@cursor/july 0.1.39 → 0.1.41

package/dist/docs/assets/{evals.md.BQXI3rXy.js → evals.md.qeoGgPw0.js}

@@ -1,4 +1,4 @@
-import{_ as e,c as a,o as i,ag as t}from"./chunks/framework.CAZyNGu9.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 i(),a("div",null,[...s[0]||(s[0]=[t(`<h1 id="evals" tabindex="-1">Evals <a class="header-anchor" href="#evals" aria-label="Permalink to &quot;Evals&quot;">​</a></h1><p>An eval is a repeatable check that runs your agent against a fixed input and gates the recorded trajectory: the run completed, the right tool ran, the reply has the right shape. Evals are how you know a prompt tweak helped, a refactor didn&#39;t regress the agent, and last month&#39;s fix is still holding.</p><p>Evals exercise the same surface your users hit. The runner starts (or targets) a real agent server, drives sessions over the public API, and grades what comes back. A passing eval means the agent started, accepted a message, and did what you asserted.</p><div class="note custom-block github-alert"><p class="custom-block-title">NOTE</p><p>Import paths here use <code>@cursor/july/evals</code>. On projects still using <code>@anysphere/agent-serve</code>, swap the import and run <code>agent-serve eval</code>. See <a href="./README.html#run-the-cli">Run the CLI</a> for the full rename table.</p></div><h2 id="define-evals-with-defineeval" tabindex="-1">Define evals with <code>defineEval</code> <a class="header-anchor" href="#define-evals-with-defineeval" aria-label="Permalink to &quot;Define evals with \`defineEval\`&quot;">​</a></h2><p>The Agent SDK discovers evals under the project-root <code>evals/</code> directory, in <code>.eval.ts</code> or <code>.eval.js</code> files. That&#39;s a sibling of <code>agent/</code>, never inside it (<code>agent/evals/</code> is silently ignored). TypeScript is the normal authoring format.</p><p>The file path is the eval&#39;s identity, so you don&#39;t author an id. Directories group related evals: <code>evals/builds/api.eval.ts</code> becomes id <code>builds/api</code>. An <code>index</code> filename collapses to its directory, so <code>evals/builds/index.eval.ts</code> becomes <code>builds</code>.</p><p>An eval is a single <code>async test(t)</code>. You drive the agent with <code>t</code> and assert on the run with the same <code>t</code>:</p><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">// evals/readiness.eval.ts</span></span>
+import{_ as e,c as a,o as i,ag as t}from"./chunks/framework.CAZyNGu9.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 i(),a("div",null,[...s[0]||(s[0]=[t(`<h1 id="evals" tabindex="-1">Evals <a class="header-anchor" href="#evals" aria-label="Permalink to &quot;Evals&quot;">​</a></h1><p>An eval is a repeatable check that runs your agent against a fixed input and gates the recorded trajectory: the run completed, the right tool ran, the reply has the right shape. Evals are how you know a prompt tweak helped, a refactor didn&#39;t regress the agent, and last month&#39;s fix is still holding.</p><p>Evals exercise the same surface your users hit. The runner starts (or targets) a real agent server, drives sessions over the public API, and grades what comes back. A passing eval means the agent started, accepted a message, and did what you asserted.</p><div class="note custom-block github-alert"><p class="custom-block-title">NOTE</p><p>Import paths here use <code>@cursor/july/evals</code>. On projects still using <code>@anysphere/agent-serve</code>, swap the import and run <code>agent-serve eval</code>. See <a href="/docs/#run-the-cli">Run the CLI</a> for the full rename table.</p></div><h2 id="define-evals-with-defineeval" tabindex="-1">Define evals with <code>defineEval</code> <a class="header-anchor" href="#define-evals-with-defineeval" aria-label="Permalink to &quot;Define evals with \`defineEval\`&quot;">​</a></h2><p>The Agent SDK discovers evals under the project-root <code>evals/</code> directory, in <code>.eval.ts</code> or <code>.eval.js</code> files. That&#39;s a sibling of <code>agent/</code>, never inside it (<code>agent/evals/</code> is silently ignored). TypeScript is the normal authoring format.</p><p>The file path is the eval&#39;s identity, so you don&#39;t author an id. Directories group related evals: <code>evals/builds/api.eval.ts</code> becomes id <code>builds/api</code>. An <code>index</code> filename collapses to its directory, so <code>evals/builds/index.eval.ts</code> becomes <code>builds</code>.</p><p>An eval is a single <code>async test(t)</code>. You drive the agent with <code>t</code> and assert on the run with the same <code>t</code>:</p><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">// evals/readiness.eval.ts</span></span>
 <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>
@@ -48,7 +48,7 @@ import{_ as e,c as a,o as i,ag as t}from"./chunks/framework.CAZyNGu9.js";const k
 <span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">  // judge: { model: &quot;...&quot; },</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;"> // default judge model for t.judge.*</span></span>
 <span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">  // reporters: [],</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;"> // destinations that observe every case</span></span>
 <span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">  // maxPlaygroundRuns: 50,</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;"> // playground /v1/dev/evals history only (default 20)</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">});</span></span></code></pre></div><p>The timeout order is case or file <code>timeoutMs</code>, CLI <code>--timeout-ms</code>, project config <code>timeoutMs</code>, then the 180-second runner default.</p><p>The optional fields:</p><table tabindex="0"><thead><tr><th>Option</th><th>Default</th><th>Meaning</th></tr></thead><tbody><tr><td><code>timeoutMs</code></td><td><code>180_000</code></td><td>Project-wide per-case timeout</td></tr><tr><td><code>judge</code></td><td>unset</td><td>Default judge model for <code>t.judge.*</code>; see <a href="#judge-free-form-output">Judge free-form output</a></td></tr><tr><td><code>reporters</code></td><td>unset</td><td>Destinations that observe every case; <code>--skip-report</code> suppresses them</td></tr><tr><td><code>maxPlaygroundRuns</code></td><td><code>20</code></td><td>Max batches in the playground / <code>/v1/dev/evals*</code> history (not CLI <code>eval</code>)</td></tr></tbody></table><p>Reporters come from <code>@cursor/july/evals/reporters</code>: <code>JUnit</code> writes a JUnit XML file for CI, <code>Artifacts</code> writes per-case files, and <code>combineReporters</code> merges several into one (<code>renderJUnitXml</code> renders the XML for a custom destination). A file or case can add its own <code>reporters</code> on top of the config list.</p><p>Playground batches live in process memory and disappear when <code>serve</code> exits. Navigating away and back still works while the process is up. To keep batches across restarts, declare an <code>evals</code> table in <code>agent/storage.ts</code>; see <a href="./storage.html#eval-and-ab-tables">Storage</a>.</p><h2 id="drive-and-assert-with-t" tabindex="-1">Drive and assert with <code>t</code> <a class="header-anchor" href="#drive-and-assert-with-t" aria-label="Permalink to &quot;Drive and assert with \`t\`&quot;">​</a></h2><p><code>t</code> is both the driver and the assertion surface. You write ordinary control flow, sending turns and asserting inline.</p><p>Drive the agent with <code>t.send(message, options?)</code>. It runs one turn and waits for the session to park or fail. Multiple sends in one case share the session, which is how you write multi-turn evals.</p><p>Each <code>t.send</code> resolves to a turn result with <code>message</code>, <code>sessionId</code>, <code>events</code>, <code>toolCalls</code>, <code>ok</code>, and <code>index</code>. The turn carries the same assertion vocabulary as <code>t</code>, scoped to that turn, so you can grade an intermediate turn before the next send overwrites <code>t.reply</code>. <code>turn.expectOk()</code> throws when the turn failed, for later steps that depend on it.</p><p>Read the full case state with <code>t.reply</code> (the last assistant text), <code>t.events</code> (every captured session event across turns), <code>t.turns</code> (settled turns, oldest first), and <code>t.sessionId</code>. <code>t.signal</code> aborts when the case hits its timeout; pass it to your own async work.</p><p>Assert with the gates:</p><table tabindex="0"><thead><tr><th>Gate</th><th>Checks</th></tr></thead><tbody><tr><td><code>t.succeeded()</code></td><td>the run did not fail and is not parked on an unanswered approval</td></tr><tr><td><code>t.parked()</code></td><td>the run cleanly parked on an unanswered approval request</td></tr><tr><td><code>t.messageIncludes(token)</code></td><td>the joined assistant text matches a string or <code>RegExp</code></td></tr><tr><td><code>t.calledTool(name, matcher?)</code></td><td>a matching call to <code>name</code> happened</td></tr><tr><td><code>t.notCalledTool(name)</code></td><td>no request for <code>name</code>, in any lifecycle state</td></tr><tr><td><code>t.loadedSkill(name)</code></td><td>the agent opened the skill&#39;s <code>SKILL.md</code> (read, grep, or shell <code>cat</code>)</td></tr><tr><td><code>t.toolOrder(names)</code></td><td>tool requests appear in this relative order (extra calls allowed)</td></tr><tr><td><code>t.usedNoTools()</code></td><td>no tool calls at all</td></tr><tr><td><code>t.maxToolCalls(max)</code></td><td>at most <code>max</code> tool calls</td></tr><tr><td><code>t.noFailedActions()</code></td><td>no tool call reported an error</td></tr><tr><td><code>t.calledSubagent(name, matcher?)</code></td><td>a matching subagent delegation happened</td></tr><tr><td><code>t.taggedArtifact(kind?, predicate?)</code></td><td>at least one <a href="./reference/artifacts.html">artifact</a> was tagged</td></tr><tr><td><code>t.event(type, matcher?)</code></td><td>at least one matching event of <code>type</code> occurred</td></tr><tr><td><code>t.notEvent(type, matcher?)</code></td><td>no matching event of <code>type</code> occurred</td></tr><tr><td><code>t.eventOrder(matchers)</code></td><td>matching event groups occur in this relative order</td></tr><tr><td><code>t.eventsSatisfy(label, predicate)</code></td><td>your predicate over the typed event stream</td></tr><tr><td><code>t.check(value, expectation)</code></td><td>any value, against a builder</td></tr><tr><td><code>t.score(name, value)</code></td><td>records a 0–1 score you computed; soft until you add a bar</td></tr><tr><td><code>t.requireToolCall(name, matcher?)</code></td><td>gates on a matching call and returns it, so later code can read its input and output</td></tr><tr><td><code>t.requireInputRequest(filter?)</code></td><td>gates on exactly one pending approval request and returns it</td></tr></tbody></table><p>Every gate returns a handle: <code>.soft()</code> demotes it to tracked-only, <code>.atLeast(0.7)</code> adds a soft score bar, and <code>.gate(0.8)</code> promotes a scored assertion into a hard gate.</p><p>With no matcher, <code>calledTool</code> is request-based: a requested call counts even when its result has not arrived. Pass <code>t.calledTool(&quot;inspect_pr&quot;, { status: &quot;completed&quot; })</code> to require the call to return. <code>input</code>, <code>output</code>, and <code>count</code> matcher fields accept a literal, a <code>RegExp</code>, or a predicate.</p><p>The expectation builders are <code>includes(string | RegExp)</code>, <code>equals(value)</code>, <code>matches(schema)</code>, <code>similarity(expected)</code>, and <code>satisfies(predicate, label)</code>. <code>includes</code> stringifies its input, <code>equals</code> compares values deeply, <code>matches</code> validates against a Standard Schema (or anything with <code>safeParse</code>, like Zod), <code>similarity</code> scores normalized text similarity, and <code>satisfies</code> runs your predicate. The plain function <code>normalizedSimilarity(actual, expected)</code> returns the same 0–1 score for use with <code>t.score</code>.</p><p>A few more context members shape a case: <code>t.require(value, expectation)</code> records a gate and stops the test body when it fails, without a duplicate execution error. <code>t.skip(reason)</code> ends the case as skipped (reported separately, never changes the exit code; call it before sending messages). <code>t.metric(name, value)</code> records a structured score for the playground case card. <code>t.log(message)</code> records a debug line for the CLI and playground result.