| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "crypto/sha256" |
| 6 | "fmt" |
| 7 | "strings" |
| 8 | "testing" |
| 9 | |
| 10 | "reasonix/internal/event" |
| 11 | "reasonix/internal/provider" |
| 12 | ) |
| 13 | |
| 14 | type recordingAsker struct { |
| 15 | questions []event.AskQuestion |
| 16 | calls int |
| 17 | } |
| 18 | |
| 19 | func (r *recordingAsker) Ask(_ context.Context, questions []event.AskQuestion) ([]event.AskAnswer, error) { |
| 20 | r.calls++ |
| 21 | r.questions = questions |
| 22 | return []event.AskAnswer{{QuestionID: "q1", Selected: []string{"Keep going"}}}, nil |
| 23 | } |
| 24 | |
| 25 | func TestAskToolReusesAcceptedDecisionAndRejectsSpoofedID(t *testing.T) { |
| 26 | turn := &turnRuntime{} |
| 27 | asker := &recordingAsker{} |
| 28 | ctx := withTurnState(withCallContext(context.Background(), "call", event.Discard, asker, false), turn) |
| 29 | args := []byte(`{"decision_id":"dec-original","questions":[{"header":"Direction","question":"Which path?","options":[{"label":"Keep going"},{"label":"Stop"}]}]}`) |
| 30 | if _, err := NewAskTool().Execute(ctx, args); err != nil { |
| 31 | t.Fatalf("first ask: %v", err) |
| 32 | } |
| 33 | if asker.calls != 1 { |
| 34 | t.Fatalf("asker calls = %d", asker.calls) |
| 35 | } |
| 36 | second, err := NewAskTool().Execute(ctx, []byte(`{"decision_id":"dec-original","questions":[{"header":"Direction","question":"Which path?","options":[{"label":"Keep going"},{"label":"Stop"}]}]}`)) |
| 37 | if err != nil || !strings.Contains(second, "dec-original") || asker.calls != 1 { |
| 38 | t.Fatalf("repeat clarification = %q err=%v calls=%d", second, err, asker.calls) |
| 39 | } |
| 40 | rephrased, err := NewAskTool().Execute(ctx, []byte(`{"questions":[{"header":"Direction","question":"Please choose the direction now.","options":[{"label":"Keep going"},{"label":"Stop"}]}]}`)) |
| 41 | if err != nil || !strings.Contains(rephrased, "same ambiguity") || !strings.Contains(rephrased, "dec-original") || asker.calls != 1 { |
| 42 | t.Fatalf("rephrased clarification = %q err=%v calls=%d", rephrased, err, asker.calls) |
| 43 | } |
| 44 | if _, err := NewAskTool().Execute(ctx, []byte(`{"questions":[{"header":"Environment","question":"Which deployment target?","options":[{"label":"Staging"},{"label":"Production"}]}]}`)); err != nil { |
| 45 | t.Fatalf("unrelated clarification should remain available: %v", err) |
| 46 | } |
| 47 | if asker.calls != 2 { |
| 48 | t.Fatalf("unrelated clarification calls = %d, want 2", asker.calls) |
| 49 | } |
| 50 | _, err = NewAskTool().Execute(ctx, []byte(`{"decision_id":"dec-spoofed","new_evidence":"changed","questions":[{"header":"Again","question":"Ask again?","options":[{"label":"Yes"},{"label":"No"}]}]}`)) |
| 51 | if err == nil || !strings.Contains(err.Error(), "unknown decision_id") || asker.calls != 2 { |
| 52 | t.Fatalf("spoofed decision: err=%v calls=%d", err, asker.calls) |
| 53 | } |
| 54 | if _, err := NewAskTool().Execute(ctx, []byte(`{"decision_id":"dec-original","new_evidence":"new failing test","questions":[{"header":"Again","question":"Reconsider?","options":[{"label":"Keep going"},{"label":"Stop"}]}]}`)); err != nil { |
| 55 | t.Fatalf("evidence-backed reopen: %v", err) |
| 56 | } |
| 57 | if asker.calls != 3 { |
| 58 | t.Fatalf("reopened asker calls = %d, want 3", asker.calls) |
| 59 | } |
| 60 | } |
| 61 | |
| 62 | func TestAskToolRejectsBlankOptionLabels(t *testing.T) { |
| 63 | _, err := NewAskTool().Execute(context.Background(), []byte(`{ |
