Feature request: AI agent reporter mode optimized for token-efficient output

[Jayllyz]

Motivation

Reduce token usage when executing by an agent.

Proposal

Add a --reporter agent mode with the following behavior:

• Suppress all passing and skipped test output
• Emit each failure immediately as it occurs (no buffering)
• Final summary line only: nextest: 3 failed, 47 passed
• No ANSI codes, no progress bars

Other test runner such as vitest use https://github.com/unjs/std-env#agent-detection to automatically enable agent reporter mode, nextest could do something similar.

Alternatives

No response

Additional context

Vitest implemented this recently (jest too): vitest.dev/guide/reporters.html#agent-reporter.

[sunshowers]

I'm okay with this in principle, but someone has to be in charge of this project in a rigorous, data-driven manner. Whoever is interested should start off by writing a design document with:

• How many tokens do current versions of nextest consume with typical output?
• Nextest already exposes knobs like --status-level: how do those affect things?
• Is --cargo-quiet necessary? Cargo produces a lot of build info.
• Does the fact that tests are no longer listed out cause the LLM to make worse decisions? (For example, accidentally not running a test and the LLM not realizing it until it's too late). (This might be mitigated by the fact that nextest exits with a non-zero code if no tests are run.)
• LLMs (at least Opus 4.6) know how to drive nextest, but will not know about a new flag unless told to. Does that mean that nextest will have to detect the environment it's in? What are the tradeoffs of this approach? (With nextest we generally hesitate to change behavior based on the environment, but there's a stronger case for reporter output being changed.)
• What should the evaluation strategy be?
• What other things need to be covered? What alternatives exist?

See https://nexte.st/docs/design/architecture/recording-runs/ for an example of the level of rigor the design document should aspire to reach. (I wouldn't focus too much on organizational or grammatical correctness -- I'm more interested in things like the principles, data, etc)

To address your points:

Suppress all passing and skipped test output

This can be tested via --status-level fail.

Emit each failure immediately as it occurs (no buffering)

This is nextest's default behavior.

Final summary line only: nextest: 3 failed, 47 passed

Nextest already outputs this kind of summary line.

No ANSI codes, no progress bars

This is already the case with things like Claude Code which don't spawn a tty. I think some coding harnesses do spawn a tty, in which case I feel like I'd put this on them and not on nextest.