diff --git a/cmd/chat_config_models.go b/cmd/chat_config_models.go
index 426e5e57..0d3f3000 100644
--- a/cmd/chat_config_models.go
+++ b/cmd/chat_config_models.go
@@ -5,6 +5,8 @@ import (
 	"strings"
 	"sync"
 
+	"github.com/GrayCodeAI/rho/internal/engine/cost"
+
 	tea "charm.land/bubbletea/v2"
 
 	rhoconfig "github.com/GrayCodeAI/rho/internal/config"
@@ -214,7 +216,7 @@ func applyModelOptionToSession(sess *engine.Session, opt configModelOption) {
 		sess.EnsureAutoCompactor()
 	}
 	if opt.PriceKnown {
-		engine.RegisterLivePricing(opt.ID, opt.InputPricePer1M, opt.OutputPricePer1M)
+		cost.RegisterLivePricing(opt.ID, opt.InputPricePer1M, opt.OutputPricePer1M)
 	}
 }
 
diff --git a/cmd/compact_ui.go b/cmd/compact_ui.go
index a8585227..5af4363c 100644
--- a/cmd/compact_ui.go
+++ b/cmd/compact_ui.go
@@ -7,6 +7,8 @@ import (
 	"strings"
 	"time"
 
+	"github.com/GrayCodeAI/rho/internal/engine/token"
+
 	tea "charm.land/bubbletea/v2"
 	lipgloss "charm.land/lipgloss/v2"
 	"github.com/mattn/go-runewidth"
@@ -174,7 +176,7 @@ func (m *chatModel) startManualCompact() (chatModel, tea.Cmd) {
 	m.compactBarWindow = m.contextWindowTokens()
 	m.compactBarUsed = sessionContextUsedTokens(m.session)
 	if m.compactBarUsed <= 0 && m.session != nil {
-		m.compactBarUsed = engine.EstimateTokens(m.session.RawMessages())
+		m.compactBarUsed = token.EstimateTokens(m.session.RawMessages())
 	}
 	if m.brailleSpinner != nil {
 		m.brailleSpinner.SetLabel("")
diff --git a/cmd/options.go b/cmd/options.go
index 66dc44bb..3dac14d0 100644
--- a/cmd/options.go
+++ b/cmd/options.go
@@ -9,6 +9,8 @@ import (
 	"os"
 	"strings"
 
+	"github.com/GrayCodeAI/rho/internal/engine/scaffold"
+
 	rhoconfig "github.com/GrayCodeAI/rho/internal/config"
 	ctxrepomap "github.com/GrayCodeAI/rho/internal/context/repomap"
 	"github.com/GrayCodeAI/rho/internal/engine"
@@ -380,7 +382,7 @@ func configureSessionStartup(sess *engine.Session, settings rhoconfig.Settings,
 	sess.LifecycleSvc().SetReflector(engine.NewReflector(sess, sess.Model()))
 
 	// Few-shot learning: collect successful patterns from sessions
-	sess.LifecycleSvc().SetFewShotStore(engine.NewFewShotStore())
+	sess.LifecycleSvc().SetFewShotStore(scaffold.NewFewShotStore())
 
 	// Adaptive prompt: learn user preferences from corrections
 	sess.LifecycleSvc().SetAdaptivePrompt(engine.NewAdaptivePrompt())
diff --git a/cmd/statusbar.go b/cmd/statusbar.go
index a0059765..f7cbc6d4 100644
--- a/cmd/statusbar.go
+++ b/cmd/statusbar.go
@@ -7,6 +7,8 @@ import (
 	"sync"
 	"time"
 
+	"github.com/GrayCodeAI/rho/internal/engine/cost"
+
 	tea "charm.land/bubbletea/v2"
 	lipgloss "charm.land/lipgloss/v2"
 	"golang.org/x/text/language"
@@ -154,7 +156,7 @@ func renderStatusBarPrimaryRight(m *chatModel) string {
 	return strings.Join(parts, statusDimStyle.Render(" · "))
 }
 
-func formatStatusCost(c *engine.Cost) string {
+func formatStatusCost(c *cost.Cost) string {
 	if c == nil {
 		return icons.Ruby() + " $0.00"
 	}
diff --git a/internal/engine/agent_intelligence.go b/internal/engine/agent_intelligence.go
index c783e613..bee371c0 100644
--- a/internal/engine/agent_intelligence.go
+++ b/internal/engine/agent_intelligence.go
@@ -3,6 +3,8 @@ package engine
 import (
 	"context"
 	"strings"
+
+	"github.com/GrayCodeAI/rho/internal/engine/agent"
 )
 
 // AgentIntelligence provides smart routing, auto-spawning, and synthesis for agents.
@@ -26,7 +28,7 @@ type SpawnDecision struct {
 type SubTask struct {
 	ID        string
 	Prompt    string
-	Mode      SubAgentMode
+	Mode      agent.SubAgentMode
 	Priority  int      // higher = run first
 	DependsOn []string // IDs of tasks this depends on
 }
@@ -78,19 +80,19 @@ func (ai *AgentIntelligence) AnalyzeForParallelism(prompt string) SpawnDecision
 }
 
 // SelectMode picks the optimal agent mode for a subtask.
-func (ai *AgentIntelligence) SelectMode(subtask string) SubAgentMode {
+func (ai *AgentIntelligence) SelectMode(subtask string) agent.SubAgentMode {
 	lower := strings.ToLower(subtask)
 
 	// Read-only tasks → explore mode (cheaper, faster)
 	readOnlyKeywords := []string{"find", "search", "list", "check", "read", "analyze", "look", "scan", "grep", "what is", "where is", "how many"}
 	for _, kw := range readOnlyKeywords {
 		if strings.Contains(lower, kw) {
-			return SubAgentExplore
+			return agent.SubAgentExplore
 		}
 	}
 
 	// Write tasks → general mode
-	return SubAgentGeneral
+	return agent.SubAgentGeneral
 }
 
 // decomposeTask splits a complex task into subtasks based on patterns.
@@ -101,17 +103,17 @@ func (ai *AgentIntelligence) decomposeTask(prompt string, scale TaskScale) []Sub
 	if (strings.Contains(lower, "research") || strings.Contains(lower, "analyze")) &&
 		(strings.Contains(lower, "implement") || strings.Contains(lower, "build") || strings.Contains(lower, "create")) {
 		return []SubTask{
-			{ID: "research", Prompt: "Research and analyze: " + prompt, Mode: SubAgentExplore},
-			{ID: "implement", Prompt: "Based on research, implement: " + prompt, Mode: SubAgentGeneral, DependsOn: []string{"research"}},
+			{ID: "research", Prompt: "Research and analyze: " + prompt, Mode: agent.SubAgentExplore},
+			{ID: "implement", Prompt: "Based on research, implement: " + prompt, Mode: agent.SubAgentGeneral, DependsOn: []string{"research"}},
 		}
 	}
 
 	// Pattern: multi-file refactor — pipeline
 	if strings.Contains(lower, "refactor") && scale >= ScaleMajor {
 		return []SubTask{
-			{ID: "scan", Prompt: "Scan and identify all files that need changes for: " + prompt, Mode: SubAgentExplore},
-			{ID: "plan", Prompt: "Create a refactoring plan based on scan results: " + prompt, Mode: SubAgentExplore, DependsOn: []string{"scan"}},
-			{ID: "execute", Prompt: "Execute the refactoring plan: " + prompt, Mode: SubAgentGeneral, DependsOn: []string{"plan"}},
+			{ID: "scan", Prompt: "Scan and identify all files that need changes for: " + prompt, Mode: agent.SubAgentExplore},
+			{ID: "plan", Prompt: "Create a refactoring plan based on scan results: " + prompt, Mode: agent.SubAgentExplore, DependsOn: []string{"scan"}},
+			{ID: "execute", Prompt: "Execute the refactoring plan: " + prompt, Mode: agent.SubAgentGeneral, DependsOn: []string{"plan"}},
 		}
 	}
 
@@ -185,7 +187,7 @@ func splitOnConjunctions(s string) []string {
 }
 
 // ExecuteWithIntelligence runs a task with smart agent routing.
-func (ai *AgentIntelligence) ExecuteWithIntelligence(ctx context.Context, prompt string, execFn func(context.Context, string, SubAgentMode) (string, error)) (string, error) {
+func (ai *AgentIntelligence) ExecuteWithIntelligence(ctx context.Context, prompt string, execFn func(context.Context, string, agent.SubAgentMode) (string, error)) (string, error) {
 	decision := ai.AnalyzeForParallelism(prompt)
 
 	if !decision.ShouldParallelize {
diff --git a/internal/engine/agent_intelligence_test.go b/internal/engine/agent_intelligence_test.go
index 426d25b3..4db5e548 100644
--- a/internal/engine/agent_intelligence_test.go
+++ b/internal/engine/agent_intelligence_test.go
@@ -3,6 +3,8 @@ package engine
 import (
 	"context"
 	"testing"
+
+	"github.com/GrayCodeAI/rho/internal/engine/agent"
 )
 
 func TestAgentIntelligence_SelectMode(t *testing.T) {
@@ -10,14 +12,14 @@ func TestAgentIntelligence_SelectMode(t *testing.T) {
 
 	tests := []struct {
 		prompt string
-		want   SubAgentMode
+		want   agent.SubAgentMode
 	}{
-		{"find all files using deprecated API", SubAgentExplore},
-		{"search for authentication code", SubAgentExplore},
-		{"analyze the dependency graph", SubAgentExplore},
-		{"implement the new endpoint", SubAgentGeneral},
-		{"refactor the auth module", SubAgentGeneral},
-		{"fix the bug in parser.go", SubAgentGeneral},
+		{"find all files using deprecated API", agent.SubAgentExplore},
+		{"search for authentication code", agent.SubAgentExplore},
+		{"analyze the dependency graph", agent.SubAgentExplore},
+		{"implement the new endpoint", agent.SubAgentGeneral},
+		{"refactor the auth module", agent.SubAgentGeneral},
+		{"fix the bug in parser.go", agent.SubAgentGeneral},
 	}
 	for _, tt := range tests {
 		got := ai.SelectMode(tt.prompt)
@@ -67,7 +69,7 @@ func TestAgentIntelligence_ExecuteWithIntelligence_Single(t *testing.T) {
 	ai := NewAgentIntelligence()
 
 	called := false
-	result, err := ai.ExecuteWithIntelligence(context.Background(), "fix typo", func(_ context.Context, prompt string, mode SubAgentMode) (string, error) {
+	result, err := ai.ExecuteWithIntelligence(context.Background(), "fix typo", func(_ context.Context, prompt string, mode agent.SubAgentMode) (string, error) {
 		called = true
 		return "fixed", nil
 	})
@@ -95,8 +97,8 @@ func TestSelfAwareness_ShouldDelegate(t *testing.T) {
 
 func TestSynthesisPrompt(t *testing.T) {
 	subtasks := []SubTask{
-		{ID: "a", Prompt: "find files", Mode: SubAgentExplore},
-		{ID: "b", Prompt: "fix them", Mode: SubAgentGeneral},
+		{ID: "a", Prompt: "find files", Mode: agent.SubAgentExplore},
+		{ID: "b", Prompt: "fix them", Mode: agent.SubAgentGeneral},
 	}
 	results := map[string]string{"a": "found 3 files", "b": "fixed all"}
 	prompt := MergeSynthesisPrompt(subtasks, results)
diff --git a/internal/engine/agent_reexports.go b/internal/engine/agent_reexports.go
deleted file mode 100644
index adee48e3..00000000
--- a/internal/engine/agent_reexports.go
+++ /dev/null
@@ -1,42 +0,0 @@
-// This file re-exports symbols from the agent sub-package so that existing
-// callers of engine.SubAgentMode, engine.NewSubAgentBudget, etc. keep compiling
-// during the Stage 2 migration. See docs/plans/engine-refactor-plan.md.
-package engine
-
-import (
-	"github.com/GrayCodeAI/rho/internal/engine/agent"
-)
-
-type (
-	SubAgentMode     = agent.SubAgentMode
-	SubAgentConfig   = agent.SubAgentConfig
-	SubAgentBudget   = agent.SubAgentBudget
-	BackgroundResult = agent.BackgroundResult
-)
-
-const (
-	SubAgentExplore     = agent.SubAgentExplore
-	SubAgentGeneral     = agent.SubAgentGeneral
-	SubAgentPlan        = agent.SubAgentPlan
-	DefaultExploreTurns = agent.DefaultExploreTurns
-	DefaultGeneralTurns = agent.DefaultGeneralTurns
-	DefaultPlanTurns    = agent.DefaultPlanTurns
-	MaxAgentDepth       = agent.MaxAgentDepth
-)
-
-var (
-	ExploreTools      = agent.ExploreTools
-	PlanTools         = agent.PlanTools
-	ModeToolAllowlist = agent.ModeToolAllowlist
-)
-
-func DefaultSubAgentConfig() SubAgentConfig { return agent.DefaultSubAgentConfig() }
-func NewSubAgentBudget(mode SubAgentMode, cfg SubAgentConfig) *SubAgentBudget {
-	return agent.NewSubAgentBudget(mode, cfg)
-}
-
-func FilterToolsForMode(mode SubAgentMode, available []string) []string {
-	return agent.FilterToolsForMode(mode, available)
-}
-func DefaultTurnsForMode(mode SubAgentMode) int { return agent.DefaultTurnsForMode(mode) }
-func IsReadOnlyMode(mode SubAgentMode) bool     { return agent.IsReadOnlyMode(mode) }
diff --git a/internal/engine/agent_session_tool.go b/internal/engine/agent_session_tool.go
index c1b0f13a..f74f6779 100644
--- a/internal/engine/agent_session_tool.go
+++ b/internal/engine/agent_session_tool.go
@@ -137,26 +137,26 @@ func (s *Session) spawnSubAgentRequest(ctx context.Context, req agentcontracts.S
 	return res, nil
 }
 
-func mapContractsType(t agentcontracts.SubagentType) SubAgentMode {
+func mapContractsType(t agentcontracts.SubagentType) engagent.SubAgentMode {
 	switch t {
 	case agentcontracts.TypePlan:
-		return SubAgentPlan
+		return engagent.SubAgentPlan
 	case agentcontracts.TypeGeneralPurpose:
-		return SubAgentGeneral
+		return engagent.SubAgentGeneral
 	default:
-		return SubAgentExplore
+		return engagent.SubAgentExplore
 	}
 }
 
 // spawnSubAgent creates a sub-agent with the given mode and depth tracking.
 // Returns (output, worktreePath, error).
-func (s *Session) spawnSubAgent(ctx context.Context, norm agentcontracts.Normalized, mode SubAgentMode, depth int) (string, string, error) {
-	if depth >= MaxAgentDepth {
-		return "", "", fmt.Errorf("max agent depth %d exceeded", MaxAgentDepth)
+func (s *Session) spawnSubAgent(ctx context.Context, norm agentcontracts.Normalized, mode engagent.SubAgentMode, depth int) (string, string, error) {
+	if depth >= engagent.MaxAgentDepth {
+		return "", "", fmt.Errorf("max agent depth %d exceeded", engagent.MaxAgentDepth)
 	}
 
-	maxTurns := DefaultTurnsForMode(mode)
-	if mode == SubAgentExplore && norm.Thoroughness != "" {
+	maxTurns := engagent.DefaultTurnsForMode(mode)
+	if mode == engagent.SubAgentExplore && norm.Thoroughness != "" {
 		maxTurns = engagent.ThoroughnessTurns(engagent.ExploreThoroughness(norm.Thoroughness))
 	}
 
@@ -174,7 +174,7 @@ func (s *Session) spawnSubAgent(ctx context.Context, norm agentcontracts.Normali
 
 	// Capability mode can further restrict tools beyond profile defaults.
 	if norm.CapabilityMode == agentcontracts.CapReadOnly {
-		registry = registry.Filter(ExploreTools)
+		registry = registry.Filter(engagent.ExploreTools)
 	}
 
 	subPromptCtx := prompts.PromptContext{
@@ -185,14 +185,14 @@ func (s *Session) spawnSubAgent(ctx context.Context, norm agentcontracts.Normali
 	if err != nil {
 		subSystemPrompt = s.Persistence().System()
 	}
-	if mode == SubAgentPlan {
+	if mode == engagent.SubAgentPlan {
 		subSystemPrompt = planSystemPrefix + "\n\n" + subSystemPrompt
 	}
 
 	sub := s.SubSession(model, subSystemPrompt, registry)
 	sub.PermSvc().SetPermissionFn(s.PermSvc().PermissionFn())
 	// Explore/plan: hard read-only bash allowlist (in addition to tool filter).
-	if IsReadOnlyMode(mode) || norm.CapabilityMode == agentcontracts.CapReadOnly {
+	if engagent.IsReadOnlyMode(mode) || norm.CapabilityMode == agentcontracts.CapReadOnly {
 		sub.Tools().SetReadOnlyBash(true)
 	}
 	// A child receives an independent snapshot of the parent's policy. This
@@ -281,30 +281,30 @@ func (s *Session) spawnSubAgent(ctx context.Context, norm agentcontracts.Normali
 const planSystemPrefix = "You are a planning sub-agent. Produce an ordered, actionable plan. " +
 	"Do not modify files. Prefer research tools (Read, Grep, Glob, LS) and only use Bash for read-only inspection."
 