| 64 | "questions":[{ |
| 65 | "header":"Direction", |
| 66 | "question":"Which path?", |
| 67 | "options":[ |
| 68 | {"label":"Keep going"}, |
| 69 | {"label":" ","description":"blank labels render as empty picker rows"} |
| 70 | ] |
| 71 | }] |
| 72 | }`)) |
| 73 | if err == nil { |
| 74 | t.Fatal("expected blank option label to be rejected") |
| 75 | } |
| 76 | if !strings.Contains(err.Error(), "option 2") || !strings.Contains(err.Error(), "label") { |
| 77 | t.Fatalf("error = %v, want it to identify the blank option label", err) |
| 78 | } |
| 79 | } |
| 80 | |
| 81 | func TestAskToolRejectsDuplicateOptionLabelsAfterTrimming(t *testing.T) { |
| 82 | _, err := NewAskTool().Execute(context.Background(), []byte(`{ |
| 83 | "questions":[{ |
| 84 | "header":"Release", |
| 85 | "question":"What should happen next?", |
| 86 | "options":[ |
| 87 | {"label":"Deploy"}, |
| 88 | {"label":" Deploy ","description":"same label after trimming"} |
| 89 | ] |
| 90 | }] |
| 91 | }`)) |
| 92 | if err == nil { |
| 93 | t.Fatal("expected duplicate trimmed option label to be rejected") |
| 94 | } |
| 95 | if !strings.Contains(err.Error(), "option 2") || !strings.Contains(err.Error(), "duplicate") || !strings.Contains(err.Error(), "Deploy") { |
| 96 | t.Fatalf("error = %v, want it to identify the duplicate option label", err) |
| 97 | } |
| 98 | } |
| 99 | |
| 100 | func TestAskToolRejectsExactDuplicateOptionLabels(t *testing.T) { |
| 101 | _, err := NewAskTool().Execute(context.Background(), []byte(`{ |
| 102 | "questions":[{ |
| 103 | "header":"Release", |
| 104 | "question":"What should happen next?", |
| 105 | "options":[ |
| 106 | {"label":"Deploy"}, |
| 107 | {"label":"Deploy"} |
| 108 | ] |
| 109 | }] |
| 110 | }`)) |
| 111 | if err == nil { |
| 112 | t.Fatal("expected duplicate option label to be rejected") |
| 113 | } |
| 114 | if !strings.Contains(err.Error(), "option 2") || !strings.Contains(err.Error(), "duplicate") || !strings.Contains(err.Error(), "Deploy") { |
| 115 | t.Fatalf("error = %v, want it to identify the duplicate option label", err) |
| 116 | } |
| 117 | } |
| 118 | |
| 119 | func TestAskToolTrimsPromptAndOptionsBeforePrompting(t *testing.T) { |
| 120 | asker := &recordingAsker{} |
| 121 | ctx := withCallContext(context.Background(), "call_1", event.Discard, asker, false) |
| 122 | out, err := NewAskTool().Execute(ctx, []byte(`{ |
| 123 | "questions":[{ |
| 124 | "header":" Direction ", |
| 125 | "question":" Which path? ", |
| 126 | "options":[ |
| 127 | {"label":" Keep going ","description":" normal path "}, |
| 128 | {"label":" Stop "} |
| 129 | ] |
| 130 | }] |
| 131 | }`)) |
| 132 | if err != nil { |
| 133 | t.Fatalf("Execute: %v", err) |
| 134 | } |
| 135 | if !strings.Contains(out, "Direction: Keep going") { |
| 136 | t.Fatalf("answer summary = %q, want trimmed header and answer", out) |
| 137 | } |
| 138 | if len(asker.questions) != 1 { |
| 139 | t.Fatalf("questions = %+v, want one", asker.questions) |
| 140 | } |
| 141 | q := asker.questions[0] |
| 142 | if q.Header != "Direction" || q.Prompt != "Which path?" { |
| 143 | t.Fatalf("prompt text not trimmed: %+v", q) |
| 144 | } |
| 145 | if q.Options[0].Label != "Keep going" || q.Options[0].Description != "normal path" { |
| 146 | t.Fatalf("option text not trimmed: %+v", q.Options[0]) |
| 147 | } |
| 148 | } |
| 149 | |
| 150 | type fixedAsker struct{ answers []event.AskAnswer } |
| 151 | |
