Joseph Doherty
3b83786b8b
Verbose roleplay-tuned narrators (Cydonia, Magnum, etc.) reliably ignore prompt-level beat-count instructions and ramble for 6-12 asterisk-action beats per turn — even with HARD CAP language and worked examples in the closing instruction. The fix is a deterministic post-stream trimmer: - New trim_to_max_beats(text, max_beats) in chat/services/prompt.py. Counts * characters in the streamed output (each beat = 2 asterisks: open + close), trims at the start of the (max_beats+1)th asterisk action, strips trailing whitespace. Idempotent and safe on under-cap input. - Wired into post_turn for both the primary stream (3-beat cap) and the optional interjection stream (2-beat cap — interjections are by definition shorter chime-ins). - Tightened the closing instruction: explicit "HARD CAP: 2-3 beats" with "After the third beat, STOP". Helps the well-behaved models self-cap; the post-processor catches the rest. - max_tokens: 250 -> 160 (lets the 3rd beat finish naturally before hitting the physical cap; trim_to_max_beats handles 4+ beat overflow). temperature: 0.85 -> 0.7 (Cydonia is more compliant with format instructions at slightly cooler sampling). - Test budgets bumped (closing grew ~15 tokens with the new wording). 6 new tests for trim_to_max_beats covering passthrough, exact-cap, 4-beat trim, 6-beat runaway, lower caps, zero cap. Verified live: 4-turn bench against chat_maya, every response is 2-3 beats consistently. Suite: 470 passed in 11.7s.
918 lines
34 KiB
Python
918 lines
34 KiB
Python
"""Tests for chat.services.prompt.assemble_narrative_prompt.
|
|
|
|
Covers Task 18 — must/should/nice trim tiers (Requirements §3.2) and
|
|
the speaker prompt assembly order (§6.3). Tests use direct event-log
|
|
seeding so the projector populates state exactly the way the runtime
|
|
will at play-time. No LLM is invoked: prompt assembly is deterministic.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import pytest
|
|
|
|
from chat.db.connection import open_db
|
|
from chat.db.migrate import apply_migrations
|
|
from chat.eventlog.log import append_and_apply, append_event
|
|
from chat.eventlog.projector import project
|
|
import chat.state.entities # noqa: F401 (registers handlers)
|
|
import chat.state.edges # noqa: F401
|
|
import chat.state.memory # noqa: F401
|
|
import chat.state.world # noqa: F401
|
|
import chat.state.events # noqa: F401
|
|
import chat.state.threads # noqa: F401
|
|
from chat.llm.client import Message
|
|
from chat.services.prompt import (
|
|
_witness_role_for,
|
|
assemble_narrative_prompt,
|
|
trim_to_max_beats,
|
|
)
|
|
|
|
|
|
def _seed_basic(conn) -> None:
|
|
"""Seed bot, you-entity, edge, chat, container, scene, activities."""
|
|
append_event(conn, kind="bot_authored", payload={
|
|
"id": "bot_a",
|
|
"name": "Aria",
|
|
"persona": "reserved coworker who notices things",
|
|
"voice_samples": ["I — sorry, I didn't mean to.", "Right. Of course."],
|
|
"traits": ["introverted", "observant"],
|
|
"backstory": "An archivist who joined the firm last spring.",
|
|
"initial_relationship_to_you": "coworker; mild crush; never voiced",
|
|
"kickoff_prose": "you stay late at the office",
|
|
})
|
|
append_event(conn, kind="you_authored", payload={
|
|
"name": "Sam",
|
|
"pronouns": "they/them",
|
|
"persona": "tired analyst",
|
|
})
|
|
append_event(conn, kind="chat_created", payload={
|
|
"id": "chat_bot_a",
|
|
"host_bot_id": "bot_a",
|
|
"guest_bot_id": None,
|
|
"initial_time": "2026-04-26T20:00:00+00:00",
|
|
"narrative_anchor": "Day 1 evening",
|
|
"weather": "clear",
|
|
})
|
|
