| 1 | package agent |
| 2 | |
| 3 | import ( |
| 4 | "strings" |
| 5 | "testing" |
| 6 | |
| 7 | "reasonix/internal/event" |
| 8 | "reasonix/internal/evidence" |
| 9 | "reasonix/internal/provider" |
| 10 | ) |
| 11 | |
| 12 | func TestFinalReadinessFallsBackToCanonicalTodos(t *testing.T) { |
| 13 | ran := evidence.Receipt{ToolName: "bash", Success: true, Command: "git push"} |
| 14 | open := []evidence.TodoItem{{Content: "push", Status: "completed"}, {Content: "rebase", Status: "pending"}} |
| 15 | |
| 16 | // Turn did work (a successful bash) but issued no todo_write this turn, so the |
| 17 | // per-turn ledger has no list — the canonical state must still gate. |
| 18 | a := &Agent{evidence: readinessLedger(ran), todoState: open} |
| 19 | if got := a.ReadinessResult(); !strings.Contains(got.Reason, "incomplete") { |
| 20 | t.Fatalf("cross-turn gate = %q, want it to report incomplete canonical todos", got.Reason) |
| 21 | } |
| 22 | |
| 23 | // A turn that touched nothing (pure Q&A) must never gate on stale canonical state. |
| 24 | idle := &Agent{evidence: evidence.NewLedger(), todoState: open} |
| 25 | if got := idle.ReadinessResult(); !got.Ready { |
| 26 | t.Fatalf("no-work turn gated on canonical todos: %+v", got) |
| 27 | } |
| 28 | |
| 29 | // All canonical items completed → no gate even after work. |
| 30 | done := &Agent{evidence: readinessLedger(ran), todoState: []evidence.TodoItem{{Content: "push", Status: "completed"}}} |
| 31 | if got := done.ReadinessResult(); !got.Ready { |
| 32 | t.Fatalf("completed canonical todos still gated: %+v", got) |
| 33 | } |
| 34 | } |
| 35 | |
| 36 | func TestAdvanceCanonicalTodoCompletesAndPromotes(t *testing.T) { |
| 37 | a := &Agent{ |
| 38 | sink: event.Discard, |
| 39 | todoState: []evidence.TodoItem{ |
| 40 | {Content: "sync branch", Status: "in_progress"}, |
| 41 | {Content: "push to origin", Status: "pending"}, |
| 42 | {Content: "rebase", Status: "pending"}, |
| 43 | }, |
| 44 | } |
| 45 | a.advanceCanonicalTodo("sync branch") |
| 46 | |
| 47 | if a.todoState[0].Status != "completed" { |
| 48 | t.Fatalf("signed-off item not completed: %+v", a.todoState[0]) |
| 49 | } |
| 50 | if a.todoState[1].Status != "in_progress" { |
| 51 | t.Fatalf("next pending item not promoted: %+v", a.todoState[1]) |
| 52 | } |
| 53 | if a.todoState[2].Status != "pending" { |
| 54 | t.Fatalf("a later item was promoted out of order: %+v", a.todoState[2]) |
| 55 | } |
| 56 | } |
| 57 | |
| 58 | func TestAdvanceCanonicalTodoRejectsPendingMatchByNumber(t *testing.T) { |
| 59 | a := &Agent{sink: event.Discard, todoState: []evidence.TodoItem{ |
| 60 | {Content: "first", Status: "in_progress"}, |
| 61 | {Content: "second", Status: "pending"}, |
| 62 | }} |
| 63 | a.advanceCanonicalTodo("2") |
| 64 | if a.todoState[1].Status != "pending" || a.todoState[0].Status != "in_progress" { |
| 65 | t.Fatalf("pending numeric step advanced out of order: %+v", a.todoState) |
| 66 | } |
| 67 | } |
| 68 | |
| 69 | func TestRebuildTodoStateIgnoresHistoricalPendingSignoff(t *testing.T) { |
| 70 | msgs := []provider.Message{ |
| 71 | {Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{{ |
| 72 | ID: "t1", Name: "todo_write", |
| 73 | Arguments: `{"todos":[{"content":"a","status":"in_progress"},{"content":"b","status":"pending"}]}`, |
| 74 | }}}, |
| 75 | {Role: provider.RoleTool, ToolCallID: "t1", Name: "todo_write", Content: "Todos updated"}, |
| 76 | {Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{{ |
| 77 | ID: "c1", Name: "complete_step", Arguments: `{"step":"b"}`, |
| 78 | }}}, |
| 79 | {Role: provider.RoleTool, ToolCallID: "c1", Name: "complete_step", Content: "signed off"}, |
| 80 | } |
| 81 | a := &Agent{} |
| 82 | a.rebuildTodoState(msgs) |
| 83 | if a.todoState[0].Status != "in_progress" || a.todoState[1].Status != "pending" { |
| 84 | t.Fatalf("historical pending signoff restored invalid state: %+v", a.todoState) |
| 85 | } |
| 86 | } |
| 87 | |
| 88 | func TestSetTodoStateNormalizesLegacyOutOfOrderSnapshot(t *testing.T) { |
