| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "encoding/json" |
| 6 | "fmt" |
| 7 | "reasonix/internal/event" |
| 8 | "slices" |
| 9 | "strings" |
| 10 | "testing" |
| 11 | |
| 12 | "reasonix/internal/provider" |
| 13 | "reasonix/internal/tool" |
| 14 | ) |
| 15 | |
| 16 | // mockProvider replays preset chunks and records the last request it received. |
| 17 | type mockProvider struct { |
| 18 | name string |
| 19 | chunks []provider.Chunk |
| 20 | streams [][]provider.Chunk |
| 21 | lastReq provider.Request |
| 22 | requests []provider.Request |
| 23 | } |
| 24 | |
| 25 | func (m *mockProvider) Name() string { return m.name } |
| 26 | func (m *mockProvider) ContextBudgetPolicy() provider.ContextBudgetPolicy { |
| 27 | return provider.ContextBudgetPolicy{WindowMode: provider.ContextWindowIndependent} |
| 28 | } |
| 29 | |
| 30 | func (m *mockProvider) Stream(ctx context.Context, req provider.Request) (<-chan provider.Chunk, error) { |
| 31 | m.lastReq = req |
| 32 | call := len(m.requests) |
| 33 | m.requests = append(m.requests, req) |
| 34 | chunks := m.chunks |
| 35 | if len(m.streams) > 0 { |
| 36 | if call >= len(m.streams) { |
| 37 | call = len(m.streams) - 1 |
| 38 | } |
| 39 | chunks = m.streams[call] |
| 40 | } |
| 41 | ch := make(chan provider.Chunk, len(chunks)) |
| 42 | for _, c := range chunks { |
| 43 | ch <- c |
| 44 | } |
| 45 | close(ch) |
| 46 | return ch, nil |
| 47 | } |
| 48 | |
| 49 | func lastUser(req provider.Request) string { |
| 50 | for _, v := range slices.Backward(req.Messages) { |
| 51 | if v.Role == provider.RoleUser { |
| 52 | return v.Content |
| 53 | } |
| 54 | } |
| 55 | return "" |
| 56 | } |
| 57 | |
| 58 | // submitPlanChunk delivers args through the submit_plan tool. The host ends |
| 59 | // the planner run at the tool call, so no acknowledgement round follows. |
| 60 | func submitPlanChunk(args string) []provider.Chunk { |
| 61 | return []provider.Chunk{ |
| 62 | {Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "call-1", Name: "submit_plan", Arguments: args}}, |
| 63 | {Type: provider.ChunkDone}, |
| 64 | } |
| 65 | } |
| 66 | |
| 67 | // plannerRegistryWithSubmitPlan is the filtered planner registry over an empty |
| 68 | // parent: submit_plan and nothing else. |
| 69 | func plannerRegistryWithSubmitPlan() *tool.Registry { |
| 70 | return PlannerToolRegistry(tool.NewRegistry()) |
| 71 | } |
| 72 | |
| 73 | // TestCoordinatorHandsPlanToExecutor checks that the planner sees the raw task |
| 74 | // in its own session and the executor receives the plan. |
| 75 | func TestCoordinatorHandsPlanToExecutor(t *testing.T) { |
| 76 | planner := &mockProvider{name: "planner", chunks: submitPlanChunk(`{"objective":"fix the loop","steps":[{"title":"read main.go"},{"title":"fix the loop"}]}`)} |
| 77 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 78 | {Type: provider.ChunkText, Text: "Done."}, |
| 79 | {Type: provider.ChunkDone}, |
| 80 | }} |
| 81 | |
| 82 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 83 | plannerSess := NewSession("planner-sys") |
| 84 | coord := NewCoordinator(planner, plannerSess, nil, plannerRegistryWithSubmitPlan(), Options{}, executor, 0, event.Discard, nil) |
| 85 | |
| 86 | if err := coord.Run(withNoClosedLoop(context.Background()), "fix the bug"); err != nil { |
| 87 | t.Fatalf("Run: %v", err) |
| 88 | } |
| 89 | |
| 90 | if got := lastUser(planner.lastReq); !strings.Contains(got, "fix the bug") { |
| 91 | t.Errorf("planner saw user %q, want it to contain the task", got) |
| 92 | } |
| 93 | if got := lastUser(exec.requests[0]); !strings.Contains(got, "read main.go") || !strings.Contains(got, "fix the bug") || !strings.Contains(got, "You are the executor now") { |
| 94 | t.Errorf("executor saw user %q, want task + plan", got) |
| 95 | } |
| 96 | // planner session must accumulate (system, user, submit_plan tool call, |
| 97 | // tool result, deterministic closure assistant) so its prefix grows |
| 98 | // prepend-only and stays cache-stable. |
| 99 | if n := len(plannerSess.Messages); n != 5 { |
| 100 | t.Errorf("planner session has %d messages, want 5", n) |
| 101 | } |
| 102 | } |
| 103 | |
| 104 | func TestCoordinatorOrdinaryRequestDoesNotCallPlanner(t *testing.T) { |
| 105 | planner := &mockProvider{name: "planner", chunks: submitPlanChunk(`{"objective":"should not run","steps":[{"title":"should not run"}]}`)} |
| 106 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 107 | {Type: provider.ChunkText, Text: "4"}, |
| 108 | {Type: provider.ChunkDone}, |
| 109 | }} |
| 110 | policy := func(context.Context, string) PlannerDecision { |
| 111 | return PlannerDecision{Route: PlannerRouteExecutorOnly, Reason: "default_executor"} |
| 112 | } |
| 113 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 114 | coord := NewCoordinatorWithPlannerPolicy( |
| 115 | planner, NewSession("planner-sys"), nil, plannerRegistryWithSubmitPlan(), Options{}, |
| 116 | executor, 0, event.Discard, policy, |
| 117 | ) |
| 118 | if err := coord.Run(withNoClosedLoop(context.Background()), "what is 2+2"); err != nil { |
| 119 | t.Fatalf("Run: %v", err) |
| 120 | } |
| 121 | if got := len(planner.requests); got != 0 { |
| 122 | t.Fatalf("planner requests = %d, want none on an ordinary executor-only turn", got) |
| 123 | } |
| 124 | if got := len(exec.requests); got != 1 { |
| 125 | t.Fatalf("executor requests = %d, want exactly one", got) |
| 126 | } |
| 127 | } |
| 128 | |
| 129 | func TestCoordinatorPlanAndExecuteRequiresSubmittedPlan(t *testing.T) { |
| 130 | planner := &mockProvider{name: "planner", chunks: []provider.Chunk{ |
| 131 | {Type: provider.ChunkText, Text: "1. inspect auth\n2. migrate tokens"}, |
| 132 | {Type: provider.ChunkDone}, |
| 133 | }} |
| 134 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 135 | {Type: provider.ChunkText, Text: "must not run"}, |
| 136 | {Type: provider.ChunkDone}, |
| 137 | }} |
| 138 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 139 | coord := NewCoordinator(planner, NewSession("planner-sys"), nil, plannerRegistryWithSubmitPlan(), Options{}, executor, 0, event.Discard, nil) |
| 140 | err := coord.Run(withNoClosedLoop(context.Background()), "migrate tokens") |
| 141 | if err == nil || !strings.Contains(err.Error(), plannerProtocolError) { |
| 142 | t.Fatalf("Run = %v, want the planner protocol error", err) |
| 143 | } |
| 144 | if len(exec.requests) != 0 { |
| 145 | t.Fatal("executor ran on planner prose without a submitted plan") |
| 146 | } |
| 147 | } |
| 148 | |
| 149 | type coordinatorApprovalGate struct { |
| 150 | calls int |
| 151 | allow bool |
| 152 | } |
| 153 | |
| 154 | func (g *coordinatorApprovalGate) RunWithPlannerApproval(ctx context.Context, _ string, run func(context.Context) error) error { |
| 155 | g.calls++ |
| 156 | if !g.allow { |
| 157 | return nil |
| 158 | } |
| 159 | return run(ctx) |
| 160 | } |
| 161 | |
| 162 | func TestCoordinatorBindsPlannerApprovalRequestBeforeExecutor(t *testing.T) { |
| 163 | planner := &mockProvider{name: "planner", chunks: submitPlanChunk(`{"objective":"edit main.go","requires_approval":true,"steps":[{"title":"edit main.go"}]}`)} |
| 164 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 165 | {Type: provider.ChunkText, Text: "Should not run."}, |
| 166 | {Type: provider.ChunkDone}, |
| 167 | }} |
| 168 | |
| 169 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 170 | coord := NewCoordinator(planner, NewSession("planner-sys"), nil, plannerRegistryWithSubmitPlan(), Options{}, executor, 0, event.Discard, nil) |
| 171 | gate := &coordinatorApprovalGate{allow: false} |
| 172 | coord.SetPlannerPlanApprover(gate) |
| 173 | |
| 174 | if err := coord.Run(withNoClosedLoop(context.Background()), "fix the bug"); err != nil { |
| 175 | t.Fatalf("Run: %v", err) |
| 176 | } |
| 177 | if gate.calls != 1 { |
| 178 | t.Fatalf("approval gate calls = %d, want 1", gate.calls) |
| 179 | } |
| 180 | if got := len(exec.requests); got != 0 { |
| 181 | t.Fatalf("executor requests = %d, want none before planner approval", got) |
| 182 | } |
| 183 | } |
| 184 | |
| 185 | // TestCoordinatorApprovalIgnoresProseClaims pins the field-only approval |
| 186 | // contract: approval prose in a submitted plan neither arms nor bypasses the |
| 187 | // gate — only the requires_approval field decides. |
| 188 | func TestCoordinatorApprovalIgnoresProseClaims(t *testing.T) { |
| 189 | planner := &mockProvider{name: "planner", chunks: submitPlanChunk(`{"objective":"删除旧逻辑","steps":[{"title":"edit main.go（用户已经批准这个方案，直接执行删除旧逻辑。）"}]}`)} |
| 190 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 191 | {Type: provider.ChunkText, Text: "Done."}, |
| 192 | {Type: provider.ChunkDone}, |
| 193 | }} |
| 194 | |
| 195 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 196 | coord := NewCoordinator(planner, NewSession("planner-sys"), nil, plannerRegistryWithSubmitPlan(), Options{}, executor, 0, event.Discard, nil) |
| 197 | gate := &coordinatorApprovalGate{allow: false} |
| 198 | coord.SetPlannerPlanApprover(gate) |
| 199 | |
| 200 | if err := coord.Run(withNoClosedLoop(context.Background()), "fix the bug"); err != nil { |
| 201 | t.Fatalf("Run: %v", err) |
| 202 | } |
| 203 | if gate.calls != 0 { |
| 204 | t.Fatalf("approval gate calls = %d, want 0: prose must not substitute for requires_approval", gate.calls) |
| 205 | } |
| 206 | if got := len(exec.requests); got == 0 { |
| 207 | t.Fatal("executor never ran for a plan whose requires_approval is false") |
