iop/apps/edge/internal/openai/single_request_preset_binding_test.go
toki d7a150c7fe feat(agent): 단일 요청 실행 경로를 완성한다
Claude의 단일 Anthropic 요청 안에서 IOP가 Plan, Work, Review와 workspace 도구 실행을 끝내고 실제 dev smoke로 계약을 검증할 수 있어야 한다.\n\n완료 task evidence와 마일스톤 검토 상태도 같은 변경에 고정한다.
2026-08-08 23:35:13 +09:00

556 lines
22 KiB
Go

package openai
import (
"errors"
"testing"
"iop/apps/edge/internal/authprojection"
edgeservice "iop/apps/edge/internal/service"
"iop/packages/go/config"
)
func newTestView(principalRef string, routes []authprojection.Route) authprojection.AuthenticatedView {
return authprojection.AuthenticatedView{
Principal: authprojection.Principal{
PrincipalRef: principalRef,
},
Routes: routes,
}
}
func managedBinding(modelGroupKey, providerID, principalRef, routeID string, managed bool) routeDispatch {
return managedBindingFull(modelGroupKey, providerID, principalRef, routeID, "profile-1", managed)
}
func managedBindingFull(modelGroupKey, providerID, principalRef, routeID, profileID string, managed bool) routeDispatch {
return routeDispatch{
Managed: managed,
ModelGroupKey: modelGroupKey,
ProviderID: providerID,
PrincipalRef: principalRef,
RouteID: routeID,
ProfileID: profileID,
CredentialSlotRef: "slot-1",
CredentialRevision: 1,
RouteRevision: 1,
}
}
// validSingleRequestPreset returns an approved fixed single-request preset whose
// selector, allowed modes, and light route exactly match the frozen
// plan→work→review policy stages: high reasoning on plan/review, none on work,
// and the selector fused to the plan stage. Each call builds fresh option maps so
// subtests may mutate one aspect in isolation.
func validSingleRequestPreset() config.ExecutionPreset {
return config.ExecutionPreset{
ID: "preset-single-request",
Selector: config.ExecutionModelBinding{
Model: "plan-model",
Options: map[string]any{"reasoning_effort": "high"},
},
AllowedModes: []string{config.ModeLight},
Routes: map[string]config.ExecutionRoute{
config.ModeLight: {
Stages: []config.ExecutionRouteStage{
{Role: "plan", Model: "plan-model", Options: map[string]any{"reasoning_effort": "high"}},
{Role: "work", Model: "work-model"},
{Role: "review", Model: "review-model", Options: map[string]any{"reasoning_effort": "high"}},
},
},
},
SingleRequest: &config.ExecutionSingleRequestPolicy{
WorkspaceRef: "ws-opaque-ref",
Limits: config.ExecutionSingleRequestLimits{
WallClockMS: 30 * 60 * 1000,
StageTimeoutMS: 10 * 60 * 1000,
MaxToolIterations: 64,
MaxOutputBytes: 16 * 1024 * 1024,
},
Stages: config.ExecutionSingleRequestStages{
Plan: config.ExecutionSingleRequestStageConfig{Model: "plan-model", Options: map[string]any{"reasoning_effort": "high"}},
Work: config.ExecutionSingleRequestStageConfig{Model: "work-model"},
Review: config.ExecutionSingleRequestStageConfig{Model: "review-model", Options: map[string]any{"reasoning_effort": "high"}},
},
},
}
}
// validSingleRequestBindings returns managed, same-principal canonical bindings
// for the plan/work/review models referenced by validSingleRequestPreset.
