Compare commits
12
Commits
| Author | SHA1 | Date | |
|---|---|---|---|
|
|
e1b9c7ca84 | ||
|
|
00af176a20 | ||
|
|
ce7000e78e | ||
|
|
6204e14177 | ||
|
|
185317ee6a | ||
|
|
2e824b3c1f | ||
|
|
f0360a8e13 | ||
|
|
5593f2ac2c | ||
|
|
216d76f1d8 | ||
|
|
499b9678e1 | ||
|
|
78a39ec011 | ||
|
|
d120cc0958 |
@@ -100,6 +100,22 @@ func (r *InteractionRepository) Utterances(ctx context.Context, id interaction.C
|
||||
return out, rows.Err()
|
||||
}
|
||||
|
||||
// UtterancesWithProjectReferences returns canonical utterances together with
|
||||
// their already-persisted Interaction-owned project references. The caller
|
||||
// uses the pair only as a read projection; this method does not infer or write
|
||||
// scope membership.
|
||||
func (r *InteractionRepository) UtterancesWithProjectReferences(ctx context.Context, id interaction.ConversationID) ([]interaction.HumanUtterance, []interaction.ProjectReference, error) {
|
||||
utterances, err := r.Utterances(ctx, id)
|
||||
if err != nil {
|
||||
return nil, nil, err
|
||||
}
|
||||
references, err := r.projectReferencesForUtterances(ctx, utterances)
|
||||
if err != nil {
|
||||
return nil, nil, err
|
||||
}
|
||||
return utterances, references, nil
|
||||
}
|
||||
|
||||
// RecentUtterancesForPrincipal provides durable conversational context when a
|
||||
// client starts a fresh conversation instead of carrying the previous ID.
|
||||
func (r *InteractionRepository) RecentUtterancesForPrincipal(ctx context.Context, principalID identity.PrincipalID, limit int) ([]interaction.HumanUtterance, error) {
|
||||
@@ -118,3 +134,52 @@ func (r *InteractionRepository) RecentUtterancesForPrincipal(ctx context.Context
|
||||
}
|
||||
return out, rows.Err()
|
||||
}
|
||||
|
||||
// RecentUtterancesForPrincipalWithProjectReferences keeps a new conversation
|
||||
// within the same durable Interaction read boundary as its scoped context.
|
||||
func (r *InteractionRepository) RecentUtterancesForPrincipalWithProjectReferences(ctx context.Context, principalID identity.PrincipalID, limit int) ([]interaction.HumanUtterance, []interaction.ProjectReference, error) {
|
||||
utterances, err := r.RecentUtterancesForPrincipal(ctx, principalID, limit)
|
||||
if err != nil {
|
||||
return nil, nil, err
|
||||
}
|
||||
references, err := r.projectReferencesForUtterances(ctx, utterances)
|
||||
if err != nil {
|
||||
return nil, nil, err
|
||||
}
|
||||
return utterances, references, nil
|
||||
}
|
||||
|
||||
func (r *InteractionRepository) projectReferencesForUtterances(ctx context.Context, utterances []interaction.HumanUtterance) ([]interaction.ProjectReference, error) {
|
||||
if len(utterances) == 0 {
|
||||
return nil, nil
|
||||
}
|
||||
ids := make([]string, 0, len(utterances))
|
||||
seen := make(map[interaction.HumanUtteranceID]struct{}, len(utterances))
|
||||
for _, utterance := range utterances {
|
||||
if utterance.ID == "" {
|
||||
continue
|
||||
}
|
||||
if _, exists := seen[utterance.ID]; exists {
|
||||
continue
|
||||
}
|
||||
seen[utterance.ID] = struct{}{}
|
||||
ids = append(ids, string(utterance.ID))
|
||||
}
|
||||
if len(ids) == 0 {
|
||||
return nil, nil
|
||||
}
|
||||
rows, err := r.pool.Query(ctx, `SELECT project_reference_id,utterance_id,project_id,project_display_name,role,provenance,confidence_millis,recorded_at_unix_millis FROM interaction.project_references WHERE utterance_id = ANY($1::text[]) ORDER BY recorded_at_unix_millis,project_reference_id`, ids)
|
||||
if err != nil {
|
||||
return nil, err
|
||||
}
|
||||
defer rows.Close()
|
||||
refs := make([]interaction.ProjectReference, 0)
|
||||
for rows.Next() {
|
||||
var ref interaction.ProjectReference
|
||||
if err := rows.Scan(&ref.ID, &ref.UtteranceID, &ref.ProjectID, &ref.ProjectDisplayName, &ref.Role, &ref.Provenance, &ref.ConfidenceMillis, &ref.RecordedAtUnixMillis); err != nil {
|
||||
return nil, err
|
||||
}
|
||||
refs = append(refs, ref)
|
||||
}
|
||||
return refs, rows.Err()
|
||||
}
|
||||
|
||||
@@ -16,15 +16,17 @@ func (p *admissionTestProvider) Advise(context.Context, cognitive.AdvisoryReques
|
||||
}
|
||||
|
||||
type admissionTestClient struct {
|
||||
decision capability.AdmissionDecision
|
||||
releases int
|
||||
decision capability.AdmissionDecision
|
||||
releases int
|
||||
releaseContextErr error
|
||||
}
|
||||
|
||||
func (c *admissionTestClient) Admit(context.Context, string, capability.AdmissionRequest) (capability.AdmissionDecision, capability.AdmissionLease, error) {
|
||||
return c.decision, capability.AdmissionLease{ID: "lease-1", PolicyID: "policy-1"}, nil
|
||||
}
|
||||
func (c *admissionTestClient) Release(context.Context, capability.AdmissionLease) error {
|
||||
func (c *admissionTestClient) Release(ctx context.Context, _ capability.AdmissionLease) error {
|
||||
c.releases++
|
||||
c.releaseContextErr = ctx.Err()
|
||||
return nil
|
||||
}
|
||||
|
||||
@@ -47,3 +49,14 @@ func TestInvokeAdmittedReleasesDurableLeaseAfterProvider(t *testing.T) {
|
||||
t.Fatalf("admitted lease was not released: response=%#v err=%v calls=%d releases=%d", response, err, provider.calls, client.releases)
|
||||
}
|
||||
}
|
||||
|
||||
func TestInvokeAdmittedReleasesLeaseAfterCallerCancellation(t *testing.T) {
|
||||
provider := &admissionTestProvider{}
|
||||
client := &admissionTestClient{decision: capability.AdmissionDecision{State: capability.AdmissionAdmitted, Reason: capability.AdmissionReasonNone, PolicyID: "policy-1"}}
|
||||
ctx, cancel := context.WithCancel(context.Background())
|
||||
cancel()
|
||||
_, _ = (&cognitiveGateway{admission: client}).invokeAdmitted(ctx, "request-1", provider, cognitive.AdvisoryRequest{Message: "hello"}, 100)
|
||||
if client.releases != 1 || client.releaseContextErr != nil {
|
||||
t.Fatalf("cancelled caller leaked durable lease: releases=%d release_context_err=%v", client.releases, client.releaseContextErr)
|
||||
}
|
||||
}
|
||||
|
||||
@@ -1,7 +1,9 @@
|
||||
package main
|
||||
|
||||
import (
|
||||
"context"
|
||||
"encoding/json"
|
||||
"fmt"
|
||||
"net/http"
|
||||
"sort"
|
||||
"time"
|
||||
@@ -22,17 +24,7 @@ func (s discoveryService) handler(w http.ResponseWriter, r *http.Request) {
|
||||
s.statusHandler(w, r)
|
||||
return
|
||||
}
|
||||
registration, ok := s.providers[r.PathValue("provider")]
|
||||
if !ok || registration.Discovery == nil {
|
||||
http.Error(w, "provider discovery not found", http.StatusNotFound)
|
||||
return
|
||||
}
|
||||
discovered, probe := registration.Discovery.Discover(r.Context())
|
||||
snapshot := capability.CognitiveDiscoverySnapshot{ProviderInstanceID: registration.ProviderInstanceID, ProviderKind: registration.Kind, EndpointRef: registration.EndpointRef, CredentialRef: registration.CredentialRef, AuthMode: registration.AuthMode, BillingMode: registration.BillingMode, Enabled: true, OwnerPrincipalID: s.owner, CapabilityID: creatorAdvisoryCapabilityID, DataPolicy: registration.DataPolicy, LimitPolicyRef: "creator-advisory-local-v1", Health: probe.Health, FailureClass: capability.CognitiveFailureType(probe.Failure), ObservedAtUnixMillis: s.now().UnixMilli()}
|
||||
for _, model := range discovered {
|
||||
snapshot.Models = append(snapshot.Models, capability.DiscoveredModelSnapshot{ProviderModelID: model.ProviderModelID, Fingerprint: model.Fingerprint, MaxContextBytes: model.MaxContextBytes, Modalities: model.Modalities, Capabilities: model.Capabilities, Installed: capability.OperationalYes, Loaded: capability.OperationalNotObserved, Invokable: capability.OperationalNotObserved})
|
||||
}
|
||||
result, err := s.resolver.RegisterDiscovery(r.Context(), snapshot)
|
||||
result, err := s.register(r.Context(), r.PathValue("provider"))
|
||||
if err != nil {
|
||||
http.Error(w, "provider discovery registration unavailable", http.StatusBadGateway)
|
||||
return
|
||||
@@ -41,6 +33,19 @@ func (s discoveryService) handler(w http.ResponseWriter, r *http.Request) {
|
||||
_ = json.NewEncoder(w).Encode(result)
|
||||
}
|
||||
|
||||
func (s discoveryService) register(ctx context.Context, provider string) (capability.CognitiveDiscoveryRegistration, error) {
|
||||
registration, ok := s.providers[provider]
|
||||
if !ok || registration.Discovery == nil {
|
||||
return capability.CognitiveDiscoveryRegistration{}, fmt.Errorf("provider discovery not found")
|
||||
}
|
||||
discovered, probe := registration.Discovery.Discover(ctx)
|
||||
snapshot := capability.CognitiveDiscoverySnapshot{ProviderInstanceID: registration.ProviderInstanceID, ProviderKind: registration.Kind, EndpointRef: registration.EndpointRef, CredentialRef: registration.CredentialRef, AuthMode: registration.AuthMode, BillingMode: registration.BillingMode, Enabled: true, OwnerPrincipalID: s.owner, CapabilityID: creatorAdvisoryCapabilityID, DataPolicy: registration.DataPolicy, LimitPolicyRef: "creator-advisory-local-v1", Health: probe.Health, FailureClass: capability.CognitiveFailureType(probe.Failure), ObservedAtUnixMillis: s.now().UnixMilli()}
|
||||
for _, model := range discovered {
|
||||
snapshot.Models = append(snapshot.Models, capability.DiscoveredModelSnapshot{ProviderModelID: model.ProviderModelID, Fingerprint: model.Fingerprint, MaxContextBytes: model.MaxContextBytes, Modalities: model.Modalities, Capabilities: model.Capabilities, Installed: capability.OperationalYes, Loaded: capability.OperationalNotObserved, Invokable: capability.OperationalNotObserved})
|
||||
}
|
||||
return s.resolver.RegisterDiscovery(ctx, snapshot)
|
||||
}
|
||||
|
||||
type discoveryStatus struct {
|
||||
Provider string `json:"provider"`
|
||||
ProviderKind capability.ProviderKind `json:"provider_kind"`
|
||||
|
||||
@@ -4,6 +4,7 @@ import (
|
||||
"bytes"
|
||||
"context"
|
||||
"encoding/json"
|
||||
"errors"
|
||||
"fmt"
|
||||
"io"
|
||||
"net/http"
|
||||
@@ -20,6 +21,8 @@ import (
|
||||
|
||||
const creatorAdvisoryCapabilityID capability.CapabilityID = "creator-advisory"
|
||||
|
||||
var errNoInvocableCognitiveBinding = errors.New("no invocable cognitive binding")
|
||||
|
||||
type bindingResolver interface {
|
||||
Resolve(context.Context, capability.CognitiveRequirement) (capability.ResolvedCognitiveBinding, error)
|
||||
Observe(context.Context, capability.CognitiveUsageObservation) error
|
||||
@@ -56,6 +59,7 @@ type cognitiveGateway struct {
|
||||
allowCanary bool
|
||||
canaryMaxOutputBytes int
|
||||
admission admissionClient
|
||||
revalidatePrimary func(context.Context) error
|
||||
}
|
||||
|
||||
func (g cognitiveGateway) policyHandler(w http.ResponseWriter, r *http.Request) {
|
||||
@@ -89,7 +93,7 @@ func (g cognitiveGateway) advisoryHandler(w http.ResponseWriter, r *http.Request
|
||||
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)
|
||||
selected, err := g.resolveWithRevalidation(r.Context(), requirement)
|
||||
if err != nil {
|
||||
writeGatewayResponse(w, g.invokeBuiltinAdvisory(r.Context(), request, maxOutputBytes))
|
||||
return
|
||||
@@ -145,11 +149,33 @@ func (g cognitiveGateway) invokeAdmitted(ctx context.Context, leaseID string, pr
|
||||
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) }()
|
||||
defer func() {
|
||||
releaseCtx, cancel := context.WithTimeout(context.WithoutCancel(ctx), 5*time.Second)
|
||||
defer cancel()
|
||||
_ = g.admission.Release(releaseCtx, lease)
|
||||
}()
|
||||
}
|
||||
return provider.Advise(ctx, invocation)
|
||||
}
|
||||
|
||||
// resolveWithRevalidation makes one bounded recovery attempt when Capability
|
||||
// has an active/validated binding but its durable health projection rejected it
|
||||
// as non-invocable. Discovery remains the canonical Capability write path; the
|
||||
// retry observes the configured primary local provider and re-runs selection
|
||||
// once. Provider invocation, admission, and fallback remain governed elsewhere.
