <code>delete</code> removes the deployment and waits until it is gone. See the <a href="./reference/cli.html#deploy">CLI reference</a> for the full command reference.</p><h2 id="self-host-the-agent-sdk" tabindex="-1">Self-host the Agent SDK <a class="header-anchor" href="#self-host-the-agent-sdk" aria-label="Permalink to &quot;Self-host the Agent SDK&quot;">​</a></h2><p>The Agent SDK runs as a Node HTTP server on Node 22.13 or newer. You can host it on a VM or container platform.</p><h3 id="the-security-model-in-one-minute" tabindex="-1">The security model in one minute <a class="header-anchor" href="#the-security-model-in-one-minute" aria-label="Permalink to &quot;The security model in one minute&quot;">​</a></h3><p><code>serve</code> binds to loopback and admits direct local callers by default. Choose one of these options before exposing it:</p><ol><li>Pass <code>--bearer-token &lt;secret&gt;</code> for a shared host.</li><li>Define channel-specific auth for routes with their own credentials or signatures.</li><li>Use <code>--allow-anonymous</code> only behind an authenticating proxy.</li></ol><p>A static bearer token maps every holder to one principal. Use authored auth when callers need separate identities. See <a href="./reference/channels.html#auth-policies">Channels</a> for policy details.</p><h3 id="credentials" tabindex="-1">Credentials <a class="header-anchor" href="#credentials" aria-label="Permalink to &quot;Credentials&quot;">​</a></h3><p>A self-hosted server can read these credentials.</p><table tabindex="0"><thead><tr><th>Credential</th><th>Used for</th><th>Provide it as</th></tr></thead><tbody><tr><td>Cursor API key</td><td>model turns, cloud runtime, Cursor account MCP connections</td><td><code>agent-sdk login</code> (stores a revocable key), <code>CURSOR_API_KEY</code>, or <code>--api-key</code> / <code>serve({ apiKey })</code></td></tr><tr><td>Slack tokens</td><td>Slack channels</td><td><code>&lt;PREFIX&gt;_SLACK_BOT_TOKEN</code> + <code>&lt;PREFIX&gt;_SLACK_APP_TOKEN</code> per agent</td></tr><tr><td>GitHub webhook secret</td><td>delivery signature verification</td><td><code>GITHUB_WEBHOOK_SECRET</code>, same value on server and signer</td></tr><tr><td>GitHub API</td><td>outbound API calls</td><td>a GitHub App (<code>GITHUB_APP_ID</code> + <code>GITHUB_APP_PRIVATE_KEY</code> + installation id) or <code>GITHUB_TOKEN</code> / <code>gh auth login</code></td></tr><tr><td>MCP connection tokens</td><td>authored MCP connections</td><td>env vars your <code>mcp-connections/*.ts</code> read, or host OAuth secrets from <code>agent-sdk mcp oauth &lt;name&gt; --store</code> (<code>MCP_OAUTH_*</code>; see <a href="./guides/mcp-oauth.html">Host MCP OAuth</a>)</td></tr></tbody></table><p>Use a dedicated Cursor key per host. <code>agent-sdk whoami</code> shows the active credential. <code>logout</code> removes the stored key from the host; revoke the key in the Cursor dashboard to invalidate it. See <a href="./reference/cli.html#login-logout-whoami">CLI authentication</a> for credential resolution.</p><h3 id="state" tabindex="-1">State <a class="header-anchor" href="#state" aria-label="Permalink to &quot;State&quot;">​</a></h3><p>Place <code>--state-root</code> on a persistent volume outside the agent repository, and back it up. Sessions survive restarts only when their state does. See <a href="./storage.html">Storage</a> and <a href="./reference/sessions.html">Sessions</a> for persistence and layout details.</p><h3 id="a-single-box" tabindex="-1">A single box <a class="header-anchor" href="#a-single-box" aria-label="Permalink to &quot;A single box&quot;">​</a></h3><p>A single-host deployment needs one supervised <code>serve</code> process on a private network. Export the Cursor key and a generated bearer token in the supervisor environment:</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">export</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> CURSOR_API_KEY</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">=</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;&lt;cursor-api-key&gt;&quot;</span></span>
 <span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">export</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> AGENT_SDK_BEARER_TOKEN</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">=</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;$(</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">openssl</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> rand </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">-hex</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> 32</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">)&quot;</span></span>
 <span class="line"></span>
 <span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;"># the server: all agents under one port</span></span>
 <span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> serve</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> /srv/agents</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --port</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> 3000</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --host</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> 127.0.0.1</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> \\</span></span>
-<span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  --state-root</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> /var/lib/agent-serve</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> \\</span></span>
+<span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  --state-root</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> /var/lib/agent-sdk</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> \\</span></span>
 <span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  --bearer-token</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">$AGENT_SDK_BEARER_TOKEN</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;</span></span></code></pre></div><p>Slack Socket Mode needs no inbound network. For GitHub, prefer <code>--cursor-events --repo owner/repo</code> on <code>serve</code> so the host pulls events through Cursor without a public webhook URL.</p><p>Webhook forwarding is the fallback. It needs one additional process. Before starting or restarting <code>serve</code>, export the same strong <code>GITHUB_WEBHOOK_SECRET</code> in both supervisor environments. Then install the extension, authenticate <code>gh</code>, and start the forwarder. Repository forwarding requires repo-admin access; organization forwarding with <code>--org</code> requires org-owner access.</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">export</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> GITHUB_WEBHOOK_SECRET</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">=</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;$(</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">openssl</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> rand </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">-hex</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> 32</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">)&quot;</span></span>
 <span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> github</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> doctor</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --install</span></span>
 <span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">gh</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> auth</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> login</span></span>
 <span class="line"></span>
 <span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;"># GitHub agents only: ONE forwarder relaying live deliveries to loopback</span></span>
 <span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">GITHUB_TOKEN</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">=</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> GH_TOKEN</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">=</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> \\</span></span>
-<span class="line"><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">  agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> github</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> forward</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> /srv/agents</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --repo</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> owner/repo</span></span></code></pre></div><p>Run long-lived processes under a supervisor. systemd survives reboots; tmux survives only SSH disconnects. On a TTY, press Enter to reload agent code. Humans reach the playground through a private network or tunnel. Keep <code>--bearer-token</code> on because tunneled requests arrive from loopback and IP-based policies can&#39;t tell them apart.</p><p>Health checks: <code>GET /v1/health</code> at the host level (made for ALB and ECS checks), and each agent also serves <code>/&lt;slug&gt;/v1/health</code>.</p><h3 id="containers" tabindex="-1">Containers <a class="header-anchor" href="#containers" aria-label="Permalink to &quot;Containers&quot;">​</a></h3><p>Build the image with Node 22.13 or newer, the agent source, and its package dependencies. Run <code>agent-sdk serve</code> as a non-root user:</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> serve</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> /srv/agents</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --mode</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> multi</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> \\</span></span>
+<span class="line"><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">  agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> github</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> forward</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> /srv/agents</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --repo</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> owner/repo</span></span></code></pre></div><p>Run long-lived processes under a supervisor. systemd survives reboots; tmux survives only SSH disconnects. On a TTY, press Enter to reload agent code. Humans reach the playground through a private network or tunnel. Keep <code>--bearer-token</code> on because tunneled requests arrive from loopback and IP-based policies can&#39;t tell them apart.</p><p>Health checks: <code>GET /v1/health</code> at the host level, and each agent also serves <code>/&lt;slug&gt;/v1/health</code>.</p><h3 id="containers" tabindex="-1">Containers <a class="header-anchor" href="#containers" aria-label="Permalink to &quot;Containers&quot;">​</a></h3><p>Build the image with Node 22.13 or newer, the agent source, and its package dependencies. Run <code>agent-sdk serve</code> as a non-root user:</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> serve</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> /srv/agents</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --mode</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> multi</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> \\</span></span>
 <span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  --host</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> 0.0.0.0</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --port</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> 3000</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> \\</span></span>
-<span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  --state-root</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> /var/lib/agent-serve</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> \\</span></span>
+<span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  --state-root</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> /var/lib/agent-sdk</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> \\</span></span>
