feat(research): expose truth runtime status

feat(capability): consume qualification evidence in selection

adapters/models/openai_compatible_local.go

@@ -32,14 +32,25 @@ func NewOpenAICompatibleLocalAdapter(raw string, transport http.RoundTripper) (*
 // NewOpenAICompatibleLocalAdapterWithAPIKey creates an adapter for a gateway
 // whose bearer credential was resolved by the trusted runtime boundary.
 func NewOpenAICompatibleLocalAdapterWithAPIKey(raw, apiKey string, transport http.RoundTripper) (*OpenAICompatibleLocalAdapter, error) {
+	return NewOpenAICompatibleLocalAdapterWithAPIKeyAndTimeout(raw, apiKey, transport, 10*time.Second)
+}
+
+// NewOpenAICompatibleLocalAdapterWithAPIKeyAndTimeout creates a local gateway
+// adapter with an explicit invocation deadline. The request context remains
+// authoritative for per-call cancellation; this client deadline prevents a
+// provider from holding a worker indefinitely.
+func NewOpenAICompatibleLocalAdapterWithAPIKeyAndTimeout(raw, apiKey string, transport http.RoundTripper, timeout time.Duration) (*OpenAICompatibleLocalAdapter, error) {
 	base, err := url.Parse(raw)
 	if err != nil || base.Scheme != "http" || base.Host == "" || strings.TrimSuffix(base.Path, "/") != "/v1" || base.RawQuery != "" || base.Fragment != "" || base.User != nil {
 		return nil, errors.New("invalid local compatible endpoint")
 	}
+	if timeout <= 0 {
+		return nil, errors.New("local compatible timeout must be positive")
+	}
 	if transport == nil {
 		transport = http.DefaultTransport
 	}
-	return &OpenAICompatibleLocalAdapter{base: base, client: &http.Client{Transport: transport, Timeout: 10 * time.Second, CheckRedirect: func(*http.Request, []*http.Request) error { return http.ErrUseLastResponse }}, apiKey: strings.TrimSpace(apiKey)}, nil
+	return &OpenAICompatibleLocalAdapter{base: base, client: &http.Client{Transport: transport, Timeout: timeout, CheckRedirect: func(*http.Request, []*http.Request) error { return http.ErrUseLastResponse }}, apiKey: strings.TrimSpace(apiKey)}, nil
 }
 
 type compatibleModelsResponse struct {
@@ -85,11 +96,15 @@ func (a *OpenAICompatibleLocalAdapter) Discover(ctx context.Context) ([]Discover
 }
 
 type compatibleChatRequest struct {
-	Model       string              `json:"model"`
-	Messages    []compatibleMessage `json:"messages"`
-	Temperature int                 `json:"temperature"`
-	Stream      bool                `json:"stream"`
-	MaxTokens   int                 `json:"max_tokens"`
+	Model              string                   `json:"model"`
+	Messages           []compatibleMessage      `json:"messages"`
+	Temperature        int                      `json:"temperature"`
+	Stream             bool                     `json:"stream"`
+	MaxTokens          int                      `json:"max_tokens"`
+	ChatTemplateKwargs compatibleTemplateKwargs `json:"chat_template_kwargs"`
+}
+type compatibleTemplateKwargs struct {
+	EnableThinking bool `json:"enable_thinking"`
 }
 type compatibleMessage struct {
 	Role    string `json:"role"`
@@ -110,7 +125,7 @@ func (a *OpenAICompatibleLocalAdapter) Advise(ctx context.Context, request cogni
 	if err := ValidateAdvisoryRequest(request); err != nil || request.ProviderModelID == "" || len(request.Context) != 0 {
 		return cognitive.AdvisoryResponse{}, ErrInvalidAdvisory
 	}
-	payload := compatibleChatRequest{Model: string(request.ProviderModelID), Messages: []compatibleMessage{{Role: "system", Content: "Provide non-canonical advisory text only. Never claim that an action occurred."}, {Role: "user", Content: request.Message}}, Temperature: 0, Stream: false, MaxTokens: request.MaxOutputBytes / 4}
+	payload := compatibleChatRequest{Model: string(request.ProviderModelID), Messages: []compatibleMessage{{Role: "system", Content: "Provide non-canonical advisory text only. Never claim that an action occurred."}, {Role: "user", Content: request.Message}}, Temperature: 0, Stream: false, MaxTokens: request.MaxOutputBytes / 4, ChatTemplateKwargs: compatibleTemplateKwargs{EnableThinking: false}}
 	var result compatibleChatResponse
 	status, err := a.requestJSON(ctx, http.MethodPost, "/chat/completions", payload, &result)
 	if err != nil {

adapters/models/openai_compatible_local_test.go

@@ -2,19 +2,48 @@ package models
 
 import (
 	"context"
+	"encoding/json"
 	"io"
 	"net/http"
 	"strings"
 	"testing"
+	"time"
+
+	"git.cin.su/m/Assistant/contracts/capability"
+	"git.cin.su/m/Assistant/contracts/cognitive"
 )
 
 type captureRoundTripper struct {
 	authorization string
+	body          []byte
+}
+
+func TestOpenAICompatibleLocalAdapterUsesConfiguredTimeout(t *testing.T) {
+	adapter, err := NewOpenAICompatibleLocalAdapterWithAPIKeyAndTimeout("http://gateway/v1", "", &captureRoundTripper{}, 47*time.Second)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if adapter.client.Timeout != 47*time.Second {
+		t.Fatalf("unexpected client timeout: %s", adapter.client.Timeout)
+	}
+}
+
+func TestOpenAICompatibleLocalAdapterRejectsNonPositiveTimeout(t *testing.T) {
+	if _, err := NewOpenAICompatibleLocalAdapterWithAPIKeyAndTimeout("http://gateway/v1", "", nil, 0); err == nil {
+		t.Fatal("expected non-positive timeout to be rejected")
+	}
 }
 
 func (t *captureRoundTripper) RoundTrip(request *http.Request) (*http.Response, error) {
 	t.authorization = request.Header.Get("Authorization")
-	return &http.Response{StatusCode: http.StatusOK, Body: io.NopCloser(strings.NewReader(`{"data":[]}`)), Header: make(http.Header), Request: request}, nil
+	if request.Body != nil {
+		t.body, _ = io.ReadAll(request.Body)
+	}
+	body := `{"data":[]}`
+	if strings.HasSuffix(request.URL.Path, "/chat/completions") {
+		body = `{"choices":[{"finish_reason":"stop","message":{"role":"assistant","content":"ok"}}],"usage":{"prompt_tokens":1,"completion_tokens":1}}`
+	}
+	return &http.Response{StatusCode: http.StatusOK, Body: io.NopCloser(strings.NewReader(body)), Header: make(http.Header), Request: request}, nil
 }
 
 func TestOpenAICompatibleLocalAdapterUsesResolvedAPIKey(t *testing.T) {
@@ -40,3 +69,26 @@ func TestOpenAICompatibleLocalAdapterDoesNotSendAuthorizationWithoutKey(t *testi
 		t.Fatalf("unexpected authorization header")
 	}
 }
+
+func TestOpenAICompatibleLocalAdapterDisablesTemplateThinkingForBoundedAdvisory(t *testing.T) {
+	transport := &captureRoundTripper{}
+	adapter, err := NewOpenAICompatibleLocalAdapter("http://gateway/v1", transport)
+	if err != nil {
+		t.Fatal(err)
+	}
+	_, err = adapter.Advise(context.Background(), cognitive.AdvisoryRequest{Version: cognitive.CreatorAdvisoryV1, RequestID: "request-1", ActorID: "actor", PrincipalID: "principal", ScopeID: "scope", Message: "help", Locale: "ru-RU", MaxOutputBytes: 256, ProviderBindingID: "binding-1", ProviderBindingRevision: 1, ProviderInstanceID: "provider-1", ModelDescriptorID: "model-1", ProviderModelID: capability.ProviderModelID("medium-model")})
+	if err != nil {
+		t.Fatal(err)
+	}
+	var payload struct {
+		ChatTemplateKwargs struct {
+			EnableThinking bool `json:"enable_thinking"`
+		} `json:"chat_template_kwargs"`
+	}
+	if err := json.Unmarshal(transport.body, &payload); err != nil {
+		t.Fatal(err)
+	}
+	if payload.ChatTemplateKwargs.EnableThinking {
+		t.Fatal("expected bounded advisory to disable template thinking")
+	}
+}

adapters/postgres/capability_repository.go

@@ -430,6 +430,32 @@ func (r *CapabilityRepository) SaveCognitiveBinding(ctx context.Context, value c
 }
 
 func (r *CapabilityRepository) LoadCognitiveCandidates(ctx context.Context, capabilityID capability.CapabilityID, owner identity.PrincipalID) ([]capability.ResolvedCognitiveBinding, error) {
+	return r.loadCognitiveCandidates(ctx, capabilityID, owner, "")
+}
+
+func (r *CapabilityRepository) LoadCognitiveCandidatesForProfile(ctx context.Context, capabilityID capability.CapabilityID, owner identity.PrincipalID, profileID capability.TaskEvaluationProfileID) ([]capability.ResolvedCognitiveBinding, error) {
+	return r.loadCognitiveCandidates(ctx, capabilityID, owner, profileID)
+}
+
+func (r *CapabilityRepository) LoadCognitiveCandidate(ctx context.Context, bindingID capability.CapabilityBindingID, revision int64) (capability.ResolvedCognitiveBinding, error) {
+	var capabilityID capability.CapabilityID
+	var owner identity.PrincipalID
+	if err := r.pool.QueryRow(ctx, `SELECT capability_id,owner_principal_id FROM capability.cognitive_bindings WHERE binding_id=$1 AND revision=$2`, bindingID, revision).Scan(&capabilityID, &owner); err != nil {
+		return capability.ResolvedCognitiveBinding{}, err
+	}
+	candidates, err := r.loadCognitiveCandidates(ctx, capabilityID, owner, "")
+	if err != nil {
+		return capability.ResolvedCognitiveBinding{}, err
+	}
+	for _, candidate := range candidates {
+		if candidate.Binding.ID == bindingID && candidate.Binding.Revision == revision {
+			return candidate, nil
+		}
+	}
+	return capability.ResolvedCognitiveBinding{}, pgx.ErrNoRows
+}
+
+func (r *CapabilityRepository) loadCognitiveCandidates(ctx context.Context, capabilityID capability.CapabilityID, owner identity.PrincipalID, profileID capability.TaskEvaluationProfileID) ([]capability.ResolvedCognitiveBinding, error) {
 	rows, err := r.pool.Query(ctx, `SELECT
 		b.binding_id,b.revision,b.capability_id,b.owner_principal_id,b.provider_instance_id,b.provider_revision,b.model_descriptor_id,b.model_revision,b.data_policy,b.qualification,b.activation,b.health,b.limit_policy_ref,
 		p.provider_instance_id,p.revision,p.provider_kind,p.endpoint_ref,COALESCE(p.credential_ref,''),p.enabled,p.qualification,COALESCE(ph.health,p.health),GREATEST(p.observed_at_unix_millis,COALESCE(ph.observed_at_unix_millis,0)),p.billing_mode,p.auth_mode,p.input_cost_micros,p.output_cost_micros,p.request_cost_micros,
@@ -454,7 +480,7 @@ func (r *CapabilityRepository) LoadCognitiveCandidates(ctx context.Context, capa
 		if err != nil {
 			return nil, err
 		}
-		if err = r.loadCognitiveEvidence(ctx, &value); err != nil {
+		if err = r.loadCognitiveEvidence(ctx, &value, profileID); err != nil {
 			return nil, err
 		}
 		result = append(result, value)
@@ -462,7 +488,7 @@ func (r *CapabilityRepository) LoadCognitiveCandidates(ctx context.Context, capa
 	return result, rows.Err()
 }
 
-func (r *CapabilityRepository) loadCognitiveEvidence(ctx context.Context, value *capability.ResolvedCognitiveBinding) error {
+func (r *CapabilityRepository) loadCognitiveEvidence(ctx context.Context, value *capability.ResolvedCognitiveBinding, profileID capability.TaskEvaluationProfileID) error {
 	var evaluation capability.EvaluatedCapability
 	err := r.pool.QueryRow(ctx, `SELECT evaluation_id,binding_id,binding_revision,capability_id,result,quality_score_millis,critical_failure_millis,benchmark_version,artifact_id,observed_at_unix_millis FROM capability.cognitive_evaluated_capabilities WHERE binding_id=$1 AND binding_revision=$2 ORDER BY observed_at_unix_millis DESC LIMIT 1`, value.Binding.ID, value.Binding.Revision).Scan(&evaluation.ID, &evaluation.BindingID, &evaluation.BindingRevision, &evaluation.CapabilityID, &evaluation.Result, &evaluation.QualityScoreMillis, &evaluation.CriticalFailureMillis, &evaluation.BenchmarkVersion, &evaluation.ArtifactID, &evaluation.ObservedAtUnixMillis)
 	if err != nil && !errors.Is(err, pgx.ErrNoRows) {
@@ -479,6 +505,24 @@ func (r *CapabilityRepository) loadCognitiveEvidence(ctx context.Context, value
 	if err == nil {
 		value.Performance = &performance
 	}
+	var qualification capability.ModelQualificationObservation
+	err = r.pool.QueryRow(ctx, `SELECT observation_id,provider_instance_id,provider_model_id,status,finish_reason,latency_millis,output_observed,quality_assessed,category,observed_at_unix_millis FROM capability.cognitive_model_qualification_observations WHERE provider_instance_id=$1 AND provider_model_id=$2 ORDER BY observed_at_unix_millis DESC LIMIT 1`, value.Provider.ID, value.Model.ProviderModelID).Scan(&qualification.ID, &qualification.ProviderInstanceID, &qualification.ProviderModelID, &qualification.Status, &qualification.FinishReason, &qualification.LatencyMillis, &qualification.OutputObserved, &qualification.QualityAssessed, &qualification.Category, &qualification.ObservedAtUnixMillis)
+	if err != nil && !errors.Is(err, pgx.ErrNoRows) {
+		return err
+	}
+	if err == nil {
+		value.Qualification = &qualification
+	}
+	if profileID != "" {
+		var taskEvaluation capability.TaskEvaluation
+		err = r.pool.QueryRow(ctx, `SELECT evaluation_id,task_profile_id,binding_id,binding_revision,capability_id,result,quality_score_millis,reliability_score_millis,latency_score_millis,overall_score_millis,critical_failure_millis,benchmark_version,artifact_id,observed_at_unix_millis FROM capability.cognitive_task_evaluations WHERE task_profile_id=$1 AND binding_id=$2 AND binding_revision=$3 ORDER BY observed_at_unix_millis DESC LIMIT 1`, profileID, value.Binding.ID, value.Binding.Revision).Scan(&taskEvaluation.ID, &taskEvaluation.ProfileID, &taskEvaluation.BindingID, &taskEvaluation.BindingRevision, &taskEvaluation.CapabilityID, &taskEvaluation.Result, &taskEvaluation.QualityScoreMillis, &taskEvaluation.ReliabilityScoreMillis, &taskEvaluation.LatencyScoreMillis, &taskEvaluation.OverallScoreMillis, &taskEvaluation.CriticalFailureMillis, &taskEvaluation.BenchmarkVersion, &taskEvaluation.ArtifactID, &taskEvaluation.ObservedAtUnixMillis)
+		if err != nil && !errors.Is(err, pgx.ErrNoRows) {
+			return err
+		}
+		if err == nil {
+			value.TaskEvaluation = &taskEvaluation
+		}
+	}
 	return nil
 }
 
@@ -592,3 +636,13 @@ func (r *CapabilityRepository) SavePerformanceEvidence(ctx context.Context, valu
 	_, err := r.pool.Exec(ctx, `INSERT INTO capability.cognitive_performance_evidence (evidence_id,binding_id,binding_revision,model_descriptor_id,model_revision,hardware_profile,runtime,quantization,context_units,thinking_enabled,generation_tps_millis,prompt_tps_millis,ttft_micros,latency_micros,memory_bytes,reliability_millis,benchmark_version,artifact_id,observed_at_unix_millis) VALUES ($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13,$14,$15,$16,$17,$18,$19) ON CONFLICT DO NOTHING`, value.ID, value.BindingID, value.BindingRevision, value.ModelDescriptorID, value.ModelRevision, value.HardwareProfile, value.Runtime, value.Quantization, value.ContextUnits, value.ThinkingEnabled, value.GenerationTPSMillis, value.PromptTPSMillis, value.TTFTMicros, value.LatencyMicros, value.MemoryBytes, value.ReliabilityMillis, value.BenchmarkVersion, value.ArtifactID, value.ObservedAtUnixMillis)
 	return err
 }
+
+func (r *CapabilityRepository) SaveTaskEvaluation(ctx context.Context, value capability.TaskEvaluation) error {
+	_, err := r.pool.Exec(ctx, `INSERT INTO capability.cognitive_task_evaluations (evaluation_id,task_profile_id,binding_id,binding_revision,capability_id,result,quality_score_millis,reliability_score_millis,latency_score_millis,overall_score_millis,critical_failure_millis,benchmark_version,artifact_id,observed_at_unix_millis) VALUES ($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13,$14) ON CONFLICT DO NOTHING`, value.ID, value.ProfileID, value.BindingID, value.BindingRevision, value.CapabilityID, value.Result, value.QualityScoreMillis, value.ReliabilityScoreMillis, value.LatencyScoreMillis, value.OverallScoreMillis, value.CriticalFailureMillis, value.BenchmarkVersion, value.ArtifactID, value.ObservedAtUnixMillis)
+	return err
+}
+
+func (r *CapabilityRepository) SaveModelQualificationObservation(ctx context.Context, value capability.ModelQualificationObservation) error {
+	_, err := r.pool.Exec(ctx, `INSERT INTO capability.cognitive_model_qualification_observations (observation_id,provider_instance_id,provider_model_id,status,finish_reason,latency_millis,output_observed,quality_assessed,category,observed_at_unix_millis) VALUES ($1,$2,$3,$4,$5,$6,$7,$8,$9,$10) ON CONFLICT DO NOTHING`, value.ID, value.ProviderInstanceID, value.ProviderModelID, value.Status, value.FinishReason, value.LatencyMillis, value.OutputObserved, value.QualityAssessed, value.Category, value.ObservedAtUnixMillis)
+	return err
+}

adapters/postgres/epistemics_repository.go

@@ -11,6 +11,14 @@ import (
 
 type EpistemicsRepository struct{ pool *pgxpool.Pool }
 
+type TruthRuntimeStatus struct {
+	BundleCount          int64            `json:"bundle_count"`
+	ClaimsByVerification map[string]int64 `json:"claims_by_verification"`
+	BeliefsByState       map[string]int64 `json:"beliefs_by_state"`
+	SourcesByFreshness   map[string]int64 `json:"sources_by_freshness"`
+	ObservedAtUnixMillis int64            `json:"observed_at_unix_millis"`
+}
+
 func OpenEpistemics(ctx context.Context, config ContinuityConfig) (*EpistemicsRepository, error) {
 	pool, err := openPool(ctx, config)
 	if err != nil {
@@ -25,6 +33,49 @@ func (r *EpistemicsRepository) Ping(ctx context.Context) error {
 	return r.pool.Ping(ctx)
 }
 
+// TruthRuntimeStatus is a read-only projection of durable truth bundles.
+// It does not promote claims, revise beliefs, or mutate epistemic state.
+func (r *EpistemicsRepository) TruthRuntimeStatus(ctx context.Context) (TruthRuntimeStatus, error) {
+	status := TruthRuntimeStatus{
+		ClaimsByVerification: map[string]int64{},
+		BeliefsByState:       map[string]int64{},
+		SourcesByFreshness:   map[string]int64{},
+		ObservedAtUnixMillis: time.Now().UnixMilli(),
+	}
+	if err := r.pool.QueryRow(ctx, `SELECT count(*) FROM epistemics.truth_bundles`).Scan(&status.BundleCount); err != nil {
+		return TruthRuntimeStatus{}, err
+	}
+	queries := []struct {
+		query       string
+		destination map[string]int64
+	}{
+		{`SELECT claim_verification, count(*) FROM epistemics.truth_bundles GROUP BY claim_verification`, status.ClaimsByVerification},
+		{`SELECT belief_state, count(*) FROM epistemics.truth_bundles GROUP BY belief_state`, status.BeliefsByState},
+		{`SELECT snapshot_freshness, count(*) FROM epistemics.truth_bundles GROUP BY snapshot_freshness`, status.SourcesByFreshness},
+	}
+	for _, item := range queries {
+		rows, err := r.pool.Query(ctx, item.query)
+		if err != nil {
+			return TruthRuntimeStatus{}, err
+		}
+		for rows.Next() {
+			var state string
+			var count int64
+			if err := rows.Scan(&state, &count); err != nil {
+				rows.Close()
+				return TruthRuntimeStatus{}, err
+			}
+			item.destination[state] = count
+		}
+		if err := rows.Err(); err != nil {
+			rows.Close()
+			return TruthRuntimeStatus{}, err
+		}
+		rows.Close()
+	}
+	return status, nil
+}
+
 func (r *EpistemicsRepository) SaveSystemExpectation(ctx context.Context, value epistemics.SystemExpectation) error {
 	_, err := r.pool.Exec(ctx, `INSERT INTO epistemics.system_expectations (
 		expectation_id, original_intent_id, source_utterance_id, expected_outcome, expected_effect,

adapters/postgres/work_repository.go

@@ -2,10 +2,14 @@ package postgres
 
 import (
 	"context"
+	"encoding/json"
 	"errors"
 	"fmt"
+	"reflect"
+	"slices"
 	"time"
 
+	"git.cin.su/m/Assistant/contracts/capability"
 	"git.cin.su/m/Assistant/contracts/work"
 	"github.com/jackc/pgx/v5"
 	"github.com/jackc/pgx/v5/pgxpool"
@@ -14,9 +18,11 @@ import (
 type WorkRepository struct{ pool *pgxpool.Pool }
 
 var (
-	ErrDispatchIntentConflict = errors.New("dispatch intent idempotency conflict")
-	ErrAgentLeaseUnavailable  = errors.New("agent lease unavailable")
-	ErrAgentCallbackRejected  = errors.New("agent callback rejected")
+	ErrDispatchIntentConflict       = errors.New("dispatch intent idempotency conflict")
+	ErrAgentLeaseUnavailable        = errors.New("agent lease unavailable")
+	ErrAgentCallbackRejected        = errors.New("agent callback rejected")
+	ErrDevelopmentGateConflict      = errors.New("development gate revision conflict")
+	ErrDevelopmentCandidateConflict = errors.New("development candidate conflict")
 )
 
 func OpenWork(ctx context.Context, config ContinuityConfig) (*WorkRepository, error) {
@@ -26,6 +32,54 @@ func OpenWork(ctx context.Context, config ContinuityConfig) (*WorkRepository, er
 	}
 	return &WorkRepository{pool}, nil
 }
+
+// WorkGraphRuntimeStatus is a read-only operational projection owned by Work.
+// It is not canonical graph state and cannot advance any lifecycle.
+type WorkGraphRuntimeStatus struct {
+	GraphsByState        map[string]int64 `json:"graphs_by_state"`
+	NodesByState         map[string]int64 `json:"nodes_by_state"`
+	QueueByState         map[string]int64 `json:"queue_by_state"`
+	ObservedAtUnixMillis int64            `json:"observed_at_unix_millis"`
+}
+
+func (r *WorkRepository) RuntimeStatus(ctx context.Context) (WorkGraphRuntimeStatus, error) {
+	status := WorkGraphRuntimeStatus{
+		GraphsByState:        map[string]int64{},
+		NodesByState:         map[string]int64{},
+		QueueByState:         map[string]int64{},
+		ObservedAtUnixMillis: time.Now().UnixMilli(),
+	}
+	queries := []struct {
+		query       string
+		destination map[string]int64
+	}{
+		{`SELECT state, count(*) FROM work.graphs GROUP BY state`, status.GraphsByState},
+		{`SELECT state, count(*) FROM work.graph_nodes GROUP BY state`, status.NodesByState},
+		{`SELECT state, count(*) FROM work.agent_dispatch_queue GROUP BY state`, status.QueueByState},
+	}
+	for _, item := range queries {
+		rows, err := r.pool.Query(ctx, item.query)
+		if err != nil {
+			return WorkGraphRuntimeStatus{}, err
+		}
+		for rows.Next() {
+			var state string
+			var count int64
+			if err := rows.Scan(&state, &count); err != nil {
+				rows.Close()
+				return WorkGraphRuntimeStatus{}, err
+			}
+			item.destination[state] = count
+		}
+		if err := rows.Err(); err != nil {
+			rows.Close()
+			return WorkGraphRuntimeStatus{}, err
+		}
+		rows.Close()
+	}
+	return status, nil
+}
+
 func (r *WorkRepository) Close()                         { r.pool.Close() }
 func (r *WorkRepository) Ping(ctx context.Context) error { return r.pool.Ping(ctx) }
 func (r *WorkRepository) Create(ctx context.Context, target work.TaskTarget) error {
@@ -49,33 +103,135 @@ func (r *WorkRepository) Load(ctx context.Context, id work.TaskID) (work.TaskTar
 	return target, nil
 }
 
+func (r *WorkRepository) LoadDevelopmentGate(ctx context.Context, id string) (work.DevelopmentGate, error) {
+	var gate work.DevelopmentGate
+	var evidence, producer, verifier []string
+	err := r.pool.QueryRow(ctx, `SELECT gate_id,revision,updated_at_unix_millis,stage,evidence_refs,proposal_ref,task_packet_ref,worktree_ref,producer_run_refs,verifier_run_refs,verification_evidence,canary_evidence,promotion_ref,rollback_ref FROM work.development_gates WHERE gate_id=$1`, id).Scan(
+		&gate.ID, &gate.Revision, &gate.UpdatedAtUnixMillis, &gate.Stage, &evidence, &gate.ProposalRef, &gate.TaskPacketRef, &gate.WorktreeRef, &producer, &verifier, &gate.VerificationEvidence, &gate.CanaryEvidence, &gate.PromotionRef, &gate.RollbackRef)
+	if err != nil {
+		return work.DevelopmentGate{}, fmt.Errorf("load development gate: %w", err)
+	}
+	gate.EvidenceRefs = append([]string(nil), evidence...)
+	for _, run := range producer {
+		gate.ProducerRunRefs = append(gate.ProducerRunRefs, work.AgentRunID(run))
+	}
+	for _, run := range verifier {
+		gate.VerifierRunRefs = append(gate.VerifierRunRefs, work.AgentRunID(run))
+	}
+	return gate, nil
+}
+
+func (r *WorkRepository) SaveDevelopmentGate(ctx context.Context, gate work.DevelopmentGate, expectedRevision int64) error {
+	if gate.ID == "" || expectedRevision < 0 || gate.Revision < 1 || gate.UpdatedAtUnixMillis < 1 {
+		return ErrDevelopmentGateConflict
+	}
+	producer := make([]string, 0, len(gate.ProducerRunRefs))
+	for _, run := range gate.ProducerRunRefs {
+		producer = append(producer, string(run))
+	}
+	verifier := make([]string, 0, len(gate.VerifierRunRefs))
+	for _, run := range gate.VerifierRunRefs {
+		verifier = append(verifier, string(run))
+	}
+	if expectedRevision == 0 {
+		command, err := r.pool.Exec(ctx, `INSERT INTO work.development_gates(gate_id,revision,updated_at_unix_millis,stage,evidence_refs,proposal_ref,task_packet_ref,worktree_ref,producer_run_refs,verifier_run_refs,verification_evidence,canary_evidence,promotion_ref,rollback_ref) VALUES($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13,$14) ON CONFLICT(gate_id) DO NOTHING`, gate.ID, gate.Revision, gate.UpdatedAtUnixMillis, gate.Stage, gate.EvidenceRefs, gate.ProposalRef, gate.TaskPacketRef, gate.WorktreeRef, producer, verifier, gate.VerificationEvidence, gate.CanaryEvidence, gate.PromotionRef, gate.RollbackRef)
+		if err != nil {
+			return err
+		}
+		if command.RowsAffected() != 1 {
+			return ErrDevelopmentGateConflict
+		}
+		return nil
+	}
+	command, err := r.pool.Exec(ctx, `UPDATE work.development_gates SET revision=$1,updated_at_unix_millis=$2,stage=$3,evidence_refs=$4,proposal_ref=$5,task_packet_ref=$6,worktree_ref=$7,producer_run_refs=$8,verifier_run_refs=$9,verification_evidence=$10,canary_evidence=$11,promotion_ref=$12,rollback_ref=$13 WHERE gate_id=$14 AND revision=$15`, gate.Revision, gate.UpdatedAtUnixMillis, gate.Stage, gate.EvidenceRefs, gate.ProposalRef, gate.TaskPacketRef, gate.WorktreeRef, producer, verifier, gate.VerificationEvidence, gate.CanaryEvidence, gate.PromotionRef, gate.RollbackRef, gate.ID, expectedRevision)
+	if err != nil {
+		return err
+	}
+	if command.RowsAffected() != 1 {
+		return ErrDevelopmentGateConflict
+	}
+	return nil
+}
+
+func (r *WorkRepository) SaveExtensionProposal(ctx context.Context, value work.ExtensionProposal) error {
+	result, err := r.pool.Exec(ctx, `INSERT INTO work.extension_proposals(proposal_id,requested_by_principal_id,stage,evidence_refs,deduplicated_signal_keys,desired_outcome,smallest_change_layer,affected_owners,affected_contracts,risk_tier,evaluation_plan,alternatives,rollback_plan,assumptions,created_at_unix_millis) VALUES($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13,$14,$15) ON CONFLICT(proposal_id) DO NOTHING`, value.ID, value.RequestedByPrincipalID, value.Stage, value.EvidenceRefs, value.DeduplicatedSignalKeys, value.DesiredOutcome, value.SmallestChangeLayer, value.AffectedOwners, value.AffectedContracts, value.RiskTier, value.EvaluationPlan, value.Alternatives, value.RollbackPlan, value.Assumptions, value.CreatedAtUnixMillis)
+	if err != nil {
+		return err
+	}
+	if result.RowsAffected() != 1 {
+		return ErrDevelopmentCandidateConflict
+	}
+	return nil
+}
+
+func (r *WorkRepository) SaveDevelopmentScaffold(ctx context.Context, value work.DevelopmentScaffold) error {
+	result, err := r.pool.Exec(ctx, `INSERT INTO work.development_scaffolds(scaffold_id,requested_by_principal_id,stage,proposal_ref,base_repo_revision,worktree_ref,task_intent,modules,allowed_paths,forbidden_paths,required_tests,artifact_lineage_refs,protected_main_access,production_access,secret_access,created_at_unix_millis) VALUES($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13,$14,$15,$16) ON CONFLICT(scaffold_id) DO NOTHING`, value.ID, value.RequestedByPrincipalID, value.Stage, value.ProposalRef, value.BaseRepoRevision, value.WorktreeRef, value.TaskIntent, value.Modules, value.AllowedPaths, value.ForbiddenPaths, value.RequiredTests, value.ArtifactLineageRefs, value.ProtectedMainAccess, value.ProductionAccess, value.SecretAccess, value.CreatedAtUnixMillis)
+	if err != nil {
+		return err
+	}
+	if result.RowsAffected() != 1 {
+		return ErrDevelopmentCandidateConflict
+	}
+	return nil
+}
+
 func (r *WorkRepository) SaveDispatchIntent(ctx context.Context, intent work.AgentDispatchIntent) error {
 	tx, err := r.pool.Begin(ctx)
 	if err != nil {
 		return err
 	}
 	defer tx.Rollback(ctx)
-	_, err = tx.Exec(ctx, `INSERT INTO work.agent_dispatch_intents(dispatch_id,idempotency_key,delegation_id,run_id,implementation_id,implementation_revision,source_revision,worktree_id,verification_plan_ref,created_at_unix_millis) VALUES($1,$2,$3,$4,$5,$6,$7,$8,$9,$10) ON CONFLICT(idempotency_key) DO NOTHING`, intent.DispatchID, intent.IdempotencyKey, intent.DelegationID, intent.RunID, intent.ImplementationID, intent.ImplementationRevision, intent.SourceRevision, intent.WorktreeID, intent.VerificationPlanRef, intent.CreatedAtUnixMillis)
+	profileJSON, casesJSON, err := marshalTaskPayload(intent.TaskEvaluationProfile, intent.TaskEvaluationCases)
+	if err != nil {
+		return err
+	}
+	executionJSON, err := json.Marshal(intent.TaskEvaluationExecution)
+	if err != nil {
+		return err
+	}
+	_, err = tx.Exec(ctx, `INSERT INTO work.agent_dispatch_intents(dispatch_id,idempotency_key,delegation_id,run_id,implementation_id,implementation_revision,source_revision,worktree_id,verification_plan_ref,created_at_unix_millis,specialist_purpose,task_evaluation_profile_id,task_evaluation_binding_id,task_evaluation_case_ids,task_evaluation_profile,task_evaluation_cases,task_evaluation_execution) VALUES($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13,$14,$15,$16,$17) ON CONFLICT(idempotency_key) DO NOTHING`, intent.DispatchID, intent.IdempotencyKey, intent.DelegationID, intent.RunID, intent.ImplementationID, intent.ImplementationRevision, intent.SourceRevision, intent.WorktreeID, intent.VerificationPlanRef, intent.CreatedAtUnixMillis, intent.SpecialistPurpose, intent.TaskEvaluationProfileID, intent.TaskEvaluationBindingID, intent.TaskEvaluationCaseIDs, profileJSON, casesJSON, executionJSON)
 	if err != nil {
 		return fmt.Errorf("save dispatch intent: %w", err)
 	}
 	var existing work.AgentDispatchIntent
-	err = tx.QueryRow(ctx, `SELECT dispatch_id,idempotency_key,delegation_id,run_id,implementation_id,implementation_revision,source_revision,worktree_id,verification_plan_ref,created_at_unix_millis FROM work.agent_dispatch_intents WHERE idempotency_key=$1`, intent.IdempotencyKey).Scan(&existing.DispatchID, &existing.IdempotencyKey, &existing.DelegationID, &existing.RunID, &existing.ImplementationID, &existing.ImplementationRevision, &existing.SourceRevision, &existing.WorktreeID, &existing.VerificationPlanRef, &existing.CreatedAtUnixMillis)
+	var profileJSONExisting, casesJSONExisting, executionJSONExisting []byte
+	err = tx.QueryRow(ctx, `SELECT dispatch_id,idempotency_key,delegation_id,run_id,implementation_id,implementation_revision,source_revision,worktree_id,verification_plan_ref,created_at_unix_millis,specialist_purpose,task_evaluation_profile_id,task_evaluation_binding_id,task_evaluation_case_ids,task_evaluation_profile,task_evaluation_cases,task_evaluation_execution FROM work.agent_dispatch_intents WHERE idempotency_key=$1`, intent.IdempotencyKey).Scan(&existing.DispatchID, &existing.IdempotencyKey, &existing.DelegationID, &existing.RunID, &existing.ImplementationID, &existing.ImplementationRevision, &existing.SourceRevision, &existing.WorktreeID, &existing.VerificationPlanRef, &existing.CreatedAtUnixMillis, &existing.SpecialistPurpose, &existing.TaskEvaluationProfileID, &existing.TaskEvaluationBindingID, &existing.TaskEvaluationCaseIDs, &profileJSONExisting, &casesJSONExisting, &executionJSONExisting)
 	if err != nil {
 		return fmt.Errorf("read dispatch intent: %w", err)
 	}
-	if existing != intent {
+	if err := unmarshalTaskPayload(profileJSONExisting, casesJSONExisting, &existing.TaskEvaluationProfile, &existing.TaskEvaluationCases); err != nil {
+		return err
+	}
+	if len(executionJSONExisting) > 0 && string(executionJSONExisting) != "null" {
+		if err := json.Unmarshal(executionJSONExisting, &existing.TaskEvaluationExecution); err != nil {
+			return err
+		}
+	}
+	if !sameDispatchIntent(existing, intent) {
 		return ErrDispatchIntentConflict
 	}
 	return tx.Commit(ctx)
 }
 