-func (s *Session) resolveSubAgentModel(mode SubAgentMode) string {
+func (s *Session) resolveSubAgentModel(mode engagent.SubAgentMode) string {
 	current := s.ChatLLM().Model()
 	if s.LifecycleSvc().Cascade() == nil {
 		return current
 	}
 	switch mode {
-	case SubAgentExplore:
+	case engagent.SubAgentExplore:
 		return s.LifecycleSvc().Cascade().SelectModel("summarize", current, s.LifecycleSvc().Cascade().Roles.Explorer)
-	case SubAgentPlan:
+	case engagent.SubAgentPlan:
 		return s.LifecycleSvc().Cascade().SelectModel("planning", current, s.LifecycleSvc().Cascade().Roles.Planner)
-	case SubAgentGeneral:
+	case engagent.SubAgentGeneral:
 		return s.LifecycleSvc().Cascade().SelectModel("implement", current, "")
 	default:
 		return current
 	}
 }
 
-func (s *Session) resolveSubAgentTools(mode SubAgentMode) *tool.Registry {
+func (s *Session) resolveSubAgentTools(mode engagent.SubAgentMode) *tool.Registry {
 	registry := s.Tools().Registry()
 	switch mode {
-	case SubAgentExplore:
-		return registry.Filter(ExploreTools)
-	case SubAgentPlan:
-		return registry.Filter(PlanTools)
+	case engagent.SubAgentExplore:
+		return registry.Filter(engagent.ExploreTools)
+	case engagent.SubAgentPlan:
+		return registry.Filter(engagent.PlanTools)
 	default:
 		return registry
 	}
diff --git a/internal/engine/bmad_features_test.go b/internal/engine/bmad_features_test.go
index 446052cc..1e2c385a 100644
--- a/internal/engine/bmad_features_test.go
+++ b/internal/engine/bmad_features_test.go
@@ -4,6 +4,8 @@ import (
 	"os"
 	"path/filepath"
 	"testing"
+
+	"github.com/GrayCodeAI/rho/internal/engine/project"
 )
 
 func TestClassifyScale(t *testing.T) {
@@ -78,7 +80,7 @@ func TestProjectContext_Load(t *testing.T) {
 	dir := t.TempDir()
 	os.WriteFile(filepath.Join(dir, "PROJECT_CONTEXT.md"), []byte("## Stack\n- Go 1.21\n- PostgreSQL"), 0o644)
 
-	pc := NewProjectContext(dir)
+	pc := project.NewProjectContext(dir)
 	content := pc.Load()
 	if !hasSubstr(content, "Go 1.21") {
 		t.Error("expected project context content")
@@ -89,7 +91,7 @@ func TestProjectContext_Load(t *testing.T) {
 }
 
 func TestProjectContext_NoFiles(t *testing.T) {
-	pc := NewProjectContext(t.TempDir())
+	pc := project.NewProjectContext(t.TempDir())
 	content := pc.Load()
 	if content != "" {
 		t.Error("expected empty content when no files exist")
diff --git a/internal/engine/compact.go b/internal/engine/compact.go
index 52b63162..5f777cf9 100644
--- a/internal/engine/compact.go
+++ b/internal/engine/compact.go
@@ -77,7 +77,7 @@ func (s *Session) smartCompactBody(ctx context.Context) {
 
 	// Extract file tracking from messages being compacted
 	if s.Persistence().Files() == nil {
-		s.Persistence().SetFiles(NewFileTracker())
+		s.Persistence().SetFiles(compact.NewFileTracker())
 	}
 	files := s.Persistence().Files()
 	compactedMsgs := raw[:len(raw)-keepEnd]
@@ -138,12 +138,12 @@ func (s *Session) generateSummary(ctx context.Context, raw []types.FluxMessage)
 	// compaction, merge the NEW messages into it rather than re-summarizing the
 	// entire conversation from scratch. This preserves already-captured context
 	// and avoids the cost of re-deriving it.
-	prior := ExtractPriorSummary(raw)
+	prior := compact.ExtractPriorSummary(raw)
 	newMsgs := raw
 	if prior != "" {
 		// Only the messages that came after the persisted summary are "new".
 		for i, m := range raw {
-			if m.Role == "user" && strings.HasPrefix(m.Content, PriorSummaryPrefix) {
+			if m.Role == "user" && strings.HasPrefix(m.Content, compact.PriorSummaryPrefix) {
 				newMsgs = raw[i+1:]
 				break
 			}
@@ -153,7 +153,7 @@ func (s *Session) generateSummary(ctx context.Context, raw []types.FluxMessage)
 	// Build a compact version of the conversation for summarization
 	// using the structured compaction prompt from compact_prompt.go
 	var summaryMsgs []types.FluxMessage
-	compactPrompt := BuildIncrementalCompactPrompt(prior)
+	compactPrompt := compact.BuildIncrementalCompactPrompt(prior)
 	summaryMsgs = append(summaryMsgs, types.FluxMessage{
 		Role:    "user",
 		Content: compactPrompt + "\n\nConversation:\n",
@@ -209,7 +209,7 @@ func (s *Session) generateSummary(ctx context.Context, raw []types.FluxMessage)
 		return ""
 	}
 	// Extract structured summary, stripping analysis block
-	return FormatCompactSummary(resp.Content)
+	return compact.FormatCompactSummary(resp.Content)
 }
 
 // extractSummaryFromCompressed pulls key information from token-compressed text
diff --git a/internal/engine/compact_auto.go b/internal/engine/compact_auto.go
index 0bc85a39..bea8551f 100644
--- a/internal/engine/compact_auto.go
+++ b/internal/engine/compact_auto.go
@@ -6,6 +6,10 @@ import (
 	"sync"
 	"time"
 
+	"github.com/GrayCodeAI/rho/internal/engine/token"
+
+	"github.com/GrayCodeAI/rho/internal/engine/compact"
+
 	"github.com/GrayCodeAI/rho/internal/circuitbreaker"
 	"github.com/GrayCodeAI/rho/internal/types"
 )
@@ -14,13 +18,13 @@ import (
 type AutoCompactor struct {
 	mu           sync.Mutex
 	registry     *StrategyRegistry
-	config       CompactConfig
+	config       compact.CompactConfig
 	breaker      *circuitbreaker.Breaker
 	lastStrategy string
 }
 
 // NewAutoCompactor creates an auto-compactor with the given config.
-func NewAutoCompactor(config CompactConfig) *AutoCompactor {
+func NewAutoCompactor(config compact.CompactConfig) *AutoCompactor {
 	return &AutoCompactor{
 		registry: NewStrategyRegistry(config),
 		config:   config,
@@ -29,7 +33,7 @@ func NewAutoCompactor(config CompactConfig) *AutoCompactor {
 }
 
 // Configure updates compaction settings and rebuilds the strategy registry.
-func (ac *AutoCompactor) Configure(config CompactConfig) {
+func (ac *AutoCompactor) Configure(config compact.CompactConfig) {
 	if ac == nil {
 		return
 	}
@@ -61,7 +65,7 @@ func (ac *AutoCompactor) ShouldAutoCompact(sess *Session) bool {
 		return false
 	}
 
-	tokenCount := EstimateTokens(sess.Persistence().RawMessages())
+	tokenCount := token.EstimateTokens(sess.Persistence().RawMessages())
 	threshold := ac.GetAutoCompactThreshold()
 	return tokenCount >= threshold
 }
@@ -73,7 +77,7 @@ func (ac *AutoCompactor) AutoCompactIfNeeded(ctx context.Context, sess *Session)
 		return "", false
 	}
 
-	tokensBefore := EstimateTokens(sess.Persistence().RawMessages())
+	tokensBefore := token.EstimateTokens(sess.Persistence().RawMessages())
 	strategy, err := ac.RunCompaction(ctx, sess)
 	if err != nil {
 		ac.mu.Lock()
@@ -84,12 +88,12 @@ func (ac *AutoCompactor) AutoCompactIfNeeded(ctx context.Context, sess *Session)
 			"failures": ac.breaker.ConsecutiveFailures(),
 		})
 		sess.compact(ctx)
-		tokensAfter := EstimateTokens(sess.Persistence().RawMessages())
+		tokensAfter := token.EstimateTokens(sess.Persistence().RawMessages())
 		sess.recordCompaction("truncate_fallback", tokensBefore, tokensAfter, false)
 		return "truncate_fallback", true
 	}
 
-	tokensAfter := EstimateTokens(sess.Persistence().RawMessages())
+	tokensAfter := token.EstimateTokens(sess.Persistence().RawMessages())
 	if tokensAfter >= tokensBefore {
 		// Strategy ran but produced no reduction (e.g. LLM summary was
 		// rejected, or messages were not reduced); fall back to truncation.
@@ -101,7 +105,7 @@ func (ac *AutoCompactor) AutoCompactIfNeeded(ctx context.Context, sess *Session)
 			"tokens_after":  tokensAfter,
 		})
 		sess.compact(ctx)
-		tokensAfter = EstimateTokens(sess.Persistence().RawMessages())
+		tokensAfter = token.EstimateTokens(sess.Persistence().RawMessages())
 		sess.recordCompaction("truncate_fallback", tokensBefore, tokensAfter, false)
 		return "truncate_fallback", true
 	}
@@ -122,7 +126,7 @@ func (ac *AutoCompactor) RunCompaction(ctx context.Context, sess *Session) (stri
 	ac.mu.Unlock()
 
 	messages := sess.Persistence().RawMessages()
-	tokenCount := EstimateTokens(messages)
+	tokenCount := token.EstimateTokens(messages)
 	strategy := registry.SelectStrategy(sess, messages, tokenCount)
 	if strategy == nil {
 		return "", errors.New("no compaction strategy available")
@@ -177,13 +181,13 @@ func (s *SmartCompactStrategy) ShouldTrigger(msgs []types.FluxMessage, tokenCoun
 	return tokenCount >= threshold && len(msgs) > 20
 }
 
-func (s *SmartCompactStrategy) Compact(ctx context.Context, sess *Session) (*CompactResult, error) {
-	tokensBefore := EstimateTokens(sess.Persistence().RawMessages())
+func (s *SmartCompactStrategy) Compact(ctx context.Context, sess *Session) (*compact.CompactResult, error) {
+	tokensBefore := token.EstimateTokens(sess.Persistence().RawMessages())
 	sess.smartCompact(ctx)
 	messages := sess.Persistence().RawMessages()
-	tokensAfter := EstimateTokens(messages)
+	tokensAfter := token.EstimateTokens(messages)
 
-	return &CompactResult{
+	return &compact.CompactResult{
 		Messages:     messages,
 		TokensBefore: tokensBefore,
 		TokensAfter:  tokensAfter,
@@ -200,13 +204,13 @@ func (s *TruncateStrategy) ShouldTrigger(_ []types.FluxMessage, tokenCount, thre
 	return tokenCount >= threshold
 }
 
-func (s *TruncateStrategy) Compact(ctx context.Context, sess *Session) (*CompactResult, error) {
-	tokensBefore := EstimateTokens(sess.Persistence().RawMessages())
+func (s *TruncateStrategy) Compact(ctx context.Context, sess *Session) (*compact.CompactResult, error) {
+	tokensBefore := token.EstimateTokens(sess.Persistence().RawMessages())
 	sess.compact(ctx)
 	messages := sess.Persistence().RawMessages()
-	tokensAfter := EstimateTokens(messages)
+	tokensAfter := token.EstimateTokens(messages)
 
-	return &CompactResult{
+	return &compact.CompactResult{
 		Messages:     messages,
 		TokensBefore: tokensBefore,
 		TokensAfter:  tokensAfter,
diff --git a/internal/engine/compact_clear.go b/internal/engine/compact_clear.go
index abe30faf..1fa0f366 100644
--- a/internal/engine/compact_clear.go
+++ b/internal/engine/compact_clear.go
@@ -4,6 +4,8 @@ import (
 	"context"
 	"sort"
 
+	"github.com/GrayCodeAI/rho/internal/engine/token"
+
 	"github.com/GrayCodeAI/rho/internal/types"
 )
 
@@ -80,7 +82,7 @@ func (s *Session) ClearOldToolResults(ctx context.Context) bool {
 	if window <= 0 {
 		return false
 	}
-	tokens := EstimateTokens(raw)
+	tokens := token.EstimateTokens(raw)
 	threshold := int(float64(window) * clearThresholdFraction)
 	if tokens < threshold {
 		return false
diff --git a/internal/engine/compact_micro_engine.go b/internal/engine/compact_micro_engine.go
index 9a085183..7bf6d3f2 100644
--- a/internal/engine/compact_micro_engine.go
+++ b/internal/engine/compact_micro_engine.go
@@ -4,6 +4,8 @@ import (
 	"context"
 	"time"
 
+	"github.com/GrayCodeAI/rho/internal/engine/token"
+
 	"github.com/GrayCodeAI/rho/internal/types"
 
 	"github.com/GrayCodeAI/rho/internal/engine/compact"
@@ -29,13 +31,13 @@ func (s *MicroCompactStrategy) ShouldTrigger(msgs []types.FluxMessage, tokenCoun
 	return compact.HasTimeGap(msgs, 60*time.Minute)
 }
 
-func (s *MicroCompactStrategy) Compact(ctx context.Context, sess *Session) (*CompactResult, error) {
+func (s *MicroCompactStrategy) Compact(ctx context.Context, sess *Session) (*compact.CompactResult, error) {
 	messages := sess.Persistence().RawMessages()
-	tokensBefore := EstimateTokens(messages)
-	result := compact.MicrocompactMessages(messages, DefaultMicroCompactConfig())
-	tokensAfter := EstimateTokens(result)
+	tokensBefore := token.EstimateTokens(messages)
+	result := compact.MicrocompactMessages(messages, compact.DefaultMicroCompactConfig())
+	tokensAfter := token.EstimateTokens(result)
 
-	return &CompactResult{
+	return &compact.CompactResult{
 		Messages:     result,
 		TokensBefore: tokensBefore,
 		TokensAfter:  tokensAfter,
diff --git a/internal/engine/compact_provider_native.go b/internal/engine/compact_provider_native.go
index 066c020d..830f2352 100644
--- a/internal/engine/compact_provider_native.go
+++ b/internal/engine/compact_provider_native.go
@@ -4,6 +4,10 @@ import (
 	"context"
 	"fmt"
 
+	"github.com/GrayCodeAI/rho/internal/engine/token"
+
+	"github.com/GrayCodeAI/rho/internal/engine/compact"
+
 	"github.com/GrayCodeAI/flux/llm"
 	"github.com/GrayCodeAI/rho/internal/types"
 )
@@ -19,7 +23,7 @@ func (s *ProviderNativeCompactStrategy) ShouldTrigger(msgs []types.FluxMessage,
 	return tokenCount >= threshold && len(msgs) >= 8
 }
 
