| 1 | package cli |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "os" |
| 6 | "path/filepath" |
| 7 | "strings" |
| 8 | "testing" |
| 9 | |
| 10 | "reasonix/internal/acp" |
| 11 | "reasonix/internal/config" |
| 12 | "reasonix/internal/event" |
| 13 | "reasonix/internal/netclient" |
| 14 | "reasonix/internal/provider" |
| 15 | "reasonix/internal/tool" |
| 16 | |
| 17 | _ "reasonix/internal/tool/builtin" |
| 18 | ) |
| 19 | |
| 20 | const acpTestProviderKind = "acp-test-provider" |
| 21 | |
| 22 | func init() { |
| 23 | provider.Register(acpTestProviderKind, func(cfg provider.Config) (provider.Provider, error) { |
| 24 | return &acpTestProvider{cfg: cfg}, nil |
| 25 | }) |
| 26 | } |
| 27 | |
| 28 | func TestACPBuiltinToolsKeepSessionLevelBuiltins(t *testing.T) { |
| 29 | dir := t.TempDir() |
| 30 | tools := toolMap(acpBuiltinTools(&config.Config{}, dir, []string{dir})) |
| 31 | for _, name := range []string{ |
| 32 | "todo_write", |
| 33 | "complete_step", |
| 34 | "bash_output", |
| 35 | "kill_shell", |
| 36 | "wait", |
| 37 | "move_file", |
| 38 | "notebook_edit", |
| 39 | } { |
| 40 | if tools[name] == nil { |
| 41 | t.Fatalf("ACP workspace tools missing %q; got %v", name, toolNames(tools)) |
| 42 | } |
| 43 | } |
| 44 | } |
| 45 | |
| 46 | func TestACPInitializesWithoutAPIKey(t *testing.T) { |
| 47 | isolateCLIConfigHome(t) |
| 48 | t.Setenv("DEEPSEEK_API_KEY", "") |
| 49 | oldStdin := os.Stdin |
| 50 | r, w, err := os.Pipe() |
| 51 | if err != nil { |
| 52 | t.Fatal(err) |
| 53 | } |
| 54 | _, _ = w.WriteString(`{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":1}}` + "\n") |
| 55 | _ = w.Close() |
| 56 | os.Stdin = r |
| 57 | t.Cleanup(func() { |
| 58 | os.Stdin = oldStdin |
| 59 | _ = r.Close() |
| 60 | }) |
| 61 | |
| 62 | out := captureStdout(t, func() { |
| 63 | if rc := Run([]string{"--acp"}, "test-version"); rc != 0 { |
| 64 | t.Fatalf("Run --acp initialize rc = %d, want 0", rc) |
| 65 | } |
| 66 | }) |
| 67 | if !strings.Contains(out, `"protocolVersion":1`) || !strings.Contains(out, `"name":"reasonix"`) { |
| 68 | t.Fatalf("initialize output = %s", out) |
| 69 | } |
| 70 | } |
| 71 | |
| 72 | func TestACPRejectsInvalidSupervisorFlags(t *testing.T) { |
| 73 | for _, args := range [][]string{ |
| 74 | {"--planner=maybe"}, |
| 75 | {"--sandbox-network=maybe"}, |
| 76 | {"--sandbox-bash=maybe"}, |
| 77 | } { |
| 78 | if rc := acpCommand(args, "test-version"); rc != 2 { |
| 79 | t.Fatalf("acpCommand(%v) rc = %d, want 2", args, rc) |
| 80 | } |
| 81 | } |
| 82 | } |
| 83 | |
| 84 | func TestACPSupervisorRuntimeStateUsesHardOverrides(t *testing.T) { |
| 85 | isolateCLIConfigHome(t) |
| 86 | project := t.TempDir() |
| 87 | if err := os.WriteFile(filepath.Join(project, "reasonix.toml"), []byte(` |
| 88 | [agent] |
| 89 | planner_model = "configured-planner" |
| 90 | |
| 91 | [sandbox] |
| 92 | network = true |
| 93 | bash = "off" |
| 94 | |
| 95 | allow_write = ["../outside"] |
| 96 | `), 0o644); err != nil { |
| 97 | t.Fatal(err) |
| 98 | } |
| 99 | off := false |
| 100 | factory := &acpFactory{ |
| 101 | plannerOff: true, networkOverride: &off, bashOverride: "enforce", workspaceOnly: true, |
| 102 | } |
| 103 | state, err := factory.SessionRuntimeState(context.Background(), acp.SessionRuntimeStateParams{ |
