| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "encoding/json" |
| 6 | "errors" |
| 7 | "strings" |
| 8 | "testing" |
| 9 | |
| 10 | "reasonix/internal/agent/testutil" |
| 11 | "reasonix/internal/event" |
| 12 | "reasonix/internal/provider" |
| 13 | "reasonix/internal/tool" |
| 14 | ) |
| 15 | |
| 16 | // steerThenCancelTool queues a steer while the turn is running, then cancels |
| 17 | // the turn so Run exits before the loop's per-iteration consume can deliver it. |
| 18 | type steerThenCancelTool struct { |
| 19 | agent *Agent |
| 20 | cancel context.CancelFunc |
| 21 | steerText string |
| 22 | accepted bool |
| 23 | } |
| 24 | |
| 25 | func (t *steerThenCancelTool) Name() string { return "steer_then_cancel" } |
| 26 | func (t *steerThenCancelTool) Description() string { return "queues a steer and cancels the turn" } |
| 27 | func (t *steerThenCancelTool) Schema() json.RawMessage { |
| 28 | return json.RawMessage(`{"type":"object","properties":{}}`) |
| 29 | } |
| 30 | func (t *steerThenCancelTool) ReadOnly() bool { return true } |
| 31 | func (t *steerThenCancelTool) Execute(context.Context, json.RawMessage) (string, error) { |
| 32 | t.accepted = t.agent.Steer(t.steerText) |
| 33 | t.cancel() |
| 34 | return "ok", nil |
| 35 | } |
| 36 | |
| 37 | // TestRunFlushesUnconsumedSteersOnCancel proves a steer that is still queued |
| 38 | // when the turn is cancelled survives in local history but not the next model |
| 39 | // context, and emits an explicit warning instead of presenting it as |
| 40 | // successfully applied guidance. |
| 41 | func TestRunFlushesUnconsumedSteersOnCancel(t *testing.T) { |
| 42 | mp := testutil.NewMock("m", |
| 43 | testutil.Turn{ToolCalls: []provider.ToolCall{{ID: "call-1", Name: "steer_then_cancel", Arguments: `{}`}}}, |
| 44 | testutil.Turn{Text: "never reached"}, |
| 45 | ) |
| 46 | hijack := &steerThenCancelTool{steerText: "use plan B"} |
| 47 | reg := tool.NewRegistry() |
| 48 | reg.Add(hijack) |
| 49 | var notices []event.Event |
| 50 | sink := event.FuncSink(func(e event.Event) { |
| 51 | if e.Kind == event.Notice && e.Code == event.NoticeCodeUnappliedSteer { |
| 52 | notices = append(notices, e) |
| 53 | } |
| 54 | }) |
| 55 | a := New(mp, reg, NewSession(""), Options{}, sink) |
| 56 | ctx, cancel := context.WithCancel(context.Background()) |
| 57 | defer cancel() |
| 58 | hijack.agent = a |
| 59 | hijack.cancel = cancel |
| 60 | |
| 61 | err := a.Run(ctx, "go") |
| 62 | if !errors.Is(err, context.Canceled) { |
| 63 | t.Fatalf("Run should exit on the cancelled context, got %v", err) |
| 64 | } |
| 65 | if !hijack.accepted { |
| 66 | t.Fatalf("Steer during an active turn should be accepted") |
| 67 | } |
| 68 | |
| 69 | var persisted []string |
| 70 | var localOnly bool |
| 71 | var persistedID string |
| 72 | for _, m := range a.Session().Messages { |
| 73 | if text, ok := SteerText(m.Content); ok { |
| 74 | persisted = append(persisted, text) |
| 75 | persistedID = m.ID |
| 76 | localOnly = m.LocalOnly && m.Role == provider.RoleTool && |
| 77 | m.ToolCallID == provider.LocalOnlyToolID && m.Name == provider.LocalOnlyToolName |
| 78 | } |
| 79 | } |
| 80 | if len(persisted) != 1 || persisted[0] != "use plan B" { |
