@cursor/july 0.1.108 → 0.1.109

package/dist/docs/assets/{skills_evals.md.BhovOvrl.js → skills_evals.md.DFxYPErF.js}

@@ -1,7 +1,7 @@
-import{_ as i,c as a,o as e,ag as t}from"./chunks/framework.BCISBCiQ.js";const c=JSON.parse('{"title":"Agent SDK evals","description":"Author and run defineEval cases. Use when writing, fixing, seeding, or hillclimbing evals. Ask whether to generate samples or take uploaded data. Materialize API-backed fixtures first. Live A/B is defineAB (skills/ab), not defineEval.","frontmatter":{"name":"agentsdk-evals","description":"Author and run defineEval cases. Use when writing, fixing, seeding, or hillclimbing evals. Ask whether to generate samples or take uploaded data. Materialize API-backed fixtures first. Live A/B is defineAB (skills/ab), not defineEval."},"headers":[],"relativePath":"../skills/evals/SKILL.md","filePath":"../skills/evals/SKILL.md"}'),n={name:"../skills/evals/SKILL.md"};function l(h,s,d,p,k,r){return e(),a("div",null,[...s[0]||(s[0]=[t(`<h1 id="agent-sdk-evals" tabindex="-1">Agent SDK evals <a class="header-anchor" href="#agent-sdk-evals" aria-label="Permalink to &quot;Agent SDK evals&quot;">​</a></h1><p>Fixed input, model turn, gates on the trajectory. Files live at project-root <code>evals/**/*.eval.ts</code>. <code>agent/evals/</code> is ignored.</p><p>Live traffic variants: <code>skills/ab/SKILL.md</code>. That is not a test runner.</p><div class="language-bash vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">bash</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> eval</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --list</span></span>
+import{_ as e,c as a,o as i,ag as t}from"./chunks/framework.BCISBCiQ.js";const k=JSON.parse('{"title":"Agent SDK evals","description":"Author and run defineEval cases. Use when writing, fixing, seeding, or hillclimbing evals. Ask whether to generate samples or take uploaded data. Materialize API-backed fixtures first.","frontmatter":{"name":"agentsdk-evals","description":"Author and run defineEval cases. Use when writing, fixing, seeding, or hillclimbing evals. Ask whether to generate samples or take uploaded data. Materialize API-backed fixtures first."},"headers":[],"relativePath":"../skills/evals/SKILL.md","filePath":"../skills/evals/SKILL.md"}'),n={name:"../skills/evals/SKILL.md"};function d(l,s,o,h,r,p){return i(),a("div",null,[...s[0]||(s[0]=[t(`<h1 id="agent-sdk-evals" tabindex="-1">Agent SDK evals <a class="header-anchor" href="#agent-sdk-evals" aria-label="Permalink to &quot;Agent SDK evals&quot;">​</a></h1><p>Fixed input, model turn, gates on the trajectory. Files live at project-root <code>evals/**/*.eval.ts</code>. <code>agent/evals/</code> is ignored. Guide: <code>docs/evals.md</code>.</p><p>Observing production: hooks. Tool logic without a model: <code>agent-sdk call &lt;tool&gt; --input &#39;{...}&#39;</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;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --list</span></span>
 <span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> eval</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --json</span></span>
 <span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> eval</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> weather/nyc</span></span>
-<span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> eval</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --tag</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> smoke</span></span></code></pre></div><table tabindex="0"><thead><tr><th>Form</th><th>Case id</th></tr></thead><tbody><tr><td><code>evals/weather.eval.ts</code> + <code>test</code></td><td><code>weather</code></td></tr><tr><td><code>evals/weather/nyc.eval.ts</code> + <code>test</code></td><td><code>weather/nyc</code></td></tr><tr><td><code>evals/weather.eval.ts</code> + <code>{ id: &quot;nyc&quot; }</code></td><td><code>weather/nyc</code></td></tr></tbody></table><p><code>eval</code> boots an ephemeral server and a temp state root. <code>--url</code> points at a running agent. Model turns need a Cursor credential (<code>CURSOR_API_KEY</code>, then <code>CURSOR_API_KEY_FILE</code>, then <code>CURSOR_SERVICE_ACCOUNT_KEY</code>, then <code>agent-sdk login</code>).</p><h2 id="seeding" tabindex="-1">Seeding <a class="header-anchor" href="#seeding" aria-label="Permalink to &quot;Seeding&quot;">​</a></h2><p>Creating or expanding cases: <code>AskQuestion</code> first.</p><table tabindex="0"><thead><tr><th>Question</th><th>Options</th></tr></thead><tbody><tr><td><strong>How should we get eval samples?</strong></td><td>Generate test eval samples for me <code>(recommended)</code> · I will add / upload the data manually</td></tr></tbody></table><ol><li><strong>Manual.</strong> They provide files or paste. Show the shape below. Do not invent cases. Then <code>eval --list</code> and wire gates. API-backed pointers (PR URLs, SHAs, gold labels): materialize under <code>fixtures/</code> first.</li><li><strong>Generated.</strong> Ask count (<code>3</code> recommended). Append to an existing <code>cases</code> array when it fits. Never overwrite or weaken a datapoint. Create <code>evals/evals.config.ts</code> if missing (<code>maxConcurrency: 20</code>; cap 200).</li></ol><h2 id="api" tabindex="-1">API <a class="header-anchor" href="#api" aria-label="Permalink to &quot;API&quot;">​</a></h2><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { defineEval, includes, satisfies } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;@cursor/july/evals&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
