Files
carbon-lang/docs/project/pull_request_workflow.md
T
Chandler Carruth b7ea32c3a4 Support re-checking, overriding, and requiring the PR dependency check (#8120)
There are a world of reasons why the dependent state check can be wrong
or get stuck. We need ways to manually override inaccurate findings or
trigger a re-check when things change.

Overriding: removing the `dependent` label causes the job to run and
immediately report success to unblock merging. The `workflow_dispatch`
can also run with an `override` mode. This requires adding more tracked
state in the form of _ignored_ PRs that would otherwise be considered
dependencies.

Re-checking: the job now makes the status have a `Details` link that can
be used to re-run it, and editing the PR description also re-runs it.
However, adding the `dependent` label or doing an explicit
`workflow_dispatch` with mode `recheck` goes further: it drops any
overrides and recomputes the open dependencies from the current commits,
keeping the record of merged dependencies.

Hiding the comment still clears _all_ of the tracked state for the PR,
so the next check starts completely from scratch.

Lastly, this extends the workflow to support being marked as required by
letting it run in the merge queue and report success.

Works to keep the changes to `_process_pr` organized so that we first
collect and analyze the state, and only at the end mutate the PR's state
to the new configuration. Adds relevant unit tests for the new features.
Also added some basic documentation of this so that folks have a
reference when it has issues.

I considered the alternative of parsing mentions out of comments
(dependabot-style), but it ended up significantly more complicated and
brittle, so instead went with basing each change on an operation on the
PR that is already mechanically available.

Assisted-by: Antigravity with Opus
2026-11-13 05:19:44 +00:00

7.6 KiB

Trunk-based pull-request GitHub workflow

Table of contents

Overview

Carbon repositories follow a few basic principles:

  • Development directly on the trunk branch and revert to green.
  • Always use pull requests, rather than pushing directly.
  • Changes should be small, incremental, and review-optimized.
  • Preserve linear history by squashing pull requests rather than using unsquashed merge commits.

These principles try to optimize for several different uses or activities with version control:

  • Continuous integration and bisection to identify failures and revert to green.
  • Code review both at the time of commit and follow-up review after commit.
  • Understanding how things evolve over time, which can manifest in different ways:
    • When were things introduced?
    • How does the main branch and project evolve over time?
    • How was a bug or surprising thing introduced?

Note that this document focuses on the mechanical workflow and branch management. Details of the code review process are in their own document.

Trunk based development

We work in a simple trunk-based development model. This means all development activity takes place on a single common trunk branch in the repository (our default branch). We focus on small, incremental changes rather than feature branches or the "scaled" variations of this workflow.

Green tests

The trunk branch should always stay "green". That means that if tests fail or if we discover bugs or errors, we revert to a "green" state by default, where the failure or bug is no longer present. Fixing forward is fine if that will be comparably fast and efficient. The goal isn't to dogmatically avoid fixing forward, but to prioritize getting back to green quickly. We hope to eventually tool this through automatic continuous-integration powered submit queues, but even those can fail and the principle remains.

Always use pull requests (with review) rather than pushing directly

We want to ensure that changes to Carbon are always reviewed, and the simplest way to do this is to consistently follow a pull request workflow. Even if the change seems trivial, still go through a pull request -- it'll likely be trivial to review. Always wait for someone else to review your pull request rather than just merging it, even if you have permission to do so.

Our GitHub repositories are configured to require pull requests and review before they are merged, so this rule is enforced automatically.

Small, incremental changes

Developing in small, incremental changes improves code review time, continuous integration, and bisection. This means we typically squash pull requests into a single commit when landing. We use two fundamental guides for deciding how to split up pull requests:

  1. Ensure that each pull request builds and passes any tests cleanly when you request review and when it lands. This will ensure bisection and continuous integration can effectively process them.

  2. Without violating the first point, try to get each pull request to be "just right": not too big, not too small. You don't want to separate a pattern of tightly related changes into separate requests when they're easier to review as a set or batch, and you don't want to bundle unrelated changes together. Typically you should try to keep the pull request as small as you can without breaking apart tightly coupled changes. However, listen to your code reviewer if they ask to split things up or combine them.

While the default is to squash pull requests into a single commit, during the review you typically want to leave the development history undisturbed until the end so that comments on any particular increment aren't lost. We typically use the GitHub squash-and-merge functionality to land things.

Dependent pull requests

When splitting work into a series of changes that build on each other, open each pull request against trunk with the commits from earlier unmerged pull requests included in its branch.

The PR dependencies check status tracks these dependencies based on shared commits:

  • While a pull request depends on other open pull requests, the check adds the dependent label, posts a comment linking to the diff of just that pull request's commits, and leaves its status pending so it doesn't merge before its dependencies.
  • Once those pull requests merge, the check removes the dependent label, updates the comment, and marks the status success.

If the check becomes stale, editing the pull request description runs it again. Contributors with write access can also re-run the workflow job from the check's Details link, or use the dependent label:

  • Re-check: Adding the label drops any override and recomputes the open dependencies from the current commits. If the label is already present, remove and re-add it.
  • Override: Removing the label ignores the currently detected dependencies and marks the status success.

Both are also available by manually running the Check Dependent PRs workflow with the recheck or override mode. Hiding the check's comment discards all of its tracked state, including merged dependencies, so the next check starts from scratch.

Linear history

We want the history of the trunk branch of each repository to be as simple and easy to understand as possible. While Git has strong support for managing complex history and merge patterns, we find understanding and reasoning about the history -- especially for humans -- to be at least somewhat simplified by sticking to a linear progression. As a consequence, pull requests are squashed when merging them.

Merging a pull request

Once approved, and even if all checks have not finished running yet, a pull request can be added to a merge queue. When queued, and after all pull request checks pass, this will automatically create a squashed version of the commits in a temporary branch on top of trunk and any previously queued merge. Checks run on this commit help to ensure that the trunk stays "green" in presence of conflicting merges.

After the merge queue checks pass, the trunk branch pointer is updated to include the now merged changeset. The resulting commit title and description will match exactly those of the pull request, so it is important to set those appropriately before merging. Co-authors are preserved by this operation. If a failure happens at any point, the merge fails, with both trunk and the pull request branch kept in their original state.