docs: record authenticated mobile acceptance

feat(capability): consume qualification evidence in selection

adapters/developmentexecutor/executor.go

@@ -0,0 +1,197 @@
+// Package developmentexecutor contains the local, read-only verification
+// adapter for governed development candidates. It is deliberately not a
+// source mutator, promotion authority, secret resolver, or deployment tool.
+package developmentexecutor
+
+import (
+	"context"
+	"crypto/sha256"
+	"encoding/hex"
+	"errors"
+	"fmt"
+	"io"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"strings"
+	"time"
+)
+
+type CheckKind string
+
+const (
+	CheckGoTest       CheckKind = "GO_TEST"
+	CheckGoVet        CheckKind = "GO_VET"
+	CheckGitDiffCheck CheckKind = "GIT_DIFF_CHECK"
+)
+
+type Check struct {
+	Kind     CheckKind
+	Packages []string
+}
+
+type Plan struct {
+	Worktree            string
+	WorktreeID          string
+	BaseRepoRevision    string
+	ProtectedMainAccess string
+	ProductionAccess    string
+	SecretAccess        string
+	Checks              []Check
+}
+
+type CheckResult struct {
+	Kind           CheckKind `json:"kind"`
+	Passed         bool      `json:"passed"`
+	ExitCode       int       `json:"exit_code"`
+	DurationMillis int64     `json:"duration_millis"`
+}
+
+type Result struct {
+	WorktreeID       string        `json:"worktree_id"`
+	BaseRepoRevision string        `json:"base_repo_revision"`
+	Passed           bool          `json:"passed"`
+	Checks           []CheckResult `json:"checks"`
+	EvidenceRef      string        `json:"evidence_ref"`
+}
+
+var (
+	ErrInvalidPlan       = errors.New("invalid governed verification plan")
+	ErrUnsupportedCheck  = errors.New("unsupported governed verification check")
+	ErrProtectedWorktree = errors.New("protected worktree is not executable")
+)
+
+type CommandRunner interface {
+	Run(context.Context, string, string, []string) int
+}
+
+type OSCommandRunner struct{}
+
+func (OSCommandRunner) Run(ctx context.Context, dir, name string, args []string) int {
+	command := exec.CommandContext(ctx, name, args...)
+	command.Dir = dir
+	command.Stdin = nil
+	command.Stdout = io.Discard
+	command.Stderr = io.Discard
+	command.Env = safeEnvironment()
+	if err := command.Run(); err != nil {
+		if exitErr, ok := err.(*exec.ExitError); ok {
+			return exitErr.ExitCode()
+		}
+		return -1
+	}
+	return 0
+}
+
+func ValidatePlan(plan Plan) error {
+	if strings.TrimSpace(plan.Worktree) == "" || strings.TrimSpace(plan.WorktreeID) == "" || strings.TrimSpace(plan.BaseRepoRevision) == "" || len(plan.Checks) == 0 {
+		return ErrInvalidPlan
+	}
+	if !filepath.IsAbs(plan.Worktree) {
+		return ErrInvalidPlan
+	}
+	if plan.ProtectedMainAccess != "DENIED" || plan.ProductionAccess != "DENIED" || plan.SecretAccess != "DENIED" {
+		return ErrProtectedWorktree
+	}
+	switch strings.ToLower(filepath.Base(filepath.Clean(plan.WorktreeID))) {
+	case "main", "protected-main", "production":
+		return ErrProtectedWorktree
+	}
+	for _, check := range plan.Checks {
+		if err := validateCheck(check); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+func validateCheck(check Check) error {
+	switch check.Kind {
+	case CheckGoTest, CheckGoVet:
+		if len(check.Packages) == 0 {
+			return ErrInvalidPlan
+		}
+		for _, packagePath := range check.Packages {
+			if strings.TrimSpace(packagePath) == "" || strings.ContainsAny(packagePath, ";&|`$()<>\r\n") || unsafePackagePath(packagePath) {
+				return ErrInvalidPlan
+			}
+		}
+	case CheckGitDiffCheck:
+		if len(check.Packages) != 0 {
+			return ErrInvalidPlan
+		}
+	default:
+		return ErrUnsupportedCheck
+	}
+	return nil
+}
+
+func unsafePackagePath(value string) bool {
+	value = strings.ReplaceAll(value, "\\", "/")
+	for _, part := range strings.Split(value, "/") {
+		if part == ".." {
+			return true
+		}
+	}
+	return strings.HasPrefix(value, "../") || strings.HasPrefix(value, "/") || (len(value) > 1 && value[1] == ':')
+}
+
+func Execute(ctx context.Context, plan Plan, runner CommandRunner) (Result, error) {
+	if err := ValidatePlan(plan); err != nil {
+		return Result{}, err
+	}
+	if runner == nil {
+		runner = OSCommandRunner{}
+	}
+	if _, err := os.Stat(plan.Worktree); err != nil {
+		return Result{}, fmt.Errorf("worktree unavailable: %w", err)
+	}
+	result := Result{WorktreeID: plan.WorktreeID, BaseRepoRevision: plan.BaseRepoRevision, Passed: true, Checks: make([]CheckResult, 0, len(plan.Checks))}
+	for _, check := range plan.Checks {
+		started := time.Now()
+		name, args := commandFor(check)
+		exitCode := runner.Run(ctx, plan.Worktree, name, args)
+		passed := exitCode == 0
+		result.Checks = append(result.Checks, CheckResult{Kind: check.Kind, Passed: passed, ExitCode: exitCode, DurationMillis: time.Since(started).Milliseconds()})
+		if !passed {
+			result.Passed = false
+		}
+		if ctx.Err() != nil {
+			result.Passed = false
+			break
+		}
+	}
+	result.EvidenceRef = evidenceRef(result)
+	return result, nil
+}
+
+func commandFor(check Check) (string, []string) {
+	switch check.Kind {
+	case CheckGoTest:
+		return "go", append([]string{"test"}, check.Packages...)
+	case CheckGoVet:
+		return "go", append([]string{"vet"}, check.Packages...)
+	default:
+		return "git", []string{"diff", "--check"}
+	}
+}
+
+func evidenceRef(result Result) string {
+	hash := sha256.New()
+	_, _ = fmt.Fprintf(hash, "%s|%s|%t|", result.WorktreeID, result.BaseRepoRevision, result.Passed)
+	for _, check := range result.Checks {
+		_, _ = fmt.Fprintf(hash, "%s|%t|%d|", check.Kind, check.Passed, check.ExitCode)
+	}
+	return "verification:" + hex.EncodeToString(hash.Sum(nil))
+}
+
+func safeEnvironment() []string {
+	allowed := []string{"PATH", "SystemRoot", "TEMP", "TMP", "GOROOT", "GOPATH", "GOMODCACHE"}
+	result := make([]string, 0, len(allowed))
+	for _, name := range allowed {
+		if value, ok := os.LookupEnv(name); ok && value != "" {
+			result = append(result, name+"="+value)
+		}
+	}
+	return result
+}

adapters/developmentexecutor/executor_test.go

@@ -0,0 +1,70 @@
+package developmentexecutor
+
+import (
+	"context"
+	"errors"
+	"path/filepath"
+	"testing"
+)
+
+type fakeRunner struct {
+	calls []string
+	exits []int
+}
+
+func (r *fakeRunner) Run(_ context.Context, _ string, name string, args []string) int {
+	call := name
+	for _, arg := range args {
+		call += " " + arg
+	}
+	r.calls = append(r.calls, call)
+	if len(r.exits) == 0 {
+		return 0
+	}
+	exit := r.exits[0]
+	r.exits = r.exits[1:]
+	return exit
+}
+
+func validPlan(t *testing.T) Plan {
+	t.Helper()
+	worktree, err := filepath.Abs(".")
+	if err != nil {
+		t.Fatal(err)
+	}
+	return Plan{Worktree: worktree, WorktreeID: "sandbox-1", BaseRepoRevision: "revision-1", ProtectedMainAccess: "DENIED", ProductionAccess: "DENIED", SecretAccess: "DENIED", Checks: []Check{{Kind: CheckGoTest, Packages: []string{"./modules/work/..."}}, {Kind: CheckGoVet, Packages: []string{"./modules/work/..."}}, {Kind: CheckGitDiffCheck}}}
+}
+
+func TestExecuteRunsOnlyTypedChecksAndReturnsEvidence(t *testing.T) {
+	runner := &fakeRunner{}
+	result, err := Execute(context.Background(), validPlan(t), runner)
+	if err != nil || !result.Passed || len(result.Checks) != 3 || result.EvidenceRef == "" {
+		t.Fatalf("result=%+v err=%v", result, err)
+	}
+	if len(runner.calls) != 3 || runner.calls[0] != "go test ./modules/work/..." || runner.calls[2] != "git diff --check" {
+		t.Fatalf("calls=%v", runner.calls)
+	}
+}
+
+func TestExecuteFailureCannotBecomeVerificationSuccess(t *testing.T) {
+	runner := &fakeRunner{exits: []int{1, 0}}
+	plan := validPlan(t)
+	plan.Checks = []Check{{Kind: CheckGoTest, Packages: []string{"./..."}}, {Kind: CheckGitDiffCheck}}
+	result, err := Execute(context.Background(), plan, runner)
+	if err != nil || result.Passed || result.Checks[0].Passed || result.EvidenceRef == "" {
+		t.Fatalf("failed result was overclaimed: %+v err=%v", result, err)
+	}
+}
+
+func TestValidatePlanRejectsProtectedAndShellLikeInputs(t *testing.T) {
+	protected := validPlan(t)
+	protected.WorktreeID = "production"
+	if !errors.Is(ValidatePlan(protected), ErrProtectedWorktree) {
+		t.Fatal("protected worktree accepted")
+	}
+	injection := validPlan(t)
+	injection.Checks = []Check{{Kind: CheckGoTest, Packages: []string{"./...; echo leaked"}}}
+	if !errors.Is(ValidatePlan(injection), ErrInvalidPlan) {
+		t.Fatal("shell-like package accepted")
+	}
+}

adapters/postgres/capability_repository.go

@@ -437,6 +437,24 @@ func (r *CapabilityRepository) LoadCognitiveCandidatesForProfile(ctx context.Con
 	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,
@@ -487,6 +505,14 @@ func (r *CapabilityRepository) loadCognitiveEvidence(ctx context.Context, value
 	if err == nil {
 		value.Performance = &performance
 	}
+	var qualification capability.ModelQualificationObservation
+	err = r.pool.QueryRow(ctx, `SELECT observation_id,provider_instance_id,provider_model_id,status,finish_reason,latency_millis,output_observed,quality_assessed,category,observed_at_unix_millis FROM capability.cognitive_model_qualification_observations WHERE provider_instance_id=$1 AND provider_model_id=$2 ORDER BY observed_at_unix_millis DESC LIMIT 1`, value.Provider.ID, value.Model.ProviderModelID).Scan(&qualification.ID, &qualification.ProviderInstanceID, &qualification.ProviderModelID, &qualification.Status, &qualification.FinishReason, &qualification.LatencyMillis, &qualification.OutputObserved, &qualification.QualityAssessed, &qualification.Category, &qualification.ObservedAtUnixMillis)
+	if err != nil && !errors.Is(err, pgx.ErrNoRows) {
+		return err
+	}
+	if err == nil {
+		value.Qualification = &qualification
+	}
 	if profileID != "" {
 		var taskEvaluation capability.TaskEvaluation
 		err = r.pool.QueryRow(ctx, `SELECT evaluation_id,task_profile_id,binding_id,binding_revision,capability_id,result,quality_score_millis,reliability_score_millis,latency_score_millis,overall_score_millis,critical_failure_millis,benchmark_version,artifact_id,observed_at_unix_millis FROM capability.cognitive_task_evaluations WHERE task_profile_id=$1 AND binding_id=$2 AND binding_revision=$3 ORDER BY observed_at_unix_millis DESC LIMIT 1`, profileID, value.Binding.ID, value.Binding.Revision).Scan(&taskEvaluation.ID, &taskEvaluation.ProfileID, &taskEvaluation.BindingID, &taskEvaluation.BindingRevision, &taskEvaluation.CapabilityID, &taskEvaluation.Result, &taskEvaluation.QualityScoreMillis, &taskEvaluation.ReliabilityScoreMillis, &taskEvaluation.LatencyScoreMillis, &taskEvaluation.OverallScoreMillis, &taskEvaluation.CriticalFailureMillis, &taskEvaluation.BenchmarkVersion, &taskEvaluation.ArtifactID, &taskEvaluation.ObservedAtUnixMillis)
@@ -615,3 +641,8 @@ func (r *CapabilityRepository) SaveTaskEvaluation(ctx context.Context, value cap
 	_, err := r.pool.Exec(ctx, `INSERT INTO capability.cognitive_task_evaluations (evaluation_id,task_profile_id,binding_id,binding_revision,capability_id,result,quality_score_millis,reliability_score_millis,latency_score_millis,overall_score_millis,critical_failure_millis,benchmark_version,artifact_id,observed_at_unix_millis) VALUES ($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13,$14) ON CONFLICT DO NOTHING`, value.ID, value.ProfileID, value.BindingID, value.BindingRevision, value.CapabilityID, value.Result, value.QualityScoreMillis, value.ReliabilityScoreMillis, value.LatencyScoreMillis, value.OverallScoreMillis, value.CriticalFailureMillis, value.BenchmarkVersion, value.ArtifactID, value.ObservedAtUnixMillis)
 	return err
 }
+
+func (r *CapabilityRepository) SaveModelQualificationObservation(ctx context.Context, value capability.ModelQualificationObservation) error {
+	_, err := r.pool.Exec(ctx, `INSERT INTO capability.cognitive_model_qualification_observations (observation_id,provider_instance_id,provider_model_id,status,finish_reason,latency_millis,output_observed,quality_assessed,category,observed_at_unix_millis) VALUES ($1,$2,$3,$4,$5,$6,$7,$8,$9,$10) ON CONFLICT DO NOTHING`, value.ID, value.ProviderInstanceID, value.ProviderModelID, value.Status, value.FinishReason, value.LatencyMillis, value.OutputObserved, value.QualityAssessed, value.Category, value.ObservedAtUnixMillis)
+	return err
+}

adapters/postgres/epistemics_repository.go

@@ -11,6 +11,14 @@ import (
 
 type EpistemicsRepository struct{ pool *pgxpool.Pool }
 
+type TruthRuntimeStatus struct {
+	BundleCount          int64            `json:"bundle_count"`
+	ClaimsByVerification map[string]int64 `json:"claims_by_verification"`
+	BeliefsByState       map[string]int64 `json:"beliefs_by_state"`
+	SourcesByFreshness   map[string]int64 `json:"sources_by_freshness"`
+	ObservedAtUnixMillis int64            `json:"observed_at_unix_millis"`
+}
+
 func OpenEpistemics(ctx context.Context, config ContinuityConfig) (*EpistemicsRepository, error) {
 	pool, err := openPool(ctx, config)
 	if err != nil {
@@ -25,6 +33,49 @@ func (r *EpistemicsRepository) Ping(ctx context.Context) error {
 	return r.pool.Ping(ctx)
 }
 
+// TruthRuntimeStatus is a read-only projection of durable truth bundles.
+// It does not promote claims, revise beliefs, or mutate epistemic state.
+func (r *EpistemicsRepository) TruthRuntimeStatus(ctx context.Context) (TruthRuntimeStatus, error) {
+	status := TruthRuntimeStatus{
+		ClaimsByVerification: map[string]int64{},
+		BeliefsByState:       map[string]int64{},
+		SourcesByFreshness:   map[string]int64{},
+		ObservedAtUnixMillis: time.Now().UnixMilli(),
+	}
+	if err := r.pool.QueryRow(ctx, `SELECT count(*) FROM epistemics.truth_bundles`).Scan(&status.BundleCount); err != nil {
+		return TruthRuntimeStatus{}, err
+	}
+	queries := []struct {
+		query       string
+		destination map[string]int64
+	}{
+		{`SELECT claim_verification, count(*) FROM epistemics.truth_bundles GROUP BY claim_verification`, status.ClaimsByVerification},
+		{`SELECT belief_state, count(*) FROM epistemics.truth_bundles GROUP BY belief_state`, status.BeliefsByState},
+		{`SELECT snapshot_freshness, count(*) FROM epistemics.truth_bundles GROUP BY snapshot_freshness`, status.SourcesByFreshness},
+	}
+	for _, item := range queries {
+		rows, err := r.pool.Query(ctx, item.query)
+		if err != nil {
+			return TruthRuntimeStatus{}, err
+		}
+		for rows.Next() {
+			var state string
+			var count int64
+			if err := rows.Scan(&state, &count); err != nil {
+				rows.Close()
+				return TruthRuntimeStatus{}, err
+			}
+			item.destination[state] = count
+		}
+		if err := rows.Err(); err != nil {
+			rows.Close()
+			return TruthRuntimeStatus{}, err
+		}
+		rows.Close()
+	}
+	return status, nil
+}
+
 func (r *EpistemicsRepository) SaveSystemExpectation(ctx context.Context, value epistemics.SystemExpectation) error {
 	_, err := r.pool.Exec(ctx, `INSERT INTO epistemics.system_expectations (
 		expectation_id, original_intent_id, source_utterance_id, expected_outcome, expected_effect,
@@ -99,7 +150,7 @@ func (r *EpistemicsRepository) ClaimReflectionCandidate(ctx context.Context, wor
 	var value epistemics.ReflectionCandidate
 	err = tx.QueryRow(ctx, `SELECT reflection_candidate_id,trigger,state,source_experience_id,original_intent_id,source_utterance_id,summary,evidence_refs,requires_cognition,recorded_at_unix_millis,revision,claimed_by,claim_expires_at_unix_millis
 		FROM epistemics.reflection_candidates
-		WHERE state='OPEN' AND (claim_expires_at_unix_millis=0 OR claim_expires_at_unix_millis < $1)
+		WHERE state='OPEN' AND requires_cognition=false AND (claim_expires_at_unix_millis=0 OR claim_expires_at_unix_millis < $1)
 		ORDER BY recorded_at_unix_millis,reflection_candidate_id
 		FOR UPDATE SKIP LOCKED LIMIT 1`, nowUnixMillis).Scan(&value.ID, &value.Trigger, &value.State, &value.SourceExperienceID, &value.OriginalIntentID, &value.SourceUtteranceID, &value.Summary, &value.EvidenceRefs, &value.RequiresCognition, &value.RecordedAtUnixMillis, &value.Revision, &value.ClaimedBy, &value.ClaimExpiresAtUnixMillis)
 	if err != nil {

adapters/postgres/work_repository.go

@@ -2,10 +2,14 @@ package postgres
 
 import (
 	"context"
+	"encoding/json"
 	"errors"
 	"fmt"
+	"reflect"
+	"slices"
 	"time"
 
+	"git.cin.su/m/Assistant/contracts/capability"
 	"git.cin.su/m/Assistant/contracts/work"
 	"github.com/jackc/pgx/v5"
 	"github.com/jackc/pgx/v5/pgxpool"
@@ -14,10 +18,11 @@ import (
 type WorkRepository struct{ pool *pgxpool.Pool }
 
 var (
-	ErrDispatchIntentConflict  = errors.New("dispatch intent idempotency conflict")
-	ErrAgentLeaseUnavailable   = errors.New("agent lease unavailable")
-	ErrAgentCallbackRejected   = errors.New("agent callback rejected")
-	ErrDevelopmentGateConflict = errors.New("development gate revision conflict")
+	ErrDispatchIntentConflict       = errors.New("dispatch intent idempotency conflict")
+	ErrAgentLeaseUnavailable        = errors.New("agent lease unavailable")
+	ErrAgentCallbackRejected        = errors.New("agent callback rejected")
+	ErrDevelopmentGateConflict      = errors.New("development gate revision conflict")
+	ErrDevelopmentCandidateConflict = errors.New("development candidate conflict")
 )
 
 func OpenWork(ctx context.Context, config ContinuityConfig) (*WorkRepository, error) {
@@ -27,6 +32,54 @@ func OpenWork(ctx context.Context, config ContinuityConfig) (*WorkRepository, er
 	}
 	return &WorkRepository{pool}, nil
 }
+
+// WorkGraphRuntimeStatus is a read-only operational projection owned by Work.
+// It is not canonical graph state and cannot advance any lifecycle.
+type WorkGraphRuntimeStatus struct {
+	GraphsByState        map[string]int64 `json:"graphs_by_state"`
+	NodesByState         map[string]int64 `json:"nodes_by_state"`
+	QueueByState         map[string]int64 `json:"queue_by_state"`
+	ObservedAtUnixMillis int64            `json:"observed_at_unix_millis"`
+}
+
+func (r *WorkRepository) RuntimeStatus(ctx context.Context) (WorkGraphRuntimeStatus, error) {
+	status := WorkGraphRuntimeStatus{
+		GraphsByState:        map[string]int64{},
+		NodesByState:         map[string]int64{},
+		QueueByState:         map[string]int64{},
+		ObservedAtUnixMillis: time.Now().UnixMilli(),
+	}
+	queries := []struct {
+		query       string
+		destination map[string]int64
+	}{
+		{`SELECT state, count(*) FROM work.graphs GROUP BY state`, status.GraphsByState},
+		{`SELECT state, count(*) FROM work.graph_nodes GROUP BY state`, status.NodesByState},
+		{`SELECT state, count(*) FROM work.agent_dispatch_queue GROUP BY state`, status.QueueByState},
+	}
+	for _, item := range queries {
+		rows, err := r.pool.Query(ctx, item.query)
+		if err != nil {
+			return WorkGraphRuntimeStatus{}, err
+		}
+		for rows.Next() {
+			var state string
+			var count int64
+			if err := rows.Scan(&state, &count); err != nil {
+				rows.Close()
+				return WorkGraphRuntimeStatus{}, err
+			}
+			item.destination[state] = count
+		}
+		if err := rows.Err(); err != nil {
+			rows.Close()
+			return WorkGraphRuntimeStatus{}, err
+		}
+		rows.Close()
+	}
+	return status, nil
+}
+
 func (r *WorkRepository) Close()                         { r.pool.Close() }
 func (r *WorkRepository) Ping(ctx context.Context) error { return r.pool.Ping(ctx) }
 func (r *WorkRepository) Create(ctx context.Context, target work.TaskTarget) error {
@@ -100,33 +153,85 @@ func (r *WorkRepository) SaveDevelopmentGate(ctx context.Context, gate work.Deve
 	return nil
 }
 
+func (r *WorkRepository) SaveExtensionProposal(ctx context.Context, value work.ExtensionProposal) error {
+	result, err := r.pool.Exec(ctx, `INSERT INTO work.extension_proposals(proposal_id,requested_by_principal_id,stage,evidence_refs,deduplicated_signal_keys,desired_outcome,smallest_change_layer,affected_owners,affected_contracts,risk_tier,evaluation_plan,alternatives,rollback_plan,assumptions,created_at_unix_millis) VALUES($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13,$14,$15) ON CONFLICT(proposal_id) DO NOTHING`, value.ID, value.RequestedByPrincipalID, value.Stage, value.EvidenceRefs, value.DeduplicatedSignalKeys, value.DesiredOutcome, value.SmallestChangeLayer, value.AffectedOwners, value.AffectedContracts, value.RiskTier, value.EvaluationPlan, value.Alternatives, value.RollbackPlan, value.Assumptions, value.CreatedAtUnixMillis)
+	if err != nil {
+		return err
+	}
+	if result.RowsAffected() != 1 {
+		return ErrDevelopmentCandidateConflict
+	}
+	return nil
+}
+
+func (r *WorkRepository) SaveDevelopmentScaffold(ctx context.Context, value work.DevelopmentScaffold) error {
+	result, err := r.pool.Exec(ctx, `INSERT INTO work.development_scaffolds(scaffold_id,requested_by_principal_id,stage,proposal_ref,base_repo_revision,worktree_ref,task_intent,modules,allowed_paths,forbidden_paths,required_tests,artifact_lineage_refs,protected_main_access,production_access,secret_access,created_at_unix_millis) VALUES($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13,$14,$15,$16) ON CONFLICT(scaffold_id) DO NOTHING`, value.ID, value.RequestedByPrincipalID, value.Stage, value.ProposalRef, value.BaseRepoRevision, value.WorktreeRef, value.TaskIntent, value.Modules, value.AllowedPaths, value.ForbiddenPaths, value.RequiredTests, value.ArtifactLineageRefs, value.ProtectedMainAccess, value.ProductionAccess, value.SecretAccess, value.CreatedAtUnixMillis)
+	if err != nil {
+		return err
+	}
+	if result.RowsAffected() != 1 {
+		return ErrDevelopmentCandidateConflict
+	}
+	return nil
+}
+
 func (r *WorkRepository) SaveDispatchIntent(ctx context.Context, intent work.AgentDispatchIntent) error {
 	tx, err := r.pool.Begin(ctx)
 	if err != nil {
 		return err
 	}
 	defer tx.Rollback(ctx)
-	_, err = tx.Exec(ctx, `INSERT INTO work.agent_dispatch_intents(dispatch_id,idempotency_key,delegation_id,run_id,implementation_id,implementation_revision,source_revision,worktree_id,verification_plan_ref,created_at_unix_millis) VALUES($1,$2,$3,$4,$5,$6,$7,$8,$9,$10) ON CONFLICT(idempotency_key) DO NOTHING`, intent.DispatchID, intent.IdempotencyKey, intent.DelegationID, intent.RunID, intent.ImplementationID, intent.ImplementationRevision, intent.SourceRevision, intent.WorktreeID, intent.VerificationPlanRef, intent.CreatedAtUnixMillis)
+	profileJSON, casesJSON, err := marshalTaskPayload(intent.TaskEvaluationProfile, intent.TaskEvaluationCases)
+	if err != nil {
+		return err
+	}
+	executionJSON, err := json.Marshal(intent.TaskEvaluationExecution)
+	if err != nil {
+		return err
+	}
+	_, err = tx.Exec(ctx, `INSERT INTO work.agent_dispatch_intents(dispatch_id,idempotency_key,delegation_id,run_id,implementation_id,implementation_revision,source_revision,worktree_id,verification_plan_ref,created_at_unix_millis,specialist_purpose,task_evaluation_profile_id,task_evaluation_binding_id,task_evaluation_case_ids,task_evaluation_profile,task_evaluation_cases,task_evaluation_execution) VALUES($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13,$14,$15,$16,$17) ON CONFLICT(idempotency_key) DO NOTHING`, intent.DispatchID, intent.IdempotencyKey, intent.DelegationID, intent.RunID, intent.ImplementationID, intent.ImplementationRevision, intent.SourceRevision, intent.WorktreeID, intent.VerificationPlanRef, intent.CreatedAtUnixMillis, intent.SpecialistPurpose, intent.TaskEvaluationProfileID, intent.TaskEvaluationBindingID, intent.TaskEvaluationCaseIDs, profileJSON, casesJSON, executionJSON)
 	if err != nil {
 		return fmt.Errorf("save dispatch intent: %w", err)
 	}
 	var existing work.AgentDispatchIntent
-	err = tx.QueryRow(ctx, `SELECT dispatch_id,idempotency_key,delegation_id,run_id,implementation_id,implementation_revision,source_revision,worktree_id,verification_plan_ref,created_at_unix_millis FROM work.agent_dispatch_intents WHERE idempotency_key=$1`, intent.IdempotencyKey).Scan(&existing.DispatchID, &existing.IdempotencyKey, &existing.DelegationID, &existing.RunID, &existing.ImplementationID, &existing.ImplementationRevision, &existing.SourceRevision, &existing.WorktreeID, &existing.VerificationPlanRef, &existing.CreatedAtUnixMillis)
+	var profileJSONExisting, casesJSONExisting, executionJSONExisting []byte
+	err = tx.QueryRow(ctx, `SELECT dispatch_id,idempotency_key,delegation_id,run_id,implementation_id,implementation_revision,source_revision,worktree_id,verification_plan_ref,created_at_unix_millis,specialist_purpose,task_evaluation_profile_id,task_evaluation_binding_id,task_evaluation_case_ids,task_evaluation_profile,task_evaluation_cases,task_evaluation_execution FROM work.agent_dispatch_intents WHERE idempotency_key=$1`, intent.IdempotencyKey).Scan(&existing.DispatchID, &existing.IdempotencyKey, &existing.DelegationID, &existing.RunID, &existing.ImplementationID, &existing.ImplementationRevision, &existing.SourceRevision, &existing.WorktreeID, &existing.VerificationPlanRef, &existing.CreatedAtUnixMillis, &existing.SpecialistPurpose, &existing.TaskEvaluationProfileID, &existing.TaskEvaluationBindingID, &existing.TaskEvaluationCaseIDs, &profileJSONExisting, &casesJSONExisting, &executionJSONExisting)
 	if err != nil {
 		return fmt.Errorf("read dispatch intent: %w", err)
 	}
-	if existing != intent {
+	if err := unmarshalTaskPayload(profileJSONExisting, casesJSONExisting, &existing.TaskEvaluationProfile, &existing.TaskEvaluationCases); err != nil {
+		return err
+	}
+	if len(executionJSONExisting) > 0 && string(executionJSONExisting) != "null" {
+		if err := json.Unmarshal(executionJSONExisting, &existing.TaskEvaluationExecution); err != nil {
+			return err
+		}
+	}
+	if !sameDispatchIntent(existing, intent) {
 		return ErrDispatchIntentConflict
 	}
 	return tx.Commit(ctx)
 }
 
+func sameDispatchIntent(left, right work.AgentDispatchIntent) bool {
+	return left.DispatchID == right.DispatchID && left.IdempotencyKey == right.IdempotencyKey && left.DelegationID == right.DelegationID && left.RunID == right.RunID && left.ImplementationID == right.ImplementationID && left.ImplementationRevision == right.ImplementationRevision && left.SourceRevision == right.SourceRevision && left.WorktreeID == right.WorktreeID && left.VerificationPlanRef == right.VerificationPlanRef && left.CreatedAtUnixMillis == right.CreatedAtUnixMillis && left.SpecialistPurpose == right.SpecialistPurpose && left.TaskEvaluationProfileID == right.TaskEvaluationProfileID && left.TaskEvaluationBindingID == right.TaskEvaluationBindingID && slices.Equal(left.TaskEvaluationCaseIDs, right.TaskEvaluationCaseIDs) && reflect.DeepEqual(left.TaskEvaluationProfile, right.TaskEvaluationProfile) && slices.Equal(left.TaskEvaluationCases, right.TaskEvaluationCases) && reflect.DeepEqual(left.TaskEvaluationExecution, right.TaskEvaluationExecution)
+}
+
 func (r *WorkRepository) LoadDispatchIntent(ctx context.Context, dispatchID string) (work.AgentDispatchIntent, error) {
 	var intent work.AgentDispatchIntent
-	err := r.pool.QueryRow(ctx, `SELECT dispatch_id,idempotency_key,delegation_id,run_id,implementation_id,implementation_revision,source_revision,worktree_id,verification_plan_ref,created_at_unix_millis FROM work.agent_dispatch_intents WHERE dispatch_id=$1`, dispatchID).Scan(&intent.DispatchID, &intent.IdempotencyKey, &intent.DelegationID, &intent.RunID, &intent.ImplementationID, &intent.ImplementationRevision, &intent.SourceRevision, &intent.WorktreeID, &intent.VerificationPlanRef, &intent.CreatedAtUnixMillis)
+	var profileJSON, casesJSON, executionJSON []byte
+	err := r.pool.QueryRow(ctx, `SELECT dispatch_id,idempotency_key,delegation_id,run_id,implementation_id,implementation_revision,source_revision,worktree_id,verification_plan_ref,created_at_unix_millis,specialist_purpose,task_evaluation_profile_id,task_evaluation_binding_id,task_evaluation_case_ids,task_evaluation_profile,task_evaluation_cases,task_evaluation_execution FROM work.agent_dispatch_intents WHERE dispatch_id=$1`, dispatchID).Scan(&intent.DispatchID, &intent.IdempotencyKey, &intent.DelegationID, &intent.RunID, &intent.ImplementationID, &intent.ImplementationRevision, &intent.SourceRevision, &intent.WorktreeID, &intent.VerificationPlanRef, &intent.CreatedAtUnixMillis, &intent.SpecialistPurpose, &intent.TaskEvaluationProfileID, &intent.TaskEvaluationBindingID, &intent.TaskEvaluationCaseIDs, &profileJSON, &casesJSON, &executionJSON)
 	if err != nil {
 		return work.AgentDispatchIntent{}, fmt.Errorf("load dispatch intent: %w", err)
 	}
+	if err := unmarshalTaskPayload(profileJSON, casesJSON, &intent.TaskEvaluationProfile, &intent.TaskEvaluationCases); err != nil {
+		return work.AgentDispatchIntent{}, err
+	}
+	if len(executionJSON) > 0 && string(executionJSON) != "null" {
+		if err := json.Unmarshal(executionJSON, &intent.TaskEvaluationExecution); err != nil {
+			return work.AgentDispatchIntent{}, err
+		}
+	}
 	return intent, nil
 }
 
