@cursor/july 0.1.97 → 0.1.100

package/dist/docs/assets/hillclimbing.md.DhESf3OO.js

@@ -1,4 +0,0 @@
-import{_ as t,c as i,o as a,ag as l}from"./chunks/framework.BCISBCiQ.js";const m=JSON.parse('{"title":"Hillclimbing","description":"Improve an agent one measured round at a time, with the hillclimb skill running the loop with you.","frontmatter":{"title":"Hillclimbing","description":"Improve an agent one measured round at a time, with the hillclimb skill running the loop with you."},"headers":[],"relativePath":"hillclimbing.md","filePath":"hillclimbing.md"}'),o={name:"hillclimbing.md"};function n(r,e,s,h,d,c){return a(),i("div",null,[...e[0]||(e[0]=[l(`<h1 id="hillclimbing" tabindex="-1">Hillclimbing <a class="header-anchor" href="#hillclimbing" aria-label="Permalink to &quot;Hillclimbing&quot;">​</a></h1><p>Make an agent better on fixed inputs: measure, change one lever, remeasure, and lock every kept win with an eval.</p><h2 id="what-is-hillclimbing" tabindex="-1">What is hillclimbing? <a class="header-anchor" href="#what-is-hillclimbing" aria-label="Permalink to &quot;What is hillclimbing?&quot;">​</a></h2><p>Hillclimbing is a measured improvement loop. You pin a few fixtures, name the one dominant problem in the run, change one lever, and check the same fixtures again. Keep only what helps. Every kept change lands an <a href="./evals.html">eval</a> so the win stays put.</p><p>You don&#39;t have to run the loop alone. The package ships a coding-agent skill that drives it with you.</p><div class="language-mermaid vp-adaptive-theme"><button title="Copy Code" class="copy"></button><span class="lang">mermaid</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;">flowchart LR</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  measure[Measure] --&gt; change[Change one lever]</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  change --&gt; remeasure[Remeasure]</span></span>
-<span class="line"><span style="--shiki-light:#24292E;--shiki-dark:#E1E4E8;">  remeasure --&gt; measure</span></span></code></pre></div><h2 id="how-do-i-hillclimb-an-agent-with-a-coding-agent" tabindex="-1">How do I hillclimb an agent with a coding agent? <a class="header-anchor" href="#how-do-i-hillclimb-an-agent-with-a-coding-agent" aria-label="Permalink to &quot;How do I hillclimb an agent with a coding agent?&quot;">​</a></h2><p>Have Cursor read <a href="./../skills/hillclimb/SKILL.html"><code>skills/hillclimb/SKILL.md</code></a>.</p><p>Tell it:</p><ol><li><strong>Which agent</strong> you&#39;re improving (path or slug)</li><li><strong>One to three fixtures</strong> you&#39;ll reuse every round: a PR URL, a saved webhook body, or a canonical chat prompt</li><li><strong>What &quot;better&quot; means</strong> this round: correct tool choice, fewer tools, lower latency, or output quality. Name the freeze line too: API shape, public output, and existing evals that must stay green</li></ol><p>The skill serves the agent, hits your fixtures, reads the session trajectory, proposes one change, remeasures, and checks with you before the next round.</p><p>Other skills cover the edges:</p><table tabindex="0"><thead><tr><th>When you need…</th><th>Skill</th></tr></thead><tbody><tr><td>The measured improvement loop</td><td><a href="./../skills/hillclimb/SKILL.html"><code>skills/hillclimb/SKILL.md</code></a></td></tr><tr><td>An eval that locks a kept win</td><td><a href="./../skills/evals/SKILL.html"><code>skills/evals/SKILL.md</code></a></td></tr><tr><td>Repeatable GitHub webhook inputs</td><td><a href="./../skills/github/SKILL.html"><code>skills/github/SKILL.md</code></a></td></tr><tr><td>A run that misbehaves</td><td><a href="./../skills/debug/SKILL.html"><code>skills/debug/SKILL.md</code></a></td></tr></tbody></table><p>See <a href="./building-with-agents.html">Building agents with agents</a> for every framework skill and a good first prompt.</p><h2 id="what-do-i-need-before-a-hillclimb-round" tabindex="-1">What do I need before a hillclimb round? <a class="header-anchor" href="#what-do-i-need-before-a-hillclimb-round" aria-label="Permalink to &quot;What do I need before a hillclimb round?