| 81 | t.Fatalf("unconsumed steer should be persisted once and round-trip through SteerText, got %v", persisted) |
| 82 | } |
| 83 | if !localOnly { |
| 84 | t.Fatal("unconsumed steer must use the provider-excluded local-only sentinel") |
| 85 | } |
| 86 | for _, m := range provider.ModelMessages(a.Session().Snapshot()) { |
| 87 | if text, ok := SteerText(m.Content); ok { |
| 88 | t.Fatalf("unconsumed steer %q leaked into the next model context", text) |
| 89 | } |
| 90 | } |
| 91 | if len(notices) != 1 || notices[0].Level != event.LevelWarn || |
| 92 | !strings.Contains(notices[0].Text, "use plan B") || |
| 93 | !strings.Contains(notices[0].Text, "not applied") { |
| 94 | t.Fatalf("flushed steer should emit an explicit warning, got %+v", notices) |
| 95 | } |
| 96 | if persistedID == "" || notices[0].MessageID != persistedID { |
| 97 | t.Fatalf("warning and durable steer have different identities: message=%q notice=%q", persistedID, notices[0].MessageID) |
| 98 | } |
| 99 | if n := a.steerQueueLen(); n != 0 { |
| 100 | t.Fatalf("steer queue should be empty after the turn, len=%d", n) |
| 101 | } |
| 102 | if !a.HasUnappliedSteer() { |
| 103 | t.Fatal("host should observe that the cancelled turn left unapplied guidance") |
| 104 | } |
| 105 | if a.Steer("after the turn") { |
| 106 | t.Fatalf("Steer must be rejected once the turn has exited") |
| 107 | } |
| 108 | } |
| 109 | |
| 110 | // TestCloseSteerIntakeIfIdleMakesAdmissionLinearizable pins the normal turn |
| 111 | // exit boundary: once the final queue check observes no pending guidance, a |
| 112 | // later steer must be rejected rather than accepted and flushed as unapplied. |
| 113 | func TestCloseSteerIntakeIfIdleMakesAdmissionLinearizable(t *testing.T) { |
| 114 | a := New(nil, tool.NewRegistry(), NewSession(""), Options{}, event.Discard) |
| 115 | a.steerMu.Lock() |
| 116 | a.steerRunActive = true |
| 117 | a.steerMu.Unlock() |
| 118 | |
| 119 | if !a.closeSteerIntakeIfIdle() { |
| 120 | t.Fatal("empty steer intake should close") |
| 121 | } |
| 122 | if a.Steer("too late") { |
| 123 | t.Fatal("steer after the final queue check must be rejected") |
| 124 | } |
| 125 | if n := a.steerQueueLen(); n != 0 { |
| 126 | t.Fatalf("rejected steer remained queued, len=%d", n) |
| 127 | } |
| 128 | if a.HasUnappliedSteer() { |
| 129 | t.Fatal("closing an empty steer intake must not report unapplied guidance") |
| 130 | } |
| 131 | } |
| 132 | |
| 133 | func TestWithdrawnDurableSteerDoesNotEmitUnappliedNotice(t *testing.T) { |
| 134 | var notices int |
| 135 | a := New(nil, tool.NewRegistry(), NewSession(""), Options{}, event.FuncSink(func(e event.Event) { |
| 136 | if e.Kind == event.Notice && e.Code == event.NoticeCodeUnappliedSteer { |
| 137 | notices++ |
| 138 | } |
| 139 | })) |
| 140 | a.steerMu.Lock() |
| 141 | a.steerRunActive = true |
| 142 | a.steerMu.Unlock() |
| 143 | if !a.SteerItem("withdrawn-consume", func() (string, error) { return "", ErrSteerWithdrawn }) { |
| 144 | t.Fatal("active steer should be accepted") |
| 145 | } |
| 146 | if text, itemID, ok := a.consumeSteer(); ok || text != "" || itemID != "" { |
| 147 | t.Fatalf("withdrawn consume = (%q, %q, %v), want silent miss", text, itemID, ok) |
| 148 | } |
