Dumitru Picalau
5056722e1a
Pack builder: Fix all and updates say why nothing went out while the web search card waits; page follow-up lines are not request words; Only rows narrow only the tag groups they rule on; answers are not request gaps; code-wired story beats skip the lookup lint; left-out rules read plainly and get Fix with AI; a rule overwriting a pick becomes an append; model labels read "For Illustrious"; roomier plan head. Chat: same-name edits refused with the shown item named; search by category for organizing; empty edit answers carry a note; prompt syntax coloured inside cards; every chat format named in the prompt; pointer captions avoid each other; one shared working line. Search: what an allowed search found shows as a card in the Pack builder and the Draft with AI panels; Use what fits sends all results for the model to keep what fits, ticking sends only those, None of these sends nothing.
252 lines
13 KiB
Python
252 lines
13 KiB
Python
"""The chat's `ask` tool: questions that pause the turn until the user
|
|
answers them, settled answers, and editing the last message."""
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
|
|
from engine.ai import config as ai_config
|
|
from tests.wp_api.test_ai_api import _chat, _two_wildcards
|
|
|
|
|
|
def _on(fake_llm, **extra):
|
|
ai_config.update({"enabled": True, "provider": "lmstudio",
|
|
"base_url": str(fake_llm.make_url("/v1")), "model": "m1", **extra})
|
|
|
|
|
|
async def _json(resp, status=200):
|
|
assert resp.status == status, await resp.text()
|
|
return await resp.json()
|
|
|
|
|
|
async def _resume(wp_client, chat_id, results, status=200):
|
|
return await _json(await wp_client.post("/wp/api/ai/chat", json={
|
|
"chat_id": chat_id, "results": results}), status)
|
|
|
|
|
|
QUESTIONS = {"questions": [
|
|
{"header": "Style", "question": "Which art style should the pack use?",
|
|
"choices": ["Anime", "Realistic: photo-like", "Painterly (oil and brush)"]},
|
|
{"q": "Add negatives to options?", "options": ["Yes", "No"]},
|
|
{"header": "Extras", "question": "Which extras?", "multi": True,
|
|
"choices": [{"label": "Hats"}, {"label": "Glasses", "hint": "round or square"}, "Scarves"]},
|
|
]}
|
|
|
|
|
|
def _ask_step(args=None, message=""):
|
|
return {"message": message, "calls": [
|
|
{"tool": "ask", "args": json.dumps(args or QUESTIONS)}]}
|
|
|
|
|
|
def _last_user(fake_llm) -> str:
|
|
return fake_llm.seen[-1]["body"]["messages"][-1]["content"]
|
|
|
|
|
|
async def test_ask_pauses_and_the_answers_carry_the_turn_on(wp_client, fake_llm):
|
|
_on(fake_llm)
|
|
fake_llm.chat_steps.extend([_ask_step(), {"message": "Building it.", "calls": []}])
|
|
first = await _chat(wp_client, "make me a fantasy character pack")
|
|
assert [c["tool"] for c in first["pending"]] == ["ask"]
|
|
call = first["pending"][0]
|
|
assert call["label"] == "Asked 3 questions"
|
|
style, neg, extras = call["args"]["questions"]
|
|
assert style == {"header": "Style", "question": "Which art style should the pack use?",
|
|
"multi": False, "choices": [
|
|
{"label": "Anime"}, {"label": "Realistic", "hint": "photo-like"},
|
|
{"label": "Painterly", "hint": "oil and brush"}]}
|
|
# Aliases (q, options) are read, and a missing header comes from the question.
|
|
assert neg["header"] == "Add negatives to" and neg["question"].startswith("Add")
|
|
assert extras["multi"] is True and extras["choices"][1] == {
|
|
"label": "Glasses", "hint": "round or square"}
|
|
chat = await _json(await wp_client.get(f"/wp/api/ai/chats/{first['chat_id']}"))