@@ -142,7 +247,7 @@ func (r *WorkRepository) ClaimNextDispatch(ctx context.Context, workerID string,
 	}
 	defer tx.Rollback(ctx)
 	var item work.AgentDispatchQueueItem
-	err = tx.QueryRow(ctx, `SELECT dispatch_id,state,available_at_unix_millis,attempts,claimed_by,updated_at_unix_millis FROM work.agent_dispatch_queue WHERE state='QUEUED' AND available_at_unix_millis <= $1 ORDER BY available_at_unix_millis,dispatch_id FOR UPDATE SKIP LOCKED LIMIT 1`, nowUnixMillis).Scan(&item.DispatchID, &item.State, &item.AvailableAtUnixMillis, &item.Attempts, &item.ClaimedBy, &item.UpdatedAtUnixMillis)
+	err = tx.QueryRow(ctx, `SELECT dispatch_id,state,available_at_unix_millis,attempts,claimed_by,updated_at_unix_millis FROM work.agent_dispatch_queue WHERE state='QUEUED' AND available_at_unix_millis <= $1 AND attempts < retry_limit ORDER BY available_at_unix_millis,dispatch_id FOR UPDATE SKIP LOCKED LIMIT 1`, nowUnixMillis).Scan(&item.DispatchID, &item.State, &item.AvailableAtUnixMillis, &item.Attempts, &item.ClaimedBy, &item.UpdatedAtUnixMillis)
 	if err != nil {
 		if errors.Is(err, pgx.ErrNoRows) {
 			return work.AgentDispatchQueueItem{}, nil
@@ -185,17 +290,36 @@ func (r *WorkRepository) CompleteDispatch(ctx context.Context, dispatchID, worke
 	return nil
 }
 
+func (r *WorkRepository) RetryDispatch(ctx context.Context, dispatchID, workerID string, nowUnixMillis int64, delay time.Duration) error {
+	if delay < 0 {
+		return fmt.Errorf("invalid retry delay")
+	}
+	availableAt := nowUnixMillis + delay.Milliseconds()
+	command, err := r.pool.Exec(ctx, `UPDATE work.agent_dispatch_queue SET state=CASE WHEN attempts >= retry_limit THEN 'FAILED' ELSE 'QUEUED' END,available_at_unix_millis=$1,claimed_by='',updated_at_unix_millis=$2 WHERE dispatch_id=$3 AND state='CLAIMED' AND claimed_by=$4`, availableAt, nowUnixMillis, dispatchID, workerID)
+	if err != nil {
+		return err
+	}
+	if command.RowsAffected() != 1 {
+		return ErrAgentCallbackRejected
+	}
+	return nil
+}
+
 func (r *WorkRepository) SaveWorkGraph(ctx context.Context, graph work.WorkGraph) error {
 	tx, err := r.pool.Begin(ctx)
 	if err != nil {
 		return err
 	}
 	defer tx.Rollback(ctx)
-	if _, err := tx.Exec(ctx, `INSERT INTO work.graphs(graph_id,revision,state,max_parallel,max_attempts) VALUES($1,$2,$3,$4,$5) ON CONFLICT(graph_id,revision) DO NOTHING`, graph.ID, graph.Revision, graph.State, graph.Budget.MaxParallel, graph.Budget.MaxAttempts); err != nil {
+	if _, err := tx.Exec(ctx, `INSERT INTO work.graphs(graph_id,revision,state,max_parallel,max_attempts,parent_graph_id,parent_graph_revision,trigger_evidence_ref,orchestration_bound,orchestration_goal_id,orchestration_goal_state,orchestration_search_round,orchestration_max_round,orchestration_cancel_requested,orchestration_verification_ready,orchestration_unresolved) VALUES($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13,$14,$15,$16) ON CONFLICT(graph_id,revision) DO NOTHING`, graph.ID, graph.Revision, graph.State, graph.Budget.MaxParallel, graph.Budget.MaxAttempts, graph.ParentGraphID, graph.ParentGraphRevision, graph.TriggerEvidenceRef, graph.SearchInputBound, graph.SearchInput.GoalID, graph.SearchInput.GoalState, graph.SearchInput.SearchRound, graph.SearchInput.MaxRound, graph.SearchInput.CancelRequested, graph.SearchInput.VerificationReady, graph.SearchInput.Unresolved); err != nil {
 		return err
 	}
 	for _, node := range graph.Nodes {
-		if _, err := tx.Exec(ctx, `INSERT INTO work.graph_nodes(graph_id,revision,node_id,dispatch_id,idempotency_key,state,attempts) VALUES($1,$2,$3,$4,$5,$6,$7) ON CONFLICT(graph_id,revision,node_id) DO NOTHING`, graph.ID, graph.Revision, node.ID, node.DispatchID, node.IdempotencyKey, node.State, node.Attempts); err != nil {
+		profileJSON, casesJSON, err := marshalTaskPayload(node.TaskProfile, node.TaskCases)
+		if err != nil {
+			return err
+		}
+		if _, err := tx.Exec(ctx, `INSERT INTO work.graph_nodes(graph_id,revision,node_id,dispatch_id,idempotency_key,state,attempts,purpose,task_profile_id,binding_id,case_ids,task_evaluation_profile,task_evaluation_cases) VALUES($1,$2,$3,$4,$5,$6,$7,$8,$9,$10,$11,$12,$13) ON CONFLICT(graph_id,revision,node_id) DO NOTHING`, graph.ID, graph.Revision, node.ID, node.DispatchID, node.IdempotencyKey, node.State, node.Attempts, node.Purpose, node.TaskProfileID, node.BindingID, node.CaseIDs, profileJSON, casesJSON); err != nil {
 			return err
 		}
 	}
@@ -209,17 +333,22 @@ func (r *WorkRepository) SaveWorkGraph(ctx context.Context, graph work.WorkGraph
 
 func (r *WorkRepository) LoadWorkGraph(ctx context.Context, graphID string, revision int64) (work.WorkGraph, error) {
 	var graph work.WorkGraph
-	err := r.pool.QueryRow(ctx, `SELECT graph_id,revision,state,max_parallel,max_attempts FROM work.graphs WHERE graph_id=$1 AND revision=$2`, graphID, revision).Scan(&graph.ID, &graph.Revision, &graph.State, &graph.Budget.MaxParallel, &graph.Budget.MaxAttempts)
+	err := r.pool.QueryRow(ctx, `SELECT graph_id,revision,state,max_parallel,max_attempts,parent_graph_id,parent_graph_revision,trigger_evidence_ref,orchestration_bound,orchestration_goal_id,orchestration_goal_state,orchestration_search_round,orchestration_max_round,orchestration_cancel_requested,orchestration_verification_ready,orchestration_unresolved FROM work.graphs WHERE graph_id=$1 AND revision=$2`, graphID, revision).Scan(&graph.ID, &graph.Revision, &graph.State, &graph.Budget.MaxParallel, &graph.Budget.MaxAttempts, &graph.ParentGraphID, &graph.ParentGraphRevision, &graph.TriggerEvidenceRef, &graph.SearchInputBound, &graph.SearchInput.GoalID, &graph.SearchInput.GoalState, &graph.SearchInput.SearchRound, &graph.SearchInput.MaxRound, &graph.SearchInput.CancelRequested, &graph.SearchInput.VerificationReady, &graph.SearchInput.Unresolved)
 	if err != nil {
 		return work.WorkGraph{}, fmt.Errorf("load graph: %w", err)
 	}
-	rows, err := r.pool.Query(ctx, `SELECT node_id,dispatch_id,idempotency_key,state,attempts FROM work.graph_nodes WHERE graph_id=$1 AND revision=$2 ORDER BY node_id`, graphID, revision)
+	rows, err := r.pool.Query(ctx, `SELECT node_id,dispatch_id,idempotency_key,state,attempts,purpose,task_profile_id,binding_id,case_ids,task_evaluation_profile,task_evaluation_cases FROM work.graph_nodes WHERE graph_id=$1 AND revision=$2 ORDER BY node_id`, graphID, revision)
 	if err != nil {
 		return work.WorkGraph{}, err
 	}
 	for rows.Next() {
 		var node work.GraphNode
-		if err := rows.Scan(&node.ID, &node.DispatchID, &node.IdempotencyKey, &node.State, &node.Attempts); err != nil {
+		var profileJSON, casesJSON []byte
+		if err := rows.Scan(&node.ID, &node.DispatchID, &node.IdempotencyKey, &node.State, &node.Attempts, &node.Purpose, &node.TaskProfileID, &node.BindingID, &node.CaseIDs, &profileJSON, &casesJSON); err != nil {
+			rows.Close()
+			return work.WorkGraph{}, err
+		}
+		if err := unmarshalTaskPayload(profileJSON, casesJSON, &node.TaskProfile, &node.TaskCases); err != nil {
 			rows.Close()
 			return work.WorkGraph{}, err
 		}
@@ -270,11 +399,30 @@ func (r *WorkRepository) SaveGraphSynthesis(ctx context.Context, conflict work.C
 	return tx.Commit(ctx)
 }
 
+func (r *WorkRepository) SaveGraphResultCandidate(ctx context.Context, candidate work.ResultCandidate) error {
+	if candidate.ID == "" || candidate.GraphID == "" || candidate.GraphRevision < 1 || candidate.NodeID == "" || candidate.AuthorRunID == "" || candidate.EvidenceSummary == "" || candidate.Summary == "" {
+		return errors.New("invalid graph result candidate")
+	}
+	_, err := r.pool.Exec(ctx, `INSERT INTO work.graph_result_candidates(result_candidate_id,graph_id,graph_revision,node_id,author_run_id,artifact_refs,evidence_summary,summary) VALUES($1,$2,$3,$4,$5,$6,$7,$8) ON CONFLICT(result_candidate_id) DO NOTHING`, candidate.ID, candidate.GraphID, candidate.GraphRevision, candidate.NodeID, candidate.AuthorRunID, candidate.ArtifactRefs, candidate.EvidenceSummary, candidate.Summary)
+	return err
+}
+
 func (r *WorkRepository) RecoverStaleClaims(ctx context.Context, nowUnixMillis, staleAfterMillis int64) (int64, error) {
-	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)
+	tx, err := r.pool.Begin(ctx)
 	if err != nil {
 		return 0, err
 	}
+	defer tx.Rollback(ctx)
+	if _, err := tx.Exec(ctx, `UPDATE work.agent_leases AS l SET state='EXPIRED' FROM work.agent_dispatch_queue AS q JOIN work.agent_dispatch_intents AS i ON i.dispatch_id=q.dispatch_id WHERE q.state='CLAIMED' AND q.updated_at_unix_millis <= $1 AND l.run_id=i.run_id AND l.state='ACTIVE'`, nowUnixMillis-staleAfterMillis); err != nil {
+		return 0, err
+	}
+	command, err := tx.Exec(ctx, `UPDATE work.agent_dispatch_queue SET state=CASE WHEN attempts >= retry_limit THEN 'FAILED' ELSE 'QUEUED' END,available_at_unix_millis=$1,claimed_by='',updated_at_unix_millis=$2 WHERE state='CLAIMED' AND updated_at_unix_millis <= $3`, nowUnixMillis, nowUnixMillis, nowUnixMillis-staleAfterMillis)
+	if err != nil {
+		return 0, err
+	}
+	if err := tx.Commit(ctx); err != nil {
+		return 0, err
+	}
 	return command.RowsAffected(), nil
 }
 
@@ -565,3 +713,32 @@ func latestRowErr(err error) error {
 	}
 	return err
 }
+
+func marshalTaskPayload(profile *capability.TaskEvaluationProfile, cases []capability.TaskEvaluationCase) ([]byte, []byte, error) {
+	var profileJSON []byte
+	var err error
+	if profile != nil {
+		profileJSON, err = json.Marshal(profile)
+		if err != nil {
+			return nil, nil, err
+		}
+	}
+	casesJSON, err := json.Marshal(cases)
+	return profileJSON, casesJSON, err
+}
+
+func unmarshalTaskPayload(profileJSON, casesJSON []byte, profile **capability.TaskEvaluationProfile, cases *[]capability.TaskEvaluationCase) error {
+	if len(profileJSON) > 0 && string(profileJSON) != "null" {
+		var value capability.TaskEvaluationProfile
+		if err := json.Unmarshal(profileJSON, &value); err != nil {
+			return err
+		}
+		*profile = &value
+	}
+	if len(casesJSON) > 0 && string(casesJSON) != "null" {
+		if err := json.Unmarshal(casesJSON, cases); err != nil {
+			return err
+		}
+	}
+	return nil
+}

adapters/research/search_connector.go

@@ -0,0 +1,150 @@
+package research
+
+import (
+	"context"
+	"crypto/sha256"
+	"encoding/hex"
+	"errors"
+	"html"
+	"io"
+	"net/http"
+	"net/url"
+	"regexp"
+	"strconv"
+	"strings"
+
+	"git.cin.su/m/Assistant/contracts/epistemics"
+)
+
+var ErrSearchRejected = errors.New("web search request rejected")
+
+var (
+	searchLinkPattern    = regexp.MustCompile(`(?is)<a[^>]+class=["'][^"']*result__a[^"']*["'][^>]*>(.*?)</a>`)
+	searchSnippetPattern = regexp.MustCompile(`(?is)<a[^>]+class=["'][^"']*result__snippet[^"']*["'][^>]*>(.*?)</a>`)
+	bingLinkPattern      = regexp.MustCompile(`(?is)<li[^>]+class=["'][^"']*b_algo[^"']*["'][^>]*>.*?<h2[^>]*><a[^>]*>(.*?)</a>`)
+	bingSnippetPattern   = regexp.MustCompile(`(?is)<li[^>]+class=["'][^"']*b_algo[^"']*["'][^>]*>.*?<p[^>]*>(.*?)</p>`)
+	tagPattern           = regexp.MustCompile(`(?s)<[^>]*>`)
+	termPattern          = regexp.MustCompile(`(?i)[\p{L}\p{N}]{2,}`)
+)
+
+var searchStopTerms = map[string]struct{}{
+	"какая": {}, "какое": {}, "какие": {}, "какой": {}, "что": {}, "как": {}, "где": {}, "когда": {}, "почему": {}, "мне": {}, "пожалуйста": {},
+	"what": {}, "which": {}, "where": {}, "when": {}, "why": {}, "how": {}, "please": {}, "the": {}, "this": {},
+}
+
+// SearchConnector uses a configured HTTPS search endpoint. It does not know
+// cities, weather terms, entities or model names: the complete semantic query
+// comes from the interpretation layer.
+type SearchConnector struct {
+	Client           *http.Client
+	EndpointTemplate string
+	AllowedHosts     map[string]struct{}
+	MaxBytes         int64
+	MaxResults       int
+}
+
+func (c SearchConnector) Collect(ctx context.Context, request epistemics.WebSearchRequest, nowUnixMillis int64) ([]epistemics.ResearchCandidate, error) {
+	if request.Plan.Validate() != nil || strings.TrimSpace(request.Query) == "" || nowUnixMillis < 1 || !strings.Contains(c.EndpointTemplate, "%s") {
+		return nil, ErrSearchRejected
+	}
+	endpoint, err := url.Parse(strings.Replace(c.EndpointTemplate, "%s", url.QueryEscape(strings.TrimSpace(request.Query)), 1))
+	if err != nil || endpoint.Scheme != "https" || endpoint.Host == "" || !c.allowedHost(strings.ToLower(endpoint.Hostname())) {
+		return nil, ErrSearchRejected
+	}
+	client := c.Client
+	if client == nil {
+		client = &http.Client{}
+	}
+	maxBytes := c.MaxBytes
+	if maxBytes <= 0 || maxBytes > 4<<20 {
+		maxBytes = 512 << 10
+	}
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, endpoint.String(), nil)
+	if err != nil {
+		return nil, ErrSearchRejected
+	}
+	response, err := client.Do(req)
+	if err != nil {
+		return nil, err
+	}
+	defer response.Body.Close()
+	if response.StatusCode < 200 || response.StatusCode >= 300 {
+		return nil, ErrSearchRejected
+	}
+	body, err := io.ReadAll(io.LimitReader(response.Body, maxBytes+1))
+	if err != nil || len(body) == 0 || int64(len(body)) > maxBytes {
+		return nil, ErrSearchRejected
+	}
+	links := searchLinkPattern.FindAllSubmatch(body, c.limit())
+	snippets := searchSnippetPattern.FindAllSubmatch(body, c.limit())
+	if len(links) == 0 {
+		links = bingLinkPattern.FindAllSubmatch(body, c.limit())
+		snippets = bingSnippetPattern.FindAllSubmatch(body, c.limit())
+	}
+	if len(links) == 0 {
+		return nil, nil
+	}
+	digest := sha256.Sum256(body)
+	digestText := "sha256:" + hex.EncodeToString(digest[:])
+	base := hex.EncodeToString(digest[:])
+	sourceID := epistemics.SourceID("source:search:" + base)
+	snapshotID := epistemics.SourceSnapshotID("snapshot:search:" + base)
+	results := make([]epistemics.ResearchCandidate, 0, len(links))
+	for index, link := range links {
+		title := cleanHTML(string(link[1]))
+		snippet := ""
+		if index < len(snippets) {
+			snippet = cleanHTML(string(snippets[index][1]))
+		}
+		statement := strings.TrimSpace(strings.Trim(title+" — "+snippet, " —"))
+		if statement == "" || !searchStatementRelatesToQuery(request.Query, statement) {
+			continue
+		}
+		ordinal := strconv.Itoa(index)
+		excerptID := epistemics.SourceExcerptID("excerpt:search:" + base + ":" + ordinal)
+		results = append(results, epistemics.ResearchCandidate{
+			PropositionKey: request.PropositionKey + ":result:" + ordinal,
+			Statement:      statement,
+			Source:         epistemics.Source{ID: sourceID, Kind: "web-search", Locator: endpoint.String(), Title: title, Publisher: endpoint.Hostname()},
+			Snapshot:       epistemics.SourceSnapshot{ID: snapshotID, SourceID: sourceID, ContentDigest: digestText, CapturedAtUnixMillis: nowUnixMillis, Freshness: epistemics.FreshnessCurrent},
+			Excerpt:        epistemics.SourceExcerpt{ID: excerptID, SnapshotID: snapshotID, Text: statement, Start: 0, End: int64(len(statement))},
+			Evidence:       epistemics.EvidenceObserved, Coverage: epistemics.CoverageComplete, Conflict: epistemics.ConflictNone,
+		})
+	}
+	return results, nil
+}
+
+// searchStatementRelatesToQuery is a bounded relevance gate, not a truth
+// check. It prevents generic search pages from becoming apparent evidence
+// when their content has no substantive relation to the requested query.
+func searchStatementRelatesToQuery(query, statement string) bool {
+	terms := termPattern.FindAllString(strings.ToLower(query), -1)
+	content := strings.ToLower(statement)
+	usable := 0
+	for _, term := range terms {
+		if _, stop := searchStopTerms[term]; stop || len([]rune(term)) < 4 {
+			continue
+		}
+		usable++
+		runes := []rune(term)
+		prefixLen := 5
+		if len(runes) < prefixLen {
+			prefixLen = len(runes)
+		}
+		if strings.Contains(content, string(runes[:prefixLen])) {
+			return true
+		}
+	}
+	return usable == 0
+}
+
+func (c SearchConnector) allowedHost(host string) bool { _, ok := c.AllowedHosts[host]; return ok }
+func (c SearchConnector) limit() int {
+	if c.MaxResults <= 0 || c.MaxResults > 8 {
+		return 3
+	}
+	return c.MaxResults
+}
+func cleanHTML(value string) string {
+	return strings.Join(strings.Fields(html.UnescapeString(tagPattern.ReplaceAllString(value, ""))), " ")
+}

cmd/agent-canary/main.go

@@ -24,7 +24,8 @@ func (s *memoryStore) SaveDispatchIntent(context.Context, work.AgentDispatchInte
 	s.writes++
 	return nil
 }
-func (s *memoryStore) SaveAgentLease(context.Context, work.AgentLease) error { s.writes++; return nil }
+func (s *memoryStore) SaveAgentLease(context.Context, work.AgentLease) error             { s.writes++; return nil }
+func (s *memoryStore) RenewAgentLease(context.Context, work.AgentLease, time.Time) error { return nil }
 func (s *memoryStore) AcceptAgentCallback(context.Context, work.AgentCallback) error {
 	s.writes++
 	return nil

cmd/agent-execution-bridge/main.go

@@ -0,0 +1,211 @@
+package main
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"os"
+	"strings"
+	"time"
+
+	"git.cin.su/m/Assistant/contracts/artifact"
+	"git.cin.su/m/Assistant/contracts/capability"
+	"git.cin.su/m/Assistant/contracts/cognitive"
+	"git.cin.su/m/Assistant/contracts/work"
+	capabilitydomain "git.cin.su/m/Assistant/modules/capability/domain"
+)
+
+type bridge struct {
+	cognitiveURL string
+	callbackURL  string
+	verifierURL  string
+	client       *http.Client
+}
+
+func main() {
+	b := bridge{
+		cognitiveURL: strings.TrimRight(envOr("COGNITIVE_WORKER_URL", "http://cognitive-worker:8083"), "/"),
+		callbackURL:  strings.TrimRight(envOr("WORK_GRAPH_RUNTIME_URL", "http://work-graph-runtime:8088"), "/"),
+		verifierURL:  strings.TrimRight(envOr("TASK_VERIFIER_URL", "http://task-verifier:8091"), "/"),
+		client:       &http.Client{Timeout: 10 * time.Minute},
+	}
+	mux := http.NewServeMux()
+	mux.HandleFunc("/healthz", healthHandler)
+	mux.HandleFunc("/readyz", healthHandler)
+	mux.HandleFunc("/v1/execute", b.executeHandler)
+	addr := envOr("LISTEN_ADDR", ":8092")
+	if err := http.ListenAndServe(addr, mux); err != nil {
+		fmt.Fprintln(os.Stderr, err)
+		os.Exit(1)
+	}
+}
+
+func (b bridge) executeHandler(w http.ResponseWriter, r *http.Request) {
+	if r.Method != http.MethodPost {
+		http.Error(w, "method not allowed", http.StatusMethodNotAllowed)
+		return
+	}
+	var request work.AgentExecutionRequest
+	if err := json.NewDecoder(http.MaxBytesReader(w, r.Body, 1<<20)).Decode(&request); err != nil {
+		http.Error(w, "invalid execution request", http.StatusBadRequest)
+		return
+	}
+	result, err := b.execute(r.Context(), request)
+	if err != nil {
+		http.Error(w, err.Error(), http.StatusBadGateway)
+		return
+	}
+	w.Header().Set("Content-Type", "application/json")
+	_ = json.NewEncoder(w).Encode(result)
+}
+
+func (b bridge) execute(ctx context.Context, request work.AgentExecutionRequest) (work.AgentExecutionResult, error) {
+	now := time.Now().UnixMilli()
+	intent := request.Intent
+	if intent.TaskEvaluationExecution != nil && intent.TaskEvaluationExecution.GenericMessage != "" {
+		return b.executeGenericCognitiveWork(ctx, request, *intent.TaskEvaluationExecution)
+	}
+	if intent.TaskEvaluationExecution == nil || intent.TaskEvaluationProfile == nil || len(intent.TaskEvaluationCases) == 0 {
+		return work.AgentExecutionResult{State: work.AgentExecutionWaiting, Summary: "локальный bridge не поддерживает это намерение; работа сохранена и ожидает подходящего исполнителя", ObservedAtUnixMillis: now}, nil
+	}
+	execution := intent.TaskEvaluationExecution
+	if execution.GraphID == "" || execution.NodeID == "" || execution.BindingRevision < 1 || execution.ActorID == "" || execution.PrincipalID == "" || execution.ScopeID == "" {
+		return work.AgentExecutionResult{State: work.AgentExecutionWaiting, Summary: "контекст task-evaluation неполон; работа сохранена для повторной маршрутизации", ObservedAtUnixMillis: now}, nil
+	}
+	payload := capability.TaskEvaluationExecutionRequest{Profile: *intent.TaskEvaluationProfile, CapabilityID: intent.TaskEvaluationProfile.CapabilityID, BindingID: intent.TaskEvaluationBindingID, BindingRevision: execution.BindingRevision, ActorID: execution.ActorID, PrincipalID: execution.PrincipalID, ScopeID: execution.ScopeID, Cases: intent.TaskEvaluationCases, MaxOutputBytes: 4096}
+	response, err := postJSON[capability.TaskEvaluationExecutionResponse](ctx, b.client, b.cognitiveURL+"/v1/task-evaluations/execute", payload)
+	if err != nil {
+		return work.AgentExecutionResult{}, err
+	}
+	if len(response.Responses) == 0 || response.ObservedAtUnixMillis < 1 {
+		return work.AgentExecutionResult{}, fmt.Errorf("task evaluation returned no observations")
+	}
+	caseByID := make(map[string]capability.TaskEvaluationCase, len(intent.TaskEvaluationCases))
+	for _, item := range intent.TaskEvaluationCases {
+		caseByID[item.CaseID] = item
+	}
+	observations := response.Observations
+	if !hasCriticalFailure(observations) {
+		observations = make([]capability.TaskCaseObservation, 0, len(response.Responses))
+		for _, item := range response.Responses {
+			caseValue, ok := caseByID[item.CaseID]
+			if !ok || caseValue.Verification.Mode == "" {
+				return work.AgentExecutionResult{State: work.AgentExecutionWaiting, Summary: "task-evaluation case has no bounded verifier; work remains waiting", ObservedAtUnixMillis: time.Now().UnixMilli()}, nil
+			}
+			verified, verifyErr := postJSON[capability.TaskCaseVerificationResponse](ctx, b.client, b.verifierURL+"/v1/task-evaluations/verify", capability.TaskCaseVerificationRequest{CaseID: item.CaseID, ResponseText: item.ResponseText, LatencyMicros: item.LatencyMicros, EvidenceArtifactRef: item.EvidenceArtifactRef, Verification: caseValue.Verification})
+			if verifyErr != nil {
+				return work.AgentExecutionResult{}, verifyErr
+			}
+			observations = append(observations, capability.TaskCaseObservation{CaseID: verified.CaseID, Result: verified.Result, QualityScoreMillis: verified.QualityScoreMillis, LatencyMicros: verified.LatencyMicros, CriticalFailure: verified.CriticalFailure, EvidenceArtifactRef: verified.EvidenceArtifactRef, VerificationEvidenceRef: verified.VerificationEvidenceRef, VerifierIdentity: verified.VerifierIdentity})
+		}
+	}
+	evaluation, err := capabilitydomain.BuildTaskEvaluation(*intent.TaskEvaluationProfile, capability.TaskEvaluationID("evaluation:"+intent.DispatchID), response.BindingID, response.BindingRevision, intent.TaskEvaluationProfile.CapabilityID, observations, artifact.ArtifactID(intent.TaskEvaluationCases[0].EvidenceArtifactRef), response.ObservedAtUnixMillis)
+	if err != nil {
+		return work.AgentExecutionResult{}, err
+	}
+	callback := work.TaskEvaluationEvidenceCallback{GraphID: execution.GraphID, Revision: execution.GraphRevision, NodeID: execution.NodeID, Evaluation: evaluation, Succeeded: true}
+	if _, err := postJSON[json.RawMessage](ctx, b.client, b.callbackURL+"/v1/work-graphs/task-evaluations/callback", callback); err != nil {
+		return work.AgentExecutionResult{}, err
+	}
+	return work.AgentExecutionResult{State: work.AgentExecutionCompleted, Summary: "local bridge executed task evaluation and persisted verifier callback", OutcomeRef: string(evaluation.ID), ArtifactRefs: []string{string(evaluation.ArtifactID)}, ObservedAtUnixMillis: time.Now().UnixMilli()}, nil
+}
+
+func (b bridge) executeGenericCognitiveWork(ctx context.Context, request work.AgentExecutionRequest, execution work.TaskEvaluationDispatchContext) (work.AgentExecutionResult, error) {
+	if execution.GraphID == "" || execution.NodeID == "" || execution.ActorID == "" || execution.PrincipalID == "" || execution.ScopeID == "" {
+		return work.AgentExecutionResult{State: work.AgentExecutionWaiting, Summary: "generic cognitive work context is incomplete; durable work remains waiting for rerouting", ObservedAtUnixMillis: time.Now().UnixMilli()}, nil
+	}
+	maxOutputBytes := execution.GenericMaxOutputBytes
+	if maxOutputBytes < 1 {
+		maxOutputBytes = 4096
+	}
+	dataPolicy := capability.DataPolicy(execution.GenericDataPolicy)
+	if dataPolicy == "" {
+		dataPolicy = capability.DataPolicyLocalOnly
+	}
+	response, err := postJSON[cognitive.GatewayAdvisoryResponse](ctx, b.client, b.cognitiveURL+"/v1/advisories", cognitive.GatewayAdvisoryRequest{
+		Version: cognitive.CreatorAdvisoryV1, RequestID: cognitive.RequestID("agent-work:" + request.Intent.DispatchID), ActorID: execution.ActorID,
+		PrincipalID: execution.PrincipalID, ScopeID: execution.ScopeID, Message: execution.GenericMessage, Locale: execution.GenericLocale,
+		MaxOutputBytes: maxOutputBytes, DataPolicy: dataPolicy, Meaning: cognitive.TaskMeaning{SpeechAct: "DIRECTIVE", RequestedOutcome: execution.GenericPurpose},
+	})
+	if err != nil {
+		return work.AgentExecutionResult{}, err
+	}
+	if response.FinishReason != cognitive.FinishCompleted || strings.TrimSpace(response.AdvisoryText) == "" {
+		return work.AgentExecutionResult{}, fmt.Errorf("generic cognitive work did not complete: %s", response.FinishReason)
+	}
+	candidateID := "candidate:" + request.Intent.DispatchID
+	artifactRef := work.EvidenceRef("agent-output:" + request.Intent.DispatchID)
+	verification, err := postJSON[work.GenericCandidateVerificationResponse](ctx, b.client, b.verifierURL+"/v1/work-graphs/candidates/verify", work.GenericCandidateVerificationRequest{
+		GraphID: execution.GraphID, GraphRevision: execution.GraphRevision, NodeID: execution.NodeID, AuthorRunID: request.Intent.RunID, CandidateID: candidateID, Summary: response.AdvisoryText, EvidenceRefs: []work.EvidenceRef{artifactRef},
+	})
+	if err != nil {
+		return work.AgentExecutionResult{}, err
+	}
+	if !verification.Verified || verification.VerificationEvidenceRef == "" || verification.VerifierIdentity == "" {
+		return work.AgentExecutionResult{State: work.AgentExecutionWaiting, Summary: "generic candidate envelope was not independently verified; work remains waiting", ObservedAtUnixMillis: time.Now().UnixMilli()}, nil
+	}
+	candidate := work.GraphResultCallback{GraphID: execution.GraphID, Revision: execution.GraphRevision, Candidate: work.ResultCandidate{
+		ID: candidateID, GraphID: execution.GraphID, GraphRevision: execution.GraphRevision, NodeID: execution.NodeID,
+		AuthorRunID: request.Intent.RunID, ArtifactRefs: []work.EvidenceRef{artifactRef, work.EvidenceRef(verification.VerificationEvidenceRef)}, EvidenceSummary: "candidate attribution and evidence envelope independently verified; semantic content remains unverified",
+		Summary: response.AdvisoryText,
+	}}
+	if _, err := postJSON[json.RawMessage](ctx, b.client, b.callbackURL+"/v1/work-graphs/result-callback", candidate); err != nil {
+		return work.AgentExecutionResult{}, err
+	}
+	return work.AgentExecutionResult{State: work.AgentExecutionCompleted, Summary: "local cognitive specialist produced a candidate for independent verification", OutcomeRef: "candidate:" + request.Intent.DispatchID, ArtifactRefs: []string{"agent-output:" + request.Intent.DispatchID}, ObservedAtUnixMillis: time.Now().UnixMilli()}, nil
+}
+
+func hasCriticalFailure(observations []capability.TaskCaseObservation) bool {
+	for _, observation := range observations {
+		if observation.CriticalFailure {
+			return true
+		}
+	}
+	return false
+}
+
+func postJSON[T any](ctx context.Context, client *http.Client, url string, payload any) (T, error) {
+	var result T
+	body, err := json.Marshal(payload)
+	if err != nil {
+		return result, err
+	}
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
+	if err != nil {
+		return result, err
+	}
+	req.Header.Set("Content-Type", "application/json")
+	resp, err := client.Do(req)
+	if err != nil {
+		return result, err
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return result, fmt.Errorf("POST %s: status %s", url, resp.Status)
+	}
+	if resp.StatusCode == http.StatusNoContent {
+		return result, nil
+	}
+	if err := json.NewDecoder(resp.Body).Decode(&result); err != nil {
+		return result, err
+	}
+	return result, nil
+}
+
+func healthHandler(w http.ResponseWriter, r *http.Request) {
+	if r.Method != http.MethodGet {
+		http.Error(w, "method not allowed", http.StatusMethodNotAllowed)
+		return
+	}
+	w.Header().Set("Content-Type", "application/json")
+	_, _ = w.Write([]byte(`{"runtime":"agent-execution-bridge","status":"ok","capabilities":["task-evaluation","generic-cognitive-work"]}`))
+}
+
+func envOr(name, fallback string) string {
+	if value := strings.TrimSpace(os.Getenv(name)); value != "" {
+		return value
+	}
+	return fallback
+}