 <span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  --bearer-token</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">$AGENT_SDK_BEARER_TOKEN</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;</span></span></code></pre></div><p>Mount the state root as a persistent volume and inject secrets at startup. Install <code>git</code> and <code>gh</code> when channels need host-side GitHub work. Don&#39;t put secrets in the image.</p><h3 id="serve-many-agents-from-one-process" tabindex="-1">Serve many agents from one process <a class="header-anchor" href="#serve-many-agents-from-one-process" aria-label="Permalink to &quot;Serve many agents from one process&quot;">​</a></h3><p>Point <code>serve</code> at a folder of agent projects and every child mounts under its directory name on one port. One process, one state root, one credential:</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> serve</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> /srv/agents</span></span>
-<span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;"># index at /, each agent at /&lt;slug&gt;/v1/*, /&lt;slug&gt;/playground</span></span></code></pre></div><p>Only mount what you mean to run. Every mounted agent&#39;s channels are live, and webhook-driven agents spend model budget on every wake. <code>--mode single</code> serves exactly one agent at the unslugged <code>/v1/*</code> when the agent is the whole host. See the <a href="./reference/http-api.html">HTTP API</a> for route layout and the <a href="./guides/slack.html">Slack guide</a> for multi-agent token setup.</p><h3 id="the-production-flags" tabindex="-1">The production flags <a class="header-anchor" href="#the-production-flags" aria-label="Permalink to &quot;The production flags&quot;">​</a></h3><p>Use these settings in production:</p><table tabindex="0"><thead><tr><th>Flag</th><th>In production</th></tr></thead><tbody><tr><td><code>--dev</code></td><td>Leave off. Dev mode admits unsigned loopback GitHub deliveries, widens playground session listing on loopback, and never auto-fires schedules.</td></tr><tr><td><code>--bearer-token</code></td><td>Set on shared hosts unless an authenticating proxy is the trust boundary and you use <code>--allow-anonymous</code> instead.</td></tr><tr><td><code>--allow-anonymous</code></td><td>Use only behind an authenticating network boundary. It also widens playground session access so Slack and webhook sessions appear.</td></tr><tr><td><code>--state-root</code></td><td>Place on a persistent volume outside any repo.</td></tr><tr><td><code>--public-url</code></td><td>Set when cloud-runtime turns must call back into peers on this host.</td></tr><tr><td><code>--no-playground</code></td><td>Set when no human needs the UI.</td></tr><tr><td><code>--no-docs</code></td><td>Set to remove the documentation site at <code>/docs</code>.</td></tr><tr><td><code>--no-schedules</code></td><td>Set on secondary hosts so schedules run exactly once.</td></tr></tbody></table><p>Schedules fire on their cron cadence (UTC) in production mode. They have no cross-host coordination, so enable them on exactly one serving process per project.</p><h3 id="restarts-and-upgrades" tabindex="-1">Restarts and upgrades <a class="header-anchor" href="#restarts-and-upgrades" aria-label="Permalink to &quot;Restarts and upgrades&quot;">​</a></h3><p>Restarts preserve sessions, event streams, and SDK conversation state under the state root. Parked approvals and in-memory reminders don&#39;t survive a restart; re-run or recreate them afterward.</p><h3 id="observability" tabindex="-1">Observability <a class="header-anchor" href="#observability" aria-label="Permalink to &quot;Observability&quot;">​</a></h3><p>Use <a href="./reference/cli.html#logs"><code>agent-sdk logs</code></a> for runtime output, <a href="./guides/opentelemetry.html">OpenTelemetry</a> for OTLP traces and metrics, <a href="./reference/hooks.html">hooks</a> for in-process subscribers, and <a href="./reference/sessions.html#how-do-i-inspect-a-saved-event-stream">session traces</a> for incident review.</p><h2 id="what-s-next" tabindex="-1">What&#39;s next <a class="header-anchor" href="#what-s-next" aria-label="Permalink to &quot;What&#39;s next&quot;">​</a></h2><p>Continue with these pages:</p><ul><li><a href="./reference/cli.html#deploy">CLI reference</a>: deploy, inspect, stop, and rotate hosted agents</li><li><a href="./storage.html">Storage</a>: preserve supported records across engine replacements</li><li><a href="./reference/channels.html#auth-policies">Channels</a>: the auth policies in detail</li><li><a href="./guides/github.html">GitHub guide</a>: delivery paths without a public URL</li><li><a href="./troubleshooting.html">Troubleshooting</a>: the symptom table for when a deploy misbehaves</li></ul>`,108)])])}const g=e(n,[["render",h]]);export{c as __pageData,g as default};
+<span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;"># index at /, each agent at /&lt;slug&gt;/v1/*, /&lt;slug&gt;/playground</span></span></code></pre></div><p>Only mount what you mean to run. Every mounted agent&#39;s channels are live, and webhook-driven agents spend model budget on every wake. <code>--mode single</code> serves exactly one agent at the unslugged <code>/v1/*</code> when the agent is the whole host. See the <a href="./reference/http-api.html">HTTP API</a> for route layout and the <a href="./guides/slack.html">Slack guide</a> for multi-agent token setup.</p><h3 id="the-production-flags" tabindex="-1">The production flags <a class="header-anchor" href="#the-production-flags" aria-label="Permalink to &quot;The production flags&quot;">​</a></h3><p>Use these settings in production:</p><table tabindex="0"><thead><tr><th>Flag</th><th>In production</th></tr></thead><tbody><tr><td><code>--dev</code></td><td>Leave off. Dev mode admits unsigned loopback GitHub deliveries, widens playground session listing on loopback, and never auto-fires schedules.</td></tr><tr><td><code>--bearer-token</code></td><td>Set on shared hosts unless an authenticating proxy is the trust boundary and you use <code>--allow-anonymous</code> instead.</td></tr><tr><td><code>--allow-anonymous</code></td><td>Use only behind an authenticating network boundary. It also widens playground session access so Slack and webhook sessions appear.</td></tr><tr><td><code>--state-root</code></td><td>Place on a persistent volume outside any repo.</td></tr><tr><td><code>--public-url</code></td><td>Set when cloud-runtime turns must call back into peers on this host.</td></tr><tr><td><code>--no-playground</code></td><td>Set when no human needs the UI.</td></tr><tr><td><code>--no-docs</code></td><td>Set to remove the documentation site at <code>/docs</code>.</td></tr><tr><td><code>--no-schedules</code></td><td>Set on secondary hosts so schedules run exactly once.</td></tr></tbody></table><p>Schedules fire on their cron cadence (UTC) in production mode. They have no cross-host coordination, so enable them on exactly one serving process per project.</p><h3 id="restarts-and-upgrades" tabindex="-1">Restarts and upgrades <a class="header-anchor" href="#restarts-and-upgrades" aria-label="Permalink to &quot;Restarts and upgrades&quot;">​</a></h3><p>Restarts preserve sessions, event streams, and SDK conversation state under the state root. Parked approvals and in-memory reminders don&#39;t survive a restart; re-run or recreate them afterward.</p><h3 id="observability" tabindex="-1">Observability <a class="header-anchor" href="#observability" aria-label="Permalink to &quot;Observability&quot;">​</a></h3><p>Use <a href="./reference/cli.html#logs"><code>agent-sdk logs</code></a> for runtime output, <a href="./guides/opentelemetry.html">OpenTelemetry</a> for OTLP traces and metrics, <a href="./reference/hooks.html">hooks</a> for in-process subscribers, and <a href="./reference/sessions.html#how-do-i-inspect-a-saved-event-stream">session traces</a> for incident review.</p><h2 id="what-s-next" tabindex="-1">What&#39;s next <a class="header-anchor" href="#what-s-next" aria-label="Permalink to &quot;What&#39;s next&quot;">​</a></h2><p>Continue with these pages:</p><ul><li><a href="./reference/cli.html#deploy">CLI reference</a>: deploy, inspect, stop, and rotate hosted agents</li><li><a href="./storage.html">Storage</a>: preserve supported records across engine replacements</li><li><a href="./reference/channels.html#auth-policies">Channels</a>: the auth policies in detail</li><li><a href="./guides/github.html">GitHub guide</a>: delivery paths without a public URL</li><li><a href="./troubleshooting.html">Troubleshooting</a>: the symptom table for when a deploy misbehaves</li></ul>`,104)])])}const g=s(n,[["render",o]]);export{c as __pageData,g as default};

package/dist/docs/assets/{deployment.md.DX_hc3ze.lean.js → deployment.md.D9msOFOW.lean.js}

@@ -1 +1 @@
-import{_ as e,c as a,o as t,ag as i}from"./chunks/framework.BCISBCiQ.js";const c=JSON.parse('{"title":"Deployment","description":"Deploy the Agent SDK with Cursor-managed hosting or on infrastructure you control.","frontmatter":{"title":"Deployment","description":"Deploy the Agent SDK with Cursor-managed hosting or on infrastructure you control."},"headers":[],"relativePath":"deployment.md","filePath":"deployment.md"}'),n={name:"deployment.md"};function h(o,s,l,r,d,p){return t(),a("div",null,[...s[0]||(s[0]=[i("",108)])])}const g=e(n,[["render",h]]);export{c as __pageData,g as default};
+import{_ as s,c as a,o as t,ag as i}from"./chunks/framework.BCISBCiQ.js";const c=JSON.parse('{"title":"Deployment","description":"Deploy the Agent SDK with Cursor-managed hosting or on infrastructure you control.","frontmatter":{"title":"Deployment","description":"Deploy the Agent SDK with Cursor-managed hosting or on infrastructure you control."},"headers":[],"relativePath":"deployment.md","filePath":"deployment.md"}'),n={name:"deployment.md"};function o(h,e,l,r,d,p){return t(),a("div",null,[...e[0]||(e[0]=[i("",104)])])}const g=s(n,[["render",o]]);export{c as __pageData,g as default};

package/dist/docs/assets/{evals.md.a0SMN6r9.js → evals.md.lfJoEVc8.js}

@@ -1,4 +1,4 @@
-import{_ as e,c as a,o as i,ag as t}from"./chunks/framework.BCISBCiQ.js";const k=JSON.parse('{"title":"Evals","description":"Define repeatable checks with defineEval, run them with agent-sdk eval, and use them as regression checks.","frontmatter":{"title":"Evals","description":"Define repeatable checks with defineEval, run them with agent-sdk eval, and use them as regression checks."},"headers":[],"relativePath":"evals.md","filePath":"evals.md"}'),n={name:"evals.md"};function l(d,s,o,h,r,p){return i(),a("div",null,[...s[0]||(s[0]=[t(`<h1 id="evals" tabindex="-1">Evals <a class="header-anchor" href="#evals" aria-label="Permalink to &quot;Evals&quot;">​</a></h1><p>An eval is a repeatable check that runs your agent against a fixed input and gates the recorded trajectory: the run completed, the right tool ran, the reply has the right shape. Evals are how you know a prompt tweak helped, a refactor didn&#39;t regress the agent, and last month&#39;s fix is still holding.</p><p>Evals exercise the same surface your users hit. The runner starts (or targets) a real agent server, drives sessions over the public API, and grades what comes back. A passing eval means the agent started, accepted a message, and did what you asserted.</p><div class="note custom-block github-alert"><p class="custom-block-title">NOTE</p><p>Import paths here use <code>@cursor/july/evals</code>. On projects still using <code>@anysphere/agent-serve</code>, swap the import and run <code>agent-serve eval</code>. See <a href="/docs/#run-the-cli">Run the CLI</a> for the full rename table.</p></div><h2 id="define-evals-with-defineeval" tabindex="-1">Define evals with <code>defineEval</code> <a class="header-anchor" href="#define-evals-with-defineeval" aria-label="Permalink to &quot;Define evals with \`defineEval\`&quot;">​</a></h2><p>The Agent SDK discovers evals under the project-root <code>evals/</code> directory, in <code>.eval.ts</code> or <code>.eval.js</code> files. That&#39;s a sibling of <code>agent/</code>, never inside it (<code>agent/evals/</code> is silently ignored). TypeScript is the normal authoring format.</p><p>The file path is the eval&#39;s identity, so you don&#39;t author an id. Directories group related evals: <code>evals/builds/api.eval.ts</code> becomes id <code>builds/api</code>. An <code>index</code> filename collapses to its directory, so <code>evals/builds/index.eval.ts</code> becomes <code>builds</code>.</p><p>An eval is a single <code>async test(t)</code>. You drive the agent with <code>t</code> and assert on the run with the same <code>t</code>:</p><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">// evals/readiness.eval.ts</span></span>
