Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/goulinkh/code-review-harness/llms.txt

Use this file to discover all available pages before exploring further.

CRH exposes a full TypeScript API so you can embed AI code reviews directly into your own pipelines without using the CLI. You import a provider (where to read the merge proposal from), a sink (where to send the review), and createReviewSession from @code-review-harness/core, then drive the session with a single prompt.

​
Install packages

npm install @code-review-harness/core @code-review-harness/launchpad-provider @code-review-harness/stdout-sink

​
Import the API

import { createReviewSession } from "@code-review-harness/core";
import { createLaunchpadProvider } from "@code-review-harness/launchpad-provider";
import { createStdoutSink } from "@code-review-harness/stdout-sink";

​
Minimal example

The following script (from examples/launchpad-mp.ts) is all you need to run a full review and print the result to stdout: 
const provider = createLaunchpadProvider({ url: process.argv[2] ?? "" });
const sink = createStdoutSink();
const { session } = await createReviewSession({ provider, sink });
session.subscribe((event) => process.stderr.write(`${JSON.stringify(event)}\n`));
await session.prompt("Review merge proposal. Submit final review with submit_review.");
Run it with: 
node --import tsx your-script.ts "https://api.launchpad.net/devel/~user/+git/repo/+merge/123"

​
createReviewSession options

createReviewSession accepts a CreateReviewSessionOptions object and returns a Promise<ReviewSessionHandle>.

​
CreateReviewSessionOptions

FieldTypeRequiredDescription
providerReviewProviderYesSource of merge proposal data (diffs, comments, CI).
sinkReviewSinkYesDestination for the finished review.
workspacestringNoPath to use as the agent workspace. A temp directory is created automatically when omitted.
gitdirstringNoPath to use as the bare git clone directory. A temp directory is created automatically when omitted.
modelModel<any>NoOverride the default LLM model.
systemPromptstringNoOverride the orchestrator agent’s system prompt.
subAgentSystemPromptstringNoOverride sub-agent system prompts used by delegate_review.
workspaceProgress(message: string) => voidNoCallback invoked with progress messages while the workspace is being prepared.
onChildEvent(event: AgentSessionEvent, ctx: DelegateChildContext) => voidNoCallback for events emitted by sub-agents spawned by delegate_review.

​
ReviewSessionHandle

createReviewSession resolves to an object with these fields:
FieldTypeDescription
sessionAgentSessionThe live agent session. Call session.prompt() to send a message and session.subscribe() to observe events.
workspacestringAbsolute path to the prepared workspace directory.
gitdirstringAbsolute path to the bare git clone directory.
changedFilesstring[]List of file paths changed in the latest preview diff.

​
Subscribing to session events

Call session.subscribe(listener) before session.prompt() to receive a stream of AgentSessionEvent objects as the agent works: 
session.subscribe((event) => {
  // event contains tool calls, assistant messages, errors, etc.
  process.stderr.write(`${JSON.stringify(event)}\n`);
});

await session.prompt("Review merge proposal. Submit final review with submit_review.");
session.prompt() resolves when the agent has finished the turn (including calling submit_review).
For CLI users, pass --debug to print the same event stream to stderr without writing any code. For programmatic users, subscribing to events is the equivalent way to get full visibility into what the agent is doing.

Build docs developers (and LLMs) love

Get started for freeTalk to us