func validSingleRequestBindings() map[string]routeDispatch {
return map[string]routeDispatch{
"plan-model": managedBinding("plan-model", "prov-1", "principal-1", "route-plan", true),
"work-model": managedBinding("work-model", "prov-1", "principal-1", "route-work", true),
"review-model": managedBinding("review-model", "prov-1", "principal-1", "route-review", true),
}
}
func TestSingleRequestPresetBindingManaged(t *testing.T) {
preset := validSingleRequestPreset()
bindings := validSingleRequestBindings()
view := newTestView("principal-1", nil)
binding, err := compileSingleRequestBinding("virtual-public-model", preset, bindings, view)
if err != nil {
t.Fatalf("managed compilation failed: %v", err)
}
if binding == nil {
t.Fatal("expected non-nil binding")
}
if binding.PublicModel != "virtual-public-model" {
t.Errorf("PublicModel=%q, want virtual-public-model", binding.PublicModel)
}
if binding.WorkspaceRef != "ws-opaque-ref" {
t.Errorf("WorkspaceRef=%q, want ws-opaque-ref", binding.WorkspaceRef)
}
if binding.Plan.Model != "plan-model" {
t.Errorf("Plan.Model=%q, want plan-model", binding.Plan.Model)
}
if binding.Work.Model != "work-model" {
t.Errorf("Work.Model=%q, want work-model", binding.Work.Model)
}
if binding.Review.Model != "review-model" {
t.Errorf("Review.Model=%q, want review-model", binding.Review.Model)
}
if binding.Limits.WallClockMS != 30*60*1000 {
t.Errorf("WallClockMS=%d, want 1800000", binding.Limits.WallClockMS)
}
// Approved fixed options survive admission: high reasoning on plan/review, and
// the work stage carries no reasoning option.
if binding.Plan.Options["reasoning_effort"] != "high" {
t.Errorf("Plan.Options[reasoning_effort]=%v, want high", binding.Plan.Options["reasoning_effort"])
}
if binding.Review.Options["reasoning_effort"] != "high" {
t.Errorf("Review.Options[reasoning_effort]=%v, want high", binding.Review.Options["reasoning_effort"])
}
if _, present := binding.Work.Options["reasoning_effort"]; present {
t.Errorf("Work.Options unexpectedly declares reasoning_effort: %v", binding.Work.Options)
}
}
func TestSingleRequestPresetBindingAllowsInitialManagedRouteRevision(t *testing.T) {
bindings := validSingleRequestBindings()
for model, dispatch := range bindings {
dispatch.RouteRevision = 0
bindings[model] = dispatch
}
binding, err := compileSingleRequestBinding(
"virtual-public-model",
validSingleRequestPreset(),
bindings,
newTestView("principal-1", nil),
)
if err != nil {
t.Fatalf("initial managed route revision rejected: %v", err)
}
for role, stage := range map[string]edgeservice.SingleRequestStageBinding{
"plan": binding.Plan, "work": binding.Work, "review": binding.Review,
} {
if stage.Dispatch.RouteRevision != 0 {
t.Fatalf("%s RouteRevision=%d, want 0", role, stage.Dispatch.RouteRevision)
}
}
}
func TestSingleRequestPresetBindingUnmanaged(t *testing.T) {
preset := validSingleRequestPreset()
_, err := compileSingleRequestBindingForUnmanaged("virtual-model", preset)
if !errors.Is(err, errSingleRequestBindingUnauthorized) {
t.Fatalf("expected errSingleRequestBindingUnauthorized, got %v", err)
}
}
func TestSingleRequestPresetBindingRejectsInvalidDefenseInDepth(t *testing.T) {
view := newTestView("principal-1", nil)
// Binding-resolution defenses: the fixed shape is valid, so compilation reaches
// the per-stage authorization checks against the managed bindings.
