feat(cognitive): execute bounded task evaluation cases

docs: record gpu adapter invocation qualification

adapters/models/openai_compatible_local.go

@@ -22,17 +22,35 @@ import (
 type OpenAICompatibleLocalAdapter struct {
 	base   *url.URL
 	client *http.Client
+	apiKey string
 }
 
 func NewOpenAICompatibleLocalAdapter(raw string, transport http.RoundTripper) (*OpenAICompatibleLocalAdapter, error) {
+	return NewOpenAICompatibleLocalAdapterWithAPIKey(raw, "", transport)
+}
+
+// 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 }}}, 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 {
@@ -78,11 +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"`
@@ -103,7 +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 {
@@ -145,6 +167,9 @@ func (a *OpenAICompatibleLocalAdapter) requestJSON(ctx context.Context, method,
 		return 0, err
 	}
 	request.Header.Set("Accept", "application/json")
+	if a.apiKey != "" {
+		request.Header.Set("Authorization", "Bearer "+a.apiKey)
+	}
 	if input != nil {
 		request.Header.Set("Content-Type", "application/json")
 	}

adapters/models/openai_compatible_local_test.go

@@ -0,0 +1,94 @@
+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")
+	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) {
+	transport := &captureRoundTripper{}
+	adapter, err := NewOpenAICompatibleLocalAdapterWithAPIKey("http://gateway/v1", "resolved-key", transport)
+	if err != nil {
+		t.Fatal(err)
+	}
+	_, _ = adapter.Discover(context.Background())
+	if transport.authorization != "Bearer resolved-key" {
+		t.Fatalf("unexpected authorization header: %q", transport.authorization)
+	}
+}
+
+func TestOpenAICompatibleLocalAdapterDoesNotSendAuthorizationWithoutKey(t *testing.T) {
+	transport := &captureRoundTripper{}
+	adapter, err := NewOpenAICompatibleLocalAdapter("http://gateway/v1", transport)
+	if err != nil {
+		t.Fatal(err)
+	}
+	_, _ = adapter.Discover(context.Background())
+	if transport.authorization != "" {
+		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 +505,16 @@ func (r *CapabilityRepository) loadCognitiveEvidence(ctx context.Context, value
 	if err == nil {
 		value.Performance = &performance
 	}
+	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 +628,8 @@ 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
+}

adapters/postgres/work_repository.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"slices"
 	"time"
 
 	"git.cin.su/m/Assistant/contracts/work"
@@ -14,9 +15,10 @@ 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")
 )
 
 func OpenWork(ctx context.Context, config ContinuityConfig) (*WorkRepository, error) {
@@ -49,30 +51,84 @@ 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) 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)
+	_, 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) VALUES($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13,$14) 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)
 	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)
+	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 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)
 	if err != nil {
 		return fmt.Errorf("read dispatch intent: %w", err)
 	}
-	if existing != intent {
+	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)
+}
+
 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)
+	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 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)
 	if err != nil {
 		return work.AgentDispatchIntent{}, fmt.Errorf("load dispatch intent: %w", err)
 	}
@@ -140,11 +196,11 @@ func (r *WorkRepository) SaveWorkGraph(ctx context.Context, graph work.WorkGraph
 		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) VALUES($1,$2,$3,$4,$5,$6,$7,$8) 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); 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 {
+		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) VALUES($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11) 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); err != nil {
 			return err
 		}
 	}
@@ -158,17 +214,17 @@ 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 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)
 	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 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 {
+		if err := rows.Scan(&node.ID, &node.DispatchID, &node.IdempotencyKey, &node.State, &node.Attempts, &node.Purpose, &node.TaskProfileID, &node.BindingID, &node.CaseIDs); err != nil {
 			rows.Close()
 			return work.WorkGraph{}, err
 		}
@@ -199,6 +255,26 @@ func (r *WorkRepository) LoadWorkGraph(ctx context.Context, graphID string, revi
 	return graph, nil
 }
 
+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) 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)
 	if err != 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 }
@@ -30,14 +34,24 @@ 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 +81,27 @@ 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"
+)
+
+// localFallbackExecutor keeps the durable work path alive when no external
+// executor is reachable. It never claims that an external action happened.
+// The work owner records the WAITING callback and can retry after a bridge is
+// configured or a human chooses the next step.
+type localFallbackExecutor struct {
+	now func() time.Time
+}
+
+func newLocalFallbackExecutor() localFallbackExecutor {
+	return localFallbackExecutor{now: time.Now}
+}
+
+func (e localFallbackExecutor) 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:              "Внешний Agent Execution Bridge недоступен; локальный fallback сохранил durable работу и ожидает 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 TestLocalFallbackNeverClaimsExternalExecution(t *testing.T) {
+	when := time.UnixMilli(1_700_000_000_000)
+	executor := localFallbackExecutor{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("fallback 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"
@@ -20,6 +22,7 @@ import (
 
 func main() {
 	workerID := flag.String("worker-id", "agent-worker-1", "durable worker identity")
+	statusAddr := flag.String("status-addr", ":8090", "internal status HTTP address")
 	staleAfter := flag.Duration("stale-after", 10*time.Minute, "age after which a claimed item is recovered")
 	leaseTTL := flag.Duration("lease-ttl", 15*time.Minute, "executor lease duration")
 	pollInterval := flag.Duration("poll-interval", 5*time.Second, "database-owned idle wake-up interval")
@@ -38,15 +41,23 @@ 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")
+		executor = newLocalFallbackExecutor()
+		localFallback = true
+		fmt.Fprintln(os.Stderr, "protected agent execution bridge URL is not configured; local fallback preserves durable work without external mutation")
 	}
+	go func() {
+		if serveErr := http.ListenAndServe(*statusAddr, workerStatusHandler(executor != nil && !localFallback, localFallback)); serveErr != nil {
+			fmt.Fprintln(os.Stderr, "agent worker status server:", serveErr)
+		}
+	}()
 	ctx := context.Background()
 	for {
 		var result application.WorkerCycleResult
@@ -75,6 +86,27 @@ 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 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

@@ -0,0 +1,47 @@
+package main
+
+import (
+	"encoding/json"
+	"net/http"
+)
+
+type workerStatusResponse struct {
+	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"`
+}
+
+func workerStatusHandler(executorConfigured bool, localFallback ...bool) http.Handler {
+	local := len(localFallback) > 0 && localFallback[0]
+	status := workerStatusResponse{
+		Runtime:            "agent-worker",
+		ExecutorConfigured: executorConfigured,
+		Mode:               "recovery-only",
+		Ready:              false,
+		Reason:             "protected agent execution bridge URL is not configured",
+		LocalFallback:      local,
+	}
+	if local {
+		status.Mode = "local-fallback"
+		status.Reason = "external bridge unavailable; durable work is retained without external mutation"
+	}
+	if executorConfigured {
+		status.Mode = "execution"
+		status.Ready = true
+		status.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") {
+			http.NotFound(w, r)
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		if r.URL.Path == "/readyz" && !status.Ready {
+			w.WriteHeader(http.StatusServiceUnavailable)
+		}
+		_ = json.NewEncoder(w).Encode(status)
+	})
+}

cmd/agent-worker/status_test.go

@@ -0,0 +1,64 @@
+package main
+
+import (
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"testing"
+)
+
+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))
+	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 != "recovery-only" || value.ExecutorConfigured {
+		t.Fatalf("unexpected status: code=%d value=%+v", response.Code, value)
+	}
+}
+
+func TestWorkerStatusReportsLocalFallback(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-fallback" || !value.LocalFallback {
+		t.Fatalf("unexpected status: code=%d value=%+v", response.Code, value)
+	}
+}
+
+func TestWorkerReadyzRequiresExecutor(t *testing.T) {
+	response := httptest.NewRecorder()
+	workerStatusHandler(false).ServeHTTP(response, httptest.NewRequest(http.MethodGet, "/readyz", nil))
+	if response.Code != http.StatusServiceUnavailable {
+		t.Fatalf("readyz status=%d", response.Code)
+	}
+	response = httptest.NewRecorder()
+	workerStatusHandler(true).ServeHTTP(response, httptest.NewRequest(http.MethodGet, "/readyz", nil))
+	if response.Code != http.StatusOK {
+		t.Fatalf("configured readyz status=%d", response.Code)
+	}
+}

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/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
@@ -341,6 +345,26 @@ 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/main.go

