fix(benchmark): resolve select agent_dirs against the suite directory

BENCHMARKING.md

@@ -327,7 +327,7 @@ recipes directly into `pipeline.SelectionOnly`:
 | `stages.selection.model` | `SelectionModelOverride` |
 | `stages.selection.effort` | `SelectionEffortOverride` |
 | `stages.selection.prompt` | file contents loaded into `SelectionPromptInstructions` |
-| `stages.reviewers.agent_dirs[]` | `AgentDirs` |
+| `stages.reviewers.agent_dirs[]` | `AgentDirs`, each resolved against the suite file directory |
 | `review_base_sha` / `review_head_sha` | pinned review SHAs |
 
 Selector-only benchmarks do not run reviewer agents, do not run synthesis, do

internal/cmd/benchmarkcmd/select.go

@@ -288,7 +288,7 @@ func executeBenchmarkSelectRun(ctx context.Context, logger *progress.Logger, sui
 		ProfileName:                 state.profileName,
 		Profile:                     state.profile,
 		PostingIdentity:             postingIdentity,
-		AgentDirs:                   append([]string(nil), candidate.Stages.Reviewers.AgentDirs...),
+		AgentDirs:                   resolveSuiteAgentDirs(suiteDir, candidate.Stages.Reviewers.AgentDirs),
 		ArtifactDir:                 runDir,
 		ReviewBaseSHA:               benchCase.ReviewBaseSHA,
 		ReviewHeadSHA:               benchCase.ReviewHeadSHA,
@@ -389,6 +389,20 @@ func selectionRunArtifacts(runDir string) (runArtifacts, error) {
 	}, nil
 }
 
+// resolveSuiteAgentDirs resolves candidate agent directories against the suite
+// file's directory, as `benchmark run` does for its child reviews, so a suite
+// selects from the same catalog whichever directory it is invoked from.
+func resolveSuiteAgentDirs(suiteDir string, dirs []string) []string {
+	if dirs == nil {
+		return nil
+	}
+	out := make([]string, 0, len(dirs))
+	for _, dir := range dirs {
+		out = append(out, benchmark.ResolveSuitePath(suiteDir, dir))
+	}
+	return out
+}
+
 func loadSelectionPromptInstructions(suiteDir, configured string) (string, error) {
 	configured = strings.TrimSpace(configured)
 	if configured == "" {

internal/cmd/benchmarkcmd/select_test.go

@@ -7,6 +7,7 @@ import (
 	"fmt"
 	"os"
 	"path/filepath"
+	"reflect"
 	"strings"
 	"testing"
 	"time"
@@ -583,3 +584,60 @@ func TestSelectionReportMarkdownListsSelectedReviewers(t *testing.T) {
 		t.Fatalf("report = %s, want selected reviewer table", report)
 	}
 }
+
+func TestSelectResolvesRelativeAgentDirsAgainstSuiteDirectory(t *testing.T) {
+	cmd, _ := newTestCommand(t)
+	suitePath := writeBenchmarkSuite(t, `
+suite:
+  id: suite1
+  name: Suite One
+  version: 1
+candidates:
+  - id: first
+    profile: home
+    stages:
+      selection:
+        model: claude-sonnet-4-6
+        effort: high
+      reviewers:
+        model: claude-sonnet-4-6
+        agent_dirs:
+          - ../agents
+cases:
+  - id: case_one
+    pr: https://github.com/open-cli-collective/codereview-cli/pull/1
+`)
+	// Invoke from an unrelated directory: a cwd-relative "../agents" would
+	// point somewhere else entirely.
+	t.Chdir(t.TempDir())
+	var requests []pipeline.SelectionRequest
+	withBenchmarkSelectSeams(t,
+		func(context.Context, string, bool, config.File, config.Profile, gitprovider.PRRef) (app.SelectionRuntime, error) {
+			return app.SelectionRuntime{Cleanup: func() {}}, nil
+		},
+		func(_ context.Context, _ pipeline.Options, req pipeline.SelectionRequest) (pipeline.SelectionResult, error) {
+			requests = append(requests, req)
+			return pipeline.SelectionResult{
+				Artifacts: pipeline.ArtifactPathsFromDir(req.ArtifactDir),
+				SelectionSession: pipeline.SelectionSession{Response: llm.Response{
+					StructuredOutput: []byte(`{"schema_version":1,"selected_agents":[],"thread_actions":[],"reasoning":"ok"}`),
+				}},
+			}, nil
+		},
+	)
+
+	if err := root.Execute(cmd, []string{
+		"benchmark", "select", suitePath,
+		"--results-dir", filepath.Join(t.TempDir(), "results"),
+		"--json",
+	}); err != nil {
+		t.Fatalf("Execute: %v", err)
+	}
+	want := []string{filepath.Join(filepath.Dir(filepath.Dir(suitePath)), "agents")}
+	if len(requests) != 1 {
+		t.Fatalf("selection requests = %d, want 1", len(requests))
+	}
+	if got := requests[0].AgentDirs; !reflect.DeepEqual(got, want) {
+		t.Fatalf("selection agent dirs = %#v, want suite-relative %#v", got, want)
+	}
+}