t.Run("missing binding", func(t *testing.T) {
bindings := validSingleRequestBindings()
delete(bindings, "plan-model")
_, err := compileSingleRequestBinding("virtual-model", validSingleRequestPreset(), bindings, view)
if !errors.Is(err, errSingleRequestBindingMissingStage) {
t.Fatalf("expected missing stage error, got %v", err)
}
})
t.Run("unmanaged binding", func(t *testing.T) {
bindings := validSingleRequestBindings()
bindings["plan-model"] = managedBinding("plan-model", "prov-1", "principal-1", "route-plan", false)
_, err := compileSingleRequestBinding("virtual-model", validSingleRequestPreset(), bindings, view)
if !errors.Is(err, errSingleRequestBindingUnauthorized) {
t.Fatalf("expected unauthorized error, got %v", err)
}
})
t.Run("wrong principal", func(t *testing.T) {
bindings := validSingleRequestBindings()
bindings["plan-model"] = managedBinding("plan-model", "prov-1", "principal-2", "route-plan", true)
_, err := compileSingleRequestBinding("virtual-model", validSingleRequestPreset(), bindings, view)
if !errors.Is(err, errSingleRequestBindingUnauthorized) {
t.Fatalf("expected unauthorized error, got %v", err)
}
})
t.Run("model group mismatch", func(t *testing.T) {
bindings := validSingleRequestBindings()
bindings["plan-model"] = managedBinding("wrong-group", "prov-1", "principal-1", "route-plan", true)
_, err := compileSingleRequestBinding("virtual-model", validSingleRequestPreset(), bindings, view)
if !errors.Is(err, errSingleRequestBindingInconsistent) {
t.Fatalf("expected inconsistent error, got %v", err)
}
})
// Fixed-shape defenses: these fail before any binding is resolved, so the
// bindings map is valid to prove the rejection comes from the frozen shape.
t.Run("allowed modes not light", func(t *testing.T) {
preset := validSingleRequestPreset()
preset.AllowedModes = []string{config.ModeDirect}
_, err := compileSingleRequestBinding("virtual-model", preset, validSingleRequestBindings(), view)
if !errors.Is(err, errSingleRequestBindingInconsistent) {
t.Fatalf("expected inconsistent error, got %v", err)
}
})
t.Run("selector model mismatch", func(t *testing.T) {
preset := validSingleRequestPreset()
preset.Selector.Model = "other-model"
_, err := compileSingleRequestBinding("virtual-model", preset, validSingleRequestBindings(), view)
if !errors.Is(err, errSingleRequestBindingInconsistent) {
t.Fatalf("expected inconsistent error, got %v", err)
}
})
t.Run("selector options mismatch", func(t *testing.T) {
preset := validSingleRequestPreset()
preset.Selector.Options = map[string]any{"reasoning_effort": "low"}
_, err := compileSingleRequestBinding("virtual-model", preset, validSingleRequestBindings(), view)
if !errors.Is(err, errSingleRequestBindingInconsistent) {
t.Fatalf("expected inconsistent error, got %v", err)
}
})
t.Run("duplicate role", func(t *testing.T) {
duplicateSequences := [][]config.ExecutionRouteStage{
{
{Role: "plan", Model: "plan-model", Options: map[string]any{"reasoning_effort": "high"}},
{Role: "plan", Model: "plan-model", Options: map[string]any{"reasoning_effort": "high"}},
{Role: "review", Model: "review-model", Options: map[string]any{"reasoning_effort": "high"}},
},
{
{Role: "work", Model: "work-model"},
{Role: "work", Model: "work-model"},
{Role: "review", Model: "review-model", Options: map[string]any{"reasoning_effort": "high"}},
},
{
{Role: "plan", Model: "plan-model", Options: map[string]any{"reasoning_effort": "high"}},
{Role: "review", Model: "review-model", Options: map[string]any{"reasoning_effort": "high"}},
{Role: "review", Model: "review-model", Options: map[string]any{"reasoning_effort": "high"}},
},
}
for _, seq := range duplicateSequences {
preset := validSingleRequestPreset()
route := preset.Routes[config.ModeLight]
route.Stages = seq
preset.Routes[config.ModeLight] = route
_, err := compileSingleRequestBinding("virtual-model", preset, validSingleRequestBindings(), view)
if !errors.Is(err, errSingleRequestBindingDuplicate) {
