Files
Assistant/cmd/cognitive-worker/gateway.go
m 4c4a85aec6
CI / test (push) Successful in 24s
CI / test (pull_request) Successful in 22s
feat(cognitive): enforce durable admission leases
2026-08-22 14:33:53 +03:00

380 lines
22 KiB
Go

package main
import (
"bytes"
"context"
"encoding/json"
"fmt"
"io"
"net/http"
"net/url"
"strings"
"time"
"git.cin.su/m/Assistant/adapters/models"
"git.cin.su/m/Assistant/contracts/capability"
"git.cin.su/m/Assistant/contracts/cognitive"
"git.cin.su/m/Assistant/contracts/interaction"
"git.cin.su/m/Assistant/platform/ids"
)
const creatorAdvisoryCapabilityID capability.CapabilityID = "creator-advisory"
type bindingResolver interface {
Resolve(context.Context, capability.CognitiveRequirement) (capability.ResolvedCognitiveBinding, error)
Observe(context.Context, capability.CognitiveUsageObservation) error
RegisterDiscovery(context.Context, capability.CognitiveDiscoverySnapshot) (capability.CognitiveDiscoveryRegistration, error)
}
type admissionClient interface {
Admit(context.Context, string, capability.AdmissionRequest) (capability.AdmissionDecision, capability.AdmissionLease, error)
Release(context.Context, capability.AdmissionLease) error
}
type providerRegistration struct {
ProviderInstanceID capability.ProviderInstanceID
Kind capability.ProviderKind
EndpointRef capability.EndpointBindingRef
CredentialRef capability.CredentialRef
AuthMode capability.AuthMode
BillingMode capability.BillingMode
DataPolicy capability.DataPolicy
Provider advisoryProvider
Discovery models.LocalProvider
}
type cognitiveGateway struct {
resolver bindingResolver
providers map[capability.ProviderKind][]providerRegistration
builtinLocal *providerRegistration
externalRecovery []providerRegistration
now func() time.Time
allowCanary bool
canaryMaxOutputBytes int
admission admissionClient
}
func (g cognitiveGateway) policyHandler(w http.ResponseWriter, r *http.Request) {
var request cognitive.GatewayAdvisoryPolicyRequest
if !decodeJSON(w, r, 8<<10, &request) || request.PrincipalID == "" {
return
}
requirement := capability.CognitiveRequirement{CapabilityID: creatorAdvisoryCapabilityID, OwnerPrincipalID: request.PrincipalID, DataPolicy: request.DataPolicy, RequiredModalities: []capability.ModelModality{capability.ModalityText}, RequiredCapabilities: []capability.ModelCapability{capability.CapabilityAdvisory}, AllowCanary: g.allowCanary, MaxOutputBytes: 1}
selected, err := g.resolver.Resolve(r.Context(), requirement)
policy := cognitive.GatewayAdvisoryPolicy{}
if err == nil && selected.Binding.DataPolicy != capability.DataPolicyLocalOnly {
policy.ExternalEgress = true
policy.AcknowledgementRequired = true
policy.Disclosure = "Текст вопроса будет передан внешнему когнитивному провайдеру; секреты и внутренний контекст не передаются."
}
w.Header().Set("Content-Type", "application/json")
_ = json.NewEncoder(w).Encode(policy)
}
func (g cognitiveGateway) advisoryHandler(w http.ResponseWriter, r *http.Request) {
var request cognitive.GatewayAdvisoryRequest
if !decodeJSON(w, r, 80<<10, &request) {
return
}
if request.Version != cognitive.CreatorAdvisoryV1 || request.RequestID == "" || request.ActorID == "" || request.PrincipalID == "" || request.ScopeID == "" || request.Message == "" || request.MaxOutputBytes < 1 {
http.Error(w, "invalid advisory request", http.StatusBadRequest)
return
}
maxOutputBytes := request.MaxOutputBytes
if g.allowCanary && g.canaryMaxOutputBytes > 0 && maxOutputBytes > g.canaryMaxOutputBytes {
maxOutputBytes = g.canaryMaxOutputBytes
}
requirement := capability.CognitiveRequirement{CapabilityID: creatorAdvisoryCapabilityID, OwnerPrincipalID: request.PrincipalID, DataPolicy: request.DataPolicy, RequiredModalities: []capability.ModelModality{capability.ModalityText}, RequiredCapabilities: []capability.ModelCapability{capability.CapabilityAdvisory}, AllowCanary: g.allowCanary, InputBytes: int64(len(request.Message)), MaxOutputBytes: int64(maxOutputBytes), Objective: selectionObjectiveForMeaning(request.Meaning, request.Message)}
