Run a project’s tests

Goal: run a tree’s own test suite on a runner. Same pipeline as a build, no artifact, an honest verdict.

Basic usage

$ saggar submit ./my-project --test
$ saggar submit ./my-project --test-cmd "make check"
$ saggar submit ./my-project --test --arch arm64 --env trixie

--test runs the tree’s own test suite instead of building: same pipeline, same target overrides (--arch, --env), same live logs, same failure metadata. The suite’s exit code is the verdict.

--test-cmd overrides detection with an explicit command (API: "test_cmd", which implies "mode": "test").

What gets autodetected

Per build system: cargo test, go test ./..., ctest and meson test (after the build those runners need), a make -n probe for a check or test target, npm run test / pnpm / bun when the manifest declares the script, deno task test, mix test (with a test/ directory), pytest through the project’s tool, swift test (with a Tests/ directory), gradle and maven (with a real suite), and the rake convention through bundler.

Honesty rules

• A tree with no detectable suite fails the job, never a green no-op, naming what was probed; the answer is an explicit --test-cmd.

• Green means success with nothing collected: a test job produces no artifact and is never registry material.

• Red is a real failure in the test phase. Infra noise still retries silently; a red test never does.

Structured results

When the suite’s runner writes report files, saggar parses them and serves what failed; no log scraping:

$ curl -H "X-Auth-Token: $SAGGAR_TOKEN" \
    https://saggar.dev/api/v1/jobs/j-…/results

The results field on the job object carries per-file suite/case counts and the failing case names with their failure messages. Each build system declares where its own runner drops reports: surefire XML for maven, the Test task’s XML for gradle, nextest’s conventional target/nextest/*/junit.xml for cargo (plain cargo test writes no report; the project’s nextest profile opts in), and meson’s testlog.json.

Results are collected green or red (a failing suite’s report is the most valuable one), but they are a bonus layer: a malformed report is skipped with one log line, absence is normal, and the verdict stays the exit code.

Tests on every push

By default a watch builds the tree and also runs its tests: a triggered push produces the formats×arches group plus a test member sharing the group id. See Set up CI: watched repos, webhooks, statuses. A watch pinned to a formats list cannot also be tests-only; the group composition is how both coexist.