refactor(opsctl): update main and tasks

cmd/opsctl/main.go

@@ -125,6 +125,7 @@ func defaultRunEnvKeys() []string {
 		"DRAGONSCALE_EVAL_BASE_CONFIG",
 		"DRAGONSCALE_EVAL_CONFIG",
 		"DRAGONSCALE_EVAL_DEBUG",
+		"DRAGONSCALE_PROMPTFOO_ARGS",
 		"DEVCONTAINER_EXEC",
 		"PLATFORM",
 		"ARCH",

cmd/opsctl/main_test.go

@@ -136,6 +136,13 @@ func TestRunHelpIncludesGlobalOutputFlags(t *testing.T) {
 	require.Contains(t, out.String(), "--timeout")
 }
 
+func TestDefaultRunEnvKeysIncludesPromptfooArgs(t *testing.T) {
+	t.Parallel()
+
+	keys := defaultRunEnvKeys()
+	require.Contains(t, keys, "DRAGONSCALE_PROMPTFOO_ARGS")
+}
+
 func TestMakefilePhonyTargetsAreAvailableInOpsctl(t *testing.T) {
 	makeTargets := parseMakefilePhonyTargets(t)
 	out := &bytes.Buffer{}

internal/opsctl/tasks/tasks.go

@@ -302,7 +302,7 @@ func NewRegistry(_ string) []app.Task {
 		NewCommandTask("eval-fixtures", "Prepare eval fixture workspace", evalFixturesSpecs, nil, nil),
 		NewCommandTask("eval-view", "Open the promptfoo results viewer", evalViewSpecs, nil, nil),
 		NewShellTask("eval-clean", "Cleanup eval artifacts", simpleScript("rm -rf eval/results eval/bin"), nil),
-		NewShellTask("eval-compare", "Run A/B comparison of current branch vs main", simpleScript("cd eval && DEVCONTAINER_EXEC= EVAL_NPM_CMD=$(npx --yes) ./scripts/compare.sh --repeat 3"), nil),
+		NewShellTask("eval-compare", "Run A/B comparison of current branch vs main", simpleScript("cd eval && DEVCONTAINER_EXEC= EVAL_NPM_CMD=\"npx --yes\" ./scripts/compare.sh --repeat 3"), nil),
 		NewShellTask("eval-test", "Run Go-native component evals", staticGoScript("-v ./eval/go_evals/..."), nil),
 	}
 	return tasks
@@ -332,6 +332,7 @@ func defaultEnv(c *app.Context) []string {
 	appendIfSet("DRAGONSCALE_EVAL_BASE_CONFIG", cEnv(c, "DRAGONSCALE_EVAL_BASE_CONFIG", ""))
 	appendIfSet("DRAGONSCALE_EVAL_CONFIG", cEnv(c, "DRAGONSCALE_EVAL_CONFIG", ""))
 	appendIfSet("DRAGONSCALE_EVAL_DEBUG", cEnv(c, "DRAGONSCALE_EVAL_DEBUG", ""))
+	appendIfSet("DRAGONSCALE_PROMPTFOO_ARGS", cEnv(c, "DRAGONSCALE_PROMPTFOO_ARGS", ""))
 	appendIfSet("VERSION", cEnv(c, "VERSION", ""))
 	appendIfSet("FANTASY_VERSION", cEnv(c, "FANTASY_VERSION", ""))
 	appendIfSet("NAME", cEnv(c, "NAME", ""))

internal/opsctl/tasks/tasks_test.go

@@ -238,6 +238,25 @@ func TestEvalRunSpecsUsesBaseConfigWhenSetAndDebugEnabled(t *testing.T) {
 	require.Equal(t, []string{"--yes", "promptfoo", "eval", "--config", "promptfooconfig.yaml", "--no-cache", "--no-progress-bar"}, specs[2].Args)
 }
 
+func TestEvalRunSpecsUsesPromptfooArgsOverride(t *testing.T) {
+	t.Parallel()
+
+	specs := evalRunSpecs(&app.Context{
+		Root: t.TempDir(),
+		ExtraEnv: map[string]string{
+			"DRAGONSCALE_PROMPTFOO_ARGS": "--no-cache --max-concurrency 1",
+		},
+	})
+
+	require.NotEmpty(t, specs)
+	last := specs[len(specs)-1]
+	require.Equal(t, "npx", last.Name)
+	require.Equal(t, []string{
+		"--yes", "promptfoo", "eval", "--config", "promptfooconfig.yaml",
+		"--no-cache", "--max-concurrency", "1",
+	}, last.Args)
+}
+
 func TestEvalCompareTaskDisablesNestedDevcontainerExecution(t *testing.T) {
 	t.Parallel()
 
@@ -265,7 +284,7 @@ func TestEvalCompareTaskDisablesNestedDevcontainerExecution(t *testing.T) {
 	require.NoError(t, err)
 	require.Len(t, fake.Calls, 1)
 	script := strings.Join(fake.Calls[0].Args, " ")
-	require.Contains(t, script, "cd eval && DEVCONTAINER_EXEC= EVAL_NPM_CMD=$(npx --yes) ./scripts/compare.sh --repeat 3")
+	require.Contains(t, script, "cd eval && DEVCONTAINER_EXEC= EVAL_NPM_CMD=\"npx --yes\" ./scripts/compare.sh --repeat 3")
 	joinedEnv := strings.Join(fake.Calls[0].Env, " ")
 	require.Contains(t, joinedEnv, "DEVCONTAINER_EXEC=npx --yes @devcontainers/cli exec --workspace-folder \"$PWD\" --")