+import{_ as e,c as i,o as a,ag as t}from"./chunks/framework.BCISBCiQ.js";const k=JSON.parse('{"title":"Evals","description":"Define repeatable checks with defineEval, run them with agent-sdk eval, and use them as regression checks.","frontmatter":{"title":"Evals","description":"Define repeatable checks with defineEval, run them with agent-sdk eval, and use them as regression checks."},"headers":[],"relativePath":"evals.md","filePath":"evals.md"}'),n={name:"evals.md"};function l(d,s,h,o,r,p){return a(),i("div",null,[...s[0]||(s[0]=[t(`<h1 id="evals" tabindex="-1">Evals <a class="header-anchor" href="#evals" aria-label="Permalink to &quot;Evals&quot;">​</a></h1><p>An eval is a repeatable check that runs your agent against a fixed input and gates the recorded trajectory: the run completed, the right tool ran, the reply has the right shape. Evals are how you know a prompt tweak helped, a refactor didn&#39;t regress the agent, and last month&#39;s fix is still holding.</p><p>Evals exercise the same surface your users hit. The runner starts (or targets) a real agent server, drives sessions over the public API, and grades what comes back. A passing eval means the agent started, accepted a message, and did what you asserted.</p><h2 id="define-evals-with-defineeval" tabindex="-1">Define evals with <code>defineEval</code> <a class="header-anchor" href="#define-evals-with-defineeval" aria-label="Permalink to &quot;Define evals with \`defineEval\`&quot;">​</a></h2><p>The Agent SDK discovers evals under the project-root <code>evals/</code> directory, in <code>.eval.ts</code> or <code>.eval.js</code> files. That&#39;s a sibling of <code>agent/</code>, never inside it (<code>agent/evals/</code> is silently ignored). TypeScript is the normal authoring format.</p><p>The file path is the eval&#39;s identity, so you don&#39;t author an id. Directories group related evals: <code>evals/builds/api.eval.ts</code> becomes id <code>builds/api</code>. An <code>index</code> filename collapses to its directory, so <code>evals/builds/index.eval.ts</code> becomes <code>builds</code>.</p><p>An eval is a single <code>async test(t)</code>. You drive the agent with <code>t</code> and assert on the run with the same <code>t</code>:</p><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">// evals/readiness.eval.ts</span></span>
 <span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { defineEval, includes } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;@cursor/july/evals&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
 <span class="line"></span>
 <span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">export</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> default</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> defineEval</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">({</span></span>
@@ -47,8 +47,8 @@ import{_ as e,c as a,o as i,ag as t}from"./chunks/framework.BCISBCiQ.js";const k
 <span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">  // timeoutMs: 180_000,</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;"> // optional project-wide default</span></span>
 <span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">  // judge: { model: &quot;...&quot; },</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;"> // default judge model for t.judge.*</span></span>
 <span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">  // reporters: [],</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;"> // destinations that observe every case</span></span>
-<span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">  // maxPlaygroundRuns: 50,</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;"> // playground /v1/dev/evals history only (default 20)</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">});</span></span></code></pre></div><p>The timeout order is case or file <code>timeoutMs</code>, CLI <code>--timeout-ms</code>, project config <code>timeoutMs</code>, then the 180-second runner default.</p><p>The optional fields:</p><table tabindex="0"><thead><tr><th>Option</th><th>Default</th><th>Meaning</th></tr></thead><tbody><tr><td><code>timeoutMs</code></td><td><code>180_000</code></td><td>Project-wide per-case timeout</td></tr><tr><td><code>judge</code></td><td>unset</td><td>Default judge model for <code>t.judge.*</code>; see <a href="#judge-free-form-output">Judge free-form output</a></td></tr><tr><td><code>reporters</code></td><td>unset</td><td>Destinations that observe every case; <code>--skip-report</code> suppresses them</td></tr><tr><td><code>maxPlaygroundRuns</code></td><td><code>20</code></td><td>Max batches in the playground / <code>/v1/dev/evals*</code> history (not CLI <code>eval</code>)</td></tr></tbody></table><p>Reporters come from <code>@cursor/july/evals/reporters</code>: <code>JUnit</code> writes a JUnit XML file for CI, <code>Artifacts</code> writes per-case files, and <code>combineReporters</code> merges several into one (<code>renderJUnitXml</code> renders the XML for a custom destination). A file or case can add its own <code>reporters</code> on top of the config list.</p><p>Playground batches survive restarts whenever <code>agent/storage.ts</code> exists with an <code>evals</code> table or a KV core providing <code>delete</code> and <code>list</code> (the table is derived over the core); see <a href="./storage.html#eval-and-a-b-tables">Storage</a>. Without storage they live in process memory and disappear when <code>serve</code> exits — navigating away and back still works while the process is up.</p><h2 id="drive-and-assert-with-t" tabindex="-1">Drive and assert with <code>t</code> <a class="header-anchor" href="#drive-and-assert-with-t" aria-label="Permalink to &quot;Drive and assert with \`t\`&quot;">​</a></h2><p><code>t</code> is both the driver and the assertion surface. You write ordinary control flow, sending turns and asserting inline.</p><p>Drive the agent with <code>t.send(message, options?)</code>. It runs one turn and waits for the session to park or fail. Multiple sends in one case share the session, which is how you write multi-turn evals.</p><p>Each <code>t.send</code> resolves to a turn result with <code>message</code>, <code>sessionId</code>, <code>events</code>, <code>toolCalls</code>, <code>ok</code>, and <code>index</code>. The turn carries the same assertion vocabulary as <code>t</code>, scoped to that turn, so you can grade an intermediate turn before the next send overwrites <code>t.reply</code>. <code>turn.expectOk()</code> throws when the turn failed, for later steps that depend on it.</p><p>Read the full case state with <code>t.reply</code> (the last assistant text), <code>t.events</code> (every captured session event across turns), <code>t.turns</code> (settled turns, oldest first), and <code>t.sessionId</code>. <code>t.signal</code> aborts when the case hits its timeout; pass it to your own async work.</p><p>Assert with the gates:</p><table tabindex="0"><thead><tr><th>Gate</th><th>Checks</th></tr></thead><tbody><tr><td><code>t.succeeded()</code></td><td>the run did not fail and is not parked on an unanswered approval</td></tr><tr><td><code>t.parked()</code></td><td>the run cleanly parked on an unanswered approval request</td></tr><tr><td><code>t.messageIncludes(token)</code></td><td>the joined assistant text matches a string or <code>RegExp</code></td></tr><tr><td><code>t.calledTool(name, matcher?)</code></td><td>a matching call to <code>name</code> happened</td></tr><tr><td><code>t.notCalledTool(name)</code></td><td>no request for <code>name</code>, in any lifecycle state</td></tr><tr><td><code>t.loadedSkill(name)</code></td><td>the agent opened the skill&#39;s <code>SKILL.md</code> (read, grep, or shell <code>cat</code>)</td></tr><tr><td><code>t.toolOrder(names)</code></td><td>tool requests appear in this relative order (extra calls allowed)</td></tr><tr><td><code>t.usedNoTools()</code></td><td>no tool calls at all</td></tr><tr><td><code>t.maxToolCalls(max)</code></td><td>at most <code>max</code> tool calls</td></tr><tr><td><code>t.noFailedActions()</code></td><td>no tool call reported an error</td></tr><tr><td><code>t.calledSubagent(name, matcher?)</code></td><td>a matching subagent delegation happened</td></tr><tr><td><code>t.taggedArtifact(kind?, predicate?)</code></td><td>at least one <a href="./reference/artifacts.html">artifact</a> was tagged</td></tr><tr><td><code>t.event(type, matcher?)</code></td><td>at least one matching event of <code>type</code> occurred</td></tr><tr><td><code>t.notEvent(type, matcher?)</code></td><td>no matching event of <code>type</code> occurred</td></tr><tr><td><code>t.eventOrder(matchers)</code></td><td>matching event groups occur in this relative order</td></tr><tr><td><code>t.eventsSatisfy(label, predicate)</code></td><td>your predicate over the typed event stream</td></tr><tr><td><code>t.check(value, expectation)</code></td><td>any value, against a builder</td></tr><tr><td><code>t.score(name, value)</code></td><td>records a 0–1 score you computed; soft until you add a bar</td></tr><tr><td><code>t.requireToolCall(name, matcher?)</code></td><td>gates on a matching call and returns it, so later code can read its input and output</td></tr><tr><td><code>t.requireInputRequest(filter?)</code></td><td>gates on exactly one pending approval request and returns it</td></tr></tbody></table><p>Every gate returns a handle: <code>.soft()</code> demotes it to tracked-only, <code>.atLeast(0.7)</code> adds a soft score bar, and <code>.gate(0.8)</code> promotes a scored assertion into a hard gate.</p><p>With no matcher, <code>calledTool</code> is request-based: a requested call counts even when its result has not arrived. Pass <code>t.calledTool(&quot;inspect_pr&quot;, { status: &quot;completed&quot; })</code> to require the call to return. <code>input</code>, <code>output</code>, and <code>count</code> matcher fields accept a literal, a <code>RegExp</code>, or a predicate.</p><p>The expectation builders are <code>includes(string | RegExp)</code>, <code>equals(value)</code>, <code>matches(schema)</code>, <code>similarity(expected)</code>, and <code>satisfies(predicate, label)</code>. <code>includes</code> stringifies its input, <code>equals</code> compares values deeply, <code>matches</code> validates against a Standard Schema (or anything with <code>safeParse</code>, like Zod), <code>similarity</code> scores normalized text similarity, and <code>satisfies</code> runs your predicate. The plain function <code>normalizedSimilarity(actual, expected)</code> returns the same 0–1 score for use with <code>t.score</code>.</p><p>A few more context members shape a case: <code>t.require(value, expectation)</code> records a gate and stops the test body when it fails, without a duplicate execution error. <code>t.skip(reason)</code> ends the case as skipped (reported separately, never changes the exit code; call it before sending messages). <code>t.metric(name, value)</code> records a structured score for the playground case card. <code>t.log(message)</code> records a debug line for the CLI and playground result.</p><p>Three <code>t.send</code> options apply on session create (first <code>t.send</code> only):</p><ul><li><code>workspaceFiles</code> — <code>{ path: contents }</code>, seeded into the local session workspace. Prefer this over machine-local paths.</li><li><code>workspaceDir</code> — absolute harness cwd (local runtime).</li><li><code>cloud</code> — per-session cloud options merged over the agent&#39;s static <code>cloud</code> config (repos / env / …). Use a pinned <code>repos</code> override to attach a fixture repo for cloud evals without putting it on the agent&#39;s default <code>cloud.repos</code>. Cloud ignores <code>workspaceFiles</code> seeds.</li></ul><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">const</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> toolResults</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> =</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> t.events.</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">filter</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">((</span><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">e</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">) </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">=&gt;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> e.type </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">===</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;action.result&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">);</span></span>