cmd/agent-execution-bridge/main_test.go

@@ -0,0 +1,88 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"git.cin.su/m/Assistant/contracts/capability"
+	"git.cin.su/m/Assistant/contracts/cognitive"
+	"git.cin.su/m/Assistant/contracts/work"
+)
+
+func TestBridgeWaitsForUnsupportedIntent(t *testing.T) {
+	b := bridge{client: http.DefaultClient}
+	result, err := b.execute(context.Background(), work.AgentExecutionRequest{Intent: work.AgentDispatchIntent{DispatchID: "dispatch-unsupported"}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	if result.State != work.AgentExecutionWaiting || result.ObservedAtUnixMillis < 1 {
+		t.Fatalf("result=%+v", result)
+	}
+}
+
+func TestHealthHandlerExposesBoundedCapability(t *testing.T) {
+	recorder := httptest.NewRecorder()
+	request := httptest.NewRequest(http.MethodGet, "/healthz", nil)
+	healthHandler(recorder, request)
+	if recorder.Code != http.StatusOK || recorder.Body.String() == "" {
+		t.Fatalf("status=%d body=%q", recorder.Code, recorder.Body.String())
+	}
+}
+
+func TestBridgeRunsTaskEvaluationThroughVerifierAndCallback(t *testing.T) {
+	profile := capability.TaskEvaluationProfile{ID: "profile-1", Revision: 1, CapabilityID: "capability-1", Stage: capability.TaskEvaluationGeneral, BenchmarkVersion: "general-v1", RequiredCaseIDs: []string{"case-1"}, QualityWeightMillis: 600, ReliabilityWeightMillis: 300, LatencyWeightMillis: 100}
+	verification := capability.TaskCaseVerification{Mode: capability.TaskCaseVerificationExactText, ExpectedText: "answer", MinimumQualityMillis: 900}
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/v1/task-evaluations/execute":
+			_ = json.NewEncoder(w).Encode(capability.TaskEvaluationExecutionResponse{ProfileID: profile.ID, BindingID: "binding-1", BindingRevision: 3, Responses: []capability.TaskCaseResponse{{CaseID: "case-1", ResponseText: "answer", LatencyMicros: 2, EvidenceArtifactRef: "artifact-1"}}, ObservedAtUnixMillis: 10})
+		case "/v1/task-evaluations/verify":
+			_ = json.NewEncoder(w).Encode(capability.TaskCaseVerificationResponse{CaseID: "case-1", Result: capability.EvaluationSupported, QualityScoreMillis: 900, LatencyMicros: 2, EvidenceArtifactRef: "artifact-1", VerificationEvidenceRef: "verify-1", VerifierIdentity: "verifier-1"})
+		case "/v1/work-graphs/task-evaluations/callback":
+			w.WriteHeader(http.StatusNoContent)
+		default:
+			http.NotFound(w, r)
+		}
+	}))
+	defer server.Close()
+	b := bridge{cognitiveURL: server.URL, callbackURL: server.URL, verifierURL: server.URL, client: server.Client()}
+	result, err := b.execute(context.Background(), work.AgentExecutionRequest{Intent: work.AgentDispatchIntent{DispatchID: "dispatch-1", TaskEvaluationBindingID: "binding-1", TaskEvaluationProfile: &profile, TaskEvaluationCases: []capability.TaskEvaluationCase{{CaseID: "case-1", Prompt: "answer", EvidenceArtifactRef: "artifact-1", Verification: verification}}, TaskEvaluationExecution: &work.TaskEvaluationDispatchContext{GraphID: "graph-1", GraphRevision: 1, NodeID: "node-1", BindingRevision: 3, ActorID: "actor-1", PrincipalID: "principal-1", ScopeID: "scope-1"}}})
+	if err != nil || result.State != work.AgentExecutionCompleted || result.OutcomeRef == "" {
+		t.Fatalf("result=%+v err=%v", result, err)
+	}
+}
+
+func TestBridgeRunsGenericCognitiveWorkAsUnverifiedGraphCandidate(t *testing.T) {
+	var callback work.GraphResultCallback
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch r.URL.Path {
+		case "/v1/advisories":
+			var request cognitive.GatewayAdvisoryRequest
+			if err := json.NewDecoder(r.Body).Decode(&request); err != nil || request.Message != "find the bounded solution" {
+				t.Fatalf("unexpected advisory request: %+v err=%v", request, err)
+			}
+			_ = json.NewEncoder(w).Encode(cognitive.GatewayAdvisoryResponse{Version: cognitive.CreatorAdvisoryV1, RequestID: request.RequestID, AdvisoryText: "candidate answer", FinishReason: cognitive.FinishCompleted})
+		case "/v1/work-graphs/result-callback":
+			if err := json.NewDecoder(r.Body).Decode(&callback); err != nil {
+				t.Fatal(err)
+			}
+			w.WriteHeader(http.StatusNoContent)
+		case "/v1/work-graphs/candidates/verify":
+			_ = json.NewEncoder(w).Encode(work.GenericCandidateVerificationResponse{Verified: true, VerificationEvidenceRef: "verification:generic-1", VerifierIdentity: "deterministic-generic-envelope-verifier-v1", Summary: "envelope verified"})
+		default:
+			http.NotFound(w, r)
+		}
+	}))
+	defer server.Close()
+	b := bridge{cognitiveURL: server.URL, callbackURL: server.URL, verifierURL: server.URL, client: server.Client()}
+	result, err := b.execute(context.Background(), work.AgentExecutionRequest{Intent: work.AgentDispatchIntent{
+		DispatchID: "dispatch-generic", RunID: "run-generic",
+		TaskEvaluationExecution: &work.TaskEvaluationDispatchContext{GraphID: "graph-1", GraphRevision: 2, NodeID: "node-1", ActorID: "actor-1", PrincipalID: "principal-1", ScopeID: "scope-1", GenericMessage: "find the bounded solution", GenericPurpose: "investigate", GenericDataPolicy: string(capability.DataPolicyLocalOnly)},
+	}})
+	if err != nil || result.State != work.AgentExecutionCompleted || callback.Candidate.Summary != "candidate answer" || callback.Candidate.AuthorRunID != "run-generic" || len(callback.Candidate.ArtifactRefs) != 2 {
+		t.Fatalf("result=%+v callback=%+v err=%v", result, callback, err)
+	}
+}

cmd/agent-worker/local_fallback.go

@@ -7,26 +7,26 @@ import (
 	"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
+// localExecutor provides the bounded local execution capability. It never
+// claims that an unsupported or external action happened. The work owner
+// records the WAITING callback and can retry after a suitable executor is
 // configured or a human chooses the next step.
-type localFallbackExecutor struct {
+type localExecutor struct {
 	now func() time.Time
 }
 
-func newLocalFallbackExecutor() localFallbackExecutor {
-	return localFallbackExecutor{now: time.Now}
+func newLocalExecutor() localExecutor {
+	return localExecutor{now: time.Now}
 }
 
-func (e localFallbackExecutor) Execute(_ context.Context, request work.AgentExecutionRequest) (work.AgentExecutionResult, error) {
+func (e localExecutor) Execute(_ context.Context, request work.AgentExecutionRequest) (work.AgentExecutionResult, error) {
 	now := time.Now()
 	if e.now != nil {
 		now = e.now()
 	}
 	return work.AgentExecutionResult{
 		State:                work.AgentExecutionWaiting,
-		Summary:              "Внешний Agent Execution Bridge недоступен; локальный fallback сохранил durable работу и ожидает исполнителя.",
+		Summary:              "Локальный исполнитель не поддерживает эту операцию; durable работа сохранена и ожидает подходящего исполнителя.",
 		ObservedAtUnixMillis: now.UnixMilli(),
 	}, nil
 }

cmd/agent-worker/local_fallback_test.go

@@ -8,9 +8,9 @@ import (
 	"git.cin.su/m/Assistant/contracts/work"
 )
 
-func TestLocalFallbackNeverClaimsExternalExecution(t *testing.T) {
+func TestLocalExecutorNeverClaimsUnsupportedExecution(t *testing.T) {
 	when := time.UnixMilli(1_700_000_000_000)
-	executor := localFallbackExecutor{now: func() time.Time { return when }}
+	executor := localExecutor{now: func() time.Time { return when }}
 	result, err := executor.Execute(context.Background(), work.AgentExecutionRequest{Intent: work.AgentDispatchIntent{DispatchID: "dispatch-1"}})
 	if err != nil {
 		t.Fatal(err)
@@ -19,6 +19,6 @@ func TestLocalFallbackNeverClaimsExternalExecution(t *testing.T) {
 		t.Fatalf("unexpected fallback result: %+v", result)
 	}
 	if result.OutcomeRef != "" || result.Summary == "" {
-		t.Fatalf("fallback must not claim an outcome: %+v", result)
+		t.Fatalf("local executor must not claim an outcome: %+v", result)
 	}
 }

cmd/agent-worker/main.go

@@ -42,24 +42,29 @@ func main() {
 	defer repository.Close()
 	var executor ports.AgentExecutor
 	localFallback := false
-	if bridgeURL, bridgeConfigured := loadProtectedBridgeURL(); bridgeConfigured {
+	if bridgeURL, bridgeConfigured := loadConfiguredBridgeURL(); bridgeConfigured {
 		executor, err = agentbridge.NewHTTPExecutor(bridgeURL, &http.Client{Timeout: *leaseTTL})
 		if err != nil {
 			fmt.Fprintln(os.Stderr, err)
 			os.Exit(1)
 		}
 	} else {
-		executor = newLocalFallbackExecutor()
+		// The local executor keeps supported task-qualification work alive, but
+		// it is not an external Agent Execution Bridge and must never be reported
+		// as one. Unsupported work remains durably WAITING.
 		localFallback = true
-		fmt.Fprintln(os.Stderr, "protected agent execution bridge URL is not configured; local fallback preserves durable work without external mutation")
+		executor = newTaskEvaluationExecutor(envOr("COGNITIVE_WORKER_URL", "http://cognitive-worker:8083"), envOr("WORK_GRAPH_RUNTIME_URL", "http://work-graph-runtime:8088"), envOr("TASK_VERIFIER_URL", "http://task-verifier:8091"), &http.Client{Timeout: *leaseTTL})
+		fmt.Fprintln(os.Stderr, "protected agent execution bridge URL is not configured; task evaluations use the local durable evaluator, other work remains in fallback")
 	}
+	runtimeStatus := newAgentWorkerRuntimeStatus()
 	go func() {
-		if serveErr := http.ListenAndServe(*statusAddr, workerStatusHandler(executor != nil && !localFallback, localFallback)); serveErr != nil {
+		if serveErr := http.ListenAndServe(*statusAddr, workerStatusHandler(executor != nil && !localFallback, localFallback, runtimeStatus)); serveErr != nil {
 			fmt.Fprintln(os.Stderr, "agent worker status server:", serveErr)
 		}
 	}()
 	ctx := context.Background()
 	for {
+		runtimeStatus.cycleStarted(time.Now())
 		var result application.WorkerCycleResult
 		var cycleErr error
 		if executor == nil {
@@ -68,11 +73,19 @@ func main() {
 			result, cycleErr = application.ExecuteWorkerCycleWithExecutor(ctx, repository, executor, *workerID, work.AgentFencingToken(1), time.Now().UTC(), *staleAfter, *leaseTTL)
 		}
 		if cycleErr != nil {
+			runtimeStatus.cycleFailed(time.Now(), cycleErr)
 			fmt.Fprintln(os.Stderr, cycleErr)
 			if *once {
 				os.Exit(1)
 			}
 		}
+		cycleResult := "idle"
+		if result.Completed {
+			cycleResult = "completed"
+		} else if result.RecoveredClaims > 0 {
+			cycleResult = "recovered"
+		}
+		runtimeStatus.cycleFinished(time.Now(), result.RecoveredClaims, result.DispatchID, cycleResult)
 		if cycleErr == nil && (result.Completed || result.RecoveredClaims > 0) {
 			_ = json.NewEncoder(os.Stdout).Encode(result)
 		}
@@ -86,6 +99,19 @@ func main() {
 	}
 }
 
+func loadConfiguredBridgeURL() (string, bool) {
+	if bridgeURL, configured := loadProtectedBridgeURL(); configured {
+		return bridgeURL, true
+	}
+	// The URL is routing configuration, not a credential. A protected file
+	// remains preferred when an independently operated bridge is supplied.
+	value := strings.TrimSpace(os.Getenv("AGENT_EXECUTION_BRIDGE_URL"))
+	if value == "" || !strings.HasPrefix(value, "http://") && !strings.HasPrefix(value, "https://") {
+		return "", false
+	}
+	return value, true
+}
+
 func loadProtectedBridgeURL() (string, bool) {
 	return loadProtectedBridgeURLFrom("/run/secrets", "/run/secrets/agent_execution_bridge_url")
 }
@@ -107,6 +133,13 @@ func loadProtectedBridgeURLFrom(root, path string) (string, bool) {
 	return value, true
 }
 
+func envOr(name, fallback string) string {
+	if value := strings.TrimSpace(os.Getenv(name)); value != "" {
+		return value
+	}
+	return fallback
+}
+
 func resolveDatabasePassword(name, refName string, consumer secretcontracts.RuntimeConsumer) (string, error) {
 	if _, ok := os.LookupEnv(name); ok {
 		return "", fmt.Errorf("%s direct environment secret is forbidden", name)

cmd/agent-worker/status.go

@@ -3,35 +3,106 @@ package main
 import (
 	"encoding/json"
 	"net/http"
+	"sync"
+	"time"
 )
 
 type workerStatusResponse struct {
-	Runtime            string `json:"runtime"`
-	ExecutorConfigured bool   `json:"executor_configured"`
-	Mode               string `json:"mode"`
-	Ready              bool   `json:"ready"`
-	Reason             string `json:"reason"`
-	LocalFallback      bool   `json:"local_fallback"`
+	Runtime              string    `json:"runtime"`
+	ExecutorConfigured   bool      `json:"executor_configured"`
+	Mode                 string    `json:"mode"`
+	Ready                bool      `json:"ready"`
+	Reason               string    `json:"reason"`
+	LocalFallback        bool      `json:"local_fallback"`
+	LocalExecutor        bool      `json:"local_executor"`
+	Capabilities         []string  `json:"capabilities,omitempty"`
+	StartedAt            time.Time `json:"started_at,omitempty"`
+	LastCycleAt          time.Time `json:"last_cycle_at,omitempty"`
+	LastCompletedAt      time.Time `json:"last_completed_at,omitempty"`
+	LastRecoveredClaims  int64     `json:"last_recovered_claims"`
+	TotalRecoveredClaims int64     `json:"total_recovered_claims"`
+	LastDispatchID       string    `json:"last_dispatch_id,omitempty"`
+	LastResult           string    `json:"last_result,omitempty"`
+	LastError            string    `json:"last_error,omitempty"`
 }
 
-func workerStatusHandler(executorConfigured bool, localFallback ...bool) http.Handler {
-	local := len(localFallback) > 0 && localFallback[0]
-	status := workerStatusResponse{
+type agentWorkerRuntimeStatus struct {
+	mu                   sync.RWMutex
+	startedAt            time.Time
+	lastCycleAt          time.Time
+	lastCompletedAt      time.Time
+	lastRecoveredClaims  int64
+	totalRecoveredClaims int64
+	lastDispatchID       string
+	lastResult           string
+	lastError            string
+}
+
+func newAgentWorkerRuntimeStatus() *agentWorkerRuntimeStatus {
+	return &agentWorkerRuntimeStatus{startedAt: time.Now().UTC()}
+}
+
+func (s *agentWorkerRuntimeStatus) cycleStarted(at time.Time) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.lastCycleAt = at.UTC()
+	s.lastResult = "running"
+	s.lastError = ""
+}
+
+func (s *agentWorkerRuntimeStatus) cycleFinished(at time.Time, recoveredClaims int64, dispatchID string, result string) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.lastCompletedAt = at.UTC()
+	s.lastRecoveredClaims = recoveredClaims
+	s.totalRecoveredClaims += recoveredClaims
+	s.lastDispatchID = dispatchID
+	s.lastResult = result
+	s.lastError = ""
+}
+
+func (s *agentWorkerRuntimeStatus) cycleFailed(at time.Time, err error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	s.lastCompletedAt = at.UTC()
+	s.lastResult = "failed"
+	s.lastError = err.Error()
+}
+
+func workerStatusHandler(executorConfigured bool, localFallback ...interface{}) http.Handler {
+	local := false
+	var runtime *agentWorkerRuntimeStatus
+	for _, argument := range localFallback {
+		switch value := argument.(type) {
+		case bool:
+			local = value
+		case *agentWorkerRuntimeStatus:
+			runtime = value
+		}
+	}
+	if runtime == nil {
+		runtime = newAgentWorkerRuntimeStatus()
+	}
+	baseStatus := workerStatusResponse{
 		Runtime:            "agent-worker",
 		ExecutorConfigured: executorConfigured,
 		Mode:               "recovery-only",
 		Ready:              false,
 		Reason:             "protected agent execution bridge URL is not configured",
 		LocalFallback:      local,
+		LocalExecutor:      local,
 	}
 	if local {
-		status.Mode = "local-fallback"
-		status.Reason = "external bridge unavailable; durable work is retained without external mutation"
+		baseStatus.Mode = "local-executor"
+		baseStatus.Ready = true
+		baseStatus.Reason = "bounded local task-evaluation capability is ready; unsupported work remains durably waiting"
+		baseStatus.Capabilities = []string{"task-evaluation"}
 	}
 	if executorConfigured {
-		status.Mode = "execution"
-		status.Ready = true
-		status.Reason = "typed agent execution bridge configured"
+		baseStatus.Mode = "execution"
+		baseStatus.Ready = true
+		baseStatus.Reason = "typed agent execution bridge configured"
+		baseStatus.Capabilities = []string{"task-evaluation", "generic-cognitive-work"}
 	}
 	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") {
@@ -39,9 +110,20 @@ func workerStatusHandler(executorConfigured bool, localFallback ...bool) http.Ha
 			return
 		}
 		w.Header().Set("Content-Type", "application/json")
-		if r.URL.Path == "/readyz" && !status.Ready {
+		if r.URL.Path == "/readyz" && !baseStatus.Ready {
 			w.WriteHeader(http.StatusServiceUnavailable)
 		}
-		_ = json.NewEncoder(w).Encode(status)
+		response := baseStatus
+		runtime.mu.RLock()
+		response.StartedAt = runtime.startedAt
+		response.LastCycleAt = runtime.lastCycleAt
+		response.LastCompletedAt = runtime.lastCompletedAt
+		response.LastRecoveredClaims = runtime.lastRecoveredClaims
+		response.TotalRecoveredClaims = runtime.totalRecoveredClaims
+		response.LastDispatchID = runtime.lastDispatchID
+		response.LastResult = runtime.lastResult
+		response.LastError = runtime.lastError
+		runtime.mu.RUnlock()
+		_ = json.NewEncoder(w).Encode(response)
 	})
 }

cmd/agent-worker/status_test.go

@@ -7,6 +7,7 @@ import (
 	"os"
 	"path/filepath"
 	"testing"
+	"time"
 )
 
 func TestLoadProtectedBridgeURL(t *testing.T) {
@@ -38,27 +39,76 @@ func TestWorkerStatusReportsRecoveryOnlyWithoutExecutor(t *testing.T) {
 	}
 }
 
-func TestWorkerStatusReportsLocalFallback(t *testing.T) {
+func TestWorkerStatusReportsLocalExecutorCapabilityReady(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 {
+	if response.Code != http.StatusOK || !value.Ready || value.Mode != "local-executor" || !value.LocalFallback || !value.LocalExecutor {
 		t.Fatalf("unexpected status: code=%d value=%+v", response.Code, value)
 	}
+	if len(value.Capabilities) != 1 || value.Capabilities[0] != "task-evaluation" {
+		t.Fatalf("unexpected local capabilities: %+v", value.Capabilities)
+	}
 }
 
-func TestWorkerReadyzRequiresExecutor(t *testing.T) {
+func TestWorkerStatusDoesNotCallLocalFallbackAnExternalExecutor(t *testing.T) {
+	response := httptest.NewRecorder()
+	workerStatusHandler(false, true).ServeHTTP(response, httptest.NewRequest(http.MethodGet, "/status", nil))
+	var value workerStatusResponse
+	if err := json.NewDecoder(response.Body).Decode(&value); err != nil {
+		t.Fatal(err)
+	}
+	if value.ExecutorConfigured || !value.Ready || !value.LocalFallback || !value.LocalExecutor || value.Mode != "local-executor" {
+		t.Fatalf("local fallback was presented as external execution: %+v", value)
+	}
+}
+
+func TestWorkerReadyzRequiresAnAdvertisedCapability(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(false, true).ServeHTTP(response, httptest.NewRequest(http.MethodGet, "/readyz", nil))
+	if response.Code != http.StatusOK {
+		t.Fatalf("local capability 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)
 	}
 }
+
+func TestWorkerStatusReportsRecoveryCycle(t *testing.T) {
+	runtime := newAgentWorkerRuntimeStatus()
+	runtime.cycleStarted(time.Date(2026, 8, 23, 10, 0, 0, 0, time.UTC))
+	runtime.cycleFinished(time.Date(2026, 8, 23, 10, 0, 1, 0, time.UTC), 3, "dispatch-1", "recovered")
+	response := httptest.NewRecorder()
+	workerStatusHandler(false, true, runtime).ServeHTTP(response, httptest.NewRequest(http.MethodGet, "/status", nil))
+	var value workerStatusResponse
+	if err := json.NewDecoder(response.Body).Decode(&value); err != nil {
+		t.Fatal(err)
+	}
+	if value.LastRecoveredClaims != 3 || value.TotalRecoveredClaims != 3 || value.LastDispatchID != "dispatch-1" || value.LastResult != "recovered" {
+		t.Fatalf("recovery evidence missing: %+v", value)
+	}
+}
+
+func TestWorkerStatusReportsCycleError(t *testing.T) {
+	runtime := newAgentWorkerRuntimeStatus()
+	runtime.cycleFailed(time.Date(2026, 8, 23, 10, 0, 1, 0, time.UTC), os.ErrClosed)
+	response := httptest.NewRecorder()
+	workerStatusHandler(false, runtime).ServeHTTP(response, httptest.NewRequest(http.MethodGet, "/status", nil))
+	var value workerStatusResponse
+	if err := json.NewDecoder(response.Body).Decode(&value); err != nil {
+		t.Fatal(err)
+	}
+	if value.LastResult != "failed" || value.LastError == "" {
+		t.Fatalf("cycle error missing: %+v", value)
+	}
+}

cmd/agent-worker/task_evaluation_executor.go

@@ -0,0 +1,129 @@
+package main
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"strings"
+	"time"
+
+	"git.cin.su/m/Assistant/contracts/artifact"
+	"git.cin.su/m/Assistant/contracts/capability"
+	"git.cin.su/m/Assistant/contracts/work"
+	capabilitydomain "git.cin.su/m/Assistant/modules/capability/domain"
+)
+
+type taskEvaluationExecutor struct {
+	cognitiveURL string
+	callbackURL  string
+	verifierURL  string
+	client       *http.Client
+	fallback     localExecutor
+}
+
+func newTaskEvaluationExecutor(cognitiveURL, callbackURL, verifierURL string, client *http.Client) taskEvaluationExecutor {
+	if client == nil {
+		client = http.DefaultClient
+	}
+	return taskEvaluationExecutor{cognitiveURL: strings.TrimRight(cognitiveURL, "/"), callbackURL: strings.TrimRight(callbackURL, "/"), verifierURL: strings.TrimRight(verifierURL, "/"), client: client, fallback: newLocalExecutor()}
+}
+
+func (e taskEvaluationExecutor) Execute(ctx context.Context, request work.AgentExecutionRequest) (work.AgentExecutionResult, error) {
+	intent := request.Intent
+	if intent.TaskEvaluationExecution == nil || intent.TaskEvaluationProfile == nil || len(intent.TaskEvaluationCases) == 0 {
+		return e.fallback.Execute(ctx, request)
+	}
+	execution := intent.TaskEvaluationExecution
+	if execution.GraphID == "" || execution.NodeID == "" || execution.BindingRevision < 1 || execution.ActorID == "" || execution.PrincipalID == "" || execution.ScopeID == "" {
+		return e.fallback.Execute(ctx, request)
+	}
+	if e.cognitiveURL == "" || e.callbackURL == "" || e.verifierURL == "" {
+		return e.fallback.Execute(ctx, request)
+	}
+	payload := capability.TaskEvaluationExecutionRequest{Profile: *intent.TaskEvaluationProfile, CapabilityID: intent.TaskEvaluationProfile.CapabilityID, BindingID: intent.TaskEvaluationBindingID, BindingRevision: execution.BindingRevision, ActorID: execution.ActorID, PrincipalID: execution.PrincipalID, ScopeID: execution.ScopeID, Cases: intent.TaskEvaluationCases, MaxOutputBytes: 4096}
+	response, err := postJSON[capability.TaskEvaluationExecutionResponse](ctx, e.client, e.cognitiveURL+"/v1/task-evaluations/execute", payload)
+	if err != nil {
+		return work.AgentExecutionResult{}, err
+	}
+	if len(response.Responses) == 0 || response.ObservedAtUnixMillis < 1 {
+		return work.AgentExecutionResult{}, fmt.Errorf("task evaluation returned no observations")
+	}
+	if len(response.Observations) == len(intent.TaskEvaluationCases) && taskEvaluationHasCriticalFailure(response.Observations) {
+		evaluation, err := capabilitydomain.BuildTaskEvaluation(*intent.TaskEvaluationProfile, capability.TaskEvaluationID("evaluation:"+intent.DispatchID), response.BindingID, response.BindingRevision, intent.TaskEvaluationProfile.CapabilityID, response.Observations, artifact.ArtifactID(intent.TaskEvaluationCases[0].EvidenceArtifactRef), response.ObservedAtUnixMillis)
+		if err != nil {
+			return work.AgentExecutionResult{}, err
+		}
+		callback := work.TaskEvaluationEvidenceCallback{GraphID: execution.GraphID, Revision: execution.GraphRevision, NodeID: execution.NodeID, Evaluation: evaluation, Succeeded: true}
+		if _, err := postJSON[json.RawMessage](ctx, e.client, e.callbackURL+"/v1/work-graphs/task-evaluations/callback", callback); err != nil {
+			return work.AgentExecutionResult{}, err
+		}
+		return work.AgentExecutionResult{State: work.AgentExecutionCompleted, Summary: "task evaluation recorded as UNKNOWN after bounded provider recovery", OutcomeRef: string(evaluation.ID), ArtifactRefs: []string{string(evaluation.ArtifactID)}, ObservedAtUnixMillis: time.Now().UnixMilli()}, nil
+	}
+	if len(intent.TaskEvaluationCases) == 0 {
+		return work.AgentExecutionResult{}, fmt.Errorf("task evaluation cases are missing")
+	}
+	caseByID := make(map[string]capability.TaskEvaluationCase, len(intent.TaskEvaluationCases))
+	for _, item := range intent.TaskEvaluationCases {
+		caseByID[item.CaseID] = item
+	}
+	observations := make([]capability.TaskCaseObservation, 0, len(response.Responses))
+	for _, item := range response.Responses {
+		caseValue, ok := caseByID[item.CaseID]
+		if !ok || caseValue.Verification.Mode == "" {
+			return e.fallback.Execute(ctx, request)
+		}
+		verified, verifyErr := postJSON[capability.TaskCaseVerificationResponse](ctx, e.client, e.verifierURL+"/v1/task-evaluations/verify", capability.TaskCaseVerificationRequest{CaseID: item.CaseID, ResponseText: item.ResponseText, LatencyMicros: item.LatencyMicros, EvidenceArtifactRef: item.EvidenceArtifactRef, Verification: caseValue.Verification})
+		if verifyErr != nil {
+			return e.fallback.Execute(ctx, request)
+		}
+		observations = append(observations, capability.TaskCaseObservation{CaseID: verified.CaseID, Result: verified.Result, QualityScoreMillis: verified.QualityScoreMillis, LatencyMicros: verified.LatencyMicros, CriticalFailure: verified.CriticalFailure, EvidenceArtifactRef: verified.EvidenceArtifactRef, VerificationEvidenceRef: verified.VerificationEvidenceRef, VerifierIdentity: verified.VerifierIdentity})
+	}
+	evaluation, err := capabilitydomain.BuildTaskEvaluation(*intent.TaskEvaluationProfile, capability.TaskEvaluationID("evaluation:"+intent.DispatchID), response.BindingID, response.BindingRevision, intent.TaskEvaluationProfile.CapabilityID, observations, artifact.ArtifactID(intent.TaskEvaluationCases[0].EvidenceArtifactRef), response.ObservedAtUnixMillis)
+	if err != nil {
+		return work.AgentExecutionResult{}, err
+	}
+	callback := work.TaskEvaluationEvidenceCallback{GraphID: execution.GraphID, Revision: execution.GraphRevision, NodeID: execution.NodeID, Evaluation: evaluation, Succeeded: true}
+	if _, err := postJSON[json.RawMessage](ctx, e.client, e.callbackURL+"/v1/work-graphs/task-evaluations/callback", callback); err != nil {
+		return work.AgentExecutionResult{}, err
+	}
+	return work.AgentExecutionResult{State: work.AgentExecutionCompleted, Summary: "task evaluation executed, aggregated and accepted by Work Graph callback", OutcomeRef: string(evaluation.ID), ArtifactRefs: []string{string(evaluation.ArtifactID)}, ObservedAtUnixMillis: time.Now().UnixMilli()}, nil
+}
+
+func taskEvaluationHasCriticalFailure(observations []capability.TaskCaseObservation) bool {
+	for _, observation := range observations {
+		if observation.CriticalFailure {
+			return true
+		}
+	}
+	return false
+}
+
+func postJSON[T any](ctx context.Context, client *http.Client, url string, payload any) (T, error) {
+	var result T
+	body, err := json.Marshal(payload)
+	if err != nil {
+		return result, err
+	}
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
+	if err != nil {
+		return result, err
+	}
+	req.Header.Set("Content-Type", "application/json")
+	resp, err := client.Do(req)
+	if err != nil {
+		return result, err
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode < 200 || resp.StatusCode >= 300 {
+		return result, fmt.Errorf("POST %s: status %s", url, resp.Status)
+	}
+	if resp.StatusCode == http.StatusNoContent {
+		return result, nil
+	}
+	if err := json.NewDecoder(resp.Body).Decode(&result); err != nil {
+		return result, err
+	}
+	return result, nil
+}

cmd/agent-worker/task_evaluation_executor_test.go

@@ -0,0 +1,72 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"git.cin.su/m/Assistant/contracts/capability"
+	"git.cin.su/m/Assistant/contracts/work"
+)
+
+func TestTaskEvaluationExecutorInvokesAggregatesAndCallbacks(t *testing.T) {
+	profile := capability.TaskEvaluationProfile{ID: "profile-1", Revision: 1, CapabilityID: "capability-1", Stage: capability.TaskEvaluationGeneral, BenchmarkVersion: "general-v1", RequiredCaseIDs: []string{"case-1"}, QualityWeightMillis: 600, ReliabilityWeightMillis: 300, LatencyWeightMillis: 100}
+	verification := capability.TaskCaseVerification{Mode: capability.TaskCaseVerificationExactText, ExpectedText: "answer", MinimumQualityMillis: 900}
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path == "/v1/task-evaluations/execute" {
+			_ = json.NewEncoder(w).Encode(capability.TaskEvaluationExecutionResponse{ProfileID: profile.ID, BindingID: "binding-1", BindingRevision: 3, Responses: []capability.TaskCaseResponse{{CaseID: "case-1", ResponseText: "answer", LatencyMicros: 2, EvidenceArtifactRef: "artifact-1"}}, ObservedAtUnixMillis: 10})
+			return
+		}
+		if r.URL.Path == "/v1/task-evaluations/verify" {
+			_ = json.NewEncoder(w).Encode(capability.TaskCaseVerificationResponse{CaseID: "case-1", Result: capability.EvaluationSupported, QualityScoreMillis: 900, LatencyMicros: 2, EvidenceArtifactRef: "artifact-1", VerificationEvidenceRef: "verify-1", VerifierIdentity: "verifier-1"})
+			return
+		}
+		if r.URL.Path == "/v1/work-graphs/task-evaluations/callback" {
+			var callback work.TaskEvaluationEvidenceCallback
+			if err := json.NewDecoder(r.Body).Decode(&callback); err != nil || callback.GraphID != "graph-1" || callback.NodeID != "node-1" || callback.Evaluation.Result != capability.EvaluationSupported {
+				http.Error(w, "invalid callback", http.StatusBadRequest)
+				return
+			}
+			w.WriteHeader(http.StatusNoContent)
+			return
+		}
+		http.NotFound(w, r)
+	}))
+	defer server.Close()
+	executor := newTaskEvaluationExecutor(server.URL, server.URL, server.URL, server.Client())
+	result, err := executor.Execute(context.Background(), work.AgentExecutionRequest{Intent: work.AgentDispatchIntent{DispatchID: "dispatch-1", TaskEvaluationBindingID: "binding-1", TaskEvaluationProfile: &profile, TaskEvaluationCases: []capability.TaskEvaluationCase{{CaseID: "case-1", Prompt: "answer", EvidenceArtifactRef: "artifact-1", Verification: verification}}, TaskEvaluationExecution: &work.TaskEvaluationDispatchContext{GraphID: "graph-1", GraphRevision: 1, NodeID: "node-1", BindingRevision: 3, ActorID: "actor-1", PrincipalID: "principal-1", ScopeID: "scope-1"}}})
+	if err != nil || result.State != work.AgentExecutionCompleted || result.OutcomeRef == "" {
+		t.Fatalf("result=%+v err=%v", result, err)
+	}
+}
+
+func TestTaskEvaluationExecutorPersistsUnknownAfterProviderFailure(t *testing.T) {
+	profile := capability.TaskEvaluationProfile{ID: "profile-unknown", Revision: 1, CapabilityID: "capability-1", Stage: capability.TaskEvaluationGeneral, BenchmarkVersion: "unknown-v1", RequiredCaseIDs: []string{"case-1"}, QualityWeightMillis: 600, ReliabilityWeightMillis: 300, LatencyWeightMillis: 100}
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if r.URL.Path == "/v1/task-evaluations/execute" {
+			_ = json.NewEncoder(w).Encode(capability.TaskEvaluationExecutionResponse{ProfileID: profile.ID, BindingID: "binding-1", BindingRevision: 3, Observations: []capability.TaskCaseObservation{{CaseID: "case-1", Result: capability.EvaluationUnknown, CriticalFailure: true, EvidenceArtifactRef: "artifact-1"}}, Responses: []capability.TaskCaseResponse{{CaseID: "case-1", EvidenceArtifactRef: "artifact-1"}}, ObservedAtUnixMillis: 10})
+			return
+		}
+		if r.URL.Path == "/v1/work-graphs/task-evaluations/callback" {
+			var callback work.TaskEvaluationEvidenceCallback
+			if err := json.NewDecoder(r.Body).Decode(&callback); err != nil || callback.Evaluation.Result != capability.EvaluationUnknown || !callback.Succeeded {
+				http.Error(w, "invalid unknown callback", http.StatusBadRequest)
+				return
+			}
+			w.WriteHeader(http.StatusNoContent)
+			return
+		}
+		if r.URL.Path == "/v1/task-evaluations/verify" {
+			t.Errorf("verifier must not be called for a provider failure")
+		}
+		http.NotFound(w, r)
+	}))
+	defer server.Close()
+	executor := newTaskEvaluationExecutor(server.URL, server.URL, server.URL, server.Client())
+	result, err := executor.Execute(context.Background(), work.AgentExecutionRequest{Intent: work.AgentDispatchIntent{DispatchID: "dispatch-unknown", TaskEvaluationBindingID: "binding-1", TaskEvaluationProfile: &profile, TaskEvaluationCases: []capability.TaskEvaluationCase{{CaseID: "case-1", Prompt: "answer", EvidenceArtifactRef: "artifact-1"}}, TaskEvaluationExecution: &work.TaskEvaluationDispatchContext{GraphID: "graph-1", GraphRevision: 1, NodeID: "node-1", BindingRevision: 3, ActorID: "actor-1", PrincipalID: "principal-1", ScopeID: "scope-1"}}})
+	if err != nil || result.State != work.AgentExecutionCompleted || result.OutcomeRef == "" {
+		t.Fatalf("result=%+v err=%v", result, err)
+	}
+}

cmd/cognitive-worker/discovery.go

@@ -3,6 +3,7 @@ package main
 import (
 	"encoding/json"
 	"net/http"
+	"sort"
 	"time"
 