@@ -10,11 +10,14 @@ import (
 	"git.cin.su/m/Assistant/contracts/capability"
 	"git.cin.su/m/Assistant/contracts/cognitive"
 	"git.cin.su/m/Assistant/contracts/identity"
+	secretcontracts "git.cin.su/m/Assistant/contracts/secrets"
+	platformsecrets "git.cin.su/m/Assistant/platform/secrets"
 	"git.cin.su/m/Assistant/runtime/health"
 	"io"
 	"log"
 	"net/http"
 	"os"
+	"path/filepath"
 	"strings"
 	"time"
 )
@@ -43,13 +46,45 @@ func main() {
 	discoveries := map[string]providerRegistration{}
 	var builtinLocal *providerRegistration
 	externalRecovery := []providerRegistration{}
-	cpuAdapter, err := models.NewOpenAICompatibleLocalAdapter(envOr("LOCAL_CPU_PROVIDER_URL", "http://docker.cin.su:18082/v1"), nil)
+	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)
 	}
-	cpu := providerRegistration{ProviderInstanceID: "local-cpu-primary", Kind: "openai-compatible-chat-v1", EndpointRef: "local-cpu-primary-v1", AuthMode: capability.AuthNone, BillingMode: capability.BillingFree, DataPolicy: capability.DataPolicyLocalOnly, Provider: cpuAdapter, Discovery: cpuAdapter}
+	cpuAdapter, err := models.NewOpenAICompatibleLocalAdapterWithAPIKey(envOr("LOCAL_CPU_PROVIDER_URL", "http://docker.cin.su:18082/v1"), cpuAPIKey, nil)
+	if err != nil {
+		log.Fatal(err)
+	}
+	cpuAuthMode := capability.AuthNone
+	cpuCredentialRef := capability.CredentialRef("")
+	if cpuAPIKey != "" {
+		cpuAuthMode = capability.AuthKey
+		cpuCredentialRef = capability.CredentialRef(envOr("LOCAL_CPU_PROVIDER_CREDENTIAL_REF", "credential://domovoy/local-cpu/provider-api-key/v1"))
+	}
+	cpu := providerRegistration{ProviderInstanceID: "local-cpu-primary", Kind: "openai-compatible-chat-v1", EndpointRef: "local-cpu-primary-v1", CredentialRef: cpuCredentialRef, AuthMode: cpuAuthMode, BillingMode: capability.BillingFree, DataPolicy: capability.DataPolicyLocalOnly, Provider: cpuAdapter, Discovery: cpuAdapter}
 	providers[cpu.Kind] = append(providers[cpu.Kind], cpu)
 	discoveries["cpu"] = cpu
+	mediumURL := strings.TrimSpace(os.Getenv("LOCAL_MEDIUM_PROVIDER_URL"))
+	mediumAPIKey := ""
+	if mediumURL != "" {
+		mediumAPIKey, err = loadProviderAPIKey("LOCAL_MEDIUM_PROVIDER_API_KEY", "LOCAL_MEDIUM_PROVIDER_API_KEY_FILE", "LOCAL_MEDIUM_PROVIDER_CREDENTIAL_REF", "credential://domovoy/local-medium/provider-api-key/v1")
+		if err != nil {
+			log.Fatal(err)
+		}
+		mediumTimeout := durationEnv("LOCAL_MEDIUM_PROVIDER_TIMEOUT", 60*time.Second)
+		mediumAdapter, adapterErr := models.NewOpenAICompatibleLocalAdapterWithAPIKeyAndTimeout(strings.TrimRight(mediumURL, "/"), mediumAPIKey, nil, mediumTimeout)
+		if adapterErr != nil {
+			log.Fatal(adapterErr)
+		}
+		mediumAuthMode := capability.AuthNone
+		mediumCredentialRef := capability.CredentialRef("")
+		if mediumAPIKey != "" {
+			mediumAuthMode = capability.AuthKey
+			mediumCredentialRef = capability.CredentialRef(envOr("LOCAL_MEDIUM_PROVIDER_CREDENTIAL_REF", "credential://domovoy/local-medium/provider-api-key/v1"))
+		}
+		medium := providerRegistration{ProviderInstanceID: "local-medium-primary", Kind: "openai-compatible-chat-v1", EndpointRef: "local-medium-primary-v1", CredentialRef: mediumCredentialRef, AuthMode: mediumAuthMode, BillingMode: capability.BillingFree, DataPolicy: capability.DataPolicyLocalOnly, Provider: mediumAdapter, Discovery: mediumAdapter}
+		providers[medium.Kind] = append(providers[medium.Kind], medium)
+		discoveries["medium"] = medium
+	}
 	if lifeboat, enabled, lifeboatErr := newLifeboatProvider(os.Getenv("LOCAL_LIFEBOAT_PROVIDER_URL")); lifeboatErr != nil {
 		log.Fatal(lifeboatErr)
 	} else if enabled {
@@ -58,14 +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})
@@ -73,25 +114,59 @@ 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)
 	refreshProviderDiscoveries(resolver, discovery)
@@ -101,6 +176,19 @@ 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)
@@ -145,6 +233,72 @@ func refreshProviderDiscoveries(resolver bindingResolver, service discoveryServi
 	}
 }
 
+func loadProviderAPIKey(directEnvName, fileEnvName, refEnvName, defaultRef string) (string, error) {
+	if _, set := os.LookupEnv(directEnvName); set {
+		return "", fmt.Errorf("%s direct environment secret is forbidden; use %s", directEnvName, fileEnvName)
+	}
+	fileName := strings.TrimSpace(os.Getenv(fileEnvName))
+	if fileName == "" {
+		return "", nil
+	}
+	ref := strings.TrimSpace(os.Getenv(refEnvName))
+	if ref == "" {
+		ref = defaultRef
+	}
+	if ref == "" {
+		return "", fmt.Errorf("%s is required when %s is configured", refEnvName, fileEnvName)
+	}
+	resolver, err := platformsecrets.NewFileResolver(envOr("SECRET_BACKEND_ROOT", "/run/secrets"))
+	if err != nil {
+		return "", err
+	}
+	return resolver.ResolveFile(secretcontracts.SecretFileRef{CredentialRef: secretcontracts.CredentialRef(ref), Purpose: secretcontracts.SecretPurposeProviderAPIKey, Consumer: secretcontracts.RuntimeConsumerCognitive, Path: fileName})
+}
+
+func loadLocalCPUProviderAPIKey() (string, error) {
+	return loadProviderAPIKey("LOCAL_CPU_PROVIDER_API_KEY", "LOCAL_CPU_PROVIDER_API_KEY_FILE", "LOCAL_CPU_PROVIDER_CREDENTIAL_REF", "credential://domovoy/local-cpu/provider-api-key/v1")
+}
+
+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

@@ -1,6 +1,10 @@
 package main
 
-import "testing"
+import (
+	"os"
+	"path/filepath"
+	"testing"
+)
 
 func TestLifeboatProviderIsOptionalAndLocalAdapterBacked(t *testing.T) {
 	if _, enabled, err := newLifeboatProvider(""); err != nil || enabled {
@@ -14,3 +18,59 @@ func TestLifeboatProviderIsOptionalAndLocalAdapterBacked(t *testing.T) {
 		t.Fatalf("unexpected lifeboat registration: %+v", provider)
 	}
 }
+
+func TestLoadLocalCPUProviderAPIKeyUsesProtectedFileReference(t *testing.T) {
+	t.Setenv("LOCAL_CPU_PROVIDER_API_KEY", "")
+	_ = os.Unsetenv("LOCAL_CPU_PROVIDER_API_KEY")
+	root := t.TempDir()
+	path := filepath.Join(root, "local_cpu_provider_api_key")
+	if err := os.WriteFile(path, []byte("resolved-key\n"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	t.Setenv("LOCAL_CPU_PROVIDER_API_KEY_FILE", path)
+	t.Setenv("LOCAL_CPU_PROVIDER_CREDENTIAL_REF", "credential://domovoy/local-cpu/provider-api-key/v1")
+	t.Setenv("SECRET_BACKEND_ROOT", root)
+	key, err := loadLocalCPUProviderAPIKey()
+	if err != nil || key != "resolved-key" {
+		t.Fatalf("unexpected key resolution result: err=%v", err)
+	}
+}
+
+func TestLoadLocalCPUProviderAPIKeyRejectsDirectEnvironmentSecret(t *testing.T) {
+	t.Setenv("LOCAL_CPU_PROVIDER_API_KEY", "must-not-be-used")
+	t.Setenv("LOCAL_CPU_PROVIDER_API_KEY_FILE", "")
+	if _, err := loadLocalCPUProviderAPIKey(); err == nil {
+		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/task_evaluation.go

@@ -0,0 +1,93 @@
+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))
+	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}
+			if invokeErr == nil && response.FinishReason == cognitive.FinishCompleted && response.AdvisoryText != "" {
+				observation.Result = capability.EvaluationSupported
+				// Invocation success proves reachability/support for this case,
+				// not semantic answer quality. A verifier may score quality later.
+				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, 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 +1,53 @@
+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.EvaluationSupported || response.Observations[1].Result != capability.EvaluationSupported {
+		t.Fatalf("unexpected observations: %+v", response.Observations)
+	}
+}

cmd/core-runtime/main.go

@@ -834,6 +834,20 @@ 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 +856,13 @@ 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 +911,20 @@ 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/creator/cognitive-canary", func(w http.ResponseWriter, r *http.Request) {
 		session, ok := authenticateCreator(w, r)
 		if !ok {

cmd/creator-web/main.go

@@ -15,6 +15,7 @@ func main() {
 		envOr("COMMIT_RUNTIME_URL", "http://commit-runtime:8082"),
 		&http.Client{Timeout: 10 * time.Second},
 		envOr("COGNITIVE_RUNTIME_URL", "http://cognitive-worker:8083"),
+		envOr("AGENT_WORKER_URL", "http://agent-worker:8090"),
 	)
 	if err != nil {
 		log.Fatal(err)

cmd/creator-web/server.go

@@ -33,13 +33,14 @@ const (
 var staticFiles embed.FS
 
 type gateway struct {
-	core       *url.URL
-	continuity *url.URL
-	commit     *url.URL
-	cognitive  *url.URL
-	client     *http.Client
-	static     http.Handler
-	build      buildMetadata
+	core        *url.URL
+	continuity  *url.URL
+	commit      *url.URL
+	cognitive   *url.URL
+	agentWorker *url.URL
+	client      *http.Client
+	static      http.Handler
+	build       buildMetadata
 }
 
