GEN2-022: govern self evolution through development gates

cmd/supervisor/main.go

@@ -0,0 +1,10 @@
+package main
+
+import "net/http"
+
+func main() {
+	_ = http.ListenAndServe(":8088", http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.WriteHeader(http.StatusOK)
+		_, _ = w.Write([]byte("lifeboat-ready\n"))
+	}))
+}

contracts/commit/execution_receipt.go

@@ -0,0 +1,32 @@
+package commit
+
+type ExecutorResult string
+type AmbiguityState string
+
+const (
+	ExecutorAccepted        ExecutorResult = "ACCEPTED"
+	ExecutorReportedSuccess ExecutorResult = "REPORTED_SUCCESS"
+	ExecutorReportedFailure ExecutorResult = "REPORTED_FAILURE"
+	ExecutorNoResponse      ExecutorResult = "NO_RESPONSE"
+	ExecutorAmbiguous       ExecutorResult = "AMBIGUOUS"
+)
+
+type ExecutionReceipt struct {
+	ID                        string
+	ActionAttemptRef          ActionAttemptID
+	ExecutorImplementationRef string
+	ExternalOperationRef      string
+	ExternalRequestID         string
+	RequestDigest             string
+	ResponseDigest            string
+	DispatchAtUnixMillis      int64
+	CompletedAtUnixMillis     *int64
+	ExecutorResult            ExecutorResult
+	Ambiguity                 AmbiguityState
+	ReconciliationKey         string
+	RawArtifactRef            string
+}
+
+func (r ExecutionReceipt) Valid() bool {
+	return r.ID != "" && r.ActionAttemptRef != "" && r.ExecutorImplementationRef != "" && r.RequestDigest != "" && r.ExecutorResult != "" && r.ReconciliationKey != "" && r.DispatchAtUnixMillis > 0
+}

contracts/epistemics/outcome.go

@@ -0,0 +1,30 @@
+package epistemics
+
+type WorldOutcomeState string
+
+const (
+	OutcomeVerifiedSuccess WorldOutcomeState = "VERIFIED_SUCCESS"
+	OutcomeVerifiedPartial WorldOutcomeState = "VERIFIED_PARTIAL"
+	OutcomeVerifiedFailure WorldOutcomeState = "VERIFIED_FAILURE"
+	OutcomeUnverified      WorldOutcomeState = "UNVERIFIED"
+	OutcomeUnknownState    WorldOutcomeState = "UNKNOWN"
+	OutcomeConflicted      WorldOutcomeState = "CONFLICTED"
+)
+
+type WorldOutcomeAssessment struct {
+	ID                   string
+	PrincipalID          string
+	ActionAttemptRef     string
+	ExecutionReceiptRefs []string
+	ExpectedEffectRef    string
+	ObservationRefs      []ObservationID
+	ClaimRefs            []ClaimID
+	State                WorldOutcomeState
+	ConfidenceMillis     int64
+	EvidenceRefs         []EvidenceRef
+	AssessedAt           int64
+}
+
+func (a WorldOutcomeAssessment) Valid() bool {
+	return a.ID != "" && a.PrincipalID != "" && a.ActionAttemptRef != "" && a.State != "" && a.ConfidenceMillis >= 0 && a.ConfidenceMillis <= 1000 && a.AssessedAt > 0
+}

contracts/work/consequence.go

@@ -0,0 +1,24 @@
+package work
+
+type ConsequenceAssessment struct {
+	ID                      string
+	TargetRefs              []string
+	Impact                  int64
+	Reversibility           int64
+	Uncertainty             int64
+	Scope                   int64
+	Recoverability          int64
+	CostOfError             int64
+	ExpectedConsequences    []string
+	MitigationOptions       []string
+	RequiresDryRun          bool
+	RequiresBackup          bool
+	RecommendedConfirmation string
+	OwnerAccepted           bool
+	EvidenceRefs            []EvidenceRef
+	RecordedAtUnixMillis    int64
+}
+
+func (a ConsequenceAssessment) Valid() bool {
+	return a.ID != "" && len(a.TargetRefs) > 0 && a.Impact >= 0 && a.Impact <= 1000 && a.Reversibility >= 0 && a.Reversibility <= 1000 && a.Uncertainty >= 0 && a.Uncertainty <= 1000 && a.Recoverability >= 0 && a.Recoverability <= 1000 && a.RecordedAtUnixMillis > 0
+}