-func (s *ProviderNativeCompactStrategy) Compact(ctx context.Context, sess *Session) (*CompactResult, error) {
+func (s *ProviderNativeCompactStrategy) Compact(ctx context.Context, sess *Session) (*compact.CompactResult, error) {
 	if sess == nil {
 		return nil, fmt.Errorf("no session")
 	}
@@ -32,7 +36,7 @@ func (s *ProviderNativeCompactStrategy) Compact(ctx context.Context, sess *Sessi
 	}
 
 	messagesBefore := sess.Persistence().RawMessages()
-	tokensBefore := EstimateTokens(messagesBefore)
+	tokensBefore := token.EstimateTokens(messagesBefore)
 	summary, err := compactor.CompactNative(ctx, llm.NativeCompactionRequest{
 		Provider:        sess.ChatLLM().Provider(),
 		Model:           sess.ChatLLM().Model(),
@@ -50,11 +54,11 @@ func (s *ProviderNativeCompactStrategy) Compact(ctx context.Context, sess *Sessi
 		keepEnd = len(messagesBefore)
 	}
 	tail := append([]types.FluxMessage(nil), messagesBefore[len(messagesBefore)-keepEnd:]...)
-	messages := append([]types.FluxMessage{{Role: "user", Content: FormatCompactSummary(summary)}}, tail...)
-	compact := &CompactResult{
+	messages := append([]types.FluxMessage{{Role: "user", Content: compact.FormatCompactSummary(summary)}}, tail...)
+	compact := &compact.CompactResult{
 		Messages:     messages,
 		TokensBefore: tokensBefore,
-		TokensAfter:  EstimateTokens(messages),
+		TokensAfter:  token.EstimateTokens(messages),
 		Strategy:     "provider_native",
 	}
 	sess.Persistence().SetMessages(compact.Messages)
diff --git a/internal/engine/compact_reexports.go b/internal/engine/compact_reexports.go
deleted file mode 100644
index 0c9014e3..00000000
--- a/internal/engine/compact_reexports.go
+++ /dev/null
@@ -1,57 +0,0 @@
-// This file re-exports symbols from the compact sub-package so that existing
-// callers of engine.* keep compiling during the Stage 2 migration.
-// See docs/plans/engine-refactor-plan.md.
-package engine
-
-import (
-	"github.com/GrayCodeAI/rho/internal/types"
-
-	"github.com/GrayCodeAI/rho/internal/engine/compact"
-)
-
-type CompactVariant = compact.CompactVariant
-
-const (
-	CompactBase    = compact.CompactBase
-	CompactPartial = compact.CompactPartial
-	CompactUpTo    = compact.CompactUpTo
-)
-
-type (
-	CompactResult       = compact.CompactResult
-	CompactConfig       = compact.CompactConfig
-	FileTracker         = compact.FileTracker
-	MicroCompactConfig  = compact.MicroCompactConfig
-	APICompactConfig    = compact.APICompactConfig
-	SessionMemoryConfig = compact.SessionMemoryConfig
-	CompactionTrigger   = compact.CompactionTrigger
-)
-
-func DefaultCompactConfig() CompactConfig             { return compact.DefaultCompactConfig() }
-func DefaultMicroCompactConfig() MicroCompactConfig   { return compact.DefaultMicroCompactConfig() }
-func DefaultAPICompactConfig() APICompactConfig       { return compact.DefaultAPICompactConfig() }
-func DefaultSessionMemoryConfig() SessionMemoryConfig { return compact.DefaultSessionMemoryConfig() }
-func NewFileTracker() *FileTracker                    { return compact.NewFileTracker() }
-func NewCompactionTrigger(windowSize int) *CompactionTrigger {
-	return compact.NewCompactionTrigger(windowSize)
-}
-
-func BuildCompactPrompt(variant CompactVariant) string { return compact.BuildCompactPrompt(variant) }
-func FormatCompactSummary(raw string) string           { return compact.FormatCompactSummary(raw) }
-func BuildIncrementalCompactPrompt(priorSummary string) string {
-	return compact.BuildIncrementalCompactPrompt(priorSummary)
-}
-
-func ExtractPriorSummary(msgs []types.FluxMessage) string {
-	return compact.ExtractPriorSummary(msgs)
-}
-
-// PriorSummaryPrefix is the marker prefix rho prepends to a persisted
-// conversation summary message.
-const PriorSummaryPrefix = compact.PriorSummaryPrefix
-
-func IsCompactableTool(name string) bool { return compact.IsCompactableTool(name) }
-func AdjustIndexToPreserveAPIInvariants(msgs []types.FluxMessage, startIdx int) int {
-	return compact.AdjustIndexToPreserveAPIInvariants(msgs, startIdx)
-}
-func HasTextContent(m types.FluxMessage) bool { return compact.HasTextContent(m) }
diff --git a/internal/engine/compact_relevance.go b/internal/engine/compact_relevance.go
index e2ca5444..3bba9155 100644
--- a/internal/engine/compact_relevance.go
+++ b/internal/engine/compact_relevance.go
@@ -4,6 +4,10 @@ import (
 	"context"
 	"time"
 
+	"github.com/GrayCodeAI/rho/internal/engine/token"
+
+	"github.com/GrayCodeAI/rho/internal/engine/compact"
+
 	"github.com/GrayCodeAI/rho/internal/relevanceprune"
 	"github.com/GrayCodeAI/rho/internal/types"
 )
@@ -27,8 +31,8 @@ func (s *RelevancePruneStrategy) ShouldTrigger(msgs []types.FluxMessage, tokenCo
 	return tokenCount >= threshold && len(msgs) >= 20
 }
 
-func (s *RelevancePruneStrategy) Compact(ctx context.Context, sess *Session) (*CompactResult, error) {
-	tokensBefore := EstimateTokens(sess.Persistence().RawMessages())
+func (s *RelevancePruneStrategy) Compact(ctx context.Context, sess *Session) (*compact.CompactResult, error) {
+	tokensBefore := token.EstimateTokens(sess.Persistence().RawMessages())
 	raw := sess.Persistence().RawMessages()
 
 	taskContext := lastUserText(raw)
@@ -45,8 +49,8 @@ func (s *RelevancePruneStrategy) Compact(ctx context.Context, sess *Session) (*C
 
 	out := toFluxMessages(pruned)
 	sess.Persistence().SetMessages(out)
-	tokensAfter := EstimateTokens(sess.Persistence().RawMessages())
-	return &CompactResult{
+	tokensAfter := token.EstimateTokens(sess.Persistence().RawMessages())
+	return &compact.CompactResult{
 		Messages:     out,
 		TokensBefore: tokensBefore,
 		TokensAfter:  tokensAfter,
diff --git a/internal/engine/compact_session_memory_engine.go b/internal/engine/compact_session_memory_engine.go
index c4592eea..1f5b338d 100644
--- a/internal/engine/compact_session_memory_engine.go
+++ b/internal/engine/compact_session_memory_engine.go
@@ -6,6 +6,8 @@ import (
 	"os"
 	"strings"
 
+	"github.com/GrayCodeAI/rho/internal/engine/token"
+
 	"github.com/GrayCodeAI/rho/internal/types"
 
 	"github.com/GrayCodeAI/rho/internal/engine/compact"
@@ -27,7 +29,7 @@ func (s *SessionMemoryStrategy) ShouldTrigger(msgs []types.FluxMessage, tokenCou
 	return true
 }
 
-func (s *SessionMemoryStrategy) Compact(ctx context.Context, sess *Session) (*CompactResult, error) {
+func (s *SessionMemoryStrategy) Compact(ctx context.Context, sess *Session) (*compact.CompactResult, error) {
 	memContent, err := compact.ReadSessionMemory("")
 	if err != nil {
 		return nil, fmt.Errorf("reading session memory: %w", err)
@@ -37,9 +39,9 @@ func (s *SessionMemoryStrategy) Compact(ctx context.Context, sess *Session) (*Co
 	}
 
 	messages := sess.Persistence().RawMessages()
-	tokensBefore := EstimateTokens(messages)
+	tokensBefore := token.EstimateTokens(messages)
 
-	cfg := DefaultSessionMemoryConfig()
+	cfg := compact.DefaultSessionMemoryConfig()
 	keepIdx := compact.CalculateMessagesToKeepIndex(messages, cfg)
 	keepIdx = compact.AdjustIndexToPreserveAPIInvariants(messages, keepIdx)
 
@@ -61,9 +63,9 @@ func (s *SessionMemoryStrategy) Compact(ctx context.Context, sess *Session) (*Co
 	})
 	result = append(result, kept...)
 
-	tokensAfter := EstimateTokens(result)
+	tokensAfter := token.EstimateTokens(result)
 
-	return &CompactResult{
+	return &compact.CompactResult{
 		Messages:     result,
 		Summary:      memContent,
 		TokensBefore: tokensBefore,
diff --git a/internal/engine/compact_split.go b/internal/engine/compact_split.go
index c65d3b14..0c811edb 100644
--- a/internal/engine/compact_split.go
+++ b/internal/engine/compact_split.go
@@ -5,6 +5,10 @@ import (
 	"strings"
 	"time"
 
+	"github.com/GrayCodeAI/rho/internal/engine/token"
+
+	"github.com/GrayCodeAI/rho/internal/engine/compact"
+
 	"github.com/GrayCodeAI/rho/internal/types"
 )
 
@@ -27,7 +31,7 @@ func turnTokenBudget(tail []types.FluxMessage) int {
 	}
 	totalTokens := 0
 	for _, msg := range tail {
-		totalTokens += EstimateMessageTokens(msg)
+		totalTokens += token.EstimateMessageTokens(msg)
 	}
 	budget := totalTokens / len(tail) * 3
 	if budget < 2000 {
@@ -51,7 +55,7 @@ func (s *Session) SplitTurnNeeded(keepCount int) bool {
 
 	// Check if any single message in the tail exceeds the budget
 	for _, msg := range tail {
-		if EstimateMessageTokens(msg) > budget {
+		if token.EstimateMessageTokens(msg) > budget {
 			return true
 		}
 	}
@@ -80,7 +84,7 @@ func (s *Session) splitTurnCompact(ctx context.Context) {
 
 	oversizedIdx := -1
 	for i, msg := range tail {
-		if EstimateMessageTokens(msg) > budget {
+		if token.EstimateMessageTokens(msg) > budget {
 			oversizedIdx = i
 			break
 		}
@@ -98,7 +102,7 @@ func (s *Session) splitTurnCompact(ctx context.Context) {
 
 	// Extract file tracking before compaction
 	if s.Persistence().Files() == nil {
-		s.Persistence().SetFiles(NewFileTracker())
+		s.Persistence().SetFiles(compact.NewFileTracker())
 	}
 	files := s.Persistence().Files()
 	files.ExtractFromMessages(raw[:splitPoint])
@@ -176,7 +180,7 @@ func (s *Session) generatePartialSummary(ctx context.Context, messages []types.F
 	}
 
 	var summaryMsgs []types.FluxMessage
-	compactPrompt := BuildCompactPrompt(CompactPartial)
+	compactPrompt := compact.BuildCompactPrompt(compact.CompactPartial)
 	var content strings.Builder
 	content.WriteString(compactPrompt)
 	content.WriteString("\n\nConversation:\n")
@@ -211,7 +215,7 @@ func (s *Session) generatePartialSummary(ctx context.Context, messages []types.F
 	if err != nil {
 		return ""
 	}
-	return FormatCompactSummary(resp.Content)
+	return compact.FormatCompactSummary(resp.Content)
 }
 
 // summarizeOversizedTurn summarizes the content of a single oversized message.
@@ -237,7 +241,7 @@ func (s *Session) summarizeOversizedTurn(ctx context.Context, msg types.FluxMess
 	var summaryMsgs []types.FluxMessage
 	summaryMsgs = append(summaryMsgs, types.FluxMessage{
 		Role: "user",
-		Content: BuildCompactPrompt(CompactUpTo) + "\n\nSummarize the key information from this content prefix (the rest will be retained verbatim):\n\n" +
+		Content: compact.BuildCompactPrompt(compact.CompactUpTo) + "\n\nSummarize the key information from this content prefix (the rest will be retained verbatim):\n\n" +
 			msg.Role + ": " + truncateRunes(content, halfLen),
 	})
 
@@ -256,7 +260,7 @@ func (s *Session) summarizeOversizedTurn(ctx context.Context, msg types.FluxMess
 	if err != nil {
 		return ""
 	}
-	return FormatCompactSummary(resp.Content)
+	return compact.FormatCompactSummary(resp.Content)
 }
 
 // smartCompactFallback is reached when split-turn compaction is detected but
diff --git a/internal/engine/compact_strategy_engine.go b/internal/engine/compact_strategy_engine.go
index 34897489..957bc4fa 100644
--- a/internal/engine/compact_strategy_engine.go
+++ b/internal/engine/compact_strategy_engine.go
@@ -3,21 +3,23 @@ package engine
 import (
 	"context"
 
+	"github.com/GrayCodeAI/rho/internal/engine/compact"
+
 	"github.com/GrayCodeAI/rho/internal/types"
 )
 
 type CompactStrategy interface {
 	Name() string
 	ShouldTrigger(msgs []types.FluxMessage, tokenCount, threshold int) bool
-	Compact(ctx context.Context, s *Session) (*CompactResult, error)
+	Compact(ctx context.Context, s *Session) (*compact.CompactResult, error)
 }
 
 type StrategyRegistry struct {
 	strategies []CompactStrategy
-	config     CompactConfig
+	config     compact.CompactConfig
 }
 
-func NewStrategyRegistry(config CompactConfig) *StrategyRegistry {
+func NewStrategyRegistry(config compact.CompactConfig) *StrategyRegistry {
 	r := &StrategyRegistry{config: config}
 	target := config.ContextWindowSize - config.AutoCompactBuffer - config.MaxOutputTokens
 	r.strategies = []CompactStrategy{
diff --git a/internal/engine/compact_strategy_test.go b/internal/engine/compact_strategy_test.go
index 33fdf6cb..84e3467d 100644
--- a/internal/engine/compact_strategy_test.go
+++ b/internal/engine/compact_strategy_test.go
@@ -6,6 +6,10 @@ import (
 	"testing"
 	"time"
 
+	"github.com/GrayCodeAI/rho/internal/engine/token"
+
+	"github.com/GrayCodeAI/rho/internal/engine/compact"
+
 	"github.com/GrayCodeAI/rho/internal/types"
 )
 
@@ -18,7 +22,7 @@ func TestSessionMemoryStrategy_ShouldTrigger(t *testing.T) {
 }
 
 func TestAutoCompactor_CircuitBreaker(t *testing.T) {
-	cfg := DefaultCompactConfig()
+	cfg := compact.DefaultCompactConfig()
 	cfg.MaxFailures = 2
 	cfg.ContextWindowSize = 1000
 	cfg.AutoCompactBuffer = 100
@@ -43,7 +47,7 @@ func TestAutoCompactor_CircuitBreaker(t *testing.T) {
 }
 
 func TestStrategyRegistry_SelectStrategy(t *testing.T) {
-	cfg := DefaultCompactConfig()
+	cfg := compact.DefaultCompactConfig()
 	registry := NewStrategyRegistry(cfg)
 
 	msgs := makeMessages(5)
@@ -109,7 +113,7 @@ func TestTurnTokenBudget(t *testing.T) {
 	for i := range big {
 		big[i] = types.FluxMessage{Role: "user", Content: strings.Repeat("x", 10000)}
 	}
-	avg := EstimateMessageTokens(big[0])
+	avg := token.EstimateMessageTokens(big[0])
 	if avg*3 <= 2000 {
 		t.Fatalf("test messages too small to exceed the 2000 floor: avg=%d", avg)
 	}
diff --git a/internal/engine/context_compaction.go b/internal/engine/context_compaction.go
index 5aaffbf8..a211664c 100644
--- a/internal/engine/context_compaction.go
+++ b/internal/engine/context_compaction.go
@@ -3,6 +3,8 @@ package engine
 import (
 	"path/filepath"
 
+	"github.com/GrayCodeAI/rho/internal/engine/token"
+
 	"github.com/GrayCodeAI/rho/internal/eventlog"
 	"github.com/GrayCodeAI/rho/internal/session"
 	"github.com/GrayCodeAI/rho/internal/storage"
@@ -73,7 +75,7 @@ func (s *Session) ContextUsedTokens() int {
 		return cache
 	}
 
-	est := EstimateTokens(msgs)
+	est := token.EstimateTokens(msgs)
 	persist.SetTokenEstimateCache(est, count, lastLen)
 
 	return est
diff --git a/internal/engine/context_governor.go b/internal/engine/context_governor.go
index 9ba2e339..a49510b2 100644
--- a/internal/engine/context_governor.go
+++ b/internal/engine/context_governor.go
@@ -4,6 +4,10 @@ import (
 	"context"
 	"fmt"
 
+	"github.com/GrayCodeAI/rho/internal/engine/token"
+
+	"github.com/GrayCodeAI/rho/internal/engine/compact"
+
 	"github.com/GrayCodeAI/rho/internal/engine/ctxmgr"
 	modelPkg "github.com/GrayCodeAI/rho/internal/provider/routing"
 )
@@ -71,11 +75,11 @@ func (s *Session) compactThresholdPct() int {
 	return pct
 }
 
-func (s *Session) compactConfig() CompactConfig {
+func (s *Session) compactConfig() compact.CompactConfig {
 	window := s.ContextWindowSize()
 	pct := s.compactThresholdPct()
 	target := window * pct / 100
-	cfg := DefaultCompactConfig()
+	cfg := compact.DefaultCompactConfig()
 	cfg.AutoEnabled = true
 	cfg.ContextWindowSize = window
 	cfg.MaxOutputTokens = 0
@@ -116,7 +120,7 @@ func (s *Session) WillCompactBeforeTurn() bool {
 	if len(s.Persistence().RawMessagesView()) > maxContextMessages {
 		return true
 	}
-	convTokens := EstimateTokens(s.Persistence().RawMessagesView())
+	convTokens := token.EstimateTokens(s.Persistence().RawMessagesView())
 	budget := ctxmgr.NewContextBudget(s.ContextWindowSize())
 	return budget.ShouldCompact(convTokens)
 }
@@ -140,25 +144,25 @@ func (s *Session) ManageContextBeforeTurn(ctx context.Context) (strategy string,
 	}
 
 	if len(s.Persistence().RawMessagesView()) > maxContextMessages {
-		before := EstimateTokens(s.Persistence().RawMessagesView())
+		before := token.EstimateTokens(s.Persistence().RawMessagesView())
 		beforeMsgs := len(s.Persistence().RawMessagesView())
 		s.smartCompact(ctx)
 		s.lastCompactionMsgDelta = beforeMsgs - len(s.Persistence().RawMessagesView())
 		s.lastCompactionSummary = "" // smartCompact may set it via the summary path
-		s.recordCompaction("smart_message_cap", before, EstimateTokens(s.Persistence().RawMessagesView()), false)
+		s.recordCompaction("smart_message_cap", before, token.EstimateTokens(s.Persistence().RawMessagesView()), false)
 		return "smart_message_cap", true
 	}
 
