@cursor/july 0.1.107 → 0.1.109

package/dist/docs/assets/concepts.md.DgEcZOfT.js

@@ -1 +0,0 @@
-import{_ as t,c as a,o,ag as s}from"./chunks/framework.BCISBCiQ.js";const p=JSON.parse('{"title":"How the Agent SDK works","description":"Understand projects, sessions, event streams, channels, and runtimes in plain language.","frontmatter":{"title":"How the Agent SDK works","description":"Understand projects, sessions, event streams, channels, and runtimes in plain language."},"headers":[],"relativePath":"concepts.md","filePath":"concepts.md"}'),n={name:"concepts.md"};function r(d,e,l,i,c,h){return o(),a("div",null,[...e[0]||(e[0]=[s('<h1 id="how-the-agent-sdk-works" tabindex="-1">How the Agent SDK works <a class="header-anchor" href="#how-the-agent-sdk-works" aria-label="Permalink to &quot;How the Agent SDK works&quot;">​</a></h1><p>An agent is a folder of instructions and capabilities. The Agent SDK discovers those files, runs conversations, and records what happened.</p><h2 id="what-happens-when-someone-sends-a-message" tabindex="-1">What happens when someone sends a message? <a class="header-anchor" href="#what-happens-when-someone-sends-a-message" aria-label="Permalink to &quot;What happens when someone sends a message?&quot;">​</a></h2><p>Follow one message through the system:</p><ol><li>A channel receives the message from HTTP, Slack, GitHub, or another webhook.</li><li>The channel starts a session or continues an existing one.</li><li>The runtime gives the model its instructions, tools, and workspace.</li><li>The model replies and can call tools along the way.</li><li>The Agent SDK appends every message and tool call to the session&#39;s event stream.</li></ol><p>The channel is the front door. The runtime does the work. The event stream is the record you inspect later.</p><h2 id="how-do-files-become-an-agent" tabindex="-1">How do files become an agent? <a class="header-anchor" href="#how-do-files-become-an-agent" aria-label="Permalink to &quot;How do files become an agent?&quot;">​</a></h2><p>Each capability has a home in the project. The path tells the Agent SDK what to load. The filename becomes the capability&#39;s name. For example, <code>agent/tools/get_weather.ts</code> creates a tool named <code>get_weather</code>.</p><table tabindex="0"><thead><tr><th>Path</th><th>What it is</th></tr></thead><tbody><tr><td><code>agent/agent.ts</code></td><td>Model and runtime settings</td></tr><tr><td><code>agent/instructions.md</code></td><td>The always-on system prompt</td></tr><tr><td><code>agent/tools/&lt;name&gt;.ts</code></td><td>Typed actions the model can call</td></tr><tr><td><code>agent/skills/*</code></td><td>Procedures loaded when needed</td></tr><tr><td><code>agent/mcp-connections/&lt;name&gt;.ts</code></td><td>Tools from external MCP servers</td></tr><tr><td><code>agent/host-connections/&lt;name&gt;.ts</code></td><td>Privileged MCP servers for host tools only</td></tr><tr><td><code>agent/channels/*.ts</code></td><td>HTTP, Slack, and GitHub entry points</td></tr><tr><td><code>agent/ab.ts</code> or <code>agent/ab/*.ts</code></td><td>Sticky variants and live performance metrics</td></tr><tr><td><code>agent/result.ts</code></td><td>Optional host <code>commit</code> on the final assistant text</td></tr><tr><td><code>evals/**/*.eval.ts</code></td><td>Repeatable checks at the project root</td></tr></tbody></table><p>Other folders add subagents, extensions, hooks, schedules, and workspace files. You don&#39;t register them elsewhere. Run <code>agent-sdk validate</code> to catch invalid files before serving the project.</p><p>See <a href="./reference/project-layout.html">Project layout</a> for the folder structure.</p><h2 id="how-does-the-agent-sdk-identify-a-conversation" tabindex="-1">How does the Agent SDK identify a conversation? <a class="header-anchor" href="#how-does-the-agent-sdk-identify-a-conversation" aria-label="Permalink to &quot;How does the Agent SDK identify a conversation?&quot;">​</a></h2><p>A session is one durable conversation. It has two identifiers:</p><ul><li><strong><code>continuationToken</code></strong> tells a channel which conversation to resume. A Slack channel can use its thread ID. A GitHub channel can use the pull request. The built-in HTTP API returns an opaque token and rotates it after each accepted follow-up.</li><li><strong><code>sessionId</code></strong> identifies the stored session. Use it to stream events, inspect the session, resolve approvals, or bind a tool call to the session.</li></ul><p>Use the continuation token to keep talking. Use the session ID to observe or manage the conversation.</p><h2 id="how-do-i-see-what-an-agent-did" tabindex="-1">How do I see what an agent did? <a class="header-anchor" href="#how-do-i-see-what-an-agent-did" aria-label="Permalink to &quot;How do I see what an agent did?&quot;">​</a></h2><p>Each session records an append-only event stream. It includes:</p><ul><li>Messages and streamed text</li><li>Requested tool calls and their results</li><li>Approval requests and decisions</li><li>Turn completion and token usage</li></ul><p>Sessions and their event streams survive server restarts. The playground renders the stream. Evals assert against it. The <code>agent-sdk trajectory</code> command turns a saved stream into a short summary.</p><p>When a run surprises you, inspect its event stream first. See <a href="./reference/sessions.html">Sessions and streaming</a> for the event vocabulary.</p><h2 id="what-does-a-channel-control" tabindex="-1">What does a channel control? <a class="header-anchor" href="#what-does-a-channel-control" aria-label="Permalink to &quot;What does a channel control?&quot;">​</a></h2><p>A channel connects the agent to a surface such as HTTP, Slack, GitHub, or a custom webhook. It controls:</p><ul><li>Routes and input schemas</li><li>Authentication</li><li>Conversation identity</li><li>How replies return to the user</li></ul><p>The built-in HTTP session API is always available. Custom routes accept loopback callers by default. Add an auth policy before sharing them over a network.</p><p>Channels should also prepare deterministic input for the model. For example, a GitHub channel can fetch the pull request, collect the diff, and seed the workspace before the turn starts. The model can then focus on the review instead of gathering files.</p><p>See <a href="./reference/channels.html">Channels</a> for route and authentication details.</p><h2 id="where-does-a-turn-run" tabindex="-1">Where does a turn run? <a class="header-anchor" href="#where-does-a-turn-run" aria-label="Permalink to &quot;Where does a turn run?&quot;">​</a></h2><p>Choose a runtime in <code>agent/agent.ts</code>:</p><table tabindex="0"><thead><tr><th></th><th>Local (default)</th><th>Cloud</th></tr></thead><tbody><tr><td>Turn runs on</td><td>The server host</td><td>A Cursor cloud agent</td></tr><tr><td>Server tools</td><td>Supported</td><td>Supported when the server has <code>--public-url</code> or <code>--cloud-tools-url</code>; the cloud turn reaches them over authenticated HTTP MCP. Without one of those flags, the server warns and cloud turns omit them.</td></tr><tr><td>Approvals (<code>needsApproval</code>)</td><td>Supported</td><td>Not supported (local runtime only)</td></tr><tr><td>Agent tool scripts</td><td>Supported</td><td>Supported</td></tr><tr><td>Skills</td><td>Added to the session workspace</td><td>Available automatically</td></tr><tr><td>Seeded files</td><td>Added to the session workspace</td><td>Available in a shared directory the first prompt names</td></tr><tr><td>Repository</td><td>You provide it</td><td>The cloud agent checks it out</td></tr></tbody></table><p>Use the local runtime when the host has the tools and files the agent needs. Use the cloud runtime when each turn needs an isolated repository checkout. <code>agent-sdk validate</code> warns when a cloud agent uses a local-only capability.</p><h2 id="what-files-can-a-local-session-access" tabindex="-1">What files can a local session access? <a class="header-anchor" href="#what-files-can-a-local-session-access" aria-label="Permalink to &quot;What files can a local session access?&quot;">​</a></h2><p>Each local session gets its own workspace. The Agent SDK writes the instructions as <code>AGENTS.md</code>, installs authored skills, copies sandbox files, and adds agent tool scripts.</p><p>The workspace is a real Cursor project. It can inherit <code>AGENTS.md</code> and <code>.cursor</code> settings from parent directories. Nested git checkouts default <code>local.cwd</code> to a per-project cache directory under <code>~/.cache</code>. Point <code>cwd</code> at a checkout only when the agent should inherit that tree. <code>run</code> and <code>eval</code> already use a temporary state root.</p><p>Session files live under <code>--state-root</code>. See <a href="./reference/sessions.html#where-does-the-agent-sdk-store-session-data">Sessions</a> for the layout.</p><h2 id="how-can-one-agent-call-another" tabindex="-1">How can one agent call another? <a class="header-anchor" href="#how-can-one-agent-call-another" aria-label="Permalink to &quot;How can one agent call another?&quot;">​</a></h2><p>Every mounted agent also serves MCP at <code>/&lt;slug&gt;/v1/mcp</code>. Another agent or MCP client can use <code>ask</code>, <code>check</code>, and <code>call_tool</code> to delegate work. A peer MCP connection such as <code>defineConnection({ agent: &quot;weather-agent&quot; })</code> adds those tools to the calling agent.</p><p>See <a href="./guides/agent-to-agent.html">Agent-to-agent</a> for a complete example.</p><h2 id="which-rules-prevent-common-setup-problems" tabindex="-1">Which rules prevent common setup problems? <a class="header-anchor" href="#which-rules-prevent-common-setup-problems" aria-label="Permalink to &quot;Which rules prevent common setup problems?&quot;">​</a></h2><ul><li>Use Node 22.13 or newer. Bun isn&#39;t supported.</li><li>Put evals under the project-root <code>evals/</code> directory, not <code>agent/evals/</code>.</li><li>Run a TypeScript check before shipping. <code>validate</code> and <code>run</code> execute TypeScript but don&#39;t type-check it.</li><li>Return JSON-shaped values from tool <code>execute</code> functions.</li><li>Keep local session workspaces away from parent rules you don&#39;t want the agent to inherit.</li><li>Sign in or set <code>CURSOR_API_KEY</code> before starting a model turn. Discovery, validation, direct tool calls, and server startup work without a credential.</li></ul><h2 id="related" tabindex="-1">Related <a class="header-anchor" href="#related" aria-label="Permalink to &quot;Related&quot;">​</a></h2><ul><li><a href="./quickstart.html">Quickstart</a></li><li><a href="./reference/project-layout.html">Project layout</a></li><li><a href="./reference/sessions.html">Sessions and streaming</a></li><li><a href="./reference/channels.html">Channels</a></li><li><a href="./ab.html">Live A/B metrics</a></li></ul>',41)])])}const m=t(n,[["render",r]]);export{p as __pageData,m as default};