+<span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">agent-sdk</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> eval</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --dir</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> .</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> --tag</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> smoke</span></span></code></pre></div><table tabindex="0"><thead><tr><th>Form</th><th>Case id</th></tr></thead><tbody><tr><td><code>evals/weather.eval.ts</code> + <code>test</code></td><td><code>weather</code></td></tr><tr><td><code>evals/weather/nyc.eval.ts</code> + <code>test</code></td><td><code>weather/nyc</code></td></tr><tr><td><code>evals/weather.eval.ts</code> + <code>{ id: &quot;nyc&quot; }</code></td><td><code>weather/nyc</code></td></tr><tr><td><code>evals/sql.eval.ts</code> exporting an array of <code>defineEval</code> calls</td><td><code>sql/0000</code>, <code>sql/0001</code>, ...</td></tr></tbody></table><p><code>eval</code> boots an ephemeral server and a temp state root. <code>--url</code> points at a running agent. Model turns need a Cursor credential (<code>CURSOR_API_KEY</code>, then <code>CURSOR_API_KEY_FILE</code>, then <code>CURSOR_SERVICE_ACCOUNT_KEY</code>, then <code>agent-sdk login</code>); <code>--list</code> does not. Node 22.13+, never Bun.</p><h2 id="seeding" tabindex="-1">Seeding <a class="header-anchor" href="#seeding" aria-label="Permalink to &quot;Seeding&quot;">​</a></h2><p>Creating or expanding cases: <code>AskQuestion</code> first.</p><table tabindex="0"><thead><tr><th>Question</th><th>Options</th></tr></thead><tbody><tr><td><strong>How should we get eval samples?</strong></td><td>Generate test eval samples for me <code>(recommended)</code> · I will add / upload the data manually</td></tr></tbody></table><ol><li><strong>Manual.</strong> They provide files or paste. Show the shape below. Do not invent cases. Then <code>eval --list</code> and wire gates. API-backed pointers (PR URLs, SHAs, gold labels): materialize under <code>fixtures/</code> first.</li><li><strong>Generated.</strong> Ask count (<code>3</code> recommended). Append to an existing <code>cases</code> array when it fits. Never overwrite or weaken a datapoint. Create <code>evals/evals.config.ts</code> if missing (<code>maxConcurrency: 20</code>; cap 200).</li></ol><h2 id="api" tabindex="-1">API <a class="header-anchor" href="#api" aria-label="Permalink to &quot;API&quot;">​</a></h2><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { defineEval, includes, satisfies } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;@cursor/july/evals&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">;</span></span>
 <span class="line"></span>
 <span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">export</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> default</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> defineEval</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">({</span></span>
 <span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  tags: [</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;smoke&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">, </span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;">&quot;weather&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">],</span></span>
@@ -22,4 +22,4 @@ import{_ as i,c as a,o as e,ag as t}from"./chunks/framework.BCISBCiQ.js";const c
 <span class="line"></span>
 <span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">export</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> default</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> defineEvalConfig</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">({</span></span>
 <span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  maxConcurrency: </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">20</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">});</span></span></code></pre></div><p>Either <code>test(t)</code> or <code>cases</code>, not both. <code>t.send</code> waits for park/fail. <code>workspaceFiles</code> seeds the first turn. Assert with <code>t.succeeded()</code>, <code>calledTool</code> / <code>notCalledTool</code>, <code>t.check(t.reply, …)</code>, <code>t.metric</code>.</p><h2 id="what-to-gate" tabindex="-1">What to gate <a class="header-anchor" href="#what-to-gate" aria-label="Permalink to &quot;What to gate&quot;">​</a></h2><p>Decisions and shape, not prose.</p><ol><li><code>t.succeeded()</code> first</li><li>Intended tool + the tempting wrong one</li><li>A shape regex or <code>satisfies</code> on parsed fields</li><li>If formatting keeps failing, tighten instructions. Do not move rendering into a host tool.