t.Fatalf("expected duplicate error for sequence %+v, got %v", seq, err)
}
}
})
t.Run("extra route key", func(t *testing.T) {
preset := validSingleRequestPreset()
preset.Routes[config.ModeDirect] = config.ExecutionRoute{
Stages: []config.ExecutionRouteStage{
{Role: "plan", Model: "plan-model"},
},
}
_, err := compileSingleRequestBinding("virtual-model", preset, validSingleRequestBindings(), view)
if !errors.Is(err, errSingleRequestBindingInconsistent) {
t.Fatalf("expected inconsistent error for extra route key, got %v", err)
}
})
t.Run("route policy model mismatch", func(t *testing.T) {
preset := validSingleRequestPreset()
route := preset.Routes[config.ModeLight]
route.Stages[1].Model = "other-work-model"
preset.Routes[config.ModeLight] = route
_, err := compileSingleRequestBinding("virtual-model", preset, validSingleRequestBindings(), view)
if !errors.Is(err, errSingleRequestBindingDynamic) {
t.Fatalf("expected dynamic error, got %v", err)
}
})
t.Run("plan option mismatch", func(t *testing.T) {
preset := validSingleRequestPreset()
route := preset.Routes[config.ModeLight]
route.Stages[0].Options = map[string]any{"reasoning_effort": "low"}
preset.Routes[config.ModeLight] = route
_, err := compileSingleRequestBinding("virtual-model", preset, validSingleRequestBindings(), view)
if !errors.Is(err, errSingleRequestBindingInconsistent) {
t.Fatalf("expected inconsistent error, got %v", err)
}
})
t.Run("review option mismatch", func(t *testing.T) {
preset := validSingleRequestPreset()
route := preset.Routes[config.ModeLight]
route.Stages[2].Options = map[string]any{"reasoning_effort": "low"}
preset.Routes[config.ModeLight] = route
_, err := compileSingleRequestBinding("virtual-model", preset, validSingleRequestBindings(), view)
if !errors.Is(err, errSingleRequestBindingInconsistent) {
t.Fatalf("expected inconsistent error, got %v", err)
}
})
t.Run("work reasoning option", func(t *testing.T) {
preset := validSingleRequestPreset()
preset.SingleRequest.Stages.Work.Options = map[string]any{"reasoning_effort": "high"}
_, err := compileSingleRequestBinding("virtual-model", preset, validSingleRequestBindings(), view)
if !errors.Is(err, errSingleRequestBindingInconsistent) {
t.Fatalf("expected inconsistent error, got %v", err)
}
})
}
func TestSingleRequestPresetBindingNoPresetPolicy(t *testing.T) {
preset := config.ExecutionPreset{
ID: "preset-no-single-request",
Selector: config.ExecutionModelBinding{Model: "selector-model"},
AllowedModes: []string{config.ModeLight},
Routes: map[string]config.ExecutionRoute{
config.ModeLight: {
Stages: []config.ExecutionRouteStage{
{Role: "local", Model: "local-model"},
{Role: "review", Model: "review-model"},
},
},
},
}
bindings := map[string]routeDispatch{
"selector-model": managedBinding("selector-model", "prov-1", "principal-1", "route-s", true),
}
view := newTestView("principal-1", nil)
binding, err := compileSingleRequestBinding("virtual-model", preset, bindings, view)
if err != nil {
t.Fatalf("unexpected error: %v", err)
}
if binding != nil {
t.Fatalf("expected nil binding for preset without SingleRequest policy, got %+v", binding)
}
}
func TestSingleRequestPresetBindingRefreshIsolation(t *testing.T) {
preset := validSingleRequestPreset()
bindings := validSingleRequestBindings()
view := newTestView("principal-1", nil)
binding, err := compileSingleRequestBinding("virtual-public-model", preset, bindings, view)
if err != nil {
t.Fatalf("compilation failed: %v", err)
}
// The admitted binding must start with non-empty approved options.
if binding.Plan.Options["reasoning_effort"] != "high" || binding.Review.Options["reasoning_effort"] != "high" {
t.Fatalf("expected admitted high options, got plan=%v review=%v", binding.Plan.Options, binding.Review.Options)
}
// Simulate a config refresh: mutate the approved stage option values after
// compilation. The already admitted binding must not reflect the mutation.