package/dist/docs/assets/evals.md.CbMoebP1.js

@@ -1,85 +0,0 @@
-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>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  description: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;Inspects a PR without approving it.&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  tags: [</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;smoke&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">],</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  timeoutMs: </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">120_000</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
-<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">  async</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> test</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">t</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">) {</span></span>
-<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">    await</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> t.</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">send</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span></span>
-<span class="line"><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">      &quot;Is https://github.com/acme/checkout/pull/42 ready to approve?&quot;</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;">    t.</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">succeeded</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;">calledTool</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;inspect_pr&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;">notCalledTool</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;approve_pr&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;">(t.reply, </span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">includes</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">/</span><span style="--shiki-light:#032F62;--shiki-dark:#DBEDFF;">ready</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">|</span><span style="--shiki-light:#032F62;--shiki-dark:#DBEDFF;">approve</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">/</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">i</span><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>One file can also hold several datapoints through <code>cases</code> (provide either <code>test</code> or <code>cases</code>, not both). Each case id becomes <code>&lt;fileId&gt;/&lt;case.id&gt;</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/prs.eval.ts → prs/checkout, prs/search</span></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>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  tags: [</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;smoke&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">, </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;prs&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">],</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  cases: [</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;">      id: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;checkout&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">      description: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;Checkout PR readiness.&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
-<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">      async</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> test</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">t</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">) {</span></span>
-<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">        await</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> t.</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">send</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span></span>
-<span class="line"><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">          &quot;Is https://github.com/acme/checkout/pull/42 ready to approve?&quot;</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;">        t.</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">succeeded</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;">calledTool</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;inspect_pr&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>
-<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;">      id: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;search&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
-<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">      async</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> test</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">t</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">) {</span></span>
-<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">        await</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> t.</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">send</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span></span>
-<span class="line"><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">          &quot;Check https://github.com/acme/search/pull/7 before approval.&quot;</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;">        t.</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">succeeded</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;">calledTool</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;inspect_pr&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>
-<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>Case ids must be single path segments, unique within the file. Each case can set its own <code>description</code>, <code>tags</code>, <code>timeoutMs</code>, and <code>iterations</code>. A case-level value replaces the file-level value for that datapoint.</p><h3 id="iterations" tabindex="-1">Iterations <a class="header-anchor" href="#iterations" aria-label="Permalink to &quot;Iterations&quot;">​</a></h3><p><code>iterations</code> (file or case, default <code>1</code>) runs a datapoint repeatedly. Discovery expands <code>iterations: 3</code> on case <code>nyc</code> to runnable ids <code>weather/nyc/1</code>, <code>weather/nyc/2</code>, <code>weather/nyc/3</code> (filter prefix <code>weather/nyc</code> still selects all three). Each expanded case exposes <code>t.iteration</code> / <code>t.iterations</code> on the test context. Cap is 100.</p><p><code>maxConcurrency</code> counts <strong>authored datapoints</strong>, not expanded iterations: siblings <code>…/1</code>…<code>…/n</code> share one concurrency slot and run sequentially. A suite with 11 cases × 3 iterations and <code>maxConcurrency: 20</code> therefore has at most 11 cases in flight, not 33.</p><h2 id="configure-eval-runs" tabindex="-1">Configure eval runs <a class="header-anchor" href="#configure-eval-runs" aria-label="Permalink to &quot;Configure eval runs&quot;">​</a></h2><p>Each project with evals needs <code>evals/evals.config.ts</code> or <code>evals/evals.config.js</code>, and it must set <code>maxConcurrency</code>. Each case issues real model-provider requests, so concurrency is capped hard at 200. Existing projects use 20. Discovery with <code>eval --list</code> works without this file, but running a case does not.</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;"> { defineEvalConfig } </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;"> defineEvalConfig</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">({</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  maxConcurrency: </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">20</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">, </span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">// required</span></span>
-<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 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. A thrown <a href="./reference/result.html">turn result</a> <code>commit</code> fails the turn, so <code>t.succeeded()</code> fails too.</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>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">);</span></span></code></pre></div><p>A case with no explicit gates falls back to whether at least one turn completed successfully. Add <code>t.succeeded()</code> and behavior-specific gates anyway. They make the contract visible during review.</p><h3 id="judge-free-form-output" tabindex="-1">Judge free-form output <a class="header-anchor" href="#judge-free-form-output" aria-label="Permalink to &quot;Judge free-form output&quot;">​</a></h3><p>When wording matters and no regex captures it, <code>t.judge</code> grades the reply with an LLM. The built-in graders are <code>factuality(expected)</code>, <code>summarizes(expected)</code>, <code>closedQA(criteria)</code>, and <code>sql(expected)</code>. Each scores <code>t.reply</code> by default; pass <code>{ on }</code> to grade another value.</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:#24292E;--shiki-dark:#E1E4E8;">t.judge.</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">factuality</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;It is 54°F in NYC right now.&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">).</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">atLeast</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">0.7</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">);</span></span></code></pre></div><p>Judge assertions are soft by default, so a judge never fails a build until you give it a bar with <code>.atLeast(0.7)</code> or promote it with <code>.gate(0.8)</code>. The judge model comes from <code>defineEvalConfig({ judge })</code>, <code>defineEval({ judge })</code>, a case-level <code>judge</code>, or a per-call <code>{ model }</code> override; the nearest one wins. For a domain-specific judge whose verdict is not a single score, <code>t.judge.model(prompt)</code> sends a raw prompt to the same model and returns the reply. You then record the parsed result with <code>t.score</code> or <code>t.check</code>.</p><h2 id="run-evals-from-the-cli" tabindex="-1">Run evals from the CLI <a class="header-anchor" href="#run-evals-from-the-cli" aria-label="Permalink to &quot;Run evals from the CLI&quot;">​</a></h2><p>The <code>eval</code> command discovers, filters, and runs cases.</p><p>Run the CLI under Node 22.13 or newer. Do not use Bun. Its HTTP/2 client breaks tool-result streams and causes eval turns to fail.</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;"> --list</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">                         # discover only</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:#6A737D;--shiki-dark:#6A737D;">                                # run all</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:#032F62;--shiki-dark:#9ECBFF;"> builds/checkout</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">                # one datapoint</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:#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 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>
-<span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  &quot;passed&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">1</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
-<span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  &quot;failed&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">0</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
-<span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  &quot;results&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>
-<span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">      &quot;id&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;readiness&quot;</span><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>
-<span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">      &quot;assertions&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: [{ </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">&quot;name&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;succeeded&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">, </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">&quot;passed&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>
-<span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">      &quot;sessionId&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;ses_123&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
-<span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">      &quot;inputs&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: [</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;Is checkout pull request 42 ready to approve?&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">],</span></span>
-<span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">      &quot;toolCalls&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: [{ </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">&quot;toolName&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;inspect_pr&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">, </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">&quot;isError&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">false</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> }],</span></span>
-<span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">      &quot;logs&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: [],</span></span>
-<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, 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><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.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.CbMoebP1.lean.js