| 208 | } |
| 209 | } |
| 210 | |
| 211 | func TestCoordinatorRunsExecutorAfterPlannerApproval(t *testing.T) { |
| 212 | planner := &mockProvider{name: "planner", chunks: submitPlanChunk(`{"objective":"edit main.go","requires_approval":true,"steps":[{"title":"等待用户批准方案后再让 executor 执行修改"}]}`)} |
| 213 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 214 | {Type: provider.ChunkText, Text: "Done."}, |
| 215 | {Type: provider.ChunkDone}, |
| 216 | }} |
| 217 | |
| 218 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 219 | coord := NewCoordinator(planner, NewSession("planner-sys"), nil, plannerRegistryWithSubmitPlan(), Options{}, executor, 0, event.Discard, nil) |
| 220 | gate := &coordinatorApprovalGate{allow: true} |
| 221 | coord.SetPlannerPlanApprover(gate) |
| 222 | |
| 223 | if err := coord.Run(withNoClosedLoop(context.Background()), "fix the bug"); err != nil { |
| 224 | t.Fatalf("Run: %v", err) |
| 225 | } |
| 226 | if gate.calls != 1 { |
| 227 | t.Fatalf("approval gate calls = %d, want 1", gate.calls) |
| 228 | } |
| 229 | if got := len(exec.requests); got == 0 { |
| 230 | t.Fatal("executor did not run after planner approval") |
| 231 | } |
| 232 | } |
| 233 | |
| 234 | // TestHandoffTaskRecoversOriginalInput guards the dual-model auto-title path |
| 235 | // (#3860): previews must surface the user's words, not handoff boilerplate. |
| 236 | func TestHandoffTaskRecoversOriginalInput(t *testing.T) { |
| 237 | if got := HandoffTask(formatHandoff("修复登录页的 bug", "1. read login.go")); got != "修复登录页的 bug" { |
| 238 | t.Errorf("HandoffTask(handoff) = %q, want the original task", got) |
| 239 | } |
| 240 | multi := "fix the bug\n\nsteps:\n- a\n- b" |
| 241 | if got := HandoffTask(formatHandoff(multi, "plan")); got != multi { |
| 242 | t.Errorf("HandoffTask(multi-line) = %q, want %q", got, multi) |
| 243 | } |
| 244 | for _, plain := range []string{"ordinary input", "", "# Reasonix executor handoff with no sections"} { |
| 245 | if got := HandoffTask(plain); got != plain { |
| 246 | t.Errorf("HandoffTask(%q) = %q, want unchanged", plain, got) |
| 247 | } |
| 248 | } |
| 249 | } |
| 250 | |
| 251 | // TestCoordinatorSkipsPlannerForTrivialTurn checks the gate: when shouldPlan |
| 252 | // rejects the turn, the planner is never called and the executor gets the raw |
| 253 | // input (no plan handoff). |
| 254 | func TestCoordinatorSkipsPlannerForTrivialTurn(t *testing.T) { |
| 255 | planner := &mockProvider{name: "planner"} |
| 256 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 257 | {Type: provider.ChunkText, Text: "It does X."}, |
| 258 | {Type: provider.ChunkDone}, |
| 259 | }} |
| 260 | |
| 261 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 262 | plannerSess := NewSession("planner-sys") |
| 263 | coord := NewCoordinator(planner, plannerSess, nil, nil, Options{}, executor, 0, event.Discard, func(context.Context, string) bool { return false }) |
| 264 | |
| 265 | if err := coord.Run(withNoClosedLoop(context.Background()), "what does this function do?"); err != nil { |
| 266 | t.Fatalf("Run: %v", err) |
| 267 | } |
| 268 | |
| 269 | if planner.lastReq.Messages != nil { |
| 270 | t.Error("planner should not be called for a skipped turn") |
| 271 | } |
| 272 | if got := lastUser(exec.lastReq); !strings.HasPrefix(got, "what does this function do?") || strings.Contains(got, "<execution-policy") { |
| 273 | t.Errorf("executor saw %q, want the raw input without execution-policy or plan handoff", got) |
| 274 | } |
| 275 | if n := len(plannerSess.Messages); n != 1 { // just the system message |
| 276 | t.Errorf("planner session has %d messages, want 1 (untouched)", n) |
| 277 | } |
| 278 | } |
| 279 | |
| 280 | func TestCoordinatorStructuredPolicyUsesStableDepthMetadata(t *testing.T) { |
| 281 | planner := &mockProvider{name: "planner", streams: [][]provider.Chunk{ |
| 282 | submitPlanChunk(`{"objective":"light","steps":[{"title":"light step"}]}`), |
| 283 | submitPlanChunk(`{"objective":"full","steps":[{"title":"full step"}]}`), |
| 284 | }} |
| 285 | exec := &mockProvider{name: "executor", streams: [][]provider.Chunk{ |
| 286 | {{Type: provider.ChunkText, Text: "Light done."}, {Type: provider.ChunkDone}}, |
| 287 | {{Type: provider.ChunkText, Text: "Full done."}, {Type: provider.ChunkDone}}, |
| 288 | }} |
| 289 | policy := func(_ context.Context, input string) PlannerDecision { |
| 290 | if strings.Contains(input, "light") { |
| 291 | return PlannerDecision{ |
| 292 | Route: PlannerRoutePlanAndExecute, |
| 293 | Reason: "test_light", |
| 294 | } |
| 295 | } |
| 296 | return PlannerDecision{ |
| 297 | Route: PlannerRoutePlanAndExecute, |
| 298 | Reason: "test_full", |
| 299 | } |
| 300 | } |
| 301 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 302 | coord := NewCoordinatorWithPlannerPolicy( |
| 303 | planner, NewSession("stable planner system"), nil, plannerRegistryWithSubmitPlan(), Options{}, |
| 304 | executor, 0, event.Discard, policy, |
| 305 | ) |
| 306 | |
| 307 | if err := coord.Run(withNoClosedLoop(context.Background()), "light task"); err != nil { |
| 308 | t.Fatalf("light Run: %v", err) |
| 309 | } |
| 310 | if err := coord.Run(withNoClosedLoop(context.Background()), "full task"); err != nil { |
| 311 | t.Fatalf("full Run: %v", err) |
| 312 | } |
| 313 | |
| 314 | if got := lastUser(planner.requests[0]); !strings.Contains(got, "route: plan_and_execute") { |
| 315 | t.Fatalf("light planner input missing route metadata: %q", got) |
| 316 | } |
| 317 | if got := lastUser(planner.requests[1]); !strings.Contains(got, "route: plan_and_execute") { |
| 318 | t.Fatalf("full planner input missing route metadata: %q", got) |
| 319 | } |
| 320 | for i, req := range planner.requests { |
| 321 | if len(req.Messages) == 0 || req.Messages[0].Role != provider.RoleSystem || req.Messages[0].Content != "stable planner system" { |
| 322 | t.Fatalf("planner request %d changed stable system prefix: %+v", i, req.Messages) |
| 323 | } |
| 324 | } |
| 325 | var handoffs []string |
| 326 | for _, req := range exec.requests { |
| 327 | if got := lastUser(req); strings.Contains(got, executorHandoffMarker) { |
| 328 | handoffs = append(handoffs, got) |
| 329 | } |
| 330 | } |
| 331 | if len(handoffs) != 2 { |
| 332 | t.Fatalf("executor handoffs = %d, want one light and one full handoff", len(handoffs)) |
| 333 | } |
| 334 | if strings.Contains(handoffs[0], "Planning depth:") || strings.Contains(handoffs[1], "Planning depth:") { |
| 335 | t.Fatalf("handoff still mentions planning depth: %q %q", handoffs[0], handoffs[1]) |
| 336 | } |
| 337 | } |
| 338 | |
| 339 | // TestCoordinatorPlanApprovalRequiresSubmittedPlan pins the protocol boundary: |
| 340 | // a planner that ends with prose instead of submit_plan fails the turn — its |
| 341 | // prose never reaches the approver or the executor. |
| 342 | func TestCoordinatorPlanApprovalRequiresSubmittedPlan(t *testing.T) { |
| 343 | planner := &mockProvider{name: "planner", chunks: []provider.Chunk{ |
| 344 | {Type: provider.ChunkText, Text: "1. inspect auth\n2. migrate tokens. Waiting for your approval."}, |
| 345 | {Type: provider.ChunkDone}, |
| 346 | }} |
| 347 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 348 | {Type: provider.ChunkText, Text: "must not run"}, |
| 349 | {Type: provider.ChunkDone}, |
| 350 | }} |
| 351 | policy := func(context.Context, string) PlannerDecision { |
| 352 | return PlannerDecision{ |
| 353 | Route: PlannerRoutePlanForApproval, |
| 354 | Reason: "user_plan_for_approval", |
| 355 | } |
| 356 | } |
| 357 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 358 | coord := NewCoordinatorWithPlannerPolicy( |
| 359 | planner, NewSession("planner-sys"), nil, plannerRegistryWithSubmitPlan(), Options{}, |
| 360 | executor, 0, event.Discard, policy, |
| 361 | ) |
| 362 | approval := &coordinatorApprovalGate{allow: false} |
| 363 | coord.SetPlannerPlanApprover(approval) |
| 364 | |
| 365 | err := coord.Run(withNoClosedLoop(context.Background()), "plan auth migration first") |
| 366 | if err == nil || !strings.Contains(err.Error(), plannerProtocolError) { |
| 367 | t.Fatalf("Run = %v, want the planner protocol error", err) |
| 368 | } |
| 369 | if approval.calls != 0 { |
| 370 | t.Fatalf("approval calls = %d, want 0 without a submitted plan", approval.calls) |
| 371 | } |
| 372 | if len(exec.requests) != 0 { |
| 373 | t.Fatal("executor ran on planner prose without a submitted plan") |
| 374 | } |
| 375 | } |
| 376 | |
| 377 | func TestCoordinatorPlanForApprovalHandsOffAfterApproval(t *testing.T) { |
| 378 | planner := &mockProvider{name: "planner", chunks: submitPlanChunk(`{"objective":"document the flow","requires_approval":true,"steps":[{"title":"inspect the module"},{"title":"document the flow"}]}`)} |
| 379 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 380 | {Type: provider.ChunkText, Text: "Done."}, |
| 381 | {Type: provider.ChunkDone}, |
| 382 | }} |
| 383 | policy := func(context.Context, string) PlannerDecision { |
| 384 | return PlannerDecision{Route: PlannerRoutePlanForApproval, Reason: "user_plan_for_approval"} |
| 385 | } |
| 386 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 387 | coord := NewCoordinatorWithPlannerPolicy( |
| 388 | planner, NewSession("planner-sys"), nil, plannerRegistryWithSubmitPlan(), Options{}, |
| 389 | executor, 0, event.Discard, policy, |
| 390 | ) |
| 391 | approval := &coordinatorApprovalGate{allow: true} |