| 89 | a := &Agent{} |
| 90 | a.setTodoState([]evidence.TodoItem{ |
| 91 | {Content: "first", Status: "in_progress"}, |
| 92 | {Content: "second", Status: "completed"}, |
| 93 | }) |
| 94 | if a.todoState[0].Status != "in_progress" || a.todoState[1].Status != "pending" { |
| 95 | t.Fatalf("legacy snapshot was not normalized: %+v", a.todoState) |
| 96 | } |
| 97 | } |
| 98 | |
| 99 | func TestRebuildTodoStateReplaysCompleteSteps(t *testing.T) { |
| 100 | msgs := []provider.Message{ |
| 101 | {Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{{ |
| 102 | ID: "t1", Name: "todo_write", |
| 103 | Arguments: `{"todos":[{"content":"a","status":"in_progress"},{"content":"b","status":"pending"}]}`, |
| 104 | }}}, |
| 105 | {Role: provider.RoleTool, ToolCallID: "t1", Name: "todo_write", Content: "Todos updated"}, |
| 106 | {Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{{ |
| 107 | ID: "c1", Name: "complete_step", Arguments: `{"step":"a"}`, |
| 108 | }}}, |
| 109 | {Role: provider.RoleTool, ToolCallID: "c1", Name: "complete_step", Content: "signed off"}, |
| 110 | } |
| 111 | a := &Agent{} |
| 112 | a.rebuildTodoState(msgs) |
| 113 | |
| 114 | if len(a.todoState) != 2 { |
| 115 | t.Fatalf("rebuilt %d todos, want 2", len(a.todoState)) |
| 116 | } |
| 117 | if a.todoState[0].Status != "completed" { |
| 118 | t.Fatalf("complete_step not replayed onto canonical state: %+v", a.todoState[0]) |
| 119 | } |
| 120 | if a.todoState[1].Status != "in_progress" { |
| 121 | t.Fatalf("next item not promoted on replay: %+v", a.todoState[1]) |
| 122 | } |
| 123 | } |
| 124 | |
| 125 | func TestRebuildTodoStateSkipsFailedCompleteStep(t *testing.T) { |
| 126 | msgs := []provider.Message{ |
| 127 | {Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{{ |
| 128 | ID: "t1", Name: "todo_write", |
| 129 | Arguments: `{"todos":[{"content":"a","status":"in_progress"}]}`, |
| 130 | }}}, |
| 131 | {Role: provider.RoleTool, ToolCallID: "t1", Name: "todo_write", Content: "Todos updated"}, |
| 132 | {Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{{ |
| 133 | ID: "c1", Name: "complete_step", Arguments: `{"step":"a"}`, |
| 134 | }}}, |
| 135 | {Role: provider.RoleTool, ToolCallID: "c1", Name: "complete_step", Content: "error: no evidence"}, |
| 136 | } |
| 137 | a := &Agent{} |
| 138 | a.rebuildTodoState(msgs) |
| 139 | |
| 140 | if a.todoState[0].Status == "completed" { |
| 141 | t.Fatalf("a failed complete_step must not advance canonical state: %+v", a.todoState[0]) |
| 142 | } |
| 143 | } |
| 144 | |
| 145 | func TestRebuildTodoStateRequiresToolResults(t *testing.T) { |
| 146 | msgs := []provider.Message{ |
| 147 | {Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{{ |
| 148 | ID: "t1", Name: "todo_write", |
| 149 | Arguments: `{"todos":[{"content":"a","status":"in_progress"}]}`, |
| 150 | }}}, |
| 151 | {Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{{ |
| 152 | ID: "c1", Name: "complete_step", Arguments: `{"step":"a"}`, |
| 153 | }}}, |
| 154 | } |
| 155 | a := &Agent{} |
| 156 | a.rebuildTodoState(msgs) |
| 157 | if len(a.todoState) != 0 { |
| 158 | t.Fatalf("todo_write without tool result rebuilt canonical state: %+v", a.todoState) |
| 159 | } |
| 160 | |
| 161 | msgs = append(msgs[:1], |
| 162 | provider.Message{Role: provider.RoleTool, ToolCallID: "t1", Name: "todo_write", Content: "Todos updated"}, |
| 163 | msgs[1], |
| 164 | ) |
| 165 | a.rebuildTodoState(msgs) |
| 166 | if len(a.todoState) != 1 { |
| 167 | t.Fatalf("successful todo_write did not rebuild canonical state: %+v", a.todoState) |
| 168 | } |
| 169 | if got := a.todoState[0].Status; got != "in_progress" { |
| 170 | t.Fatalf("complete_step without tool result changed status to %q", got) |
| 171 | } |
| 172 | } |
| 173 | |
| 174 | func TestRebuildTodoStateHonorsEmptyTodoWriteClear(t *testing.T) { |
| 175 | msgs := []provider.Message{ |
| 176 | {Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{{ |
| 177 | ID: "t1", |
| 178 | Name: "todo_write", |