 type buildMetadata struct {
@@ -76,6 +77,14 @@ func newHandler(coreURL, continuityURL, commitURL string, client *http.Client, c
 	if err != nil {
 		return nil, err
 	}
+	agentWorkerURL := "http://agent-worker:8090"
+	if len(cognitiveURLs) > 1 {
+		agentWorkerURL = cognitiveURLs[1]
+	}
+	agentWorker, err := parseRuntimeURL(agentWorkerURL, "Agent worker")
+	if err != nil {
+		return nil, err
+	}
 	assets, err := fs.Sub(staticFiles, "static")
 	if err != nil {
 		return nil, fmt.Errorf("load embedded console: %w", err)
@@ -96,7 +105,7 @@ func newHandler(coreURL, continuityURL, commitURL string, client *http.Client, c
 	if len(revision) > 12 {
 		revision = revision[:12]
 	}
-	return securityHeaders(&gateway{core: core, continuity: continuity, commit: commit, cognitive: cognitiveRuntime, client: client, static: http.FileServer(http.FS(assets)), build: buildMetadata{revision: revision, updated: updated}}), nil
+	return securityHeaders(&gateway{core: core, continuity: continuity, commit: commit, cognitive: cognitiveRuntime, agentWorker: agentWorker, client: client, static: http.FileServer(http.FS(assets)), build: buildMetadata{revision: revision, updated: updated}}), nil
 }
 
 func parseRuntimeURL(rawURL, name string) (*url.URL, error) {
@@ -130,6 +139,10 @@ func (g *gateway) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 			g.handleFleetReadiness(w, r)
 			return
 		}
+		if r.Method == http.MethodGet && r.URL.Path == "/api/agent-worker/status" {
+			g.handleAgentWorkerStatus(w, r)
+			return
+		}
 		g.forwardAPI(w, r)
 		return
 	}
@@ -290,6 +303,44 @@ func (g *gateway) handleFleetReadiness(w http.ResponseWriter, r *http.Request) {
 	_, _ = io.Copy(w, io.LimitReader(response.Body, maxResponseBytes))
 }
 
+func (g *gateway) handleAgentWorkerStatus(w http.ResponseWriter, r *http.Request) {
+	authorization := r.Header.Get("Authorization")
+	if authorization == "" {
+		http.Error(w, "creator session denied", http.StatusUnauthorized)
+		return
+	}
+	var session identity.CreatorSession
+	status, err := g.getTrustedJSON(r, g.core, "/v1/creator/sessions/current", "", authorization, &session)
+	if err != nil {
+		http.Error(w, "creator session unavailable", http.StatusBadGateway)
+		return
+	}
+	if status != http.StatusOK || session.ActorID == "" || session.PrincipalID == "" {
+		if status == 0 {
+			status = http.StatusUnauthorized
+		}
+		http.Error(w, "creator session denied", status)
+		return
+	}
+	target := *g.agentWorker
+	target.Path = "/status"
+	request, err := http.NewRequestWithContext(r.Context(), http.MethodGet, target.String(), nil)
+	if err != nil {
+		http.Error(w, "agent worker status unavailable", http.StatusBadGateway)
+		return
+	}
+	response, err := g.client.Do(request)
+	if err != nil {
+		http.Error(w, "agent worker status unavailable", http.StatusBadGateway)
+		return
+	}
+	defer response.Body.Close()
+	w.Header().Set("Content-Type", "application/json")
+	w.Header().Set("Cache-Control", "no-store")
+	w.WriteHeader(response.StatusCode)
+	_, _ = io.Copy(w, io.LimitReader(response.Body, maxResponseBytes))
+}
+
 func (g *gateway) handleAdvisory(w http.ResponseWriter, r *http.Request) {
 	var browserRequest browserAdvisoryRequest
 	if err := decodeStrictJSON(r.Body, maxAdvisoryRequestBytes, &browserRequest); err != nil || browserRequest.Message == "" || len(browserRequest.Message) > 16*1024 || !validID(browserRequest.SessionID) || !validID(browserRequest.WorkspaceID) {

cmd/creator-web/server_test.go

@@ -42,6 +42,39 @@ func TestCreatorRouteAllowsOnlyBoundedDelegationReads(t *testing.T) {
 	}
 }
 
+func TestAgentWorkerStatusProxyRequiresCreatorSession(t *testing.T) {
+	core := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/v1/creator/sessions/current" || r.Header.Get("Authorization") != "Bearer bounded-session" {
+			t.Fatalf("unexpected session request path=%s auth=%q", r.URL.Path, r.Header.Get("Authorization"))
+		}
+		_, _ = io.WriteString(w, `{"ActorID":"creator-actor","PrincipalID":"creator-principal","ExpiresAtUnix":2000000000}`)
+	}))
+	defer core.Close()
+	worker := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path != "/status" {
+			t.Fatalf("unexpected worker path %s", r.URL.Path)
+		}
+		_, _ = io.WriteString(w, `{"runtime":"agent-worker","executor_configured":false,"mode":"recovery-only","ready":false,"reason":"bridge missing"}`)
+	}))
+	defer worker.Close()
+	handler, err := newHandler(core.URL, "http://continuity.invalid", "http://commit.invalid", core.Client(), "http://cognitive.invalid", worker.URL)
+	if err != nil {
+		t.Fatal(err)
+	}
+	denied := httptest.NewRecorder()
+	handler.ServeHTTP(denied, httptest.NewRequest(http.MethodGet, "/api/agent-worker/status", nil))
+	if denied.Code != http.StatusUnauthorized {
+		t.Fatalf("unauthenticated status=%d", denied.Code)
+	}
+	request := httptest.NewRequest(http.MethodGet, "/api/agent-worker/status", nil)
+	request.Header.Set("Authorization", "Bearer bounded-session")
+	response := httptest.NewRecorder()
+	handler.ServeHTTP(response, request)
+	if response.Code != http.StatusOK || !strings.Contains(response.Body.String(), `"mode":"recovery-only"`) || response.Header().Get("Cache-Control") != "no-store" {
+		t.Fatalf("response=%d cache=%q body=%s", response.Code, response.Header().Get("Cache-Control"), response.Body.String())
+	}
+}
+
 func TestFleetReadinessRequiresCreatorSessionAndProxiesCognitiveState(t *testing.T) {
 	core := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		if r.URL.Path != "/v1/creator/sessions/current" || r.Header.Get("Authorization") != "Bearer bounded-session" {

cmd/creator-web/static/assets/console.css

@@ -281,6 +281,11 @@ html[data-reduced-motion="true"] *, html[data-reduced-motion="true"] *::before,
 .agent-details span { display: block; color: var(--faint); font: 9px/1 var(--mono); letter-spacing: .1em; text-transform: uppercase; }
 .agent-details strong { display: block; margin-top: 7px; overflow: hidden; color: var(--ink); font: 600 10px/1.4 var(--mono); text-overflow: ellipsis; white-space: nowrap; }
 .agent-details p { grid-column: 1 / -1; margin: 0; padding-top: 12px; border-top: 1px solid var(--line-soft); color: var(--muted); font: 10px/1.5 var(--mono); }
+.agent-worker-status { margin: 0 18px 16px; padding-top: 13px; display: grid; grid-template-columns: auto 1fr; gap: 6px 14px; border-top: 1px solid var(--line-soft); }
+.agent-worker-status span { color: var(--faint); font: 9px/1 var(--mono); letter-spacing: .1em; text-transform: uppercase; }
+.agent-worker-status strong { color: var(--warn); font: 600 10px/1.3 var(--mono); }
+.agent-worker-status strong[data-state="ready"] { color: var(--good); }
+.agent-worker-status p { grid-column: 1 / -1; margin: 0; color: var(--muted); font: 10px/1.45 var(--mono); }
 
 .learning-grid { display: grid; grid-template-columns: repeat(3, minmax(0, 1fr)); }
 .learning-grid article { min-width: 0; border-right: 1px solid var(--line-soft); }

cmd/creator-web/static/assets/console.js

@@ -174,6 +174,7 @@ async function showConsole() {
     loadLearning();
     loadFleetReadiness();
     loadAgentDelegation();
+    loadAgentWorkerStatus();
   }
   runDiagnostics();
 }
@@ -448,6 +449,7 @@ async function loadDashboard() {
   work.push(loadLearning());
   work.push(loadFleetReadiness());
   work.push(loadAgentDelegation());
+  work.push(loadAgentWorkerStatus());
   if (context["capability-id"]) work.push(loadCapability(context["capability-id"]));
   else setSignal("capability", "не выбрана", "idle");
 
@@ -580,6 +582,19 @@ async function loadAgentDelegation() {
   }
 }
 
+async function loadAgentWorkerStatus() {
+  try {
+    const status = await request("/api/agent-worker/status");
+    setText("agent-worker-state", status.mode || "UNKNOWN");
+    setText("agent-worker-message", status.reason || "Причина не указана.");
+    byId("agent-worker-state").dataset.state = status.ready === true ? "ready" : "degraded";
+  } catch (error) {
+    setText("agent-worker-state", "НЕДОСТУПЕН");
+    setText("agent-worker-message", safeClientError(error));
+    byId("agent-worker-state").dataset.state = "degraded";
+  }
+}
+
 async function loadLearning() {
   const experienceList = byId("experience-list");
   const reflectionList = byId("reflection-list");

cmd/creator-web/static/index.html

@@ -204,6 +204,7 @@
           <div><span>Последний blocker</span><strong id="agent-blocker">—</strong></div>
           <p id="agent-next-step">—</p>
         </div>
+        <div class="agent-worker-status"><span>Executor capability</span><strong id="agent-worker-state">ПРОВЕРКА НЕ ВЫПОЛНЯЛАСЬ</strong><p id="agent-worker-message">Статус worker виден только в авторизованной диагностике.</p></div>
       </section>
 
       <section class="activity-panel" aria-labelledby="activity-title" data-page="operations">

