| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "encoding/json" |
| 6 | "fmt" |
| 7 | "strings" |
| 8 | "testing" |
| 9 | "time" |
| 10 | |
| 11 | "reasonix/internal/event" |
| 12 | "reasonix/internal/evidence" |
| 13 | "reasonix/internal/provider" |
| 14 | "reasonix/internal/tool" |
| 15 | ) |
| 16 | |
| 17 | type recordingRecoveryGate struct { |
| 18 | observation RecoveryObservation |
| 19 | proposals []RecoveryProposal |
| 20 | decision RecoveryDecision |
| 21 | } |
| 22 | |
| 23 | func TestRecoveryPlanTransitionDetectsOnlyStructuralRewriteOfActivePlan(t *testing.T) { |
| 24 | a := &Agent{} |
| 25 | initial := json.RawMessage(`{"todos":[{"content":"Implement parser","status":"in_progress"}]}`) |
| 26 | if changed, _, _ := a.recoveryPlanTransition("todo_write", initial); changed { |
| 27 | t.Fatal("initial plan must stay on the fast path") |
| 28 | } |
| 29 | |
| 30 | a.setTodoState([]evidence.TodoItem{ |
| 31 | {Content: "Implement parser", Status: "in_progress"}, |
| 32 | {Content: "Run tests", Status: "pending"}, |
| 33 | }) |
| 34 | progressOnly := json.RawMessage(`{"todos":[{"content":"Implement parser","status":"completed"},{"content":"Run tests","status":"in_progress"}]}`) |
| 35 | if changed, _, _ := a.recoveryPlanTransition("todo_write", progressOnly); changed { |
| 36 | t.Fatal("progress-only update must not invoke the plan reviewer") |
| 37 | } |
| 38 | |
| 39 | replacement := json.RawMessage(`{"todos":[{"content":"Replace parser architecture","status":"in_progress"},{"content":"Run tests","status":"pending"}]}`) |
| 40 | changed, before, after := a.recoveryPlanTransition("todo_write", replacement) |
| 41 | if !changed { |
| 42 | t.Fatal("structural rewrite of active plan was not detected") |
| 43 | } |
| 44 | if !strings.Contains(before, "Implement parser") || !strings.Contains(after, "Replace parser architecture") { |
| 45 | t.Fatalf("plan evidence before=%q after=%q", before, after) |
| 46 | } |
| 47 | } |
| 48 | |
| 49 | func TestRecoveryPlanTransitionIgnoresCompletedPriorPlan(t *testing.T) { |
| 50 | a := &Agent{} |
| 51 | a.setTodoState([]evidence.TodoItem{{Content: "Old task", Status: "completed"}}) |
| 52 | next := json.RawMessage(`{"todos":[{"content":"New user task","status":"in_progress"}]}`) |
| 53 | if changed, _, _ := a.recoveryPlanTransition("todo_write", next); changed { |
| 54 | t.Fatal("a new task after a completed plan is not a mid-plan transition") |
| 55 | } |
| 56 | } |
| 57 | |
| 58 | func (g *recordingRecoveryGate) ObserveResult(_ context.Context, observation RecoveryObservation) string { |
| 59 | g.observation = observation |
| 60 | return "" |
| 61 | } |
| 62 | |
| 63 | func (g *recordingRecoveryGate) BeforeMutation(_ context.Context, proposal RecoveryProposal) (RecoveryDecision, error) { |
| 64 | g.proposals = append(g.proposals, proposal) |
| 65 | decision := g.decision |
| 66 | if decision == (RecoveryDecision{}) { |
| 67 | decision.Allow = true |
| 68 | } |
| 69 | return decision, nil |
| 70 | } |
| 71 | |
| 72 | func TestAuthorizedRecoveryPlanTransitionCanReplaceCurrentTodo(t *testing.T) { |
| 73 | reg := tool.NewRegistry() |
| 74 | reg.Add(mustBuiltinTool(t, "todo_write")) |
| 75 | gate := &recordingRecoveryGate{decision: RecoveryDecision{ |
| 76 | Allow: true, AuthorizePlanReplacement: true, |
| 77 | }} |
| 78 | a := New(nil, reg, NewSession(""), Options{RecoveryGate: gate}, event.Discard) |
| 79 | a.SeedTodoState([]evidence.TodoItem{ |
| 80 | {Content: "Inspect environment", Status: "completed"}, |
| 81 | {Content: "Implement parser", Status: "in_progress"}, |
| 82 | {Content: "Run tests", Status: "pending"}, |
| 83 | }) |
| 84 | |
| 85 | out := a.executeOne(context.Background(), provider.ToolCall{ |
| 86 | ID: "replace-plan", |
| 87 | Name: "todo_write", |
| 88 | Arguments: `{"todos":[ |
| 89 | {"content":"Inspect environment","status":"completed"}, |
| 90 | {"content":"Replace parser architecture","status":"in_progress"}, |
| 91 | {"content":"Run tests","status":"pending"} |
| 92 | ]}`, |
| 93 | }) |
| 94 | if out.errMsg != "" { |
| 95 | t.Fatalf("authorized plan replacement was blocked: %+v", out) |
| 96 | } |
| 97 | if len(gate.proposals) != 1 || !gate.proposals[0].PlanTransition { |
| 98 | t.Fatalf("recovery proposals = %+v, want one plan transition", gate.proposals) |
| 99 | } |
| 100 | got := a.CanonicalTodoState() |
