| 1 | package acp |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "encoding/json" |
| 6 | "fmt" |
| 7 | "path/filepath" |
| 8 | "strings" |
| 9 | "sync" |
| 10 | "testing" |
| 11 | "time" |
| 12 | |
| 13 | "reasonix/internal/agent" |
| 14 | "reasonix/internal/control" |
| 15 | "reasonix/internal/event" |
| 16 | "reasonix/internal/permission" |
| 17 | "reasonix/internal/provider" |
| 18 | ) |
| 19 | |
| 20 | // scriptedRequester answers agent → client requests from a per-method script |
| 21 | // and records the call order. |
| 22 | type scriptedRequester struct { |
| 23 | mu sync.Mutex |
| 24 | calls []string |
| 25 | params map[string]json.RawMessage |
| 26 | results map[string]any |
| 27 | errs map[string]error |
| 28 | } |
| 29 | |
| 30 | func newScriptedRequester() *scriptedRequester { |
| 31 | return &scriptedRequester{ |
| 32 | params: map[string]json.RawMessage{}, |
| 33 | results: map[string]any{}, |
| 34 | errs: map[string]error{}, |
| 35 | } |
| 36 | } |
| 37 | |
| 38 | func (r *scriptedRequester) Request(_ context.Context, method string, params any) (json.RawMessage, error) { |
| 39 | raw, _ := json.Marshal(params) |
| 40 | r.mu.Lock() |
| 41 | r.calls = append(r.calls, method) |
| 42 | r.params[method] = raw |
| 43 | res, err := r.results[method], r.errs[method] |
| 44 | r.mu.Unlock() |
| 45 | if err != nil { |
| 46 | return nil, err |
| 47 | } |
| 48 | out, _ := json.Marshal(res) |
| 49 | return out, nil |
| 50 | } |
| 51 | |
| 52 | func (r *scriptedRequester) callOrder() []string { |
| 53 | r.mu.Lock() |
| 54 | defer r.mu.Unlock() |
| 55 | return append([]string(nil), r.calls...) |
| 56 | } |
| 57 | |
| 58 | func TestClientIOReadWriteTextFile(t *testing.T) { |
| 59 | req := newScriptedRequester() |
| 60 | req.results["fs/read_text_file"] = FSReadTextFileResult{Content: "buffer content"} |
| 61 | req.results["fs/write_text_file"] = struct{}{} |
| 62 | io := newClientIO(req, "sess-1", ClientCapabilities{FS: FSCapabilities{ReadTextFile: true, WriteTextFile: true}}) |
| 63 | |
| 64 | content, ok := io.ReadTextFile(context.Background(), "/proj/a.go") |
| 65 | if !ok || content != "buffer content" { |
| 66 | t.Fatalf("ReadTextFile = %q, %v; want buffer content, true", content, ok) |
| 67 | } |
| 68 | var readParams FSReadTextFileParams |
| 69 | json.Unmarshal(req.params["fs/read_text_file"], &readParams) |
| 70 | if readParams.SessionID != "sess-1" || readParams.Path != "/proj/a.go" { |
| 71 | t.Fatalf("fs/read_text_file params = %+v", readParams) |
| 72 | } |
| 73 | |
| 74 | handled, err := io.WriteTextFile(context.Background(), "/proj/a.go", "new content") |
| 75 | if !handled || err != nil { |
| 76 | t.Fatalf("WriteTextFile = %v, %v; want true, nil", handled, err) |
| 77 | } |
| 78 | |
| 79 | // Without the capability, both degrade to unhandled so tools use the disk. |
| 80 | none := newClientIO(req, "sess-1", ClientCapabilities{}) |
| 81 | if _, ok := none.ReadTextFile(context.Background(), "/proj/a.go"); ok { |
| 82 | t.Fatal("ReadTextFile without capability must report ok=false") |
| 83 | } |
| 84 | if handled, _ := none.WriteTextFile(context.Background(), "/proj/a.go", "x"); handled { |