-	convTokens := EstimateTokens(s.Persistence().RawMessagesView())
+	convTokens := token.EstimateTokens(s.Persistence().RawMessagesView())
 	window := s.ContextWindowSize()
 	budget := ctxmgr.NewContextBudget(window)
 	if budget.ShouldCompact(convTokens) {
-		before := EstimateTokens(s.Persistence().RawMessagesView())
+		before := token.EstimateTokens(s.Persistence().RawMessagesView())
 		beforeMsgs := len(s.Persistence().RawMessagesView())
 		s.smartCompact(ctx)
 		s.lastCompactionMsgDelta = beforeMsgs - len(s.Persistence().RawMessagesView())
 		s.lastCompactionSummary = ""
-		s.recordCompaction("smart_budget", before, EstimateTokens(s.Persistence().RawMessagesView()), false)
+		s.recordCompaction("smart_budget", before, token.EstimateTokens(s.Persistence().RawMessagesView()), false)
 		return "smart_budget", true
 	}
 
@@ -173,13 +177,13 @@ func (s *Session) CompactConversation(ctx context.Context) (strategy string, tok
 	s.Persistence().SetRawMessages(ctxmgr.CollapseRepeatedMessages(s.Persistence().RawMessages()))
 	s.EnsureAutoCompactor()
 	beforeMsgs := len(s.Persistence().RawMessages())
-	tokensBefore = EstimateTokens(s.Persistence().RawMessages())
+	tokensBefore = token.EstimateTokens(s.Persistence().RawMessages())
 	strategy, err = s.Persistence().AutoCompactor().RunCompaction(ctx, s)
 	if err != nil {
 		s.smartCompact(ctx)
 		strategy = "smart_fallback"
 	}
-	tokensAfter = EstimateTokens(s.Persistence().RawMessages())
+	tokensAfter = token.EstimateTokens(s.Persistence().RawMessages())
 	s.lastCompactionMsgDelta = beforeMsgs - len(s.Persistence().RawMessages())
 	s.recordCompaction(strategy, tokensBefore, tokensAfter, true)
 	return strategy, tokensBefore, tokensAfter, nil
@@ -188,6 +192,6 @@ func (s *Session) CompactConversation(ctx context.Context) (strategy string, tok
 // ShouldCompactByBudget reports whether conversation tokens exceed the configured % of window.
 func (s *Session) ShouldCompactByBudget() bool {
 	window := s.ContextWindowSize()
-	conv := EstimateTokens(s.Persistence().RawMessages())
+	conv := token.EstimateTokens(s.Persistence().RawMessages())
 	return conv >= window*s.compactThresholdPct()/100
 }
diff --git a/internal/engine/control_reexports.go b/internal/engine/control_reexports.go
deleted file mode 100644
index 5fb30009..00000000
--- a/internal/engine/control_reexports.go
+++ /dev/null
@@ -1,20 +0,0 @@
-package engine
-
-import "github.com/GrayCodeAI/rho/internal/engine/control"
-
-type (
-	LoopDetector    = control.LoopDetector
-	StallEntry      = control.StallEntry
-	StallResult     = control.StallResult
-	StallDetector   = control.StallDetector
-	DecisionPoint   = control.DecisionPoint
-	BacktrackEngine = control.BacktrackEngine
-)
-
-const DoomLoopThreshold = control.DoomLoopThreshold
-
-var (
-	NewLoopDetector    = control.NewLoopDetector
-	NewStallDetector   = control.NewStallDetector
-	NewBacktrackEngine = control.NewBacktrackEngine
-)
diff --git a/internal/engine/cost_reexports.go b/internal/engine/cost_reexports.go
deleted file mode 100644
index c42adfd9..00000000
--- a/internal/engine/cost_reexports.go
+++ /dev/null
@@ -1,28 +0,0 @@
-package engine
-
-import (
-	"github.com/GrayCodeAI/rho/internal/engine/cost"
-	analytics "github.com/GrayCodeAI/rho/internal/observability"
-)
-
-type (
-	Cost           = cost.Cost
-	CostSnapshot   = cost.Snapshot
-	CostOptimizer  = cost.CostOptimizer
-	CostTracker    = cost.CostTracker
-	RequestCost    = cost.RequestCost
-	ModelPrice     = cost.ModelPrice
-	Recommendation = cost.Recommendation
-)
-
-func NewCostOptimizer() *CostOptimizer                { return cost.NewCostOptimizer() }
-func NewCostTracker(sessionID string) *CostTracker    { return cost.NewCostTracker(sessionID) }
-func LoadCostHistory() ([]analytics.CostEntry, error) { return cost.LoadCostHistory() }
-func FormatCostDisplay(totalUSD float64) string       { return cost.FormatCostDisplay(totalUSD) }
-func ModelPricing(modelName string) (inputPricePerM, outputPricePerM float64) {
-	return cost.ModelPricing(modelName)
-}
-
-func RegisterLivePricing(model string, inputPerM, outputPerM float64) {
-	cost.RegisterLivePricing(model, inputPerM, outputPerM)
-}
diff --git a/internal/engine/developer_features_test.go b/internal/engine/developer_features_test.go
index fc6bd0fd..e631e4da 100644
--- a/internal/engine/developer_features_test.go
+++ b/internal/engine/developer_features_test.go
@@ -5,6 +5,8 @@ import (
 	"strings"
 	"testing"
 	"time"
+
+	"github.com/GrayCodeAI/rho/internal/engine/compact"
 )
 
 func TestBackgroundRunner_Delegate(t *testing.T) {
@@ -49,7 +51,7 @@ func TestBackgroundRunner_CollectWhileRunning(t *testing.T) {
 }
 
 func TestCompactionTrigger(t *testing.T) {
-	ct := NewCompactionTrigger(100000)
+	ct := compact.NewCompactionTrigger(100000)
 	if ct.ShouldCompact(50000) {
 		t.Error("50% should not trigger at 75% threshold")
 	}
diff --git a/internal/engine/diff_reexports.go b/internal/engine/diff_reexports.go
deleted file mode 100644
index acc6934e..00000000
--- a/internal/engine/diff_reexports.go
+++ /dev/null
@@ -1,55 +0,0 @@
-package engine
-
-import "github.com/GrayCodeAI/rho/internal/engine/diff"
-
-// Types from diff sub-package.
-
-type (
-	PendingChange  = diff.PendingChange
-	DiffSandbox    = diff.DiffSandbox
-	StagingArea    = diff.StagingArea
-	StagedChange   = diff.StagedChange
-	StagedHunk     = diff.StagedHunk
-	DiffPreview    = diff.DiffPreview
-	FileChange     = diff.FileChange
-	DiffHunk       = diff.DiffHunk
-	DiffLine       = diff.DiffLine
-	ChangeStats    = diff.ChangeStats
-	DiffSummary    = diff.DiffSummary
-	FileSummary    = diff.FileSummary
-	DiffSummarizer = diff.DiffSummarizer
-	TestSelector   = diff.TestSelector
-	SelectedTests  = diff.SelectedTests
-	Diff3Result    = diff.Diff3Result
-	Diff3Conflict  = diff.Diff3Conflict
-	Diff3Stats     = diff.Diff3Stats
-	Diff3Region    = diff.Diff3Region
-	Edit           = diff.Edit
-)
-
-// Short-name aliases.
-
-type (
-	Preview    = diff.DiffPreview
-	Summarizer = diff.DiffSummarizer
-)
-
-// Functions.
-
-var (
-	NewDiffSandbox        = diff.NewDiffSandbox
-	NewStagingArea        = diff.NewStagingArea
-	NewDiffPreview        = diff.NewDiffPreview
-	NewDiffSummarizer     = diff.NewDiffSummarizer
-	NewTestSelector       = diff.NewTestSelector
-	ComputeDiff           = diff.ComputeDiff
-	ComputeMyersDiff      = diff.ComputeMyersDiff
-	RenderUnified         = diff.RenderUnified
-	Merge3                = diff.Merge3
-	MergeClean            = diff.MergeClean
-	FormatConflictMarkers = diff.FormatConflictMarkers
-	LCS                   = diff.LCS
-	EditScript            = diff.EditScript
-	BuildDependencyGraph  = diff.BuildDependencyGraph
-	GenerateTestCommand   = diff.GenerateTestCommand
-)
diff --git a/internal/engine/engine_integration_test.go b/internal/engine/engine_integration_test.go
index 00d566fc..2372ed48 100644
--- a/internal/engine/engine_integration_test.go
+++ b/internal/engine/engine_integration_test.go
@@ -5,6 +5,8 @@ import (
 	"testing"
 	"time"
 
+	"github.com/GrayCodeAI/rho/internal/engine/cost"
+
 	contracts "github.com/GrayCodeAI/rho/internal/contracts/policy"
 	"github.com/GrayCodeAI/rho/internal/types"
 
@@ -365,7 +367,7 @@ func TestIntegration_MaxTurns(t *testing.T) {
 	}
 
 	// Simulate cost accumulation to test budget checking.
-	sess.Cost = Cost{Model: "gpt-4o"}
+	sess.Cost = cost.Cost{Model: "gpt-4o"}
 	sess.Cost.Add(500_000, 200_000) // ~$3.25 which exceeds $1.0
 	if !sess.exceededBudget() {
 		t.Fatal("expected budget to be exceeded")
@@ -374,7 +376,7 @@ func TestIntegration_MaxTurns(t *testing.T) {
 	// Under-budget should not trigger.
 	sess2 := newTestSession()
 	sess2.SetMaxBudgetUSD(100.0)
-	sess2.Cost = Cost{Model: "gpt-4o"}
+	sess2.Cost = cost.Cost{Model: "gpt-4o"}
 	sess2.Cost.Add(100, 50)
 	if sess2.exceededBudget() {
 		t.Fatal("should not exceed $100 budget with minimal tokens")
diff --git a/internal/engine/engine_test.go b/internal/engine/engine_test.go
index f8ed5fc1..42e76d46 100644
--- a/internal/engine/engine_test.go
+++ b/internal/engine/engine_test.go
@@ -3,6 +3,8 @@ package engine
 import (
 	"context"
 	"testing"
+
+	"github.com/GrayCodeAI/rho/internal/engine/cost"
 )
 
 func TestPermissionMemoryAlwaysAllow(t *testing.T) {
@@ -135,7 +137,7 @@ func TestToolNeedsPermission(t *testing.T) {
 }
 
 func TestCostPricing(t *testing.T) {
-	c := Cost{Model: "claude-3-5-sonnet-20241022"}
+	c := cost.Cost{Model: "claude-3-5-sonnet-20241022"}
 	c.Add(1000, 500)
 	if c.PromptTokens != 1000 {
 		t.Fatalf("got %d prompt tokens", c.PromptTokens)
@@ -153,7 +155,7 @@ func TestCostPricing(t *testing.T) {
 }
 
 func TestCostSummary(t *testing.T) {
-	c := Cost{Model: "gpt-4o"}
+	c := cost.Cost{Model: "gpt-4o"}
 	c.Add(100, 50)
 	s := c.Summary()
 	if s == "" {
@@ -162,7 +164,7 @@ func TestCostSummary(t *testing.T) {
 }
 
 func TestCostTotal(t *testing.T) {
-	c := Cost{Model: "gpt-4o"}
+	c := cost.Cost{Model: "gpt-4o"}
 	c.Add(100, 50)
 	if c.Total() <= 0 {
 		t.Fatal("expected positive total")
diff --git a/internal/engine/errs_reexports.go b/internal/engine/errs_reexports.go
deleted file mode 100644
index 6eb21669..00000000
--- a/internal/engine/errs_reexports.go
+++ /dev/null
@@ -1,31 +0,0 @@
-package engine
-
-import "github.com/GrayCodeAI/rho/internal/engine/errs"
-
-type (
-	ErrorContext      = errs.ErrorContext
-	ErrorHelp         = errs.ErrorHelp
-	EnrichedError     = errs.EnrichedError
-	ErrorGroup        = errs.ErrorGroup
-	ErrorInstance     = errs.ErrorInstance
-	ErrorGrouper      = errs.ErrorGrouper
-	ErrorPattern      = errs.ErrorPattern
-	ErrorPatternDB    = errs.ErrorPatternDB
-	ErrorLearner      = errs.ErrorLearner
-	ErrorLearnerStats = errs.ErrorLearnerStats
-	LearnedPattern    = errs.LearnedPattern
-	ErrorRecovery     = errs.ErrorRecovery
-	RecoveryStrategy  = errs.RecoveryStrategy
-	RecoveryContext   = errs.RecoveryContext
-	RecoveryResult    = errs.RecoveryResult
-	RecoveryAttempt   = errs.RecoveryAttempt
-)
-
-func NewErrorContext() *ErrorContext             { return errs.NewErrorContext() }
-func FormatError(enriched *EnrichedError) string { return errs.FormatError(enriched) }
-func NewErrorGrouper() *ErrorGrouper             { return errs.NewErrorGrouper() }
-func NormalizeError(msg string) string           { return errs.NormalizeError(msg) }
-func NewErrorLearner() *ErrorLearner             { return errs.NewErrorLearner() }
-func ExtractPattern(errorMsg string) string      { return errs.ExtractPattern(errorMsg) }
-func NewErrorPatternDB() *ErrorPatternDB         { return errs.NewErrorPatternDB() }
-func NewErrorRecovery() *ErrorRecovery           { return errs.NewErrorRecovery() }
diff --git a/internal/engine/history_reexports.go b/internal/engine/history_reexports.go
deleted file mode 100644
index f6720c02..00000000
--- a/internal/engine/history_reexports.go
+++ /dev/null
@@ -1,39 +0,0 @@
-package engine
-
-import "github.com/GrayCodeAI/rho/internal/engine/history"
-
-type (
-	CommandRecord          = history.CommandRecord
-	CommandFrequency       = history.CommandFrequency
-	AliasSuggestion        = history.AliasSuggestion
-	CommandHistory         = history.CommandHistory
-	SummaryLevel           = history.SummaryLevel
-	SumMessage             = history.SumMessage
-	Summary                = history.Summary
-	ConversationSummarizer = history.ConversationSummarizer
-	DistillExample         = history.DistillExample
-	DistillStats           = history.DistillStats
-	DistillationPipeline   = history.DistillationPipeline
-	HeadTailWindow         = history.HeadTailWindow
-	WindowResult           = history.WindowResult
-	WindowMessage          = history.WindowMessage
-	FileMentionDetector    = history.FileMentionDetector
-	Annotation             = history.Annotation
-	AnnotationManager      = history.AnnotationManager
-)
-
-var (
-	NewCommandHistory         = history.NewCommandHistory
-	NewConversationSummarizer = history.NewConversationSummarizer
-	NewDistillationPipeline   = history.NewDistillationPipeline
-	NewHeadTailWindow         = history.NewHeadTailWindow
-	AdaptiveSizes             = history.AdaptiveSizes
-	PreserveToolPairs         = history.PreserveToolPairs
-	FormatWindow              = history.FormatWindow
-	ShouldApply               = history.ShouldApply
-	NewFileMentionDetector    = history.NewFileMentionDetector
-	NewAnnotationManager      = history.NewAnnotationManager
-	StripAnnotations          = history.StripAnnotations
-	DetectAnnotations         = history.DetectAnnotations
-	FormatAnnotations         = history.FormatAnnotations
-)
diff --git a/internal/engine/integration.go b/internal/engine/integration.go
index 3e45ca07..76562bbb 100644
--- a/internal/engine/integration.go
+++ b/internal/engine/integration.go
@@ -8,6 +8,18 @@ import (
 	"sync"
 	"time"
 