selected, err := g.resolver.Resolve(r.Context(), requirement)
if err != nil {
writeGatewayResponse(w, g.invokeBuiltinAdvisory(r.Context(), request, maxOutputBytes))
return
}
if selected.Binding.DataPolicy != capability.DataPolicyLocalOnly && !request.EgressAcknowledged {
http.Error(w, "external egress acknowledgement required", http.StatusBadRequest)
return
}
provider := g.provider(selected)
if provider == nil {
writeGatewayResponse(w, g.invokeBuiltinAdvisory(r.Context(), request, maxOutputBytes))
return
}
invocation := cognitive.AdvisoryRequest{Version: request.Version, RequestID: request.RequestID, ActorID: request.ActorID, PrincipalID: request.PrincipalID, ScopeID: request.ScopeID, Message: request.Message, Context: request.Context, Locale: request.Locale, MaxOutputBytes: maxOutputBytes, ProviderBindingID: selected.Binding.ID, ProviderBindingRevision: selected.Binding.Revision, ProviderInstanceID: selected.Provider.ID, ModelDescriptorID: selected.Model.ID, ProviderModelID: selected.Model.ProviderModelID}
started := g.now()
response, invokeErr := g.invokeAdmitted(r.Context(), string(request.RequestID), provider, invocation, maxOutputBytes)
if invokeErr != nil {
response = cognitive.AdvisoryResponse{Version: request.Version, RequestID: request.RequestID, ProviderBindingID: selected.Binding.ID, ProviderBindingRevision: selected.Binding.Revision, FinishReason: cognitive.FinishProviderUnavailable}
}
g.observe(r.Context(), selected, response, started)
if fallbackEligible(response.FinishReason) {
requirement.ExcludedBindingIDs = append(requirement.ExcludedBindingIDs, selected.Binding.ID)
fallback, resolveErr := g.resolver.Resolve(r.Context(), requirement)
fallbackProvider := g.provider(fallback)
if resolveErr == nil && fallbackProvider != nil && (fallback.Binding.DataPolicy == capability.DataPolicyLocalOnly || request.EgressAcknowledged) {
fallbackRequest := invocation
fallbackRequest.ProviderBindingID, fallbackRequest.ProviderBindingRevision = fallback.Binding.ID, fallback.Binding.Revision
fallbackRequest.ProviderInstanceID, fallbackRequest.ModelDescriptorID, fallbackRequest.ProviderModelID = fallback.Provider.ID, fallback.Model.ID, fallback.Model.ProviderModelID
fallbackStarted := g.now()
fallbackResponse, fallbackErr := g.invokeAdmitted(r.Context(), string(request.RequestID)+":fallback", fallbackProvider, fallbackRequest, maxOutputBytes)
if fallbackErr != nil {
fallbackResponse = cognitive.AdvisoryResponse{Version: request.Version, RequestID: request.RequestID, ProviderBindingID: fallback.Binding.ID, ProviderBindingRevision: fallback.Binding.Revision, FinishReason: cognitive.FinishProviderUnavailable}
}
g.observe(r.Context(), fallback, fallbackResponse, fallbackStarted)
response = fallbackResponse
}
}
if response.FinishReason != cognitive.FinishCompleted && fallbackEligible(response.FinishReason) {
writeGatewayResponse(w, g.invokeBuiltinAdvisory(r.Context(), request, maxOutputBytes))
return
}
writeGatewayResponse(w, cognitive.GatewayAdvisoryResponse{Version: response.Version, RequestID: response.RequestID, AdvisoryText: response.AdvisoryText, ReferencedExcerptIDs: response.ReferencedExcerptIDs, FinishReason: response.FinishReason, Usage: response.Usage})
}
func (g cognitiveGateway) invokeAdmitted(ctx context.Context, leaseID string, provider advisoryProvider, invocation cognitive.AdvisoryRequest, maxOutputBytes int) (cognitive.AdvisoryResponse, error) {
var lease capability.AdmissionLease
if g.admission != nil {
decision, admittedLease, err := g.admission.Admit(ctx, leaseID, capability.AdmissionRequest{InputBytes: len(invocation.Message), MaxOutputBytes: maxOutputBytes, DeadlineMillis: 0})
if err != nil {
return cognitive.AdvisoryResponse{}, err
}
if decision.State != capability.AdmissionAdmitted {