| 85 | t.Fatal("WriteTextFile without capability must report handled=false") |
| 86 | } |
| 87 | |
| 88 | // A client read error falls back (ok=false); a client write error surfaces |
| 89 | // (falling back could double-apply). |
| 90 | req.errs["fs/read_text_file"] = fmt.Errorf("not open") |
| 91 | if _, ok := io.ReadTextFile(context.Background(), "/proj/a.go"); ok { |
| 92 | t.Fatal("ReadTextFile client error must report ok=false") |
| 93 | } |
| 94 | req.errs["fs/write_text_file"] = fmt.Errorf("readonly buffer") |
| 95 | handled, err = io.WriteTextFile(context.Background(), "/proj/a.go", "x") |
| 96 | if !handled || err == nil { |
| 97 | t.Fatalf("WriteTextFile client error = %v, %v; want handled=true with error", handled, err) |
| 98 | } |
| 99 | } |
| 100 | |
| 101 | func TestClientIORunCommandLifecycle(t *testing.T) { |
| 102 | req := newScriptedRequester() |
| 103 | req.results["terminal/create"] = TerminalCreateResult{TerminalID: "term-1"} |
| 104 | req.results["terminal/wait_for_exit"] = TerminalWaitResult{} |
| 105 | exitZero := 0 |
| 106 | req.results["terminal/output"] = TerminalOutputResult{Output: "hello from client", ExitStatus: &TerminalExitStatus{ExitCode: &exitZero}} |
| 107 | req.results["terminal/release"] = struct{}{} |
| 108 | io := newClientIO(req, "sess-1", ClientCapabilities{Terminal: true}) |
| 109 | |
| 110 | out, ok, err := io.RunCommand(context.Background(), "echo hello", "/proj", time.Minute, nil) |
| 111 | if !ok || err != nil || out != "hello from client" { |
| 112 | t.Fatalf("RunCommand = %q, %v, %v", out, ok, err) |
| 113 | } |
| 114 | order := io2str(req.callOrder()) |
| 115 | if order != "terminal/create,terminal/wait_for_exit,terminal/output,terminal/release" { |
| 116 | t.Fatalf("call order = %s", order) |
| 117 | } |
| 118 | |
| 119 | // A nonzero exit surfaces as an error alongside the captured output. |
| 120 | exitOne := 1 |
| 121 | req.results["terminal/output"] = TerminalOutputResult{Output: "boom", ExitStatus: &TerminalExitStatus{ExitCode: &exitOne}} |
| 122 | out, ok, err = io.RunCommand(context.Background(), "false", "/proj", time.Minute, nil) |
| 123 | if !ok || err == nil || !strings.Contains(err.Error(), "exit status 1") || out != "boom" { |
| 124 | t.Fatalf("RunCommand nonzero exit = %q, %v, %v", out, ok, err) |
| 125 | } |
| 126 | |
| 127 | // No terminal capability → unhandled, local bash runs instead. |
| 128 | none := newClientIO(req, "sess-1", ClientCapabilities{}) |
| 129 | if _, ok, _ := none.RunCommand(context.Background(), "echo", "/proj", time.Minute, nil); ok { |
| 130 | t.Fatal("RunCommand without capability must report ok=false") |
| 131 | } |
| 132 | |
| 133 | // terminal/create failure degrades to unhandled rather than failing the call. |
| 134 | req.errs["terminal/create"] = fmt.Errorf("client rejected") |
| 135 | if _, ok, _ := io.RunCommand(context.Background(), "echo", "/proj", time.Minute, nil); ok { |
| 136 | t.Fatal("RunCommand with failed create must report ok=false") |
| 137 | } |
| 138 | } |
| 139 | |
| 140 | func io2str(calls []string) string { return strings.Join(calls, ",") } |
| 141 | |
| 142 | func TestUpdateSinkToolLocations(t *testing.T) { |