 	"git.cin.su/m/Assistant/contracts/capability"
@@ -17,6 +18,10 @@ type discoveryService struct {
 }
 
 func (s discoveryService) handler(w http.ResponseWriter, r *http.Request) {
+	if r.Method == http.MethodGet && r.URL.Path == "/v1/provider-discovery/status" {
+		s.statusHandler(w, r)
+		return
+	}
 	registration, ok := s.providers[r.PathValue("provider")]
 	if !ok || registration.Discovery == nil {
 		http.Error(w, "provider discovery not found", http.StatusNotFound)
@@ -35,3 +40,42 @@ func (s discoveryService) handler(w http.ResponseWriter, r *http.Request) {
 	w.Header().Set("Content-Type", "application/json")
 	_ = json.NewEncoder(w).Encode(result)
 }
+
+type discoveryStatus struct {
+	Provider             string                          `json:"provider"`
+	ProviderKind         capability.ProviderKind         `json:"provider_kind"`
+	Health               capability.HealthState          `json:"health"`
+	FailureClass         capability.CognitiveFailureType `json:"failure_class,omitempty"`
+	ObservedAtUnixMillis int64                           `json:"observed_at_unix_millis"`
+	Models               []discoveredModelStatus         `json:"models"`
+}
+
+type discoveredModelStatus struct {
+	ProviderModelID capability.ProviderModelID   `json:"provider_model_id"`
+	Fingerprint     capability.ModelFingerprint  `json:"fingerprint,omitempty"`
+	Capabilities    []capability.ModelCapability `json:"capabilities,omitempty"`
+}
+
+func (s discoveryService) statusHandler(w http.ResponseWriter, r *http.Request) {
+	providers := make([]string, 0, len(s.providers))
+	for name := range s.providers {
+		providers = append(providers, name)
+	}
+	sort.Strings(providers)
+	result := make([]discoveryStatus, 0, len(providers))
+	for _, name := range providers {
+		registration := s.providers[name]
+		status := discoveryStatus{Provider: name, ProviderKind: registration.Kind, Health: capability.HealthUnknown, ObservedAtUnixMillis: s.now().UnixMilli(), Models: []discoveredModelStatus{}}
+		if registration.Discovery != nil {
+			models, probe := registration.Discovery.Discover(r.Context())
+			status.Health = probe.Health
+			status.FailureClass = capability.CognitiveFailureType(probe.Failure)
+			for _, model := range models {
+				status.Models = append(status.Models, discoveredModelStatus{ProviderModelID: model.ProviderModelID, Fingerprint: model.Fingerprint, Capabilities: model.Capabilities})
+			}
+		}
+		result = append(result, status)
+	}
+	w.Header().Set("Content-Type", "application/json")
+	_ = json.NewEncoder(w).Encode(result)
+}

cmd/cognitive-worker/discovery_test.go

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

cmd/cognitive-worker/gateway.go

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

cmd/cognitive-worker/gateway_registry_test.go

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

cmd/cognitive-worker/interpretation.go

@@ -163,8 +163,8 @@ func interpretationPrompt(request cognitive.GatewayInterpretationRequest) string
 Allowed speech acts: DIRECTIVE, ASSERTION, DECISION, PREFERENCE, QUESTION, HYPOTHESIS, CORRECTION, REQUEST_TO_INVESTIGATE, CONFIRMATION, REJECTION, CANCELLATION, CONVERSATIONAL, UNKNOWN.
 Allowed applicability: APPLICABLE, LOW_CONFIDENCE, MIXED_INTENT, NEEDS_BROADER_INTERPRETATION, CLARIFICATION_REQUIRED.
 Allowed clarification: NOT_REQUIRED or REQUIRED.
-Capability hints may only contain known capability kinds like COGNITIVE_REASONING, MODIFY_CODE, RUN_TESTS, SEND_EMAIL, SCHEDULE_REMINDER, VERIFY_FACT, INSPECT_PROJECT, READ_DATABASE, QUERY_LOGS.
-Interpret the human utterance semantically. Do not claim action execution. Do not treat punctuation alone as decisive. If uncertain, set applicability accordingly.
+Capability hints may only contain known capability kinds like COGNITIVE_REASONING, WEB_RESEARCH, MODIFY_CODE, RUN_TESTS, SEND_EMAIL, SCHEDULE_REMINDER, VERIFY_FACT, INSPECT_PROJECT, READ_DATABASE, QUERY_LOGS.
+Interpret the human utterance semantically. Do not claim action execution. A question asking for a current, external or otherwise source-dependent fact should receive WEB_RESEARCH; do not answer that fact from model memory. A request to discuss hypotheses or options is different and may receive COGNITIVE_REASONING. Do not treat punctuation alone as decisive. If uncertain, set applicability accordingly.
 Message: %q
 Deterministic primary: %s
 Deterministic secondary: %s

cmd/cognitive-worker/main.go

@@ -97,7 +97,8 @@ func main() {
 	if gpuErr != nil {
 		log.Printf("optional GPU provider disabled: %v", gpuErr)
 	} else if gpuConfigured {
-		gpuAdapter, adapterErr := models.NewOpenAICompatibleLocalAdapterWithAPIKey(gpuURL, gpuKey, nil)
+		gpuTimeout := durationEnv("LOCAL_GPU_PROVIDER_TIMEOUT", 60*time.Second)
+		gpuAdapter, adapterErr := models.NewOpenAICompatibleLocalAdapterWithAPIKeyAndTimeout(gpuURL, gpuKey, nil, gpuTimeout)
 		if adapterErr != nil {
 			log.Printf("optional GPU provider disabled: %v", adapterErr)
 		} else {
@@ -166,8 +167,11 @@ func main() {
 	mux.HandleFunc("POST /v1/advisory-policy", gateway.policyHandler)
 	mux.HandleFunc("POST /v1/advisories", gateway.advisoryHandler)
 	mux.HandleFunc("POST /v1/interpretations", gateway.interpretationHandler)
+	mux.HandleFunc("POST /v1/task-evaluations/execute", gateway.taskEvaluationHandler)
 	discovery := discoveryService{resolver: resolver, providers: discoveries, owner: identity.PrincipalID(envOr("COGNITIVE_OWNER_PRINCIPAL_ID", "domovoy-creator-principal")), now: time.Now}
 	mux.HandleFunc("POST /v1/provider-discovery/{provider}", discovery.handler)
+	mux.Handle("POST /v1/provider-discovery/{provider}/qualify", modelQualificationHandler(discoveries, string(envOr("COGNITIVE_OWNER_PRINCIPAL_ID", "domovoy-creator-principal")), httpQualificationObserver{resolver: resolver}))
+	mux.HandleFunc("GET /v1/provider-discovery/status", discovery.statusHandler)
 	refreshProviderDiscoveries(resolver, discovery)
 	log.Printf("cognitive-worker listening on %s", addr)
 	if err := http.ListenAndServe(addr, mux); err != nil {

cmd/cognitive-worker/main_test.go

@@ -4,6 +4,7 @@ import (
 	"os"
 	"path/filepath"
 	"testing"
+	"time"
 )
 
 func TestLifeboatProviderIsOptionalAndLocalAdapterBacked(t *testing.T) {
@@ -19,6 +20,21 @@ func TestLifeboatProviderIsOptionalAndLocalAdapterBacked(t *testing.T) {
 	}
 }
 
+func TestGPUProviderTimeoutUsesBoundedConfigurableDefault(t *testing.T) {
+	t.Setenv("LOCAL_GPU_PROVIDER_TIMEOUT", "")
+	if got := durationEnv("LOCAL_GPU_PROVIDER_TIMEOUT", 60*time.Second); got != 60*time.Second {
+		t.Fatalf("default GPU timeout=%s", got)
+	}
+	t.Setenv("LOCAL_GPU_PROVIDER_TIMEOUT", "31s")
+	if got := durationEnv("LOCAL_GPU_PROVIDER_TIMEOUT", 60*time.Second); got != 31*time.Second {
+		t.Fatalf("configured GPU timeout=%s", got)
+	}
+	t.Setenv("LOCAL_GPU_PROVIDER_TIMEOUT", "invalid")
+	if got := durationEnv("LOCAL_GPU_PROVIDER_TIMEOUT", 60*time.Second); got != 60*time.Second {
+		t.Fatalf("invalid GPU timeout=%s", got)
+	}
+}
+
 func TestLoadLocalCPUProviderAPIKeyUsesProtectedFileReference(t *testing.T) {
 	t.Setenv("LOCAL_CPU_PROVIDER_API_KEY", "")
 	_ = os.Unsetenv("LOCAL_CPU_PROVIDER_API_KEY")

cmd/cognitive-worker/model_qualification.go

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

cmd/cognitive-worker/model_qualification_test.go

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

cmd/cognitive-worker/task_evaluation.go

@@ -0,0 +1,139 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"strings"
+	"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
+	}
+
+	caseSet := make(map[string]struct{}, len(request.Profile.RequiredCaseIDs))
+	seenCases := make(map[string]struct{}, len(request.Cases))
+	for _, id := range request.Profile.RequiredCaseIDs {
+		caseSet[id] = struct{}{}
+	}
+	for _, 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{}{}
+	}
+	if len(seenCases) != len(caseSet) {
+		http.Error(w, "task evaluation case set is incomplete", http.StatusBadRequest)
+		return
+	}
+
+	// One bounded fan-out wave uses one binding snapshot. If that candidate is
+	// unavailable, only this binding is excluded and the existing resolver
+	// chooses the next eligible candidate. Provider output remains untrusted
+	// until the independent verifier accepts it downstream.
+	requirement := capability.CognitiveRequirement{CapabilityID: request.CapabilityID, OwnerPrincipalID: request.PrincipalID, DataPolicy: selected.Binding.DataPolicy, RequiredModalities: []capability.ModelModality{capability.ModalityText}, RequiredCapabilities: []capability.ModelCapability{capability.CapabilityAdvisory}, AllowCanary: g.allowCanary, MaxOutputBytes: request.MaxOutputBytes}
+	var observations []capability.TaskCaseObservation
+	var responses []capability.TaskCaseResponse
+	for attempt := 0; attempt < 3; attempt++ {
+		provider := g.provider(selected)
+		if provider == nil {
+			observations, responses = taskEvaluationUnavailable(request.Cases)
+		} else {
+			observations, responses = g.executeTaskEvaluationCases(r.Context(), request, selected, provider)
+		}
+		if taskEvaluationResponsesUsable(responses) {
+			break
+		}
+		requirement.ExcludedBindingIDs = append(requirement.ExcludedBindingIDs, selected.Binding.ID)
+		fallback, resolveErr := g.resolver.Resolve(r.Context(), requirement)
+		fallbackProvider := g.provider(fallback)
+		if resolveErr != nil || fallbackProvider == nil || fallback.Binding.ID == selected.Binding.ID || (fallback.Binding.DataPolicy != capability.DataPolicyLocalOnly && !request.EgressAcknowledged) {
+			break
+		}
+		selected = fallback
+	}
+
+	writeTaskEvaluationJSON(w, capability.TaskEvaluationExecutionResponse{ProfileID: request.Profile.ID, BindingID: selected.Binding.ID, BindingRevision: selected.Binding.Revision, Observations: observations, Responses: responses, ObservedAtUnixMillis: time.Now().UnixMilli()})
+}
+
+func (g cognitiveGateway) executeTaskEvaluationCases(ctx context.Context, request capability.TaskEvaluationExecutionRequest, selected capability.ResolvedCognitiveBinding, provider advisoryProvider) ([]capability.TaskCaseObservation, []capability.TaskCaseResponse) {
+	observations := make([]capability.TaskCaseObservation, len(request.Cases))
+	responses := make([]capability.TaskCaseResponse, len(request.Cases))
+	var wg sync.WaitGroup
+	for i, item := range request.Cases {
+		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(ctx, string(invocation.RequestID), provider, invocation, int(request.MaxOutputBytes))
+			latency := time.Since(started).Microseconds()
+			observations[index] = capability.TaskCaseObservation{CaseID: item.CaseID, Result: capability.EvaluationUnknown, CriticalFailure: invokeErr != nil || response.FinishReason != cognitive.FinishCompleted || strings.TrimSpace(response.AdvisoryText) == "", EvidenceArtifactRef: item.EvidenceArtifactRef, LatencyMicros: latency}
+			responses[index] = capability.TaskCaseResponse{CaseID: item.CaseID, ResponseText: response.AdvisoryText, LatencyMicros: latency, EvidenceArtifactRef: item.EvidenceArtifactRef}
+		}(i, item)
+	}
+	wg.Wait()
+	return observations, responses
+}
+
+func taskEvaluationUnavailable(cases []capability.TaskEvaluationCase) ([]capability.TaskCaseObservation, []capability.TaskCaseResponse) {
+	observations := make([]capability.TaskCaseObservation, len(cases))
+	responses := make([]capability.TaskCaseResponse, len(cases))
+	for i, item := range cases {
+		observations[i] = capability.TaskCaseObservation{CaseID: item.CaseID, Result: capability.EvaluationUnknown, CriticalFailure: true, EvidenceArtifactRef: item.EvidenceArtifactRef}
+		responses[i] = capability.TaskCaseResponse{CaseID: item.CaseID, EvidenceArtifactRef: item.EvidenceArtifactRef}
+	}
+	return observations, responses
+}
+
+func taskEvaluationResponsesUsable(responses []capability.TaskCaseResponse) bool {
+	if len(responses) == 0 {
+		return false
+	}
+	for _, response := range responses {
+		if strings.TrimSpace(response.ResponseText) == "" {
+			return false
+		}
+	}
+	return true
+}
+
+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 +1,145 @@
+package main
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+	"time"
+
+	"git.cin.su/m/Assistant/contracts/authority"
+	"git.cin.su/m/Assistant/contracts/capability"
+	"git.cin.su/m/Assistant/contracts/cognitive"
+	"git.cin.su/m/Assistant/contracts/identity"
+)
+
+type candidateRegistryStub struct {
+	registryStub
+	selected capability.ResolvedCognitiveBinding
+}
+
+func (s *candidateRegistryStub) ResolveCandidate(context.Context, capability.CapabilityBindingID, int64) (capability.ResolvedCognitiveBinding, error) {
+	return s.selected, nil
+}
+
+func TestTaskEvaluationHandlerInvokesEveryBoundedCase(t *testing.T) {
+	selected := gatewaySelection()
+	registry := &candidateRegistryStub{selected: selected}
+	provider := &stubAdvisoryProvider{response: cognitive.AdvisoryResponse{Version: cognitive.CreatorAdvisoryV1, FinishReason: cognitive.FinishCompleted, AdvisoryText: "observed response"}}
+	gateway := cognitiveGateway{
+		resolver: registry,
+		providers: map[capability.ProviderKind][]providerRegistration{
+			selected.Provider.Kind: {{ProviderInstanceID: selected.Provider.ID, Kind: selected.Provider.Kind, EndpointRef: selected.Provider.EndpointRef, CredentialRef: selected.Provider.CredentialRef, Provider: provider}},
+		},
+		now: func() time.Time { return time.UnixMilli(1000) },
+	}
+	profile := capability.TaskEvaluationProfile{ID: "profile-general", Revision: 1, CapabilityID: selected.Binding.CapabilityID, Stage: capability.TaskEvaluationGeneral, BenchmarkVersion: "general-v1", RequiredCaseIDs: []string{"case-a", "case-b"}, QualityWeightMillis: 600, ReliabilityWeightMillis: 300, LatencyWeightMillis: 100}
+	request := capability.TaskEvaluationExecutionRequest{Profile: profile, CapabilityID: selected.Binding.CapabilityID, BindingID: selected.Binding.ID, BindingRevision: selected.Binding.Revision, ActorID: identity.ActorID("actor-1"), PrincipalID: selected.Binding.OwnerPrincipalID, ScopeID: authority.ScopeID("scope-1"), MaxOutputBytes: 512, EgressAcknowledged: true, Cases: []capability.TaskEvaluationCase{{CaseID: "case-a", Prompt: "one", EvidenceArtifactRef: "artifact:a"}, {CaseID: "case-b", Prompt: "two", EvidenceArtifactRef: "artifact:b"}}}
+	payload, _ := json.Marshal(request)
+	recorder := httptest.NewRecorder()
+	gateway.taskEvaluationHandler(recorder, httptest.NewRequest(http.MethodPost, "/v1/task-evaluations/execute", bytes.NewReader(payload)))
+	if recorder.Code != http.StatusOK {
+		t.Fatalf("status=%d body=%s", recorder.Code, recorder.Body.String())
+	}
+	var response capability.TaskEvaluationExecutionResponse
+	if err := json.Unmarshal(recorder.Body.Bytes(), &response); err != nil {
+		t.Fatal(err)
+	}
+	if len(response.Observations) != 2 || response.Observations[0].Result != capability.EvaluationUnknown || response.Observations[1].Result != capability.EvaluationUnknown {
+		t.Fatalf("unexpected observations: %+v", response.Observations)
+	}
+	for _, observation := range response.Observations {
+		if observation.QualityScoreMillis != 0 || observation.CriticalFailure {
+			t.Fatalf("reachability-only evidence must not carry quality or failure claims: %+v", observation)
+		}
+	}
+}
+
+type taskEvaluationFailoverRegistry struct {
+	primary, fallback capability.ResolvedCognitiveBinding
+	resolveCalls      int
+}
+
+func (s *taskEvaluationFailoverRegistry) ResolveCandidate(context.Context, capability.CapabilityBindingID, int64) (capability.ResolvedCognitiveBinding, error) {
+	return s.primary, nil
+}
+
+func (s *taskEvaluationFailoverRegistry) Resolve(_ context.Context, requirement capability.CognitiveRequirement) (capability.ResolvedCognitiveBinding, error) {
+	s.resolveCalls++
+	if len(requirement.ExcludedBindingIDs) == 0 {
+		return s.primary, nil
+	}
+	return s.fallback, nil
+}
+
+func (s *taskEvaluationFailoverRegistry) Observe(context.Context, capability.CognitiveUsageObservation) error {
+	return nil
+}
+
+func (s *taskEvaluationFailoverRegistry) RegisterDiscovery(context.Context, capability.CognitiveDiscoverySnapshot) (capability.CognitiveDiscoveryRegistration, error) {
+	return capability.CognitiveDiscoveryRegistration{}, nil
+}
+
+func TestTaskEvaluationExcludesUnavailableCandidateAndUsesFallback(t *testing.T) {
+	primary := gatewaySelection()
+	fallback := primary
+	fallback.Binding.ID = "binding-fallback"
+	fallback.Binding.Revision = 7
+	fallback.Binding.ProviderInstanceID = "provider-fallback"
+	fallback.Provider.ID = "provider-fallback"
+	fallback.Provider.Kind = "provider-fallback-kind"
+	fallback.Provider.EndpointRef = "fallback-endpoint"
+	fallback.Provider.CredentialRef = "fallback-credential"
+	fallback.Model.ProviderInstanceID = "provider-fallback"
+	registry := &taskEvaluationFailoverRegistry{primary: primary, fallback: fallback}
+	primaryProvider := &stubAdvisoryProvider{response: cognitive.AdvisoryResponse{Version: cognitive.CreatorAdvisoryV1, FinishReason: cognitive.FinishProviderUnavailable}}
+	fallbackProvider := &stubAdvisoryProvider{response: cognitive.AdvisoryResponse{Version: cognitive.CreatorAdvisoryV1, FinishReason: cognitive.FinishCompleted, AdvisoryText: "fallback response"}}
+	gateway := cognitiveGateway{resolver: registry, providers: map[capability.ProviderKind][]providerRegistration{
+		primary.Provider.Kind:  {{ProviderInstanceID: primary.Provider.ID, Kind: primary.Provider.Kind, EndpointRef: primary.Provider.EndpointRef, CredentialRef: primary.Provider.CredentialRef, Provider: primaryProvider}},
+		fallback.Provider.Kind: {{ProviderInstanceID: fallback.Provider.ID, Kind: fallback.Provider.Kind, EndpointRef: fallback.Provider.EndpointRef, CredentialRef: fallback.Provider.CredentialRef, Provider: fallbackProvider}},
+	}, now: func() time.Time { return time.UnixMilli(1000) }}
+	profile := capability.TaskEvaluationProfile{ID: "profile-failover", Revision: 1, CapabilityID: primary.Binding.CapabilityID, Stage: capability.TaskEvaluationGeneral, BenchmarkVersion: "failover-v1", RequiredCaseIDs: []string{"case-a"}, QualityWeightMillis: 600, ReliabilityWeightMillis: 300, LatencyWeightMillis: 100}
+	request := capability.TaskEvaluationExecutionRequest{Profile: profile, CapabilityID: profile.CapabilityID, BindingID: primary.Binding.ID, BindingRevision: primary.Binding.Revision, ActorID: identity.ActorID("actor-1"), PrincipalID: primary.Binding.OwnerPrincipalID, ScopeID: authority.ScopeID("scope-1"), MaxOutputBytes: 512, EgressAcknowledged: true, Cases: []capability.TaskEvaluationCase{{CaseID: "case-a", Prompt: "one", EvidenceArtifactRef: "artifact:a"}}}
+	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 response.BindingID != fallback.Binding.ID || !primaryProvider.called || !fallbackProvider.called || registry.resolveCalls != 1 {
+		t.Fatalf("failover binding=%s primary_called=%t fallback_called=%t resolve_calls=%d", response.BindingID, primaryProvider.called, fallbackProvider.called, registry.resolveCalls)
+	}
+	if response.Observations[0].CriticalFailure || response.Responses[0].ResponseText != "fallback response" {
+		t.Fatalf("fallback response was not usable: %+v %+v", response.Observations, response.Responses)
+	}
+}
+
+func TestTaskEvaluationUnavailableResponseIsMarkedCriticalUnknown(t *testing.T) {
+	selected := gatewaySelection()
+	registry := &taskEvaluationFailoverRegistry{primary: selected, fallback: selected}
+	provider := &stubAdvisoryProvider{response: cognitive.AdvisoryResponse{Version: cognitive.CreatorAdvisoryV1, FinishReason: cognitive.FinishTimeout}}
+	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-unknown", Revision: 1, CapabilityID: selected.Binding.CapabilityID, Stage: capability.TaskEvaluationGeneral, BenchmarkVersion: "unknown-v1", RequiredCaseIDs: []string{"case-a"}, QualityWeightMillis: 600, ReliabilityWeightMillis: 300, LatencyWeightMillis: 100}
+	request := capability.TaskEvaluationExecutionRequest{Profile: profile, CapabilityID: profile.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"}}}
+	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) != 1 || response.Observations[0].Result != capability.EvaluationUnknown || !response.Observations[0].CriticalFailure {
+		t.Fatalf("unavailable response was overclaimed: %+v", response.Observations)
+	}
+}

cmd/core-runtime/cognitive_interpretation.go

@@ -61,13 +61,6 @@ func (c cognitiveInterpretationClient) Interpret(ctx context.Context, request co
 }
 
 func shouldRequestBroaderInterpretation(result interactiondomain.SemanticResult) bool {
-	// The built-in advisory channel is the default conversational path. A
-	// simple question must not first wait for the slower, strict JSON
-	// interpretation route: that route exists for ambiguity and work routing,
-	// not as a prerequisite for talking to Domovoy.
-	if isSimpleLocalQuestion(result) {
-		return false
-	}
 	switch result.Interpretation.Applicability {
 	case interaction.SemanticNeedsBroaderInterpret, interaction.SemanticLowConfidence, interaction.SemanticMixedIntent:
 		return true

cmd/core-runtime/cognitive_interpretation_test.go

@@ -7,7 +7,7 @@ import (
 	interactiondomain "git.cin.su/m/Assistant/modules/interaction/domain"
 )
 
-func TestSimpleQuestionUsesBuiltInAdvisoryWithoutStrictInterpretation(t *testing.T) {
+func TestSimpleQuestionUsesStrictInterpretationForSemanticRouting(t *testing.T) {
 	result := interactiondomain.SemanticResult{
 		Interpretation: interaction.UtteranceInterpretation{
 			SpeechAct:     interaction.SpeechActQuestion,
@@ -16,8 +16,8 @@ func TestSimpleQuestionUsesBuiltInAdvisoryWithoutStrictInterpretation(t *testing
 		},
 		CapabilityHints: []string{"COGNITIVE_REASONING", "VERIFY_FACT"},
 	}
-	if shouldRequestBroaderInterpretation(result) {
-		t.Fatal("simple question must go directly to the built-in advisory path")
+	if !shouldRequestBroaderInterpretation(result) {
+		t.Fatal("simple question must reach semantic interpretation before any answer")
 	}
 }
 

cmd/core-runtime/development_gate.go

@@ -0,0 +1,129 @@
+package main
+
+import (
+	"context"
+	"errors"
+	"strings"
+	"time"
+
+	"git.cin.su/m/Assistant/adapters/developmentexecutor"
+	"git.cin.su/m/Assistant/contracts/work"
+	workdomain "git.cin.su/m/Assistant/modules/work/domain"
+	"git.cin.su/m/Assistant/modules/work/ports"
+)
+
+var errDevelopmentGateCreate = errors.New("development gate must start at evidence collected with revision one")
+var errDevelopmentGateNotIsolated = errors.New("development gate is not isolated")
+var errDevelopmentVerifierNotIndependent = errors.New("development verifier is not independent")
+var errDevelopmentWorktreeMismatch = errors.New("verification worktree does not match gate")
+var errDevelopmentChecksFailed = errors.New("development verification checks failed")
+var errDevelopmentEvidenceRequired = errors.New("development transition evidence is required")
+var errDevelopmentEvidenceStage = errors.New("development transition evidence does not match stage")
+
+type developmentGateVerificationRequest struct {
+	GateID           string
+	ExpectedRevision int64
+	VerifierRunID    work.AgentRunID
+	Plan             developmentexecutor.Plan
+}
+
+type developmentGateEvidenceRequest struct {
+	GateID           string
+	ExpectedRevision int64
+	NextStage        work.DevelopmentStage
+	EvidenceRef      string
+}
+
+func recordDevelopmentGateEvidence(ctx context.Context, store ports.DevelopmentGateStore, request developmentGateEvidenceRequest, now time.Time) (work.DevelopmentGate, error) {
+	gate, err := store.LoadDevelopmentGate(ctx, request.GateID)
+	if err != nil {
+		return work.DevelopmentGate{}, err
+	}
+	if request.ExpectedRevision < 1 || gate.Revision != request.ExpectedRevision {
+		return gate, errDevelopmentGateNotIsolated
+	}
+	if strings.TrimSpace(request.EvidenceRef) == "" {
+		return gate, errDevelopmentEvidenceRequired
+	}
+	switch request.NextStage {
+	case work.DevelopmentCanary:
+		gate.CanaryEvidence = strings.TrimSpace(request.EvidenceRef)
+	case work.DevelopmentPromoted:
+		gate.PromotionRef = strings.TrimSpace(request.EvidenceRef)
+	case work.DevelopmentRolledBack:
+		gate.RollbackRef = strings.TrimSpace(request.EvidenceRef)
+	default:
+		return gate, errDevelopmentEvidenceStage
+	}
+	advanced, err := advanceDevelopmentGate(gate, request.NextStage, now)
+	if err != nil {
+		return gate, err
+	}
+	if err := store.SaveDevelopmentGate(ctx, advanced, request.ExpectedRevision); err != nil {
+		return gate, err
+	}
+	return advanced, nil
+}
+
+func verifyDevelopmentGate(ctx context.Context, store ports.DevelopmentGateStore, request developmentGateVerificationRequest, now time.Time, runner developmentexecutor.CommandRunner) (work.DevelopmentGate, developmentexecutor.Result, error) {
+	gate, err := store.LoadDevelopmentGate(ctx, request.GateID)
+	if err != nil {
+		return work.DevelopmentGate{}, developmentexecutor.Result{}, err
+	}
+	if request.ExpectedRevision < 1 || gate.Revision != request.ExpectedRevision || gate.Stage != work.DevelopmentIsolated {
+		return gate, developmentexecutor.Result{}, errDevelopmentGateNotIsolated
+	}
+	if request.VerifierRunID == "" || sameRun(request.VerifierRunID, gate.ProducerRunRefs) {
+		return gate, developmentexecutor.Result{}, errDevelopmentVerifierNotIndependent
+	}
+	if request.Plan.WorktreeID != gate.WorktreeRef {
+		return gate, developmentexecutor.Result{}, errDevelopmentWorktreeMismatch
+	}
+	result, err := developmentexecutor.Execute(ctx, request.Plan, runner)
+	if err != nil {
+		return gate, result, err
+	}
+	if !result.Passed {
+		return gate, result, errDevelopmentChecksFailed
+	}
+	gate.VerifierRunRefs = append(append([]work.AgentRunID(nil), gate.VerifierRunRefs...), request.VerifierRunID)
+	gate.VerificationEvidence = result.EvidenceRef
+	advanced, err := advanceDevelopmentGate(gate, work.DevelopmentVerified, now)
+	if err != nil {
+		return gate, result, err
+	}
+	if err := store.SaveDevelopmentGate(ctx, advanced, request.ExpectedRevision); err != nil {
+		return gate, result, err
+	}
+	return advanced, result, nil
+}
+
+func sameRun(value work.AgentRunID, runs []work.AgentRunID) bool {
+	for _, run := range runs {
+		if run == value {
+			return true
+		}
+	}
+	return false
+}
+
+func prepareDevelopmentGate(value work.DevelopmentGate, now time.Time) (work.DevelopmentGate, error) {
+	if value.Stage != work.DevelopmentEvidenceCollected || value.Revision != 1 {
+		return work.DevelopmentGate{}, errDevelopmentGateCreate
+	}
+	value.UpdatedAtUnixMillis = now.UnixMilli()
+	if err := workdomain.ValidateDevelopmentGate(value); err != nil {
+		return work.DevelopmentGate{}, err
+	}
+	return value, nil
+}
+
+func advanceDevelopmentGate(value work.DevelopmentGate, next work.DevelopmentStage, now time.Time) (work.DevelopmentGate, error) {
+	advanced, err := workdomain.AdvanceDevelopmentGate(value, next)
+	if err != nil {
+		return work.DevelopmentGate{}, err
+	}
+	advanced.Revision = value.Revision + 1
+	advanced.UpdatedAtUnixMillis = now.UnixMilli()
+	return advanced, nil
+}