@@ -1 +0,0 @@
-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_cloud-agents.md.DS8RIjwx.js

@@ -1,9 +0,0 @@
-import{_ as t,c as a,o as s,ag as n}from"./chunks/framework.BCISBCiQ.js";const p=JSON.parse('{"title":"Cursor cloud agents","description":"Let an agent-sdk agent hand coding work to Cursor cloud agents and steer them: mount the extension, pick the repositories it may touch, and decide what needs approval.","frontmatter":{"title":"Cursor cloud agents","description":"Let an agent-sdk agent hand coding work to Cursor cloud agents and steer them: mount the extension, pick the repositories it may touch, and decide what needs approval."},"headers":[],"relativePath":"guides/cloud-agents.md","filePath":"guides/cloud-agents.md"}'),o={name:"guides/cloud-agents.md"};function i(r,e,h,l,d,u){return s(),a("div",null,[...e[0]||(e[0]=[n(`<h1 id="cursor-cloud-agents" tabindex="-1">Cursor cloud agents <a class="header-anchor" href="#cursor-cloud-agents" aria-label="Permalink to &quot;Cursor cloud agents&quot;">​</a></h1><p>The <code>cursor-cloud-agents</code> extension lets an agent delegate coding work to Cursor cloud agents and stay in the conversation while they do it. The cloud agents work on real repository checkouts and push branches and pull requests; your agent hands off the task, reports back, and answers questions in the meantime. A Slack triage agent turns a bug report into a PR, a planner fans a refactor out across repositories, a reviewer asks a cloud agent to address its own comments.</p><p>This is different from <a href="./../reference/agent-config.html#choose-a-runtime"><code>runtime: &quot;cloud&quot;</code></a>. The runtime moves <em>this</em> agent&#39;s turns onto a cloud VM. The extension keeps the agent where it is and lets it start <em>other</em> cloud agents.</p><h2 id="mount-it" tabindex="-1">Mount it <a class="header-anchor" href="#mount-it" aria-label="Permalink to &quot;Mount it&quot;">​</a></h2><p>The extension ships with <code>@cursor/july</code>. Mount it with one file.</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;">// agent/extensions/cloud.ts</span></span>
-<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> cursorCloudAgents </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;@cursor/july/extensions/cursor-cloud-agents&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;"> cursorCloudAgents</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/acme/api&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:#24292E;--shiki-dark:#E1E4E8;">    autoCreatePR: </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;">  },</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">});</span></span></code></pre></div><p>Launches run under the host&#39;s signed-in account (<code>agent-sdk login</code> or <code>CURSOR_API_KEY</code>) and spend its cloud-agent budget. Name the repositories the agent may target, and keep approval on when you can&#39;t.</p><p>Mount it on an agent that runs on the local runtime; that is where launches park for approval. On the cloud runtime approvals don&#39;t resolve.</p><h2 id="what-the-agent-can-do" tabindex="-1">What the agent can do <a class="header-anchor" href="#what-the-agent-can-do" aria-label="Permalink to &quot;What the agent can do&quot;">​</a></h2><p>With the extension mounted, the agent can hand off coding work and stay in the conversation while it happens:</p><ul><li><strong>Hand off a task.</strong> Launch a cloud agent on one of your repositories, or on a pull request, with a self-contained brief. It gets back an agent id and a URL to share.</li><li><strong>Check on it.</strong> See whether the agent is still running, what it produced, and which branches or pull requests it pushed.</li><li><strong>Read what was done.</strong> Pull the agent&#39;s full conversation into the workspace, quote its final report, or trace the exact edits before summarizing for a person.</li><li><strong>Steer or stop.</strong> Send a follow-up to an agent it launched; it queues behind the current run, or interrupts it on request. Cancel a run that went sideways.</li><li><strong>Keep track.</strong> List the agents launched in this conversation, or everything it has launched before.</li></ul><p>A skill teaches the model when to delegate, how to write a brief a cloud agent can act on alone, and how to follow up. Everything the extension launches stays bound to whoever launched it: an agent started by another user, or outside the extension, is out of reach. That separation is only as fine as your host&#39;s auth; a shared bearer token or <code>--allow-anonymous</code> makes every caller one principal, so use <a href="./../reference/channels.html#auth-policies">authored auth</a> when callers need their own.</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><table tabindex="0"><thead><tr><th>Option</th><th>What it controls</th></tr></thead><tbody><tr><td><code>cloud</code></td><td>Launch defaults, in the same shape as the agent&#39;s own <a href="./../reference/agent-config.html#cloud-options">cloud options</a>.</td></tr><tr><td><code>cloud.repos</code></td><td>The repositories the agent may hand work to, with a default branch or pull request for each. Without it, launches run with no repository and the model can&#39;t name one, so secrets in <code>cloud.envVars</code> never reach a repository you didn&#39;t list.</td></tr><tr><td><code>model</code></td><td>The model launched agents run. Omit it to use the account default.</td></tr><tr><td><code>needsApproval</code></td><td>Park launches and follow-ups for a <a href="./human-in-the-loop.html">human decision</a>. Defaults to <code>true</code> when no <code>cloud.repos</code> are configured, <code>false</code> otherwise.</td></tr><tr><td><code>apiKey</code></td><td>A Cursor API key to launch with instead of the host&#39;s own.</td></tr></tbody></table><h2 id="adjust-it" tabindex="-1">Adjust it <a class="header-anchor" href="#adjust-it" aria-label="Permalink to &quot;Adjust it&quot;">​</a></h2><p>To take a capability away, such as sending follow-ups, mount as a directory and disable that tool; see <a href="./../reference/extensions.html#adjust-a-mounted-extension">overlays</a>.</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/extensions.html">Extensions</a>: how mounts, namespaces, and overrides work</li><li><a href="./human-in-the-loop.html">Human-in-the-loop approvals</a>: where parked launches get approved</li></ul>`,19)])])}const g=t(o,[["render",i]]);export{p as __pageData,g as default};