|
||||
func (g cognitiveGateway) resolveWithRevalidation(ctx context.Context, requirement capability.CognitiveRequirement) (capability.ResolvedCognitiveBinding, error) {
|
||||
selected, err := g.resolver.Resolve(ctx, requirement)
|
||||
if err == nil || !errors.Is(err, errNoInvocableCognitiveBinding) || g.revalidatePrimary == nil {
|
||||
return selected, err
|
||||
}
|
||||
revalidationCtx, cancel := context.WithTimeout(ctx, 45*time.Second)
|
||||
defer cancel()
|
||||
if revalidationErr := g.revalidatePrimary(revalidationCtx); revalidationErr != nil {
|
||||
return selected, err
|
||||
}
|
||||
return g.resolver.Resolve(ctx, requirement)
|
||||
}
|
||||
|
||||
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") {
|
||||
@@ -374,6 +400,10 @@ type httpBindingResolver struct {
|
||||
client *http.Client
|
||||
}
|
||||
|
||||
type httpStatusError struct{ status int }
|
||||
|
||||
func (e httpStatusError) Error() string { return fmt.Sprintf("registry status %d", e.status) }
|
||||
|
||||
func newHTTPBindingResolver(raw string, client *http.Client) (*httpBindingResolver, error) {
|
||||
base, err := url.Parse(raw)
|
||||
if err != nil || base.Scheme == "" || base.Host == "" {
|
||||
@@ -385,6 +415,10 @@ func newHTTPBindingResolver(raw string, client *http.Client) (*httpBindingResolv
|
||||
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)
|
||||
var statusErr httpStatusError
|
||||
if errors.As(err, &statusErr) && statusErr.status == http.StatusNotFound {
|
||||
return result, errNoInvocableCognitiveBinding
|
||||
}
|
||||
return result, err
|
||||
}
|
||||
|
||||
@@ -437,7 +471,7 @@ func (r *httpBindingResolver) post(ctx context.Context, path string, value, dest
|
||||
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)
|
||||
return httpStatusError{status: response.StatusCode}
|
||||
}
|
||||
if destination != nil {
|
||||
return json.NewDecoder(io.LimitReader(response.Body, 64<<10)).Decode(destination)
|
||||
|
||||
@@ -22,6 +22,26 @@ type registryStub struct {
|
||||
observation *capability.CognitiveUsageObservation
|
||||
}
|
||||
|
||||
type revalidatingRegistry struct {
|
||||
selected capability.ResolvedCognitiveBinding
|
||||
calls int
|
||||
alwaysNo bool
|
||||
}
|
||||
|
||||
func (s *revalidatingRegistry) Resolve(context.Context, capability.CognitiveRequirement) (capability.ResolvedCognitiveBinding, error) {
|
||||
s.calls++
|
||||
if s.alwaysNo || s.calls == 1 {
|
||||
return capability.ResolvedCognitiveBinding{}, errNoInvocableCognitiveBinding
|
||||
}
|
||||
return s.selected, nil
|
||||
}
|
||||
func (*revalidatingRegistry) Observe(context.Context, capability.CognitiveUsageObservation) error {
|
||||
return nil
|
||||
}
|
||||
func (*revalidatingRegistry) RegisterDiscovery(context.Context, capability.CognitiveDiscoverySnapshot) (capability.CognitiveDiscoveryRegistration, error) {
|
||||
return capability.CognitiveDiscoveryRegistration{}, nil
|
||||
}
|
||||
|
||||
type failoverRegistry struct {
|
||||
primary, fallback capability.ResolvedCognitiveBinding
|
||||
observations []capability.CognitiveUsageObservation
|
||||
@@ -88,6 +108,59 @@ func TestActiveRegistryBindingReachesProviderAdapter(t *testing.T) {
|
||||
}
|
||||
}
|
||||
|
||||
func TestStaleUnavailableBindingRevalidatesOnceBeforeInvocation(t *testing.T) {
|
||||
selected := gatewaySelection()
|
||||
selected.Binding.DataPolicy = capability.DataPolicyLocalOnly
|
||||
registry := &revalidatingRegistry{selected: selected}
|
||||
provider := &stubAdvisoryProvider{response: cognitive.AdvisoryResponse{Version: cognitive.CreatorAdvisoryV1, RequestID: "request-1", AdvisoryText: "ok", FinishReason: cognitive.FinishCompleted}}
|
||||
revalidations := 0
|
||||
gateway := cognitiveGateway{resolver: registry, providers: map[capability.ProviderKind][]providerRegistration{"provider-kind-1": {{EndpointRef: "endpoint-1", CredentialRef: "credential-1", Provider: provider}}}, now: time.Now, revalidatePrimary: func(context.Context) error { revalidations++; return nil }}
|
||||
response := httptest.NewRecorder()
|
||||
gateway.advisoryHandler(response, httptest.NewRequest(http.MethodPost, "/v1/advisories", strings.NewReader(gatewayRequest())))
|
||||
if response.Code != http.StatusOK || !provider.called || revalidations != 1 || registry.calls != 2 {
|
||||
t.Fatalf("status=%d provider_called=%t revalidations=%d resolves=%d", response.Code, provider.called, revalidations, registry.calls)
|
||||
}
|
||||
}
|
||||
|
||||
func TestRevalidationFailureDoesNotMakeBindingInvocable(t *testing.T) {
|
||||
registry := ®istryStub{err: errNoInvocableCognitiveBinding}
|
||||
provider := &stubAdvisoryProvider{}
|
||||
revalidations := 0
|
||||
gateway := cognitiveGateway{resolver: registry, providers: map[capability.ProviderKind][]providerRegistration{"provider-kind-1": {{Provider: provider}}}, now: time.Now, revalidatePrimary: func(context.Context) error { revalidations++; return errors.New("provider unavailable") }}
|
||||
gateway.advisoryHandler(httptest.NewRecorder(), httptest.NewRequest(http.MethodPost, "/v1/advisories", strings.NewReader(gatewayRequest())))
|
||||
if provider.called || revalidations != 1 {
|
||||
t.Fatalf("failed provider revalidation was treated as invocable: provider_called=%t revalidations=%d", provider.called, revalidations)
|
||||
}
|
||||
}
|
||||
|
||||
func TestRevalidationIsNotRetriedInALoop(t *testing.T) {
|
||||
registry := &revalidatingRegistry{alwaysNo: true}
|
||||
revalidations := 0
|
||||
gateway := cognitiveGateway{resolver: registry, now: time.Now, revalidatePrimary: func(context.Context) error { revalidations++; return nil }}
|
||||
gateway.advisoryHandler(httptest.NewRecorder(), httptest.NewRequest(http.MethodPost, "/v1/advisories", strings.NewReader(gatewayRequest())))
|
||||
if revalidations != 1 || registry.calls != 2 {
|
||||
t.Fatalf("revalidation loop: revalidations=%d resolves=%d", revalidations, registry.calls)
|
||||
}
|
||||
}
|
||||
|
||||
func TestHTTPResolverMarksNoInvocableBindingForSingleRevalidation(t *testing.T) {
|
||||
server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
|
||||
if r.URL.Path != "/v1/cognitive-bindings/resolve" {
|
||||
t.Fatalf("path=%s", r.URL.Path)
|
||||
}
|
||||
http.Error(w, "no invocable cognitive binding", http.StatusNotFound)
|
||||
}))
|
||||
defer server.Close()
|
||||
resolver, err := newHTTPBindingResolver(server.URL, server.Client())
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
_, err = resolver.Resolve(context.Background(), capability.CognitiveRequirement{CapabilityID: creatorAdvisoryCapabilityID, OwnerPrincipalID: "principal-1", MaxOutputBytes: 1})
|
||||
if !errors.Is(err, errNoInvocableCognitiveBinding) {
|
||||
t.Fatalf("resolve error=%v, want no-invocable sentinel", err)
|
||||
}
|
||||
}
|
||||
|
||||
func TestProviderUsesStableInstanceWhenCredentialMetadataIsStale(t *testing.T) {
|
||||
provider := &stubAdvisoryProvider{}
|
||||
gateway := cognitiveGateway{providers: map[capability.ProviderKind][]providerRegistration{
|
||||
|
||||
@@ -113,11 +113,15 @@ func main() {
|
||||
kind := capability.ProviderKind(envOr("OPENAI_PROVIDER_KIND", "openai-responses-v1"))
|
||||
providers[kind] = append(providers[kind], providerRegistration{ProviderInstanceID: "openai-bootstrap", Kind: kind, EndpointRef: capability.EndpointBindingRef(envOr("OPENAI_ENDPOINT_REF", "openai-responses-default")), CredentialRef: capability.CredentialRef(envOr("OPENAI_CREDENTIAL_REF", "openai-advisory-v1")), AuthMode: capability.AuthKey, BillingMode: capability.BillingPerInput, DataPolicy: capability.DataPolicyExternalMessageOnly, Provider: openAIProvider})
|
||||
}
|
||||
gateway := cognitiveGateway{resolver: resolver, providers: providers, builtinLocal: builtinLocal, externalRecovery: externalRecovery, now: time.Now, allowCanary: os.Getenv("COGNITIVE_ALLOW_CANARY") == "true", canaryMaxOutputBytes: 4096, admission: newHTTPAdmissionClient(envOr("CORE_RUNTIME_URL", "http://core-runtime:8080"), &http.Client{Timeout: 5 * time.Second})}
|
||||
discovery := discoveryService{resolver: resolver, providers: discoveries, owner: identity.PrincipalID(envOr("COGNITIVE_OWNER_PRINCIPAL_ID", "domovoy-creator-principal")), now: time.Now}
|
||||
gateway := cognitiveGateway{resolver: resolver, providers: providers, builtinLocal: builtinLocal, externalRecovery: externalRecovery, now: time.Now, allowCanary: os.Getenv("COGNITIVE_ALLOW_CANARY") == "true", canaryMaxOutputBytes: 4096, admission: newHTTPAdmissionClient(envOr("CORE_RUNTIME_URL", "http://core-runtime:8080"), &http.Client{Timeout: 5 * time.Second}), revalidatePrimary: func(ctx context.Context) error {
|
||||
_, err := discovery.register(ctx, "medium")
|
||||
return err
|
||||
}}
|
||||
mux.Handle("/healthz", health.Handler{Runtime: "cognitive-worker", SemanticGeneration: 1, ArchitectureGeneration: 1})
|
||||
readinessClient := &http.Client{Timeout: 2 * time.Second}
|
||||
cpuReadinessClient := &http.Client{Timeout: 10 * time.Second}
|
||||
mediumReadinessTimeout := 120 * time.Second
|
||||
mediumReadinessTimeout := 120 * time.Second
|
||||
mediumReadinessClient := &http.Client{Timeout: mediumReadinessTimeout}
|
||||
lifeboatURL := strings.TrimSuffix(strings.TrimRight(envOr("LOCAL_LIFEBOAT_PROVIDER_URL", "http://cognitive-lifeboat:8086/v1"), "/"), "/v1") + "/health"
|
||||
fleetClient := &http.Client{Timeout: 2 * time.Second}
|
||||
@@ -199,7 +203,6 @@ func main() {
|
||||
mux.HandleFunc("POST /v1/multimodal/ocr", multimodalGateway.handler(capability.MultimodalRecognizeOCR))
|
||||
mux.HandleFunc("POST /v1/multimodal/video-understand", multimodalGateway.handler(capability.MultimodalUnderstandVideo))
|
||||
mux.HandleFunc("POST /v1/multimodal/generate-image", multimodalGateway.handler(capability.MultimodalGenerateImage))
|
||||
discovery := discoveryService{resolver: resolver, providers: discoveries, owner: identity.PrincipalID(envOr("COGNITIVE_OWNER_PRINCIPAL_ID", "domovoy-creator-principal")), now: time.Now}
|
||||
mux.HandleFunc("POST /v1/provider-discovery/{provider}", discovery.handler)
|
||||
mux.Handle("POST /v1/provider-discovery/{provider}/qualify", modelQualificationHandler(discoveries, string(envOr("COGNITIVE_OWNER_PRINCIPAL_ID", "domovoy-creator-principal")), httpQualificationObserver{resolver: resolver}))
|
||||
mux.HandleFunc("GET /v1/provider-discovery/status", discovery.statusHandler)
|
||||
|
||||
@@ -28,6 +28,35 @@ type cognitiveInterpretationClient struct {
|
||||
media *http.Client
|
||||
}
|
||||
|
||||
// compileInteractionContext is the Core-side handoff from the existing
|
||||
// Interaction retrieval projection to the existing Context Compiler. The
|
||||
// compiler receives only already-authorized, typed fragments and applies its
|
||||
// scope and byte-budget guards before the cognitive gateway sees them.