cmd/core-runtime/development_gate_test.go

@@ -0,0 +1,182 @@
+package main
+
+import (
+	"context"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"git.cin.su/m/Assistant/adapters/developmentexecutor"
+	"git.cin.su/m/Assistant/contracts/work"
+)
+
+type developmentGateStoreStub struct {
+	gate     work.DevelopmentGate
+	saved    bool
+	expected int64
+}
+
+func (s *developmentGateStoreStub) LoadDevelopmentGate(context.Context, string) (work.DevelopmentGate, error) {
+	return s.gate, nil
+}
+
+func (s *developmentGateStoreStub) SaveDevelopmentGate(_ context.Context, gate work.DevelopmentGate, expected int64) error {
+	s.gate = gate
+	s.saved = true
+	s.expected = expected
+	return nil
+}
+
+type developmentCommandRunnerStub struct {
+	exitCodes []int
+	commands  int
+}
+
+func (r *developmentCommandRunnerStub) Run(context.Context, string, string, []string) int {
+	code := 0
+	if r.commands < len(r.exitCodes) {
+		code = r.exitCodes[r.commands]
+	}
+	r.commands++
+	return code
+}
+
+func validDevelopmentVerificationPlan(t *testing.T, worktreeID string) developmentexecutor.Plan {
+	t.Helper()
+	worktree, err := filepath.Abs(".")
+	if err != nil {
+		t.Fatal(err)
+	}
+	return developmentexecutor.Plan{
+		Worktree:            worktree,
+		WorktreeID:          worktreeID,
+		BaseRepoRevision:    "base-revision",
+		ProtectedMainAccess: "DENIED",
+		ProductionAccess:    "DENIED",
+		SecretAccess:        "DENIED",
+		Checks:              []developmentexecutor.Check{{Kind: developmentexecutor.CheckGitDiffCheck}},
+	}
+}
+
+func isolatedDevelopmentGate() work.DevelopmentGate {
+	return work.DevelopmentGate{
+		ID:                   "gate-1",
+		Revision:             4,
+		Stage:                work.DevelopmentIsolated,
+		EvidenceRefs:         []string{"evidence-1"},
+		ProposalRef:          "proposal-1",
+		TaskPacketRef:        "task-1",
+		WorktreeRef:          "sandbox-1",
+		ProducerRunRefs:      []work.AgentRunID{"producer-1"},
+		VerifierRunRefs:      nil,
+		VerificationEvidence: "",
+	}
+}
+
+func TestVerifyDevelopmentGatePersistsVerifiedWithIndependentEvidence(t *testing.T) {
+	store := &developmentGateStoreStub{gate: isolatedDevelopmentGate()}
+	runner := &developmentCommandRunnerStub{exitCodes: []int{0}}
+	gate, result, err := verifyDevelopmentGate(context.Background(), store, developmentGateVerificationRequest{
+		GateID:           "gate-1",
+		ExpectedRevision: 4,
+		VerifierRunID:    "verifier-1",
+		Plan:             validDevelopmentVerificationPlan(t, "sandbox-1"),
+	}, time.UnixMilli(20), runner)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !result.Passed || gate.Stage != work.DevelopmentVerified || gate.Revision != 5 {
+		t.Fatalf("verification did not advance gate: gate=%#v result=%#v", gate, result)
+	}
+	if !store.saved || store.expected != 4 || gate.VerificationEvidence == "" {
+		t.Fatalf("verification was not durably persisted: store=%#v gate=%#v", store, gate)
+	}
+}
+
+func TestVerifyDevelopmentGateLeavesIsolatedOnFailedChecks(t *testing.T) {
+	store := &developmentGateStoreStub{gate: isolatedDevelopmentGate()}
+	gate, result, err := verifyDevelopmentGate(context.Background(), store, developmentGateVerificationRequest{
+		GateID:           "gate-1",
+		ExpectedRevision: 4,
+		VerifierRunID:    "verifier-1",
+		Plan:             validDevelopmentVerificationPlan(t, "sandbox-1"),
+	}, time.UnixMilli(20), &developmentCommandRunnerStub{exitCodes: []int{1}})
+	if err != errDevelopmentChecksFailed || result.Passed || gate.Stage != work.DevelopmentIsolated || store.saved {
+		t.Fatalf("failed checks changed gate unexpectedly: gate=%#v result=%#v err=%v store=%#v", gate, result, err, store)
+	}
+}
+
+func TestVerifyDevelopmentGateRejectsProducerAsVerifier(t *testing.T) {
+	store := &developmentGateStoreStub{gate: isolatedDevelopmentGate()}
+	_, _, err := verifyDevelopmentGate(context.Background(), store, developmentGateVerificationRequest{
+		GateID:           "gate-1",
+		ExpectedRevision: 4,
+		VerifierRunID:    "producer-1",
+		Plan:             validDevelopmentVerificationPlan(t, "sandbox-1"),
+	}, time.UnixMilli(20), &developmentCommandRunnerStub{exitCodes: []int{0}})
+	if err != errDevelopmentVerifierNotIndependent {
+		t.Fatalf("producer was accepted as verifier: %v", err)
+	}
+}
+
+func TestRecordDevelopmentGateEvidenceAdvancesCanaryAndPromotion(t *testing.T) {
+	store := &developmentGateStoreStub{gate: work.DevelopmentGate{
+		ID: "gate-1", Revision: 6, Stage: work.DevelopmentVerified,
+		EvidenceRefs: []string{"evidence-1"}, ProposalRef: "proposal-1", TaskPacketRef: "task-1", WorktreeRef: "sandbox-1",
+		ProducerRunRefs: []work.AgentRunID{"producer-1"}, VerifierRunRefs: []work.AgentRunID{"verifier-1"}, VerificationEvidence: "verification-1",
+	}}
+	gate, err := recordDevelopmentGateEvidence(context.Background(), store, developmentGateEvidenceRequest{GateID: "gate-1", ExpectedRevision: 6, NextStage: work.DevelopmentCanary, EvidenceRef: "canary-1"}, time.UnixMilli(30))
+	if err != nil || gate.Stage != work.DevelopmentCanary || gate.Revision != 7 || gate.CanaryEvidence != "canary-1" {
+		t.Fatalf("canary transition failed: gate=%#v err=%v", gate, err)
+	}
+	gate, err = recordDevelopmentGateEvidence(context.Background(), store, developmentGateEvidenceRequest{GateID: "gate-1", ExpectedRevision: 7, NextStage: work.DevelopmentPromoted, EvidenceRef: "promotion-1"}, time.UnixMilli(31))
+	if err != nil || gate.Stage != work.DevelopmentPromoted || gate.Revision != 8 || gate.PromotionRef != "promotion-1" {
+		t.Fatalf("promotion transition failed: gate=%#v err=%v", gate, err)
+	}
+}
+
+func TestRecordDevelopmentGateEvidenceSupportsCanaryRollbackAndRejectsMissingEvidence(t *testing.T) {
+	store := &developmentGateStoreStub{gate: work.DevelopmentGate{ID: "gate-1", Revision: 2, Stage: work.DevelopmentCanary, EvidenceRefs: []string{"evidence-1"}, CanaryEvidence: "canary-1"}}
+	gate, err := recordDevelopmentGateEvidence(context.Background(), store, developmentGateEvidenceRequest{GateID: "gate-1", ExpectedRevision: 2, NextStage: work.DevelopmentRolledBack, EvidenceRef: "rollback-1"}, time.UnixMilli(40))
+	if err != nil || gate.Stage != work.DevelopmentRolledBack || gate.RollbackRef != "rollback-1" {
+		t.Fatalf("rollback transition failed: gate=%#v err=%v", gate, err)
+	}
+	store = &developmentGateStoreStub{gate: work.DevelopmentGate{ID: "gate-2", Revision: 2, Stage: work.DevelopmentCanary, EvidenceRefs: []string{"evidence-1"}, CanaryEvidence: "canary-1"}}
+	gate, err = recordDevelopmentGateEvidence(context.Background(), store, developmentGateEvidenceRequest{GateID: "gate-2", ExpectedRevision: 2, NextStage: work.DevelopmentRolledBack}, time.UnixMilli(40))
+	if err != errDevelopmentEvidenceRequired || gate.Stage != work.DevelopmentCanary || store.saved {
+		t.Fatalf("missing rollback evidence changed gate: gate=%#v err=%v store=%#v", gate, err, store)
+	}
+}
+
+func TestDevelopmentGateControllerRequiresSequentialVerifiedEvidence(t *testing.T) {
+	now := time.UnixMilli(10)
+	gate, err := prepareDevelopmentGate(work.DevelopmentGate{ID: "gate-1", Revision: 1, Stage: work.DevelopmentEvidenceCollected, EvidenceRefs: []string{"evidence-1"}}, now)
+	if err != nil {
+		t.Fatal(err)
+	}
+	gate.ProposalRef = "proposal-1"
+	gate, err = advanceDevelopmentGate(gate, work.DevelopmentProposalCreated, now)
+	if err != nil || gate.Revision != 2 {
+		t.Fatalf("proposal transition failed: %#v %v", gate, err)
+	}
+	gate.TaskPacketRef = "task-1"
+	gate, err = advanceDevelopmentGate(gate, work.DevelopmentTaskAdmitted, now)
+	if err != nil {
+		t.Fatal(err)
+	}
+	gate.WorktreeRef = "isolated-worktree"
+	gate, err = advanceDevelopmentGate(gate, work.DevelopmentIsolated, now)
+	if err != nil {
+		t.Fatal(err)
+	}
+	gate.ProducerRunRefs = []work.AgentRunID{"producer"}
+	gate.VerifierRunRefs = []work.AgentRunID{"verifier"}
+	gate.VerificationEvidence = "verification-1"
+	gate, err = advanceDevelopmentGate(gate, work.DevelopmentVerified, now)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if _, err := advanceDevelopmentGate(gate, work.DevelopmentPromoted, now); err == nil {
+		t.Fatal("promotion bypassed canary")
+	}
+}

cmd/core-runtime/main.go

@@ -5,6 +5,7 @@ import (
 	"encoding/json"
 	"errors"
 	"fmt"
+	"git.cin.su/m/Assistant/adapters/developmentexecutor"
 	"git.cin.su/m/Assistant/adapters/postgres"
 	"git.cin.su/m/Assistant/contracts/authority"
 	"git.cin.su/m/Assistant/contracts/capability"
@@ -21,6 +22,7 @@ import (
 	interactiondomain "git.cin.su/m/Assistant/modules/interaction/domain"
 	situationdomain "git.cin.su/m/Assistant/modules/situation/domain"
 	workdomain "git.cin.su/m/Assistant/modules/work/domain"
+	"git.cin.su/m/Assistant/platform/ids"
 	platformsecrets "git.cin.su/m/Assistant/platform/secrets"
 	"git.cin.su/m/Assistant/runtime/health"
 	"github.com/jackc/pgx/v5"
@@ -116,6 +118,10 @@ func main() {
 	if err != nil {
 		log.Fatal(err)
 	}
+	researchClient, err := newResearchRuntimeClient(envOr("RESEARCH_RUNTIME_URL", "http://research-runtime:8087"), &http.Client{Timeout: 12 * time.Second})
+	if err != nil {
+		log.Fatal(err)
+	}
 	learning := learningRecorder{epistemics: epistemicsRepository, capability: capabilityRepository}
 	mux := http.NewServeMux()
 	authenticateCreator := func(w http.ResponseWriter, r *http.Request) (identity.CreatorSession, bool) {
@@ -178,6 +184,185 @@ func main() {
 		w.Header().Set("Content-Type", "application/json")
 		_ = json.NewEncoder(w).Encode(session)
 	})
+	mux.HandleFunc("POST /v1/creator/research/web", func(w http.ResponseWriter, r *http.Request) {
+		if _, ok := authenticateCreator(w, r); !ok {
+			return
+		}
+		var request epistemics.WebResearchRequest
+		decoder := json.NewDecoder(http.MaxBytesReader(w, r.Body, 128<<10))
+		decoder.DisallowUnknownFields()
+		if err := decoder.Decode(&request); err != nil {
+			http.Error(w, "invalid web research request", http.StatusBadRequest)
+			return
+		}
+		result, err := researchClient.EvaluateWeb(r.Context(), request)
+		if err != nil {
+			http.Error(w, "web research unavailable", http.StatusBadGateway)
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_ = json.NewEncoder(w).Encode(result)
+	})
+	mux.HandleFunc("POST /v1/creator/development/proposals", func(w http.ResponseWriter, r *http.Request) {
+		session, ok := authenticateCreator(w, r)
+		if !ok {
+			return
+		}
+		var request struct {
+			Evidence []work.ExtensionEvidence      `json:"evidence"`
+			Proposal work.ExtensionProposalRequest `json:"proposal"`
+		}
+		decoder := json.NewDecoder(http.MaxBytesReader(w, r.Body, 64<<10))
+		decoder.DisallowUnknownFields()
+		if err := decoder.Decode(&request); err != nil {
+			http.Error(w, "invalid development proposal", http.StatusBadRequest)
+			return
+		}
+		proposal, err := workdomain.BuildExtensionProposal(request.Evidence, request.Proposal)
+		if err != nil {
+			http.Error(w, "development proposal rejected", http.StatusUnprocessableEntity)
+			return
+		}
+		proposal.RequestedByPrincipalID = string(session.PrincipalID)
+		if err := workRepository.SaveExtensionProposal(r.Context(), proposal); err != nil {
+			http.Error(w, "development proposal conflict", http.StatusConflict)
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusCreated)
+		_ = json.NewEncoder(w).Encode(proposal)
+	})
+	mux.HandleFunc("POST /v1/creator/development/scaffolds", func(w http.ResponseWriter, r *http.Request) {
+		session, ok := authenticateCreator(w, r)
+		if !ok {
+			return
+		}
+		var request work.DevelopmentScaffoldRequest
+		decoder := json.NewDecoder(http.MaxBytesReader(w, r.Body, 64<<10))
+		decoder.DisallowUnknownFields()
+		if err := decoder.Decode(&request); err != nil {
+			http.Error(w, "invalid development scaffold", http.StatusBadRequest)
+			return
+		}
+		scaffold, err := workdomain.BuildDevelopmentScaffold(request)
+		if err != nil {
+			http.Error(w, "development scaffold rejected", http.StatusUnprocessableEntity)
+			return
+		}
+		scaffold.RequestedByPrincipalID = string(session.PrincipalID)
+		if err := workRepository.SaveDevelopmentScaffold(r.Context(), scaffold); err != nil {
+			http.Error(w, "development scaffold conflict", http.StatusConflict)
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusCreated)
+		_ = json.NewEncoder(w).Encode(scaffold)
+	})
+	mux.HandleFunc("POST /v1/creator/development/gates", func(w http.ResponseWriter, r *http.Request) {
+		if _, ok := authenticateCreator(w, r); !ok {
+			return
+		}
+		var gate work.DevelopmentGate
+		decoder := json.NewDecoder(http.MaxBytesReader(w, r.Body, 64<<10))
+		decoder.DisallowUnknownFields()
+		if err := decoder.Decode(&gate); err != nil {
+			http.Error(w, "invalid development gate", http.StatusBadRequest)
+			return
+		}
+		gate, err := prepareDevelopmentGate(gate, time.Now())
+		if err != nil {
+			http.Error(w, "development gate rejected", http.StatusUnprocessableEntity)
+			return
+		}
+		if err := workRepository.SaveDevelopmentGate(r.Context(), gate, 0); err != nil {
+			http.Error(w, "development gate conflict", http.StatusConflict)
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusCreated)
+		_ = json.NewEncoder(w).Encode(gate)
+	})
+	mux.HandleFunc("POST /v1/creator/development/gates/{id}/advance", func(w http.ResponseWriter, r *http.Request) {
+		if _, ok := authenticateCreator(w, r); !ok {
+			return
+		}
+		var request struct {
+			ExpectedRevision int64                 `json:"expected_revision"`
+			NextStage        work.DevelopmentStage `json:"next_stage"`
+		}
+		decoder := json.NewDecoder(http.MaxBytesReader(w, r.Body, 16<<10))
+		decoder.DisallowUnknownFields()
+		if err := decoder.Decode(&request); err != nil || request.ExpectedRevision < 1 || request.NextStage == "" {
+			http.Error(w, "invalid development gate transition", http.StatusBadRequest)
+			return
+		}
+		gate, err := workRepository.LoadDevelopmentGate(r.Context(), r.PathValue("id"))
+		if err != nil || gate.Revision != request.ExpectedRevision {
+			http.Error(w, "development gate revision conflict", http.StatusConflict)
+			return
+		}
+		gate, err = advanceDevelopmentGate(gate, request.NextStage, time.Now())
+		if err != nil {
+			http.Error(w, "development gate transition rejected", http.StatusUnprocessableEntity)
+			return
+		}
+		if err := workRepository.SaveDevelopmentGate(r.Context(), gate, request.ExpectedRevision); err != nil {
+			http.Error(w, "development gate conflict", http.StatusConflict)
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_ = json.NewEncoder(w).Encode(gate)
+	})
+	mux.HandleFunc("POST /v1/creator/development/gates/{id}/verify", func(w http.ResponseWriter, r *http.Request) {
+		if _, ok := authenticateCreator(w, r); !ok {
+			return
+		}
+		var request struct {
+			ExpectedRevision int64                    `json:"expected_revision"`
+			VerifierRunID    work.AgentRunID          `json:"verifier_run_id"`
+			Plan             developmentexecutor.Plan `json:"plan"`
+		}
+		decoder := json.NewDecoder(http.MaxBytesReader(w, r.Body, 64<<10))
+		decoder.DisallowUnknownFields()
+		if err := decoder.Decode(&request); err != nil || request.ExpectedRevision < 1 || request.VerifierRunID == "" {
+			http.Error(w, "invalid development verification request", http.StatusBadRequest)
+			return
+		}
+		gate, result, err := verifyDevelopmentGate(r.Context(), workRepository, developmentGateVerificationRequest{GateID: r.PathValue("id"), ExpectedRevision: request.ExpectedRevision, VerifierRunID: request.VerifierRunID, Plan: request.Plan}, time.Now(), developmentexecutor.OSCommandRunner{})
+		if err != nil {
+			http.Error(w, err.Error(), http.StatusUnprocessableEntity)
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_ = json.NewEncoder(w).Encode(map[string]any{"gate": gate, "verification": result})
+	})
+	registerDevelopmentGateEvidenceRoute := func(path string, next work.DevelopmentStage, fieldName string) {
+		mux.HandleFunc("POST /v1/creator/development/gates/{id}/"+path, func(w http.ResponseWriter, r *http.Request) {
+			if _, ok := authenticateCreator(w, r); !ok {
+				return
+			}
+			var request struct {
+				ExpectedRevision int64  `json:"expected_revision"`
+				EvidenceRef      string `json:"evidence_ref"`
+			}
+			decoder := json.NewDecoder(http.MaxBytesReader(w, r.Body, 16<<10))
+			decoder.DisallowUnknownFields()
+			if err := decoder.Decode(&request); err != nil || request.ExpectedRevision < 1 || request.EvidenceRef == "" {
+				http.Error(w, "invalid development transition evidence", http.StatusBadRequest)
+				return
+			}
+			gate, err := recordDevelopmentGateEvidence(r.Context(), workRepository, developmentGateEvidenceRequest{GateID: r.PathValue("id"), ExpectedRevision: request.ExpectedRevision, NextStage: next, EvidenceRef: request.EvidenceRef}, time.Now())
+			if err != nil {
+				http.Error(w, "development "+fieldName+" transition rejected", http.StatusUnprocessableEntity)
+				return
+			}
+			w.Header().Set("Content-Type", "application/json")
+			_ = json.NewEncoder(w).Encode(gate)
+		})
+	}
+	registerDevelopmentGateEvidenceRoute("canary", work.DevelopmentCanary, "canary")
+	registerDevelopmentGateEvidenceRoute("promote", work.DevelopmentPromoted, "promotion")
+	registerDevelopmentGateEvidenceRoute("rollback", work.DevelopmentRolledBack, "rollback")
 	mux.HandleFunc("GET /v1/creator/focus", func(w http.ResponseWriter, r *http.Request) {
 		session, ok := authenticateCreator(w, r)
 		if !ok {
@@ -437,6 +622,17 @@ func main() {
 			planningReply = humanCognitiveChannelReply(semantic.InformationRequestKind, semantic.Interpretation.SpeechAct, health, healthErr)
 			needsCognition = false
 		}
+		researchBacked := false
+		if hasCapabilityHint(semantic.CapabilityHints, "WEB_RESEARCH") && semantic.Interpretation.Clarification != interaction.ClarificationRequired {
+			if planID, idErr := ids.New(); idErr == nil {
+				result, researchErr := researchClient.SearchWeb(r.Context(), epistemics.WebSearchRequest{
+					Plan:           epistemics.ResearchPlan{ID: epistemics.ResearchPlanID("plan:" + planID), Question: utterance.Content, SourceClasses: []string{"web-search"}, FreshnessPolicy: "current"},
+					PropositionKey: "question:" + string(utterance.ID), Query: utterance.Content,
+				})
+				planningReply, researchBacked = renderResearchReply(result, researchErr)
+				needsCognition = false
+			}
+		}
 		if helpPath.Reply != "" {
 			planningReply = helpPath.Reply
 			needsCognition = false
@@ -534,6 +730,7 @@ func main() {
 			Clarification:                  semantic.Interpretation.Clarification,
 			WorkPlanned:                    workPlanned,
 			PlanningReply:                  planningReply,
+			ResearchBacked:                 researchBacked,
 			CapabilityStatus:               capabilityStatus,
 			CapabilityKind:                 capabilityKind,
 			CapabilityHints:                toCapabilityKinds(semantic.CapabilityHints),
@@ -834,6 +1031,20 @@ 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))
@@ -911,6 +1122,20 @@ func main() {
 		}
 		w.WriteHeader(http.StatusCreated)
 	})
+	mux.HandleFunc("POST /v1/cognitive-registry/model-qualification-observations", func(w http.ResponseWriter, r *http.Request) {
+		var value capability.ModelQualificationObservation
+		decoder := json.NewDecoder(io.LimitReader(r.Body, 16<<10))
+		decoder.DisallowUnknownFields()
+		if err := decoder.Decode(&value); err != nil || value.Validate() != nil {
+			http.Error(w, "invalid model qualification observation", http.StatusBadRequest)
+			return
+		}
+		if err := capabilityRepository.SaveModelQualificationObservation(r.Context(), value); err != nil {
+			http.Error(w, "save model qualification observation", http.StatusInternalServerError)
+			return
+		}
+		w.WriteHeader(http.StatusCreated)
+	})
 	mux.HandleFunc("POST /v1/creator/cognitive-canary", func(w http.ResponseWriter, r *http.Request) {
 		session, ok := authenticateCreator(w, r)
 		if !ok {
@@ -1215,6 +1440,32 @@ func envInt(name string, fallback int) int {
 	return value
 }
 
+func hasCapabilityHint(hints []string, wanted string) bool {
+	for _, hint := range hints {
+		if hint == wanted {
+			return true
+		}
+	}
+	return false
+}
+
+func renderResearchReply(result epistemics.ResearchResult, researchErr error) (string, bool) {
+	if researchErr != nil {
+		return "Я не смог получить проверяемые данные из research-источника, поэтому не буду придумывать ответ.", false
+	}
+	lines := make([]string, 0, len(result.Findings))
+	for _, finding := range result.Findings {
+		if finding.State != epistemics.FindingSupported || strings.TrimSpace(finding.Excerpt.Text) == "" || strings.TrimSpace(finding.Source.Locator) == "" {
+			continue
+		}
+		lines = append(lines, "- "+finding.Excerpt.Text+"\n  Источник: "+finding.Source.Locator)
+	}
+	if len(lines) == 0 {
+		return "Я не получил достаточно согласованных данных из проверяемых источников, поэтому не буду выдавать предположение как факт.", false
+	}
+	return "Нашёл точные выдержки из подключённого источника. Я не заменяю их догадкой:\n" + strings.Join(lines, "\n"), true
+}
+
 func humanCognitiveModelInventory(candidates []capability.ResolvedCognitiveBinding, inventoryErr error) string {
 	if inventoryErr != nil {
 		return "Я не смогла прочитать реестр подключённых моделей. Встроенный канал Домового остаётся доступен; настройки можно проверить повторно после восстановления реестра. Ключи и пароли не выводятся в чат."

cmd/core-runtime/research_client.go

@@ -0,0 +1,78 @@
+package main
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"strings"
+
+	"git.cin.su/m/Assistant/contracts/epistemics"
+)
+
+type researchRuntimeClient struct {
+	baseURL string
+	client  *http.Client
+}
+
+func newResearchRuntimeClient(rawURL string, client *http.Client) (researchRuntimeClient, error) {
+	base := strings.TrimRight(strings.TrimSpace(rawURL), "/")
+	if base == "" {
+		return researchRuntimeClient{}, fmt.Errorf("research runtime URL is empty")
+	}
+	if client == nil {
+		client = http.DefaultClient
+	}
+	return researchRuntimeClient{baseURL: base, client: client}, nil
+}
+
+func (c researchRuntimeClient) EvaluateWeb(ctx context.Context, request epistemics.WebResearchRequest) (epistemics.ResearchResult, error) {
+	payload, err := json.Marshal(request)
+	if err != nil {
+		return epistemics.ResearchResult{}, err
+	}
+	httpRequest, err := http.NewRequestWithContext(ctx, http.MethodPost, c.baseURL+"/v1/research/web/evaluate", bytes.NewReader(payload))
+	if err != nil {
+		return epistemics.ResearchResult{}, err
+	}
+	httpRequest.Header.Set("Content-Type", "application/json")
+	response, err := c.client.Do(httpRequest)
+	if err != nil {
+		return epistemics.ResearchResult{}, err
+	}
+	defer response.Body.Close()
+	if response.StatusCode < 200 || response.StatusCode >= 300 {
+		return epistemics.ResearchResult{}, fmt.Errorf("research runtime status %d", response.StatusCode)
+	}
+	var result epistemics.ResearchResult
+	if err := json.NewDecoder(response.Body).Decode(&result); err != nil {
+		return epistemics.ResearchResult{}, err
+	}
+	return result, nil
+}
+
+func (c researchRuntimeClient) SearchWeb(ctx context.Context, request epistemics.WebSearchRequest) (epistemics.ResearchResult, error) {
+	payload, err := json.Marshal(request)
+	if err != nil {
+		return epistemics.ResearchResult{}, err
+	}
+	httpRequest, err := http.NewRequestWithContext(ctx, http.MethodPost, c.baseURL+"/v1/research/web/search", bytes.NewReader(payload))
+	if err != nil {
+		return epistemics.ResearchResult{}, err
+	}
+	httpRequest.Header.Set("Content-Type", "application/json")
+	response, err := c.client.Do(httpRequest)
+	if err != nil {
+		return epistemics.ResearchResult{}, err
+	}
+	defer response.Body.Close()
+	if response.StatusCode < 200 || response.StatusCode >= 300 {
+		return epistemics.ResearchResult{}, fmt.Errorf("research search status %d", response.StatusCode)
+	}
+	var result epistemics.ResearchResult
+	if err := json.NewDecoder(response.Body).Decode(&result); err != nil {
+		return epistemics.ResearchResult{}, err
+	}
+	return result, nil
+}

cmd/core-runtime/research_client_test.go

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

cmd/creator-web/chat_test.go

@@ -151,6 +151,44 @@ func TestChatMessageUsesCoreAndLocalOnlyCognitiveGateway(t *testing.T) {
 	}
 }
 
+func TestChatMessageDoesNotExposeUnverifiedQuestionAsFact(t *testing.T) {
+	core := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		switch {
+		case r.Method == http.MethodGet && r.URL.Path == "/v1/creator/sessions/current":
+			_ = json.NewEncoder(w).Encode(identity.CreatorSession{ActorID: "actor-1", PrincipalID: "principal-1"})
+		case r.Method == http.MethodPost && r.URL.Path == "/v1/creator/chat/messages":
+			w.WriteHeader(http.StatusCreated)
+			_ = json.NewEncoder(w).Encode(interaction.ChatMessageResponse{ConversationID: "conversation-1", UtteranceID: "utterance-1", SpeechAct: interaction.SpeechActQuestion, NeedsCognition: true, FallbackText: "небезопасная модельная догадка"})
+		default:
+			t.Fatalf("unexpected core route %s %s", r.Method, r.URL.Path)
+		}
+	}))
+	defer core.Close()
+	cognitiveRuntime := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		t.Fatalf("unverified question must not reach advisory: %s %s", r.Method, r.URL.Path)
+	}))
+	defer cognitiveRuntime.Close()
+	handler, err := newHandler(core.URL, "http://continuity", "http://commit", core.Client(), cognitiveRuntime.URL)
+	if err != nil {
+		t.Fatal(err)
+	}
+	request := httptest.NewRequest(http.MethodPost, "/api/chat/messages", strings.NewReader(`{"message":"Какая погода?"}`))
+	request.Header.Set("Authorization", "Bearer session-token")
+	request.Header.Set("Content-Type", "application/json")
+	recorder := httptest.NewRecorder()
+	handler.ServeHTTP(recorder, request)
+	if recorder.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d body=%s", recorder.Code, recorder.Body.String())
+	}
+	var response browserChatResponse
+	if err := json.NewDecoder(recorder.Body).Decode(&response); err != nil {
+		t.Fatal(err)
+	}
+	if response.Cognition != "verification_required" || strings.Contains(response.ReplyText, "небезопасная") || !strings.Contains(response.ReplyText, "не буду придумывать") {
+		t.Fatalf("unexpected truth-first response: %#v", response)
+	}
+}
+
 func TestChatMessageReturnsHumanCapabilityMetadata(t *testing.T) {
 	core := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		switch {

cmd/creator-web/responsive_test.go

@@ -0,0 +1,53 @@
+package main
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestCreatorConsoleResponsiveGuardrails(t *testing.T) {
+	html, err := staticFiles.ReadFile("static/index.html")
+	if err != nil {
+		t.Fatal(err)
+	}
+	if !strings.Contains(string(html), `<meta name="viewport" content="width=device-width, initial-scale=1">`) {
+		t.Fatal("creator console is missing responsive viewport metadata")
+	}
+	css, err := staticFiles.ReadFile("static/assets/console.css")
+	if err != nil {
+		t.Fatal(err)
+	}
+	stylesheet := string(css)
+	for _, guard := range []string{
+		"@media (max-width: 680px)",
+		".console { display: block; }",
+		".diagnostic-list { grid-template-columns: 1fr; }",
+		".agent-details { grid-template-columns: 1fr; gap: 14px; }",
+		".fleet-list { grid-template-columns: repeat(2, minmax(0, 1fr)); }",
+		".section-nav { display: flex;",
+		"overflow-x: auto",
+	} {
+		if !strings.Contains(stylesheet, guard) {
+			t.Fatalf("responsive guardrail missing: %s", guard)
+		}
+	}
+}
+
+func TestChatResponsiveGuardrailsConstrainMobileGridItem(t *testing.T) {
+	css, err := staticFiles.ReadFile("static/assets/chat.css")
+	if err != nil {
+		t.Fatal(err)
+	}
+	stylesheet := string(css)
+	for _, guard := range []string{
+		".chat-view { width: 100%; min-width: 0;",
+		"@media (max-width: 640px)",
+		".chat-layout { height: auto; display: block; }",
+		".chat-header > .operator-link, .build-meta { display: none; }",
+		".chat-header .cognitive-console-link { flex: 0 1 110px;",
+	} {
+		if !strings.Contains(stylesheet, guard) {
+			t.Fatalf("chat responsive guardrail missing: %s", guard)
+		}
+	}
+}

cmd/creator-web/server.go

@@ -150,7 +150,7 @@ func (g *gateway) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 		http.Error(w, "method not allowed", http.StatusMethodNotAllowed)
 		return
 	}
-	if r.URL.Path == "/" || r.URL.Path == "/chat" || r.URL.Path == "/chat.html" {
+	if r.URL.Path == "/" || r.URL.Path == "/login" || r.URL.Path == "/chat" || r.URL.Path == "/chat.html" {
 		g.serveEmbeddedHTML(w, "chat.html")
 		return
 	}
@@ -209,6 +209,7 @@ type browserChatResponse struct {
 	CapabilityKind                 string                      `json:"capability_kind"`
 	CapabilityHints                []capability.CapabilityKind `json:"capability_hints,omitempty"`
 	RequestedOutcome               string                      `json:"requested_outcome,omitempty"`
+	ResearchBacked                 bool                        `json:"research_backed"`
 	InterpretationConfidenceMillis int64                       `json:"interpretation_confidence_millis,omitempty"`
 	DelegationID                   string                      `json:"delegation_id"`
 	DelegationState                string                      `json:"delegation_state"`
@@ -467,7 +468,11 @@ func (g *gateway) handleChatMessage(w http.ResponseWriter, r *http.Request) {
 			saved.NeedsCognition = false
 		}
 	}
-	if saved.NeedsCognition && !saved.WorkPlanned && saved.Clarification != interaction.ClarificationRequired {
+	if saved.SpeechAct == interaction.SpeechActQuestion && saved.InformationRequestKind == "" && !saved.ResearchBacked && !saved.WorkPlanned && saved.Clarification != interaction.ClarificationRequired && !asksAttention(browserRequest.Message) {
+		reply = truthfulVerificationRequiredReply()
+		cognition = "verification_required"
+		saved.NeedsCognition = false
+	} else if saved.NeedsCognition && !saved.WorkPlanned && saved.Clarification != interaction.ClarificationRequired {
 		cognition = "unavailable"
 		requestID, idErr := ids.New()
 		if idErr == nil {
@@ -501,6 +506,7 @@ func (g *gateway) handleChatMessage(w http.ResponseWriter, r *http.Request) {
 		CapabilityKind:                 string(saved.CapabilityKind),
 		CapabilityHints:                saved.CapabilityHints,
 		RequestedOutcome:               saved.RequestedOutcome,
+		ResearchBacked:                 saved.ResearchBacked,
 		InterpretationConfidenceMillis: saved.InterpretationConfidenceMillis,
 		DelegationID:                   saved.DelegationID,
 		DelegationState:                saved.DelegationState,
@@ -508,6 +514,10 @@ func (g *gateway) handleChatMessage(w http.ResponseWriter, r *http.Request) {
 	})
 }
 