preset.SingleRequest.Stages.Plan.Options["reasoning_effort"] = "low"
preset.SingleRequest.Stages.Review.Options["reasoning_effort"] = "low"
if binding.Plan.Options["reasoning_effort"] != "high" {
t.Errorf("binding Plan.Options reflected refresh mutation: %v", binding.Plan.Options["reasoning_effort"])
}
if binding.Review.Options["reasoning_effort"] != "high" {
t.Errorf("binding Review.Options reflected refresh mutation: %v", binding.Review.Options["reasoning_effort"])
}
// Simulate a catalog refresh: add a new entry to the bindings map. The
// compiled binding must still reference the original models.
bindings["extra-model"] = managedBinding("extra-model", "prov-2", "principal-1", "route-extra", true)
if binding.Plan.Model != "plan-model" || binding.Work.Model != "work-model" || binding.Review.Model != "review-model" {
t.Errorf("binding models changed after refresh: plan=%q work=%q review=%q",
binding.Plan.Model, binding.Work.Model, binding.Review.Model)
}
}
func TestSingleRequestPresetBindingPublicModelEcho(t *testing.T) {
preset := validSingleRequestPreset()
bindings := validSingleRequestBindings()
view := newTestView("principal-1", nil)
// Public model echo: the binding's PublicModel equals the requested virtual
// model ID, not the selector's route ID or canonical model group.
expectedPublicModels := []string{"virtual-gpt-combo", "my-cool-preset", "preset-alpha"}
for _, publicModel := range expectedPublicModels {
binding, err := compileSingleRequestBinding(publicModel, preset, bindings, view)
if err != nil {
t.Fatalf("compilation failed for %q: %v", publicModel, err)
}
if binding.PublicModel != publicModel {
t.Errorf("PublicModel=%q, want %q", binding.PublicModel, publicModel)
}
// The canonical stage models must never equal the public model.
if binding.Plan.Model == publicModel || binding.Work.Model == publicModel || binding.Review.Model == publicModel {
t.Errorf("stage model unexpectedly equals public model %q", publicModel)
}
}
}
func TestSingleRequestPresetBindingDefensiveCopies(t *testing.T) {
preset := validSingleRequestPreset()
bindings := validSingleRequestBindings()
view := newTestView("principal-1", nil)
binding, err := compileSingleRequestBinding("virtual-model", preset, bindings, view)
if err != nil {
t.Fatalf("compilation failed: %v", err)
}
// The admitted binding carries the approved non-empty plan/review options.
if binding.Plan.Options["reasoning_effort"] != "high" {
t.Fatalf("Plan.Options[reasoning_effort]=%v, want high", binding.Plan.Options["reasoning_effort"])
}
// Clone the binding and verify deep-copy isolation of both structural values
// and the stage option maps in both mutation directions.
clone := binding.Clone()
if clone == nil {
t.Fatal("Clone returned nil")
}
if clone.PublicModel != binding.PublicModel {
t.Errorf("clone PublicModel mismatch")
}
if clone.Plan.Model != binding.Plan.Model || clone.Work.Model != binding.Work.Model || clone.Review.Model != binding.Review.Model {
t.Errorf("clone stage model mismatch: %+v", clone)
}
// Mutating the clone's options must not affect the original.
clone.Plan.Options["reasoning_effort"] = "low"
if binding.Plan.Options["reasoning_effort"] != "high" {
t.Errorf("original Plan.Options mutated through clone: %v", binding.Plan.Options["reasoning_effort"])
}
// Mutating the original's options must not affect the clone.
binding.Review.Options["reasoning_effort"] = "medium"
if clone.Review.Options["reasoning_effort"] != "high" {
t.Errorf("clone Review.Options mutated through original: %v", clone.Review.Options["reasoning_effort"])
}
}
// ensure edgeservice import is used
var _ = edgeservice.SingleRequestBinding{}
func TestSingleRequestPresetBindingDispatchSnapshot(t *testing.T) {
preset := validSingleRequestPreset()
bindings := validSingleRequestBindings()
view := newTestView("principal-1", nil)
binding, err := compileSingleRequestBinding("virtual-model", preset, bindings, view)
if err != nil {
t.Fatalf("compilation failed: %v", err)
}
// Each stage must carry a non-nil dispatch snapshot with the expected facts.