package/dist/docs/assets/index.md.BW_6tOgR.js

@@ -1,5 +0,0 @@
-import{_ as t,c as a,o as s,ag as r}from"./chunks/framework.BCISBCiQ.js";const g=JSON.parse('{"title":"Agent SDK documentation","description":"Build and run Cursor agents with tools, approvals, channels, and evals.","frontmatter":{"title":"Agent SDK documentation","description":"Build and run Cursor agents with tools, approvals, channels, and evals."},"headers":[],"relativePath":"index.md","filePath":"README.md"}'),i={name:"index.md"};function n(o,e,l,h,d,c){return s(),a("div",null,[...e[0]||(e[0]=[r(`<h1 id="agent-sdk-documentation" tabindex="-1">Agent SDK documentation <a class="header-anchor" href="#agent-sdk-documentation" aria-label="Permalink to &quot;Agent SDK documentation&quot;">​</a></h1><p>Use the Agent SDK to define Cursor agents in TypeScript and Markdown. See <a href="./reference/project-layout.html">Project layout</a> for the directory structure.</p><p>Use Node 22.13 or newer. Bun isn&#39;t supported.</p><p>Create a project:</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;">npx</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> @cursor/july</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> init</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> ./my-agent</span></span>
-<span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">cd</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> my-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;"> dev</span></span></code></pre></div><p>If <code>agent-sdk</code> isn&#39;t on <code>PATH</code>, use <code>npx @cursor/july &lt;command&gt;</code>.</p><p>Open the docs locally:</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;">npx</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> @cursor/july</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> docs</span></span></code></pre></div><h2 id="where-to-start" tabindex="-1">Where to start <a class="header-anchor" href="#where-to-start" aria-label="Permalink to &quot;Where to start&quot;">​</a></h2><table tabindex="0"><thead><tr><th>You are...</th><th>Start with</th></tr></thead><tbody><tr><td>New to the Agent SDK</td><td><a href="./quickstart.html">Quickstart</a> (PR reviewer), then <a href="./concepts.html">Concepts</a></td></tr><tr><td>Building a new agent with Cursor</td><td><a href="./scaffolding-agents.html">Scaffold an agent with Cursor</a></td></tr><tr><td>Turning a Cursor Automation into a project</td><td><a href="./guides/convert-automation.html">Convert a Cursor Automation</a></td></tr><tr><td>Wiring an agent to Slack</td><td><a href="./guides/slack.html">Slack guide</a></td></tr><tr><td>Starting from a packaged template</td><td><a href="./templates/demo.html">Demo</a>, <a href="./templates/grokbot-agents.html">Grok Bot agents</a>, <a href="./templates/code-wiki.html">Code wiki</a>, <a href="./templates/agents-md.html">Living AGENTS.md</a>, <a href="./templates/security-reviewer.html">Security reviewer</a>, <a href="./templates/security-help.html">Security help</a>, <a href="./templates/triage.html">Triage</a>, or <a href="./templates/agentic-owners.html">Agentic Owners</a></td></tr><tr><td>Wiring an agent to GitHub webhooks</td><td><a href="./guides/github.html">GitHub guide</a></td></tr><tr><td>Driving PRs from a cloud VM</td><td><a href="./templates/pr-autofixer.html">PR autofixer template</a></td></tr><tr><td>Handing coding work to Cursor cloud agents</td><td><a href="./guides/cloud-agents.html">Cursor cloud agents</a></td></tr><tr><td>Talking to your Grok Bot agents</td><td><a href="./guides/grokbot-agents.html">Cursor Grok Bot agents</a></td></tr><tr><td>Driving an agent from Linear (or another tracker)</td><td><a href="./guides/webhooks.html#example-linear-as-the-control-plane">Webhooks guide: Linear example</a></td></tr><tr><td>Making an existing agent measurably better</td><td><a href="./evals.html">Evals</a>, then <a href="./hillclimbing.html">Hillclimbing</a></td></tr><tr><td>Comparing variants on live traffic</td><td><a href="./ab.html">Live A/B metrics</a></td></tr><tr><td>Deploying with Cursor or on your own infrastructure</td><td><a href="./deployment.html">Deployment</a></td></tr><tr><td>Debugging something that misbehaves</td><td><a href="./troubleshooting.html">Fix common agent problems</a></td></tr></tbody></table><h2 id="documentation" tabindex="-1">Documentation <a class="header-anchor" href="#documentation" aria-label="Permalink to &quot;Documentation&quot;">​</a></h2><p><strong>Core</strong></p><ul><li><a href="./quickstart.html">Quickstart</a>: build a PR reviewer that classifies changes by complexity and handles GitHub webhook events.</li><li><a href="./scaffolding-agents.html">Scaffold an agent with Cursor</a>: use the bundled skill for a guided build.</li><li><a href="./guides/convert-automation.html">Convert a Cursor Automation</a>: export a dashboard Automation into an Agent SDK project.</li><li><a href="./concepts.html">Concepts</a>: agent discovery, sessions, channels, runtimes, and observability.</li></ul><p><strong>Templates</strong></p><ul><li><a href="./templates/demo.html">Record a walkthrough from a collected PR</a>: host collects the PR, the model records, then comments.</li><li><a href="./templates/grokbot-agents.html">Talk to your Grok Bot agents</a>: name the bots, then consult them from this agent&#39;s own conversation.</li><li><a href="./templates/code-wiki.html">Keep a code wiki current after merge</a>: update in-repo <code>wiki/</code> pages when a pull request merges.</li><li><a href="./templates/agents-md.html">Keep AGENTS.md current from last week&#39;s work</a>: pull rules out of merged PRs and Slack, then open a follow-up PR.</li><li><a href="./templates/security-reviewer.html">Security reviewer</a>: review pull requests for exploitable bugs and post one comment.</li><li><a href="./templates/security-help.html">Answer security questions in Slack</a>: answer from your skills, or take a first look at a pasted PR.</li><li><a href="./templates/triage.html">Triage Linear or Jira issues in place</a>: classify existing tickets and comment on them.</li><li><a href="./templates/agentic-owners.html">Review pull requests with owners policies</a>: request owners and approve changes allowed by repository policy.</li><li><a href="./templates/pr-autofixer.html">Fix pull requests on a Cursor cloud VM</a></li></ul><p><strong>Self-improving Agents</strong></p><ul><li><a href="./building-with-agents.html">Building agents with agents</a>: use a coding agent to scaffold, run, and iterate on your agent.</li><li><a href="./evals.html">Evals</a>: author <code>defineEval</code> cases, pick fixtures, and use evals as regression checks.</li><li><a href="./ab.html">Live A/B metrics</a>: assign sticky variants and compare cumulative metrics on live sessions.</li><li><a href="./storage.html">Storage</a>: point durable storage at a backend you own with <code>defineStorage</code>.</li><li><a href="./hillclimbing.html">Hillclimbing</a>: measure and improve an agent iteratively.</li></ul><p><strong>Guides</strong></p><ul><li><a href="./guides/webhooks.html">Webhooks and custom channels</a>: give the agent its own HTTP surface.</li><li><a href="./guides/github.html">GitHub</a>: trigger the agent from pull requests, CI, and comments.</li><li><a href="./guides/slack.html">Slack</a>: put the agent in Slack over Socket Mode.</li><li><a href="./guides/human-in-the-loop.html">Human-in-the-loop approvals</a>: park a tool call until a person signs off.</li><li><a href="./guides/mcp-oauth.html">Host MCP OAuth</a>: authorize <code>oauth: true</code> connections, store tokens locally, and persist them on hosted deployments (<code>--store</code>).</li><li><a href="./guides/agent-to-agent.html">Agent-to-agent</a>: every agent is an MCP server; agents can delegate to each other.</li><li><a href="./guides/cloud-agents.html">Cursor cloud agents</a>: let the agent launch and steer cloud agents through the <code>cursor-cloud-agents</code> extension.</li><li><a href="./guides/grokbot-agents.html">Cursor Grok Bot agents</a>: let the agent talk to Grok Bot agents through the <code>cursor-grokbot-agents</code> extension.</li><li><a href="./guides/opentelemetry.html">OpenTelemetry</a>: push session, turn, and tool traces to an OTLP collector you run.</li></ul><p><strong>Operating</strong></p><ul><li><a href="./deployment.html">Deployment</a>: Cursor-managed hosting, self-hosting, auth, state, and operations.</li><li><a href="./troubleshooting.html">Fix common agent problems</a>: diagnose common failures by symptom.</li></ul><p><strong>Reference</strong></p><ul><li><a href="./reference/project-layout.html">Project layout</a>: the full folder structure.</li><li><a href="./reference/agent-config.html">Agent config</a> · <a href="./reference/instructions.html">Instructions</a> · <a href="./reference/tools.html">Tools</a> · <a href="./reference/prompt.html"><code>prompt</code></a> · <a href="./reference/skills.html">Skills</a> · <a href="./reference/connections.html">MCP connections</a> · <a href="./reference/subagents.html">Subagents</a> · <a href="./reference/extensions.html">Extensions</a></li><li><a href="./reference/channels.html">Channels</a> · <a href="./reference/schedules.html">Schedules and reminders</a> · <a href="./reference/hooks.html">Hooks</a> · <a href="./reference/artifacts.html">Artifacts</a> · <a href="./reference/result.html">Turn result</a> · <a href="./reference/sessions.html">Sessions and streaming</a> · <a href="./reference/playground.html">Playground</a></li><li><a href="./reference/cli.html">CLI</a> · <a href="./reference/http-api.html">HTTP API</a></li></ul><h2 id="run-the-cli" tabindex="-1">Run the CLI <a class="header-anchor" href="#run-the-cli" aria-label="Permalink to &quot;Run the CLI&quot;">​</a></h2><p>Docs use <code>agent-sdk &lt;command&gt;</code>. If it isn&#39;t on <code>PATH</code>, use <code>npx @cursor/july &lt;command&gt;</code>.</p><h2 id="credentials" tabindex="-1">Credentials <a class="header-anchor" href="#credentials" aria-label="Permalink to &quot;Credentials&quot;">​</a></h2><p>Sign in to Cursor or set <code>CURSOR_API_KEY</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;"> login</span></span>
-<span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;"># or: export CURSOR_API_KEY=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;"> whoami</span></span></code></pre></div><p>Confirm <code>agent-sdk whoami</code> shows the expected account.</p><h2 id="related-documentation" tabindex="-1">Related documentation <a class="header-anchor" href="#related-documentation" aria-label="Permalink to &quot;Related documentation&quot;">​</a></h2><ul><li>Package reference: <a href="./../README.html"><code>README.md</code></a></li><li>Coding-agent workflows: <a href="./skills/"><code>skills/</code></a></li></ul>`,31)])])}const p=t(i,[["render",n]]);export{g as __pageData,p as default};