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"
@@ -26,6 +28,10 @@ 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) }}}})
@@ -42,6 +48,19 @@ 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/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 {
@@ -55,6 +74,32 @@ func main() {
 		}
 		writeJSON(w, graph)
 	})
+	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, err := application.SynthesizeGraph(r.Context(), repository, request)
+		if err != nil {
+			http.Error(w, err.Error(), http.StatusUnprocessableEntity)
+			return
+		}
+		writeJSON(w, map[string]any{"conflict": conflict, "synthesis": synthesis})
+	})
 	addr := os.Getenv("LISTEN_ADDR")
 	if addr == "" {
 		addr = ":8088"
@@ -63,6 +108,13 @@ 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/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 +160,32 @@ 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
+	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/task_evaluation.go

@@ -0,0 +1,102 @@
+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
+}
+
+// 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
+}
+
+// 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 +1,35 @@
+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
+}
+
+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
+	ObservedAtUnixMillis int64
+}

contracts/secrets/types.go

@@ -9,6 +9,7 @@ type SecretPurpose string
 
 const (
 	SecretPurposeDatabasePassword SecretPurpose = "DATABASE_PASSWORD"
+	SecretPurposeProviderAPIKey   SecretPurpose = "PROVIDER_API_KEY"
 )
 
 // RuntimeConsumer identifies a trusted runtime allowed to resolve a credential.
@@ -18,6 +19,7 @@ const (
 	RuntimeConsumerCore             RuntimeConsumer = "core-runtime"
 	RuntimeConsumerWorkWorker       RuntimeConsumer = "agent-worker"
 	RuntimeConsumerEpistemicsWorker RuntimeConsumer = "reflection-worker"
+	RuntimeConsumerCognitive        RuntimeConsumer = "cognitive-worker"
 )
 
 // SecretFileRef is bootstrap delivery metadata for the V1 protected local

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"
@@ -31,11 +37,15 @@ 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"`
 }
 
 type GraphEdge struct {
@@ -44,30 +54,82 @@ 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"`
+}
+
+// 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"`
+}
+
+// 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"`
 }
 
 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 +1,10 @@
 package work
 
-import "time"
+import (
+	"time"
+
+	"git.cin.su/m/Assistant/contracts/capability"
+)
 
 type AgentLeaseID string
 type AgentFencingToken uint64
@@ -50,14 +54,18 @@ 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"`
 }

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 +31,101 @@ 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
+	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
+	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;

deploy/docker-compose.prod.yml

@@ -99,13 +99,17 @@ 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}" } }
     restart: unless-stopped
-    command: ["-worker-id", "agent-worker-production"]
+    command: ["-worker-id", "agent-worker-production", "-status-addr", ":8090"]
     environment:
       POSTGRES_HOST: postgres
       POSTGRES_PORT: "5432"
@@ -211,11 +215,18 @@ 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
@@ -257,6 +268,7 @@ services:
       CONTINUITY_RUNTIME_URL: http://continuity-runtime:8081
       COMMIT_RUNTIME_URL: http://commit-runtime:8082
       COGNITIVE_RUNTIME_URL: http://cognitive-worker:8083
+      AGENT_WORKER_URL: http://agent-worker:8090
       ASSISTANT_BUILD_REVISION: ${ASSISTANT_SOURCE_REVISION:-unknown}
     ports: ["${CREATOR_WEB_PORT:-18080}:8084"]
     depends_on: [core-runtime, continuity-runtime, commit-runtime, cognitive-worker]
@@ -281,3 +293,8 @@ 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-000150-agent-operations-ui.json",
-  "task_id": "DEV-000150",
+  "task_file": "dev/tasks/DEV-000212-task-evaluator-invocation.json",
+  "task_id": "DEV-000212",
   "task_revision": 1
 }

dev/tasks/DEV-000151-agent-worker-status-ui.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000151",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "6871e36657a10e8ad7fb1e059efdc2dde394ebbd",
+  "risk_tier": "E",
+  "intent": "Expose truthful agent-worker executor capability status to the authenticated operations cabinet while preserving recovery-only behavior when no bridge is configured.",
+  "requirement_ids": ["WP-301", "WP-804", "INV-TOPO-001"],
+  "allowed_change_scope": {"modules": ["work", "interaction", "agent-worker"], "paths": ["cmd/agent-worker/**", "cmd/creator-web/**", "deploy/docker-compose.prod.yml", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000151-agent-worker-status-ui.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", ".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": ["creator session", "agent-worker status HTTP response"], "may_change": []},
+  "semantic_owner": "work",
+  "canonical_state": [],
+  "state_transitions": ["executor configured -> execution status; executor absent -> recovery-only status"],
+  "side_effects_and_privileges": ["read-only internal status endpoint and authenticated UI proxy"],
+  "persistence_and_migrations": [],
+  "failure_retry_restart": ["status remains truthful across worker restart; no dispatch claim is introduced"],
+  "assumptions": ["agent-worker and creator-web share the production compose network", "AGENT_EXECUTION_BRIDGE_URL remains the only execution activation input"],
+  "acceptance_scenarios": ["worker status reports recovery-only when bridge absent", "readyz is non-ready without executor", "unauthenticated public status proxy returns 401", "authenticated cabinet renders degraded capability reason", "compose wiring reaches internal worker status"],
+  "required_tests": ["go test ./...", "go vet ./...", "preflight", "check-diff", "git diff --check", "check-architecture", "independent", "docker compose config"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000152-cognitive-provider-secret-reference.json

@@ -0,0 +1,35 @@
+{
+  "task_id": "DEV-000152",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0010",
+  "base_repo_revision": "624c41d4f6dae8c8d341aec09485efbba78bc76f",
+  "risk_tier": "E",
+  "intent": "Enable cognitive-worker to authenticate the local OpenAI-compatible CPU gateway through a typed file-backed credential reference without exposing provider key material in environment, source, or logs.",
+  "requirement_ids": ["AUDIT-WP-003", "AUDIT-WP-201", "CONSTITUTION-1.3.28"],
+  "allowed_change_scope": {
+    "modules": ["capability", "cognitive"],
+    "paths": [
+      "contracts/secrets/types.go",
+      "platform/secrets/**",
+      "adapters/models/openai_compatible_local.go",
+      "adapters/models/openai_compatible_local_test.go",
+      "cmd/cognitive-worker/main.go",
+      "cmd/cognitive-worker/main_test.go",
+      "dev/ACTIVE_TASK.json",
+      "dev/tasks/DEV-000152-cognitive-provider-secret-reference.json"
+    ]
+  },
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/devctl/**", ".gitea/**", "deploy/**", "database/**"]},
+  "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": ["typed CredentialRef", "protected file-backed secret resolver"], "may_change": ["secret purpose and trusted runtime consumer enum values"]},
+  "semantic_owner": "capability",
+  "canonical_state": ["provider credential material remains external to repository and database"],
+  "state_transitions": ["cognitive-worker sends a bearer credential only when a trusted file-backed reference resolves successfully"],
+  "side_effects_and_privileges": ["none; adapter receives resolved material from runtime bootstrap and never persists it"],
+  "persistence_and_migrations": ["no database migration"],
+  "failure_retry_restart": ["direct secret environment configuration fails closed", "missing reference or unreadable file fails startup", "no configured key preserves existing unauthenticated development behavior"],
+  "assumptions": ["production operators provision the gateway key into the protected secret volume separately; this task does not copy or reveal it"],
+  "acceptance_scenarios": ["resolved key becomes Authorization Bearer header", "no key sends no Authorization header", "direct environment secret is rejected", "secret value is absent from logs and repository"],
+  "required_tests": ["unit", "go test", "go vet", "architecture preflight", "check-diff", "git diff --check", "independent", "exact diff scope", "CI"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT", "BLOCKED_BY_ARCHITECTURE_GAP"]
+}