+<span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">  // maxPlaygroundRuns: 50,</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;"> // playground history only (default 20)</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">});</span></span></code></pre></div><p>The timeout order is case or file <code>timeoutMs</code>, CLI <code>--timeout-ms</code>, project config <code>timeoutMs</code>, then the 180-second runner default.</p><p>The optional fields:</p><table tabindex="0"><thead><tr><th>Option</th><th>Default</th><th>Meaning</th></tr></thead><tbody><tr><td><code>timeoutMs</code></td><td><code>180_000</code></td><td>Project-wide per-case timeout</td></tr><tr><td><code>judge</code></td><td>unset</td><td>Default judge model for <code>t.judge.*</code>; see <a href="#judge-free-form-output">Judge free-form output</a></td></tr><tr><td><code>reporters</code></td><td>unset</td><td>Destinations that observe every case; <code>--skip-report</code> suppresses them</td></tr><tr><td><code>maxPlaygroundRuns</code></td><td><code>20</code></td><td>Max batches in the playground / <code>/v1/dev/evals*</code> history (not CLI <code>eval</code>). Hard-capped at 500.</td></tr></tbody></table><p>Reporters come from <code>@cursor/july/evals/reporters</code>: <code>JUnit</code> writes a JUnit XML file for CI, <code>Artifacts</code> writes per-case files, and <code>combineReporters</code> merges several into one (<code>renderJUnitXml</code> renders the XML for a custom destination). A file or case can add its own <code>reporters</code> on top of the config list.</p><p>Playground batches survive restarts whenever <code>agent/storage.ts</code> exists with an <code>evals</code> table or a KV core providing <code>delete</code> and <code>list</code> (the table is derived over the core); see <a href="./storage.html#eval-and-a-b-tables">Storage</a>. Without storage they live in process memory and disappear when <code>serve</code> exits. Navigating away and back still works while the process is up.</p><h2 id="drive-and-assert-with-t" tabindex="-1">Drive and assert with <code>t</code> <a class="header-anchor" href="#drive-and-assert-with-t" aria-label="Permalink to &quot;Drive and assert with \`t\`&quot;">​</a></h2><p><code>t</code> is both the driver and the assertion surface. You write ordinary control flow, sending turns and asserting inline.</p><p>Drive the agent with <code>t.send(message, options?)</code>. It runs one turn and waits for the session to park or fail. Multiple sends in one case share the session, which is how you write multi-turn evals.</p><p>Each <code>t.send</code> resolves to a turn result with <code>message</code>, <code>sessionId</code>, <code>events</code>, <code>toolCalls</code>, <code>ok</code>, and <code>index</code>. The turn carries the same assertion vocabulary as <code>t</code>, scoped to that turn, so you can grade an intermediate turn before the next send overwrites <code>t.reply</code>. <code>turn.expectOk()</code> throws when the turn failed, for later steps that depend on it.</p><p>Read the full case state with <code>t.reply</code> (the last assistant text), <code>t.events</code> (session events captured so far), <code>t.turns</code> (settled turns, oldest first), and <code>t.sessionId</code>. <code>t.signal</code> aborts when the case hits its timeout; pass it to your own async work.</p><p>Assert with the gates:</p><table tabindex="0"><thead><tr><th>Gate</th><th>Checks</th></tr></thead><tbody><tr><td><code>t.succeeded()</code></td><td>the run did not fail and is not parked on an unanswered approval</td></tr><tr><td><code>t.parked()</code></td><td>the run cleanly parked on an unanswered approval request</td></tr><tr><td><code>t.messageIncludes(token)</code></td><td>the joined assistant text matches a string or <code>RegExp</code></td></tr><tr><td><code>t.calledTool(name, matcher?)</code></td><td>a matching call to <code>name</code> happened</td></tr><tr><td><code>t.notCalledTool(name)</code></td><td>no request for <code>name</code>, in any lifecycle state</td></tr><tr><td><code>t.loadedSkill(name)</code></td><td>the agent opened the skill&#39;s <code>SKILL.md</code> (read, grep, or shell <code>cat</code>)</td></tr><tr><td><code>t.toolOrder(names)</code></td><td>tool requests appear in this relative order (extra calls allowed)</td></tr><tr><td><code>t.usedNoTools()</code></td><td>no tool calls at all</td></tr><tr><td><code>t.maxToolCalls(max)</code></td><td>at most <code>max</code> tool calls</td></tr><tr><td><code>t.noFailedActions()</code></td><td>no tool call reported an error</td></tr><tr><td><code>t.calledSubagent(name, matcher?)</code></td><td>a matching subagent delegation happened</td></tr><tr><td><code>t.taggedArtifact(kind?, predicate?)</code></td><td>at least one <a href="./reference/artifacts.html">artifact</a> was tagged</td></tr><tr><td><code>t.event(type, matcher?)</code></td><td>at least one matching event of <code>type</code> occurred</td></tr><tr><td><code>t.notEvent(type, matcher?)</code></td><td>no matching event of <code>type</code> occurred</td></tr><tr><td><code>t.eventOrder(matchers)</code></td><td>matching event groups occur in this relative order</td></tr><tr><td><code>t.eventsSatisfy(label, predicate)</code></td><td>your predicate over the typed event stream</td></tr><tr><td><code>t.check(value, expectation)</code></td><td>any value, against a builder</td></tr><tr><td><code>t.score(name, value)</code></td><td>records a 0–1 score you computed; soft until you add a bar</td></tr><tr><td><code>t.requireToolCall(name, matcher?)</code></td><td>gates on a matching call and returns it, so later code can read its input and output</td></tr><tr><td><code>t.requireInputRequest(filter?)</code></td><td>gates on exactly one pending approval request and returns it</td></tr></tbody></table><p>Every gate returns a handle: <code>.soft()</code> demotes it to tracked-only, <code>.atLeast(0.7)</code> adds a soft score bar, and <code>.gate(0.8)</code> promotes a scored assertion into a hard gate.</p><p>With no matcher, <code>calledTool</code> is request-based: a requested call counts even when its result has not arrived. Pass <code>t.calledTool(&quot;inspect_pr&quot;, { status: &quot;completed&quot; })</code> to require the call to return. <code>input</code>, <code>output</code>, and <code>count</code> matcher fields accept a literal, a <code>RegExp</code>, or a predicate.</p><p>The expectation builders are <code>includes(string | RegExp)</code>, <code>equals(value)</code>, <code>matches(schema)</code>, <code>similarity(expected)</code>, and <code>satisfies(predicate, label)</code>. <code>includes</code> stringifies its input, <code>equals</code> compares values deeply, <code>matches</code> validates against a Standard Schema (or anything with <code>safeParse</code>, like Zod), <code>similarity</code> scores normalized text similarity, and <code>satisfies</code> runs your predicate. The plain function <code>normalizedSimilarity(actual, expected)</code> returns the same 0–1 score for use with <code>t.score</code>.</p><p>A few more context members shape a case: <code>t.require(value, expectation)</code> records a gate and stops the test body when it fails, without a duplicate execution error. <code>t.skip(reason)</code> ends the case as skipped (reported separately, never changes the exit code; call it before sending messages). <code>t.metric(name, value)</code> records a structured score for the playground case card. <code>t.log(message)</code> records a debug line for the CLI and playground result.</p><p>Three <code>t.send</code> options apply on session create (first <code>t.send</code> only):</p><ul><li><code>workspaceFiles</code>: <code>{ path: contents }</code>, seeded into the local session workspace. Prefer this over machine-local paths.</li><li><code>workspaceDir</code>: absolute harness cwd (local runtime).</li><li><code>cloud</code>: per-session cloud options merged over the agent&#39;s static <code>cloud</code> config (repos / env / …). Use a pinned <code>repos</code> override to attach a fixture repo for cloud evals without putting it on the agent&#39;s default <code>cloud.repos</code>. Cloud ignores <code>workspaceFiles</code> seeds.</li></ul><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">const</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> toolResults</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> =</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> t.events.</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">filter</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">((</span><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">e</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">) </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">=&gt;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> e.type </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">===</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;action.result&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">);</span></span>
 <span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">t.</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">check</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span></span>
 <span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  toolResults.</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">length</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
 <span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">  satisfies</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">((</span><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">n</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">) </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">=&gt;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> (n </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">as</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> number</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">) </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">&lt;=</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> 4</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">, </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;at most 4 tool calls&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">)</span></span>
@@ -58,7 +58,7 @@ import{_ as e,c as a,o as i,ag as t}from"./chunks/framework.BCISBCiQ.js";const k
 <span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> eval</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> builds</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> search</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">                  # several ids or prefixes</span></span>
 <span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> eval</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --tag</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> smoke</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --tag</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> pull-request</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;"> # any matching tag</span></span>
 <span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> eval</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --json</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --no-stream</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">             # machine-readable results</span></span>
-<span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> eval</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --verbose</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">                      # logs + reply snippets</span></span></code></pre></div><p>Id filters use OR semantics. Each filter selects an exact id and its descendants. For example, <code>builds</code> selects <code>builds</code>, <code>builds/checkout</code>, and every other case below that path. Repeated tags also use OR semantics. When you provide both ids and tags, a case must match both groups.</p><p><code>eval</code> boots an ephemeral server on port 0 with a temp state root outside the project, so cases don&#39;t inherit ambient monorepo rules and don&#39;t pollute <code>.agent-serve/</code>. Point <code>--url</code> at a running server to eval a live agent instead:</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> eval</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> \\</span></span>