+	"github.com/GrayCodeAI/rho/internal/engine/compression"
+	"github.com/GrayCodeAI/rho/internal/engine/control"
+	"github.com/GrayCodeAI/rho/internal/engine/errs"
+	"github.com/GrayCodeAI/rho/internal/engine/history"
+	"github.com/GrayCodeAI/rho/internal/engine/intelligence"
+	"github.com/GrayCodeAI/rho/internal/engine/memory"
+	"github.com/GrayCodeAI/rho/internal/engine/observability"
+	"github.com/GrayCodeAI/rho/internal/engine/review"
+	"github.com/GrayCodeAI/rho/internal/engine/streaming"
+	"github.com/GrayCodeAI/rho/internal/engine/validation"
+	"github.com/GrayCodeAI/rho/internal/engine/workflow"
+
 	"github.com/GrayCodeAI/rho/internal/engine/ctxmgr"
 	"github.com/GrayCodeAI/rho/internal/engine/token"
 	"github.com/GrayCodeAI/rho/internal/eventlog"
@@ -27,40 +39,40 @@ import (
 // through the request lifecycle.
 type IntegrationPipeline struct {
 	// Pre-query pipeline
-	IntentClassifier *IntentClassifier
-	ToolSelector     *ToolSelector
+	IntentClassifier *intelligence.IntentClassifier
+	ToolSelector     *intelligence.ToolSelector
 	ContextDecay     *ctxmgr.ContextDecay
 	BudgetAllocator  *BudgetAllocator
-	TokenPredictor   *TokenPredictor
+	TokenPredictor   *token.TokenPredictor
 	AdaptivePrompt   *SystemPromptBuilder
 
 	// Post-response pipeline
-	ResponseFormatter   *ResponseFormatter
-	QualityScorer       *QualityScorer
-	FileMentionDetector *FileMentionDetector
+	ResponseFormatter   *streaming.ResponseFormatter
+	QualityScorer       *review.QualityScorer
+	FileMentionDetector *history.FileMentionDetector
 
 	// Post-tool pipeline
-	LintLoop      *LintLoop
-	TestLoop      *TestLoop
-	StallDetector *StallDetector
-	ErrorRecovery *ErrorRecovery
+	LintLoop      *validation.LintLoop
+	TestLoop      *validation.TestLoop
+	StallDetector *control.StallDetector
+	ErrorRecovery *errs.ErrorRecovery
 
 	// Security pipeline
 	InjectionScanner *InjectionScanner
 	OutputRedactor   *OutputRedactor
 
 	// Learning pipeline
-	ExperienceStore   *ExperienceStore
-	KnowledgeBase     *KnowledgeBase
-	FeedbackCollector *FeedbackCollector
-	SelfAssessor      *SelfAssessor
+	ExperienceStore   *memory.ExperienceStore
+	KnowledgeBase     *memory.KnowledgeBase
+	FeedbackCollector *observability.FeedbackCollector
+	SelfAssessor      *review.SelfAssessor
 
 	// Session management
-	Timeline       *Timeline
-	TokenReporter  *TokenReporter
-	WorkspaceState *WorkspaceState
-	CommandHistory *CommandHistory
-	ResponseCache  *ResponseCache
+	Timeline       *compression.Timeline
+	TokenReporter  *token.TokenReporter
+	WorkspaceState *workflow.WorkspaceState
+	CommandHistory *history.CommandHistory
+	ResponseCache  *streaming.ResponseCache
 
 	// Internal
 	sessionStart time.Time
@@ -149,7 +161,7 @@ func (is *InjectionScanner) Scan(input string) *ScanResult {
 
 // PreQueryResult holds the aggregated results of the pre-query pipeline.
 type PreQueryResult struct {
-	Intent           *Intent
+	Intent           *intelligence.Intent
 	SuggestedTools   []string
 	BudgetAllocation map[string]int
 	PredictedCost    float64
@@ -180,8 +192,8 @@ type PostToolResult struct {
 
 // SessionSummary captures the final assessment when a session ends.
 type SessionSummary struct {
-	Assessment    *Assessment
-	Experience    *Experience
+	Assessment    *review.Assessment
+	Experience    *memory.Experience
 	Summary       string
 	TokensTotal   int
 	Duration      time.Duration
@@ -205,40 +217,40 @@ func NewIntegrationPipeline() *IntegrationPipeline {
 
 	return &IntegrationPipeline{
 		// Pre-query
-		IntentClassifier: NewIntentClassifier(),
-		ToolSelector:     NewToolSelector(defaultToolSet()),
+		IntentClassifier: intelligence.NewIntentClassifier(),
+		ToolSelector:     intelligence.NewToolSelector(defaultToolSet()),
 		ContextDecay:     ctxmgr.NewContextDecay(30 * time.Minute),
 		BudgetAllocator:  newDefaultBudgetAllocator(),
-		TokenPredictor:   NewTokenPredictor(),
+		TokenPredictor:   token.NewTokenPredictor(),
 		AdaptivePrompt:   NewSystemPromptBuilder("", 4096),
 
 		// Post-response
-		ResponseFormatter:   NewResponseFormatter(),
-		QualityScorer:       NewQualityScorer(),
-		FileMentionDetector: NewFileMentionDetector("."),
+		ResponseFormatter:   streaming.NewResponseFormatter(),
+		QualityScorer:       review.NewQualityScorer(),
+		FileMentionDetector: history.NewFileMentionDetector("."),
 
 		// Post-tool
-		LintLoop:      NewLintLoop(),
-		TestLoop:      NewTestLoop(),
-		StallDetector: NewStallDetector(),
-		ErrorRecovery: NewErrorRecovery(),
+		LintLoop:      validation.NewLintLoop(),
+		TestLoop:      validation.NewTestLoop(),
+		StallDetector: control.NewStallDetector(),
+		ErrorRecovery: errs.NewErrorRecovery(),
 
 		// Security
 		InjectionScanner: NewInjectionScanner(),
 		OutputRedactor:   NewOutputRedactor(),
 
 		// Learning
-		ExperienceStore:   NewExperienceStore(filepath.Join(stateRoot, "experience")),
-		KnowledgeBase:     NewKnowledgeBase(filepath.Join(stateRoot, "knowledge")),
-		FeedbackCollector: NewFeedbackCollector(filepath.Join(stateRoot, "feedback")),
-		SelfAssessor:      NewSelfAssessor(),
+		ExperienceStore:   memory.NewExperienceStore(filepath.Join(stateRoot, "experience")),
+		KnowledgeBase:     memory.NewKnowledgeBase(filepath.Join(stateRoot, "knowledge")),
+		FeedbackCollector: observability.NewFeedbackCollector(filepath.Join(stateRoot, "feedback")),
+		SelfAssessor:      review.NewSelfAssessor(),
 
 		// Session management
-		Timeline:       NewTimeline("default"),
-		TokenReporter:  NewTokenReporter(200000),
-		WorkspaceState: NewWorkspaceState("."),
-		CommandHistory: NewCommandHistory(),
-		ResponseCache:  NewResponseCache(1000, 24*time.Hour),
+		Timeline:       compression.NewTimeline("default"),
+		TokenReporter:  token.NewTokenReporter(200000),
+		WorkspaceState: workflow.NewWorkspaceState("."),
+		CommandHistory: history.NewCommandHistory(),
+		ResponseCache:  streaming.NewResponseCache(1000, 24*time.Hour),
 
 		// Internal
 		sessionStart: time.Now(),
@@ -246,8 +258,8 @@ func NewIntegrationPipeline() *IntegrationPipeline {
 }
 
 // defaultToolSet returns the standard set of tools available to the agent.
-func defaultToolSet() []ToolInfo {
-	return []ToolInfo{
+func defaultToolSet() []intelligence.ToolInfo {
+	return []intelligence.ToolInfo{
 		{Name: "Read", Category: "file", Cost: "free", ReadOnly: true},
 		{Name: "Write", Category: "file", Cost: "cheap", ReadOnly: false},
 		{Name: "Edit", Category: "file", Cost: "cheap", ReadOnly: false},
@@ -328,7 +340,7 @@ func (p *IntegrationPipeline) PreQuery(messages []types.FluxMessage, userInput s
 	result.InjectionRisk = p.InjectionScanner.Scan(userInput)
 
 	// 8. Check response cache
-	if ShouldCache(userInput) {
+	if streaming.ShouldCache(userInput) {
 		if entry, ok := p.ResponseCache.Get(userInput, ""); ok {
 			result.CacheHit = true
 			result.CachedResponse = entry.Response
@@ -364,7 +376,7 @@ func (p *IntegrationPipeline) PostResponse(response string, messages []types.Flu
 
 	// 2. Score quality
 	userPrompt := lastUserMessage(messages)
-	scoreCtx := ResponseContext{
+	scoreCtx := review.ResponseContext{
 		UserPrompt:        userPrompt,
 		AssistantResponse: result.FormattedResponse,
 		TokensUsed:        EstimateStringTokens(result.FormattedResponse),
@@ -400,12 +412,12 @@ func (p *IntegrationPipeline) PostResponse(response string, messages []types.Flu
 	p.TokenReporter.Record(0, tokensUsed, "", "", 0)
 
 	// 7. Cache the response
-	if ShouldCache(userPrompt) && result.QualityScore >= 0.7 {
+	if streaming.ShouldCache(userPrompt) && result.QualityScore >= 0.7 {
 		p.ResponseCache.Set(userPrompt, result.FormattedResponse, "", tokensUsed)
 	}
 
 	// 8. Update experience store with implicit signal
-	_ = p.FeedbackCollector.RecordImplicit(ImplicitSignal{
+	_ = p.FeedbackCollector.RecordImplicit(observability.ImplicitSignal{
 		Type:      "accepted",
 		SessionID: p.Timeline.SessionID,
 		Timestamp: time.Now(),
@@ -457,7 +469,7 @@ func (p *IntegrationPipeline) PostToolExecution(toolName string, args map[string
 
 	// 4. Attempt error recovery if the tool failed
 	if err != nil {
-		recoveryCtx := &RecoveryContext{
+		recoveryCtx := &errs.RecoveryContext{
 			Error:        err,
 			ErrorMsg:     err.Error(),
 			LastToolCall: toolName,
@@ -521,7 +533,7 @@ func (p *IntegrationPipeline) EndSession(ctx context.Context, success bool, task
 	}
 
 	// 1. Self-assess performance
-	taskCtx := TaskContext{
+	taskCtx := review.TaskContext{
 		Goal:          taskGoal,
 		FilesModified: len(summary.FilesModified),
 		Duration:      summary.Duration,
@@ -547,7 +559,7 @@ func (p *IntegrationPipeline) EndSession(ctx context.Context, success bool, task
 
 	// 3. Update knowledge base
 	if success && taskGoal != "" {
-		_ = p.KnowledgeBase.Add(&KnowledgeEntry{
+		_ = p.KnowledgeBase.Add(&memory.KnowledgeEntry{
 			Title:      fmt.Sprintf("Completed: %s", truncateString(taskGoal, 60)),
 			Content:    fmt.Sprintf("Task completed in %s with %d files modified.", summary.Duration.Round(time.Second), len(summary.FilesModified)),
 			Category:   "session_outcome",
@@ -558,7 +570,7 @@ func (p *IntegrationPipeline) EndSession(ctx context.Context, success bool, task
 	}
 
 	// 4. Collect implicit feedback
-	_ = p.FeedbackCollector.RecordImplicit(ImplicitSignal{
+	_ = p.FeedbackCollector.RecordImplicit(observability.ImplicitSignal{
 		Type:      "accepted",
 		SessionID: p.Timeline.SessionID,
 		Timestamp: time.Now(),
@@ -645,7 +657,7 @@ func formatStatus(name string, active bool) string {
 }
 
 // intentCategory safely extracts the category from a potentially-nil intent.
-func intentCategory(intent *Intent) string {
+func intentCategory(intent *intelligence.Intent) string {
 	if intent == nil {
 		return "unknown"
 	}
diff --git a/internal/engine/integration_test.go b/internal/engine/integration_test.go
index 02b651dc..93beed32 100644
--- a/internal/engine/integration_test.go
+++ b/internal/engine/integration_test.go
@@ -5,6 +5,8 @@ import (
 	"testing"
 	"time"
 
+	"github.com/GrayCodeAI/rho/internal/engine/cost"
+
 	"github.com/GrayCodeAI/rho/internal/tool"
 )
 
@@ -136,7 +138,7 @@ func TestCompact(t *testing.T) {
 
 // TestCostTracking tests cost tracking.
 func TestCostTracking(t *testing.T) {
-	c := Cost{Model: "gpt-4o"}
+	c := cost.Cost{Model: "gpt-4o"}
 	c.Add(1000, 500)
 
 	if c.PromptTokens != 1000 {
diff --git a/internal/engine/intelligence_reexports.go b/internal/engine/intelligence_reexports.go
deleted file mode 100644
index 18651b88..00000000
--- a/internal/engine/intelligence_reexports.go
+++ /dev/null
@@ -1,35 +0,0 @@
-// This file re-exports symbols from the intelligence sub-package so that existing
-// callers of engine.Intent, engine.NewIntentClassifier, etc. keep compiling
-// during the Stage 2 migration. See docs/plans/engine-refactor-plan.md.
-package engine
-
-import "github.com/GrayCodeAI/rho/internal/engine/intelligence"
-
-type (
-	Intent             = intelligence.Intent
-	IntentRule         = intelligence.IntentRule
-	ClassifiedInput    = intelligence.ClassifiedInput
-	IntentClassifier   = intelligence.IntentClassifier
-	Capability         = intelligence.Capability
-	CapabilityRegistry = intelligence.CapabilityRegistry
-	LanguageConfig     = intelligence.LanguageConfig
-	LanguageRegistry   = intelligence.LanguageRegistry
-	ToolInfo           = intelligence.ToolInfo
-	ToolSelection      = intelligence.ToolSelection
-	ToolSelector       = intelligence.ToolSelector
-	CommandSuggestion  = intelligence.CommandSuggestion
-	SuggestionRule     = intelligence.SuggestionRule
-	SuggestionEngine   = intelligence.SuggestionEngine
-)
-
-var (
-	NewIntentClassifier      = intelligence.NewIntentClassifier
-	FormatIntent             = intelligence.FormatIntent
-	NewCapabilityRegistry    = intelligence.NewCapabilityRegistry
-	NewLanguageRegistry      = intelligence.NewLanguageRegistry
-	FormatLanguages          = intelligence.FormatLanguages
-	NewToolSelector          = intelligence.NewToolSelector
-	FormatToolSelection      = intelligence.FormatToolSelection
-	NewSuggestionEngine      = intelligence.NewSuggestionEngine
-	FormatCommandSuggestions = intelligence.FormatCommandSuggestions
-)
diff --git a/internal/engine/io_reexports.go b/internal/engine/io_reexports.go
deleted file mode 100644
index a8831883..00000000
--- a/internal/engine/io_reexports.go
+++ /dev/null
@@ -1,42 +0,0 @@
-package engine
-
-import "github.com/GrayCodeAI/rho/internal/engine/io"
-
-type (
-	ClipboardMonitor  = io.ClipboardMonitor
-	AIComment         = io.AIComment
-	AIWatcher         = io.AIWatcher
-	FileEvent         = io.FileEvent
-	WatcherConfig     = io.WatcherConfig
-	FileWatcher       = io.FileWatcher
-	SingleFileWatcher = io.SingleFileWatcher
-	CronJob           = io.CronJob
-	CronScheduler     = io.CronScheduler
-	CronExpr          = io.CronExpr
-	ClipboardBridge   = io.ClipboardBridge
-)
-
-var (
-	NewClipboardMonitor   = io.NewClipboardMonitor
-	ReadClipboard         = io.ReadClipboard
-	WriteClipboard        = io.WriteClipboard
-	DetectContentType     = io.DetectContentType
-	DetectLanguage        = io.DetectLanguage
-	NewAIWatcher          = io.NewAIWatcher
-	ScanFile              = io.ScanFile
-	ScanDirectory         = io.ScanDirectory
-	BuildPrompt           = io.BuildPrompt
-	RemoveComment         = io.RemoveComment
-	NewFileWatcher        = io.NewFileWatcher
-	WatchSingle           = io.WatchSingle
-	DefaultIgnorePatterns = io.DefaultIgnorePatterns
-	MatchesPattern        = io.MatchesPattern
-	DedupEvents           = io.DedupEvents
-	FormatEvents          = io.FormatEvents
-	NewCronScheduler      = io.NewCronScheduler
-	ParseCron             = io.ParseCron
-	IsDue                 = io.IsDue
-	NextRunTime           = io.NextRunTime
-	SummarizeClipboard    = io.SummarizeClipboard
-	FormatForContext      = io.FormatForContext
-)
diff --git a/internal/engine/lifecycle_finalize_test.go b/internal/engine/lifecycle_finalize_test.go
index 526b619c..a224d9a9 100644
--- a/internal/engine/lifecycle_finalize_test.go
+++ b/internal/engine/lifecycle_finalize_test.go
@@ -5,6 +5,8 @@ import (
 	"testing"
 	"time"
 
+	"github.com/GrayCodeAI/rho/internal/engine/lifecycle"
+
 	"github.com/GrayCodeAI/rho/internal/types"
 )
 