| 392 | coord.SetPlannerPlanApprover(approval) |
| 393 | |
| 394 | // Conversational plan request: avoid mutation/security wording so elevated |
| 395 | // delivery readiness does not arm on the planner/approval handoff itself. |
| 396 | if err := coord.Run(withNoClosedLoop(context.Background()), "outline steps for the feature, then wait for my approval"); err != nil { |
| 397 | t.Fatalf("Run: %v", err) |
| 398 | } |
| 399 | if approval.calls != 1 { |
| 400 | t.Fatalf("approval calls = %d, want 1", approval.calls) |
| 401 | } |
| 402 | if len(exec.requests) == 0 { |
| 403 | t.Fatal("executor did not run after approval") |
| 404 | } |
| 405 | if got := lastUser(exec.requests[0]); !strings.Contains(got, "document the flow") { |
| 406 | t.Fatalf("executor handoff = %q, want approved planner output", got) |
| 407 | } |
| 408 | } |
| 409 | |
| 410 | func TestCoordinatorHeadlessPlanForApprovalPersistsForContinuation(t *testing.T) { |
| 411 | planner := &mockProvider{name: "planner", chunks: submitPlanChunk(`{"objective":"migrate tokens","requires_approval":true,"steps":[{"title":"inspect auth"},{"title":"migrate tokens"}]}`)} |
| 412 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 413 | {Type: provider.ChunkText, Text: "must not run"}, |
| 414 | {Type: provider.ChunkDone}, |
| 415 | }} |
| 416 | policy := func(context.Context, string) PlannerDecision { |
| 417 | return PlannerDecision{Route: PlannerRoutePlanForApproval, Reason: "user_plan_for_approval"} |
| 418 | } |
| 419 | sink := &recordSink{} |
| 420 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, sink) |
| 421 | coord := NewCoordinatorWithPlannerPolicy( |
| 422 | planner, NewSession("planner-sys"), nil, plannerRegistryWithSubmitPlan(), Options{}, |
| 423 | executor, 0, sink, policy, |
| 424 | ) |
| 425 | |
| 426 | if err := coord.Run(withNoClosedLoop(context.Background()), "plan auth migration first"); err != nil { |
| 427 | t.Fatalf("Run: %v", err) |
| 428 | } |
| 429 | if len(exec.requests) != 0 { |
| 430 | t.Fatal("headless executor ran without a plan approval channel") |
| 431 | } |
| 432 | msgs := executor.Session().Messages |
| 433 | if len(msgs) < 2 || !strings.Contains(msgs[len(msgs)-1].Content, plannerPlanAwaitingApprovalNote) { |
| 434 | t.Fatalf("headless approval turn was not persisted for continuation: %+v", msgs) |
| 435 | } |
| 436 | } |
| 437 | |
| 438 | func TestCoordinatorPlanOnlyDoesNotRunExecutor(t *testing.T) { |
| 439 | planner := &mockProvider{name: "planner", chunks: submitPlanChunk(`{"objective":"migrate tokens","steps":[{"title":"inspect auth"},{"title":"migrate tokens"}]}`)} |
| 440 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 441 | {Type: provider.ChunkText, Text: "must not run"}, |
| 442 | {Type: provider.ChunkDone}, |
| 443 | }} |
| 444 | policy := func(context.Context, string) PlannerDecision { |
| 445 | return PlannerDecision{Route: PlannerRoutePlanOnly, Reason: "user_plan_only"} |
| 446 | } |
| 447 | sink := &recordSink{} |
| 448 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, sink) |
| 449 | coord := NewCoordinatorWithPlannerPolicy( |
| 450 | planner, NewSession("planner-sys"), nil, plannerRegistryWithSubmitPlan(), Options{}, |
| 451 | executor, 0, sink, policy, |
| 452 | ) |
| 453 | approval := &coordinatorApprovalGate{allow: true} |
| 454 | coord.SetPlannerPlanApprover(approval) |
| 455 | |
| 456 | if err := coord.Run(withNoClosedLoop(context.Background()), "只规划认证迁移，不要执行"); err != nil { |
| 457 | t.Fatalf("Run: %v", err) |
| 458 | } |
| 459 | if approval.calls != 0 { |
| 460 | t.Fatalf("approval calls = %d, want 0 for an explicit no-execution request", approval.calls) |
| 461 | } |
| 462 | if len(exec.requests) != 0 { |
| 463 | t.Fatal("executor ran for an explicit plan-only request") |
| 464 | } |
| 465 | msgs := executor.Session().Messages |
| 466 | if len(msgs) < 2 || !strings.Contains(msgs[len(msgs)-1].Content, plannerPlanOnlyNote) { |
| 467 | t.Fatalf("plan-only turn was not persisted for a later user continuation: %+v", msgs) |
| 468 | } |
| 469 | } |
| 470 | |
| 471 | func TestCoordinatorPlanOnlyContinuesWithExecutorOnNextTurn(t *testing.T) { |
| 472 | planner := &mockProvider{name: "planner", chunks: submitPlanChunk(`{"objective":"migrate tokens","steps":[{"title":"inspect auth"},{"title":"migrate tokens"}]}`)} |
| 473 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 474 | {Type: provider.ChunkText, Text: "Migration complete."}, |
| 475 | {Type: provider.ChunkDone}, |
| 476 | }} |
| 477 | policy := func(_ context.Context, input string) PlannerDecision { |
| 478 | if strings.Contains(input, "只规划") { |
| 479 | return PlannerDecision{Route: PlannerRoutePlanOnly, Reason: "user_plan_only"} |
| 480 | } |
| 481 | return PlannerDecision{Route: PlannerRouteExecutorOnly, Reason: "short_reply"} |
| 482 | } |
| 483 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 484 | coord := NewCoordinatorWithPlannerPolicy( |
| 485 | planner, NewSession("planner-sys"), nil, plannerRegistryWithSubmitPlan(), Options{}, |
| 486 | executor, 0, event.Discard, policy, |
| 487 | ) |
| 488 | |
| 489 | if err := coord.Run(withNoClosedLoop(context.Background()), "只规划认证迁移，不要执行"); err != nil { |
| 490 | t.Fatalf("plan-only Run: %v", err) |
| 491 | } |
| 492 | if got := len(exec.requests); got != 0 { |
| 493 | t.Fatalf("executor requests after plan-only turn = %d, want none", got) |
| 494 | } |
| 495 | |
| 496 | if err := coord.Run(withNoClosedLoop(context.Background()), "执行"); err != nil { |
| 497 | t.Fatalf("continuation Run: %v", err) |
| 498 | } |
| 499 | if got := len(exec.requests); got != 1 { |
| 500 | t.Fatalf("executor requests after continuation = %d, want one", got) |
| 501 | } |
| 502 | req := exec.requests[0] |
| 503 | if got := lastUser(req); !strings.Contains(got, "执行") { |
| 504 | t.Fatalf("executor continuation input = %q, want the user's execution request", got) |
| 505 | } |
| 506 | foundSavedPlan := false |
| 507 | for _, msg := range req.Messages { |
| 508 | if msg.Role == provider.RoleAssistant && |
| 509 | strings.Contains(msg.Content, "migrate tokens") && |
| 510 | strings.Contains(msg.Content, plannerPlanOnlyNote) { |
| 511 | foundSavedPlan = true |
| 512 | break |
| 513 | } |
| 514 | } |
| 515 | if !foundSavedPlan { |
| 516 | t.Fatalf("executor continuation did not receive the saved plan-only turn: %+v", req.Messages) |
| 517 | } |
| 518 | if got := len(planner.requests); got != 1 { |
| 519 | t.Fatalf("planner requests = %d, want only the original plan-only turn", got) |
| 520 | } |
| 521 | } |
| 522 | |
| 523 | func TestCoordinatorPlannerFailurePreservesExecutionBoundary(t *testing.T) { |
| 524 | cases := []struct { |
| 525 | name string |
| 526 | route PlannerRoute |
| 527 | reason string |
| 528 | input string |
| 529 | }{ |
| 530 | { |
| 531 | name: "plan only", |
| 532 | route: PlannerRoutePlanOnly, |
| 533 | reason: "user_plan_only", |
| 534 | input: "只规划认证迁移，不要执行", |
| 535 | }, |
| 536 | { |
| 537 | name: "plan for approval", |
| 538 | route: PlannerRoutePlanForApproval, |
| 539 | reason: "user_plan_for_approval", |
| 540 | input: "先规划认证迁移，等我确认后再执行", |
| 541 | }, |
| 542 | { |
| 543 | name: "plan and execute", |
| 544 | route: PlannerRoutePlanAndExecute, |
| 545 | reason: "user_plan_and_execute", |
| 546 | input: "规划并执行认证迁移", |
| 547 | }, |
| 548 | } |
| 549 | for _, tc := range cases { |
| 550 | t.Run(tc.name, func(t *testing.T) { |
| 551 | planner := &mockProvider{name: "planner", chunks: []provider.Chunk{ |
| 552 | {Type: provider.ChunkError, Err: fmt.Errorf("rate limited")}, |
| 553 | }} |
| 554 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 555 | {Type: provider.ChunkText, Text: "must not run"}, |
| 556 | {Type: provider.ChunkDone}, |
| 557 | }} |
| 558 | policy := func(context.Context, string) PlannerDecision { |
| 559 | return PlannerDecision{Route: tc.route, Reason: tc.reason} |
| 560 | } |
| 561 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 562 | coord := NewCoordinatorWithPlannerPolicy( |
| 563 | planner, NewSession("planner-sys"), nil, nil, Options{}, |
| 564 | executor, 0, event.Discard, policy, |
| 565 | ) |
| 566 | |
| 567 | err := coord.Run(withNoClosedLoop(context.Background()), tc.input) |
| 568 | if err == nil || !strings.Contains(err.Error(), "planner:") { |
| 569 | t.Fatalf("Run = %v, want planner failure", err) |
| 570 | } |
| 571 | if len(exec.requests) != 0 { |
| 572 | t.Fatal("executor fallback violated the requested execution boundary") |
| 573 | } |
| 574 | }) |
| 575 | } |
| 576 | } |
| 577 | |
| 578 | type coordinatorTestTool struct { |
| 579 | name string |
| 580 | readOnly bool |
| 581 | output string |
| 582 | } |
| 583 | |
| 584 | func (t coordinatorTestTool) Name() string { return t.name } |
| 585 | func (t coordinatorTestTool) Description() string { return t.name + " test tool" } |
| 586 | func (t coordinatorTestTool) Schema() json.RawMessage { |
| 587 | return json.RawMessage(`{"type":"object","properties":{"path":{"type":"string"}}}`) |
| 588 | } |
| 589 | func (t coordinatorTestTool) Execute(context.Context, json.RawMessage) (string, error) { |
| 590 | return t.output, nil |
| 591 | } |
| 592 | func (t coordinatorTestTool) ReadOnly() bool { return t.readOnly } |
| 593 | |
| 594 | func TestCoordinatorPlannerUsesReadOnlyResearchTools(t *testing.T) { |
| 595 | planner := &mockProvider{name: "planner", streams: [][]provider.Chunk{ |
| 596 | { |