+<span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> eval</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --verbose</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">                      # logs + reply snippets</span></span></code></pre></div><p>Id filters use OR semantics. Each filter selects an exact id and its descendants. For example, <code>builds</code> selects <code>builds</code>, <code>builds/checkout</code>, and every other case below that path. Repeated tags also use OR semantics. When you provide both ids and tags, a case must match both groups.</p><p><code>eval</code> boots an ephemeral server on port 0 with a temp state root outside the project, so cases don&#39;t inherit ambient monorepo rules and don&#39;t write into the project state directory. Point <code>--url</code> at a running server to eval a live agent instead:</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> eval</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> \\</span></span>
 <span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  --url</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> http://127.0.0.1:3000/weather-agent</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> \\</span></span>
 <span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  --bearer-token</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">$AGENT_TOKEN</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;</span></span></code></pre></div><p>The eval definitions still come from <code>--dir</code>; <code>--url</code> only changes the agent that receives the turns. For a locally mounted multi-agent directory, <code>--slug weather-agent</code> chooses the target. Use <code>--state-root</code> to keep ephemeral session state at a chosen path, <code>--timeout-ms</code> to override the project timeout, and <code>--no-stream</code> to keep live progress off stderr. A TTY streams turn progress by default. <code>--verbose</code> still writes <code>t.log</code> lines to stderr and adds reply snippets to text results.</p><p>Model turns need a Cursor credential from <code>agent-sdk login</code> or <code>CURSOR_API_KEY</code>.</p><p>See <a href="./reference/cli.html#eval">CLI: eval</a> for flags and exit codes.</p><h3 id="json-results" tabindex="-1">JSON results <a class="header-anchor" href="#json-results" aria-label="Permalink to &quot;JSON results&quot;">​</a></h3><p>Use <code>--json --no-stream</code> in scripts and CI. The top-level result carries the totals and one result per case:</p><div class="language-json vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">json</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">{</span></span>
 <span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  &quot;ok&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">true</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
@@ -76,10 +76,10 @@ import{_ as e,c as a,o as i,ag as t}from"./chunks/framework.BCISBCiQ.js";const k
 <span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">      &quot;durationMs&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">12340</span></span>
 <span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">    }</span></span>
 <span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  ]</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">}</span></span></code></pre></div><p>Each case result can also include <code>description</code>, <code>finalText</code>, <code>tools</code>, <code>error</code>, and tool arguments or output. This shape lets CI report the failed assertion without parsing terminal text.</p><h2 id="run-evals-in-the-playground" tabindex="-1">Run evals in the playground <a class="header-anchor" href="#run-evals-in-the-playground" aria-label="Permalink to &quot;Run evals in the playground&quot;">​</a></h2><p>Start the server with <code>--dev</code>, open the playground, and choose <strong>Evals</strong>. You can run every case or one case, watch progress, and open the resulting session trace.</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> serve</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dev</span></span></code></pre></div><p>Playground runs target the live server instead of an ephemeral one. Their sessions appear in the session list. One eval batch can run at a time. Batches persist across restarts whenever <code>agent/storage.ts</code> provides an <code>evals</code> table or a KV core with <code>delete</code> and <code>list</code> (the table is derived over the core); without storage they are <strong>in-memory only</strong> (capped by <code>maxPlaygroundRuns</code>) — see <a href="./storage.html#eval-and-a-b-tables">Storage</a>.</p><p>The UI uses the playground eval routes (available without <code>--dev</code>): <code>GET /v1/dev/evals</code> lists datapoints and config (includes <code>maxPlaygroundRuns</code> / <code>durableRuns</code>), <code>GET /v1/dev/evals/runs</code> rehydrates recent batches after navigation, <code>POST /v1/dev/evals/runs</code> starts a batch (returns an <strong>Eval ID</strong> / <code>runId</code>), <code>GET /v1/dev/evals/runs/:runId</code> polls it, and <code>POST /v1/dev/evals/runs/:runId/cancel</code> cancels a running batch. See <a href="./reference/http-api.html#playground-eval-routes">Playground eval routes</a>. The start request returns <code>202</code> while cases run in the background. Poll until the snapshot status becomes <code>completed</code>, <code>failed</code>, or <code>cancelled</code>. Configuration errors appear on a failed snapshot.</p><p>On <code>--prod</code> / <code>--url</code>, the CLI prints the Eval ID as soon as the batch is accepted (and a Playground deep link with <code>?view=evals&amp;evalRunId=…</code>):</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> eval</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --prod</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --slug</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> vulnerability-scanner</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --tag</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> deepsec</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">}</span></span></code></pre></div><p>Each case result can also include <code>description</code>, <code>finalText</code>, <code>tools</code>, <code>error</code>, and tool arguments or output. This shape lets CI report the failed assertion without parsing terminal text.</p><h2 id="run-evals-in-the-playground" tabindex="-1">Run evals in the playground <a class="header-anchor" href="#run-evals-in-the-playground" aria-label="Permalink to &quot;Run evals in the playground&quot;">​</a></h2><p>Start the server, open the playground, and choose <strong>Evals</strong>. You can run every case or one case, watch progress, and open the resulting session trace. The Evals tab works on a normal <code>serve</code>.</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> serve</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span></span></code></pre></div><p>Playground runs target the live server instead of an ephemeral one. Their sessions appear in the session list. One eval batch can run at a time. Persistence follows the rule under <a href="#configure-eval-runs">Configure eval runs</a>. See <a href="./reference/http-api.html#playground-eval-routes">Playground eval routes</a>. The start request returns <code>202</code> while cases run in the background. Poll until the snapshot status becomes <code>completed</code>, <code>failed</code>, or <code>cancelled</code>. Configuration errors appear on a failed snapshot.</p><p>On <code>--prod</code> / <code>--url</code>, the CLI prints the Eval ID as soon as the batch is accepted (and a Playground deep link with <code>?view=evals&amp;evalRunId=…</code>):</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> eval</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --prod</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --slug</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> vulnerability-scanner</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --tag</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> deepsec</span></span>
 <span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;"># Eval ID: evalrun_…</span></span>
 <span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;"># Cancel:  agent-sdk eval cancel evalrun_… --prod --slug vulnerability-scanner</span></span>
 <span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;"># Playground: https://…/playground?view=evals&amp;evalRunId=evalrun_…</span></span>
 <span class="line"></span>
 <span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> eval</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> cancel</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> evalrun_…</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --prod</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --slug</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> vulnerability-scanner</span></span>
-<span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> eval</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> status</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> evalrun_…</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --prod</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --slug</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> vulnerability-scanner</span></span></code></pre></div><p>The Evals tab prefers the server’s in-flight batch (<code>activeRunId</code>) over a stale tab-local remembered id, so CLI / Slack kicks show up without an incognito window.</p><h2 id="what-good-cases-assert" tabindex="-1">What good cases assert <a class="header-anchor" href="#what-good-cases-assert" aria-label="Permalink to &quot;What good cases assert&quot;">​</a></h2><p>Gate decisions and shape, not prose. Model wording varies run to run. Tool choice, tool avoidance, and output structure are the stable contract.</p><ol><li><code>t.succeeded()</code>: always, first.</li><li>The tool decision: <code>calledTool</code> for the intended path, <code>notCalledTool</code> for the likely wrong alternative. The pair is stronger than either alone.</li><li>Output shape: a regex for the contract (<code>/ready|blocked/i</code>, a JSON marker, a findings-block fence), never exact sentences.</li><li>For structured output, parse <code>t.reply</code> and check fields with <code>satisfies</code> instead of substring-matching JSON.</li></ol><p>The common failure modes: asserting exact phrasing, packing more than about five gates into one case (split it), and cases that depend on live external state that drifts (pin the input; see fixtures).</p><h2 id="pick-fixtures-by-agent-type" tabindex="-1">Pick fixtures by agent type <a class="header-anchor" href="#pick-fixtures-by-agent-type" aria-label="Permalink to &quot;Pick fixtures by agent type&quot;">​</a></h2><p>The right fixture depends on the surface under test.</p><table tabindex="0"><thead><tr><th>Agent surface</th><th>Fixture</th></tr></thead><tbody><tr><td>Chat / domain assistant</td><td>A canonical prompt string, chosen once and frozen</td></tr><tr><td>Tool-heavy</td><td>Run <code>agent-sdk call &lt;tool&gt;</code> first to pin what the tool returns, then freeze the prompt that triggers it</td></tr><tr><td>GitHub webhook</td><td><code>agent-sdk github replay &lt;pr&gt; --events &#39;*&#39; --dry-run --out fixtures/github</code> snapshots real payloads for offline replay (<a href="./guides/github.html">GitHub guide</a>)</td></tr><tr><td>PR reviewer with host preparation</td><td>Diff, metadata, and gold labels pinned to commit SHAs; keep any live PR matrix small</td></tr><tr><td>Workspace-dependent</td><td><code>workspaceFiles</code> in <code>t.send</code> options, never developer-machine paths</td></tr></tbody></table><p>Tag the fast, reliably passing core <code>smoke</code> and run <code>--tag smoke</code> in the inner loop. Leave slow or flaky-prone cases untagged for explicit runs.</p><h3 id="materialize-api-backed-fixtures" tabindex="-1">Materialize API-backed fixtures <a class="header-anchor" href="#materialize-api-backed-fixtures" aria-label="Permalink to &quot;Materialize API-backed fixtures&quot;">​</a></h3><p>An input that only points at external data, such as a pull request URL, snapshot id, or pair of commit SHAs, is not self-contained. Fetch it once and commit the rendered fixture before you expand the suite.