</li></ol><p>Anti-patterns: exact phrasing; more than ~5 gates (split); live drifting inputs (pin them).</p><table tabindex="0"><thead><tr><th>Surface</th><th>Fixture</th></tr></thead><tbody><tr><td>Chat</td><td>One frozen prompt</td></tr><tr><td>Tool-heavy</td><td><code>agent-sdk call</code> first, then the prompt</td></tr><tr><td>GitHub</td><td><code>github replay … --dry-run --out fixtures/github</code></td></tr><tr><td>Host-prep PR review</td><td>A team-owned PR; gate findings shape, not counts</td></tr><tr><td>Workspace</td><td><code>workspaceFiles</code> in <code>t.send</code></td></tr></tbody></table><p>Every kept hillclimb change lands an eval that would have failed before it. Never weaken a gate to pass a round.</p>`,20)])])}const E=i(n,[["render",l]]);export{c as __pageData,E as default};
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">});</span></span></code></pre></div><p>Either <code>test(t)</code> or <code>cases</code>, not both. <code>t.send</code> waits for the turn to settle (complete, park, or fail) and returns it; <code>turn.calledTool(...)</code> reads only that turn, <code>t.*</code> reads the whole run. First-send options: <code>workspaceFiles</code>, <code>workspaceDir</code>, <code>cloud</code>.</p><p>Gates: <code>t.succeeded()</code> / <code>t.parked()</code>, <code>calledTool(name, { input, output, status, count })</code> / <code>notCalledTool</code>, <code>toolOrder</code>, <code>maxToolCalls</code>, <code>taggedArtifact</code>, <code>event</code> / <code>notEvent</code>, <code>t.check(value, includes | equals | matches | similarity | satisfies)</code>. Records: <code>t.metric</code>, <code>t.log</code>, <code>t.score</code> (soft 0-1).</p><p>Severity is on the handle: default gate; <code>.soft()</code> tracked; <code>.atLeast(0.8)</code> soft with a bar (verdict <code>scored</code>, exit 1 only under <code>--strict</code>); <code>.gate(0.8)</code> hard. <code>t.judge.factuality</code>, <code>summarizes</code>, <code>closedQA</code>, and <code>sql</code> are soft by default and need a judge model (<code>judge</code> in <code>evals.config.ts</code>, on the eval or case, or per-call <code>{ model }</code>).</p><h2 id="side-effects" tabindex="-1">Side effects <a class="header-anchor" href="#side-effects" aria-label="Permalink to &quot;Side effects&quot;">​</a></h2><p>Eval sessions run real tools. Guard actuation with <code>ctx.session.purpose === &quot;eval&quot;</code> in the tool, hook, or <code>defineResult</code> and return a shaped result so <code>calledTool</code> can still assert the decision.</p><h2 id="what-to-gate" tabindex="-1">What to gate <a class="header-anchor" href="#what-to-gate" aria-label="Permalink to &quot;What to gate&quot;">​</a></h2><p>Decisions and shape, not prose.</p><ol><li><code>t.succeeded()</code> first</li><li>Intended tool + the tempting wrong one</li><li>A shape regex or <code>satisfies</code> on parsed fields</li><li>If formatting keeps failing, tighten instructions. Do not move rendering into a host tool.</li></ol><p>Anti-patterns: exact phrasing; more than ~5 gates (split); live drifting inputs (pin them).</p><table tabindex="0"><thead><tr><th>Surface</th><th>Fixture</th></tr></thead><tbody><tr><td>Chat</td><td>One frozen prompt</td></tr><tr><td>Tool-heavy</td><td><code>agent-sdk call</code> first, then the prompt</td></tr><tr><td>GitHub</td><td><code>github replay … --dry-run --out fixtures/github</code></td></tr><tr><td>Host-prep PR review</td><td>A team-owned PR; gate findings shape, not counts</td></tr><tr><td>Workspace</td><td><code>workspaceFiles</code> in <code>t.send</code></td></tr><tr><td>Dataset</td><td><code>loadJson</code> / <code>loadJsonl</code> / <code>loadYaml</code> from <code>@cursor/july/evals/loaders</code>, paths from the project root</td></tr></tbody></table><h2 id="debug-and-ci" tabindex="-1">Debug and CI <a class="header-anchor" href="#debug-and-ci" aria-label="Permalink to &quot;Debug and CI&quot;">​</a></h2><p>A failed local run leaves <code>evals/&lt;case-id&gt;.json</code> (assertions, inputs, tool I/O, final text, logs) under <code>evals/&lt;stamp&gt;/</code> in the project state directory (the CLI prints the path); read it before editing the eval. <code>--state-root &lt;path&gt;</code> keeps the raw session events too.</p><p>CI: <code>agent-sdk eval --dir . --json --no-stream --junit reports/evals.xml</code>. Exit <code>1</code> on a failed gate, <code>2</code> when nothing matched, <code>--strict</code> to fail on <code>scored</code>.</p><p>Every kept hillclimb change lands an eval that would have failed before it. Never weaken a gate to pass a round.</p>`,27)])])}const g=e(n,[["render",d]]);export{k as __pageData,g as default};

package/dist/docs/assets/skills_evals.md.DFxYPErF.lean.js

@@ -0,0 +1 @@
