|
| 1 | +"""Regression: the agent tool-RAG domain classifier had no image/media domain, |
| 2 | +so image-generation requests matched no domain, were flagged low_signal, and had |
| 3 | +tool retrieval SKIPPED entirely — the model only received ALWAYS_AVAILABLE tools |
| 4 | +(manage_memory, ask_user, update_plan) and never `generate_image`/`edit_image`, |
| 5 | +so it could not generate images (and tended to loop on manage_memory). |
| 6 | +
|
| 7 | +Root cause: `_classify_agent_request` in src/agent_loop.py sets |
| 8 | +`low_signal = not continuation and not domains`; with no `images` domain, prompts |
| 9 | +like "generate two images of X" matched nothing -> low_signal -> retrieval skipped. |
| 10 | +
|
| 11 | +The classifier is deterministic string matching (no embeddings / no DB), so it |
| 12 | +can be exercised directly. |
| 13 | +""" |
| 14 | + |
| 15 | +from src.agent_loop import ( |
| 16 | + _classify_agent_request, |
| 17 | + _DOMAIN_TOOL_MAP, |
| 18 | + _DOMAIN_RULES, |
| 19 | + _domain_rules_for_tools, |
| 20 | +) |
| 21 | + |
| 22 | + |
| 23 | +def _classify(text): |
| 24 | + return _classify_agent_request([{"role": "user", "content": text}], text) |
| 25 | + |
| 26 | + |
| 27 | +def test_image_generation_requests_get_image_domain(): |
| 28 | + """Image-generation phrasings must match the `images` domain and NOT be |
| 29 | + treated as low-signal (which would skip tool retrieval).""" |
| 30 | + prompts = [ |
| 31 | + "generate two images of this character: one action pose, one relaxed", |
| 32 | + "draw me a picture of a cat", |
| 33 | + "make an illustration of a spaceship", |
| 34 | + "create an image of a sunset over mountains", |
| 35 | + "design a logo for my coffee shop", |
| 36 | + ] |
| 37 | + for p in prompts: |
| 38 | + intent = _classify(p) |
| 39 | + assert "images" in intent["domains"], f"expected images domain for: {p!r}" |
| 40 | + assert intent["low_signal"] is False, f"must not be low_signal: {p!r}" |
| 41 | + |
| 42 | + |
| 43 | +def test_image_edit_requests_get_image_domain(): |
| 44 | + """Edit/upscale/background phrasings also resolve to the image domain.""" |
| 45 | + for p in ("upscale image 5", "remove the background from this photo", "inpaint the selected area"): |
| 46 | + intent = _classify(p) |
| 47 | + assert "images" in intent["domains"], f"expected images domain for: {p!r}" |
| 48 | + |
| 49 | + |
| 50 | +def test_image_domain_seeds_generate_and_edit_image(): |
| 51 | + """The domain must seed the actual image tools so they are offered even when |
| 52 | + semantic retrieval misses.""" |
| 53 | + assert _DOMAIN_TOOL_MAP["images"] == {"generate_image", "edit_image"} |
| 54 | + |
| 55 | + |
| 56 | +def test_image_domain_has_a_rule_pack(): |
| 57 | + """Every domain in _DOMAIN_TOOL_MAP needs a matching _DOMAIN_RULES entry, |
| 58 | + otherwise _domain_rules_for_tools raises KeyError when the tools are selected.""" |
| 59 | + assert "images" in _DOMAIN_RULES |
| 60 | + rules = _domain_rules_for_tools({"generate_image"}) |
| 61 | + assert any("Image rules" in r for r in rules) |
| 62 | + |
| 63 | + |
| 64 | +def test_non_image_requests_do_not_match_image_domain(): |
| 65 | + """Guard against over-triggering: ordinary prompts must not be flagged image.""" |
| 66 | + assert "images" not in _classify("what is the capital of France")["domains"] |
| 67 | + assert "images" not in _classify("reply to the latest email in my inbox")["domains"] |
0 commit comments