append_event(conn, kind="container_created", payload={
|
|
"chat_id": "chat_bot_a",
|
|
"name": "office bullpen",
|
|
"type": "workplace",
|
|
"properties": {"public": False, "moving": False, "audible_range": "room"},
|
|
})
|
|
append_event(conn, kind="edge_update", payload={
|
|
"source_id": "bot_a",
|
|
"target_id": "you",
|
|
"affinity_delta": 12,
|
|
"trust_delta": 5,
|
|
"knowledge_facts": [
|
|
"they work on the same floor",
|
|
"they've stayed late twice this week",
|
|
],
|
|
})
|
|
append_event(conn, kind="activity_change", payload={
|
|
"entity_id": "you",
|
|
"container_id": 1,
|
|
"posture": "sitting at your desk",
|
|
"action": {"verb": "finishing emails"},
|
|
"attention": "the screen",
|
|
"holding": ["coffee mug"],
|
|
})
|
|
append_event(conn, kind="activity_change", payload={
|
|
"entity_id": "bot_a",
|
|
"container_id": 1,
|
|
"posture": "sitting at her desk",
|
|
"action": {"verb": "pretending to work"},
|
|
"attention": "you, in glances",
|
|
})
|
|
append_event(conn, kind="scene_opened", payload={
|
|
"chat_id": "chat_bot_a",
|
|
"container_id": 1,
|
|
"started_at": "2026-04-26T20:00:00+00:00",
|
|
"participants": ["you", "bot_a"],
|
|
})
|
|
project(conn)
|
|
|
|
|
|
def test_basic_assembly_returns_system_message_with_all_must_blocks(tmp_path):
|
|
db = tmp_path / "t.db"
|
|
apply_migrations(db)
|
|
with open_db(db) as conn:
|
|
_seed_basic(conn)
|
|
msgs = assemble_narrative_prompt(
|
|
conn,
|
|
chat_id="chat_bot_a",
|
|
speaker_bot_id="bot_a",
|
|
recent_dialogue=[],
|
|
retrieved_memory_summaries=[],
|
|
)
|
|
assert isinstance(msgs, list)
|
|
assert len(msgs) == 1
|
|
sys_msg = msgs[0]
|
|
assert isinstance(sys_msg, Message)
|
|
assert sys_msg.role == "system"
|
|
body = sys_msg.content
|
|
# Must-include markers
|
|
assert "Aria" in body
|
|
assert "PERSONA" in body
|
|
assert "ACTIVITIES" in body
|
|
assert "CURRENT SCENE" in body
|
|
# Edge to addressee — name + numeric values (default affinity 50, +12 = 62)
|
|
assert "Sam" in body
|
|
assert "62/100" in body
|
|
|
|
|
|
def test_user_turn_appended_as_user_message(tmp_path):
|
|
db = tmp_path / "t.db"
|
|
apply_migrations(db)
|
|
with open_db(db) as conn:
|
|
_seed_basic(conn)
|
|
msgs = assemble_narrative_prompt(
|
|
conn,
|
|
chat_id="chat_bot_a",
|
|
speaker_bot_id="bot_a",
|
|
user_turn_prose="*looks up* Hey.",
|
|
recent_dialogue=[],
|
|
retrieved_memory_summaries=[],
|
|
)
|
|
assert len(msgs) == 2
|
|
assert msgs[0].role == "system"
|
|
assert msgs[1].role == "user"
|
|
assert msgs[1].content == "*looks up* Hey."
|
|
|
|
|
|
def test_must_only_succeeds_with_empty_optional_blocks(tmp_path):
|
|
"""No dialogue, memories, other edges, or previous scene summary — should not raise."""
|
|
db = tmp_path / "t.db"
|
|
apply_migrations(db)
|
|
with open_db(db) as conn:
|
|
_seed_basic(conn)
|
|
msgs = assemble_narrative_prompt(
|
|
conn,
|
|
chat_id="chat_bot_a",
|
|
speaker_bot_id="bot_a",
|
|
recent_dialogue=None, # default → nothing
|
|
retrieved_memory_summaries=None,
|
|
user_turn_prose=None,
|
|
)
|
|
assert len(msgs) == 1
|
|
body = msgs[0].content
|
|
# Must blocks present
|
|
assert "PERSONA" in body
|
|
assert "ACTIVITIES" in body
|
|
# Optional blocks not in body (nothing to render)
|
|
assert "OTHER EDGES" not in body
|
|
assert "PREVIOUS SCENE SUMMARY" not in body
|
|
assert "RELEVANT MEMORIES" not in body
|
|
|
|
|
|
def test_long_dialogue_keeps_last_4_verbatim_and_summarizes_earlier(tmp_path):
|
|
"""Stuff a huge dialogue history under budget pressure; older turns
|
|
must be elided to a placeholder, the last 4 verbatim, and earlier
|
|
unique markers gone.