+func sameDispatchIntent(left, right work.AgentDispatchIntent) bool {
+	return left.DispatchID == right.DispatchID && left.IdempotencyKey == right.IdempotencyKey && left.DelegationID == right.DelegationID && left.RunID == right.RunID && left.ImplementationID == right.ImplementationID && left.ImplementationRevision == right.ImplementationRevision && left.SourceRevision == right.SourceRevision && left.WorktreeID == right.WorktreeID && left.VerificationPlanRef == right.VerificationPlanRef && left.CreatedAtUnixMillis == right.CreatedAtUnixMillis && left.SpecialistPurpose == right.SpecialistPurpose && left.TaskEvaluationProfileID == right.TaskEvaluationProfileID && left.TaskEvaluationBindingID == right.TaskEvaluationBindingID && slices.Equal(left.TaskEvaluationCaseIDs, right.TaskEvaluationCaseIDs) && reflect.DeepEqual(left.TaskEvaluationProfile, right.TaskEvaluationProfile) && slices.Equal(left.TaskEvaluationCases, right.TaskEvaluationCases) && reflect.DeepEqual(left.TaskEvaluationExecution, right.TaskEvaluationExecution)
+}
+
 func (r *WorkRepository) LoadDispatchIntent(ctx context.Context, dispatchID string) (work.AgentDispatchIntent, error) {
 	var intent work.AgentDispatchIntent
-	err := r.pool.QueryRow(ctx, `SELECT dispatch_id,idempotency_key,delegation_id,run_id,implementation_id,implementation_revision,source_revision,worktree_id,verification_plan_ref,created_at_unix_millis FROM work.agent_dispatch_intents WHERE dispatch_id=$1`, dispatchID).Scan(&intent.DispatchID, &intent.IdempotencyKey, &intent.DelegationID, &intent.RunID, &intent.ImplementationID, &intent.ImplementationRevision, &intent.SourceRevision, &intent.WorktreeID, &intent.VerificationPlanRef, &intent.CreatedAtUnixMillis)
+	var profileJSON, casesJSON, executionJSON []byte
+	err := r.pool.QueryRow(ctx, `SELECT dispatch_id,idempotency_key,delegation_id,run_id,implementation_id,implementation_revision,source_revision,worktree_id,verification_plan_ref,created_at_unix_millis,specialist_purpose,task_evaluation_profile_id,task_evaluation_binding_id,task_evaluation_case_ids,task_evaluation_profile,task_evaluation_cases,task_evaluation_execution FROM work.agent_dispatch_intents WHERE dispatch_id=$1`, dispatchID).Scan(&intent.DispatchID, &intent.IdempotencyKey, &intent.DelegationID, &intent.RunID, &intent.ImplementationID, &intent.ImplementationRevision, &intent.SourceRevision, &intent.WorktreeID, &intent.VerificationPlanRef, &intent.CreatedAtUnixMillis, &intent.SpecialistPurpose, &intent.TaskEvaluationProfileID, &intent.TaskEvaluationBindingID, &intent.TaskEvaluationCaseIDs, &profileJSON, &casesJSON, &executionJSON)
 	if err != nil {
 		return work.AgentDispatchIntent{}, fmt.Errorf("load dispatch intent: %w", err)
 	}
+	if err := unmarshalTaskPayload(profileJSON, casesJSON, &intent.TaskEvaluationProfile, &intent.TaskEvaluationCases); err != nil {
+		return work.AgentDispatchIntent{}, err
+	}
+	if len(executionJSON) > 0 && string(executionJSON) != "null" {
+		if err := json.Unmarshal(executionJSON, &intent.TaskEvaluationExecution); err != nil {
+			return work.AgentDispatchIntent{}, err
+		}
+	}
 	return intent, nil
 }
 
@@ -91,7 +247,7 @@ func (r *WorkRepository) ClaimNextDispatch(ctx context.Context, workerID string,
 	}
 	defer tx.Rollback(ctx)
 	var item work.AgentDispatchQueueItem
-	err = tx.QueryRow(ctx, `SELECT dispatch_id,state,available_at_unix_millis,attempts,claimed_by,updated_at_unix_millis FROM work.agent_dispatch_queue WHERE state='QUEUED' AND available_at_unix_millis <= $1 ORDER BY available_at_unix_millis,dispatch_id FOR UPDATE SKIP LOCKED LIMIT 1`, nowUnixMillis).Scan(&item.DispatchID, &item.State, &item.AvailableAtUnixMillis, &item.Attempts, &item.ClaimedBy, &item.UpdatedAtUnixMillis)
+	err = tx.QueryRow(ctx, `SELECT dispatch_id,state,available_at_unix_millis,attempts,claimed_by,updated_at_unix_millis FROM work.agent_dispatch_queue WHERE state='QUEUED' AND available_at_unix_millis <= $1 AND attempts < retry_limit ORDER BY available_at_unix_millis,dispatch_id FOR UPDATE SKIP LOCKED LIMIT 1`, nowUnixMillis).Scan(&item.DispatchID, &item.State, &item.AvailableAtUnixMillis, &item.Attempts, &item.ClaimedBy, &item.UpdatedAtUnixMillis)
 	if err != nil {
 		if errors.Is(err, pgx.ErrNoRows) {
 			return work.AgentDispatchQueueItem{}, nil
@@ -134,17 +290,36 @@ func (r *WorkRepository) CompleteDispatch(ctx context.Context, dispatchID, worke
 	return nil
 }
 
+func (r *WorkRepository) RetryDispatch(ctx context.Context, dispatchID, workerID string, nowUnixMillis int64, delay time.Duration) error {
+	if delay < 0 {
+		return fmt.Errorf("invalid retry delay")
+	}
+	availableAt := nowUnixMillis + delay.Milliseconds()
+	command, err := r.pool.Exec(ctx, `UPDATE work.agent_dispatch_queue SET state=CASE WHEN attempts >= retry_limit THEN 'FAILED' ELSE 'QUEUED' END,available_at_unix_millis=$1,claimed_by='',updated_at_unix_millis=$2 WHERE dispatch_id=$3 AND state='CLAIMED' AND claimed_by=$4`, availableAt, nowUnixMillis, dispatchID, workerID)
+	if err != nil {
+		return err
+	}
+	if command.RowsAffected() != 1 {
+		return ErrAgentCallbackRejected
+	}
+	return nil
+}
+
 func (r *WorkRepository) SaveWorkGraph(ctx context.Context, graph work.WorkGraph) error {
 	tx, err := r.pool.Begin(ctx)
 	if err != nil {
 		return err
 	}
 	defer tx.Rollback(ctx)
-	if _, err := tx.Exec(ctx, `INSERT INTO work.graphs(graph_id,revision,state,max_parallel,max_attempts) VALUES($1,$2,$3,$4,$5) ON CONFLICT(graph_id,revision) DO NOTHING`, graph.ID, graph.Revision, graph.State, graph.Budget.MaxParallel, graph.Budget.MaxAttempts); err != nil {
+	if _, err := tx.Exec(ctx, `INSERT INTO work.graphs(graph_id,revision,state,max_parallel,max_attempts,parent_graph_id,parent_graph_revision,trigger_evidence_ref,orchestration_bound,orchestration_goal_id,orchestration_goal_state,orchestration_search_round,orchestration_max_round,orchestration_cancel_requested,orchestration_verification_ready,orchestration_unresolved) VALUES($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13,$14,$15,$16) ON CONFLICT(graph_id,revision) DO NOTHING`, graph.ID, graph.Revision, graph.State, graph.Budget.MaxParallel, graph.Budget.MaxAttempts, graph.ParentGraphID, graph.ParentGraphRevision, graph.TriggerEvidenceRef, graph.SearchInputBound, graph.SearchInput.GoalID, graph.SearchInput.GoalState, graph.SearchInput.SearchRound, graph.SearchInput.MaxRound, graph.SearchInput.CancelRequested, graph.SearchInput.VerificationReady, graph.SearchInput.Unresolved); err != nil {
 		return err
 	}
 	for _, node := range graph.Nodes {
-		if _, err := tx.Exec(ctx, `INSERT INTO work.graph_nodes(graph_id,revision,node_id,dispatch_id,idempotency_key,state,attempts) VALUES($1,$2,$3,$4,$5,$6,$7) ON CONFLICT(graph_id,revision,node_id) DO NOTHING`, graph.ID, graph.Revision, node.ID, node.DispatchID, node.IdempotencyKey, node.State, node.Attempts); err != nil {
+		profileJSON, casesJSON, err := marshalTaskPayload(node.TaskProfile, node.TaskCases)
+		if err != nil {
+			return err
+		}
+		if _, err := tx.Exec(ctx, `INSERT INTO work.graph_nodes(graph_id,revision,node_id,dispatch_id,idempotency_key,state,attempts,purpose,task_profile_id,binding_id,case_ids,task_evaluation_profile,task_evaluation_cases) VALUES($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13) ON CONFLICT(graph_id,revision,node_id) DO NOTHING`, graph.ID, graph.Revision, node.ID, node.DispatchID, node.IdempotencyKey, node.State, node.Attempts, node.Purpose, node.TaskProfileID, node.BindingID, node.CaseIDs, profileJSON, casesJSON); err != nil {
 			return err
 		}
 	}
@@ -158,17 +333,22 @@ func (r *WorkRepository) SaveWorkGraph(ctx context.Context, graph work.WorkGraph
 
 func (r *WorkRepository) LoadWorkGraph(ctx context.Context, graphID string, revision int64) (work.WorkGraph, error) {
 	var graph work.WorkGraph
-	err := r.pool.QueryRow(ctx, `SELECT graph_id,revision,state,max_parallel,max_attempts FROM work.graphs WHERE graph_id=$1 AND revision=$2`, graphID, revision).Scan(&graph.ID, &graph.Revision, &graph.State, &graph.Budget.MaxParallel, &graph.Budget.MaxAttempts)
+	err := r.pool.QueryRow(ctx, `SELECT graph_id,revision,state,max_parallel,max_attempts,parent_graph_id,parent_graph_revision,trigger_evidence_ref,orchestration_bound,orchestration_goal_id,orchestration_goal_state,orchestration_search_round,orchestration_max_round,orchestration_cancel_requested,orchestration_verification_ready,orchestration_unresolved FROM work.graphs WHERE graph_id=$1 AND revision=$2`, graphID, revision).Scan(&graph.ID, &graph.Revision, &graph.State, &graph.Budget.MaxParallel, &graph.Budget.MaxAttempts, &graph.ParentGraphID, &graph.ParentGraphRevision, &graph.TriggerEvidenceRef, &graph.SearchInputBound, &graph.SearchInput.GoalID, &graph.SearchInput.GoalState, &graph.SearchInput.SearchRound, &graph.SearchInput.MaxRound, &graph.SearchInput.CancelRequested, &graph.SearchInput.VerificationReady, &graph.SearchInput.Unresolved)
 	if err != nil {
 		return work.WorkGraph{}, fmt.Errorf("load graph: %w", err)
 	}
-	rows, err := r.pool.Query(ctx, `SELECT node_id,dispatch_id,idempotency_key,state,attempts FROM work.graph_nodes WHERE graph_id=$1 AND revision=$2 ORDER BY node_id`, graphID, revision)
+	rows, err := r.pool.Query(ctx, `SELECT node_id,dispatch_id,idempotency_key,state,attempts,purpose,task_profile_id,binding_id,case_ids,task_evaluation_profile,task_evaluation_cases FROM work.graph_nodes WHERE graph_id=$1 AND revision=$2 ORDER BY node_id`, graphID, revision)
 	if err != nil {
 		return work.WorkGraph{}, err
 	}
 	for rows.Next() {
 		var node work.GraphNode
-		if err := rows.Scan(&node.ID, &node.DispatchID, &node.IdempotencyKey, &node.State, &node.Attempts); err != nil {
+		var profileJSON, casesJSON []byte
+		if err := rows.Scan(&node.ID, &node.DispatchID, &node.IdempotencyKey, &node.State, &node.Attempts, &node.Purpose, &node.TaskProfileID, &node.BindingID, &node.CaseIDs, &profileJSON, &casesJSON); err != nil {
+			rows.Close()
+			return work.WorkGraph{}, err
+		}
+		if err := unmarshalTaskPayload(profileJSON, casesJSON, &node.TaskProfile, &node.TaskCases); err != nil {
 			rows.Close()
 			return work.WorkGraph{}, err
 		}
@@ -199,9 +379,48 @@ func (r *WorkRepository) LoadWorkGraph(ctx context.Context, graphID string, revi
 	return graph, nil
 }
 
-func (r *WorkRepository) RecoverStaleClaims(ctx context.Context, nowUnixMillis, staleAfterMillis int64) (int64, error) {
-	command, err := r.pool.Exec(ctx, `UPDATE work.agent_dispatch_queue SET state='QUEUED',claimed_by='',updated_at_unix_millis=$1 WHERE state='CLAIMED' AND updated_at_unix_millis <= $2`, nowUnixMillis, nowUnixMillis-staleAfterMillis)
+func (r *WorkRepository) SaveGraphSynthesis(ctx context.Context, conflict work.ConflictSet, candidates []work.ResultCandidate, synthesis work.GraphSynthesisCandidate) error {
+	tx, err := r.pool.Begin(ctx)
 	if err != nil {
+		return err
+	}
+	defer tx.Rollback(ctx)
+	for _, candidate := range candidates {
+		if _, err := tx.Exec(ctx, `INSERT INTO work.graph_result_candidates(result_candidate_id,graph_id,graph_revision,node_id,author_run_id,artifact_refs,evidence_summary,summary) VALUES($1,$2,$3,$4,$5,$6,$7,$8) ON CONFLICT(result_candidate_id) DO NOTHING`, candidate.ID, candidate.GraphID, candidate.GraphRevision, candidate.NodeID, candidate.AuthorRunID, candidate.ArtifactRefs, candidate.EvidenceSummary, candidate.Summary); err != nil {
+			return err
+		}
+	}
+	if _, err := tx.Exec(ctx, `INSERT INTO work.graph_conflicts(conflict_id,graph_id,graph_revision,node_id,candidate_ids,resolved) VALUES($1,$2,$3,$4,$5,$6) ON CONFLICT(conflict_id) DO NOTHING`, conflict.ID, conflict.GraphID, conflict.GraphRevision, conflict.NodeID, conflict.CandidateIDs, conflict.Resolved); err != nil {
+		return err
+	}
+	if _, err := tx.Exec(ctx, `INSERT INTO work.graph_synthesis_candidates(synthesis_id,graph_id,graph_revision,node_id,candidate_ids,selected_candidate_id,dissent_candidate_ids,verifier_run_id,verifier_identity,evidence_refs,summary,state) VALUES($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12) ON CONFLICT(synthesis_id) DO NOTHING`, synthesis.ID, synthesis.GraphID, synthesis.GraphRevision, synthesis.NodeID, synthesis.CandidateIDs, synthesis.SelectedCandidateID, synthesis.DissentCandidateIDs, synthesis.VerifierRunID, synthesis.VerifierIdentity, synthesis.EvidenceRefs, synthesis.Summary, synthesis.State); err != nil {
+		return err
+	}
+	return tx.Commit(ctx)
+}
+
+func (r *WorkRepository) SaveGraphResultCandidate(ctx context.Context, candidate work.ResultCandidate) error {
+	if candidate.ID == "" || candidate.GraphID == "" || candidate.GraphRevision < 1 || candidate.NodeID == "" || candidate.AuthorRunID == "" || candidate.EvidenceSummary == "" || candidate.Summary == "" {
+		return errors.New("invalid graph result candidate")
+	}
+	_, err := r.pool.Exec(ctx, `INSERT INTO work.graph_result_candidates(result_candidate_id,graph_id,graph_revision,node_id,author_run_id,artifact_refs,evidence_summary,summary) VALUES($1,$2,$3,$4,$5,$6,$7,$8) ON CONFLICT(result_candidate_id) DO NOTHING`, candidate.ID, candidate.GraphID, candidate.GraphRevision, candidate.NodeID, candidate.AuthorRunID, candidate.ArtifactRefs, candidate.EvidenceSummary, candidate.Summary)
+	return err
+}
+
+func (r *WorkRepository) RecoverStaleClaims(ctx context.Context, nowUnixMillis, staleAfterMillis int64) (int64, error) {
+	tx, err := r.pool.Begin(ctx)
+	if err != nil {
+		return 0, err
+	}
+	defer tx.Rollback(ctx)
+	if _, err := tx.Exec(ctx, `UPDATE work.agent_leases AS l SET state='EXPIRED' FROM work.agent_dispatch_queue AS q JOIN work.agent_dispatch_intents AS i ON i.dispatch_id=q.dispatch_id WHERE q.state='CLAIMED' AND q.updated_at_unix_millis <= $1 AND l.run_id=i.run_id AND l.state='ACTIVE'`, nowUnixMillis-staleAfterMillis); err != nil {
+		return 0, err
+	}
+	command, err := tx.Exec(ctx, `UPDATE work.agent_dispatch_queue SET state=CASE WHEN attempts >= retry_limit THEN 'FAILED' ELSE 'QUEUED' END,available_at_unix_millis=$1,claimed_by='',updated_at_unix_millis=$2 WHERE state='CLAIMED' AND updated_at_unix_millis <= $3`, nowUnixMillis, nowUnixMillis, nowUnixMillis-staleAfterMillis)
+	if err != nil {
+		return 0, err
+	}
+	if err := tx.Commit(ctx); err != nil {
 		return 0, err
 	}
 	return command.RowsAffected(), nil
@@ -494,3 +713,32 @@ func latestRowErr(err error) error {
 	}
 	return err
 }
+
+func marshalTaskPayload(profile *capability.TaskEvaluationProfile, cases []capability.TaskEvaluationCase) ([]byte, []byte, error) {
+	var profileJSON []byte
+	var err error
+	if profile != nil {
+		profileJSON, err = json.Marshal(profile)
+		if err != nil {
+			return nil, nil, err
+		}
+	}
+	casesJSON, err := json.Marshal(cases)
+	return profileJSON, casesJSON, err
+}
+
+func unmarshalTaskPayload(profileJSON, casesJSON []byte, profile **capability.TaskEvaluationProfile, cases *[]capability.TaskEvaluationCase) error {
+	if len(profileJSON) > 0 && string(profileJSON) != "null" {
+		var value capability.TaskEvaluationProfile
+		if err := json.Unmarshal(profileJSON, &value); err != nil {
+			return err
+		}
+		*profile = &value
+	}
+	if len(casesJSON) > 0 && string(casesJSON) != "null" {
+		if err := json.Unmarshal(casesJSON, cases); err != nil {
+			return err
+		}
+	}
+	return nil
+}

cmd/agent-canary/main.go

@@ -6,12 +6,16 @@ import (
 	"flag"
 	"fmt"
 	"os"
+	"path/filepath"
+	"strings"
 	"time"
 
 	"git.cin.su/m/Assistant/adapters/postgres"
+	secretcontracts "git.cin.su/m/Assistant/contracts/secrets"
 	"git.cin.su/m/Assistant/contracts/work"
 	"git.cin.su/m/Assistant/modules/work/application"
 	"git.cin.su/m/Assistant/modules/work/ports"
+	platformsecrets "git.cin.su/m/Assistant/platform/secrets"
 )
 
 type memoryStore struct{ writes int }
@@ -20,7 +24,8 @@ func (s *memoryStore) SaveDispatchIntent(context.Context, work.AgentDispatchInte
 	s.writes++
 	return nil
 }
-func (s *memoryStore) SaveAgentLease(context.Context, work.AgentLease) error { s.writes++; return nil }
+func (s *memoryStore) SaveAgentLease(context.Context, work.AgentLease) error             { s.writes++; return nil }
+func (s *memoryStore) RenewAgentLease(context.Context, work.AgentLease, time.Time) error { return nil }
 func (s *memoryStore) AcceptAgentCallback(context.Context, work.AgentCallback) error {
 	s.writes++
 	return nil
@@ -30,14 +35,24 @@ func main() {
 	enabled := flag.Bool("enable-canary", false, "explicitly enable the manual canary")
 	implementation := flag.String("implementation", "fake-worker", "allowlisted deterministic implementation")
 	postgresMode := flag.Bool("postgres", false, "use the Work PostgreSQL repository instead of in-memory storage")
+	runID := flag.String("run-id", "", "optional unique canary run suffix; generated when omitted")
 	flag.Parse()
 
-	now := time.UnixMilli(1_700_000_000_000)
-	intent := work.AgentDispatchIntent{DispatchID: "manual-dispatch-1", IdempotencyKey: "manual-idempotency-1", DelegationID: "manual-delegation-1", RunID: "manual-run-1", ImplementationID: *implementation, ImplementationRevision: 1, SourceRevision: "manual-source-1", WorktreeID: "manual-worktree-1", VerificationPlanRef: "manual-verification-1", CreatedAtUnixMillis: now.UnixMilli()}
+	suffix := *runID
+	if suffix == "" {
+		suffix = fmt.Sprintf("%d", time.Now().UnixNano())
+	}
+	now := time.Now().UTC()
+	intent := work.AgentDispatchIntent{DispatchID: "manual-dispatch-" + suffix, IdempotencyKey: "manual-idempotency-" + suffix, DelegationID: work.DelegationID("manual-delegation-" + suffix), RunID: work.AgentRunID("manual-run-" + suffix), ImplementationID: *implementation, ImplementationRevision: 1, SourceRevision: "manual-source-" + suffix, WorktreeID: "manual-worktree-" + suffix, VerificationPlanRef: "manual-verification-" + suffix, CreatedAtUnixMillis: now.UnixMilli()}
 	var store ports.ExecutionStore = &memoryStore{}
 	var closeStore func()
 	if *postgresMode {
-		repository, err := postgres.OpenWork(context.Background(), postgres.ContinuityConfig{Host: os.Getenv("POSTGRES_HOST"), Port: os.Getenv("POSTGRES_PORT"), Database: os.Getenv("POSTGRES_DATABASE"), User: os.Getenv("WORK_POSTGRES_USER"), Password: os.Getenv("WORK_POSTGRES_PASSWORD")})
+		password, err := resolveDatabasePassword()
+		if err != nil {
+			fmt.Fprintln(os.Stderr, err)
+			os.Exit(1)
+		}
+		repository, err := postgres.OpenWork(context.Background(), postgres.ContinuityConfig{Host: os.Getenv("POSTGRES_HOST"), Port: os.Getenv("POSTGRES_PORT"), Database: os.Getenv("POSTGRES_DATABASE"), User: os.Getenv("WORK_POSTGRES_USER"), Password: password})
 		if err != nil {
 			fmt.Fprintln(os.Stderr, err)
 			os.Exit(1)
@@ -67,3 +82,27 @@ func main() {
 		os.Exit(1)
 	}
 }
+
+func resolveDatabasePassword() (string, error) {
+	if _, ok := os.LookupEnv("WORK_POSTGRES_PASSWORD"); ok {
+		return "", fmt.Errorf("WORK_POSTGRES_PASSWORD direct environment secret is forbidden")
+	}
+	path, ref := os.Getenv("WORK_POSTGRES_PASSWORD_FILE"), os.Getenv("WORK_POSTGRES_CREDENTIAL_REF")
+	if path == "" || ref == "" {
+		return "", fmt.Errorf("WORK_POSTGRES_PASSWORD_FILE and WORK_POSTGRES_CREDENTIAL_REF are required")
+	}
+	root := os.Getenv("SECRET_BACKEND_ROOT")
+	if root == "" {
+		root = "/run/secrets"
+	}
+	cleanRoot := filepath.Clean(root)
+	cleanPath := filepath.Clean(path)
+	if cleanPath != cleanRoot && !strings.HasPrefix(cleanPath, cleanRoot+string(os.PathSeparator)) {
+		return "", fmt.Errorf("database password file escapes protected secret root")
+	}
+	resolver, err := platformsecrets.NewFileResolver(cleanRoot)
+	if err != nil {
+		return "", err
+	}
+	return resolver.ResolveFile(secretcontracts.SecretFileRef{CredentialRef: secretcontracts.CredentialRef(ref), Purpose: secretcontracts.SecretPurposeDatabasePassword, Consumer: secretcontracts.RuntimeConsumerWorkWorker, Path: cleanPath})
+}

cmd/agent-canary/main_test.go

@@ -0,0 +1,23 @@
+package main
+
+import (
+	"path/filepath"
+	"testing"
+)
+
+func TestResolveDatabasePasswordRejectsDirectEnvironmentSecret(t *testing.T) {
+	t.Setenv("WORK_POSTGRES_PASSWORD", "literal-secret")
+	if _, err := resolveDatabasePassword(); err == nil {
+		t.Fatal("direct database password environment variable was accepted")
+	}
+}
+
+func TestResolveDatabasePasswordRejectsPathOutsideProtectedRoot(t *testing.T) {
+	root := t.TempDir()
+	t.Setenv("SECRET_BACKEND_ROOT", root)
+	t.Setenv("WORK_POSTGRES_PASSWORD_FILE", filepath.Join(root, "..", "outside"))
+	t.Setenv("WORK_POSTGRES_CREDENTIAL_REF", "credential://domovoy/work/db-password/v1")
+	if _, err := resolveDatabasePassword(); err == nil {
+		t.Fatal("database password path outside protected root was accepted")
+	}
+}

cmd/agent-worker/local_fallback.go

@@ -0,0 +1,32 @@
+package main
+
+import (
+	"context"
+	"time"
+
+	"git.cin.su/m/Assistant/contracts/work"
+)
+
+// localExecutor provides the bounded local execution capability. It never
+// claims that an unsupported or external action happened. The work owner
+// records the WAITING callback and can retry after a suitable executor is
+// configured or a human chooses the next step.
+type localExecutor struct {
+	now func() time.Time
+}
+
+func newLocalExecutor() localExecutor {
+	return localExecutor{now: time.Now}
+}
+
+func (e localExecutor) Execute(_ context.Context, request work.AgentExecutionRequest) (work.AgentExecutionResult, error) {
+	now := time.Now()
+	if e.now != nil {
+		now = e.now()
+	}
+	return work.AgentExecutionResult{
+		State:                work.AgentExecutionWaiting,
+		Summary:              "Локальный исполнитель не поддерживает эту операцию; durable работа сохранена и ожидает подходящего исполнителя.",
+		ObservedAtUnixMillis: now.UnixMilli(),
+	}, nil
+}

cmd/agent-worker/local_fallback_test.go

@@ -0,0 +1,24 @@
+package main
+
+import (
+	"context"
+	"testing"
+	"time"
+
+	"git.cin.su/m/Assistant/contracts/work"
+)
+
+func TestLocalExecutorNeverClaimsUnsupportedExecution(t *testing.T) {
+	when := time.UnixMilli(1_700_000_000_000)
+	executor := localExecutor{now: func() time.Time { return when }}
+	result, err := executor.Execute(context.Background(), work.AgentExecutionRequest{Intent: work.AgentDispatchIntent{DispatchID: "dispatch-1"}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.State != work.AgentExecutionWaiting || result.ObservedAtUnixMillis != when.UnixMilli() {
+		t.Fatalf("unexpected fallback result: %+v", result)
+	}
+	if result.OutcomeRef != "" || result.Summary == "" {
+		t.Fatalf("local executor must not claim an outcome: %+v", result)
+	}
+}

cmd/agent-worker/main.go

@@ -7,6 +7,8 @@ import (
 	"fmt"
 	"net/http"
 	"os"
+	"path/filepath"
+	"strings"
 	"time"
 
 	"git.cin.su/m/Assistant/adapters/agentbridge"
@@ -39,22 +41,30 @@ func main() {
 	}
 	defer repository.Close()
 	var executor ports.AgentExecutor
-	if bridgeURL := os.Getenv("AGENT_EXECUTION_BRIDGE_URL"); bridgeURL != "" {
+	localFallback := false
+	if bridgeURL, bridgeConfigured := loadProtectedBridgeURL(); bridgeConfigured {
 		executor, err = agentbridge.NewHTTPExecutor(bridgeURL, &http.Client{Timeout: *leaseTTL})
 		if err != nil {
 			fmt.Fprintln(os.Stderr, err)
 			os.Exit(1)
 		}
 	} else {
-		fmt.Fprintln(os.Stderr, "agent execution bridge is not configured; dispatch claims are disabled; recovery-only mode")
+		// The local executor keeps supported task-qualification work alive, but
+		// it is not an external Agent Execution Bridge and must never be reported
+		// as one. Unsupported work remains durably WAITING.
+		localFallback = true
+		executor = newTaskEvaluationExecutor(envOr("COGNITIVE_WORKER_URL", "http://cognitive-worker:8083"), envOr("WORK_GRAPH_RUNTIME_URL", "http://work-graph-runtime:8088"), envOr("TASK_VERIFIER_URL", "http://task-verifier:8091"), &http.Client{Timeout: *leaseTTL})
+		fmt.Fprintln(os.Stderr, "protected agent execution bridge URL is not configured; task evaluations use the local durable evaluator, other work remains in fallback")
 	}
+	runtimeStatus := newAgentWorkerRuntimeStatus()
 	go func() {
-		if serveErr := http.ListenAndServe(*statusAddr, workerStatusHandler(executor != nil)); serveErr != nil {
+		if serveErr := http.ListenAndServe(*statusAddr, workerStatusHandler(executor != nil && !localFallback, localFallback, runtimeStatus)); serveErr != nil {
 			fmt.Fprintln(os.Stderr, "agent worker status server:", serveErr)
 		}
 	}()
 	ctx := context.Background()
 	for {
+		runtimeStatus.cycleStarted(time.Now())
 		var result application.WorkerCycleResult
 		var cycleErr error
 		if executor == nil {
@@ -63,11 +73,19 @@ func main() {
 			result, cycleErr = application.ExecuteWorkerCycleWithExecutor(ctx, repository, executor, *workerID, work.AgentFencingToken(1), time.Now().UTC(), *staleAfter, *leaseTTL)
 		}
 		if cycleErr != nil {
+			runtimeStatus.cycleFailed(time.Now(), cycleErr)
 			fmt.Fprintln(os.Stderr, cycleErr)
 			if *once {
 				os.Exit(1)
 			}
 		}
+		cycleResult := "idle"
+		if result.Completed {
+			cycleResult = "completed"
+		} else if result.RecoveredClaims > 0 {
+			cycleResult = "recovered"
+		}
+		runtimeStatus.cycleFinished(time.Now(), result.RecoveredClaims, result.DispatchID, cycleResult)
 		if cycleErr == nil && (result.Completed || result.RecoveredClaims > 0) {
 			_ = json.NewEncoder(os.Stdout).Encode(result)
 		}
@@ -81,6 +99,34 @@ func main() {
 	}
 }
 
+func loadProtectedBridgeURL() (string, bool) {
+	return loadProtectedBridgeURLFrom("/run/secrets", "/run/secrets/agent_execution_bridge_url")
+}
+
+func loadProtectedBridgeURLFrom(root, path string) (string, bool) {
+	clean := filepath.Clean(path)
+	cleanRoot := filepath.Clean(root)
+	if clean != cleanRoot && !strings.HasPrefix(clean, cleanRoot+string(os.PathSeparator)) {
+		return "", false
+	}
+	payload, err := os.ReadFile(clean)
+	if err != nil {
+		return "", false
+	}
+	value := strings.TrimSpace(string(payload))
+	if value == "" || strings.ContainsAny(value, "\r\n") {
+		return "", false
+	}
+	return value, true
+}
+
+func envOr(name, fallback string) string {
+	if value := strings.TrimSpace(os.Getenv(name)); value != "" {
+		return value
+	}
+	return fallback
+}
+
 func resolveDatabasePassword(name, refName string, consumer secretcontracts.RuntimeConsumer) (string, error) {
 	if _, ok := os.LookupEnv(name); ok {
 		return "", fmt.Errorf("%s direct environment secret is forbidden", name)

cmd/agent-worker/status.go

@@ -3,28 +3,104 @@ package main
 import (
 	"encoding/json"
 	"net/http"
+	"sync"
+	"time"
 )
 
 type workerStatusResponse struct {
-	Runtime            string `json:"runtime"`
-	ExecutorConfigured bool   `json:"executor_configured"`
-	Mode               string `json:"mode"`
-	Ready              bool   `json:"ready"`
-	Reason             string `json:"reason"`
+	Runtime              string    `json:"runtime"`
+	ExecutorConfigured   bool      `json:"executor_configured"`
+	Mode                 string    `json:"mode"`
+	Ready                bool      `json:"ready"`
+	Reason               string    `json:"reason"`
+	LocalFallback        bool      `json:"local_fallback"`
+	LocalExecutor        bool      `json:"local_executor"`
+	Capabilities         []string  `json:"capabilities,omitempty"`
+	StartedAt            time.Time `json:"started_at,omitempty"`
+	LastCycleAt          time.Time `json:"last_cycle_at,omitempty"`
+	LastCompletedAt      time.Time `json:"last_completed_at,omitempty"`
+	LastRecoveredClaims  int64     `json:"last_recovered_claims"`
+	TotalRecoveredClaims int64     `json:"total_recovered_claims"`
+	LastDispatchID       string    `json:"last_dispatch_id,omitempty"`
+	LastResult           string    `json:"last_result,omitempty"`
+	LastError            string    `json:"last_error,omitempty"`
 }
 
-func workerStatusHandler(executorConfigured bool) http.Handler {
-	status := workerStatusResponse{
+type agentWorkerRuntimeStatus struct {
+	mu                   sync.RWMutex
+	startedAt            time.Time
+	lastCycleAt          time.Time
+	lastCompletedAt      time.Time
+	lastRecoveredClaims  int64
+	totalRecoveredClaims int64
+	lastDispatchID       string
+	lastResult           string
+	lastError            string
+}
+
+func newAgentWorkerRuntimeStatus() *agentWorkerRuntimeStatus {
+	return &agentWorkerRuntimeStatus{startedAt: time.Now().UTC()}
+}
+
+func (s *agentWorkerRuntimeStatus) cycleStarted(at time.Time) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.lastCycleAt = at.UTC()
+	s.lastResult = "running"
+	s.lastError = ""
+}
+
+func (s *agentWorkerRuntimeStatus) cycleFinished(at time.Time, recoveredClaims int64, dispatchID string, result string) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.lastCompletedAt = at.UTC()
+	s.lastRecoveredClaims = recoveredClaims
+	s.totalRecoveredClaims += recoveredClaims
+	s.lastDispatchID = dispatchID
+	s.lastResult = result
+	s.lastError = ""
+}
+
+func (s *agentWorkerRuntimeStatus) cycleFailed(at time.Time, err error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.lastCompletedAt = at.UTC()
+	s.lastResult = "failed"
+	s.lastError = err.Error()
+}
+
+func workerStatusHandler(executorConfigured bool, localFallback ...interface{}) http.Handler {
+	local := false
+	var runtime *agentWorkerRuntimeStatus
+	for _, argument := range localFallback {
+		switch value := argument.(type) {
+		case bool:
+			local = value
+		case *agentWorkerRuntimeStatus:
+			runtime = value
+		}
+	}
+	if runtime == nil {
+		runtime = newAgentWorkerRuntimeStatus()
+	}
+	baseStatus := workerStatusResponse{
 		Runtime:            "agent-worker",
 		ExecutorConfigured: executorConfigured,
 		Mode:               "recovery-only",
 		Ready:              false,
-		Reason:             "AGENT_EXECUTION_BRIDGE_URL is not configured",
+		Reason:             "protected agent execution bridge URL is not configured",
+		LocalFallback:      local,
+		LocalExecutor:      local,
+	}
+	if local {
+		baseStatus.Mode = "local-executor"
+		baseStatus.Reason = "bounded local task-evaluation executor is active; unsupported work remains durably waiting"
+		baseStatus.Capabilities = []string{"task-evaluation"}
 	}
 	if executorConfigured {
-		status.Mode = "execution"
-		status.Ready = true
-		status.Reason = "typed agent execution bridge configured"
+		baseStatus.Mode = "execution"
+		baseStatus.Ready = true
+		baseStatus.Reason = "typed agent execution bridge configured"
 	}
 	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		if r.Method != http.MethodGet || (r.URL.Path != "/healthz" && r.URL.Path != "/readyz" && r.URL.Path != "/status") {
@@ -32,9 +108,20 @@ func workerStatusHandler(executorConfigured bool) http.Handler {
 			return
 		}
 		w.Header().Set("Content-Type", "application/json")
-		if r.URL.Path == "/readyz" && !status.Ready {
+		if r.URL.Path == "/readyz" && !baseStatus.Ready {
 			w.WriteHeader(http.StatusServiceUnavailable)
 		}
-		_ = json.NewEncoder(w).Encode(status)
+		response := baseStatus
+		runtime.mu.RLock()
+		response.StartedAt = runtime.startedAt
+		response.LastCycleAt = runtime.lastCycleAt
+		response.LastCompletedAt = runtime.lastCompletedAt
+		response.LastRecoveredClaims = runtime.lastRecoveredClaims
+		response.TotalRecoveredClaims = runtime.totalRecoveredClaims
+		response.LastDispatchID = runtime.lastDispatchID
+		response.LastResult = runtime.lastResult
+		response.LastError = runtime.lastError
+		runtime.mu.RUnlock()
+		_ = json.NewEncoder(w).Encode(response)
 	})
 }

cmd/agent-worker/status_test.go

@@ -4,9 +4,29 @@ import (
 	"encoding/json"
 	"net/http"
 	"net/http/httptest"
+	"os"
+	"path/filepath"
 	"testing"
+	"time"
 )
 
