|
| 1 | +package anthropic |
| 2 | + |
| 3 | +import ( |
| 4 | + "testing" |
| 5 | + |
| 6 | + "github.com/stretchr/testify/assert" |
| 7 | + "github.com/stretchr/testify/require" |
| 8 | + |
| 9 | + "github.com/docker/cagent/pkg/chat" |
| 10 | + "github.com/docker/cagent/pkg/tools" |
| 11 | +) |
| 12 | + |
| 13 | +func TestConvertBetaMessages_MergesConsecutiveToolMessages(t *testing.T) { |
| 14 | + // Simulates the roast battle scenario where: |
| 15 | + // - Assistant message has 2 tool_use blocks (transfer_task calls) |
| 16 | + // - Two separate tool messages follow (one for each transfer_task result) |
| 17 | + // - These should be merged into a single user message with 2 tool_result blocks |
| 18 | + |
| 19 | + messages := []chat.Message{ |
| 20 | + { |
| 21 | + Role: chat.MessageRoleUser, |
| 22 | + Content: "Start the roast battle", |
| 23 | + }, |
| 24 | + { |
| 25 | + Role: chat.MessageRoleAssistant, |
| 26 | + Content: "Let me transfer tasks to the comedians", |
| 27 | + ToolCalls: []tools.ToolCall{ |
| 28 | + { |
| 29 | + ID: "tool_call_1", |
| 30 | + Type: "function", |
| 31 | + Function: tools.FunctionCall{ |
| 32 | + Name: "transfer_task", |
| 33 | + Arguments: `{"agent":"roaster_a","task":"Write roast"}`, |
| 34 | + }, |
| 35 | + }, |
| 36 | + { |
| 37 | + ID: "tool_call_2", |
| 38 | + Type: "function", |
| 39 | + Function: tools.FunctionCall{ |
| 40 | + Name: "transfer_task", |
| 41 | + Arguments: `{"agent":"roaster_b","task":"Write counter-roast"}`, |
| 42 | + }, |
| 43 | + }, |
| 44 | + }, |
| 45 | + }, |
| 46 | + { |
| 47 | + Role: chat.MessageRoleTool, |
| 48 | + Content: "Roast A completed", |
| 49 | + ToolCallID: "tool_call_1", |
| 50 | + }, |
| 51 | + { |
| 52 | + Role: chat.MessageRoleTool, |
| 53 | + Content: "Roast B completed", |
| 54 | + ToolCallID: "tool_call_2", |
| 55 | + }, |
| 56 | + { |
| 57 | + Role: chat.MessageRoleAssistant, |
| 58 | + Content: "Both roasts are complete!", |
| 59 | + }, |
| 60 | + } |
| 61 | + |
| 62 | + // Convert to Beta format |
| 63 | + betaMessages := convertBetaMessages(messages) |
| 64 | + |
| 65 | + // Verify structure: User -> Assistant (with 2 tool_use) -> User (with 2 tool_result) -> Assistant |
| 66 | + require.Len(t, betaMessages, 4, "Should have 4 messages after conversion") |
| 67 | + |
| 68 | + // Verify roles |
| 69 | + msg0Map, _ := marshalToMapBeta(betaMessages[0]) |
| 70 | + msg1Map, _ := marshalToMapBeta(betaMessages[1]) |
| 71 | + msg2Map, _ := marshalToMapBeta(betaMessages[2]) |
| 72 | + msg3Map, _ := marshalToMapBeta(betaMessages[3]) |
| 73 | + assert.Equal(t, "user", msg0Map["role"]) |
| 74 | + assert.Equal(t, "assistant", msg1Map["role"]) |
| 75 | + assert.Equal(t, "user", msg2Map["role"]) |
| 76 | + assert.Equal(t, "assistant", msg3Map["role"]) |
| 77 | + |
| 78 | + // Verify the second user message (tool results) has both tool_result blocks |
| 79 | + userMsg2Map, ok := marshalToMapBeta(betaMessages[2]) |
| 80 | + require.True(t, ok) |
| 81 | + content := contentArrayBeta(userMsg2Map) |
| 82 | + require.Len(t, content, 2, "User message should have 2 tool_result blocks") |
| 83 | + |
| 84 | + // Verify both tool_result IDs are present |
| 85 | + toolResultIDs := collectToolResultIDs(content) |
| 86 | + assert.Contains(t, toolResultIDs, "tool_call_1") |
| 87 | + assert.Contains(t, toolResultIDs, "tool_call_2") |