</p><ol><li>Save the diff, metadata, and labels under <code>fixtures/</code> at pinned revisions.</li><li>Seed those files with <code>workspaceFiles</code>, or read them from the fixture directory.</li><li>Assert decisions and output shape against the saved evidence.</li><li>Keep a small <code>smoke</code> subset for any remaining live pipeline checks.</li></ol><p>Read committed fixtures with <code>@cursor/july/evals/loaders</code>: <code>loadJson</code>, <code>loadJsonl</code>, and <code>loadYaml</code> resolve relative paths against the project root the runner discovered, not the cwd the CLI was invoked from (<code>resolveFixturePath</code> and <code>evalFixtureRoot</code> expose the same resolution for other file formats).</p><p><code>maxConcurrency</code> limits parallel datapoints. It does not limit model or API fan-out inside one datapoint. Materialized fixtures prevent a large suite from exhausting provider and GitHub rate limits. The <a href="./../skills/evals/SKILL.html">evals skill</a> has the full fixture workflow.</p><h2 id="keep-improvements-with-regression-evals" tabindex="-1">Keep improvements with regression evals <a class="header-anchor" href="#keep-improvements-with-regression-evals" aria-label="Permalink to &quot;Keep improvements with regression evals&quot;">​</a></h2><p>Every <a href="./hillclimbing.html">hillclimb</a> round that keeps a change must land an eval that would have failed before the change. If you can&#39;t express the improvement as a gate (a <code>calledTool</code> shift, a bounded <code>action.result</code> count, an output-shape regex), the improvement is unverified, and it&#39;ll regress silently.</p><p>The rule cuts the other way too: never weaken an existing gate to make a round pass. That&#39;s the freeze line moving, and it turns your regression suite into a list of checks that no longer protect anything.</p><h2 id="compare-variants-on-live-traffic" tabindex="-1">Compare variants on live traffic <a class="header-anchor" href="#compare-variants-on-live-traffic" aria-label="Permalink to &quot;Compare variants on live traffic&quot;">​</a></h2><p>Use <code>defineAB</code> to compare variant metrics on live sessions. It is not a test runner and has no <code>agent-sdk ab</code> command. Keep <code>defineEval</code> as the regression ratchet. Eval sessions do not enroll or change live metrics. See <a href="./ab.html">Live A/B metrics</a> for assignment, behavior, collection, and inspection.</p><h2 id="what-s-next" tabindex="-1">What&#39;s next <a class="header-anchor" href="#what-s-next" aria-label="Permalink to &quot;What&#39;s next&quot;">​</a></h2><p>Continue with these pages:</p><ul><li><a href="./ab.html">Live A/B metrics</a>: sticky variants and cumulative metrics on live sessions</li><li><a href="./hillclimbing.html">Hillclimbing</a>: the loop evals make trustworthy</li><li><a href="./building-with-agents.html">Building agents with agents</a>: have a coding agent write the first suite</li><li><a href="./guides/github.html">GitHub guide</a>: deterministic webhook fixtures with <code>github replay</code></li><li><a href="./reference/sessions.html">Sessions and streaming</a>: the events <code>t.events</code> contains</li></ul>`,85)])])}const g=e(n,[["render",l]]);export{k as __pageData,g as default};
+<span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> eval</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> status</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> evalrun_…</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --prod</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --slug</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> vulnerability-scanner</span></span></code></pre></div><h2 id="what-good-cases-assert" tabindex="-1">What good cases assert <a class="header-anchor" href="#what-good-cases-assert" aria-label="Permalink to &quot;What good cases assert&quot;">​</a></h2><p>Gate decisions and shape, not prose. Model wording varies run to run. Tool choice, tool avoidance, and output structure are the stable contract.</p><ol><li><code>t.succeeded()</code>: always, first.</li><li>The tool decision: <code>calledTool</code> for the intended path, <code>notCalledTool</code> for the likely wrong alternative. The pair is stronger than either alone.</li><li>Output shape: a regex for the contract (<code>/ready|blocked/i</code>, a JSON marker, a findings-block fence), never exact sentences.</li><li>For structured output, parse <code>t.reply</code> and check fields with <code>satisfies</code> instead of substring-matching JSON.</li></ol><p>The common failure modes: asserting exact phrasing, packing more than about five gates into one case (split it), and cases that depend on live external state that drifts (pin the input; see fixtures).</p><h2 id="pick-fixtures-by-agent-type" tabindex="-1">Pick fixtures by agent type <a class="header-anchor" href="#pick-fixtures-by-agent-type" aria-label="Permalink to &quot;Pick fixtures by agent type&quot;">​</a></h2><p>The right fixture depends on the surface under test.</p><table tabindex="0"><thead><tr><th>Agent surface</th><th>Fixture</th></tr></thead><tbody><tr><td>Chat / domain assistant</td><td>A canonical prompt string, chosen once and frozen</td></tr><tr><td>Tool-heavy</td><td>Run <code>agent-sdk call &lt;tool&gt;</code> first to pin what the tool returns, then freeze the prompt that triggers it</td></tr><tr><td>GitHub webhook</td><td><code>agent-sdk github replay &lt;pr&gt; --events &#39;*&#39; --dry-run --out fixtures/github</code> snapshots real payloads for offline replay (<a href="./guides/github.html">GitHub guide</a>)</td></tr><tr><td>PR reviewer with host preparation</td><td>Diff, metadata, and gold labels pinned to commit SHAs; keep any live PR matrix small</td></tr><tr><td>Workspace-dependent</td><td><code>workspaceFiles</code> in <code>t.send</code> options, never developer-machine paths</td></tr></tbody></table><p>Tag the fast, reliably passing core <code>smoke</code> and run <code>--tag smoke</code> in the inner loop. Leave slow or flaky-prone cases untagged for explicit runs.</p><h3 id="materialize-api-backed-fixtures" tabindex="-1">Materialize API-backed fixtures <a class="header-anchor" href="#materialize-api-backed-fixtures" aria-label="Permalink to &quot;Materialize API-backed fixtures&quot;">​</a></h3><p>An input that only points at external data, such as a pull request URL, snapshot id, or pair of commit SHAs, is not self-contained. Fetch it once and commit the rendered fixture before you expand the suite.</p><ol><li>Save the diff, metadata, and labels under <code>fixtures/</code> at pinned revisions.</li><li>Seed those files with <code>workspaceFiles</code>, or read them from the fixture directory.</li><li>Assert decisions and output shape against the saved evidence.</li><li>Keep a small <code>smoke</code> subset for any remaining live pipeline checks.</li></ol><p>Read committed fixtures with <code>@cursor/july/evals/loaders</code>: <code>loadJson</code>, <code>loadJsonl</code>, and <code>loadYaml</code> resolve relative paths against the project root the runner discovered, not the cwd the CLI was invoked from (<code>resolveFixturePath</code> and <code>evalFixtureRoot</code> expose the same resolution for other file formats).</p><p><code>maxConcurrency</code> limits parallel datapoints. It does not limit model or API fan-out inside one datapoint. Materialized fixtures prevent a large suite from exhausting provider and GitHub rate limits. The <a href="./../skills/evals/SKILL.html">evals skill</a> has the full fixture workflow.</p><h2 id="keep-improvements-with-regression-evals" tabindex="-1">Keep improvements with regression evals <a class="header-anchor" href="#keep-improvements-with-regression-evals" aria-label="Permalink to &quot;Keep improvements with regression evals&quot;">​</a></h2><p>Every <a href="./hillclimbing.html">hillclimb</a> round that keeps a change must land an eval that would have failed before the change. If you can&#39;t express the improvement as a gate (a <code>calledTool</code> shift, a bounded <code>action.result</code> count, an output-shape regex), the improvement is unverified, and it&#39;ll regress silently.</p><p>The rule cuts the other way too: never weaken an existing gate to make a round pass. That&#39;s the freeze line moving, and it turns your regression suite into a list of checks that no longer protect anything.</p><h2 id="compare-variants-on-live-traffic" tabindex="-1">Compare variants on live traffic <a class="header-anchor" href="#compare-variants-on-live-traffic" aria-label="Permalink to &quot;Compare variants on live traffic&quot;">​</a></h2><p>Use <code>defineAB</code> to compare variant metrics on live sessions. It is not a test runner and has no <code>agent-sdk ab</code> command. Keep <code>defineEval</code> as the regression ratchet. Eval sessions do not enroll or change live metrics. See <a href="./ab.html">Live A/B metrics</a> for assignment, behavior, collection, and inspection.</p><h2 id="what-s-next" tabindex="-1">What&#39;s next <a class="header-anchor" href="#what-s-next" aria-label="Permalink to &quot;What&#39;s next&quot;">​</a></h2><p>Continue with these pages:</p><ul><li><a href="./ab.html">Live A/B metrics</a>: sticky variants and cumulative metrics on live sessions</li><li><a href="./hillclimbing.html">Hillclimbing</a>: the loop evals make trustworthy</li><li><a href="./building-with-agents.html">Building agents with agents</a>: have a coding agent write the first suite</li><li><a href="./guides/github.html">GitHub guide</a>: deterministic webhook fixtures with <code>github replay</code></li><li><a href="./reference/sessions.html">Sessions and streaming</a>: the events <code>t.events</code> contains</li></ul>`,82)])])}const g=e(n,[["render",l]]);export{k as __pageData,g as default};

package/dist/docs/assets/{evals.md.a0SMN6r9.lean.js → evals.md.lfJoEVc8.lean.js}

@@ -1 +1 @@
-import{_ as e,c as a,o as i,ag as t}from"./chunks/framework.BCISBCiQ.js";const k=JSON.parse('{"title":"Evals","description":"Define repeatable checks with defineEval, run them with agent-sdk eval, and use them as regression checks.","frontmatter":{"title":"Evals","description":"Define repeatable checks with defineEval, run them with agent-sdk eval, and use them as regression checks."},"headers":[],"relativePath":"evals.md","filePath":"evals.md"}'),n={name:"evals.md"};function l(d,s,o,h,r,p){return i(),a("div",null,[...s[0]||(s[0]=[t("",85)])])}const g=e(n,[["render",l]]);export{k as __pageData,g as default};
+import{_ as e,c as i,o as a,ag as t}from"./chunks/framework.BCISBCiQ.js";const k=JSON.parse('{"title":"Evals","description":"Define repeatable checks with defineEval, run them with agent-sdk eval, and use them as regression checks.","frontmatter":{"title":"Evals","description":"Define repeatable checks with defineEval, run them with agent-sdk eval, and use them as regression checks."},"headers":[],"relativePath":"evals.md","filePath":"evals.md"}'),n={name:"evals.md"};function l(d,s,h,o,r,p){return a(),i("div",null,[...s[0]||(s[0]=[t("",82)])])}const g=e(n,[["render",l]]);export{k as __pageData,g as default};

package/dist/docs/assets/{guides_agent-to-agent.md.BCeVdJRJ.js → guides_agent-to-agent.md.BDb0t1QV.js}

@@ -1,4 +1,4 @@
