Skip to main content
|

Overview

Every starting point produces the same thing: a coding task bound to a repository and to either a pull request or a branch. What differs is the context the task starts with, which providers offer it, and whether the first turn is free. Pick the starting point that matches where you already are. If you’re looking at a pull request, start from the pull request. If you’re starting something new, use the app.
If your repository needs a specific toolchain, private dependencies, or a build step, set up a coding environment first. Without one, the sandbox is prepared from what CodeRabbit can detect on its own.

Ways to start a task

From the Coding Agent app

Go to app.coderabbit.ai/code and write a request in the composer. Use it when the work isn’t already attached to a pull request or a CodeRabbit finding. Choose the repository and the branch or pull request the task should work against, then compose the request. The composer accepts:
  • File mentions — reference specific files so the agent starts with them in context.
  • Attachments — images, logs, specs, and other supporting files. See Attachments.
  • Skills — choose a reusable procedure from the catalog or your library. See Coding Agent skills.
  • Voice input — dictate the request instead of typing it.
  • Slash commands — see the table below.
A request is capped at 65,536 characters.
The composer's slash-command menu, showing /plan, /upload, and /private

The composer's slash-command menu, showing /plan, /upload, and /private

From a Linear or Jira issue

Under Start with on the new-task page, choose Linear issue or Jira issue. Search your connected tracker and select the issue to include as context, then choose the repository and branch and describe what you want implemented. If the tracker is not connected, the app offers a connection flow where your permissions allow it. See Issue trackers for setup. You can start from issue context without first creating an Issue Planner plan.

From CodeRabbit review findings

On a pull request, select the CodeRabbit review findings you want addressed and start one task for all of them. This is Autofix. The task is bound to the pull request and works on its branch.
Requesting Autofix on a CodeRabbit review finding by replying @coderabbitai autofix

Requesting Autofix on a CodeRabbit review finding by replying @coderabbitai autofix

From a finishing touch

Finishing Touches are one-click actions on a pull request. Each one starts a coding task with a fixed goal: Which of these are offered on a given pull request depends on your platform, your plan, and your configuration. The CodeRabbit walkthrough comment on the pull request lists the ones actually available for that review — see Finishing Touches.
The Finishing Touches menu, expanding Fix failing CI checks to choose delivery as a stacked PR or a commit on the current branch

The Finishing Touches menu, expanding Fix failing CI checks to choose delivery as a stacked PR or a commit on the current branch

From a merge conflict

A pull request with merge conflicts can hand the reconciliation to a coding task, which works out the resolution against the current state of the target branch and pushes the reconciled branch. See Resolve merge conflicts. Automated conflict handling covers up to 500 conflicted files.

From a security finding

CodeRabbit Security’s Fix with AI starts a coding task carrying the finding and its fix instructions. See CodeRabbit Security.
The IDE extension’s identically named Fix with AI is a different feature and creates no coding task. See Names that collide.

From CodeRabbit’s chat

When you ask CodeRabbit’s chat for a code change, it can hand the request into a coding task bound to the repository you were discussing. From there the task behaves like any other: watch it, steer it, and choose how its changes land. See Deliver a task’s changes.

From Change Stack

Change Stack can start a task from the review panel without leaving the pull request, and it is the one place where you can collect several review findings into a single task before dispatching it. The run it produces is an ordinary coding task.

From an approved plan

When an implementation plan has been written and approved, you can switch the task to code mode and have Coding Agent implement exactly the approved plan. See Review a plan.

Where each starting point is available

Coding Agent’s task contract covers GitHub, GitLab, Azure DevOps, and Bitbucket Cloud, but individual starting points are narrower than that. Check the table below rather than reading support off the badges. Each finishing touch states its own platform and plan support on its own page. Two things this table does not tell you. Provider support is not the same as being switched on: a row saying GitHub means the starting point works on GitHub, not that your organization has it enabled yet. And these rows describe CodeRabbit’s cloud service — support on self-hosted Git providers and self-hosted CodeRabbit deployments varies and isn’t covered here, so contact support before planning around it.

Attachments

A request can carry supporting files. Attachments are bounded:
  • At most 10 attachments per message.
  • At most 26,214,400 bytes (25 MiB) per attachment.
  • Stored attachments are retained for 30 days.
Accepted content types:

What does not start a coding task

These look like they should, and don’t:
  • The IDE extension’s Fix with AI. It routes the fix to your own local agent or your clipboard.
  • A file in your repository, by itself. Committing an instructions file does not start a task. Once a task runs, it can use repository instructions and relevant skills alongside its coding environment.

When task creation fails

Creating a task runs several checks first — authorization, billing, repository access, and dispatch. Any of them can fail before a durable task exists, which means there is no task page to open and nothing to retry from. If that happens, retry from wherever you started: the composer, the pull request, or the finding. Nothing was left half-created.

What’s next

Work with a running task

Follow the work, steer it, answer its questions, and review its plan or patch.

Deliver a task's changes

Choose how changes land, and understand who pushes them.

Coding Agent skills

Choose a reusable procedure or create a skill for your team.