| 101 | if len(got) != 3 || got[0].Status != "completed" || got[1].Content != "Replace parser architecture" { |
| 102 | t.Fatalf("canonical todo state = %+v, want preserved history plus replacement", got) |
| 103 | } |
| 104 | } |
| 105 | |
| 106 | func TestPlanTransitionNeedsDedicatedReplacementAuthorization(t *testing.T) { |
| 107 | reg := tool.NewRegistry() |
| 108 | reg.Add(mustBuiltinTool(t, "todo_write")) |
| 109 | gate := &recordingRecoveryGate{decision: RecoveryDecision{Allow: true}} |
| 110 | a := New(nil, reg, NewSession(""), Options{RecoveryGate: gate}, event.Discard) |
| 111 | a.SeedTodoState([]evidence.TodoItem{{Content: "Implement parser", Status: "in_progress"}}) |
| 112 | |
| 113 | out := a.executeOne(context.Background(), provider.ToolCall{ |
| 114 | ID: "replace-plan-without-authorization", |
| 115 | Name: "todo_write", |
| 116 | Arguments: `{"todos":[{"content":"Replace parser architecture","status":"in_progress"}]}`, |
| 117 | }) |
| 118 | if out.errMsg == "" || !strings.Contains(out.output, "cannot be removed or replaced") { |
| 119 | t.Fatalf("plain allow unexpectedly replaced current todo: %+v", out) |
| 120 | } |
| 121 | } |
| 122 | |
| 123 | func TestObserveRecoveryResultMarksCancellation(t *testing.T) { |
| 124 | gate := &recordingRecoveryGate{} |
| 125 | a := &Agent{recoveryGate: gate} |
| 126 | a.observeRecoveryResult( |
| 127 | context.Background(), |
| 128 | "write_file", |
| 129 | json.RawMessage(`{"path":"a.go"}`), |
| 130 | false, |
| 131 | true, |
| 132 | "cancelled", |
| 133 | context.Canceled, |
| 134 | false, |
| 135 | false, |
| 136 | 0, |
| 137 | ) |
| 138 | if !gate.observation.Cancelled { |
| 139 | t.Fatalf("observation = %+v, want cancellation marked", gate.observation) |
| 140 | } |
| 141 | if gate.observation.TaskScopeID == "" { |
| 142 | t.Fatalf("observation = %+v, want a host-owned recovery scope", gate.observation) |
| 143 | } |
| 144 | } |
| 145 | |
| 146 | func TestObserveRecoveryResultKeepsToolOwnedDeadlineAsFailure(t *testing.T) { |
| 147 | gate := &recordingRecoveryGate{} |
| 148 | a := &Agent{recoveryGate: gate} |
| 149 | a.observeRecoveryResult( |
| 150 | context.Background(), |
| 151 | "mcp__server__write", |
| 152 | json.RawMessage(`{"value":"x"}`), |
| 153 | false, |
| 154 | true, |
| 155 | "", |
| 156 | fmt.Errorf("MCP tool timed out after 30s: %w", context.DeadlineExceeded), |
| 157 | false, |
| 158 | false, |
| 159 | 0, |
| 160 | ) |
| 161 | if gate.observation.Cancelled { |
| 162 | t.Fatalf("observation = %+v, tool-owned deadline must remain a qualifying transient failure", gate.observation) |
| 163 | } |
| 164 | if !strings.Contains(gate.observation.ErrSummary, "timed out") { |
| 165 | t.Fatalf("observation = %+v, want timeout evidence preserved", gate.observation) |
| 166 | } |
| 167 | } |
| 168 | |
| 169 | func TestObserveRecoveryResultMarksParentDeadlineCancellation(t *testing.T) { |
| 170 | gate := &recordingRecoveryGate{} |
| 171 | a := &Agent{recoveryGate: gate} |
| 172 | ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(-time.Second)) |
| 173 | defer cancel() |
| 174 | a.observeRecoveryResult( |
| 175 | ctx, |
| 176 | "mcp__server__write", |
| 177 | json.RawMessage(`{"value":"x"}`), |
| 178 | false, |
| 179 | true, |
| 180 | "", |
| 181 | context.DeadlineExceeded, |
| 182 | false, |
| 183 | false, |
| 184 | 0, |
| 185 | ) |
| 186 | if !gate.observation.Cancelled { |
| 187 | t.Fatalf("observation = %+v, parent deadline must remain a cancellation", gate.observation) |
| 188 | } |
| 189 | } |
| 190 | |
| 191 | func TestRecoveryBlockSurfacesConcreteReason(t *testing.T) { |
| 192 | reg := tool.NewRegistry() |
| 193 | reg.Add(mustBuiltinTool(t, "write_file")) |
| 194 | gate := &recordingRecoveryGate{decision: RecoveryDecision{ |
| 195 | Blocked: true, |
| 196 | Message: "blocked: Auto stopped repeating this operation after 3 consecutive failures: write a.go. Other operations remain available.", |
| 197 | }} |
| 198 | a := New(nil, reg, NewSession(""), Options{RecoveryGate: gate}, event.Discard) |
| 199 | out := a.executeOne(context.Background(), provider.ToolCall{ |
| 200 | ID: "blocked-write", Name: "write_file", Arguments: `{"path":"a.go","content":"x"}`, |
| 201 | }) |
| 202 | if !out.blocked || !strings.Contains(out.errMsg, "stopped repeating this operation") || strings.Contains(out.errMsg, "Auto Guard") { |
| 203 | t.Fatalf("recovery failure card = %+v", out) |
| 204 | } |
| 205 | } |
| 206 |