&quot;">​</a></h2><p>Agree on four things before you edit:</p><ol><li><strong>The target agent</strong>: the project you&#39;re improving</li><li><strong>Fixtures</strong>: one to three fixed inputs you can compare across runs</li><li><strong>Success criteria</strong>: what better means this round</li><li><strong>The freeze line</strong>: what must not change</li></ol><p>Pin the input first. A moving fixture is noise. For GitHub agents, use <code>agent-sdk github replay</code> (see the <a href="./guides/github.html">GitHub guide</a>). For a single tool without a model turn, use <code>agent-sdk call</code>. For a chat turn, use <code>agent-sdk run --dir . --message &quot;…&quot;</code>.</p><h2 id="how-do-i-run-one-hillclimb-round" tabindex="-1">How do I run one hillclimb round? <a class="header-anchor" href="#how-do-i-run-one-hillclimb-round" aria-label="Permalink to &quot;How do I run one hillclimb round?&quot;">​</a></h2><p><strong>Measure.</strong> Hit the agent the way a user would: playground, channel HTTP, or Slack in <code>--dev</code>. Or ask the hillclimb skill to do it. <code>agent-sdk run</code> returns a JSON trajectory and writes a trace under the project state directory.</p><p><strong>Reflect.</strong> Score the trajectory, not impressions. Was the answer right? Did the model thrash (too many tools, fat evidence, grep loops)? Did it invent work the host should have prepared? Name the single dominant problem for this round in one sentence. Example: &quot;Full-file dumps trigger grep loops.&quot;</p><p><strong>Change one lever.</strong> Prefer the smallest change that addresses that problem:</p><ol><li>Host prep: seed what the model needs so it doesn&#39;t hunt</li><li>Evidence shape: trim or reorder artifacts</li><li>Instructions and skills: tighten the procedure</li><li>Tool surface: remove or gate tools that invite wandering</li><li>Framework changes: only when the agent can&#39;t express the fix</li></ol><p><strong>Remeasure.</strong> Same fixtures. Diff tools, wall time, and quality side by side. Keep the change only if the target metric improves and the freeze line holds.</p><h2 id="how-do-i-lock-a-hillclimb-improvement-with-an-eval" tabindex="-1">How do I lock a hillclimb improvement with an eval? <a class="header-anchor" href="#how-do-i-lock-a-hillclimb-improvement-with-an-eval" aria-label="Permalink to &quot;How do I lock a hillclimb improvement with an eval?&quot;">​</a></h2><p>Every kept change needs an eval that would have failed before the change: a tool-choice gate, an <code>action.result</code> count bound, or an output-shape check. Run <code>agent-sdk eval --dir . --json</code> between rounds. Never weaken an existing gate to pass the round.</p><p>Details live in <a href="./evals.html">Evals</a>. The evals skill will author the case with you.</p><h2 id="what-habits-help-hillclimbing-stay-reliable" tabindex="-1">What habits help hillclimbing stay reliable? <a class="header-anchor" href="#what-habits-help-hillclimbing-stay-reliable" aria-label="Permalink to &quot;What habits help hillclimbing stay reliable?&quot;">​</a></h2><ul><li>One problem per round. Don&#39;t bundle &quot;trim evidence and rewrite instructions&quot; unless you chose that on purpose.</li><li>Keep fixtures fixed until you deliberately need a harder case.</li><li>Separate host work from model tools when you blame latency. Moving deterministic prep onto the host is often the biggest win. In one PR reviewer, host-prepared evidence cut turns from about 8 minutes to about 1 minute.</li><li>Spot-check quality on at least one fixture against a known-good answer. Efficiency-only climbs quietly drop findings.</li><li>Treat <code>turn.failed</code> with <code>&quot;turn interrupted&quot;</code> as expected when a follow-up or stop preempted the turn.</li><li>Don&#39;t deploy, post to real surfaces, or weaken evals as part of a climb.</li></ul><h2 id="related" tabindex="-1">Related <a class="header-anchor" href="#related" aria-label="Permalink to &quot;Related&quot;">​</a></h2><ul><li><a href="./evals.html">Evals</a></li><li><a href="./building-with-agents.html">Building agents with agents</a></li><li><a href="./guides/github.html">GitHub guide</a></li><li><a href="./troubleshooting.html">Fix common agent problems</a></li></ul>`,31)])])}const p=t(o,[["render",n]]);export{m as __pageData,p as default};