|
||||
func compileInteractionContext(projection interactiondomain.ContextProjection, focusProjectID string) (string, error) {
|
||||
if len(projection.Fragments) == 0 {
|
||||
return projection.Text, nil
|
||||
}
|
||||
candidates := make([]cognitive.ContextCandidate, 0, len(projection.Fragments))
|
||||
for _, fragment := range projection.Fragments {
|
||||
candidates = append(candidates, cognitive.ContextCandidate{
|
||||
Excerpt: cognitive.ContextExcerpt{ID: cognitive.ContextExcerptID("interaction:" + string(fragment.UtteranceID)), SourceKind: "interaction-gen2-retrieval", SourceID: "interaction:utterances", Revision: 1, Sensitivity: cognitive.SensitivityPrivateAllowed, Text: fragment.Text},
|
||||
Relevance: fragment.Relevance,
|
||||
Lane: "interaction",
|
||||
ContextScopeRef: fragment.ScopeRef,
|
||||
FreshnessState: "CURRENT",
|
||||
})
|
||||
}
|
||||
bundle, err := cognitive.CompileContext(cognitive.ContextCompileInput{Candidates: candidates, DataPolicy: capability.DataPolicyLocalOnly, MaxBytes: 2048, ContextScopeRef: focusProjectID, FreshnessRequirement: "conversation-current", RiskClass: "low"})
|
||||
if err != nil {
|
||||
return "", err
|
||||
}
|
||||
parts := make([]string, 0, len(bundle.Excerpts))
|
||||
for _, excerpt := range bundle.Excerpts {
|
||||
parts = append(parts, excerpt.Text)
|
||||
}
|
||||
return strings.Join(parts, "\n"), nil
|
||||
}
|
||||
|
||||
const (
|
||||
// Household interpretation and response review are intentionally compact
|
||||
// calls. Keeping their output bounded prevents a loaded local model from
|
||||
|
||||
@@ -70,6 +70,16 @@ func TestAnswerStableKnowledgeUsesLocalOnlyCognitiveRoute(t *testing.T) {
|
||||
}
|
||||
}
|
||||
|
||||
func TestCompileInteractionContextAppliesTypedScopeBeforeGateway(t *testing.T) {
|
||||
context, err := compileInteractionContext(interactiondomain.ContextProjection{Fragments: []interactiondomain.ContextFragment{
|
||||
{UtteranceID: "a", Text: "контрольный код = ALPHA", ScopeRef: "company-a", Relevance: 2},
|
||||
{UtteranceID: "b", Text: "контрольный код = BETA", ScopeRef: "company-b", Relevance: 3},
|
||||
}}, "company-a")
|
||||
if err != nil || context != "контрольный код = ALPHA" {
|
||||
t.Fatalf("context=%q err=%v", context, err)
|
||||
}
|
||||
}
|
||||
|
||||
func TestGenerateImageUsesDedicatedCognitiveMediaRoute(t *testing.T) {
|
||||
server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
|
||||
if r.Method != http.MethodPost || r.URL.Path != "/v1/multimodal/generate-image" {
|
||||
|
||||
@@ -571,9 +571,10 @@ func main() {
|
||||
travelContinuation := false
|
||||
var previousTask *interaction.HumanUtterance
|
||||
var conversationHistory []interaction.HumanUtterance
|
||||
var contextProjectReferences []interaction.ProjectReference
|
||||
if request.ConversationID != "" {
|
||||
var historyErr error
|
||||
conversationHistory, historyErr = interactionRepository.Utterances(r.Context(), conversation.ID)
|
||||
conversationHistory, contextProjectReferences, historyErr = interactionRepository.UtterancesWithProjectReferences(r.Context(), conversation.ID)
|
||||
if historyErr != nil {
|
||||
log.Printf("interaction conversation history unavailable: %v", historyErr)
|
||||
}
|
||||
@@ -581,14 +582,24 @@ func main() {
|
||||
// Keep durable user context available even when the web client carries a
|
||||
// conversation ID; a new or rotated conversation must not erase learned
|
||||
// preferences.
|
||||
recentUserHistory, historyErr := interactionRepository.RecentUtterancesForPrincipal(r.Context(), session.PrincipalID, 64)
|
||||
recentUserHistory, recentProjectReferences, historyErr := interactionRepository.RecentUtterancesForPrincipalWithProjectReferences(r.Context(), session.PrincipalID, 64)
|
||||
if historyErr != nil {
|
||||
log.Printf("interaction principal history unavailable: %v", historyErr)
|
||||
} else {
|
||||
log.Printf("interaction principal history loaded: %d utterances", len(recentUserHistory))
|
||||
}
|
||||
conversationHistory = append(conversationHistory, recentUserHistory...)
|
||||
gen2Context := interactiondomain.BuildScopedGen2Context(conversationHistory, string(session.PrincipalID))
|
||||
contextProjectReferences = append(contextProjectReferences, recentProjectReferences...)
|
||||
focusProjectID := ""
|
||||
if previousFocus != nil {
|
||||
focusProjectID = string(previousFocus.ProjectID)
|
||||
}
|
||||
gen2Context := interactiondomain.BuildScopedGen2ContextForQuery(conversationHistory, contextProjectReferences, string(session.PrincipalID), focusProjectID, effectiveTaskMessage)
|
||||
if compiledContext, contextErr := compileInteractionContext(gen2Context, focusProjectID); contextErr != nil {
|
||||
log.Printf("interaction context compilation unavailable: %v", contextErr)
|
||||
} else {
|
||||
gen2Context.Text = compiledContext
|
||||
}
|
||||
if gen2Context.UsedPersonalMemoryFallback {
|
||||
telemetry.RecordCompatibility("personal-memory-fallback", "core-runtime", "read", "gen2-context-empty", true, true, false, "USED", utterance.CorrelationID, string(session.PrincipalID), now)
|
||||
}
|
||||
@@ -887,7 +898,7 @@ func main() {
|
||||
planningReply = retrievedMemoryReply
|
||||
needsCognition = false
|
||||
}
|
||||
if hasChatRoute(routePlan, chatRouteResearch) && semantic.Interpretation.Clarification != interaction.ClarificationRequired && shouldRunGeneralResearch(semantic.InformationRequestKind) {
|
||||
if hasChatRoute(routePlan, chatRouteResearch) && len(gen2Context.Fragments) == 0 && semantic.Interpretation.Clarification != interaction.ClarificationRequired && shouldRunGeneralResearch(semantic.InformationRequestKind) {
|
||||
researchAttempted = true
|
||||
if planID, idErr := ids.New(); idErr == nil {
|
||||
parts := interactiondomain.IndependentRequestParts(effectiveTaskMessage)
|
||||
@@ -954,7 +965,7 @@ func main() {
|
||||
}
|
||||
}
|
||||
canGenerateTextReply := shouldGenerateNonMutatingTextDirective(semantic, routePlan, helpPath, creativeRequest) && planningReply == deterministicReply
|
||||
if (planningReply == "" && needsCognition && semantic.Interpretation.SpeechAct == interaction.SpeechActQuestion || canGenerateTextReply) && !hasChatRoute(routePlan, chatRouteResearch) && !creativeRequest {
|
||||
if (planningReply == "" && needsCognition && semantic.Interpretation.SpeechAct == interaction.SpeechActQuestion || canGenerateTextReply) && (!hasChatRoute(routePlan, chatRouteResearch) || len(gen2Context.Fragments) > 0) && !creativeRequest {
|
||||
answerCtx, cancelAnswer := context.WithTimeout(r.Context(), 12*time.Second)
|
||||
answer, answerErr := cognitiveClient.AnswerStableKnowledge(answerCtx, session, effectiveTaskMessage, gen2Context.Text, cognitive.TaskMeaning{
|
||||
SpeechAct: string(semantic.Interpretation.SpeechAct), CapabilityHints: toCapabilityKinds(semantic.CapabilityHints), RequestedOutcome: semantic.Interpretation.RequestedOutcome, InformationRequestKind: semantic.InformationRequestKind, ConfidenceMillis: semantic.Interpretation.ConfidenceMillis,
|
||||
@@ -1483,6 +1494,10 @@ func main() {
|
||||
candidates, err = capabilityRepository.LoadCognitiveCandidates(r.Context(), requirement.CapabilityID, requirement.OwnerPrincipalID)
|
||||
}
|
||||
if err != nil {
|
||||
// The resolver is an internal capability boundary. Keep request identity
|
||||
// and content out of logs, but preserve the repository error so a 503
|
||||
// can be distinguished from a true empty/unsuitable binding set.
|
||||
log.Printf("cognitive registry resolve unavailable: %v", err)
|
||||
http.Error(w, "cognitive registry unavailable", http.StatusServiceUnavailable)
|
||||
return
|
||||
}
|
||||
|
||||
+14
-1
@@ -261,7 +261,20 @@ func (s *server) promoteTraffic(w http.ResponseWriter, _ *http.Request) {
|
||||
respond(w, ok, err)
|
||||
}
|
||||
func (s *server) rollbackTraffic(w http.ResponseWriter, _ *http.Request) {
|
||||
ok, err := s.lifeboat.RollbackTraffic()
|
||||
state, snapshotErr := s.lifeboat.Snapshot()
|
||||
if snapshotErr != nil {
|
||||
respond(w, false, snapshotErr)
|
||||
return
|
||||
}
|
||||
// A ReleaseSet must roll back as a whole. The legacy release-only method
|
||||
// would merely flip the traffic mode while retaining a candidate upstream.
|
||||
var ok bool
|
||||
var err error
|
||||
if state.StableReleaseSet != nil {
|
||||
ok, err = s.lifeboat.RollbackSet()
|
||||
} else {
|
||||
ok, err = s.lifeboat.RollbackTraffic()
|
||||
}
|
||||
respond(w, ok, err)
|
||||
}
|
||||
|
||||
|
||||
@@ -0,0 +1,149 @@
|
||||
# This file deliberately owns only one inactive release slot. It joins the
|
||||
# stable project's shared network as an external resource, but has a distinct
|
||||
# Compose project name supplied by run-production-canary.sh. Do not add
|
||||
# durable services or stable services here.
|
||||
services:
|
||||
research-runtime:
|
||||
build:
|
||||
context: ..
|
||||
dockerfile: deploy/images/runtime.Dockerfile
|
||||
args: { TARGET: research-runtime, SOURCE_REVISION: "${ASSISTANT_SOURCE_REVISION:-unknown}" }
|
||||
restart: "no"
|
||||
environment:
|
||||
LISTEN_ADDR: :8087
|
||||
RESEARCH_SOURCE_ROOT: /var/lib/research-sources
|
||||
RESEARCH_HTTP_ALLOWED_HOSTS: assistant.cin.su,git.cin.su
|
||||
RESEARCH_SEARCH_URL_TEMPLATE: ${RESEARCH_SEARCH_URL_TEMPLATE:-https://www.bing.com/search?q=%s}
|
||||
RESEARCH_SEARCH_ALLOWED_HOSTS: ${RESEARCH_SEARCH_ALLOWED_HOSTS:-www.bing.com}
|
||||
POSTGRES_HOST: postgres
|
||||
POSTGRES_PORT: "5432"
|
||||
POSTGRES_DATABASE: ${POSTGRES_DB:-assistant}
|
||||
EPISTEMICS_POSTGRES_USER: epistemics_role
|
||||
EPISTEMICS_POSTGRES_CREDENTIAL_REF: credential://domovoy/epistemics/db-password/v1
|
||||
SECRET_BACKEND_ROOT: /run/secrets
|
||||
volumes:
|
||||
- assistant_core_secrets:/run/secrets:ro
|
||||
- assistant_research_sources:/var/lib/research-sources:ro
|
||||
networks:
|
||||
shared:
|
||||
aliases: [research-runtime-candidate]
|
||||
healthcheck:
|
||||
test: ["CMD-SHELL", "wget -qO- http://localhost:8087/healthz >/dev/null && wget -qO- http://localhost:8087/readyz >/dev/null"]
|
||||
interval: 10s
|
||||
timeout: 10s
|
||||
retries: 20
|
||||
|
||||
core-runtime:
|
||||
build:
|
||||
context: ..
|
||||
dockerfile: deploy/images/runtime.Dockerfile
|
||||
args: { TARGET: core-runtime, SOURCE_REVISION: "${ASSISTANT_SOURCE_REVISION:-unknown}" }
|
||||
restart: "no"
|
||||
environment:
|
||||
LISTEN_ADDR: ":8080"
|
||||
POSTGRES_HOST: postgres
|
||||
POSTGRES_PORT: "5432"
|
||||
POSTGRES_DATABASE: ${POSTGRES_DB:-assistant}
|
||||
POSTGRES_USER: situation_role
|
||||
POSTGRES_PASSWORD: ${SITUATION_DB_PASSWORD}
|
||||
IDENTITY_POSTGRES_USER: identity_role
|
||||
IDENTITY_POSTGRES_PASSWORD: ${IDENTITY_DB_PASSWORD}
|
||||
CREATOR_USERNAME: ${CREATOR_USERNAME}
|
||||
WORK_POSTGRES_USER: work_role
|
||||
WORK_POSTGRES_PASSWORD: ${WORK_DB_PASSWORD}
|
||||
CAPABILITY_POSTGRES_USER: capability_role
|
||||
CAPABILITY_POSTGRES_PASSWORD: ${CAPABILITY_DB_PASSWORD}
|
||||
EPISTEMICS_POSTGRES_USER: epistemics_role
|
||||
EPISTEMICS_POSTGRES_CREDENTIAL_REF: credential://domovoy/epistemics/db-password/v1
|
||||
INTERACTION_POSTGRES_USER: interaction_role
|
||||
INTERACTION_POSTGRES_CREDENTIAL_REF: credential://domovoy/interaction/db-password/v1
|
||||
SECRET_BACKEND_ROOT: /run/secrets
|
||||
COGNITIVE_ADMISSION_POLICY_ID: creator-advisory-local-v1
|
||||
COGNITIVE_ADMISSION_MAX_CONCURRENT: ${COGNITIVE_ADMISSION_MAX_CONCURRENT:-1}
|
||||
COGNITIVE_ADMISSION_MAX_ACCEPTED_PER_HOUR: ${COGNITIVE_ADMISSION_MAX_ACCEPTED_PER_HOUR:-60}
|
||||
COGNITIVE_ADMISSION_MAX_INPUT_BYTES: ${COGNITIVE_ADMISSION_MAX_INPUT_BYTES:-81920}
|
||||
COGNITIVE_ADMISSION_MAX_OUTPUT_BYTES: ${COGNITIVE_ADMISSION_MAX_OUTPUT_BYTES:-16384}
|
||||
COGNITIVE_RUNTIME_URL: http://cognitive-worker-candidate:8083
|
||||
RESEARCH_RUNTIME_URL: http://research-runtime-candidate:8087
|
||||
volumes:
|
||||
- assistant_core_secrets:/run/secrets:ro
|
||||
networks:
|
||||
shared:
|
||||
aliases: [core-runtime-candidate]
|
||||
healthcheck:
|
||||
test: ["CMD-SHELL", "wget -qO- http://localhost:8080/healthz >/dev/null && wget -qO- http://localhost:8080/readyz >/dev/null"]
|
||||
interval: 10s
|
||||
timeout: 20s
|
||||
retries: 20
|
||||
|
||||
cognitive-worker:
|
||||
build:
|
||||
context: ..
