Files
Assistant/cmd/cognitive-worker/main.go
m 6ecb431ccd
CI / test (push) Failing after 28s
CI / test (pull_request) Failing after 26s
feat: configure medium provider invocation timeout
2026-08-22 22:08:32 +03:00

368 lines
16 KiB
Go

package main
import (
"bytes"
"context"
"encoding/json"
"errors"
"fmt"
"git.cin.su/m/Assistant/adapters/models"
"git.cin.su/m/Assistant/contracts/capability"
"git.cin.su/m/Assistant/contracts/cognitive"
"git.cin.su/m/Assistant/contracts/identity"
secretcontracts "git.cin.su/m/Assistant/contracts/secrets"
platformsecrets "git.cin.su/m/Assistant/platform/secrets"
"git.cin.su/m/Assistant/runtime/health"
"io"
"log"
"net/http"
"os"
"strings"
"time"
)
const maxOpenAISecretBytes = 8 << 10
type advisoryProvider interface {
Advise(context.Context, cognitive.AdvisoryRequest) (cognitive.AdvisoryResponse, error)
}
func main() {
addr := os.Getenv("LISTEN_ADDR")
if addr == "" {
addr = ":8083"
}
openAIProvider, err := loadOpenAIAdvisoryProvider(os.Getenv("OPENAI_API_KEY_FILE"))
if err != nil {
log.Fatal(err)
}
mux := http.NewServeMux()
resolver, err := newHTTPBindingResolver(envOr("CORE_RUNTIME_URL", "http://core-runtime:8080"), &http.Client{Timeout: 12 * time.Second})
if err != nil {
log.Fatal(err)
}
providers := map[capability.ProviderKind][]providerRegistration{}
discoveries := map[string]providerRegistration{}
var builtinLocal *providerRegistration
externalRecovery := []providerRegistration{}
cpuAPIKey, err := loadProviderAPIKey("LOCAL_CPU_PROVIDER_API_KEY", "LOCAL_CPU_PROVIDER_API_KEY_FILE", "LOCAL_CPU_PROVIDER_CREDENTIAL_REF", "credential://domovoy/local-cpu/provider-api-key/v1")
if err != nil {
log.Fatal(err)
}
cpuAdapter, err := models.NewOpenAICompatibleLocalAdapterWithAPIKey(envOr("LOCAL_CPU_PROVIDER_URL", "http://docker.cin.su:18082/v1"), cpuAPIKey, nil)
if err != nil {
log.Fatal(err)
}
cpuAuthMode := capability.AuthNone
cpuCredentialRef := capability.CredentialRef("")
if cpuAPIKey != "" {
cpuAuthMode = capability.AuthKey
cpuCredentialRef = capability.CredentialRef(envOr("LOCAL_CPU_PROVIDER_CREDENTIAL_REF", "credential://domovoy/local-cpu/provider-api-key/v1"))
}
cpu := providerRegistration{ProviderInstanceID: "local-cpu-primary", Kind: "openai-compatible-chat-v1", EndpointRef: "local-cpu-primary-v1", CredentialRef: cpuCredentialRef, AuthMode: cpuAuthMode, BillingMode: capability.BillingFree, DataPolicy: capability.DataPolicyLocalOnly, Provider: cpuAdapter, Discovery: cpuAdapter}
providers[cpu.Kind] = append(providers[cpu.Kind], cpu)
discoveries["cpu"] = cpu
mediumURL := strings.TrimSpace(os.Getenv("LOCAL_MEDIUM_PROVIDER_URL"))
mediumAPIKey := ""
if mediumURL != "" {
mediumAPIKey, err = loadProviderAPIKey("LOCAL_MEDIUM_PROVIDER_API_KEY", "LOCAL_MEDIUM_PROVIDER_API_KEY_FILE", "LOCAL_MEDIUM_PROVIDER_CREDENTIAL_REF", "credential://domovoy/local-medium/provider-api-key/v1")
if err != nil {
log.Fatal(err)
}
mediumTimeout := durationEnv("LOCAL_MEDIUM_PROVIDER_TIMEOUT", 60*time.Second)
mediumAdapter, adapterErr := models.NewOpenAICompatibleLocalAdapterWithAPIKeyAndTimeout(strings.TrimRight(mediumURL, "/"), mediumAPIKey, nil, mediumTimeout)
if adapterErr != nil {
log.Fatal(adapterErr)
}
mediumAuthMode := capability.AuthNone
mediumCredentialRef := capability.CredentialRef("")
if mediumAPIKey != "" {
mediumAuthMode = capability.AuthKey
mediumCredentialRef = capability.CredentialRef(envOr("LOCAL_MEDIUM_PROVIDER_CREDENTIAL_REF", "credential://domovoy/local-medium/provider-api-key/v1"))
}