| 104 | Cwd: project, Model: "configured-planner", RuntimeProfile: "balanced", |
| 105 | }) |
| 106 | if err != nil { |
| 107 | t.Fatalf("SessionRuntimeState: %v", err) |
| 108 | } |
| 109 | if state.PlannerMode != "off" || state.Sandbox.Mode != "enforce" || state.Sandbox.NetworkEnabled { |
| 110 | t.Fatalf("runtime overrides = %+v", state) |
| 111 | } |
| 112 | if len(state.Sandbox.WriteRoots) != 1 || state.Sandbox.WriteRoots[0] != project || state.Sandbox.WorkspaceRoot != project { |
| 113 | t.Fatalf("workspace confinement = %+v", state.Sandbox) |
| 114 | } |
| 115 | } |
| 116 | |
| 117 | func TestACPSupervisorRuntimeStateDegradesWhenSandboxIsUnavailable(t *testing.T) { |
| 118 | isolateCLIConfigHome(t) |
| 119 | project := t.TempDir() |
| 120 | if err := os.WriteFile(filepath.Join(project, "reasonix.toml"), []byte("[sandbox]\nbash = \"enforce\"\n"), 0o644); err != nil { |
| 121 | t.Fatal(err) |
| 122 | } |
| 123 | unavailable := func() bool { return false } |
| 124 | params := acp.SessionRuntimeStateParams{Cwd: project, RuntimeProfile: "balanced"} |
| 125 | |
| 126 | state, err := (&acpFactory{ |
| 127 | bashOverride: "enforce", sandboxAvailable: unavailable, |
| 128 | }).SessionRuntimeState(context.Background(), params) |
| 129 | if err != nil { |
| 130 | t.Fatalf("default ACP startup must report unavailable sandbox instead of failing: %v", err) |
| 131 | } |
| 132 | if state.Sandbox.Mode != "enforce" || state.Sandbox.Available { |
| 133 | t.Fatalf("degraded sandbox state = %+v", state.Sandbox) |
| 134 | } |
| 135 | |
| 136 | _, err = (&acpFactory{ |
| 137 | bashOverride: "enforce", requireSandbox: true, sandboxAvailable: unavailable, |
| 138 | }).SessionRuntimeState(context.Background(), params) |
| 139 | if err == nil || !strings.Contains(err.Error(), "sandbox unavailable") { |
| 140 | t.Fatalf("explicit enforce must fail closed, got %v", err) |
| 141 | } |
| 142 | } |
| 143 | |
| 144 | func TestEffectiveACPPlannerModeMatchesSelectedRuntime(t *testing.T) { |
| 145 | cfg := config.Default() |
| 146 | cfg.Agent.PlannerModel = "planner/planner-model" |
| 147 | cfg.Providers = []config.ProviderEntry{ |
| 148 | {Name: "executor", Kind: "openai", Model: "executor-model"}, |
| 149 | {Name: "planner", Kind: "openai", Model: "planner-model"}, |
| 150 | } |
| 151 | if got := effectiveACPPlannerMode(cfg, false, "executor/executor-model", "balanced"); got != "on" { |
| 152 | t.Fatalf("balanced split-model planner mode = %q, want on", got) |
| 153 | } |
| 154 | if got := effectiveACPPlannerMode(cfg, false, "executor/executor-model", "economy"); got != "off" { |
| 155 | t.Fatalf("economy planner mode = %q, want off", got) |
| 156 | } |
| 157 | if got := effectiveACPPlannerMode(cfg, false, "planner/planner-model", "balanced"); got != "off" { |
| 158 | t.Fatalf("same-model planner mode = %q, want off", got) |
| 159 | } |
| 160 | } |
| 161 | |
| 162 | func TestACPFactoryLoadsSessionCwdProjectConfig(t *testing.T) { |
| 163 | home := isolateCLIConfigHome(t) |
| 164 | if _, err := config.SetCredential("REASONIX_TEST_KEY", "test-key"); err != nil { |
| 165 | t.Fatalf("SetCredential: %v", err) |
| 166 | } |
| 167 | project := t.TempDir() |
| 168 | if err := os.WriteFile(filepath.Join(project, "reasonix.toml"), []byte(` |
| 169 | default_model = "local" |
| 170 | |
| 171 | [[providers]] |