package/dist/docs/assets/scaffolding-agents.md.D7UUkWw0.js

@@ -1 +0,0 @@
-import{_ as o,c as a,o as t,ag as l}from"./chunks/framework.BCISBCiQ.js";const f=JSON.parse('{"title":"Scaffold an agent with Cursor","description":"Use the bundled create-agent skill to plan, build, and verify a new agent.","frontmatter":{"title":"Scaffold an agent with Cursor","description":"Use the bundled create-agent skill to plan, build, and verify a new agent."},"headers":[],"relativePath":"scaffolding-agents.md","filePath":"scaffolding-agents.md"}'),s={name:"scaffolding-agents.md"};function i(r,e,n,d,c,h){return t(),a("div",null,[...e[0]||(e[0]=[l('<h1 id="scaffold-an-agent-with-cursor" tabindex="-1">Scaffold an agent with Cursor <a class="header-anchor" href="#scaffold-an-agent-with-cursor" aria-label="Permalink to &quot;Scaffold an agent with Cursor&quot;">​</a></h1><p>Turn an idea into a verified agent while Cursor guides you through each decision.</p><h2 id="what-does-the-create-agent-skill-do" tabindex="-1">What does the create-agent skill do? <a class="header-anchor" href="#what-does-the-create-agent-skill-do" aria-label="Permalink to &quot;What does the create-agent skill do?&quot;">​</a></h2><p>The bundled <a href="./../skills/create-agent/SKILL.html"><code>create-agent</code> skill</a> turns your goal into a small working project. Have Cursor read that file and follow it.</p><p>Where to find the file depends on how you got the package:</p><ul><li>Installing <code>@cursor/july</code> (<code>npm install</code>, <code>npx @cursor/july</code>, a version bump) copies every package skill into <code>~/.cursor/skills/agentsdk/</code> with <code>alwaysApply: true</code>, so Cursor injects the skill body into context instead of waiting for the model to pick it from the catalog. The <code>/</code> menu lists them as <code>/agentsdk-create-agent</code>, <code>/agentsdk-hillclimb</code>, and the rest. Re-installing overwrites those copies with the package version.</li><li>Installed <code>@cursor/july</code> as a dependency? The skill also ships inside the package at <code>node_modules/@cursor/july/skills/create-agent/SKILL.md</code>.</li><li>Working from this package&#39;s source? The skill is at <a href="./../skills/create-agent/SKILL.html"><code>skills/create-agent/SKILL.md</code></a>. Run <code>agent-sdk install-skills</code> if you want the same copies in <code>~/.cursor/skills/agentsdk/</code> (the package postinstall skips the source checkout).</li></ul><p>Cursor will:</p><ul><li>Ask only for choices missing from your prompt</li><li>Recommend defaults based on what you want to build</li><li>Show you the plan and file tree before writing files</li><li>Create the agent after you confirm the plan</li><li>Run structural checks, a real turn, and a smoke eval</li></ul><p>Use this skill for a new agent. Use <a href="./guides/convert-automation.html">convert-automation</a> when the starting point is a Cursor Automation in the dashboard. Use <a href="./hillclimbing.html"><code>hillclimb</code></a> (<code>skills/hillclimb/SKILL.md</code>) when an existing agent works but needs better results.</p><h2 id="how-do-i-start-a-guided-scaffold" tabindex="-1">How do I start a guided scaffold? <a class="header-anchor" href="#how-do-i-start-a-guided-scaffold" aria-label="Permalink to &quot;How do I start a guided scaffold?&quot;">​</a></h2><p>Describe the outcome and any constraints you already know:</p><blockquote><p>Build a local weather agent for the playground. Give it one tool for current conditions and add a smoke eval. Guide me through the remaining decisions.</p></blockquote><p>More detail means fewer questions. Include a channel, runtime, model, or required integration when those choices are fixed.</p><h2 id="which-choices-will-cursor-ask-me-to-make" tabindex="-1">Which choices will Cursor ask me to make? <a class="header-anchor" href="#which-choices-will-cursor-ask-me-to-make" aria-label="Permalink to &quot;Which choices will Cursor ask me to make?