@@ -36,7 +38,7 @@ func (m *finalizeMockSkillStore) Retrieve(query string) []string { return nil }
 // distillation was dead code.
 func TestFinalizePopulatesToolsAndFiles_TriggersSkillDistill(t *testing.T) {
 	sk := &finalizeMockSkillStore{}
-	svc := &LifecycleService{lifecycle: &SessionLifecycle{SkillStore: sk}}
+	svc := &LifecycleService{lifecycle: &lifecycle.SessionLifecycle{SkillStore: sk}}
 
 	messages := []types.FluxMessage{
 		{Role: "user", Content: "Refactor the auth module"},
@@ -73,7 +75,7 @@ func TestFinalizePopulatesToolsAndFiles_TriggersSkillDistill(t *testing.T) {
 // tools/files does not trigger distillation (regression guard).
 func TestFinalizeSimpleTaskNoSkillDistill(t *testing.T) {
 	sk := &finalizeMockSkillStore{}
-	svc := &LifecycleService{lifecycle: &SessionLifecycle{SkillStore: sk}}
+	svc := &LifecycleService{lifecycle: &lifecycle.SessionLifecycle{SkillStore: sk}}
 
 	messages := []types.FluxMessage{
 		{Role: "user", Content: "What is 2+2?"},
diff --git a/internal/engine/lifecycle_reexports.go b/internal/engine/lifecycle_reexports.go
deleted file mode 100644
index cc75d8b6..00000000
--- a/internal/engine/lifecycle_reexports.go
+++ /dev/null
@@ -1,31 +0,0 @@
-// This file re-exports symbols from the lifecycle sub-package so that existing
-// callers of engine.SessionLifecycle, engine.NewLimitTracker, etc. keep compiling
-// during the Stage 2 migration. See docs/plans/engine-refactor-plan.md.
-package engine
-
-import "github.com/GrayCodeAI/rho/internal/engine/lifecycle"
-
-type (
-	SessionLifecycle        = lifecycle.SessionLifecycle
-	EvolvingMemoryInterface = lifecycle.EvolvingMemoryInterface
-	SkillStoreInterface     = lifecycle.SkillStoreInterface
-	CostTrackerInterface    = lifecycle.CostTrackerInterface
-	CostEntry               = lifecycle.CostEntry
-	SessionOutcome          = lifecycle.SessionOutcome
-	EvolvingMemoryAdapter   = lifecycle.EvolvingMemoryAdapter
-	SkillDistillerAdapter   = lifecycle.SkillDistillerAdapter
-	SafetyLimits            = lifecycle.SafetyLimits
-	LimitTracker            = lifecycle.LimitTracker
-	TimeoutConfig           = lifecycle.TimeoutConfig
-)
-
-var (
-	NewLimitTracker      = lifecycle.NewLimitTracker
-	DefaultLimits        = lifecycle.DefaultLimits
-	VibeLimits           = lifecycle.VibeLimits
-	ResearchLimits       = lifecycle.ResearchLimits
-	DefaultTimeoutConfig = lifecycle.DefaultTimeoutConfig
-	WithTimeout          = lifecycle.WithTimeout
-	RemainingTime        = lifecycle.RemainingTime
-	TimeoutMessage       = lifecycle.TimeoutMessage
-)
diff --git a/internal/engine/lifecycle_service.go b/internal/engine/lifecycle_service.go
index 4b5691e3..1784b702 100644
--- a/internal/engine/lifecycle_service.go
+++ b/internal/engine/lifecycle_service.go
@@ -5,6 +5,15 @@ import (
 	"sync"
 	"time"
 
+	"github.com/GrayCodeAI/rho/internal/engine/cost"
+	"github.com/GrayCodeAI/rho/internal/engine/scaffold"
+
+	"github.com/GrayCodeAI/rho/internal/engine/control"
+	"github.com/GrayCodeAI/rho/internal/engine/lifecycle"
+	"github.com/GrayCodeAI/rho/internal/engine/review"
+	"github.com/GrayCodeAI/rho/internal/engine/streaming"
+	"github.com/GrayCodeAI/rho/internal/engine/validation"
+
 	"github.com/GrayCodeAI/rho/internal/engine/branching"
 	"github.com/GrayCodeAI/rho/internal/engine/token"
 	"github.com/GrayCodeAI/rho/internal/intelligence/memory"
@@ -34,40 +43,40 @@ type LifecycleService struct {
 	// smart turn routing (simple/strong per turn).
 	smartRouting *smartrouting.Config
 	// limit tracking.
-	limits *LimitTracker
+	limits *lifecycle.LimitTracker
 	// doom-loop / snowball / loop detection.
-	loopDet  *LoopDetector
+	loopDet  *control.LoopDetector
 	snowball *branching.SnowballDetector
 	// beliefs.
 	beliefs *BeliefState
 	// decision recording.
-	backtrack *BacktrackEngine
+	backtrack *control.BacktrackEngine
 	// post-write critics.
-	critic *Critic
+	critic *review.Critic
 	// pre-edit shadow validation.
 	shadow *branching.ShadowWorkspace
 	// verbal self-reflection on tool failure.
 	reflector *Reflector
 	// few-shot + adaptive prompt.
-	fewShotStore   *FewShotStore
+	fewShotStore   *scaffold.FewShotStore
 	adaptivePrompt *AdaptivePrompt
 	// activity tracker.
 	activity *memory.ActivityTracker
 	// agents accumulator.
 	agentsAccum *prompts.AgentsAccumulator
 	// response cache (used in agentLoop for cache hits).
-	responseCache *ResponseCache
+	responseCache *streaming.ResponseCache
 	// integration pipeline (pre-query / post-response / end-session).
 	pipeline *IntegrationPipeline
 	// steering queue.
-	steering *SteeringQueue
+	steering *streaming.SteeringQueue
 	// local quality loops run after write tools and belong to lifecycle
 	// feedback rather than transport or persistence.
-	lintLoop *LintLoop
-	testLoop *TestLoop
+	lintLoop *validation.LintLoop
+	testLoop *validation.TestLoop
 	// session-level lifecycle hook.
-	lifecycle   *SessionLifecycle
-	costTracker *CostTracker
+	lifecycle   *lifecycle.SessionLifecycle
+	costTracker *cost.CostTracker
 	teach       TeachConfig
 	trajectory  *TrajectoryDistiller
 	// smartSkills caches loaded SmartSkills for auto-discovery per-turn.
@@ -86,13 +95,13 @@ func NewLifecycleService(log *logger.Logger) *LifecycleService {
 		log = logger.Default()
 	}
 	return &LifecycleService{
-		limits:         NewLimitTracker(DefaultLimits()),
-		loopDet:        NewLoopDetector(10, DoomLoopThreshold),
+		limits:         lifecycle.NewLimitTracker(lifecycle.DefaultLimits()),
+		loopDet:        control.NewLoopDetector(10, control.DoomLoopThreshold),
 		snowball:       branching.NewSnowballDetector(500000),
 		beliefs:        NewBeliefState(),
-		backtrack:      NewBacktrackEngine(),
+		backtrack:      control.NewBacktrackEngine(),
 		lifecycle:      nil, // constructed in New() with cwd
-		responseCache:  NewResponseCache(1000, 24*time.Hour),
+		responseCache:  streaming.NewResponseCache(1000, 24*time.Hour),
 		pipeline:       NewIntegrationPipeline(),
 		log:            log,
 		fewShotStore:   nil, // lazy
@@ -119,7 +128,7 @@ func (s *LifecycleService) OnSessionStart(ctx context.Context, s2 *Session, last
 // adaptive-prompt learning feedback.
 func (s *LifecycleService) OnSessionEnd(ctx context.Context, s2 *Session, success bool, duration time.Duration) {
 	if s.lifecycle != nil {
-		outcome := SessionOutcome{Success: success, Duration: duration}
+		outcome := lifecycle.SessionOutcome{Success: success, Duration: duration}
 		messages := s2.Persistence().RawMessages()
 		if len(messages) > 0 {
 			for _, m := range messages {
@@ -155,7 +164,7 @@ func (s *LifecycleService) Finalize(ctx context.Context, messages []types.FluxMe
 	if s == nil {
 		return
 	}
-	outcome := SessionOutcome{Success: success, Duration: duration, TotalCost: totalCost}
+	outcome := lifecycle.SessionOutcome{Success: success, Duration: duration, TotalCost: totalCost}
 	for _, message := range messages {
 		if message.Role == "user" && len(message.ToolResults) == 0 && outcome.TaskGoal == "" {
 			outcome.TaskGoal = message.Content
@@ -252,28 +261,28 @@ func (s *LifecycleService) SnapshotTurnProgress(tokens int, progress float64) {
 
 func (s *LifecycleService) SetCascade(c *branching.CascadeRouter)       { s.cascade = c }
 func (s *LifecycleService) SetSmartRouting(c *smartrouting.Config)      { s.smartRouting = c }
-func (s *LifecycleService) SetLifecycle(l *SessionLifecycle)            { s.lifecycle = l }
+func (s *LifecycleService) SetLifecycle(l *lifecycle.SessionLifecycle)  { s.lifecycle = l }
 func (s *LifecycleService) SetReflector(r *Reflector)                   { s.reflector = r }
-func (s *LifecycleService) SetCritic(c *Critic)                         { s.critic = c }
+func (s *LifecycleService) SetCritic(c *review.Critic)                  { s.critic = c }
 func (s *LifecycleService) SetShadow(sh *branching.ShadowWorkspace)     { s.shadow = sh }
-func (s *LifecycleService) SetFewShotStore(f *FewShotStore)             { s.fewShotStore = f }
+func (s *LifecycleService) SetFewShotStore(f *scaffold.FewShotStore)    { s.fewShotStore = f }
 func (s *LifecycleService) SetAdaptivePrompt(a *AdaptivePrompt)         { s.adaptivePrompt = a }
 func (s *LifecycleService) SetActivity(act *memory.ActivityTracker)     { s.activity = act }
 func (s *LifecycleService) SetAgentsAccum(a *prompts.AgentsAccumulator) { s.agentsAccum = a }
-func (s *LifecycleService) SetSteering(st *SteeringQueue)               { s.steering = st }
-func (s *LifecycleService) SetLintLoop(loop *LintLoop)                  { s.lintLoop = loop }
-func (s *LifecycleService) SetTestLoop(loop *TestLoop)                  { s.testLoop = loop }
+func (s *LifecycleService) SetSteering(st *streaming.SteeringQueue)     { s.steering = st }
+func (s *LifecycleService) SetLintLoop(loop *validation.LintLoop)       { s.lintLoop = loop }
+func (s *LifecycleService) SetTestLoop(loop *validation.TestLoop)       { s.testLoop = loop }
 
 // Accessors used by stream.go and the agent loop. nil-safe.
 func (s *LifecycleService) Beliefs() *BeliefState                   { return s.beliefs }
-func (s *LifecycleService) Backtrack() *BacktrackEngine             { return s.backtrack }
-func (s *LifecycleService) Limits() *LimitTracker                   { return s.limits }
-func (s *LifecycleService) Critic() *Critic                         { return s.critic }
+func (s *LifecycleService) Backtrack() *control.BacktrackEngine     { return s.backtrack }
+func (s *LifecycleService) Limits() *lifecycle.LimitTracker         { return s.limits }
+func (s *LifecycleService) Critic() *review.Critic                  { return s.critic }
 func (s *LifecycleService) Shadow() *branching.ShadowWorkspace      { return s.shadow }
 func (s *LifecycleService) Reflector() *Reflector                   { return s.reflector }
 func (s *LifecycleService) Cascade() *branching.CascadeRouter       { return s.cascade }
 func (s *LifecycleService) SmartRouting() *smartrouting.Config      { return s.smartRouting }
-func (s *LifecycleService) FewShotStore() *FewShotStore             { return s.fewShotStore }
+func (s *LifecycleService) FewShotStore() *scaffold.FewShotStore    { return s.fewShotStore }
 func (s *LifecycleService) AdaptivePrompt() *AdaptivePrompt         { return s.adaptivePrompt }
 func (s *LifecycleService) Activity() *memory.ActivityTracker       { return s.activity }
 func (s *LifecycleService) AgentsAccum() *prompts.AgentsAccumulator { return s.agentsAccum }
@@ -281,16 +290,16 @@ func (s *LifecycleService) AgentsAccum() *prompts.AgentsAccumulator { return s.a
 // SetAgentsAccumulator attaches the project-learning accumulator.
 func (s *LifecycleService) SetAgentsAccumulator(a *prompts.AgentsAccumulator) { s.agentsAccum = a }
 
-func (s *LifecycleService) ResponseCache() *ResponseCache        { return s.responseCache }
-func (s *LifecycleService) Pipeline() *IntegrationPipeline       { return s.pipeline }
-func (s *LifecycleService) Steering() *SteeringQueue             { return s.steering }
-func (s *LifecycleService) Lifecycle() *SessionLifecycle         { return s.lifecycle }
-func (s *LifecycleService) CostTracker() *CostTracker            { return s.costTracker }
-func (s *LifecycleService) SetCostTracker(c *CostTracker)        { s.costTracker = c }
-func (s *LifecycleService) Teach() TeachConfig                   { return s.teach }
-func (s *LifecycleService) SetTeach(t TeachConfig)               { s.teach = t }
-func (s *LifecycleService) Trajectory() *TrajectoryDistiller     { return s.trajectory }
-func (s *LifecycleService) SetTrajectory(t *TrajectoryDistiller) { s.trajectory = t }
+func (s *LifecycleService) ResponseCache() *streaming.ResponseCache { return s.responseCache }
+func (s *LifecycleService) Pipeline() *IntegrationPipeline          { return s.pipeline }
+func (s *LifecycleService) Steering() *streaming.SteeringQueue      { return s.steering }
+func (s *LifecycleService) Lifecycle() *lifecycle.SessionLifecycle  { return s.lifecycle }
+func (s *LifecycleService) CostTracker() *cost.CostTracker          { return s.costTracker }
+func (s *LifecycleService) SetCostTracker(c *cost.CostTracker)      { s.costTracker = c }
+func (s *LifecycleService) Teach() TeachConfig                      { return s.teach }
+func (s *LifecycleService) SetTeach(t TeachConfig)                  { s.teach = t }
+func (s *LifecycleService) Trajectory() *TrajectoryDistiller        { return s.trajectory }
+func (s *LifecycleService) SetTrajectory(t *TrajectoryDistiller)    { s.trajectory = t }
 
 // LoadSmartSkills loads the session's auto-discovery skills once.
 func (s *LifecycleService) LoadSmartSkills() {
@@ -359,9 +368,9 @@ func (s *LifecycleService) ToggleVerbose() bool {
 	s.verbose = !s.verbose
 	return s.verbose
 }
-func (s *LifecycleService) Verbose() bool       { return s != nil && s.verbose }
-func (s *LifecycleService) LintLoop() *LintLoop { return s.lintLoop }
-func (s *LifecycleService) TestLoop() *TestLoop { return s.testLoop }
+func (s *LifecycleService) Verbose() bool                  { return s != nil && s.verbose }
+func (s *LifecycleService) LintLoop() *validation.LintLoop { return s.lintLoop }
+func (s *LifecycleService) TestLoop() *validation.TestLoop { return s.testLoop }
 