| 172 | name = "local" |
| 173 | kind = "acp-test-provider" |
| 174 | base_url = "http://example.invalid" |
| 175 | model = "fake-model" |
| 176 | api_key_env = "REASONIX_TEST_KEY" |
| 177 | `), 0o644); err != nil { |
| 178 | t.Fatal(err) |
| 179 | } |
| 180 | cmdDir := filepath.Join(project, ".reasonix", "commands") |
| 181 | if err := os.MkdirAll(cmdDir, 0o755); err != nil { |
| 182 | t.Fatal(err) |
| 183 | } |
| 184 | if err := os.WriteFile(filepath.Join(cmdDir, "acp-only.md"), []byte("ACP project command"), 0o644); err != nil { |
| 185 | t.Fatal(err) |
| 186 | } |
| 187 | if err := os.Chdir(home); err != nil { |
| 188 | t.Fatal(err) |
| 189 | } |
| 190 | |
| 191 | ctrl, err := (&acpFactory{}).NewSession(context.Background(), acp.SessionParams{Cwd: project, Sink: event.Discard}) |
| 192 | if err != nil { |
| 193 | t.Fatalf("NewSession: %v", err) |
| 194 | } |
| 195 | defer ctrl.Close() |
| 196 | |
| 197 | for _, cmd := range ctrl.Commands() { |
| 198 | if cmd.Name == "acp-only" { |
| 199 | return |
| 200 | } |
| 201 | } |
| 202 | t.Fatalf("ACP session did not load project command from cwd; commands=%v", ctrl.Commands()) |
| 203 | } |
| 204 | |
| 205 | func TestACPFactoryClearsEffortOverrideForUnsupportedModel(t *testing.T) { |
| 206 | isolateCLIConfigHome(t) |
| 207 | if _, err := config.SetCredential("REASONIX_TEST_KEY", "test-key"); err != nil { |
| 208 | t.Fatalf("SetCredential: %v", err) |
| 209 | } |
| 210 | project := t.TempDir() |
| 211 | if err := os.WriteFile(filepath.Join(project, "reasonix.toml"), []byte(` |
| 212 | default_model = "reasoner/reasoning-model" |
| 213 | |
| 214 | [[providers]] |
| 215 | name = "reasoner" |
| 216 | kind = "acp-test-provider" |
| 217 | base_url = "http://example.invalid" |
| 218 | model = "reasoning-model" |
| 219 | api_key_env = "REASONIX_TEST_KEY" |
| 220 | supported_efforts = ["low", "high"] |
| 221 | |
| 222 | [[providers]] |
| 223 | name = "plain" |
| 224 | kind = "acp-test-provider" |
| 225 | base_url = "http://example.invalid" |
| 226 | model = "plain-model" |
| 227 | api_key_env = "REASONIX_TEST_KEY" |
| 228 | effort = "high" |
| 229 | `), 0o644); err != nil { |
| 230 | t.Fatal(err) |
| 231 | } |
| 232 | |
| 233 | high := "high" |
| 234 | state, err := (&acpFactory{}).SessionConfigState(context.Background(), acp.SessionConfigStateParams{ |
| 235 | Cwd: project, |
| 236 | Model: "reasoner/reasoning-model", |
| 237 | EffortOverride: &high, |
| 238 | }) |
| 239 | if err != nil { |
| 240 | t.Fatalf("reasoning SessionConfigState: %v", err) |
| 241 | } |
| 242 | if state.EffortOverride == nil || *state.EffortOverride != "high" { |
| 243 | t.Fatalf("reasoning effort override = %v, want high", state.EffortOverride) |
| 244 | } |
| 245 | |
| 246 | state, err = (&acpFactory{}).SessionConfigState(context.Background(), acp.SessionConfigStateParams{ |
| 247 | Cwd: project, |
| 248 | Model: "plain/plain-model", |
| 249 | EffortOverride: &high, |
| 250 | }) |
| 251 | if err != nil { |
| 252 | t.Fatalf("plain SessionConfigState: %v", err) |
| 253 | } |
| 254 | if _, ok := findACPConfigOption(state.ConfigOptions, "effort"); ok { |
| 255 | t.Fatalf("plain model should not advertise effort option: %+v", state.ConfigOptions) |
| 256 | } |
| 257 | if state.EffortOverride == nil || *state.EffortOverride != "" { |