+func TestLoadProtectedBridgeURL(t *testing.T) {
+	root := t.TempDir()
+	path := filepath.Join(root, "agent_execution_bridge_url")
+	if err := os.WriteFile(path, []byte("http://bridge.local\n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	if value, configured := loadProtectedBridgeURLFrom(root, path); !configured || value != "http://bridge.local" {
+		t.Fatalf("protected bridge was not loaded: %q %v", value, configured)
+	}
+	if value, configured := loadProtectedBridgeURLFrom(root, filepath.Join(root, "missing")); configured || value != "" {
+		t.Fatalf("missing bridge was accepted: %q %v", value, configured)
+	}
+	if value, configured := loadProtectedBridgeURLFrom(root, "/tmp/not-protected"); configured || value != "" {
+		t.Fatalf("path outside protected root was accepted: %q %v", value, configured)
+	}
+}
+
 func TestWorkerStatusReportsRecoveryOnlyWithoutExecutor(t *testing.T) {
 	response := httptest.NewRecorder()
 	workerStatusHandler(false).ServeHTTP(response, httptest.NewRequest(http.MethodGet, "/status", nil))
@@ -19,6 +39,33 @@ func TestWorkerStatusReportsRecoveryOnlyWithoutExecutor(t *testing.T) {
 	}
 }
 
+func TestWorkerStatusReportsLocalExecutor(t *testing.T) {
+	response := httptest.NewRecorder()
+	workerStatusHandler(false, true).ServeHTTP(response, httptest.NewRequest(http.MethodGet, "/status", nil))
+	var value workerStatusResponse
+	if err := json.NewDecoder(response.Body).Decode(&value); err != nil {
+		t.Fatal(err)
+	}
+	if response.Code != http.StatusOK || value.Ready || value.Mode != "local-executor" || !value.LocalFallback || !value.LocalExecutor {
+		t.Fatalf("unexpected status: code=%d value=%+v", response.Code, value)
+	}
+	if len(value.Capabilities) != 1 || value.Capabilities[0] != "task-evaluation" {
+		t.Fatalf("unexpected local capabilities: %+v", value.Capabilities)
+	}
+}
+
+func TestWorkerStatusDoesNotCallLocalFallbackAnExternalExecutor(t *testing.T) {
+	response := httptest.NewRecorder()
+	workerStatusHandler(false, true).ServeHTTP(response, httptest.NewRequest(http.MethodGet, "/status", nil))
+	var value workerStatusResponse
+	if err := json.NewDecoder(response.Body).Decode(&value); err != nil {
+		t.Fatal(err)
+	}
+	if value.ExecutorConfigured || value.Ready || !value.LocalFallback || !value.LocalExecutor || value.Mode != "local-executor" {
+		t.Fatalf("local fallback was presented as external execution: %+v", value)
+	}
+}
+
 func TestWorkerReadyzRequiresExecutor(t *testing.T) {
 	response := httptest.NewRecorder()
 	workerStatusHandler(false).ServeHTTP(response, httptest.NewRequest(http.MethodGet, "/readyz", nil))
@@ -31,3 +78,32 @@ func TestWorkerReadyzRequiresExecutor(t *testing.T) {
 		t.Fatalf("configured readyz status=%d", response.Code)
 	}
 }
+
+func TestWorkerStatusReportsRecoveryCycle(t *testing.T) {
+	runtime := newAgentWorkerRuntimeStatus()
+	runtime.cycleStarted(time.Date(2026, 8, 23, 10, 0, 0, 0, time.UTC))
+	runtime.cycleFinished(time.Date(2026, 8, 23, 10, 0, 1, 0, time.UTC), 3, "dispatch-1", "recovered")
+	response := httptest.NewRecorder()
+	workerStatusHandler(false, true, runtime).ServeHTTP(response, httptest.NewRequest(http.MethodGet, "/status", nil))
+	var value workerStatusResponse
+	if err := json.NewDecoder(response.Body).Decode(&value); err != nil {
+		t.Fatal(err)
+	}
+	if value.LastRecoveredClaims != 3 || value.TotalRecoveredClaims != 3 || value.LastDispatchID != "dispatch-1" || value.LastResult != "recovered" {
+		t.Fatalf("recovery evidence missing: %+v", value)
+	}
+}
+
+func TestWorkerStatusReportsCycleError(t *testing.T) {
+	runtime := newAgentWorkerRuntimeStatus()
+	runtime.cycleFailed(time.Date(2026, 8, 23, 10, 0, 1, 0, time.UTC), os.ErrClosed)
+	response := httptest.NewRecorder()
+	workerStatusHandler(false, runtime).ServeHTTP(response, httptest.NewRequest(http.MethodGet, "/status", nil))
+	var value workerStatusResponse
+	if err := json.NewDecoder(response.Body).Decode(&value); err != nil {
+		t.Fatal(err)
+	}
+	if value.LastResult != "failed" || value.LastError == "" {
+		t.Fatalf("cycle error missing: %+v", value)
+	}
+}

cmd/agent-worker/task_evaluation_executor.go

@@ -0,0 +1,109 @@
+package main
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"strings"
+	"time"
+
+	"git.cin.su/m/Assistant/contracts/artifact"
+	"git.cin.su/m/Assistant/contracts/capability"
+	"git.cin.su/m/Assistant/contracts/work"
+	capabilitydomain "git.cin.su/m/Assistant/modules/capability/domain"
+)
+
+type taskEvaluationExecutor struct {
+	cognitiveURL string
+	callbackURL  string
+	verifierURL  string
+	client       *http.Client
+	fallback     localExecutor
+}
+
+func newTaskEvaluationExecutor(cognitiveURL, callbackURL, verifierURL string, client *http.Client) taskEvaluationExecutor {
+	if client == nil {
+		client = http.DefaultClient
+	}
+	return taskEvaluationExecutor{cognitiveURL: strings.TrimRight(cognitiveURL, "/"), callbackURL: strings.TrimRight(callbackURL, "/"), verifierURL: strings.TrimRight(verifierURL, "/"), client: client, fallback: newLocalExecutor()}
+}
+
+func (e taskEvaluationExecutor) Execute(ctx context.Context, request work.AgentExecutionRequest) (work.AgentExecutionResult, error) {
+	intent := request.Intent
+	if intent.TaskEvaluationExecution == nil || intent.TaskEvaluationProfile == nil || len(intent.TaskEvaluationCases) == 0 {
+		return e.fallback.Execute(ctx, request)
+	}
+	execution := intent.TaskEvaluationExecution
+	if execution.GraphID == "" || execution.NodeID == "" || execution.BindingRevision < 1 || execution.ActorID == "" || execution.PrincipalID == "" || execution.ScopeID == "" {
+		return e.fallback.Execute(ctx, request)
+	}
+	if e.cognitiveURL == "" || e.callbackURL == "" || e.verifierURL == "" {
+		return e.fallback.Execute(ctx, request)
+	}
+	payload := capability.TaskEvaluationExecutionRequest{Profile: *intent.TaskEvaluationProfile, CapabilityID: intent.TaskEvaluationProfile.CapabilityID, BindingID: intent.TaskEvaluationBindingID, BindingRevision: execution.BindingRevision, ActorID: execution.ActorID, PrincipalID: execution.PrincipalID, ScopeID: execution.ScopeID, Cases: intent.TaskEvaluationCases, MaxOutputBytes: 4096}
+	response, err := postJSON[capability.TaskEvaluationExecutionResponse](ctx, e.client, e.cognitiveURL+"/v1/task-evaluations/execute", payload)
+	if err != nil {
+		return work.AgentExecutionResult{}, err
+	}
+	if len(response.Responses) == 0 || response.ObservedAtUnixMillis < 1 {
+		return work.AgentExecutionResult{}, fmt.Errorf("task evaluation returned no observations")
+	}
+	if len(intent.TaskEvaluationCases) == 0 {
+		return work.AgentExecutionResult{}, fmt.Errorf("task evaluation cases are missing")
+	}
+	caseByID := make(map[string]capability.TaskEvaluationCase, len(intent.TaskEvaluationCases))
+	for _, item := range intent.TaskEvaluationCases {
+		caseByID[item.CaseID] = item
+	}
+	observations := make([]capability.TaskCaseObservation, 0, len(response.Responses))
+	for _, item := range response.Responses {
+		caseValue, ok := caseByID[item.CaseID]
+		if !ok || caseValue.Verification.Mode == "" {
+			return e.fallback.Execute(ctx, request)
+		}
+		verified, verifyErr := postJSON[capability.TaskCaseVerificationResponse](ctx, e.client, e.verifierURL+"/v1/task-evaluations/verify", capability.TaskCaseVerificationRequest{CaseID: item.CaseID, ResponseText: item.ResponseText, LatencyMicros: item.LatencyMicros, EvidenceArtifactRef: item.EvidenceArtifactRef, Verification: caseValue.Verification})
+		if verifyErr != nil {
+			return e.fallback.Execute(ctx, request)
+		}
+		observations = append(observations, capability.TaskCaseObservation{CaseID: verified.CaseID, Result: verified.Result, QualityScoreMillis: verified.QualityScoreMillis, LatencyMicros: verified.LatencyMicros, CriticalFailure: verified.CriticalFailure, EvidenceArtifactRef: verified.EvidenceArtifactRef, VerificationEvidenceRef: verified.VerificationEvidenceRef, VerifierIdentity: verified.VerifierIdentity})
+	}
+	evaluation, err := capabilitydomain.BuildTaskEvaluation(*intent.TaskEvaluationProfile, capability.TaskEvaluationID("evaluation:"+intent.DispatchID), response.BindingID, response.BindingRevision, intent.TaskEvaluationProfile.CapabilityID, observations, artifact.ArtifactID(intent.TaskEvaluationCases[0].EvidenceArtifactRef), response.ObservedAtUnixMillis)
+	if err != nil {
+		return work.AgentExecutionResult{}, err
+	}
+	callback := work.TaskEvaluationEvidenceCallback{GraphID: execution.GraphID, Revision: execution.GraphRevision, NodeID: execution.NodeID, Evaluation: evaluation, Succeeded: true}
+	if _, err := postJSON[json.RawMessage](ctx, e.client, e.callbackURL+"/v1/work-graphs/task-evaluations/callback", callback); err != nil {
+		return work.AgentExecutionResult{}, err
+	}
+	return work.AgentExecutionResult{State: work.AgentExecutionCompleted, Summary: "task evaluation executed, aggregated and accepted by Work Graph callback", OutcomeRef: string(evaluation.ID), ArtifactRefs: []string{string(evaluation.ArtifactID)}, ObservedAtUnixMillis: time.Now().UnixMilli()}, nil
+}
+
+func postJSON[T any](ctx context.Context, client *http.Client, url string, payload any) (T, error) {
+	var result T
+	body, err := json.Marshal(payload)
+	if err != nil {
+		return result, err
+	}
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
+	if err != nil {
+		return result, err
+	}
+	req.Header.Set("Content-Type", "application/json")
+	resp, err := client.Do(req)
+	if err != nil {
+		return result, err
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return result, fmt.Errorf("POST %s: status %s", url, resp.Status)
+	}
+	if resp.StatusCode == http.StatusNoContent {
+		return result, nil
+	}
+	if err := json.NewDecoder(resp.Body).Decode(&result); err != nil {
+		return result, err
+	}
+	return result, nil
+}

cmd/agent-worker/task_evaluation_executor_test.go

@@ -0,0 +1,43 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"git.cin.su/m/Assistant/contracts/capability"
+	"git.cin.su/m/Assistant/contracts/work"
+)
+
+func TestTaskEvaluationExecutorInvokesAggregatesAndCallbacks(t *testing.T) {
+	profile := capability.TaskEvaluationProfile{ID: "profile-1", Revision: 1, CapabilityID: "capability-1", Stage: capability.TaskEvaluationGeneral, BenchmarkVersion: "general-v1", RequiredCaseIDs: []string{"case-1"}, QualityWeightMillis: 600, ReliabilityWeightMillis: 300, LatencyWeightMillis: 100}
+	verification := capability.TaskCaseVerification{Mode: capability.TaskCaseVerificationExactText, ExpectedText: "answer", MinimumQualityMillis: 900}
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path == "/v1/task-evaluations/execute" {
+			_ = json.NewEncoder(w).Encode(capability.TaskEvaluationExecutionResponse{ProfileID: profile.ID, BindingID: "binding-1", BindingRevision: 3, Responses: []capability.TaskCaseResponse{{CaseID: "case-1", ResponseText: "answer", LatencyMicros: 2, EvidenceArtifactRef: "artifact-1"}}, ObservedAtUnixMillis: 10})
+			return
+		}
+		if r.URL.Path == "/v1/task-evaluations/verify" {
+			_ = json.NewEncoder(w).Encode(capability.TaskCaseVerificationResponse{CaseID: "case-1", Result: capability.EvaluationSupported, QualityScoreMillis: 900, LatencyMicros: 2, EvidenceArtifactRef: "artifact-1", VerificationEvidenceRef: "verify-1", VerifierIdentity: "verifier-1"})
+			return
+		}
+		if r.URL.Path == "/v1/work-graphs/task-evaluations/callback" {
+			var callback work.TaskEvaluationEvidenceCallback
+			if err := json.NewDecoder(r.Body).Decode(&callback); err != nil || callback.GraphID != "graph-1" || callback.NodeID != "node-1" || callback.Evaluation.Result != capability.EvaluationSupported {
+				http.Error(w, "invalid callback", http.StatusBadRequest)
+				return
+			}
+			w.WriteHeader(http.StatusNoContent)
+			return
+		}
+		http.NotFound(w, r)
+	}))
+	defer server.Close()
+	executor := newTaskEvaluationExecutor(server.URL, server.URL, server.URL, server.Client())
+	result, err := executor.Execute(context.Background(), work.AgentExecutionRequest{Intent: work.AgentDispatchIntent{DispatchID: "dispatch-1", TaskEvaluationBindingID: "binding-1", TaskEvaluationProfile: &profile, TaskEvaluationCases: []capability.TaskEvaluationCase{{CaseID: "case-1", Prompt: "answer", EvidenceArtifactRef: "artifact-1", Verification: verification}}, TaskEvaluationExecution: &work.TaskEvaluationDispatchContext{GraphID: "graph-1", GraphRevision: 1, NodeID: "node-1", BindingRevision: 3, ActorID: "actor-1", PrincipalID: "principal-1", ScopeID: "scope-1"}}})
+	if err != nil || result.State != work.AgentExecutionCompleted || result.OutcomeRef == "" {
+		t.Fatalf("result=%+v err=%v", result, err)
+	}
+}

cmd/cognitive-benchmark/main.go

@@ -98,8 +98,9 @@ func main() {
 	manifestPath := flag.String("manifest", "", "benchmark manifest JSON")
 	outputPath := flag.String("output", "", "output artifact JSON")
 	timeout := flag.Duration("case-timeout", 30*time.Second, "deadline for each case")
+	maxOutputBytes := flag.Int64("max-output-bytes", 4096, "maximum advisory output bytes per case")
 	flag.Parse()
-	if *manifestPath == "" || *outputPath == "" || *timeout <= 0 {
+	if *manifestPath == "" || *outputPath == "" || *timeout <= 0 || *maxOutputBytes < 1 {
 		fatal(errors.New("manifest, output and positive case-timeout are required"))
 	}
 	input, err := os.ReadFile(*manifestPath)
@@ -110,7 +111,7 @@ func main() {
 	if err := json.Unmarshal(input, &plan); err != nil {
 		fatal(err)
 	}
-	artifact, err := run(context.Background(), plan, *timeout, time.Now().UTC())
+	artifact, err := run(context.Background(), plan, *timeout, *maxOutputBytes, time.Now().UTC())
 	if err != nil {
 		fatal(err)
 	}
@@ -125,7 +126,7 @@ func main() {
 	fmt.Printf("artifact=%s models=%d output=%s\n", artifact.ArtifactID, len(artifact.Results), *outputPath)
 }
 
-func run(ctx context.Context, plan manifest, timeout time.Duration, observed time.Time) (benchmarkArtifact, error) {
+func run(ctx context.Context, plan manifest, timeout time.Duration, maxOutputBytes int64, observed time.Time) (benchmarkArtifact, error) {
 	if plan.ArtifactID == "" || len(plan.Providers) == 0 {
 		return benchmarkArtifact{}, errors.New("invalid benchmark manifest")
 	}
@@ -134,7 +135,7 @@ func run(ctx context.Context, plan manifest, timeout time.Duration, observed tim
 		artifact.Corpus = append(artifact.Corpus, corpusRecord{ID: item.id, Prompt: item.prompt, PromptSHA256: digest(item.prompt), Critical: item.critical})
 	}
 	for _, provider := range plan.Providers {
-		adapter, err := newProvider(provider)
+		adapter, err := newProvider(provider, timeout)
 		if err != nil {
 			return benchmarkArtifact{}, err
 		}
@@ -142,7 +143,7 @@ func run(ctx context.Context, plan manifest, timeout time.Duration, observed tim
 			return provider.Candidates[i].ProviderModelID < provider.Candidates[j].ProviderModelID
 		})
 		for _, candidate := range provider.Candidates {
-			result, err := runModel(ctx, adapter, provider.ProviderInstanceID, candidate, timeout)
+			result, err := runModel(ctx, adapter, provider.ProviderInstanceID, candidate, timeout, maxOutputBytes)
 			if err != nil {
 				return benchmarkArtifact{}, err
 			}
@@ -158,10 +159,10 @@ func run(ctx context.Context, plan manifest, timeout time.Duration, observed tim
 	return artifact, nil
 }
 
-func newProvider(candidate providerCandidate) (models.LocalProvider, error) {
+func newProvider(candidate providerCandidate, timeout time.Duration) (models.LocalProvider, error) {
 	switch candidate.ProviderKind {
 	case "openai-compatible-chat-v1":
-		return models.NewOpenAICompatibleLocalAdapter(candidate.Endpoint, nil)
+		return models.NewOpenAICompatibleLocalAdapterWithAPIKeyAndTimeout(candidate.Endpoint, "", nil, timeout)
 	case "ollama-v1":
 		return models.NewOllamaAdapter(candidate.Endpoint, nil)
 	default:
@@ -169,7 +170,7 @@ func newProvider(candidate providerCandidate) (models.LocalProvider, error) {
 	}
 }
 
-func runModel(ctx context.Context, adapter models.LocalProvider, providerID string, candidate modelCandidate, timeout time.Duration) (modelResult, error) {
+func runModel(ctx context.Context, adapter models.LocalProvider, providerID string, candidate modelCandidate, timeout time.Duration, maxOutputBytes int64) (modelResult, error) {
 	if providerID == "" || candidate.ProviderModelID == "" || candidate.ModelDescriptorID == "" || candidate.ModelRevision < 1 || candidate.BindingID == "" || candidate.BindingRevision < 1 {
 		return modelResult{}, errors.New("invalid benchmark candidate")
 	}
@@ -178,7 +179,7 @@ func runModel(ctx context.Context, adapter models.LocalProvider, providerID stri
 	for index, item := range corpus {
 		caseCtx, cancel := context.WithTimeout(ctx, timeout)
 		started := time.Now()
-		response, err := adapter.Advise(caseCtx, cognitive.AdvisoryRequest{Version: cognitive.CreatorAdvisoryV1, RequestID: cognitive.RequestID(fmt.Sprintf("bench-%s-%d", digest(candidate.BindingID)[:12], index)), ActorID: "benchmark-actor", PrincipalID: "domovoy-creator-principal", ScopeID: "benchmark-public-synthetic", Message: item.prompt, Locale: "ru-RU", MaxOutputBytes: 4096, ProviderBindingID: capability.CapabilityBindingID(candidate.BindingID), ProviderBindingRevision: candidate.BindingRevision, ProviderInstanceID: capability.ProviderInstanceID(providerID), ModelDescriptorID: capability.ModelDescriptorID(candidate.ModelDescriptorID), ProviderModelID: capability.ProviderModelID(candidate.ProviderModelID)})
+		response, err := adapter.Advise(caseCtx, cognitive.AdvisoryRequest{Version: cognitive.CreatorAdvisoryV1, RequestID: cognitive.RequestID(fmt.Sprintf("bench-%s-%d", digest(candidate.BindingID)[:12], index)), ActorID: "benchmark-actor", PrincipalID: "domovoy-creator-principal", ScopeID: "benchmark-public-synthetic", Message: item.prompt, Locale: "ru-RU", MaxOutputBytes: int(maxOutputBytes), ProviderBindingID: capability.CapabilityBindingID(candidate.BindingID), ProviderBindingRevision: candidate.BindingRevision, ProviderInstanceID: capability.ProviderInstanceID(providerID), ModelDescriptorID: capability.ModelDescriptorID(candidate.ModelDescriptorID), ProviderModelID: capability.ProviderModelID(candidate.ProviderModelID)})
 		cancel()
 		caseResult := caseResult{CaseID: item.id, LatencyMillis: time.Since(started).Milliseconds()}
 		if err != nil {

cmd/cognitive-worker/discovery.go

@@ -3,6 +3,7 @@ package main
 import (
 	"encoding/json"
 	"net/http"
+	"sort"
 	"time"
 
 	"git.cin.su/m/Assistant/contracts/capability"
@@ -17,6 +18,10 @@ type discoveryService struct {
 }
 
 func (s discoveryService) handler(w http.ResponseWriter, r *http.Request) {
+	if r.Method == http.MethodGet && r.URL.Path == "/v1/provider-discovery/status" {
+		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)
@@ -35,3 +40,42 @@ func (s discoveryService) handler(w http.ResponseWriter, r *http.Request) {
 	w.Header().Set("Content-Type", "application/json")
 	_ = json.NewEncoder(w).Encode(result)
 }
+
+type discoveryStatus struct {
+	Provider             string                          `json:"provider"`
+	ProviderKind         capability.ProviderKind         `json:"provider_kind"`
+	Health               capability.HealthState          `json:"health"`
+	FailureClass         capability.CognitiveFailureType `json:"failure_class,omitempty"`
+	ObservedAtUnixMillis int64                           `json:"observed_at_unix_millis"`
+	Models               []discoveredModelStatus         `json:"models"`
+}
+
+type discoveredModelStatus struct {
+	ProviderModelID capability.ProviderModelID   `json:"provider_model_id"`
+	Fingerprint     capability.ModelFingerprint  `json:"fingerprint,omitempty"`
+	Capabilities    []capability.ModelCapability `json:"capabilities,omitempty"`
+}
+
+func (s discoveryService) statusHandler(w http.ResponseWriter, r *http.Request) {
+	providers := make([]string, 0, len(s.providers))
+	for name := range s.providers {
+		providers = append(providers, name)
+	}
+	sort.Strings(providers)
+	result := make([]discoveryStatus, 0, len(providers))
+	for _, name := range providers {
+		registration := s.providers[name]
+		status := discoveryStatus{Provider: name, ProviderKind: registration.Kind, Health: capability.HealthUnknown, ObservedAtUnixMillis: s.now().UnixMilli(), Models: []discoveredModelStatus{}}
+		if registration.Discovery != nil {
+			models, probe := registration.Discovery.Discover(r.Context())
+			status.Health = probe.Health
+			status.FailureClass = capability.CognitiveFailureType(probe.Failure)
+			for _, model := range models {
+				status.Models = append(status.Models, discoveredModelStatus{ProviderModelID: model.ProviderModelID, Fingerprint: model.Fingerprint, Capabilities: model.Capabilities})
+			}
+		}
+		result = append(result, status)
+	}
+	w.Header().Set("Content-Type", "application/json")
+	_ = json.NewEncoder(w).Encode(result)
+}

cmd/cognitive-worker/discovery_test.go

@@ -0,0 +1,52 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+	"time"
+
+	"git.cin.su/m/Assistant/adapters/models"
+	"git.cin.su/m/Assistant/contracts/capability"
+	"git.cin.su/m/Assistant/contracts/cognitive"
+	"git.cin.su/m/Assistant/contracts/identity"
+)
+
+type discoveryTestProvider struct{}
+
+func (discoveryTestProvider) Probe(context.Context) models.ProviderProbe {
+	return models.ProviderProbe{Health: capability.HealthAvailable}
+}
+func (discoveryTestProvider) Discover(context.Context) ([]models.DiscoveredModel, models.ProviderProbe) {
+	return []models.DiscoveredModel{{ProviderModelID: "local-model", Fingerprint: "fingerprint-1", Capabilities: []capability.ModelCapability{capability.CapabilityAdvisory}}}, models.ProviderProbe{Health: capability.HealthAvailable}
+}
+func (discoveryTestProvider) Advise(context.Context, cognitive.AdvisoryRequest) (cognitive.AdvisoryResponse, error) {
+	return cognitive.AdvisoryResponse{}, nil
+}
+
+func TestDiscoveryStatusIsReadOnlyAndDoesNotExposeCredentialMetadata(t *testing.T) {
+	service := discoveryService{
+		providers: map[string]providerRegistration{"cpu": {ProviderInstanceID: "provider-1", Kind: "openai-compatible-chat-v1", CredentialRef: "credential://must-not-be-returned", Discovery: discoveryTestProvider{}}},
+		owner:     identity.PrincipalID("owner-1"),
+		now:       func() time.Time { return time.UnixMilli(1234) },
+	}
+	response := httptest.NewRecorder()
+	service.handler(response, httptest.NewRequest(http.MethodGet, "/v1/provider-discovery/status", nil))
+	if response.Code != http.StatusOK {
+		t.Fatalf("status=%d", response.Code)
+	}
+	body := response.Body.String()
+	var result []discoveryStatus
+	if err := json.NewDecoder(strings.NewReader(body)).Decode(&result); err != nil {
+		t.Fatal(err)
+	}
+	if len(result) != 1 || result[0].Provider != "cpu" || result[0].Health != capability.HealthAvailable || len(result[0].Models) != 1 || result[0].ObservedAtUnixMillis != 1234 {
+		t.Fatalf("unexpected discovery status: %+v", result)
+	}
+	if strings.Contains(body, "must-not-be-returned") || strings.Contains(body, "credential://") {
+		t.Fatalf("credential metadata leaked in discovery status: %s", body)
+	}
+}

cmd/cognitive-worker/fleet_status.go

@@ -17,6 +17,8 @@ type fleetEndpoint struct {
 	ModelID    string
 	Tier       capability.FleetTier
 	URL        string
+	Headers    map[string]string
+	Client     *http.Client
 }
 
 type fleetStatusResponse struct {
@@ -37,7 +39,14 @@ func fleetReadinessHandler(client *http.Client, preferred capability.FleetTier,
 			}
 			request, err := http.NewRequestWithContext(r.Context(), http.MethodGet, endpoint.URL, nil)
 			if err == nil {
-				response, requestErr := client.Do(request)
+				for name, value := range endpoint.Headers {
+					request.Header.Set(name, value)
+				}
+				probeClient := endpoint.Client
+				if probeClient == nil {
+					probeClient = client
+				}
+				response, requestErr := probeClient.Do(request)
 				if requestErr == nil {
 					response.Body.Close()
 					if response.StatusCode >= 200 && response.StatusCode < 300 {

cmd/cognitive-worker/fleet_status_test.go

@@ -9,7 +9,7 @@ import (
 	"git.cin.su/m/Assistant/contracts/capability"
 )
 
-func TestFleetReadinessSelectsLifeboatWhenPrimaryTiersUnavailable(t *testing.T) {
+func TestFleetReadinessSelectsBestAvailablePrimaryTierBeforeLifeboat(t *testing.T) {
 	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) { w.WriteHeader(http.StatusOK) }))
 	defer server.Close()
 	handler := fleetReadinessHandler(server.Client(), capability.FleetTierHeavy, []fleetEndpoint{
@@ -21,7 +21,26 @@ func TestFleetReadinessSelectsLifeboatWhenPrimaryTiersUnavailable(t *testing.T)
 	recorder := httptest.NewRecorder()
 	request := httptest.NewRequest(http.MethodGet, "/v1/fleet/readiness", nil)
 	handler.ServeHTTP(recorder, request)
-	if recorder.Code != http.StatusOK || !strings.Contains(recorder.Body.String(), `"selected_tier":"LIFEBOAT"`) || !strings.Contains(recorder.Body.String(), `"tier":"MEDIUM"`) {
+	if recorder.Code != http.StatusOK || !strings.Contains(recorder.Body.String(), `"selected_tier":"FAST"`) || !strings.Contains(recorder.Body.String(), `"tier":"MEDIUM"`) {
+		t.Fatalf("status=%d body=%s", recorder.Code, recorder.Body.String())
+	}
+}
+
+func TestFleetReadinessSendsConfiguredProviderHeaders(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.Header.Get("Authorization") != "Bearer fleet-key" {
+			t.Fatalf("authorization=%q", r.Header.Get("Authorization"))
+		}
+		w.WriteHeader(http.StatusOK)
+	}))
+	defer server.Close()
+	handler := fleetReadinessHandler(server.Client(), capability.FleetTierFast, []fleetEndpoint{{
+		ProviderID: "cpu", ModelID: "cpu-model", Tier: capability.FleetTierFast, URL: server.URL,
+		Headers: map[string]string{"Authorization": "Bearer fleet-key"},
+	}})
+	recorder := httptest.NewRecorder()
+	handler.ServeHTTP(recorder, httptest.NewRequest(http.MethodGet, "/v1/fleet/readiness", nil))
+	if recorder.Code != http.StatusOK || !strings.Contains(recorder.Body.String(), `"ready":true`) {
 		t.Fatalf("status=%d body=%s", recorder.Code, recorder.Body.String())
 	}
 }

cmd/cognitive-worker/gateway.go

@@ -26,6 +26,10 @@ type bindingResolver interface {
 	RegisterDiscovery(context.Context, capability.CognitiveDiscoverySnapshot) (capability.CognitiveDiscoveryRegistration, error)
 }
 
+type candidateBindingResolver interface {
+	ResolveCandidate(context.Context, capability.CapabilityBindingID, int64) (capability.ResolvedCognitiveBinding, error)
+}
+
 type admissionClient interface {
 	Admit(context.Context, string, capability.AdmissionRequest) (capability.AdmissionDecision, capability.AdmissionLease, error)
 	Release(context.Context, capability.AdmissionLease) error
@@ -294,12 +298,21 @@ func fallbackEligible(reason cognitive.FinishReason) bool {
 }
 
 func (g cognitiveGateway) provider(selected capability.ResolvedCognitiveBinding) advisoryProvider {
+	var byStableID advisoryProvider
 	for _, registration := range g.providers[selected.Provider.Kind] {
 		if registration.EndpointRef == selected.Provider.EndpointRef && registration.CredentialRef == selected.Provider.CredentialRef {
 			return registration.Provider
 		}
+		// Capability revisions can legitimately retain an older credential
+		// reference while the protected runtime has rotated or reloaded the
+		// current credential. ProviderInstanceID is the stable owner identity;
+		// use it after exact metadata matching so stale non-secret metadata does
+		// not force an unnecessary builtin fallback.
+		if byStableID == nil && registration.ProviderInstanceID == selected.Provider.ID && registration.Kind == selected.Provider.Kind {
+			byStableID = registration.Provider
+		}
 	}
-	return nil
+	return byStableID
 }
 
 func writeGatewayResponse(w http.ResponseWriter, response cognitive.GatewayAdvisoryResponse) {
@@ -341,6 +354,26 @@ func (r *httpBindingResolver) Resolve(ctx context.Context, value capability.Cogn
 	return result, err
 }
 
+func (r *httpBindingResolver) ResolveCandidate(ctx context.Context, bindingID capability.CapabilityBindingID, revision int64) (capability.ResolvedCognitiveBinding, error) {
+	var result capability.ResolvedCognitiveBinding
+	target := *r.base
+	target.Path = "/v1/cognitive-bindings/" + url.PathEscape(string(bindingID)) + "/revisions/" + fmt.Sprint(revision) + "/snapshot"
+	request, err := http.NewRequestWithContext(ctx, http.MethodGet, target.String(), nil)
+	if err != nil {
+		return result, err
+	}
+	response, err := r.client.Do(request)
+	if err != nil {
+		return result, err
+	}
+	defer response.Body.Close()
+	if response.StatusCode < 200 || response.StatusCode >= 300 {
+		return result, fmt.Errorf("candidate status %d", response.StatusCode)
+	}
+	err = json.NewDecoder(io.LimitReader(response.Body, 64<<10)).Decode(&result)
+	return result, err
+}
+
 func (r *httpBindingResolver) Observe(ctx context.Context, value capability.CognitiveUsageObservation) error {
 	return r.post(ctx, "/v1/cognitive-usage-observations", value, nil)
 }

cmd/cognitive-worker/gateway_registry_test.go

@@ -88,6 +88,19 @@ func TestActiveRegistryBindingReachesProviderAdapter(t *testing.T) {
 	}
 }
 
+func TestProviderUsesStableInstanceWhenCredentialMetadataIsStale(t *testing.T) {
+	provider := &stubAdvisoryProvider{}
+	gateway := cognitiveGateway{providers: map[capability.ProviderKind][]providerRegistration{
+		"provider-kind-1": {{ProviderInstanceID: "provider-1", Kind: "provider-kind-1", EndpointRef: "current-endpoint", CredentialRef: "current-credential", Provider: provider}},
+	}}
+	selected := gatewaySelection()
+	selected.Provider.EndpointRef = "old-endpoint"
+	selected.Provider.CredentialRef = ""
+	if got := gateway.provider(selected); got != provider {
+		t.Fatalf("stable provider identity was not used for stale metadata")
+	}
+}
+
 func TestCanaryFlagIsForwardedToCapabilityResolver(t *testing.T) {
 	registry := &registryStub{err: errors.New("stop after resolve")}
 	gateway := cognitiveGateway{resolver: registry, providers: map[capability.ProviderKind][]providerRegistration{}, now: time.Now, allowCanary: true}

cmd/cognitive-worker/main.go

@@ -17,6 +17,7 @@ import (
 	"log"
 	"net/http"
 	"os"
+	"path/filepath"
 	"strings"
 	"time"
 )
@@ -45,7 +46,7 @@ func main() {
 	discoveries := map[string]providerRegistration{}
 	var builtinLocal *providerRegistration
 	externalRecovery := []providerRegistration{}
-	cpuAPIKey, err := loadLocalCPUProviderAPIKey()
+	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)
 	}
@@ -62,6 +63,28 @@ func main() {
 	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 {
@@ -70,14 +93,20 @@ func main() {
 		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)
+	gpuURL, gpuKey, gpuConfigured, gpuErr := loadProtectedGPUConfiguration()
+	if gpuErr != nil {
+		log.Printf("optional GPU provider disabled: %v", gpuErr)
+	} else if gpuConfigured {
+		gpuAdapter, adapterErr := models.NewOpenAICompatibleLocalAdapterWithAPIKey(gpuURL, gpuKey, nil)
+		if adapterErr != nil {
+			log.Printf("optional GPU provider disabled: %v", adapterErr)
+		} else {
+			gpu := providerRegistration{ProviderInstanceID: "local-gpu-opportunistic", Kind: "openai-compatible-chat-v1", EndpointRef: "local-gpu-opportunistic-v1", CredentialRef: "credential://domovoy/gpu/provider-api-key/v1", AuthMode: capability.AuthKey, 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)
+		}
 	}
-	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})
@@ -85,27 +114,63 @@ func main() {
 	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"
-	mux.Handle("/readyz", health.ReadinessHandler{Runtime: "cognitive-worker", Probes: []health.Probe{
-		health.HTTPProbe(readinessClient, strings.TrimRight(envOr("CORE_RUNTIME_URL", "http://core-runtime:8080"), "/")+"/healthz"),
-		health.OptionalHTTPProbe(readinessClient, strings.TrimRight(envOr("LOCAL_CPU_PROVIDER_URL", "http://docker.cin.su:18082/v1"), "/")+"/models"),
-		health.HTTPProbe(readinessClient, lifeboatURL),
-		health.OptionalHTTPProbe(readinessClient, strings.TrimRight(envOr("LOCAL_GPU_PROVIDER_URL", "http://192.168.220.91:11434"), "/")+"/api/tags"),
-	}})
 	fleetClient := &http.Client{Timeout: 2 * time.Second}
-	preferredFleet := capability.FleetTier(envOr("COGNITIVE_PREFERRED_FLEET_TIER", string(capability.FleetTierHeavy)))
-	mux.Handle("/v1/fleet/readiness", fleetReadinessHandler(fleetClient, preferredFleet, []fleetEndpoint{
-		{ProviderID: "local-cpu-primary", ModelID: "local-cpu-model", Tier: capability.FleetTierFast, URL: strings.TrimRight(envOr("LOCAL_CPU_PROVIDER_URL", "http://docker.cin.su:18082/v1"), "/") + "/models"},
-		{ProviderID: "local-medium-unconfigured", ModelID: "medium-unconfigured", Tier: capability.FleetTierMedium},
-		{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"},
+	preferredFleet := capability.FleetTier(envOr("COGNITIVE_PREFERRED_FLEET_TIER", string(capability.FleetTierFast)))
+	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))
+	if gpuConfigured {
+		gpuHeaders := map[string]string{"Authorization": "Bearer " + gpuKey}
+		gpuProbe := health.OptionalHTTPProbeWithHeaders(readinessClient, strings.TrimRight(gpuURL, "/")+"/models", gpuHeaders)
+		gpuProbe.Timeout = 10 * time.Second
+		readinessProbes = append(readinessProbes, gpuProbe)
+	}
+
+	fleetEndpoints := []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: "cognitive-lifeboat", ModelID: "qwen3-8b-lifeboat", Tier: capability.FleetTierLifeboat, URL: lifeboatURL},
-	}))
+	}
+	if gpuConfigured {
+		fleetEndpoints = append(fleetEndpoints, fleetEndpoint{ProviderID: "local-gpu-opportunistic", ModelID: "gpu-opportunistic", Tier: capability.FleetTierHeavy, URL: strings.TrimRight(gpuURL, "/") + "/models", Headers: map[string]string{"Authorization": "Bearer " + gpuKey}})
+	}
+	mux.Handle("/readyz", health.ReadinessHandler{Runtime: "cognitive-worker", Probes: readinessProbes})
+	mux.Handle("/v1/fleet/readiness", fleetReadinessHandler(fleetClient, preferredFleet, fleetEndpoints))
 	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)
+	mux.HandleFunc("POST /v1/task-evaluations/execute", gateway.taskEvaluationHandler)
 	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)
 	refreshProviderDiscoveries(resolver, discovery)
 	log.Printf("cognitive-worker listening on %s", addr)
 	if err := http.ListenAndServe(addr, mux); err != nil {
@@ -113,6 +178,19 @@ func main() {
 	}
 }
 