</p><p>Three <code>t.send</code> options apply on session create (first <code>t.send</code> only):</p><ul><li><code>workspaceFiles</code> — <code>{ path: contents }</code>, seeded into the local session workspace. Prefer this over machine-local paths.</li><li><code>workspaceDir</code> — absolute harness cwd (local runtime).</li><li><code>cloud</code> — per-session cloud options merged over the agent&#39;s static <code>cloud</code> config (repos / env / …). Use a pinned <code>repos</code> override to attach a fixture repo for cloud evals without putting it on the agent&#39;s default <code>cloud.repos</code>. Cloud ignores <code>workspaceFiles</code> seeds.</li></ul><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">const</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> toolResults</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> =</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> t.events.</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">filter</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">((</span><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">e</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">) </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">=&gt;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> e.type </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">===</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;action.result&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">);</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">});</span></span></code></pre></div><p>The timeout order is case or file <code>timeoutMs</code>, CLI <code>--timeout-ms</code>, project config <code>timeoutMs</code>, then the 180-second runner default.</p><p>The optional fields:</p><table tabindex="0"><thead><tr><th>Option</th><th>Default</th><th>Meaning</th></tr></thead><tbody><tr><td><code>timeoutMs</code></td><td><code>180_000</code></td><td>Project-wide per-case timeout</td></tr><tr><td><code>judge</code></td><td>unset</td><td>Default judge model for <code>t.judge.*</code>; see <a href="#judge-free-form-output">Judge free-form output</a></td></tr><tr><td><code>reporters</code></td><td>unset</td><td>Destinations that observe every case; <code>--skip-report</code> suppresses them</td></tr><tr><td><code>maxPlaygroundRuns</code></td><td><code>20</code></td><td>Max batches in the playground / <code>/v1/dev/evals*</code> history (not CLI <code>eval</code>)</td></tr></tbody></table><p>Reporters come from <code>@cursor/july/evals/reporters</code>: <code>JUnit</code> writes a JUnit XML file for CI, <code>Artifacts</code> writes per-case files, and <code>combineReporters</code> merges several into one (<code>renderJUnitXml</code> renders the XML for a custom destination). A file or case can add its own <code>reporters</code> on top of the config list.</p><p>Playground batches live in process memory and disappear when <code>serve</code> exits. Navigating away and back still works while the process is up. To keep batches across restarts, declare an <code>evals</code> table in <code>agent/storage.ts</code>; see <a href="./storage.html#eval-and-a-b-tables">Storage</a>.</p><h2 id="drive-and-assert-with-t" tabindex="-1">Drive and assert with <code>t</code> <a class="header-anchor" href="#drive-and-assert-with-t" aria-label="Permalink to &quot;Drive and assert with \`t\`&quot;">​</a></h2><p><code>t</code> is both the driver and the assertion surface. You write ordinary control flow, sending turns and asserting inline.</p><p>Drive the agent with <code>t.send(message, options?)</code>. It runs one turn and waits for the session to park or fail. Multiple sends in one case share the session, which is how you write multi-turn evals.</p><p>Each <code>t.send</code> resolves to a turn result with <code>message</code>, <code>sessionId</code>, <code>events</code>, <code>toolCalls</code>, <code>ok</code>, and <code>index</code>. The turn carries the same assertion vocabulary as <code>t</code>, scoped to that turn, so you can grade an intermediate turn before the next send overwrites <code>t.reply</code>. <code>turn.expectOk()</code> throws when the turn failed, for later steps that depend on it.</p><p>Read the full case state with <code>t.reply</code> (the last assistant text), <code>t.events</code> (every captured session event across turns), <code>t.turns</code> (settled turns, oldest first), and <code>t.sessionId</code>. <code>t.signal</code> aborts when the case hits its timeout; pass it to your own async work.</p><p>Assert with the gates:</p><table tabindex="0"><thead><tr><th>Gate</th><th>Checks</th></tr></thead><tbody><tr><td><code>t.succeeded()</code></td><td>the run did not fail and is not parked on an unanswered approval</td></tr><tr><td><code>t.parked()</code></td><td>the run cleanly parked on an unanswered approval request</td></tr><tr><td><code>t.messageIncludes(token)</code></td><td>the joined assistant text matches a string or <code>RegExp</code></td></tr><tr><td><code>t.calledTool(name, matcher?)</code></td><td>a matching call to <code>name</code> happened</td></tr><tr><td><code>t.notCalledTool(name)</code></td><td>no request for <code>name</code>, in any lifecycle state</td></tr><tr><td><code>t.loadedSkill(name)</code></td><td>the agent opened the skill&#39;s <code>SKILL.md</code> (read, grep, or shell <code>cat</code>)</td></tr><tr><td><code>t.toolOrder(names)</code></td><td>tool requests appear in this relative order (extra calls allowed)</td></tr><tr><td><code>t.usedNoTools()</code></td><td>no tool calls at all</td></tr><tr><td><code>t.maxToolCalls(max)</code></td><td>at most <code>max</code> tool calls</td></tr><tr><td><code>t.noFailedActions()</code></td><td>no tool call reported an error</td></tr><tr><td><code>t.calledSubagent(name, matcher?)</code></td><td>a matching subagent delegation happened</td></tr><tr><td><code>t.taggedArtifact(kind?, predicate?)</code></td><td>at least one <a href="./reference/artifacts.html">artifact</a> was tagged</td></tr><tr><td><code>t.event(type, matcher?)</code></td><td>at least one matching event of <code>type</code> occurred</td></tr><tr><td><code>t.notEvent(type, matcher?)</code></td><td>no matching event of <code>type</code> occurred</td></tr><tr><td><code>t.eventOrder(matchers)</code></td><td>matching event groups occur in this relative order</td></tr><tr><td><code>t.eventsSatisfy(label, predicate)</code></td><td>your predicate over the typed event stream</td></tr><tr><td><code>t.check(value, expectation)</code></td><td>any value, against a builder</td></tr><tr><td><code>t.score(name, value)</code></td><td>records a 0–1 score you computed; soft until you add a bar</td></tr><tr><td><code>t.requireToolCall(name, matcher?)</code></td><td>gates on a matching call and returns it, so later code can read its input and output</td></tr><tr><td><code>t.requireInputRequest(filter?)</code></td><td>gates on exactly one pending approval request and returns it</td></tr></tbody></table><p>Every gate returns a handle: <code>.soft()</code> demotes it to tracked-only, <code>.atLeast(0.7)</code> adds a soft score bar, and <code>.gate(0.8)</code> promotes a scored assertion into a hard gate.</p><p>With no matcher, <code>calledTool</code> is request-based: a requested call counts even when its result has not arrived. Pass <code>t.calledTool(&quot;inspect_pr&quot;, { status: &quot;completed&quot; })</code> to require the call to return. <code>input</code>, <code>output</code>, and <code>count</code> matcher fields accept a literal, a <code>RegExp</code>, or a predicate.</p><p>The expectation builders are <code>includes(string | RegExp)</code>, <code>equals(value)</code>, <code>matches(schema)</code>, <code>similarity(expected)</code>, and <code>satisfies(predicate, label)</code>. <code>includes</code> stringifies its input, <code>equals</code> compares values deeply, <code>matches</code> validates against a Standard Schema (or anything with <code>safeParse</code>, like Zod), <code>similarity</code> scores normalized text similarity, and <code>satisfies</code> runs your predicate. The plain function <code>normalizedSimilarity(actual, expected)</code> returns the same 0–1 score for use with <code>t.score</code>.</p><p>A few more context members shape a case: <code>t.require(value, expectation)</code> records a gate and stops the test body when it fails, without a duplicate execution error. <code>t.skip(reason)</code> ends the case as skipped (reported separately, never changes the exit code; call it before sending messages). <code>t.metric(name, value)</code> records a structured score for the playground case card. <code>t.log(message)</code> records a debug line for the CLI and playground result.</p><p>Three <code>t.send</code> options apply on session create (first <code>t.send</code> only):</p><ul><li><code>workspaceFiles</code> — <code>{ path: contents }</code>, seeded into the local session workspace. Prefer this over machine-local paths.</li><li><code>workspaceDir</code> — absolute harness cwd (local runtime).</li><li><code>cloud</code> — per-session cloud options merged over the agent&#39;s static <code>cloud</code> config (repos / env / …). Use a pinned <code>repos</code> override to attach a fixture repo for cloud evals without putting it on the agent&#39;s default <code>cloud.repos</code>. Cloud ignores <code>workspaceFiles</code> seeds.</li></ul><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">const</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> toolResults</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> =</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> t.events.</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">filter</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">((</span><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">e</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">) </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">=&gt;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> e.type </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">===</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;action.result&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">);</span></span>
 <span class="line"><span style="--shiki-light:#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>
@@ -76,7 +76,7 @@ import{_ as e,c as a,o as i,ag as t}from"./chunks/framework.CAZyNGu9.js";const k
 <span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">      &quot;durationMs&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">12340</span></span>
 <span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">    }</span></span>
 <span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  ]</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">}</span></span></code></pre></div><p>Each case result can also include <code>description</code>, <code>finalText</code>, <code>tools</code>, <code>error</code>, and tool arguments or output. This shape lets CI report the failed assertion without parsing terminal text.</p><h2 id="run-evals-in-the-playground" tabindex="-1">Run evals in the playground <a class="header-anchor" href="#run-evals-in-the-playground" aria-label="Permalink to &quot;Run evals in the playground&quot;">​</a></h2><p>Start the server with <code>--dev</code>, open the playground, and choose <strong>Evals</strong>. You can run every case or one case, watch progress, and open the resulting session trace.</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> serve</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dev</span></span></code></pre></div><p>Playground runs target the live server instead of an ephemeral one. Their sessions appear in the session list. One eval batch can run at a time. By default those batches are <strong>in-memory only</strong> (capped by <code>maxPlaygroundRuns</code>); declare an <code>evals</code> table in <code>agent/storage.ts</code> if you need them after a serve restart — see <a href="./storage.html#eval-and-ab-tables">Storage</a>.</p><p>The UI uses the playground eval routes (available without <code>--dev</code>): <code>GET /v1/dev/evals</code> lists datapoints and config (includes <code>maxPlaygroundRuns</code> / <code>durableRuns</code>), <code>GET /v1/dev/evals/runs</code> rehydrates recent batches after navigation, <code>POST /v1/dev/evals/runs</code> starts a batch (returns an <strong>Eval ID</strong> / <code>runId</code>), <code>GET /v1/dev/evals/runs/:runId</code> polls it, and <code>POST /v1/dev/evals/runs/:runId/cancel</code> cancels a running batch. See <a href="./reference/http-api.html#playground-eval-routes">Playground eval routes</a>. The start request returns <code>202</code> while cases run in the background. Poll until the snapshot status becomes <code>completed</code>, <code>failed</code>, or <code>cancelled</code>. Configuration errors appear on a failed snapshot.