 // containsStringVec reports whether s is present in the slice.
 func containsStringVec(slice []string, s string) bool {
diff --git a/internal/engine/memory_reexports.go b/internal/engine/memory_reexports.go
deleted file mode 100644
index a5494117..00000000
--- a/internal/engine/memory_reexports.go
+++ /dev/null
@@ -1,22 +0,0 @@
-package engine
-
-import "github.com/GrayCodeAI/rho/internal/engine/memory"
-
-type (
-	KnowledgeEntry     = memory.KnowledgeEntry
-	KnowledgeStats     = memory.KnowledgeStats
-	KnowledgeBase      = memory.KnowledgeBase
-	Experience         = memory.Experience
-	ExperienceStats    = memory.ExperienceStats
-	ExperienceStore    = memory.ExperienceStore
-	RawMemory          = memory.RawMemory
-	ConsolidatedMemory = memory.ConsolidatedMemory
-	ConsolidatorStats  = memory.ConsolidatorStats
-	MemoryConsolidator = memory.MemoryConsolidator
-)
-
-var (
-	NewKnowledgeBase      = memory.NewKnowledgeBase
-	NewExperienceStore    = memory.NewExperienceStore
-	NewMemoryConsolidator = memory.NewMemoryConsolidator
-)
diff --git a/internal/engine/observability_reexports.go b/internal/engine/observability_reexports.go
deleted file mode 100644
index da3b6e66..00000000
--- a/internal/engine/observability_reexports.go
+++ /dev/null
@@ -1,34 +0,0 @@
-// This file re-exports symbols from the observability sub-package so that existing
-// callers of engine.Profiler, engine.NewProfiler, etc. keep compiling
-// during the Stage 2 migration. See docs/plans/engine-refactor-plan.md.
-package engine
-
-import "github.com/GrayCodeAI/rho/internal/engine/observability"
-
-type (
-	Profiler          = observability.Profiler
-	ProfileSpan       = observability.ProfileSpan
-	Counter           = observability.Counter
-	Timer             = observability.Timer
-	DebugSession      = observability.DebugSession
-	DebugStep         = observability.DebugStep
-	Hypothesis        = observability.Hypothesis
-	DebugRecorder     = observability.DebugRecorder
-	LogLevel          = observability.LogLevel
-	LogEntry          = observability.LogEntry
-	StructuredLogger  = observability.StructuredLogger
-	AgentLogger       = observability.AgentLogger
-	RotatingWriter    = observability.RotatingWriter
-	Feedback          = observability.Feedback
-	FeedbackCollector = observability.FeedbackCollector
-	ImplicitSignal    = observability.ImplicitSignal
-)
-
-var (
-	NewProfiler          = observability.NewProfiler
-	NewDebugRecorder     = observability.NewDebugRecorder
-	NewStructuredLogger  = observability.NewStructuredLogger
-	ParseLevel           = observability.ParseLevel
-	NewRotatingWriter    = observability.NewRotatingWriter
-	NewFeedbackCollector = observability.NewFeedbackCollector
-)
diff --git a/internal/engine/persistence_service.go b/internal/engine/persistence_service.go
index 1f49a339..2e6c75f8 100644
--- a/internal/engine/persistence_service.go
+++ b/internal/engine/persistence_service.go
@@ -4,6 +4,9 @@ import (
 	"strings"
 	"sync"
 
+	"github.com/GrayCodeAI/rho/internal/engine/compact"
+	"github.com/GrayCodeAI/rho/internal/engine/streaming"
+
 	"github.com/GrayCodeAI/rho/internal/eventlog"
 	"github.com/GrayCodeAI/rho/internal/observability/logger"
 	"github.com/GrayCodeAI/rho/internal/session"
@@ -46,13 +49,13 @@ type PersistenceService struct {
 	// at explicit Save calls).
 	writeBehind *session.WriteBehind
 	// steering is the per-iteration user-guidance queue.
-	steering *SteeringQueue
+	steering *streaming.SteeringQueue
 	// logger.
 	log *logger.Logger
 	// Compaction and checkpoint state belongs to persistence. Session keeps
 	// deprecated aliases only for source compatibility with older callers.
 	autoCompactor        *AutoCompactor
-	files                *FileTracker
+	files                *compact.FileTracker
 	persistID            string
 	lastPromptTokens     int
 	lastCompletionTokens int
@@ -154,10 +157,10 @@ func (s *PersistenceService) SetGraph(graph *session.ConversationGraph) { s.grap
 
 // Steering returns the per-iteration user-guidance queue. New
 // code should access this through s.Persistence().Steering().
-func (s *PersistenceService) Steering() *SteeringQueue { return s.steering }
+func (s *PersistenceService) Steering() *streaming.SteeringQueue { return s.steering }
 
 // SetSteering attaches the user-guidance queue.
-func (s *PersistenceService) SetSteering(sq *SteeringQueue) { s.steering = sq }
+func (s *PersistenceService) SetSteering(sq *streaming.SteeringQueue) { s.steering = sq }
 
 // AddAssistant appends an assistant message.
 func (s *PersistenceService) AddAssistant(content string) {
@@ -538,13 +541,13 @@ func (s *PersistenceService) SetAutoCompactor(ac *AutoCompactor) {
 	s.stateMu.Unlock()
 }
 
-func (s *PersistenceService) Files() *FileTracker {
+func (s *PersistenceService) Files() *compact.FileTracker {
 	s.stateMu.RLock()
 	defer s.stateMu.RUnlock()
 	return s.files
 }
 
-func (s *PersistenceService) SetFiles(files *FileTracker) {
+func (s *PersistenceService) SetFiles(files *compact.FileTracker) {
 	s.stateMu.Lock()
 	s.files = files
 	s.stateMu.Unlock()
diff --git a/internal/engine/project_reexports.go b/internal/engine/project_reexports.go
deleted file mode 100644
index c853db2f..00000000
--- a/internal/engine/project_reexports.go
+++ /dev/null
@@ -1,59 +0,0 @@
-// This file re-exports symbols from the project sub-package so that existing
-// callers of engine.ProjectAnalysis, engine.NewProjectAnalyzer, etc. keep compiling
-// during the Stage 2 migration. See docs/plans/engine-refactor-plan.md.
-package engine
-
-import "github.com/GrayCodeAI/rho/internal/engine/project"
-
-type (
-	ProjectAnalysis      = project.ProjectAnalysis
-	ModuleInfo           = project.ModuleInfo
-	Pattern              = project.Pattern
-	ProjectAnalyzer      = project.ProjectAnalyzer
-	ProjectContext       = project.ProjectContext
-	ProjectSnapshot      = project.ProjectSnapshot
-	ProjectSnapshotCache = project.ProjectSnapshotCache
-	ImpactAnalysis       = project.ImpactAnalysis
-	ImpactAnalyzer       = project.ImpactAnalyzer
-	DependencyUpdater    = project.DependencyUpdater
-	Dependency           = project.Dependency
-	UpdatePlan           = project.UpdatePlan
-	MigrationPlan        = project.MigrationPlan
-	MigrationStep        = project.MigrationStep
-	MigrationResult      = project.MigrationResult
-	MigrationPlanner     = project.MigrationPlanner
-	ReleaseManager       = project.ReleaseManager
-	Release              = project.Release
-	ChangeEntry          = project.ChangeEntry
-	ReleaseStats         = project.ReleaseStats
-	Convention           = project.Convention
-	ConventionSet        = project.ConventionSet
-	Violation            = project.Violation
-)
-
-var (
-	NewProjectAnalyzer      = project.NewProjectAnalyzer
-	DetectArchitecture      = project.DetectArchitecture
-	DetectPatterns          = project.DetectPatterns
-	GenerateOnboardingDoc   = project.GenerateOnboardingDoc
-	FormatAnalysis          = project.FormatAnalysis
-	NewProjectContext       = project.NewProjectContext
-	NewProjectSnapshotCache = project.NewProjectSnapshotCache
-	NewImpactAnalyzer       = project.NewImpactAnalyzer
-	BuildImportGraph        = project.BuildImportGraph
-	FormatImpact            = project.FormatImpact
-	NewDependencyUpdater    = project.NewDependencyUpdater
-	ClassifyUpdate          = project.ClassifyUpdate
-	ParseSemver             = project.ParseSemver
-	FormatOutdated          = project.FormatOutdated
-	FormatPlan              = project.FormatPlan
-	NewMigrationPlanner     = project.NewMigrationPlanner
-	NewReleaseManager       = project.NewReleaseManager
-	ParseConventionalCommit = project.ParseConventionalCommit
-	BumpVersion             = project.BumpVersion
-	GenerateChangelog       = project.GenerateChangelog
-	FormatReleaseNotes      = project.FormatReleaseNotes
-	UpdateVersionFile       = project.UpdateVersionFile
-	NewConventionSet        = project.NewConventionSet
-	FormatViolations        = project.FormatViolations
-)
diff --git a/internal/engine/review_reexports.go b/internal/engine/review_reexports.go
deleted file mode 100644
index fc7e47e8..00000000
--- a/internal/engine/review_reexports.go
+++ /dev/null
@@ -1,62 +0,0 @@
-package engine
-
-import "github.com/GrayCodeAI/rho/internal/engine/review"
-
-// Types from review sub-package.
-
-type (
-	ReviewBot        = review.ReviewBot
-	ReviewRule       = review.ReviewRule
-	ReviewComment    = review.ReviewComment
-	ReviewReport     = review.ReviewReport
-	PatchVerdict     = review.PatchVerdict
-	Critic           = review.Critic
-	Assessment       = review.Assessment
-	SelfAssessor     = review.SelfAssessor
-	TaskContext      = review.TaskContext
-	ConsensusSampler = review.ConsensusSampler
-	Sample           = review.Sample
-	ConsensusResult  = review.ConsensusResult
-	QualityScorer    = review.QualityScorer
-	ScoreWeights     = review.ScoreWeights
-	ScoredResponse   = review.ScoredResponse
-	ResponseContext  = review.ResponseContext
-	SolutionReviewer = review.SolutionReviewer
-	Solution         = review.Solution
-	ReviewResult     = review.ReviewResult
-)
-
-// Short-name aliases.
-
-type (
-	Bot     = review.ReviewBot
-	Rule    = review.ReviewRule
-	Comment = review.ReviewComment
-	Report  = review.ReviewReport
-)
-
-// Functions.
-
-var (
-	NewReviewBot         = review.NewReviewBot
-	NewCritic            = review.NewCritic
-	NewSelfAssessor      = review.NewSelfAssessor
-	NewConsensusSampler  = review.NewConsensusSampler
-	NewQualityScorer     = review.NewQualityScorer
-	NewSolutionReviewer  = review.NewSolutionReviewer
-	DefaultWeights       = review.DefaultWeights
-	FormatSelfAssessment = review.FormatSelfAssessment
-	FormatConsensus      = review.FormatConsensus
-	CompareApproaches    = review.CompareApproaches
-	FormatReview         = review.FormatReview
-	ShouldRetry          = review.ShouldRetry
-	FormatReport         = review.FormatReport
-	FormatInline         = review.FormatInline
-	FilterBySeverity     = review.FilterBySeverity
-)
-
-// NewBot delegates to review.NewBot.
-func NewBot() *Bot { return review.NewBot() }
-
-// ReviewBeforeWrite, LLMClient, SelfReviewResult, ConfidenceThreshold
-// are defined directly in package engine (self_review.go, client_interface.go).
diff --git a/internal/engine/scaffold_reexports.go b/internal/engine/scaffold_reexports.go
deleted file mode 100644
index b55d352d..00000000
--- a/internal/engine/scaffold_reexports.go
+++ /dev/null
@@ -1,28 +0,0 @@
-package engine
-
-import "github.com/GrayCodeAI/rho/internal/engine/scaffold"
-
-type (
-	Template         = scaffold.Template
-	TemplateFile     = scaffold.TemplateFile
-	TemplateVariable = scaffold.TemplateVariable
-	Scaffolder       = scaffold.Scaffolder
-	Recipe           = scaffold.Recipe
-	RecipeRegistry   = scaffold.RecipeRegistry
-	PromptPattern    = scaffold.PromptPattern
-	PatternLibrary   = scaffold.PatternLibrary
-	Skill            = scaffold.Skill
-	SkillStep        = scaffold.SkillStep
-	SkillResult      = scaffold.SkillResult
-	SkillRegistry    = scaffold.SkillRegistry
-	FewShotStore     = scaffold.FewShotStore
-	FewShotExample   = scaffold.FewShotExample
-)
-
-func NewScaffolder() *Scaffolder                   { return scaffold.NewScaffolder() }
-func NewRecipeRegistry(dir string) *RecipeRegistry { return scaffold.NewRecipeRegistry(dir) }
-func NewPatternLibrary(dir string) *PatternLibrary { return scaffold.NewPatternLibrary(dir) }
-func NewSkillRegistry(dir string) *SkillRegistry   { return scaffold.NewSkillRegistry(dir) }
-func NewFewShotStore() *FewShotStore               { return scaffold.NewFewShotStore() }
-func FormatPattern(p *PromptPattern) string        { return scaffold.FormatPattern(p) }
-func FormatSkill(s *Skill) string                  { return scaffold.FormatSkill(s) }
diff --git a/internal/engine/search_reexports.go b/internal/engine/search_reexports.go
deleted file mode 100644
index ed2a87fa..00000000
--- a/internal/engine/search_reexports.go
+++ /dev/null
@@ -1,28 +0,0 @@
-package engine
-
-import "github.com/GrayCodeAI/rho/internal/engine/search"
-
-type (
-	URLScraper      = search.URLScraper
-	ScrapeResult    = search.ScrapeResult
-	Issue           = search.Issue
-	SimilarIssue    = search.SimilarIssue
-	IssueIndex      = search.IssueIndex
-	ResearchAgent   = search.ResearchAgent
-	ResearchQuery   = search.ResearchQuery
-	ResearchResult  = search.ResearchResult
-	ResearchFinding = search.ResearchFinding
-)
-
-var (
-	NewURLScraper      = search.NewURLScraper
-	NewIssueIndex      = search.NewIssueIndex
-	NewResearchAgent   = search.NewResearchAgent
-	ExtractHTML        = search.ExtractHTML
-	ExtractJSON        = search.ExtractJSON
-	ExtractMarkdown    = search.ExtractMarkdown
-	ExtractCode        = search.ExtractCode
-	SuggestResolution  = search.SuggestResolution
-	FormatIssueResults = search.FormatIssueResults
-	BuildSearchContext = search.BuildSearchContext
-)
diff --git a/internal/engine/session.go b/internal/engine/session.go
index 2ad62654..6c38b242 100644
--- a/internal/engine/session.go
+++ b/internal/engine/session.go
@@ -10,6 +10,11 @@ import (
 	"sync"
 	"time"
 
+	"github.com/GrayCodeAI/rho/internal/engine/cost"
+
+	"github.com/GrayCodeAI/rho/internal/engine/streaming"
+	"github.com/GrayCodeAI/rho/internal/engine/validation"
+
 	agentcontracts "github.com/GrayCodeAI/rho/internal/contracts/agent"
 	"github.com/GrayCodeAI/rho/internal/conversationarc"
 	"github.com/GrayCodeAI/rho/internal/engine/planning"
@@ -64,7 +69,7 @@ type SnapshotTracker interface {
 // and lifecycle state are owned by the corresponding services below.
 type Session struct {
 	mu   sync.RWMutex
-	Cost Cost
+	Cost cost.Cost
 
 	// llm is the LLM transport service (Phase 1 extraction). All new
 	// code should go through s.llm.* rather than duplicating transport state.
@@ -247,8 +252,8 @@ func NewSessionWithClient(chat ChatClient, provider, model, systemPrompt string,
 	})
 	s.refreshContextWindowCache()
 	s.life.SetAgentsAccumulator(agentsAccum)
-	s.life.SetLintLoop(NewLintLoop())
-	s.life.SetTestLoop(NewTestLoop())
+	s.life.SetLintLoop(validation.NewLintLoop())
+	s.life.SetTestLoop(validation.NewTestLoop())
 	s.promptQueue = NewPromptQueue()
 	s.announcements = NewAnnouncementFeed()
 	return s
@@ -1037,7 +1042,7 @@ func (s *Session) EnsureSkillCatalogStatement() string {
 // CostValue returns the session's cost accumulator (a pointer
 // to a value type, so its methods can be called). New code
 // should call this instead of reading s.Cost directly.
-func (s *Session) CostValue() *Cost {
+func (s *Session) CostValue() *cost.Cost {
 	return &s.Cost
 }
 
@@ -1088,7 +1093,7 @@ func (s *Session) Schedule() *schedule.Manager {
 			content := fmt.Sprintf("[Scheduled Reminder: %s]\n%s", item.ID, item.Prompt)
 			if p := s.Persistence(); p != nil {
 				if sq := p.Steering(); sq != nil {
-					sq.Enqueue(SteeringMessage{
+					sq.Enqueue(streaming.SteeringMessage{
 						Content:  content,
 						Priority: 1,
 					})
diff --git a/internal/engine/session_reexports.go b/internal/engine/session_reexports.go
deleted file mode 100644
index 2558fe30..00000000
--- a/internal/engine/session_reexports.go
+++ /dev/null
@@ -1,27 +0,0 @@
-package engine
-
-import "github.com/GrayCodeAI/rho/internal/engine/compression"
-
-type (
-	Timeline            = compression.Timeline
-	TimelineEvent       = compression.TimelineEvent
-	SessionCompressor   = compression.SessionCompressor
-	CompressStrategy    = compression.CompressStrategy
-	CompressMessage     = compression.CompressMessage
-	CompressedBlock     = compression.CompressedBlock
-	CompressionResult   = compression.CompressionResult
-	CrossSessionLearner = compression.CrossSessionLearner
-	Insight             = compression.Insight
-	FailurePattern      = compression.FailurePattern
-	SessionConvention   = compression.SessionConvention
-	LearnerStats        = compression.LearnerStats
-)
-
-func NewTimeline(sessionID string) *Timeline { return compression.NewTimeline(sessionID) }
-func NewSessionCompressor(strategy CompressStrategy) *SessionCompressor {
-	return compression.NewSessionCompressor(strategy)
-}
-
-func NewCrossSessionLearner(dir string) *CrossSessionLearner {
-	return compression.NewCrossSessionLearner(dir)
-}
diff --git a/internal/engine/stream.go b/internal/engine/stream.go
index 0b406b6f..8a3d865f 100644
--- a/internal/engine/stream.go
+++ b/internal/engine/stream.go
@@ -8,6 +8,11 @@ import (
 	"strings"
 	"time"
 