|
|
"""
|
|
db = tmp_path / "t.db"
|
|
apply_migrations(db)
|
|
with open_db(db) as conn:
|
|
_seed_basic(conn)
|
|
dialogue = []
|
|
for i in range(20):
|
|
speaker = "you" if i % 2 == 0 else "bot_a"
|
|
# Each line ~250 tokens of filler => 20 turns ≈ 5000 tokens,
|
|
# which together with MUST blocks pushes over soft (1500).
|
|
dialogue.append({
|
|
"speaker": speaker,
|
|
"text": f"unique-line-marker-{i:02d} " + ("filler " * 200),
|
|
})
|
|
msgs = assemble_narrative_prompt(
|
|
conn,
|
|
chat_id="chat_bot_a",
|
|
speaker_bot_id="bot_a",
|
|
recent_dialogue=dialogue,
|
|
retrieved_memory_summaries=[],
|
|
# Soft small enough to force NICE trim but hard fits MUST + 4.
|
|
budget_soft=1200,
|
|
budget_hard=8000,
|
|
)
|
|
body = msgs[0].content
|
|
# The last 4 unique markers (16, 17, 18, 19) must be present verbatim.
|
|
for i in range(16, 20):
|
|
assert f"unique-line-marker-{i:02d}" in body, f"expected last-4 marker {i} in body"
|
|
# Older markers must be dropped (replaced by elision placeholder).
|
|
for i in range(0, 16):
|
|
assert f"unique-line-marker-{i:02d}" not in body
|
|
# An "earlier" summary line must be present.
|
|
assert "earlier" in body.lower()
|
|
# Token count of system message respects hard budget.
|
|
import tiktoken
|
|
enc = tiktoken.get_encoding("cl100k_base")
|
|
assert len(enc.encode(body)) <= 8000
|
|
|
|
|
|
def test_memories_drop_to_top_2_under_budget_pressure(tmp_path):
|
|
"""4 memory summaries, each large; under tight soft budget only 2 should appear."""
|
|
db = tmp_path / "t.db"
|
|
apply_migrations(db)
|
|
with open_db(db) as conn:
|
|
_seed_basic(conn)
|
|
# Each ~1500 tokens of repeated text; drop tier should kick in.
|
|
long_chunk = "alpha beta gamma delta " * 400
|
|
memories = [
|
|
f"MEMORY-A {long_chunk}",
|
|
f"MEMORY-B {long_chunk}",
|
|
f"MEMORY-C {long_chunk}",
|
|
f"MEMORY-D {long_chunk}",
|
|
]
|
|
msgs = assemble_narrative_prompt(
|
|
conn,
|
|
chat_id="chat_bot_a",
|
|
speaker_bot_id="bot_a",
|
|
recent_dialogue=[],
|
|
retrieved_memory_summaries=memories,
|
|
# Pressure: budgets that allow MUST + 2 memories but not 4.
|
|
budget_soft=4000,
|
|
budget_hard=5000,
|
|
)
|
|
body = msgs[0].content
|
|
# MEMORY-A and MEMORY-B are the top-2 and should remain; C & D dropped.
|
|
assert "MEMORY-A" in body
|
|
assert "MEMORY-B" in body
|
|
assert "MEMORY-C" not in body
|
|
assert "MEMORY-D" not in body
|
|
# Token count fits the hard budget.
|
|
import tiktoken
|
|
enc = tiktoken.get_encoding("cl100k_base")
|
|
assert len(enc.encode(body)) <= 5000
|
|
|
|
|
|
def test_must_exceeds_budget_hard_raises_value_error(tmp_path):
|
|
db = tmp_path / "t.db"
|
|
apply_migrations(db)
|
|
with open_db(db) as conn:
|
|
_seed_basic(conn)
|
|
with pytest.raises(ValueError):
|
|
assemble_narrative_prompt(
|
|
conn,
|
|
chat_id="chat_bot_a",
|
|
speaker_bot_id="bot_a",
|
|
recent_dialogue=[],
|
|
retrieved_memory_summaries=[],
|
|
budget_soft=5,
|
|
budget_hard=10,
|
|
)
|
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# Task 43: multi-entity prompt assembly (guest_id support)
|
|
# ---------------------------------------------------------------------------
|
|
|
|
|
|
def _seed_with_guest(conn) -> None:
|
|
"""Seed a 3-entity scene: you (Sam) + host (Aria, bot_a) + guest (Iris, bot_b).