</p><p>On <code>--prod</code> / <code>--url</code>, the CLI prints the Eval ID as soon as the batch is accepted (and a Playground deep link with <code>?view=evals&amp;evalRunId=…</code>):</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> eval</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --prod</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --slug</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> vulnerability-scanner</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --tag</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> deepsec</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">}</span></span></code></pre></div><p>Each case result can also include <code>description</code>, <code>finalText</code>, <code>tools</code>, <code>error</code>, and tool arguments or output. This shape lets CI report the failed assertion without parsing terminal text.</p><h2 id="run-evals-in-the-playground" tabindex="-1">Run evals in the playground <a class="header-anchor" href="#run-evals-in-the-playground" aria-label="Permalink to &quot;Run evals in the playground&quot;">​</a></h2><p>Start the server with <code>--dev</code>, open the playground, and choose <strong>Evals</strong>. You can run every case or one case, watch progress, and open the resulting session trace.</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> serve</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dev</span></span></code></pre></div><p>Playground runs target the live server instead of an ephemeral one. Their sessions appear in the session list. One eval batch can run at a time. By default those batches are <strong>in-memory only</strong> (capped by <code>maxPlaygroundRuns</code>); declare an <code>evals</code> table in <code>agent/storage.ts</code> if you need them after a serve restart — see <a href="./storage.html#eval-and-a-b-tables">Storage</a>.</p><p>The UI uses the playground eval routes (available without <code>--dev</code>): <code>GET /v1/dev/evals</code> lists datapoints and config (includes <code>maxPlaygroundRuns</code> / <code>durableRuns</code>), <code>GET /v1/dev/evals/runs</code> rehydrates recent batches after navigation, <code>POST /v1/dev/evals/runs</code> starts a batch (returns an <strong>Eval ID</strong> / <code>runId</code>), <code>GET /v1/dev/evals/runs/:runId</code> polls it, and <code>POST /v1/dev/evals/runs/:runId/cancel</code> cancels a running batch. See <a href="./reference/http-api.html#playground-eval-routes">Playground eval routes</a>. The start request returns <code>202</code> while cases run in the background. Poll until the snapshot status becomes <code>completed</code>, <code>failed</code>, or <code>cancelled</code>. Configuration errors appear on a failed snapshot.</p><p>On <code>--prod</code> / <code>--url</code>, the CLI prints the Eval ID as soon as the batch is accepted (and a Playground deep link with <code>?view=evals&amp;evalRunId=…</code>):</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> eval</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --prod</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --slug</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> vulnerability-scanner</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --tag</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> deepsec</span></span>
 <span class="line"><span style="--shiki-light:#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>

package/dist/docs/assets/example-agents_index.md.xmJ87d_3.js

@@ -0,0 +1,2 @@
+import{_ as t,c as a,o as r,ag as o}from"./chunks/framework.CAZyNGu9.js";const m=JSON.parse('{"title":"Choose the right Agent SDK example","description":"Compare all twelve example agents by runtime, channels, tools, state, and the framework pattern each one teaches.","frontmatter":{"title":"Choose the right Agent SDK example","description":"Compare all twelve example agents by runtime, channels, tools, state, and the framework pattern each one teaches."},"headers":[],"relativePath":"example-agents/index.md","filePath":"example-agents/index.md"}'),s={name:"example-agents/index.md"};function n(i,e,d,l,h,c){return r(),a("div",null,[...e[0]||(e[0]=[o(`<h1 id="choose-the-right-agent-sdk-example" tabindex="-1">Choose the right Agent SDK example <a class="header-anchor" href="#choose-the-right-agent-sdk-example" aria-label="Permalink to &quot;Choose the right Agent SDK example&quot;">​</a></h1><p>The examples progress from one-channel assistants to durable, event-driven workflows. Start with the smallest agent for your use case. Each guide explains its request flow, framework features, verification path, and reusable design.</p><p>The source projects live under <a href="./../../examples/"><code>examples/</code></a>. Run the commands below from <code>packages/agent-serve</code>. See <a href="/docs/#run-the-cli">Run the CLI</a> if the <code>agent-sdk</code> command isn&#39;t installed.</p><h2 id="compare-the-examples" tabindex="-1">Compare the examples <a class="header-anchor" href="#compare-the-examples" aria-label="Permalink to &quot;Compare the examples&quot;">​</a></h2><table tabindex="0"><thead><tr><th>Agent</th><th>Runtime</th><th>Intake</th><th>Framework focus</th><th>What sets it apart</th></tr></thead><tbody><tr><td><a href="./weather-agent.html">Weather agent</a></td><td>Local</td><td>HTTP and two Slack transports</td><td>Tools, stdio MCP, approvals, skill, subagent, schedule, hook, A/B, and evals</td><td>It demonstrates the broad local-runtime surface in one domain.</td></tr><tr><td><a href="./slack-agent.html">Slack agent</a></td><td>Local</td><td>Account-linked Slack</td><td>Channel identity, threads, and suggested prompts</td><td>It reaches Slack without authored tools.</td></tr><tr><td><a href="./concierge.html">Concierge</a></td><td>Local</td><td>Built-in HTTP</td><td>Peer MCP and multi-agent serving</td><td>It delegates to a separate agent with its own tools, sessions, and context.</td></tr><tr><td><a href="./benny.html">Playbook router</a></td><td>Local with repo context</td><td>Two Slack transports</td><td>Channel watching, inherited skills, custom cwd, and an eval</td><td>An allowlisted Slack channel becomes an intake queue for repo playbooks.</td></tr><tr><td><a href="./oncall.html">Alert investigator</a></td><td>Local</td><td>Watched Slack alerts channel</td><td>Bot-post channel watching, per-thread debounce, reminder tools, and host Slack calls</td><td>Every alert gets a thread-pinned investigation that schedules its own re-checks.</td></tr><tr><td><a href="./bugbot.html">PR evidence reviewer</a></td><td>Local</td><td>Custom HTTP and Slack</td><td>Host tool, skill, seeded workspaces, and an eval</td><td>The model receives a prepared diff-first evidence tree instead of a checkout.</td></tr><tr><td><a href="./approval-buddy.html">Approval Buddy</a></td><td>Local</td><td>GitHub and Slack</td><td>Policy tools, two subagents, durable storage, and evals</td><td>Code decides whether a PR may be approved. Reviews stay informational.</td></tr><tr><td><a href="./security-reviewer.html">Security Reviewer</a></td><td>Local host pipeline</td><td>GitHub and chat</td><td>Staged tools, parallel SDK agents, progress UI, durable storage, A/B, and evals</td><td>Lives in <code>factory/security-reviewer/</code>. Reviewers and triage overlap while the playground shows every stage.</td></tr><tr><td><a href="./fsd.html">Remote PR coordinator</a></td><td>Local coordinator and remote PR sessions</td><td>HTTP, GitHub, and Slack</td><td>Remote handoff, hooks, affinity, buffering, reminders, and workflow MCP</td><td>One remote conversation follows a PR across chat, webhooks, and timed wakes.</td></tr><tr><td><a href="./knowledge-base.html">Knowledge base</a></td><td>Local</td><td>Built-in HTTP chat</td><td>Durable host-side state, a conventions skill, a schedule, unit tests, and evals</td><td>People curate shared facts in chat, and fresh sessions retrieve them from markdown.</td></tr><tr><td><a href="./codebase-wiki.html">Codebase wiki</a></td><td>Local</td><td>GitHub and chat</td><td>Task-dispatch webhooks, seeded digests, a mapping skill, a schedule, and evals</td><td>Merged PRs accumulate into per-feature wiki pages with a daily digest.</td></tr><tr><td><a href="./codeowners-review.html">Codeowners review</a></td><td>Local</td><td>GitHub, chat, and fixtures</td><td>Ownership routing in code, playbook data files, parallel subagents, and evals</td><td>Each product area reviews with its own playbook, and verdicts aggregate mechanically.</td></tr></tbody></table><h2 id="pick-a-learning-path" tabindex="-1">Pick a learning path <a class="header-anchor" href="#pick-a-learning-path" aria-label="Permalink to &quot;Pick a learning path&quot;">​</a></h2><p>Use this order when you want to learn the Agent SDK one capability at a time:</p><ol><li>Start with <a href="./weather-agent.html">Weather agent</a> to explore the filesystem conventions and local runtime.</li><li>Strip the project back to <a href="./slack-agent.html">Slack agent</a> to see the minimum channel surface.</li><li>Read <a href="./benny.html">Playbook router</a> when Slack should route requests into repo playbooks.</li><li>Continue to <a href="./oncall.html">Alert investigator</a> when the intake is bot posts and the agent must pace its own engagement and re-checks.</li><li>Add composition with <a href="./concierge.html">Concierge</a>.</li><li>Study <a href="./bugbot.html">PR evidence reviewer</a> before giving a model repository evidence.</li><li>Move policy into code with <a href="./approval-buddy.html">Approval Buddy</a>.</li><li>Compare <a href="./security-reviewer.html">Security Reviewer</a> and <a href="./fsd.html">Remote PR coordinator</a> for host-side versus remote PR work.</li><li>See parallel subagent delegation carry team judgment in <a href="./codeowners-review.html">Codeowners review</a>.</li><li>Curate team context through conversation with <a href="./knowledge-base.html">Knowledge base</a>, then let GitHub events maintain product documentation in <a href="./codebase-wiki.html">Codebase wiki</a>.</li></ol><h2 id="common-prerequisites" tabindex="-1">Common prerequisites <a class="header-anchor" href="#common-prerequisites" aria-label="Permalink to &quot;Common prerequisites&quot;">​</a></h2><p>All examples require:</p><ul><li>Node 22.13 or newer. Don&#39;t run the Agent SDK under Bun.</li><li>Workspace dependencies installed.</li><li>An agent-runtime credential for model turns.</li></ul><p>Several examples need more:</p><ul><li>Account-linked Slack channels require a connected host account.</li><li>Alert investigator needs a dedicated Socket Mode app with channel-post events and membership in the watched alerts channel.</li><li>GitHub examples require access to the target repository. Codebase wiki and Codeowners review call the host <code>gh</code> CLI for PR data; the codeowners fixtures run without network.</li><li>Example agents use <code>cursorHostedStorage</code> (<code>agent/storage.ts</code>) for Cursor-hosted session storage (control-plane proxy).</li><li>Remote PR coordinator starts remote agent sessions and needs access to its workflow backend.</li></ul><p>Each guide lists its own credentials, services, and side effects.</p><h2 id="validate-any-example" tabindex="-1">Validate any example <a class="header-anchor" href="#validate-any-example" aria-label="Permalink to &quot;Validate any example&quot;">​</a></h2><p>Discovery commands don&#39;t start a model turn:</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;"> validate</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> examples/weather-agent</span></span>
+<span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> info</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> examples/weather-agent</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --json</span></span></code></pre></div><p>Start one development server with <code>agent-sdk dev examples/&lt;name&gt;</code>. Concierge depends on Weather agent, so its guide creates an isolated two-project mount. Don&#39;t mount the whole examples directory to test one agent; several advanced examples subscribe to live GitHub events.</p><h2 id="read-by-framework-feature" tabindex="-1">Read by framework feature <a class="header-anchor" href="#read-by-framework-feature" aria-label="Permalink to &quot;Read by framework feature&quot;">​</a></h2><ul><li><a href="./../concepts.html">Concepts</a> explains filesystem discovery and runtime boundaries.</li><li><a href="./../reference/project-layout.html">Project layout</a> lists every authored folder.</li><li><a href="./../reference/tools.html">Tools</a>, <a href="./../reference/channels.html">channels</a>, and <a href="./../reference/connections.html">MCP connections</a> cover the core extension points.</li><li><a href="./../evals.html">Evals</a> and <a href="./../ab.html">live A/B metrics</a> cover measured iteration.</li><li><a href="./../deployment.html">Deployment</a> covers credentials, auth, storage, and hosting.</li></ul>`,20)])])}const u=t(s,[["render",n]]);export{m as __pageData,u as default};