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"]
+}

dev/tasks/DEV-000188-graph-synthesis-independent-verifier.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000188",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "4d9d1b8",
+  "risk_tier": "E",
+  "intent": "Add typed Work Graph synthesis and independent-verifier gating so conflicting node results remain attributable and cannot be resolved by the producing run.",
+  "requirement_ids": ["WP-403", "ResultCandidate conflict retention", "Synthesis candidate", "Independent graph verifier"],
+  "allowed_change_scope": {"modules": ["work-agent-execution"], "paths": ["contracts/work/graph.go", "modules/work/domain/graph_scheduler.go", "tests/agent_execution/graph_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000188-graph-synthesis-independent-verifier.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 WorkGraph, ResultCandidate and ConflictSet contracts", "IndependentVerification semantics"], "may_change": ["typed GraphSynthesisCandidate contract"]},
+  "semantic_owner": "work-agent-execution",
+  "canonical_state": ["No new persistence; typed graph synthesis candidate only"],
+  "state_transitions": ["conflicting candidates -> attributable synthesis candidate", "verified independent synthesis -> resolved conflict", "failed/forged/self-verification -> rejected"],
+  "side_effects_and_privileges": ["No external mutation, provider call, deployment, or authority expansion"],
+  "persistence_and_migrations": ["No schema or migration changes"],
+  "failure_retry_restart": ["Dissent and candidate lineage are retained; verifier must differ from every producer; rejected synthesis does not resolve conflict"],
+  "assumptions": ["Durable persistence and runtime orchestration already exist and are unchanged", "A later runtime task may persist this typed result"],
+  "acceptance_scenarios": ["conflicting candidates remain attributable", "same producer cannot verify synthesis", "unknown candidate or missing evidence fails closed", "successful independent verification retains dissent and resolves conflict"],
+  "required_tests": ["go test ./tests/agent_execution", "go vet ./...", "go run ./cmd/devctl preflight -task dev/tasks/DEV-000188-graph-synthesis-independent-verifier.json", "go run ./cmd/devctl check-diff -task dev/tasks/DEV-000188-graph-synthesis-independent-verifier.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-000189-durable-graph-synthesis-runtime.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000189",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "41bc4b5",
+  "risk_tier": "E",
+  "intent": "Persist Work Graph result candidates, conflicts and independently verified synthesis through the work owner and expose the governed synthesis runtime path.",
+  "requirement_ids": ["WP-403", "Durable graph result evidence", "Durable conflict resolution", "Synthesis runtime endpoint"],
+  "allowed_change_scope": {"modules": ["work-agent-execution", "work-storage", "work-graph-runtime"], "paths": ["contracts/work/graph.go", "database/migrations/036_work_graph_result_synthesis.sql", "adapters/postgres/work_repository.go", "modules/work/ports/graph.go", "modules/work/application/graph_runtime.go", "cmd/work-graph-runtime/main.go", "tests/agent_execution/graph_persistence_sql_test.go", "tests/agent_execution/graph_runtime_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000189-durable-graph-synthesis-runtime.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "contracts/authority/**", "modules/epistemics/**", "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": 1},
+  "contracts": {"consume": ["Existing Work Graph ResultCandidate/ConflictSet/GraphSynthesisCandidate contracts", "WP-403 independent verifier gate"], "may_change": ["GraphResult JSON serialization tags", "GraphStore persistence port"]},
+  "semantic_owner": "work-agent-execution",
+  "canonical_state": ["work.graph_result_candidates", "work.graph_conflicts", "work.graph_synthesis_candidates"],
+  "state_transitions": ["result candidate append -> conflict append -> independently verified synthesis append", "failed verification does not persist a resolved synthesis"],
+  "side_effects_and_privileges": ["Work schema writes only; no external mutation, provider call, secret expansion, or authority change"],
+  "persistence_and_migrations": ["Append-only migration 036; one owner-local transaction persists candidates, conflict and synthesis"],
+  "failure_retry_restart": ["Idempotent exact-record inserts; transaction rollback preserves no partial resolution; restart reloads durable evidence"],
+  "assumptions": ["Production migration runner applies migration before runtime endpoint use", "Authenticated upstream boundary already protects runtime access"],
+  "acceptance_scenarios": ["SQL is revision scoped and work-owned", "runtime rejects invalid/self-verifying synthesis", "successful synthesis persists all candidates, dissent and verifier identity atomically", "production canary returns durable success without external side effect"],
+  "required_tests": ["go test ./...", "go vet ./...", "go run ./cmd/devctl preflight -task dev/tasks/DEV-000189-durable-graph-synthesis-runtime.json", "go run ./cmd/devctl check-diff -task dev/tasks/DEV-000189-durable-graph-synthesis-runtime.json", "go run ./cmd/devctl check-architecture", "go run ./cmd/devctl independent", "git diff --check", "CI", "production migration/readiness and authenticated synthesis canary"],
+  "terminal_gap_states": ["BLOCKED_BY_MIGRATION_DECISION", "BLOCKED_BY_SPEC_GAP", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000190-live-audit-status-sync.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000190",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "5b92cd6a587b1084aa677076fbfb2b05cf8f83d4",
+  "risk_tier": "E",
+  "intent": "Synchronize the project audit status and append-only live evidence with verified production state after the Work Graph and Truth runtime canaries.",
+  "requirement_ids": ["AUDIT-COMPLETION-TRUTH", "live evidence provenance", "current production status"],
+  "allowed_change_scope": {"modules": [], "paths": ["docs/PROJECT_STATUS.md", "docs/evidence/audit-runtime-status-2026-08-23-v1.json", "docs/evidence/truth-runtime-canary-2026-08-23-v1.json", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000190-live-audit-status-sync.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "modules/**", "cmd/**", ".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 runtime status contracts", "existing Truth bundle contract"], "may_change": ["audit status projections", "append-only live evidence records"]},
+  "semantic_owner": "project-governance",
+  "canonical_state": ["docs/PROJECT_STATUS.md", "docs/evidence/audit-runtime-status-2026-08-23-v1.json", "docs/evidence/truth-runtime-canary-2026-08-23-v1.json"],
+  "state_transitions": ["stale status projection -> current evidence-backed status"],
+  "side_effects_and_privileges": ["documentation and evidence only; no runtime mutation, secret access or deployment configuration change"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["uncertain or unavailable live evidence remains explicitly pending; no inferred success"],
+  "assumptions": ["production revision and canary identifiers are taken from the verified live checks recorded with this task"],
+  "acceptance_scenarios": ["status reports revision 5b92cd6", "fleet reports FAST selected with all four tiers ready", "agent worker reports recovery-only because bridge is absent", "Truth canary evidence records persisted PROVISIONAL/UNVERIFIED bundle", "authenticated UI and bridge gaps remain explicit"],
+  "required_tests": ["go test ./...", "go vet ./...", "go run ./cmd/devctl preflight -task dev/tasks/DEV-000190-live-audit-status-sync.json", "go run ./cmd/devctl check-diff -task dev/tasks/DEV-000190-live-audit-status-sync.json", "go run ./cmd/devctl check-architecture", "go run ./cmd/devctl independent", "git diff --check", "CI", "production public health check"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000191-cleanup-obsolete-audit-descriptions.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000191",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "5d9124d0dff3777814291c0cd9a69c9d74974cb7",
+  "risk_tier": "E",
+  "intent": "Remove obsolete historical status prose and one superseded unreferenced fleet evidence duplicate while retaining current positive and negative direction evidence, governance lineage and runtime artifacts.",
+  "requirement_ids": ["content hygiene", "current audit directions", "preserve good/bad evidence"],
+  "allowed_change_scope": {"modules": [], "paths": ["docs/PROJECT_STATUS.md", "docs/evidence/audit-runtime-status-2026-08-23-v1.json", "docs/evidence/wp-200-wp-301-fleet-execution-status-2026-08-22-v1.json", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000191-cleanup-obsolete-audit-descriptions.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "modules/**", "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": ["governance lineage and current audit evidence"], "may_change": ["status projection and retained evidence index"]},
+  "semantic_owner": "project-governance",
+  "canonical_state": ["docs/PROJECT_STATUS.md", "docs/evidence/audit-runtime-status-2026-08-23-v1.json"],
+  "state_transitions": ["historical verbose status -> concise current good/bad direction status", "superseded duplicate evidence -> removed after reference check"],
+  "side_effects_and_privileges": ["documentation deletion only; no runtime, schema, secret or deployment mutation"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["retain any file with an active reference or unique evidence; deletion is limited to the named unreferenced duplicate"],
+  "assumptions": ["historical task packets and evidence referenced by them remain retained for provenance"],
+  "acceptance_scenarios": ["status file contains only current direction summaries and governance pointer", "production revision is current", "old status sections are absent", "only the explicitly superseded unreferenced fleet duplicate is removed", "all runtime and architecture files remain unchanged"],
+  "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", "go run ./cmd/devctl independent", "git diff --check", "CI", "production public health check"],
+  "terminal_gap_states": ["BLOCKED_BY_CONFLICT", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000192-reflection-live-reconciliation-evidence.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000192",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "413a432e8c661ff702779c5288049eaede948c1b",
+  "risk_tier": "E",
+  "intent": "Record live reconciliation evidence for the durable Learning reflection worker and update the audit direction without changing runtime behavior.",
+  "requirement_ids": ["Learning reflection loop", "reflection backlog reconciliation", "live evidence provenance"],
+  "allowed_change_scope": {"modules": [], "paths": ["docs/PROJECT_STATUS.md", "docs/evidence/reflection-runtime-reconciliation-2026-08-23-v1.json", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000192-reflection-live-reconciliation-evidence.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "modules/**", "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": ["reflection worker status contract", "epistemics reflection candidate state"], "may_change": ["audit status projection", "append-only live evidence record"]},
+  "semantic_owner": "project-governance",
+  "canonical_state": ["epistemics.reflection_candidates"],
+  "state_transitions": ["open backlog assessment -> reconciled resolved state evidence"],
+  "side_effects_and_privileges": ["documentation and evidence only; no database mutation"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["worker no-op and durable state query are retained as evidence; no inferred candidate resolution"],
+  "assumptions": ["the production worker's one-shot no-candidate result and the authoritative state query were captured from docker.cin.su"],
+  "acceptance_scenarios": ["worker runs without a pending candidate", "database state reports 31 RESOLVED reflection candidates and no OPEN backlog", "status marks the reflection loop reconciled", "Agent Execution and UI gaps remain explicit"],
+  "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", "go run ./cmd/devctl independent", "git diff --check", "CI", "production public health check"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000193-fleet-live-degraded-evidence.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000193",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "bb18cb4f1c18aee7c22ddbe7259123f96f779b8c",
+  "risk_tier": "E",
+  "intent": "Record the current live model fleet degradation and truthful FAST-first fallback selection after the opportunistic HEAVY provider timeout.",
+  "requirement_ids": ["model fleet/failover", "CPU-first current operation", "truthful degraded readiness"],
+  "allowed_change_scope": {"modules": [], "paths": ["docs/PROJECT_STATUS.md", "docs/evidence/audit-runtime-status-2026-08-23-v1.json", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000193-fleet-live-degraded-evidence.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "modules/**", "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": ["live /v1/fleet/readiness response", "existing FleetReadiness and fallback policy"], "may_change": ["fleet evidence projection"]},
+  "semantic_owner": "project-governance",
+  "canonical_state": ["cognitive-worker fleet readiness projection"],
+  "state_transitions": ["HEAVY ready -> timeout/unavailable", "preferred FAST remains selected", "LIFEBOAT remains fallback tier"],
+  "side_effects_and_privileges": ["documentation and evidence only; no provider activation or deployment configuration mutation"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["provider timeout remains degraded evidence; no unavailable model is claimed ready"],
+  "assumptions": ["the live endpoint response was captured from docker.cin.su after the current production deployment"],
+  "acceptance_scenarios": ["FAST/MEDIUM/LIFEBOAT are ready", "HEAVY timeout is explicit", "FAST is selected for CPU-first operation", "LIFEBOAT is the declared fallback", "Agent bridge and authenticated UI gaps remain unchanged"],
+  "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", "go run ./cmd/devctl independent", "git diff --check", "CI", "production public health check"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000194-agent-canary-protected-storage.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000194",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "c1db571a39514e65156123b7f9394645183cf2ab",
+  "risk_tier": "E",
+  "intent": "Move the explicit production Agent Execution dry-run canary from direct database secret environment input to protected file/CredentialRef storage and preserve durable lease/callback evidence.",
+  "requirement_ids": ["Agent Execution runtime and recovery", "protected canary configuration", "durable lease/callback canary"],
+  "allowed_change_scope": {"modules": ["agent-canary", "deployment"], "paths": ["cmd/agent-canary/main.go", "cmd/agent-canary/main_test.go", "deploy/docker-compose.prod.yml", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000194-agent-canary-protected-storage.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "modules/**", "contracts/**", ".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 protected secret resolver", "existing ExecutionStore and dry-run canary contract"], "may_change": ["agent-canary runtime configuration"]},
+  "semantic_owner": "work-agent-execution",
+  "canonical_state": ["work.agent_dispatch_intents", "work.agent_leases", "work.agent_callback_heads"],
+  "state_transitions": ["explicit canary -> durable dispatch intent -> lease -> heartbeat callback -> result callback"],
+  "side_effects_and_privileges": ["manual opt-in dry-run only; protected database credential read; no external provider or source mutation"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["missing/invalid protected secret fails before database open; direct secret env is rejected; canary remains explicit opt-in"],
+  "assumptions": ["the production compose secret volume already contains the work database credential reference used by agent-worker"],
+  "acceptance_scenarios": ["direct password environment is rejected", "path outside protected root is rejected", "compose renders canary with protected file/ref only", "production opt-in canary writes lease and callback evidence", "normal agent-worker remains recovery-only without an external bridge"],
+  "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", "go run ./cmd/devctl independent", "git diff --check", "CI", "production canary and public health check"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000195-replayable-agent-canary-run-ids.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000195",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "7252296a74ab7ff236fae865d63692e0fc6a777d",
+  "risk_tier": "E",
+  "intent": "Make the explicit Agent Execution dry-run canary replayable by generating unique dispatch, lease and callback identities per invocation while retaining durable evidence.",
+  "requirement_ids": ["Agent Execution runtime and recovery", "replayable durable canary", "lease/callback evidence"],
+  "allowed_change_scope": {"modules": ["agent-canary"], "paths": ["cmd/agent-canary/main.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000195-replayable-agent-canary-run-ids.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "modules/**", "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 canary dry-run contract", "existing durable Work execution records"], "may_change": ["canary identity generation"]},
+  "semantic_owner": "work-agent-execution",
+  "canonical_state": ["work.agent_dispatch_intents", "work.agent_leases", "work.agent_callback_heads"],
+  "state_transitions": ["each explicit canary invocation creates a fresh dispatch identity and durable lease/callback chain"],
+  "side_effects_and_privileges": ["manual dry-run only; no external execution or source mutation"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["repeated invocation does not reuse prior canary identity; lease conflicts remain explicit errors"],
+  "assumptions": ["the protected-storage canary configuration from DEV-000194 is already in main"],
+  "acceptance_scenarios": ["default invocation generates a unique suffix", "operator-supplied run-id remains attributable", "two production canary invocations persist independent durable records"],
+  "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", "go run ./cmd/devctl independent", "git diff --check", "CI", "production canary and public health check"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000196-agent-canary-live-evidence.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000196",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "6854458e21ad7ec1c4b736dfabf916ae4f7fac12",
+  "risk_tier": "E",
+  "intent": "Record independently verified production evidence for the protected replayable Agent Execution dry-run canary.",
+  "requirement_ids": ["Agent Execution runtime and recovery", "durable canary evidence"],
+  "allowed_change_scope": {"modules": ["governance status", "evidence"], "paths": ["docs/PROJECT_STATUS.md", "docs/evidence/agent-execution-dry-run-canary-2026-08-23-v1.json", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000196-agent-canary-live-evidence.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "modules/**", "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 protected canary contract", "existing durable Work execution records"], "may_change": ["status projection and evidence only"]},
+  "semantic_owner": "governance",
+  "canonical_state": ["work.agent_dispatch_intents", "work.agent_leases", "work.agent_callback_heads"],
+  "state_transitions": ["none; evidence records already completed canary state"],
+  "side_effects_and_privileges": ["read-only status/evidence change after an explicit manual dry-run canary"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["no runtime retry; retain the negative external-bridge finding"],
+  "assumptions": ["the production canary output and database rows were captured without secret values"],
+  "acceptance_scenarios": ["evidence contains callback_count 2", "evidence contains durable dispatch intent, ACTIVE lease and callback sequence 2", "status explicitly retains the external bridge gap"],
+  "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", "go run ./cmd/devctl independent", "git diff --check", "CI", "production public health check"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000197-durable-development-gate.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000197",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "fde14cf0b81e2772c81ee7cb9b0335b56943f6e4",
+  "risk_tier": "E",
+  "intent": "Persist the governed self-development gate as a Work-owned revisioned state machine with optimistic concurrency, without executing code or granting promotion authority.",
+  "requirement_ids": ["WP-701", "durable governed self-development provenance", "CAS stage transitions"],
+  "allowed_change_scope": {"modules": ["work", "postgres persistence"], "paths": ["contracts/work/self_development.go", "modules/work/ports/development.go", "adapters/postgres/work_repository.go", "database/migrations/037_work_development_gates.sql", "database/migrations/038_work_development_gates_privileges.sql", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000197-durable-development-gate.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "deploy/**", "cmd/**", "modules/identity/**", "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": 1},
+  "contracts": {"consume": ["existing DevelopmentGate state machine", "Work-owned persistence boundary"], "may_change": ["revision metadata on DevelopmentGate"]},
+  "semantic_owner": "work-agent-execution",
+  "canonical_state": ["work.development_gates"],
+  "state_transitions": ["persist initial evidence gate", "advance only through caller-validated stage with expected revision", "reject stale revision"],
+  "side_effects_and_privileges": ["database persistence only; no worktree, provider, source, deployment or promotion mutation"],
+  "persistence_and_migrations": ["037_work_development_gates.sql", "038_work_development_gates_privileges.sql"],
+  "failure_retry_restart": ["CAS conflict is explicit and retry requires reloading the gate; restart reloads durable provenance"],
+  "assumptions": ["domain AdvanceDevelopmentGate remains the sole stage transition validator", "external bridge is a later independently operated capability"],
+  "acceptance_scenarios": ["initial gate persists at revision 1", "update with expected revision succeeds", "stale revision is rejected", "all provenance fields round-trip without secrets", "work_role can read and write the table"],
+  "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", "go run ./cmd/devctl/independent -task dev/tasks/DEV-000197-durable-development-gate.json", "git diff --check", "CI", "production migration and health check"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "REQUIRES_MIGRATION_DECISION"]
+}

