fix: phase 3 medium findings (token estimate, seatbelt, pool errors)

internal/engine/conversation_adapter_test.go

@@ -272,7 +272,7 @@ func TestExportMessagesFormat(t *testing.T) {
 func TestTokenEstimateRoughAccuracy(t *testing.T) {
 	cm := NewConversationManager(ConversationConfig{})
 
-	// estimateTokens uses len(text)/4
+	// estimateTokens uses the BPE-based tok tokenizer.
 	content := "this is a test string that should produce some tokens"
 	cm.AddMessage("user", content)
 
@@ -283,9 +283,13 @@ func TestTokenEstimateRoughAccuracy(t *testing.T) {
 		t.Errorf("expected token estimate %d, got %d", expected, got)
 	}
 
-	// Verify the formula: roughly 1 token per 4 characters.
-	if got != len(content)/4 {
-		t.Errorf("token estimate should be len/4 = %d, got %d", len(content)/4, got)
+	// The estimate must be non-zero and stay within a sane range for English
+	// text (BPE typically lands between ~3 and ~7 chars per token).
+	if got <= 0 {
+		t.Errorf("token estimate should be positive, got %d", got)
+	}
+	if cpt := float64(len(content)) / float64(got); cpt < 3 || cpt > 7 {
+		t.Errorf("chars-per-token %0.2f outside expected range (3-7) for %d chars and %d tokens", cpt, len(content), got)
 	}
 }
 

internal/engine/selective_rag.go

@@ -4,6 +4,8 @@ import (
 	"strings"
 	"sync"
 	"time"
+
+	"github.com/GrayCodeAI/hawk/internal/engine/token"
 )
 
 // SelectiveRAG implements the Repoformer-style selective retrieval mechanism.
@@ -171,8 +173,9 @@ func (r *SelectiveRAG) Stats() map[string]interface{} {
 // Helper functions for query classification
 
 func estimateTokens(text string) int {
-	// Rough estimate: 1 token per 4 chars
-	return len(text) / 4
+	// BPE-based estimate via the tok tokenizer instead of the len/4 char
+	// heuristic, which systematically undercounts code-heavy text.
+	return token.CountTokensFast(text)
 }
 
 func isNavigationQuery(query string) bool {

internal/multiagent/parallel/parallel.go

@@ -2,6 +2,7 @@ package parallel
 
 import (
 	"context"
+	"errors"
 	"fmt"
 	"sync"
 )
@@ -87,7 +88,9 @@ func (p *Pool) AddTask(description string) *Task {
 
 // Run executes all queued tasks in parallel, each in its own git worktree.
 // workFn receives the worktree path and task, and returns a result summary.
-// Tasks that fail do not prevent other tasks from completing.
+// Tasks that fail do not prevent other tasks from completing; the returned
+// error aggregates every failed task so callers are not left with silent
+// failures.
 func (p *Pool) Run(ctx context.Context, workFn func(ctx context.Context, worktreePath string, task *Task) (string, error)) error {
 	p.mu.Lock()
 	tasks := make([]*Task, len(p.tasks))
@@ -161,6 +164,22 @@ func (p *Pool) Run(ctx context.Context, workFn func(ctx context.Context, worktre
 	}
 
 	wg.Wait()
+
+	// Aggregate failures instead of returning nil unconditionally, so a run
+	// where every task failed is distinguishable from a fully successful run.
+	var failed []*Task
+	for _, t := range tasks {
+		if t.Status == StatusFailed {
+			failed = append(failed, t)
+		}
+	}
+	if len(failed) > 0 {
+		errs := make([]error, 0, len(failed))
+		for _, t := range failed {
+			errs = append(errs, fmt.Errorf("task %q failed: %w", t.Description, t.Error))
+		}
+		return errors.Join(errs...)
+	}
 	return nil
 }
 

internal/multiagent/parallel/parallel_test.go

@@ -6,6 +6,7 @@ import (
 	"os"
 	"os/exec"
 	"path/filepath"
+	"strings"
 	"sync"
 	"sync/atomic"
 	"testing"
@@ -202,8 +203,11 @@ func TestErrorHandling(t *testing.T) {
 		}
 		return "success", nil
 	})
-	if err != nil {
-		t.Fatalf("Run: %v", err)
+	if err == nil {
+		t.Fatal("Run should return an aggregate error when a task fails")
+	}
+	if !strings.Contains(err.Error(), "fail-1") || !strings.Contains(err.Error(), "intentional failure") {
+		t.Errorf("aggregate error should name the failed task and its error, got: %v", err)
 	}
 	defer pool.Cleanup()
 