package/dist/docs/assets/{guides_webhooks.md.BERuBSJW.js → guides_webhooks.md.DB-r_er9.js}

@@ -1,4 +1,4 @@
-import{_ as i,c as a,o as n,ag as h}from"./chunks/framework.CAZyNGu9.js";const o=JSON.parse('{"title":"Webhooks & custom channels","description":"Give the agent its own HTTP surface: validated routes, sessions keyed to your domain, and replies delivered back out.","frontmatter":{"title":"Webhooks & custom channels","description":"Give the agent its own HTTP surface: validated routes, sessions keyed to your domain, and replies delivered back out."},"headers":[],"relativePath":"guides/webhooks.md","filePath":"guides/webhooks.md"}'),t={name:"guides/webhooks.md"};function e(l,s,k,p,r,d){return n(),a("div",null,[...s[0]||(s[0]=[h(`<h1 id="webhooks-and-custom-channels" tabindex="-1">Webhooks and custom channels <a class="header-anchor" href="#webhooks-and-custom-channels" aria-label="Permalink to &quot;Webhooks and custom channels&quot;">​</a></h1><p>A custom channel gives the agent its own HTTP surface. You get routes with validated payloads, sessions keyed to something in your domain (a thread, a ticket, a PR), and replies delivered back to the caller. The <a href="./slack.html">Slack</a> and <a href="./github.html">GitHub</a> packs build on this mechanism. This page is the mechanism itself.</p><h2 id="what-you-already-have" tabindex="-1">What you already have <a class="header-anchor" href="#what-you-already-have" aria-label="Permalink to &quot;What you already have&quot;">​</a></h2><p>The built-in HTTP channel is always mounted (under <code>/&lt;slug&gt;</code> in the default multi-agent layout). <code>POST /v1/session</code> starts a conversation, <code>POST /v1/session/:id</code> follows up, and <code>GET /v1/session/:id/stream</code> streams NDJSON events, plus sessions, approvals, and tool routes. The full list is in the <a href="./../reference/http-api.html">HTTP API reference</a>.</p><p>Write a custom channel when that shape doesn&#39;t fit: a webhook with its own payload contract, a surface that keys sessions by a domain id, or a flow that does host-side work before (or instead of) a model turn.</p><h2 id="define-a-channel" tabindex="-1">Define a channel <a class="header-anchor" href="#define-a-channel" aria-label="Permalink to &quot;Define a channel&quot;">​</a></h2><p>Author <code>agent/channels/&lt;id&gt;.ts</code> with <code>defineChannel</code>. The filename is the channel id, and routes mount under <code>/v1/channels/&lt;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:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { defineChannel, POST } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;@cursor/july/channels&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
+import{_ as i,c as a,o as n,ag as h}from"./chunks/framework.CAZyNGu9.js";const o=JSON.parse('{"title":"Webhooks & custom channels","description":"Give the agent its own HTTP surface: validated routes, sessions keyed to your domain, and replies delivered back out.","frontmatter":{"title":"Webhooks & custom channels","description":"Give the agent its own HTTP surface: validated routes, sessions keyed to your domain, and replies delivered back out."},"headers":[],"relativePath":"guides/webhooks.md","filePath":"guides/webhooks.md"}'),t={name:"guides/webhooks.md"};function e(k,s,l,p,r,E){return n(),a("div",null,[...s[0]||(s[0]=[h(`<h1 id="webhooks-and-custom-channels" tabindex="-1">Webhooks and custom channels <a class="header-anchor" href="#webhooks-and-custom-channels" aria-label="Permalink to &quot;Webhooks and custom channels&quot;">​</a></h1><p>A custom channel gives the agent its own HTTP surface. You get routes with validated payloads, sessions keyed to something in your domain (a thread, a ticket, a PR), and replies delivered back to the caller. The <a href="./slack.html">Slack</a> and <a href="./github.html">GitHub</a> packs build on this mechanism. This page is the mechanism itself.</p><h2 id="what-you-already-have" tabindex="-1">What you already have <a class="header-anchor" href="#what-you-already-have" aria-label="Permalink to &quot;What you already have&quot;">​</a></h2><p>The built-in HTTP channel is always mounted (under <code>/&lt;slug&gt;</code> in the default multi-agent layout). <code>POST /v1/session</code> starts a conversation, <code>POST /v1/session/:id</code> follows up, and <code>GET /v1/session/:id/stream</code> streams NDJSON events, plus sessions, approvals, and tool routes. The full list is in the <a href="./../reference/http-api.html">HTTP API reference</a>.</p><p>Write a custom channel when that shape doesn&#39;t fit: a webhook with its own payload contract, a surface that keys sessions by a domain id, or a flow that does host-side work before (or instead of) a model turn.</p><h2 id="define-a-channel" tabindex="-1">Define a channel <a class="header-anchor" href="#define-a-channel" aria-label="Permalink to &quot;Define a channel&quot;">​</a></h2><p>Author <code>agent/channels/&lt;id&gt;.ts</code> with <code>defineChannel</code>. The filename is the channel id, and routes mount under <code>/v1/channels/&lt;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:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { defineChannel, POST } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;@cursor/july/channels&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
 <span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { z } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;zod&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;"> defineChannel</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">({</span></span>
@@ -47,7 +47,7 @@ import{_ as i,c as a,o as n,ag as h}from"./chunks/framework.CAZyNGu9.js";const o
 <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;"> defineChannel</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">({</span></span>
 <span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  auth: [</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">localDevStrict</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(), </span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">bearerAuth</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(process.env.</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">WEBHOOK_TOKEN</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> ??</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">)],</span></span>
 <span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  routes: [</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">/* … */</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">],</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">});</span></span></code></pre></div><p>The built-in policies are <code>localDevStrict()</code> (the default), <code>localDev()</code>, <code>loopbackOnly()</code>, <code>bearerAuth(tokenOrVerify)</code>, and <code>allowAll()</code>, and any <code>(request, info) =&gt; AuthContext | null</code> function composes with them. Signature-verified surfaces usually use <code>allowAll()</code> at the route and verify the HMAC in the channel; the GitHub channel with a webhook secret works this way. The details live in <a href="./../reference/channels.html#auth-policies">Channels</a>.</p><p>Server-level flags interact with channel auth. <code>--bearer-token &lt;secret&gt;</code> swaps the default for <code>bearerAuth</code> on every channel that doesn&#39;t author its own chain, and <code>--allow-anonymous</code> swaps it for <code>allowAll()</code> (trusted networks only). Authored <code>auth</code> arrays always win over both.</p><h2 id="example-linear-as-the-control-plane" tabindex="-1">Example: Linear as the control plane <a class="header-anchor" href="#example-linear-as-the-control-plane" aria-label="Permalink to &quot;Example: Linear as the control plane&quot;">​</a></h2><p>Everything above composes into a working ticket-driven agent. This example wires Linear to the agent: new issues and comments start or resume sessions, and replies land back on the issue as comments. The same shape works for any tracker with signed webhooks.</p><p>Create the webhook in Linear under Settings → API → &quot;New webhook&quot;, pointed at <code>https://&lt;your-host&gt;/&lt;slug&gt;/v1/channels/linear</code>, and copy the signing secret. Linear requires a public HTTPS URL, so use a tunnel during local development or test with signed fixtures (below). Set three environment variables:</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:#24292E;--shiki-dark:#E1E4E8;">LINEAR_WEBHOOK_SECRET</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">=</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">lin_wh_...</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">   # the webhook&#39;s signing secret</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">});</span></span></code></pre></div><p>The built-in policies are <code>localDevStrict()</code> (the default), <code>localDev()</code>, <code>loopbackOnly()</code>, <code>bearerAuth(tokenOrVerify)</code>, and <code>allowAll()</code>, and any <code>(request, info) =&gt; AuthContext | null</code> function composes with them. Signature-verified surfaces usually use <code>allowAll()</code> at the route and verify the HMAC in the channel; the GitHub channel with a webhook secret works this way. The details live in <a href="./../reference/channels.html#auth-policies">Channels</a>.</p><p>Server-level flags interact with channel auth. <code>--bearer-token &lt;secret&gt;</code> swaps the default for <code>bearerAuth</code> on every channel that doesn&#39;t author its own chain, and <code>--allow-anonymous</code> swaps it for <code>allowAll()</code> (trusted networks only). Authored <code>auth</code> arrays always win over both.</p><h2 id="hosted-aliases-need-an-alias-token-relay" tabindex="-1">Hosted aliases need an alias-token relay <a class="header-anchor" href="#hosted-aliases-need-an-alias-token-relay" aria-label="Permalink to &quot;Hosted aliases need an alias-token relay&quot;">​</a></h2><p>Cursor-managed deployments expose a stable alias URL. External callers must send <code>X-Agent-Alias-Token</code> on every request to that URL. Channel auth still runs after the alias gate.</p><p>Webhook providers that cannot attach custom headers cannot POST straight at a managed alias. Put a relay (or other authenticating intermediary) in front, use a Cursor-managed ingress path that does not use the public alias (for example Cursor Slack / GitHub event relay), or self-host. See <a href="./../deployment.html#use-the-hosted-agent">Deployment</a>.</p><h2 id="example-linear-as-the-control-plane" tabindex="-1">Example: Linear as the control plane <a class="header-anchor" href="#example-linear-as-the-control-plane" aria-label="Permalink to &quot;Example: Linear as the control plane&quot;">​</a></h2><p>Everything above composes into a working ticket-driven agent. This example wires Linear to the agent: new issues and comments start or resume sessions, and replies land back on the issue as comments. The same shape works for any tracker with signed webhooks.</p><p>Create the webhook in Linear under Settings → API → &quot;New webhook&quot;, pointed at <code>https://&lt;your-host&gt;/&lt;slug&gt;/v1/channels/linear</code>, and copy the signing secret. Linear requires a public HTTPS URL, so use a tunnel during local development or test with signed fixtures (below). Set three environment variables:</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:#24292E;--shiki-dark:#E1E4E8;">LINEAR_WEBHOOK_SECRET</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">=</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">lin_wh_...</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">   # the webhook&#39;s signing secret</span></span>
 <span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">LINEAR_API_KEY</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">=</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">lin_api_...</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">         # posts replies as comments</span></span>