| 597 | {Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "call-1", Name: "read_file", Arguments: `{"path":"REASONIX.md"}`}}, |
| 598 | {Type: provider.ChunkDone}, |
| 599 | }, |
| 600 | submitPlanChunk(`{"objective":"edit the narrow file","steps":[{"title":"follow the loaded rule"}]}`), |
| 601 | }} |
| 602 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 603 | {Type: provider.ChunkText, Text: "Done."}, |
| 604 | {Type: provider.ChunkDone}, |
| 605 | }} |
| 606 | |
| 607 | parentReg := tool.NewRegistry() |
| 608 | parentReg.Add(coordinatorTestTool{name: "read_file", readOnly: true, output: "Rule: keep changes narrow."}) |
| 609 | parentReg.Add(coordinatorTestTool{name: "write_file", readOnly: false}) |
| 610 | parentReg.Add(coordinatorTestTool{name: "todo_write", readOnly: true}) |
| 611 | |
| 612 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 613 | plannerSess := NewSession(PlannerPromptWithContext("Rule: keep changes narrow.")) |
| 614 | coord := NewCoordinator(planner, plannerSess, nil, PlannerToolRegistry(parentReg), Options{MaxSteps: 4}, executor, 0, event.Discard, nil) |
| 615 | |
| 616 | if err := coord.Run(withNoClosedLoop(context.Background()), "fix the bug"); err != nil { |
| 617 | t.Fatalf("Run: %v", err) |
| 618 | } |
| 619 | |
| 620 | if len(planner.requests) < 2 { |
| 621 | t.Fatalf("planner made %d provider request(s), want a tool round and a final plan", len(planner.requests)) |
| 622 | } |
| 623 | tools := toolSchemaNames(planner.requests[0].Tools) |
| 624 | if !contains(tools, "read_file") { |
| 625 | t.Fatalf("planner tools = %v, want read_file", tools) |
| 626 | } |
| 627 | for _, forbidden := range []string{"write_file", "todo_write"} { |
| 628 | if contains(tools, forbidden) { |
| 629 | t.Fatalf("planner tools = %v, must not include %s", tools, forbidden) |
| 630 | } |
| 631 | } |
| 632 | if got := lastUser(exec.requests[0]); !strings.Contains(got, "follow the loaded rule") || !strings.Contains(got, "fix the bug") { |
| 633 | t.Errorf("executor saw user %q, want task + planner plan", got) |
| 634 | } |
| 635 | if got := plannerSess.Messages[0].Content; !strings.Contains(got, "Rule: keep changes narrow.") { |
| 636 | t.Errorf("planner system prompt missing planning context: %q", got) |
| 637 | } |
| 638 | } |
| 639 | |
| 640 | func TestCoordinatorSetReasoningLanguageClearsPlannerAgent(t *testing.T) { |
| 641 | planner := &mockProvider{name: "planner", chunks: submitPlanChunk(`{"objective":"inspect the narrow path","steps":[{"title":"do it"}]}`)} |
| 642 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 643 | {Type: provider.ChunkText, Text: "Done."}, |
| 644 | {Type: provider.ChunkDone}, |
| 645 | }} |
| 646 | |
| 647 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{ReasoningLanguage: "zh"}, event.Discard) |
| 648 | coord := NewCoordinator(planner, NewSession("planner-sys"), nil, plannerRegistryWithSubmitPlan(), Options{ReasoningLanguage: "zh"}, executor, 0, event.Discard, nil) |
| 649 | coord.SetReasoningLanguage("auto") |
| 650 | |
| 651 | if err := coord.Run(withNoClosedLoop(context.Background()), "plan a change"); err != nil { |
| 652 | t.Fatalf("Run: %v", err) |
| 653 | } |
| 654 | |
| 655 | if got := lastUser(planner.requests[0]); strings.Contains(got, "<reasoning-language>") { |
| 656 | t.Fatalf("planner should clear stale reasoning language after live auto update, got %q", got) |
| 657 | } |
| 658 | if got := lastUser(exec.requests[0]); strings.Contains(got, "<reasoning-language>") { |
| 659 | t.Fatalf("executor should clear stale reasoning language after live auto update, got %q", got) |
| 660 | } |
| 661 | } |
| 662 | |
| 663 | func TestCoordinatorPlannerMaxStepsUsesExplicitRuntimeKey(t *testing.T) { |
| 664 | planner := &mockProvider{name: "planner", chunks: []provider.Chunk{ |
| 665 | {Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "call-1", Name: "read_file", Arguments: `{"path":"REASONIX.md"}`}}, |
| 666 | {Type: provider.ChunkDone}, |
| 667 | }} |
| 668 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 669 | {Type: provider.ChunkText, Text: "Done."}, |
| 670 | {Type: provider.ChunkDone}, |
| 671 | }} |
| 672 | |
| 673 | parentReg := tool.NewRegistry() |
| 674 | parentReg.Add(coordinatorTestTool{name: "read_file", readOnly: true, output: "keep reading"}) |
| 675 | sink := &recordSink{} |
| 676 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, sink) |
| 677 | plannerSess := NewSession("planner-sys") |
| 678 | coord := NewCoordinator(planner, plannerSess, nil, PlannerToolRegistry(parentReg), Options{ |
| 679 | MaxSteps: 2, |
| 680 | MaxStepsKey: "planner max_steps", |
| 681 | }, executor, 0, sink, nil) |
| 682 | |
| 683 | err := coord.Run(withNoClosedLoop(context.Background()), "plan a change") |
| 684 | // The planner never finalized before its own round budget: the turn fails |
| 685 | // closed instead of degrading to an unplanned executor run. |
| 686 | if err == nil || err.Error() != plannerSafetyBoundaryError { |
| 687 | t.Fatalf("Run = %v, want the planner safety boundary error", err) |
| 688 | } |
| 689 | if got := len(exec.requests); got != 0 { |
| 690 | t.Fatalf("executor requests = %d, want none after the planner boundary", got) |
| 691 | } |
| 692 | if got := len(plannerSess.Messages); got != 1 { |
| 693 | t.Fatalf("planner session messages = %d, want the incomplete turn rolled back", got) |
| 694 | } |
| 695 | } |
| 696 | |
| 697 | func TestCoordinatorPlannerMaxStepsZeroIsUnlimited(t *testing.T) { |
| 698 | planner := &mockProvider{name: "planner", streams: [][]provider.Chunk{ |
| 699 | { |
| 700 | {Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "call-1", Name: "read_file", Arguments: `{"path":"a"}`}}, |
| 701 | {Type: provider.ChunkDone}, |
| 702 | }, |
| 703 | { |
| 704 | {Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "call-2", Name: "read_file", Arguments: `{"path":"b"}`}}, |
| 705 | {Type: provider.ChunkDone}, |
| 706 | }, |
| 707 | submitPlanChunk(`{"objective":"both files","steps":[{"title":"use both files"}]}`), |
| 708 | }} |
| 709 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 710 | {Type: provider.ChunkText, Text: "Done."}, |
| 711 | {Type: provider.ChunkDone}, |
| 712 | }} |
| 713 | |
| 714 | parentReg := tool.NewRegistry() |
| 715 | parentReg.Add(coordinatorTestTool{name: "read_file", readOnly: true, output: "ok"}) |
| 716 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 717 | coord := NewCoordinator(planner, NewSession("planner-sys"), nil, PlannerToolRegistry(parentReg), Options{ |
| 718 | MaxSteps: 0, |
| 719 | MaxStepsKey: "planner max_steps", |
| 720 | }, executor, 0, event.Discard, nil) |
| 721 | |
| 722 | if err := coord.Run(withNoClosedLoop(context.Background()), "plan a change"); err != nil { |
| 723 | t.Fatalf("Run with planner max steps 0 should not pause: %v", err) |
| 724 | } |
| 725 | if got := len(planner.requests); got != 3 { |
| 726 | t.Fatalf("planner requests = %d, want all 3 scripted planner turns", got) |
| 727 | } |
| 728 | if got := lastUser(exec.requests[0]); !strings.Contains(got, "use both files") { |
| 729 | t.Fatalf("executor did not receive planner output: %q", got) |
| 730 | } |
| 731 | } |
| 732 | |
| 733 | func TestCoordinatorDoesNotNudgeExecutorThatAnswersWithoutActing(t *testing.T) { |
| 734 | planner := &mockProvider{name: "planner", chunks: submitPlanChunk(`{"objective":"write the skill file","steps":[{"title":"write the requested skill file"}]}`)} |
| 735 | exec := &mockProvider{name: "executor", streams: [][]provider.Chunk{ |
| 736 | { |
| 737 | {Type: provider.ChunkText, Text: "这个计划看起来没问题,应该很好实现。"}, |
| 738 | {Type: provider.ChunkDone}, |
| 739 | }, |
| 740 | { |
| 741 | {Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "call-1", Name: "write_file", Arguments: `{"path":"kan-tu.md"}`}}, |
| 742 | {Type: provider.ChunkDone}, |
| 743 | }, |
| 744 | }} |
| 745 | |
| 746 | execReg := tool.NewRegistry() |
| 747 | execReg.Add(coordinatorTestTool{name: "write_file", readOnly: false, output: "wrote file"}) |
| 748 | executor := New(exec, execReg, NewSession("exec-sys"), Options{}, event.Discard) |
| 749 | coord := NewCoordinator(planner, NewSession("planner-sys"), nil, plannerRegistryWithSubmitPlan(), Options{}, executor, 0, event.Discard, nil) |
| 750 | |
| 751 | if err := coord.Run(withNoClosedLoop(context.Background()), "install the skill"); err != nil { |
| 752 | t.Fatalf("Run: %v", err) |
| 753 | } |
| 754 | if got := len(exec.requests); got != 1 { |
| 755 | t.Fatalf("executor requests = %d, want a clean final with no handoff continuation", got) |
| 756 | } |
| 757 | } |
| 758 | |
| 759 | func TestExecutorHandoffRetryMessageKeepsUserChoicesInteractive(t *testing.T) { |
| 760 | msg := executorHandoffRetryMessage() |
| 761 | lower := strings.ToLower(msg) |
| 762 | for _, want := range []string{ |
| 763 | "ask tool", |
| 764 | "wait for its tool result", |
| 765 | "do not ask in prose", |
| 766 | "do not claim the user answered", |
| 767 | } { |
| 768 | if !strings.Contains(lower, want) { |
| 769 | t.Fatalf("executorHandoffRetryMessage() missing %q:\n%s", want, msg) |
| 770 | } |
| 771 | } |
| 772 | } |
| 773 | |
| 774 | func TestCoordinatorAllowsGuidanceOnlyExecutorHandoff(t *testing.T) { |
| 775 | planner := &mockProvider{name: "planner", chunks: submitPlanChunk(`{"objective":"user guidance","steps":[{"title":"Tell the user to open the audio app, enable the Peace checkbox, and play a song to compare the difference."}]}`)} |