+import{_ as e,c as a,o as i,ag as t}from"./chunks/framework.BCISBCiQ.js";const k=JSON.parse('{"title":"Agent SDK evals","description":"Author and run defineEval cases. Use when writing, fixing, seeding, or hillclimbing evals. Ask whether to generate samples or take uploaded data. Materialize API-backed fixtures first.","frontmatter":{"name":"agentsdk-evals","description":"Author and run defineEval cases. Use when writing, fixing, seeding, or hillclimbing evals. Ask whether to generate samples or take uploaded data. Materialize API-backed fixtures first."},"headers":[],"relativePath":"../skills/evals/SKILL.md","filePath":"../skills/evals/SKILL.md"}'),n={name:"../skills/evals/SKILL.md"};function d(l,s,o,h,r,p){return i(),a("div",null,[...s[0]||(s[0]=[t("",27)])])}const g=e(n,[["render",d]]);export{k as __pageData,g as default};

package/dist/docs/assets/{skills_framework-map.md.D-tFZhFS.js → skills_framework-map.md.BxLSOhSY.js}

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

package/dist/docs/assets/skills_index.md.DL7EHaQ-.js

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

package/dist/docs/assets/skills_index.md.DL7EHaQ-.lean.js

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

package/dist/docs/assets/{storage.md.BOHeqk2M.js → storage.md.BUrhJ-Zz.js}

@@ -1,4 +1,4 @@
-import{_ as e,c as i,o as t,ag as a}from"./chunks/framework.BCISBCiQ.js";const c=JSON.parse(`{"title":"Storage","description":"Point the Agent SDK's durable storage at a backend you own with defineStorage.","frontmatter":{"title":"Storage","description":"Point the Agent SDK's durable storage at a backend you own with defineStorage."},"headers":[],"relativePath":"storage.md","filePath":"storage.md"}`),n={name:"storage.md"};function o(d,s,h,r,l,p){return t(),i("div",null,[...s[0]||(s[0]=[a(`<h1 id="storage" tabindex="-1">Storage <a class="header-anchor" href="#storage" aria-label="Permalink to &quot;Storage&quot;">​</a></h1><p>The Agent SDK owns durable storage for sessions, continuation tokens, reminders, playground eval history, and live A/B samples. It chooses the keys, when to read and write, and how to restore after restart.</p><p>The Agent SDK owns key encoding. Backends must accept the keys they are given. Do not fail <code>put</code> to enforce a shorter cap.</p><p>By default that storage lives under <code>--state-root</code> on local disk. Fine for one machine; it does not survive replacing the host.</p><p>To keep the same framework storage across hosts, plug in a key-value backend with <code>agent/storage.ts</code>. You provide <code>put</code> / <code>get</code> / <code>delete</code> / <code>list</code>. The optional <code>cas</code> group adds conditional writes (see <a href="#conditional-writes-the-cas-group">Conditional writes</a>).</p><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">// agent/storage.ts</span></span>
+import{_ as i,c as e,o as t,ag as a}from"./chunks/framework.BCISBCiQ.js";const c=JSON.parse(`{"title":"Storage","description":"Point the Agent SDK's durable storage at a backend you own with defineStorage.","frontmatter":{"title":"Storage","description":"Point the Agent SDK's durable storage at a backend you own with defineStorage."},"headers":[],"relativePath":"storage.md","filePath":"storage.md"}`),n={name:"storage.md"};function o(h,s,r,l,d,p){return t(),e("div",null,[...s[0]||(s[0]=[a(`<h1 id="storage" tabindex="-1">Storage <a class="header-anchor" href="#storage" aria-label="Permalink to &quot;Storage&quot;">​</a></h1><p>The Agent SDK owns durable storage for sessions, continuation tokens, reminders, and playground eval history. It chooses the keys, when to read and write, and how to restore after restart.</p><p>The Agent SDK owns key encoding. Backends must accept the keys they are given. Do not fail <code>put</code> to enforce a shorter cap.</p><p>By default that storage lives under <code>--state-root</code> on local disk. Fine for one machine; it does not survive replacing the host.</p><p>To keep the same framework storage across hosts, plug in a key-value backend with <code>agent/storage.ts</code>. You provide <code>put</code> / <code>get</code> / <code>delete</code> / <code>list</code>. The optional <code>cas</code> group adds conditional writes (see <a href="#conditional-writes-the-cas-group">Conditional writes</a>).</p><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">// agent/storage.ts</span></span>
 <span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { defineStorage } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;@cursor/july/storage&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;"> defineStorage</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">({</span></span>
@@ -12,7 +12,7 @@ import{_ as e,c as i,o as t,ag as a}from"./chunks/framework.BCISBCiQ.js";const c
 <span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">    putIfVersion</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: (</span><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">key</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">, </span><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">value</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">, </span><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">expected</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;"> db.</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">casUpdate</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(key, value, expected),</span></span>