package/dist/docs/assets/reference_sessions.md.CueyOHSL.js

@@ -1 +0,0 @@
-import{_ as t,c as s,o as a,ag as o}from"./chunks/framework.BCISBCiQ.js";const u=JSON.parse('{"title":"Sessions, events, and streaming","description":"Understand how conversations continue, how events stream, and where session data lives.","frontmatter":{"title":"Sessions, events, and streaming","description":"Understand how conversations continue, how events stream, and where session data lives."},"headers":[],"relativePath":"reference/sessions.md","filePath":"reference/sessions.md"}'),n={name:"reference/sessions.md"};function d(i,e,r,c,l,h){return a(),s("div",null,[...e[0]||(e[0]=[o('<h1 id="sessions-events-and-streaming" tabindex="-1">Sessions, events, and streaming <a class="header-anchor" href="#sessions-events-and-streaming" aria-label="Permalink to &quot;Sessions, events, and streaming&quot;">​</a></h1><p>A session keeps one conversation, its workspace, and an append-only record of every message and tool call.</p><h2 id="what-does-a-session-contain" tabindex="-1">What does a session contain? <a class="header-anchor" href="#what-does-a-session-contain" aria-label="Permalink to &quot;What does a session contain?&quot;">​</a></h2><p>Each session combines:</p><ul><li>A channel and authenticated caller</li><li>A conversation the caller can continue</li><li>A workspace for local turns</li><li>An NDJSON event stream</li><li>Runtime state needed to resume after a server restart</li></ul><p>Sessions belong to the principal that created them. Follow-up, stream, and list routes return <code>403</code> when another caller tries to access one.</p><h2 id="which-session-identifier-should-i-use" tabindex="-1">Which session identifier should I use? <a class="header-anchor" href="#which-session-identifier-should-i-use" aria-label="Permalink to &quot;Which session identifier should I use?&quot;">​</a></h2><p>Sessions have two identifiers because conversation routing and inspection are different jobs.</p><table tabindex="0"><thead><tr><th>Identifier</th><th>Use it for</th></tr></thead><tbody><tr><td><code>continuationToken</code></td><td>Continue a conversation through its channel</td></tr><tr><td><code>sessionId</code></td><td>Stream events, inspect state, resolve approvals, or run a session-bound tool call</td></tr></tbody></table><p>Channels decide what a continuation token looks like. Slack uses its thread identity. A PR channel can use a key such as <code>pr:owner/repo#1</code>. The built-in HTTP API returns an opaque token and rotates it after each accepted follow-up. Reusing a stale HTTP token returns <code>409</code>.</p><p>Use the continuation token to keep talking. Use the session ID to observe or manage the stored session.</p><h2 id="which-session-modes-are-available" tabindex="-1">Which session modes are available? <a class="header-anchor" href="#which-session-modes-are-available" aria-label="Permalink to &quot;Which session modes are available?&quot;">​</a></h2><table tabindex="0"><thead><tr><th>Mode</th><th>Created by</th><th>What happens after a turn</th></tr></thead><tbody><tr><td><code>chat</code></td><td>HTTP sessions, channel <code>send</code>, Slack, or MCP <code>ask</code></td><td>Waits in <code>session.waiting</code> and accepts follow-ups</td></tr><tr><td><code>task</code></td><td>Markdown schedules and fire-and-forget dispatch</td><td>Ends in <code>session.completed</code> or <code>session.failed</code></td></tr></tbody></table><p>Task sessions don&#39;t accept follow-ups. Trying one returns <code>409</code>.</p><h2 id="what-happens-when-i-send-a-follow-up" tabindex="-1">What happens when I send a follow-up? <a class="header-anchor" href="#what-happens-when-i-send-a-follow-up" aria-label="Permalink to &quot;What happens when I send a follow-up?&quot;">​</a></h2><p>A follow-up to an idle chat session starts another turn. Admission when the session is already busy depends on the channel:</p><table tabindex="0"><thead><tr><th>Path</th><th>Busy-session policy</th></tr></thead><tbody><tr><td>HTTP playground / <code>POST /v1/session/:id</code> / MCP <code>ask</code></td><td><strong>Preempt</strong> (default): interrupt the in-flight turn, wait for it to settle, then run the new message</td></tr><tr><td>Slack mentions / DMs / alert-watch</td><td><strong>Coalesce</strong>: leave the active turn running, enqueue the follow-up, and drain queued asks into one follow-up turn when the active turn finishes (no mid-turn tool/hook inject)</td></tr></tbody></table><p>Pass <code>admission: &quot;coalesce&quot;</code> on <code>send()</code> to opt into the Slack policy from other callers. Omit it (or pass <code>&quot;preempt&quot;</code>) to keep interrupt semantics.</p><p><code>POST /v1/session/:id/stop</code> interrupts a turn without sending a new message. Interrupted turns record <code>turn.failed</code> with <code>&quot;turn interrupted&quot;</code>. This means the turn was preempted. A whole-message Slack <code>stop</code> / <code>@agent stop</code> does the same for that thread and clears pending coalesced nudges.</p><p>Session-bound deterministic tool calls share the lock only for writes: a write-effect call returns <code>409 session_busy</code> while a model turn is running, a read-effect call runs alongside the turn (see <a href="./tools.html#call-a-tool-without-a-model-turn">Tools</a>).</p><h2 id="which-events-can-i-stream" tabindex="-1">Which events can I stream? <a class="header-anchor" href="#which-events-can-i-stream" aria-label="Permalink to &quot;Which events can I stream?&quot;">​</a></h2><p>Each NDJSON line uses this envelope: <code>{ type, index, sessionId, turnId?, at, data }</code>. The <code>index</code> increases within one session. The <code>at</code> field is an ISO-8601 timestamp.</p><table tabindex="0"><thead><tr><th>Phase</th><th>Events</th><th>What they tell you</th></tr></thead><tbody><tr><td>Session</td><td><code>session.started</code>, <a href="./../ab.html#assign-sticky-variants"><code>ab.assigned</code></a>, <code>session.waiting</code>, <code>session.completed</code>, <code>session.failed</code></td><td>Session creation, A/B enrollment, readiness, and task completion</td></tr><tr><td>Agent</td><td><code>agent.bound</code></td><td>Cloud conversation URL</td></tr><tr><td>Input</td><td><code>message.received</code></td><td>A user message was accepted</td></tr><tr><td>Turn</td><td><code>turn.queued</code>, <code>turn.started</code>, <code>turn.completed</code>, <code>turn.failed</code></td><td>Queue position under a <a href="./agent-config.html#concurrency"><code>maxRunningTurns</code> cap</a>, then turn status, final result, and token usage</td></tr><tr><td>Steps</td><td><code>step.started</code>, <code>step.completed</code></td><td>Model step boundaries and duration</td></tr><tr><td>Reasoning</td><td><code>reasoning.appended</code>, <code>reasoning.completed</code></td><td>Streamed reasoning blocks</td></tr><tr><td>Reply</td><td><code>message.appended</code>, <code>message.completed</code></td><td>Text deltas and finalized assistant messages</td></tr><tr><td>Tools</td><td><code>actions.requested</code>, <code>action.result</code></td><td>Tool names, validated arguments, outputs, and errors</td></tr><tr><td>Approvals</td><td><code>action.approval_requested</code>, <code>action.approval_resolved</code></td><td>A parked tool call and the human decision</td></tr><tr><td>Subagents</td><td><code>subagent.called</code>, <code>subagent.completed</code></td><td>Delegated work</td></tr><tr><td>Artifacts</td><td><code>artifact.tagged</code></td><td>A durable <a href="./artifacts.html">artifact</a> was tagged for this session, by host code or <code>tag_artifact</code></td></tr></tbody></table><p>Pair <code>actions.requested</code> with <code>action.result</code> to reconstruct the tool trajectory. Read <code>turn.completed.data.usage</code> for input, output, and cache token counts. When <code>agent/result.ts</code> is authored, a thrown <code>commit</code> or empty assistant text emits <code>turn.failed</code> instead of <code>turn.completed</code>. See <a href="./result.html">Turn result</a>.</p><h2 id="how-do-i-stream-or-replay-session-events" tabindex="-1">How do I stream or replay session events? <a class="header-anchor" href="#how-do-i-stream-or-replay-session-events" aria-label="Permalink to &quot;How do I stream or replay session events?&quot;">​</a></h2><p>One endpoint handles both live streaming and replay:</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;">curl</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> -N</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &#39;http://127.0.0.1:3000/&lt;slug&gt;/v1/session/ses_…/stream?startIndex=0&#39;</span></span></code></pre></div><p>Pass <code>startIndex</code> to continue after the last event you received. Omit it or pass <code>0</code> to replay the full session before following new events. <code>GET /v1/session/:id/events</code> returns a one-time dump without staying connected.</p><p>Event streams replay from disk after a server restart. Conversation state resumes from the Cursor SDK store.</p><h2 id="what-goes-into-a-local-session-workspace" tabindex="-1">What goes into a local session workspace? <a class="header-anchor" href="#what-goes-into-a-local-session-workspace" aria-label="Permalink to &quot;What goes into a local session workspace?&quot;">​</a></h2><p>The Agent SDK creates a workspace before the first local turn:</p><table tabindex="0"><thead><tr><th>Source path</th><th>Lands as</th></tr></thead><tbody><tr><td><code>instructions.*</code></td><td><code>AGENTS.md</code></td></tr><tr><td><code>skills/*</code></td><td><code>.cursor/skills/&lt;name&gt;/SKILL.md</code></td></tr><tr><td>agent tools (<code>execution: &quot;agent&quot;</code>)</td><td>scripts in the session workspace, with a catalog in <code>AGENTS.md</code></td></tr><tr><td><code>sandbox/workspace/**</code></td><td>copied in as seed files</td></tr><tr><td>per-send <code>workspaceFiles</code></td><td>written before the turn</td></tr></tbody></table><p>The local harness uses this workspace as its working directory. Parent directories can contribute <code>AGENTS.md</code> and <code>.cursor</code> settings. Set <code>local.cwd</code> when you need a clean parent directory. A channel can also provide a different working directory for one session, such as a PR worktree.</p><p>See <a href="./agent-config.html#local-cwd">Agent config: local cwd</a> for the inheritance rules.</p><h2 id="where-does-the-agent-sdk-store-session-data" tabindex="-1">Where does the Agent SDK store session data? <a class="header-anchor" href="#where-does-the-agent-sdk-store-session-data" aria-label="Permalink to &quot;Where does the Agent SDK store session data?&quot;">​</a></h2><p>Local state lives under <code>--state-root</code>. Slugged mounts store it under a subdirectory named for the slug.</p><p>Deleting a session directory removes the session from the server: it disappears from listings and can no longer be streamed or continued. Cloud conversations remain on the Cursor backend.</p><p>Nested git checkouts already default <code>local.cwd</code> outside the enclosing repo. See <a href="./../concepts.html#what-files-can-a-local-session-access">local session workspaces</a>.</p><h2 id="how-do-i-inspect-a-saved-event-stream" tabindex="-1">How do I inspect a saved event stream? <a class="header-anchor" href="#how-do-i-inspect-a-saved-event-stream" aria-label="Permalink to &quot;How do I inspect a saved event stream?&quot;">​</a></h2><p>Use <code>trajectory</code> with a saved 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;"> trajectory</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --events</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> &lt;</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">state-roo</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">t</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">&gt;</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">/traces/</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">&lt;</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">sessionI</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">d</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">&gt;</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">.ndjson</span></span></code></pre></div><p>The command prints tool calls, the reply, and token usage in the same JSON shape as <code>run</code>. Use <strong>Open trace</strong> in the playground for a visual view.</p><h2 id="related" tabindex="-1">Related <a class="header-anchor" href="#related" aria-label="Permalink to &quot;Related&quot;">​</a></h2><ul><li><a href="./http-api.html">HTTP API</a></li><li><a href="./hooks.html">Hooks</a></li><li><a href="./../ab.html">Live A/B metrics</a></li><li><a href="./../concepts.html">How the Agent SDK works</a></li></ul>',44)])])}const m=t(n,[["render",d]]);export{u as __pageData,m as default};

