CI for coding agents

An agent's mistake in a KiCI pipeline is a type error before the push, not a red pipeline ten minutes later. The pipeline is TypeScript, so the compiler reads it. The agent reads the compiler, fixes the line, and runs the workflow on your machine — all before anything leaves the laptop.

The loop closes on your machine

Here is a workflow with one wrong key. The agent wrote runOn; the SDK spells it runsOn. The compiler says so:

// .kici/workflows/site-demo.ts
import { workflow, job, step, push } from '@kici-dev/sdk';

const test = step('test', async ({ $ }) => {
  await $`pnpm test`;
});

export default workflow('site-demo', {
  on: [push({ branches: 'main' })],
  jobs: [job('test', { runOn: 'kici:os:linux', steps: [test] })],
});

site-demo.ts(10,24): error TS2769: No overload matches this call.
  The last overload gave the following error.
    Object literal may only specify known properties, but 'runOn' does not exist in type 'JobOptions'. Did you mean to write 'runsOn'?

Nothing was pushed to produce that error. These commands take an agent from a blank repository to a green run:

1. kici compile — shows the error.
2. Fix it.
3. kici run remote — runs the working tree on your agents with test-scoped secrets, and streams the logs back.
4. Push.

kici run --local runs the same workflow on your laptop. Start here, or work offline.

Here is that same workflow, fixed, running on real agents — with one test that exists only in the working tree.

$ git status --short
 M test/ci.test.js

$ kici run remote --workflow site-demo
kici v0.8.0
✓ Compiled workflows → .kici/kici.lock.json (1 workflow)
Types generated ~/site-demo/.kici/types/secrets.d.ts
Running workflow "site-demo" directly (bypassing triggers)
Creating overlay tarball...
Running your local working tree (overlay includes .git, so git steps work)
55 files changed, 0 new, 0 deleted (46.2 KB compressed)
Initializing upload...
Uploading overlay...
Run started: 00000000-0000-0000-0000-000000000000
> test
> node --test
✔ the committed test passes (1.195345ms)
✔ this test is not committed, and it still ran on the agent (0.333915ms)
ℹ tests 2
ℹ suites 0
ℹ pass 2
ℹ fail 0
ℹ cancelled 0
ℹ skipped 0
ℹ todo 0
ℹ duration_ms 106.56215
npm notice
npm notice New major version of npm available! 11.19.1 -> 12.0.2
npm notice Changelog: https://github.com/npm/cli/releases/tag/v12.0.2
npm notice To update run: npm install -g npm@12.0.2
npm notice

┌──────┬────────┬──────────┐
│ Job  │ Status │ Duration │
├──────┼────────┼──────────┤
│ test │ ✓ pass │ 5.5s     │
└──────┴────────┴──────────┘

Result: PASSED (7.9s)

Brief your agent before it writes a line

The bundles below are task-scoped. Fetch the one the task needs instead of the whole corpus. This table is read from docs.kici.dev/llms.txt when the site is built, so it cannot drift from what the docs publish.

Offline, kici docs llm sdk prints the same bundle from the installed package.

Let the agent drive the real pipeline

Set up the MCP server →

What this is not

KiCI has no model in the loop. It does not write a pipeline for you, it does not read your repository, and it holds no key to any AI provider. The agent you already run writes the TypeScript on your machine. KiCI gives that agent types to write against, a compiler to answer it, and a local run to prove the answer.

Questions