 <span class="line"><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">    listKeys</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">: (</span><span style="--shiki-light:#E36209;--shiki-dark:#FFAB70;">prefix</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;"> db.</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">listKeysByPrefix</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">(prefix),</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><h2 id="which-fields-to-provide" tabindex="-1">Which fields to provide <a class="header-anchor" href="#which-fields-to-provide" aria-label="Permalink to &quot;Which fields to provide&quot;">​</a></h2><p>Implement the small KV core (<code>put</code>/<code>get</code>/<code>delete</code>/<code>list</code> plus the <code>cas</code> group) and you get <strong>full functionality</strong>: eval-run and A/B history are derived over the core automatically. The dedicated <code>evals</code> / <code>abs</code> groups are backend-native optimizations, not required-or-lose-history hooks.</p><table tabindex="0"><thead><tr><th>Field</th><th>Required</th><th>Role</th></tr></thead><tbody><tr><td><code>put</code></td><td>Yes</td><td>Write or update a value</td></tr><tr><td><code>cas</code></td><td>Optional</td><td>Conditional writes; see <a href="#conditional-writes-the-cas-group">Conditional writes</a></td></tr><tr><td><code>get</code></td><td>For restore</td><td>Look up one key</td></tr><tr><td><code>list</code></td><td>For restore</td><td>Return entries under a prefix, in key order</td></tr><tr><td><code>delete</code></td><td>For cleanup</td><td>Remove a key</td></tr><tr><td><code>name</code></td><td>No</td><td>Label surfaced on <code>GET /v1/info</code> diagnostics</td></tr><tr><td><code>policy</code></td><td>No</td><td>Timing knobs; see <a href="#policy">Policy</a></td></tr><tr><td><code>evals</code></td><td>No</td><td>Backend-native eval-runs table; derived over the core when omitted. See <a href="#eval-and-a-b-tables">Eval and A/B tables</a></td></tr><tr><td><code>abs</code></td><td>No</td><td>Backend-native A/B metrics table; derived over the core when omitted. See <a href="#eval-and-a-b-tables">Eval and A/B tables</a></td></tr></tbody></table><p>A throwing <code>put</code> is logged and dropped. It never fails a turn. When resolving a missing continuation token, a throwing <code>get</code> fails the follow-up so a store outage does not open a new session. Return <code>undefined</code> only for a real miss.</p><h2 id="eval-and-a-b-tables" tabindex="-1">Eval and A/B tables <a class="header-anchor" href="#eval-and-a-b-tables" aria-label="Permalink to &quot;Eval and A/B tables&quot;">​</a></h2><p>Two dedicated table groups carry structured rows instead of opaque KV values. Both are <strong>optional optimizations</strong>: when a group is not authored, <code>defineStorage</code> derives it over the KV core, so a backend that implements only the core loses nothing. Author a group only when the backend has a better native shape (a real database table, an analytics pipeline). The built-in <code>fileKv</code> and <code>cursorHostedStorage</code> both do.</p><p><code>evals</code> keeps playground eval batches across restarts (<code>put</code>, <code>delete</code>, <code>list</code> over run snapshots keyed by <code>runId</code>). A core missing <code>delete</code> or <code>list</code> leaves eval history in memory until restart. See <a href="./evals.html#configure-eval-runs">Evals</a>.</p><p><code>abs</code> exports live A/B metrics: <code>putSample</code> appends one cumulative metric sample per enrolled experiment on each completed or failed turn; optional <code>putSnapshot</code> / <code>getSnapshot</code> store and serve back the latest aggregate so a replacement host can still serve the A/Bs surface. <code>putSample</code> and <code>putSnapshot</code> need only core <code>put</code>; <code>getSnapshot</code> needs core <code>get</code>. Session event logs remain the assignment source of truth either way. See <a href="./ab.html">Live A/B metrics</a>.</p><h2 id="policy" tabindex="-1">Policy <a class="header-anchor" href="#policy" aria-label="Permalink to &quot;Policy&quot;">​</a></h2><p>Two knobs change behavior:</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;">export</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> default</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> defineStorage</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><h2 id="which-fields-to-provide" tabindex="-1">Which fields to provide <a class="header-anchor" href="#which-fields-to-provide" aria-label="Permalink to &quot;Which fields to provide&quot;">​</a></h2><p>Implement the small KV core (<code>put</code>/<code>get</code>/<code>delete</code>/<code>list</code> plus the <code>cas</code> group) and you get <strong>full functionality</strong>: eval-run history is derived over the core automatically. The dedicated <code>evals</code> group is a backend-native optimization, not a required-or-lose-history hook.