+	"github.com/GrayCodeAI/rho/internal/engine/cost"
+	"github.com/GrayCodeAI/rho/internal/engine/token"
+
+	"github.com/GrayCodeAI/rho/internal/engine/control"
+
 	"github.com/GrayCodeAI/rho/internal/provider/gateway"
 	"github.com/GrayCodeAI/rho/internal/smartrouting"
 	"github.com/GrayCodeAI/rho/internal/types"
@@ -228,8 +233,8 @@ func (s *Session) agentLoop(ctx context.Context, ch chan<- StreamEvent) {
 	toolTurns := 0 // turns that used tools (for skill distillation)
 	var toolsUsedSet map[string]bool
 	var filesModifiedSet map[string]bool
-	snowball := branching.NewSnowballDetector(500000) // 500K token ceiling
-	loopDet := NewLoopDetector(10, DoomLoopThreshold) // 10-step window, 3 repeats = doom loop
+	snowball := branching.NewSnowballDetector(500000)                 // 500K token ceiling
+	loopDet := control.NewLoopDetector(10, control.DoomLoopThreshold) // 10-step window, 3 repeats = doom loop
 
 	for {
 		// Close the boundary of the previous turn before evaluating whether to
@@ -259,12 +264,12 @@ func (s *Session) agentLoop(ctx context.Context, ch chan<- StreamEvent) {
 			s.LifecycleSvc().Beliefs().Prune(turnCount)
 		}
 		// Context governor: collapse → micro/smart/truncate (settings threshold %).
-		tokensBefore := EstimateTokens(s.Persistence().RawMessages())
+		tokensBefore := token.EstimateTokens(s.Persistence().RawMessages())
 		if s.WillCompactBeforeTurn() {
 			emit(StreamEvent{Type: "compact_start"})
 		}
 		if compactStrategy, didCompact := s.ManageContextBeforeTurn(ctx); didCompact {
-			tokensAfter := EstimateTokens(s.Persistence().RawMessages())
+			tokensAfter := token.EstimateTokens(s.Persistence().RawMessages())
 			s.Logger().Info("context compacted", map[string]interface{}{
 				"strategy": compactStrategy,
 				"messages": len(s.Persistence().RawMessages()),
@@ -337,7 +342,7 @@ func (s *Session) agentLoop(ctx context.Context, ch chan<- StreamEvent) {
 		// Dynamic max_tokens based on task type and recent tool patterns
 		taskType := classifyPromptForBudget(s.Persistence().RawMessages())
 		contextSize := s.ContextWindowSize()
-		maxTok := DynamicMaxTokens(s.Persistence().RawMessages(), contextSize, taskType)
+		maxTok := token.DynamicMaxTokens(s.Persistence().RawMessages(), contextSize, taskType)
 
 		// Model cascade: select optimal model for this request
 		activeModel := strings.TrimSpace(s.ChatLLM().Model())
@@ -426,16 +431,16 @@ func (s *Session) agentLoop(ctx context.Context, ch chan<- StreamEvent) {
 		// Count actual input tokens for precise budget tracking
 		inputTokens := 0
 		for _, msg := range s.Persistence().RawMessages() {
-			inputTokens += CountTokensFast(msg.Content)
+			inputTokens += token.CountTokensFast(msg.Content)
 			for _, tr := range msg.ToolResults {
-				inputTokens += CountTokensFast(tr.Content)
+				inputTokens += token.CountTokensFast(tr.Content)
 			}
 		}
-		inputTokens += CountTokensFast(s.Persistence().System())
+		inputTokens += token.CountTokensFast(s.Persistence().System())
 		s.Logger().Info("token count", map[string]interface{}{"input_tokens": inputTokens, "model": s.ChatLLM().Model()})
 
 		// Cost warning for expensive calls
-		inPrice, outPrice := ModelPricing(s.ChatLLM().Model())
+		inPrice, outPrice := cost.ModelPricing(s.ChatLLM().Model())
 		estCost := float64(inputTokens)*inPrice/1_000_000 + float64(maxTok)*outPrice/1_000_000
 		if estCost > 0.50 {
 			emit(StreamEvent{Type: "blast_radius", Content: fmt.Sprintf("%s This request will use ~%d tokens (~$%.2f). Continue? The agent will proceed automatically.", icons.Alert(), inputTokens+maxTok, estCost)})
diff --git a/internal/engine/stream_guards.go b/internal/engine/stream_guards.go
index c71f2df6..4ee218d9 100644
--- a/internal/engine/stream_guards.go
+++ b/internal/engine/stream_guards.go
@@ -6,6 +6,8 @@ import (
 	"fmt"
 	"os"
 
+	"github.com/GrayCodeAI/rho/internal/engine/control"
+
 	"github.com/GrayCodeAI/rho/internal/engine/branching"
 )
 
@@ -17,7 +19,7 @@ const gracefulExhaustionEnv = "RHO_GRACEFUL_EXHAUSTION"
 // checkGuardConditions runs all pre-turn guard checks.
 // Returns false when the loop should stop (abort conditions met).
 // On first-stage loop detection, injects a break-loop message and continues.
-func (s *Session) checkGuardConditions(ctx context.Context, ch chan<- StreamEvent, turnCount int, snowball *branching.SnowballDetector, loopDet *LoopDetector) bool {
+func (s *Session) checkGuardConditions(ctx context.Context, ch chan<- StreamEvent, turnCount int, snowball *branching.SnowballDetector, loopDet *control.LoopDetector) bool {
 	if ctx.Err() != nil {
 		msg := "Request cancelled."
 		if errors.Is(ctx.Err(), context.DeadlineExceeded) {
diff --git a/internal/engine/stream_usage.go b/internal/engine/stream_usage.go
index 19bf24bf..a66f300e 100644
--- a/internal/engine/stream_usage.go
+++ b/internal/engine/stream_usage.go
@@ -6,6 +6,8 @@ import (
 	"strings"
 	"time"
 
+	"github.com/GrayCodeAI/rho/internal/engine/token"
+
 	"github.com/GrayCodeAI/rho/internal/types"
 	"github.com/GrayCodeAI/rho/internal/ui/icons"
 	"github.com/GrayCodeAI/rho/internal/usage"
@@ -131,12 +133,12 @@ func (s *Session) recordStreamUsage(ch chan<- StreamEvent, prompt, completion in
 }
 
 func estimateStreamCompletionTokens(text string, toolCalls []types.ToolCall) int {
-	n := CountTokensFast(text)
+	n := token.CountTokensFast(text)
 	for _, tc := range toolCalls {
 		if b, err := json.Marshal(tc.Arguments); err == nil {
-			n += CountTokensFast(string(b))
+			n += token.CountTokensFast(string(b))
 		}
-		n += CountTokensFast(tc.Name)
+		n += token.CountTokensFast(tc.Name)
 	}
 	return n
 }
diff --git a/internal/engine/stream_usage_test.go b/internal/engine/stream_usage_test.go
index 72602ef9..4830cda4 100644
--- a/internal/engine/stream_usage_test.go
+++ b/internal/engine/stream_usage_test.go
@@ -4,6 +4,8 @@ import (
 	"testing"
 	"time"
 
+	"github.com/GrayCodeAI/rho/internal/engine/cost"
+
 	"github.com/GrayCodeAI/rho/internal/types"
 )
 
@@ -53,7 +55,7 @@ func TestRecordStreamUsageAttributesResolvedRoute(t *testing.T) {
 		resolvedProvider = "openai"
 		resolvedModel    = "openai/fallback-test-model"
 	)
-	RegisterLivePricing(resolvedModel, 2, 4)
+	cost.RegisterLivePricing(resolvedModel, 2, 4)
 	sess := NewSessionWithClient(NewMockClientForTest(), "anthropic", "anthropic/requested-test-model", "", nil, true)
 	ch := make(chan StreamEvent, 1)
 
diff --git a/internal/engine/streaming_reexports.go b/internal/engine/streaming_reexports.go
deleted file mode 100644
index 10b56051..00000000
--- a/internal/engine/streaming_reexports.go
+++ /dev/null
@@ -1,37 +0,0 @@
-// This file re-exports symbols from the streaming sub-package so that existing
-// callers of engine.ResponseCache, engine.NewResponseCache, etc. keep compiling
-// during the Stage 2 migration. See docs/plans/engine-refactor-plan.md.
-package engine
-
-import "github.com/GrayCodeAI/rho/internal/engine/streaming"
-
-type (
-	CacheEntry        = streaming.CacheEntry
-	CacheStats        = streaming.CacheStats
-	ResponseCache     = streaming.ResponseCache
-	FormatRule        = streaming.FormatRule
-	FormattedResponse = streaming.FormattedResponse
-	ResponseFormatter = streaming.ResponseFormatter
-	StreamOptimizer   = streaming.StreamOptimizer
-	StreamStats       = streaming.StreamStats
-	ThinkingPhase     = streaming.ThinkingPhase
-	ThinkingStep      = streaming.ThinkingStep
-	ThinkingProtocol  = streaming.ThinkingProtocol
-	SteeringQueue     = streaming.SteeringQueue
-	SteeringMessage   = streaming.SteeringMessage
-)
-
-const DefaultMaxEntries = streaming.DefaultMaxEntries
-
-var (
-	NewResponseCache     = streaming.NewResponseCache
-	NewResponseFormatter = streaming.NewResponseFormatter
-	NewStreamOptimizer   = streaming.NewStreamOptimizer
-	NewThinkingProtocol  = streaming.NewThinkingProtocol
-	NewSteeringQueue     = streaming.NewSteeringQueue
-	HashPrompt           = streaming.HashPrompt
-	ShouldCache          = streaming.ShouldCache
-	FixCodeFences        = streaming.FixCodeFences
-	RemoveFluff          = streaming.RemoveFluff
-	FixMarkdown          = streaming.FixMarkdown
-)
diff --git a/internal/engine/token_reexports.go b/internal/engine/token_reexports.go
deleted file mode 100644
index c970cde6..00000000
--- a/internal/engine/token_reexports.go
+++ /dev/null
@@ -1,37 +0,0 @@
-package engine
-
-import (
-	"github.com/GrayCodeAI/rho/internal/engine/token"
-	"github.com/GrayCodeAI/rho/internal/types"
-)
-
-type (
-	TokenPredictor   = token.TokenPredictor
-	PredictionRecord = token.PredictionRecord
-	Prediction       = token.Prediction
-	TokenEntry       = token.TokenEntry
-	BudgetAlert      = token.BudgetAlert
-	TokenReporter    = token.TokenReporter
-)
-
-func NewTokenPredictor() *TokenPredictor                { return token.NewTokenPredictor() }
-func NewTokenReporter(sessionBudget int) *TokenReporter { return token.NewTokenReporter(sessionBudget) }
-
-func DynamicMaxTokens(messages []types.FluxMessage, contextSize int, taskType string) int {
-	return token.DynamicMaxTokens(messages, contextSize, taskType)
-}
-func ClassifyTaskComplexity(task string) string { return token.ClassifyTaskComplexity(task) }
-func FormatPrediction(pred *Prediction, model string) string {
-	return token.FormatPrediction(pred, model)
-}
-
-func WarnIfExpensive(pred *Prediction, budgetUSD float64) string {
-	return token.WarnIfExpensive(pred, budgetUSD)
-}
-func CountTokens(text string) int     { return token.CountTokens(text) }
-func CountTokensFast(text string) int { return token.CountTokensFast(text) }
-func CompressForContext(text string, budget int) (string, int) {
-	return token.CompressForContext(text, budget)
-}
-func EstimateTokens(msgs []types.FluxMessage) int   { return token.EstimateTokens(msgs) }
-func EstimateMessageTokens(m types.FluxMessage) int { return token.EstimateMessageTokens(m) }
diff --git a/internal/engine/tool_service.go b/internal/engine/tool_service.go
index 158e048b..953282d3 100644
--- a/internal/engine/tool_service.go
+++ b/internal/engine/tool_service.go
@@ -10,6 +10,8 @@ import (
 	"strings"
 	"sync"
 
+	"github.com/GrayCodeAI/rho/internal/engine/token"
+
 	"github.com/GrayCodeAI/rho/internal/engine/diff"
 	"github.com/GrayCodeAI/rho/internal/engine/planning"
 	"github.com/GrayCodeAI/rho/internal/hooks"
@@ -595,8 +597,8 @@ func (s *ToolService) NormalizeOutput(output, canonicalTool, toolID string, cont
 	}
 	compressBudget := maxChars / 2
 	if len(output) > compressBudget {
-		compressed, tokens := CompressForContext(output, compressBudget/4)
-		if tokens > 0 && tokens < CountTokensFast(output) {
+		compressed, tokens := token.CompressForContext(output, compressBudget/4)
+		if tokens > 0 && tokens < token.CountTokensFast(output) {
 			output = compressed
 		}
 	}
diff --git a/internal/engine/validation_reexports.go b/internal/engine/validation_reexports.go
deleted file mode 100644
index 2abc0069..00000000
--- a/internal/engine/validation_reexports.go
+++ /dev/null
@@ -1,24 +0,0 @@
-package engine
-
-import "github.com/GrayCodeAI/rho/internal/engine/validation"
-
-type (
-	GenValidator           = validation.GenValidator
-	GenCheck               = validation.GenCheck
-	GenIssue               = validation.GenIssue
-	GenValidation          = validation.GenValidation
-	SchemaValidator        = validation.SchemaValidator
-	Schema                 = validation.Schema
-	FieldSpec              = validation.FieldSpec
-	SchemaValidationResult = validation.SchemaValidationResult
-	SchemaValidationError  = validation.SchemaValidationError
-	TestLoop               = validation.TestLoop
-	TestResult             = validation.TestResult
-	LintLoop               = validation.LintLoop
-	LintResult             = validation.LintResult
-)
-
-func NewGenValidator() *GenValidator       { return validation.NewGenValidator() }
-func NewSchemaValidator() *SchemaValidator { return validation.NewSchemaValidator() }
-func NewTestLoop() *TestLoop               { return validation.NewTestLoop() }
-func NewLintLoop() *LintLoop               { return validation.NewLintLoop() }
diff --git a/internal/engine/workflow_reexports.go b/internal/engine/workflow_reexports.go
deleted file mode 100644
index 8efd031b..00000000
--- a/internal/engine/workflow_reexports.go
+++ /dev/null
@@ -1,57 +0,0 @@
-package engine
-
-import (
-	"context"
-
-	"github.com/GrayCodeAI/rho/internal/engine/workflow"
-)
-
-// Types from workflow sub-package.
-
-type (
-	Workflow            = workflow.Workflow
-	WorkflowStep        = workflow.WorkflowStep
-	WorkflowResult      = workflow.WorkflowResult
-	StepResult          = workflow.StepResult
-	WorkflowEngine      = workflow.WorkflowEngine
-	WorkspaceState      = workflow.WorkspaceState
-	FileState           = workflow.FileState
-	WorkspaceDiffReport = workflow.WorkspaceDiffReport
-	FileDiffReport      = workflow.FileDiffReport
-	DiffReporter        = workflow.DiffReporter
-	TrajectoryEvent     = workflow.TrajectoryEvent
-	TrajectoryInspector = workflow.TrajectoryInspector
-)
-
-// Short-name aliases provided by the workflow sub-package.
-
-type (
-	Step   = workflow.WorkflowStep
-	Result = workflow.WorkflowResult
-	Engine = workflow.WorkflowEngine
-)
-
-// Functions.
-
-var (
-	NewWorkflowEngine      = workflow.NewWorkflowEngine
-	NewWorkspaceState      = workflow.NewWorkspaceState
-	NewDiffReporter        = workflow.NewDiffReporter
-	NewTrajectoryInspector = workflow.NewTrajectoryInspector
-	SubstituteVars         = workflow.SubstituteVars
-	EvalCondition          = workflow.EvalCondition
-	ValidateWorkflow       = workflow.ValidateWorkflow
-	BuiltinWorkflows       = workflow.BuiltinWorkflows
-	FormatAsMarkdown       = workflow.FormatAsMarkdown
-	FormatAsTerminal       = workflow.FormatAsTerminal
-	FormatForCommit        = workflow.FormatForCommit
-	CompareReports         = workflow.CompareReports
-)
-
-// NewEngine delegates to workflow.NewEngine.
-func NewEngine(executeFn func(ctx context.Context, agent, prompt string) (string, error)) *Engine {
-	return workflow.NewEngine(executeFn)
-}
-
-// TrajectoryRun, TrajectoryDistiller, NewTrajectoryDistiller, SummarizeTrajectory
-// are defined directly in package engine (trajectory.go).