|
||||
dockerfile: deploy/images/runtime.Dockerfile
|
||||
args: { TARGET: cognitive-worker, SOURCE_REVISION: "${ASSISTANT_SOURCE_REVISION:-unknown}" }
|
||||
restart: "no"
|
||||
environment:
|
||||
LISTEN_ADDR: ":8083"
|
||||
CORE_RUNTIME_URL: http://core-runtime-candidate:8080
|
||||
LOCAL_CPU_PROVIDER_URL: http://docker.cin.su:18082/v1
|
||||
LOCAL_CPU_PROVIDER_API_KEY_FILE: ${LOCAL_CPU_PROVIDER_API_KEY_FILE:-}
|
||||
LOCAL_CPU_PROVIDER_CREDENTIAL_REF: ${LOCAL_CPU_PROVIDER_CREDENTIAL_REF:-credential://domovoy/local-cpu/provider-api-key/v1}
|
||||
LOCAL_MEDIUM_PROVIDER_URL: ${LOCAL_MEDIUM_PROVIDER_URL:-http://llm-local-gateway:8080/v1}
|
||||
LOCAL_MEDIUM_PROVIDER_API_KEY_FILE: ${LOCAL_MEDIUM_PROVIDER_API_KEY_FILE:-/run/secrets/llm_local_gateway_api_key}
|
||||
LOCAL_MEDIUM_PROVIDER_CREDENTIAL_REF: ${LOCAL_MEDIUM_PROVIDER_CREDENTIAL_REF:-credential://domovoy/local-medium/provider-api-key/v1}
|
||||
LOCAL_GPU_PROVIDER_TIMEOUT: ${LOCAL_GPU_PROVIDER_TIMEOUT:-60s}
|
||||
LOCAL_LIFEBOAT_PROVIDER_URL: http://cognitive-lifeboat:8086/v1
|
||||
LOCAL_IMAGE_PROVIDER_KIND: ${LOCAL_IMAGE_PROVIDER_KIND:-comfyui}
|
||||
LOCAL_IMAGE_PROVIDER_URL: ${LOCAL_IMAGE_PROVIDER_URL:-http://comfyui:8188}
|
||||
LOCAL_IMAGE_MODEL: ${LOCAL_IMAGE_MODEL:-z_image_turbo_bf16.safetensors}
|
||||
COGNITIVE_OWNER_PRINCIPAL_ID: domovoy-creator-principal
|
||||
COGNITIVE_ALLOW_CANARY: "true"
|
||||
volumes:
|
||||
- assistant_core_secrets:/run/secrets:ro
|
||||
networks:
|
||||
shared:
|
||||
aliases: [cognitive-worker-candidate]
|
||||
llm-local: {}
|
||||
healthcheck:
|
||||
test: ["CMD-SHELL", "wget -qO- http://localhost:8083/healthz >/dev/null && wget -qO- http://localhost:8083/readyz >/dev/null"]
|
||||
interval: 10s
|
||||
timeout: 20s
|
||||
retries: 20
|
||||
|
||||
creator-web:
|
||||
build:
|
||||
context: ..
|
||||
dockerfile: deploy/images/runtime.Dockerfile
|
||||
args: { TARGET: creator-web, SOURCE_REVISION: "${ASSISTANT_SOURCE_REVISION:-unknown}" }
|
||||
restart: "no"
|
||||
environment:
|
||||
LISTEN_ADDR: ":8084"
|
||||
CORE_RUNTIME_URL: http://core-runtime-candidate:8080
|
||||
CONTINUITY_RUNTIME_URL: http://continuity-runtime:8081
|
||||
COMMIT_RUNTIME_URL: http://commit-runtime:8082
|
||||
COGNITIVE_RUNTIME_URL: http://cognitive-worker-candidate:8083
|
||||
AGENT_WORKER_URL: http://agent-worker:8090
|
||||
ASSISTANT_BUILD_REVISION: ${ASSISTANT_SOURCE_REVISION:-unknown}
|
||||
networks:
|
||||
shared:
|
||||
aliases: [creator-web-candidate]
|
||||
healthcheck:
|
||||
test: ["CMD-SHELL", "wget -qO- http://localhost:8084/healthz >/dev/null"]
|
||||
interval: 10s
|
||||
timeout: 5s
|
||||
retries: 20
|
||||
|
||||
volumes:
|
||||
assistant_core_secrets:
|
||||
external: true
|
||||
assistant_research_sources:
|
||||
name: assistant_research_sources
|
||||
external: true
|
||||
|
||||
networks:
|
||||
shared:
|
||||
name: assistant_default
|
||||
external: true
|
||||
llm-local:
|
||||
name: llm-local_default
|
||||
external: true
|
||||
@@ -89,17 +89,6 @@ services:
|
||||
timeout: 20s
|
||||
retries: 20
|
||||
|
||||
core-runtime-candidate:
|
||||
profiles: ["assistant-canary"]
|
||||
extends:
|
||||
service: core-runtime
|
||||
restart: "no"
|
||||
environment:
|
||||
RESEARCH_RUNTIME_URL: http://research-runtime-candidate:8087
|
||||
depends_on:
|
||||
postgres: { condition: service_healthy }
|
||||
research-runtime-candidate: { condition: service_started }
|
||||
|
||||
agent-canary:
|
||||
profiles: ["agent-canary"]
|
||||
build: { context: .., dockerfile: deploy/images/runtime.Dockerfile, args: { TARGET: agent-canary, SOURCE_REVISION: "${ASSISTANT_SOURCE_REVISION:-unknown}" } }
|
||||
@@ -223,14 +212,6 @@ services:
|
||||
timeout: 10s
|
||||
retries: 20
|
||||
|
||||
research-runtime-candidate:
|
||||
profiles: ["assistant-canary"]
|
||||
extends:
|
||||
service: research-runtime
|
||||
restart: "no"
|
||||
depends_on:
|
||||
postgres: { condition: service_healthy }
|
||||
|
||||
continuity-runtime:
|
||||
build: { context: .., dockerfile: deploy/images/runtime.Dockerfile, args: { TARGET: continuity-runtime, SOURCE_REVISION: "${ASSISTANT_SOURCE_REVISION:-unknown}" } }
|
||||
restart: unless-stopped
|
||||
@@ -299,17 +280,6 @@ services:
|
||||
timeout: 20s
|
||||
retries: 20
|
||||
|
||||
cognitive-worker-candidate:
|
||||
profiles: ["assistant-canary"]
|
||||
extends:
|
||||
service: cognitive-worker
|
||||
restart: "no"
|
||||
environment:
|
||||
CORE_RUNTIME_URL: http://core-runtime-candidate:8080
|
||||
depends_on:
|
||||
core-runtime-candidate: { condition: service_started }
|
||||
cognitive-lifeboat: { condition: service_healthy }
|
||||
|
||||
cognitive-lifeboat:
|
||||
image: ghcr.io/ggml-org/llama.cpp:server
|
||||
restart: unless-stopped
|
||||
@@ -386,28 +356,6 @@ services:
|
||||
timeout: 3s
|
||||
retries: 10
|
||||
|
||||
creator-web-candidate:
|
||||
profiles: ["assistant-canary"]
|
||||
extends:
|
||||
service: creator-web
|
||||
restart: "no"
|
||||
environment:
|
||||
CORE_RUNTIME_URL: http://core-runtime-candidate:8080
|
||||
CONTINUITY_RUNTIME_URL: http://continuity-runtime:8081
|
||||
COMMIT_RUNTIME_URL: http://commit-runtime:8082
|
||||
COGNITIVE_RUNTIME_URL: http://cognitive-worker-candidate:8083
|
||||
AGENT_WORKER_URL: http://agent-worker:8090
|
||||
depends_on:
|
||||
core-runtime-candidate: { condition: service_started }
|
||||
continuity-runtime: { condition: service_started }
|
||||
commit-runtime: { condition: service_started }
|
||||
cognitive-worker-candidate: { condition: service_healthy }
|
||||
healthcheck:
|
||||
test: ["CMD-SHELL", "wget -qO- http://localhost:8084/healthz >/dev/null"]
|
||||
interval: 10s
|
||||
timeout: 5s
|
||||
retries: 20
|
||||
|
||||
volumes:
|
||||
assistant_pg:
|
||||
assistant_nats:
|
||||
|
||||
@@ -0,0 +1,55 @@
|
||||
#!/usr/bin/env bash
|
||||
set -euo pipefail
|
||||
|
||||
# Stable-only production updater. Candidate slots are managed exclusively by
|
||||
# their isolated release command and Supervisor; this controller never starts
|
||||
# or reconciles a candidate service.
|
||||
BASE_DIR="${BASE_DIR:-/home/m/assistant-deploy}"
|
||||
REPO_DIR="${REPO_DIR:-$BASE_DIR/repo}"
|
||||
ENV_FILE="${ENV_FILE:-$BASE_DIR/production-compose.env}"
|
||||
COMPOSE_FILE="$REPO_DIR/deploy/docker-compose.prod.yml"
|
||||
LOG_FILE="${LOG_FILE:-$BASE_DIR/deploy.log}"
|
||||
LOCK_FILE="${LOCK_FILE:-/tmp/assistant-stable-deploy.lock}"
|
||||
stable_services=(postgres core-runtime continuity-runtime commit-runtime cognitive-worker creator-web notification-receiver research-runtime agent-worker work-graph-runtime)
|
||||
|
||||
exec 9>"$LOCK_FILE"
|
||||
flock -n 9 || exit 0
|
||||
log() { printf '%s %s\n' "$(date -u +%Y-%m-%dT%H:%M:%SZ)" "$*" >>"$LOG_FILE"; }
|
||||
fail() { log "FAILED: $*"; exit 1; }
|
||||
compose() { docker compose --env-file "$ENV_FILE" -f "$COMPOSE_FILE" "$@"; }
|
||||
|
||||
cd "$REPO_DIR"
|
||||
git fetch --prune origin main >>"$LOG_FILE" 2>&1
|
||||
git merge --ff-only origin/main >>"$LOG_FILE" 2>&1 || fail "stable checkout is not a fast-forward to origin/main"
|
||||
revision="$(git rev-parse HEAD)"
|
||||
[[ "$revision" =~ ^[0-9a-f]{40}$ ]] || fail "invalid source revision"
|
||||
|
||||
if ! REPO_DIR="$REPO_DIR" ENV_FILE="$ENV_FILE" bash "$REPO_DIR/deploy/apply-migrations.sh" >>"$LOG_FILE" 2>&1; then
|
||||
fail "migration application failed"
|
||||
fi
|
||||
compose config >/dev/null 2>>"$LOG_FILE" || fail "stable compose config validation failed"
|
||||
|
||||
all_current=1
|
||||
for service in "${stable_services[@]}"; do
|
||||
container="assistant-${service}-1"
|
||||
label="$(docker inspect "$container" --format '{{index .Config.Labels "org.opencontainers.image.revision"}}' 2>/dev/null || true)"
|
||||
[[ "$label" == "$revision" ]] || all_current=0
|
||||
done
|
||||
if [[ "$all_current" == 1 ]]; then log "SKIP stable revision=$revision"; exit 0; fi
|
||||
|
||||
log "BUILD stable revision=$revision"
|
||||
ASSISTANT_SOURCE_REVISION="$revision" compose build --build-arg "SOURCE_REVISION=$revision" "${stable_services[@]}" >>"$LOG_FILE" 2>&1
|
||||
log "UP stable revision=$revision"
|
||||
ASSISTANT_SOURCE_REVISION="$revision" compose up -d --no-build "${stable_services[@]}" >>"$LOG_FILE" 2>&1
|
||||
|
||||
for attempt in $(seq 1 30); do
|
||||
if curl -fsS --max-time 5 https://assistant.cin.su/healthz >/dev/null 2>&1; then break; fi
|
||||
[[ "$attempt" == 30 ]] && fail "public health failed"
|
||||
sleep 10
|
||||
done
|
||||
for service in "${stable_services[@]}"; do
|
||||
container="assistant-${service}-1"
|
||||
label="$(docker inspect "$container" --format '{{index .Config.Labels "org.opencontainers.image.revision"}}' 2>/dev/null || true)"
|
||||
[[ "$label" == "$revision" ]] || fail "artifact drift container=$container"
|
||||
done
|
||||
log "OK stable revision=$revision"