+func truthfulVerificationRequiredReply() string {
+	return "Я не могу сейчас подтвердить этот факт по проверенному источнику, поэтому не буду придумывать ответ. Для проверки используйте официальный сайт соответствующей службы или надёжный сервис с текущими данными; когда подключён research-маршрут, Домовой сможет проверить это сам."
+}
+
 func localChatContext(saved interaction.ChatMessageResponse) []cognitive.ContextExcerpt {
 	if strings.TrimSpace(saved.ActiveProjectName) == "" {
 		return []cognitive.ContextExcerpt{}

cmd/creator-web/server_test.go

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

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

@@ -29,7 +29,7 @@ textarea { resize: vertical; min-height: 84px; max-height: 240px; padding: 14px;
 .login-form button, .composer button { border: 0; color: #11150f; background: var(--signal); font-weight: 700; }
 .login-form button { width: 100%; height: 50px; margin-top: 18px; }
 #login-error { min-height: 20px; color: var(--bad); font-size: 13px; }
-.chat-view { height: 100vh; min-height: 0; display: grid; grid-template-rows: 64px minmax(0, 1fr); overflow: hidden; }
+.chat-view { width: 100%; min-width: 0; height: 100vh; min-height: 0; display: grid; grid-template-rows: 64px minmax(0, 1fr); overflow: hidden; }
 .chat-header { min-width: 0; display: flex; align-items: center; gap: 14px; padding: 0 22px; border-bottom: 1px solid var(--line); background: rgba(16, 23, 20, .98); }
 .chat-header a { color: var(--ink); text-decoration: none; font: 400 24px/1 var(--serif); }
 #active-focus { color: var(--good); }
@@ -107,9 +107,10 @@ textarea { resize: vertical; min-height: 84px; max-height: 240px; padding: 14px;
   .conversation-starters { align-items: stretch; }
   .conversation-starters span { width: 100%; }
   .chat-header { gap: 10px; padding: 0 12px; }
-  .chat-header .cognitive-status { display: inline-flex; max-width: 150px; overflow: hidden; text-overflow: ellipsis; }
-  .build-meta { font-size: 8px; letter-spacing: .04em; }
-  .build-meta time { font-size: 7px; }
+  .chat-header .cognitive-status { display: inline-flex; flex: 1 1 110px; min-width: 0; max-width: 110px; overflow: hidden; text-overflow: ellipsis; }
+  .chat-header .cognitive-console-link { flex: 0 1 110px; min-width: 0; overflow: hidden; text-overflow: ellipsis; }
+  .chat-header > .operator-link, .build-meta { display: none; }
+  .chat-header > #logout { flex: 0 0 38px; }
   .chat-layout { height: auto; display: block; }
   .chat-side, .chat-main { min-height: auto; max-height: none; overflow: visible; }
   .chat-left, .chat-right { border: 0; border-bottom: 1px solid var(--line); }

cmd/reflection-worker/main.go

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

cmd/reflection-worker/status.go

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

cmd/reflection-worker/status_test.go

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

cmd/research-runtime/main.go

@@ -34,6 +34,7 @@ func main() {
 	mux := http.NewServeMux()
 	mux.Handle("/healthz", health.Handler{Runtime: "research-runtime", SemanticGeneration: 1, ArchitectureGeneration: 1})
 	mux.Handle("/readyz", health.ReadinessHandler{Runtime: "research-runtime", Probes: []health.Probe{{Name: "epistemics-postgres", Check: func(ctx context.Context) error { return repository.Ping(ctx) }}}})
+	mux.Handle("/v1/research/status", truthRuntimeStatusHandler(repository))
 	mux.HandleFunc("POST /v1/research/evaluate", func(w http.ResponseWriter, r *http.Request) {
 		var request epistemics.ResearchEvaluationRequest
 		if err := json.NewDecoder(http.MaxBytesReader(w, r.Body, 2<<20)).Decode(&request); err != nil {
@@ -88,6 +89,26 @@ func main() {
 		w.Header().Set("Content-Type", "application/json")
 		_ = json.NewEncoder(w).Encode(result)
 	})
+	searchConnector := researchadapter.SearchConnector{EndpointTemplate: os.Getenv("RESEARCH_SEARCH_URL_TEMPLATE"), AllowedHosts: allowedHosts(os.Getenv("RESEARCH_SEARCH_ALLOWED_HOSTS")), Client: &http.Client{Timeout: 8 * time.Second}, MaxResults: 3}
+	mux.HandleFunc("POST /v1/research/web/search", func(w http.ResponseWriter, r *http.Request) {
+		var request epistemics.WebSearchRequest
+		if err := json.NewDecoder(http.MaxBytesReader(w, r.Body, 64<<10)).Decode(&request); err != nil {
+			http.Error(w, "invalid web search request", http.StatusBadRequest)
+			return
+		}
+		candidates, err := searchConnector.Collect(r.Context(), request, time.Now().UnixMilli())
+		if err != nil {
+			http.Error(w, err.Error(), http.StatusUnprocessableEntity)
+			return
+		}
+		result, err := application.EvaluateAndPersist(r.Context(), epistemics.ResearchEvaluationRequest{Plan: request.Plan, Candidates: candidates}, repository, time.Now().UnixMilli())
+		if err != nil {
+			http.Error(w, err.Error(), http.StatusUnprocessableEntity)
+			return
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_ = json.NewEncoder(w).Encode(result)
+	})
 	addr := os.Getenv("LISTEN_ADDR")
 	if addr == "" {
 		addr = ":8087"

cmd/research-runtime/status.go

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

cmd/research-runtime/status_test.go

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

cmd/task-verifier/main.go

@@ -0,0 +1,100 @@
+package main
+
+import (
+	"crypto/sha256"
+	"encoding/hex"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"os"
+	"strings"
+
+	"git.cin.su/m/Assistant/contracts/capability"
+	"git.cin.su/m/Assistant/contracts/work"
+)
+
+const verifierIdentity = "deterministic-task-verifier-v1"
+const genericVerifierIdentity = "deterministic-generic-envelope-verifier-v1"
+
+func main() {
+	mux := http.NewServeMux()
+	mux.HandleFunc("GET /healthz", func(w http.ResponseWriter, _ *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{"runtime":"task-verifier","status":"ok"}`))
+	})
+	mux.HandleFunc("POST /v1/task-evaluations/verify", verifyHandler)
+	mux.HandleFunc("POST /v1/work-graphs/candidates/verify", genericCandidateVerifyHandler)
+	addr := ":8091"
+	if value := strings.TrimSpace(os.Getenv("LISTEN_ADDR")); value != "" {
+		addr = value
+	}
+	if err := http.ListenAndServe(addr, mux); err != nil {
+		panic(err)
+	}
+}
+
+func genericCandidateVerifyHandler(w http.ResponseWriter, r *http.Request) {
+	var request work.GenericCandidateVerificationRequest
+	decoder := json.NewDecoder(http.MaxBytesReader(w, r.Body, 64<<10))
+	decoder.DisallowUnknownFields()
+	if err := decoder.Decode(&request); err != nil || request.GraphID == "" || request.GraphRevision < 1 || request.NodeID == "" || request.AuthorRunID == "" || request.CandidateID == "" || strings.TrimSpace(request.Summary) == "" || len(request.EvidenceRefs) == 0 {
+		http.Error(w, "invalid generic candidate envelope", http.StatusBadRequest)
+		return
+	}
+	payload := fmt.Sprintf("%s|%s|%d|%s|%s|%s|%s", genericVerifierIdentity, request.GraphID, request.GraphRevision, request.NodeID, request.AuthorRunID, request.CandidateID, strings.Join(evidenceRefStrings(request.EvidenceRefs), "|"))
+	digest := sha256.Sum256([]byte(payload))
+	writeJSON(w, work.GenericCandidateVerificationResponse{Verified: true, VerificationEvidenceRef: "verification:" + hex.EncodeToString(digest[:]), VerifierIdentity: genericVerifierIdentity, Summary: "candidate attribution and evidence envelope verified; semantic content remains unverified"})
+}
+
+func evidenceRefStrings(values []work.EvidenceRef) []string {
+	result := make([]string, len(values))
+	for i, value := range values {
+		result[i] = string(value)
+	}
+	return result
+}
+
+func verifyHandler(w http.ResponseWriter, r *http.Request) {
+	var request capability.TaskCaseVerificationRequest
+	decoder := json.NewDecoder(http.MaxBytesReader(w, r.Body, 64<<10))
+	decoder.DisallowUnknownFields()
+	if err := decoder.Decode(&request); err != nil || request.CaseID == "" || request.ResponseText == "" || request.EvidenceArtifactRef == "" || request.LatencyMicros < 0 || (request.Verification.Mode != capability.TaskCaseVerificationExactText && request.Verification.Mode != capability.TaskCaseVerificationJSONFieldEquals) || (request.Verification.Mode == capability.TaskCaseVerificationJSONFieldEquals && request.Verification.FieldPath == "") || request.Verification.ExpectedText == "" || request.Verification.MinimumQualityMillis < 0 || request.Verification.MinimumQualityMillis > 1000 {
+		http.Error(w, "invalid task verification request", http.StatusBadRequest)
+		return
+	}
+	matched := matches(request)
+	quality := int64(0)
+	result := capability.EvaluationUnsupported
+	if matched {
+		quality = request.Verification.MinimumQualityMillis
+		if quality == 0 {
+			quality = 1000
+		}
+		result = capability.EvaluationSupported
+	}
+	evidence := verificationEvidence(request, matched)
+	writeJSON(w, capability.TaskCaseVerificationResponse{CaseID: request.CaseID, Result: result, QualityScoreMillis: quality, LatencyMicros: request.LatencyMicros, EvidenceArtifactRef: request.EvidenceArtifactRef, VerificationEvidenceRef: evidence, VerifierIdentity: verifierIdentity})
+}
+
+func verificationEvidence(request capability.TaskCaseVerificationRequest, matched bool) string {
+	payload := fmt.Sprintf("%s|%s|%s|%s|%s|%t", verifierIdentity, request.CaseID, request.EvidenceArtifactRef, request.Verification.Mode, request.Verification.FieldPath, matched)
+	digest := sha256.Sum256([]byte(payload))
+	return "verification:" + hex.EncodeToString(digest[:])
+}
+
+func matches(request capability.TaskCaseVerificationRequest) bool {
+	if request.Verification.Mode == capability.TaskCaseVerificationExactText {
+		return strings.TrimSpace(request.ResponseText) == strings.TrimSpace(request.Verification.ExpectedText)
+	}
+	var document map[string]json.RawMessage
+	if json.Unmarshal([]byte(request.ResponseText), &document) != nil {
+		return false
+	}
+	value, ok := document[request.Verification.FieldPath]
+	return ok && strings.TrimSpace(string(value)) == strings.TrimSpace(request.Verification.ExpectedText)
+}
+
+func writeJSON(w http.ResponseWriter, value any) {
+	w.Header().Set("Content-Type", "application/json")
+	_ = json.NewEncoder(w).Encode(value)
+}

cmd/task-verifier/main_test.go

@@ -0,0 +1,53 @@
+package main
+
+import (
+	"bytes"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"git.cin.su/m/Assistant/contracts/capability"
+	"git.cin.su/m/Assistant/contracts/work"
+)
+
+func TestVerifyExactTextProducesIndependentEvidence(t *testing.T) {
+	payload, _ := json.Marshal(capability.TaskCaseVerificationRequest{CaseID: "case-1", ResponseText: " answer ", EvidenceArtifactRef: "case-artifact", Verification: capability.TaskCaseVerification{Mode: capability.TaskCaseVerificationExactText, ExpectedText: "answer", MinimumQualityMillis: 900}})
+	recorder := httptest.NewRecorder()
+	verifyHandler(recorder, httptest.NewRequest(http.MethodPost, "/v1/task-evaluations/verify", bytes.NewReader(payload)))
+	var response capability.TaskCaseVerificationResponse
+	if recorder.Code != http.StatusOK || json.NewDecoder(recorder.Body).Decode(&response) != nil || response.Result != capability.EvaluationSupported || response.QualityScoreMillis != 900 || response.VerifierIdentity == "" || response.VerificationEvidenceRef == "" {
+		t.Fatalf("unexpected verification response: code=%d body=%s value=%+v", recorder.Code, recorder.Body.String(), response)
+	}
+}
+
+func TestVerifyMismatchIsUnsupportedWithEvidence(t *testing.T) {
+	payload, _ := json.Marshal(capability.TaskCaseVerificationRequest{CaseID: "case-1", ResponseText: "wrong", EvidenceArtifactRef: "case-artifact", Verification: capability.TaskCaseVerification{Mode: capability.TaskCaseVerificationExactText, ExpectedText: "answer", MinimumQualityMillis: 900}})
+	recorder := httptest.NewRecorder()
+	verifyHandler(recorder, httptest.NewRequest(http.MethodPost, "/v1/task-evaluations/verify", bytes.NewReader(payload)))
+	var response capability.TaskCaseVerificationResponse
+	if recorder.Code != http.StatusOK || json.NewDecoder(recorder.Body).Decode(&response) != nil || response.Result != capability.EvaluationUnsupported || response.QualityScoreMillis != 0 || response.VerificationEvidenceRef == "" {
+		t.Fatalf("unexpected mismatch response: code=%d body=%s value=%+v", recorder.Code, recorder.Body.String(), response)
+	}
+}
+
+func TestVerifyJSONFieldEquals(t *testing.T) {
+	payload, _ := json.Marshal(capability.TaskCaseVerificationRequest{CaseID: "json-1", ResponseText: `{"answer":42}`, EvidenceArtifactRef: "case-artifact", Verification: capability.TaskCaseVerification{Mode: capability.TaskCaseVerificationJSONFieldEquals, FieldPath: "answer", ExpectedText: "42", MinimumQualityMillis: 850}})
+	recorder := httptest.NewRecorder()
+	verifyHandler(recorder, httptest.NewRequest(http.MethodPost, "/v1/task-evaluations/verify", bytes.NewReader(payload)))
+	var response capability.TaskCaseVerificationResponse
+	if recorder.Code != http.StatusOK || json.NewDecoder(recorder.Body).Decode(&response) != nil || response.Result != capability.EvaluationSupported || response.QualityScoreMillis != 850 {
+		t.Fatalf("unexpected JSON verification response: code=%d body=%s value=%+v", recorder.Code, recorder.Body.String(), response)
+	}
+}
+
+func TestVerifyGenericCandidateEnvelopeDoesNotClaimTruth(t *testing.T) {
+	payload, _ := json.Marshal(work.GenericCandidateVerificationRequest{GraphID: "graph-1", GraphRevision: 2, NodeID: "node-1", AuthorRunID: "run-1", CandidateID: "candidate-1", Summary: "model proposal", EvidenceRefs: []work.EvidenceRef{"agent-output:1"}})
+	recorder := httptest.NewRecorder()
+	genericCandidateVerifyHandler(recorder, httptest.NewRequest(http.MethodPost, "/v1/work-graphs/candidates/verify", bytes.NewReader(payload)))
+	var response work.GenericCandidateVerificationResponse
+	if recorder.Code != http.StatusOK || json.NewDecoder(recorder.Body).Decode(&response) != nil || !response.Verified || response.VerifierIdentity == "" || response.VerificationEvidenceRef == "" || !strings.Contains(response.Summary, "semantic content remains unverified") {
+		t.Fatalf("unexpected generic verification response: code=%d body=%s value=%+v", recorder.Code, recorder.Body.String(), response)
+	}
+}

cmd/work-graph-runtime/main.go

@@ -7,9 +7,11 @@ import (
 	"log"
 	"net/http"
 	"os"
+	"time"
 
 	"git.cin.su/m/Assistant/adapters/postgres"
 	secretcontracts "git.cin.su/m/Assistant/contracts/secrets"
+	"git.cin.su/m/Assistant/contracts/work"
 	"git.cin.su/m/Assistant/modules/work/application"
 	"git.cin.su/m/Assistant/platform/secrets"
 	"git.cin.su/m/Assistant/runtime/health"
@@ -26,9 +28,14 @@ func main() {
 		log.Fatal(err)
 	}
 	defer repository.Close()
+	evidenceSink, err := newCoreTaskEvaluationSink(envOr("CORE_RUNTIME_URL", "http://core-runtime:8080"), &http.Client{Timeout: 12 * time.Second})
+	if err != nil {
+		log.Fatal(err)
+	}
 	mux := http.NewServeMux()
 	mux.Handle("/healthz", health.Handler{Runtime: "work-graph-runtime", SemanticGeneration: 1, ArchitectureGeneration: 1})
 	mux.Handle("/readyz", health.ReadinessHandler{Runtime: "work-graph-runtime", Probes: []health.Probe{{Name: "work-postgres", Check: func(ctx context.Context) error { return repository.Ping(ctx) }}}})
+	mux.Handle("/v1/work-graphs/status", workGraphStatusHandler(repository))
 	mux.HandleFunc("POST /v1/work-graphs/schedule", func(w http.ResponseWriter, r *http.Request) {
 		var request application.GraphScheduleRequest
 		if err := json.NewDecoder(http.MaxBytesReader(w, r.Body, 4<<20)).Decode(&request); err != nil {
@@ -37,6 +44,7 @@ func main() {
 		}
 		graph, ready, err := application.AdmitAndSchedule(r.Context(), repository, request)
 		if err != nil {
+			log.Printf("work graph schedule rejected: %v", err)
 			http.Error(w, err.Error(), http.StatusUnprocessableEntity)
 			return
 		}
@@ -50,18 +58,85 @@ func main() {
 		}
 		graph, ready, err := application.AdmitAndScheduleTaskEvaluation(r.Context(), repository, request)
 		if err != nil {
+			log.Printf("task evaluation graph schedule rejected: %v", err)
 			http.Error(w, err.Error(), http.StatusUnprocessableEntity)
 			return
 		}
 		writeJSON(w, map[string]any{"graph": graph, "scheduled_nodes": ready})
 	})
+	mux.HandleFunc("POST /v1/work-graphs/task-evaluations/additional/schedule", func(w http.ResponseWriter, r *http.Request) {
+		var request application.TaskEvaluationExpansionScheduleRequest
+		if err := json.NewDecoder(http.MaxBytesReader(w, r.Body, 4<<20)).Decode(&request); err != nil {
+			http.Error(w, "invalid additional task evaluation request", http.StatusBadRequest)
+			return
+		}
+		graph, ready, err := application.ScheduleAdditionalTaskEvaluation(r.Context(), repository, request)
+		if err != nil {
+			log.Printf("additional task evaluation schedule rejected: %v", err)
+			http.Error(w, err.Error(), http.StatusUnprocessableEntity)
+			return
+		}
+		writeJSON(w, map[string]any{"graph": graph, "scheduled_nodes": ready})
+	})
+	mux.HandleFunc("POST /v1/work-graphs/orchestration/next", func(w http.ResponseWriter, r *http.Request) {
+		var request work.SolutionSearchInput
+		if err := json.NewDecoder(http.MaxBytesReader(w, r.Body, 32<<10)).Decode(&request); err != nil {
+			http.Error(w, "invalid orchestration input", http.StatusBadRequest)
+			return
+		}
+		decision, err := application.DecideSolutionSearch(request)
+		if err != nil {
+			http.Error(w, "invalid orchestration input", http.StatusUnprocessableEntity)
+			return
+		}
+		writeJSON(w, decision)
+	})
+	mux.HandleFunc("POST /v1/work-graphs/orchestration/wave", func(w http.ResponseWriter, r *http.Request) {
+		var request work.SolutionSearchWaveRequest
+		if err := json.NewDecoder(http.MaxBytesReader(w, r.Body, 1<<20)).Decode(&request); err != nil {
+			http.Error(w, "invalid orchestration wave request", http.StatusBadRequest)
+			return
+		}
+		result, err := application.OrchestrateSolutionSearch(r.Context(), repository, request)
+		if err != nil {
+			http.Error(w, "invalid orchestration wave request", http.StatusUnprocessableEntity)
+			return
+		}
+		writeJSON(w, result)
+	})
+	mux.HandleFunc("POST /v1/work-graphs/result-callback", func(w http.ResponseWriter, r *http.Request) {
+		var request work.GraphResultCallback
+		if err := json.NewDecoder(http.MaxBytesReader(w, r.Body, 1<<20)).Decode(&request); err != nil {
+			http.Error(w, "invalid graph result callback", http.StatusBadRequest)
+			return
+		}
+		graph, err := application.RecordGraphResultCandidate(r.Context(), repository, repository, request)
+		if err != nil {
+			http.Error(w, "invalid graph result callback", http.StatusUnprocessableEntity)
+			return
+		}
+		writeJSON(w, graph)
+	})
 	mux.HandleFunc("POST /v1/work-graphs/complete", func(w http.ResponseWriter, r *http.Request) {
 		var request application.GraphCompleteRequest
 		if err := json.NewDecoder(http.MaxBytesReader(w, r.Body, 64<<10)).Decode(&request); err != nil {
 			http.Error(w, "invalid graph completion request", http.StatusBadRequest)
 			return
 		}
-		graph, err := application.CompleteGraphNode(r.Context(), repository, request.GraphID, request.Revision, request.NodeID, request.Succeeded)
+		result, err := application.CompleteGraphNodeWithReplan(r.Context(), repository, request)
+		if err != nil {
+			http.Error(w, err.Error(), http.StatusUnprocessableEntity)
+			return
+		}
+		writeJSON(w, result)
+	})
+	mux.HandleFunc("POST /v1/work-graphs/task-evaluations/callback", func(w http.ResponseWriter, r *http.Request) {
+		var request work.TaskEvaluationEvidenceCallback
+		if err := json.NewDecoder(http.MaxBytesReader(w, r.Body, 1<<20)).Decode(&request); err != nil {
+			http.Error(w, "invalid task evaluation callback", http.StatusBadRequest)
+			return
+		}
+		graph, err := application.CompleteTaskEvaluationNode(r.Context(), repository, evidenceSink, request)
 		if err != nil {
 			http.Error(w, err.Error(), http.StatusUnprocessableEntity)
 			return
@@ -74,12 +149,12 @@ func main() {
 			http.Error(w, "invalid graph synthesis request", http.StatusBadRequest)
 			return
 		}
-		conflict, synthesis, err := application.SynthesizeGraph(r.Context(), repository, request)
+		conflict, synthesis, graph, err := application.SynthesizeAndCompleteGraph(r.Context(), repository, repository, request)
 		if err != nil {
 			http.Error(w, err.Error(), http.StatusUnprocessableEntity)
 			return
 		}
-		writeJSON(w, map[string]any{"conflict": conflict, "synthesis": synthesis})
+		writeJSON(w, map[string]any{"conflict": conflict, "synthesis": synthesis, "graph": graph})
 	})
 	addr := os.Getenv("LISTEN_ADDR")
 	if addr == "" {
@@ -89,6 +164,13 @@ func main() {
 	log.Fatal(http.ListenAndServe(addr, mux))
 }
 
+func envOr(name, fallback string) string {
+	if value := os.Getenv(name); value != "" {
+		return value
+	}
+	return fallback
+}
+
 func writeJSON(w http.ResponseWriter, value any) {
 	w.Header().Set("Content-Type", "application/json")
 	_ = json.NewEncoder(w).Encode(value)

cmd/work-graph-runtime/status.go

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

cmd/work-graph-runtime/status_test.go

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

cmd/work-graph-runtime/task_evaluation_sink.go

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

contracts/capability/cognitive_registry.go

@@ -172,6 +172,7 @@ type ResolvedCognitiveBinding struct {
 	Evaluation     *EvaluatedCapability
 	Performance    *PerformanceEvidence
 	TaskEvaluation *TaskEvaluation
+	Qualification  *ModelQualificationObservation
 	Rating         *ModelRating
 }
 

contracts/capability/model_qualification.go

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

contracts/capability/task_evaluation.go

@@ -70,6 +70,22 @@ type TaskEvaluationExpansionRequest struct {
 	BenchmarkVersion   string
 }
 
+type TaskCaseVerificationMode string
+
+const (
+	TaskCaseVerificationExactText       TaskCaseVerificationMode = "EXACT_TEXT"
+	TaskCaseVerificationJSONFieldEquals TaskCaseVerificationMode = "JSON_FIELD_EQUALS"
+)
+
+// TaskCaseVerification is a bounded, task-owned rubric. The expected text is
+// test data, not model output, and is never sent to the cognitive provider.
+type TaskCaseVerification struct {
+	Mode                 TaskCaseVerificationMode
+	FieldPath            string
+	ExpectedText         string
+	MinimumQualityMillis int64
+}
+
 // TaskCaseObservation is produced by the bounded evaluator and verifier. A
 // score without a case result and evidence reference is not accepted as
 // qualification evidence.
@@ -80,6 +96,10 @@ type TaskCaseObservation struct {
 	LatencyMicros       int64
 	CriticalFailure     bool
 	EvidenceArtifactRef string
+	// VerificationEvidenceRef and VerifierIdentity are required whenever an
+	// observation claims semantic support or a non-zero quality score.
+	VerificationEvidenceRef string
+	VerifierIdentity        string
 }
 
 // TaskEvaluation is durable evidence for one model/binding revision under one

contracts/capability/task_evaluation_execution.go

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

contracts/epistemics/research.go

@@ -96,6 +96,8 @@ type ResearchFinding struct {
 	Statement      string
 	Claim          Claim
 	Citations      []Citation
+	Source         Source
+	Excerpt        SourceExcerpt
 	State          ResearchFindingState
 	Reason         string
 }
@@ -131,6 +133,15 @@ type WebResearchRequest struct {
 	Statement      string
 }
 
+// WebSearchRequest asks the research runtime to discover bounded source
+// evidence for a human question. The provider is configured by the runtime;
+// the caller supplies meaning, not a provider URL or model choice.
+type WebSearchRequest struct {
+	Plan           ResearchPlan
+	PropositionKey string
+	Query          string
+}
+
 // ResolveCitation creates evidence only when all identifiers form one exact
 // source chain. It deliberately does not fetch, infer or rewrite source text.
 func ResolveCitation(source Source, snapshot SourceSnapshot, excerpt SourceExcerpt, evidence EvidenceKind, coverage CoverageState, conflict ConflictState, resolvedAtUnixMillis int64) (Citation, error) {

contracts/interaction/types.go

@@ -173,6 +173,7 @@ type ChatMessageResponse struct {
 	Clarification                  ClarificationState
 	WorkPlanned                    bool
 	PlanningReply                  string
+	ResearchBacked                 bool
 	CapabilityStatus               capability.CapabilityResolutionStatus
 	CapabilityKind                 capability.CapabilityKind
 	CapabilityHints                []capability.CapabilityKind

contracts/work/graph.go

@@ -21,14 +21,15 @@ const (
 	GraphFailed    GraphState = "FAILED"
 	GraphUnknown   GraphState = "UNKNOWN"
 
-	GraphNodeBlocked GraphNodeState = "BLOCKED_PREREQUISITE"
-	GraphNodeReady   GraphNodeState = "READY"
-	GraphNodeLeased  GraphNodeState = "LEASED"
-	GraphNodeRunning GraphNodeState = "RUNNING"
-	GraphNodeSuccess GraphNodeState = "SUCCEEDED"
-	GraphNodeFailed  GraphNodeState = "FAILED"
-	GraphNodeCancel  GraphNodeState = "CANCELLED"
-	GraphNodeUnknown GraphNodeState = "UNKNOWN"
+	GraphNodeBlocked   GraphNodeState = "BLOCKED_PREREQUISITE"
+	GraphNodeReady     GraphNodeState = "READY"
+	GraphNodeLeased    GraphNodeState = "LEASED"
+	GraphNodeRunning   GraphNodeState = "RUNNING"
+	GraphNodeVerifying GraphNodeState = "VERIFYING"
+	GraphNodeSuccess   GraphNodeState = "SUCCEEDED"
+	GraphNodeFailed    GraphNodeState = "FAILED"
+	GraphNodeCancel    GraphNodeState = "CANCELLED"
+	GraphNodeUnknown   GraphNodeState = "UNKNOWN"
 )
 
 type GraphBudget struct {
@@ -46,6 +47,8 @@ type GraphNode struct {
 	TaskProfileID  capability.TaskEvaluationProfileID `json:"task_profile_id,omitempty"`
 	BindingID      capability.CapabilityBindingID     `json:"binding_id,omitempty"`
 	CaseIDs        []string                           `json:"case_ids,omitempty"`
+	TaskProfile    *capability.TaskEvaluationProfile  `json:"task_profile,omitempty"`
+	TaskCases      []capability.TaskEvaluationCase    `json:"task_cases,omitempty"`
 }
 
 type GraphEdge struct {
@@ -54,15 +57,17 @@ 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"`
-	ParentGraphID       string      `json:"parent_graph_id,omitempty"`
-	ParentGraphRevision int64       `json:"parent_graph_revision,omitempty"`
-	TriggerEvidenceRef  string      `json:"trigger_evidence_ref,omitempty"`
+	ID                  string              `json:"id"`
+	Revision            int64               `json:"revision"`
+	State               GraphState          `json:"state"`
+	Nodes               []GraphNode         `json:"nodes"`
+	Edges               []GraphEdge         `json:"edges"`
+	Budget              GraphBudget         `json:"budget"`
+	ParentGraphID       string              `json:"parent_graph_id,omitempty"`
+	ParentGraphRevision int64               `json:"parent_graph_revision,omitempty"`
+	TriggerEvidenceRef  string              `json:"trigger_evidence_ref,omitempty"`
+	SearchInputBound    bool                `json:"search_input_bound,omitempty"`
+	SearchInput         SolutionSearchInput `json:"search_input,omitempty"`
 }
 
 // TaskEvaluationGraphRequest describes a durable fan-out. The graph contains
@@ -79,6 +84,22 @@ type TaskEvaluationGraphRequest struct {
 	ParentGraphRevision int64                            `json:"parent_graph_revision,omitempty"`
 	ParentGraphState    GraphState                       `json:"parent_graph_state,omitempty"`
 	TriggerEvidenceRef  string                           `json:"trigger_evidence_ref,omitempty"`
+	Cases               []capability.TaskEvaluationCase  `json:"cases,omitempty"`
+	BindingRevision     int64                            `json:"binding_revision,omitempty"`
+	ActorID             string                           `json:"actor_id,omitempty"`
+	PrincipalID         string                           `json:"principal_id,omitempty"`
+	ScopeID             string                           `json:"scope_id,omitempty"`
+}
+
+// TaskEvaluationEvidenceCallback is emitted by a specialist after it has
+// completed the bounded case set. Capability remains the evidence owner;
+// Work only advances its matching graph node after the evidence is accepted.
+type TaskEvaluationEvidenceCallback struct {
+	GraphID    string                    `json:"graph_id"`
+	Revision   int64                     `json:"revision"`
+	NodeID     GraphNodeID               `json:"node_id"`
+	Evaluation capability.TaskEvaluation `json:"evaluation"`
+	Succeeded  bool                      `json:"succeeded"`
 }
 
 type ResultCandidate struct {
@@ -92,6 +113,35 @@ type ResultCandidate struct {
 	Summary         string        `json:"summary"`
 }
 
