Always format JSON via Grape::Json.dump

[ericproulx]

Closes #2486.

Summary

Grape::Formatter::Json and Grape::Formatter::SerializableHash short-circuited to object.to_json whenever the object responded to it. Since virtually every object responds to to_json, Grape::Json.dump was effectively never reached during formatting — so a configured multi_json back-end (e.g. oj) was silently bypassed, inconsistent with Grape::ErrorFormatter::Json and Grape::Parser::Json, which already go through Grape::Json.

This removes the to_json short-circuit from both formatters so JSON is always serialized through Grape::Json.dump, honoring the configured back-end consistently.

Changes

• lib/grape/formatter/json.rb: drop return object.to_json if object.respond_to?(:to_json).
• lib/grape/formatter/serializable_hash.rb: drop the same to_json fallback.
• README.md: update the :json formatter description.
• UPGRADING.md: document the behavior change (breaking).

🤖 Generated with Claude Code

[github-actions]

Danger Report

No issues found.

View run

[dblock]

Test failures look legit.

[ericproulx]

Closing this — I'm keeping the to_json short-circuit.

The premise was that object.to_json and Grape::Json.dump(object) are equivalent and the short-circuit only had the side effect of bypassing the configured multi_json back-end. That's not the case: they produce different output, so this isn't a no-op consistency fix — it's a silent serialization change.

to_json routes through each object's own JSON representation (ActiveSupport's as_json), while Grape::Json.dump (→ JSON.generate) only handles JSON primitives natively and falls back to to_s for everything else. Removing the short-circuit would change the output for Time, Date, BigDecimal, symbols, and any object defining as_json/to_json — and can even flip the string encoding.

Concretely, for a Time:

t = Time.at(0).getlocal("+00:00")

t.to_json            # => "1970-01-01T00:00:00.000+00:00"  (ISO 8601, honors as_json)
Grape::Json.dump(t)  # => "1970-01-01 00:00:00 +0000"      (Time#to_s fallback)

That's a behavior regression for anyone relying on as_json/ISO 8601 output (which is the common case with ActiveSupport loaded), and a far wider contract break than "honor the configured back-end." The consistency gain with ErrorFormatter::Json / Parser::Json doesn't justify regressing established serialization semantics, so I'd rather leave the formatters as-is.

#2486 should be revisited with an approach that preserves as_json semantics if we still want the back-end to apply in the formatter path.