</p><table tabindex="0"><thead><tr><th>Field</th><th>Required</th><th>Role</th></tr></thead><tbody><tr><td><code>put</code></td><td>Yes</td><td>Write or update a value</td></tr><tr><td><code>cas</code></td><td>Optional</td><td>Conditional writes; see <a href="#conditional-writes-the-cas-group">Conditional writes</a></td></tr><tr><td><code>get</code></td><td>For restore</td><td>Look up one key</td></tr><tr><td><code>list</code></td><td>For restore</td><td>Return entries under a prefix, in key order</td></tr><tr><td><code>delete</code></td><td>For cleanup</td><td>Remove a key</td></tr><tr><td><code>name</code></td><td>No</td><td>Label surfaced on <code>GET /v1/info</code> diagnostics</td></tr><tr><td><code>policy</code></td><td>No</td><td>Timing knobs; see <a href="#policy">Policy</a></td></tr><tr><td><code>evals</code></td><td>No</td><td>Backend-native eval-runs table; derived over the core when omitted. See <a href="#eval-table">Eval table</a></td></tr></tbody></table><p>A throwing <code>put</code> is logged and dropped. It never fails a turn. When resolving a missing continuation token, a throwing <code>get</code> fails the follow-up so a store outage does not open a new session. Return <code>undefined</code> only for a real miss.</p><h2 id="eval-table" tabindex="-1">Eval table <a class="header-anchor" href="#eval-table" aria-label="Permalink to &quot;Eval table&quot;">​</a></h2><p>The dedicated <code>evals</code> table group carries structured rows instead of opaque KV values. It is an <strong>optional optimization</strong>: when the group is not authored, <code>defineStorage</code> derives it over the KV core, so a backend that implements only the core loses nothing. Author the group only when the backend has a better native shape (a real database table, an analytics pipeline). The built-in <code>fileKv</code> and <code>cursorHostedStorage</code> both do.</p><p><code>evals</code> keeps playground eval batches across restarts (<code>put</code>, <code>delete</code>, <code>list</code> over run snapshots keyed by <code>runId</code>). A core missing <code>delete</code> or <code>list</code> leaves eval history in memory until restart. See <a href="./evals.html#configure-eval-runs">Evals</a>.</p><h2 id="policy" tabindex="-1">Policy <a class="header-anchor" href="#policy" aria-label="Permalink to &quot;Policy&quot;">​</a></h2><p>Two knobs change behavior:</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;">export</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> default</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;"> defineStorage</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">({</span></span>
 <span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  policy: {</span></span>
 <span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">    // Batch event writes while the session is busy (default: once per turn)</span></span>
 <span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">    debounceMs: </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">30_000</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">,</span></span>
@@ -20,11 +20,11 @@ import{_ as e,c as i,o as t,ag as a}from"./chunks/framework.BCISBCiQ.js";const c
 <span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">    restore: { maxSessions: </span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;">500</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:#6A737D;--shiki-dark:#6A737D;">  // put / get / delete / list …</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">});</span></span></code></pre></div><table tabindex="0"><thead><tr><th>Knob</th><th>Default</th><th>Meaning</th></tr></thead><tbody><tr><td><code>debounceMs</code></td><td>unset (once per turn)</td><td>Wait this long after activity before writing event batches</td></tr><tr><td><code>restore</code></td><td>caps below</td><td>How much to load at serve start</td></tr><tr><td><code>restore.maxSessions</code></td><td><code>1000</code></td><td>Max sessions loaded at serve start</td></tr><tr><td><code>restore.maxAgeMs</code></td><td>30 days</td><td>Skip older sessions at serve start</td></tr><tr><td><code>restore.maxTotalBytes</code></td><td>1 GiB</td><td>Stop loading once this budget is reached</td></tr></tbody></table><p>Set <code>restore: &quot;off&quot;</code> on high-traffic hosts. Sessions then load when a follow-up arrives instead of at startup.</p><h2 id="restore-after-restart" tabindex="-1">Restore after restart <a class="header-anchor" href="#restore-after-restart" aria-label="Permalink to &quot;Restore after restart&quot;">​</a></h2><p>With <code>get</code> and <code>list</code>, serve can rebuild local state from your store:</p><ul><li>At startup, the Agent SDK loads recent sessions up to the restore caps.</li><li>On demand, a missing continuation token resolves through the store and resumes that session.</li><li>Playground eval history and A/B aggregates can load from the same sink.