| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "strings" |
| 6 | "testing" |
| 7 | |
| 8 | "reasonix/internal/event" |
| 9 | "reasonix/internal/provider" |
| 10 | "reasonix/internal/tool" |
| 11 | ) |
| 12 | |
| 13 | func TestCoordinatorPlannerDepthDoesNotCapResearchRounds(t *testing.T) { |
| 14 | planner := &mockProvider{name: "planner", streams: [][]provider.Chunk{ |
| 15 | {{Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "call-1", Name: "read_file", Arguments: `{"path":"a"}`}}, {Type: provider.ChunkDone}}, |
| 16 | {{Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "call-2", Name: "read_file", Arguments: `{"path":"b"}`}}, {Type: provider.ChunkDone}}, |
| 17 | {{Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "call-3", Name: "read_file", Arguments: `{"path":"c"}`}}, {Type: provider.ChunkDone}}, |
| 18 | {{Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "plan-1", Name: "submit_plan", Arguments: `{"objective":"apply the narrow change","steps":[{"title":"run the focused test","verified_files":["a","b","c"]}]}`}}, {Type: provider.ChunkDone}}, |
| 19 | }} |
| 20 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{{Type: provider.ChunkText, Text: "Done."}, {Type: provider.ChunkDone}}} |
| 21 | reg := tool.NewRegistry() |
| 22 | reg.Add(coordinatorTestTool{name: "read_file", readOnly: true, output: "ok"}) |
| 23 | policy := func(context.Context, string) PlannerDecision { |
| 24 | return PlannerDecision{Route: PlannerRoutePlanAndExecute, Reason: "adaptive_work"} |
| 25 | } |
| 26 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 27 | coord := NewCoordinatorWithPlannerPolicy(planner, NewSession("planner-sys"), nil, |
| 28 | PlannerToolRegistry(reg), Options{}, executor, 0, event.Discard, policy) |
| 29 | |
| 30 | if err := coord.Run(withNoClosedLoop(context.Background()), "make the adaptive change"); err != nil { |
| 31 | t.Fatalf("Run: %v", err) |
| 32 | } |
| 33 | if got := len(planner.requests); got != 4 { |
| 34 | t.Fatalf("planner requests = %d, want three evidence rounds plus submit_plan", got) |
| 35 | } |
| 36 | for i, req := range planner.requests { |
| 37 | if got := lastUser(req); strings.Contains(got, "research rounds") || strings.Contains(got, "safety boundary") { |
| 38 | t.Fatalf("planner request %d received a fixed-round nudge: %q", i, got) |
| 39 | } |
| 40 | } |
| 41 | if len(exec.requests) == 0 || !strings.Contains(lastUser(exec.requests[0]), executorHandoffMarker) { |
| 42 | t.Fatal("executor did not receive the submitted plan") |
| 43 | } |
| 44 | } |
| 45 | |
| 46 | func TestCoordinatorEmergencyBoundedPlannerCanSubmitPlanInFinalizationRound(t *testing.T) { |
| 47 | planner := &mockProvider{name: "planner", streams: [][]provider.Chunk{ |
| 48 | {{Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "call-1", Name: "read_file", Arguments: `{"path":"a"}`}}, {Type: provider.ChunkDone}}, |
| 49 | {{Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "call-2", Name: "read_file", Arguments: `{"path":"b"}`}}, {Type: provider.ChunkDone}}, |
| 50 | {{Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "plan-1", Name: "submit_plan", Arguments: `{"objective":"fix the bounded planner","steps":[{"title":"apply the owner-level fix","verified_files":["a","b"]}]}`}}, {Type: provider.ChunkDone}}, |
| 51 | }} |
| 52 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{{Type: provider.ChunkText, Text: "Done."}, {Type: provider.ChunkDone}}} |
| 53 | reg := tool.NewRegistry() |
| 54 | reg.Add(coordinatorTestTool{name: "read_file", readOnly: true, output: strings.Repeat("research evidence\n", 10_000)}) |
| 55 | var notices []event.Event |
| 56 | sink := event.FuncSink(func(e event.Event) { |
| 57 | if e.Kind == event.Notice { |
| 58 | notices = append(notices, e) |
| 59 | } |
| 60 | }) |
| 61 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 62 | plannerSess := NewSession("planner-sys") |
| 63 | coord := NewCoordinatorWithPlannerPolicy(planner, plannerSess, nil, PlannerToolRegistry(reg), |