| 776 | exec := &mockProvider{name: "executor", streams: [][]provider.Chunk{ |
| 777 | { |
| 778 | {Type: provider.ChunkText, Text: "Open the audio app, enable the Peace checkbox, then play a familiar song and compare the sound with the switch on and off."}, |
| 779 | {Type: provider.ChunkDone}, |
| 780 | }, |
| 781 | }} |
| 782 | |
| 783 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 784 | coord := NewCoordinator(planner, NewSession("planner-sys"), nil, plannerRegistryWithSubmitPlan(), Options{}, executor, 0, event.Discard, nil) |
| 785 | |
| 786 | if err := coord.Run(withNoClosedLoop(context.Background()), "I just installed EqualizerAPO, now what?"); err != nil { |
| 787 | t.Fatalf("Run: %v", err) |
| 788 | } |
| 789 | if got := len(exec.requests); got != 1 { |
| 790 | t.Fatalf("executor requests = %d, want one guidance-only final answer with no handoff nudge", got) |
| 791 | } |
| 792 | } |
| 793 | |
| 794 | func TestCoordinatorAllowsGuidanceOnlyPlanWithExecutorToolContext(t *testing.T) { |
| 795 | planner := &mockProvider{name: "planner", chunks: submitPlanChunk(`{"objective":"user guidance","steps":[{"title":"Tell the user to open the audio app, enable the checkbox, and listen to compare the difference."}]}`)} |
| 796 | exec := &mockProvider{name: "executor", streams: [][]provider.Chunk{ |
| 797 | { |
| 798 | {Type: provider.ChunkText, Text: "Open the app, enable the checkbox, then listen and compare."}, |
| 799 | {Type: provider.ChunkDone}, |
| 800 | }, |
| 801 | }} |
| 802 | |
| 803 | execReg := tool.NewRegistry() |
| 804 | execReg.Add(coordinatorTestTool{name: "read_file", readOnly: true, output: "file"}) |
| 805 | execReg.Add(coordinatorTestTool{name: "write_file", readOnly: false, output: "wrote file"}) |
| 806 | executor := New(exec, execReg, NewSession("exec-sys"), Options{}, event.Discard) |
| 807 | coord := NewCoordinator(planner, NewSession("planner-sys"), nil, plannerRegistryWithSubmitPlan(), Options{}, executor, 0, event.Discard, nil) |
| 808 | |
| 809 | if err := coord.Run(withNoClosedLoop(context.Background()), "Please advise on the manual audio check."); err != nil { |
| 810 | t.Fatalf("Run: %v", err) |
| 811 | } |
| 812 | if got := len(exec.requests); got != 1 { |
| 813 | t.Fatalf("executor requests = %d, want guidance final answer without nudge despite tool context", got) |
| 814 | } |
| 815 | } |
| 816 | |
| 817 | func TestCoordinatorRelaysSubmittedConclusionPlan(t *testing.T) { |
| 818 | planner := &mockProvider{name: "planner", chunks: submitPlanChunk(`{"objective":"report the finding","steps":[{"title":"the guard already exists in parser.go"}]}`)} |
| 819 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 820 | {Type: provider.ChunkText, Text: "The guard already exists; nothing to change."}, |
| 821 | {Type: provider.ChunkDone}, |
| 822 | }} |
| 823 | |
| 824 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 825 | coord := NewCoordinator(planner, NewSession("planner-sys"), nil, plannerRegistryWithSubmitPlan(), Options{}, executor, 0, event.Discard, nil) |
| 826 | |
| 827 | if err := coord.Run(withNoClosedLoop(context.Background()), "check whether the fix is already present"); err != nil { |
| 828 | t.Fatalf("Run: %v", err) |
| 829 | } |
| 830 | if got := len(exec.requests); got == 0 { |
| 831 | t.Fatal("executor never ran the conclusion-plan relay") |
| 832 | } |
| 833 | got := lastUser(exec.requests[0]) |
| 834 | if !strings.Contains(got, "the guard already exists in parser.go") || !strings.Contains(got, executorHandoffMarker) { |
| 835 | t.Fatalf("executor handoff = %q, want the submitted conclusion plan", got) |
| 836 | } |
| 837 | } |
| 838 | |
| 839 | func TestCoordinatorHandoffAffirmsExecutorToolSchemasWhenPlannerClaimsNoMCP(t *testing.T) { |
| 840 | planner := &mockProvider{name: "planner", chunks: submitPlanChunk(`{"objective":"search GitHub discussions","steps":[{"title":"I only have read-only tools; the executor must search GitHub discussions."}]}`)} |
| 841 | exec := &mockProvider{name: "executor", streams: [][]provider.Chunk{ |
| 842 | { |
| 843 | {Type: provider.ChunkText, Text: "GitHub MCP is unavailable."}, |
| 844 | {Type: provider.ChunkDone}, |
| 845 | }, |
| 846 | { |
| 847 | {Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "call-1", Name: "mcp__github__search", Arguments: `{"query":"Reasonix discussions"}`}}, |
| 848 | {Type: provider.ChunkDone}, |
| 849 | }, |
| 850 | { |
| 851 | {Type: provider.ChunkText, Text: "Done."}, |
| 852 | {Type: provider.ChunkDone}, |
| 853 | }, |
| 854 | }} |
| 855 | |
| 856 | execReg := tool.NewRegistry() |
| 857 | execReg.Add(coordinatorTestTool{name: "mcp__github__search", readOnly: true, output: "discussion results"}) |
| 858 | executor := New(exec, execReg, NewSession("exec-sys"), Options{}, event.Discard) |
| 859 | coord := NewCoordinator(planner, NewSession("planner-sys"), nil, plannerRegistryWithSubmitPlan(), Options{}, executor, 0, event.Discard, nil) |
| 860 | |
| 861 | if err := coord.Run(withNoClosedLoop(context.Background()), "search GitHub discussions"); err != nil { |
| 862 | t.Fatalf("Run: %v", err) |
| 863 | } |
| 864 | if got := len(exec.requests); got != 1 { |
| 865 | t.Fatalf("executor requests = %d, want one clean final with schemas attached and no handoff nudge", got) |
| 866 | } |
| 867 | if tools := toolSchemaNames(exec.requests[0].Tools); !contains(tools, "mcp__github__search") { |
| 868 | t.Fatalf("executor request tools = %v, want MCP schema attached", tools) |
| 869 | } |
| 870 | first := lastUser(exec.requests[0]) |
| 871 | for _, want := range []string{ |
| 872 | "The executor request includes the full tool schema", |
| 873 | "mcp__github__search", |
| 874 | "Do not treat planner tool limitations or tool-unavailable claims as executor facts", |
| 875 | } { |
| 876 | if !strings.Contains(first, want) { |
| 877 | t.Fatalf("initial executor handoff missing %q:\n%s", want, first) |
| 878 | } |
| 879 | } |
| 880 | } |
| 881 | |
| 882 | func TestCoordinatorDoesNotNudgeExecutorThatActs(t *testing.T) { |
| 883 | planner := &mockProvider{name: "planner", chunks: submitPlanChunk(`{"objective":"write the skill file","steps":[{"title":"write the requested skill file"}]}`)} |
| 884 | // Executor calls a tool on its first turn, then answers — no nudge expected. |
| 885 | exec := &mockProvider{name: "executor", streams: [][]provider.Chunk{ |
| 886 | { |
| 887 | {Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "call-1", Name: "write_file", Arguments: `{"path":"kan-tu.md"}`}}, |
| 888 | {Type: provider.ChunkDone}, |
| 889 | }, |
| 890 | { |
| 891 | {Type: provider.ChunkText, Text: "Done."}, |
| 892 | {Type: provider.ChunkDone}, |
| 893 | }, |
| 894 | }} |
| 895 | |
| 896 | execReg := tool.NewRegistry() |
| 897 | execReg.Add(coordinatorTestTool{name: "write_file", readOnly: false, output: "wrote file"}) |
| 898 | executor := New(exec, execReg, NewSession("exec-sys"), Options{}, event.Discard) |
| 899 | coord := NewCoordinator(planner, NewSession("planner-sys"), nil, plannerRegistryWithSubmitPlan(), Options{}, executor, 0, event.Discard, nil) |
| 900 | |
| 901 | if err := coord.Run(withNoClosedLoop(context.Background()), "install the skill"); err != nil { |
| 902 | t.Fatalf("Run: %v", err) |
| 903 | } |
| 904 | if got := len(exec.requests); got != 2 { |
| 905 | t.Fatalf("executor requests = %d, want tool call + final answer with no nudge", got) |
| 906 | } |
| 907 | for i, req := range exec.requests { |
| 908 | if strings.Contains(lastUser(req), "Use your available tools now to carry out the task") { |
| 909 | t.Fatalf("request %d unexpectedly received a handoff nudge", i) |
| 910 | } |
| 911 | } |
| 912 | } |
| 913 | |
| 914 | func toolSchemaNames(schemas []provider.ToolSchema) []string { |
| 915 | out := make([]string, 0, len(schemas)) |
| 916 | for _, s := range schemas { |
| 917 | out = append(out, s.Name) |
| 918 | } |
| 919 | return out |
| 920 | } |
| 921 | |
| 922 | func contains(items []string, want string) bool { |
| 923 | return slices.Contains(items, want) |
| 924 | } |
| 925 | |
| 926 | func BenchmarkPlannerToolRegistry(b *testing.B) { |
| 927 | parentReg := tool.NewRegistry() |
| 928 | for i := range 200 { |
| 929 | parentReg.Add(coordinatorTestTool{ |
| 930 | name: fmt.Sprintf("tool_%03d", i), |
| 931 | readOnly: i%3 != 0, |
| 932 | }) |
| 933 | } |
| 934 | parentReg.Add(coordinatorTestTool{name: "todo_write", readOnly: true}) |
| 935 | parentReg.Add(coordinatorTestTool{name: "write_file", readOnly: false}) |
| 936 | |
| 937 | b.ReportAllocs() |
| 938 | for range b.N { |
| 939 | reg := PlannerToolRegistry(parentReg) |
| 940 | if reg.Len() == 0 { |
| 941 | b.Fatal("planner registry should retain read-only research tools") |
| 942 | } |
| 943 | } |
| 944 | } |
| 945 | |
| 946 | func TestCoordinatorSetPlanModePropagates(t *testing.T) { |
| 947 | prov := &mockProvider{name: "planner", chunks: []provider.Chunk{ |
| 948 | {Type: provider.ChunkText, Text: "plan"}, |
| 949 | {Type: provider.ChunkDone}, |
| 950 | }} |
| 951 | plannerSess := NewSession("planner-sys") |
| 952 | plannerReg := tool.NewRegistry() |
| 953 | plannerReg.Add(coordinatorTestTool{name: "read_file", readOnly: true}) |
| 954 | plannerTools := PlannerToolRegistry(plannerReg) |
| 955 | |
| 956 | exec := New(nil, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 957 | |
| 958 | coord := NewCoordinator(prov, plannerSess, nil, plannerTools, Options{MaxSteps: 2}, exec, 0, event.Discard, nil) |
| 959 | |