+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)
@@ -157,17 +235,20 @@ func refreshProviderDiscoveries(resolver bindingResolver, service discoveryServi
 	}
 }
 
-func loadLocalCPUProviderAPIKey() (string, error) {
-	if _, set := os.LookupEnv("LOCAL_CPU_PROVIDER_API_KEY"); set {
-		return "", fmt.Errorf("LOCAL_CPU_PROVIDER_API_KEY direct environment secret is forbidden; use LOCAL_CPU_PROVIDER_API_KEY_FILE")
+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("LOCAL_CPU_PROVIDER_API_KEY_FILE"))
+	fileName := strings.TrimSpace(os.Getenv(fileEnvName))
 	if fileName == "" {
 		return "", nil
 	}
-	ref := strings.TrimSpace(os.Getenv("LOCAL_CPU_PROVIDER_CREDENTIAL_REF"))
+	ref := strings.TrimSpace(os.Getenv(refEnvName))
 	if ref == "" {
-		return "", fmt.Errorf("LOCAL_CPU_PROVIDER_CREDENTIAL_REF is required when LOCAL_CPU_PROVIDER_API_KEY_FILE is configured")
+		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 {
@@ -176,6 +257,50 @@ func loadLocalCPUProviderAPIKey() (string, error) {
 	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")
+}
+
+const (
+	gpuProviderURLFile = "/run/secrets/llm_gpu_gateway_url"
+	gpuProviderKeyFile = "/run/secrets/llm_gpu_gateway_api_key"
+)
+
+func loadProtectedGPUConfiguration() (string, string, bool, error) {
+	urlValue, err := readProtectedRuntimeValue(gpuProviderURLFile)
+	if errors.Is(err, os.ErrNotExist) {
+		return "", "", false, nil
+	}
+	if err != nil {
+		return "", "", false, err
+	}
+	key, err := readProtectedRuntimeValue(gpuProviderKeyFile)
+	if errors.Is(err, os.ErrNotExist) {
+		return "", "", false, fmt.Errorf("GPU provider key is not configured")
+	}
+	if err != nil {
+		return "", "", false, err
+	}
+	return urlValue, key, true, nil
+}
+
+func readProtectedRuntimeValue(path string) (string, error) {
+	const root = "/run/secrets"
+	clean := filepath.Clean(path)
+	if clean != root && !strings.HasPrefix(clean, root+string(os.PathSeparator)) {
+		return "", fmt.Errorf("protected runtime path outside secret root")
+	}
+	payload, err := os.ReadFile(clean)
+	if err != nil {
+		return "", err
+	}
+	value := strings.TrimSpace(string(payload))
+	if value == "" || strings.ContainsAny(value, "\r\n") {
+		return "", fmt.Errorf("protected runtime value is empty or multiline")
+	}
+	return value, nil
+}
+
 func deterministicAdvisory(w http.ResponseWriter, r *http.Request) {
 	request, ok := decodeAdvisoryRequest(w, r)
 	if !ok {

cmd/cognitive-worker/main_test.go

@@ -43,3 +43,34 @@ func TestLoadLocalCPUProviderAPIKeyRejectsDirectEnvironmentSecret(t *testing.T)
 		t.Fatal("expected direct environment secret to be rejected")
 	}
 }
+
+func TestLoadProviderAPIKeySupportsMediumCredentialRef(t *testing.T) {
+	t.Setenv("LOCAL_MEDIUM_PROVIDER_API_KEY", "")
+	_ = os.Unsetenv("LOCAL_MEDIUM_PROVIDER_API_KEY")
+	root := t.TempDir()
+	path := filepath.Join(root, "medium_provider_api_key")
+	if err := os.WriteFile(path, []byte("medium-key\n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	t.Setenv("LOCAL_MEDIUM_PROVIDER_API_KEY_FILE", path)
+	t.Setenv("LOCAL_MEDIUM_PROVIDER_CREDENTIAL_REF", "credential://domovoy/local-medium/provider-api-key/v1")
+	t.Setenv("SECRET_BACKEND_ROOT", root)
+	key, 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 || key != "medium-key" {
+		t.Fatalf("unexpected medium key resolution result: err=%v", err)
+	}
+}
+
+func TestLoadProviderAPIKeyRejectsMediumDirectEnvironmentSecret(t *testing.T) {
+	t.Setenv("LOCAL_MEDIUM_PROVIDER_API_KEY", "must-not-be-used")
+	t.Setenv("LOCAL_MEDIUM_PROVIDER_API_KEY_FILE", "")
+	if _, 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"); err == nil {
+		t.Fatal("expected direct medium environment secret to be rejected")
+	}
+}
+
+func TestReadProtectedRuntimeValueRejectsPathOutsideSecretRoot(t *testing.T) {
+	if _, err := readProtectedRuntimeValue("/tmp/not-a-domovoy-secret"); err == nil {
+		t.Fatal("expected protected path rejection")
+	}
+}

cmd/cognitive-worker/model_qualification.go

@@ -0,0 +1,167 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"slices"
+	"sync"
+	"time"
+
+	"git.cin.su/m/Assistant/adapters/models"
+	"git.cin.su/m/Assistant/contracts/authority"
+	"git.cin.su/m/Assistant/contracts/capability"
+	"git.cin.su/m/Assistant/contracts/cognitive"
+	"git.cin.su/m/Assistant/contracts/identity"
+	"git.cin.su/m/Assistant/platform/ids"
+)
+
+type qualificationObserver interface {
+	ObserveQualification(context.Context, capability.ModelQualificationObservation) error
+}
+
+type httpQualificationObserver struct{ resolver *httpBindingResolver }
+
+func (o httpQualificationObserver) ObserveQualification(ctx context.Context, value capability.ModelQualificationObservation) error {
+	if o.resolver == nil {
+		return fmt.Errorf("qualification observer resolver is nil")
+	}
+	return o.resolver.post(ctx, "/v1/cognitive-registry/model-qualification-observations", value, nil)
+}
+
+type modelQualificationRequest struct {
+	Prompt         string                       `json:"prompt"`
+	MaxOutputBytes int                          `json:"max_output_bytes"`
+	ModelIDs       []capability.ProviderModelID `json:"model_ids,omitempty"`
+}
+
+type modelQualificationResult struct {
+	ProviderModelID capability.ProviderModelID     `json:"provider_model_id"`
+	Status          string                         `json:"status"`
+	FinishReason    cognitive.FinishReason         `json:"finish_reason,omitempty"`
+	LatencyMillis   int64                          `json:"latency_millis"`
+	OutputObserved  bool                           `json:"output_observed"`
+	QualityAssessed bool                           `json:"quality_assessed"`
+	Category        capability.ModelRatingCategory `json:"category"`
+	Error           string                         `json:"error,omitempty"`
+}
+
+type modelQualificationResponse struct {
+	Provider             capability.ProviderInstanceID `json:"provider"`
+	ObservedAtUnixMillis int64                         `json:"observed_at_unix_millis"`
+	Results              []modelQualificationResult    `json:"results"`
+}
+
+func modelQualificationHandler(providers map[string]providerRegistration, owner string, observers ...qualificationObserver) http.Handler {
+	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		name := r.PathValue("provider")
+		registration, ok := providers[name]
+		if !ok || registration.Discovery == nil || registration.Provider == nil {
+			http.Error(w, "provider qualification not found", http.StatusNotFound)
+			return
+		}
+		var request modelQualificationRequest
+		if !decodeJSON(w, r, 16<<10, &request) || request.Prompt == "" {
+			return
+		}
+		if request.MaxOutputBytes < 1 {
+			request.MaxOutputBytes = 512
+		}
+		if request.MaxOutputBytes > 4096 {
+			http.Error(w, "max_output_bytes exceeds qualification bound", http.StatusBadRequest)
+			return
+		}
+		modelsFound, probe := registration.Discovery.Discover(r.Context())
+		if probe.Health != capability.HealthAvailable || len(modelsFound) == 0 {
+			http.Error(w, "provider unavailable for qualification", http.StatusServiceUnavailable)
+			return
+		}
+		if len(request.ModelIDs) > 0 {
+			modelsFound = slices.DeleteFunc(modelsFound, func(model models.DiscoveredModel) bool {
+				return !slices.Contains(request.ModelIDs, model.ProviderModelID)
+			})
+		}
+		if len(modelsFound) == 0 {
+			http.Error(w, "requested model not discovered", http.StatusNotFound)
+			return
+		}
+		if len(modelsFound) > 16 {
+			modelsFound = modelsFound[:16]
+		}
+		results := qualifyModels(r.Context(), registration, owner, request.Prompt, request.MaxOutputBytes, modelsFound)
+		if len(observers) > 0 && observers[0] != nil {
+			if err := persistQualificationResults(r.Context(), observers[0], registration, results); err != nil {
+				http.Error(w, "persist model qualification observation", http.StatusBadGateway)
+				return
+			}
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_ = jsonEncode(w, modelQualificationResponse{Provider: registration.ProviderInstanceID, ObservedAtUnixMillis: time.Now().UnixMilli(), Results: results})
+	})
+}
+
+func qualifyModels(ctx context.Context, registration providerRegistration, owner, prompt string, maxOutputBytes int, discovered []models.DiscoveredModel) []modelQualificationResult {
+	results := make([]modelQualificationResult, len(discovered))
+	var wg sync.WaitGroup
+	for index, model := range discovered {
+		wg.Add(1)
+		go func(index int, model models.DiscoveredModel) {
+			defer wg.Done()
+			started := time.Now()
+			request := cognitive.AdvisoryRequest{Version: cognitive.CreatorAdvisoryV1, RequestID: cognitive.RequestID(fmt.Sprintf("qualification:%s:%d", model.ProviderModelID, started.UnixNano())), ActorID: identity.ActorID("qualification-runtime"), PrincipalID: identity.PrincipalID(owner), ScopeID: authority.ScopeID("qualification-scope"), Message: prompt, Locale: "en-US", MaxOutputBytes: maxOutputBytes, ProviderBindingID: "qualification-binding-v1", ProviderBindingRevision: 1, ProviderInstanceID: registration.ProviderInstanceID, ModelDescriptorID: capability.ModelDescriptorID("qualification-" + string(model.ProviderModelID)), ProviderModelID: model.ProviderModelID}
+			response, err := registration.Provider.Advise(ctx, request)
+			result := modelQualificationResult{ProviderModelID: model.ProviderModelID, LatencyMillis: time.Since(started).Milliseconds(), QualityAssessed: false, Category: capability.ModelRatingUnrated}
+			if err != nil {
+				result.Status = "UNAVAILABLE"
+				result.FinishReason = cognitive.FinishProviderUnavailable
+				result.Error = err.Error()
+				results[index] = result
+				return
+			}
+			result.FinishReason = response.FinishReason
+			result.OutputObserved = response.AdvisoryText != ""
+			if response.FinishReason == cognitive.FinishCompleted && result.OutputObserved {
+				result.Status = "REACHABLE"
+			} else {
+				result.Status = "RESPONDED_WITHOUT_QUALIFICATION"
+			}
+			results[index] = result
+		}(index, model)
+	}
+	wg.Wait()
+	return results
+}
+
+func persistQualificationResults(ctx context.Context, observer qualificationObserver, registration providerRegistration, results []modelQualificationResult) error {
+	for _, result := range results {
+		observationID, err := ids.New()
+		if err != nil {
+			return err
+		}
+		status := capability.ModelQualificationRespondedWithoutRating
+		switch result.Status {
+		case "REACHABLE":
+			status = capability.ModelQualificationReachable
+		case "UNAVAILABLE":
+			status = capability.ModelQualificationUnavailable
+		}
+		observation := capability.ModelQualificationObservation{
+			ID: capability.ModelQualificationObservationID(observationID), ProviderInstanceID: registration.ProviderInstanceID, ProviderModelID: result.ProviderModelID,
+			Status: status, FinishReason: string(result.FinishReason), LatencyMillis: result.LatencyMillis,
+			OutputObserved: result.OutputObserved, QualityAssessed: false, Category: capability.ModelRatingUnrated,
+			ObservedAtUnixMillis: time.Now().UnixMilli(),
+		}
+		if err := observation.Validate(); err != nil {
+			return err
+		}
+		if err := observer.ObserveQualification(ctx, observation); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func jsonEncode(w http.ResponseWriter, value any) error {
+	return json.NewEncoder(w).Encode(value)
+}

cmd/cognitive-worker/model_qualification_test.go

@@ -0,0 +1,85 @@
+package main
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"git.cin.su/m/Assistant/adapters/models"
+	"git.cin.su/m/Assistant/contracts/capability"
+	"git.cin.su/m/Assistant/contracts/cognitive"
+)
+
+type qualificationProvider struct{}
+
+func (qualificationProvider) Probe(context.Context) models.ProviderProbe {
+	return models.ProviderProbe{Health: capability.HealthAvailable}
+}
+func (qualificationProvider) Discover(context.Context) ([]models.DiscoveredModel, models.ProviderProbe) {
+	return []models.DiscoveredModel{{ProviderModelID: "model-a"}, {ProviderModelID: "model-b"}}, models.ProviderProbe{Health: capability.HealthAvailable}
+}
+func (qualificationProvider) Advise(_ context.Context, request cognitive.AdvisoryRequest) (cognitive.AdvisoryResponse, error) {
+	if err := models.ValidateAdvisoryRequest(request); err != nil {
+		return cognitive.AdvisoryResponse{}, err
+	}
+	return cognitive.AdvisoryResponse{Version: request.Version, RequestID: request.RequestID, AdvisoryText: "reachable", FinishReason: cognitive.FinishCompleted}, nil
+}
+
+func TestModelQualificationUsesDiscoveredModelsAndLeavesQualityUnrated(t *testing.T) {
+	provider := qualificationProvider{}
+	mux := http.NewServeMux()
+	mux.Handle("POST /v1/provider-discovery/{provider}/qualify", modelQualificationHandler(map[string]providerRegistration{"gpu": {ProviderInstanceID: "gpu", Provider: provider, Discovery: provider}}, "owner"))
+	server := httptest.NewServer(mux)
+	defer server.Close()
+	request := httptest.NewRequest(http.MethodPost, server.URL+"/v1/provider-discovery/gpu/qualify", strings.NewReader(`{"prompt":"ready","max_output_bytes":128}`))
+	request.Header.Set("Content-Type", "application/json")
+	response := httptest.NewRecorder()
+	mux.ServeHTTP(response, request)
+	if response.Code != http.StatusOK {
+		t.Fatalf("qualification status=%d body=%s", response.Code, response.Body.String())
+	}
+	if !strings.Contains(response.Body.String(), `"status":"REACHABLE"`) || !strings.Contains(response.Body.String(), `"category":"UNRATED"`) || strings.Contains(response.Body.String(), `"quality_assessed":true`) {
+		t.Fatalf("qualification overclaimed quality: %s", response.Body.String())
+	}
+}
+
+func TestModelQualificationRejectsUnboundedOutput(t *testing.T) {
+	provider := qualificationProvider{}
+	mux := http.NewServeMux()
+	mux.Handle("POST /v1/provider-discovery/{provider}/qualify", modelQualificationHandler(map[string]providerRegistration{"cpu": {ProviderInstanceID: "cpu", Provider: provider, Discovery: provider}}, "owner"))
+	request := httptest.NewRequest(http.MethodPost, "/v1/provider-discovery/cpu/qualify", strings.NewReader(`{"prompt":"ready","max_output_bytes":8192}`))
+	request.Header.Set("Content-Type", "application/json")
+	response := httptest.NewRecorder()
+	mux.ServeHTTP(response, request)
+	if response.Code != http.StatusBadRequest {
+		t.Fatalf("unbounded qualification status=%d", response.Code)
+	}
+}
+
+type qualificationObserverStub struct {
+	observations []capability.ModelQualificationObservation
+}
+
+func (s *qualificationObserverStub) ObserveQualification(_ context.Context, value capability.ModelQualificationObservation) error {
+	s.observations = append(s.observations, value)
+	return nil
+}
+
+func TestModelQualificationPersistsReachabilityWithoutQuality(t *testing.T) {
+	provider := qualificationProvider{}
+	observer := &qualificationObserverStub{}
+	mux := http.NewServeMux()
+	mux.Handle("POST /v1/provider-discovery/{provider}/qualify", modelQualificationHandler(map[string]providerRegistration{"cpu": {ProviderInstanceID: "cpu", Provider: provider, Discovery: provider}}, "owner", observer))
+	request := httptest.NewRequest(http.MethodPost, "/v1/provider-discovery/cpu/qualify", strings.NewReader(`{"prompt":"ready","max_output_bytes":128,"model_ids":["model-a"]}`))
+	request.Header.Set("Content-Type", "application/json")
+	response := httptest.NewRecorder()
+	mux.ServeHTTP(response, request)
+	if response.Code != http.StatusOK || len(observer.observations) != 1 {
+		t.Fatalf("qualification persistence status=%d observations=%d body=%s", response.Code, len(observer.observations), response.Body.String())
+	}
+	if err := observer.observations[0].Validate(); err != nil || observer.observations[0].Status != capability.ModelQualificationReachable {
+		t.Fatalf("invalid qualification observation: %+v err=%v", observer.observations[0], err)
+	}
+}

cmd/cognitive-worker/task_evaluation.go

@@ -0,0 +1,96 @@
+package main
+
+import (
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"sync"
+	"time"
+
+	"git.cin.su/m/Assistant/contracts/capability"
+	"git.cin.su/m/Assistant/contracts/cognitive"
+)
+
+func (g cognitiveGateway) taskEvaluationHandler(w http.ResponseWriter, r *http.Request) {
+	var request capability.TaskEvaluationExecutionRequest
+	if !decodeJSON(w, r, 256<<10, &request) {
+		return
+	}
+	if err := request.Profile.Validate(); err != nil || request.CapabilityID == "" || request.CapabilityID != request.Profile.CapabilityID || request.BindingID == "" || request.BindingRevision < 1 || request.ActorID == "" || request.PrincipalID == "" || request.ScopeID == "" || request.MaxOutputBytes < 1 || len(request.Cases) != len(request.Profile.RequiredCaseIDs) {
+		http.Error(w, "invalid task evaluation execution request", http.StatusBadRequest)
+		return
+	}
+	resolver, ok := g.resolver.(candidateBindingResolver)
+	if !ok {
+		http.Error(w, "exact candidate resolver unavailable", http.StatusServiceUnavailable)
+		return
+	}
+	selected, err := resolver.ResolveCandidate(r.Context(), request.BindingID, request.BindingRevision)
+	if err != nil || selected.Binding.ID != request.BindingID || selected.Binding.Revision != request.BindingRevision || selected.Binding.CapabilityID != request.CapabilityID || selected.Binding.OwnerPrincipalID != request.PrincipalID {
+		http.Error(w, "cognitive binding snapshot unavailable", http.StatusConflict)
+		return
+	}
+	if selected.Binding.DataPolicy != capability.DataPolicyLocalOnly && !request.EgressAcknowledged {
+		http.Error(w, "external egress acknowledgement required", http.StatusBadRequest)
+		return
+	}
+	provider := g.provider(selected)
+	if provider == nil {
+		http.Error(w, "cognitive provider unavailable", http.StatusServiceUnavailable)
+		return
+	}
+	caseSet := make(map[string]struct{}, len(request.Profile.RequiredCaseIDs))
+	for _, id := range request.Profile.RequiredCaseIDs {
+		caseSet[id] = struct{}{}
+	}
+	observations := make([]capability.TaskCaseObservation, len(request.Cases))
+	responses := make([]capability.TaskCaseResponse, len(request.Cases))
+	var wg sync.WaitGroup
+	seenCases := make(map[string]struct{}, len(request.Cases))
+	for i, item := range request.Cases {
+		if item.CaseID == "" || item.Prompt == "" || item.EvidenceArtifactRef == "" {
+			http.Error(w, "invalid task evaluation case", http.StatusBadRequest)
+			return
+		}
+		if _, exists := caseSet[item.CaseID]; !exists {
+			http.Error(w, "task evaluation case is not in profile", http.StatusBadRequest)
+			return
+		}
+		if _, exists := seenCases[item.CaseID]; exists {
+			http.Error(w, "duplicate task evaluation case", http.StatusBadRequest)
+			return
+		}
+		seenCases[item.CaseID] = struct{}{}
+		wg.Add(1)
+		go func(index int, item capability.TaskEvaluationCase) {
+			defer wg.Done()
+			started := time.Now()
+			invocation := cognitive.AdvisoryRequest{Version: cognitive.CreatorAdvisoryV1, RequestID: cognitive.RequestID(fmt.Sprintf("task-evaluation:%s:%s", request.Profile.ID, item.CaseID)), ActorID: request.ActorID, PrincipalID: request.PrincipalID, ScopeID: request.ScopeID, Message: item.Prompt, MaxOutputBytes: int(request.MaxOutputBytes), ProviderBindingID: selected.Binding.ID, ProviderBindingRevision: selected.Binding.Revision, ProviderInstanceID: selected.Provider.ID, ModelDescriptorID: selected.Model.ID, ProviderModelID: selected.Model.ProviderModelID}
+			response, invokeErr := g.invokeAdmitted(r.Context(), string(invocation.RequestID), provider, invocation, int(request.MaxOutputBytes))
+			observation := capability.TaskCaseObservation{CaseID: item.CaseID, LatencyMicros: time.Since(started).Microseconds(), EvidenceArtifactRef: item.EvidenceArtifactRef}
+			responses[index] = capability.TaskCaseResponse{CaseID: item.CaseID, ResponseText: response.AdvisoryText, LatencyMicros: observation.LatencyMicros, EvidenceArtifactRef: item.EvidenceArtifactRef}
+			if invokeErr == nil && response.FinishReason == cognitive.FinishCompleted && response.AdvisoryText != "" {
+				// Invocation success proves reachability only. Without an
+				// independent verifier and answer evidence, semantic support and
+				// quality must remain UNKNOWN rather than becoming false evidence.
+				observation.Result = capability.EvaluationUnknown
+				observation.QualityScoreMillis = 0
+			} else {
+				observation.Result = capability.EvaluationUnknown
+				observation.CriticalFailure = invokeErr != nil
+			}
+			observations[index] = observation
+		}(i, item)
+	}
+	if len(seenCases) != len(caseSet) {
+		http.Error(w, "task evaluation case set is incomplete", http.StatusBadRequest)
+		return
+	}
+	wg.Wait()
+	writeTaskEvaluationJSON(w, capability.TaskEvaluationExecutionResponse{ProfileID: request.Profile.ID, BindingID: selected.Binding.ID, BindingRevision: selected.Binding.Revision, Observations: observations, Responses: responses, ObservedAtUnixMillis: time.Now().UnixMilli()})
+}
+
+func writeTaskEvaluationJSON(w http.ResponseWriter, value any) {
+	w.Header().Set("Content-Type", "application/json")
+	_ = json.NewEncoder(w).Encode(value)
+}

cmd/cognitive-worker/task_evaluation_test.go

@@ -0,0 +1,58 @@
+package main
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+	"time"
+
+	"git.cin.su/m/Assistant/contracts/authority"
+	"git.cin.su/m/Assistant/contracts/capability"
+	"git.cin.su/m/Assistant/contracts/cognitive"
+	"git.cin.su/m/Assistant/contracts/identity"
+)
+
+type candidateRegistryStub struct {
+	registryStub
+	selected capability.ResolvedCognitiveBinding
+}
+
+func (s *candidateRegistryStub) ResolveCandidate(context.Context, capability.CapabilityBindingID, int64) (capability.ResolvedCognitiveBinding, error) {
+	return s.selected, nil
+}
+
+func TestTaskEvaluationHandlerInvokesEveryBoundedCase(t *testing.T) {
+	selected := gatewaySelection()
+	registry := &candidateRegistryStub{selected: selected}
+	provider := &stubAdvisoryProvider{response: cognitive.AdvisoryResponse{Version: cognitive.CreatorAdvisoryV1, FinishReason: cognitive.FinishCompleted, AdvisoryText: "observed response"}}
+	gateway := cognitiveGateway{
+		resolver: registry,
+		providers: map[capability.ProviderKind][]providerRegistration{
+			selected.Provider.Kind: {{ProviderInstanceID: selected.Provider.ID, Kind: selected.Provider.Kind, EndpointRef: selected.Provider.EndpointRef, CredentialRef: selected.Provider.CredentialRef, Provider: provider}},
+		},
+		now: func() time.Time { return time.UnixMilli(1000) },
+	}
+	profile := capability.TaskEvaluationProfile{ID: "profile-general", Revision: 1, CapabilityID: selected.Binding.CapabilityID, Stage: capability.TaskEvaluationGeneral, BenchmarkVersion: "general-v1", RequiredCaseIDs: []string{"case-a", "case-b"}, QualityWeightMillis: 600, ReliabilityWeightMillis: 300, LatencyWeightMillis: 100}
+	request := capability.TaskEvaluationExecutionRequest{Profile: profile, CapabilityID: selected.Binding.CapabilityID, BindingID: selected.Binding.ID, BindingRevision: selected.Binding.Revision, ActorID: identity.ActorID("actor-1"), PrincipalID: selected.Binding.OwnerPrincipalID, ScopeID: authority.ScopeID("scope-1"), MaxOutputBytes: 512, EgressAcknowledged: true, Cases: []capability.TaskEvaluationCase{{CaseID: "case-a", Prompt: "one", EvidenceArtifactRef: "artifact:a"}, {CaseID: "case-b", Prompt: "two", EvidenceArtifactRef: "artifact:b"}}}
+	payload, _ := json.Marshal(request)
+	recorder := httptest.NewRecorder()
+	gateway.taskEvaluationHandler(recorder, httptest.NewRequest(http.MethodPost, "/v1/task-evaluations/execute", bytes.NewReader(payload)))
+	if recorder.Code != http.StatusOK {
+		t.Fatalf("status=%d body=%s", recorder.Code, recorder.Body.String())
+	}
+	var response capability.TaskEvaluationExecutionResponse
+	if err := json.Unmarshal(recorder.Body.Bytes(), &response); err != nil {
+		t.Fatal(err)
+	}
+	if len(response.Observations) != 2 || response.Observations[0].Result != capability.EvaluationUnknown || response.Observations[1].Result != capability.EvaluationUnknown {
+		t.Fatalf("unexpected observations: %+v", response.Observations)
+	}
+	for _, observation := range response.Observations {
+		if observation.QualityScoreMillis != 0 || observation.CriticalFailure {
+			t.Fatalf("reachability-only evidence must not carry quality or failure claims: %+v", observation)
+		}
+	}
+}

cmd/core-runtime/development_gate.go

@@ -0,0 +1,32 @@
+package main
+
+import (
+	"errors"
+	"time"
+
+	"git.cin.su/m/Assistant/contracts/work"
+	workdomain "git.cin.su/m/Assistant/modules/work/domain"
+)
+
+var errDevelopmentGateCreate = errors.New("development gate must start at evidence collected with revision one")
+
+func prepareDevelopmentGate(value work.DevelopmentGate, now time.Time) (work.DevelopmentGate, error) {
+	if value.Stage != work.DevelopmentEvidenceCollected || value.Revision != 1 {
+		return work.DevelopmentGate{}, errDevelopmentGateCreate
+	}
+	value.UpdatedAtUnixMillis = now.UnixMilli()
+	if err := workdomain.ValidateDevelopmentGate(value); err != nil {
+		return work.DevelopmentGate{}, err
+	}
+	return value, nil
+}
+
+func advanceDevelopmentGate(value work.DevelopmentGate, next work.DevelopmentStage, now time.Time) (work.DevelopmentGate, error) {
+	advanced, err := workdomain.AdvanceDevelopmentGate(value, next)
+	if err != nil {
+		return work.DevelopmentGate{}, err
+	}
+	advanced.Revision = value.Revision + 1
+	advanced.UpdatedAtUnixMillis = now.UnixMilli()
+	return advanced, nil
+}

cmd/core-runtime/development_gate_test.go

@@ -0,0 +1,41 @@
+package main
+
+import (
+	"testing"
+	"time"
+
+	"git.cin.su/m/Assistant/contracts/work"
+)
+
+func TestDevelopmentGateControllerRequiresSequentialVerifiedEvidence(t *testing.T) {
+	now := time.UnixMilli(10)
+	gate, err := prepareDevelopmentGate(work.DevelopmentGate{ID: "gate-1", Revision: 1, Stage: work.DevelopmentEvidenceCollected, EvidenceRefs: []string{"evidence-1"}}, now)
+	if err != nil {
+		t.Fatal(err)
+	}
+	gate.ProposalRef = "proposal-1"
+	gate, err = advanceDevelopmentGate(gate, work.DevelopmentProposalCreated, now)
+	if err != nil || gate.Revision != 2 {
+		t.Fatalf("proposal transition failed: %#v %v", gate, err)
+	}
+	gate.TaskPacketRef = "task-1"
+	gate, err = advanceDevelopmentGate(gate, work.DevelopmentTaskAdmitted, now)
+	if err != nil {
+		t.Fatal(err)
+	}
+	gate.WorktreeRef = "isolated-worktree"
+	gate, err = advanceDevelopmentGate(gate, work.DevelopmentIsolated, now)
+	if err != nil {
+		t.Fatal(err)
+	}
+	gate.ProducerRunRefs = []work.AgentRunID{"producer"}
+	gate.VerifierRunRefs = []work.AgentRunID{"verifier"}
+	gate.VerificationEvidence = "verification-1"
+	gate, err = advanceDevelopmentGate(gate, work.DevelopmentVerified, now)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := advanceDevelopmentGate(gate, work.DevelopmentPromoted, now); err == nil {
+		t.Fatal("promotion bypassed canary")
+	}
+}

cmd/core-runtime/main.go

@@ -116,6 +116,10 @@ func main() {
 	if err != nil {
 		log.Fatal(err)
 	}
+	researchClient, err := newResearchRuntimeClient(envOr("RESEARCH_RUNTIME_URL", "http://research-runtime:8087"), &http.Client{Timeout: 12 * time.Second})
+	if err != nil {
+		log.Fatal(err)
+	}
 	learning := learningRecorder{epistemics: epistemicsRepository, capability: capabilityRepository}
 	mux := http.NewServeMux()
 	authenticateCreator := func(w http.ResponseWriter, r *http.Request) (identity.CreatorSession, bool) {
@@ -178,6 +182,135 @@ func main() {
 		w.Header().Set("Content-Type", "application/json")
 		_ = json.NewEncoder(w).Encode(session)
 	})
+	mux.HandleFunc("POST /v1/creator/research/web", func(w http.ResponseWriter, r *http.Request) {
+		if _, ok := authenticateCreator(w, r); !ok {
+			return
+		}
+		var request epistemics.WebResearchRequest
+		decoder := json.NewDecoder(http.MaxBytesReader(w, r.Body, 128<<10))
+		decoder.DisallowUnknownFields()
+		if err := decoder.Decode(&request); err != nil {
+			http.Error(w, "invalid web research request", http.StatusBadRequest)
+			return
+		}
+		result, err := researchClient.EvaluateWeb(r.Context(), request)
+		if err != nil {
+			http.Error(w, "web research unavailable", http.StatusBadGateway)
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_ = json.NewEncoder(w).Encode(result)
+	})
+	mux.HandleFunc("POST /v1/creator/development/proposals", func(w http.ResponseWriter, r *http.Request) {
+		session, ok := authenticateCreator(w, r)
+		if !ok {
+			return
+		}
+		var request struct {
+			Evidence []work.ExtensionEvidence      `json:"evidence"`
+			Proposal work.ExtensionProposalRequest `json:"proposal"`
+		}
+		decoder := json.NewDecoder(http.MaxBytesReader(w, r.Body, 64<<10))
+		decoder.DisallowUnknownFields()
+		if err := decoder.Decode(&request); err != nil {
+			http.Error(w, "invalid development proposal", http.StatusBadRequest)
+			return
+		}
+		proposal, err := workdomain.BuildExtensionProposal(request.Evidence, request.Proposal)
+		if err != nil {
+			http.Error(w, "development proposal rejected", http.StatusUnprocessableEntity)
+			return
+		}
+		proposal.RequestedByPrincipalID = string(session.PrincipalID)
+		if err := workRepository.SaveExtensionProposal(r.Context(), proposal); err != nil {
+			http.Error(w, "development proposal conflict", http.StatusConflict)
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusCreated)
+		_ = json.NewEncoder(w).Encode(proposal)
+	})
+	mux.HandleFunc("POST /v1/creator/development/scaffolds", func(w http.ResponseWriter, r *http.Request) {
+		session, ok := authenticateCreator(w, r)
+		if !ok {
+			return
+		}
+		var request work.DevelopmentScaffoldRequest
+		decoder := json.NewDecoder(http.MaxBytesReader(w, r.Body, 64<<10))
+		decoder.DisallowUnknownFields()
+		if err := decoder.Decode(&request); err != nil {
+			http.Error(w, "invalid development scaffold", http.StatusBadRequest)
+			return
+		}
+		scaffold, err := workdomain.BuildDevelopmentScaffold(request)
+		if err != nil {
+			http.Error(w, "development scaffold rejected", http.StatusUnprocessableEntity)
+			return
+		}
+		scaffold.RequestedByPrincipalID = string(session.PrincipalID)
+		if err := workRepository.SaveDevelopmentScaffold(r.Context(), scaffold); err != nil {
+			http.Error(w, "development scaffold conflict", http.StatusConflict)
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusCreated)
+		_ = json.NewEncoder(w).Encode(scaffold)
+	})
+	mux.HandleFunc("POST /v1/creator/development/gates", func(w http.ResponseWriter, r *http.Request) {
+		if _, ok := authenticateCreator(w, r); !ok {
+			return
+		}
+		var gate work.DevelopmentGate
+		decoder := json.NewDecoder(http.MaxBytesReader(w, r.Body, 64<<10))
+		decoder.DisallowUnknownFields()
+		if err := decoder.Decode(&gate); err != nil {
+			http.Error(w, "invalid development gate", http.StatusBadRequest)
+			return
+		}
+		gate, err := prepareDevelopmentGate(gate, time.Now())
+		if err != nil {
+			http.Error(w, "development gate rejected", http.StatusUnprocessableEntity)
+			return
+		}
+		if err := workRepository.SaveDevelopmentGate(r.Context(), gate, 0); err != nil {
+			http.Error(w, "development gate conflict", http.StatusConflict)
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusCreated)
+		_ = json.NewEncoder(w).Encode(gate)
+	})
+	mux.HandleFunc("POST /v1/creator/development/gates/{id}/advance", func(w http.ResponseWriter, r *http.Request) {
+		if _, ok := authenticateCreator(w, r); !ok {
+			return
+		}
+		var request struct {
+			ExpectedRevision int64                 `json:"expected_revision"`
+			NextStage        work.DevelopmentStage `json:"next_stage"`
+		}
+		decoder := json.NewDecoder(http.MaxBytesReader(w, r.Body, 16<<10))
+		decoder.DisallowUnknownFields()
+		if err := decoder.Decode(&request); err != nil || request.ExpectedRevision < 1 || request.NextStage == "" {
+			http.Error(w, "invalid development gate transition", http.StatusBadRequest)
+			return
+		}
+		gate, err := workRepository.LoadDevelopmentGate(r.Context(), r.PathValue("id"))
+		if err != nil || gate.Revision != request.ExpectedRevision {
+			http.Error(w, "development gate revision conflict", http.StatusConflict)
+			return
+		}
+		gate, err = advanceDevelopmentGate(gate, request.NextStage, time.Now())
+		if err != nil {
+			http.Error(w, "development gate transition rejected", http.StatusUnprocessableEntity)
+			return
+		}
+		if err := workRepository.SaveDevelopmentGate(r.Context(), gate, request.ExpectedRevision); err != nil {
+			http.Error(w, "development gate conflict", http.StatusConflict)
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_ = json.NewEncoder(w).Encode(gate)
+	})
 	mux.HandleFunc("GET /v1/creator/focus", func(w http.ResponseWriter, r *http.Request) {
 		session, ok := authenticateCreator(w, r)
 		if !ok {
@@ -834,6 +967,20 @@ func main() {
 		w.Header().Set("Content-Type", "application/json")
 		_ = json.NewEncoder(w).Encode(binding)
 	})
+	mux.HandleFunc("GET /v1/cognitive-bindings/{id}/revisions/{revision}/snapshot", func(w http.ResponseWriter, r *http.Request) {
+		revision, err := strconv.ParseInt(r.PathValue("revision"), 10, 64)
+		if err != nil || revision < 1 {
+			http.Error(w, "invalid cognitive binding revision", http.StatusBadRequest)
+			return
+		}
+		snapshot, err := capabilityRepository.LoadCognitiveCandidate(r.Context(), capability.CapabilityBindingID(r.PathValue("id")), revision)
+		if err != nil {
+			http.Error(w, "cognitive binding snapshot not found", http.StatusNotFound)
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_ = json.NewEncoder(w).Encode(snapshot)
+	})
 	mux.HandleFunc("POST /v1/cognitive-bindings/resolve", func(w http.ResponseWriter, r *http.Request) {
 		var requirement capability.CognitiveRequirement
 		decoder := json.NewDecoder(io.LimitReader(r.Body, 16*1024+1))
@@ -842,7 +989,13 @@ func main() {
 			http.Error(w, "invalid cognitive requirement", http.StatusBadRequest)
 			return
 		}
-		candidates, err := capabilityRepository.LoadCognitiveCandidates(r.Context(), requirement.CapabilityID, requirement.OwnerPrincipalID)
+		var candidates []capability.ResolvedCognitiveBinding
+		var err error
+		if requirement.TaskProfileID != "" {
+			candidates, err = capabilityRepository.LoadCognitiveCandidatesForProfile(r.Context(), requirement.CapabilityID, requirement.OwnerPrincipalID, requirement.TaskProfileID)
+		} else {
+			candidates, err = capabilityRepository.LoadCognitiveCandidates(r.Context(), requirement.CapabilityID, requirement.OwnerPrincipalID)
+		}
 		if err != nil {
 			http.Error(w, "cognitive registry unavailable", http.StatusServiceUnavailable)
 			return
@@ -891,6 +1044,34 @@ func main() {
 		}
 		w.WriteHeader(http.StatusCreated)
 	})
+	mux.HandleFunc("POST /v1/cognitive-registry/task-evaluations", func(w http.ResponseWriter, r *http.Request) {
+		var value capability.TaskEvaluation
+		decoder := json.NewDecoder(io.LimitReader(r.Body, 64<<10))
+		decoder.DisallowUnknownFields()
+		if err := decoder.Decode(&value); err != nil || capabilitydomain.ValidateTaskEvaluation(value) != nil {
+			http.Error(w, "invalid task evaluation", http.StatusBadRequest)
+			return
+		}
+		if err := capabilityRepository.SaveTaskEvaluation(r.Context(), value); err != nil {
+			http.Error(w, "save task evaluation", http.StatusInternalServerError)
+			return
+		}
+		w.WriteHeader(http.StatusCreated)
+	})
+	mux.HandleFunc("POST /v1/cognitive-registry/model-qualification-observations", func(w http.ResponseWriter, r *http.Request) {
+		var value capability.ModelQualificationObservation
+		decoder := json.NewDecoder(io.LimitReader(r.Body, 16<<10))
+		decoder.DisallowUnknownFields()
+		if err := decoder.Decode(&value); err != nil || value.Validate() != nil {
+			http.Error(w, "invalid model qualification observation", http.StatusBadRequest)
+			return
+		}
+		if err := capabilityRepository.SaveModelQualificationObservation(r.Context(), value); err != nil {
+			http.Error(w, "save model qualification observation", http.StatusInternalServerError)
+			return
+		}
+		w.WriteHeader(http.StatusCreated)
+	})
 	mux.HandleFunc("POST /v1/creator/cognitive-canary", func(w http.ResponseWriter, r *http.Request) {
 		session, ok := authenticateCreator(w, r)
 		if !ok {