| 179 | Arguments: `{"todos":[{"content":"a","status":"in_progress"}]}`, |
| 180 | }}}, |
| 181 | {Role: provider.RoleTool, ToolCallID: "t1", Name: "todo_write", Content: "Todos updated"}, |
| 182 | {Role: provider.RoleAssistant, ToolCalls: []provider.ToolCall{{ |
| 183 | ID: "t2", |
| 184 | Name: "todo_write", |
| 185 | Arguments: `{"todos":[]}`, |
| 186 | }}}, |
| 187 | {Role: provider.RoleTool, ToolCallID: "t2", Name: "todo_write", Content: "Todos updated"}, |
| 188 | } |
| 189 | a := &Agent{} |
| 190 | a.rebuildTodoState(msgs) |
| 191 | if len(a.todoState) != 0 { |
| 192 | t.Fatalf("empty todo_write should clear rebuilt canonical state: %+v", a.todoState) |
| 193 | } |
| 194 | } |
| 195 | |
| 196 | func TestSeedTodoState(t *testing.T) { |
| 197 | a := &Agent{sink: event.Discard} |
| 198 | todos := []evidence.TodoItem{ |
| 199 | {Content: "step 1", Status: "in_progress"}, |
| 200 | {Content: "step 2", Status: "pending"}, |
| 201 | } |
| 202 | a.SeedTodoState(todos) |
| 203 | if len(a.todoState) != 2 { |
| 204 | t.Fatalf("SeedTodoState: got %d items, want 2", len(a.todoState)) |
| 205 | } |
| 206 | if a.todoState[0].Status != "in_progress" { |
| 207 | t.Fatalf("SeedTodoState: first item status = %q, want in_progress", a.todoState[0].Status) |
| 208 | } |
| 209 | } |
| 210 | |
| 211 | func TestSeedTodoStateReplacesExisting(t *testing.T) { |
| 212 | a := &Agent{sink: event.Discard, todoState: []evidence.TodoItem{ |
| 213 | {Content: "existing", Status: "in_progress"}, |
| 214 | }} |
| 215 | a.SeedTodoState([]evidence.TodoItem{ |
| 216 | {Content: "new", Status: "in_progress"}, |
| 217 | }) |
| 218 | if len(a.todoState) != 1 || a.todoState[0].Content != "new" { |
| 219 | t.Fatalf("SeedTodoState did not replace existing state: %+v", a.todoState) |
| 220 | } |
| 221 | } |
| 222 | |
| 223 | func TestSeedTodoStateAllowsAdvanceAfterSeed(t *testing.T) { |
| 224 | a := &Agent{sink: event.Discard} |
| 225 | a.SeedTodoState([]evidence.TodoItem{ |
| 226 | {Content: "step 1", Status: "in_progress"}, |
| 227 | {Content: "step 2", Status: "pending"}, |
| 228 | }) |
| 229 | a.advanceCanonicalTodo("step 1") |
| 230 | if a.todoState[0].Status != "completed" { |
| 231 | t.Fatalf("advance after seed: item 0 status = %q, want completed", a.todoState[0].Status) |
| 232 | } |
| 233 | if a.todoState[1].Status != "in_progress" { |
| 234 | t.Fatalf("advance after seed: item 1 status = %q, want in_progress", a.todoState[1].Status) |
| 235 | } |
| 236 | } |
| 237 | |
| 238 | func TestAdvanceCanonicalTodoWalksPhaseChain(t *testing.T) { |
| 239 | a := &Agent{sink: event.Discard, todoState: []evidence.TodoItem{ |
| 240 | {Content: "Port the parser", Status: "pending"}, |
| 241 | {Content: "move files", Status: "in_progress", Level: 1}, |
| 242 | {Content: "fix imports", Status: "pending", Level: 1}, |
| 243 | {Content: "Ship it", Status: "pending"}, |
| 244 | {Content: "run tests", Status: "pending", Level: 1}, |
| 245 | }} |
| 246 | |
| 247 | a.advanceCanonicalTodo("Port the parser") |
| 248 | if a.todoState[0].Status != "pending" { |
| 249 | t.Fatalf("pending phase advanced ahead of its sub-steps: %+v", a.todoState) |
| 250 | } |
| 251 | |
| 252 | a.advanceCanonicalTodo("move files") |
| 253 | if a.todoState[1].Status != "completed" || a.todoState[2].Status != "in_progress" { |
| 254 | t.Fatalf("completing a sub-step should promote its sibling: %+v", a.todoState) |
| 255 | } |
| 256 | |
| 257 | a.advanceCanonicalTodo("fix imports") |
| 258 | if a.todoState[2].Status != "completed" || a.todoState[0].Status != "in_progress" { |
| 259 | t.Fatalf("after the last sub-step the phase should become the signable item: %+v", a.todoState) |
| 260 | } |
| 261 | |
| 262 | a.advanceCanonicalTodo("Port the parser") |
| 263 | if a.todoState[0].Status != "completed" || a.todoState[3].Status != "pending" || a.todoState[4].Status != "in_progress" { |
| 264 | t.Fatalf("phase sign-off should promote the next phase's first sub-step: %+v", a.todoState) |
| 265 | } |
| 266 | } |
| 267 |