 <span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">LINEAR_AGENT_USER_ID</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">=</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">...</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">           # the API key&#39;s user: query { viewer { id } }</span></span></code></pre></div><p><code>LINEAR_AGENT_USER_ID</code> matters: replies posted with the API key trigger the Comment webhook again, so the channel must recognize and skip its own comments. Without the guard, every reply starts another turn.</p><p>Author <code>agent/channels/linear.ts</code>:</p><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { Buffer } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;node:buffer&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
 <span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { createHmac, timingSafeEqual } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;node:crypto&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
@@ -175,9 +175,52 @@ import{_ as i,c as a,o as n,ag as h}from"./chunks/framework.CAZyNGu9.js";const o
 <span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { defineConnection } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;@cursor/july/connections&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
 <span class="line"></span>
 <span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">// Uses the signed-in Cursor account&#39;s Linear connector.</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;"> defineConnection</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">({ cursorAccount: </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">true</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">, servers: [</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;Linear&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">] });</span></span></code></pre></div><p>See <a href="./../reference/connections.html">MCP connections</a> for the direct <code>https://mcp.linear.app/mcp</code> form when the host isn&#39;t signed in to Cursor.</p><p>To test without a public URL, save a payload from the webhook&#39;s delivery log (or the sample in <a href="https://linear.app/developers/webhooks" target="_blank" rel="noreferrer">Linear&#39;s webhook docs</a>) under <code>fixtures/</code>, refresh its <code>webhookTimestamp</code>, and sign it yourself:</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:#24292E;--shiki-dark:#E1E4E8;">SIG</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">=</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">$(</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">node</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> -e</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &#39;const {createHmac}=require(&quot;node:crypto&quot;);const fs=require(&quot;node:fs&quot;);</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;"> defineConnection</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">({ cursorAccount: </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">true</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">, servers: [</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;Linear&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">] });</span></span></code></pre></div><p>See <a href="./../reference/connections.html">MCP connections</a> for the direct <code>https://mcp.linear.app/mcp</code> form when the host isn&#39;t signed in to Cursor.</p><h3 id="production-linear-agents-expiring-oauth-tokens" tabindex="-1">Production Linear agents: expiring OAuth tokens <a class="header-anchor" href="#production-linear-agents-expiring-oauth-tokens" aria-label="Permalink to &quot;Production Linear agents: expiring OAuth tokens&quot;">​</a></h3><p>A plain API key fits a personal integration. A production Linear agent is an OAuth application: installing it in a workspace mints an access token with an expiry plus a refresh token. The serve host can be replaced at any time, so the pair can&#39;t live in process memory. Persist it in <a href="./../storage.html#author-kv-ctx-host-kv"><code>host.kv</code></a> instead. Route handlers receive <code>host</code> in their args, and event handlers get the same services on <code>ctx</code>.</p><p>Two prerequisites:</p><ul><li>Keep the OAuth client id and client secret in <a href="./../deployment.html">deployment secrets</a>. They&#39;re static, so deploy-time env vars fit them.</li><li>Back kv with durable storage. On Cursor-managed hosting, use <code>cursorHostedStorage</code> in <code>agent/storage.ts</code> (<a href="./../storage.html">Storage</a>). Without it, kv falls back to local disk and a host replacement drops the tokens.</li></ul><p>Save the pair when the install flow completes, then refresh on demand:</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:#D73A49;--shiki-dark:#F97583;"> type</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { HostKvApi } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;@cursor/july/kv&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;">interface</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> LinearTokens</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> {</span></span>
+<span class="line"><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">  accessToken</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">:</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> string</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
+<span class="line"><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">  refreshToken</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">:</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> string</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
+<span class="line"><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">  expiresAtMs</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">:</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> number</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>
+<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">const</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> TOKENS_KEY</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> =</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;linear/oauth-tokens&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;"> async</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> function</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> linearAccessToken</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">kv</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">:</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> HostKvApi</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">)</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">:</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> Promise</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">&lt;</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">string</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">&gt; {</span></span>
+<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">  const</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> tokens</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> =</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> (</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">await</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> kv.</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">get</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">TOKENS_KEY</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">)) </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">as</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> LinearTokens</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> |</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> undefined</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
+<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">  if</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> (tokens </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">===</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> undefined</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">) {</span></span>
+<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">    throw</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> new</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> Error</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;No Linear tokens stored. Complete the install flow first.&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:#D73A49;--shiki-dark:#F97583;">  if</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> (Date.</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">now</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">() </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">&lt;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> tokens.expiresAtMs </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">-</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> 60_000</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">) {</span></span>
+<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">    return</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> tokens.accessToken;</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  }</span></span>
+<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">  const</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> response</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> =</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> await</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> fetch</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;https://api.linear.app/oauth/token&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">, {</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">    method: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;POST&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">    headers: { </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;content-type&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;application/x-www-form-urlencoded&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> },</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">    body: </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">new</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> URLSearchParams</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">({</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">      grant_type: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;refresh_token&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">      refresh_token: tokens.refreshToken,</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">      client_id: process.env.</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">LINEAR_CLIENT_ID</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> ??</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">      client_secret: process.env.</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">LINEAR_CLIENT_SECRET</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> ??</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;&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:#D73A49;--shiki-dark:#F97583;">  if</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> (</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">!</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">response.ok) {</span></span>
+<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">    throw</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> new</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> Error</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">\`Linear token refresh failed: \${</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">response</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">.</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">status</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:#D73A49;--shiki-dark:#F97583;">  const</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> next</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> =</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> (</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">await</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> response.</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">json</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">()) </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">as</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> {</span></span>
+<span class="line"><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">    access_token</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">:</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> string</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
+<span class="line"><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">    refresh_token</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">?:</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> string</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