| 143 | s := newUpdateSink(&fakeNotifier{}, "sess") |
| 144 | cwd := t.TempDir() |
| 145 | s.bindCwd(cwd) |
| 146 | |
| 147 | locs := s.toolLocations("read_file", `{"path":"pkg/a.go","offset":41}`) |
| 148 | if len(locs) != 1 || locs[0].Path != filepath.Join(cwd, "pkg", "a.go") { |
| 149 | t.Fatalf("read_file locations = %+v", locs) |
| 150 | } |
| 151 | if locs[0].Line == nil || *locs[0].Line != 42 { |
| 152 | t.Fatalf("read_file line = %v, want 42 (offset is 0-based)", locs[0].Line) |
| 153 | } |
| 154 | // A platform-absolute path passes through untouched. |
| 155 | abs := filepath.Join(t.TempDir(), "b.go") |
| 156 | absArgs, _ := json.Marshal(map[string]string{"path": abs}) |
| 157 | if locs := s.toolLocations("edit_file", string(absArgs)); len(locs) != 1 || locs[0].Path != abs || locs[0].Line != nil { |
| 158 | t.Fatalf("edit_file locations = %+v, want %s", locs, abs) |
| 159 | } |
| 160 | if locs := s.toolLocations("bash", `{"command":"ls"}`); locs != nil { |
| 161 | t.Fatalf("bash should have no locations, got %+v", locs) |
| 162 | } |
| 163 | if locs := s.toolLocations("grep", `{"path":"pkg","pattern":"x"}`); locs != nil { |
| 164 | t.Fatalf("grep (directory scope) should have no locations, got %+v", locs) |
| 165 | } |
| 166 | if locs := s.toolLocations("read_file", `{"offset":1}`); locs != nil { |
| 167 | t.Fatalf("path-less args should have no locations, got %+v", locs) |
| 168 | } |
| 169 | } |
| 170 | |
| 171 | func TestPlanEntriesFromTodoArgs(t *testing.T) { |
| 172 | entries, ok := planEntriesFromTodoArgs(`{"todos":[ |
| 173 | {"content":"Phase one","status":"in_progress","level":0}, |
| 174 | {"content":"Sub step","status":"pending","level":1}, |
| 175 | {"content":"Done step","status":"completed"}, |
| 176 | {"content":"","status":"pending"}, |
| 177 | {"content":"Weird","status":"???"} |
| 178 | ]}`) |
| 179 | if !ok || len(entries) != 4 { |
| 180 | t.Fatalf("entries = %+v, ok=%v; want 4 entries", entries, ok) |
| 181 | } |
| 182 | if entries[0].Priority != "high" || entries[0].Status != "in_progress" { |
| 183 | t.Fatalf("phase entry = %+v", entries[0]) |
| 184 | } |
| 185 | if entries[1].Priority != "medium" { |
| 186 | t.Fatalf("sub-step entry = %+v", entries[1]) |
| 187 | } |
| 188 | if entries[3].Status != "pending" { |
| 189 | t.Fatalf("unknown status must degrade to pending, got %+v", entries[3]) |
| 190 | } |
| 191 | if _, ok := planEntriesFromTodoArgs(`{"todos":[]}`); ok { |
| 192 | t.Fatal("empty todos must not produce a plan update") |
| 193 | } |
| 194 | if _, ok := planEntriesFromTodoArgs(`not json`); ok { |
| 195 | t.Fatal("malformed args must not produce a plan update") |
| 196 | } |
| 197 | } |
| 198 | |
| 199 | func TestUpdateSinkEmitsPlanForTodoWrite(t *testing.T) { |
| 200 | n := &fakeNotifier{} |
| 201 | s := newUpdateSink(n, "sess-1") |
| 202 | s.Emit(event.Event{Kind: event.ToolDispatch, Tool: event.Tool{ |
| 203 | ID: "t1", Name: "todo_write", Args: `{"todos":[{"content":"Do it","status":"pending"}]}`, |
| 204 | }}) |
| 205 | // The plan update precedes the tool_call for the same dispatch. |
| 206 | plan := n.updateMap(t, 0) |
| 207 | if plan["sessionUpdate"] != "plan" { |