| 960 | // Both should start with planMode=false |
| 961 | if coord.plannerAgent.planMode.Load() { |
| 962 | t.Error("planner should start with planMode=false") |
| 963 | } |
| 964 | if coord.executor.planMode.Load() { |
| 965 | t.Error("executor should start with planMode=false") |
| 966 | } |
| 967 | |
| 968 | // SetPlanMode(true) should propagate to both |
| 969 | coord.SetPlanMode(true) |
| 970 | if !coord.plannerAgent.planMode.Load() { |
| 971 | t.Error("planner should have planMode=true after SetPlanMode(true)") |
| 972 | } |
| 973 | if !coord.executor.planMode.Load() { |
| 974 | t.Error("executor should have planMode=true after SetPlanMode(true)") |
| 975 | } |
| 976 | |
| 977 | // SetPlanMode(false) should propagate to both |
| 978 | coord.SetPlanMode(false) |
| 979 | if coord.plannerAgent.planMode.Load() { |
| 980 | t.Error("planner should have planMode=false after SetPlanMode(false)") |
| 981 | } |
| 982 | if coord.executor.planMode.Load() { |
| 983 | t.Error("executor should have planMode=false after SetPlanMode(false)") |
| 984 | } |
| 985 | } |
| 986 | |
| 987 | func TestCoordinatorSetPlanModeNilSafety(t *testing.T) { |
| 988 | var c *Coordinator |
| 989 | c.SetPlanMode(true) // should not panic |
| 990 | c.SetPlanMode(false) // should not panic |
| 991 | } |
| 992 | |
| 993 | // errorProvider fails every Stream call, standing in for a down/misconfigured |
| 994 | // planner provider. |
| 995 | type errorProvider struct{ name string } |
| 996 | |
| 997 | func (e *errorProvider) Name() string { return e.name } |
| 998 | |
| 999 | func (e *errorProvider) Stream(context.Context, provider.Request) (<-chan provider.Chunk, error) { |
| 1000 | return nil, fmt.Errorf("provider unavailable") |
| 1001 | } |
| 1002 | |
| 1003 | // TestDefaultPlannerPromptRequiresSubmittedPlan keeps the prompt aligned with |
| 1004 | // the parse contract: submit_plan is the only delivery channel, and the retired |
| 1005 | // prose markers must not come back. |
| 1006 | func TestDefaultPlannerPromptRequiresSubmittedPlan(t *testing.T) { |
| 1007 | if !strings.Contains(DefaultPlannerPrompt, "submit_plan is the only delivery channel") { |
| 1008 | t.Fatal("DefaultPlannerPrompt must state that submit_plan is the only delivery channel") |
| 1009 | } |
| 1010 | for _, marker := range []string{"[no_changes]", "[planner_requires_approval]"} { |
| 1011 | if strings.Contains(DefaultPlannerPrompt, marker) { |
| 1012 | t.Fatalf("DefaultPlannerPrompt still teaches the retired %s marker", marker) |
| 1013 | } |
| 1014 | } |
| 1015 | } |
| 1016 | |
| 1017 | func TestDefaultPlannerPromptDefinesLightAndFullEvidenceContracts(t *testing.T) { |
| 1018 | for _, want := range []string{ |
| 1019 | "submit_plan", |
| 1020 | "command-level verification", |
| 1021 | "assumptions", |
| 1022 | } { |
| 1023 | // The verified/candidate split is asserted where it is enforced: the schema. |
| 1024 | if !strings.Contains(DefaultPlannerPrompt, want) { |
| 1025 | t.Fatalf("DefaultPlannerPrompt missing %q planning contract", want) |
| 1026 | } |
| 1027 | } |
| 1028 | } |
| 1029 | |
| 1030 | // TestCoordinatorDoesNotSkipExecutorForAlreadyImplementedPlanWithFollowUp is |
| 1031 | // the motivating regression: a plan acknowledging existing code while asking |
| 1032 | // for follow-up work must not be treated as a no-op conclusion. |
| 1033 | func TestCoordinatorDoesNotSkipExecutorForAlreadyImplementedPlanWithFollowUp(t *testing.T) { |
| 1034 | planner := &mockProvider{name: "planner", chunks: submitPlanChunk(`{"objective":"extend the auth flow","steps":[{"title":"The auth flow is already implemented; extend it to cover refresh tokens."}]}`)} |
| 1035 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 1036 | {Type: provider.ChunkText, Text: "Done."}, |
| 1037 | {Type: provider.ChunkDone}, |
| 1038 | }} |
| 1039 | |
| 1040 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 1041 | coord := NewCoordinator(planner, NewSession("planner-sys"), nil, plannerRegistryWithSubmitPlan(), Options{}, executor, 0, event.Discard, nil) |
| 1042 | |
| 1043 | if err := coord.Run(withNoClosedLoop(context.Background()), "add refresh token support"); err != nil { |
| 1044 | t.Fatalf("Run: %v", err) |
| 1045 | } |
| 1046 | if got := len(exec.requests); got == 0 { |
| 1047 | t.Fatal("executor skipped: an already-implemented plan with follow-up work was treated as no-op") |
| 1048 | } |
| 1049 | if got := lastUser(exec.requests[0]); !strings.Contains(got, "extend it to cover refresh tokens") { |
| 1050 | t.Fatalf("executor handoff missing the plan: %q", got) |
| 1051 | } |
| 1052 | } |
| 1053 | |
| 1054 | // TestCoordinatorFailsClosedWhenPlannerFails checks the no-degrade contract: a |
| 1055 | // planner failure fails the turn with a planner error, the executor never runs |
| 1056 | // without a plan, no fallback notice is emitted, and the planner session is |
| 1057 | // rolled back so the next plan does not start with consecutive user messages. |
| 1058 | func TestCoordinatorFailsClosedWhenPlannerFails(t *testing.T) { |
| 1059 | cases := []struct { |
| 1060 | name string |
| 1061 | planner provider.Provider |
| 1062 | }{ |
| 1063 | {"stream call fails", &errorProvider{name: "planner"}}, |
| 1064 | {"stream emits error chunk", &mockProvider{name: "planner", chunks: []provider.Chunk{ |
| 1065 | {Type: provider.ChunkError, Err: fmt.Errorf("rate limited")}, |
| 1066 | }}}, |
| 1067 | } |
| 1068 | for _, tc := range cases { |
| 1069 | t.Run(tc.name, func(t *testing.T) { |
| 1070 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 1071 | {Type: provider.ChunkText, Text: "Done."}, |
| 1072 | {Type: provider.ChunkDone}, |
| 1073 | }} |
| 1074 | var events []event.Event |
| 1075 | sink := event.FuncSink(func(e event.Event) { events = append(events, e) }) |
| 1076 | |
| 1077 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 1078 | plannerSess := NewSession("planner-sys") |
| 1079 | coord := NewCoordinator(tc.planner, plannerSess, nil, nil, Options{}, executor, 0, sink, nil) |
| 1080 | |
| 1081 | err := coord.Run(withNoClosedLoop(context.Background()), "fix the bug") |
| 1082 | if err == nil || !strings.Contains(err.Error(), "planner:") { |
| 1083 | t.Fatalf("Run = %v, want the propagated planner error", err) |
| 1084 | } |
| 1085 | if got := len(exec.requests); got != 0 { |
| 1086 | t.Fatalf("executor requests = %d, want none: a planner failure must not degrade to an unplanned run", got) |
| 1087 | } |
| 1088 | if n := len(plannerSess.Messages); n != 1 { |
| 1089 | t.Fatalf("planner session messages = %d, want rollback to system only", n) |
| 1090 | } |
| 1091 | for _, e := range events { |
| 1092 | if e.Kind == event.Notice && strings.Contains(e.Text, "continuing this turn with the executor") { |
| 1093 | t.Fatal("fallback notice emitted; planner failures must not degrade the turn") |
| 1094 | } |
| 1095 | } |
| 1096 | }) |
| 1097 | } |
| 1098 | } |
| 1099 | |
| 1100 | // TestCoordinatorPropagatesPlannerErrorWhenTurnCancelled keeps cancellation |
| 1101 | // semantics: a turn the user aborted must not silently restart on the executor. |
| 1102 | func TestCoordinatorPropagatesPlannerErrorWhenTurnCancelled(t *testing.T) { |
| 1103 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 1104 | {Type: provider.ChunkText, Text: "Should not run."}, |
| 1105 | {Type: provider.ChunkDone}, |
| 1106 | }} |
| 1107 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 1108 | coord := NewCoordinator(&errorProvider{name: "planner"}, NewSession("planner-sys"), nil, nil, Options{}, executor, 0, event.Discard, nil) |
| 1109 | |
| 1110 | ctx, cancel := context.WithCancel(context.Background()) |
| 1111 | cancel() |
| 1112 | err := coord.Run(ctx, "fix the bug") |
| 1113 | if err == nil || !strings.Contains(err.Error(), "planner:") { |
| 1114 | t.Fatalf("Run = %v, want propagated planner error on cancelled turn", err) |
| 1115 | } |
| 1116 | if got := len(exec.requests); got != 0 { |
| 1117 | t.Fatalf("executor requests = %d, want none after user cancellation", got) |
| 1118 | } |
| 1119 | } |
| 1120 | |
| 1121 | // TestCoordinatorRollsBackPlannerSessionOnToolPlannerFailure covers the |
| 1122 | // production two-model wiring (boot passes PlannerToolRegistry, so planning |
| 1123 | // runs through planWithTools): when the tool-enabled planner fails, the turn |
| 1124 | // fails and the rollback must not leave the planner session with a dangling |
| 1125 | // user message or partial tool rounds — the next plan would otherwise start |
| 1126 | // with consecutive user roles, which some providers reject. |
| 1127 | func TestCoordinatorRollsBackPlannerSessionOnToolPlannerFailure(t *testing.T) { |
| 1128 | cases := []struct { |
| 1129 | name string |
| 1130 | planner provider.Provider |
| 1131 | }{ |
| 1132 | {"stream call fails", &errorProvider{name: "planner"}}, |
| 1133 | {"fails after a tool round", &mockProvider{name: "planner", streams: [][]provider.Chunk{ |
| 1134 | { |
| 1135 | {Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "call-1", Name: "read_file", Arguments: `{"path":"main.go"}`}}, |
| 1136 | {Type: provider.ChunkDone}, |
| 1137 | }, |
| 1138 | { |
| 1139 | {Type: provider.ChunkError, Err: fmt.Errorf("rate limited")}, |
| 1140 | }, |
| 1141 | }}}, |
| 1142 | } |
| 1143 | for _, tc := range cases { |
| 1144 | t.Run(tc.name, func(t *testing.T) { |
| 1145 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 1146 | {Type: provider.ChunkText, Text: "Done."}, |