contracts/work/long_running.go

@@ -0,0 +1,35 @@
+package work
+
+type LongRunningWorkID string
+type LongRunningState string
+
+const (
+	LongRunningProposed  LongRunningState = "PROPOSED"
+	LongRunningRunning   LongRunningState = "RUNNING"
+	LongRunningPaused    LongRunningState = "PAUSED"
+	LongRunningBlocked   LongRunningState = "BLOCKED"
+	LongRunningCompleted LongRunningState = "COMPLETED"
+	LongRunningExhausted LongRunningState = "EXHAUSTED"
+	LongRunningCancelled LongRunningState = "CANCELLED"
+)
+
+type Investigation struct {
+	ID                   LongRunningWorkID
+	Objective            string
+	HypothesisRefs       []EvidenceRef
+	AttemptRefs          []AgentRunID
+	FindingRefs          []EvidenceRef
+	Blockers             []string
+	CheckpointRefs       []AgentCheckpointID
+	NextStep             string
+	State                LongRunningState
+	Revision             int64
+	TimeBudgetMillis     int64
+	CostBudgetMicros     int64
+	MaxAttempts          int
+	RecordedAtUnixMillis int64
+}
+
+func (w Investigation) Valid() bool {
+	return w.ID != "" && w.Objective != "" && w.State != "" && w.Revision > 0 && w.TimeBudgetMillis >= 0 && w.CostBudgetMicros >= 0 && w.MaxAttempts >= 0 && w.RecordedAtUnixMillis > 0
+}

contracts/work/perspective.go

@@ -0,0 +1,26 @@
+package work
+
+type PerspectiveID string
+type PerspectiveSpec struct {
+	ID                     PerspectiveID
+	ObjectiveLens          string
+	RetrievalProfileRef    string
+	EvaluationCriteria     []string
+	CapabilityRequirements []string
+	ExpectedOutputShape    string
+	Critic                 bool
+}
+type SpecialistAssignment struct {
+	Perspective  PerspectiveSpec
+	SubtaskRef   string
+	EvidenceRefs []EvidenceRef
+}
+type TeamDecision struct {
+	UseTeam     bool
+	Reason      string
+	Assignments []SpecialistAssignment
+}
+
+func (p PerspectiveSpec) Valid() bool {
+	return p.ID != "" && p.ObjectiveLens != "" && len(p.EvaluationCriteria) > 0 && p.ExpectedOutputShape != ""
+}

contracts/work/self_evolution.go

@@ -0,0 +1,36 @@
+package work
+
+type EvolutionChangeLevel string
+
+const (
+	EvolutionKnowledge  EvolutionChangeLevel = "KNOWLEDGE"
+	EvolutionBehavior   EvolutionChangeLevel = "BEHAVIOR"
+	EvolutionCapability EvolutionChangeLevel = "CAPABILITY"
+	EvolutionCore       EvolutionChangeLevel = "CORE"
+)
+
+type EvolutionCandidate struct {
+	ID                    string
+	ProposalRef           string
+	ChangeLevel           EvolutionChangeLevel
+	CreatorIdentity       string
+	BaseVersionRef        string
+	CandidateVersionRef   string
+	MustNotWorsen         []string
+	SandboxEvidenceRefs   []EvidenceRef
+	BenchmarkEvidenceRefs []EvidenceRef
+	RollbackRef           string
+	State                 string
+}
+type EvolutionEvaluation struct {
+	CandidateRef         string
+	EvaluatorIdentity    string
+	Approved             bool
+	EvidenceRefs         []EvidenceRef
+	MustNotWorsenPassed  bool
+	RecordedAtUnixMillis int64
+}
+
+func (e EvolutionEvaluation) ValidFor(c EvolutionCandidate) bool {
+	return e.CandidateRef == c.ID && e.EvaluatorIdentity != "" && e.EvaluatorIdentity != c.CreatorIdentity && e.Approved && e.MustNotWorsenPassed && len(e.EvidenceRefs) > 0 && e.RecordedAtUnixMillis > 0
+}

dev/ACTIVE_TASK.json

@@ -1 +1 @@
-{"task_file":"dev/tasks/GEN2-016-intelligence-strategy-resource-cost-policy-and-personal-benchmark.json","task_id":"GEN2-016","task_revision":1}
+{"task_file":"dev/tasks/GEN2-022-wire-governed-self-evolution-loop-using-existing-development-gates.json","task_id":"GEN2-022","task_revision":1}

modules/work/domain/consequence.go

@@ -0,0 +1,14 @@
+package domain
+
+import "git.cin.su/m/Assistant/contracts/work"
+
+func AssessConsequences(id string, target string, impact, reversibility, uncertainty, recoverability, cost int64, at int64) work.ConsequenceAssessment {
+	a := work.ConsequenceAssessment{ID: id, TargetRefs: []string{target}, Impact: impact, Reversibility: reversibility, Uncertainty: uncertainty, Recoverability: recoverability, CostOfError: cost, RecordedAtUnixMillis: at}
+	a.RequiresDryRun = impact >= 700 || reversibility >= 700
+	a.RequiresBackup = reversibility >= 500
+	a.RecommendedConfirmation = "STANDARD"
+	if uncertainty*cost > 500000 {
+		a.RecommendedConfirmation = "EXPLICIT_OWNER_CONFIRMATION"
+	}
+	return a
+}