|
||||
@@ -1,57 +1,54 @@
|
||||
#!/usr/bin/env bash
|
||||
set -euo pipefail
|
||||
|
||||
# Candidate services are never part of the stable Compose project. The slot
|
||||
# Compose file may use only declared external networks and durable resources.
|
||||
BASE_DIR="${REPO_DIR:-$(CDPATH= cd -- "$(dirname -- "$0")/.." && pwd)}"
|
||||
ENV_FILE="${ENV_FILE:-/home/m/assistant-deploy/production-compose.env}"
|
||||
COMPOSE_FILE="$BASE_DIR/deploy/docker-compose.prod.yml"
|
||||
CANDIDATE_FILE="$BASE_DIR/deploy/docker-compose.candidate-slot.yml"
|
||||
SUPERVISOR_CONTROL_URL="${SUPERVISOR_CONTROL_URL:-http://127.0.0.1:${SUPERVISOR_CONTROL_PORT:-18085}}"
|
||||
REVISION="${1:?candidate revision is required}"
|
||||
SLOT="${CANDIDATE_SLOT:-b}"
|
||||
STABLE_PROJECT="${STABLE_PROJECT:-assistant}"
|
||||
CANDIDATE_PROJECT="${CANDIDATE_PROJECT:-assistant-slot-$SLOT}"
|
||||
HEALTH_ATTEMPTS="${HEALTH_ATTEMPTS:-360}"
|
||||
PROMOTE_AFTER_CANARY="${PROMOTE_AFTER_CANARY:-true}"
|
||||
|
||||
set -a
|
||||
. "$ENV_FILE"
|
||||
set +a
|
||||
[[ "$SLOT" =~ ^[a-z0-9-]+$ ]] || { echo "invalid candidate slot" >&2; exit 2; }
|
||||
[[ "$CANDIDATE_PROJECT" != "$STABLE_PROJECT" ]] || { echo "candidate and stable Compose projects must differ" >&2; exit 2; }
|
||||
|
||||
compose() { docker compose --env-file "$ENV_FILE" -f "$COMPOSE_FILE" "$@"; }
|
||||
candidate_compose() { docker compose -p "$CANDIDATE_PROJECT" --env-file "$ENV_FILE" -f "$CANDIDATE_FILE" "$@"; }
|
||||
control() { curl -fsS --max-time 10 -X POST "$SUPERVISOR_CONTROL_URL$1" "${@:2}" >/dev/null; }
|
||||
fail_rollback() {
|
||||
control /v1/traffic/rollback || true
|
||||
control /v1/releases/rollback || true
|
||||
echo "production ReleaseSet canary failed; stable ReleaseSet rollback requested" >&2
|
||||
echo "isolated candidate ReleaseSet failed; stable ReleaseSet rollback requested" >&2
|
||||
exit 1
|
||||
}
|
||||
|
||||
services=(creator-web-candidate core-runtime-candidate cognitive-worker-candidate research-runtime-candidate)
|
||||
containers=(assistant-creator-web-candidate-1 assistant-core-runtime-candidate-1 assistant-cognitive-worker-candidate-1 assistant-research-runtime-candidate-1)
|
||||
services=(creator-web core-runtime cognitive-worker research-runtime)
|
||||
stable_services=(creator-web core-runtime cognitive-worker research-runtime)
|
||||
|
||||
stable_ids() {
|
||||
local service container id
|
||||
for service in "${stable_services[@]}"; do
|
||||
container="assistant-${service}-1"
|
||||
container="${STABLE_PROJECT}-${service}-1"
|
||||
id=$(docker inspect -f '{{.Id}}' "$container" 2>/dev/null || true)
|
||||
[[ -n "$id" ]] || { echo "stable container missing: $container" >&2; return 1; }
|
||||
printf '%s=%s\n' "$service" "$id"
|
||||
done
|
||||
}
|
||||
|
||||
assert_stable_ids_unchanged() {
|
||||
local after
|
||||
after=$(stable_ids) || fail_rollback
|
||||
[[ "$after" == "$stable_ids_before" ]] || {
|
||||
echo "stable container IDs changed during candidate operation" >&2
|
||||
echo "stable container IDs changed during isolated candidate operation" >&2
|
||||
diff -u <(printf '%s\n' "$stable_ids_before") <(printf '%s\n' "$after") >&2 || true
|
||||
fail_rollback
|
||||
}
|
||||
}
|
||||
|
||||
stable_ids_before=$(stable_ids) || exit 1
|
||||
|
||||
compose --profile assistant-canary build --build-arg "SOURCE_REVISION=$REVISION" "${services[@]}"
|
||||
|
||||
candidate_container() { printf '%s-%s-1' "$CANDIDATE_PROJECT" "$1"; }
|
||||
wait_healthy() {
|
||||
local container="$1" status="" ready=""
|
||||
ready=""
|
||||
for attempt in $(seq 1 "$HEALTH_ATTEMPTS"); do
|
||||
status=$(docker inspect -f '{{if .State.Health}}{{.State.Health.Status}}{{else}}unknown{{end}}' "$container" 2>/dev/null || true)
|
||||
if [[ "$status" == healthy ]]; then ready=true; break; fi
|
||||
@@ -60,68 +57,57 @@ wait_healthy() {
|
||||
[[ "$ready" == true ]] || fail_rollback
|
||||
}
|
||||
|
||||
compose --profile assistant-canary up -d --no-build --no-deps research-runtime-candidate core-runtime-candidate
|
||||
wait_healthy assistant-research-runtime-candidate-1
|
||||
wait_healthy assistant-core-runtime-candidate-1
|
||||
compose --profile assistant-canary up -d --no-build --no-deps cognitive-worker-candidate
|
||||
wait_healthy assistant-cognitive-worker-candidate-1
|
||||
compose --profile assistant-canary up -d --no-build --no-deps creator-web-candidate
|
||||
wait_healthy assistant-creator-web-candidate-1
|
||||
stable_ids_before=$(stable_ids) || exit 1
|
||||
candidate_compose config >/dev/null
|
||||
candidate_compose build --build-arg "SOURCE_REVISION=$REVISION" "${services[@]}"
|
||||
assert_stable_ids_unchanged
|
||||
|
||||
for container in "${containers[@]}"; do
|
||||
image_revision=$(docker inspect -f '{{index .Config.Labels "org.opencontainers.image.revision"}}' "$container" 2>/dev/null || true)
|
||||
candidate_compose up -d --no-build --no-deps research-runtime core-runtime
|
||||
wait_healthy "$(candidate_container research-runtime)"
|
||||
wait_healthy "$(candidate_container core-runtime)"
|
||||
candidate_compose up -d --no-build --no-deps cognitive-worker
|
||||
wait_healthy "$(candidate_container cognitive-worker)"
|
||||
candidate_compose up -d --no-build --no-deps creator-web
|
||||
wait_healthy "$(candidate_container creator-web)"
|
||||
assert_stable_ids_unchanged
|
||||
for service in "${services[@]}"; do
|
||||
image_revision=$(docker inspect -f '{{index .Config.Labels "org.opencontainers.image.revision"}}' "$(candidate_container "$service")" 2>/dev/null || true)
|
||||
[[ "$image_revision" == "$REVISION" ]] || fail_rollback
|
||||
done
|
||||
assert_stable_ids_unchanged
|
||||
|
||||
if [[ "${PREPARE_ONLY:-false}" == true ]]; then
|
||||
echo "production ReleaseSet candidate prepared revision=$REVISION"
|
||||
echo "isolated candidate ReleaseSet prepared slot=$SLOT project=$CANDIDATE_PROJECT revision=$REVISION"
|
||||
exit 0
|
||||
fi
|
||||
|
||||
release_json=$(cat <<JSON
|
||||
{
|
||||
"id":"candidate-$REVISION",
|
||||
"revision":"$REVISION",
|
||||
"components":{
|
||||
"creator-web":{"name":"creator-web","revision":"$REVISION","upstream":"http://creator-web-candidate:8084","health_url":"http://creator-web-candidate:8084/healthz","ready_url":"http://creator-web-candidate:8084/healthz"},
|
||||
"core-runtime":{"name":"core-runtime","revision":"$REVISION","upstream":"http://core-runtime-candidate:8080","health_url":"http://core-runtime-candidate:8080/healthz","ready_url":"http://core-runtime-candidate:8080/readyz"},
|
||||
"cognitive-worker":{"name":"cognitive-worker","revision":"$REVISION","upstream":"http://cognitive-worker-candidate:8083","health_url":"http://cognitive-worker-candidate:8083/healthz","ready_url":"http://cognitive-worker-candidate:8083/readyz"},
|
||||
"research-runtime":{"name":"research-runtime","revision":"$REVISION","upstream":"http://research-runtime-candidate:8087","health_url":"http://research-runtime-candidate:8087/healthz","ready_url":"http://research-runtime-candidate:8087/readyz"}
|
||||
}
|
||||
}
|
||||
{"id":"candidate-$SLOT-$REVISION","slot":"$SLOT","revision":"$REVISION","components":{"creator-web":{"name":"creator-web","revision":"$REVISION","upstream":"http://creator-web-candidate:8084","health_url":"http://creator-web-candidate:8084/healthz","ready_url":"http://creator-web-candidate:8084/healthz"},"core-runtime":{"name":"core-runtime","revision":"$REVISION","upstream":"http://core-runtime-candidate:8080","health_url":"http://core-runtime-candidate:8080/healthz","ready_url":"http://core-runtime-candidate:8080/readyz"},"cognitive-worker":{"name":"cognitive-worker","revision":"$REVISION","upstream":"http://cognitive-worker-candidate:8083","health_url":"http://cognitive-worker-candidate:8083/healthz","ready_url":"http://cognitive-worker-candidate:8083/readyz"},"research-runtime":{"name":"research-runtime","revision":"$REVISION","upstream":"http://research-runtime-candidate:8087","health_url":"http://research-runtime-candidate:8087/healthz","ready_url":"http://research-runtime-candidate:8087/readyz"}}}
|
||||
JSON
|
||||
)
|
||||
|
||||
stable_json=$(curl -fsS --max-time 10 "$SUPERVISOR_CONTROL_URL/v1/state")
|
||||
stable_set=$(printf '%s' "$stable_json" | jq -c '.stable_release_set // empty')
|
||||
if [[ -z "$stable_set" ]]; then
|
||||
stable_revision=$(printf '%s' "$stable_json" | jq -r '.current.version // "unknown"')
|
||||
stable_set=$(cat <<JSON
|
||||
{"id":"stable-$stable_revision","revision":"$stable_revision","components":{
|
||||
"creator-web":{"name":"creator-web","revision":"$stable_revision","upstream":"http://creator-web:8084","health_url":"http://creator-web:8084/healthz","ready_url":"http://creator-web:8084/readyz"},
|
||||
"core-runtime":{"name":"core-runtime","revision":"$stable_revision","upstream":"http://core-runtime:8080","health_url":"http://core-runtime:8080/healthz","ready_url":"http://core-runtime:8080/readyz"},
|
||||
"cognitive-worker":{"name":"cognitive-worker","revision":"$stable_revision","upstream":"http://cognitive-worker:8083","health_url":"http://cognitive-worker:8083/healthz","ready_url":"http://cognitive-worker:8083/readyz"},
|
||||
"research-runtime":{"name":"research-runtime","revision":"$stable_revision","upstream":"http://research-runtime:8087","health_url":"http://research-runtime:8087/healthz","ready_url":"http://research-runtime:8087/readyz"}}}
|
||||
{"id":"stable-$stable_revision","slot":"legacy-stable","revision":"$stable_revision","components":{"creator-web":{"name":"creator-web","revision":"$stable_revision","upstream":"http://creator-web:8084","health_url":"http://creator-web:8084/healthz","ready_url":"http://creator-web:8084/healthz"},"core-runtime":{"name":"core-runtime","revision":"$stable_revision","upstream":"http://core-runtime:8080","health_url":"http://core-runtime:8080/healthz","ready_url":"http://core-runtime:8080/readyz"},"cognitive-worker":{"name":"cognitive-worker","revision":"$stable_revision","upstream":"http://cognitive-worker:8083","health_url":"http://cognitive-worker:8083/healthz","ready_url":"http://cognitive-worker:8083/readyz"},"research-runtime":{"name":"research-runtime","revision":"$stable_revision","upstream":"http://research-runtime:8087","health_url":"http://research-runtime:8087/healthz","ready_url":"http://research-runtime:8087/readyz"}}}
|
||||
JSON
|
||||
)
|
||||
curl -fsS --max-time 10 -X POST "$SUPERVISOR_CONTROL_URL/v1/releases/baseline-set" -H 'Content-Type: application/json' --data "$stable_set" >/dev/null || fail_rollback
|
||||
fi
|
||||
|
||||
curl -fsS --max-time 10 -X POST "$SUPERVISOR_CONTROL_URL/v1/releases/start-set" -H 'Content-Type: application/json' --data "$release_json" >/dev/null || fail_rollback
|
||||
assert_stable_ids_unchanged
|
||||
observation=$(printf '{"release_id":"candidate-%s","ready":true,"reason":"all ReleaseSet healthchecks and image revisions passed","observed_at_unix_millis":%s,"components":{"creator-web":{"revision":"%s","health":true,"ready":true},"core-runtime":{"revision":"%s","health":true,"ready":true},"cognitive-worker":{"revision":"%s","health":true,"ready":true},"research-runtime":{"revision":"%s","health":true,"ready":true}}}\n' "$REVISION" "$(date +%s%3N)" "$REVISION" "$REVISION" "$REVISION" "$REVISION")
|
||||
observation=$(printf '{"release_id":"candidate-%s-%s","ready":true,"reason":"isolated slot health and image revisions passed","observed_at_unix_millis":%s,"components":{"creator-web":{"revision":"%s","health":true,"ready":true},"core-runtime":{"revision":"%s","health":true,"ready":true},"cognitive-worker":{"revision":"%s","health":true,"ready":true},"research-runtime":{"revision":"%s","health":true,"ready":true}}}\n' "$SLOT" "$REVISION" "$(date +%s%3N)" "$REVISION" "$REVISION" "$REVISION" "$REVISION")
|
||||
curl -fsS --max-time 10 -X POST "$SUPERVISOR_CONTROL_URL/v1/releases/observe-set" -H 'Content-Type: application/json' --data "$observation" >/dev/null || fail_rollback
|
||||
assert_stable_ids_unchanged
|
||||
control /v1/traffic/canary || fail_rollback
|
||||
assert_stable_ids_unchanged
|
||||
|
||||
headers=$(mktemp)
|
||||
trap 'rm -f "$headers"' EXIT
|
||||
if ! curl -fsS --max-time 10 -D "$headers" -H 'X-Assistant-Canary: true' "https://assistant.cin.su/login" -o /dev/null; then fail_rollback; fi
|
||||
grep -qi '^X-Assistant-Traffic-Target: candidate' "$headers" || fail_rollback
|
||||
|
||||
if [[ "$PROMOTE_AFTER_CANARY" != true ]]; then
|
||||
echo "isolated candidate ReleaseSet canary ready slot=$SLOT project=$CANDIDATE_PROJECT revision=$REVISION"
|
||||
exit 0
|
||||
fi
|
||||
control /v1/releases/promote || fail_rollback
|
||||
assert_stable_ids_unchanged
|
||||
echo "production ReleaseSet canary promoted revision=$REVISION"
|
||||
echo "isolated candidate ReleaseSet promoted slot=$SLOT project=$CANDIDATE_PROJECT revision=$REVISION"
|
||||
|
||||
@@ -1 +1 @@
|
||||
{"task_file":"dev/tasks/GEN2-COMPAT-OBS-001-production-compatibility-observation.json","task_id":"GEN2-COMPAT-OBS-001","task_revision":2}
|
||||
{"task_file":"dev/tasks/GEN2-COGNITIVE-002-restore-production-cognitive-route.json","task_id":"GEN2-COGNITIVE-002","task_revision":3}
|
||||
|
||||
@@ -0,0 +1,64 @@
|
||||
{
|
||||
"task_id": "GEN2-COGNITIVE-002",
|
||||
"task_revision": 3,
|
||||
"architecture_revision": "ARCH-GEN2-000",
|
||||
"base_repo_revision": "gen2-integration",
|
||||
"risk_tier": "E",
|
||||