return cognitive.AdvisoryResponse{Version: invocation.Version, RequestID: invocation.RequestID, ProviderBindingID: invocation.ProviderBindingID, ProviderBindingRevision: invocation.ProviderBindingRevision, FinishReason: cognitive.FinishRateLimited}, nil
}
lease = admittedLease
defer func() { _ = g.admission.Release(ctx, lease) }()
}
return provider.Advise(ctx, invocation)
}
func selectionObjectiveForMessage(message string) capability.SelectionObjective {
lower := strings.ToLower(message)
if strings.Contains(lower, "дешев") || strings.Contains(lower, "бесплат") || strings.Contains(lower, "эконом") || strings.Contains(lower, "cost") || strings.Contains(lower, "cheap") {
return capability.SelectionEconomy
}
if strings.Contains(lower, "быстр") || strings.Contains(lower, "срочно") || strings.Contains(lower, "статус") || strings.Contains(lower, "quick") || strings.Contains(lower, "fast") {
return capability.SelectionSpeed
}
if strings.Contains(lower, "сложн") || strings.Contains(lower, "анализ") || strings.Contains(lower, "код") || strings.Contains(lower, "план") || strings.Contains(lower, "глубок") || strings.Contains(lower, "reason") || strings.Contains(lower, "code") {
return capability.SelectionQuality
}
return capability.SelectionBalanced
}
func selectionObjectiveForMeaning(meaning cognitive.TaskMeaning, message string) capability.SelectionObjective {
if meaning.CapabilityKind != "" || len(meaning.CapabilityHints) > 0 || meaning.SpeechAct != "" || meaning.RequestedOutcome != "" {
kinds := append([]capability.CapabilityKind{meaning.CapabilityKind}, meaning.CapabilityHints...)
for _, kind := range kinds {
switch kind {
case capability.CapabilityKindModifyCode, capability.CapabilityKindRunTests, capability.CapabilityKindVerifyFact, capability.CapabilityKindInspectProject, capability.CapabilityKindReadDatabase, capability.CapabilityKindQueryLogs:
return capability.SelectionQuality
case capability.CapabilityKindScheduleReminder:
return capability.SelectionSpeed
}
}
if meaning.RequestedOutcome != "" || meaning.SpeechAct == "REQUEST_TO_INVESTIGATE" || meaning.SpeechAct == "DIRECTIVE" {
return capability.SelectionQuality
}
if meaning.ConfidenceMillis > 0 {
return capability.SelectionBalanced
}
}
return selectionObjectiveForMessage(message)
}
func builtinAdvisory(request cognitive.GatewayAdvisoryRequest) cognitive.GatewayAdvisoryResponse {
text := "Я рядом и продолжаю работу во встроенном режиме Домового. Внешний когнитивный маршрут сейчас недоступен, поэтому опираюсь на уже установленный смысл сообщения и локальные возможности."
switch request.Meaning.InformationRequestKind {
case interaction.InformationRequestCognitiveModelInventory:
text = "Запрос понят как проверка моделей. Встроенный режим Домового доступен; я могу показать подключённые модели по назначению, качеству, скорости и стоимости. Ключи и пароли в разговор не нужны."
case interaction.InformationRequestCapabilityInventory:
text = "Запрос понят как проверка возможностей. Я могу объяснить доступные capabilities и отличить то, что уже подключено, от того, что пока является только планом или capability gap."
case interaction.InformationRequestCognitiveChannelStatus:
text = "Запрос понят как проверка когнитивного канала. Локальный режим остаётся доступен; я объясню наблюдаемое состояние и что можно сделать без ожидания внешней модели."
case interaction.InformationRequestCognitiveRecovery:
text = "Запрос понят как восстановление ИИ. Начинаю с локального маршрута: фиксирую, что недоступно, проверяю безопасные варианты восстановления и сформулирую следующий обратимый шаг."
default:
switch request.Meaning.SpeechAct {
case "REQUEST_TO_INVESTIGATE":
text = "Я принял это как исследование. Даже без внешней модели могу сформулировать гипотезы, отделить известное от неизвестного и предложить следующий проверяемый шаг."
case "QUESTION":
text = "Я отвечаю во встроенном режиме Домового. Если для полного ответа не хватает внешней модели, объясню это прямо и продолжу с тем, что можно проверить локально."