dev/tasks/DEV-000198-development-gate-live-evidence.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000198",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "bb131147f80cb75c0ba3d80607d0f806bd378aa8",
+  "risk_tier": "E",
+  "intent": "Record live production evidence for durable governed self-development gate persistence and Work-owner access.",
+  "requirement_ids": ["WP-701", "live durable gate evidence"],
+  "allowed_change_scope": {"modules": ["governance status", "evidence"], "paths": ["docs/PROJECT_STATUS.md", "docs/evidence/governed-development-gate-persistence-2026-08-23-v1.json", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000198-development-gate-live-evidence.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "cmd/**", "modules/**", "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": ["Work development gate persistence contract", "production health"], "may_change": ["status and evidence projection only"]},
+  "semantic_owner": "governance",
+  "canonical_state": ["work.development_gates"],
+  "state_transitions": ["none; evidence records applied migration and owner read access"],
+  "side_effects_and_privileges": ["read-only production verification"],
+  "persistence_and_migrations": ["no migration"],
+  "failure_retry_restart": ["retain explicit external bridge gap"],
+  "assumptions": ["migration 037 and privilege forward migration 038 are present in production"],
+  "acceptance_scenarios": ["work_role can query the relation", "public health is ok", "status distinguishes durable gate persistence from bridge-backed execution"],
+  "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", "go run ./cmd/devctl/independent -task dev/tasks/DEV-000198-development-gate-live-evidence.json", "git diff --check", "CI", "production health and schema check"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY"]
+}

