Find the users of the default#
A default has users you cannot see in the diff: callers that omit the parameter, configuration files that leave the key out, deployments that rely on the old value without knowing it. The comment is worth a code change when those users exist and the new value would surprise them. The checks are different from a breaking-change comment, because nothing stops compiling:
| What you find | Decision |
|---|---|
| The default moved as a side effect of a refactor, such as a new constructor or a reordered fallback | Change: restore the old value and add a test that asserts it |
| The new default is the point of the change, and every caller sets the option explicitly or lives in this repository | Keep it, and record the new default in the changelog |
| The new default is the point, and released users or downstream systems may rely on the old one | Make the new behavior opt-in, or ship it in a major release with a release note that says what to set to keep the old behavior |
| The bot asks you to change a default you kept on purpose | Apply the same questions to the current default, and answer with the evidence you used |
Two kinds of evidence help most. A search of your own code and configuration for explicit settings tells you who opted out of the default already. The release that introduced the old default tells you how long outside users have had to depend on it.
A default the maintainers declined to flip#
tRPC's httpBatchStreamLink sends trpc-accept: application/jsonl, while the server answered streamed batches with content-type: application/json. An earlier pull request proposed changing the response header to application/jsonl for everyone, and maintainers discussed it in March 2026. Pull request
A maintainer, Nick-Lucas, raised the default's invisible users in a review comment: "ChatGPT has got me worried that CDN et al might start buffering it because the MIME type isn't standard". Review comment Six days later, the same maintainer wrote: "I'm uncomfortable pushing this out to every user as a default when I have no evidence about its side effects and it works fine today", and "Like you say nothing is really "correct" here, so I'll vote for stability without significant evidence". Comment
The users at risk here were not callers of a function. They were proxies and clients between a tRPC server and its users, which no search of the tRPC repository can find. When the users of a default cannot be enumerated, the reasoning shifts to evidence about side effects, and this thread had none.
A bot that argued for flipping it#
A second pull request, opened in April 2026, took the opt-in route. It added a jsonl.contentType option to initTRPC.create() and kept application/json as the default. CodeRabbit's review objected: "Default still ships the original bug; confirm this is the intended resolution of #6819." It offered two ways forward: "flipping the default to 'application/jsonl' in a major release (and calling it out as breaking), or" "updating the PR title/description to make clear this is an opt-in configuration knob, not a default fix." Review
Both options were reasonable, and the merged pull request matches the second. Its title is "fix(server): make batch stream content-type configurable." In the merged diff, the server code falls back to the old value:
config.jsonl?.contentType ?? 'application/json',
The option's doc comment says @default 'application/json', the client documentation says "The default contentType remains application/json for backwards compatibility", and the existing streaming test now asserts that a server with no option set returns only application/json. A new parameterized test runs once per opt-in value. Pull request KATT approved and merged it on 19 October 2026 and closed the earlier proposal the same day: "The default stays application/json to avoid breaking existing proxies and clients." Closing comment
The bot was right that the original bug still ships by default. It did not know what the maintainers knew about the default's users, which the earlier thread recorded. Its second option, a plain statement that the new behavior is opt-in, is what the merged title and documentation say. The default itself stayed as it was.
Pin the default in a test#
Whichever way you decide, a default that matters should have a test that fails when it moves. tRPC's test reads the header from the response, not the constant, so a refactor that changes the fallback fails the test even if the constant's name stays the same. A test like that turns the next changed-default comment into a question CI answers. The reviewer can still ask whether the default should change, and the test makes any change to it deliberate.
When you keep a default the bot wanted changed, link the thread or decision that explains why, as tRPC's closing comment does. A later reader, human or bot, then sees the reason with the code. If you defer the new default to a major release, record it in an issue that names the release. If the bot repeats the comment on the next push after you answered it, the follow-through guide shows how to count that cost. If the default you change belongs to a released API, the breaking-change guide covers how to find its callers.