"intent": "Восстановить доказанно рабочий production cognitive route для обычного household и memory-recall запроса через существующие Capability и Cognitive Gateway boundaries, без подмены primary route Lifeboat-ом.",
|
||||
"requirement_ids": ["REQ-GEN2-COGNITIVE-002-TRACE", "REQ-GEN2-COGNITIVE-002-ROUTING", "REQ-GEN2-COGNITIVE-002-RECOVERY"],
|
||||
"allowed_change_scope": {
|
||||
"modules": ["capability", "cognitive-worker", "core-runtime", "interaction", "production deployment"],
|
||||
"paths": [
|
||||
"cmd/cognitive-worker/**",
|
||||
"cmd/core-runtime/**",
|
||||
"modules/interaction/domain/gen2_context.go",
|
||||
"modules/interaction/domain/gen2_context_test.go",
|
||||
"adapters/postgres/interaction_repository.go",
|
||||
"contracts/capability/**",
|
||||
"contracts/cognitive/**",
|
||||
"modules/capability/**",
|
||||
"adapters/models/**",
|
||||
"adapters/postgres/capability_repository.go",
|
||||
"deploy/run-production-canary.sh",
|
||||
"deploy/docker-compose.candidate-slot.yml",
|
||||
"tests/**",
|
||||
"docs/evidence/**",
|
||||
"docs/**",
|
||||
"dev/ACTIVE_TASK.json",
|
||||
"dev/tasks/GEN2-COGNITIVE-002-restore-production-cognitive-route.json"
|
||||
]
|
||||
},
|
||||
"forbidden_change_scope": {
|
||||
"paths": ["architecture/**", "security/**", "database/**", "go.mod", "go.sum", ".gitea/workflows/**", "GEN2-028", "GEN2-029", "GEN2-030"]
|
||||
},
|
||||
"allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 1, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
|
||||
"contracts": {
|
||||
"consume": ["existing Capability binding lifecycle", "existing cognitive gateway/advisory contracts", "existing Supervisor ReleaseSet control plane"],
|
||||
"may_change": ["typed capability/provider selection and invocation wiring only within declared paths", "the existing Interaction Gen2 retrieval projection for bounded, scoped human assertions; no new Memory owner or reader"]
|
||||
},
|
||||
"semantic_owner": "Capability remains the sole owner of provider/model/binding lifecycle; cognitive-worker invokes providers through adapters; Core consumes the governed gateway; Supervisor alone controls traffic promotion and rollback.",
|
||||
"canonical_state": ["existing Capability provider/model/binding/evaluation evidence", "canonical Interaction human utterance history and existing typed ProjectReference read through the existing Gen2 retrieval projection", "no new canonical writer, owner, schema, role, or migration"],
|
||||
"state_transitions": ["trace stable and candidate loaded binding heads -> correct stale/divergent lifecycle or invocation defect -> governed repair/qualification -> scoped Interaction projection selects bounded current human assertions -> isolated candidate validation -> Supervisor canary -> promote or rollback"],
|
||||
"side_effects_and_privileges": ["production candidate deployment and Supervisor canary/promotion are explicitly authorized", "adapters/models/** may change only after trace proves a model adapter, provider endpoint, injection, timeout, rate-limit, context budget, or invocation defect there", "cmd/core-runtime/** may change only after trace proves Core-to-worker request construction, route selection, or response handling is causal", "modules/interaction/domain/gen2_context.go may change only to extend the existing bounded Gen2 read projection over canonical Interaction utterances with Principal hard filtering, current applicability and relevance selection; it must not write state, become a Memory owner, or promote PersonalMemory V1 to primary", "adapters/postgres/interaction_repository.go may only read the already-persisted typed interaction.ProjectReference records for the selected canonical Interaction utterances; it may not create schema, write state, or bypass Interaction ownership", "no direct SQL semantic mutation", "no credential, cookie, token, prompt, or user content read/emitted", "Lifeboat may be observed as fallback but cannot mask Local MEDIUM primary failure"],
|
||||
"persistence_and_migrations": ["no schema/migration change is authorized", "adapters/postgres/capability_repository.go may change only after trace proves durable Capability head/state loading diverges from the correct persisted state; reuse its existing repository boundary"],
|
||||
"failure_retry_restart": ["discovery health, invocation health, and availability remain distinct", "CPU/GPU gaps must not suppress a valid Local MEDIUM", "restart reloads a durable active/validated binding head", "failed candidate rolls traffic to Supervisor LKG without stable recreation"],
|
||||
"assumptions": ["isolated candidate slot B and Supervisor controls from GEN2-DEPLOY-003 are available", "an existing authenticated production session may be used only through ordinary runtime interaction", "GEN2-COMPAT-OBS-001 remains paused and V1 paths remain unchanged"],
|
||||
"acceptance_scenarios": [
|
||||
"COGNITIVE-002-SCN-01: factual stable/candidate trace records Core request, requirement, binding selection, loaded binding head, fleet availability, provider eligibility, invocation result, and response/error without sensitive content",
|
||||
"COGNITIVE-002-SCN-02: Local MEDIUM is shown ACTIVE, VALIDATED, AVAILABLE, and actually invoked for bounded OK, Russian household, and prepared-context recall requests",
|
||||
"COGNITIVE-002-SCN-03: health endpoint success alone is not used as inference proof; discovery and invocation failures are separately classified",
|
||||
"COGNITIVE-002-SCN-04: candidate four-component health/revision check, real canary household request, scoped bounded arbitrary-assertion memory write/multiple turns/recall, and candidate Core plus cognitive-worker restart/recovery pass",
|
||||
"COGNITIVE-002-SCN-05: only after candidate success, Supervisor promotes it; evidence retains LKG and rollback readiness",
|
||||
"COGNITIVE-002-EXIT: ordinary household and memory-recall response use Local MEDIUM through the governed route and remain usable after candidate restart"
|
||||
],
|
||||
"required_tests": [
|
||||
"FAST: targeted go test ./cmd/cognitive-worker/... ./cmd/core-runtime/... ./modules/capability/... ./adapters/models/... ./adapters/postgres/... only for changed packages",
|
||||
"FAST: targeted go vet for changed packages",
|
||||
"FAST governance: go run ./cmd/devctl preflight -active dev/ACTIVE_TASK.json",
|
||||
"FAST scope: go run ./cmd/devctl check-diff -active dev/ACTIVE_TASK.json",
|
||||
"FAST independent: go run ./cmd/devctl/independent",
|
||||
"FAST whitespace: git diff --check gen2-integration --",
|
||||
"production: stable/candidate trace, isolated candidate health/canary, Local MEDIUM invocation, restart recovery, and Supervisor promotion/rollback evidence"
|
||||
],
|
||||
"terminal_gap_states": ["BLOCKED_BY_SPEC_GAP", "BLOCKED_BY_ARCHITECTURE_GAP", "BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT", "REQUIRES_MIGRATION_DECISION"]
|
||||
}
|
||||
@@ -0,0 +1,59 @@
|
||||
{
|
||||
"task_id": "GEN2-DEPLOY-003",
|
||||
"task_revision": 1,
|
||||
"architecture_revision": "ARCH-GEN2-000",
|
||||
"base_repo_revision": "gen2-integration",
|
||||
"risk_tier": "E",
|
||||
"intent": "Изолировать stable и candidate ReleaseSet в независимые Compose deployment projects/slots, чтобы подготовка candidate не могла пересоздать stable containers, сохранив Supervisor как единственный traffic/promotion/rollback control plane.",
|
||||
"requirement_ids": ["REQ-GEN2-DEPLOY-003-SLOT-ISOLATION", "REQ-GEN2-DEPLOY-003-UPDATER-OWNERSHIP", "REQ-GEN2-DEPLOY-003-ROLLBACK"],
|
||||
"allowed_change_scope": {
|
||||
"modules": ["deployment control", "runtime supervisor"],
|
||||
"paths": [
|
||||
"deploy/**",
|
||||
"runtime/supervisor/**",
|
||||
"cmd/supervisor/**",
|
||||
"docs/evidence/**",
|
||||
"docs/DEPLOYMENT.md",
|
||||
"tests/deployment/**",
|
||||
"dev/ACTIVE_TASK.json",
|
||||
"dev/tasks/GEN2-DEPLOY-003-isolated-release-slots.json"
|
||||
]
|
||||
},
|
||||
"forbidden_change_scope": {
|
||||
"paths": ["architecture/**", "database/**", "modules/**", "contracts/**", "security/**", "adapters/**", "go.mod", "go.sum", ".gitea/workflows/**", "GEN2-028", "GEN2-029", "GEN2-030"]
|
||||
},
|
||||
"allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 1, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
|
||||
"contracts": {
|
||||
"consume": ["existing Supervisor durable ReleaseSet state", "existing production compose stack", "existing shared durable PostgreSQL/volumes"],
|
||||
"may_change": ["deployment slot topology", "deployment-controller scripts", "Supervisor operational ReleaseSet identity fields"]
|
||||
},
|
||||
"semantic_owner": "Operations/Supervisor; slot and ReleaseSet identity are runtime-only operational control-plane state, not semantic canonical state.",
|
||||
"canonical_state": ["no semantic owner, canonical writer, schema or data migration change", "Supervisor durable release evidence/LKG remains retained outside candidate containers"],
|
||||
"state_transitions": ["stable slot retained -> isolated candidate slot start -> four-component health/revision validation -> Supervisor canary -> promote or traffic rollback -> inactive slot retained for rollback"],
|
||||
"side_effects_and_privileges": ["production operations only through Docker/Compose and Supervisor control plane explicitly authorized for this remediation", "candidate preparation must never directly replace/recreate stable", "no credential reads or outputs", "external updater pause is allowed only as an explicit bounded operation with restore evidence"],
|
||||
"persistence_and_migrations": ["no database migration", "shared PostgreSQL, named volumes and durable services are attached as existing external resources and are never duplicated or deleted by candidate cleanup"],
|
||||
"failure_retry_restart": ["capture stable container IDs before every candidate operation and abort before canary when any changes", "candidate component failure returns traffic to LKG without container recreation", "Supervisor restart reloads ReleaseSet/slot/LKG evidence", "updater ownership must be project-scoped so it cannot reconcile the other slot"],
|
||||
"assumptions": ["current stable may remain in its existing project during the one-time transition; it must not be recreated merely to rename a slot", "external deployment controller can be changed only when its exact owner/process is evidenced", "GEN2-COMPAT-OBS-001, V1 retirement and Gen3 remain stopped until this task passes"],
|
||||
"acceptance_scenarios": [
|
||||
"DEPLOY-003-SCN-01: stable and candidate Compose projects, service/container identities and updater ownership are distinct",
|
||||
"DEPLOY-003-SCN-02: candidate dependencies resolve only to candidate services plus declared shared external resources",
|
||||
"DEPLOY-003-SCN-03: candidate operation leaves every captured stable container ID unchanged; deviation aborts before canary",
|
||||
"DEPLOY-003-SCN-04: all four candidate components report health/readiness and one revision",
|
||||
"DEPLOY-003-SCN-05: Supervisor records ReleaseSet slot identity/revision/LKG and canaries the candidate",
|
||||
"DEPLOY-003-SCN-06: forced candidate component failure rolls all traffic back to previous stable without stable recreate",
|
||||
"DEPLOY-003-SCN-07: Supervisor restart preserves LKG/rollback target and candidate recovery does not alter stable IDs",
|
||||
"DEPLOY-003-EXIT: isolated production ReleaseSet drill passes, after which GEN2-COMPAT-OBS-001 may resume without V1 retirement"
|
||||
],
|
||||
"required_tests": [
|
||||
"FAST: compose config for stable and candidate slots",
|
||||
"FAST: targeted go test ./runtime/supervisor ./cmd/supervisor",
|
||||
"FAST: targeted go vet ./runtime/supervisor ./cmd/supervisor",
|
||||
"FAST: deployment shell regression for project/identity isolation and stable ID invariance",
|
||||
"FAST governance: go run ./cmd/devctl preflight -active dev/ACTIVE_TASK.json",
|
||||
"FAST scope: go run ./cmd/devctl check-diff -active dev/ACTIVE_TASK.json",
|
||||
"FAST independent: go run ./cmd/devctl/independent",
|
||||
"FAST whitespace: git diff --check gen2-integration --",
|
||||
"PRODUCTION: isolated candidate ReleaseSet health, canary, forced rollback, Supervisor restart/LKG and stable-ID drill"
|
||||
],
|
||||
"terminal_gap_states": ["BLOCKED_BY_DEPLOYMENT_CONTROL", "BLOCKED_BY_SPEC_GAP", "BLOCKED_BY_ARCHITECTURE_GAP", "BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT", "REQUIRES_MIGRATION_DECISION"]
|
||||
}
|
||||
@@ -9,8 +9,9 @@
|
||||
- `PROVIDER_GAPS = CPU_FAST_UNAVAILABLE; GPU_UNAVAILABLE`; Local MEDIUM remains the primary route and Lifeboat remains fallback per existing M5 evidence.
|
||||
- `GEN2_M6_FINAL_STATUS = NOT_FINAL_PASS_RETIREMENT_GAP_REMAINS`; FULL_FINAL technical checks passed, but no final M6 baseline tag is created because active V1 compatibility callers were not proven safe to retire.
|
||||
- `GEN2_M6_FINAL_EVIDENCE = docs/evidence/gen2-030-final-acceptance-2026-08-30.json`.
|
||||
- `GEN2_FINAL_STATUS = GEN2_CORE_ACCEPTED / V1_COMPATIBILITY_RETAINED / UI_DEFERRED`.
|
||||
- `GEN2-COMPAT-OBS-001 = INSTRUMENTATION_COMMITTED; PRODUCTION_OBSERVATION_BLOCKED_BY_MISSING_CAPABILITY`; evidence: `docs/evidence/gen2-compat-obs-001-blocked-2026-08-30.json`. No V1 path is `SAFE_TO_RETIRE`; no retirement task was created.
|
||||
- `GEN2-COGNITIVE-002 = IMPLEMENTED / CORE_VERIFIED / BROWSER_ACCEPTANCE_DEFERRED_BY_OWNER`; non-browser evidence: `docs/evidence/gen2-cognitive-002-core-verified-2026-08-30.json`.