| 208 | t.Fatalf("first update = %v, want plan", plan["sessionUpdate"]) |
| 209 | } |
| 210 | raw, _ := json.Marshal(plan) |
| 211 | if !strings.Contains(string(raw), `"content":"Do it"`) { |
| 212 | t.Fatalf("plan update missing entry: %s", raw) |
| 213 | } |
| 214 | call := n.updateMap(t, 1) |
| 215 | if call["sessionUpdate"] != "tool_call" { |
| 216 | t.Fatalf("second update = %v, want tool_call", call["sessionUpdate"]) |
| 217 | } |
| 218 | } |
| 219 | |
| 220 | // recordingFactory wraps e2eFactory and captures the SessionParams the service |
| 221 | // hands to NewSession, so tests can assert the client-capability wiring. |
| 222 | type recordingFactory struct { |
| 223 | inner *e2eFactory |
| 224 | mu sync.Mutex |
| 225 | params []SessionParams |
| 226 | } |
| 227 | |
| 228 | func (f *recordingFactory) SessionDir() string { return f.inner.SessionDir() } |
| 229 | |
| 230 | func (f *recordingFactory) NewSession(ctx context.Context, p SessionParams) (*control.Controller, error) { |
| 231 | f.mu.Lock() |
| 232 | f.params = append(f.params, p) |
| 233 | f.mu.Unlock() |
| 234 | return f.inner.NewSession(ctx, p) |
| 235 | } |
| 236 | |
| 237 | func (f *recordingFactory) last() SessionParams { |
| 238 | f.mu.Lock() |
| 239 | defer f.mu.Unlock() |
| 240 | return f.params[len(f.params)-1] |
| 241 | } |
| 242 | |
| 243 | func TestSessionParamsCarryClientIOFromInitializeCaps(t *testing.T) { |
| 244 | dir := t.TempDir() |
| 245 | prov := &scriptedProvider{name: "fake", responses: [][]provider.Chunk{ |
| 246 | {{Type: provider.ChunkText, Text: "done"}, {Type: provider.ChunkDone}}, |
| 247 | }} |
| 248 | factory := &recordingFactory{inner: &e2eFactory{ |
| 249 | prov: prov, |
| 250 | tool: fakeTool{name: "peek", ro: true, out: "ok"}, |
| 251 | policy: permission.New("ask", nil, nil, nil), |
| 252 | sessionDir: dir, |
| 253 | }} |
| 254 | client, stop := startServer(t, factory) |
| 255 | defer stop() |
| 256 | |
| 257 | client.call(t, "initialize", InitializeParams{ |
| 258 | ProtocolVersion: 1, |
| 259 | ClientCapabilities: ClientCapabilities{ |
| 260 | FS: FSCapabilities{ReadTextFile: true, WriteTextFile: true}, |
| 261 | Terminal: true, |
| 262 | }, |
| 263 | }) |
| 264 | client.call(t, "session/new", SessionNewParams{Cwd: t.TempDir()}) |
| 265 | p := factory.last() |
| 266 | if p.FileOverlay == nil { |
| 267 | t.Fatal("SessionParams.FileOverlay should be bound when the client declares fs capabilities") |
| 268 | } |
| 269 | if p.Terminal == nil { |
| 270 | t.Fatal("SessionParams.Terminal should be bound when the client declares the terminal capability") |
| 271 | } |
| 272 | } |
| 273 | |
| 274 | func TestSessionParamsNilClientIOWithoutCaps(t *testing.T) { |
| 275 | dir := t.TempDir() |
| 276 | prov := &scriptedProvider{name: "fake", responses: [][]provider.Chunk{ |
| 277 | {{Type: provider.ChunkText, Text: "done"}, {Type: provider.ChunkDone}}, |
| 278 | }} |
| 279 | factory := &recordingFactory{inner: &e2eFactory{ |
| 280 | prov: prov, |
| 281 | tool: fakeTool{name: "peek", ro: true, out: "ok"}, |
| 282 | policy: permission.New("ask", nil, nil, nil), |
| 283 | sessionDir: dir, |
| 284 | }} |
| 285 | client, stop := startServer(t, factory) |