+<span class="line"><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">    expires_in</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">:</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> number</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:#D73A49;--shiki-dark:#F97583;">  const</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> rotated</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">:</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> LinearTokens</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> =</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> {</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">    accessToken: next.access_token,</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">    refreshToken: next.refresh_token </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">??</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> tokens.refreshToken,</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">    expiresAtMs: Date.</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">now</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">() </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">+</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> next.expires_in </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">*</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> 1000</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:#D73A49;--shiki-dark:#F97583;">  await</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> kv.</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">put</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">TOKENS_KEY</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">, rotated);</span></span>
+<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">  return</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> rotated.accessToken;</span></span>
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">}</span></span></code></pre></div><p>Then authorize GraphQL calls with <code>Bearer \${await linearAccessToken(host.kv)}</code> in place of the static <code>apiKey</code> above. Writes to <code>host.kv</code> await the durable sink and propagate errors, so a failed save surfaces instead of silently losing the rotated refresh token.</p><p>To test without a public URL, save a payload from the webhook&#39;s delivery log (or the sample in <a href="https://linear.app/developers/webhooks" target="_blank" rel="noreferrer">Linear&#39;s webhook docs</a>) under <code>fixtures/</code>, refresh its <code>webhookTimestamp</code>, and sign it yourself:</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:#24292E;--shiki-dark:#E1E4E8;">SIG</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">=</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">$(</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">node</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> -e</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &#39;const {createHmac}=require(&quot;node:crypto&quot;);const fs=require(&quot;node:fs&quot;);</span></span>
 <span class="line"><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">process.stdout.write(createHmac(&quot;sha256&quot;,process.env.LINEAR_WEBHOOK_SECRET)</span></span>
 <span class="line"><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">  .update(fs.readFileSync(&quot;fixtures/issue-create.json&quot;)).digest(&quot;hex&quot;))&#39;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">)</span></span>
 <span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">curl</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> -X</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> POST</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> http://127.0.0.1:3000/</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">&lt;</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">slu</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">g</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">&gt;</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">/v1/channels/linear/</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> \\</span></span>
 <span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  -H</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &#39;content-type: application/json&#39;</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> -H</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;linear-signature: </span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">$SIG</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> \\</span></span>
-<span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  --data-binary</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> @fixtures/issue-create.json</span></span></code></pre></div><h2 id="test-a-channel" tabindex="-1">Test a channel <a class="header-anchor" href="#test-a-channel" aria-label="Permalink to &quot;Test a channel&quot;">​</a></h2><p>Start with curl and saved payloads under <code>fixtures/</code>. The playground&#39;s <strong>Try</strong> modal covers manual probes. It remembers your last body per endpoint, has Copy curl, and opens the created session on a successful Try. For regression coverage, drive the same behavior through an eval, or keep channel logic deterministic in <code>agent/lib/</code> and unit-test it there. When something looks wrong, read the session&#39;s <code>events.ndjson</code>. The stream is the record of what happened.</p><p>For GitHub specifically, don&#39;t hand-roll fixtures. <code>agent-sdk github replay</code> synthesizes real-shaped, signed payloads from any PR you can read. See the <a href="./github.html">GitHub guide</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/channels.html">Channels reference</a>: the full authoring API</li><li><a href="./github.html">GitHub</a> and <a href="./slack.html">Slack</a>: the packaged channels</li><li><a href="./../reference/sessions.html">Sessions and streaming</a>: events your channel can subscribe to</li></ul>`,49)])])}const g=i(t,[["render",e]]);export{o as __pageData,g as default};
+<span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  --data-binary</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> @fixtures/issue-create.json</span></span></code></pre></div><h2 id="test-a-channel" tabindex="-1">Test a channel <a class="header-anchor" href="#test-a-channel" aria-label="Permalink to &quot;Test a channel&quot;">​</a></h2><p>Start with curl and saved payloads under <code>fixtures/</code>. The playground&#39;s <strong>Try</strong> modal covers manual probes. It remembers your last body per endpoint, has Copy curl, and opens the created session on a successful Try. For regression coverage, drive the same behavior through an eval, or keep channel logic deterministic in <code>agent/lib/</code> and unit-test it there. When something looks wrong, read the session&#39;s <code>events.ndjson</code>. The stream is the record of what happened.</p><p>For GitHub specifically, don&#39;t hand-roll fixtures. <code>agent-sdk github replay</code> synthesizes real-shaped, signed payloads from any PR you can read. See the <a href="./github.html">GitHub guide</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/channels.html">Channels reference</a>: the full authoring API</li><li><a href="./github.html">GitHub</a> and <a href="./slack.html">Slack</a>: the packaged channels</li><li><a href="./../reference/sessions.html">Sessions and streaming</a>: events your channel can subscribe to</li></ul>`,59)])])}const g=i(t,[["render",e]]);export{o as __pageData,g as default};

package/dist/docs/assets/{guides_webhooks.md.BERuBSJW.lean.js → guides_webhooks.md.DB-r_er9.lean.js}

@@ -1 +1 @@
-import{_ as i,c as a,o as n,ag as h}from"./chunks/framework.CAZyNGu9.js";const o=JSON.parse('{"title":"Webhooks & custom channels","description":"Give the agent its own HTTP surface: validated routes, sessions keyed to your domain, and replies delivered back out.","frontmatter":{"title":"Webhooks & custom channels","description":"Give the agent its own HTTP surface: validated routes, sessions keyed to your domain, and replies delivered back out."},"headers":[],"relativePath":"guides/webhooks.md","filePath":"guides/webhooks.md"}'),t={name:"guides/webhooks.md"};function e(l,s,k,p,r,d){return n(),a("div",null,[...s[0]||(s[0]=[h("",49)])])}const g=i(t,[["render",e]]);export{o as __pageData,g as default};
+import{_ as i,c as a,o as n,ag as h}from"./chunks/framework.CAZyNGu9.js";const o=JSON.parse('{"title":"Webhooks & custom channels","description":"Give the agent its own HTTP surface: validated routes, sessions keyed to your domain, and replies delivered back out.","frontmatter":{"title":"Webhooks & custom channels","description":"Give the agent its own HTTP surface: validated routes, sessions keyed to your domain, and replies delivered back out."},"headers":[],"relativePath":"guides/webhooks.md","filePath":"guides/webhooks.md"}'),t={name:"guides/webhooks.md"};function e(k,s,l,p,r,E){return n(),a("div",null,[...s[0]||(s[0]=[h("",59)])])}const g=i(t,[["render",e]]);export{o as __pageData,g as default};

package/dist/docs/assets/index.md.BKpbkCa6.js

@@ -0,0 +1,24 @@
+import{_ as a,c as t,o as s,ag as n}from"./chunks/framework.CAZyNGu9.js";const g=JSON.parse('{"title":"Agent SDK documentation","description":"Build your own software factory with Cursor agents as ordinary files: tools, approvals, channels, and evals.","frontmatter":{"title":"Agent SDK documentation","description":"Build your own software factory with Cursor agents as ordinary files: tools, approvals, channels, and evals."},"headers":[],"relativePath":"index.md","filePath":"README.md"}'),i={name:"index.md"};function o(l,e,r,d,h,c){return s(),t("div",null,[...e[0]||(e[0]=[n(`<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>The Agent SDK helps you build your own software factory: agents that inspect builds, review pull requests, gate promotions, and wake from Slack or GitHub when work arrives. You author each agent as ordinary files in a TypeScript project under <code>agent/</code>: markdown for agent instruction prompts, TypeScript for typed behavior. The framework discovers those files, and serves the agent over channels. The Cursor SDK and the Cursor harness run the turns.</p><p>You write the tools, instructions, channels, and evals. In return you get a factory you can version, test, and ship: side effects stay behind human approvals, and every change stays regression-checked.</p><div class="language-text vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">text</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span>my-agent/</span></span>
+<span class="line"><span>├── package.json</span></span>
+<span class="line"><span>├── agent/</span></span>
+<span class="line"><span>│   ├── agent.ts             # runtime config: model, local/cloud runtime</span></span>
+<span class="line"><span>│   ├── instructions.md      # the always-on system prompt</span></span>
+<span class="line"><span>│   ├── tools/               # one typed tool per file</span></span>
+<span class="line"><span>│   ├── skills/              # on-demand procedures (SKILL.md convention)</span></span>
+<span class="line"><span>│   ├── mcp-connections/     # tools from external MCP servers</span></span>
+<span class="line"><span>│   ├── subagents/           # specialist child agents</span></span>
+<span class="line"><span>│   ├── channels/            # HTTP / Slack / GitHub surfaces</span></span>
+<span class="line"><span>│   ├── hooks/               # observe the runtime event stream</span></span>
+<span class="line"><span>│   ├── ab.ts                # optional live A/B experiment</span></span>
+<span class="line"><span>│   ├── ab/                  # optional: more experiments</span></span>
+<span class="line"><span>│   ├── schedules/           # cron-driven runs</span></span>
+<span class="line"><span>│   ├── sandbox/workspace/   # files seeded into each session workspace</span></span>
+<span class="line"><span>│   └── lib/                 # shared code (import-only, never discovered)</span></span>