medium := providerRegistration{ProviderInstanceID: "local-medium-primary", Kind: "openai-compatible-chat-v1", EndpointRef: "local-medium-primary-v1", CredentialRef: mediumCredentialRef, AuthMode: mediumAuthMode, BillingMode: capability.BillingFree, DataPolicy: capability.DataPolicyLocalOnly, Provider: mediumAdapter, Discovery: mediumAdapter}
providers[medium.Kind] = append(providers[medium.Kind], medium)
discoveries["medium"] = medium
}
if lifeboat, enabled, lifeboatErr := newLifeboatProvider(os.Getenv("LOCAL_LIFEBOAT_PROVIDER_URL")); lifeboatErr != nil {
log.Fatal(lifeboatErr)
} else if enabled {
lifeboat.DataPolicy = capability.DataPolicyLocalOnly
providers[lifeboat.Kind] = append(providers[lifeboat.Kind], lifeboat)
discoveries["lifeboat"] = lifeboat
builtinLocal = &lifeboat
}
gpuAdapter, err := models.NewOllamaAdapter(envOr("LOCAL_GPU_PROVIDER_URL", "http://192.168.220.91:11434"), nil)
if err != nil {
log.Fatal(err)
}
gpu := providerRegistration{ProviderInstanceID: "local-gpu-opportunistic", Kind: "ollama-v1", EndpointRef: "local-gpu-opportunistic-v1", AuthMode: capability.AuthNone, BillingMode: capability.BillingFree, DataPolicy: capability.DataPolicyExternalMessageOnly, Provider: gpuAdapter, Discovery: gpuAdapter}
providers[gpu.Kind] = append(providers[gpu.Kind], gpu)
discoveries["gpu"] = gpu
externalRecovery = append(externalRecovery, gpu)
if openAIProvider != nil {
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})}
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}
mediumReadinessClient := &http.Client{Timeout: 10 * time.Second}
lifeboatURL := strings.TrimSuffix(strings.TrimRight(envOr("LOCAL_LIFEBOAT_PROVIDER_URL", "http://cognitive-lifeboat:8086/v1"), "/"), "/v1") + "/health"
fleetClient := &http.Client{Timeout: 2 * time.Second}
preferredFleet := capability.FleetTier(envOr("COGNITIVE_PREFERRED_FLEET_TIER", string(capability.FleetTierHeavy)))
cpuFleetHeaders := map[string]string{}
if cpuAPIKey != "" {
cpuFleetHeaders["Authorization"] = "Bearer " + cpuAPIKey
}
cpuEndpoint := strings.TrimRight(envOr("LOCAL_CPU_PROVIDER_URL", "http://docker.cin.su:18082/v1"), "/") + "/models"
cpuHealthProbe := health.OptionalHTTPProbeWithHeaders(cpuReadinessClient, cpuEndpoint, cpuFleetHeaders)
cpuHealthProbe.Timeout = 10 * time.Second
mediumEndpoint := ""
mediumHeaders := map[string]string{}
mediumHealthProbe := health.Probe{}
if mediumURL != "" {
mediumEndpoint = strings.TrimRight(mediumURL, "/") + "/models"
if mediumAPIKey != "" {
mediumHeaders["Authorization"] = "Bearer " + mediumAPIKey
}
mediumHealthProbe = health.OptionalHTTPProbeWithHeaders(mediumReadinessClient, mediumEndpoint, mediumHeaders)
mediumHealthProbe.Timeout = 10 * time.Second
}
readinessProbes := []health.Probe{
health.HTTPProbe(readinessClient, strings.TrimRight(envOr("CORE_RUNTIME_URL", "http://core-runtime:8080"), "/")+"/healthz"),
cpuHealthProbe,
}
if mediumURL != "" {
readinessProbes = append(readinessProbes, mediumHealthProbe)
}
readinessProbes = append(readinessProbes,
health.HTTPProbe(readinessClient, lifeboatURL),
health.OptionalHTTPProbe(readinessClient, strings.TrimRight(envOr("LOCAL_GPU_PROVIDER_URL", "http://192.168.220.91:11434"), "/")+"/api/tags"),
)
mux.Handle("/readyz", health.ReadinessHandler{Runtime: "cognitive-worker", Probes: readinessProbes})
mux.Handle("/v1/fleet/readiness", fleetReadinessHandler(fleetClient, preferredFleet, []fleetEndpoint{
{ProviderID: "local-cpu-primary", ModelID: "local-cpu-model", Tier: capability.FleetTierFast, URL: cpuEndpoint, Headers: cpuFleetHeaders, Client: &http.Client{Timeout: 10 * time.Second}},
{ProviderID: "local-medium-primary", ModelID: "local-medium-model", Tier: capability.FleetTierMedium, URL: mediumEndpoint, Headers: mediumHeaders, Client: mediumReadinessClient},