@@ -310,8 +314,8 @@ func TestContextCancellation(t *testing.T) {
 			return "", ctx.Err()
 		}
 	})
-	if err != nil {
-		t.Fatalf("Run: %v", err)
+	if err == nil {
+		t.Fatal("Run should report an aggregate error when tasks fail")
 	}
 	defer pool.Cleanup()
 

internal/sandbox/seatbelt.go

@@ -31,8 +31,25 @@ func GenerateSeatbeltProfile(policy *SeatbeltPolicy) string {
 	b.WriteString("(version 1)\n")
 	b.WriteString("(deny default)\n")
 
-	// Always allow basic mach-lookup for system functionality.
-	b.WriteString("(allow mach-lookup)\n")
+	// mach-lookup grants access to macOS XPC services. The legacy TierOff
+	// behavior allows every service; all other tiers are restricted to the
+	// minimal set of services required for normal tooling to work. Network
+	// resolution services are only reachable when network is allowed.
+	switch policy.Tier {
+	case TierOff:
+		b.WriteString("(allow mach-lookup)\n")
+	default:
+		b.WriteString("(allow mach-lookup\n")
+		b.WriteString("  (global-name \"com.apple.system.opendirectoryd.api\")\n")
+		b.WriteString("  (global-name \"com.apple.system.notification_center\")\n")
+		b.WriteString("  (global-name \"com.apple.cfprefsd\")\n")
+		b.WriteString("  (global-name \"com.apple.system.logger\")\n")
+		if policy.AllowNetwork {
+			b.WriteString("  (global-name \"com.apple.mDNSResponder\")\n")
+			b.WriteString("  (global-name \"com.apple.system.systemconfigurationd\")\n")
+		}
+		b.WriteString(")\n")
+	}
 
 	// Sysctl read for basic system queries.
 	if policy.AllowSysctl {

internal/sandbox/seatbelt_test.go

@@ -169,6 +169,35 @@ func TestDefaultHawkPolicy_ProfileProducesValidSBPL(t *testing.T) {
 	}
 }
 
+func TestGenerateSeatbeltProfile_MachLookupTiered(t *testing.T) {
+	// TierOff keeps the legacy broad mach-lookup rule.
+	off := GenerateSeatbeltProfile(&SeatbeltPolicy{Tier: TierOff})
+	if !strings.Contains(off, "(allow mach-lookup)\n") {
+		t.Error("TierOff profile should allow all mach-lookup services")
+	}
+
+	// All other tiers restrict mach-lookup to the service allowlist.
+	strict := GenerateSeatbeltProfile(&SeatbeltPolicy{Tier: TierStrict})
+	if strings.Contains(strict, "(allow mach-lookup)\n") {
+		t.Error("restricted profile should not contain the broad mach-lookup rule")
+	}
+	if !strings.Contains(strict, `(global-name "com.apple.system.opendirectoryd.api")`) {
+		t.Error("restricted profile should allow opendirectoryd.api")
+	}
+	if strings.Contains(strict, `(global-name "com.apple.mDNSResponder")`) {
+		t.Error("network service should not be reachable when AllowNetwork is false")
+	}
+
+	// Network-enabled tiers additionally allow resolution services.
+	net := GenerateSeatbeltProfile(&SeatbeltPolicy{Tier: TierWorkspace, AllowNetwork: true})
+	if !strings.Contains(net, `(global-name "com.apple.mDNSResponder")`) {
+		t.Error("network-enabled profile should allow mDNSResponder")
+	}
+	if !strings.Contains(net, `(global-name "com.apple.system.systemconfigurationd")`) {
+		t.Error("network-enabled profile should allow systemconfigurationd")
+	}
+}
+
 func TestRunSeatbelted_ReturnsCmd(t *testing.T) {
 	policy := &SeatbeltPolicy{
 		AllowNetwork:  false,