&quot;">​</a></h2><p>Cursor fills gaps in two short rounds:</p><ul><li><strong>Identity:</strong> purpose, project name, and location</li><li><strong>Runtime:</strong> local or cloud</li><li><strong>Model:</strong> the default model or another Cursor model</li><li><strong>Channels:</strong> playground and HTTP, Slack, GitHub, a webhook, or a schedule</li><li><strong>MCP connections:</strong> remote or local MCP servers</li><li><strong>Capabilities:</strong> tools, skills, subagents, hooks, seed files, approvals, and evals</li></ul><p>Questions adapt to your goal. A playground chat agent won&#39;t get cloud-repository questions. A local agent won&#39;t get cloud setup questions.</p><h2 id="what-happens-before-cursor-writes-files" tabindex="-1">What happens before Cursor writes files? <a class="header-anchor" href="#what-happens-before-cursor-writes-files" aria-label="Permalink to &quot;What happens before Cursor writes files?&quot;">​</a></h2><p>Cursor shows one plan with the choices it made and the folders it will create. Choose <strong>Scaffold it</strong> to continue or <strong>Adjust something</strong> to change the plan.</p><p>No files change before you approve this step.</p><h2 id="what-will-cursor-create" tabindex="-1">What will Cursor create? <a class="header-anchor" href="#what-will-cursor-create" aria-label="Permalink to &quot;What will Cursor create?&quot;">​</a></h2><p>A first version usually includes:</p><ul><li><code>AGENTS.md</code>, <code>.gitignore</code>, <code>package.json</code>, and <code>tsconfig.json</code></li><li><code>agent/agent.ts</code> for the model and runtime</li><li><code>agent/instructions.md</code> for the always-on prompt</li><li><code>agent/hooks/memory.ts</code> for memory guidance</li><li>One or two tools under <code>agent/tools/</code></li><li>Any channels or MCP connections you selected</li><li><code>evals/evals.config.ts</code> and one smoke eval</li></ul><p>Cursor keeps the first version small. Side-effecting server tools use <code>needsApproval</code>. Deterministic setup, such as fetching a pull request, stays in host code instead of model instructions.</p><p>See <a href="./reference/project-layout.html">Project layout</a> for every supported folder.</p><h2 id="how-does-cursor-verify-the-scaffold" tabindex="-1">How does Cursor verify the scaffold? <a class="header-anchor" href="#how-does-cursor-verify-the-scaffold" aria-label="Permalink to &quot;How does Cursor verify the scaffold?&quot;">​</a></h2><p>Cursor checks the project in this order:</p><ol><li>Run <code>agent-sdk validate</code> and fix every error</li><li>Inspect the discovered surface with <code>info --json</code></li><li>Call each server tool directly with validated sample input</li><li>Run one end-to-end model turn</li><li>Run the smoke eval</li><li>Run the project&#39;s TypeScript check</li></ol><p>Validation, discovery, direct server-tool calls, and server startup work without a Cursor credential. Model turns and evals need <code>CURSOR_API_KEY</code> or a saved <code>agent-sdk login</code>.</p><h2 id="what-happens-after-the-first-agent-works" tabindex="-1">What happens after the first agent works? <a class="header-anchor" href="#what-happens-after-the-first-agent-works" aria-label="Permalink to &quot;What happens after the first agent works?&quot;">​</a></h2><p>Choose one to three fixed inputs, define what should improve, and name what must stay unchanged. Then have Cursor follow <a href="./../skills/hillclimb/SKILL.html"><code>skills/hillclimb/SKILL.md</code></a>.</p><p>The hillclimb skill measures a baseline, changes one lever, runs the same inputs again, and adds an eval for each improvement you keep.</p><h2 id="related" tabindex="-1">Related <a class="header-anchor" href="#related" aria-label="Permalink to &quot;Related&quot;">​</a></h2><ul><li><a href="./quickstart.html">Build your first PR reviewer</a></li><li><a href="./guides/convert-automation.html">Convert a Cursor Automation</a></li><li><a href="./building-with-agents.html">Building agents with agents</a></li><li><a href="./evals.html">Evals</a></li><li><a href="./hillclimbing.html">Hillclimbing</a></li><li><a href="./reference/project-layout.html">Project layout</a></li></ul>',34)])])}const p=o(s,[["render",i]]);export{f as __pageData,p as default};