| 88 | + |
| 89 | + // Most importantly: validate that the sequence is valid for Anthropic API |
| 90 | + err := validateAnthropicSequencingBeta(betaMessages) |
| 91 | + require.NoError(t, err, "Converted messages should pass Anthropic sequencing validation") |
| 92 | +} |
| 93 | + |
| 94 | +func TestConvertBetaMessages_SingleToolMessage(t *testing.T) { |
| 95 | + // When there's only one tool message, it should still work correctly |
| 96 | + messages := []chat.Message{ |
| 97 | + { |
| 98 | + Role: chat.MessageRoleUser, |
| 99 | + Content: "Test", |
| 100 | + }, |
| 101 | + { |
| 102 | + Role: chat.MessageRoleAssistant, |
| 103 | + Content: "", |
| 104 | + ToolCalls: []tools.ToolCall{ |
| 105 | + { |
| 106 | + ID: "tool_1", |
| 107 | + Type: "function", |
| 108 | + Function: tools.FunctionCall{ |
| 109 | + Name: "test_tool", |
| 110 | + Arguments: `{}`, |
| 111 | + }, |
| 112 | + }, |
| 113 | + }, |
| 114 | + }, |
| 115 | + { |
| 116 | + Role: chat.MessageRoleTool, |
| 117 | + Content: "Tool result", |
| 118 | + ToolCallID: "tool_1", |
| 119 | + }, |
| 120 | + { |
| 121 | + Role: chat.MessageRoleAssistant, |
| 122 | + Content: "Done", |
| 123 | + }, |
| 124 | + } |
| 125 | + |
| 126 | + betaMessages := convertBetaMessages(messages) |
| 127 | + require.Len(t, betaMessages, 4) |
| 128 | + |
| 129 | + // Validate sequence |
| 130 | + err := validateAnthropicSequencingBeta(betaMessages) |
| 131 | + require.NoError(t, err) |
| 132 | +} |
| 133 | + |
| 134 | +func TestConvertBetaMessages_NonConsecutiveToolMessages(t *testing.T) { |
| 135 | + // When tool messages are separated by other messages (edge case) |
| 136 | + // Each tool message group should be handled independently |
| 137 | + messages := []chat.Message{ |
| 138 | + { |
| 139 | + Role: chat.MessageRoleUser, |
| 140 | + Content: "First request", |
| 141 | + }, |
| 142 | + { |
| 143 | + Role: chat.MessageRoleAssistant, |
| 144 | + Content: "", |
| 145 | + ToolCalls: []tools.ToolCall{ |
| 146 | + { |
| 147 | + ID: "tool_1", |
| 148 | + Type: "function", |
| 149 | + Function: tools.FunctionCall{ |
| 150 | + Name: "test_tool", |
| 151 | + Arguments: `{}`, |
| 152 | + }, |
| 153 | + }, |
| 154 | + }, |
| 155 | + }, |
| 156 | + { |
| 157 | + Role: chat.MessageRoleTool, |
| 158 | + Content: "Tool result 1", |
| 159 | + ToolCallID: "tool_1", |
| 160 | + }, |
| 161 | + { |
| 162 | + Role: chat.MessageRoleAssistant, |
| 163 | + Content: "Intermediate response", |
| 164 | + ToolCalls: []tools.ToolCall{ |
| 165 | + { |
| 166 | + ID: "tool_2", |
| 167 | + Type: "function", |
| 168 | + Function: tools.FunctionCall{ |
| 169 | + Name: "test_tool", |
| 170 | + Arguments: `{}`, |
| 171 | + }, |
| 172 | + }, |
| 173 | + }, |
| 174 | + }, |
| 175 | + { |
| 176 | + Role: chat.MessageRoleTool, |
| 177 | + Content: "Tool result 2", |
| 178 | + ToolCallID: "tool_2", |
| 179 | + }, |
| 180 | + { |
| 181 | + Role: chat.MessageRoleAssistant, |
| 182 | + Content: "Final response", |
| 183 | + }, |
| 184 | + } |
| 185 | + |
| 186 | + betaMessages := convertBetaMessages(messages) |
| 187 | + |
| 188 | + // Validate the entire sequence |
| 189 | + err := validateAnthropicSequencingBeta(betaMessages) |
| 190 | + require.NoError(t, err, "Messages with non-consecutive tool calls should still validate") |
| 191 | +} |
0 commit comments