| 1 | package control |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "encoding/json" |
| 6 | "strings" |
| 7 | "testing" |
| 8 | |
| 9 | "reasonix/internal/agent" |
| 10 | "reasonix/internal/event" |
| 11 | "reasonix/internal/provider" |
| 12 | "reasonix/internal/tool" |
| 13 | ) |
| 14 | |
| 15 | // approvalPlanTurn is planTurn with requires_approval set, gating execution |
| 16 | // behind the host approval gate. |
| 17 | func approvalPlanTurn(text string) []provider.Chunk { |
| 18 | args, _ := json.Marshal(map[string]any{ |
| 19 | "objective": text, |
| 20 | "requires_approval": true, |
| 21 | "steps": []map[string]string{{"title": text}}, |
| 22 | }) |
| 23 | return []provider.Chunk{ |
| 24 | {Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "plan-1", Name: "submit_plan", Arguments: string(args)}}, |
| 25 | {Type: provider.ChunkDone}, |
| 26 | } |
| 27 | } |
| 28 | |
| 29 | // scriptedTurns is a provider that replays a distinct chunk set per Stream call, |
| 30 | // so a controller turn that re-enters the agent (plan turn, then approved |
| 31 | // execution turn) sees a different model response each time. |
| 32 | type scriptedTurns struct { |
| 33 | turns [][]provider.Chunk |
| 34 | call int |
| 35 | } |
| 36 | |
| 37 | func (s *scriptedTurns) Name() string { return "scripted" } |
| 38 | |
| 39 | func (s *scriptedTurns) Stream(_ context.Context, _ provider.Request) (<-chan provider.Chunk, error) { |
| 40 | i := s.call |
| 41 | if i >= len(s.turns) { |
| 42 | i = len(s.turns) - 1 |
| 43 | } |
| 44 | s.call++ |
| 45 | ch := make(chan provider.Chunk, len(s.turns[i])) |
| 46 | for _, c := range s.turns[i] { |
| 47 | ch <- c |
| 48 | } |
| 49 | close(ch) |
| 50 | return ch, nil |
| 51 | } |
| 52 | |
| 53 | func firstUserMessage(msgs []provider.Message) string { |
| 54 | for _, m := range msgs { |
| 55 | if agent.IsUserAuthoredTurnMessage(m) { |
| 56 | if m.ProviderContent != "" { |
| 57 | return m.ProviderContent |
| 58 | } |
| 59 | return m.Content |
| 60 | } |
| 61 | } |
| 62 | return "" |
| 63 | } |
| 64 | |
| 65 | // planTurn delivers the plan text through submit_plan, the planner's only |
| 66 | // delivery channel; prose plans fail the turn as a protocol error. |
| 67 | func planTurn(text string) []provider.Chunk { |
| 68 | args, _ := json.Marshal(map[string]any{ |
| 69 | "objective": text, |
| 70 | "steps": []map[string]string{{"title": text}}, |
| 71 | }) |
| 72 | return []provider.Chunk{ |
| 73 | {Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "plan-1", Name: "submit_plan", Arguments: string(args)}}, |
| 74 | {Type: provider.ChunkDone}, |
| 75 | } |
| 76 | } |
| 77 | |
| 78 | func textTurn(text string) []provider.Chunk { |
| 79 | return []provider.Chunk{{Type: provider.ChunkText, Text: text}, {Type: provider.ChunkDone}} |
| 80 | } |
| 81 | |
| 82 | func readFileTurn() []provider.Chunk { |
| 83 | return []provider.Chunk{ |
| 84 | {Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "r1", Name: "read_file", Arguments: `{"path":"README.md"}`}}, |
| 85 | {Type: provider.ChunkDone}, |
| 86 | } |
| 87 | } |
| 88 | |
| 89 | func planThenExecuteTurns(plan, answer string) [][]provider.Chunk { |
| 90 | return [][]provider.Chunk{textTurn(plan), readFileTurn(), textTurn(answer)} |
| 91 | } |
| 92 | |
| 93 | func newPlanTestAgent(prov provider.Provider) *agent.Agent { |
| 94 | reg := tool.NewRegistry() |
| 95 | reg.Add(fakeControlTool{name: "read_file"}) |
| 96 | return agent.New(prov, reg, agent.NewSession(""), agent.Options{}, event.Discard) |
| 97 | } |
| 98 | |
| 99 | // TestPlanGateEndToEnd drives explicit Plan Mode through a real agent: the plan |
| 100 | // marker reaches the model, the controller asks for approval, and approval exits |
| 101 | // Plan Mode, seeds the task list, and runs the execution turn. |
| 102 | func TestPlanGateEndToEnd(t *testing.T) { |
| 103 | prov := &scriptedTurns{turns: planThenExecuteTurns( |
| 104 | "Plan:\n1. Add the config field\n2. Wire it into boot\n3. Add tests", |
| 105 | "Done — implemented the plan.", |
| 106 | )} |
| 107 | ag := newPlanTestAgent(prov) |
| 108 | |
| 109 | approvalID := make(chan string, 1) |
| 110 | c := newOwnedTestController(t, Options{ |