|
||||
- `GEN2-COMPAT-OBS-001 = OBSERVATION_DEFERRED_BY_OWNER`; instrumentation remains committed, but no production observation window or retirement decision is claimed. No V1 path is `SAFE_TO_RETIRE`; no retirement task was created.
|
||||
- `GEN2_FINAL_STATUS = GEN2_CORE_ACCEPTED / V1_COMPATIBILITY_RETAINED / BROWSER_ACCEPTANCE_DEFERRED / COMPAT_OBSERVATION_DEFERRED`.
|
||||
- `NEXT_EVOLUTION_QUEUE = instrument bounded compatibility usage, observe, then retire only individually proven paths under a future FULL_RETIREMENT gate; no Gen3 or new feature work is started.
|
||||
|
||||
## Current audit delta — 2026-08-24
|
||||
|
||||
@@ -0,0 +1,43 @@
|
||||
{
|
||||
"evidence_id": "gen2-cognitive-002-core-verified-2026-08-30",
|
||||
"task_id": "GEN2-COGNITIVE-002",
|
||||
"task_revision": 3,
|
||||
"commit": "00af176a200142bd3de67ba914e542b2299f43d5",
|
||||
"status": "IMPLEMENTED / CORE_VERIFIED / BROWSER_ACCEPTANCE_DEFERRED_BY_OWNER",
|
||||
"owner_amendment": {
|
||||
"browser_authenticated_acceptance": "DEFERRED_BY_OWNER",
|
||||
"compatibility_observation": "DEFERRED_BY_OWNER",
|
||||
"v1_retirement": "NOT_AUTHORIZED"
|
||||
},
|
||||
"non_browser_acceptance": {
|
||||
"targeted_tests": "PASS: cmd/cognitive-worker, cmd/core-runtime, modules/interaction/domain, adapters/postgres",
|
||||
"targeted_vet": "PASS: same packages",
|
||||
"governance": ["preflight PASS", "check-diff PASS", "independent PASS", "git diff --check PASS"],
|
||||
"scoped_retrieval": [
|
||||
"arbitrary assertion retrieval",
|
||||
"current value selected while historical value is retained",
|
||||
"ProjectReference hard scope isolation before relevance ranking",
|
||||
"irrelevant assertions excluded from bounded context"
|
||||
],
|
||||
"context_compiler": "PASS: bounded typed Interaction fragments are compiled before the cognitive gateway request",
|
||||
"candidate_slot_b": {
|
||||
"components": "creator-web, core-runtime, cognitive-worker, research-runtime healthy",
|
||||
"restart": "candidate core-runtime and cognitive-worker restarted; both health and ready endpoints recovered",
|
||||
"local_medium": "governed gateway advisory completed after restart"
|
||||
}
|
||||
},
|
||||
"preserved_invariants": [
|
||||
"Capability remains canonical owner of binding lifecycle",
|
||||
"Interaction remains canonical owner of utterances and ProjectReference",
|
||||
"no direct semantic SQL update",
|
||||
"PersonalMemory V1 is not restored as primary retrieval",
|
||||
"Lifeboat is fallback, not proof of Local MEDIUM"
|
||||
],
|
||||
"not_claimed": [
|
||||
"browser/authenticated production acceptance",
|
||||
"production promotion",
|
||||
"compatibility telemetry observation",
|
||||
"V1 retirement readiness"
|
||||
],
|
||||
"rollback": "Supervisor release rollback returns traffic to the stable last-known-good ReleaseSet; stable containers are not recreated."
|
||||
}
|
||||
@@ -0,0 +1,51 @@
|
||||
{
|
||||
"task": "GEN2-DEPLOY-003",
|
||||
"status": "PASS",
|
||||
"timestamp_utc": "2026-08-30T11:09:35Z",
|
||||
"root_cause": {
|
||||
"stable_and_candidate_project": "assistant",
|
||||
"stable_and_candidate_working_directory": "/home/m/assistant-deploy/repo/deploy",
|
||||
"updater": "/home/m/assistant-deploy/update.sh",
|
||||
"updater_schedule": "cron every five minutes",
|
||||
"unsafe_operation": "the updater ran full docker compose up for the stable project and invoked the old in-project candidate canary"
|
||||
},
|
||||
"remediation": {
|
||||
"stable_project": "assistant",
|
||||
"candidate_project": "assistant-slot-b",
|
||||
"candidate_file": "deploy/docker-compose.candidate-slot.yml",
|
||||
"shared_resources": ["assistant_default external network", "llm-local_default external network", "existing secret and research volumes"],
|
||||
"updater_ownership": "stable-only updater; candidate release command is not invoked by the updater",
|
||||
"supervisor_release_set_slot": "ReleaseSet.slot persisted in durable Supervisor state"
|
||||
},
|
||||
"candidate_revision": "216d76f1d8b4eeba6ee9427da4c87c3e5c728c02",
|
||||
"stable_revision": "2ee5da81145cdad2da1c92c24a91145ad4af61a5",
|
||||
"candidate_components": {
|
||||
"creator-web": "healthy",
|
||||
"core-runtime": "healthy",
|
||||
"cognitive-worker": "healthy",
|
||||
"research-runtime": "healthy"
|
||||
},
|
||||
"drill": {
|
||||
"stable_ids_captured_before_candidate": true,
|
||||
"stable_ids_unchanged_after_candidate_start": true,
|
||||
"header_canary": "candidate slot b",
|
||||
"forced_failure": "candidate cognitive-worker stopped",
|
||||
"rollback": "Supervisor release rollback returned traffic to legacy stable without recreating stable containers",
|
||||
"stable_ids_unchanged_after_rollback": true,
|
||||
"supervisor_restart": "last-known-good legacy stable release set preserved",
|
||||
"candidate_recovery": "candidate cognitive-worker restored healthy",
|
||||
"stable_ids_unchanged_after_candidate_recovery": true
|
||||
},
|
||||
"final_state": {
|
||||
"traffic_mode": "STABLE",
|
||||
"stable_upstream": "http://creator-web:8084",
|
||||
"last_known_good_slot": "legacy-stable",
|
||||
"public_health": 200,
|
||||
"public_ready": 200
|
||||
},
|
||||
"limits": [
|
||||
"The existing stable project remains legacy-stable during the non-disruptive transition; it was not recreated merely to rename it.",
|
||||
"Local MEDIUM discovery was slow/unavailable during candidate cognitive startup; it delayed startup but did not affect stable traffic or the isolated release-set drill.",
|
||||
"GEN2-COMPAT-OBS-001 has not resumed and no V1 path has been retired."
|
||||
]
|
||||
}
|
||||
@@ -1,11 +1,31 @@
|
||||
package domain
|
||||
|
||||
import "git.cin.su/m/Assistant/contracts/interaction"
|
||||
import (
|
||||
"sort"
|
||||
"strings"
|
||||
|
||||
"git.cin.su/m/Assistant/contracts/interaction"
|
||||
)
|
||||
|
||||
const (
|
||||
gen2ContextMaxCandidates = 4
|
||||
gen2ContextMaxBytes = 2048
|
||||
)
|
||||
|
||||
type ContextFragment struct {
|
||||
UtteranceID interaction.HumanUtteranceID
|
||||
Text string
|
||||
ScopeRef string
|
||||
Relevance int
|
||||
Historical bool
|
||||
}
|
||||
|
||||
type ContextProjection struct {
|
||||
Text string
|
||||
Source string
|
||||
UsedPersonalMemoryFallback bool
|
||||
Fragments []ContextFragment
|
||||
Historical []ContextFragment
|
||||
}
|
||||
|
||||
// BuildGen2Context is a read-side projection over canonical utterance history.
|
||||
@@ -19,6 +39,14 @@ func BuildGen2Context(history []interaction.HumanUtterance) ContextProjection {
|
||||
// filter prevents accidental cross-principal projection when a mixed source
|
||||
// is passed during degraded operation.
|
||||
func BuildScopedGen2Context(history []interaction.HumanUtterance, principalID string) ContextProjection {
|
||||
return BuildScopedGen2ContextForQuery(history, nil, principalID, "", "")
|
||||
}
|
||||
|
||||
// BuildScopedGen2ContextForQuery projects bounded, explicit human assertions
|
||||
// from the canonical Interaction history. Principal and typed project refs are
|
||||
// hard filters before relevance ranking. Older values remain lineage in
|
||||
// Historical and are never represented as the current value.
|
||||
func BuildScopedGen2ContextForQuery(history []interaction.HumanUtterance, references []interaction.ProjectReference, principalID, focusProjectID, query string) ContextProjection {
|
||||
if principalID != "" {
|
||||
scoped := make([]interaction.HumanUtterance, 0, len(history))
|
||||
for _, item := range history {
|
||||
@@ -28,9 +56,182 @@ func BuildScopedGen2Context(history []interaction.HumanUtterance, principalID st
|
||||
}
|
||||
history = scoped
|
||||
}
|
||||
projectRefs := make(map[interaction.HumanUtteranceID][]interaction.ProjectReference, len(references))
|
||||
for _, ref := range references {
|
||||
projectRefs[ref.UtteranceID] = append(projectRefs[ref.UtteranceID], ref)
|
||||
}
|
||||
assertions := explicitAssertionCandidates(history, projectRefs, focusProjectID, query)
|
||||
if len(assertions) > 0 {
|
||||
current, historical := selectCurrentAssertions(assertions)
|
||||
return contextProjectionFromAssertions(current, historical)
|
||||
}
|
||||
text := CompilePersonalMemory(history).ContextText()
|
||||
if text != "" {
|
||||
return ContextProjection{Text: text, Source: "retrieval:interaction-utterances"}
|
||||
}
|
||||
return ContextProjection{Text: text, Source: "personal-memory-v1-fallback", UsedPersonalMemoryFallback: true}
|
||||
}
|
||||
|
||||
type assertionCandidate struct {
|
||||
fragment ContextFragment
|
||||
subject string
|
||||
observed int64
|
||||
}
|
||||
|
||||
func explicitAssertionCandidates(history []interaction.HumanUtterance, projectRefs map[interaction.HumanUtteranceID][]interaction.ProjectReference, focusProjectID, query string) []assertionCandidate {
|
||||
queryTokens := significantContextTokens(query)
|
||||
result := make([]assertionCandidate, 0)
|
||||
for _, utterance := range history {
|
||||
subject, value, ok := explicitAssertion(utterance.Content)
|
||||
if !ok {
|
||||
continue
|
||||
}
|
||||
refs := projectRefs[utterance.ID]
|
||||
scopeRef := ""
|
||||
if focusProjectID != "" {
|
||||
matched := false
|
||||
for _, ref := range refs {
|
||||
if string(ref.ProjectID) == focusProjectID {
|
||||
matched = true
|
||||
scopeRef = focusProjectID
|
||||
break
|
||||
}
|
||||
}
|
||||
if !matched {
|
||||
continue
|
||||
}
|
||||
} else if len(refs) == 1 {
|
||||
scopeRef = string(refs[0].ProjectID)
|
||||
}
|
||||
relevance := contextRelevance(subject, queryTokens)
|
||||
if len(queryTokens) > 0 && relevance == 0 {
|
||||
continue
|
||||
}
|
||||
result = append(result, assertionCandidate{
|
||||
fragment: ContextFragment{UtteranceID: utterance.ID, Text: subject + " = " + value, ScopeRef: scopeRef, Relevance: relevance},
|
||||
subject: normalizedContextSubject(subject),
|
||||
observed: utterance.ReceivedAtUnixMillis,
|
||||
})
|
||||
}
|
||||
return result
|
||||
}
|
||||
|
||||
func selectCurrentAssertions(candidates []assertionCandidate) ([]ContextFragment, []ContextFragment) {
|
||||
latest := make(map[string]assertionCandidate)
|
||||
for _, candidate := range candidates {
|
||||
current, exists := latest[candidate.subject]
|
||||
if !exists || candidate.observed > current.observed || candidate.observed == current.observed && candidate.fragment.UtteranceID > current.fragment.UtteranceID {
|
||||
latest[candidate.subject] = candidate
|
||||
}
|
||||
}
|
||||
current := make([]ContextFragment, 0, len(latest))
|
||||
historical := make([]ContextFragment, 0, len(candidates)-len(latest))
|
||||
for _, candidate := range candidates {
|
||||
selected := latest[candidate.subject]
|
||||
if candidate.fragment.UtteranceID == selected.fragment.UtteranceID {
|
||||
current = append(current, candidate.fragment)
|
||||
continue
|
||||
}
|
||||
candidate.fragment.Historical = true
|
||||
historical = append(historical, candidate.fragment)
|
||||
}
|
||||
sort.SliceStable(current, func(i, j int) bool {
|
||||
if current[i].Relevance != current[j].Relevance {
|
||||
return current[i].Relevance > current[j].Relevance
|
||||
}
|
||||
return current[i].UtteranceID > current[j].UtteranceID
|
||||
})
|
||||
return current, historical
|
||||
}
|
||||
|
||||
func contextProjectionFromAssertions(current, historical []ContextFragment) ContextProjection {
|
||||
selected := make([]ContextFragment, 0, len(current))
|
||||
text := make([]string, 0, len(current))
|
||||
usedBytes := 0
|
||||
for _, candidate := range current {
|
||||
if len(selected) == gen2ContextMaxCandidates {
|
||||
break
|
||||
}
|
||||
bytes := len([]byte(candidate.Text))
|
||||
if bytes == 0 || usedBytes+bytes > gen2ContextMaxBytes {