|
|
|
|
Group node row is initialized with summary + dynamic, edges in all
|
|
relevant directions are seeded, and activities are recorded for all
|
|
three entities.
|
|
"""
|
|
append_event(conn, kind="bot_authored", payload={
|
|
"id": "bot_a",
|
|
"name": "Aria",
|
|
"persona": "reserved coworker who notices things",
|
|
"voice_samples": ["I — sorry, I didn't mean to.", "Right. Of course."],
|
|
"traits": ["introverted", "observant"],
|
|
"backstory": "An archivist who joined the firm last spring.",
|
|
"initial_relationship_to_you": "coworker; mild crush; never voiced",
|
|
"kickoff_prose": "you stay late at the office",
|
|
})
|
|
append_event(conn, kind="bot_authored", payload={
|
|
"id": "bot_b",
|
|
"name": "Iris",
|
|
"persona": "wry transplant from the Boston office",
|
|
"voice_samples": ["Oh, please.", "Don't make me say it twice."],
|
|
"traits": ["sardonic", "loyal"],
|
|
"backstory": "Met Aria at a conference two years back.",
|
|
"initial_relationship_to_you": "stranger; curious",
|
|
"kickoff_prose": "",
|
|
})
|
|
append_event(conn, kind="you_authored", payload={
|
|
"name": "Sam",
|
|
"pronouns": "they/them",
|
|
"persona": "tired analyst",
|
|
})
|
|
append_event(conn, kind="chat_created", payload={
|
|
"id": "chat_bot_a",
|
|
"host_bot_id": "bot_a",
|
|
"guest_bot_id": "bot_b",
|
|
"initial_time": "2026-04-26T20:00:00+00:00",
|
|
"narrative_anchor": "Day 1 evening",
|
|
"weather": "clear",
|
|
})
|
|
append_event(conn, kind="container_created", payload={
|
|
"chat_id": "chat_bot_a",
|
|
"name": "office bullpen",
|
|
"type": "workplace",
|
|
"properties": {"public": False, "moving": False, "audible_range": "room"},
|
|
})
|
|
# Edges: host -> you, guest -> you, host -> guest, guest -> host.
|
|
append_event(conn, kind="edge_update", payload={
|
|
"source_id": "bot_a",
|
|
"target_id": "you",
|
|
"affinity_delta": 12,
|
|
"trust_delta": 5,
|
|
"knowledge_facts": ["they work on the same floor"],
|
|
})
|
|
append_event(conn, kind="edge_update", payload={
|
|
"source_id": "bot_a",
|
|
"target_id": "bot_b",
|
|
"affinity_delta": 20,
|
|
"trust_delta": 15,
|
|
"knowledge_facts": ["studied physics together"],
|
|
})
|
|
append_event(conn, kind="edge_update", payload={
|
|
"source_id": "bot_b",
|
|
"target_id": "you",
|
|
"affinity_delta": 4,
|
|
"trust_delta": 0,
|
|
"knowledge_facts": ["Aria's coworker"],
|
|
})
|
|
append_event(conn, kind="edge_update", payload={
|
|
"source_id": "bot_b",
|
|
"target_id": "bot_a",
|
|
"affinity_delta": 18,
|
|
"trust_delta": 12,
|
|
"knowledge_facts": ["former roommate"],
|
|
})