+// GenericCandidateVerificationRequest asks the independent deterministic
+// verifier to validate attribution and evidence-envelope integrity. It does
+// not assert that the candidate's semantic content is true.
+type GenericCandidateVerificationRequest struct {
+	GraphID       string        `json:"graph_id"`
+	GraphRevision int64         `json:"graph_revision"`
+	NodeID        GraphNodeID   `json:"node_id"`
+	AuthorRunID   AgentRunID    `json:"author_run_id"`
+	CandidateID   string        `json:"candidate_id"`
+	Summary       string        `json:"summary"`
+	EvidenceRefs  []EvidenceRef `json:"evidence_refs"`
+}
+
+type GenericCandidateVerificationResponse struct {
+	Verified                bool   `json:"verified"`
+	VerificationEvidenceRef string `json:"verification_evidence_ref"`
+	VerifierIdentity        string `json:"verifier_identity"`
+	Summary                 string `json:"summary"`
+}
+
+// GraphResultCallback carries one agent's attributable candidate to Work.
+// Recording a candidate never makes the node successful; the node enters
+// VERIFYING until an independent synthesis/verifier accepts it.
+type GraphResultCallback struct {
+	GraphID   string          `json:"graph_id"`
+	Revision  int64           `json:"revision"`
+	Candidate ResultCandidate `json:"candidate"`
+}
+
 type ConflictSet struct {
 	ID            string      `json:"id"`
 	GraphID       string      `json:"graph_id"`

contracts/work/lease.go

@@ -1,6 +1,12 @@
 package work
 
-import "time"
+import (
+	"time"
+
+	"git.cin.su/m/Assistant/contracts/authority"
+	"git.cin.su/m/Assistant/contracts/capability"
+	"git.cin.su/m/Assistant/contracts/identity"
+)
 
 type AgentLeaseID string
 type AgentFencingToken uint64
@@ -50,14 +56,39 @@ type IndependentVerification struct {
 }
 
 type AgentDispatchIntent struct {
-	DispatchID             string       `json:"dispatch_id"`
-	IdempotencyKey         string       `json:"idempotency_key"`
-	DelegationID           DelegationID `json:"delegation_id"`
-	RunID                  AgentRunID   `json:"run_id"`
-	ImplementationID       string       `json:"implementation_id"`
-	ImplementationRevision int64        `json:"implementation_revision"`
-	SourceRevision         string       `json:"source_revision"`
-	WorktreeID             string       `json:"worktree_id"`
-	VerificationPlanRef    string       `json:"verification_plan_ref"`
-	CreatedAtUnixMillis    int64        `json:"created_at_unix_millis"`
+	DispatchID              string                             `json:"dispatch_id"`
+	IdempotencyKey          string                             `json:"idempotency_key"`
+	DelegationID            DelegationID                       `json:"delegation_id"`
+	RunID                   AgentRunID                         `json:"run_id"`
+	ImplementationID        string                             `json:"implementation_id"`
+	ImplementationRevision  int64                              `json:"implementation_revision"`
+	SourceRevision          string                             `json:"source_revision"`
+	WorktreeID              string                             `json:"worktree_id"`
+	VerificationPlanRef     string                             `json:"verification_plan_ref"`
+	CreatedAtUnixMillis     int64                              `json:"created_at_unix_millis"`
+	SpecialistPurpose       string                             `json:"specialist_purpose,omitempty"`
+	TaskEvaluationProfileID capability.TaskEvaluationProfileID `json:"task_evaluation_profile_id,omitempty"`
+	TaskEvaluationBindingID capability.CapabilityBindingID     `json:"task_evaluation_binding_id,omitempty"`
+	TaskEvaluationCaseIDs   []string                           `json:"task_evaluation_case_ids,omitempty"`
+	TaskEvaluationProfile   *capability.TaskEvaluationProfile  `json:"task_evaluation_profile,omitempty"`
+	TaskEvaluationCases     []capability.TaskEvaluationCase    `json:"task_evaluation_cases,omitempty"`
+	TaskEvaluationExecution *TaskEvaluationDispatchContext     `json:"task_evaluation_execution,omitempty"`
+}
+
+type TaskEvaluationDispatchContext struct {
+	GraphID         string               `json:"graph_id"`
+	GraphRevision   int64                `json:"graph_revision"`
+	NodeID          GraphNodeID          `json:"node_id"`
+	BindingRevision int64                `json:"binding_revision"`
+	ActorID         identity.ActorID     `json:"actor_id"`
+	PrincipalID     identity.PrincipalID `json:"principal_id"`
+	ScopeID         authority.ScopeID    `json:"scope_id"`
+	// GenericMessage turns the existing durable dispatch context into a
+	// discriminated local cognitive-work payload without adding a second
+	// persistence column. It is ignored by task-evaluation executors.
+	GenericMessage        string `json:"generic_message,omitempty"`
+	GenericPurpose        string `json:"generic_purpose,omitempty"`
+	GenericLocale         string `json:"generic_locale,omitempty"`
+	GenericDataPolicy     string `json:"generic_data_policy,omitempty"`
+	GenericMaxOutputBytes int    `json:"generic_max_output_bytes,omitempty"`
 }

contracts/work/orchestration.go

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

contracts/work/self_development.go

@@ -70,6 +70,7 @@ type ExtensionProposalRequest struct {
 // authority grant, implementation body or promotion decision.
 type ExtensionProposal struct {
 	ID                     string
+	RequestedByPrincipalID string
 	Stage                  ExtensionProposalStage
 	EvidenceRefs           []string
 	DeduplicatedSignalKeys []string
@@ -113,19 +114,20 @@ type DevelopmentScaffoldRequest struct {
 // 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
+	ID                     string
+	RequestedByPrincipalID string
+	Stage                  DevelopmentScaffoldStage
+	ProposalRef            string
+	BaseRepoRevision       string
+	WorktreeRef            string
+	TaskIntent             string
+	Modules                []string
+	AllowedPaths           []string
+	ForbiddenPaths         []string
+	RequiredTests          []string
+	ArtifactLineageRefs    []string
+	ProtectedMainAccess    string
+	ProductionAccess       string
+	SecretAccess           string
+	CreatedAtUnixMillis    int64
 }

database/migrations/040_task_evaluation_dispatch.sql

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

database/migrations/041_task_evaluation_payload.sql

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

database/migrations/042_agent_execution_recovery.sql

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

database/migrations/043_governed_development_candidates.sql

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

database/migrations/044_work_graph_orchestration_snapshot.sql

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

database/migrations/045_capability_model_qualification_observations.sql

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

deploy/docker-compose.prod.yml

@@ -118,10 +118,44 @@ services:
       WORK_POSTGRES_CREDENTIAL_REF: credential://domovoy/work/db-password/v1
       WORK_POSTGRES_PASSWORD_FILE: /run/secrets/work_db_password
       SECRET_BACKEND_ROOT: /run/secrets
-    depends_on: { postgres: { condition: service_healthy } }
+      TASK_VERIFIER_URL: http://task-verifier:8091
+      AGENT_EXECUTION_BRIDGE_URL: http://agent-execution-bridge:8092
+    depends_on:
+      postgres: { condition: service_healthy }
+      task-verifier: { condition: service_healthy }
+      agent-execution-bridge: { condition: service_healthy }
     volumes:
       - assistant_core_secrets:/run/secrets:ro
 
+  agent-execution-bridge:
+    build: { context: .., dockerfile: deploy/images/runtime.Dockerfile, args: { TARGET: agent-execution-bridge, SOURCE_REVISION: "${ASSISTANT_SOURCE_REVISION:-unknown}" } }
+    restart: unless-stopped
+    environment:
+      LISTEN_ADDR: :8092
+      COGNITIVE_WORKER_URL: http://cognitive-worker:8083
+      WORK_GRAPH_RUNTIME_URL: http://work-graph-runtime:8088
+      TASK_VERIFIER_URL: http://task-verifier:8091
+    depends_on:
+      cognitive-worker: { condition: service_started }
+      task-verifier: { condition: service_healthy }
+      work-graph-runtime: { condition: service_started }
+    healthcheck:
+      test: ["CMD-SHELL", "wget -qO- http://localhost:8092/healthz >/dev/null"]
+      interval: 10s
+      timeout: 3s
+      retries: 10
+
+  task-verifier:
+    build: { context: .., dockerfile: deploy/images/runtime.Dockerfile, args: { TARGET: task-verifier, SOURCE_REVISION: "${ASSISTANT_SOURCE_REVISION:-unknown}" } }
+    restart: unless-stopped
+    environment:
+      LISTEN_ADDR: :8091
+    healthcheck:
+      test: ["CMD-SHELL", "wget -qO- http://localhost:8091/healthz >/dev/null"]
+      interval: 10s
+      timeout: 3s
+      retries: 10
+
   work-graph-runtime:
     build: { context: .., dockerfile: deploy/images/runtime.Dockerfile, args: { TARGET: work-graph-runtime, SOURCE_REVISION: "${ASSISTANT_SOURCE_REVISION:-unknown}" } }
     restart: unless-stopped
@@ -161,6 +195,8 @@ services:
       LISTEN_ADDR: :8087
       RESEARCH_SOURCE_ROOT: /var/lib/research-sources
       RESEARCH_HTTP_ALLOWED_HOSTS: assistant.cin.su,git.cin.su
+      RESEARCH_SEARCH_URL_TEMPLATE: ${RESEARCH_SEARCH_URL_TEMPLATE:-https://www.bing.com/search?q=%s}
+      RESEARCH_SEARCH_ALLOWED_HOSTS: ${RESEARCH_SEARCH_ALLOWED_HOSTS:-www.bing.com}
       POSTGRES_HOST: postgres
       POSTGRES_PORT: "5432"
       POSTGRES_DATABASE: ${POSTGRES_DB:-assistant}
@@ -220,6 +256,7 @@ services:
       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_GPU_PROVIDER_TIMEOUT: ${LOCAL_GPU_PROVIDER_TIMEOUT:-60s}
       LOCAL_LIFEBOAT_PROVIDER_URL: http://cognitive-lifeboat:8086/v1
       COGNITIVE_OWNER_PRINCIPAL_ID: domovoy-creator-principal
       COGNITIVE_ALLOW_CANARY: "true"

dev/ACTIVE_TASK.json

@@ -1,5 +1,5 @@
 {
-  "task_file": "dev/tasks/DEV-000209-task-evaluator-work-graph.json",
-  "task_id": "DEV-000209",
+  "task_file": "dev/tasks/DEV-000291-mobile-acceptance-evidence.json",
+  "task_id": "DEV-000291",
   "task_revision": 1
 }

dev/tasks/DEV-000210-task-evaluator-worker.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000210",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "9a6ab6a9caf68e65f302c95ab0859b259122dacf",
+  "risk_tier": "E",
+  "intent": "Carry task-evaluation specialist metadata through durable dispatch, persistence and recovery so workers can execute bounded model cases and return evidence without losing profile lineage.",
+  "requirement_ids": ["worker execution context", "durable task metadata", "recovery-safe dispatch", "bounded case set", "no synchronous chat blocking"],
+  "allowed_change_scope": {"modules": ["work"], "paths": ["contracts/work/**", "modules/work/**", "adapters/postgres/work_repository.go", "cmd/agent-worker/**", "database/migrations/040_task_evaluation_dispatch.sql", "tests/model_qualification/**", "tests/agent_execution/**", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000210-task-evaluator-worker.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "deploy/**", "cmd/devctl/**", "modules/identity/**", "modules/capability/**"]},
+  "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": ["Work Graph task-evaluation node metadata", "AgentExecutionRequest/Result", "existing lease and recovery contract"], "may_change": ["typed dispatch task-evaluation metadata"]},
+  "semantic_owner": "work",
+  "canonical_state": ["work-owned dispatch metadata and graph lineage; capability remains owner of evaluation evidence"],
+  "state_transitions": ["graph specialist node -> enriched dispatch intent", "claimed dispatch -> executor receives profile/binding/cases", "stale claim -> recovery preserves task metadata for retry"],
+  "side_effects_and_privileges": ["work role migration only; no provider calls and no secret access added"],
+  "persistence_and_migrations": ["040 adds typed task metadata columns to existing work dispatch/graph tables"],
+  "failure_retry_restart": ["metadata is persisted before enqueue; idempotency conflict rejects divergent task context", "lease timeout/recovery does not erase profile or case references"],
+  "assumptions": ["an upcoming cognitive specialist executor uses the typed metadata to invoke the model gateway and post Capability evidence", "external bridge absence remains typed WAITING/local-fallback"],
+  "acceptance_scenarios": ["task graph dispatch carries profile, binding and all cases", "persist/load round trip preserves metadata", "divergent idempotency retry is rejected", "legacy non-task dispatch remains valid", "recovery path does not claim success"],
+  "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-000211-task-evaluator-callback.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000211",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "6dec47c79163f456cf6b18085a268ebb2619c7bb",
+  "risk_tier": "E",
+  "intent": "Accept specialist task-evaluation evidence through a Work Graph callback: persist evidence through the Capability owner and complete the matching graph node only after typed lineage and revision checks pass.",
+  "requirement_ids": ["evidence callback", "fan-in completion", "owner-local persistence", "revision safety", "no cross-owner DB access"],
+  "allowed_change_scope": {"modules": ["work"], "paths": ["contracts/work/**", "modules/work/**", "cmd/work-graph-runtime/**", "tests/model_qualification/**", "tests/agent_execution/**", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000211-task-evaluator-callback.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "deploy/**", "cmd/devctl/**", "modules/identity/**", "database/**", "modules/capability/**"]},
+  "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.TaskEvaluation", "existing Work Graph revision/completion contract", "Core Runtime capability evaluation HTTP boundary"], "may_change": ["typed task-evaluation callback contract"]},
+  "semantic_owner": "work for graph lifecycle; capability for evaluation evidence",
+  "canonical_state": ["Capability owns evaluation evidence; Work owns graph/node lifecycle and callback lineage"],
+  "state_transitions": ["specialist result -> callback received", "callback -> Capability evidence persisted", "persisted evidence -> matching node completion and graph fan-in"],
+  "side_effects_and_privileges": ["remote Core Runtime call through existing internal boundary; no direct foreign DB access and no new secrets"],
+  "persistence_and_migrations": ["uses existing capability task-evaluation and work graph persistence; no migration"],
+  "failure_retry_restart": ["Capability write failure leaves node incomplete and callback retryable", "stale graph revision is rejected; unknown/unsupported evaluation is not promoted by callback"],
+  "assumptions": ["specialist executor has already produced evidence and artifact references", "the callback is asynchronous and never blocks the chat request"],
+  "acceptance_scenarios": ["callback rejects non-task graph node", "callback rejects profile/binding lineage mismatch", "Capability write occurs before node completion", "sink failure leaves graph node unfinished", "valid callback advances graph fan-in", "duplicate evidence remains idempotent at Capability owner"],
+  "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"]
+}

dev/tasks/DEV-000212-task-evaluator-invocation.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000212",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "dff6c91c8e08611d73231a04a1af67efeebed22b",
+  "risk_tier": "E",
+  "intent": "Execute bounded task-evaluation cases against an exact Capability binding through Cognitive Gateway and return typed observations for the Work Graph callback.",
+  "requirement_ids": ["actual model invocation", "exact binding evaluation", "bounded cases", "provider-neutral evidence", "unknown on provider failure"],
+  "allowed_change_scope": {"modules": ["capability", "core-runtime", "cognitive-worker"], "paths": ["contracts/capability/**", "adapters/postgres/capability_repository.go", "cmd/core-runtime/**", "cmd/cognitive-worker/**", "tests/model_qualification/**", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000212-task-evaluator-invocation.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "deploy/**", "cmd/devctl/**", "database/**", "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": 0},
+  "contracts": {"consume": ["TaskEvaluationProfile", "TaskCaseObservation", "Cognitive Gateway provider boundary", "Capability cognitive binding snapshot boundary"], "may_change": ["task evaluation execution request/response"]},
+  "semantic_owner": "capability owns evaluation meaning; cognitive-worker owns provider invocation adapter boundary",
+  "canonical_state": ["none; this task returns observations only and does not promote or persist activation"],
+  "state_transitions": ["execution request -> exact binding snapshot", "case -> provider invocation -> SUPPORTED or UNKNOWN observation", "all observations -> Work Graph callback input"],
+  "side_effects_and_privileges": ["provider calls only through existing cognitive adapter; no direct DB and no new secret access"],
+  "persistence_and_migrations": ["none; durable persistence remains callback/Core Capability owner"],
+  "failure_retry_restart": ["provider timeout/auth/unavailable maps to UNKNOWN observation with evidence reference; caller may retry as a new attempt", "partial case set is never returned as supported aggregate"],
+  "assumptions": ["specialist supplies case prompts and artifact references from the approved bounded profile", "quality scoring beyond bounded invocation is performed by a verifier or later evaluator and is not fabricated here"],
+  "acceptance_scenarios": ["exact binding snapshot is required", "all supplied cases are invoked", "successful non-empty response yields supported observation with latency", "provider failure yields unknown observation", "unknown is never silently converted to supported", "no provider or hardware field participates in scoring"],
+  "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"]
+}

dev/tasks/DEV-000213-worker-evaluator-link.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000213",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "a86d1994c3f48db86592eb87fb54533297e0671c",
+  "risk_tier": "E",
+  "intent": "Carry the complete bounded task-evaluation profile and case payload through durable Work Graph dispatch so agent-worker can invoke Cognitive Worker, aggregate observations, and return the existing Work Graph callback after restart-safe recovery.",
+  "requirement_ids": ["dispatch-to-evaluator link", "payload durability", "worker recovery", "observation aggregation", "callback integration"],
+  "allowed_change_scope": {"modules": ["work", "agent-worker"], "paths": ["contracts/work/**", "modules/work/**", "adapters/postgres/work_repository.go", "cmd/agent-worker/**", "cmd/work-graph-runtime/**", "database/migrations/041_task_evaluation_payload.sql", "tests/agent_execution/**", "tests/model_qualification/**", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000213-worker-evaluator-link.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "deploy/**", "cmd/devctl/**", "modules/identity/**", "modules/capability/**"]},
+  "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": ["TaskEvaluationExecutionRequest/Response", "Work Graph callback", "existing AgentExecutor and recovery contract"], "may_change": ["typed dispatch payload and worker evaluator executor"]},
+  "semantic_owner": "work owns dispatch lifecycle; capability owns evaluation evidence; cognitive-worker owns provider invocation",
+  "canonical_state": ["work dispatch payload is durable input; Capability remains evidence owner"],
+  "state_transitions": ["task graph node -> durable payload dispatch", "worker claim -> evaluator invocation", "response -> aggregate TaskEvaluation", "aggregate -> Work Graph evidence callback"],
+  "side_effects_and_privileges": ["HTTP calls to internal cognitive/core/work boundaries; no direct foreign DB and no new secret"],
+  "persistence_and_migrations": ["041 adds JSONB payload columns to work graph nodes and dispatch intents"],
+  "failure_retry_restart": ["payload is loaded after stale claim recovery; evaluator failure leaves dispatch retryable", "callback only follows a persisted aggregate; unknown remains unknown"],
+  "assumptions": ["Cognitive Worker endpoint is available as optional local execution path", "external bridge remains preferred for non-evaluation agent tasks"],
+  "acceptance_scenarios": ["profile/cases survive save-load", "task dispatch invokes evaluator with original prompt payload", "all observations aggregate deterministically", "callback receives matching binding/profile/artifact", "evaluator or callback failure does not complete dispatch", "legacy dispatch remains unchanged"],
+  "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-000214-agent-execution-recovery.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000214",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "f9b9765100df02bead16280d2db9c1af478a4589",
+  "risk_tier": "E",
+  "intent": "Make Agent Execution recovery restart-safe: renew active leases with managed heartbeats, fence stale workers, cap retries, and requeue transient execution failures without losing durable dispatch state.",
+  "requirement_ids": ["durable heartbeat", "lease fencing", "bounded retry", "restart recovery", "transient failure requeue"],
+  "allowed_change_scope": {"modules": ["work", "agent-worker"], "paths": ["contracts/work/**", "modules/work/**", "adapters/postgres/work_repository.go", "cmd/agent-worker/**", "cmd/agent-canary/**", "database/migrations/042_agent_execution_recovery.sql", "tests/agent_execution/**", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000214-agent-execution-recovery.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "deploy/**", "cmd/devctl/**", "modules/identity/**", "modules/capability/**", "LLM-local/**"]},
+  "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 AgentExecutor", "existing lease/callback contract", "durable dispatch queue"], "may_change": ["WorkerStore retry/recovery port"]},
+  "semantic_owner": "work owns dispatch, lease, callback and retry lifecycle",
+  "canonical_state": ["work agent dispatch queue and lease tables remain canonical; heartbeat is durable evidence of liveness"],
+  "state_transitions": ["CLAIMED -> QUEUED on transient failure or stale recovery", "CLAIMED -> FAILED after retry limit", "ACTIVE lease -> EXPIRED before stale claim requeue", "heartbeat renews lease and callback sequence"],
+  "side_effects_and_privileges": ["work_role writes only work-owned queue and lease state; no external mutation"],
+  "persistence_and_migrations": ["042 adds bounded retry limit to dispatch queue"],
+  "failure_retry_restart": ["stale recovery expires the old lease before requeue", "transient execution failure uses durable delayed retry", "retry limit prevents hot-loop poison dispatch", "heartbeat failure cancels execution and leaves a retryable dispatch"],
+  "assumptions": ["operational heartbeat ticker is lifecycle-managed by the bounded worker call and durable state remains authoritative"],
+  "acceptance_scenarios": ["long execution renews lease and emits ordered heartbeats", "stale claim is fenced and requeued", "third failed attempt becomes FAILED", "waiting/error dispatch is not left CLAIMED", "legacy dispatch behavior remains compatible"],
+  "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-000215-governed-development-control-plane.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000215",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "8d3e7f7cac236aa50919bd1cd854012dcc8500a5",
+  "risk_tier": "E",
+  "intent": "Expose a durable, authenticated control-plane boundary for governed self-development proposals and isolated scaffold candidates without executing source mutation, bypassing verification, or granting promotion authority.",
+  "requirement_ids": ["durable proposal intake", "isolated scaffold contract", "authenticated operator boundary", "no implicit promotion"],
+  "allowed_change_scope": {"modules": ["work", "core-runtime"], "paths": ["contracts/work/self_development.go", "modules/work/domain/**", "cmd/core-runtime/main.go", "adapters/postgres/work_repository.go", "database/migrations/043_governed_development_candidates.sql", "tests/semantic_evolution/**", "tests/work/**", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000215-governed-development-control-plane.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "deploy/**", "cmd/devctl/**", "modules/identity/**", "modules/capability/**", "LLM-local/**"]},
+  "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": ["ADR-0022 governed self-development boundary", "creator authentication", "work development gate"], "may_change": ["work-owned proposal/scaffold persistence and authenticated control-plane endpoints"]},
+  "semantic_owner": "work owns proposal, scaffold and development gate lineage",
+  "canonical_state": ["work-owned extension proposal and scaffold candidates are durable; source tree and production remain outside this task"],
+  "state_transitions": ["repeated evidence -> PROPOSED", "proposal -> SANDBOXED scaffold candidate", "gate remains explicit and cannot be promoted by these endpoints"],
+  "side_effects_and_privileges": ["authenticated creator session required; database writes stay within work owner; no checkout, source mutation, deployment or secret access"],
+  "persistence_and_migrations": ["043 adds durable work-owned proposal/scaffold candidate records"],
+  "failure_retry_restart": ["idempotent candidate IDs preserve lineage; duplicate IDs are rejected; no endpoint advances promotion stage"],
+  "assumptions": ["an independent executor and verifier remain separate future capabilities"],
+  "acceptance_scenarios": ["unauthenticated requests are rejected", "repeated evidence creates a durable proposal", "protected paths cannot create a scaffold", "valid scaffold is durable and remains SANDBOXED", "no promotion or source mutation occurs"],
+  "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-000216-agent-worker-status-truth.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000216",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "62f0c529b3f8b3c92e4a69fcc6298d59b3015fde",
+  "risk_tier": "E",
+  "intent": "Keep agent-worker status truthful when the protected external execution bridge is absent while retaining local task-evaluation fallback.",
+  "requirement_ids": ["WP-301", "truthful degraded status", "local fallback"],
+  "allowed_change_scope": {"modules": ["agent-worker"], "paths": ["cmd/agent-worker/main.go", "cmd/agent-worker/status_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000216-agent-worker-status-truth.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "database/**", "deploy/**", "security/**", "contracts/**", "modules/**", "LLM-local/**"]},
+  "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 bridge file", "agent-worker status response", "local task-evaluation fallback"], "may_change": []},
+  "semantic_owner": "work",
+  "canonical_state": ["existing worker status projection and durable dispatch lifecycle"],
+  "state_transitions": ["bridge absent -> local-fallback status", "bridge present -> execution status"],
+  "side_effects_and_privileges": ["no external mutation; no new credentials; no persistence change"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["bridge absence remains explicit after restart; local evaluator remains available for bounded qualification work"],
+  "assumptions": ["task-evaluation fallback is not general external agent execution"],
+  "acceptance_scenarios": ["worker marks local_fallback when protected bridge URL is absent", "status never reports executor_configured=true for local fallback", "configured bridge path remains execution mode", "no secret value is exposed"],
+  "required_tests": ["go test ./cmd/agent-worker", "go vet ./cmd/agent-worker", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production worker status"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000217-current-status-evidence.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000217",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "05a246ea3f240a77be769969ad62a236fba33bac",
+  "risk_tier": "E",
+  "intent": "Refresh the project status projection and redacted live evidence after the agent-worker truthful fallback fix without changing runtime behavior.",
+  "requirement_ids": ["current audit evidence", "task lineage accuracy", "negative-state transparency"],
+  "allowed_change_scope": {"modules": ["operations-evidence"], "paths": ["docs/PROJECT_STATUS.md", "docs/evidence/audit-runtime-status-2026-08-23-v2.json", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000217-current-status-evidence.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/**", "modules/**", "contracts/**", "database/**", "deploy/**", "LLM-local/**"]},
+  "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 responses", "merged task lineage"], "may_change": ["operations status projection and redacted evidence artifact"]},
+  "semantic_owner": "operations-evidence projection",
+  "canonical_state": ["none; docs and evidence are rebuildable projections"],
+  "state_transitions": ["stale status projection -> current evidence projection"],
+  "side_effects_and_privileges": ["read-only live checks; no secrets, credentials or runtime mutation"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["evidence records unavailable/degraded states explicitly; no success inferred from missing data"],
+  "assumptions": ["historical evidence remains immutable and is not rewritten"],
+  "acceptance_scenarios": ["active task points to the current merged packet", "bridge absence is reported as local fallback", "GPU timeout is reported as degraded without changing FAST preference", "authenticated UI availability is not claimed without a current signed-in session", "no secret value appears in the projection"],
+  "required_tests": ["go test", "go vet", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production public health and runtime status"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000218-automatic-expansion-policy.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000218",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "ab840878fe7d21a0a6aeb69193a3d58ea23a56e0",
+  "risk_tier": "E",
+  "intent": "Add a conservative deterministic policy gate for creating additional model-evaluation profiles only after general evaluation evidence remains unresolved.",
+  "requirement_ids": ["general-first evaluation", "bounded additional evaluation", "no unnecessary fleet-wide tests"],
+  "allowed_change_scope": {"modules": ["capability"], "paths": ["modules/capability/domain/task_evaluation.go", "modules/capability/domain/task_evaluation_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000218-automatic-expansion-policy.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "cmd/**", "contracts/**", "LLM-local/**"]},
+  "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": ["TaskEvaluationProfile", "TaskEvaluation", "TaskEvaluationExpansionRequest"], "may_change": []},
+  "semantic_owner": "capability owns model qualification evidence and expansion policy",
+  "canonical_state": ["none; this packet creates no persistence or activation authority"],
+  "state_transitions": ["general COMPLETED/PARTIAL unresolved evidence -> bounded additional profile candidate", "qualified candidate -> no expansion"],
+  "side_effects_and_privileges": ["pure deterministic decision; no provider call, no model invocation, no mutation"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["missing trigger evidence or additional cases fails closed; repeated decision is deterministic"],
+  "assumptions": ["a later task supplies task-specific additional cases from an unresolved outcome and schedules them through the existing Work Graph"],
+  "acceptance_scenarios": ["general graph not completed/partial does not expand", "unknown/unsupported/low-scoring candidates expand only with evidence and new cases", "one qualified candidate suppresses expansion", "no provider or hardware label changes the decision"],
+  "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_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000219-reachability-not-quality.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000219",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "7476f43ffa57d55f5a5a9bddc7f1c74e9b87ae0e",
+  "risk_tier": "E",
+  "intent": "Prevent task-evaluation invocation success from being recorded as semantic model quality or support without independent verifier evidence.",
+  "requirement_ids": ["truthful model evidence", "independent verification", "no false qualification"],
+  "allowed_change_scope": {"modules": ["cognitive-worker"], "paths": ["cmd/cognitive-worker/task_evaluation.go", "cmd/cognitive-worker/task_evaluation_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000219-reachability-not-quality.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "contracts/**", "modules/**", "LLM-local/**"]},
+  "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": ["TaskEvaluationExecutionRequest/Response", "TaskCaseObservation", "Cognitive Gateway"], "may_change": []},
+  "semantic_owner": "capability evidence remains owner; cognitive-worker only records invocation observation",
+  "canonical_state": ["existing task-evaluation evidence; no new persistence"],
+  "state_transitions": ["successful invocation -> UNKNOWN reachability observation", "independent verifier evidence -> future supported/quality observation"],
+  "side_effects_and_privileges": ["no additional provider calls or privileges"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["unknown evidence remains retryable/eligible for a separately verified evaluation; no promotion follows invocation alone"],
+  "assumptions": ["a later verifier task will attach answer-quality evidence and score cases"],
+  "acceptance_scenarios": ["nonempty provider response is not SUPPORTED", "quality score remains zero", "provider invocation failure remains UNKNOWN with critical failure", "no model/hardware label can manufacture quality evidence"],
+  "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_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000220-verifier-evidence-boundary.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000220",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "66ed523248ea6cf0a41383e2206d9a70eeaafda8",
+  "risk_tier": "E",
+  "intent": "Require explicit independent verifier identity and evidence before task evaluation can claim semantic support or non-zero quality.",
+  "requirement_ids": ["independent verifier evidence", "no false quality", "typed model qualification"],
+  "allowed_change_scope": {"modules": ["capability", "agent-worker"], "paths": ["contracts/capability/task_evaluation.go", "modules/capability/domain/task_evaluation.go", "modules/capability/domain/task_evaluation_test.go", "cmd/agent-worker/task_evaluation_executor_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000220-verifier-evidence-boundary.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "cmd/cognitive-worker/**", "LLM-local/**"]},
+  "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": ["TaskCaseObservation", "TaskEvaluation", "existing Work Graph callback"], "may_change": ["typed observation fields for verifier evidence"]},
+  "semantic_owner": "capability owns qualification evidence; verifier remains independent of provider invocation",
+  "canonical_state": ["existing capability task-evaluation evidence"],
+  "state_transitions": ["reachability UNKNOWN -> separately verified SUPPORTED only with evidence"],
+  "side_effects_and_privileges": ["no provider invocation or external mutation"],
+  "persistence_and_migrations": ["existing task evaluation persistence is unchanged"],
+  "failure_retry_restart": ["missing verifier fields fail closed; unknown reachability remains non-promotable"],
+  "assumptions": ["a later verifier runtime will produce VerificationEvidenceRef and VerifierIdentity"],
+  "acceptance_scenarios": ["supported without verifier is rejected", "unknown with zero quality remains valid", "verified supported observation aggregates normally", "no hardware/provider name bypasses the boundary"],
+  "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_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000221-task-verifier-runtime.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000221",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "0aab10863cb23824688029c0bab6734f100b4610",
+  "risk_tier": "E",
+  "intent": "Provide an independent deterministic task-verifier runtime and wire model task evaluation through response evidence rather than provider self-assertion.",
+  "requirement_ids": ["independent verifier runtime", "typed answer evidence", "quality qualification boundary", "degraded verifier fallback"],
+  "allowed_change_scope": {"modules": ["capability", "cognitive-worker", "agent-worker", "task-verifier", "deployment"], "paths": ["contracts/capability/task_evaluation.go", "contracts/capability/task_evaluation_execution.go", "cmd/cognitive-worker/task_evaluation.go", "cmd/agent-worker/main.go", "cmd/agent-worker/task_evaluation_executor.go", "cmd/agent-worker/task_evaluation_executor_test.go", "cmd/task-verifier/**", "deploy/docker-compose.prod.yml", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000221-task-verifier-runtime.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "modules/**", "LLM-local/**"]},
+  "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": ["TaskEvaluationCase", "TaskEvaluationExecutionRequest/Response", "TaskCaseObservation", "existing worker callback"], "may_change": ["typed verification spec, response envelope and verifier result" ]},
+  "semantic_owner": "capability owns evidence; task-verifier is a replaceable independent verifier adapter and does not own canonical state",
+  "canonical_state": ["existing capability task-evaluation evidence only; verifier has no database"],
+  "state_transitions": ["provider response -> bounded response evidence -> verifier result -> TaskCaseObservation -> TaskEvaluation"],
+  "side_effects_and_privileges": ["local network calls only; no secrets; no external mutation; no authority expansion"],
+  "persistence_and_migrations": ["none; verifier evidence is carried through existing callback and capability persistence"],
+  "failure_retry_restart": ["verifier unavailable returns local fallback/waiting; no quality or promotion claim is made"],
+  "assumptions": ["initial verifier mode is exact-text for bounded deterministic cases; future verifiers may add typed modes under separate task packets"],
+  "acceptance_scenarios": ["exact response produces supported evidence with verifier identity", "mismatch produces unsupported evidence", "cognitive worker never asserts support from invocation alone", "worker wires verifier result before callback", "verifier process has health endpoint and production compose dependency", "verifier loss does not produce false quality"],
+  "required_tests": ["go test", "go vet", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production public health and verifier health check"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000222-verifier-modes-and-case-builder.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000222",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "d736500d9818a7748c3715c442edb695c5aed1100",
+  "risk_tier": "E",
+  "intent": "Extend the independent task verifier with a structured JSON field mode and add a bounded case builder driven by unresolved evidence.",
+  "requirement_ids": ["task-aware verification", "additional cases during lifetime", "general-first bounded expansion"],
+  "allowed_change_scope": {"modules": ["capability", "task-verifier"], "paths": ["contracts/capability/task_evaluation.go", "cmd/task-verifier/main.go", "cmd/task-verifier/main_test.go", "modules/capability/domain/task_evaluation.go", "modules/capability/domain/task_evaluation_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000222-verifier-modes-and-case-builder.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "cmd/core-runtime/**", "cmd/creator-web/**", "cmd/agent-worker/**", "cmd/cognitive-worker/**", "modules/identity/**", "LLM-local/**"]},
+  "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": ["TaskCaseVerification", "TaskEvaluationCase", "unresolved outcome evidence"], "may_change": ["verification modes and bounded case construction" ]},
+  "semantic_owner": "capability owns evaluation policy; verifier remains stateless adapter",
+  "canonical_state": ["none; case builder returns candidates and does not persist or activate them"],
+  "state_transitions": ["unresolved evidence -> two bounded additional case candidates", "structured response -> verifier evidence -> supported/unsupported"],
+  "side_effects_and_privileges": ["no model calls, no external mutation, no secrets"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["invalid unresolved evidence fails closed; verifier remains stateless and repeatable"],
+  "assumptions": ["generated cases still require an explicit verification spec before quality can be claimed"],
+  "acceptance_scenarios": ["JSON field match verifies with evidence", "JSON mismatch is unsupported", "unresolved evidence yields exactly two bounded probes", "generated cases do not bypass verifier or promotion gate"],
+  "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_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000223-status-reconciliation.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000223",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "ee68e2207eff5535c79cc594637a971cea03d71c",
+  "risk_tier": "E",
+  "intent": "Reconcile the status projection with the merged DEV-000222 verifier stage and preserve explicit remaining runtime and UI gaps.",
+  "requirement_ids": ["current audit evidence", "task lineage accuracy", "negative-state transparency"],
+  "allowed_change_scope": {"modules": ["operations-evidence"], "paths": ["docs/PROJECT_STATUS.md", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000223-status-reconciliation.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/**", "modules/**", "contracts/**", "database/**", "deploy/**", "LLM-local/**"]},
+  "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": ["merged task lineage", "existing redacted live evidence"], "may_change": ["operations status projection"]},
+  "semantic_owner": "operations-evidence projection",
+  "canonical_state": ["none; docs are rebuildable projections"],
+  "state_transitions": ["stale task pointer -> merged verifier stage pointer"],
+  "side_effects_and_privileges": ["none; documentation only"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["historical evidence remains immutable"],
+  "assumptions": ["DEV-000222 main revision ee68e22 is the current deployed baseline"],
+  "acceptance_scenarios": ["status names DEV-000222 as last completed", "remaining bridge and current authenticated UI gaps remain explicit", "no secret value is added"],
+  "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_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000224-live-fleet-evidence.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000224",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "c4411afaaf21886cba9b5ead586afb26d697a355",
+  "risk_tier": "E",
+  "intent": "Record the current production model-fleet readiness and dynamic tier selection without claiming quality or promotion.",
+  "requirement_ids": ["model fleet/failover", "GPU is not intelligence priority", "truthful operational evidence"],
+  "allowed_change_scope": {"modules": ["operations-evidence"], "paths": ["docs/PROJECT_STATUS.md", "docs/evidence/fleet-readiness-live-2026-08-23-v2.json", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000224-live-fleet-evidence.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/**", "modules/**", "contracts/**", "database/**", "deploy/**", "LLM-local/**"]},
+  "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 cognitive-worker fleet readiness response"], "may_change": ["redacted evidence and status projection"]},
+  "semantic_owner": "operations-evidence projection",
+  "canonical_state": ["none; evidence is a rebuildable projection"],
+  "state_transitions": ["live fleet readiness -> redacted evidence record"],
+  "side_effects_and_privileges": ["read-only production inspection"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["unavailable providers remain explicit; no missing data is promoted to ready"],
+  "assumptions": ["fleet endpoint response is authoritative for this observation window"],
+  "acceptance_scenarios": ["FAST unavailable selects available MEDIUM", "HEAVY availability does not imply priority", "LIFEBOAT remains fallback", "secret values are absent"],
+  "required_tests": ["go test", "go vet", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production public health and fleet live check"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000225-login-route-fix.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000225",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "e080c34a772caa5d256ab9f4ae0d2568bb8a7fb9",
+  "risk_tier": "E",
+  "intent": "Serve the existing chat/login shell at the direct /login route so users can reach authentication without a 404.",
+  "requirement_ids": ["UI login direct route", "authenticated acceptance entrypoint"],
+  "allowed_change_scope": {"modules": ["creator-web"], "paths": ["cmd/creator-web/server.go", "cmd/creator-web/server_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000225-login-route-fix.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "contracts/**", "modules/**", "LLM-local/**"]},
+  "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 embedded chat/login shell"], "may_change": ["HTTP route mapping only"]},
+  "semantic_owner": "creator-web route projection",
+  "canonical_state": ["none"],
+  "state_transitions": ["GET /login -> existing chat shell"],
+  "side_effects_and_privileges": ["none; GET only"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["route remains deterministic and stateless"],
+  "assumptions": ["authentication behavior remains unchanged and is handled by the existing chat shell"],
+  "acceptance_scenarios": ["GET /login returns 200", "login inputs are present", "root route remains unchanged", "no credential values are added"],
+  "required_tests": ["go test", "go vet", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production public health and /login check"],
+  "terminal_gap_states": ["BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000226-ui-acceptance-evidence.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000226",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "40b56798e23d852d26156b0bc2abbc518fd6b9b6",
+  "risk_tier": "E",
+  "intent": "Record current authenticated desktop/mobile UI and operations acceptance evidence without mutating production state.",
+  "requirement_ids": ["UI operations acceptance", "authenticated browser evidence", "responsive acceptance"],
+  "allowed_change_scope": {"modules": ["operations-evidence"], "paths": ["docs/PROJECT_STATUS.md", "docs/evidence/authenticated-ui-acceptance-2026-08-23-v2.json", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000226-ui-acceptance-evidence.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "cmd/**", "modules/**", "contracts/**", "database/**", "deploy/**", "LLM-local/**"]},
+  "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 authenticated browser DOM and visual state", "live production UI responses"], "may_change": ["redacted UI acceptance evidence and status projection"]},
+  "semantic_owner": "operations-evidence projection",
+  "canonical_state": ["none; evidence is rebuildable and redacted"],
+  "state_transitions": ["current browser acceptance -> evidence record"],
+  "side_effects_and_privileges": ["read-only navigation and section inspection; no production mutation"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["missing authenticated session or unavailable section must remain explicit"],
+  "assumptions": ["browser evidence reflects the deployed revision shown in the page metadata"],
+  "acceptance_scenarios": ["root chat authenticated", "diagnostics 7/7", "fleet 4/4 with fallback visible", "learning/reflection visible", "390x844 layout readable", "no secret values retained"],
+  "required_tests": ["go test", "go vet", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production public health and authenticated browser evidence"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000227-local-executor.json