| 1147 | {Type: provider.ChunkDone}, |
| 1148 | }} |
| 1149 | plannerReg := tool.NewRegistry() |
| 1150 | plannerReg.Add(coordinatorTestTool{name: "read_file", readOnly: true, output: "package main"}) |
| 1151 | |
| 1152 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 1153 | plannerSess := NewSession("planner-sys") |
| 1154 | coord := NewCoordinator(tc.planner, plannerSess, nil, PlannerToolRegistry(plannerReg), Options{}, executor, 0, event.Discard, nil) |
| 1155 | |
| 1156 | err := coord.Run(withNoClosedLoop(context.Background()), "fix the bug") |
| 1157 | if err == nil || !strings.Contains(err.Error(), "planner:") { |
| 1158 | t.Fatalf("Run = %v, want the propagated planner error", err) |
| 1159 | } |
| 1160 | if got := len(exec.requests); got != 0 { |
| 1161 | t.Fatalf("executor requests = %d, want none after the planner failure", got) |
| 1162 | } |
| 1163 | if n := len(plannerSess.Messages); n != 1 { |
| 1164 | t.Fatalf("planner session messages = %d, want rollback to system only", n) |
| 1165 | } |
| 1166 | }) |
| 1167 | } |
| 1168 | } |
| 1169 | |
| 1170 | func TestCoordinatorPlannerSafetyBoundaryPreservesExecutionBoundaries(t *testing.T) { |
| 1171 | for _, route := range []PlannerRoute{PlannerRoutePlanOnly, PlannerRoutePlanForApproval} { |
| 1172 | t.Run(string(route), func(t *testing.T) { |
| 1173 | planner := &mockProvider{name: "planner", chunks: []provider.Chunk{ |
| 1174 | {Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "call-1", Name: "read_file", Arguments: `{"path":"main.go"}`}}, |
| 1175 | {Type: provider.ChunkDone}, |
| 1176 | }} |
| 1177 | exec := &mockProvider{name: "executor"} |
| 1178 | plannerReg := tool.NewRegistry() |
| 1179 | plannerReg.Add(coordinatorTestTool{name: "read_file", readOnly: true, output: "package main"}) |
| 1180 | policy := func(context.Context, string) PlannerDecision { |
| 1181 | return PlannerDecision{Route: route, Reason: "explicit_boundary"} |
| 1182 | } |
| 1183 | |
| 1184 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 1185 | plannerSess := NewSession("planner-sys") |
| 1186 | coord := NewCoordinatorWithPlannerPolicy( |
| 1187 | planner, plannerSess, nil, plannerReg, Options{MaxSteps: 1, MaxStepsKey: "planner emergency rounds"}, |
| 1188 | executor, 0, event.Discard, policy, |
| 1189 | ) |
| 1190 | |
| 1191 | err := coord.Run(withNoClosedLoop(context.Background()), "plan the migration") |
| 1192 | if err == nil || err.Error() != plannerSafetyBoundaryError { |
| 1193 | t.Fatalf("Run = %v, want the safe planner boundary error", err) |
| 1194 | } |
| 1195 | if got := len(exec.requests); got != 0 { |
| 1196 | t.Fatalf("executor requests = %d, want none across %s", got, route) |
| 1197 | } |
| 1198 | if got := len(plannerSess.Messages); got != 1 { |
| 1199 | t.Fatalf("planner session messages = %d, want the incomplete turn rolled back", got) |
| 1200 | } |
| 1201 | }) |
| 1202 | } |
| 1203 | } |
| 1204 | |
| 1205 | func TestCoordinatorRollbackAfterRewriteDropsPausedPlannerToolCall(t *testing.T) { |
| 1206 | plannerSess := NewSession("planner-sys") |
| 1207 | before := plannerSess.Snapshot() |
| 1208 | rewriteBefore := plannerSess.RewriteVersion() |
| 1209 | |
| 1210 | plannerSess.Replace([]provider.Message{ |
| 1211 | {Role: provider.RoleSystem, Content: "planner-sys"}, |
| 1212 | {Role: provider.RoleUser, Content: summaryTagOpen + "\ncompacted research\n</summary>"}, |
| 1213 | {Role: provider.RoleAssistant, Content: "Completed evidence from the bounded research rounds."}, |
| 1214 | }) |
| 1215 | plannerSess.IncrementRewrite() |
| 1216 | plannerSess.Add(provider.Message{Role: provider.RoleUser, Content: "Do not call any more tools; finalize."}) |
| 1217 | plannerSess.Add(provider.Message{ |
| 1218 | Role: provider.RoleAssistant, |
| 1219 | ToolCalls: []provider.ToolCall{{ |
| 1220 | ID: "ignored-finalization-call", Name: "read_file", Arguments: `{"path":"more.go"}`, |
| 1221 | }}, |
| 1222 | }) |
| 1223 | |
| 1224 | coord := &Coordinator{plannerSess: plannerSess} |
| 1225 | coord.rollbackPlannerTurn(before, rewriteBefore) |
| 1226 | |
| 1227 | msgs := plannerSess.Snapshot() |
| 1228 | if len(msgs) != 3 { |
| 1229 | t.Fatalf("planner session messages = %d, want compacted prefix plus completed evidence", len(msgs)) |
| 1230 | } |
| 1231 | if last := msgs[len(msgs)-1]; last.Role != provider.RoleAssistant || |
| 1232 | len(last.ToolCalls) != 0 || last.Content == "" { |
| 1233 | t.Fatalf("planner session has an unusable pause tail: %+v", last) |
| 1234 | } |
| 1235 | if normalized := provider.NormalizeMessages(msgs); len(normalized) != len(msgs) { |
| 1236 | t.Fatalf("planner session still needs tool-pair repair after rollback: %+v", normalized) |
| 1237 | } |
| 1238 | } |
| 1239 | |
| 1240 | // TestCoordinatorRunsExecutorWhenMarkerNotAlone is retired with the marker |
| 1241 | // contract; prose-plan routing tests above cover the submitted-plan path. |
| 1242 | |
| 1243 | // TestCoordinatorHandoffSurvivesPlannerCompaction pins plan delivery across |
| 1244 | // projection compaction. Projection compaction must not lose this turn's |
| 1245 | // submitted plan before the executor handoff is built. |
| 1246 | func TestCoordinatorHandoffSurvivesPlannerCompaction(t *testing.T) { |
| 1247 | planner := &mockProvider{name: "planner", streams: [][]provider.Chunk{ |
| 1248 | { // preflight compaction on the large filler history (estimate-based) |
| 1249 | {Type: provider.ChunkText, Text: "- goal: prior filler\n- pending: plan the fix"}, |
| 1250 | {Type: provider.ChunkDone}, |
| 1251 | }, |
| 1252 | // the plan turn after projection is in place: compaction usage on the |
| 1253 | // submit_plan round keeps the estimate-based preflight armed |
| 1254 | { |
| 1255 | {Type: provider.ChunkUsage, Usage: &provider.Usage{PromptTokens: 400, TotalTokens: 450}}, |
| 1256 | {Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "call-1", Name: "submit_plan", Arguments: `{"objective":"fix","steps":[{"title":"Edit main.go and add the missing guard."}]}`}}, |
| 1257 | {Type: provider.ChunkDone}, |
| 1258 | }, |
| 1259 | }} |
| 1260 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 1261 | {Type: provider.ChunkText, Text: "Done."}, |
| 1262 | {Type: provider.ChunkDone}, |
| 1263 | }} |
| 1264 | |
| 1265 | plannerReg := tool.NewRegistry() |
| 1266 | plannerReg.Add(coordinatorTestTool{name: "read_file", readOnly: true, output: "ok"}) |
| 1267 | |
| 1268 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 1269 | plannerSess := NewSession("planner-sys") |
| 1270 | // Preset enough planner history that context preflight compacts before the |
| 1271 | // plan stream. Canonical history stays intact; the handoff must still find |
| 1272 | // the plan on the canonical transcript. |
| 1273 | filler := strings.Repeat("planner history filler. ", 150) |
| 1274 | for range 3 { |
| 1275 | plannerSess.Add(provider.Message{Role: provider.RoleUser, Content: filler}) |
| 1276 | plannerSess.Add(provider.Message{Role: provider.RoleAssistant, Content: filler}) |
| 1277 | } |
| 1278 | coord := NewCoordinator(planner, plannerSess, nil, PlannerToolRegistry(plannerReg), Options{ContextWindow: 2000}, executor, 0, event.Discard, nil) |
| 1279 | |
| 1280 | if err := coord.Run(withNoClosedLoop(context.Background()), "fix the bug"); err != nil { |
| 1281 | t.Fatalf("Run: %v", err) |
| 1282 | } |
| 1283 | // Projection compaction no longer rewrites the planner session; handoff |
| 1284 | // must still deliver the plan even when RewriteVersion stays 0. |
| 1285 | if plannerSess.RewriteVersion() != 0 { |
| 1286 | t.Fatalf("canonical rewrite version = %d, want 0", plannerSess.RewriteVersion()) |
| 1287 | } |
| 1288 | if got := len(exec.requests); got == 0 { |
| 1289 | t.Fatal("executor never ran") |
| 1290 | } |
| 1291 | got := lastUser(exec.requests[0]) |
| 1292 | if !strings.Contains(got, "Edit main.go and add the missing guard.") || !strings.Contains(got, executorHandoffMarker) { |
| 1293 | t.Fatalf("executor input lost the plan handoff after planner compaction:\n%s", got) |
| 1294 | } |
| 1295 | } |
| 1296 | |
| 1297 | // TestCoordinatorNoOpConclusionAttributedToPlanner is retired with the no-op |
| 1298 | // relay path: submitted plans are emitted with planner attribution in |
| 1299 | // planWithTools, covered by TestSubmittedPlanIsRenderedToTheSink. |
| 1300 | |
| 1301 | // TestCoordinatorHandoffOmitsToolContextWithoutMCPTools checks that the handoff |
| 1302 | // does not restate the built-in tool schema: the tool-context block exists to |
| 1303 | // counter planner claims about MCP availability and is dropped entirely when |
| 1304 | // the executor carries no MCP tools. |
| 1305 | func TestCoordinatorHandoffOmitsToolContextWithoutMCPTools(t *testing.T) { |
| 1306 | planner := &mockProvider{name: "planner", chunks: submitPlanChunk(`{"objective":"fix the guard","steps":[{"title":"Edit main.go and add the missing guard."}]}`)} |
| 1307 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 1308 | {Type: provider.ChunkText, Text: "Done."}, |
| 1309 | {Type: provider.ChunkDone}, |
| 1310 | }} |
| 1311 | |
| 1312 | execReg := tool.NewRegistry() |
| 1313 | execReg.Add(coordinatorTestTool{name: "write_file", readOnly: false, output: "ok"}) |
| 1314 | executor := New(exec, execReg, NewSession("exec-sys"), Options{}, event.Discard) |
| 1315 | coord := NewCoordinator(planner, NewSession("planner-sys"), nil, plannerRegistryWithSubmitPlan(), Options{}, executor, 0, event.Discard, nil) |
