Try to improve prompt adherence for low variety image models

vibe_bot/tests/test_main.py

@@ -1030,6 +1030,10 @@ def test_image_prompt_system_prompt_covers_key_details() -> None:
     assert "style" in lowered
     assert "canada goose" in lowered
     assert "only the image generation prompt" in lowered
+    assert "exactly as written" in lowered
+    assert "fountain pen wearing pants" in lowered
+    assert "centaur" in lowered
+    assert "not a person riding a horse" in lowered
 
 
 def test_doodlebob_prompt_rewrite_uses_detailed_system_prompt(
@@ -1046,6 +1050,7 @@ def test_doodlebob_prompt_rewrite_uses_detailed_system_prompt(
     mock_llama_wrapper.chat_completion_instruct.side_effect = [
         "landscape",  # layout selection
         "a very detailed prompt",  # prompt rewrite
+        "PASS",  # prompt verification
     ]
     mock_llama_wrapper.image_generation.return_value = "aW1hZ2U="
     mock_database.get_image_generation_time_estimate.return_value = None
@@ -1064,6 +1069,55 @@ def test_doodlebob_prompt_rewrite_uses_detailed_system_prompt(
     )
 
 
+def test_verify_image_prompt_passes_unchanged(
+    mock_llama_wrapper: MagicMock,
+) -> None:
+    """A PASS verdict keeps the original prompt and sends both inputs."""
+    import vibe_bot.main as main_module
+
+    mock_llama_wrapper.chat_completion_instruct.return_value = "PASS"
+    result = main_module.verify_image_prompt("a centaur", "a detailed prompt")
+    assert result == "a detailed prompt"
+    call = mock_llama_wrapper.chat_completion_instruct.call_args
+    assert call.kwargs["system_prompt"] == main_module.IMAGE_PROMPT_VERIFY_SYSTEM_PROMPT
+    assert "a centaur" in call.kwargs["user_prompt"]
+    assert "a detailed prompt" in call.kwargs["user_prompt"]
+
+
+def test_verify_image_prompt_case_insensitive_pass(
+    mock_llama_wrapper: MagicMock,
+) -> None:
+    """A lower-case, punctuated 'pass' verdict also keeps the original."""
+    import vibe_bot.main as main_module
+
+    mock_llama_wrapper.chat_completion_instruct.return_value = "pass."
+    assert main_module.verify_image_prompt("a centaur", "p") == "p"
+
+
+def test_verify_image_prompt_correction_replaces_prompt(
+    mock_llama_wrapper: MagicMock,
+) -> None:
+    """A non-PASS response is used as the corrected prompt."""
+    import vibe_bot.main as main_module
+
+    corrected = (
+        "one fused creature, a human torso joined to a horse's front half, "
+        "human arms raised, standing in a rocky field"
+    )
+    mock_llama_wrapper.chat_completion_instruct.return_value = corrected
+    assert main_module.verify_image_prompt("a centaur", "a horse") == corrected
+
+
+def test_verify_image_prompt_empty_falls_back(
+    mock_llama_wrapper: MagicMock,
+) -> None:
+    """An empty verification response falls back to the drafted prompt."""
+    import vibe_bot.main as main_module
+
+    mock_llama_wrapper.chat_completion_instruct.return_value = ""
+    assert main_module.verify_image_prompt("a centaur", "drafted") == "drafted"
+
+
 def test_doodlebob_selects_portrait(
     mock_ctx: MagicMock,
     mock_llama_wrapper: MagicMock,
@@ -1078,6 +1132,7 @@ def test_doodlebob_selects_portrait(
     mock_llama_wrapper.chat_completion_instruct.side_effect = [
         "portrait",  # layout selection
         "a tall portrait of a lighthouse",  # prompt rewrite
+        "PASS",  # prompt verification
     ]
     mock_llama_wrapper.image_generation.return_value = "aW1hZ2U="
     mock_database.get_image_generation_time_estimate.return_value = None
@@ -1113,6 +1168,7 @@ def test_doodlebob_selects_landscape(
     mock_llama_wrapper.chat_completion_instruct.side_effect = [
         "landscape",
         "a wide panoramic coastline",
+        "PASS",  # prompt verification
     ]
     mock_llama_wrapper.image_generation.return_value = "aW1hZ2U="
     mock_database.get_image_generation_time_estimate.return_value = None
@@ -1137,6 +1193,7 @@ def test_doodlebob_malformed_layout_defaults_square(
     mock_llama_wrapper.chat_completion_instruct.side_effect = [
         "banana",  # malformed layout
         "a balanced composition",
+        "PASS",  # prompt verification
     ]
     mock_llama_wrapper.image_generation.return_value = "aW1hZ2U="
     mock_database.get_image_generation_time_estimate.return_value = None
@@ -1163,6 +1220,7 @@ def test_doodlebob_empty_layout_defaults_square(
     mock_llama_wrapper.chat_completion_instruct.side_effect = [
         "",  # empty layout response
         "a balanced composition",
+        "PASS",  # prompt verification
     ]
     mock_llama_wrapper.image_generation.return_value = "aW1hZ2U="
     mock_database.get_image_generation_time_estimate.return_value = None
@@ -1188,6 +1246,7 @@ def test_doodlebob_reports_estimate_and_elapsed(
     mock_llama_wrapper.chat_completion_instruct.side_effect = [
         "square",
         "a test scene",
+        "PASS",  # prompt verification
     ]
     mock_llama_wrapper.image_generation.return_value = "aW1hZ2U="
     mock_database.get_image_generation_time_estimate.return_value = 12.34
@@ -1220,6 +1279,7 @@ def test_doodlebob_no_estimate_without_history(
     mock_llama_wrapper.chat_completion_instruct.side_effect = [
         "square",
         "a test scene",
+        "PASS",  # prompt verification
     ]
     mock_llama_wrapper.image_generation.return_value = "aW1hZ2U="
     mock_database.get_image_generation_time_estimate.return_value = None
@@ -1247,6 +1307,7 @@ def test_doodlebob_failed_generation_not_recorded(
     mock_llama_wrapper.chat_completion_instruct.side_effect = [
         "square",
         "a test scene",
+        "PASS",  # prompt verification
     ]
     mock_llama_wrapper.image_generation.return_value = ""
     mock_database.get_image_generation_time_estimate.return_value = 8.0