| 149 | if !a.SteerItem("withdrawn-flush", func() (string, error) { return "", ErrSteerWithdrawn }) { |
| 150 | t.Fatal("second active steer should be accepted") |
| 151 | } |
| 152 | a.flushSteerQueue() |
| 153 | if notices != 0 { |
| 154 | t.Fatalf("withdrawn steer emitted %d unapplied notices", notices) |
| 155 | } |
| 156 | if len(a.Session().Messages) != 0 { |
| 157 | t.Fatalf("withdrawn steer wrote transcript messages: %+v", a.Session().Messages) |
| 158 | } |
| 159 | } |
| 160 | |
| 161 | // TestSteerTextSurvivesTurnPreferenceWrapping pins replay: steers are |
| 162 | // persisted through withTurnPreferences, which prepends transient language |
| 163 | // blocks (for Chinese text even in auto mode, and for any text under an |
| 164 | // explicit language) ahead of the steer prefix. SteerText must skip the |
| 165 | // wrapping and return the user's exact original text, or replay degrades the |
| 166 | // steer into a plain user message. |
| 167 | func TestSteerTextSurvivesTurnPreferenceWrapping(t *testing.T) { |
| 168 | plain := New(nil, nil, NewSession(""), Options{}, event.Discard) |
| 169 | explicit := New(nil, nil, NewSession(""), Options{}, event.Discard) |
| 170 | explicit.SetReasoningLanguage("zh") |
| 171 | explicit.SetResponseLanguage("zh") |
| 172 | |
| 173 | cases := []struct { |
| 174 | name string |
| 175 | agent *Agent |
| 176 | text string |
| 177 | }{ |
| 178 | {"english auto (no blocks)", plain, "use plan B"}, |
| 179 | {"chinese auto (reasoning block)", plain, "请改用方案B"}, |
| 180 | {"explicit zh (both blocks)", explicit, "switch to plan B"}, |
| 181 | {"exact text preserved", plain, " spaced\ttext "}, |
| 182 | } |
| 183 | for _, tc := range cases { |
| 184 | persisted := tc.agent.withTurnPreferences(midTurnSteerMessage(tc.text)) |
| 185 | got, ok := SteerText(persisted) |
| 186 | if !ok { |
| 187 | t.Fatalf("%s: SteerText failed to recognize the persisted steer (head %.80q)", tc.name, persisted) |
| 188 | } |
| 189 | if got != tc.text { |
| 190 | t.Fatalf("%s: SteerText = %q, want %q", tc.name, got, tc.text) |
| 191 | } |
| 192 | } |
| 193 | |
| 194 | if _, ok := SteerText(plain.withTurnPreferences("请总结一下这个文件")); ok { |
| 195 | t.Fatalf("a wrapped ordinary user message must not be detected as a steer") |
| 196 | } |
| 197 | } |
| 198 | |
| 199 | // TestSteerRejectedWithoutActiveTurn proves a steer arriving when no turn is |
| 200 | // running is rejected instead of parked in a queue no loop will consume, so |
| 201 | // the controller can convert it into a regular turn. |
| 202 | func TestSteerRejectedWithoutActiveTurn(t *testing.T) { |
| 203 | a := New(testutil.NewMock("m", testutil.Turn{Text: "done"}), tool.NewRegistry(), NewSession(""), Options{}, event.Discard) |
| 204 | if a.Steer("early") { |
| 205 | t.Fatalf("Steer with no active turn must be rejected") |
| 206 | } |
| 207 | if n := a.steerQueueLen(); n != 0 { |
| 208 | t.Fatalf("rejected steer must not linger in the queue, len=%d", n) |
| 209 | } |
| 210 | if err := a.Run(context.Background(), "go"); err != nil { |
| 211 | t.Fatalf("Run: %v", err) |
| 212 | } |
| 213 | if a.Steer("between turns") { |
| 214 | t.Fatalf("Steer between turns must be rejected") |
| 215 | } |
| 216 | } |
| 217 |