</li></ul><p>A turn in flight at crash time is not replayed. The next follow-up resumes from the last flushed state.</p><h2 id="author-kv-ctx-host-kv" tabindex="-1">Author KV (<code>ctx.host.kv</code>) <a class="header-anchor" href="#author-kv-ctx-host-kv" aria-label="Permalink to &quot;Author KV (\`ctx.host.kv\`)&quot;">​</a></h2><p>Handlers can store their own JSON under the same sink without minting framework keys:</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;">await</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> ctx.host.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:#032F62;--shiki-dark:#9ECBFF;">&quot;alert-memory/abc&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">, { updated: </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></code></pre></div><table tabindex="0"><thead><tr><th>Knob</th><th>Default</th><th>Meaning</th></tr></thead><tbody><tr><td><code>debounceMs</code></td><td>unset (once per turn)</td><td>Wait this long after activity before writing event batches</td></tr><tr><td><code>restore</code></td><td>caps below</td><td>How much to load at serve start</td></tr><tr><td><code>restore.maxSessions</code></td><td><code>1000</code></td><td>Max sessions loaded at serve start</td></tr><tr><td><code>restore.maxAgeMs</code></td><td>30 days</td><td>Skip older sessions at serve start</td></tr><tr><td><code>restore.maxTotalBytes</code></td><td>1 GiB</td><td>Stop loading once this budget is reached</td></tr></tbody></table><p>Set <code>restore: &quot;off&quot;</code> on high-traffic hosts. Sessions then load when a follow-up arrives instead of at startup.</p><h2 id="restore-after-restart" tabindex="-1">Restore after restart <a class="header-anchor" href="#restore-after-restart" aria-label="Permalink to &quot;Restore after restart&quot;">​</a></h2><p>With <code>get</code> and <code>list</code>, serve can rebuild local state from your store:</p><ul><li>At startup, the Agent SDK loads recent sessions up to the restore caps.</li><li>On demand, a missing continuation token resolves through the store and resumes that session.</li><li>Playground eval history can load from the same sink.</li></ul><p>A turn in flight at crash time is not replayed. The next follow-up resumes from the last flushed state.</p><h2 id="author-kv-ctx-host-kv" tabindex="-1">Author KV (<code>ctx.host.kv</code>) <a class="header-anchor" href="#author-kv-ctx-host-kv" aria-label="Permalink to &quot;Author KV (\`ctx.host.kv\`)&quot;">​</a></h2><p>Handlers can store their own JSON under the same sink without minting framework keys:</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;">await</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> ctx.host.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:#032F62;--shiki-dark:#9ECBFF;">&quot;alert-memory/abc&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">, { updated: </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:#D73A49;--shiki-dark:#F97583;">const</span><span style="--shiki-light:#005CC5;--shiki-dark:#79B8FF;"> prior</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> =</span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;"> await</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> ctx.host.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:#032F62;--shiki-dark:#9ECBFF;">&quot;alert-memory/abc&quot;</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">);</span></span></code></pre></div><p>Writes <strong>await</strong> the sink and propagate errors, unlike session mirrors, which are at-most-once.</p><p>Without <code>agent/storage.ts</code>, <code>host.kv</code> is local-only and does not survive replacing the host. Hosted agents should author <code>cursorHostedStorage()</code> so session and <code>host.kv</code> records survive replace.</p><div class="language-ts vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">ts</span><pre class="shiki shiki-themes github-light github-dark vp-code" tabindex="0"><code><span class="line"><span style="--shiki-light:#6A737D;--shiki-dark:#6A737D;">// agent/storage.ts: Cursor-managed hosting</span></span>
 <span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">import</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;"> { defineStorage } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;@cursor/july/storage&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;"> { cursorHostedStorage } </span><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">from</span><span style="--shiki-light:#032F62;--shiki-dark:#9ECBFF;"> &quot;@cursor/july/storage/cursor-hosted&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;"> defineStorage</span><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">({</span></span>