package/dist/docs/assets/skills_ab.md.CsFNatVx.js

@@ -1,26 +0,0 @@
-import{_ as i,c as a,o as n,ag as t}from"./chunks/framework.BCISBCiQ.js";const o=JSON.parse('{"title":"Agent SDK A/B metrics (defineAB)","description":"Author defineAB live A/B metrics under agent/ab. Sticky variants and onSample on live runs. Not a test runner. Use when comparing variants on traffic; use defineEval for regression gates.","frontmatter":{"name":"agentsdk-ab","description":"Author defineAB live A/B metrics under agent/ab. Sticky variants and onSample on live runs. Not a test runner. Use when comparing variants on traffic; use defineEval for regression gates."},"headers":[],"relativePath":"../skills/ab/SKILL.md","filePath":"../skills/ab/SKILL.md"}'),e={name:"../skills/ab/SKILL.md"};function l(p,s,h,k,d,r){return n(),a("div",null,[...s[0]||(s[0]=[t(`<h1 id="agent-sdk-a-b-metrics-defineab" tabindex="-1">Agent SDK A/B metrics (<code>defineAB</code>) <a class="header-anchor" href="#agent-sdk-a-b-metrics-defineab" aria-label="Permalink to &quot;Agent SDK A/B metrics (\`defineAB\`)&quot;">​</a></h1><p>Live metrics plug-in. No <code>agent-sdk ab</code> CLI. No assertion API. Reference: <code>docs/ab.md</code>.</p><table tabindex="0"><thead><tr><th></th><th><code>defineEval</code></th><th><code>defineAB</code></th></tr></thead><tbody><tr><td>Job</td><td>Gates on frozen fixtures</td><td>Metrics on live runs</td></tr><tr><td>Location</td><td><code>evals/**/*.eval.ts</code></td><td><code>agent/ab.ts</code> or <code>agent/ab/&lt;name&gt;.ts</code></td></tr><tr><td>How it runs</td><td><code>agent-sdk eval</code></td><td>Under <code>serve</code> / <code>run</code></td></tr></tbody></table><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;"> { defineAB, splitBySessionHash } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;@cursor/july/ab&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;"> defineAB</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">({</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  name: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;concise-instructions&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  variants: {</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">    control: { label: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;Baseline&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> },</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">    treatment: {</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">      label: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;Shorter&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">      instructions: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;Keep replies to one short paragraph.&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>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  },</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  split: </span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">splitBySessionHash</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">({ holdout: </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">0.1</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> }),</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  derive: {</span></span>
-<span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">    weatherCalls</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: (</span><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">event</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">) </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">=&gt;</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">      event.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:#D73A49;--shiki-dark:#F97583;"> &amp;&amp;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> event.data.toolName </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">===</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;get_weather&quot;</span></span>
-<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">        ?</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> 1</span></span>
-<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">        :</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> null</span><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:#6F42C1;--shiki-dark:#B392F0;">  onSample</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">sample</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">) {</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">    console.</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">log</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(sample.variant, sample.metrics.toolCalls, sample.metrics.wallTimeMs);</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><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:#24292E;--shiki-dark:#E1E4E8;">async </span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">execute</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(input, ctx) {</span></span>
-<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">  if</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> (ctx.session.abs?.[</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;concise-instructions&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">] </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">===</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;treatment&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">) {</span></span>
-<span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">    // treatment-specific behavior</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>Enrollment is at session creation. Eval sessions skip it. Do not use <code>splitIf</code> to filter evals. Split helpers and <code>onSample</code> fields: <code>docs/ab.md</code>.</p><p>Pick a name, arm labels, a split, and a real <code>onSample</code> sink. Do not invent credentials.</p>`,7)])])}const c=i(e,[["render",l]]);export{o as __pageData,c as default};

package/dist/docs/assets/skills_ab.md.CsFNatVx.lean.js

@@ -1 +0,0 @@
-import{_ as i,c as a,o as n,ag as t}from"./chunks/framework.BCISBCiQ.js";const o=JSON.parse('{"title":"Agent SDK A/B metrics (defineAB)","description":"Author defineAB live A/B metrics under agent/ab. Sticky variants and onSample on live runs. Not a test runner. Use when comparing variants on traffic; use defineEval for regression gates.","frontmatter":{"name":"agentsdk-ab","description":"Author defineAB live A/B metrics under agent/ab. Sticky variants and onSample on live runs. Not a test runner. Use when comparing variants on traffic; use defineEval for regression gates."},"headers":[],"relativePath":"../skills/ab/SKILL.md","filePath":"../skills/ab/SKILL.md"}'),e={name:"../skills/ab/SKILL.md"};function l(p,s,h,k,d,r){return n(),a("div",null,[...s[0]||(s[0]=[t("",7)])])}const c=i(e,[["render",l]]);export{o as __pageData,c as default};