for name, stage := range map[string]edgeservice.SingleRequestStageBinding{
"plan": binding.Plan,
"work": binding.Work,
"review": binding.Review,
} {
if stage.Dispatch == nil {
t.Errorf("%s.Dispatch is nil", name)
continue
}
if !stage.Dispatch.Managed {
t.Errorf("%s.Dispatch.Managed=false", name)
}
if stage.Dispatch.PrincipalRef != "principal-1" {
t.Errorf("%s.Dispatch.PrincipalRef=%q, want principal-1", name, stage.Dispatch.PrincipalRef)
}
if stage.Dispatch.CredentialSlotRef == "" {
t.Errorf("%s.Dispatch.CredentialSlotRef is empty", name)
}
if stage.Dispatch.RouteRevision < 1 {
t.Errorf("%s.Dispatch.RouteRevision=%d, want >= 1", name, stage.Dispatch.RouteRevision)
}
if stage.Dispatch.CredentialRevision < 1 {
t.Errorf("%s.Dispatch.CredentialRevision=%d, want >= 1", name, stage.Dispatch.CredentialRevision)
}
}
// CredentialBindingSnapshot must return a non-nil, non-empty binding.
cb := binding.Plan.Dispatch.CredentialBindingSnapshot()
if cb == nil {
t.Fatal("Plan credential binding snapshot is nil")
}
if cb.PrincipalRef != "principal-1" {
t.Errorf("CredentialBindingSnapshot.PrincipalRef=%q", cb.PrincipalRef)
}
}
func TestSingleRequestPresetBindingDispatchRefreshIsolation(t *testing.T) {
preset := validSingleRequestPreset()
bindings := validSingleRequestBindings()
view := newTestView("principal-1", nil)
binding, err := compileSingleRequestBinding("virtual-model", preset, bindings, view)
if err != nil {
t.Fatalf("compilation failed: %v", err)
}
originalProviderID := binding.Plan.Dispatch.ProviderID
originalRouteID := binding.Plan.Dispatch.RouteID
// Simulate a catalog refresh: mutate the binding source after compilation.
planBinding := bindings["plan-model"]
planBinding.ProviderID = "mutated-provider"
planBinding.RouteID = "mutated-route"
bindings["plan-model"] = planBinding
if binding.Plan.Dispatch.ProviderID != originalProviderID {
t.Errorf("admitted ProviderID reflected catalog refresh mutation: got %q, want %q",
binding.Plan.Dispatch.ProviderID, originalProviderID)
}
if binding.Plan.Dispatch.RouteID != originalRouteID {
t.Errorf("admitted RouteID reflected catalog refresh mutation: got %q, want %q",
binding.Plan.Dispatch.RouteID, originalRouteID)
}
}
func TestSingleRequestPresetBindingDispatchCloneIndependence(t *testing.T) {
preset := validSingleRequestPreset()
bindings := validSingleRequestBindings()
view := newTestView("principal-1", nil)
binding, err := compileSingleRequestBinding("virtual-model", preset, bindings, view)
if err != nil {
t.Fatalf("compilation failed: %v", err)
}
clone := binding.Clone()
// Mutate the clone's dispatch; the original must be unchanged.
clone.Plan.Dispatch.ProviderID = "clone-mutated"
if binding.Plan.Dispatch.ProviderID != "prov-1" {
t.Errorf("original ProviderID mutated through clone: got %q", binding.Plan.Dispatch.ProviderID)
}
// Mutate the original's dispatch; the clone must retain its own mutation,
// not reflect the original's new value.
binding.Plan.Dispatch.ProviderID = "orig-mutated"
if clone.Plan.Dispatch.ProviderID != "clone-mutated" {
t.Errorf("clone ProviderID unexpectedly changed: got %q, want clone-mutated", clone.Plan.Dispatch.ProviderID)
}
if binding.Plan.Dispatch.ProviderID != "orig-mutated" {
t.Errorf("original ProviderID not updated: got %q, want orig-mutated", binding.Plan.Dispatch.ProviderID)
}
}