@@ -0,0 +1,47 @@
+{
+  "task_id": "DEV-000227",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "1f6ca363ee241b41a89bff1f1b603f15e9873b00",
+  "risk_tier": "E",
+  "intent": "Make the existing bounded local task-evaluation path an explicit local executor capability while preserving truthful durable waiting for unsupported work and keeping the external bridge optional.",
+  "requirement_ids": ["local executor", "truthful degraded execution", "optional external bridge"],
+  "allowed_change_scope": {
+    "modules": ["agent-worker", "operations-status"],
+    "paths": [
+      "cmd/agent-worker/local_fallback.go",
+      "cmd/agent-worker/local_fallback_test.go",
+      "cmd/agent-worker/task_evaluation_executor.go",
+      "cmd/agent-worker/status.go",
+      "cmd/agent-worker/status_test.go",
+      "cmd/agent-worker/main.go",
+      "dev/ACTIVE_TASK.json",
+      "dev/tasks/DEV-000227-local-executor.json",
+      "docs/PROJECT_STATUS.md"
+    ]
+  },
+  "forbidden_change_scope": {
+    "paths": ["architecture/**", "security/**", "contracts/**", "database/**", "deploy/**", "LLM-local/**"]
+  },
+  "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.AgentExecutor", "typed task-evaluation runtime endpoints"],
+    "may_change": ["agent-worker status projection adds local executor capability metadata"]
+  },
+  "semantic_owner": "work agent-worker runtime",
+  "canonical_state": ["existing durable Work dispatch, lease and callback state"],
+  "state_transitions": ["unsupported local request -> WAITING", "supported task evaluation -> verified callback"],
+  "side_effects_and_privileges": ["existing bounded internal task-evaluation calls and Work callback only; no external mutation authority"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["unsupported work remains durable WAITING; existing lease recovery remains authoritative"],
+  "assumptions": ["task-evaluation execution and verification endpoints are already deployed internal capabilities", "external bridge may remain absent"],
+  "acceptance_scenarios": ["local status names bounded local executor and capability", "unsupported work never claims completion", "task evaluation still executes, verifies and callbacks", "external bridge status remains separate", "no secret or new privilege is introduced"],
+  "required_tests": ["go test ./cmd/agent-worker/...", "go vet ./cmd/agent-worker/...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production status smoke"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000228-controlled-discovery.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000228",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "2609f31e59d3914cdf4adcf23ec618705277d12a",
+  "risk_tier": "E",
+  "intent": "Expose a read-only, provider-neutral discovery status view so Domovoy can inspect available optional providers and models before requesting qualification, without exposing credential metadata or activating anything.",
+  "requirement_ids": ["controlled discovery", "provider qualification visibility", "optional external tools"],
+  "allowed_change_scope": {"modules": ["cognitive-worker", "operations-status"], "paths": ["cmd/cognitive-worker/discovery.go", "cmd/cognitive-worker/discovery_test.go", "cmd/cognitive-worker/main.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000228-controlled-discovery.json", "docs/PROJECT_STATUS.md"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "contracts/**", "database/**", "deploy/**", "LLM-local/**"]},
+  "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 LocalProvider discovery probes", "existing capability model metadata"], "may_change": ["read-only cognitive-worker discovery status projection"]},
+  "semantic_owner": "cognitive-worker observation projection",
+  "canonical_state": ["existing Capability discovery registration remains the canonical persisted state"],
+  "state_transitions": ["provider probe -> observed discovery status", "discovery registration remains DISCOVERED/INACTIVE until qualification and promotion"],
+  "side_effects_and_privileges": ["read-only provider probe on explicit status request; no activation, mutation or credential disclosure"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["unavailable provider is reported with health/failure state; persisted discovery remains available through existing owner"],
+  "assumptions": ["provider discovery adapters already enforce bounded requests", "qualification remains a separate task-evaluation and promotion step"],
+  "acceptance_scenarios": ["status lists registered providers deterministically", "model inventory and probe health are visible", "credential references are not returned", "status does not activate or persist a provider", "unknown provider remains absent or unavailable"],
+  "required_tests": ["go test ./cmd/cognitive-worker/...", "go vet ./cmd/cognitive-worker/...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production status smoke"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000229-research-runtime-bridge.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000229",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "22f839df99d8f6abd95f00cf99d34b4b45ace692",
+  "risk_tier": "E",
+  "intent": "Connect authenticated Domovoy/Core requests to the existing bounded Truth/Research web runtime while preserving source freshness, exact citations, conflicts and read-only connector policy.",
+  "requirement_ids": ["Truth/Research runtime", "authenticated internet research", "source evidence"],
+  "allowed_change_scope": {"modules": ["core-runtime", "research-runtime-bridge"], "paths": ["cmd/core-runtime/main.go", "cmd/core-runtime/research_client.go", "cmd/core-runtime/research_client_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000229-research-runtime-bridge.json", "docs/PROJECT_STATUS.md"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "contracts/**", "database/**", "deploy/**", "LLM-local/**"]},
+  "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 epistemics.WebResearchRequest and ResearchResult", "existing authenticated Creator session"], "may_change": ["authenticated Core-to-research runtime adapter route"]},
+  "semantic_owner": "epistemics research runtime; Core is an authenticated transport boundary",
+  "canonical_state": ["epistemics runtime remains sole owner of persisted truth bundles, claims, beliefs and citations"],
+  "state_transitions": ["authenticated request -> bounded web connector -> evaluated persisted finding"],
+  "side_effects_and_privileges": ["read-only outbound source fetches to preconfigured allowlisted hosts; authenticated user request persists evidence in epistemics owner"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["research runtime unavailable returns explicit gateway failure; no claim is emitted by Core"],
+  "assumptions": ["research runtime allowlist is configured outside code", "web connector retains exact source snapshot and does not invent citations"],
+  "acceptance_scenarios": ["unauthenticated research request is denied", "authenticated request reaches research runtime", "source evidence and finding are returned without fabricated claims", "research runtime remains bounded by allowlist", "no secret is forwarded or returned"],
+  "required_tests": ["go test ./cmd/core-runtime/...", "go vet ./cmd/core-runtime/...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production health and research-runtime readiness smoke"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000230-development-gate-controller.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000230",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "ec79f19a7935e1eb8f0f59c4c9ce2940b97c4855",
+  "risk_tier": "E",
+  "intent": "Expose the existing Work-owned governed development state machine through authenticated optimistic-concurrency routes without executing code, changing protected source or bypassing promotion gates.",
+  "requirement_ids": ["governed self-development", "durable gate progression", "canary rollback control"],
+  "allowed_change_scope": {"modules": ["core-runtime", "governed-development"], "paths": ["cmd/core-runtime/main.go", "cmd/core-runtime/development_gate.go", "cmd/core-runtime/development_gate_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000230-development-gate-controller.json", "docs/PROJECT_STATUS.md"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "contracts/**", "database/**", "deploy/**", "LLM-local/**"]},
+  "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 DevelopmentGate state machine", "authenticated Creator session", "optimistic revision persistence"], "may_change": ["authenticated Core gate controller routes"]},
+  "semantic_owner": "Work governed development gate",
+  "canonical_state": ["Work-owned development gate and existing proposal/scaffold records"],
+  "state_transitions": ["EVIDENCE_COLLECTED -> PROPOSAL_CREATED -> TASK_ADMITTED -> ISOLATED -> VERIFIED -> CANARY -> PROMOTED", "CANARY -> ROLLED_BACK"],
+  "side_effects_and_privileges": ["authenticated durable gate creation/transition only; no checkout, code execution, secret access, production mutation or automatic promotion"],
+  "persistence_and_migrations": ["none; existing development_gates table and repository are reused"],
+  "failure_retry_restart": ["optimistic revision conflict is explicit; invalid or skipped transitions are rejected; rollback remains available only from canary"],
+  "assumptions": ["Creator authentication is the explicit governance boundary for these operator routes", "actual worktree execution and deployment remain separate controlled systems"],
+  "acceptance_scenarios": ["gate creation requires evidence-collected revision one", "sequential transitions require their typed references", "verified cannot jump directly to promoted", "stale revision is rejected", "canary rollback is allowed", "unauthenticated requests are denied"],
+  "required_tests": ["go test ./cmd/core-runtime/...", "go vet ./cmd/core-runtime/...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production health smoke"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000231-solution-orchestrator.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000231",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "45874f218646d7879dfcbc814782fe56bf0aa06a",
+  "risk_tier": "E",
+  "intent": "Provide a durable, parallel-first solution-search decision policy for Work Graph goals so unresolved work continues through independent agent, capability, research and reflection avenues instead of dead-ending after one bounded wave.",
+  "requirement_ids": ["solution orchestration", "parallel agent team", "durable retry waves", "independent verification"],
+  "allowed_change_scope": {"modules": ["work-graph", "work-domain"], "paths": ["contracts/work/orchestration.go", "modules/work/domain/orchestration.go", "modules/work/domain/orchestration_test.go", "modules/work/application/orchestration.go", "cmd/work-graph-runtime/main.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000231-solution-orchestrator.json", "docs/PROJECT_STATUS.md"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "LLM-local/**", "modules/work/secrets/**"]},
+  "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 Goal state", "durable Work Graph scheduling", "independent verification result"], "may_change": ["typed solution-search decision contract", "Work Graph orchestration decision endpoint"]},
+  "semantic_owner": "Work solution orchestration policy",
+  "canonical_state": ["existing Work Goal and Work Graph records; caller persists the returned wave decision"],
+  "state_transitions": ["SEARCHING -> EXECUTING -> VERIFYING -> COMPLETED", "SEARCHING -> WAITING -> SEARCHING", "any unresolved state -> next durable search wave", "any active state -> CANCELLED"],
+  "side_effects_and_privileges": ["pure policy decision and unauthenticated internal runtime route only; no provider call, secret access, code mutation, promotion or deployment"],
+  "persistence_and_migrations": ["none; this stage returns a typed decision for the existing durable scheduler"],
+  "failure_retry_restart": ["bounded wave exhaustion becomes WAITING with a new wave number and human-attention avenue; it is never represented as failure; cancellation and independently verified completion are terminal"],
+  "assumptions": ["the existing Work Graph dispatcher or next integration stage will persist the decision and fan out options", "model quality and provider routing remain separate capability decisions"],
+  "acceptance_scenarios": ["first unresolved goal selects all independent search avenues", "active wave remains EXECUTING", "exhausted wave continues with a higher durable round and human attention", "only cancellation or independently verified completion stops search", "invalid input is rejected", "the decision endpoint does not mutate source, secrets or production"],
+  "required_tests": ["go test ./modules/work/domain ./modules/work/application ./cmd/work-graph-runtime/...", "go vet ./modules/work/domain ./modules/work/application ./cmd/work-graph-runtime/...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production health smoke"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000232-durable-solution-wave-dispatch.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000232",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "89694047b1e6c7729b00c20a52acb5a2487e6dd6",
+  "risk_tier": "E",
+  "intent": "Connect the solution-search policy to the existing durable Work Graph scheduler so each unresolved wave is materialized as parallel dispatch intents and queued nodes.",
+  "requirement_ids": ["durable solution waves", "parallel fan-out", "agent/reflex orchestration", "restartable scheduling"],
+  "allowed_change_scope": {"modules": ["work-graph", "work-domain"], "paths": ["contracts/work/orchestration.go", "modules/work/domain/orchestration.go", "modules/work/domain/orchestration_test.go", "modules/work/application/orchestration.go", "cmd/work-graph-runtime/main.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000232-durable-solution-wave-dispatch.json", "docs/PROJECT_STATUS.md"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "LLM-local/**", "modules/work/secrets/**"]},
+  "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 solution-search decision", "existing Work Graph and dispatch queue ports"], "may_change": ["solution-search wave materialization request/result", "Work Graph orchestration wave route"]},
+  "semantic_owner": "Work solution orchestration and Work Graph",
+  "canonical_state": ["existing work.graphs, graph_nodes, dispatch_intents and dispatch_queue stores"],
+  "state_transitions": ["policy SEARCHING/WAITING -> graph ADMITTED/RUNNING", "each option -> one queued dispatch node", "human attention remains decision metadata"],
+  "side_effects_and_privileges": ["durable Work-owned graph and queue writes only; no provider call, secret access, external mutation, promotion or deployment"],
+  "persistence_and_migrations": ["none; existing graph and queue tables are reused"],
+  "failure_retry_restart": ["existing queue idempotency and graph revision persistence govern restart; invalid wave shape is rejected; no unresolved goal is marked failed by wave exhaustion"],
+  "assumptions": ["dispatch intents are supplied by the upstream authority-aware planner", "provider/model selection and execution remain worker responsibilities"],
+  "acceptance_scenarios": ["four independent solution avenues materialize as four graph nodes", "all ready nodes are queued up to graph MaxParallel", "human attention is not silently executed as an agent", "terminal policy decisions do not create a graph", "invalid or duplicate options are rejected"],
+  "required_tests": ["go test ./modules/work/domain ./modules/work/application ./cmd/work-graph-runtime/...", "go vet ./modules/work/domain ./modules/work/application ./cmd/work-graph-runtime/...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production health smoke"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000233-graph-result-fanin-verification.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000233",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "ee16103e03a5fd911119e2bc45e5904255ab7752",
+  "risk_tier": "E",
+  "intent": "Close the multi-agent Work Graph result lifecycle: persist attributable agent candidates, prevent candidate text from completing a node, and complete only after independently verified synthesis.",
+  "requirement_ids": ["fan-in", "independent verifier", "result provenance", "no false completion"],
+  "allowed_change_scope": {"modules": ["work-graph", "agent-execution", "postgres-work-adapter"], "paths": ["contracts/work/graph.go", "modules/work/ports/graph.go", "modules/work/application/graph_runtime.go", "adapters/postgres/work_repository.go", "cmd/work-graph-runtime/main.go", "tests/agent_execution/graph_runtime_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000233-graph-result-fanin-verification.json", "docs/PROJECT_STATUS.md"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "LLM-local/**", "modules/work/secrets/**"]},
+  "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 Work Graph revision store", "existing graph result candidate/synthesis tables", "agent callback evidence"], "may_change": ["typed graph result callback", "Work result callback route", "verified synthesis completion path"]},
+  "semantic_owner": "Work Graph result and verification lifecycle",
+  "canonical_state": ["work.graphs, work.graph_nodes, work.graph_result_candidates, work.graph_conflicts and work.graph_synthesis_candidates"],
+  "state_transitions": ["agent result -> durable ResultCandidate", "candidate -> no completion", "independent verified synthesis -> graph node SUCCEEDED", "all verified nodes -> graph COMPLETED"],
+  "side_effects_and_privileges": ["Work-owned candidate/graph persistence only; no provider call, secret access, external mutation, promotion or deployment"],
+  "persistence_and_migrations": ["none; existing result candidate and synthesis tables are reused"],
+  "failure_retry_restart": ["immutable graph revisions remain attributable; duplicate candidate IDs are idempotent at storage; missing or unverified synthesis cannot mark success"],
+  "assumptions": ["the verifier supplies all competing candidates and evidence in the synthesis request", "automatic replan selection remains the next stage because a graph does not yet own a persisted goal-search snapshot"],
+  "acceptance_scenarios": ["result callback persists candidate", "result callback leaves graph running and not completed", "verified independent synthesis advances the node", "verifier cannot be a producer", "fan-in completion requires all graph nodes successful"],
+  "required_tests": ["go test ./modules/work/domain ./modules/work/application ./tests/agent_execution ./cmd/work-graph-runtime/...", "go vet ./modules/work/domain ./modules/work/application ./tests/agent_execution ./cmd/work-graph-runtime/...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production health smoke"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000234-orchestration-snapshot-replan.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000234",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "eac39a3fd40e9f8d0752af6f9e7893cb0b03055b",
+  "risk_tier": "E",
+  "intent": "Persist the solution-search snapshot with every orchestration Work Graph revision and return the next durable search decision after verified graph completion.",
+  "requirement_ids": ["durable goal continuity", "automatic replan signal", "restart recovery", "no invented execution authority"],
+  "allowed_change_scope": {"modules": ["work-graph", "postgres-work-adapter"], "paths": ["contracts/work/graph.go", "modules/work/domain/orchestration.go", "modules/work/application/graph_runtime.go", "adapters/postgres/work_repository.go", "database/migrations/044_work_graph_orchestration_snapshot.sql", "cmd/work-graph-runtime/main.go", "tests/agent_execution/graph_runtime_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000234-orchestration-snapshot-replan.json", "docs/PROJECT_STATUS.md"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "deploy/**", "LLM-local/**", "modules/work/secrets/**"]},
+  "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": ["Work Graph immutable revision store", "SolutionSearchDecision policy"], "may_change": ["orchestration snapshot fields on WorkGraph", "graph completion result with next search decision"]},
+  "semantic_owner": "Work Graph orchestration continuity",
+  "canonical_state": ["work.graphs revision snapshot including GoalID and search round"],
+  "state_transitions": ["wave materialization persists search snapshot", "verified graph completion -> next search decision", "next decision remains nonterminal when unresolved"],
+  "side_effects_and_privileges": ["Work-owned graph persistence and queue orchestration only; no model/provider call, secret access, external mutation or promotion"],
+  "persistence_and_migrations": ["migration 044 adds nullable-by-contract orchestration snapshot columns with backward-compatible defaults"],
+  "failure_retry_restart": ["restarting from any immutable graph revision reconstructs the same search input; missing snapshot does not invent a replan"],
+  "assumptions": ["dispatch-intent discovery/materialization is supplied by the next wave planner; this stage returns the durable hand-off rather than fabricating agent authority"],
+  "acceptance_scenarios": ["new solution wave persists goal and round", "verified completion returns next round decision", "cancelled or verified terminal goal does not replan", "legacy graph without snapshot completes without invented replan"],
+  "required_tests": ["go test ./modules/work/domain ./modules/work/application ./tests/agent_execution ./cmd/work-graph-runtime/...", "go vet ./modules/work/domain ./modules/work/application ./tests/agent_execution ./cmd/work-graph-runtime/...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production health smoke"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP", "REQUIRES_MIGRATION_DECISION"]
+}

dev/tasks/DEV-000235-model-qualification-probe.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000235",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "60e8292c1df947b887079413b3b35ce8b4212086",
+  "risk_tier": "E",
+  "intent": "Add a bounded provider-neutral model qualification probe that discovers actual model IDs, invokes them in parallel, records reachability and latency, and never overclaims quality or hardware superiority.",
+  "requirement_ids": ["model fleet/failover", "dynamic model discovery", "quality separated from reachability", "parallel qualification"],
+  "allowed_change_scope": {"modules": ["cognitive-worker", "capability-observation"], "paths": ["cmd/cognitive-worker/main.go", "cmd/cognitive-worker/model_qualification.go", "cmd/cognitive-worker/model_qualification_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000235-model-qualification-probe.json", "docs/PROJECT_STATUS.md"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "LLM-local/**", "modules/work/secrets/**"]},
+  "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": ["OpenAI-compatible LocalProvider discovery/advisory adapter", "Capability model rating categories"], "may_change": ["internal bounded provider qualification route"]},
+  "semantic_owner": "Capability cognitive runtime observation",
+  "canonical_state": ["none; probe response is an operational observation and does not promote a binding"],
+  "state_transitions": ["DISCOVERED model -> bounded probe -> REACHABLE/UNAVAILABLE/RESPONDED_WITHOUT_QUALIFICATION", "quality remains UNRATED until independent task tests"],
+  "side_effects_and_privileges": ["read-only provider model discovery and bounded chat invocation; no external mutation, secret exposure, activation, promotion or routing priority"],
+  "persistence_and_migrations": ["none; existing discovery/observation and promotion flows remain unchanged"],
+  "failure_retry_restart": ["each probe has provider adapter timeout; one model failure does not cancel sibling probes; rerun is bounded and idempotent in effect"],
+  "assumptions": ["provider credentials are resolved inside cognitive-worker protected runtime", "additional task-specific tests can be added later without changing the primary probe"],
+  "acceptance_scenarios": ["actual discovered model IDs are used instead of static placeholders", "models are probed in parallel", "successful reachability reports latency", "quality is explicitly UNRATED", "GPU presence does not alter category or promotion", "output bound and model selection are constrained"],
+  "required_tests": ["go test ./cmd/cognitive-worker/... ./modules/capability/domain ./adapters/models/...", "go vet ./cmd/cognitive-worker/... ./modules/capability/domain ./adapters/models/...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production health smoke"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000236-qualification-locale-fix.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000236",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "risk_tier": "E",
+  "base_repo_revision": "b20611590a075e3574efddbdba6e1ace312d11a8",
+  "intent": "Make the bounded model qualification invocation satisfy the canonical advisory contract and keep a regression test for the required locale.",
+  "requirement_ids": ["model qualification smoke", "advisory contract"],
+  "allowed_change_scope": {"modules": ["cognitive-worker"], "paths": ["cmd/cognitive-worker/model_qualification.go", "cmd/cognitive-worker/model_qualification_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000236-qualification-locale-fix.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "LLM-local/**", "modules/work/secrets/**", "contracts/**"]},
+  "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 request validation"], "may_change": []},
+  "semantic_owner": "Cognitive worker qualification observation",
+  "canonical_state": ["none; qualification remains an operational observation"],
+  "state_transitions": ["invalid advisory invocation -> valid bounded invocation"],
+  "side_effects_and_privileges": ["read-only bounded provider invocation; no promotion or routing mutation"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["one bounded probe failure remains isolated"],
+  "assumptions": ["provider credentials remain protected inside cognitive-worker"],
+  "acceptance_scenarios": ["live qualification invocation passes advisory validation", "quality remains UNRATED"],
+  "required_tests": ["go test ./cmd/cognitive-worker/...", "go vet ./cmd/cognitive-worker/...", "preflight", "check-diff", "check-architecture", "independent", "git diff --check", "CI", "production health smoke"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000237-ui-authenticated-section-smoke.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000237",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "be01b6b883c2ce6e9bb47bc851505fe222c46d14",
+  "risk_tier": "E",
+  "intent": "Record current authenticated quick acceptance for the five creator-console sections without mutating production state.",
+  "requirement_ids": ["WP-800", "WP-804", "authenticated UI quick acceptance"],
+  "allowed_change_scope": {"modules": ["creator-web", "evidence"], "paths": ["docs/evidence/authenticated-ui-section-smoke-2026-08-23-v1.json", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000237-ui-authenticated-section-smoke.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "LLM-local/**", "modules/work/secrets/**", "contracts/**"]},
+  "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 creator-console UI"], "may_change": []},
+  "semantic_owner": "Creator-web acceptance evidence",
+  "canonical_state": ["evidence only; no runtime state mutation"],
+  "state_transitions": ["unverified current section -> observed authenticated quick acceptance"],
+  "side_effects_and_privileges": ["read-only browser inspection; no mutation"],
+  "persistence_and_migrations": ["evidence artifact only"],
+  "failure_retry_restart": ["recheck is idempotent and does not change product state"],
+  "assumptions": ["the authenticated browser session belongs to the authorized creator"],
+  "acceptance_scenarios": ["five admin sections load authenticated", "no visible section error is present", "full mobile/accessibility/chaos acceptance remains explicitly unclaimed"],
+  "required_tests": ["go test ./cmd/creator-web/...", "go vet ./cmd/creator-web/...", "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", "git diff --check", "independent", "CI", "production health smoke"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000238-local-fallback-evidence.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000238",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "0609561dfefb6dce302f56d50bb56bd5a222fb64",
+  "risk_tier": "E",
+  "intent": "Record a bounded live proof that the advisory gateway keeps a local lifeboat path when the normal model route cannot be used.",
+  "requirement_ids": ["Agent Execution local recovery", "model fleet fallback", "false completion prevention"],
+  "allowed_change_scope": {"modules": ["evidence"], "paths": ["docs/evidence/local-fallback-gateway-live-2026-08-23-v1.json", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000238-local-fallback-evidence.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "LLM-local/**", "modules/work/secrets/**", "contracts/**"]},
+  "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 gateway", "durable cognitive usage observation"], "may_change": []},
+  "semantic_owner": "Evidence only",
+  "canonical_state": ["no runtime state mutation; read-only evidence capture"],
+  "state_transitions": ["normal route unavailable or not eligible -> builtin local lifeboat response"],
+  "side_effects_and_privileges": ["bounded local advisory request; no external egress and no secret exposure"],
+  "persistence_and_migrations": ["existing cognitive usage observation only"],
+  "failure_retry_restart": ["live request is bounded and idempotent in effect"],
+  "assumptions": ["the production database observation is the authoritative route attribution"],
+  "acceptance_scenarios": ["local response is COMPLETED", "usage evidence identifies builtin-local-lifeboat-v1", "no claim of external bridge or multi-model failover"],
+  "required_tests": ["go test ./cmd/core-runtime/...", "go vet ./cmd/core-runtime/...", "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", "git diff --check", "independent", "CI", "production health smoke"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000239-stable-provider-resolution.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000239",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "6061eb5c13e46cec1e60fc104b022394f0d94dee",
+  "risk_tier": "E",
+  "intent": "Resolve a registered provider by its stable instance identity when a persisted Capability revision contains stale non-secret endpoint credential metadata, while preserving exact metadata matching first.",
+  "requirement_ids": ["model fleet/failover", "credential rotation compatibility", "local fallback attribution"],
+  "allowed_change_scope": {"modules": ["cognitive-worker"], "paths": ["cmd/cognitive-worker/gateway.go", "cmd/cognitive-worker/gateway_registry_test.go", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000239-stable-provider-resolution.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "LLM-local/**", "modules/work/secrets/**", "contracts/**"]},
+  "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 ResolvedCognitiveBinding", "protected provider registration"], "may_change": ["internal provider lookup fallback order"]},
+  "semantic_owner": "Cognitive worker provider adapter boundary",
+  "canonical_state": ["no new durable state"],
+  "state_transitions": ["exact provider metadata match -> stable provider identity match -> no provider"],
+  "side_effects_and_privileges": ["provider lookup only; secret values remain in protected runtime"],
+  "persistence_and_migrations": ["none"],
+  "failure_retry_restart": ["missing provider remains explicit and uses existing builtin fallback"],
+  "assumptions": ["ProviderInstanceID is the stable Capability-owned provider identity"],
+  "acceptance_scenarios": ["stale credential metadata still reaches current registered adapter", "exact metadata match retains priority", "no provider remains nil"],
+  "required_tests": ["go test ./cmd/cognitive-worker/...", "go vet ./cmd/cognitive-worker/...", "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", "git diff --check", "independent", "CI", "production health smoke"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}

dev/tasks/DEV-000240-cpu-lifeboat-failover-evidence.json

@@ -0,0 +1,23 @@
+{
+  "task_id": "DEV-000240",
+  "task_revision": 1,
+  "architecture_revision": "ADR-0018",
+  "base_repo_revision": "383d4c6e4fd53bb0c34ed358e0d1be6a416da6cc",
+  "risk_tier": "E",
+  "intent": "Record live durable evidence of CPU model failure followed by successful local Lifeboat failover after stable provider resolution was deployed.",
+  "requirement_ids": ["model fleet/failover", "durable usage evidence", "false completion prevention"],
+  "allowed_change_scope": {"modules": ["evidence"], "paths": ["docs/evidence/cpu-lifeboat-failover-2026-08-23-v1.json", "dev/ACTIVE_TASK.json", "dev/tasks/DEV-000240-cpu-lifeboat-failover-evidence.json"]},
+  "forbidden_change_scope": {"paths": ["architecture/**", "security/**", "database/**", "deploy/**", "LLM-local/**", "modules/work/secrets/**", "contracts/**"]},
+  "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": ["cognitive advisory gateway", "cognitive usage observations"], "may_change": []},
+  "semantic_owner": "Evidence only",
+  "canonical_state": ["no runtime state mutation; evidence captures existing durable observations"],
+  "state_transitions": ["CPU MODEL_UNAVAILABLE -> builtin local Lifeboat COMPLETED"],
+  "side_effects_and_privileges": ["bounded LOCAL_ONLY advisory request; no external egress"],
+  "persistence_and_migrations": ["existing usage observation persistence only"],
+  "failure_retry_restart": ["the fallback is selected within the bounded gateway request"],
+  "assumptions": ["provider_instance_id and result_class in capability usage observations are authoritative route attribution"],
+  "acceptance_scenarios": ["primary failure is durably observed", "fallback completion is durably observed", "HTTP gateway returns COMPLETED", "public health remains OK"],
+  "required_tests": ["go test ./cmd/cognitive-worker/...", "go vet ./cmd/cognitive-worker/...", "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", "git diff --check", "independent", "CI", "production health smoke"],
+  "terminal_gap_states": ["BLOCKED_BY_MISSING_CAPABILITY", "BLOCKED_BY_SPEC_GAP"]
+}