|
||||
continue
|
||||
}
|
||||
usedBytes += bytes
|
||||
selected = append(selected, candidate)
|
||||
text = append(text, candidate.Text)
|
||||
}
|
||||
if len(selected) == 0 {
|
||||
return ContextProjection{Source: "retrieval:interaction-assertions", Historical: historical}
|
||||
}
|
||||
return ContextProjection{Text: strings.Join(text, "\n"), Source: "retrieval:interaction-assertions", Fragments: selected, Historical: historical}
|
||||
}
|
||||
|
||||
func explicitAssertion(message string) (string, string, bool) {
|
||||
text := strings.TrimSpace(StripAssistantVocative(message))
|
||||
if text == "" || strings.Contains(text, "?") {
|
||||
return "", "", false
|
||||
}
|
||||
lower := strings.ToLower(text)
|
||||
if index := strings.Index(lower, "запомни:"); index >= 0 {
|
||||
text = strings.TrimSpace(text[index+len("запомни:"):])
|
||||
lower = strings.ToLower(text)
|
||||
}
|
||||
for _, prefix := range []string{"запомни:", "запомни", "помни:"} {
|
||||
if strings.HasPrefix(lower, prefix) {
|
||||
text = strings.TrimSpace(text[len(prefix):])
|
||||
break
|
||||
}
|
||||
}
|
||||
for _, separator := range []string{"=", "—"} {
|
||||
subject, value, found := strings.Cut(text, separator)
|
||||
if !found {
|
||||
continue
|
||||
}
|
||||
subject, value = strings.TrimSpace(subject), strings.TrimSpace(value)
|
||||
if subject != "" && value != "" && len([]rune(subject)) <= 160 && len([]rune(value)) <= 512 {
|
||||
return subject, value, true
|
||||
}
|
||||
}
|
||||
return "", "", false
|
||||
}
|
||||
|
||||
func normalizedContextSubject(subject string) string {
|
||||
return strings.Join(significantContextTokens(subject), " ")
|
||||
}
|
||||
|
||||
func significantContextTokens(value string) []string {
|
||||
ignored := map[string]struct{}{"какой": {}, "какая": {}, "какое": {}, "какие": {}, "текущий": {}, "текущая": {}, "текущее": {}, "я": {}, "ты": {}, "назвал": {}, "назвала": {}, "сообщил": {}, "сообщила": {}}
|
||||
result := make([]string, 0)
|
||||
seen := make(map[string]struct{})
|
||||
for _, token := range semanticTokens(value) {
|
||||
if len([]rune(token)) < 3 {
|
||||
continue
|
||||
}
|
||||
if _, skip := ignored[token]; skip {
|
||||
continue
|
||||
}
|
||||
if _, duplicate := seen[token]; duplicate {
|
||||
continue
|
||||
}
|
||||
seen[token] = struct{}{}
|
||||
result = append(result, token)
|
||||
}
|
||||
return result
|
||||
}
|
||||
|
||||
func contextRelevance(subject string, queryTokens []string) int {
|
||||
if len(queryTokens) == 0 {
|
||||
return 1
|
||||
}
|
||||
subjectTokens := significantContextTokens(subject)
|
||||
matched := 0
|
||||
for _, query := range queryTokens {
|
||||
for _, token := range subjectTokens {
|
||||
if query == token {
|
||||
matched++
|
||||
break
|
||||
}
|
||||
}
|
||||
}
|
||||
return matched
|
||||
}
|
||||
|
||||
@@ -32,7 +32,61 @@ func TestGen2ContextProjectsExplicitRememberedAssertion(t *testing.T) {
|
||||
got := domain.BuildScopedGen2Context([]interaction.HumanUtterance{
|
||||
{ID: "assertion", PrincipalID: "principal-1", Content: "Для этого разговора запомни: мой тестовый цвет — синий.", ReceivedAtUnixMillis: 1},
|
||||
}, "principal-1")
|
||||
if got.Source != "retrieval:interaction-utterances" || got.UsedPersonalMemoryFallback || !strings.Contains(got.Text, "тестовый цвет — синий") {
|
||||
if got.Source != "retrieval:interaction-assertions" || got.UsedPersonalMemoryFallback || !strings.Contains(got.Text, "тестовый цвет = синий") {
|
||||
t.Fatalf("remembered assertion was not projected: %#v", got)
|
||||
}
|
||||
}
|
||||
|
||||
func TestGen2ContextSelectsArbitraryRelevantCurrentAssertion(t *testing.T) {
|
||||
history := []interaction.HumanUtterance{
|
||||
{ID: "assertion", PrincipalID: "principal-1", Content: "контрольный код = ORBIT-731", ReceivedAtUnixMillis: 1},
|
||||
{ID: "turn-1", PrincipalID: "principal-1", Content: "Спасибо", ReceivedAtUnixMillis: 2},
|
||||
{ID: "turn-2", PrincipalID: "principal-1", Content: "Продолжим", ReceivedAtUnixMillis: 3},
|
||||
}
|
||||
got := domain.BuildScopedGen2ContextForQuery(history, nil, "principal-1", "", "Какой контрольный код я назвал?")
|
||||
if got.Source != "retrieval:interaction-assertions" || !strings.Contains(got.Text, "ORBIT-731") || len(got.Fragments) != 1 {
|
||||
t.Fatalf("arbitrary assertion was not selected: %#v", got)
|
||||
}
|
||||
}
|
||||
|
||||
func TestGen2ContextKeepsHistoricalAssertionOutOfCurrentProjection(t *testing.T) {
|
||||
history := []interaction.HumanUtterance{
|
||||
{ID: "old", PrincipalID: "principal-1", Content: "контрольный код = OLD-111", ReceivedAtUnixMillis: 1},
|
||||
{ID: "new", PrincipalID: "principal-1", Content: "контрольный код = NEW-222", ReceivedAtUnixMillis: 2},
|
||||
}
|
||||
got := domain.BuildScopedGen2ContextForQuery(history, nil, "principal-1", "", "Какой контрольный код?")
|
||||
if strings.Contains(got.Text, "OLD-111") || !strings.Contains(got.Text, "NEW-222") || len(got.Historical) != 1 || !strings.Contains(got.Historical[0].Text, "OLD-111") {
|
||||
t.Fatalf("current/historical assertion separation failed: %#v", got)
|
||||
}
|
||||
}
|
||||
|
||||
func TestGen2ContextHardFiltersProjectReferencesBeforeRanking(t *testing.T) {
|
||||
history := []interaction.HumanUtterance{
|
||||
{ID: "a", PrincipalID: "principal-1", Content: "контрольный код = ALPHA", ReceivedAtUnixMillis: 1},
|
||||
{ID: "b", PrincipalID: "principal-1", Content: "контрольный код = BETA", ReceivedAtUnixMillis: 2},
|
||||
}
|
||||
references := []interaction.ProjectReference{
|
||||
{ID: "ref-a", UtteranceID: "a", ProjectID: "company-a"},
|
||||
{ID: "ref-b", UtteranceID: "b", ProjectID: "company-b"},
|
||||
}
|
||||
for projectID, expected := range map[string]struct{ want, absent string }{
|
||||
"company-a": {want: "ALPHA", absent: "BETA"},
|
||||
"company-b": {want: "BETA", absent: "ALPHA"},
|
||||
} {
|
||||
got := domain.BuildScopedGen2ContextForQuery(history, references, "principal-1", projectID, "Какой контрольный код?")
|
||||
if !strings.Contains(got.Text, expected.want) || strings.Contains(got.Text, expected.absent) || len(got.Fragments) != 1 || got.Fragments[0].ScopeRef != projectID {
|
||||
t.Fatalf("scope %s leaked context: %#v", projectID, got)
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
func TestGen2ContextExcludesIrrelevantAssertionsFromBoundedProjection(t *testing.T) {
|
||||
history := []interaction.HumanUtterance{
|
||||
{ID: "code", PrincipalID: "principal-1", Content: "контрольный код = ORBIT-731", ReceivedAtUnixMillis: 1},
|
||||
{ID: "city", PrincipalID: "principal-1", Content: "город встречи = Казань", ReceivedAtUnixMillis: 2},
|
||||
}
|
||||
got := domain.BuildScopedGen2ContextForQuery(history, nil, "principal-1", "", "Какой контрольный код?")
|
||||
if strings.Contains(got.Text, "Казань") || len(got.Fragments) != 1 {
|
||||
t.Fatalf("irrelevant assertion entered bounded projection: %#v", got)
|
||||
}
|
||||
}
|
||||
|
||||
@@ -46,7 +46,10 @@ type ComponentRelease struct {
|
||||
|
||||
// ReleaseSet is the atomic deployment unit supervised by the control plane.
|
||||
type ReleaseSet struct {
|
||||
ID string `json:"id"`
|
||||
ID string `json:"id"`
|
||||
// Slot identifies the independently managed deployment slot. An empty value
|
||||
// remains readable for release state recorded before slot isolation.
|
||||
Slot string `json:"slot,omitempty"`
|
||||
Revision string `json:"revision"`
|
||||
Components map[string]ComponentRelease `json:"components"`
|
||||
}
|
||||
@@ -207,6 +210,9 @@ func (l *Lifeboat) PromoteSet() bool {
|
||||
l.Current.Set = &set
|
||||
l.Current.LastKnownGood = true
|
||||
l.CandidateReleaseSet = nil
|
||||
// The promoted ReleaseSet is now the stable target. Leaving header-canary
|
||||
// mode active would keep an obsolete candidate route in the public proxy.
|
||||
l.TrafficMode = TrafficStable
|
||||
return l.setSetUpstreams(&set, nil)
|
||||
}
|
||||
|
||||
|
||||
@@ -110,7 +110,9 @@ func TestPersistentReleaseSetRollbackPreservesWholeLastKnownGood(t *testing.T) {
|
||||
}
|
||||
first := supervisor.NewPersistentLifeboat(store)
|
||||
good := testReleaseSet("stable-r1", "r1", "http://web", "http://core", "http://cognitive", "http://research")
|
||||
good.Slot = "a"
|
||||
candidate := testReleaseSet("candidate-r2", "r2", "http://web-c", "http://core-c", "http://cognitive-c", "http://research-c")
|
||||
candidate.Slot = "b"
|
||||
if ok, err := first.EstablishBaselineSet(good); err != nil || !ok {
|
||||
t.Fatalf("baseline: %v %v", ok, err)
|
||||
}
|
||||
@@ -135,7 +137,7 @@ func TestPersistentReleaseSetRollbackPreservesWholeLastKnownGood(t *testing.T) {
|
||||
if err != nil {
|
||||
t.Fatal(err)
|
||||
}
|
||||
if state.StableReleaseSet == nil || state.StableReleaseSet.Revision != "r1" || state.TrafficMode != supervisor.TrafficStable {
|
||||
if state.StableReleaseSet == nil || state.StableReleaseSet.Revision != "r1" || state.StableReleaseSet.Slot != "a" || state.TrafficMode != supervisor.TrafficStable {
|
||||
t.Fatalf("state=%+v", state)
|
||||
}
|
||||
}
|
||||
@@ -154,6 +156,9 @@ func TestReleaseSetPromotionRetainsPreviousStableForRollback(t *testing.T) {
|
||||
if !l.RecordSet(observation) || !l.PromoteSet() {
|
||||
t.Fatal("promotion failed")
|
||||
}
|
||||
if l.TrafficMode != supervisor.TrafficStable {
|
||||
t.Fatalf("promoted ReleaseSet traffic=%s, want stable", l.TrafficMode)
|
||||
}
|
||||
if l.RollbackTargetReleaseSet == nil || l.RollbackTargetReleaseSet.Revision != "r1" {
|
||||
t.Fatalf("rollback target=%+v", l.RollbackTargetReleaseSet)
|
||||
}
|
||||
|
||||
@@ -4,10 +4,16 @@ set -euo pipefail
|
||||
root="$(CDPATH= cd -- "$(dirname -- "$0")/../.." && pwd)"
|
||||
canary="$root/deploy/run-production-canary.sh"
|
||||
supervisor="$root/deploy/update-supervisor.sh"
|
||||
candidate_compose="$root/deploy/docker-compose.candidate-slot.yml"
|
||||
stable_compose="$root/deploy/docker-compose.prod.yml"
|
||||
stable_updater="$root/deploy/production-stable-updater.sh"
|
||||
|
||||
grep -Fq 'up -d --no-build --no-deps research-runtime-candidate core-runtime-candidate' "$canary"
|
||||
grep -Fq 'up -d --no-build --no-deps cognitive-worker-candidate' "$canary"
|
||||
grep -Fq 'up -d --no-build --no-deps creator-web-candidate' "$canary"
|
||||
grep -Fq 'CANDIDATE_PROJECT="${CANDIDATE_PROJECT:-assistant-slot-$SLOT}"' "$canary"
|
||||
grep -Fq '[[ "$CANDIDATE_PROJECT" != "$STABLE_PROJECT" ]]' "$canary"
|
||||
grep -Fq 'docker compose -p "$CANDIDATE_PROJECT"' "$canary"
|
||||
grep -Fq 'up -d --no-build --no-deps research-runtime core-runtime' "$canary"
|
||||
grep -Fq 'up -d --no-build --no-deps cognitive-worker' "$canary"
|
||||
grep -Fq 'up -d --no-build --no-deps creator-web' "$canary"
|
||||
grep -Fq 'compose up -d --no-build --no-deps supervisor' "$supervisor"
|
||||
grep -Fq 'stable_ids_before=$(stable_ids)' "$canary"
|
||||
grep -Fq 'assert_stable_ids_unchanged' "$canary"
|
||||
@@ -16,4 +22,11 @@ grep -Fq 'after=$(stable_ids)' "$supervisor"
|
||||
grep -Fq 'stable container IDs changed' "$supervisor"
|
||||
grep -Fq 'HEALTH_ATTEMPTS="${HEALTH_ATTEMPTS:-360}"' "$canary"
|
||||
grep -Fq 'seq 1 "$HEALTH_ATTEMPTS"' "$canary"
|
||||
grep -Fq 'PROMOTE_AFTER_CANARY="${PROMOTE_AFTER_CANARY:-true}"' "$canary"
|
||||
grep -Fq 'if [[ "$PROMOTE_AFTER_CANARY" != true ]]' "$canary"
|
||||
grep -Fq 'name: assistant_default' "$candidate_compose"
|
||||
grep -Fq 'external: true' "$candidate_compose"
|
||||
! grep -Fq 'candidate:' "$stable_compose"
|
||||
! grep -Fq 'run-production-canary.sh' "$stable_updater"
|
||||
grep -Fq 'compose up -d --no-build "${stable_services[@]}"' "$stable_updater"
|
||||
echo 'stable release-set guard: PASS'
|
||||
|
||||
Reference in New Issue
Block a user