|
|
# Activity for all three entities — note distinct verbs so we can
|
|
# check whose activity got dropped under tight budget.
|
|
append_event(conn, kind="activity_change", payload={
|
|
"entity_id": "you",
|
|
"container_id": 1,
|
|
"posture": "sitting at your desk",
|
|
"action": {"verb": "finishing emails"},
|
|
"attention": "the screen",
|
|
"holding": ["coffee mug"],
|
|
})
|
|
append_event(conn, kind="activity_change", payload={
|
|
"entity_id": "bot_a",
|
|
"container_id": 1,
|
|
"posture": "sitting at her desk",
|
|
"action": {"verb": "pretending to work"},
|
|
"attention": "you, in glances",
|
|
})
|
|
append_event(conn, kind="activity_change", payload={
|
|
"entity_id": "bot_b",
|
|
"container_id": 1,
|
|
"posture": "leaning against the doorframe",
|
|
"action": {"verb": "smirking-distinctively"},
|
|
"attention": "Aria",
|
|
})
|
|
append_event(conn, kind="scene_opened", payload={
|
|
"chat_id": "chat_bot_a",
|
|
"container_id": 1,
|
|
"started_at": "2026-04-26T20:00:00+00:00",
|
|
"participants": ["you", "bot_a", "bot_b"],
|
|
})
|
|
append_event(conn, kind="group_node_initialized", payload={
|
|
"chat_id": "chat_bot_a",
|
|
"members": ["you", "bot_a", "bot_b"],
|
|
"summary": "Three coworkers catching up after hours UNIQUE-GROUP-SUMMARY.",
|
|
"dynamic": "warm-but-prickly UNIQUE-GROUP-DYNAMIC",
|
|
})
|
|
project(conn)
|
|
|
|
|
|
def test_assemble_with_no_guest_matches_phase1(tmp_path):
|
|
"""Regression: 2-entity scenario without guest_id behaves exactly as Phase 1."""
|
|
db = tmp_path / "t.db"
|
|
apply_migrations(db)
|
|
with open_db(db) as conn:
|
|
_seed_basic(conn)
|
|
msgs = assemble_narrative_prompt(
|
|
conn,
|
|
chat_id="chat_bot_a",
|
|
speaker_bot_id="bot_a",
|
|
recent_dialogue=[],
|
|
retrieved_memory_summaries=[],
|
|
)
|
|
body = msgs[0].content
|
|
# Phase 1 must blocks present.
|
|
assert "Aria" in body
|
|
assert "PERSONA" in body
|
|
assert "Sam" in body
|
|
assert "ACTIVITIES" in body
|
|
assert "62/100" in body # speaker → addressee edge intact
|
|
# No guest content leaks in.
|
|
assert "Group dynamic" not in body
|
|
assert "Iris" not in body
|
|
|
|
|
|
def test_assemble_with_guest_includes_group_node_summary(tmp_path):
|
|
"""When guest is present (auto-detected via chat.guest_bot_id) and a
|
|
group_node row exists, its summary + dynamic are rendered."""
|
|
db = tmp_path / "t.db"
|
|
apply_migrations(db)
|
|
with open_db(db) as conn:
|
|
_seed_with_guest(conn)
|
|
msgs = assemble_narrative_prompt(
|
|
conn,
|
|
chat_id="chat_bot_a",
|
|
speaker_bot_id="bot_a",
|
|
recent_dialogue=[],
|
|
retrieved_memory_summaries=[],
|
|
)
|
|
body = msgs[0].content
|
|
assert "Group dynamic" in body
|
|
assert "UNIQUE-GROUP-SUMMARY" in body
|
|
assert "UNIQUE-GROUP-DYNAMIC" in body
|
|
# Guest activity also present (SHOULD-tier, fits at default budget).
|
|
assert "smirking-distinctively" in body
|
|
# Speaker's other edges include the host -> guest direction.
|
|
assert "Iris" in body
|
|
|
|
|
|
def test_assemble_when_speaker_is_guest_orients_edges_correctly(tmp_path):
|
|
"""When the guest is the speaker, identity is the guest, the
|
|
addressee edge is guest → you, and other edges include guest → host."""
|
|
db = tmp_path / "t.db"
|
|
apply_migrations(db)
|
|
with open_db(db) as conn:
|
|
_seed_with_guest(conn)
|
|
msgs = assemble_narrative_prompt(
|
|
conn,
|
|
chat_id="chat_bot_a",
|
|
speaker_bot_id="bot_b", # guest as speaker
|
|
recent_dialogue=[],
|
|
retrieved_memory_summaries=[],
|
|
)
|
|
body = msgs[0].content
|
|
# Speaker identity is the guest's persona.
|
|
assert "You are Iris." in body
|
|
assert "wry transplant from the Boston office" in body
|
|
# Edge to addressee is guest → you (Sam) with the seeded values
|
|
# (default 50 + 4 affinity = 54).
|
|
assert "YOUR EDGE TO Sam" in body
|
|
assert "54/100" in body
|
|
# Other edges include guest → host (Aria) with seeded value
|
|
# (default 50 + 18 = 68).
|
|
assert "OTHER EDGES" in body
|
|
assert "Aria" in body
|
|
assert "68/100" in body
|
|
|
|
|
|
def test_speaker_is_guest_uses_guest_witness_role(tmp_path, monkeypatch):
|
|
"""T71.1: when the guest is the speaker, ``search_memories`` is
|
|
called with ``witness_role="guest"``, not the previously-hardcoded
|
|
``"host"``. Pins the parametric witness role at the prompt call site
|
|
so guest-as-speaker honours the witness mask via Phase 2 T46.