| 152 | func (f fixedAsker) Ask(_ context.Context, _ []event.AskQuestion) ([]event.AskAnswer, error) { |
| 153 | return f.answers, nil |
| 154 | } |
| 155 | |
| 156 | func TestAskToolProviderContractStable(t *testing.T) { |
| 157 | tool := NewAskTool() |
| 158 | contract := tool.Description() + "\n" + string(provider.CanonicalizeSchema(tool.Schema())) |
| 159 | got := fmt.Sprintf("%x", sha256.Sum256([]byte(contract))) |
| 160 | const want = "0c33e37d7e6e95a300e7e77dfc38e36445c83b377e302ddfe1b1743fc3f271df" |
| 161 | if got != want { |
| 162 | t.Fatalf("ask provider contract hash = %s, want %s; tool description or canonical schema changed", got, want) |
| 163 | } |
| 164 | } |
| 165 | |
| 166 | func TestAskToolDismissTellsModelToStopNotProceed(t *testing.T) { |
| 167 | ctx := withCallContext(context.Background(), "call_1", event.Discard, fixedAsker{answers: nil}, false) |
| 168 | out, err := NewAskTool().Execute(ctx, []byte(`{ |
| 169 | "questions":[{ |
| 170 | "header":"Config", |
| 171 | "question":"Configure a statusline script?", |
| 172 | "options":[{"label":"Yes"},{"label":"No"}] |
| 173 | }] |
| 174 | }`)) |
| 175 | if err != nil { |
| 176 | t.Fatalf("Execute: %v", err) |
| 177 | } |
| 178 | if strings.Contains(out, "(no answer)") { |
| 179 | t.Fatalf("dismiss result still uses the (no answer) wording the model reads as proceed: %q", out) |
| 180 | } |
| 181 | if !strings.Contains(out, "Do not") || !strings.Contains(out, "wait for the user") { |
| 182 | t.Fatalf("dismiss result should tell the model to stop and wait, got %q", out) |
| 183 | } |
| 184 | } |
| 185 | |
| 186 | func TestAskToolPartialAnswerMarksUnansweredQuestions(t *testing.T) { |
| 187 | ctx := withCallContext(context.Background(), "call_1", event.Discard, |
| 188 | fixedAsker{answers: []event.AskAnswer{{QuestionID: "q1", Selected: []string{"Deploy"}}}}, false) |
| 189 | out, err := NewAskTool().Execute(ctx, []byte(`{ |
| 190 | "questions":[ |
| 191 | {"header":"Release","question":"What next?","options":[{"label":"Deploy"},{"label":"Hold"}]}, |
| 192 | {"header":"Notify","question":"Tell the team?","options":[{"label":"Yes"},{"label":"No"}]} |
| 193 | ] |
| 194 | }`)) |
| 195 | if err != nil { |
| 196 | t.Fatalf("Execute: %v", err) |
| 197 | } |
| 198 | if !strings.Contains(out, "Release: Deploy") { |
| 199 | t.Fatalf("answered question should be reported, got %q", out) |
| 200 | } |
| 201 | if !strings.Contains(out, "Notify:") || !strings.Contains(out, "don't assume a choice") { |
| 202 | t.Fatalf("unanswered question should be marked, got %q", out) |
| 203 | } |
| 204 | } |
| 205 | |
| 206 | func TestAskToolHeadlessFallbackIsExplicitModelAssumption(t *testing.T) { |
| 207 | out, err := NewAskTool().Execute(context.Background(), []byte(`{ |
| 208 | "questions":[{ |
| 209 | "header":"Direction", |
| 210 | "question":"Which path?", |
| 211 | "options":[ |
| 212 | {"label":"Keep going"}, |
| 213 | {"label":"Stop"} |
| 214 | ] |
| 215 | }] |
| 216 | }`)) |
| 217 | if err != nil { |
| 218 | t.Fatalf("Execute: %v", err) |
| 219 | } |
| 220 | for _, want := range []string{"No interactive user answered", "model-assumption fallback", "not a user answer"} { |
| 221 | if !strings.Contains(out, want) { |
| 222 | t.Fatalf("headless fallback = %q, want it to contain %q", out, want) |
| 223 | } |
| 224 | } |
| 225 | if strings.Contains(out, "The user answered") { |
| 226 | t.Fatalf("headless fallback must not be formatted as a user answer: %q", out) |
| 227 | } |
| 228 | } |
| 229 |