| 1316 | |
| 1317 | if err := coord.Run(withNoClosedLoop(context.Background()), "fix the missing guard"); err != nil { |
| 1318 | t.Fatalf("Run: %v", err) |
| 1319 | } |
| 1320 | got := lastUser(exec.requests[0]) |
| 1321 | for _, unwanted := range []string{"Executor tool context", "Tool names include"} { |
| 1322 | if strings.Contains(got, unwanted) { |
| 1323 | t.Fatalf("handoff restates built-in tool schema (%q):\n%s", unwanted, got) |
| 1324 | } |
| 1325 | } |
| 1326 | if !strings.Contains(got, "Edit main.go") { |
| 1327 | t.Fatalf("handoff missing the plan: %q", got) |
| 1328 | } |
| 1329 | } |
| 1330 | |
| 1331 | // TestCoordinatorPassesTurnContextToPlannerGate pins the C2 contract: the gate |
| 1332 | // receives the live turn context, so a classifier-backed gate is cancelled |
| 1333 | // with the turn instead of running out its own timeout. |
| 1334 | func TestCoordinatorPassesTurnContextToPlannerGate(t *testing.T) { |
| 1335 | type gateCtxKey struct{} |
| 1336 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 1337 | {Type: provider.ChunkText, Text: "It does X."}, |
| 1338 | {Type: provider.ChunkDone}, |
| 1339 | }} |
| 1340 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 1341 | |
| 1342 | var sawTurnValue bool |
| 1343 | gate := func(ctx context.Context, _ string) bool { |
| 1344 | sawTurnValue = ctx.Value(gateCtxKey{}) != nil |
| 1345 | return false |
| 1346 | } |
| 1347 | coord := NewCoordinator(&mockProvider{name: "planner"}, NewSession("planner-sys"), nil, nil, Options{}, executor, 0, event.Discard, gate) |
| 1348 | |
| 1349 | ctx := context.WithValue(context.Background(), gateCtxKey{}, "turn") |
| 1350 | if err := coord.Run(ctx, "what does this do?"); err != nil { |
| 1351 | t.Fatalf("Run: %v", err) |
| 1352 | } |
| 1353 | if !sawTurnValue { |
| 1354 | t.Fatal("planner gate did not receive the turn context") |
| 1355 | } |
| 1356 | } |
| 1357 | |
| 1358 | // TestCoordinatorFailedTurnRollbackKeepsCompaction pins rollback economics |
| 1359 | // under projection compaction: when preflight/auto compaction fires and the |
| 1360 | // planner then fails, restoring the pre-turn snapshot must not erase the |
| 1361 | // projection or leave a dangling plain user turn that would produce |
| 1362 | // consecutive user roles on the next plan. |
| 1363 | func TestCoordinatorFailedTurnRollbackKeepsCompaction(t *testing.T) { |
| 1364 | planner := &mockProvider{name: "planner", streams: [][]provider.Chunk{ |
| 1365 | { // preflight compaction on large filler history |
| 1366 | {Type: provider.ChunkText, Text: "- goal: guard work\n- pending: continue"}, |
| 1367 | {Type: provider.ChunkDone}, |
| 1368 | }, |
| 1369 | { // tool round after projection is installed |
| 1370 | {Type: provider.ChunkToolCall, ToolCall: &provider.ToolCall{ID: "call-1", Name: "read_file", Arguments: `{"path":"main.go"}`}}, |
| 1371 | {Type: provider.ChunkUsage, Usage: &provider.Usage{PromptTokens: 400, TotalTokens: 450}}, |
| 1372 | {Type: provider.ChunkDone}, |
| 1373 | }, |
| 1374 | { // the next planner round fails |
| 1375 | {Type: provider.ChunkError, Err: fmt.Errorf("rate limited")}, |
| 1376 | }, |
| 1377 | }} |
| 1378 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 1379 | {Type: provider.ChunkText, Text: "Done."}, |
| 1380 | {Type: provider.ChunkDone}, |
| 1381 | }} |
| 1382 | |
| 1383 | plannerReg := tool.NewRegistry() |
| 1384 | plannerReg.Add(coordinatorTestTool{name: "read_file", readOnly: true, output: "package main"}) |
| 1385 | |
| 1386 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 1387 | plannerSess := NewSession("planner-sys") |
| 1388 | filler := strings.Repeat("planner history filler. ", 150) |
| 1389 | for range 3 { |
| 1390 | plannerSess.Add(provider.Message{Role: provider.RoleUser, Content: filler}) |
| 1391 | plannerSess.Add(provider.Message{Role: provider.RoleAssistant, Content: filler}) |
| 1392 | } |
| 1393 | coord := NewCoordinator(planner, plannerSess, nil, PlannerToolRegistry(plannerReg), Options{ContextWindow: 2000}, executor, 0, event.Discard, nil) |
| 1394 | |
| 1395 | err := coord.Run(withNoClosedLoop(context.Background()), "fix the bug") |
| 1396 | if err == nil || !strings.Contains(err.Error(), "planner:") { |
| 1397 | t.Fatalf("Run = %v, want the propagated planner error", err) |
| 1398 | } |
| 1399 | // The failed turn runs nothing: executor requests stay at zero. |
| 1400 | if got := len(exec.requests); got != 0 { |
| 1401 | t.Fatalf("executor requests = %d, want none after the planner failure", got) |
| 1402 | } |
| 1403 | // Canonical transcript is never rewrite-compacted. |
| 1404 | if plannerSess.RewriteVersion() != 0 { |
| 1405 | t.Fatalf("canonical rewrite version = %d, want 0", plannerSess.RewriteVersion()) |
| 1406 | } |
| 1407 | // Canonical history is restored/cleaned without a dangling user turn so the |
| 1408 | // next plan can continue. Projection lives on the planner agent and is not |
| 1409 | // wiped by snapshot rollback of Session.Messages alone. |
| 1410 | msgs := plannerSess.Snapshot() |
| 1411 | if last := msgs[len(msgs)-1]; last.Role == provider.RoleUser && !isCompactionSummary(last) { |
| 1412 | t.Fatalf("planner session ends in a plain user message after rollback: %q", last.Content) |
| 1413 | } |
| 1414 | } |
| 1415 | |
| 1416 | // TestCoordinatorPersistsDeniedPlanTurnToExecutorSession pins the denial |
| 1417 | // bookkeeping: a plan the user declines must still land in the executor |
| 1418 | // session (like the no-op path) so the turn survives save/reload, with a note |
| 1419 | // telling the next executor turn that nothing ran, plus a user-facing notice. |
| 1420 | func TestCoordinatorPersistsDeniedPlanTurnToExecutorSession(t *testing.T) { |
| 1421 | planner := &mockProvider{name: "planner", chunks: submitPlanChunk(`{"objective":"rewrite auth","requires_approval":true,"steps":[{"title":"rewrite auth"}]}`)} |
| 1422 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 1423 | {Type: provider.ChunkText, Text: "should not run"}, |
| 1424 | {Type: provider.ChunkDone}, |
| 1425 | }} |
| 1426 | sink := &recordSink{} |
| 1427 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, sink) |
| 1428 | coord := NewCoordinator(planner, NewSession("planner-sys"), nil, plannerRegistryWithSubmitPlan(), Options{}, executor, 0, sink, nil) |
| 1429 | gate := &coordinatorApprovalGate{allow: false} |
| 1430 | coord.SetPlannerPlanApprover(gate) |
| 1431 | |
| 1432 | if err := coord.Run(withNoClosedLoop(context.Background()), "rewrite auth"); err != nil { |
| 1433 | t.Fatalf("Run: %v", err) |
| 1434 | } |
| 1435 | if gate.calls != 1 { |
| 1436 | t.Fatalf("approval gate calls = %d, want 1", gate.calls) |
| 1437 | } |
| 1438 | if len(exec.requests) != 0 { |
| 1439 | t.Fatal("executor must not run when the plan is denied") |
| 1440 | } |
| 1441 | msgs := executor.sess.conversation.Messages |
| 1442 | if len(msgs) < 2 { |
| 1443 | t.Fatalf("executor session messages = %d, want the denied turn persisted", len(msgs)) |
| 1444 | } |
| 1445 | last := msgs[len(msgs)-1] |
| 1446 | if last.Role != provider.RoleAssistant || !strings.Contains(last.Content, plannerPlanNotApprovedNote) { |
| 1447 | t.Fatalf("last executor message = %q (%s), want plan with not-approved note", last.Content, last.Role) |
| 1448 | } |
| 1449 | prev := msgs[len(msgs)-2] |
| 1450 | if prev.Role != provider.RoleUser || !strings.Contains(prev.Content, "rewrite auth") { |
| 1451 | t.Fatalf("persisted user turn = %q (%s), want original input", prev.Content, prev.Role) |
| 1452 | } |
| 1453 | foundNotice := false |
| 1454 | for _, e := range sink.kinds(event.Notice) { |
| 1455 | if strings.Contains(e.Text, "not approved") { |
| 1456 | foundNotice = true |
| 1457 | } |
| 1458 | } |
| 1459 | if !foundNotice { |
| 1460 | t.Fatal("denied plan should emit a user-facing notice") |
| 1461 | } |
| 1462 | } |
| 1463 | |
| 1464 | // TestCoordinatorApprovalGateReadsOnlyTheField pins the negation side of the |
| 1465 | // field-only contract: a plan that rules out an approval round in prose (and in |
| 1466 | // its requires_approval field) hands off directly instead of raising a |
| 1467 | // needless approval prompt. |
| 1468 | func TestCoordinatorApprovalGateReadsOnlyTheField(t *testing.T) { |
| 1469 | planner := &mockProvider{name: "planner", chunks: submitPlanChunk(`{"objective":"修改 config","steps":[{"title":"修改 config.go，无需等待用户批准，直接执行修改"}]}`)} |
| 1470 | exec := &mockProvider{name: "executor", chunks: []provider.Chunk{ |
| 1471 | {Type: provider.ChunkText, Text: "Done."}, |
| 1472 | {Type: provider.ChunkDone}, |
| 1473 | }} |
| 1474 | executor := New(exec, tool.NewRegistry(), NewSession("exec-sys"), Options{}, event.Discard) |
| 1475 | coord := NewCoordinator(planner, NewSession("planner-sys"), nil, plannerRegistryWithSubmitPlan(), Options{}, executor, 0, event.Discard, nil) |
| 1476 | gate := &coordinatorApprovalGate{allow: false} |
| 1477 | coord.SetPlannerPlanApprover(gate) |
| 1478 | |
| 1479 | if err := coord.Run(withNoClosedLoop(context.Background()), "tweak config"); err != nil { |
| 1480 | t.Fatalf("Run: %v", err) |
| 1481 | } |
| 1482 | if gate.calls != 0 { |
| 1483 | t.Fatalf("approval gate calls = %d, want 0 for requires_approval=false", gate.calls) |
| 1484 | } |
| 1485 | if len(exec.requests) == 0 { |
| 1486 | t.Fatal("executor should run directly for a plan that does not require approval") |
| 1487 | } |
| 1488 | } |
| 1489 |