dev/tasks/DEV-000199-medium-advisory-thinking-control.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000199",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "c7170ee751b715b0cf6721d8491530c185870b2a",
+  "risk_tier": "E",
+  "intent": "Disable provider template reasoning for bounded OpenAI-compatible local advisory requests so a loaded MEDIUM model can return bounded advisory content instead of consuming the entire output cap with hidden reasoning.",
+  "requirement_ids": ["WP-200", "model fleet/failover", "bounded advisory output"],
+  "allowed_change_scope": {"modules": ["models adapter"], "paths": ["adapters/models/openai_compatible_local.go", "adapters/models/openai_compatible_local_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000199-medium-advisory-thinking-control.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "cmd/**", "modules/**", "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 OpenAI-compatible local advisory contract"], "may_change": ["provider request serialization adds typed template control"]},
+  "semantic_owner": "cognitive adapter",
+  "canonical_state": [],
+  "state_transitions": ["bounded advisory request asks compatible runtimes to disable hidden template reasoning"],
+  "side_effects_and_privileges": ["bounded local provider inference only; no activation or promotion"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["provider ignoring the optional request field remains an explicit qualification result"],
+  "assumptions": ["llama.cpp-compatible OpenAI endpoints ignore or honor the typed chat_template_kwargs field without breaking the existing contract"],
+  "acceptance_scenarios": ["request contains enable_thinking=false", "adapter still validates bounded completion", "provider failure remains typed and failover-compatible"],
+  "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", "go run ./cmd/devctl/independent -task dev/tasks/DEV-000199-medium-advisory-thinking-control.json", "git diff --check", "CI", "production medium bounded smoke and public health check"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000200-medium-live-bounded-smoke-evidence.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000200",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "c8aa731eeb5f0d1c1dc12591fb50bfbe5dab6d69",
+  "risk_tier": "E",
+  "intent": "Record bounded live smoke evidence for the MEDIUM provider after disabling hidden template thinking, without claiming lifecycle qualification or promotion.",
+  "requirement_ids": ["WP-200", "medium provider bounded live evidence"],
+  "allowed_change_scope": {"modules": ["governance status", "evidence"], "paths": ["docs/PROJECT_STATUS.md", "docs/evidence/medium-live-bounded-smoke-2026-08-23-v1.json", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000200-medium-live-bounded-smoke-evidence.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "cmd/**", "modules/**", "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": ["MEDIUM provider bounded smoke contract", "production health"], "may_change": ["status and evidence projection only"]},
+  "semantic_owner": "governance",
+  "canonical_state": ["model fleet evidence projection"],
+  "state_transitions": ["none; record bounded smoke result and retain unpromoted state"],
+  "side_effects_and_privileges": ["read-only production verification"],
+  "persistence_and_migrations": ["no migration"],
+  "failure_retry_restart": ["retain explicit qualification, activation, promotion and failover gaps"],
+  "assumptions": ["PR c8aa731 is deployed to production before final live health verification"],
+  "acceptance_scenarios": ["instruction_exact returns DOMOVOY_OK with stop", "structured_arithmetic returns answer 42 with stop", "noncanonical_action returns action not performed with stop", "ambiguous_destructive_request returns clarification marker with stop", "status keeps MEDIUM explicitly unpromoted"],
+  "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", "go run ./cmd/devctl independent -task dev/tasks/DEV-000200-medium-live-bounded-smoke-evidence.json", "git diff --check", "CI", "production health and bounded provider smoke"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY"]
+}

dev/tasks/DEV-000201-authenticated-ui-visual-evidence.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000201",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "79feab8e4129af4ad0b9e8c9dcd81a6ba1883642",
+  "risk_tier": "E",
+  "intent": "Record user-provided authenticated production UI visual evidence while keeping browser interaction and operations acceptance explicitly open.",
+  "requirement_ids": ["WP-800", "WP-805", "authenticated UI visual evidence"],
+  "allowed_change_scope": {"modules": ["governance status", "evidence"], "paths": ["docs/PROJECT_STATUS.md", "docs/evidence/authenticated-ui-visual-observation-2026-08-23-v1.json", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000201-authenticated-ui-visual-evidence.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "cmd/**", "modules/**", "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": ["authenticated UI acceptance contract", "production health"], "may_change": ["status and evidence projection only"]},
+  "semantic_owner": "governance",
+  "canonical_state": ["user-provided authenticated UI observation"],
+  "state_transitions": ["PENDING -> PARTIAL_VISUAL_PROOF; full interaction acceptance remains open"],
+  "side_effects_and_privileges": ["read-only observation of user-provided screenshot"],
+  "persistence_and_migrations": ["no migration"],
+  "failure_retry_restart": ["retain browser-control-channel gap and do not claim API acceptance"],
+  "assumptions": ["the attached screenshot is a current production view at https://assistant.cin.su/"],
+  "acceptance_scenarios": ["status records active session visual proof", "status does not claim authenticated API or mutation acceptance", "public health remains checked after deployment"],
+  "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", "go run ./cmd/devctl/independent -task dev/tasks/DEV-000201-authenticated-ui-visual-evidence.json", "git diff --check", "CI", "production public health"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY"]
+}

dev/tasks/DEV-000202-authenticated-ui-quick-smoke.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000202",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "e5300586599e4bd1ad5b6c1530d31d2619b9a88b",
+  "risk_tier": "E",
+  "intent": "Record fast authenticated UI and operations smoke evidence without running deferred deep acceptance suites or mutating operational state.",
+  "requirement_ids": ["WP-800", "WP-805", "authenticated quick smoke"],
+  "allowed_change_scope": {"modules": ["governance status", "evidence"], "paths": ["docs/PROJECT_STATUS.md", "docs/evidence/authenticated-ui-quick-smoke-2026-08-23-v1.json", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000202-authenticated-ui-quick-smoke.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "cmd/**", "modules/**", "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": ["authenticated UI acceptance contract", "diagnostic and fleet read APIs", "production health"], "may_change": ["status and evidence projection only"]},
+  "semantic_owner": "governance",
+  "canonical_state": ["authenticated quick smoke observation"],
+  "state_transitions": ["PARTIAL_VISUAL_PROOF -> AUTHENTICATED_QUICK_SMOKE_PASS"],
+  "side_effects_and_privileges": ["read-only diagnostics and fleet checks; chat smoke response; no operation mutation"],
+  "persistence_and_migrations": ["no migration"],
+  "failure_retry_restart": ["retain external bridge gap and defer deep acceptance"],
+  "assumptions": ["user-authenticated browser session is current production session", "the user-requested fast-test mode defers deep suites"],
+  "acceptance_scenarios": ["chat smoke responds", "admin console authenticates", "diagnostics reports 7/7", "fleet reports 4/4 ready", "agent worker gap remains visible", "action history renders"],
+  "required_tests": ["quick browser smoke", "go test ./cmd/creator-web", "go vet ./cmd/creator-web", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "production public health"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY"]
+}

dev/tasks/DEV-000203-local-agent-execution-fallback.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000203",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "bee238b61e4760bd864355b893f1fbb188b3f4ce",
+  "risk_tier": "E",
+  "intent": "Provide a first-party local Agent Execution fallback that preserves durable work without claiming external execution when the optional bridge is unavailable.",
+  "requirement_ids": ["WP-301", "local agent execution fallback", "degraded operation"],
+  "allowed_change_scope": {"modules": ["agent-worker"], "paths": ["cmd/agent-worker/**", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000203-local-agent-execution-fallback.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "database/**", "deploy/**", "security/**", "contracts/**", "modules/**", ".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": ["AgentExecutionRequest", "AgentExecutionResult", "durable worker lease/callback path"], "may_change": ["worker degraded-mode status projection"]},
+  "semantic_owner": "work",
+  "canonical_state": ["existing work dispatch, lease and callback state"],
+  "state_transitions": ["dispatch -> durable lease/heartbeat -> explicit WAITING checkpoint when external bridge is absent"],
+  "side_effects_and_privileges": ["no external mutation; no new credentials"],
+  "persistence_and_migrations": ["reuse existing work persistence; no migration"],
+  "failure_retry_restart": ["WAITING remains evidence-bearing and stale lease recovery remains authoritative"],
+  "assumptions": ["local fallback is degraded coordination/recovery, not a claim of completed agent work", "external bridge remains optional"],
+  "acceptance_scenarios": ["missing bridge selects local-fallback mode", "fallback produces WAITING with no outcome reference", "status exposes local_fallback=true", "external bridge path remains unchanged"],
+  "required_tests": ["go test ./cmd/agent-worker", "go vet ./cmd/agent-worker", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "quick production worker status"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000204-gpu-discovery-evidence.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000204",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "b4d166480e5d06049718618ffeea0ccdbb993336",
+  "risk_tier": "E",
+  "intent": "Record protected GPU model discovery without equating hardware availability with intelligence, qualification or routing priority.",
+  "requirement_ids": ["WP-200", "model fleet/failover", "hardware-neutral routing"],
+  "allowed_change_scope": {"modules": ["governance status", "evidence"], "paths": ["docs/PROJECT_STATUS.md", "docs/evidence/gpu-discovery-2026-08-23-v1.json", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000204-gpu-discovery-evidence.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "cmd/**", "modules/**", "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": ["protected provider discovery", "model fleet policy"], "may_change": ["status and evidence projection only"]},
+  "semantic_owner": "governance",
+  "canonical_state": ["capability discovery projection"],
+  "state_transitions": ["HEAVY degraded -> provider AVAILABLE; models remain DISCOVERED/INACTIVE"],
+  "side_effects_and_privileges": ["protected read/discovery request only"],
+  "persistence_and_migrations": ["no migration"],
+  "failure_retry_restart": ["retain unqualified/unpromoted state until task-specific evidence exists"],
+  "assumptions": ["cognitive-worker uses the configured protected GPU credential"],
+  "acceptance_scenarios": ["two GPU models are observed", "both remain unpromoted", "status rejects GPU-as-intelligence inference"],
+  "required_tests": ["go test ./cmd/cognitive-worker", "go vet ./cmd/cognitive-worker", "protected provider discovery smoke", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "production public health"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY"]
+}