|
|
"""
|
|
db = tmp_path / "t.db"
|
|
apply_migrations(db)
|
|
captured: dict = {}
|
|
|
|
def _fake_search(conn, owner_id, witness_role, query, k=4):
|
|
captured["owner_id"] = owner_id
|
|
captured["witness_role"] = witness_role
|
|
captured["query"] = query
|
|
return []
|
|
|
|
# Patch the imported reference inside the prompt module so the
|
|
# production call site uses the fake.
|
|
import chat.services.prompt as prompt_mod
|
|
monkeypatch.setattr(prompt_mod, "search_memories", _fake_search)
|
|
|
|
with open_db(db) as conn:
|
|
_seed_with_guest(conn)
|
|
# Guest as speaker — must request memories with witness_role="guest".
|
|
assemble_narrative_prompt(
|
|
conn,
|
|
chat_id="chat_bot_a",
|
|
speaker_bot_id="bot_b",
|
|
recent_dialogue=[],
|
|
# retrieved_memory_summaries=None forces the search path.
|
|
retrieved_memory_summaries=None,
|
|
)
|
|
assert captured["owner_id"] == "bot_b"
|
|
assert captured["witness_role"] == "guest"
|
|
|
|
|
|
def test_speaker_is_host_uses_host_witness_role(tmp_path, monkeypatch):
|
|
"""T71.1 (regression): host-as-speaker still queries with
|
|
``witness_role="host"``."""
|
|
db = tmp_path / "t.db"
|
|
apply_migrations(db)
|
|
captured: dict = {}
|
|
|
|
def _fake_search(conn, owner_id, witness_role, query, k=4):
|
|
captured["witness_role"] = witness_role
|
|
return []
|
|
|
|
import chat.services.prompt as prompt_mod
|
|
monkeypatch.setattr(prompt_mod, "search_memories", _fake_search)
|
|
|
|
with open_db(db) as conn:
|
|
_seed_with_guest(conn)
|
|
assemble_narrative_prompt(
|
|
conn,
|
|
chat_id="chat_bot_a",
|
|
speaker_bot_id="bot_a", # host as speaker
|
|
recent_dialogue=[],
|
|
retrieved_memory_summaries=None,
|
|
)
|
|
assert captured["witness_role"] == "host"
|
|
|
|
|
|
def test_single_activities_block_with_three_bullets_when_3_entities(tmp_path):
|
|
"""T71.2: with you + host + guest present, the assembled prompt
|
|
contains exactly ONE ``ACTIVITIES:`` header and bullets for all
|
|
three entities (no duplicate header from the prior dual-block
|
|
rendering).
|
|
"""
|
|
db = tmp_path / "t.db"
|
|
apply_migrations(db)
|
|
with open_db(db) as conn:
|
|
_seed_with_guest(conn)
|
|
msgs = assemble_narrative_prompt(
|
|
conn,
|
|
chat_id="chat_bot_a",
|
|
speaker_bot_id="bot_a",
|
|
recent_dialogue=[],
|
|
retrieved_memory_summaries=[],
|
|
)
|
|
body = msgs[0].content
|
|
# Exactly one ACTIVITIES: header.
|
|
assert body.count("ACTIVITIES:") == 1
|
|
# Bullets for all three entities (you=Sam, host=Aria, guest=Iris)
|
|
# — pin on the unique action verbs from the seed data.
|
|
assert "finishing emails" in body # you bullet
|
|
assert "pretending to work" in body # speaker (host) bullet
|
|
assert "smirking-distinctively" in body # guest bullet
|
|
|
|
|
|
def test_tight_budget_drops_guest_activity_bullet_first(tmp_path):
|
|
"""T71.2: under tight budget the speaker bullet survives but the
|
|
guest activity bullet is the first ACTIVITIES: bullet to drop. The
|
|
block as a whole stays present (it's MUST-tier); only its body
|
|
contracts.