{ProviderID: "local-gpu-opportunistic", ModelID: "gpu-opportunistic", Tier: capability.FleetTierHeavy, URL: strings.TrimRight(envOr("LOCAL_GPU_PROVIDER_URL", "http://192.168.220.91:11434"), "/") + "/api/tags"},
{ProviderID: "cognitive-lifeboat", ModelID: "qwen3-8b-lifeboat", Tier: capability.FleetTierLifeboat, URL: lifeboatURL},
}))
mux.HandleFunc("POST /v1/proposals/deterministic", deterministicProposal)
mux.HandleFunc("POST /v1/advisory-policy", gateway.policyHandler)
mux.HandleFunc("POST /v1/advisories", gateway.advisoryHandler)
mux.HandleFunc("POST /v1/interpretations", gateway.interpretationHandler)
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)
refreshProviderDiscoveries(resolver, discovery)
log.Printf("cognitive-worker listening on %s", addr)
if err := http.ListenAndServe(addr, mux); err != nil {
log.Fatal(err)
}
}
func durationEnv(name string, fallback time.Duration) time.Duration {
value := strings.TrimSpace(os.Getenv(name))
if value == "" {
return fallback
}
parsed, err := time.ParseDuration(value)
if err != nil || parsed <= 0 {
log.Printf("invalid %s=%q; using %s", name, value, fallback)
return fallback
}
return parsed
}
func refreshProviderDiscoveries(resolver bindingResolver, service discoveryService) {
for name := range service.providers {
ctx, cancel := context.WithTimeout(context.Background(), 8*time.Second)
registration, ok := service.providers[name]
if !ok || registration.Discovery == nil {
cancel()
continue
}
discovered, probe := registration.Discovery.Discover(ctx)
cancel()
snapshot := capability.CognitiveDiscoverySnapshot{
ProviderInstanceID: registration.ProviderInstanceID,
ProviderKind: registration.Kind,
EndpointRef: registration.EndpointRef,
CredentialRef: registration.CredentialRef,
AuthMode: registration.AuthMode,
BillingMode: registration.BillingMode,
Enabled: true,
OwnerPrincipalID: service.owner,
CapabilityID: creatorAdvisoryCapabilityID,
DataPolicy: registration.DataPolicy,
LimitPolicyRef: "creator-advisory-local-v1",
Health: probe.Health,
FailureClass: capability.CognitiveFailureType(probe.Failure),
ObservedAtUnixMillis: time.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,
})
}
if _, err := resolver.RegisterDiscovery(context.Background(), snapshot); err != nil {
log.Printf("startup cognitive discovery refresh failed for %s: %v", name, err)
}
}
}
func loadProviderAPIKey(directEnvName, fileEnvName, refEnvName, defaultRef string) (string, error) {
if _, set := os.LookupEnv(directEnvName); set {
return "", fmt.Errorf("%s direct environment secret is forbidden; use %s", directEnvName, fileEnvName)
}
fileName := strings.TrimSpace(os.Getenv(fileEnvName))
if fileName == "" {
return "", nil
}
ref := strings.TrimSpace(os.Getenv(refEnvName))
if ref == "" {
ref = defaultRef
}
if ref == "" {
return "", fmt.Errorf("%s is required when %s is configured", refEnvName, fileEnvName)
}
resolver, err := platformsecrets.NewFileResolver(envOr("SECRET_BACKEND_ROOT", "/run/secrets"))
if err != nil {
return "", err
}
return resolver.ResolveFile(secretcontracts.SecretFileRef{CredentialRef: secretcontracts.CredentialRef(ref), Purpose: secretcontracts.SecretPurposeProviderAPIKey, Consumer: secretcontracts.RuntimeConsumerCognitive, Path: fileName})
}
func loadLocalCPUProviderAPIKey() (string, error) {
return loadProviderAPIKey("LOCAL_CPU_PROVIDER_API_KEY", "LOCAL_CPU_PROVIDER_API_KEY_FILE", "LOCAL_CPU_PROVIDER_CREDENTIAL_REF", "credential://domovoy/local-cpu/provider-api-key/v1")
}
