Testing

The suite is one kaocha suite, :unit (tests.edn), over test/ (dvergr) and benchmarks/test/ (the benchmark providers). ^:integration tests (real provider calls) and ^:security-open tests (sandbox escapes still open) are skipped by default.

The MCP relay (bin/dvergr-mcp) is a babashka script, and its tests in dvergr.mcp.smoke-test skip (with a note) when bb is not on PATH; CI installs babashka 1.12.200 so they run there. Install bb to run them locally.

Running tests locally

clojure -M:test                                    # everything
clojure -M:test --focus dvergr.jobs-test           # one namespace
clojure -M:test --focus dvergr.jobs-test/cancelling-a-job-settles-its-run-cancelled
clojure -M:test --focus-meta :integration          # the provider-calling tests
clojure -M:test --seed 1713321041                  # repeat a run's test order

Every run prints how long loading (compiling) the test namespaces took and how long the tests ran (dvergr.test-phases), and the profiling plugin lists the 40 slowest namespaces and tests. Note that --focus filters after loading: kaocha still compiles every test namespace first.

To run only some namespaces, the way one CI container does, put them in a file and point the split hook at it. Only those namespaces are loaded:

.circleci/test-namespaces.sh > /tmp/all.txt
printf 'dvergr.jobs-test\ndvergr.effects-test\n' > /tmp/mine.txt
DVERGR_TEST_ALL_NAMESPACES=/tmp/all.txt DVERGR_TEST_NAMESPACES=/tmp/mine.txt clojure -M:test

How CI splits the suite

The CircleCI test job runs with parallelism: 3. Each container:

1. lists every test namespace (.circleci/test-namespaces.sh, from the file names under test/ and benchmarks/test/);
2. takes its share with circleci tests split --split-by=timings --timings-type=classname, which balances the shares by the namespace timings of earlier runs (the JUnit results each container stores);
3. runs kaocha on that share: dvergr.test-split, a kaocha pre-load hook, narrows the suite's :ns-patterns to exactly those namespaces, so a container compiles only what it runs.

No test is lost or run twice: tests split puts each listed namespace in exactly one container, and the hook refuses to run unless the list equals the set of namespaces kaocha itself finds in the suite. A namespace the file-name derivation misses, or a stale name, fails every container before any test runs. The containers' summary lines add up to the full run's test count.

The three containers report as one check, ci/circleci: test.

Writing tests that stay green

• Wait on the condition, not on a fixed time. A wait for something that must happen (a turn starting, a reply, a Run settling) gets a generous bound (llm-test's wait-ms is 30 s), since a passing test never waits it out and a cold CI JVM can be slow. A check that something does not happen watches a short window that starts after the event that would have caused it has settled.
• A stochastic assertion needs a stated error bound: fix the seed, or size the sample so the tolerance is many standard deviations (see dvergr.sandbox.steer-test).
• Process-wide registries (tools, MCP tools, channels) are shared by every test in the JVM. Undo only what the test changed (dvergr.test-registries): restoring a whole snapshot drops what namespaces loaded during the test registered, and a later test then fails depending on order.

The REPL meta-harness contract

Before asking an LLM to invent an agentic workflow, dvergr verifies that the programming surface it will use can construct and observe one without a model.

The contract executes the real clojure_eval tool against a room-bound SCI context and ChatContext. Sandboxed Clojure creates a persistent child room through dvergr.room/create!; a second evaluation proves that the REPL definition and room view persisted; the host then checks the shared room registry and Datahike registry row.

Run the test directly:

clojure -M:test --focus dvergr.repl-meta-harness-test

Or invoke the same test fixture interactively:

;; Start with: clojure -M:dev
(require '[dvergr.repl-meta-harness-test :as contract])

(def result (contract/run-contract!))
(:passed? result)
;; => true

(:checks result)

run-contract! uses a fresh state root under the JVM temporary directory, disables file-driven agent autostart, stops its daemon, and restores the prior dvergr state root. It refuses to run if a daemon is already active in the JVM. No model provider is resolved or invoked.

To exercise an already running development daemon deliberately, use a unique child slug. This retains the child for inspection through the REPL, TUI, or web UI:

(require '[dvergr.repl-meta-harness-test :as contract])
(require '[dvergr.orchestration.daemon :as daemon])

(contract/run-against-daemon!
 @daemon/current-daemon
 {:parent-slug "boardroom"
  :child-slug "my-meta-harness-probe"})