cmd/core-runtime/research_client.go

@@ -0,0 +1,53 @@
+package main
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"strings"
+
+	"git.cin.su/m/Assistant/contracts/epistemics"
+)
+
+type researchRuntimeClient struct {
+	baseURL string
+	client  *http.Client
+}
+
+func newResearchRuntimeClient(rawURL string, client *http.Client) (researchRuntimeClient, error) {
+	base := strings.TrimRight(strings.TrimSpace(rawURL), "/")
+	if base == "" {
+		return researchRuntimeClient{}, fmt.Errorf("research runtime URL is empty")
+	}
+	if client == nil {
+		client = http.DefaultClient
+	}
+	return researchRuntimeClient{baseURL: base, client: client}, nil
+}
+
+func (c researchRuntimeClient) EvaluateWeb(ctx context.Context, request epistemics.WebResearchRequest) (epistemics.ResearchResult, error) {
+	payload, err := json.Marshal(request)
+	if err != nil {
+		return epistemics.ResearchResult{}, err
+	}
+	httpRequest, err := http.NewRequestWithContext(ctx, http.MethodPost, c.baseURL+"/v1/research/web/evaluate", bytes.NewReader(payload))
+	if err != nil {
+		return epistemics.ResearchResult{}, err
+	}
+	httpRequest.Header.Set("Content-Type", "application/json")
+	response, err := c.client.Do(httpRequest)
+	if err != nil {
+		return epistemics.ResearchResult{}, err
+	}
+	defer response.Body.Close()
+	if response.StatusCode < 200 || response.StatusCode >= 300 {
+		return epistemics.ResearchResult{}, fmt.Errorf("research runtime status %d", response.StatusCode)
+	}
+	var result epistemics.ResearchResult
+	if err := json.NewDecoder(response.Body).Decode(&result); err != nil {
+		return epistemics.ResearchResult{}, err
+	}
+	return result, nil
+}

cmd/core-runtime/research_client_test.go

@@ -0,0 +1,33 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"git.cin.su/m/Assistant/contracts/epistemics"
+)
+
+func TestResearchRuntimeClientForwardsBoundedWebRequest(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.Method != http.MethodPost || r.URL.Path != "/v1/research/web/evaluate" || r.Header.Get("Content-Type") != "application/json" {
+			t.Fatalf("unexpected request: %s %s", r.Method, r.URL.Path)
+		}
+		var request epistemics.WebResearchRequest
+		if err := json.NewDecoder(r.Body).Decode(&request); err != nil || request.URL != "https://assistant.cin.su/healthz" {
+			t.Fatalf("unexpected request body: %#v %v", request, err)
+		}
+		_ = json.NewEncoder(w).Encode(epistemics.ResearchResult{PlanID: request.Plan.ID, Findings: []epistemics.ResearchFinding{{PropositionKey: request.PropositionKey, State: epistemics.FindingSupported}}})
+	}))
+	defer server.Close()
+	client, err := newResearchRuntimeClient(server.URL, server.Client())
+	if err != nil {
+		t.Fatal(err)
+	}
+	result, err := client.EvaluateWeb(context.Background(), epistemics.WebResearchRequest{Plan: epistemics.ResearchPlan{ID: "plan-1"}, PropositionKey: "health", URL: "https://assistant.cin.su/healthz"})
+	if err != nil || result.PlanID != "plan-1" || len(result.Findings) != 1 || result.Findings[0].State != epistemics.FindingSupported {
+		t.Fatalf("unexpected result: %#v %v", result, err)
+	}
+}

cmd/creator-web/server.go

@@ -150,7 +150,7 @@ func (g *gateway) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 		http.Error(w, "method not allowed", http.StatusMethodNotAllowed)
 		return
 	}
-	if r.URL.Path == "/" || r.URL.Path == "/chat" || r.URL.Path == "/chat.html" {
+	if r.URL.Path == "/" || r.URL.Path == "/login" || r.URL.Path == "/chat" || r.URL.Path == "/chat.html" {
 		g.serveEmbeddedHTML(w, "chat.html")
 		return
 	}

cmd/creator-web/server_test.go

@@ -354,6 +354,18 @@ func TestConsoleAndSecurityHeaders(t *testing.T) {
 	}
 }
 
+func TestLoginRouteServesChatShell(t *testing.T) {
+	handler, err := newHandler("http://core-runtime:8080", "http://continuity-runtime:8081", "http://commit-runtime:8082", http.DefaultClient)
+	if err != nil {
+		t.Fatal(err)
+	}
+	response := httptest.NewRecorder()
+	handler.ServeHTTP(response, httptest.NewRequest(http.MethodGet, "/login", nil))
+	if response.Code != http.StatusOK || !strings.Contains(response.Body.String(), "id=\"username\"") || !strings.Contains(response.Body.String(), "id=\"password\"") {
+		t.Fatalf("login response = %d", response.Code)
+	}
+}
+
 func TestAdvisoryModeRejectsInvalidConfiguration(t *testing.T) {
 	_, err := newHandlerWithAdvisoryMode("http://core.invalid", "http://continuity.invalid", "http://commit.invalid", http.DefaultClient, "browser-selected")
 	if err == nil || !strings.Contains(err.Error(), "invalid advisory binding mode") {

cmd/reflection-worker/main.go

@@ -4,6 +4,8 @@ import (
 	"context"
 	"flag"
 	"fmt"
+	"log"
+	"net/http"
 	"os"
 	"time"
 
@@ -29,9 +31,21 @@ func main() {
 		panic(err)
 	}
 	defer repository.Close()
+	status := newWorkerStatus(*workerID, *once)
+	listenAddr := os.Getenv("REFLECTION_LISTEN_ADDR")
+	if listenAddr == "" {
+		listenAddr = ":8092"
+	}
+	go func() {
+		if err := http.ListenAndServe(listenAddr, status.handler()); err != nil {
+			log.Printf("reflection status server stopped: %v", err)
+		}
+	}()
 	for {
 		now := time.Now().UnixMilli()
+		status.cycleStarted(time.Now())
 		if _, err := repository.RecoverReflectionClaims(ctx, now); err != nil {
+			status.cycleFailed(time.Now(), err)
 			panic(err)
 		}
 		candidate, claimErr := repository.ClaimReflectionCandidate(ctx, *workerID, now, lease.Milliseconds())
@@ -39,16 +53,19 @@ func main() {
 			// This worker only closes the durable evaluation loop. Any cognitive
 			// proposal remains a candidate and must be admitted by a separate governed flow.
 			if err := repository.ResolveReflectionCandidate(ctx, candidate.ID, *workerID, time.Now().UnixMilli()); err != nil {
+				status.cycleFailed(time.Now(), err)
 				panic(err)
 			}
 			fmt.Printf("resolved reflection candidate %s\n", candidate.ID)
 		} else if *once {
 			fmt.Println("no reflection candidate")
 		}
+		status.cycleFinished(time.Now(), map[bool]string{true: "candidate-resolved", false: "idle"}[claimErr == nil])
 		if *once {
 			return
 		}
 		if err := repository.WaitForReflectionCandidate(ctx, *pollInterval); err != nil {
+			status.cycleFailed(time.Now(), err)
 			panic(err)
 		}
 	}

cmd/reflection-worker/status.go

@@ -0,0 +1,74 @@
+package main
+
+import (
+	"encoding/json"
+	"net/http"
+	"sync"
+	"time"
+
+	"git.cin.su/m/Assistant/runtime/health"
+)
+
+type workerStatus struct {
+	mu            sync.RWMutex
+	workerID      string
+	once          bool
+	startedAt     time.Time
+	lastCycleAt   time.Time
+	lastCompleted time.Time
+	lastResult    string
+	lastError     string
+}
+
+func newWorkerStatus(workerID string, once bool) *workerStatus {
+	return &workerStatus{workerID: workerID, once: once, startedAt: time.Now().UTC()}
+}
+
+func (s *workerStatus) cycleStarted(at time.Time) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.lastCycleAt = at.UTC()
+	s.lastResult = "running"
+	s.lastError = ""
+}
+
+func (s *workerStatus) cycleFinished(at time.Time, result string) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.lastCompleted = at.UTC()
+	s.lastResult = result
+	s.lastError = ""
+}
+
+func (s *workerStatus) cycleFailed(at time.Time, err error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.lastCompleted = at.UTC()
+	s.lastResult = "failed"
+	s.lastError = err.Error()
+}
+
+func (s *workerStatus) handler() http.Handler {
+	mux := http.NewServeMux()
+	mux.Handle("/healthz", health.Handler{Runtime: "reflection-worker", SemanticGeneration: 1, ArchitectureGeneration: 1})
+	mux.HandleFunc("/status", func(w http.ResponseWriter, _ *http.Request) {
+		s.mu.RLock()
+		response := struct {
+			Runtime       string    `json:"runtime"`
+			WorkerID      string    `json:"worker_id"`
+			Mode          string    `json:"mode"`
+			StartedAt     time.Time `json:"started_at"`
+			LastCycleAt   time.Time `json:"last_cycle_at,omitempty"`
+			LastCompleted time.Time `json:"last_completed_at,omitempty"`
+			LastResult    string    `json:"last_result"`
+			LastError     string    `json:"last_error,omitempty"`
+		}{Runtime: "reflection-worker", WorkerID: s.workerID, Mode: "durable-reflection-loop", StartedAt: s.startedAt, LastCycleAt: s.lastCycleAt, LastCompleted: s.lastCompleted, LastResult: s.lastResult, LastError: s.lastError}
+		if s.once {
+			response.Mode = "durable-reflection-once"
+		}
+		s.mu.RUnlock()
+		w.Header().Set("Content-Type", "application/json")
+		_ = json.NewEncoder(w).Encode(response)
+	})
+	return mux
+}

cmd/reflection-worker/status_test.go

@@ -0,0 +1,35 @@
+package main
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+	"time"
+)
+
+func TestWorkerStatusReportsCycle(t *testing.T) {
+	status := newWorkerStatus("test-worker", false)
+	status.cycleStarted(time.Date(2026, 8, 23, 10, 0, 0, 0, time.UTC))
+	status.cycleFinished(time.Date(2026, 8, 23, 10, 0, 1, 0, time.UTC), "idle")
+
+	recorder := httptest.NewRecorder()
+	status.handler().ServeHTTP(recorder, httptest.NewRequest(http.MethodGet, "/status", nil))
+	if recorder.Code != http.StatusOK {
+		t.Fatalf("status code = %d, want 200", recorder.Code)
+	}
+	body := recorder.Body.String()
+	for _, expected := range []string{"\"runtime\":\"reflection-worker\"", "\"worker_id\":\"test-worker\"", "\"last_result\":\"idle\""} {
+		if !strings.Contains(body, expected) {
+			t.Fatalf("status body missing %q: %s", expected, body)
+		}
+	}
+}
+
+func TestWorkerHealthIsAvailable(t *testing.T) {
+	recorder := httptest.NewRecorder()
+	newWorkerStatus("test-worker", true).handler().ServeHTTP(recorder, httptest.NewRequest(http.MethodGet, "/healthz", nil))
+	if recorder.Code != http.StatusOK {
+		t.Fatalf("health status code = %d, want 200", recorder.Code)
+	}
+}

cmd/research-runtime/main.go

@@ -34,6 +34,7 @@ func main() {
 	mux := http.NewServeMux()
 	mux.Handle("/healthz", health.Handler{Runtime: "research-runtime", SemanticGeneration: 1, ArchitectureGeneration: 1})
 	mux.Handle("/readyz", health.ReadinessHandler{Runtime: "research-runtime", Probes: []health.Probe{{Name: "epistemics-postgres", Check: func(ctx context.Context) error { return repository.Ping(ctx) }}}})
+	mux.Handle("/v1/research/status", truthRuntimeStatusHandler(repository))
 	mux.HandleFunc("POST /v1/research/evaluate", func(w http.ResponseWriter, r *http.Request) {
 		var request epistemics.ResearchEvaluationRequest
 		if err := json.NewDecoder(http.MaxBytesReader(w, r.Body, 2<<20)).Decode(&request); err != nil {

cmd/research-runtime/status.go

@@ -0,0 +1,47 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+
+	"git.cin.su/m/Assistant/adapters/postgres"
+)
+
+type truthRuntimeStatusReader interface {
+	TruthRuntimeStatus(context.Context) (postgres.TruthRuntimeStatus, error)
+}
+
+type truthRuntimeStatusResponse struct {
+	Runtime              string           `json:"runtime"`
+	Status               string           `json:"status"`
+	BundleCount          int64            `json:"bundle_count"`
+	ClaimsByVerification map[string]int64 `json:"claims_by_verification"`
+	BeliefsByState       map[string]int64 `json:"beliefs_by_state"`
+	SourcesByFreshness   map[string]int64 `json:"sources_by_freshness"`
+	ObservedAtUnixMillis int64            `json:"observed_at_unix_millis"`
+}
+
+func truthRuntimeStatusHandler(reader truthRuntimeStatusReader) http.Handler {
+	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.Method != http.MethodGet || r.URL.Path != "/v1/research/status" {
+			http.NotFound(w, r)
+			return
+		}
+		status, err := reader.TruthRuntimeStatus(r.Context())
+		if err != nil {
+			http.Error(w, "research status unavailable", http.StatusServiceUnavailable)
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_ = json.NewEncoder(w).Encode(truthRuntimeStatusResponse{
+			Runtime:              "research-runtime",
+			Status:               "ready",
+			BundleCount:          status.BundleCount,
+			ClaimsByVerification: status.ClaimsByVerification,
+			BeliefsByState:       status.BeliefsByState,
+			SourcesByFreshness:   status.SourcesByFreshness,
+			ObservedAtUnixMillis: status.ObservedAtUnixMillis,
+		})
+	})
+}

cmd/research-runtime/status_test.go

@@ -0,0 +1,48 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"git.cin.su/m/Assistant/adapters/postgres"
+)
+
+type fakeTruthRuntimeStatusReader struct {
+	status postgres.TruthRuntimeStatus
+	err    error
+}
+
+func (f fakeTruthRuntimeStatusReader) TruthRuntimeStatus(context.Context) (postgres.TruthRuntimeStatus, error) {
+	return f.status, f.err
+}
+
+func TestTruthRuntimeStatusReportsEvidenceCounts(t *testing.T) {
+	response := httptest.NewRecorder()
+	truthRuntimeStatusHandler(fakeTruthRuntimeStatusReader{status: postgres.TruthRuntimeStatus{
+		BundleCount:          4,
+		ClaimsByVerification: map[string]int64{"UNVERIFIED": 3, "VERIFIED": 1},
+		BeliefsByState:       map[string]int64{"PROVISIONAL": 4},
+		SourcesByFreshness:   map[string]int64{"FRESH": 4},
+		ObservedAtUnixMillis: 1700000000000,
+	}}).ServeHTTP(response, httptest.NewRequest(http.MethodGet, "/v1/research/status", nil))
+	if response.Code != http.StatusOK {
+		t.Fatalf("status code = %d", response.Code)
+	}
+	for _, expected := range []string{"\"status\":\"ready\"", "\"bundle_count\":4", "\"VERIFIED\":1", "\"PROVISIONAL\":4", "1700000000000"} {
+		if !strings.Contains(response.Body.String(), expected) {
+			t.Fatalf("response missing %q: %s", expected, response.Body.String())
+		}
+	}
+}
+
+func TestTruthRuntimeStatusFailsClosedOnDatabaseError(t *testing.T) {
+	response := httptest.NewRecorder()
+	truthRuntimeStatusHandler(fakeTruthRuntimeStatusReader{err: errors.New("database unavailable")}).ServeHTTP(response, httptest.NewRequest(http.MethodGet, "/v1/research/status", nil))
+	if response.Code != http.StatusServiceUnavailable {
+		t.Fatalf("status code = %d", response.Code)
+	}
+}

cmd/task-verifier/main.go

@@ -0,0 +1,76 @@
+package main
+
+import (
+	"crypto/sha256"
+	"encoding/hex"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"os"
+	"strings"
+
+	"git.cin.su/m/Assistant/contracts/capability"
+)
+
+const verifierIdentity = "deterministic-task-verifier-v1"
+
+func main() {
+	mux := http.NewServeMux()
+	mux.HandleFunc("GET /healthz", func(w http.ResponseWriter, _ *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"runtime":"task-verifier","status":"ok"}`))
+	})
+	mux.HandleFunc("POST /v1/task-evaluations/verify", verifyHandler)
+	addr := ":8091"
+	if value := strings.TrimSpace(os.Getenv("LISTEN_ADDR")); value != "" {
+		addr = value
+	}
+	if err := http.ListenAndServe(addr, mux); err != nil {
+		panic(err)
+	}
+}
+
+func verifyHandler(w http.ResponseWriter, r *http.Request) {
+	var request capability.TaskCaseVerificationRequest
+	decoder := json.NewDecoder(http.MaxBytesReader(w, r.Body, 64<<10))
+	decoder.DisallowUnknownFields()
+	if err := decoder.Decode(&request); err != nil || request.CaseID == "" || request.ResponseText == "" || request.EvidenceArtifactRef == "" || request.LatencyMicros < 0 || (request.Verification.Mode != capability.TaskCaseVerificationExactText && request.Verification.Mode != capability.TaskCaseVerificationJSONFieldEquals) || (request.Verification.Mode == capability.TaskCaseVerificationJSONFieldEquals && request.Verification.FieldPath == "") || request.Verification.ExpectedText == "" || request.Verification.MinimumQualityMillis < 0 || request.Verification.MinimumQualityMillis > 1000 {
+		http.Error(w, "invalid task verification request", http.StatusBadRequest)
+		return
+	}
+	matched := matches(request)
+	quality := int64(0)
+	result := capability.EvaluationUnsupported
+	if matched {
+		quality = request.Verification.MinimumQualityMillis
+		if quality == 0 {
+			quality = 1000
+		}
+		result = capability.EvaluationSupported
+	}
+	evidence := verificationEvidence(request, matched)
+	writeJSON(w, capability.TaskCaseVerificationResponse{CaseID: request.CaseID, Result: result, QualityScoreMillis: quality, LatencyMicros: request.LatencyMicros, EvidenceArtifactRef: request.EvidenceArtifactRef, VerificationEvidenceRef: evidence, VerifierIdentity: verifierIdentity})
+}
+
+func verificationEvidence(request capability.TaskCaseVerificationRequest, matched bool) string {
+	payload := fmt.Sprintf("%s|%s|%s|%s|%s|%t", verifierIdentity, request.CaseID, request.EvidenceArtifactRef, request.Verification.Mode, request.Verification.FieldPath, matched)
+	digest := sha256.Sum256([]byte(payload))
+	return "verification:" + hex.EncodeToString(digest[:])
+}
+
+func matches(request capability.TaskCaseVerificationRequest) bool {
+	if request.Verification.Mode == capability.TaskCaseVerificationExactText {
+		return strings.TrimSpace(request.ResponseText) == strings.TrimSpace(request.Verification.ExpectedText)
+	}
+	var document map[string]json.RawMessage
+	if json.Unmarshal([]byte(request.ResponseText), &document) != nil {
+		return false
+	}
+	value, ok := document[request.Verification.FieldPath]
+	return ok && strings.TrimSpace(string(value)) == strings.TrimSpace(request.Verification.ExpectedText)
+}
+
+func writeJSON(w http.ResponseWriter, value any) {
+	w.Header().Set("Content-Type", "application/json")
+	_ = json.NewEncoder(w).Encode(value)
+}

cmd/task-verifier/main_test.go

@@ -0,0 +1,41 @@
+package main
+
+import (
+	"bytes"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"git.cin.su/m/Assistant/contracts/capability"
+)
+
+func TestVerifyExactTextProducesIndependentEvidence(t *testing.T) {
+	payload, _ := json.Marshal(capability.TaskCaseVerificationRequest{CaseID: "case-1", ResponseText: " answer ", EvidenceArtifactRef: "case-artifact", Verification: capability.TaskCaseVerification{Mode: capability.TaskCaseVerificationExactText, ExpectedText: "answer", MinimumQualityMillis: 900}})
+	recorder := httptest.NewRecorder()
+	verifyHandler(recorder, httptest.NewRequest(http.MethodPost, "/v1/task-evaluations/verify", bytes.NewReader(payload)))
+	var response capability.TaskCaseVerificationResponse
+	if recorder.Code != http.StatusOK || json.NewDecoder(recorder.Body).Decode(&response) != nil || response.Result != capability.EvaluationSupported || response.QualityScoreMillis != 900 || response.VerifierIdentity == "" || response.VerificationEvidenceRef == "" {
+		t.Fatalf("unexpected verification response: code=%d body=%s value=%+v", recorder.Code, recorder.Body.String(), response)
+	}
+}
+
+func TestVerifyMismatchIsUnsupportedWithEvidence(t *testing.T) {
+	payload, _ := json.Marshal(capability.TaskCaseVerificationRequest{CaseID: "case-1", ResponseText: "wrong", EvidenceArtifactRef: "case-artifact", Verification: capability.TaskCaseVerification{Mode: capability.TaskCaseVerificationExactText, ExpectedText: "answer", MinimumQualityMillis: 900}})
+	recorder := httptest.NewRecorder()
+	verifyHandler(recorder, httptest.NewRequest(http.MethodPost, "/v1/task-evaluations/verify", bytes.NewReader(payload)))
+	var response capability.TaskCaseVerificationResponse
+	if recorder.Code != http.StatusOK || json.NewDecoder(recorder.Body).Decode(&response) != nil || response.Result != capability.EvaluationUnsupported || response.QualityScoreMillis != 0 || response.VerificationEvidenceRef == "" {
+		t.Fatalf("unexpected mismatch response: code=%d body=%s value=%+v", recorder.Code, recorder.Body.String(), response)
+	}
+}
+
+func TestVerifyJSONFieldEquals(t *testing.T) {
+	payload, _ := json.Marshal(capability.TaskCaseVerificationRequest{CaseID: "json-1", ResponseText: `{"answer":42}`, EvidenceArtifactRef: "case-artifact", Verification: capability.TaskCaseVerification{Mode: capability.TaskCaseVerificationJSONFieldEquals, FieldPath: "answer", ExpectedText: "42", MinimumQualityMillis: 850}})
+	recorder := httptest.NewRecorder()
+	verifyHandler(recorder, httptest.NewRequest(http.MethodPost, "/v1/task-evaluations/verify", bytes.NewReader(payload)))
+	var response capability.TaskCaseVerificationResponse
+	if recorder.Code != http.StatusOK || json.NewDecoder(recorder.Body).Decode(&response) != nil || response.Result != capability.EvaluationSupported || response.QualityScoreMillis != 850 {
+		t.Fatalf("unexpected JSON verification response: code=%d body=%s value=%+v", recorder.Code, recorder.Body.String(), response)
+	}
+}

cmd/work-graph-runtime/main.go

@@ -7,9 +7,11 @@ import (
 	"log"
 	"net/http"
 	"os"
+	"time"
 
 	"git.cin.su/m/Assistant/adapters/postgres"
 	secretcontracts "git.cin.su/m/Assistant/contracts/secrets"
+	"git.cin.su/m/Assistant/contracts/work"
 	"git.cin.su/m/Assistant/modules/work/application"
 	"git.cin.su/m/Assistant/platform/secrets"
 	"git.cin.su/m/Assistant/runtime/health"
@@ -26,9 +28,14 @@ func main() {
 		log.Fatal(err)
 	}
 	defer repository.Close()
+	evidenceSink, err := newCoreTaskEvaluationSink(envOr("CORE_RUNTIME_URL", "http://core-runtime:8080"), &http.Client{Timeout: 12 * time.Second})
+	if err != nil {
+		log.Fatal(err)
+	}
 	mux := http.NewServeMux()
 	mux.Handle("/healthz", health.Handler{Runtime: "work-graph-runtime", SemanticGeneration: 1, ArchitectureGeneration: 1})
 	mux.Handle("/readyz", health.ReadinessHandler{Runtime: "work-graph-runtime", Probes: []health.Probe{{Name: "work-postgres", Check: func(ctx context.Context) error { return repository.Ping(ctx) }}}})
+	mux.Handle("/v1/work-graphs/status", workGraphStatusHandler(repository))
 	mux.HandleFunc("POST /v1/work-graphs/schedule", func(w http.ResponseWriter, r *http.Request) {
 		var request application.GraphScheduleRequest
 		if err := json.NewDecoder(http.MaxBytesReader(w, r.Body, 4<<20)).Decode(&request); err != nil {
@@ -42,19 +49,97 @@ func main() {
 		}
 		writeJSON(w, map[string]any{"graph": graph, "scheduled_nodes": ready})
 	})
+	mux.HandleFunc("POST /v1/work-graphs/task-evaluations/schedule", func(w http.ResponseWriter, r *http.Request) {
+		var request application.TaskEvaluationScheduleRequest
+		if err := json.NewDecoder(http.MaxBytesReader(w, r.Body, 4<<20)).Decode(&request); err != nil {
+			http.Error(w, "invalid task evaluation graph request", http.StatusBadRequest)
+			return
+		}
+		graph, ready, err := application.AdmitAndScheduleTaskEvaluation(r.Context(), repository, request)
+		if err != nil {
+			http.Error(w, err.Error(), http.StatusUnprocessableEntity)
+			return
+		}
+		writeJSON(w, map[string]any{"graph": graph, "scheduled_nodes": ready})
+	})
+	mux.HandleFunc("POST /v1/work-graphs/orchestration/next", func(w http.ResponseWriter, r *http.Request) {
+		var request work.SolutionSearchInput
+		if err := json.NewDecoder(http.MaxBytesReader(w, r.Body, 32<<10)).Decode(&request); err != nil {
+			http.Error(w, "invalid orchestration input", http.StatusBadRequest)
+			return
+		}
+		decision, err := application.DecideSolutionSearch(request)
+		if err != nil {
+			http.Error(w, "invalid orchestration input", http.StatusUnprocessableEntity)
+			return
+		}
+		writeJSON(w, decision)
+	})
+	mux.HandleFunc("POST /v1/work-graphs/orchestration/wave", func(w http.ResponseWriter, r *http.Request) {
+		var request work.SolutionSearchWaveRequest
+		if err := json.NewDecoder(http.MaxBytesReader(w, r.Body, 1<<20)).Decode(&request); err != nil {
+			http.Error(w, "invalid orchestration wave request", http.StatusBadRequest)
+			return
+		}
+		result, err := application.OrchestrateSolutionSearch(r.Context(), repository, request)
+		if err != nil {
+			http.Error(w, "invalid orchestration wave request", http.StatusUnprocessableEntity)
+			return
+		}
+		writeJSON(w, result)
+	})
+	mux.HandleFunc("POST /v1/work-graphs/result-callback", func(w http.ResponseWriter, r *http.Request) {
+		var request work.GraphResultCallback
+		if err := json.NewDecoder(http.MaxBytesReader(w, r.Body, 1<<20)).Decode(&request); err != nil {
+			http.Error(w, "invalid graph result callback", http.StatusBadRequest)
+			return
+		}
+		graph, err := application.RecordGraphResultCandidate(r.Context(), repository, repository, request)
+		if err != nil {
+			http.Error(w, "invalid graph result callback", http.StatusUnprocessableEntity)
+			return
+		}
+		writeJSON(w, graph)
+	})
 	mux.HandleFunc("POST /v1/work-graphs/complete", func(w http.ResponseWriter, r *http.Request) {
 		var request application.GraphCompleteRequest
 		if err := json.NewDecoder(http.MaxBytesReader(w, r.Body, 64<<10)).Decode(&request); err != nil {
 			http.Error(w, "invalid graph completion request", http.StatusBadRequest)
 			return
 		}
-		graph, err := application.CompleteGraphNode(r.Context(), repository, request.GraphID, request.Revision, request.NodeID, request.Succeeded)
+		result, err := application.CompleteGraphNodeWithReplan(r.Context(), repository, request)
+		if err != nil {
+			http.Error(w, err.Error(), http.StatusUnprocessableEntity)
+			return
+		}
+		writeJSON(w, result)
+	})
+	mux.HandleFunc("POST /v1/work-graphs/task-evaluations/callback", func(w http.ResponseWriter, r *http.Request) {
+		var request work.TaskEvaluationEvidenceCallback
+		if err := json.NewDecoder(http.MaxBytesReader(w, r.Body, 1<<20)).Decode(&request); err != nil {
+			http.Error(w, "invalid task evaluation callback", http.StatusBadRequest)
+			return
+		}
+		graph, err := application.CompleteTaskEvaluationNode(r.Context(), repository, evidenceSink, request)
 		if err != nil {
 			http.Error(w, err.Error(), http.StatusUnprocessableEntity)
 			return
 		}
 		writeJSON(w, graph)
 	})
+	mux.HandleFunc("POST /v1/work-graphs/synthesis", func(w http.ResponseWriter, r *http.Request) {
+		var request application.GraphSynthesisRequest
+		if err := json.NewDecoder(http.MaxBytesReader(w, r.Body, 4<<20)).Decode(&request); err != nil {
+			http.Error(w, "invalid graph synthesis request", http.StatusBadRequest)
+			return
+		}
+		conflict, synthesis, graph, err := application.SynthesizeAndCompleteGraph(r.Context(), repository, repository, request)
+		if err != nil {
+			http.Error(w, err.Error(), http.StatusUnprocessableEntity)
+			return
+		}
+		writeJSON(w, map[string]any{"conflict": conflict, "synthesis": synthesis, "graph": graph})
+	})
 	addr := os.Getenv("LISTEN_ADDR")
 	if addr == "" {
 		addr = ":8088"
@@ -63,6 +148,13 @@ func main() {
 	log.Fatal(http.ListenAndServe(addr, mux))
 }
 
+func envOr(name, fallback string) string {
+	if value := os.Getenv(name); value != "" {
+		return value
+	}
+	return fallback
+}
+
 func writeJSON(w http.ResponseWriter, value any) {
 	w.Header().Set("Content-Type", "application/json")
 	_ = json.NewEncoder(w).Encode(value)

cmd/work-graph-runtime/status.go

@@ -0,0 +1,45 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+
+	"git.cin.su/m/Assistant/adapters/postgres"
+)
+
+type workGraphStatusReader interface {
+	RuntimeStatus(context.Context) (postgres.WorkGraphRuntimeStatus, error)
+}
+
+type workGraphStatusResponse struct {
+	Runtime              string           `json:"runtime"`
+	Status               string           `json:"status"`
+	GraphsByState        map[string]int64 `json:"graphs_by_state"`
+	NodesByState         map[string]int64 `json:"nodes_by_state"`
+	QueueByState         map[string]int64 `json:"queue_by_state"`
+	ObservedAtUnixMillis int64            `json:"observed_at_unix_millis"`
+}
+
+func workGraphStatusHandler(reader workGraphStatusReader) http.Handler {
+	return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.Method != http.MethodGet || r.URL.Path != "/v1/work-graphs/status" {
+			http.NotFound(w, r)
+			return
+		}
+		status, err := reader.RuntimeStatus(r.Context())
+		if err != nil {
+			http.Error(w, "work graph status unavailable", http.StatusServiceUnavailable)
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_ = json.NewEncoder(w).Encode(workGraphStatusResponse{
+			Runtime:              "work-graph-runtime",
+			Status:               "ready",
+			GraphsByState:        status.GraphsByState,
+			NodesByState:         status.NodesByState,
+			QueueByState:         status.QueueByState,
+			ObservedAtUnixMillis: status.ObservedAtUnixMillis,
+		})
+	})
+}

cmd/work-graph-runtime/status_test.go

@@ -0,0 +1,56 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"git.cin.su/m/Assistant/adapters/postgres"
+)
+
+type fakeWorkGraphStatusReader struct {
+	status postgres.WorkGraphRuntimeStatus
+	err    error
+}
+
+func (f fakeWorkGraphStatusReader) RuntimeStatus(context.Context) (postgres.WorkGraphRuntimeStatus, error) {
+	return f.status, f.err
+}
+
+func TestWorkGraphStatusReportsDurableCounts(t *testing.T) {
+	response := httptest.NewRecorder()
+	workGraphStatusHandler(fakeWorkGraphStatusReader{status: postgres.WorkGraphRuntimeStatus{
+		GraphsByState:        map[string]int64{"RUNNING": 2},
+		NodesByState:         map[string]int64{"LEASED": 3, "SUCCEEDED": 1},
+		QueueByState:         map[string]int64{"QUEUED": 3},
+		ObservedAtUnixMillis: 1700000000000,
+	}}).ServeHTTP(response, httptest.NewRequest(http.MethodGet, "/v1/work-graphs/status", nil))
+	if response.Code != http.StatusOK {
+		t.Fatalf("status code = %d", response.Code)
+	}
+	body := response.Body.String()
+	for _, expected := range []string{"\"status\":\"ready\"", "\"RUNNING\":2", "\"LEASED\":3", "\"QUEUED\":3", "1700000000000"} {
+		if !contains(body, expected) {
+			t.Fatalf("response missing %q: %s", expected, body)
+		}
+	}
+}
+
+func TestWorkGraphStatusFailsClosedOnDatabaseError(t *testing.T) {
+	response := httptest.NewRecorder()
+	workGraphStatusHandler(fakeWorkGraphStatusReader{err: errors.New("database unavailable")}).ServeHTTP(response, httptest.NewRequest(http.MethodGet, "/v1/work-graphs/status", nil))
+	if response.Code != http.StatusServiceUnavailable {
+		t.Fatalf("status code = %d", response.Code)
+	}
+}
+
+func contains(value, expected string) bool {
+	for i := 0; i+len(expected) <= len(value); i++ {
+		if value[i:i+len(expected)] == expected {
+			return true
+		}
+	}
+	return false
+}

cmd/work-graph-runtime/task_evaluation_sink.go

@@ -0,0 +1,49 @@
+package main
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"strings"
+
+	"git.cin.su/m/Assistant/contracts/capability"
+)
+
+type coreTaskEvaluationSink struct {
+	baseURL string
+	client  *http.Client
+}
+
+func newCoreTaskEvaluationSink(baseURL string, client *http.Client) (*coreTaskEvaluationSink, error) {
+	baseURL = strings.TrimRight(strings.TrimSpace(baseURL), "/")
+	if baseURL == "" {
+		return nil, fmt.Errorf("core runtime URL is required")
+	}
+	if client == nil {
+		client = http.DefaultClient
+	}
+	return &coreTaskEvaluationSink{baseURL: baseURL, client: client}, nil
+}
+
+func (s *coreTaskEvaluationSink) SaveTaskEvaluation(ctx context.Context, value capability.TaskEvaluation) error {
+	payload, err := json.Marshal(value)
+	if err != nil {
+		return err
+	}
+	request, err := http.NewRequestWithContext(ctx, http.MethodPost, s.baseURL+"/v1/cognitive-registry/task-evaluations", bytes.NewReader(payload))
+	if err != nil {
+		return err
+	}
+	request.Header.Set("Content-Type", "application/json")
+	response, err := s.client.Do(request)
+	if err != nil {
+		return err
+	}
+	defer response.Body.Close()
+	if response.StatusCode < 200 || response.StatusCode >= 300 {
+		return fmt.Errorf("save task evaluation: status %s", response.Status)
+	}
+	return nil
+}

contracts/capability/cognitive_registry.go

@@ -26,6 +26,7 @@ type PerformanceEvidenceID ids.ID
 type BillingMode string
 type AuthMode string
 type SelectionObjective string
+type ModelRatingCategory string
 
 const (
 	QualificationDiscovered QualificationState = "DISCOVERED"
@@ -73,6 +74,11 @@ const (
 	SelectionQuality  SelectionObjective = "QUALITY"
 	SelectionSpeed    SelectionObjective = "SPEED"
 	SelectionEconomy  SelectionObjective = "ECONOMY"
+
+	ModelRatingUnrated  ModelRatingCategory = "UNRATED"
+	ModelRatingFast     ModelRatingCategory = "FAST"
+	ModelRatingBalanced ModelRatingCategory = "BALANCED"
+	ModelRatingQuality  ModelRatingCategory = "QUALITY"
 )
 