| 286 | defer stop() |
| 287 | |
| 288 | client.call(t, "initialize", InitializeParams{ProtocolVersion: 1}) |
| 289 | client.call(t, "session/new", SessionNewParams{Cwd: t.TempDir()}) |
| 290 | p := factory.last() |
| 291 | if p.FileOverlay != nil || p.Terminal != nil { |
| 292 | t.Fatalf("SessionParams overlay/terminal must stay nil without client capabilities; got %v / %v", p.FileOverlay, p.Terminal) |
| 293 | } |
| 294 | } |
| 295 | |
| 296 | func TestE2ESessionModes(t *testing.T) { |
| 297 | dir := t.TempDir() |
| 298 | prov := &scriptedProvider{name: "fake", responses: [][]provider.Chunk{ |
| 299 | {{Type: provider.ChunkText, Text: "done"}, {Type: provider.ChunkDone}}, |
| 300 | }} |
| 301 | factory := &e2eFactory{ |
| 302 | prov: prov, |
| 303 | tool: fakeTool{name: "peek", ro: true, out: "ok"}, |
| 304 | policy: permission.New("ask", nil, nil, nil), |
| 305 | sessionDir: dir, |
| 306 | } |
| 307 | client, stop := startServer(t, factory) |
| 308 | defer stop() |
| 309 | |
| 310 | client.call(t, "initialize", InitializeParams{ProtocolVersion: 1}) |
| 311 | cwd := t.TempDir() |
| 312 | resp := client.call(t, "session/new", SessionNewParams{Cwd: cwd}) |
| 313 | var nr SessionNewResult |
| 314 | if err := json.Unmarshal(resp.Result, &nr); err != nil { |
| 315 | t.Fatalf("session/new result: %v", err) |
| 316 | } |
| 317 | if nr.Modes == nil || nr.Modes.CurrentModeID != sessionModeNormal || len(nr.Modes.AvailableModes) != 3 { |
| 318 | t.Fatalf("session/new modes = %+v, want normal current with 3 available", nr.Modes) |
| 319 | } |
| 320 | |
| 321 | setResp := client.call(t, "session/set_mode", SessionSetModeParams{SessionID: nr.SessionID, ModeID: sessionModePlan}) |
| 322 | if setResp.Error != nil { |
| 323 | t.Fatalf("session/set_mode: %+v", setResp.Error) |
| 324 | } |
| 325 | // The switch is confirmed with a current_mode_update notification. |
| 326 | deadline := time.After(5 * time.Second) |
| 327 | for { |
| 328 | select { |
| 329 | case n := <-client.notifs: |
| 330 | var p struct { |
| 331 | Update struct { |
| 332 | SessionUpdate string `json:"sessionUpdate"` |
| 333 | CurrentModeID string `json:"currentModeId"` |
| 334 | } `json:"update"` |
| 335 | } |
| 336 | if json.Unmarshal(n.Params, &p) == nil && p.Update.SessionUpdate == "current_mode_update" { |
| 337 | if p.Update.CurrentModeID != sessionModePlan { |
| 338 | t.Fatalf("current_mode_update = %q, want plan", p.Update.CurrentModeID) |
| 339 | } |
| 340 | goto unknownMode |
| 341 | } |
| 342 | case <-deadline: |
| 343 | t.Fatal("no current_mode_update notification after session/set_mode") |
| 344 | } |
| 345 | } |
| 346 | unknownMode: |
| 347 | bad := client.call(t, "session/set_mode", SessionSetModeParams{SessionID: nr.SessionID, ModeID: "yolo"}) |
| 348 | if bad.Error == nil { |
| 349 | t.Fatal("unknown modeId must be rejected") |
| 350 | } |
| 351 | |
| 352 | // Reconnecting to the live session must report its actual mode, not a |
| 353 | // hardcoded default — the mode picker would otherwise go stale. |
| 354 | loadResp := client.call(t, "session/load", SessionLoadParams{SessionID: nr.SessionID, Cwd: cwd}) |
| 355 | if loadResp.Error != nil { |