| 258 | t.Fatalf("plain effort override = %v, want explicit empty override", state.EffortOverride) |
| 259 | } |
| 260 | } |
| 261 | |
| 262 | func TestACPFactoryAdvertisesAndNormalizesRuntimeProfiles(t *testing.T) { |
| 263 | isolateCLIConfigHome(t) |
| 264 | if _, err := config.SetCredential("REASONIX_TEST_KEY", "test-key"); err != nil { |
| 265 | t.Fatalf("SetCredential: %v", err) |
| 266 | } |
| 267 | project := t.TempDir() |
| 268 | if err := os.WriteFile(filepath.Join(project, "reasonix.toml"), []byte(` |
| 269 | default_model = "local" |
| 270 | |
| 271 | [[providers]] |
| 272 | name = "local" |
| 273 | kind = "acp-test-provider" |
| 274 | base_url = "http://example.invalid" |
| 275 | model = "fake-model" |
| 276 | api_key_env = "REASONIX_TEST_KEY" |
| 277 | `), 0o644); err != nil { |
| 278 | t.Fatal(err) |
| 279 | } |
| 280 | |
| 281 | state, err := (&acpFactory{profile: "full"}).SessionConfigState(context.Background(), acp.SessionConfigStateParams{ |
| 282 | Cwd: project, |
| 283 | RuntimeProfile: "delivery", |
| 284 | }) |
| 285 | if err != nil { |
| 286 | t.Fatalf("SessionConfigState: %v", err) |
| 287 | } |
| 288 | work, ok := findACPConfigOption(state.ConfigOptions, "work_mode") |
| 289 | if !ok || work.CurrentValue != "delivery" || state.RuntimeProfile != "delivery" || len(work.Options) != 3 { |
| 290 | t.Fatalf("work mode state = %+v / %q, want delivery with 3 options", work, state.RuntimeProfile) |
| 291 | } |
| 292 | |
| 293 | state, err = (&acpFactory{profile: "full"}).SessionConfigState(context.Background(), acp.SessionConfigStateParams{Cwd: project}) |
| 294 | if err != nil { |
| 295 | t.Fatalf("default SessionConfigState: %v", err) |
| 296 | } |
| 297 | work, _ = findACPConfigOption(state.ConfigOptions, "work_mode") |
| 298 | if work.CurrentValue != "balanced" || state.RuntimeProfile != "balanced" { |
| 299 | t.Fatalf("legacy full profile = %+v / %q, want balanced", work, state.RuntimeProfile) |
| 300 | } |
| 301 | } |
| 302 | |
| 303 | func TestACPTaskProfileDefaults(t *testing.T) { |
| 304 | cfg := config.Default() |
| 305 | cfg.Agent.SubagentModel = "default-model" |
| 306 | cfg.Agent.SubagentEffort = "high" |
| 307 | cfg.Agent.SubagentModels = map[string]string{"task": "task-model"} |
| 308 | cfg.Agent.SubagentEfforts = map[string]string{"task": "max"} |
| 309 | |
| 310 | model, effort := acpTaskProfileDefaults(cfg) |
| 311 | if model != "task-model" || effort != "max" { |
| 312 | t.Fatalf("task profile defaults = %q/%q, want task-model/max", model, effort) |
| 313 | } |
| 314 | |
| 315 | cfg.Agent.SubagentModels = nil |
| 316 | cfg.Agent.SubagentEfforts = nil |
| 317 | model, effort = acpTaskProfileDefaults(cfg) |
| 318 | if model != "default-model" || effort != "high" { |
| 319 | t.Fatalf("fallback task profile defaults = %q/%q, want default-model/high", model, effort) |
| 320 | } |
| 321 | } |
| 322 | |
| 323 | func TestACPSubagentProviderResolverHonorsProfile(t *testing.T) { |
| 324 | cfg := config.Default() |
| 325 | cfg.Providers = []config.ProviderEntry{ |
| 326 | { |
| 327 | Name: "parent", |
| 328 | Kind: acpTestProviderKind, |
| 329 | Model: "parent-model", |
| 330 | ContextWindow: 111, |
| 331 | SupportedEfforts: []string{"low", "high"}, |
| 332 | }, |
| 333 | { |
| 334 | Name: "sub", |
| 335 | Kind: acpTestProviderKind, |