 // ProviderInstance contains only opaque references to endpoint and credentials.
@@ -154,16 +160,33 @@ type CognitiveRequirement struct {
 	ExcludedBindingIDs   []CapabilityBindingID
 	Objective            SelectionObjective
 	MaxCostMicros        *int64
+	TaskProfileID        TaskEvaluationProfileID
 }
 
 // ResolvedCognitiveBinding is a provider-neutral invocation snapshot selected
 // by Capability. It contains references and declared metadata, never secrets.
 type ResolvedCognitiveBinding struct {
-	Binding     CognitiveBinding
-	Provider    ProviderInstance
-	Model       ModelDescriptor
-	Evaluation  *EvaluatedCapability
-	Performance *PerformanceEvidence
+	Binding        CognitiveBinding
+	Provider       ProviderInstance
+	Model          ModelDescriptor
+	Evaluation     *EvaluatedCapability
+	Performance    *PerformanceEvidence
+	TaskEvaluation *TaskEvaluation
+	Qualification  *ModelQualificationObservation
+	Rating         *ModelRating
+}
+
+// ModelRating is a derived, provider-neutral explanation of why an eligible
+// model is routed. It never grants qualification or activation authority.
+type ModelRating struct {
+	TaskProfileID          TaskEvaluationProfileID `json:"task_profile_id,omitempty"`
+	TaskScoreMillis        int64                   `json:"task_score_millis,omitempty"`
+	Category               ModelRatingCategory     `json:"category"`
+	OverallScoreMillis     int64                   `json:"overall_score_millis"`
+	QualityScoreMillis     int64                   `json:"quality_score_millis"`
+	ReliabilityScoreMillis int64                   `json:"reliability_score_millis"`
+	LatencyScoreMillis     int64                   `json:"latency_score_millis"`
+	EvidenceScoreMillis    int64                   `json:"evidence_score_millis"`
 }
 
 type CognitiveResultClass string

contracts/capability/model_qualification.go

@@ -0,0 +1,47 @@
+package capability
+
+import (
+	"errors"
+
+	"git.cin.su/m/Assistant/platform/ids"
+)
+
+type ModelQualificationObservationID ids.ID
+type ModelQualificationStatus string
+
+const (
+	ModelQualificationReachable              ModelQualificationStatus = "REACHABLE"
+	ModelQualificationRespondedWithoutRating ModelQualificationStatus = "RESPONDED_WITHOUT_QUALIFICATION"
+	ModelQualificationUnavailable            ModelQualificationStatus = "UNAVAILABLE"
+)
+
+// ModelQualificationObservation is operational evidence only. It records
+// reachability and latency, never quality, activation, routing priority or
+// provider credentials.
+type ModelQualificationObservation struct {
+	ID                   ModelQualificationObservationID
+	ProviderInstanceID   ProviderInstanceID
+	ProviderModelID      ProviderModelID
+	Status               ModelQualificationStatus
+	FinishReason         string
+	LatencyMillis        int64
+	OutputObserved       bool
+	QualityAssessed      bool
+	Category             ModelRatingCategory
+	ObservedAtUnixMillis int64
+}
+
+func (o ModelQualificationObservation) Validate() error {
+	if o.ID == "" || o.ProviderInstanceID == "" || o.ProviderModelID == "" || o.FinishReason == "" || o.LatencyMillis < 0 || o.ObservedAtUnixMillis < 1 {
+		return errors.New("model qualification observation identity and timing are incomplete")
+	}
+	switch o.Status {
+	case ModelQualificationReachable, ModelQualificationRespondedWithoutRating, ModelQualificationUnavailable:
+	default:
+		return errors.New("model qualification observation status is unsupported")
+	}
+	if o.Category != ModelRatingUnrated || o.QualityAssessed {
+		return errors.New("model qualification observation cannot assert quality")
+	}
+	return nil
+}

contracts/capability/task_evaluation.go

@@ -0,0 +1,122 @@
+package capability
+
+import (
+	"errors"
+
+	"git.cin.su/m/Assistant/contracts/artifact"
+	"git.cin.su/m/Assistant/platform/ids"
+)
+
+type TaskEvaluationProfileID ids.ID
+type TaskEvaluationID ids.ID
+type TaskEvaluationProfileStage string
+
+const (
+	TaskEvaluationGeneral    TaskEvaluationProfileStage = "GENERAL"
+	TaskEvaluationAdditional TaskEvaluationProfileStage = "ADDITIONAL"
+)
+
+// TaskEvaluationProfile is the extensible, provider-neutral description of a
+// task-specific qualification run. Case payloads live in Artifact; the
+// profile stores only stable case references and scoring policy.
+type TaskEvaluationProfile struct {
+	ID                      TaskEvaluationProfileID
+	Revision                int64
+	CapabilityID            CapabilityID
+	Stage                   TaskEvaluationProfileStage
+	ParentProfileID         TaskEvaluationProfileID
+	TriggerEvidenceRef      string
+	BenchmarkVersion        string
+	RequiredCaseIDs         []string
+	QualityWeightMillis     int64
+	ReliabilityWeightMillis int64
+	LatencyWeightMillis     int64
+}
+
+func (p TaskEvaluationProfile) Validate() error {
+	if p.ID == "" || p.Revision < 1 || p.CapabilityID == "" || (p.Stage != TaskEvaluationGeneral && p.Stage != TaskEvaluationAdditional) || p.BenchmarkVersion == "" || len(p.RequiredCaseIDs) == 0 {
+		return errors.New("task evaluation profile identity and cases are incomplete")
+	}
+	if p.Stage == TaskEvaluationGeneral && (p.ParentProfileID != "" || p.TriggerEvidenceRef != "") {
+		return errors.New("general task evaluation profile cannot have an expansion parent")
+	}
+	if p.Stage == TaskEvaluationAdditional && (p.ParentProfileID == "" || p.TriggerEvidenceRef == "") {
+		return errors.New("additional task evaluation profile requires parent and trigger evidence")
+	}
+	if p.QualityWeightMillis < 0 || p.ReliabilityWeightMillis < 0 || p.LatencyWeightMillis < 0 || p.QualityWeightMillis+p.ReliabilityWeightMillis+p.LatencyWeightMillis != 1000 {
+		return errors.New("task evaluation profile weights must be non-negative and sum to 1000")
+	}
+	seen := make(map[string]struct{}, len(p.RequiredCaseIDs))
+	for _, caseID := range p.RequiredCaseIDs {
+		if caseID == "" {
+			return errors.New("task evaluation profile contains an empty case")
+		}
+		if _, ok := seen[caseID]; ok {
+			return errors.New("task evaluation profile contains duplicate cases")
+		}
+		seen[caseID] = struct{}{}
+	}
+	return nil
+}
+
+// TaskEvaluationExpansionRequest is emitted after a task remains unresolved
+// with the current best eligible candidate. It creates only a bounded new
+// profile; it never silently rewrites the general benchmark.
+type TaskEvaluationExpansionRequest struct {
+	ParentProfileID    TaskEvaluationProfileID
+	CapabilityID       CapabilityID
+	TriggerEvidenceRef string
+	NewCaseIDs         []string
+	BenchmarkVersion   string
+}
+
+type TaskCaseVerificationMode string
+
+const (
+	TaskCaseVerificationExactText       TaskCaseVerificationMode = "EXACT_TEXT"
+	TaskCaseVerificationJSONFieldEquals TaskCaseVerificationMode = "JSON_FIELD_EQUALS"
+)
+
+// TaskCaseVerification is a bounded, task-owned rubric. The expected text is
+// test data, not model output, and is never sent to the cognitive provider.
+type TaskCaseVerification struct {
+	Mode                 TaskCaseVerificationMode
+	FieldPath            string
+	ExpectedText         string
+	MinimumQualityMillis int64
+}
+
+// TaskCaseObservation is produced by the bounded evaluator and verifier. A
+// score without a case result and evidence reference is not accepted as
+// qualification evidence.
+type TaskCaseObservation struct {
+	CaseID              string
+	Result              EvaluationResult
+	QualityScoreMillis  int64
+	LatencyMicros       int64
+	CriticalFailure     bool
+	EvidenceArtifactRef string
+	// VerificationEvidenceRef and VerifierIdentity are required whenever an
+	// observation claims semantic support or a non-zero quality score.
+	VerificationEvidenceRef string
+	VerifierIdentity        string
+}
+
+// TaskEvaluation is durable evidence for one model/binding revision under one
+// task profile. It is not activation authority.
+type TaskEvaluation struct {
+	ID                     TaskEvaluationID
+	ProfileID              TaskEvaluationProfileID
+	BindingID              CapabilityBindingID
+	BindingRevision        int64
+	CapabilityID           CapabilityID
+	Result                 EvaluationResult
+	QualityScoreMillis     int64
+	ReliabilityScoreMillis int64
+	LatencyScoreMillis     int64
+	OverallScoreMillis     int64
+	CriticalFailureMillis  int64
+	BenchmarkVersion       string
+	ArtifactID             artifact.ArtifactID
+	ObservedAtUnixMillis   int64
+}

contracts/capability/task_evaluation_execution.go

@@ -0,0 +1,63 @@
+package capability
+
+import (
+	"git.cin.su/m/Assistant/contracts/authority"
+	"git.cin.su/m/Assistant/contracts/identity"
+)
+
+// TaskEvaluationCase carries one bounded prompt selected by the evaluator.
+// The prompt is an input artifact, not a model-produced fact.
+type TaskEvaluationCase struct {
+	CaseID              string
+	Prompt              string
+	EvidenceArtifactRef string
+	Verification        TaskCaseVerification
+}
+
+type TaskCaseResponse struct {
+	CaseID              string
+	ResponseText        string
+	LatencyMicros       int64
+	EvidenceArtifactRef string
+}
+
+type TaskCaseVerificationRequest struct {
+	CaseID              string
+	ResponseText        string
+	LatencyMicros       int64
+	EvidenceArtifactRef string
+	Verification        TaskCaseVerification
+}
+
+type TaskCaseVerificationResponse struct {
+	CaseID                  string
+	Result                  EvaluationResult
+	QualityScoreMillis      int64
+	LatencyMicros           int64
+	CriticalFailure         bool
+	EvidenceArtifactRef     string
+	VerificationEvidenceRef string
+	VerifierIdentity        string
+}
+
+type TaskEvaluationExecutionRequest struct {
+	Profile            TaskEvaluationProfile
+	CapabilityID       CapabilityID
+	BindingID          CapabilityBindingID
+	BindingRevision    int64
+	ActorID            identity.ActorID
+	PrincipalID        identity.PrincipalID
+	ScopeID            authority.ScopeID
+	Cases              []TaskEvaluationCase
+	MaxOutputBytes     int64
+	EgressAcknowledged bool
+}
+
+type TaskEvaluationExecutionResponse struct {
+	ProfileID            TaskEvaluationProfileID
+	BindingID            CapabilityBindingID
+	BindingRevision      int64
+	Observations         []TaskCaseObservation
+	Responses            []TaskCaseResponse
+	ObservedAtUnixMillis int64
+}

contracts/work/graph.go

@@ -1,11 +1,17 @@
 package work
 
+import "git.cin.su/m/Assistant/contracts/capability"
+
 type GraphNodeID string
 
 type GraphState string
 type GraphNodeState string
+type GraphNodePurpose string
 
 const (
+	GraphNodePurposeTaskEvaluation GraphNodePurpose = "TASK_EVALUATION_SPECIALIST"
+	GraphNodePurposeGeneric        GraphNodePurpose = "GENERIC"
+
 	GraphDraft     GraphState = "DRAFT"
 	GraphAdmitted  GraphState = "ADMITTED"
 	GraphRunning   GraphState = "RUNNING"
@@ -15,14 +21,15 @@ const (
 	GraphFailed    GraphState = "FAILED"
 	GraphUnknown   GraphState = "UNKNOWN"
 
-	GraphNodeBlocked GraphNodeState = "BLOCKED_PREREQUISITE"
-	GraphNodeReady   GraphNodeState = "READY"
-	GraphNodeLeased  GraphNodeState = "LEASED"
-	GraphNodeRunning GraphNodeState = "RUNNING"
-	GraphNodeSuccess GraphNodeState = "SUCCEEDED"
-	GraphNodeFailed  GraphNodeState = "FAILED"
-	GraphNodeCancel  GraphNodeState = "CANCELLED"
-	GraphNodeUnknown GraphNodeState = "UNKNOWN"
+	GraphNodeBlocked   GraphNodeState = "BLOCKED_PREREQUISITE"
+	GraphNodeReady     GraphNodeState = "READY"
+	GraphNodeLeased    GraphNodeState = "LEASED"
+	GraphNodeRunning   GraphNodeState = "RUNNING"
+	GraphNodeVerifying GraphNodeState = "VERIFYING"
+	GraphNodeSuccess   GraphNodeState = "SUCCEEDED"
+	GraphNodeFailed    GraphNodeState = "FAILED"
+	GraphNodeCancel    GraphNodeState = "CANCELLED"
+	GraphNodeUnknown   GraphNodeState = "UNKNOWN"
 )
 
 type GraphBudget struct {
@@ -31,11 +38,17 @@ type GraphBudget struct {
 }
 
 type GraphNode struct {
-	ID             GraphNodeID    `json:"id"`
-	DispatchID     string         `json:"dispatch_id"`
-	IdempotencyKey string         `json:"idempotency_key"`
-	State          GraphNodeState `json:"state"`
-	Attempts       int            `json:"attempts"`
+	ID             GraphNodeID                        `json:"id"`
+	DispatchID     string                             `json:"dispatch_id"`
+	IdempotencyKey string                             `json:"idempotency_key"`
+	State          GraphNodeState                     `json:"state"`
+	Attempts       int                                `json:"attempts"`
+	Purpose        GraphNodePurpose                   `json:"purpose,omitempty"`
+	TaskProfileID  capability.TaskEvaluationProfileID `json:"task_profile_id,omitempty"`
+	BindingID      capability.CapabilityBindingID     `json:"binding_id,omitempty"`
+	CaseIDs        []string                           `json:"case_ids,omitempty"`
+	TaskProfile    *capability.TaskEvaluationProfile  `json:"task_profile,omitempty"`
+	TaskCases      []capability.TaskEvaluationCase    `json:"task_cases,omitempty"`
 }
 
 type GraphEdge struct {
@@ -44,30 +57,98 @@ type GraphEdge struct {
 }
 
 type WorkGraph struct {
-	ID       string      `json:"id"`
-	Revision int64       `json:"revision"`
-	State    GraphState  `json:"state"`
-	Nodes    []GraphNode `json:"nodes"`
-	Edges    []GraphEdge `json:"edges"`
-	Budget   GraphBudget `json:"budget"`
+	ID                  string              `json:"id"`
+	Revision            int64               `json:"revision"`
+	State               GraphState          `json:"state"`
+	Nodes               []GraphNode         `json:"nodes"`
+	Edges               []GraphEdge         `json:"edges"`
+	Budget              GraphBudget         `json:"budget"`
+	ParentGraphID       string              `json:"parent_graph_id,omitempty"`
+	ParentGraphRevision int64               `json:"parent_graph_revision,omitempty"`
+	TriggerEvidenceRef  string              `json:"trigger_evidence_ref,omitempty"`
+	SearchInputBound    bool                `json:"search_input_bound,omitempty"`
+	SearchInput         SolutionSearchInput `json:"search_input,omitempty"`
+}
+
+// TaskEvaluationGraphRequest describes a durable fan-out. The graph contains
+// one specialist node per binding; each node receives the profile's bounded
+// cases and reports evidence back to Capability for fan-in aggregation.
+type TaskEvaluationGraphRequest struct {
+	GraphID             string                           `json:"graph_id"`
+	Profile             capability.TaskEvaluationProfile `json:"profile"`
+	BindingIDs          []capability.CapabilityBindingID `json:"binding_ids"`
+	DispatchIDs         []string                         `json:"dispatch_ids"`
+	IdempotencyPrefix   string                           `json:"idempotency_prefix"`
+	Budget              GraphBudget                      `json:"budget"`
+	ParentGraphID       string                           `json:"parent_graph_id,omitempty"`
+	ParentGraphRevision int64                            `json:"parent_graph_revision,omitempty"`
+	ParentGraphState    GraphState                       `json:"parent_graph_state,omitempty"`
+	TriggerEvidenceRef  string                           `json:"trigger_evidence_ref,omitempty"`
+	Cases               []capability.TaskEvaluationCase  `json:"cases,omitempty"`
+	BindingRevision     int64                            `json:"binding_revision,omitempty"`
+	ActorID             string                           `json:"actor_id,omitempty"`
+	PrincipalID         string                           `json:"principal_id,omitempty"`
+	ScopeID             string                           `json:"scope_id,omitempty"`
+}
+
+// TaskEvaluationEvidenceCallback is emitted by a specialist after it has
+// completed the bounded case set. Capability remains the evidence owner;
+// Work only advances its matching graph node after the evidence is accepted.
+type TaskEvaluationEvidenceCallback struct {
+	GraphID    string                    `json:"graph_id"`
+	Revision   int64                     `json:"revision"`
+	NodeID     GraphNodeID               `json:"node_id"`
+	Evaluation capability.TaskEvaluation `json:"evaluation"`
+	Succeeded  bool                      `json:"succeeded"`
 }
 
 type ResultCandidate struct {
-	ID              string
-	GraphID         string
-	GraphRevision   int64
-	NodeID          GraphNodeID
-	AuthorRunID     AgentRunID
-	ArtifactRefs    []EvidenceRef
-	EvidenceSummary string
-	Summary         string
+	ID              string        `json:"id"`
+	GraphID         string        `json:"graph_id"`
+	GraphRevision   int64         `json:"graph_revision"`
+	NodeID          GraphNodeID   `json:"node_id"`
+	AuthorRunID     AgentRunID    `json:"author_run_id"`
+	ArtifactRefs    []EvidenceRef `json:"artifact_refs"`
+	EvidenceSummary string        `json:"evidence_summary"`
+	Summary         string        `json:"summary"`
+}
+
+// GraphResultCallback carries one agent's attributable candidate to Work.
+// Recording a candidate never makes the node successful; the node enters
+// VERIFYING until an independent synthesis/verifier accepts it.
+type GraphResultCallback struct {
+	GraphID   string          `json:"graph_id"`
+	Revision  int64           `json:"revision"`
+	Candidate ResultCandidate `json:"candidate"`
 }
 
 type ConflictSet struct {
-	ID            string
-	GraphID       string
-	GraphRevision int64
-	NodeID        GraphNodeID
-	CandidateIDs  []string
-	Resolved      bool
+	ID            string      `json:"id"`
+	GraphID       string      `json:"graph_id"`
+	GraphRevision int64       `json:"graph_revision"`
+	NodeID        GraphNodeID `json:"node_id"`
+	CandidateIDs  []string    `json:"candidate_ids"`
+	Resolved      bool        `json:"resolved"`
+}
+
+type GraphSynthesisState string
+
+const GraphSynthesisVerified GraphSynthesisState = "VERIFIED"
+
+// GraphSynthesisCandidate retains every competing result while recording the
+// independently verified selection. It is a candidate, not an external action
+// or authority grant.
+type GraphSynthesisCandidate struct {
+	ID                  string              `json:"id"`
+	GraphID             string              `json:"graph_id"`
+	GraphRevision       int64               `json:"graph_revision"`
+	NodeID              GraphNodeID         `json:"node_id"`
+	CandidateIDs        []string            `json:"candidate_ids"`
+	SelectedCandidateID string              `json:"selected_candidate_id"`
+	DissentCandidateIDs []string            `json:"dissent_candidate_ids"`
+	VerifierRunID       AgentRunID          `json:"verifier_run_id"`
+	VerifierIdentity    string              `json:"verifier_identity"`
+	EvidenceRefs        []EvidenceRef       `json:"evidence_refs"`
+	Summary             string              `json:"summary"`
+	State               GraphSynthesisState `json:"state"`
 }

contracts/work/lease.go

@@ -1,6 +1,12 @@
 package work
 
-import "time"
+import (
+	"time"
+
+	"git.cin.su/m/Assistant/contracts/authority"
+	"git.cin.su/m/Assistant/contracts/capability"
+	"git.cin.su/m/Assistant/contracts/identity"
+)
 
 type AgentLeaseID string
 type AgentFencingToken uint64
@@ -50,14 +56,31 @@ type IndependentVerification struct {
 }
 
 type AgentDispatchIntent struct {
-	DispatchID             string       `json:"dispatch_id"`
-	IdempotencyKey         string       `json:"idempotency_key"`
-	DelegationID           DelegationID `json:"delegation_id"`
-	RunID                  AgentRunID   `json:"run_id"`
-	ImplementationID       string       `json:"implementation_id"`
-	ImplementationRevision int64        `json:"implementation_revision"`
-	SourceRevision         string       `json:"source_revision"`
-	WorktreeID             string       `json:"worktree_id"`
-	VerificationPlanRef    string       `json:"verification_plan_ref"`
-	CreatedAtUnixMillis    int64        `json:"created_at_unix_millis"`
+	DispatchID              string                             `json:"dispatch_id"`
+	IdempotencyKey          string                             `json:"idempotency_key"`
+	DelegationID            DelegationID                       `json:"delegation_id"`
+	RunID                   AgentRunID                         `json:"run_id"`
+	ImplementationID        string                             `json:"implementation_id"`
+	ImplementationRevision  int64                              `json:"implementation_revision"`
+	SourceRevision          string                             `json:"source_revision"`
+	WorktreeID              string                             `json:"worktree_id"`
+	VerificationPlanRef     string                             `json:"verification_plan_ref"`
+	CreatedAtUnixMillis     int64                              `json:"created_at_unix_millis"`
+	SpecialistPurpose       string                             `json:"specialist_purpose,omitempty"`
+	TaskEvaluationProfileID capability.TaskEvaluationProfileID `json:"task_evaluation_profile_id,omitempty"`
+	TaskEvaluationBindingID capability.CapabilityBindingID     `json:"task_evaluation_binding_id,omitempty"`
+	TaskEvaluationCaseIDs   []string                           `json:"task_evaluation_case_ids,omitempty"`
+	TaskEvaluationProfile   *capability.TaskEvaluationProfile  `json:"task_evaluation_profile,omitempty"`
+	TaskEvaluationCases     []capability.TaskEvaluationCase    `json:"task_evaluation_cases,omitempty"`
+	TaskEvaluationExecution *TaskEvaluationDispatchContext     `json:"task_evaluation_execution,omitempty"`
+}
+
+type TaskEvaluationDispatchContext struct {
+	GraphID         string               `json:"graph_id"`
+	GraphRevision   int64                `json:"graph_revision"`
+	NodeID          GraphNodeID          `json:"node_id"`
+	BindingRevision int64                `json:"binding_revision"`
+	ActorID         identity.ActorID     `json:"actor_id"`
+	PrincipalID     identity.PrincipalID `json:"principal_id"`
+	ScopeID         authority.ScopeID    `json:"scope_id"`
 }

contracts/work/orchestration.go

@@ -0,0 +1,66 @@
+package work
+
+type OrchestrationState string
+type SearchOptionKind string
+
+const (
+	OrchestrationSearching OrchestrationState = "SEARCHING"
+	OrchestrationExecuting OrchestrationState = "EXECUTING"
+	OrchestrationVerifying OrchestrationState = "VERIFYING"
+	OrchestrationWaiting   OrchestrationState = "WAITING"
+	OrchestrationCompleted OrchestrationState = "COMPLETED"
+	OrchestrationCancelled OrchestrationState = "CANCELLED"
+
+	SearchOptionAgents     SearchOptionKind = "AGENT_TEAM"
+	SearchOptionCapability SearchOptionKind = "CAPABILITY_DISCOVERY"
+	SearchOptionResearch   SearchOptionKind = "TRUTH_RESEARCH"
+	SearchOptionReflection SearchOptionKind = "REFLECTION_MEMORY"
+	SearchOptionHuman      SearchOptionKind = "HUMAN_ATTENTION"
+)
+
+// SolutionSearchInput is the durable orchestration snapshot. The scheduler
+// stores the graph/goal separately; this value only decides the next bounded
+// search wave and never grants authority by itself.
+type SolutionSearchInput struct {
+	GoalID            GoalID
+	GoalState         GoalState
+	SearchRound       int
+	MaxRound          int
+	CancelRequested   bool
+	ActiveOptionCount int
+	VerificationReady bool
+	Unresolved        bool
+}
+
+type SolutionSearchDecision struct {
+	GoalID            GoalID
+	State             OrchestrationState
+	ContinueSearching bool
+	NextSearchRound   int
+	Options           []SearchOptionKind
+	Reason            string
+}
+
+type SolutionSearchWaveRequest struct {
+	Input               SolutionSearchInput      `json:"input"`
+	GraphID             string                   `json:"graph_id"`
+	IdempotencyPrefix   string                   `json:"idempotency_prefix"`
+	Dispatches          []SolutionSearchDispatch `json:"dispatches"`
+	Budget              GraphBudget              `json:"budget"`
+	ParentGraphID       string                   `json:"parent_graph_id,omitempty"`
+	ParentGraphRevision int64                    `json:"parent_graph_revision,omitempty"`
+	TriggerEvidenceRef  string                   `json:"trigger_evidence_ref,omitempty"`
+	NowUnixMillis       int64                    `json:"now_unix_millis"`
+}
+
+type SolutionSearchDispatch struct {
+	Option SearchOptionKind    `json:"option"`
+	Intent AgentDispatchIntent `json:"intent"`
+}
+
+type SolutionSearchWaveResult struct {
+	Decision          SolutionSearchDecision `json:"decision"`
+	Graph             WorkGraph              `json:"graph"`
+	ScheduledNodeIDs  []GraphNodeID          `json:"scheduled_node_ids,omitempty"`
+	GraphMaterialized bool                   `json:"graph_materialized"`
+}

contracts/work/self_development.go

@@ -1,5 +1,7 @@
 package work
 
+import "time"
+
 type DevelopmentStage string
 
 const (
@@ -15,6 +17,8 @@ const (
 
 type DevelopmentGate struct {
 	ID                   string
+	Revision             int64
+	UpdatedAtUnixMillis  int64
 	Stage                DevelopmentStage
 	EvidenceRefs         []string
 	ProposalRef          string
@@ -27,3 +31,103 @@ type DevelopmentGate struct {
 	PromotionRef         string
 	RollbackRef          string
 }
+
+// ExtensionProposalStage is deliberately proposal-only in this contract. A
+// proposal does not authorize source changes, provider calls, deployment or
+// promotion; those stages require the existing governed development gate.
+type ExtensionProposalStage string
+
+const ExtensionProposalProposed ExtensionProposalStage = "PROPOSED"
+
+// ExtensionEvidence is an immutable reference to an observation that points
+// at a repeated or corroborated improvement signal. The payload is owned by
+// its source module; Work stores only the typed reference and stable signal
+// key.
+type ExtensionEvidence struct {
+	Ref       string
+	SignalKey string
+}
+
+// ExtensionProposalRequest contains the operator/governance information
+// needed to decide whether an evidence pattern is worth reviewing. Empty
+// fields are rejected by the proposal generator rather than inferred.
+type ExtensionProposalRequest struct {
+	ID                  string
+	DesiredOutcome      string
+	SmallestChangeLayer string
+	AffectedOwners      []string
+	AffectedContracts   []string
+	RiskTier            string
+	EvaluationPlan      []string
+	Alternatives        []string
+	RollbackPlan        string
+	Assumptions         []string
+	CreatedAt           time.Time
+}
+
+// ExtensionProposal is a bounded, evidence-bearing candidate for governed
+// development. It intentionally contains no executable command, credential,
+// authority grant, implementation body or promotion decision.
+type ExtensionProposal struct {
+	ID                     string
+	RequestedByPrincipalID string
+	Stage                  ExtensionProposalStage
+	EvidenceRefs           []string
+	DeduplicatedSignalKeys []string
+	DesiredOutcome         string
+	SmallestChangeLayer    string
+	AffectedOwners         []string
+	AffectedContracts      []string
+	RiskTier               string
+	EvaluationPlan         []string
+	Alternatives           []string
+	RollbackPlan           string
+	Assumptions            []string
+	CreatedAtUnixMillis    int64
+}
+
+type DevelopmentScaffoldStage string
+
+const DevelopmentScaffoldSandboxed DevelopmentScaffoldStage = "SANDBOXED"
+
+// DevelopmentScaffoldRequest describes the boundary a later control plane
+// would use to create an isolated worktree. This value object itself performs
+// no checkout or filesystem operation.
+type DevelopmentScaffoldRequest struct {
+	ID                  string
+	ProposalRef         string
+	BaseRepoRevision    string
+	WorktreeRef         string
+	TaskIntent          string
+	Modules             []string
+	AllowedPaths        []string
+	ForbiddenPaths      []string
+	ProtectedPaths      []string
+	RequiredTests       []string
+	ArtifactLineageRefs []string
+	ProtectedMainAccess string
+	ProductionAccess    string
+	SecretAccess        string
+	CreatedAt           time.Time
+}
+
+// DevelopmentScaffold is a provenance-linked task-packet candidate. It has
+// no executable instruction, credential, checkout handle or authority grant.
+type DevelopmentScaffold struct {
+	ID                     string
+	RequestedByPrincipalID string
+	Stage                  DevelopmentScaffoldStage
+	ProposalRef            string
+	BaseRepoRevision       string
+	WorktreeRef            string
+	TaskIntent             string
+	Modules                []string
+	AllowedPaths           []string
+	ForbiddenPaths         []string
+	RequiredTests          []string
+	ArtifactLineageRefs    []string
+	ProtectedMainAccess    string
+	ProductionAccess       string
+	SecretAccess           string
+	CreatedAtUnixMillis    int64
+}

database/migrations/036_work_graph_result_synthesis.sql

@@ -0,0 +1,58 @@
+BEGIN;
+
+SET ROLE work_role;
+
+CREATE TABLE IF NOT EXISTS work.graph_result_candidates (
+  result_candidate_id TEXT PRIMARY KEY,
+  graph_id TEXT NOT NULL,
+  graph_revision BIGINT NOT NULL,
+  node_id TEXT NOT NULL,
+  author_run_id TEXT NOT NULL,
+  artifact_refs TEXT[] NOT NULL DEFAULT ARRAY[]::TEXT[],
+  evidence_summary TEXT NOT NULL,
+  summary TEXT NOT NULL,
+  recorded_at TIMESTAMPTZ NOT NULL DEFAULT CURRENT_TIMESTAMP,
+  CHECK (result_candidate_id <> ''), CHECK (graph_id <> ''), CHECK (graph_revision > 0),
+  CHECK (node_id <> ''), CHECK (author_run_id <> ''), CHECK (evidence_summary <> ''), CHECK (summary <> ''),
+  FOREIGN KEY (graph_id, graph_revision) REFERENCES work.graphs(graph_id, revision)
+);
+
+CREATE TABLE IF NOT EXISTS work.graph_conflicts (
+  conflict_id TEXT PRIMARY KEY,
+  graph_id TEXT NOT NULL,
+  graph_revision BIGINT NOT NULL,
+  node_id TEXT NOT NULL,
+  candidate_ids TEXT[] NOT NULL,
+  resolved BOOLEAN NOT NULL DEFAULT FALSE,
+  recorded_at TIMESTAMPTZ NOT NULL DEFAULT CURRENT_TIMESTAMP,
+  CHECK (conflict_id <> ''), CHECK (graph_id <> ''), CHECK (graph_revision > 0), CHECK (node_id <> ''),
+  CHECK (cardinality(candidate_ids) >= 2),
+  FOREIGN KEY (graph_id, graph_revision) REFERENCES work.graphs(graph_id, revision)
+);
+
+CREATE TABLE IF NOT EXISTS work.graph_synthesis_candidates (
+  synthesis_id TEXT PRIMARY KEY,
+  graph_id TEXT NOT NULL,
+  graph_revision BIGINT NOT NULL,
+  node_id TEXT NOT NULL,
+  candidate_ids TEXT[] NOT NULL,
+  selected_candidate_id TEXT NOT NULL,
+  dissent_candidate_ids TEXT[] NOT NULL DEFAULT ARRAY[]::TEXT[],
+  verifier_run_id TEXT NOT NULL,
+  verifier_identity TEXT NOT NULL,
+  evidence_refs TEXT[] NOT NULL,
+  summary TEXT NOT NULL,
+  state TEXT NOT NULL,
+  recorded_at TIMESTAMPTZ NOT NULL DEFAULT CURRENT_TIMESTAMP,
+  CHECK (synthesis_id <> ''), CHECK (graph_id <> ''), CHECK (graph_revision > 0), CHECK (node_id <> ''),
+  CHECK (cardinality(candidate_ids) >= 2), CHECK (selected_candidate_id <> ''), CHECK (verifier_run_id <> ''),
+  CHECK (verifier_identity <> ''), CHECK (cardinality(evidence_refs) >= 1), CHECK (summary <> ''), CHECK (state = 'VERIFIED'),
+  FOREIGN KEY (graph_id, graph_revision) REFERENCES work.graphs(graph_id, revision)
+);
+
+CREATE INDEX IF NOT EXISTS graph_result_candidates_node_idx ON work.graph_result_candidates(graph_id, graph_revision, node_id);
+CREATE INDEX IF NOT EXISTS graph_conflicts_node_idx ON work.graph_conflicts(graph_id, graph_revision, node_id);
+CREATE INDEX IF NOT EXISTS graph_synthesis_node_idx ON work.graph_synthesis_candidates(graph_id, graph_revision, node_id);
+
+RESET ROLE;
+COMMIT;

database/migrations/037_work_development_gates.sql

@@ -0,0 +1,16 @@
+CREATE TABLE IF NOT EXISTS work.development_gates (
+  gate_id TEXT PRIMARY KEY,
+  revision BIGINT NOT NULL CHECK (revision > 0),
+  updated_at_unix_millis BIGINT NOT NULL CHECK (updated_at_unix_millis > 0),
+  stage TEXT NOT NULL,
+  evidence_refs TEXT[] NOT NULL,
+  proposal_ref TEXT NOT NULL,
+  task_packet_ref TEXT NOT NULL,
+  worktree_ref TEXT NOT NULL,
+  producer_run_refs TEXT[] NOT NULL,
+  verifier_run_refs TEXT[] NOT NULL,
+  verification_evidence TEXT NOT NULL,
+  canary_evidence TEXT NOT NULL,
+  promotion_ref TEXT NOT NULL,
+  rollback_ref TEXT NOT NULL
+);