func deterministicAdvisory(w http.ResponseWriter, r *http.Request) {
request, ok := decodeAdvisoryRequest(w, r)
if !ok {
return
}
response, err := models.DeterministicAdvisoryUnavailable(request)
if err != nil {
http.Error(w, "invalid advisory request", http.StatusBadRequest)
return
}
w.Header().Set("Content-Type", "application/json")
_ = json.NewEncoder(w).Encode(response)
}
func openAIAdvisory(provider advisoryProvider) http.HandlerFunc {
return func(w http.ResponseWriter, r *http.Request) {
if provider == nil {
http.Error(w, "provider unavailable", http.StatusServiceUnavailable)
return
}
request, ok := decodeAdvisoryRequest(w, r)
if !ok {
return
}
response, err := provider.Advise(r.Context(), request)
if err != nil {
http.Error(w, "invalid advisory request", http.StatusBadRequest)
return
}
w.Header().Set("Content-Type", "application/json")
_ = json.NewEncoder(w).Encode(response)
}
}
func decodeAdvisoryRequest(w http.ResponseWriter, r *http.Request) (cognitive.AdvisoryRequest, bool) {
r.Body = http.MaxBytesReader(w, r.Body, 80*1024)
decoder := json.NewDecoder(r.Body)
decoder.DisallowUnknownFields()
var request cognitive.AdvisoryRequest
if err := decoder.Decode(&request); err != nil {
http.Error(w, "invalid advisory request", http.StatusBadRequest)
return cognitive.AdvisoryRequest{}, false
}
if err := decoder.Decode(&struct{}{}); err != io.EOF {
http.Error(w, "invalid advisory request", http.StatusBadRequest)
return cognitive.AdvisoryRequest{}, false
}
return request, true
}
func loadOpenAIAdvisoryProvider(secretPath string) (advisoryProvider, error) {
if os.Getenv("OPENAI_API_KEY") != "" {
return nil, errors.New("direct OpenAI credential environment is forbidden; use OPENAI_API_KEY_FILE")
}
if secretPath == "" {
return nil, nil
}
file, err := os.Open(secretPath)
if err != nil {
return nil, fmt.Errorf("open OpenAI credential file: %w", err)
}
defer file.Close()
raw, err := io.ReadAll(io.LimitReader(file, maxOpenAISecretBytes+1))
if err != nil {
return nil, fmt.Errorf("read OpenAI credential file: %w", err)
}
defer clear(raw)
if len(raw) > maxOpenAISecretBytes {
return nil, errors.New("OpenAI credential file is oversized")
}
key := bytes.TrimSpace(raw)
if len(key) == 0 {
return nil, errors.New("OpenAI credential file is empty")
}
for _, character := range key {
if character < 0x21 || character > 0x7e {
return nil, errors.New("OpenAI credential file contains invalid characters")
}
}
provider, err := models.NewOpenAIAdvisoryAdapter(key, nil)
if err != nil {
return nil, errors.New("OpenAI credential file is invalid")
}
return provider, nil
}
func deterministicProposal(w http.ResponseWriter, r *http.Request) {
var request models.DeterministicRequest
if err := json.NewDecoder(r.Body).Decode(&request); err != nil {
http.Error(w, "invalid deterministic proposal request", http.StatusBadRequest)
return
}
w.Header().Set("Content-Type", "application/json")
_ = json.NewEncoder(w).Encode(models.Propose(request))
}
func envOr(name, fallback string) string {
if value := os.Getenv(name); value != "" {
return value
}
return fallback
}
func newLifeboatProvider(rawURL string) (providerRegistration, bool, error) {
if rawURL == "" {
return providerRegistration{}, false, nil
}
adapter, err := models.NewOpenAICompatibleLocalAdapter(rawURL, nil)
if err != nil {
return providerRegistration{}, false, err
}
return providerRegistration{
ProviderInstanceID: "local-lifeboat",
Kind: "openai-compatible-chat-v1",
EndpointRef: "local-lifeboat-v1",
AuthMode: capability.AuthNone,
BillingMode: capability.BillingFree,
Provider: adapter,
Discovery: adapter,
}, true, nil
}