package/dist/docs/assets/skills_evals.md.723kpUmA.js

@@ -1,25 +0,0 @@
-import{_ as i,c as a,o as e,ag as t}from"./chunks/framework.BCISBCiQ.js";const c=JSON.parse('{"title":"Agent SDK evals","description":"Author and run defineEval cases. Use when writing, fixing, seeding, or hillclimbing evals. Ask whether to generate samples or take uploaded data. Materialize API-backed fixtures first. Live A/B is defineAB (skills/ab), not defineEval.","frontmatter":{"name":"agentsdk-evals","description":"Author and run defineEval cases. Use when writing, fixing, seeding, or hillclimbing evals. Ask whether to generate samples or take uploaded data. Materialize API-backed fixtures first. Live A/B is defineAB (skills/ab), not defineEval."},"headers":[],"relativePath":"../skills/evals/SKILL.md","filePath":"../skills/evals/SKILL.md"}'),n={name:"../skills/evals/SKILL.md"};function l(h,s,d,p,k,r){return e(),a("div",null,[...s[0]||(s[0]=[t(`<h1 id="agent-sdk-evals" tabindex="-1">Agent SDK evals <a class="header-anchor" href="#agent-sdk-evals" aria-label="Permalink to &quot;Agent SDK evals&quot;">​</a></h1><p>Fixed input, model turn, gates on the trajectory. Files live at project-root <code>evals/**/*.eval.ts</code>. <code>agent/evals/</code> is ignored.</p><p>Live traffic variants: <code>skills/ab/SKILL.md</code>. That is not a test runner.</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;"> --list</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>
-<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;"> weather/nyc</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></code></pre></div><table tabindex="0"><thead><tr><th>Form</th><th>Case id</th></tr></thead><tbody><tr><td><code>evals/weather.eval.ts</code> + <code>test</code></td><td><code>weather</code></td></tr><tr><td><code>evals/weather/nyc.eval.ts</code> + <code>test</code></td><td><code>weather/nyc</code></td></tr><tr><td><code>evals/weather.eval.ts</code> + <code>{ id: &quot;nyc&quot; }</code></td><td><code>weather/nyc</code></td></tr></tbody></table><p><code>eval</code> boots an ephemeral server and a temp state root. <code>--url</code> points at a running agent. Model turns need <code>CURSOR_API_KEY</code>.</p><h2 id="seeding" tabindex="-1">Seeding <a class="header-anchor" href="#seeding" aria-label="Permalink to &quot;Seeding&quot;">​</a></h2><p>Creating or expanding cases: <code>AskQuestion</code> first.</p><table tabindex="0"><thead><tr><th>Question</th><th>Options</th></tr></thead><tbody><tr><td><strong>How should we get eval samples?</strong></td><td>Generate test eval samples for me <code>(recommended)</code> · I will add / upload the data manually</td></tr></tbody></table><ol><li><strong>Manual.</strong> They provide files or paste. Show the shape below. Do not invent cases. Then <code>eval --list</code> and wire gates. API-backed pointers (PR URLs, SHAs, gold labels): materialize under <code>fixtures/</code> first.</li><li><strong>Generated.</strong> Ask count (<code>3</code> recommended). Append to an existing <code>cases</code> array when it fits. Never overwrite or weaken a datapoint. Create <code>evals/evals.config.ts</code> if missing (<code>maxConcurrency: 20</code>; cap 200).</li></ol><h2 id="api" tabindex="-1">API <a class="header-anchor" href="#api" aria-label="Permalink to &quot;API&quot;">​</a></h2><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;"> { defineEval, includes, satisfies } </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>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  tags: [</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;smoke&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">, </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;weather&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">],</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  cases: [</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;">      id: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;nyc&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">      description: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;NYC temperature.&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
-<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">      async</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> test</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">t</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">) {</span></span>
-<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">        await</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> t.</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">send</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;What&#39;s the temperature in NYC?&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;">succeeded</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;">calledTool</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;get_weather&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;">notCalledTool</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;save_weather_note&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;">(t.reply, </span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">includes</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">/</span><span style="--shiki-light:#032F62;--shiki-dark:#DBEDFF;">°</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">|</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">[FC]</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">/</span><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>
-<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><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;"> { defineEvalConfig } </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;"> defineEvalConfig</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">({</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  maxConcurrency: </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">20</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>Either <code>test(t)</code> or <code>cases</code>, not both. <code>t.send</code> waits for park/fail. <code>workspaceFiles</code> seeds the first turn. Assert with <code>t.succeeded()</code>, <code>calledTool</code> / <code>notCalledTool</code>, <code>t.check(t.reply, …)</code>, <code>t.metric</code>.</p><h2 id="what-to-gate" tabindex="-1">What to gate <a class="header-anchor" href="#what-to-gate" aria-label="Permalink to &quot;What to gate&quot;">​</a></h2><p>Decisions and shape, not prose.</p><ol><li><code>t.succeeded()</code> first</li><li>Intended tool + the tempting wrong one</li><li>A shape regex or <code>satisfies</code> on parsed fields</li><li>If formatting keeps failing, tighten instructions. Do not move rendering into a host tool.</li></ol><p>Anti-patterns: exact phrasing; more than ~5 gates (split); live drifting inputs (pin them).</p><table tabindex="0"><thead><tr><th>Surface</th><th>Fixture</th></tr></thead><tbody><tr><td>Chat</td><td>One frozen prompt</td></tr><tr><td>Tool-heavy</td><td><code>agent-sdk call</code> first, then the prompt</td></tr><tr><td>GitHub</td><td><code>github replay … --dry-run --out fixtures/github</code></td></tr><tr><td>Host-prep PR review</td><td>A team-owned PR; gate findings shape, not counts</td></tr><tr><td>Workspace</td><td><code>workspaceFiles</code> in <code>t.send</code></td></tr></tbody></table><p>Every kept hillclimb change lands an eval that would have failed before it. Never weaken a gate to pass a round.</p>`,20)])])}const E=i(n,[["render",l]]);export{c as __pageData,E as default};

package/dist/docs/assets/skills_evals.md.723kpUmA.lean.js

@@ -1 +0,0 @@
-import{_ as i,c as a,o as e,ag as t}from"./chunks/framework.BCISBCiQ.js";const c=JSON.parse('{"title":"Agent SDK evals","description":"Author and run defineEval cases. Use when writing, fixing, seeding, or hillclimbing evals. Ask whether to generate samples or take uploaded data. Materialize API-backed fixtures first. Live A/B is defineAB (skills/ab), not defineEval.","frontmatter":{"name":"agentsdk-evals","description":"Author and run defineEval cases. Use when writing, fixing, seeding, or hillclimbing evals. Ask whether to generate samples or take uploaded data. Materialize API-backed fixtures first. Live A/B is defineAB (skills/ab), not defineEval."},"headers":[],"relativePath":"../skills/evals/SKILL.md","filePath":"../skills/evals/SKILL.md"}'),n={name:"../skills/evals/SKILL.md"};function l(h,s,d,p,k,r){return e(),a("div",null,[...s[0]||(s[0]=[t("",20)])])}const E=i(n,[["render",l]]);export{c as __pageData,E as default};