database/migrations/038_work_development_gates_privileges.sql

@@ -0,0 +1,6 @@
+BEGIN;
+
+ALTER TABLE work.development_gates OWNER TO work_role;
+GRANT SELECT, INSERT, UPDATE ON work.development_gates TO work_role;
+
+COMMIT;

database/migrations/039_capability_task_evaluations.sql

@@ -0,0 +1,35 @@
+BEGIN;
+SET ROLE capability_role;
+
+CREATE TABLE IF NOT EXISTS capability.cognitive_task_evaluations (
+  evaluation_id TEXT PRIMARY KEY,
+  task_profile_id TEXT NOT NULL,
+  binding_id TEXT NOT NULL,
+  binding_revision BIGINT NOT NULL,
+  capability_id TEXT NOT NULL,
+  result TEXT NOT NULL,
+  quality_score_millis BIGINT NOT NULL,
+  reliability_score_millis BIGINT NOT NULL,
+  latency_score_millis BIGINT NOT NULL,
+  overall_score_millis BIGINT NOT NULL,
+  critical_failure_millis BIGINT NOT NULL,
+  benchmark_version TEXT NOT NULL,
+  artifact_id TEXT NOT NULL,
+  observed_at_unix_millis BIGINT NOT NULL,
+  recorded_at TIMESTAMPTZ NOT NULL DEFAULT CURRENT_TIMESTAMP,
+  CHECK (binding_revision > 0),
+  CHECK (result IN ('UNKNOWN','SUPPORTED','UNSUPPORTED')),
+  CHECK (quality_score_millis BETWEEN 0 AND 1000),
+  CHECK (reliability_score_millis BETWEEN 0 AND 1000),
+  CHECK (latency_score_millis BETWEEN 0 AND 1000),
+  CHECK (overall_score_millis BETWEEN 0 AND 1000),
+  CHECK (critical_failure_millis BETWEEN 0 AND 1000),
+  CHECK (observed_at_unix_millis > 0)
+);
+
+CREATE INDEX IF NOT EXISTS cognitive_task_evaluations_lookup
+  ON capability.cognitive_task_evaluations(task_profile_id, binding_id, binding_revision, observed_at_unix_millis DESC);
+
+GRANT SELECT, INSERT ON capability.cognitive_task_evaluations TO capability_role;
+RESET ROLE;
+COMMIT;

database/migrations/040_task_evaluation_dispatch.sql

@@ -0,0 +1,23 @@
+BEGIN;
+
+SET ROLE work_role;
+
+ALTER TABLE work.agent_dispatch_intents
+  ADD COLUMN IF NOT EXISTS specialist_purpose TEXT NOT NULL DEFAULT '',
+  ADD COLUMN IF NOT EXISTS task_evaluation_profile_id TEXT NOT NULL DEFAULT '',
+  ADD COLUMN IF NOT EXISTS task_evaluation_binding_id TEXT NOT NULL DEFAULT '',
+  ADD COLUMN IF NOT EXISTS task_evaluation_case_ids TEXT[] NOT NULL DEFAULT '{}';
+
+ALTER TABLE work.graphs
+  ADD COLUMN IF NOT EXISTS parent_graph_id TEXT NOT NULL DEFAULT '',
+  ADD COLUMN IF NOT EXISTS parent_graph_revision BIGINT NOT NULL DEFAULT 0,
+  ADD COLUMN IF NOT EXISTS trigger_evidence_ref TEXT NOT NULL DEFAULT '';
+
+ALTER TABLE work.graph_nodes
+  ADD COLUMN IF NOT EXISTS purpose TEXT NOT NULL DEFAULT 'GENERIC',
+  ADD COLUMN IF NOT EXISTS task_profile_id TEXT NOT NULL DEFAULT '',
+  ADD COLUMN IF NOT EXISTS binding_id TEXT NOT NULL DEFAULT '',
+  ADD COLUMN IF NOT EXISTS case_ids TEXT[] NOT NULL DEFAULT '{}';
+
+RESET ROLE;
+COMMIT;

database/migrations/041_task_evaluation_payload.sql

@@ -0,0 +1,16 @@
+BEGIN;
+
+SET ROLE work_role;
+
+ALTER TABLE work.agent_dispatch_intents
+  ADD COLUMN IF NOT EXISTS task_evaluation_profile JSONB,
+  ADD COLUMN IF NOT EXISTS task_evaluation_cases JSONB NOT NULL DEFAULT '[]',
+  ADD COLUMN IF NOT EXISTS task_evaluation_execution JSONB;
+
+ALTER TABLE work.graph_nodes
+  ADD COLUMN IF NOT EXISTS task_evaluation_profile JSONB,
+  ADD COLUMN IF NOT EXISTS task_evaluation_cases JSONB NOT NULL DEFAULT '[]',
+  ADD COLUMN IF NOT EXISTS task_evaluation_execution JSONB;
+
+RESET ROLE;
+COMMIT;

database/migrations/042_agent_execution_recovery.sql

@@ -0,0 +1,15 @@
+BEGIN;
+
+SET ROLE work_role;
+
+ALTER TABLE work.agent_dispatch_queue
+  ADD COLUMN IF NOT EXISTS retry_limit INTEGER NOT NULL DEFAULT 3;
+
+ALTER TABLE work.agent_dispatch_queue
+  DROP CONSTRAINT IF EXISTS agent_dispatch_queue_retry_limit_check;
+
+ALTER TABLE work.agent_dispatch_queue
+  ADD CONSTRAINT agent_dispatch_queue_retry_limit_check CHECK (retry_limit > 0);
+
+RESET ROLE;
+COMMIT;

database/migrations/043_governed_development_candidates.sql

@@ -0,0 +1,45 @@
+BEGIN;
+
+SET ROLE work_role;
+
+CREATE TABLE IF NOT EXISTS work.extension_proposals (
+  proposal_id TEXT PRIMARY KEY,
+  requested_by_principal_id TEXT NOT NULL,
+  stage TEXT NOT NULL,
+  evidence_refs TEXT[] NOT NULL DEFAULT ARRAY[]::TEXT[],
+  deduplicated_signal_keys TEXT[] NOT NULL DEFAULT ARRAY[]::TEXT[],
+  desired_outcome TEXT NOT NULL,
+  smallest_change_layer TEXT NOT NULL,
+  affected_owners TEXT[] NOT NULL DEFAULT ARRAY[]::TEXT[],
+  affected_contracts TEXT[] NOT NULL DEFAULT ARRAY[]::TEXT[],
+  risk_tier TEXT NOT NULL,
+  evaluation_plan TEXT[] NOT NULL DEFAULT ARRAY[]::TEXT[],
+  alternatives TEXT[] NOT NULL DEFAULT ARRAY[]::TEXT[],
+  rollback_plan TEXT NOT NULL,
+  assumptions TEXT[] NOT NULL DEFAULT ARRAY[]::TEXT[],
+  created_at_unix_millis BIGINT NOT NULL,
+  CHECK (proposal_id <> ''), CHECK (requested_by_principal_id <> ''), CHECK (stage = 'PROPOSED'), CHECK (desired_outcome <> ''), CHECK (smallest_change_layer <> ''), CHECK (risk_tier <> ''), CHECK (rollback_plan <> ''), CHECK (created_at_unix_millis > 0)
+);
+
+CREATE TABLE IF NOT EXISTS work.development_scaffolds (
+  scaffold_id TEXT PRIMARY KEY,
+  requested_by_principal_id TEXT NOT NULL,
+  stage TEXT NOT NULL,
+  proposal_ref TEXT NOT NULL,
+  base_repo_revision TEXT NOT NULL,
+  worktree_ref TEXT NOT NULL,
+  task_intent TEXT NOT NULL,
+  modules TEXT[] NOT NULL DEFAULT ARRAY[]::TEXT[],
+  allowed_paths TEXT[] NOT NULL DEFAULT ARRAY[]::TEXT[],
+  forbidden_paths TEXT[] NOT NULL DEFAULT ARRAY[]::TEXT[],
+  required_tests TEXT[] NOT NULL DEFAULT ARRAY[]::TEXT[],
+  artifact_lineage_refs TEXT[] NOT NULL DEFAULT ARRAY[]::TEXT[],
+  protected_main_access TEXT NOT NULL,
+  production_access TEXT NOT NULL,
+  secret_access TEXT NOT NULL,
+  created_at_unix_millis BIGINT NOT NULL,
+  CHECK (scaffold_id <> ''), CHECK (requested_by_principal_id <> ''), CHECK (stage = 'SANDBOXED'), CHECK (proposal_ref <> ''), CHECK (base_repo_revision <> ''), CHECK (worktree_ref <> ''), CHECK (task_intent <> ''), CHECK (protected_main_access = 'DENIED'), CHECK (production_access = 'DENIED'), CHECK (secret_access = 'DENIED'), CHECK (created_at_unix_millis > 0)
+);
+
+RESET ROLE;
+COMMIT;

database/migrations/044_work_graph_orchestration_snapshot.sql

@@ -0,0 +1,21 @@
+BEGIN;
+
+SET ROLE work_role;
+
+ALTER TABLE work.graphs ADD COLUMN IF NOT EXISTS orchestration_bound BOOLEAN NOT NULL DEFAULT FALSE;
+ALTER TABLE work.graphs ADD COLUMN IF NOT EXISTS orchestration_goal_id TEXT NOT NULL DEFAULT '';
+ALTER TABLE work.graphs ADD COLUMN IF NOT EXISTS orchestration_goal_state TEXT NOT NULL DEFAULT '';
+ALTER TABLE work.graphs ADD COLUMN IF NOT EXISTS orchestration_search_round INTEGER NOT NULL DEFAULT 0;
+ALTER TABLE work.graphs ADD COLUMN IF NOT EXISTS orchestration_max_round INTEGER NOT NULL DEFAULT 0;
+ALTER TABLE work.graphs ADD COLUMN IF NOT EXISTS orchestration_cancel_requested BOOLEAN NOT NULL DEFAULT FALSE;
+ALTER TABLE work.graphs ADD COLUMN IF NOT EXISTS orchestration_verification_ready BOOLEAN NOT NULL DEFAULT FALSE;
+ALTER TABLE work.graphs ADD COLUMN IF NOT EXISTS orchestration_unresolved BOOLEAN NOT NULL DEFAULT FALSE;
+
+ALTER TABLE work.graphs DROP CONSTRAINT IF EXISTS work_graphs_orchestration_snapshot_check;
+ALTER TABLE work.graphs ADD CONSTRAINT work_graphs_orchestration_snapshot_check CHECK (
+  (NOT orchestration_bound AND orchestration_goal_id = '' AND orchestration_search_round = 0 AND orchestration_max_round = 0)
+  OR (orchestration_bound AND orchestration_goal_id <> '' AND orchestration_search_round >= 1 AND orchestration_max_round >= 1)
+);
+
+RESET ROLE;
+COMMIT;

database/migrations/045_capability_model_qualification_observations.sql

@@ -0,0 +1,24 @@
+BEGIN;
+
+SET ROLE capability_role;
+
+CREATE TABLE IF NOT EXISTS capability.cognitive_model_qualification_observations (
+  observation_id TEXT PRIMARY KEY,
+  provider_instance_id TEXT NOT NULL,
+  provider_model_id TEXT NOT NULL,
+  status TEXT NOT NULL CHECK (status IN ('REACHABLE','RESPONDED_WITHOUT_QUALIFICATION','UNAVAILABLE')),
+  finish_reason TEXT NOT NULL,
+  latency_millis BIGINT NOT NULL CHECK (latency_millis >= 0),
+  output_observed BOOLEAN NOT NULL,
+  quality_assessed BOOLEAN NOT NULL DEFAULT FALSE CHECK (quality_assessed = FALSE),
+  category TEXT NOT NULL CHECK (category = 'UNRATED'),
+  observed_at_unix_millis BIGINT NOT NULL CHECK (observed_at_unix_millis > 0)
+);
+
+CREATE INDEX IF NOT EXISTS cognitive_model_qualification_lookup
+  ON capability.cognitive_model_qualification_observations(provider_instance_id, provider_model_id, observed_at_unix_millis DESC);
+
+GRANT SELECT, INSERT ON capability.cognitive_model_qualification_observations TO capability_role;
+
+RESET ROLE;
+COMMIT;

deploy/docker-compose.prod.yml

@@ -99,8 +99,12 @@ services:
       POSTGRES_PORT: "5432"
       POSTGRES_DATABASE: ${POSTGRES_DB:-assistant}
       WORK_POSTGRES_USER: work_role
-      WORK_POSTGRES_PASSWORD: ${WORK_DB_PASSWORD}
+      WORK_POSTGRES_CREDENTIAL_REF: credential://domovoy/work/db-password/v1
+      WORK_POSTGRES_PASSWORD_FILE: /run/secrets/work_db_password
+      SECRET_BACKEND_ROOT: /run/secrets
     depends_on: { postgres: { condition: service_healthy } }
+    volumes:
+      - assistant_core_secrets:/run/secrets:ro
 
   agent-worker:
     build: { context: .., dockerfile: deploy/images/runtime.Dockerfile, args: { TARGET: agent-worker, SOURCE_REVISION: "${ASSISTANT_SOURCE_REVISION:-unknown}" } }
@@ -114,10 +118,24 @@ services:
       WORK_POSTGRES_CREDENTIAL_REF: credential://domovoy/work/db-password/v1
       WORK_POSTGRES_PASSWORD_FILE: /run/secrets/work_db_password
       SECRET_BACKEND_ROOT: /run/secrets
-    depends_on: { postgres: { condition: service_healthy } }
+      TASK_VERIFIER_URL: http://task-verifier:8091
+    depends_on:
+      postgres: { condition: service_healthy }
+      task-verifier: { condition: service_healthy }
     volumes:
       - assistant_core_secrets:/run/secrets:ro
 
+  task-verifier:
+    build: { context: .., dockerfile: deploy/images/runtime.Dockerfile, args: { TARGET: task-verifier, SOURCE_REVISION: "${ASSISTANT_SOURCE_REVISION:-unknown}" } }
+    restart: unless-stopped
+    environment:
+      LISTEN_ADDR: :8091
+    healthcheck:
+      test: ["CMD-SHELL", "wget -qO- http://localhost:8091/healthz >/dev/null"]
+      interval: 10s
+      timeout: 3s
+      retries: 10
+
   work-graph-runtime:
     build: { context: .., dockerfile: deploy/images/runtime.Dockerfile, args: { TARGET: work-graph-runtime, SOURCE_REVISION: "${ASSISTANT_SOURCE_REVISION:-unknown}" } }
     restart: unless-stopped
@@ -211,11 +229,18 @@ services:
       LISTEN_ADDR: ":8083"
       CORE_RUNTIME_URL: http://core-runtime: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-api:8080/v1}
+      LOCAL_MEDIUM_PROVIDER_API_KEY_FILE: ${LOCAL_MEDIUM_PROVIDER_API_KEY_FILE:-}
+      LOCAL_MEDIUM_PROVIDER_CREDENTIAL_REF: ${LOCAL_MEDIUM_PROVIDER_CREDENTIAL_REF:-credential://domovoy/local-medium/provider-api-key/v1}
       LOCAL_LIFEBOAT_PROVIDER_URL: http://cognitive-lifeboat:8086/v1
-      LOCAL_GPU_PROVIDER_URL: http://192.168.220.91:11434
       COGNITIVE_OWNER_PRINCIPAL_ID: domovoy-creator-principal
       COGNITIVE_ALLOW_CANARY: "true"
     depends_on: [core-runtime, cognitive-lifeboat]
+    volumes:
+      - assistant_core_secrets:/run/secrets:ro
+    networks: [default, llm-local]
 
   cognitive-lifeboat:
     image: ghcr.io/ggml-org/llama.cpp:server
@@ -282,3 +307,8 @@ volumes:
     external: true
   assistant_research_sources:
     name: assistant_research_sources
+
+networks:
+  llm-local:
+    external: true
+    name: llm-local_default

dev/ACTIVE_TASK.json

@@ -1,5 +1,5 @@
 {
-  "task_file": "dev/tasks/DEV-000152-cognitive-provider-secret-reference.json",
-  "task_id": "DEV-000152",
+  "task_file": "dev/tasks/DEV-000247-research-runtime-status.json",
+  "task_id": "DEV-000247",
   "task_revision": 1
 }

dev/tasks/DEV-000153-fleet-secret-wiring.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000153",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0010",
+  "base_repo_revision": "dbbc54171d9c8a9edaa93e5ff5870d3ec434d799",
+  "risk_tier": "E",
+  "intent": "Wire the production cognitive-worker to the existing protected secret volume for optional local CPU gateway authentication without placing credential material in compose environment values.",
+  "requirement_ids": ["AUDIT-WP-003", "AUDIT-WP-201", "CONSTITUTION-1.3.28"],
+  "allowed_change_scope": {"modules": ["cognitive"], "paths": ["deploy/docker-compose.prod.yml", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000153-fleet-secret-wiring.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/devctl/**", ".gitea/**", "database/**", "contracts/**", "modules/**", "adapters/**"]},
+  "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": ["assistant_core_secrets read-only volume", "optional LOCAL_CPU_PROVIDER_API_KEY_FILE reference"], "may_change": []},
+  "semantic_owner": "capability",
+  "canonical_state": ["provider credential material remains outside repository and compose environment"],
+  "state_transitions": ["when operators provision the referenced file, cognitive-worker resolves it and authenticates the local CPU gateway; when absent, existing recovery behavior remains available"],
+  "side_effects_and_privileges": ["read-only mount only; no secret value is added to compose"],
+  "persistence_and_migrations": ["no database migration"],
+  "failure_retry_restart": ["empty optional file reference preserves current unauthenticated startup", "configured missing credential file fails closed at worker startup"],
+  "assumptions": ["the gateway credential is provisioned by production operators into assistant_core_secrets; this task does not copy or reveal it"],
+  "acceptance_scenarios": ["compose renders with empty optional key path", "cognitive-worker receives read-only /run/secrets mount", "no literal provider credential occurs in repository or compose", "CI and production deployment succeed"],
+  "required_tests": ["go test", "go vet", "docker compose config", "architecture preflight", "check-diff", "git diff --check", "independent", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT", "BLOCKED_BY_ARCHITECTURE_GAP"]
+}

dev/tasks/DEV-000154-fast-live-smoke-evidence.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000154",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0010",
+  "base_repo_revision": "283f5c044c3bf6739c8ed859412cd736754e3177",
+  "risk_tier": "E",
+  "intent": "Record bounded production evidence that the authenticated local CPU provider is available and invokable through the cognitive boundary.",
+  "requirement_ids": ["AUDIT-WP-003", "AUDIT-WP-201", "WP-601"],
+  "allowed_change_scope": {"modules": ["capability", "cognitive"], "paths": ["docs/evidence/wp-601-fast-gateway-live-smoke-2026-08-22-v1.json", "docs/evidence/WP-601-FAST-GATEWAY-LIVE-SMOKE-2026-08-22.md", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000154-fast-live-smoke-evidence.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/devctl/**", ".gitea/**", "database/**", "deploy/**", "contracts/**", "modules/**", "adapters/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["production cognitive discovery and bounded advisory response evidence"], "may_change": []},
+  "semantic_owner": "capability",
+  "canonical_state": ["none; append-only evidence artifact only"],
+  "state_transitions": ["none; evidence does not activate or promote a binding"],
+  "side_effects_and_privileges": ["four bounded local-only synthetic advisory requests"],
+  "persistence_and_migrations": ["no database migration"],
+  "failure_retry_restart": ["not applicable to evidence artifact"],
+  "assumptions": ["provider key remains in protected production volume and is not copied to repository or logs"],
+  "acceptance_scenarios": ["evidence records available API_KEY provider", "four requests complete with expected sentinels", "limitations remain explicit", "no secret material is present"],
+  "required_tests": ["JSON validation", "go test", "go vet", "architecture preflight", "check-diff", "git diff --check", "independent", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT", "BLOCKED_BY_ARCHITECTURE_GAP"]
+}

dev/tasks/DEV-000155-current-audit-status.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000155",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0010",
+  "base_repo_revision": "5703ab3ed5e22a5424a2db7f19886c0305e54280",
+  "risk_tier": "E",
+  "intent": "Synchronize the project status projection with current live audit evidence, explicitly distinguishing completed runtime foundations from external blockers.",
+  "requirement_ids": ["AUDIT-TRUTHFUL-STATUS", "WP-003", "WP-301", "WP-800"],
+  "allowed_change_scope": {"modules": [], "paths": ["docs/PROJECT_STATUS.md", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000155-current-audit-status.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/devctl/**", ".gitea/**", "database/**", "deploy/**", "contracts/**", "modules/**", "adapters/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["live audit observations and committed evidence artifacts"], "may_change": []},
+  "semantic_owner": "none; documentation projection",
+  "canonical_state": ["none"],
+  "state_transitions": ["none"],
+  "side_effects_and_privileges": ["none"],
+  "persistence_and_migrations": ["no database migration"],
+  "failure_retry_restart": ["not applicable"],
+  "assumptions": ["status projection remains subordinate to runtime evidence and does not declare blocked items complete"],
+  "acceptance_scenarios": ["current revision and FAST evidence are referenced", "Medium/Heavy and bridge gaps are explicit", "UI acceptance remains pending without session", "no secret value appears"],
+  "required_tests": ["JSON validation", "go test", "go vet", "architecture preflight", "check-diff", "git diff --check", "independent", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT", "BLOCKED_BY_ARCHITECTURE_GAP"]
+}

dev/tasks/DEV-000156-fleet-execution-status-evidence.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000156",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "03c3484ee1488ebc3b2349ab6ba5c8005c895110",
+  "risk_tier": "E",
+  "intent": "Capture fresh production evidence for model fleet readiness and Agent Execution capability without changing runtime behavior.",
+  "requirement_ids": ["WP-200", "WP-301", "AUDIT-TRUTHFUL-STATUS"],
+  "allowed_change_scope": {"modules": [], "paths": ["docs/evidence/**", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000156-fleet-execution-status-evidence.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/**", "modules/**", "contracts/**", "adapters/**", "database/**", "deploy/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["fleet readiness response", "agent-worker status response"], "may_change": []},
+  "semantic_owner": "none; evidence projection",
+  "canonical_state": [],
+  "state_transitions": ["evidence only; no activation, promotion or configuration change"],
+  "side_effects_and_privileges": ["bounded read-only production status probes"],
+  "persistence_and_migrations": ["immutable JSON evidence artifact only"],
+  "failure_retry_restart": ["probe failures are recorded as observed facts; no retry-dependent claim"],
+  "assumptions": ["all probes are internal read-only GET requests", "no secret values are collected"],
+  "acceptance_scenarios": ["FAST readiness auth mismatch is distinguished from successful bounded gateway smoke", "MEDIUM remains NOT_CONFIGURED", "HEAVY timeout is recorded", "LIFEBOAT is the selected fallback", "Agent Execution is recovery-only", "no secret value appears"],
+  "required_tests": ["JSON validation", "go test", "go vet", "preflight", "check-diff", "git diff --check", "independent", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000157-postgres-restore-evidence.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000157",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0010",
+  "base_repo_revision": "077f16e50ca22e14dbf00214005f0fe98a1c8308",
+  "risk_tier": "E",
+  "intent": "Capture fresh production PostgreSQL backup checksum and isolated restore verification evidence without changing the production database.",
+  "requirement_ids": ["WP-002", "AUDIT-RECOVERY-TRUTH"],
+  "allowed_change_scope": {"modules": [], "paths": ["docs/evidence/**", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000157-postgres-restore-evidence.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/**", "modules/**", "contracts/**", "adapters/**", "database/**", "deploy/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["backup artifact format", "restore verification output"], "may_change": []},
+  "semantic_owner": "none; evidence projection",
+  "canonical_state": [],
+  "state_transitions": ["evidence only; no production database mutation"],
+  "side_effects_and_privileges": ["isolated temporary restore container created and removed by the existing verifier"],
+  "persistence_and_migrations": ["immutable JSON evidence artifact only"],
+  "failure_retry_restart": ["checksum or restore failure is recorded as failure; no blind retry claim"],
+  "assumptions": ["the standard verifier uses tmpfs and cleans up its temporary container", "backup contents and credentials are not copied into repository evidence"],
+  "acceptance_scenarios": ["latest backup directory is selected", "checksums pass", "isolated restore reports expected schema/table counts", "production database remains untouched"],
+  "required_tests": ["JSON validation", "go test", "go vet", "preflight", "check-diff", "git diff --check", "independent", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000158-status-evidence-consolidation.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000158",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0010",
+  "base_repo_revision": "3a52c18bef7e29a0369ad1edf05adeed1be5935e",
+  "risk_tier": "E",
+  "intent": "Synchronize the project status projection with the latest production revision and committed fleet/recovery evidence.",
+  "requirement_ids": ["AUDIT-TRUTHFUL-STATUS", "WP-002", "WP-200", "WP-301"],
+  "allowed_change_scope": {"modules": [], "paths": ["docs/PROJECT_STATUS.md", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000158-status-evidence-consolidation.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/**", "modules/**", "contracts/**", "adapters/**", "database/**", "deploy/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["production revision", "committed fleet evidence", "committed restore evidence"], "may_change": []},
+  "semantic_owner": "none; documentation projection",
+  "canonical_state": [],
+  "state_transitions": ["evidence projection only"],
+  "side_effects_and_privileges": ["none"],
+  "persistence_and_migrations": ["no database migration"],
+  "failure_retry_restart": ["not applicable"],
+  "assumptions": ["status remains subordinate to live evidence and does not declare unavailable capabilities complete"],
+  "acceptance_scenarios": ["latest production revision is shown", "fleet and restore evidence are linked", "remaining gaps remain explicit", "no secret value appears"],
+  "required_tests": ["JSON validation", "go test", "go vet", "preflight", "check-diff", "git diff --check", "independent", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000159-agent-bridge-config-wiring.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000159",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "e1186757bfcf9a349ce0ced603f4a4ebdad636fe",
+  "risk_tier": "E",
+  "intent": "Expose the already-supported optional Agent Execution bridge URL to the production agent-worker without enabling it when unset.",
+  "requirement_ids": ["WP-301", "Agent Execution runtime and recovery"],
+  "allowed_change_scope": {"modules": ["deployment", "work-agent-execution"], "paths": ["deploy/docker-compose.prod.yml", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000159-agent-bridge-config-wiring.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "contracts/**", "cmd/**", "modules/**", "adapters/**", ".gitea/**"]},
+  "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": ["AGENT_EXECUTION_BRIDGE_URL deployment variable", "agent-worker optional bridge behavior"], "may_change": []},
+  "semantic_owner": "deployment projection; work-agent-execution remains owner",
+  "canonical_state": ["none"],
+  "state_transitions": ["unset URL -> recovery-only; configured URL -> existing typed HTTP executor path"],
+  "side_effects_and_privileges": ["deployment privileged-surface review is declared; no effective new privilege; no credential forwarding; no outbound call when variable is empty"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["existing worker fail-closed and stale recovery behavior remains unchanged"],
+  "assumptions": ["an independently operated bridge is still required before enabling the variable in production", "empty default preserves current recovery-only behavior"],
+  "acceptance_scenarios": ["production compose renders with empty bridge URL", "agent-worker remains recovery-only when unset", "configured URL is available to the process without source secret", "no bridge runtime is fabricated"],
+  "required_tests": ["JSON validation", "docker compose config", "go test", "go vet", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000160-status-after-bridge-wiring.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000160",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0010",
+  "base_repo_revision": "ada2da0c6c93d630f7699bd05abd32031c11544c",
+  "risk_tier": "E",
+  "intent": "Synchronize project status with the latest bridge wiring deployment and preserve the explicit external capability gap.",
+  "requirement_ids": ["AUDIT-TRUTHFUL-STATUS", "WP-301"],
+  "allowed_change_scope": {"modules": [], "paths": ["docs/PROJECT_STATUS.md", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000160-status-after-bridge-wiring.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/**", "modules/**", "contracts/**", "adapters/**", "database/**", "deploy/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["production revision", "optional bridge configuration state"], "may_change": []},
+  "semantic_owner": "none; documentation projection",
+  "canonical_state": [],
+  "state_transitions": ["evidence projection only"],
+  "side_effects_and_privileges": ["none"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["not applicable"],
+  "assumptions": ["unset bridge URL remains truthful recovery-only state"],
+  "acceptance_scenarios": ["current production revision is shown", "bridge wiring and missing runtime are distinguished", "no secret value appears"],
+  "required_tests": ["JSON validation", "go test", "go vet", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000161-completion-audit-matrix.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000161",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0010",
+  "base_repo_revision": "3cf44ada8afd75fb008cf8d5965caf9c3f5b0462",
+  "risk_tier": "E",
+  "intent": "Publish a requirement-by-requirement completion matrix for the audit objective, distinguishing repository proof, production proof and external gaps.",
+  "requirement_ids": ["AUDIT-COMPLETION-TRUTH", "WP-301", "WP-601", "WP-800"],
+  "allowed_change_scope": {"modules": [], "paths": ["docs/evidence/**", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000161-completion-audit-matrix.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/**", "modules/**", "contracts/**", "adapters/**", "database/**", "deploy/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["audit roadmap", "committed task packets", "production evidence artifacts"], "may_change": []},
+  "semantic_owner": "none; audit evidence projection",
+  "canonical_state": [],
+  "state_transitions": ["evidence projection only"],
+  "side_effects_and_privileges": ["none"],
+  "persistence_and_migrations": ["immutable JSON evidence artifact only"],
+  "failure_retry_restart": ["not applicable"],
+  "assumptions": ["unknown or externally blocked requirements are not marked complete", "matrix claims are limited to cited evidence scope"],
+  "acceptance_scenarios": ["all seven objective areas have a status", "production and repository proof are distinguished", "remaining external gaps are explicit", "no secret value appears"],
+  "required_tests": ["JSON validation", "go test", "go vet", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000162-status-after-completion-matrix.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000162",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0010",
+  "base_repo_revision": "bda71caecece67e5ef225ac250a7a80c4bc84d94",
+  "risk_tier": "E",
+  "intent": "Refresh current production revision in project status and completion evidence after the audit matrix deployment.",
+  "requirement_ids": ["AUDIT-TRUTHFUL-STATUS"],
+  "allowed_change_scope": {"modules": [], "paths": ["docs/PROJECT_STATUS.md", "docs/evidence/audit-completion-matrix-2026-08-22-v1.json", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000162-status-after-completion-matrix.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/**", "modules/**", "contracts/**", "adapters/**", "database/**", "deploy/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["current production revision", "completion matrix evidence"], "may_change": []},
+  "semantic_owner": "none; documentation projection",
+  "canonical_state": [],
+  "state_transitions": ["evidence projection only"],
+  "side_effects_and_privileges": ["none"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["not applicable"],
+  "assumptions": ["matrix remains NOT_COMPLETE until external gaps are actually verified"],
+  "acceptance_scenarios": ["status and matrix point to bda71ca", "no completion claim is widened", "no secret value appears"],
+  "required_tests": ["JSON validation", "go test", "go vet", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000163-operations-ui-contract-refresh.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000163",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "c25a6346d0615f66ddb411450fcc7a36c62d2b48",
+  "risk_tier": "E",
+  "intent": "Extend authenticated cabinet contract tests to cover truthful model-fleet and Agent Execution status surfaces already implemented in Creator Web.",
+  "requirement_ids": ["WP-801", "WP-802", "WP-805", "Authenticated cabinet acceptance"],
+  "allowed_change_scope": {"modules": ["creator-web"], "paths": ["tests/creator_web/cabinet_acceptance_contract_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000163-operations-ui-contract-refresh.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "deploy/**", "database/**", "cmd/**", "contracts/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["existing authenticated Creator Web status routes and static cabinet"], "may_change": ["acceptance tests only"]},
+  "semantic_owner": "creator-web",
+  "canonical_state": ["none; tests inspect existing UI contract"],
+  "state_transitions": ["authenticated session gates truthful status requests"],
+  "side_effects_and_privileges": ["no external mutation; no credentials embedded"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["missing status surface fails CI"],
+  "assumptions": ["live authenticated browser visual acceptance remains a separate production verification"],
+  "acceptance_scenarios": ["fleet and executor status controls exist", "server exposes authenticated routes", "console renders actual selected tier and worker mode", "responsive/reduced-motion tests remain active"],
+  "required_tests": ["go test ./...", "go vet ./...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_AUTHENTICATED_BROWSER_SESSION", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000164-authorized-fleet-readiness.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000164",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "65a60d5f068e3d8f70733c4ee96bfbb9d9943d5e",
+  "risk_tier": "E",
+  "intent": "Make live model-fleet readiness use the existing file-backed CPU provider API key so FAST status reflects the authenticated provider path.",
+  "requirement_ids": ["WP-601", "Model fleet/failover"],
+  "allowed_change_scope": {"modules": ["cognitive-worker", "capability"], "paths": ["cmd/cognitive-worker/main.go", "cmd/cognitive-worker/fleet_status.go", "cmd/cognitive-worker/fleet_status_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000164-authorized-fleet-readiness.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "contracts/**", "adapters/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["existing provider API key resolver", "fleet readiness response"], "may_change": ["fleet readiness probe request headers"]},
+  "semantic_owner": "capability/cognitive-worker",
+  "canonical_state": [],
+  "state_transitions": ["authenticated provider probe -> truthful FAST readiness"],
+  "side_effects_and_privileges": ["uses existing in-memory resolved key only for the internal provider probe; no key logging or persistence"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["401 remains a typed unavailable result when key is absent or rejected; GPU/Medium behavior unchanged"],
+  "assumptions": ["OpenAI-compatible local provider uses Bearer authentication as the existing model adapter does"],
+  "acceptance_scenarios": ["fleet readiness sends Bearer header when CPU key is configured", "FAST is ready when authenticated /models succeeds", "no Authorization header is sent without a key", "Lifeboat fallback remains unchanged"],
+  "required_tests": ["go test ./...", "go vet ./...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production fleet readiness"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000165-authorized-cpu-health-probe.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000165",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "5d5afc88883c7f79e3b051c790642e41386666ec",
+  "risk_tier": "E",
+  "intent": "Use the existing file-backed CPU provider API key in cognitive-worker readiness probes so /readyz and fleet readiness report consistent provider state.",
+  "requirement_ids": ["WP-601", "Model fleet/failover"],
+  "allowed_change_scope": {"modules": ["cognitive-worker", "platform-health"], "paths": ["runtime/health/server.go", "runtime/health/server_test.go", "cmd/cognitive-worker/main.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000165-authorized-cpu-health-probe.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "contracts/**", "adapters/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["existing health Probe contract", "existing provider API key resolver"], "may_change": ["health probe header support"]},
+  "semantic_owner": "cognitive-worker/platform health",
+  "canonical_state": [],
+  "state_transitions": ["authenticated CPU probe -> consistent readiness evidence"],
+  "side_effects_and_privileges": ["resolved key is used only in outbound probe header; never logged or persisted"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["provider timeout or rejected key remains degraded; no fallback behavior changed"],
+  "assumptions": ["the local provider uses Bearer authentication"],
+  "acceptance_scenarios": ["health probe sends configured header", "cognitive /readyz uses authenticated CPU probe", "GPU remains optional", "no key is sent when absent"],
+  "required_tests": ["go test ./...", "go vet ./...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production /readyz and fleet readiness"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000166-per-provider-readiness-timeouts.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000166",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "4086b7522460a78afc341d996ca7874505018871",
+  "risk_tier": "E",
+  "intent": "Give the authenticated CPU readiness probe the provider adapter timeout while retaining short opportunistic GPU probing.",
+  "requirement_ids": ["WP-601", "Model fleet/failover"],
+  "allowed_change_scope": {"modules": ["cognitive-worker", "platform-health"], "paths": ["runtime/health/server.go", "runtime/health/server_test.go", "cmd/cognitive-worker/main.go", "cmd/cognitive-worker/fleet_status.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000166-per-provider-readiness-timeouts.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "contracts/**", "adapters/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["existing health Probe contract", "existing fleet readiness probe"], "may_change": ["per-probe timeout metadata"]},
+  "semantic_owner": "cognitive-worker/platform health",
+  "canonical_state": [],
+  "state_transitions": ["provider-specific probe deadline -> truthful readiness result"],
+  "side_effects_and_privileges": ["no new privilege; existing provider auth only"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["CPU gets bounded 10-second probe; GPU retains 2-second opportunistic timeout"],
+  "assumptions": ["CPU provider catalog may take longer than opportunistic probes while remaining within adapter bounds"],
+  "acceptance_scenarios": ["per-probe health timeout is honored", "CPU fleet client uses 10 seconds", "GPU remains bounded", "Lifeboat fallback remains unchanged"],
+  "required_tests": ["go test ./...", "go vet ./...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production /readyz and fleet readiness"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000167-cpu-readiness-client-timeout.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000167",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "c7b57daaf8233b0f0e79c77c02e92040378921d5",
+  "risk_tier": "E",
+  "intent": "Use the bounded CPU provider HTTP client for the authenticated readiness probe so /readyz reports the same truthful result as fleet readiness.",
+  "requirement_ids": ["WP-601", "Model fleet/failover"],
+  "allowed_change_scope": {"modules": ["cognitive-worker"], "paths": ["cmd/cognitive-worker/main.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000167-cpu-readiness-client-timeout.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "contracts/**", "adapters/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["existing readiness Probe contract", "existing authenticated CPU fleet endpoint"], "may_change": []},
+  "semantic_owner": "cognitive-worker",
+  "canonical_state": [],
+  "state_transitions": ["CPU provider response -> consistent /readyz and fleet readiness state"],
+  "side_effects_and_privileges": ["no new privilege; existing provider auth only"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["CPU readiness remains bounded at 10 seconds; opportunistic GPU remains 2 seconds"],
+  "assumptions": ["The CPU provider is authenticated with the existing file-backed API key"],
+  "acceptance_scenarios": ["/readyz can report CPU AVAILABLE when fleet readiness reports CPU AVAILABLE", "GPU remains bounded", "Lifeboat fallback remains unchanged"],
+  "required_tests": ["go test ./...", "go vet ./...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production /readyz and fleet readiness"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000168-configurable-medium-fleet-slot.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000168",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "205d8f620603ccf3e0138f6beb93156450664314",
+  "risk_tier": "E",
+  "intent": "Make the MEDIUM model-fleet slot a real optional OpenAI-compatible provider path with protected file credentials, discovery, readiness and gateway registration.",
+  "requirement_ids": ["WP-200", "WP-301", "Model fleet/failover"],
+  "allowed_change_scope": {"modules": ["cognitive-worker", "deployment-policy", "tests"], "paths": ["cmd/cognitive-worker/main.go", "cmd/cognitive-worker/main_test.go", "deploy/docker-compose.prod.yml", "tests/cognitive_advisory/cognitive_worker_env_policy.ps1", "tests/cognitive_advisory/cognitive_worker_env_policy_test.ps1", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000168-configurable-medium-fleet-slot.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "contracts/**", "adapters/**", ".gitea/**"]},
+  "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 provider adapter contract", "existing discovery registration", "existing fleet readiness contract", "existing CredentialRef file contract"], "may_change": []},
+  "semantic_owner": "cognitive-worker/capability registration boundary",
+  "canonical_state": [],
+  "state_transitions": ["MEDIUM endpoint absent -> NOT_CONFIGURED", "MEDIUM endpoint configured -> discovery/readiness registration", "available provider -> eligible gateway binding subject to core registry policy"],
+  "side_effects_and_privileges": ["declares one new deployment configuration surface for an existing read-only secret volume; provider key is resolved only through the existing protected file CredentialRef"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["medium probe has a bounded 10-second client timeout; absent endpoint remains NOT_CONFIGURED; fallback selection remains Lifeboat when preferred tiers are unavailable"],
+  "assumptions": ["A real medium OpenAI-compatible runtime may be supplied later through deployment environment injection", "No production medium runtime is claimed by this change"],
+  "acceptance_scenarios": ["empty medium URL produces NOT_CONFIGURED without startup failure", "configured medium URL is registered for discovery and fleet readiness", "direct medium secret environment is rejected", "medium CredentialRef file is accepted", "CPU/GPU/Lifeboat behavior remains unchanged"],
+  "required_tests": ["go test ./...", "go vet ./...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "environment policy tests", "CI", "production fleet readiness"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000169-optional-medium-readiness.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000169",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "8f3920519b6a228d78b6ad775a7c19838dfbb70b",
+  "risk_tier": "E",
+  "intent": "Keep an absent optional MEDIUM provider out of runtime readiness while preserving its explicit NOT_CONFIGURED fleet member.",
+  "requirement_ids": ["WP-200", "WP-301", "Model fleet/failover"],
+  "allowed_change_scope": {"modules": ["cognitive-worker"], "paths": ["cmd/cognitive-worker/main.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000169-optional-medium-readiness.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "contracts/**", "adapters/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["existing readiness handler", "existing fleet readiness contract"], "may_change": []},
+  "semantic_owner": "cognitive-worker",
+  "canonical_state": [],
+  "state_transitions": ["MEDIUM absent -> fleet NOT_CONFIGURED without runtime readiness degradation", "MEDIUM configured -> readiness probe included"],
+  "side_effects_and_privileges": ["none"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["configured MEDIUM remains bounded by its 10-second client; absent MEDIUM is not probed"],
+  "assumptions": ["Fleet status remains the explicit inventory surface for unconfigured optional tiers"],
+  "acceptance_scenarios": ["production /readyz stays ready when MEDIUM URL is empty", "fleet readiness retains MEDIUM NOT_CONFIGURED", "configured MEDIUM is included in /readyz"],
+  "required_tests": ["go test ./...", "go vet ./...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production /readyz and fleet readiness"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000170-refresh-fleet-audit-evidence.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000170",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "51427a246e1ccc115a6c2640790d9220beca95e9",
+  "risk_tier": "E",
+  "intent": "Refresh audit completion and production fleet evidence after the configurable MEDIUM slot and optional readiness fixes were deployed.",
+  "requirement_ids": ["AUDIT-COMPLETION-TRUTH", "WP-200", "WP-301"],
+  "allowed_change_scope": {"modules": [], "paths": ["docs/evidence/audit-completion-matrix-2026-08-22-v1.json", "docs/evidence/wp-200-wp-301-fleet-execution-status-2026-08-22-v2.json", "docs/PROJECT_STATUS.md", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000170-refresh-fleet-audit-evidence.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/**", "modules/**", "contracts/**", "adapters/**", "database/**", "deploy/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["live fleet readiness evidence", "audit completion matrix"], "may_change": []},
+  "semantic_owner": "none; audit evidence projection",
+  "canonical_state": [],
+  "state_transitions": ["stale evidence -> current production evidence"],
+  "side_effects_and_privileges": ["none"],
+  "persistence_and_migrations": ["immutable JSON evidence artifacts only"],
+  "failure_retry_restart": ["not applicable"],
+  "assumptions": ["live probe output was captured after production deployment of 51427a24", "missing external capabilities remain explicitly incomplete"],
+  "acceptance_scenarios": ["production revision is current", "MEDIUM configuration and readiness semantics are recorded", "agent bridge and authenticated UI gaps remain visible", "no secret value appears"],
+  "required_tests": ["JSON validation", "go test ./...", "go vet ./...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000171-connect-existing-medium-runtime.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000171",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "503fe840741ac99499e06bd93df977acf1baad41",
+  "risk_tier": "E",
+  "intent": "Connect cognitive-worker to the already-operated llm-local CPU model runtime as the configured MEDIUM fleet member through an explicit external Docker network.",
+  "requirement_ids": ["WP-200", "WP-301", "Model fleet/failover"],
+  "allowed_change_scope": {"modules": ["deployment-topology", "cognitive-worker"], "paths": ["deploy/docker-compose.prod.yml", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000171-connect-existing-medium-runtime.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "contracts/**", "adapters/**", ".gitea/**"]},
+  "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 OpenAI-compatible local provider adapter", "existing MEDIUM fleet readiness contract", "existing llm-local Docker network"], "may_change": []},
+  "semantic_owner": "deployment topology/cognitive-worker registration",
+  "canonical_state": [],
+  "state_transitions": ["MEDIUM unconfigured -> live provider discovery/readiness", "preferred HEAVY unavailable -> highest available fallback selection"],
+  "side_effects_and_privileges": ["one explicit read-only network attachment from cognitive-worker to existing llm-local runtime; no secret or database privilege added"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["if external llm-local network is absent deployment fails closed; provider timeout remains bounded; Lifeboat remains fallback"],
+  "assumptions": ["production llm-local_default network and llm-local-api are independently operated and present", "llm-local-api exposes OpenAI-compatible /v1/models and chat routes on port 8080"],
+  "acceptance_scenarios": ["compose config validates with external network declaration", "cognitive-worker resolves llm-local-api", "MEDIUM fleet member is AVAILABLE live", "runtime /readyz remains 200", "Lifeboat fallback remains available", "no secret value is added"],
+  "required_tests": ["docker compose config", "go test ./...", "go vet ./...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production /readyz and fleet readiness"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000172-fleet-failover-priority.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000172",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "2c7a85779b530e3147fbf47b71742759cc010d58",
+  "risk_tier": "E",
+  "intent": "Make fleet failover choose the best available non-Lifeboat tier below the preferred tier before using Lifeboat.",
+  "requirement_ids": ["WP-301", "Model fleet/failover"],
+  "allowed_change_scope": {"modules": ["capability", "cognitive-worker"], "paths": ["modules/capability/domain/fleet.go", "modules/capability/domain/fleet_test.go", "cmd/cognitive-worker/fleet_status_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000172-fleet-failover-priority.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "contracts/**", "adapters/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["existing FleetReadiness contract", "existing FleetTier selection contract"], "may_change": []},
+  "semantic_owner": "capability",
+  "canonical_state": [],
+  "state_transitions": ["preferred unavailable -> next lower ready tier -> Lifeboat fallback"],
+  "side_effects_and_privileges": ["none"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["selection remains deterministic from observed readiness; invalid preferred tier fails closed"],
+  "assumptions": ["HEAVY > MEDIUM > FAST is the quality/capability fallback order; LIFEBOAT is last-resort recovery"],
+  "acceptance_scenarios": ["HEAVY unavailable and MEDIUM ready selects MEDIUM", "HEAVY and MEDIUM unavailable and FAST ready selects FAST", "all primary tiers unavailable selects Lifeboat", "invalid preferred tier is rejected"],
+  "required_tests": ["go test ./...", "go vet ./...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production fleet readiness"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000173-refresh-medium-failover-evidence.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000173",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "64642f4baed28f9051efe6a269769c5838b547e4",
+  "risk_tier": "E",
+  "intent": "Refresh audit and production evidence after live MEDIUM runtime availability and deterministic fleet failover priority are deployed.",
+  "requirement_ids": ["AUDIT-COMPLETION-TRUTH", "WP-200", "WP-301"],
+  "allowed_change_scope": {"modules": [], "paths": ["docs/evidence/audit-completion-matrix-2026-08-22-v1.json", "docs/evidence/wp-200-wp-301-fleet-execution-status-2026-08-22-v3.json", "docs/PROJECT_STATUS.md", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000173-refresh-medium-failover-evidence.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/**", "modules/**", "contracts/**", "adapters/**", "database/**", "deploy/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["live fleet readiness evidence", "audit completion matrix"], "may_change": []},
+  "semantic_owner": "none; audit evidence projection",
+  "canonical_state": [],
+  "state_transitions": ["stale evidence -> current MEDIUM/failover evidence"],
+  "side_effects_and_privileges": ["none"],
+  "persistence_and_migrations": ["immutable JSON evidence artifacts only"],
+  "failure_retry_restart": ["not applicable"],
+  "assumptions": ["live probes were captured after production deployment of 64642f4b", "availability is not interpreted as qualification or activation"],
+  "acceptance_scenarios": ["production revision is current", "MEDIUM model catalog and readiness are recorded", "preferred HEAVY falls back to MEDIUM", "remaining bridge/GPU/browser gaps remain explicit", "no secret value appears"],
+  "required_tests": ["JSON validation", "go test ./...", "go vet ./...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000174-medium-qualification-evidence.json