 <span class="line"><span style="--shiki-light:#D73A49;--shiki-dark:#F97583;">  ...</span><span style="--shiki-light:#6F42C1;--shiki-dark:#B392F0;">cursorHostedStorage</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>Built-in helpers:</p><table tabindex="0"><thead><tr><th>Import</th><th>Backend</th></tr></thead><tbody><tr><td><code>@cursor/july/storage/file-kv</code></td><td>File-per-key under the project state directory</td></tr><tr><td><code>@cursor/july/storage/cursor-hosted</code></td><td>Cursor-managed hosting; records survive replace</td></tr></tbody></table><h2 id="bring-your-own-backend" tabindex="-1">Bring your own backend <a class="header-anchor" href="#bring-your-own-backend" aria-label="Permalink to &quot;Bring your own backend&quot;">​</a></h2><p>Any durable store works if it can put, get, delete, list by prefix, and optionally compare-and-set. Implement the <code>StorageConfig</code> methods (and the optional <code>cas</code> group) against that store. The contract, defined at <code>@cursor/july/kv</code>: version tokens are opaque strings that change on every successful write, including plain <code>put</code>, so a stale token fences instead of clobbering; conditional writes are atomic; <code>listKeys</code> returns every key under the prefix.</p><h2 id="conditional-writes-the-cas-group" tabindex="-1">Conditional writes (the <code>cas</code> group) <a class="header-anchor" href="#conditional-writes-the-cas-group" aria-label="Permalink to &quot;Conditional writes (the \`cas\` group)&quot;">​</a></h2><p>The <code>cas</code> group is compare-and-swap over the same keyspace: <code>getWithVersion</code> / <code>putIfAbsent</code> / <code>putIfVersion</code> / <code>listKeys</code>, defined backend-agnostically at <code>@cursor/july/kv</code>. Plain storage works without it, so an existing backend keeps working across a platform upgrade. <code>defineStorage</code> rejects a partial group. Implement all four methods or none. Version tokens are opaque strings that must change on every write (a counter column, a row version, a content hash). The built-in backends both include the group.</p>`,35)])])}const g=e(n,[["render",o]]);export{c as __pageData,g as default};
+<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">});</span></span></code></pre></div><p>Built-in helpers:</p><table tabindex="0"><thead><tr><th>Import</th><th>Backend</th></tr></thead><tbody><tr><td><code>@cursor/july/storage/file-kv</code></td><td>File-per-key under the project state directory</td></tr><tr><td><code>@cursor/july/storage/cursor-hosted</code></td><td>Cursor-managed hosting; records survive replace</td></tr></tbody></table><h2 id="bring-your-own-backend" tabindex="-1">Bring your own backend <a class="header-anchor" href="#bring-your-own-backend" aria-label="Permalink to &quot;Bring your own backend&quot;">​</a></h2><p>Any durable store works if it can put, get, delete, list by prefix, and optionally compare-and-set. Implement the <code>StorageConfig</code> methods (and the optional <code>cas</code> group) against that store. The contract, defined at <code>@cursor/july/kv</code>: version tokens are opaque strings that change on every successful write, including plain <code>put</code>, so a stale token fences instead of clobbering; conditional writes are atomic; <code>listKeys</code> returns every key under the prefix.</p><h2 id="conditional-writes-the-cas-group" tabindex="-1">Conditional writes (the <code>cas</code> group) <a class="header-anchor" href="#conditional-writes-the-cas-group" aria-label="Permalink to &quot;Conditional writes (the \`cas\` group)&quot;">​</a></h2><p>The <code>cas</code> group is compare-and-swap over the same keyspace: <code>getWithVersion</code> / <code>putIfAbsent</code> / <code>putIfVersion</code> / <code>listKeys</code>, defined backend-agnostically at <code>@cursor/july/kv</code>. Plain storage works without it, so an existing backend keeps working across a platform upgrade. <code>defineStorage</code> rejects a partial group. Implement all four methods or none. Version tokens are opaque strings that must change on every write (a counter column, a row version, a content hash). The built-in backends both include the group.</p>`,34)])])}const g=i(n,[["render",o]]);export{c as __pageData,g as default};

package/dist/docs/assets/{storage.md.BOHeqk2M.lean.js → storage.md.BUrhJ-Zz.lean.js}

@@ -1 +1 @@
-import{_ as e,c as i,o as t,ag as a}from"./chunks/framework.BCISBCiQ.js";const c=JSON.parse(`{"title":"Storage","description":"Point the Agent SDK's durable storage at a backend you own with defineStorage.","frontmatter":{"title":"Storage","description":"Point the Agent SDK's durable storage at a backend you own with defineStorage."},"headers":[],"relativePath":"storage.md","filePath":"storage.md"}`),n={name:"storage.md"};function o(d,s,h,r,l,p){return t(),i("div",null,[...s[0]||(s[0]=[a("",35)])])}const g=e(n,[["render",o]]);export{c as __pageData,g as default};
+import{_ as i,c as e,o as t,ag as a}from"./chunks/framework.BCISBCiQ.js";const c=JSON.parse(`{"title":"Storage","description":"Point the Agent SDK's durable storage at a backend you own with defineStorage.","frontmatter":{"title":"Storage","description":"Point the Agent SDK's durable storage at a backend you own with defineStorage."},"headers":[],"relativePath":"storage.md","filePath":"storage.md"}`),n={name:"storage.md"};function o(h,s,r,l,d,p){return t(),e("div",null,[...s[0]||(s[0]=[a("",34)])])}const g=i(n,[["render",o]]);export{c as __pageData,g as default};