|
|
assert chat["waiting"] is True
|
|
|
|
# Only numbers come back; labels are the server's own.
|
|
done = await _resume(wp_client, first["chat_id"], [{"id": call["id"], "answers": [
|
|
{"q": 0, "picked": [0]}, {"q": 1, "decide": True},
|
|
{"q": 2, "picked": [0, 1, 9]}, {"q": 7, "picked": [0]}]}])
|
|
assert done["message"] == "Building it." and done["pending"] == []
|
|
back = _last_user(fake_llm)
|
|
assert ("ask: The user answered: Style → Anime; Add negatives to → you decide "
|
|
"(choose, and say what you chose); Extras → Hats, Glasses") in back
|
|
|
|
# A reopened chat shows the answers under the questions.
|
|
chat = await _json(await wp_client.get(f"/wp/api/ai/chats/{first['chat_id']}"))
|
|
assert chat["waiting"] is False
|
|
asked = chat["turns"][1]["view"]
|
|
assert asked["pending"][0]["tool"] == "ask"
|
|
assert asked["answered"] == {call["id"]: {"answers": [
|
|
{"header": "Style", "answer": "Anime"},
|
|
{"header": "Add negatives to", "answer": "you decide"},
|
|
{"header": "Extras", "answer": "Hats, Glasses"}]}}
|
|
|
|
# Settled: asking Style again is answered at once, without a pause.
|
|
fake_llm.chat_steps.extend([
|
|
_ask_step({"questions": [{"header": "style", "question": "Anime or realistic?",
|
|
"choices": ["Anime", "Realistic"]}]}),
|
|
{"message": "Anime it is.", "calls": []}])
|
|
again = await _chat(wp_client, "now the outfits", chat_id=first["chat_id"])
|
|
assert again["pending"] == [] and again["message"] == "Anime it is."
|
|
assert "ask: already answered: Style → Anime" in _last_user(fake_llm)
|
|
|
|
# Partly settled: the rest is asked, and the model hears what was settled.
|
|
fake_llm.chat_steps.extend([
|
|
_ask_step({"questions": [
|
|
{"header": "Style", "question": "Style?", "choices": ["A", "B"]},
|
|
{"header": "Palette", "question": "Which palette?", "choices": ["Warm", "Cool"]}]}),
|
|
{"message": "ok", "calls": []}])
|
|
part = await _chat(wp_client, "colours too", chat_id=first["chat_id"])
|
|
pcall = part["pending"][0]
|
|
assert [q["header"] for q in pcall["args"]["questions"]] == ["Palette"]
|
|
await _resume(wp_client, part["chat_id"], [{"id": pcall["id"], "answers": [
|
|
{"q": 0, "other": "pastel only"}]}])
|
|
assert ('Palette → "pastel only" (their own words). Already answered before: '
|
|
"Style → Anime") in _last_user(fake_llm)
|
|
|
|
|
|
async def test_bad_asks_are_refused_with_a_reason(wp_client, fake_llm):
|
|
_src, tgt = await _two_wildcards(wp_client)
|
|
_on(fake_llm)
|
|
one_choice = {"questions": [{"header": "X", "question": "Which?", "choices": ["Only"]}]}
|
|
generic = {"questions": [{"header": "More", "question": "Anything else you'd like?",
|
|
"choices": ["No", "Yes"]}]}
|
|
permission = {"questions": [{"header": "Go", "question": "Do you want me to add them?",
|
|
"choices": ["Yes", "No"]}]}
|
|
too_many = {"questions": [{"question": f"Q{i}?", "choices": ["a", "b"]} for i in range(5)]}
|
|
fake_llm.chat_steps.extend([_ask_step(a) for a in (too_many, one_choice, generic, permission)]
|
|
+ [{"message": "fine", "calls": []}])
|
|
out = await _chat(wp_client, "make a pack")
|
|
assert out["pending"] == []
|
|
notes = [s["note"] for s in out["steps"] if not s["ok"]]
|
|
assert "at most 4 questions" in notes[0]
|
|
assert "give 2-5" in notes[1]
|
|
assert "too general" in notes[2]
|
|
assert "asks permission" in notes[3]