}
}
return cognitive.GatewayAdvisoryResponse{Version: request.Version, RequestID: request.RequestID, AdvisoryText: text, FinishReason: cognitive.FinishCompleted}
}
func (g cognitiveGateway) invokeBuiltinAdvisory(ctx context.Context, request cognitive.GatewayAdvisoryRequest, maxOutputBytes int) cognitive.GatewayAdvisoryResponse {
if g.builtinLocal != nil && g.builtinLocal.Provider != nil {
invocation := cognitive.AdvisoryRequest{Version: request.Version, RequestID: request.RequestID, ActorID: request.ActorID, PrincipalID: request.PrincipalID, ScopeID: request.ScopeID, Message: request.Message, Context: request.Context, Locale: request.Locale, MaxOutputBytes: maxOutputBytes, ProviderBindingID: "builtin-local-binding-v1", ProviderBindingRevision: 1, ProviderInstanceID: "builtin-local-lifeboat-v1", ModelDescriptorID: "builtin-local-model-v1", ProviderModelID: "domovoy-lifeboat"}
started := g.now()
response, err := g.builtinLocal.Provider.Advise(ctx, invocation)
if response.RequestID == "" {
response.RequestID = request.RequestID
}
if response.Version == "" {
response.Version = request.Version
}
if err != nil {
response = cognitive.AdvisoryResponse{Version: request.Version, RequestID: request.RequestID, ProviderBindingID: "builtin-local-binding-v1", ProviderBindingRevision: 1, FinishReason: cognitive.FinishInvalidProviderResponse}
}
g.observeBuiltin(ctx, g.builtinLocal, response, started)
if err == nil && response.FinishReason == cognitive.FinishCompleted && strings.TrimSpace(response.AdvisoryText) != "" {
return cognitive.GatewayAdvisoryResponse{Version: request.Version, RequestID: request.RequestID, AdvisoryText: response.AdvisoryText, ReferencedExcerptIDs: response.ReferencedExcerptIDs, FinishReason: response.FinishReason, Usage: response.Usage}
}
}
if request.EgressAcknowledged {
if response, ok := g.recoverExternalAdvisory(ctx, request, maxOutputBytes); ok {
return response
}
}
return builtinAdvisory(request)
}
func (g cognitiveGateway) recoverExternalAdvisory(ctx context.Context, request cognitive.GatewayAdvisoryRequest, maxOutputBytes int) (cognitive.GatewayAdvisoryResponse, bool) {
for _, candidate := range g.externalRecovery {
if candidate.Provider == nil || candidate.Discovery == nil {
continue
}
probeCtx, cancel := context.WithTimeout(ctx, 8*time.Second)
modelsFound, probe := candidate.Discovery.Discover(probeCtx)
cancel()
if probe.Health != capability.HealthAvailable || len(modelsFound) == 0 {
continue
}
model := modelsFound[0]
invocation := cognitive.AdvisoryRequest{Version: request.Version, RequestID: request.RequestID, ActorID: request.ActorID, PrincipalID: request.PrincipalID, ScopeID: request.ScopeID, Message: request.Message, Context: request.Context, Locale: request.Locale, MaxOutputBytes: maxOutputBytes, ProviderBindingID: "recovery-external-binding-v1", ProviderBindingRevision: 1, ProviderInstanceID: capability.ProviderInstanceID("recovery-" + string(candidate.ProviderInstanceID)), ModelDescriptorID: capability.ModelDescriptorID("recovery-model-" + string(model.ProviderModelID)), ProviderModelID: model.ProviderModelID}
response, err := candidate.Provider.Advise(ctx, invocation)
if err == nil && response.FinishReason == cognitive.FinishCompleted && strings.TrimSpace(response.AdvisoryText) != "" {
return cognitive.GatewayAdvisoryResponse{Version: request.Version, RequestID: request.RequestID, AdvisoryText: response.AdvisoryText, ReferencedExcerptIDs: response.ReferencedExcerptIDs, FinishReason: response.FinishReason, Usage: response.Usage}, true
}
}
return cognitive.GatewayAdvisoryResponse{}, false
}
func (g cognitiveGateway) observe(ctx context.Context, selected capability.ResolvedCognitiveBinding, response cognitive.AdvisoryResponse, started time.Time) {
usage := response.Usage
observationID, idErr := ids.New()
if idErr == nil {
observation := capability.CognitiveUsageObservation{ObservationID: observationID, RequestID: ids.ID(response.RequestID), BindingID: selected.Binding.ID, BindingRevision: selected.Binding.Revision, ProviderInstanceID: selected.Provider.ID, ModelDescriptorID: selected.Model.ID, ObservedAtUnixMillis: g.now().UnixMilli(), LatencyMillis: g.now().Sub(started).Milliseconds(), ResultClass: capability.CognitiveResultClass(response.FinishReason)}
if usage != nil {
observation.InputUnits, observation.OutputUnits = usage.InputUnits, usage.OutputUnits
}
if response.FinishReason != cognitive.FinishCompleted {
observation.FailureType = capability.CognitiveFailureType(response.FinishReason)
}
if selected.Provider.Kind == "openai-compatible-chat-v1" || selected.Provider.Kind == "ollama-v1" {
zero := int64(0)
observation.ActualCostMicros = &zero
}
_ = g.resolver.Observe(ctx, observation)
}
}
// observeBuiltin gives the always-available local provider the same evidence
// path as a resolved binding. The identity is stable and contains references
// only; no endpoint, credential, prompt, or response text is persisted here.