-import{_ as t,c as s,o as a,ag as n}from"./chunks/framework.BCISBCiQ.js";const g=JSON.parse('{"title":"Agent-to-agent","description":"Every agent is an MCP server. Peer MCP connections let one agent delegate to another that keeps its own tools, sessions, and context.","frontmatter":{"title":"Agent-to-agent","description":"Every agent is an MCP server. Peer MCP connections let one agent delegate to another that keeps its own tools, sessions, and context."},"headers":[],"relativePath":"guides/agent-to-agent.md","filePath":"guides/agent-to-agent.md"}'),o={name:"guides/agent-to-agent.md"};function i(r,e,h,l,d,c){return a(),s("div",null,[...e[0]||(e[0]=[n(`<h1 id="agent-to-agent" tabindex="-1">Agent-to-agent <a class="header-anchor" href="#agent-to-agent" aria-label="Permalink to &quot;Agent-to-agent&quot;">​</a></h1><p>Every mounted agent is also an MCP server. So agents can delegate to each other without extra infrastructure. One agent hands a question to another, and the peer answers in its own session with its own instructions, tools, and context. Any external MCP client can do the same. A peer MCP connection makes the wiring one line.</p><p>This guide wires a <code>concierge</code> agent that delegates weather questions to a <code>weather-agent</code> peer mounted on the same host.</p><h2 id="the-mcp-endpoint" tabindex="-1">The MCP endpoint <a class="header-anchor" href="#the-mcp-endpoint" aria-label="Permalink to &quot;The MCP endpoint&quot;">​</a></h2><p>Each agent serves the Model Context Protocol over streamable HTTP at <code>/&lt;slug&gt;/v1/mcp</code> (or <code>/v1/mcp</code> in single mode). The surface is stateless (session identity travels in tool arguments) and runs the same route auth chain as the session API. It exposes three tools:</p><table tabindex="0"><thead><tr><th>Tool</th><th>Behavior</th></tr></thead><tbody><tr><td><code>ask</code></td><td>Send a message. Runs a model turn in this agent&#39;s own session and returns <code>{ status, sessionId, reply }</code>. Omit <code>sessionId</code> for a fresh session; pass it back to follow up.</td></tr><tr><td><code>check</code></td><td>Wait for or poll a running session (<code>waitSeconds: 0</code> polls without blocking).</td></tr><tr><td><code>call_tool</code></td><td>Call one of the agent&#39;s server tools directly, no model turn. Registered only when the agent has server tools.</td></tr></tbody></table><p>Waits are bounded at roughly 50 seconds, below typical MCP client request timeouts: a long turn returns <code>status: &quot;running&quot;</code> and the caller keeps waiting with <code>check</code>. Sessions created this way live on the <code>mcp</code> channel, bind to the calling principal, and show up in the playground and <code>GET /v1/sessions</code> like any other session.</p><p>Any MCP client can attach to this endpoint. It isn&#39;t only for other Agent SDK agents.</p><h2 id="wire-a-peer-mcp-connection" tabindex="-1">Wire a peer MCP connection <a class="header-anchor" href="#wire-a-peer-mcp-connection" aria-label="Permalink to &quot;Wire a peer MCP connection&quot;">​</a></h2><p>A peer MCP connection points one agent at another mounted on the same serve host. Author an MCP connection whose transport is the peer&#39;s slug:</p><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">// agents/concierge/agent/mcp-connections/weather.ts</span></span>
+import{_ as t,c as s,o as a,ag as n}from"./chunks/framework.BCISBCiQ.js";const g=JSON.parse('{"title":"Agent-to-agent","description":"Every agent is an MCP server. Peer MCP connections let one agent delegate to another that keeps its own tools, sessions, and context.","frontmatter":{"title":"Agent-to-agent","description":"Every agent is an MCP server. Peer MCP connections let one agent delegate to another that keeps its own tools, sessions, and context."},"headers":[],"relativePath":"guides/agent-to-agent.md","filePath":"guides/agent-to-agent.md"}'),o={name:"guides/agent-to-agent.md"};function i(r,e,h,l,d,c){return a(),s("div",null,[...e[0]||(e[0]=[n(`<h1 id="agent-to-agent" tabindex="-1">Agent-to-agent <a class="header-anchor" href="#agent-to-agent" aria-label="Permalink to &quot;Agent-to-agent&quot;">​</a></h1><p>Every mounted agent is also an MCP server. So agents can delegate to each other without extra infrastructure. One agent hands a question to another, and the peer answers in its own session with its own instructions, tools, and context. Any external MCP client can do the same. A peer MCP connection makes the wiring one line.</p><p>This guide wires a <code>concierge</code> agent that delegates weather questions to a <code>weather-agent</code> peer mounted on the same host.</p><h2 id="mcp-endpoint" tabindex="-1">MCP endpoint <a class="header-anchor" href="#mcp-endpoint" aria-label="Permalink to &quot;MCP endpoint&quot;">​</a></h2><p>Each agent serves the Model Context Protocol over streamable HTTP at <code>/&lt;slug&gt;/v1/mcp</code> (or <code>/v1/mcp</code> in single mode). The surface is stateless (session identity travels in tool arguments) and runs the same route auth chain as the session API. It exposes three tools:</p><table tabindex="0"><thead><tr><th>Tool</th><th>Behavior</th></tr></thead><tbody><tr><td><code>ask</code></td><td>Send a message. Runs a model turn in this agent&#39;s own session and returns <code>{ status, sessionId, reply }</code>. Omit <code>sessionId</code> for a fresh session; pass it back to follow up.</td></tr><tr><td><code>check</code></td><td>Wait for or poll a running session (<code>waitSeconds: 0</code> polls without blocking).</td></tr><tr><td><code>call_tool</code></td><td>Call one of the agent&#39;s server tools directly, no model turn. Registered only when the agent has server tools.</td></tr></tbody></table><p>Waits are bounded at roughly 50 seconds, below typical MCP client request timeouts: a long turn returns <code>status: &quot;running&quot;</code> and the caller keeps waiting with <code>check</code>. Sessions created this way live on the <code>mcp</code> channel, bind to the calling principal, and show up in the playground and <code>GET /v1/sessions</code> like any other session.</p><p>Any MCP client can attach to this endpoint. It isn&#39;t only for other Agent SDK agents.</p><h2 id="wire-a-peer-mcp-connection" tabindex="-1">Wire a peer MCP connection <a class="header-anchor" href="#wire-a-peer-mcp-connection" aria-label="Permalink to &quot;Wire a peer MCP connection&quot;">​</a></h2><p>A peer MCP connection points one agent at another mounted on the same serve host. Author an MCP connection whose transport is the peer&#39;s slug:</p><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">// agents/concierge/agent/mcp-connections/weather.ts</span></span>
 <span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { defineConnection } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;@cursor/july/connections&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
 <span class="line"></span>
 <span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">export</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> default</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> defineConnection</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">({</span></span>
@@ -7,4 +7,4 @@ import{_ as t,c as s,o as a,ag as n}from"./chunks/framework.BCISBCiQ.js";const g
 <span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">});</span></span></code></pre></div><p>The parent model now sees the peer&#39;s <code>ask</code> and <code>check</code> (and <code>call_tool</code>) tools under the <code>weather</code> server name. It reads like subagent delegation, except the peer is a full agent that keeps its own instructions, tools, MCP connections, and sessions.</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> serve</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> ./agents</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dev</span></span>
 <span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> chat</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --url</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> http://127.0.0.1:3000/concierge</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> \\</span></span>
 <span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  --message</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;What&#39;s the weather in Paris right now?&quot;</span></span>
-<span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;"># concierge → weather.ask → weather-agent&#39;s own session/tools → reply</span></span></code></pre></div><h2 id="peer-or-subagent" tabindex="-1">Peer or subagent? <a class="header-anchor" href="#peer-or-subagent" aria-label="Permalink to &quot;Peer or subagent?&quot;">​</a></h2><p>The two delegation mechanisms solve different problems.</p><table tabindex="0"><thead><tr><th></th><th>Subagent (<code>agent/subagents/&lt;id&gt;/</code>)</th><th>Peer (<code>defineConnection({ agent })</code>)</th></tr></thead><tbody><tr><td>Runs as</td><td>an SDK custom subagent inside the parent&#39;s harness</td><td>an independent agent on the same host</td></tr><tr><td>Own tools, MCP connections, sessions</td><td>no, inherits the parent&#39;s surface</td><td>yes, everything is its own</td></tr><tr><td>Visible to others</td><td>only its parent</td><td>any MCP client, other agents, its own playground</td></tr><tr><td>Reach for it when</td><td>splitting one job into specialist roles</td><td>composing independently useful agents</td></tr></tbody></table><h2 id="how-peer-urls-resolve" tabindex="-1">How peer URLs resolve <a class="header-anchor" href="#how-peer-urls-resolve" aria-label="Permalink to &quot;How peer URLs resolve&quot;">​</a></h2><p>Peer URLs resolve when the server starts, so the ephemeral ports <code>run</code> and <code>eval</code> use work too. Local-runtime turns (and host-side <code>ctx.host.mcp</code> or channel-handler calls) reach the peer over loopback, which works under the default auth. Cloud-runtime turns execute on a VM that cannot reach this host&#39;s loopback address: pass <code>--public-url https://agent-sdk.example.com</code> (or <code>serve(dir, { publicUrl })</code>) so peers resolve to a reachable URL. Without one, peers are omitted from cloud turns and the server warns at startup. With <code>--bearer-token</code>, the token is attached to peer calls automatically so they pass the target agent&#39;s auth chain.</p><h2 id="guardrails" tabindex="-1">Guardrails <a class="header-anchor" href="#guardrails" aria-label="Permalink to &quot;Guardrails&quot;">​</a></h2><p>Unknown peer slugs and self-references fail <code>serve</code> at startup, so you find out immediately rather than at delegation time. Peers require the multi-agent layout (each agent mounted under its slug, the default).</p><p>The framework does not provide cross-host loop protection. If agent A&#39;s instructions delegate to B and B&#39;s delegate back to A, they can recurse. Scope each agent&#39;s delegation instructions narrowly. The concierge delegates <em>weather questions</em> to <code>weather-agent</code>, not everything.</p><h2 id="call-a-peer-from-host-code" tabindex="-1">Call a peer from host code <a class="header-anchor" href="#call-a-peer-from-host-code" aria-label="Permalink to &quot;Call a peer from host code&quot;">​</a></h2><p>Peer MCP connections are ordinary MCP connections, so deterministic host code can use them too. A channel handler or server tool can call <code>host.mcp.callTool(&quot;weather&quot;, &quot;ask&quot;, { message: &quot;…&quot; })</code> without any model turn deciding to. See <a href="./../reference/connections.html#every-mcp-connection-is-available-in-three-places">MCP connections</a> for the three places every MCP connection is available.</p><h2 id="what-s-next" tabindex="-1">What&#39;s next <a class="header-anchor" href="#what-s-next" aria-label="Permalink to &quot;What&#39;s next&quot;">​</a></h2><p>Continue with these pages:</p><ul><li><a href="./../reference/connections.html">MCP connections</a>: all four MCP connection transports</li><li><a href="./../reference/subagents.html">Subagents</a>: the in-harness alternative</li><li><a href="./../reference/http-api.html#mcp-endpoint">HTTP API</a>: the endpoint contract</li></ul>`,26)])])}const u=t(o,[["render",i]]);export{g as __pageData,u as default};