|
|
|
|
# Never with change tools, and last in its step.
|
|
fake_llm.chat_steps.extend([
|
|
{"message": "", "calls": [
|
|
{"tool": "ask", "args": json.dumps(QUESTIONS)},
|
|
{"tool": "update_info", "args": json.dumps({"id": tgt, "tags_add": ["gear"]})}]},
|
|
{"message": "", "calls": [
|
|
{"tool": "ask", "args": json.dumps(QUESTIONS)},
|
|
{"tool": "read", "args": json.dumps({"id": tgt})}]},
|
|
{"message": "done", "calls": []},
|
|
])
|
|
out = await _chat(wp_client, "tag my Weapon", chat_id=out["chat_id"])
|
|
assert out["pending"] == []
|
|
notes = [s["note"] for s in out["steps"] if not s["ok"]]
|
|
assert "ask first, then propose" in notes[0]
|
|
assert "put ask last" in notes[1]
|
|
|
|
|
|
async def test_a_new_message_answers_the_ask_in_its_own_words(wp_client, fake_llm):
|
|
_on(fake_llm)
|
|
fake_llm.chat_steps.extend([_ask_step(), {"message": "Got it.", "calls": []}])
|
|
first = await _chat(wp_client, "make a pack")
|
|
call_id = first["pending"][0]["id"]
|
|
await _chat(wp_client, "anime, no negatives", chat_id=first["chat_id"])
|
|
last = _last_user(fake_llm)
|
|
assert last.index("ask: the user answered in their own words in the next message") < \
|
|
last.index("anime, no negatives")
|
|
assert "not run" not in last
|
|
chat = await _json(await wp_client.get(f"/wp/api/ai/chats/{first['chat_id']}"))
|
|
assert chat["turns"][1]["view"]["answered"] == {call_id: {"own": True}}
|
|
|
|
|
|
async def test_at_most_two_asks_per_message(wp_client, fake_llm):
|
|
_on(fake_llm)
|
|
|
|
def q(head):
|
|
return _ask_step({"questions": [{"header": head, "question": f"{head}?",
|
|
"choices": ["a", "b"]}]})
|
|
|
|
fake_llm.chat_steps.extend([q("One"), q("Two"), q("Three"), {"message": "ok", "calls": []}])
|
|
first = await _chat(wp_client, "pack please")
|
|
second = await _resume(wp_client, first["chat_id"], [
|
|
{"id": first["pending"][0]["id"], "answers": [{"q": 0, "picked": [1]}]}])
|
|
assert second["pending"][0]["args"]["questions"][0]["header"] == "Two"
|
|
third = await _resume(wp_client, first["chat_id"], [
|
|
{"id": second["pending"][0]["id"], "answers": []}])
|
|
assert third["pending"] == [] and third["message"] == "ok"
|
|
assert "decide yourself and say what you assumed" in _last_user(fake_llm)
|
|
# A new message may ask again.
|
|
fake_llm.chat_steps.extend([q("Four"), {"message": "ok", "calls": []}])
|
|
fourth = await _chat(wp_client, "next", chat_id=first["chat_id"])
|
|
assert fourth["pending"][0]["tool"] == "ask"
|
|
|
|
|
|
async def test_compaction_keeps_the_settled_answers(wp_client, fake_llm):
|
|
_on(fake_llm)
|
|
fake_llm.overrides["_ctx"] = 4096
|
|
fake_llm.overrides["summary"] = {"summary": "What the user wants:\n- a pack"}
|
|
fake_llm.chat_steps.extend([_ask_step(), {"message": "ok", "calls": []}])
|
|
first = await _chat(wp_client, "pack")
|
|
chat_id = first["chat_id"]
|
|
await _resume(wp_client, chat_id, [{"id": first["pending"][0]["id"], "answers": [
|
|
{"q": 0, "picked": [1]}]}])
|
|
fake_llm.chat_steps.extend({"message": f"Answer {i}. " + "words " * 200, "calls": []}
|
|
for i in range(4))
|
|
for word in ("two", "three", "four", "five"):
|
|
await _chat(wp_client, word, chat_id=chat_id)
|
|
meta = await _json(await wp_client.get(f"/wp/api/ai/chats/{chat_id}"))
|
|
assert "Style → Realistic" in meta["summary"]
|
|
assert "never ask them again" in meta["summary"]
|
|
|
|
|
|
async def test_editing_the_last_message_runs_the_new_text(wp_client, fake_llm):
|
|
_src, tgt = await _two_wildcards(wp_client)
|
|
_on(fake_llm)
|
|
fake_llm.chat_steps.extend([{"message": "One.", "calls": []},
|
|
{"message": "Red.", "calls": []},
|
|
{"message": "Blue.", "calls": []}])
|
|
first = await _chat(wp_client, "first")
|
|
await _chat(wp_client, "make it red", chat_id=first["chat_id"])
|
|
edited = await _chat(wp_client, "make it blue", chat_id=first["chat_id"], retry=True)
|
|
assert edited["message"] == "Blue."