package/dist/docs/assets/skills_framework-map.md.haibFyoB.js

@@ -1 +0,0 @@
-import{_ as t,c as d,o,ag as r}from"./chunks/framework.BCISBCiQ.js";const g=JSON.parse('{"title":"Agent SDK framework map","description":"Orient on Agent SDK layout, local vs cloud runtime, sessions, and invariants (Node not Bun, root evals/, typecheck after tsx). Use when creating, editing, or running a project with agent/agent.ts.","frontmatter":{"name":"agentsdk-framework-map","description":"Orient on Agent SDK layout, local vs cloud runtime, sessions, and invariants (Node not Bun, root evals/, typecheck after tsx). Use when creating, editing, or running a project with agent/agent.ts."},"headers":[],"relativePath":"../skills/framework-map/SKILL.md","filePath":"../skills/framework-map/SKILL.md"}'),s={name:"../skills/framework-map/SKILL.md"};function c(a,e,n,l,i,u){return o(),d("div",null,[...e[0]||(e[0]=[r('<h1 id="agent-sdk-framework-map" tabindex="-1">Agent SDK framework map <a class="header-anchor" href="#agent-sdk-framework-map" aria-label="Permalink to &quot;Agent SDK framework map&quot;">​</a></h1><p><code>@cursor/july</code> discovers files under <code>agent/</code> and serves the agent over HTTP, Slack, and GitHub. Markdown is prose. TypeScript is typed behavior. Ground truth: package <code>README.md</code> and <code>AGENTS.md</code>.</p><p>CLI is <code>agent-sdk</code> (Node, never Bun).</p><p>Public docs: <code>node_modules/@cursor/july/dist/docs/llms.txt</code> or <code>/docs/llms.txt</code> on a running host.</p><h2 id="invariants" tabindex="-1">Invariants <a class="header-anchor" href="#invariants" aria-label="Permalink to &quot;Invariants&quot;">​</a></h2><ol><li><strong>Node 22.13+, never Bun.</strong> Bun corrupts harness tool-result streams (<code>NGHTTP2_FRAME_SIZE_ERROR</code>).</li><li><strong>Evals live at project-root <code>evals/</code>.</strong> <code>agent/evals/</code> is ignored.</li><li><strong>tsx does not typecheck.</strong> Tool <code>execute</code> must return JSON-shaped values: object literals or <code>type</code> aliases, not <code>interface</code> types.</li><li><strong>Nested git checkouts.</strong> Discovery sets <code>local.cwd</code> to a per-project cache directory under <code>~/.cache</code>. Point cwd at a checkout only when the agent must inherit that tree.</li><li><strong>Attached MCP is nameless</strong> until <code>advertiseTools: true</code>.</li><li><strong>Model turns need <code>CURSOR_API_KEY</code>.</strong> <code>validate</code>, <code>info</code>, <code>call</code>, and <code>serve</code> bring-up do not.</li></ol><h2 id="folder-structure" tabindex="-1">Folder structure <a class="header-anchor" href="#folder-structure" aria-label="Permalink to &quot;Folder structure&quot;">​</a></h2><p>Path is identity. Full list: README &quot;Folder structure&quot;.</p><table tabindex="0"><thead><tr><th>Path</th><th>Role</th></tr></thead><tbody><tr><td><code>agent/agent.ts</code></td><td><code>defineAgent({ model?, runtime?, cloud?, local? })</code></td></tr><tr><td><code>agent/instructions.md</code></td><td>Always-on system prompt (required)</td></tr><tr><td><code>agent/tools/&lt;name&gt;.ts</code></td><td>One tool. <code>execution: &quot;server&quot;</code> or <code>&quot;agent&quot;</code></td></tr><tr><td><code>agent/skills/*</code></td><td>On-demand procedures</td></tr><tr><td><code>agent/mcp-connections/&lt;name&gt;.ts</code></td><td>MCP. Never name an account file <code>cursor.ts</code>. <code>advertiseTools: true</code> for named local tools</td></tr><tr><td><code>agent/host-connections/&lt;name&gt;.ts</code></td><td>Privileged MCP for <code>ctx.host.mcp</code> / <code>mcp oauth</code></td></tr><tr><td><code>agent/subagents/&lt;id&gt;/</code></td><td>Child agent (<code>description</code> required)</td></tr><tr><td><code>agent/extensions/&lt;ns&gt;.ts</code></td><td>Mounted extension or Cursor plugin (<code>cursorPlugin</code>); contributions are <code>&lt;ns&gt;__&lt;name&gt;</code></td></tr><tr><td><code>agent/channels/*.ts</code></td><td>Slack / GitHub / custom HTTP</td></tr><tr><td><code>agent/hooks/*.ts</code></td><td>Observe-only</td></tr><tr><td><code>agent/artifacts.ts</code></td><td>Durable tagged outputs (<code>defineArtifacts</code>)</td></tr><tr><td><code>agent/result.ts</code></td><td>Host <code>commit</code> on the final assistant text (<code>defineResult</code>)</td></tr><tr><td><code>agent/ab.ts</code> or <code>agent/ab/*.ts</code></td><td>Live A/B (<code>defineAB</code>)</td></tr><tr><td><code>agent/otel.ts</code></td><td>OpenTelemetry (<code>defineOtel</code>)</td></tr><tr><td><code>agent/schedules/*</code></td><td>Cron. Never auto-fire under <code>--dev</code></td></tr><tr><td><code>agent/sandbox/workspace/</code></td><td>Session seed files (local only)</td></tr><tr><td><code>agent/lib/</code></td><td>Import-only. Never discovered</td></tr><tr><td><code>evals/**/*.eval.ts</code></td><td>Case id is the path under <code>evals/</code></td></tr></tbody></table><h2 id="local-vs-cloud" tabindex="-1">Local vs cloud <a class="header-anchor" href="#local-vs-cloud" aria-label="Permalink to &quot;Local vs cloud&quot;">​</a></h2><p><code>runtime: &quot;local&quot;</code> (default) runs on the serve host. <code>runtime: &quot;cloud&quot;</code> needs <code>cloud: { repos: [...] }</code>.</p><table tabindex="0"><thead><tr><th>Capability</th><th>local</th><th>cloud</th></tr></thead><tbody><tr><td>Server tools</td><td>yes</td><td>yes on managed hosting; self-hosted needs <code>--public-url</code></td></tr><tr><td>Tool approvals</td><td>yes</td><td>no</td></tr><tr><td>Agent tools / skills</td><td>yes</td><td>yes</td></tr><tr><td>sandbox seeds</td><td>yes</td><td>no</td></tr><tr><td>Checkout</td><td>you arrange it</td><td>the VM carries it</td></tr></tbody></table><p>Use cloud when the job needs a checkout at scale. <code>validate</code> warns when cloud is combined with local-only capabilities.</p><h2 id="sessions" tabindex="-1">Sessions <a class="header-anchor" href="#sessions" aria-label="Permalink to &quot;Sessions&quot;">​</a></h2><ul><li><strong>continuationToken</strong> continues a conversation. HTTP follow-ups rotate it. Stale tokens return <code>409</code>.</li><li><strong>sessionId</strong> is the inspect handle (<code>GET /v1/session/:id/stream?startIndex=N</code>).</li></ul><p>A follow-up to a busy HTTP/MCP session interrupts the in-flight turn. Slack coalesces. Routes: <code>docs/reference/http-api.md</code>. Session files live under the project state directory (<code>--state-root</code>).</p><h2 id="where-logic-belongs" tabindex="-1">Where logic belongs <a class="header-anchor" href="#where-logic-belongs" aria-label="Permalink to &quot;Where logic belongs&quot;">​</a></h2><p>Code: side-effect gates, write dedupe, auth, evidence seeding, hard API caps. Model: formatting, summarizing, classification, replies. Budget: <code>skills/create-agent/SKILL.md</code>.</p><p>Loop: <code>validate</code> / <code>info</code> / <code>call</code> / <code>run</code> / <code>eval</code> / <code>serve</code>. Serve only this agent&#39;s directory.</p><table tabindex="0"><thead><tr><th>Task</th><th>Skill</th></tr></thead><tbody><tr><td>Scaffold</td><td><code>skills/create-agent/SKILL.md</code></td></tr><tr><td>Evals</td><td><code>skills/evals/SKILL.md</code></td></tr><tr><td>Live A/B</td><td><code>skills/ab/SKILL.md</code></td></tr><tr><td>OpenTelemetry</td><td><code>skills/otel/SKILL.md</code></td></tr><tr><td>GitHub</td><td><code>skills/github/SKILL.md</code></td></tr><tr><td>Slack</td><td><code>skills/setup-slack/SKILL.md</code></td></tr><tr><td>Host MCP OAuth</td><td><code>skills/mcp-auth/SKILL.md</code></td></tr><tr><td>Local triage</td><td><code>skills/debug/SKILL.md</code></td></tr><tr><td>Measured improvement</td><td><code>skills/hillclimb/SKILL.md</code></td></tr></tbody></table>',20)])])}const p=t(s,[["render",c]]);export{g as __pageData,p as default};

package/dist/docs/assets/skills_index.md.DKwIxzGg.js

@@ -1 +0,0 @@
-import{_ as e,c as a,o as d,ag as l}from"./chunks/framework.BCISBCiQ.js";const f=JSON.parse('{"title":"Coding-agent skills","description":"Task guides that ship with the Agent SDK for scaffolding, evals, channels, and hillclimbing.","frontmatter":{"title":"Coding-agent skills","description":"Task guides that ship with the Agent SDK for scaffolding, evals, channels, and hillclimbing."},"headers":[],"relativePath":"skills/index.md","filePath":"skills/index.md"}'),i={name:"skills/index.md"};function r(s,t,n,o,h,c){return d(),a("div",null,[...t[0]||(t[0]=[l('<h1 id="coding-agent-skills" tabindex="-1">Coding-agent skills <a class="header-anchor" href="#coding-agent-skills" aria-label="Permalink to &quot;Coding-agent skills&quot;">​</a></h1><p>Each skill is a procedure a coding agent can follow. Installing <code>@cursor/july</code> copies them into <code>~/.cursor/skills/agentsdk/</code>. This site publishes the same files.</p><p>See <a href="./../building-with-agents.html">Building agents with agents</a> for when to use each one.</p><table tabindex="0"><thead><tr><th>Skill</th><th>Use it to</th></tr></thead><tbody><tr><td><a href="./framework-map.html">framework-map</a></td><td>Learn the project layout and runtimes</td></tr><tr><td><a href="./create-agent.html">create-agent</a></td><td>Scaffold and verify a new agent</td></tr><tr><td><a href="./evals.html">evals</a></td><td>Write fixtures and regression checks</td></tr><tr><td><a href="./ab.html">ab</a></td><td>Compare variants on live traffic</td></tr><tr><td><a href="./otel.html">otel</a></td><td>Export OpenTelemetry traces</td></tr><tr><td><a href="./hillclimb.html">hillclimb</a></td><td>Improve an agent against fixed inputs</td></tr><tr><td><a href="./github.html">github</a></td><td>Add GitHub webhooks and replay events</td></tr><tr><td><a href="./setup-slack.html">setup-slack</a></td><td>Connect an agent to Slack</td></tr><tr><td><a href="./mcp-auth.html">mcp-auth</a></td><td>Authorize host MCP OAuth</td></tr><tr><td><a href="./debug.html">debug</a></td><td>Diagnose a local run</td></tr></tbody></table>',4)])])}const m=e(i,[["render",r]]);export{f as __pageData,m as default};

package/dist/docs/assets/skills_index.md.DKwIxzGg.lean.js

@@ -1 +0,0 @@
-import{_ as e,c as a,o as d,ag as l}from"./chunks/framework.BCISBCiQ.js";const f=JSON.parse('{"title":"Coding-agent skills","description":"Task guides that ship with the Agent SDK for scaffolding, evals, channels, and hillclimbing.","frontmatter":{"title":"Coding-agent skills","description":"Task guides that ship with the Agent SDK for scaffolding, evals, channels, and hillclimbing."},"headers":[],"relativePath":"skills/index.md","filePath":"skills/index.md"}'),i={name:"skills/index.md"};function r(s,t,n,o,h,c){return d(),a("div",null,[...t[0]||(t[0]=[l("",4)])])}const m=e(i,[["render",r]]);export{f as __pageData,m as default};