+<span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;"># concierge → weather.ask → weather-agent&#39;s own session/tools → reply</span></span></code></pre></div><h2 id="peer-or-subagent" tabindex="-1">Peer or subagent? <a class="header-anchor" href="#peer-or-subagent" aria-label="Permalink to &quot;Peer or subagent?&quot;">​</a></h2><p>The two delegation mechanisms solve different problems.</p><table tabindex="0"><thead><tr><th></th><th>Subagent (<code>agent/subagents/&lt;id&gt;/</code>)</th><th>Peer (<code>defineConnection({ agent })</code>)</th></tr></thead><tbody><tr><td>Runs as</td><td>an SDK custom subagent inside the parent&#39;s harness</td><td>an independent agent on the same host</td></tr><tr><td>Own tools, MCP connections, sessions</td><td>no, inherits the parent&#39;s surface</td><td>yes, everything is its own</td></tr><tr><td>Visible to others</td><td>only its parent</td><td>any MCP client, other agents, its own playground</td></tr><tr><td>Reach for it when</td><td>splitting one job into specialist roles</td><td>composing independently useful agents</td></tr></tbody></table><h2 id="how-peer-urls-resolve" tabindex="-1">How peer URLs resolve <a class="header-anchor" href="#how-peer-urls-resolve" aria-label="Permalink to &quot;How peer URLs resolve&quot;">​</a></h2><p>Peer URLs resolve when the server starts, so the ephemeral ports <code>run</code> and <code>eval</code> use work too. Local-runtime turns (and host-side <code>ctx.host.mcp</code> or channel-handler calls) reach the peer over loopback, which works under the default auth. Cloud-runtime turns execute on a VM that cannot reach this host&#39;s loopback address: pass <code>--public-url https://agent-sdk.example.com</code> (or <code>serve(dir, { publicUrl })</code>) so peers resolve to a reachable URL. Without one, peers are omitted from cloud turns and the server warns at startup. With <code>--bearer-token</code>, the token is attached to peer calls automatically so they pass the target agent&#39;s auth chain.</p><h2 id="guardrails" tabindex="-1">Guardrails <a class="header-anchor" href="#guardrails" aria-label="Permalink to &quot;Guardrails&quot;">​</a></h2><p>Unknown peer slugs and self-references fail <code>serve</code> at startup, so you find out immediately rather than at delegation time. Peers require the multi-agent layout (each agent mounted under its slug, the default).</p><p>The framework does not provide cross-host loop protection. If agent A&#39;s instructions delegate to B and B&#39;s delegate back to A, they can recurse. Scope each agent&#39;s delegation instructions narrowly. The concierge delegates <em>weather questions</em> to <code>weather-agent</code>, not everything.</p><h2 id="call-a-peer-from-host-code" tabindex="-1">Call a peer from host code <a class="header-anchor" href="#call-a-peer-from-host-code" aria-label="Permalink to &quot;Call a peer from host code&quot;">​</a></h2><p>Peer MCP connections are ordinary MCP connections, so deterministic host code can use them too. A channel handler or server tool can call <code>host.mcp.callTool(&quot;weather&quot;, &quot;ask&quot;, { message: &quot;…&quot; })</code> without any model turn deciding to. See <a href="./../reference/connections.html#every-model-visible-mcp-connection-is-available-in-three-places">MCP connections</a> for the three places every model-visible MCP connection is available.</p><h2 id="what-s-next" tabindex="-1">What&#39;s next <a class="header-anchor" href="#what-s-next" aria-label="Permalink to &quot;What&#39;s next&quot;">​</a></h2><p>Continue with these pages:</p><ul><li><a href="./../reference/connections.html">MCP connections</a>: all four MCP connection transports</li><li><a href="./../reference/subagents.html">Subagents</a>: the in-harness alternative</li><li><a href="./../reference/http-api.html#mcp-endpoint">HTTP API</a>: the endpoint contract</li></ul>`,26)])])}const u=t(o,[["render",i]]);export{g as __pageData,u as default};

package/dist/docs/assets/guides_cloud-runtime.md.CkYbjnAX.js

@@ -0,0 +1,9 @@
+import{_ as t,c as s,o,ag as a}from"./chunks/framework.BCISBCiQ.js";const p=JSON.parse('{"title":"Cloud runtime","description":"Run turns on Cursor cloud agents instead of the local harness: when to switch, what changes, and the patterns that hold up.","frontmatter":{"title":"Cloud runtime","description":"Run turns on Cursor cloud agents instead of the local harness: when to switch, what changes, and the patterns that hold up."},"headers":[],"relativePath":"guides/cloud-runtime.md","filePath":"guides/cloud-runtime.md"}'),n={name:"guides/cloud-runtime.md"};function i(r,e,d,l,c,h){return o(),s("div",null,[...e[0]||(e[0]=[a(`<h1 id="cloud-runtime" tabindex="-1">Cloud runtime <a class="header-anchor" href="#cloud-runtime" aria-label="Permalink to &quot;Cloud runtime&quot;">​</a></h1><p>By default, turns execute on the Cursor SDK&#39;s local harness, on the same machine as the server. Set <code>runtime: &quot;cloud&quot;</code> and turns execute on Cursor cloud agents instead. They&#39;re ephemeral VMs that carry a repo checkout, run <code>gh</code>, <code>git</code>, and tests for real, and scale past what one host&#39;s disk and CPU can do. The serve host keeps handling routing, host preparation, sessions, and bookkeeping.</p><p>A canonical use is a PR driver whose triage runs on cloud VMs. The patterns in this guide come from running one against real PR traffic.</p><h2 id="when-to-switch" tabindex="-1">When to switch <a class="header-anchor" href="#when-to-switch" aria-label="Permalink to &quot;When to switch&quot;">​</a></h2><p>A guideline from running PR agents at scale: per-PR worktrees on the serve host don&#39;t scale to hundreds of engineers opening PRs. When the job needs a repo checkout at scale, use cloud. The signals:</p><ul><li>The agent must run repo commands (tests, builds, <code>git</code>) against many different refs concurrently.</li><li>Turns are long and heavy, and you don&#39;t want them competing with the server for resources.</li><li>The work product is a PR or branch the VM can push, not a local file.</li></ul><p>Stay local when the agent is conversational, tool-driven against APIs, or works over host-prepared evidence. Local turns are cheaper, start faster, and support the full authored surface.</p><h2 id="configure-it" tabindex="-1">Configure it <a class="header-anchor" href="#configure-it" aria-label="Permalink to &quot;Configure it&quot;">​</a></h2><p>Cloud runtime is two fields on the agent config.</p><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { defineAgent } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;@cursor/july&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
+<span class="line"></span>
+<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">export</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> default</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> defineAgent</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">({</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  runtime: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;cloud&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  cloud: {</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">    repos: [{ url: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;https://github.com/org/repo&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">, startingRef: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;main&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> }],</span></span>
+<span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">    // env / envVars / … forwarded to the Cursor SDK</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  },</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">});</span></span></code></pre></div><p>The host must be signed in (<code>agent-sdk login</code> or <code>CURSOR_API_KEY</code>).</p><div class="important custom-block github-alert"><p class="custom-block-title">IMPORTANT</p><p>Cloud agents run against the Cursor backend under the signed-in account, and every wake spends real cloud-agent budget. Decide explicitly what may trigger one.</p></div><h2 id="what-changes-on-cloud" tabindex="-1">What changes on cloud <a class="header-anchor" href="#what-changes-on-cloud" aria-label="Permalink to &quot;What changes on cloud&quot;">​</a></h2><p>Cloud turns run on a VM without your authored files, so the runtime mapping shifts:</p><table tabindex="0"><thead><tr><th>Folder or file</th><th>Local runtime</th><th>Cloud runtime</th></tr></thead><tbody><tr><td><code>instructions.*</code></td><td><code>AGENTS.md</code> in the session workspace</td><td>prepended to the first prompt</td></tr><tr><td>Server tools (<code>execution: &quot;server&quot;</code>)</td><td>in-process SDK custom tools</td><td>authenticated HTTP MCP back to the AgentSDK host, when <code>--public-url</code> or <code>--cloud-tools-url</code> is set</td></tr><tr><td>Agent tools (<code>execution: &quot;agent&quot;</code>)</td><td>scripts in the session workspace</td><td>catalog + script bodies on the first prompt</td></tr><tr><td><code>skills/*</code></td><td><code>.cursor/skills/</code> in the workspace</td><td>native discovery after the first turn, from the hosted store or the signed-in account</td></tr><tr><td><code>mcp-connections/*.ts</code></td><td>SDK <code>mcpServers</code></td><td>SDK <code>mcpServers</code> (peers need <code>--public-url</code>)</td></tr><tr><td><code>host-connections/*.ts</code></td><td><code>ctx.host.mcp</code> only</td><td><code>ctx.host.mcp</code> only</td></tr><tr><td><code>sandbox/workspace/**</code></td><td>seeded into the session workspace</td><td>ignored</td></tr><tr><td>Tool approvals (<code>needsApproval</code>)</td><td>supported</td><td>not supported; keep approval-gated tools on local turns</td></tr></tbody></table><p>Authored skills are discovered natively after the first cloud turn, using the hosted store or the signed-in account.</p><p>Approvals are a local-runtime contract. On cloud, a <code>needsApproval</code> tool call rides one HTTP MCP request from the VM, and a parked call would hold that request open until it times out; there is no durable approval flow for cloud turns.</p><p>Peer MCP connections need <code>--public-url</code> for cloud turns. Without one, peers are omitted and the server warns at startup.</p><h2 id="hybrid-local-agent-cloud-sessions" tabindex="-1">Hybrid: local agent, cloud sessions <a class="header-anchor" href="#hybrid-local-agent-cloud-sessions" aria-label="Permalink to &quot;Hybrid: local agent, cloud sessions&quot;">​</a></h2><p>A local-runtime agent can still open cloud-attached sessions per send. Channel handlers may pass a <code>cloud</code> block (repos pinned to a PR ref, say) in <code>send</code> options, and Slack handlers may return <code>cloud</code> from a mention hook. A PR driver works this way: chat stays local, and the <code>drive</code> flow attaches the PR to a cloud VM. The agent-level <code>cloud</code> config is the base that per-session options merge over.</p><h2 id="patterns-that-hold-up" tabindex="-1">Patterns that hold up <a class="header-anchor" href="#patterns-that-hold-up" aria-label="Permalink to &quot;Patterns that hold up&quot;">​</a></h2><p>These come from running a PR driver against real PR traffic:</p><ul><li>One cloud session per unit of work, keyed with a stable continuation token (<code>pr:owner/repo#N</code>) so every wake lands on the same conversation.</li><li>Keep the host deterministic: fetch briefs and metadata on the host, send the VM a compact prompt, and let the VM re-read source of truth with its own <code>gh</code> and <code>git</code> instead of trusting payload snapshots.</li><li>Limit exposure: add repository allowlists on webhook channels, because every wake spends the account&#39;s budget.</li></ul><h2 id="verify-cloud-agents" tabindex="-1">Verify cloud agents <a class="header-anchor" href="#verify-cloud-agents" aria-label="Permalink to &quot;Verify cloud agents&quot;">​</a></h2><p><code>agent-sdk run</code> and <code>eval</code> work unchanged. The trajectory records the same event vocabulary plus <code>agent.bound</code> with the cloud URL, so you can open the cloud conversation for any session. Cloud turns take minutes. Pass generous <code>--timeout-ms</code> values, and keep curl timeouts long when driving channels directly.</p><h2 id="what-s-next" tabindex="-1">What&#39;s next <a class="header-anchor" href="#what-s-next" aria-label="Permalink to &quot;What&#39;s next&quot;">​</a></h2><p>Continue with these pages:</p><ul><li><a href="./../reference/agent-config.html">Agent config</a>: the <code>runtime</code> and <code>cloud</code> fields</li><li><a href="./github.html">GitHub guide</a>: the webhook patterns that pair with cloud triage</li></ul>`,28)])])}const g=t(n,[["render",i]]);export{p as __pageData,g as default};

package/dist/docs/assets/guides_cloud-runtime.md.CkYbjnAX.lean.js

@@ -0,0 +1 @@
+import{_ as t,c as s,o,ag as a}from"./chunks/framework.BCISBCiQ.js";const p=JSON.parse('{"title":"Cloud runtime","description":"Run turns on Cursor cloud agents instead of the local harness: when to switch, what changes, and the patterns that hold up.","frontmatter":{"title":"Cloud runtime","description":"Run turns on Cursor cloud agents instead of the local harness: when to switch, what changes, and the patterns that hold up."},"headers":[],"relativePath":"guides/cloud-runtime.md","filePath":"guides/cloud-runtime.md"}'),n={name:"guides/cloud-runtime.md"};function i(r,e,d,l,c,h){return o(),s("div",null,[...e[0]||(e[0]=[a("",28)])])}const g=t(n,[["render",i]]);export{p as __pageData,g as default};