modules/work/domain/consequence_test.go

@@ -0,0 +1,13 @@
+package domain_test
+
+import (
+	"git.cin.su/m/Assistant/modules/work/domain"
+	"testing"
+)
+
+func TestConsequenceAdvisesMitigationWithoutCapabilityBlacklist(t *testing.T) {
+	a := domain.AssessConsequences("a", "database:a", 800, 800, 900, 300, 900, 1)
+	if !a.Valid() || !a.RequiresDryRun || !a.RequiresBackup || a.RecommendedConfirmation != "EXPLICIT_OWNER_CONFIRMATION" {
+		t.Fatalf("%#v", a)
+	}
+}

modules/work/domain/long_running.go

@@ -0,0 +1,18 @@
+package domain
+
+import "git.cin.su/m/Assistant/contracts/work"
+
+func PauseInvestigation(value work.Investigation) work.Investigation {
+	if value.Valid() {
+		value.State = work.LongRunningPaused
+		value.Revision++
+	}
+	return value
+}
+func ResumeInvestigation(value work.Investigation) work.Investigation {
+	if value.Valid() && value.State == work.LongRunningPaused {
+		value.State = work.LongRunningRunning
+		value.Revision++
+	}
+	return value
+}

modules/work/domain/long_running_test.go

@@ -0,0 +1,16 @@
+package domain_test
+
+import (
+	"git.cin.su/m/Assistant/contracts/work"
+	"git.cin.su/m/Assistant/modules/work/domain"
+	"testing"
+)
+
+func TestInvestigationPauseResumeIsDurableStateTransition(t *testing.T) {
+	w := work.Investigation{ID: "i", Objective: "diagnose", State: work.LongRunningRunning, Revision: 1, RecordedAtUnixMillis: 1}
+	w = domain.PauseInvestigation(w)
+	w = domain.ResumeInvestigation(w)
+	if w.State != work.LongRunningRunning || w.Revision != 3 {
+		t.Fatalf("%#v", w)
+	}
+}

modules/work/domain/perspective.go

@@ -0,0 +1,17 @@
+package domain
+
+import "git.cin.su/m/Assistant/contracts/work"
+
+func AssembleSpecialists(p work.PerspectiveSpec, subtasks []string) work.TeamDecision {
+	if !p.Valid() {
+		return work.TeamDecision{Reason: "invalid perspective"}
+	}
+	if len(subtasks) < 2 {
+		return work.TeamDecision{Reason: "single reasoning path is sufficient"}
+	}
+	assignments := make([]work.SpecialistAssignment, 0, len(subtasks))
+	for _, s := range subtasks {
+		assignments = append(assignments, work.SpecialistAssignment{Perspective: p, SubtaskRef: s})
+	}
+	return work.TeamDecision{UseTeam: true, Reason: "independent subtasks justify bounded perspectives", Assignments: assignments}
+}

modules/work/domain/perspective_test.go

@@ -0,0 +1,17 @@
+package domain_test
+
+import (
+	"git.cin.su/m/Assistant/contracts/work"
+	"git.cin.su/m/Assistant/modules/work/domain"
+	"testing"
+)
+
+func TestPerspectiveTeamIsBoundedByIndependentSubtasks(t *testing.T) {
+	p := work.PerspectiveSpec{ID: "architect", ObjectiveLens: "architecture", EvaluationCriteria: []string{"correctness"}, ExpectedOutputShape: "finding"}
+	if d := domain.AssembleSpecialists(p, []string{"a"}); d.UseTeam {
+		t.Fatal("team for one subtask")
+	}
+	if d := domain.AssembleSpecialists(p, []string{"a", "b"}); !d.UseTeam || len(d.Assignments) != 2 {
+		t.Fatalf("%#v", d)
+	}
+}

modules/work/domain/self_evolution.go

@@ -0,0 +1,7 @@
+package domain
+
+import "git.cin.su/m/Assistant/contracts/work"
+
+func PromoteEvolution(c work.EvolutionCandidate, e work.EvolutionEvaluation) bool {
+	return e.ValidFor(c) && c.ProposalRef != "" && c.RollbackRef != "" && len(c.SandboxEvidenceRefs) > 0 && len(c.BenchmarkEvidenceRefs) > 0
+}

