Getting started
The manual. One page per feature, and each page is the whole feature: a build-it-yourself walkthrough first, then every key, every rule, and a complete example — for a person reading top to bottom and for a model fetching one page to author a file.
In order:
- Install — the extension, and the one prerequisite that is not obvious.
- Agent SDKs — GHCP,
the default provider for AI tasks and specs, and what an agent
session is given.
claudeis the other SDK provider; eight more come in over ACP (item 11). Everything that is not an AI task works without any of them. - AI chat — the Chat sidebar and the chat editor tab: one conversation that belongs to no spec and no workflow.
- Tasks —
tasks.yml, a library of commands. Start here: values, templating and anchors are covered once, on this page, for both formats. - Workflows — one
pipeline: stages, tasks in order, an AI draft, a human gate, and
one task per item with
foreach:. - Variables — every
built-in variable on one page: the anchors,
params,vars,envandrun.context, and the one resolution order. - Spec-driven development — requirements, design, plan, implementation, with a decision between each. The shipped styles, their exact prompts, and writing your own.
- Plugins — the last, smallest thing that executes. The eight that ship, and writing your own in plain JavaScript.
package.jsonscripts — your existing npm scripts, listed beside everything else.settings.json— the engine’s one configuration file.- Agent Client Protocol —
adding an agent this build has no dedicated provider for, like Kiro,
as
provider:on anai@1task. - Workspaces — where
“here” is: folder, workspace and profile scopes, and what changes
when you open a
.code-workspace. - Steering files — your project’s rules for the AI: which ones every chat gets, and which ones the AI loads by topic only when it needs them.
- Plugin apps —
giving a plugin its own page in the run and spec panels: every file,
the whole
lwAPI, and what the page cannot do. - Evaluator — running one eval set on several runtimes and models at once, and comparing what each one scored, how long it took and what it used.