| 111 | Runner: ag, |
| 112 | Executor: ag, |
| 113 | Sink: event.FuncSink(func(e event.Event) { |
| 114 | switch e.Kind { |
| 115 | case event.ApprovalRequest: |
| 116 | approvalID <- e.Approval.ID |
| 117 | } |
| 118 | }), |
| 119 | }) |
| 120 | c.SetPlanMode(true) |
| 121 | |
| 122 | go func() { c.Approve(<-approvalID, true, false, false) }() |
| 123 | |
| 124 | input := "实现 issue #2395：新增配置项、自动判断复杂任务、补测试和文档" |
| 125 | if err := c.runTurnWithRaw(context.Background(), input, input); err != nil { |
| 126 | t.Fatalf("runTurnWithRaw: %v", err) |
| 127 | } |
| 128 | |
| 129 | msgs := ag.Session().Messages |
| 130 | if got := agent.StripTransientUserBlocks(firstUserMessage(msgs)); !strings.HasPrefix(got, PlanModeMarker) { |
| 131 | t.Fatalf("first model input = %q, want the plan marker prefixed", got) |
| 132 | } |
| 133 | if c.PlanMode() { |
| 134 | t.Fatal("plan mode should be off after approval") |
| 135 | } |
| 136 | if got := c.Todos(); len(got) != 0 { |
| 137 | t.Fatalf("approved plan populated todo state: %+v", got) |
| 138 | } |
| 139 | if got := lastAssistantText(msgs); got != "Done — implemented the plan." { |
| 140 | t.Fatalf("last assistant text = %q, want the execution turn's answer", got) |
| 141 | } |
| 142 | if prov.call != 3 { |
| 143 | t.Fatalf("provider called %d times, want 3 (plan + read + answer)", prov.call) |
| 144 | } |
| 145 | } |
| 146 | |
| 147 | func TestApprovedPlanDoesNotSeedTodosWithoutModelTodoWrite(t *testing.T) { |
| 148 | prov := &scriptedTurns{turns: planThenExecuteTurns( |
| 149 | "Plan:\n1. Add the config field\n2. Wire it into boot", |
| 150 | "Done.", |
| 151 | )} |
| 152 | ag := newPlanTestAgent(prov) |
| 153 | |
| 154 | approvalID := make(chan string, 1) |
| 155 | c := newOwnedTestController(t, Options{ |
| 156 | Runner: ag, |
| 157 | Executor: ag, |
| 158 | Sink: event.FuncSink(func(e event.Event) { |
| 159 | switch e.Kind { |
| 160 | case event.ApprovalRequest: |
| 161 | approvalID <- e.Approval.ID |
| 162 | } |
| 163 | }), |
| 164 | }) |
| 165 | c.SetPlanMode(true) |
| 166 | |
| 167 | go func() { c.Approve(<-approvalID, true, false, false) }() |
| 168 | |
| 169 | input := "Implement issue #2395: add config, wire boot, add tests and docs" |
| 170 | if err := c.runTurnWithRaw(context.Background(), input, input); err != nil { |
| 171 | t.Fatalf("runTurnWithRaw: %v", err) |
| 172 | } |
| 173 | |
| 174 | if got := c.Todos(); len(got) != 0 { |
| 175 | t.Fatalf("plan approval seeded todo state: %+v", got) |
| 176 | } |
| 177 | } |
| 178 | |
| 179 | // TestPlanGateRejectionStaysInPlan proves a rejected plan keeps plan mode on |
| 180 | // and never runs the execution turn: only the plan turn reached the model. |
| 181 | func TestPlanGateRejectionStaysInPlan(t *testing.T) { |
| 182 | prov := &scriptedTurns{turns: [][]provider.Chunk{ |
| 183 | textTurn("Plan:\n1. Add the config field\n2. Add tests"), |
| 184 | }} |
| 185 | ag := agent.New(prov, tool.NewRegistry(), agent.NewSession(""), agent.Options{}, event.Discard) |
| 186 | |
| 187 | approvalID := make(chan string, 1) |
| 188 | var seeded bool |
| 189 | c := newOwnedTestController(t, Options{ |
| 190 | Runner: ag, |
| 191 | Executor: ag, |
| 192 | Sink: event.FuncSink(func(e event.Event) { |
| 193 | switch e.Kind { |
| 194 | case event.ApprovalRequest: |
| 195 | approvalID <- e.Approval.ID |
| 196 | case event.ToolDispatch: |
| 197 | if e.Tool.ID == "plan-seed" { |
| 198 | seeded = true |
| 199 | } |
| 200 | } |
| 201 | }), |
| 202 | }) |
| 203 | c.SetPlanMode(true) |
| 204 | |
| 205 | go func() { c.Approve(<-approvalID, false, false, false) }() |
| 206 | |
| 207 | input := "实现 issue #2395：新增配置项、自动判断复杂任务、补测试和文档" |
| 208 | if err := c.runTurnWithRaw(context.Background(), input, input); err != nil { |
| 209 | t.Fatalf("runTurnWithRaw: %v", err) |
| 210 | } |
| 211 | |
| 212 | if !c.PlanMode() { |
| 213 | t.Fatal("rejected plan should keep plan mode on") |
| 214 | } |
| 215 | if seeded { |
| 216 | t.Fatal("rejected plan must not seed the task list") |
| 217 | } |
| 218 | if prov.call != 1 { |
| 219 | t.Fatalf("provider called %d times, want 1 (plan only, no execution)", prov.call) |
| 220 | } |
| 221 | } |
| 222 |