+<span class="line"><span>└── evals/                   # filesystem evals (regression checks)</span></span></code></pre></div><p>Bootstrap a project with nothing installed beyond Node:</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><code>init</code> scaffolds the project, runs <code>npm install</code>, links <code>agent-sdk</code> onto your npm global bin (so bare <code>agent-sdk</code> works on PATH), and offers a Cursor sign-in. If the global link cannot update PATH, next steps include an <code>export PATH=&quot;$PWD/node_modules/.bin:$PATH&quot;</code> fallback.</p><p>Browse the docs locally without serving an agent:</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">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><p>Every running serve host also mounts this documentation at <code>/docs</code> (disable it with <code>--no-docs</code>).</p><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><p>Pick your entry point based on your goal.</p><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 approver), 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>Learning from working agents</td><td><a href="./example-agents/">Example agents</a></td></tr><tr><td>Wiring an agent to Slack</td><td><a href="./guides/slack.html">Slack guide</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 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="the-documentation" tabindex="-1">The documentation <a class="header-anchor" href="#the-documentation" aria-label="Permalink to &quot;The documentation&quot;">​</a></h2><p><strong>Core</strong></p><ul><li><a href="./quickstart.html">Quickstart</a>: build a PR approver that reviews by complexity and wakes from webhooks.</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="./concepts.html">Concepts</a>: the mental model behind the framework.</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>: make an agent better one measured round at a time.</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>: wake 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 <code>--store</code> them on hosted deployments.</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-runtime.html">Cloud runtime</a>: run turns on Cursor cloud agents instead of the local harness.</li></ul><p><strong>Example agents</strong></p><ul><li><a href="./example-agents/">Choose the right example</a>: compare all twelve agents by runtime, channels, tools, state, and architecture.</li><li><a href="./example-agents/weather-agent.html">Weather agent</a>: explore tools, MCP, approvals, skills, subagents, schedules, hooks, A/B metrics, and evals.</li><li><a href="./example-agents/slack-agent.html">Slack agent</a>: put a minimal agent in Slack through an account-linked transport.</li><li><a href="./example-agents/concierge.html">Concierge</a>: delegate work to a peer agent with its own context and sessions.</li><li><a href="./example-agents/benny.html">Playbook router</a>: route Slack intake through inherited repository playbooks.</li><li><a href="./example-agents/oncall.html">Alert investigator</a>: watch a Slack alerts channel and pin a self-rechecking investigation to every alert thread.</li><li><a href="./example-agents/bugbot.html">PR evidence reviewer</a>: review a host-prepared, diff-first pull-request evidence tree.</li><li><a href="./example-agents/approval-buddy.html">Approval Buddy</a>: keep approval policy in code while subagents supply review findings.</li><li><a href="./example-agents/security-reviewer.html">Security Reviewer</a>: run a staged, parallel security pipeline with live playground progress.</li><li><a href="./example-agents/fsd.html">Remote PR coordinator</a>: hand PR triage from local chat and webhooks to durable remote sessions.</li><li><a href="./example-agents/knowledge-base.html">Knowledge base</a>: turn conversations about people, systems, decisions, and preferences into shared markdown.</li><li><a href="./example-agents/codebase-wiki.html">Codebase wiki</a>: ingest merged PRs into per-feature pages with a daily digest schedule.</li><li><a href="./example-agents/codeowners-review.html">Codeowners review</a>: route PR reviews by ownership to per-area playbooks and aggregate verdicts.</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>: symptom to cause, in plain language.</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></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/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>The docs write commands as <code>agent-sdk &lt;command&gt;</code>. Where that command comes from depends on where you run. Starting fresh? This works with no prior install:</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><div class="note custom-block github-alert"><p class="custom-block-title">NOTE</p><p>When running from a source checkout there is no installed bin. Alias it from the package directory:</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:#005CC5;--shiki-dark:#79B8FF;">cd</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> packages/agent-serve</span></span>
+<span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">alias</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> agent-sdk</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">=</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;pnpm exec tsx </span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">$PWD</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">/src/bin/agent-serve.ts&quot;</span></span></code></pre></div><p>When <code>@cursor/july</code> is installed as a dependency, the <code>agent-sdk</code> bin lands in <code>node_modules/.bin</code>. After <code>init</code>, the CLI is also linked onto your npm global bin so bare <code>agent-sdk</code> works. You can still use <code>npx agent-sdk</code>, <code>pnpm exec agent-sdk</code>, or the scaffold&#39;s <code>npm run dev</code>. <code>npx @cursor/july docs</code> runs the <code>july</code> bin with that command (no local install required).</p></div><div class="note custom-block github-alert"><p class="custom-block-title">NOTE</p><p>The framework is being renamed from agent-serve to the Agent SDK. The <code>agent-sdk</code> bin already ships (alongside <code>july</code>, <code>agentkit</code>, and the legacy <code>agent-serve</code> alias), and projects already import from <code>@cursor/july</code>. A few on-disk names keep their old form until the code rename ships:</p><table tabindex="0"><thead><tr><th>Future name</th><th>Current name</th></tr></thead><tbody><tr><td><code>.agent-sdk/</code> state directory</td><td><code>.agent-serve/</code></td></tr><tr><td><code>/var/lib/agent-sdk</code> (deploy state root)</td><td><code>/var/lib/agent-serve</code></td></tr><tr><td><code>CURSOR_AGENT_SDK_*</code> env vars</td><td><code>AGENT_SERVE_*</code></td></tr><tr><td>Package path <code>packages/agent-sdk</code></td><td><code>packages/agent-serve</code></td></tr></tbody></table></div><div class="warning custom-block github-alert"><p class="custom-block-title">WARNING</p><p>Run the Agent SDK with Node 22.13 or newer, and never with Bun. Bun&#39;s HTTP/2 client corrupts the Cursor SDK&#39;s tool-result streams (<code>NGHTTP2_FRAME_SIZE_ERROR</code>), so every built-in read or grep the model makes fails and turns degrade into minutes-long retry loops.</p></div><h2 id="credentials" tabindex="-1">Credentials <a class="header-anchor" href="#credentials" aria-label="Permalink to &quot;Credentials&quot;">​</a></h2><p>Model turns run on the Cursor harness, so the serving host needs a Cursor credential. Sign in once, or export an API key:</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 style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">          # browser sign-in; mints + stores a revocable API key</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 style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">         # which account powers this host, and why</span></span></code></pre></div><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><p>These docs describe behavior. The package <a href="./../README.html"><code>README.md</code></a> is the compact reference, and <a href="./../AGENTS.html"><code>AGENTS.md</code></a> is the coding-agent cheat sheet. Task-shaped guides that ship with the package live under <a href="./../skills/"><code>skills/</code></a>; point a coding agent working on an Agent SDK project at them first. When the docs and the code disagree, the code is authoritative. Fix the doc.</p>`,37)])])}const u=a(i,[["render",o]]);export{g as __pageData,u as default};

package/dist/docs/assets/{index.md.CmhptOmN.lean.js → index.md.BKpbkCa6.lean.js}

@@ -1 +1 @@
-import{_ as a,c as t,o as s,ag as n}from"./chunks/framework.CAZyNGu9.js";const g=JSON.parse('{"title":"Agent SDK documentation","description":"Build your own software factory with Cursor agents as ordinary files: tools, approvals, channels, and evals.","frontmatter":{"title":"Agent SDK documentation","description":"Build your own software factory with Cursor agents as ordinary files: tools, approvals, channels, and evals."},"headers":[],"relativePath":"index.md","filePath":"README.md"}'),i={name:"index.md"};function o(r,e,l,h,d,p){return s(),t("div",null,[...e[0]||(e[0]=[n("",37)])])}const u=a(i,[["render",o]]);export{g as __pageData,u as default};
+import{_ as a,c as t,o as s,ag as n}from"./chunks/framework.CAZyNGu9.js";const g=JSON.parse('{"title":"Agent SDK documentation","description":"Build your own software factory with Cursor agents as ordinary files: tools, approvals, channels, and evals.","frontmatter":{"title":"Agent SDK documentation","description":"Build your own software factory with Cursor agents as ordinary files: tools, approvals, channels, and evals."},"headers":[],"relativePath":"index.md","filePath":"README.md"}'),i={name:"index.md"};function o(l,e,r,d,h,c){return s(),t("div",null,[...e[0]||(e[0]=[n("",37)])])}const u=a(i,[["render",o]]);export{g as __pageData,u as default};

package/dist/docs/assets/{quickstart.md.C_b6ESpD.js → quickstart.md.rIbfahJM.js}

@@ -1,6 +1,6 @@
-import{_ as i,c as a,o as n,ag as t}from"./chunks/framework.CAZyNGu9.js";const o=JSON.parse('{"title":"Build your first PR approver","description":"Create an agent that reviews pull requests by complexity, approves the safe ones, and wakes from GitHub webhooks.","frontmatter":{"title":"Build your first PR approver","description":"Create an agent that reviews pull requests by complexity, approves the safe ones, and wakes from GitHub webhooks."},"headers":[],"relativePath":"quickstart.md","filePath":"quickstart.md"}'),e={name:"quickstart.md"};function h(l,s,p,k,r,d){return n(),a("div",null,[...s[0]||(s[0]=[t(`<h1 id="build-your-first-pr-approver" tabindex="-1">Build your first PR approver <a class="header-anchor" href="#build-your-first-pr-approver" aria-label="Permalink to &quot;Build your first PR approver&quot;">​</a></h1><p>Build an agent that reviews GitHub pull requests. It fetches the diff, rates the change&#39;s complexity in plain TypeScript, approves the safe ones, and flags the rest for a human. Then wire it to GitHub webhooks and watch a pull request wake it.</p><p>The split is the point of the exercise: deterministic policy lives in typed tools, judgment lives in the model, and every decision is inspectable in the playground.</p><h2 id="prerequisites" tabindex="-1">Prerequisites <a class="header-anchor" href="#prerequisites" aria-label="Permalink to &quot;Prerequisites&quot;">​</a></h2><p>You need:</p><ul><li>Node 22.13 or newer. Bun isn&#39;t supported.</li><li>The <code>agent-sdk</code> CLI. <code>npx @cursor/july init ./pr-approver</code> bootstraps it with no prior install: <code>init</code> scaffolds the project, runs <code>npm install</code>, and offers a Cursor sign-in, after which <code>npx agent-sdk</code> resolves from the project&#39;s own dependencies. See <a href="./README.html#run-the-cli">Run the CLI</a> for other setups, such as a monorepo source checkout.</li><li>A Cursor credential for model turns. Sign in once:</li></ul><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></code></pre></div><p>You can also set <code>CURSOR_API_KEY</code> instead of signing in.</p><ul><li>A GitHub credential. <code>gh auth login</code> is enough, or set <code>GITHUB_TOKEN</code>. The tools you write resolve either one automatically. Reading pull requests works on any public repo; posting reviews needs write access to the repo you review.</li></ul><h2 id="first-run-in-10-minutes" tabindex="-1">First run in 10 minutes <a class="header-anchor" href="#first-run-in-10-minutes" aria-label="Permalink to &quot;First run in 10 minutes&quot;">​</a></h2><p>Want a working agent before the full tutorial? Four commands get you there:</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;"> init</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> ./pr-approver</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">     # scaffold + npm install + sign-in offer</span></span>