func (g cognitiveGateway) observeBuiltin(ctx context.Context, registration *providerRegistration, response cognitive.AdvisoryResponse, started time.Time) {
if registration == nil {
return
}
selected := capability.ResolvedCognitiveBinding{
Binding: capability.CognitiveBinding{ID: "builtin-local-binding-v1", Revision: 1},
Provider: capability.ProviderInstance{ID: "builtin-local-lifeboat-v1", Revision: 1, Kind: registration.Kind},
Model: capability.ModelDescriptor{ID: "builtin-local-model-v1", Revision: 1, ProviderInstanceID: "builtin-local-lifeboat-v1", ProviderModelID: "domovoy-lifeboat"},
}
g.observe(ctx, selected, response, started)
}
func fallbackEligible(reason cognitive.FinishReason) bool {
switch reason {
case cognitive.FinishProviderUnavailable, cognitive.FinishModelUnavailable, cognitive.FinishTimeout, cognitive.FinishDegraded:
return true
}
return false
}
func (g cognitiveGateway) provider(selected capability.ResolvedCognitiveBinding) advisoryProvider {
for _, registration := range g.providers[selected.Provider.Kind] {
if registration.EndpointRef == selected.Provider.EndpointRef && registration.CredentialRef == selected.Provider.CredentialRef {
return registration.Provider
}
}
return nil
}
func writeGatewayResponse(w http.ResponseWriter, response cognitive.GatewayAdvisoryResponse) {
w.Header().Set("Content-Type", "application/json")
_ = json.NewEncoder(w).Encode(response)
}
func decodeJSON(w http.ResponseWriter, r *http.Request, limit int64, destination any) bool {
r.Body = http.MaxBytesReader(w, r.Body, limit)
decoder := json.NewDecoder(r.Body)
decoder.DisallowUnknownFields()
if err := decoder.Decode(destination); err != nil {
http.Error(w, "invalid request", http.StatusBadRequest)
return false
}
if err := decoder.Decode(&struct{}{}); err != io.EOF {
http.Error(w, "invalid request", http.StatusBadRequest)
return false
}
return true
}
type httpBindingResolver struct {
base *url.URL
client *http.Client
}
func newHTTPBindingResolver(raw string, client *http.Client) (*httpBindingResolver, error) {
base, err := url.Parse(raw)
if err != nil || base.Scheme == "" || base.Host == "" {
return nil, fmt.Errorf("invalid Core Runtime URL")
}
return &httpBindingResolver{base: base, client: client}, nil
}
func (r *httpBindingResolver) Resolve(ctx context.Context, value capability.CognitiveRequirement) (capability.ResolvedCognitiveBinding, error) {
var result capability.ResolvedCognitiveBinding
err := r.post(ctx, "/v1/cognitive-bindings/resolve", value, &result)
return result, err
}
func (r *httpBindingResolver) Observe(ctx context.Context, value capability.CognitiveUsageObservation) error {
return r.post(ctx, "/v1/cognitive-usage-observations", value, nil)
}
func (r *httpBindingResolver) RegisterDiscovery(ctx context.Context, value capability.CognitiveDiscoverySnapshot) (capability.CognitiveDiscoveryRegistration, error) {
var result capability.CognitiveDiscoveryRegistration
err := r.post(ctx, "/v1/cognitive-registry/discoveries", value, &result)
return result, err
}
func (r *httpBindingResolver) post(ctx context.Context, path string, value, destination any) error {
payload, err := json.Marshal(value)
if err != nil {
return err
}
target := *r.base
target.Path = path
request, err := http.NewRequestWithContext(ctx, http.MethodPost, target.String(), bytes.NewReader(payload))
if err != nil {
return err
}
request.Header.Set("Content-Type", "application/json")
response, err := r.client.Do(request)
if err != nil {
return err
}
defer response.Body.Close()
if response.StatusCode < 200 || response.StatusCode >= 300 {
_, _ = io.Copy(io.Discard, io.LimitReader(response.Body, 4096))
return fmt.Errorf("registry status %d", response.StatusCode)
}
if destination != nil {
return json.NewDecoder(io.LimitReader(response.Body, 64<<10)).Decode(destination)
}
return nil
}