| 1 | package main |
| 2 | |
| 3 | import ( |
| 4 | "strings" |
| 5 | "testing" |
| 6 | ) |
| 7 | |
| 8 | func TestSummarizeTrajectoryAnchorSafetyAudits(t *testing.T) { |
| 9 | path := writeTrajectory(t, "anchor-safety.trajectory.jsonl", []string{ |
| 10 | `{"seq":1,"ts":1000,"anchor_safety_audit":{"mode":"shadow","task_mode":"interactive","range_lines":3,"observation_age":2,"legacy_allowed":false,"shadow_allowed":true,"reason":"would_allow_exact_match"}}`, |
| 11 | `{"seq":2,"ts":2000,"anchor_safety_audit":{"mode":"shadow","task_mode":"interactive","range_lines":4,"observation_age":0,"legacy_allowed":true,"shadow_allowed":false,"reason":"same_batch_read_rejected","same_batch_read_rejected":true}}`, |
| 12 | `{"seq":3,"ts":3000,"anchor_safety_audit":{"mode":"shadow","task_mode":"loop","range_lines":2,"observation_age":5,"legacy_allowed":true,"shadow_allowed":false,"reason":"would_block_target_changed"}}`, |
| 13 | `{"seq":4,"ts":4000,"anchor_safety_audit":{"mode":"shadow","task_mode":"loop","reason":"native_target_invalid"}}`, |
| 14 | }) |
| 15 | s, err := summarizeTrajectory(path) |
| 16 | if err != nil { |
| 17 | t.Fatalf("summarizeTrajectory: %v", err) |
| 18 | } |
| 19 | a := s.AnchorSafety |
| 20 | if a == nil || a.Samples != 4 || a.ShadowAllows != 1 || a.LegacyAllows != 2 || a.SameBatchReads != 1 || a.TargetChanged != 1 || a.NativeInvalid != 1 { |
| 21 | t.Fatalf("anchor safety summary = %+v", a) |
| 22 | } |
| 23 | if a.ShadowOnlyAllows != 1 || a.ShadowOnlyBlocks != 2 || a.MaxObservationAge != 5 { |
| 24 | t.Fatalf("anchor safety deltas = %+v", a) |
| 25 | } |
| 26 | if a.ByTaskMode["interactive"] != 2 || a.ByTaskMode["loop"] != 2 { |
| 27 | t.Fatalf("anchor safety task modes = %+v", a.ByTaskMode) |
| 28 | } |
| 29 | got := renderAnchorSafety([]result{{Trajectory: s}}) |
| 30 | for _, want := range []string{"**Anchor safety shadow**", "**samples** 4", "**shadow-only allows** 1", "**same-batch reads** 1", "**interactive** 2", "**loop** 2"} { |
| 31 | if !strings.Contains(got, want) { |
| 32 | t.Errorf("anchor safety report missing %q:\n%s", want, got) |
| 33 | } |
| 34 | } |
| 35 | } |
| 36 | |
| 37 | func TestSummarizeTrajectoryIgnoresOldAnchorSafetyAbsence(t *testing.T) { |
| 38 | path := writeTrajectory(t, "pre-anchor-safety.trajectory.jsonl", []string{ |
| 39 | `{"seq":1,"ts":1000,"event":{"kind":"turn_started"}}`, |
| 40 | }) |
| 41 | s, err := summarizeTrajectory(path) |
| 42 | if err != nil { |
| 43 | t.Fatalf("summarizeTrajectory: %v", err) |
| 44 | } |
| 45 | if s.AnchorSafety != nil || renderAnchorSafety([]result{{Trajectory: s}}) != "" { |
| 46 | t.Fatalf("old trajectory unexpectedly reported anchor safety: %+v", s.AnchorSafety) |
| 47 | } |
| 48 | } |
| 49 |