+import{_ as i,c as a,o as n,ag as t}from"./chunks/framework.CAZyNGu9.js";const o=JSON.parse('{"title":"Build your first PR approver","description":"Create an agent that reviews pull requests by complexity, approves the safe ones, and wakes from GitHub webhooks.","frontmatter":{"title":"Build your first PR approver","description":"Create an agent that reviews pull requests by complexity, approves the safe ones, and wakes from GitHub webhooks."},"headers":[],"relativePath":"quickstart.md","filePath":"quickstart.md"}'),e={name:"quickstart.md"};function h(l,s,p,k,r,d){return n(),a("div",null,[...s[0]||(s[0]=[t(`<h1 id="build-your-first-pr-approver" tabindex="-1">Build your first PR approver <a class="header-anchor" href="#build-your-first-pr-approver" aria-label="Permalink to &quot;Build your first PR approver&quot;">​</a></h1><p>Build an agent that reviews GitHub pull requests. It fetches the diff, rates the change&#39;s complexity in plain TypeScript, approves the safe ones, and flags the rest for a human. Then wire it to GitHub webhooks and watch a pull request wake it.</p><p>The split is the point of the exercise: deterministic policy lives in typed tools, judgment lives in the model, and every decision is inspectable in the playground.</p><h2 id="prerequisites" tabindex="-1">Prerequisites <a class="header-anchor" href="#prerequisites" aria-label="Permalink to &quot;Prerequisites&quot;">​</a></h2><p>You need:</p><ul><li>Node 22.13 or newer. Bun isn&#39;t supported.</li><li>The <code>agent-sdk</code> CLI. <code>npx @cursor/july init ./pr-approver</code> bootstraps it with no prior install: <code>init</code> scaffolds the project, runs <code>npm install</code>, links <code>agent-sdk</code> onto your PATH, and offers a Cursor sign-in. See <a href="/docs/#run-the-cli">Run the CLI</a> for other setups, such as a monorepo source checkout.</li><li>A Cursor credential for model turns. Sign in once:</li></ul><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></code></pre></div><p>You can also set <code>CURSOR_API_KEY</code> instead of signing in.</p><ul><li>A GitHub credential. <code>gh auth login</code> is enough, or set <code>GITHUB_TOKEN</code>. The tools you write resolve either one automatically. Reading pull requests works on any public repo; posting reviews needs write access to the repo you review.</li></ul><h2 id="first-run-in-10-minutes" tabindex="-1">First run in 10 minutes <a class="header-anchor" href="#first-run-in-10-minutes" aria-label="Permalink to &quot;First run in 10 minutes&quot;">​</a></h2><p>Want a working agent before the full tutorial? Four commands get you there:</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;"> ./pr-approver</span><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">     # scaffold + npm install + PATH link + sign-in offer</span></span>
 <span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">cd</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> pr-approver</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 style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">                    # serve, and print the playground URL</span></span></code></pre></div><p>Open the playground URL and chat with the scaffold. Then, in a second terminal (<code>dev</code> keeps running), run one turn from the command line:</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;"> run</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;"> --message</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;Introduce yourself in one sentence.&quot;</span></span></code></pre></div><p>That&#39;s the whole loop: files become an agent, <code>dev</code> serves it, and <code>run</code> exercises it. The rest of this page turns that scaffold into a real PR approver.</p><h2 id="scaffolding-agents" tabindex="-1">Scaffolding Agents <a class="header-anchor" href="#scaffolding-agents" aria-label="Permalink to &quot;Scaffolding Agents&quot;">​</a></h2><p>Have Cursor read <a href="./../skills/create-agent/SKILL.html"><code>skills/create-agent/SKILL.md</code></a> and describe what you want:</p><blockquote><p>Build me a PR approver for the playground. Start with one tool that inspects a pull request and guide me through the remaining decisions.</p></blockquote><p>Cursor asks for missing choices, shows you the plan, then builds and verifies the agent. Continue below to do the same by hand.</p><p>See <a href="./scaffolding-agents.html">Scaffold an agent with Cursor</a> for the full guided workflow.</p><h2 id="create-your-project" tabindex="-1">Create your project <a class="header-anchor" href="#create-your-project" aria-label="Permalink to &quot;Create your project&quot;">​</a></h2><p>Start with the built-in scaffold:</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;"> init</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> ./pr-approver</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 style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">                           # serve, and print the playground URL</span></span></code></pre></div><p>Open the playground URL and chat with the scaffold. Then, in a second terminal (<code>dev</code> keeps running), run one turn from the command line:</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;"> run</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;"> --message</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;Introduce yourself in one sentence.&quot;</span></span></code></pre></div><p>That&#39;s the whole loop: files become an agent, <code>dev</code> serves it, and <code>run</code> exercises it. The rest of this page turns that scaffold into a real PR approver.</p><h2 id="scaffolding-agents" tabindex="-1">Scaffolding Agents <a class="header-anchor" href="#scaffolding-agents" aria-label="Permalink to &quot;Scaffolding Agents&quot;">​</a></h2><p>Have Cursor read <a href="./../skills/create-agent/SKILL.html"><code>skills/create-agent/SKILL.md</code></a> and describe what you want:</p><blockquote><p>Build me a PR approver for the playground. Start with one tool that inspects a pull request and guide me through the remaining decisions.</p></blockquote><p>Cursor asks for missing choices, shows you the plan, then builds and verifies the agent. Continue below to do the same by hand.</p><p>See <a href="./scaffolding-agents.html">Scaffold an agent with Cursor</a> for the full guided workflow.</p><h2 id="create-your-project" tabindex="-1">Create your project <a class="header-anchor" href="#create-your-project" aria-label="Permalink to &quot;Create your project&quot;">​</a></h2><p>Start with the built-in scaffold:</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;"> ./pr-approver</span></span>
 <span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">cd</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> pr-approver</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>The scaffold creates the files the Agent SDK discovers, plus empty capability folders (each with a <code>.gitkeep</code>) so you can drop tools, channels, and evals in place:</p><div class="language-text vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">text</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span>pr-approver/</span></span>
 <span class="line"><span>├── agent/</span></span>
@@ -128,7 +128,7 @@ import{_ as i,c as a,o as n,ag as t}from"./chunks/framework.CAZyNGu9.js";const o
 <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>The file adds one tool named <code>inspect_pr</code>:</p><ul><li><code>description</code> tells the model when to call it.</li><li><code>inputSchema</code> defines and validates the arguments.</li><li><code>execute</code> runs on the server and returns data to the model.</li></ul><p>Two details carry the design. <code>rateComplexity</code> is the review policy, and it lives in code: the model never decides what counts as a big change. And <code>ctx.host.github</code> is the shared host GitHub client, so the tool inherits whatever credential the host has (a token, <code>gh auth</code>, or a GitHub App) without parsing any of it.</p><h2 id="try-the-inspect-tool" tabindex="-1">Try the inspect tool <a class="header-anchor" href="#try-the-inspect-tool" aria-label="Permalink to &quot;Try the inspect tool&quot;">​</a></h2><p>Call the tool directly first, on a real merged pull request:</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;"> call</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> inspect_pr</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;">  --input</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &#39;{&quot;prUrl&quot;:&quot;https://github.com/react/react/pull/35623&quot;}&#39;</span></span></code></pre></div><p><code>call</code> validates the input and runs <code>execute</code> without a model turn. This PR is a one-character typo fix, so the result comes back rated <code>trivial</code> with the whole patch inline:</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;">  --input</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &#39;{&quot;prUrl&quot;:&quot;https://github.com/facebook/react/pull/35623&quot;}&#39;</span></span></code></pre></div><p><code>call</code> validates the input and runs <code>execute</code> without a model turn. This PR is a one-character typo fix, so the result comes back rated <code>trivial</code> with the whole patch inline:</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;title&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;Fix typo: accomodate -&gt; accommodate&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
 <span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  &quot;additions&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;deletions&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>
@@ -136,7 +136,7 @@ import{_ as i,c as a,o as n,ag as t}from"./chunks/framework.CAZyNGu9.js";const o
 <span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  &quot;complexity&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;trivial&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
 <span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  &quot;files&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: [{ </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">&quot;path&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;compiler/packages/...&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">, </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">&quot;patch&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;@@ -1315,7 ...&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>Now call it on the PR that added <code>experimental_useEvent</code> to React: 1,027 additions across 26 files.</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;"> call</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> inspect_pr</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;">  --input</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &#39;{&quot;prUrl&quot;:&quot;https://github.com/react/react/pull/25229&quot;}&#39;</span></span></code></pre></div><p>The rating flips to <code>large</code> and the patches disappear from the result. The policy in the tool decides how much the model gets to see, before any model turn spends a token on it.</p><h2 id="add-the-review-tool" tabindex="-1">Add the review tool <a class="header-anchor" href="#add-the-review-tool" aria-label="Permalink to &quot;Add the review tool&quot;">​</a></h2><p>The approver needs a way to act on its verdict. Create <code>agent/tools/submit_review.ts</code>:</p><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { defineTool } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;@cursor/july/tools&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
+<span class="line"><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">  --input</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &#39;{&quot;prUrl&quot;:&quot;https://github.com/facebook/react/pull/25229&quot;}&#39;</span></span></code></pre></div><p>The rating flips to <code>large</code> and the patches disappear from the result. The policy in the tool decides how much the model gets to see, before any model turn spends a token on it.</p><h2 id="add-the-review-tool" tabindex="-1">Add the review tool <a class="header-anchor" href="#add-the-review-tool" aria-label="Permalink to &quot;Add the review tool&quot;">​</a></h2><p>The approver needs a way to act on its verdict. Create <code>agent/tools/submit_review.ts</code>:</p><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { defineTool } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;@cursor/july/tools&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
 <span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { z } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;zod&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
 <span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { parsePullUrl } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;../lib/github.js&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
 <span class="line"></span>