modules/work/domain/self_evolution_test.go

@@ -0,0 +1,19 @@
+package domain_test
+
+import (
+	"git.cin.su/m/Assistant/contracts/work"
+	"git.cin.su/m/Assistant/modules/work/domain"
+	"testing"
+)
+
+func TestEvolutionRequiresIndependentEvaluatorAndRollbackEvidence(t *testing.T) {
+	c := work.EvolutionCandidate{ID: "c", ProposalRef: "p", ChangeLevel: work.EvolutionBehavior, CreatorIdentity: "creator", RollbackRef: "rollback", SandboxEvidenceRefs: []work.EvidenceRef{"sandbox"}, BenchmarkEvidenceRefs: []work.EvidenceRef{"benchmark"}}
+	e := work.EvolutionEvaluation{CandidateRef: "c", EvaluatorIdentity: "creator", Approved: true, MustNotWorsenPassed: true, EvidenceRefs: []work.EvidenceRef{"eval"}, RecordedAtUnixMillis: 1}
+	if domain.PromoteEvolution(c, e) {
+		t.Fatal("creator promoted own change")
+	}
+	e.EvaluatorIdentity = "independent"
+	if !domain.PromoteEvolution(c, e) {
+		t.Fatal("independently evaluated candidate rejected")
+	}
+}

runtime/supervisor/supervisor.go

@@ -0,0 +1,62 @@
+package supervisor
+
+type ReleaseState string
+
+const (
+	ReleaseCandidate  ReleaseState = "CANDIDATE"
+	ReleaseHealthy    ReleaseState = "HEALTHY"
+	ReleaseFailed     ReleaseState = "FAILED"
+	ReleaseRolledBack ReleaseState = "ROLLED_BACK"
+)
+
+type Release struct {
+	ID            string
+	Version       string
+	State         ReleaseState
+	LastKnownGood bool
+}
+type Observation struct {
+	ReleaseID            string
+	Ready                bool
+	Reason               string
+	ObservedAtUnixMillis int64
+}
+type Lifeboat struct {
+	Current       Release
+	LastKnownGood Release
+	Observations  []Observation
+}
+
+func (l *Lifeboat) Start(candidate Release) bool {
+	if candidate.ID == "" || candidate.Version == "" {
+		return false
+	}
+	candidate.State = ReleaseCandidate
+	l.Current = candidate
+	return true
+}
+func (l *Lifeboat) Record(o Observation) bool {
+	if o.ReleaseID == "" {
+		return false
+	}
+	l.Observations = append(l.Observations, o)
+	if o.ReleaseID != l.Current.ID {
+		return true
+	}
+	if o.Ready {
+		l.Current.State = ReleaseHealthy
+		l.Current.LastKnownGood = true
+		l.LastKnownGood = l.Current
+	} else {
+		l.Current.State = ReleaseFailed
+	}
+	return true
+}
+func (l *Lifeboat) Rollback() bool {
+	if l.LastKnownGood.ID == "" {
+		return false
+	}
+	l.Current = l.LastKnownGood
+	l.Current.State = ReleaseRolledBack
+	return true
+}

runtime/supervisor/supervisor_test.go

@@ -0,0 +1,17 @@
+package supervisor_test
+
+import (
+	"git.cin.su/m/Assistant/runtime/supervisor"
+	"testing"
+)
+
+func TestLifeboatRollsBackCandidateToLastKnownGood(t *testing.T) {
+	l := supervisor.Lifeboat{}
+	l.Start(supervisor.Release{ID: "good", Version: "1"})
+	l.Record(supervisor.Observation{ReleaseID: "good", Ready: true})
+	l.Start(supervisor.Release{ID: "candidate", Version: "2"})
+	l.Record(supervisor.Observation{ReleaseID: "candidate", Ready: false})
+	if !l.Rollback() || l.Current.ID != "good" || l.Current.State != supervisor.ReleaseRolledBack {
+		t.Fatalf("%#v", l)
+	}
+}

tests/execution_outcome/outcome_test.go

@@ -0,0 +1,18 @@
+package execution_outcome_test
+
+import (
+	"git.cin.su/m/Assistant/contracts/commit"
+	"git.cin.su/m/Assistant/contracts/epistemics"
+	"testing"
+)
+
+func TestReportedExecutionSuccessIsNotWorldVerification(t *testing.T) {
+	r := commit.ExecutionReceipt{ID: "r", ActionAttemptRef: "a", ExecutorImplementationRef: "impl", RequestDigest: "d", ExecutorResult: commit.ExecutorReportedSuccess, ReconciliationKey: "k", DispatchAtUnixMillis: 1}
+	if !r.Valid() {
+		t.Fatal("receipt invalid")
+	}
+	o := epistemics.WorldOutcomeAssessment{ID: "o", PrincipalID: "p", ActionAttemptRef: "a", State: epistemics.OutcomeUnverified, ConfidenceMillis: 0, AssessedAt: 1}
+	if !o.Valid() {
+		t.Fatal("unverified assessment invalid")
+	}
+}