| 356 | t.Fatalf("session/load: %+v", loadResp.Error) |
| 357 | } |
| 358 | var lr SessionLoadResult |
| 359 | if err := json.Unmarshal(loadResp.Result, &lr); err != nil { |
| 360 | t.Fatalf("session/load result: %v", err) |
| 361 | } |
| 362 | if lr.Modes == nil || lr.Modes.CurrentModeID != sessionModePlan { |
| 363 | t.Fatalf("session/load modes = %+v, want current plan for the live session", lr.Modes) |
| 364 | } |
| 365 | } |
| 366 | |
| 367 | // TestRebuildSessionKeepsClientIOAndMode pins two rebuild invariants: a |
| 368 | // model/effort switch must rebuild the controller with the same client |
| 369 | // capability wiring (fs overlay, host terminal) the original had, and must |
| 370 | // re-apply the session's ACP mode — a fresh controller boots with normal |
| 371 | // switches, which would silently drop a user-selected plan mode. |
| 372 | func TestRebuildSessionKeepsClientIOAndMode(t *testing.T) { |
| 373 | dir := t.TempDir() |
| 374 | path := filepath.Join(dir, "sess-rebuild.jsonl") |
| 375 | base := agent.NewSession("sys prompt") |
| 376 | base.Add(provider.Message{Role: provider.RoleUser, Content: "hi"}) |
| 377 | if err := base.Save(path); err != nil { |
| 378 | t.Fatalf("save session: %v", err) |
| 379 | } |
| 380 | |
| 381 | sink := newUpdateSink(&fakeNotifier{}, "sess-rebuild") |
| 382 | sess := &acpSession{ |
| 383 | id: "sess-rebuild", |
| 384 | sink: sink, |
| 385 | cwd: dir, |
| 386 | model: "fast", |
| 387 | transcript: path, |
| 388 | modeID: sessionModePlan, |
| 389 | } |
| 390 | lease, err := agent.TryAcquireSessionLease(path) |
| 391 | if err != nil { |
| 392 | t.Fatalf("acquire session lease: %v", err) |
| 393 | } |
| 394 | sess.lease = lease |
| 395 | t.Cleanup(sess.releaseSessionLease) |
| 396 | |
| 397 | factory := &configurableFactory{dir: dir} |
| 398 | svc := &service{ |
| 399 | factory: factory, |
| 400 | sessions: map[string]*acpSession{sess.id: sess}, |
| 401 | clientCaps: ClientCapabilities{ |
| 402 | FS: FSCapabilities{ReadTextFile: true, WriteTextFile: true}, |
| 403 | Terminal: true, |
| 404 | }, |
| 405 | } |
| 406 | oldCtrl := control.New(control.Options{ |
| 407 | Executor: agent.New(nil, nil, base, agent.Options{}, event.Discard), |
| 408 | SessionDir: dir, |
| 409 | SessionPath: path, |
| 410 | Label: "fast", |
| 411 | }) |
| 412 | sess.ctrl = oldCtrl |
| 413 | |
| 414 | if err := svc.rebuildSession(context.Background(), sess, SessionConfigState{Model: "pro"}, []sessionConfigDelta{{axis: "model", model: "pro"}}); err != nil { |
| 415 | t.Fatalf("rebuildSession: %v", err) |
| 416 | } |
| 417 | if sess.ctrl == oldCtrl { |
| 418 | t.Fatal("session controller was not replaced") |
| 419 | } |
| 420 | |
| 421 | factory.mu.Lock() |
| 422 | last := factory.builds[len(factory.builds)-1] |
| 423 | factory.mu.Unlock() |
| 424 | if last.FileOverlay == nil { |
| 425 | t.Fatal("rebuild must keep the fs overlay wiring") |
| 426 | } |
| 427 | if last.Terminal == nil { |
| 428 | t.Fatal("rebuild must keep the host terminal wiring") |
| 429 | } |
| 430 | if !sess.ctrl.PlanMode() { |
| 431 | t.Fatal("rebuild must re-apply the session's plan mode to the new controller") |
| 432 | } |
| 433 | } |
| 434 |