|
|
users = [m["content"] for m in fake_llm.seen[-1]["body"]["messages"] if m["role"] == "user"]
|
|
assert any("make it blue" in u for u in users)
|
|
assert not any("make it red" in u for u in users)
|
|
chat = await _json(await wp_client.get(f"/wp/api/ai/chats/{first['chat_id']}"))
|
|
assert [t.get("text") or t["view"]["message"] for t in chat["turns"]] == [
|
|
"first", "One.", "make it blue", "Blue."]
|
|
|
|
# Refused like a retry once a card from that answer was applied.
|
|
fake_llm.chat_steps.append({"message": "Here.", "calls": [
|
|
{"tool": "update_info", "args": json.dumps({"id": tgt, "tags_add": ["gear"]})}]})
|
|
out = await _chat(wp_client, "tag Weapon", chat_id=first["chat_id"])
|
|
await _json(await wp_client.patch(f"/wp/api/ai/chats/{first['chat_id']}", json={
|
|
"events": [{"pid": out["proposals"][0]["pid"], "status": "applied"}]}))
|
|
refused = await _json(await wp_client.post("/wp/api/ai/chat", json={
|
|
"message": "tag Class", "chat_id": first["chat_id"], "retry": True}), 409)
|
|
assert "undo" in refused["error"]
|
|
|
|
|
|
async def test_the_item_an_answer_picks_may_be_changed(wp_client, fake_llm):
|
|
"""Haiku hair eval (2026-10-21): "make a hair colour wildcard" found the
|
|
user's own, asked whether to trim it, and the user picked "trim it";
|
|
the edit was then refused as an item they never named, because an
|
|
answer by button sends no text. The item the picked answer names
|
|
counts as named."""
|
|
_src, tgt = await _two_wildcards(wp_client)
|
|
_on(fake_llm, tool_format="json")
|
|
ask = {"questions": [{"header": "Weapons", "question": "You already have a weapon list.",
|
|
"choices": [f"Trim [[{tgt}#Weapon]]", "Make a new one"]}]}
|
|
edit = {"message": "Trimmed.", "calls": [
|
|
{"tool": "edit_options", "args": json.dumps({"id": tgt, "remove": [1]})}]}
|
|
fake_llm.chat_steps.extend([_ask_step(ask), edit])
|
|
first = await _chat(wp_client, "make a weapon wildcard", page={"route": "wildcards"})
|
|
[call] = first["pending"]
|
|
done = await _resume(wp_client, first["chat_id"], [{"id": call["id"], "answers": [
|
|
{"q": 0, "picked": [0]}]}])
|
|
[card] = done["proposals"]
|
|
assert card["op"]["id"] == tgt
|
|
assert [o["value"] for o in card["op"]["body"]["payload"]["options"]] == ["longsword"]
|
|
|
|
# Leaving it to the assistant names nothing: the edit is still refused.
|
|
fake_llm.chat_steps.extend([_ask_step(ask), edit, {"message": "Asked instead.", "calls": []}])
|
|
again = await _chat(wp_client, "make another weapon wildcard", page={"route": "wildcards"})
|
|
[call2] = again["pending"]
|
|
left = await _resume(wp_client, again["chat_id"], [{"id": call2["id"], "answers": [
|
|
{"q": 0, "decide": True}]}])
|
|
assert left["proposals"] == []
|