| 64 | Options{MaxSteps: 2, MaxStepsKey: "planner emergency rounds"}, executor, 0, sink, |
| 65 | func(context.Context, string) PlannerDecision { |
| 66 | return PlannerDecision{Route: PlannerRoutePlanAndExecute, Reason: "emergency_bounded_work"} |
| 67 | }) |
| 68 | |
| 69 | if err := coord.Run(withNoClosedLoop(context.Background()), "fix the planner bug"); err != nil { |
| 70 | t.Fatalf("Run: %v", err) |
| 71 | } |
| 72 | if got := len(planner.requests); got != 3 { |
| 73 | t.Fatalf("planner requests = %d, want two research rounds plus terminal submission", got) |
| 74 | } |
| 75 | if got := lastUser(planner.requests[2]); !strings.Contains(got, "call submit_plan now") || strings.Contains(got, "Do not call any more tools") { |
| 76 | t.Fatalf("planner finalization nudge conflicts with submit_plan: %q", got) |
| 77 | } |
| 78 | sawTruncation := false |
| 79 | for _, notice := range notices { |
| 80 | sawTruncation = sawTruncation || strings.HasPrefix(notice.Text, "tool output truncated:") |
| 81 | } |
| 82 | if !sawTruncation { |
| 83 | t.Fatal("test setup did not reproduce truncated planner research") |
| 84 | } |
| 85 | if len(exec.requests) == 0 || !strings.Contains(lastUser(exec.requests[0]), "**Objective** — fix the bounded planner") { |
| 86 | t.Fatal("executor did not receive the structured bounded plan") |
| 87 | } |
| 88 | messages := plannerSess.Snapshot() |
| 89 | if last := messages[len(messages)-1]; last.Role != provider.RoleAssistant || last.Content != plannerPlanSubmittedClosure { |
| 90 | t.Fatalf("planner transcript did not close deterministically: %+v", last) |
| 91 | } |
| 92 | } |
| 93 | |
| 94 | func TestCoordinatorTaskBudgetLetsPlannerSubmitTerminalPlan(t *testing.T) { |
| 95 | planner := &mockProvider{name: "planner", streams: [][]provider.Chunk{ |
| 96 | { |
| 97 | {Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "call-1", Name: "read_file", Arguments: `{"path":"main.go"}`}}, |
| 98 | {Type: provider.ChunkUsage, Usage: &provider.Usage{PromptTokens: 10, TotalTokens: 10}}, |
| 99 | {Type: provider.ChunkDone}, |
| 100 | }, |
| 101 | {{Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "plan-1", Name: "submit_plan", Arguments: `{"objective":"finish within the safety envelope","steps":[{"title":"apply the verified change","verified_files":["main.go"]}]}`}}, {Type: provider.ChunkDone}}, |
| 102 | }} |
| 103 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{{Type: provider.ChunkText, Text: "Done."}, {Type: provider.ChunkDone}}} |
| 104 | reg := tool.NewRegistry() |
| 105 | reg.Add(coordinatorTestTool{name: "read_file", readOnly: true, output: "package main"}) |
| 106 | var notices []event.Event |
| 107 | sink := event.FuncSink(func(e event.Event) { |
| 108 | if e.Kind == event.Notice { |
| 109 | notices = append(notices, e) |
| 110 | } |
| 111 | }) |
| 112 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 113 | coord := NewCoordinatorWithPlannerPolicy(planner, NewSession("planner-sys"), nil, PlannerToolRegistry(reg), |
| 114 | Options{}, executor, 0, sink, func(context.Context, string) PlannerDecision { |
| 115 | return PlannerDecision{Route: PlannerRoutePlanAndExecute, Reason: "budgeted_work"} |
| 116 | }) |
| 117 | |
| 118 | ctx := withNoClosedLoop(WithTaskBudget(context.Background(), TaskBudget{Tokens: 1})) |
| 119 | if err := coord.Run(ctx, "plan the budgeted change"); err != nil { |
| 120 | t.Fatalf("Run: %v", err) |
| 121 | } |
| 122 | if got := len(planner.requests); got != 2 { |
| 123 | t.Fatalf("planner requests = %d, want research plus terminal submit_plan", got) |
| 124 | } |
| 125 | if got := lastUser(planner.requests[1]); !strings.Contains(got, "reached its token budget") || !strings.Contains(got, "call submit_plan now") { |
| 126 | t.Fatalf("task-budget finalization did not request terminal plan: %q", got) |
| 127 | } |
| 128 | if len(exec.requests) == 0 || !strings.Contains(lastUser(exec.requests[0]), "finish within the safety envelope") { |
| 129 | t.Fatal("executor did not receive the task-budget terminal plan") |
| 130 | } |
| 131 | } |
| 132 |