@@ -0,0 +1,32 @@
+{
+  "task_id": "DEV-000174",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "6e0d1a227918775b906e3462406c1505813f51c6",
+  "risk_tier": "E",
+  "intent": "Record the bounded production qualification result for the reachable MEDIUM model without changing lifecycle state.",
+  "requirement_ids": ["AUDIT-COMPLETION-TRUTH", "WP-200", "WP-301"],
+  "allowed_change_scope": {
+    "modules": [],
+    "paths": [
+      "docs/evidence/wp-200-medium-qualification-2026-08-22.json",
+      "docs/evidence/audit-completion-matrix-2026-08-22-v1.json",
+      "docs/PROJECT_STATUS.md",
+      "dev/ACTIVE_TASK.json",
+      "dev/tasks/DEV-000174-medium-qualification-evidence.json"
+    ]
+  },
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/**", "modules/**", "contracts/**", "adapters/**", "database/**", "deploy/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["creator-advisory synthetic benchmark", "production cognitive registry revisions"], "may_change": []},
+  "semantic_owner": "none; audit evidence projection",
+  "canonical_state": [],
+  "state_transitions": ["unqualified MEDIUM availability -> recorded qualification result"],
+  "side_effects_and_privileges": ["read-only registry lookup and one-shot benchmark invocation; no lifecycle mutation"],
+  "persistence_and_migrations": ["immutable JSON evidence artifacts only"],
+  "failure_retry_restart": ["benchmark timeout is recorded as UNSUPPORTED; no retry changes lifecycle state"],
+  "assumptions": ["the benchmark binary used the repository cognitive adapter and the production llm-local network", "binding and model revisions were obtained read-only from production registry"],
+  "acceptance_scenarios": ["all four synthetic cases are recorded", "timeout and unsupported result are explicit", "no canary or activation is performed", "no secret value appears", "audit matrix retains qualification gap"],
+  "required_tests": ["JSON validation", "go test ./...", "go vet ./...", "go run ./cmd/devctl preflight -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-diff -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-architecture -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl/independent", "git diff --check", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000175-medium-invocation-timeout.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000175",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "b29f8775e1217506098d5c9a0e99c1046494cba3",
+  "risk_tier": "E",
+  "intent": "Make local OpenAI-compatible provider invocation deadlines explicit per provider so slow but reachable fleet members can be qualified through the existing adapter path.",
+  "requirement_ids": ["WP-200", "WP-301", "AUDIT-COMPLETION-TRUTH"],
+  "allowed_change_scope": {"modules": ["cognitive adapter", "cognitive worker", "benchmark runtime"], "paths": ["adapters/models/openai_compatible_local.go", "adapters/models/openai_compatible_local_test.go", "cmd/cognitive-worker/main.go", "cmd/cognitive-benchmark/main.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000175-medium-invocation-timeout.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "contracts/**", "modules/**", "adapters/postgres/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["existing provider-neutral cognitive adapter contract", "existing provider environment configuration"], "may_change": ["provider-specific invocation timeout configuration"]},
+  "semantic_owner": "cognitive adapter/runtime",
+  "canonical_state": ["adapter request deadline"],
+  "state_transitions": ["fixed local adapter deadline -> explicit positive provider-specific deadline"],
+  "side_effects_and_privileges": ["none beyond existing provider requests"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["non-positive configuration falls back to safe default; request context remains authoritative"],
+  "assumptions": ["the existing Medium runtime is reachable but may require more than 10 seconds for a complete response", "default CPU behavior remains 10 seconds unless separately configured"],
+  "acceptance_scenarios": ["adapter accepts positive explicit timeout", "adapter rejects non-positive timeout", "benchmark passes its case timeout to the adapter", "production Medium defaults to 60 seconds", "no provider call bypasses the adapter"],
+  "required_tests": ["go test ./...", "go vet ./...", "go run ./cmd/devctl preflight -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-diff -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-architecture -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl/independent", "git diff --check", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000176-medium-qualification-retest.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000176",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "389dd0c23d818e3bde08d8353287938837179da8",
+  "risk_tier": "E",
+  "intent": "Record the post-timeout-change production qualification retest for the reachable MEDIUM model.",
+  "requirement_ids": ["WP-200", "WP-301", "AUDIT-COMPLETION-TRUTH"],
+  "allowed_change_scope": {"modules": [], "paths": ["docs/evidence/wp-200-medium-qualification-2026-08-22-v2.json", "docs/PROJECT_STATUS.md", "docs/evidence/audit-completion-matrix-2026-08-22-v1.json", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000176-medium-qualification-retest.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/**", "modules/**", "contracts/**", "adapters/**", "database/**", "deploy/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["production cognitive adapter qualification artifact"], "may_change": []},
+  "semantic_owner": "none; audit evidence projection",
+  "canonical_state": [],
+  "state_transitions": ["timeout-policy retest -> verified model-generation gap"],
+  "side_effects_and_privileges": ["one-shot benchmark only; no lifecycle mutation"],
+  "persistence_and_migrations": ["immutable JSON evidence artifacts only"],
+  "failure_retry_restart": ["45-second case deadline is recorded; unsupported result is not promoted"],
+  "assumptions": ["retest ran through the updated cognitive adapter in production networks", "no activation or canary was attempted"],
+  "acceptance_scenarios": ["all four cases record approximately 45-second timeout", "unsupported result remains explicit", "production status retains Medium qualification gap", "no secrets appear"],
+  "required_tests": ["JSON validation", "go test ./...", "go vet ./...", "go run ./cmd/devctl preflight -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-diff -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-architecture -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl/independent", "git diff --check", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000177-heavy-qualification-evidence.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000177",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "d60fb8697a577b113da2043bddca3b88e9e562fd",
+  "risk_tier": "E",
+  "intent": "Record bounded production qualification results for currently readiness-available Heavy model candidates.",
+  "requirement_ids": ["WP-200", "WP-301", "AUDIT-COMPLETION-TRUTH"],
+  "allowed_change_scope": {"modules": [], "paths": ["docs/evidence/wp-200-heavy-qualification-2026-08-22.json", "docs/PROJECT_STATUS.md", "docs/evidence/audit-completion-matrix-2026-08-22-v1.json", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000177-heavy-qualification-evidence.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/**", "modules/**", "contracts/**", "adapters/**", "database/**", "deploy/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["production Ollama adapter qualification artifact", "fleet readiness evidence"], "may_change": []},
+  "semantic_owner": "none; audit evidence projection",
+  "canonical_state": [],
+  "state_transitions": ["Heavy readiness available -> Heavy model candidates qualification recorded"],
+  "side_effects_and_privileges": ["one-shot benchmark only; no lifecycle mutation"],
+  "persistence_and_migrations": ["immutable JSON evidence artifacts only"],
+  "failure_retry_restart": ["MODEL_UNAVAILABLE is recorded as unsupported; no activation or promotion"],
+  "assumptions": ["all four candidates were taken from the production registry read-only", "readiness availability does not imply model invocability"],
+  "acceptance_scenarios": ["all four Heavy candidates are represented", "MODEL_UNAVAILABLE results are explicit", "audit status distinguishes readiness from qualification", "no secret values appear"],
+  "required_tests": ["JSON validation", "go test ./...", "go vet ./...", "go run ./cmd/devctl preflight -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-diff -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-architecture -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl/independent", "git diff --check", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000178-heavy-qualification-refresh.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000178",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "dc1ee7ff3c5e284548bf5d9e1e26c9dc9f2c88fd",
+  "risk_tier": "E",
+  "intent": "Record refreshed Heavy registry discovery and successful bounded qualification for the currently installed model candidate.",
+  "requirement_ids": ["WP-200", "WP-301", "AUDIT-COMPLETION-TRUTH"],
+  "allowed_change_scope": {"modules": [], "paths": ["docs/evidence/wp-200-heavy-qualification-2026-08-22-v2.json", "docs/PROJECT_STATUS.md", "docs/evidence/audit-completion-matrix-2026-08-22-v1.json", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000178-heavy-qualification-refresh.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/**", "modules/**", "contracts/**", "adapters/**", "database/**", "deploy/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["provider discovery endpoint", "Ollama adapter qualification artifact"], "may_change": []},
+  "semantic_owner": "none; audit evidence projection",
+  "canonical_state": [],
+  "state_transitions": ["stale Heavy inventory -> current gemma4:26b supported qualification evidence"],
+  "side_effects_and_privileges": ["provider discovery refresh and one-shot benchmark; no lifecycle promotion"],
+  "persistence_and_migrations": ["immutable JSON evidence artifacts only"],
+  "failure_retry_restart": ["stale candidates remain historical evidence; current candidate is separately identified by immutable revisions"],
+  "assumptions": ["discovery was invoked through cognitive-worker's Ollama adapter path", "qualification does not activate an INACTIVE binding"],
+  "acceptance_scenarios": ["current gemma4:26b revisions are recorded", "all four synthetic cases pass", "stale qwen candidate result is not treated as current", "activation remains explicit and unclaimed", "no secrets appear"],
+  "required_tests": ["JSON validation", "go test ./...", "go vet ./...", "go run ./cmd/devctl preflight -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-diff -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-architecture -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl/independent", "git diff --check", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000179-runtime-status-consolidation.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000179",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "5960844fcb7b9bfba9c24e18a5d15b569d0d3a08",
+  "risk_tier": "E",
+  "intent": "Synchronize audit projections with current production runtime evidence after the Heavy qualification deployment.",
+  "requirement_ids": ["AUDIT-COMPLETION-TRUTH", "WP-301", "WP-400", "WP-102", "WP-600"],
+  "allowed_change_scope": {"modules": [], "paths": ["docs/evidence/audit-runtime-status-2026-08-22-v4.json", "docs/evidence/audit-completion-matrix-2026-08-22-v1.json", "docs/PROJECT_STATUS.md", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000179-runtime-status-consolidation.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/**", "modules/**", "contracts/**", "adapters/**", "database/**", "deploy/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["live health/readiness/status endpoints"], "may_change": []},
+  "semantic_owner": "none; audit projection",
+  "canonical_state": [],
+  "state_transitions": ["stale audit projection -> current production runtime evidence"],
+  "side_effects_and_privileges": ["read-only live probes"],
+  "persistence_and_migrations": ["immutable JSON evidence only"],
+  "failure_retry_restart": ["not applicable"],
+  "assumptions": ["production revision is read from creator-web runtime environment", "runtime readiness is not treated as capability qualification"],
+  "acceptance_scenarios": ["public health is current", "research/work-graph readiness is current", "agent recovery-only state is explicit", "fleet tiers and selected tier are current", "no secrets appear"],
+  "required_tests": ["JSON validation", "go test ./...", "go vet ./...", "go run ./cmd/devctl preflight -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-diff -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-architecture -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl/independent", "git diff --check", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000180-bounded-medium-qualification-output.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000180",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "2ed2302526f6915c9ddef852ed17a8c47f716cc2",
+  "risk_tier": "E",
+  "intent": "Allow the existing synthetic qualification benchmark to bound maximum output bytes per case for slow local providers.",
+  "requirement_ids": ["WP-200", "AUDIT-COMPLETION-TRUTH"],
+  "allowed_change_scope": {"modules": ["cognitive benchmark"], "paths": ["cmd/cognitive-benchmark/main.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000180-bounded-medium-qualification-output.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/cognitive-worker/**", "modules/**", "contracts/**", "adapters/**", "database/**", "deploy/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["existing cognitive advisory request contract"], "may_change": ["benchmark CLI parameter only"]},
+  "semantic_owner": "cognitive benchmark",
+  "canonical_state": [],
+  "state_transitions": ["fixed benchmark output cap -> explicit bounded qualification run"],
+  "side_effects_and_privileges": ["none"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["non-positive output cap rejected; default preserves previous 4096-byte behavior"],
+  "assumptions": ["bounded output is valid only for synthetic exact-answer corpus and does not alter production provider configuration"],
+  "acceptance_scenarios": ["default behavior remains unchanged", "positive output cap is forwarded to advisory request", "invalid cap fails closed", "no provider boundary is bypassed"],
+  "required_tests": ["go test ./...", "go vet ./...", "go run ./cmd/devctl preflight -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-diff -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-architecture -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl/independent", "git diff --check", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000181-medium-bounded-retest-evidence.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000181",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "6f0e099632a79af90eab439b4f026cd31abaad26",
+  "risk_tier": "E",
+  "intent": "Record the final bounded Medium qualification retest after adding an explicit benchmark output cap.",
+  "requirement_ids": ["WP-200", "AUDIT-COMPLETION-TRUTH"],
+  "allowed_change_scope": {"modules": [], "paths": ["docs/evidence/wp-200-medium-qualification-2026-08-22-v3.json", "docs/PROJECT_STATUS.md", "docs/evidence/audit-completion-matrix-2026-08-22-v1.json", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000181-medium-bounded-retest-evidence.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/**", "modules/**", "contracts/**", "adapters/**", "database/**", "deploy/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["bounded cognitive benchmark artifact"], "may_change": []},
+  "semantic_owner": "none; audit evidence projection",
+  "canonical_state": [],
+  "state_transitions": ["Medium output-cap hypothesis -> verified runtime performance gap"],
+  "side_effects_and_privileges": ["one-shot benchmark only; no lifecycle mutation"],
+  "persistence_and_migrations": ["immutable JSON evidence only"],
+  "failure_retry_restart": ["60-second case timeout and 256-byte cap are recorded; unsupported result is not promoted"],
+  "assumptions": ["benchmark used the existing cognitive adapter and current production Medium revisions", "a bounded exact-answer corpus is sufficient to disprove the output-cap hypothesis"],
+  "acceptance_scenarios": ["all four cases record approximately 60-second timeout", "quality remains 0‰ and critical failure 1000‰", "Medium remains unqualified", "no secret values appear"],
+  "required_tests": ["JSON validation", "go test ./...", "go vet ./...", "go run ./cmd/devctl preflight -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-diff -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl check-architecture -active dev/ACTIVE_TASK.json", "go run ./cmd/devctl/independent", "git diff --check", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000182-model-rating-and-adaptive-routing.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000182",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "4709ef46076cf5c22522bbfea785fdb3c6220c14",
+  "risk_tier": "E",
+  "intent": "Expose an evidence-based model rating and adaptive routing policy so model category follows qualification and observed performance rather than provider or hardware names.",
+  "requirement_ids": ["WP-200", "WP-601", "adaptive model routing", "CPU-first current operation"],
+  "allowed_change_scope": {"modules": ["capability", "cognitive-worker"], "paths": ["contracts/capability/cognitive_registry.go", "modules/capability/domain/cognitive_registry.go", "modules/capability/domain/fleet.go", "modules/capability/domain/fleet_test.go", "cmd/cognitive-worker/main.go", "cmd/cognitive-worker/fleet_status_test.go", "modules/capability/domain/model_rating_test.go", "docs/MODEL_FLEET_POLICY.md", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000182-model-rating-and-adaptive-routing.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["EvaluatedCapability", "PerformanceEvidence", "CognitiveRequirement"], "may_change": ["provider-neutral model rating projection"]},
+  "semantic_owner": "capability",
+  "canonical_state": ["existing owner-local qualification and performance evidence; rating is a derived projection"],
+  "state_transitions": ["unrated -> rated only after supported qualification evidence; category is recomputed from current evidence; routing chooses an eligible rated candidate and falls back on invocation failure"],
+  "side_effects_and_privileges": ["none; no provider invocation or external mutation"],
+  "persistence_and_migrations": ["no migration; rating is derived and serializable"],
+  "failure_retry_restart": ["missing or stale evidence yields UNRATED and cannot silently claim quality; existing gateway fallback remains authoritative"],
+  "assumptions": ["CPU is the current operational choice, but routing remains provider- and hardware-neutral", "discovery never activates a binding"],
+  "acceptance_scenarios": ["supported evidence produces a visible rating and category", "missing evidence remains UNRATED", "quality objective prefers higher evidence rating", "speed objective prefers lower observed latency", "provider identity alone cannot raise a rating", "CPU is preferred now but all ready capability classes remain valid failover candidates"],
+  "required_tests": ["go test ./...", "go vet ./...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_SPEC_GAP", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000184-gpu-openai-protected-provider.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000184",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "16a0e8c4924a2e644294e12f048e747c00f940b9",
+  "risk_tier": "E",
+  "intent": "Connect an optional OpenAI-compatible GPU provider through protected Domovoy storage without literal provider credentials or hardwired environment configuration, while preserving CPU/Lifeboat degraded operation.",
+  "requirement_ids": ["GPU provider endpoint is authenticated through protected storage", "GPU absence or invalid configuration does not stop cognitive-worker", "fleet readiness and discovery use OpenAI-compatible /v1/models", "CPU and Lifeboat remain available fallback paths"],
+  "allowed_change_scope": {"modules": ["cognitive-worker", "deployment"], "paths": ["cmd/cognitive-worker/**", "deploy/docker-compose.prod.yml", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000184-gpu-openai-protected-provider.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "database/**", "contracts/**", "modules/**", "security/**", ".gitea/**"]},
+  "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": ["protected Domovoy secret files", "OpenAI-compatible /v1/models and /v1/chat/completions", "existing cognitive gateway and fleet readiness contracts"], "may_change": []},
+  "semantic_owner": "capability provider adapter projection",
+  "canonical_state": [],
+  "state_transitions": ["missing/invalid protected GPU config -> provider omitted and degraded readiness", "valid config -> authenticated discovery/readiness and optional heavy-tier candidate"],
+  "side_effects_and_privileges": ["outbound authenticated requests to the explicitly configured GPU endpoint; no credentials forwarded to Core or user responses"],
+  "persistence_and_migrations": ["no schema changes; protected file storage only"],
+  "failure_retry_restart": ["GPU timeout/unavailability falls back through existing router to CPU/Lifeboat", "restart reloads protected files", "invalid config disables only optional GPU provider"],
+  "assumptions": ["the supplied endpoint implements OpenAI-compatible /v1/models and /v1/chat/completions", "the protected secret volume is mounted read-only into cognitive-worker", "docker-test remains prohibited"],
+  "acceptance_scenarios": ["GPU /v1/models succeeds with protected key", "GPU key never appears in env/logs/responses", "GPU discovery creates provider-neutral records", "fleet readiness reports GPU heavy tier only when configured", "stopping GPU leaves CPU/Lifeboat ready", "chat remains available without GPU files"],
+  "required_tests": ["go test ./cmd/cognitive-worker ./adapters/models", "go vet ./...", "go run ./cmd/devctl preflight -task dev/tasks/DEV-000184-gpu-openai-protected-provider.json", "go run ./cmd/devctl check-diff -task dev/tasks/DEV-000184-gpu-openai-protected-provider.json", "go run ./cmd/devctl check-architecture", "go run ./cmd/devctl/independent", "git diff --check", "CI", "production protected-key models/readiness/live fallback checks"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000185-agent-bridge-protected-config.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000185",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0017",
+  "base_repo_revision": "7ab4b02afa78d5c65eb1545bfd6e117cc1cdfad7",
+  "risk_tier": "E",
+  "intent": "Load the optional external Agent Execution bridge endpoint from Domovoy protected storage and preserve recovery-only operation when it is absent or invalid.",
+  "requirement_ids": ["protected bridge configuration", "degraded recovery without bridge", "durable dispatch when bridge is configured"],
+  "allowed_change_scope": {"modules": ["agent-worker", "deployment"], "paths": ["cmd/agent-worker/**", "deploy/docker-compose.prod.yml", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000185-agent-bridge-protected-config.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "database/**", "contracts/**", "security/**", ".gitea/**"]},
+  "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 agentbridge HTTP executor", "protected runtime configuration files", "existing work dispatch contract"], "may_change": []},
+  "semantic_owner": "work agent execution adapter",
+  "canonical_state": [],
+  "state_transitions": ["missing/invalid protected bridge URL -> recovery-only", "valid protected bridge URL -> execution mode"],
+  "side_effects_and_privileges": ["outbound HTTP calls only to protected configured bridge endpoint"],
+  "persistence_and_migrations": ["no schema changes; protected file storage only"],
+  "failure_retry_restart": ["bridge errors remain handled by existing worker cycle", "restart reloads protected URL", "missing bridge never prevents recovery"],
+  "assumptions": ["protected secret volume is mounted read-only", "bridge URL file is deployment-owned"],
+  "acceptance_scenarios": ["protected URL enables execution mode", "missing URL keeps recovery-only mode", "URL is not read from environment", "URL value never appears in status"],
+  "required_tests": ["go test ./cmd/agent-worker/...", "go vet ./...", "go run ./cmd/devctl preflight -task dev/tasks/DEV-000185-agent-bridge-protected-config.json", "go run ./cmd/devctl check-diff -task dev/tasks/DEV-000185-agent-bridge-protected-config.json", "go run ./cmd/devctl check-architecture", "go run ./cmd/devctl independent", "git diff --check", "CI", "production status and recovery checks"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000186-extension-proposal-generator.json

@@ -0,0 +1,37 @@
+{
+  "task_id": "DEV-000186",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "cdbccb6",
+  "risk_tier": "E",
+  "intent": "Add a typed, proposal-only ExtensionProposal generator from repeated or corroborated evidence without granting mutation, authority, deployment, or promotion capability.",
+  "requirement_ids": ["WP-601", "Extension proposal generator", "Evidence deduplication", "Governed rollback plan"],
+  "allowed_change_scope": {
+    "modules": ["work-agent-execution"],
+    "paths": ["contracts/work/self_development.go", "modules/work/domain/self_development.go", "modules/work/domain/self_development_test.go", "tests/semantic_evolution/self_development_contract_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000186-extension-proposal-generator.json"]
+  },
+  "forbidden_change_scope": {
+    "paths": ["architecture/**", "security/**", "database/**", "cmd/**", "adapters/**", "deploy/**", "contracts/authority/**", ".gitea/**"]
+  },
+  "allowed_architecture_delta": {
+    "new_cross_module_dependencies": 0,
+    "new_canonical_owners": 0,
+    "new_privileges": 0,
+    "new_third_party_dependencies": 0,
+    "new_database_owner_roles": 0
+  },
+  "contracts": {
+    "consume": ["ADR-0022 governed self-development boundary", "existing Work DevelopmentGate"],
+    "may_change": ["typed Work ExtensionProposal value objects"]
+  },
+  "semantic_owner": "work-agent-execution",
+  "canonical_state": ["No durable canonical state; proposal value object only"],
+  "state_transitions": ["repeated evidence -> PROPOSED ExtensionProposal", "insufficient/conflicting/incomplete evidence -> rejected result"],
+  "side_effects_and_privileges": ["No source mutation, provider invocation, secret access, deployment, promotion, or authority expansion"],
+  "persistence_and_migrations": ["No schema or migration changes"],
+  "failure_retry_restart": ["Deterministic generation is rerunnable; unique evidence refs and signal keys preserve lineage; one failure cannot generate a proposal"],
+  "assumptions": ["Human governance remains the promotion authority", "A later task may add durable proposal persistence only after owner and migration review"],
+  "acceptance_scenarios": ["one failure is rejected", "repeated same-signal evidence produces one proposal with deduplicated signal keys and retained refs", "missing assumptions/evaluation/rollback fails closed", "proposal contains no executable authority or mutation field"],
+  "required_tests": ["go test ./modules/work/domain ./tests/semantic_evolution", "go vet ./...", "go run ./cmd/devctl preflight -task dev/tasks/DEV-000186-extension-proposal-generator.json", "go run ./cmd/devctl check-diff -task dev/tasks/DEV-000186-extension-proposal-generator.json", "go run ./cmd/devctl check-architecture", "go run ./cmd/devctl independent", "git diff --check", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_ARCHITECTURE_GAP", "BLOCKED_BY_SPEC_GAP", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000187-scaffold-task-packet-candidate.json

@@ -0,0 +1,26 @@
+{
+  "task_id": "DEV-000187",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "a09a7e8",
+  "risk_tier": "E",
+  "intent": "Add a fail-closed typed scaffold candidate factory that produces an isolated task-packet descriptor without creating files, reading secrets, or accessing protected main/production.",
+  "requirement_ids": ["WP-602", "Scoped task packet candidate", "Isolated worktree provenance", "Protected path denial"],
+  "allowed_change_scope": {
+    "modules": ["work-agent-execution"],
+    "paths": ["contracts/work/self_development.go", "modules/work/domain/self_development.go", "modules/work/domain/self_development_test.go", "tests/semantic_evolution/self_development_contract_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000187-scaffold-task-packet-candidate.json"]
+  },
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "cmd/**", "adapters/**", "deploy/**", "contracts/authority/**", ".gitea/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 0},
+  "contracts": {"consume": ["Existing ExtensionProposal", "ADR-0022 governed development boundary"], "may_change": ["typed Work scaffold candidate value objects"]},
+  "semantic_owner": "work-agent-execution",
+  "canonical_state": ["No durable state; isolated scaffold candidate value object only"],
+  "state_transitions": ["proposal -> SANDBOXED task-packet candidate", "invalid scope/policy -> rejected result"],
+  "side_effects_and_privileges": ["No filesystem mutation, source checkout, secret access, provider call, deployment, promotion, or authority expansion"],
+  "persistence_and_migrations": ["No schema or migration changes"],
+  "failure_retry_restart": ["Deterministic validation is rerunnable; invalid or protected scope fails closed"],
+  "assumptions": ["A later external development control plane performs actual worktree creation", "Human governance remains promotion authority"],
+  "acceptance_scenarios": ["valid proposal produces scoped SANDBOXED descriptor", "absolute/traversal/protected path is rejected", "production and secret access policies must be DENIED", "missing provenance or test plan is rejected", "descriptor contains no command or credential material"],
+  "required_tests": ["go test ./modules/work/domain ./tests/semantic_evolution", "go vet ./...", "go run ./cmd/devctl preflight -task dev/tasks/DEV-000187-scaffold-task-packet-candidate.json", "go run ./cmd/devctl check-diff -task dev/tasks/DEV-000187-scaffold-task-packet-candidate.json", "go run ./cmd/devctl check-architecture", "go run ./cmd/devctl independent", "git diff --check", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_ARCHITECTURE_GAP", "BLOCKED_BY_SPEC_GAP", "BLOCKED_BY_CONFLICT"]
+}