dev/tasks/DEV-000205-gpu-invocation-smoke-evidence.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000205",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "288d18a4fba0535e3e2d15643de5b78499beb961",
+  "risk_tier": "E",
+  "intent": "Record a bounded protected GPU invocation smoke result without promoting or activating an unqualified model.",
+  "requirement_ids": ["WP-200", "WP-201", "model fleet/failover", "truthful negative evidence"],
+  "allowed_change_scope": {"modules": ["governance status", "evidence"], "paths": ["docs/evidence/gpu-invocation-smoke-2026-08-23-v1.json", "docs/PROJECT_STATUS.md", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000205-gpu-invocation-smoke-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": ["protected GPU gateway OpenAI-compatible endpoint", "existing GPU discovery evidence"], "may_change": []},
+  "semantic_owner": "governance evidence projection",
+  "canonical_state": [],
+  "state_transitions": ["discovery available -> invocation smoke not observed; no activation"],
+  "side_effects_and_privileges": ["one bounded synthetic request through an already configured protected credential"],
+  "persistence_and_migrations": ["immutable JSON evidence only"],
+  "failure_retry_restart": ["no retry; 400 remains an explicit negative observation"],
+  "assumptions": ["the gateway credential is resolved inside the protected runtime and is not copied into evidence"],
+  "acceptance_scenarios": ["model discovery remains visible", "failed invocation is recorded without claiming quality", "GPU remains inactive and cannot become routing priority from discovery alone"],
+  "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", "independent", "git diff --check", "CI", "production public health check"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000206-completion-matrix-refresh.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000206",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "faab913227f985fc06436618b5747376d4495487",
+  "risk_tier": "E",
+  "intent": "Refresh the audit completion matrix against the current production revision and preserve explicit evidence for every remaining gap.",
+  "requirement_ids": ["AUDIT-COMPLETION-TRUTH", "all objective areas", "current production evidence"],
+  "allowed_change_scope": {"modules": ["governance status", "evidence"], "paths": ["docs/evidence/audit-completion-matrix-2026-08-23-v2.json", "docs/PROJECT_STATUS.md", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000206-completion-matrix-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": ["current production status and evidence artifacts"], "may_change": []},
+  "semantic_owner": "governance evidence projection",
+  "canonical_state": [],
+  "state_transitions": ["none; refreshes an append-only audit projection"],
+  "side_effects_and_privileges": ["none beyond repository evidence and status"],
+  "persistence_and_migrations": ["immutable JSON evidence only"],
+  "failure_retry_restart": ["retain explicit NOT_COMPLETE result when any required evidence is missing"],
+  "assumptions": ["production revision faab913 is the deployed main artifact", "deep acceptance remains deferred by current testing policy"],
+  "acceptance_scenarios": ["all seven objective areas have current statuses", "repository and production proof are distinguished", "remaining external gaps are explicit", "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", "independent", "git diff --check", "CI", "production public health check"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000207-gpu-invocation-qualification.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000207",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "9f96393fb9a7ddce19eda3f90cbdeecb024029e6",
+  "risk_tier": "E",
+  "intent": "Record a bounded production GPU invocation using the exact Domovoy adapter request shape without automatic promotion.",
+  "requirement_ids": ["WP-200", "WP-201", "model fleet/failover", "hardware-neutral routing"],
+  "allowed_change_scope": {"modules": ["governance status", "evidence"], "paths": ["docs/evidence/gpu-invocation-qualification-2026-08-23-v1.json", "docs/evidence/audit-completion-matrix-2026-08-23-v3.json", "docs/PROJECT_STATUS.md", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000207-gpu-invocation-qualification.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": ["existing OpenAI-compatible local adapter request contract", "protected GPU gateway", "current completion matrix"], "may_change": []},
+  "semantic_owner": "governance evidence projection",
+  "canonical_state": [],
+  "state_transitions": ["discovered/inactive -> bounded invocation observed; remains unpromoted"],
+  "side_effects_and_privileges": ["one bounded synthetic request through an existing protected credential"],
+  "persistence_and_migrations": ["immutable JSON evidence only"],
+  "failure_retry_restart": ["no retry and no automatic activation"],
+  "assumptions": ["the protected runtime credential is used in place and is not copied into repository artifacts"],
+  "acceptance_scenarios": ["exact adapter payload returns one bounded completion", "quality remains unassessed", "GPU does not receive routing priority merely because it is available"],
+  "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", "independent", "git diff --check", "CI", "production public health check"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_CONFLICT"]
+}

dev/tasks/DEV-000208-task-aware-model-evaluation.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000208",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "b78a026ee169047feb8468f0799fa1b9f4e27562",
+  "risk_tier": "E",
+  "intent": "Make Domovoy's model qualification and routing task-aware through parallel durable specialist work: typed evaluation profiles/evidence, durable task scores, and explainable selection without hardware/provider heuristics.",
+  "requirement_ids": ["task-specific model evaluation", "parallel specialist evaluation", "no main execution thread", "model fleet/failover", "hardware-neutral routing", "explainable model rating"],
+  "allowed_change_scope": {"modules": ["capability", "core-runtime"], "paths": ["contracts/capability/**", "modules/capability/**", "adapters/postgres/capability_repository.go", "cmd/core-runtime/main.go", "database/migrations/039_capability_task_evaluations.sql", "tests/model_qualification/**", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000208-task-aware-model-evaluation.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "deploy/**", ".gitea/**", "contracts/authority/**", "modules/identity/**"]},
+  "allowed_architecture_delta": {"new_cross_module_dependencies": 0, "new_canonical_owners": 0, "new_privileges": 0, "new_third_party_dependencies": 0, "new_database_owner_roles": 1},
+  "contracts": {"consume": ["CognitiveRequirement", "ResolvedCognitiveBinding", "existing capability evidence", "existing capability owner repository", "Work Graph durable fan-out/fan-in"], "may_change": ["task profile/evaluation contracts", "task-aware selection request/response serialization"]},
+  "semantic_owner": "capability",
+  "canonical_state": ["capability task evaluation evidence keyed by task profile and binding revision"],
+  "state_transitions": ["GENERAL profile -> first bounded evaluation", "unresolved task evidence -> ADDITIONAL profile with parent and trigger evidence", "profile result UNKNOWN -> SUPPORTED/UNSUPPORTED; evidence is append-only by binding revision", "task-aware requirement -> eligible candidates with explainable task rating", "evaluation request -> durable fan-out -> specialist evidence -> fan-in aggregate"],
+  "side_effects_and_privileges": ["core capability database role writes only; no provider calls and no secret access in this task"],
+  "persistence_and_migrations": ["039 adds capability.cognitive_task_evaluations owned by capability"],
+  "failure_retry_restart": ["missing, stale or unknown task evidence excludes a task-aware candidate; no score defaults to success", "one specialist failure does not block unrelated tasks; stale leases are recoverable and fan-in preserves partial/unknown state"],
+  "assumptions": ["a later Work Graph evaluation runner will produce per-case evidence; this task establishes the durable contract and routing gate", "the chat/request path never waits synchronously for the full evaluation graph"],
+  "acceptance_scenarios": ["general profile runs before additional profiles", "an unresolved task creates only a bounded additional profile linked to failure evidence", "task-aware requirement rejects a model without current supported task evidence", "task scores are included in explainable rating", "provider kind, endpoint and hardware do not affect score", "unknown/failed task evidence never silently becomes supported", "existing requirements without a task profile retain current behavior"],
+  "required_tests": ["go test", "go vet", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "migration/schema checks", "CI", "production public health check"],
+  "terminal_gap_states": ["BLOCKED_BY_SPEC_GAP", "BLOCKED_BY_ARCHITECTURE_GAP", "BLOCKED_BY_MISSING_CAPABILITY"]
+}

dev/tasks/DEV-000209-task-evaluator-work-graph.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000209",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "a82592a1300c06ed7a568c058faca6560e085ab3",
+  "risk_tier": "E",
+  "intent": "Create durable Work Graph plans for task-aware model evaluation: run the general profile first, fan out independent model specialists in parallel, and permit bounded additional profiles linked to unresolved evidence without blocking the chat path.",
+  "requirement_ids": ["general-first model evaluation", "parallel specialist evaluation", "bounded additional task tests", "fan-out/fan-in", "no synchronous chat blocking", "provider and hardware neutral"],
+  "allowed_change_scope": {"modules": ["work"], "paths": ["contracts/work/**", "modules/work/**", "cmd/work-graph-runtime/**", "tests/model_qualification/**", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000209-task-evaluator-work-graph.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "deploy/**", "cmd/devctl/**", "modules/identity/**", "database/**"]},
+  "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": ["capability.TaskEvaluationProfile", "capability.TaskEvaluationProfileStage", "capability.TaskEvaluationProfileID", "existing Work Graph schedule contract"], "may_change": ["typed Work Graph node purpose and task-evaluation planning contract"]},
+  "semantic_owner": "work",
+  "canonical_state": ["work-owned durable graph and node lifecycle; capability remains owner of model evaluation evidence"],
+  "state_transitions": ["GENERAL profile -> admitted evaluation graph", "eligible model bindings -> parallel specialist nodes", "all specialist nodes -> fan-in completion state", "ADDITIONAL profile -> a separate bounded graph linked by parent/trigger evidence"],
+  "side_effects_and_privileges": ["none; graph planning only, no provider calls and no secret access"],
+  "persistence_and_migrations": ["uses existing work graph persistence; no new schema"],
+  "failure_retry_restart": ["each specialist remains independently retryable through existing node attempts/leases; partial or unknown graph state is preserved", "invalid stage/parent linkage is rejected before scheduling"],
+  "assumptions": ["a later worker adapter executes each dispatched specialist and posts evidence to capability", "chat/request handlers enqueue the graph and return without waiting for all nodes"],
+  "acceptance_scenarios": ["additional profile cannot be scheduled as a replacement for general profile", "one node is created per binding and profile with independent dispatch identity", "parallel budget is retained for fan-out", "graph metadata is sufficient for a worker to identify profile, binding and specialist role", "provider kind, endpoint and hardware are absent from planning inputs"],
+  "required_tests": ["go test", "go vet", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production public health check"],
+  "terminal_gap_states": ["BLOCKED_BY_SPEC_GAP", "BLOCKED_BY_ARCHITECTURE_GAP", "BLOCKED_BY_MISSING_CAPABILITY"]
+}