| 336 | Models: []string{"sub-model"}, |
| 337 | Default: "sub-model", |
| 338 | ContextWindow: 222, |
| 339 | SupportedEfforts: []string{"low", "high"}, |
| 340 | }, |
| 341 | } |
| 342 | parent, ok := cfg.ResolveModel("parent") |
| 343 | if !ok { |
| 344 | t.Fatal("parent model did not resolve") |
| 345 | } |
| 346 | |
| 347 | resolve := newACPSubagentProviderResolver(cfg, parent, netclient.ProxySpec{}) |
| 348 | prov, _, ctxWin, err := resolve("sub/sub-model", "HIGH") |
| 349 | if err != nil { |
| 350 | t.Fatalf("resolve sub profile: %v", err) |
| 351 | } |
| 352 | got := prov.(*acpTestProvider).cfg |
| 353 | if got.Model != "sub-model" || got.Extra["effort"] != "high" || ctxWin != 222 { |
| 354 | t.Fatalf("resolved profile = model:%q effort:%v ctx:%d, want sub-model/high/222", got.Model, got.Extra["effort"], ctxWin) |
| 355 | } |
| 356 | |
| 357 | prov, _, ctxWin, err = resolve("", "low") |
| 358 | if err != nil { |
| 359 | t.Fatalf("resolve effort-only profile: %v", err) |
| 360 | } |
| 361 | got = prov.(*acpTestProvider).cfg |
| 362 | if got.Model != "parent-model" || got.Extra["effort"] != "low" || ctxWin != 111 { |
| 363 | t.Fatalf("effort-only profile = model:%q effort:%v ctx:%d, want parent-model/low/111", got.Model, got.Extra["effort"], ctxWin) |
| 364 | } |
| 365 | } |
| 366 | |
| 367 | func TestACPSubagentProviderResolverRejectsInvalidEffort(t *testing.T) { |
| 368 | cfg := config.Default() |
| 369 | cfg.Providers = []config.ProviderEntry{{ |
| 370 | Name: "parent", |
| 371 | Kind: acpTestProviderKind, |
| 372 | Model: "parent-model", |
| 373 | SupportedEfforts: []string{"low", "high"}, |
| 374 | }} |
| 375 | parent, ok := cfg.ResolveModel("parent") |
| 376 | if !ok { |
| 377 | t.Fatal("parent model did not resolve") |
| 378 | } |
| 379 | |
| 380 | resolve := newACPSubagentProviderResolver(cfg, parent, netclient.ProxySpec{}) |
| 381 | if _, _, _, err := resolve("", "max"); err == nil { |
| 382 | t.Fatal("invalid effort should fail before ACP task falls back to the parent profile") |
| 383 | } |
| 384 | } |
| 385 | |
| 386 | func findACPConfigOption(options []acp.SessionConfigOption, id string) (acp.SessionConfigOption, bool) { |
| 387 | for _, opt := range options { |
| 388 | if opt.ID == id { |
| 389 | return opt, true |
| 390 | } |
| 391 | } |
| 392 | return acp.SessionConfigOption{}, false |
| 393 | } |
| 394 | |
| 395 | func toolMap(tools []tool.Tool) map[string]tool.Tool { |
| 396 | out := make(map[string]tool.Tool, len(tools)) |
| 397 | for _, t := range tools { |
| 398 | out[t.Name()] = t |
| 399 | } |
| 400 | return out |
| 401 | } |
| 402 | |
| 403 | func toolNames(tools map[string]tool.Tool) []string { |
| 404 | out := make([]string, 0, len(tools)) |
| 405 | for name := range tools { |
| 406 | out = append(out, name) |
| 407 | } |
| 408 | return out |
| 409 | } |
| 410 | |
| 411 | type acpTestProvider struct { |
| 412 | cfg provider.Config |
| 413 | } |
| 414 | |
| 415 | func (p *acpTestProvider) Name() string { return p.cfg.Name } |
| 416 | |
| 417 | func (p *acpTestProvider) Stream(context.Context, provider.Request) (<-chan provider.Chunk, error) { |
| 418 | ch := make(chan provider.Chunk, 1) |
| 419 | ch <- provider.Chunk{Type: provider.ChunkDone} |
| 420 | close(ch) |
| 421 | return ch, nil |
| 422 | } |
| 423 |