|
|
"""
|
|
db = tmp_path / "t.db"
|
|
apply_migrations(db)
|
|
with open_db(db) as conn:
|
|
_seed_with_guest(conn)
|
|
dialogue = [
|
|
{"speaker": "you", "text": "line-16 hi there"},
|
|
{"speaker": "bot_a", "text": "line-17 hey"},
|
|
{"speaker": "you", "text": "line-18 quiet night"},
|
|
{"speaker": "bot_a", "text": "line-19 indeed"},
|
|
]
|
|
msgs = assemble_narrative_prompt(
|
|
conn,
|
|
chat_id="chat_bot_a",
|
|
speaker_bot_id="bot_a",
|
|
recent_dialogue=dialogue,
|
|
retrieved_memory_summaries=[],
|
|
# Closing instruction grew with the asterisk-format spec
|
|
# (Phase 4.6 narrative-style fix). Budget bumped enough to
|
|
# accommodate the larger MUST floor while still exercising
|
|
# the SHOULD-tier trim path.
|
|
budget_soft=480,
|
|
budget_hard=510,
|
|
)
|
|
body = msgs[0].content
|
|
# Speaker bullet survives (MUST-tier floor).
|
|
assert "pretending to work" in body
|
|
assert "ACTIVITIES:" in body
|
|
# Guest bullet is dropped first under budget pressure.
|
|
assert "smirking-distinctively" not in body
|
|
|
|
|
|
def test_nice_trim_order_documented(tmp_path):
|
|
"""T71.3: pin the NICE-tier trim order so a future refactor can't
|
|
quietly invert it.
|
|
|
|
Order under NICE pressure is:
|
|
1. previous-scene summary (dropped FIRST)
|
|
2. memories beyond top-2
|
|
3. older dialogue turns (collapsed to last-4)
|
|
|
|
We size the budget so that all-NICE-included is over soft, but
|
|
dropping ONLY previous-scene gets us back under soft. The observed
|
|
behaviour we pin: previous-scene gone, memories/dialogue intact.
|
|
"""
|
|
db = tmp_path / "t.db"
|
|
apply_migrations(db)
|
|
# Heavy previous-scene summary — large enough that dropping it
|
|
# alone clears the soft-budget overage. Defined out here so the
|
|
# marker is in scope for the assertions below.
|
|
prev_scene_blob = "PREVSCENE-MARKER " + ("filler " * 200)
|
|
with open_db(db) as conn:
|
|
# Append all events first, project once at the end (project is
|
|
# not idempotent — it replays every event in the log).
|
|
from chat.eventlog.log import append_event as _append
|
|
_append(conn, kind="bot_authored", payload={
|
|
"id": "bot_a",
|
|
"name": "Aria",
|
|
"persona": "reserved coworker who notices things",
|
|
"voice_samples": ["I — sorry, I didn't mean to."],
|
|
"traits": ["introverted"],
|
|
"backstory": "An archivist who joined the firm last spring.",
|
|
"initial_relationship_to_you": "coworker",
|
|
"kickoff_prose": "you stay late at the office",
|
|
})
|
|
_append(conn, kind="you_authored", payload={
|
|
"name": "Sam",
|
|
"pronouns": "they/them",
|
|
"persona": "tired analyst",
|
|
})
|
|
_append(conn, kind="chat_created", payload={
|
|
"id": "chat_bot_a",
|
|
"host_bot_id": "bot_a",
|
|
"guest_bot_id": None,
|
|
"initial_time": "2026-04-26T20:00:00+00:00",
|
|
"narrative_anchor": "Day 1 evening",
|
|
"weather": "clear",
|
|
})
|
|
_append(conn, kind="container_created", payload={
|
|
"chat_id": "chat_bot_a",
|
|
"name": "office bullpen",
|
|
"type": "workplace",
|
|
"properties": {"public": False, "moving": False, "audible_range": "room"},
|
|
})
|
|
_append(conn, kind="edge_update", payload={
|
|
"source_id": "bot_a",
|
|
"target_id": "you",
|
|
"affinity_delta": 12,
|
|
"trust_delta": 5,
|
|
"knowledge_facts": ["they work on the same floor"],
|
|
})
|
|
_append(conn, kind="activity_change", payload={
|
|
"entity_id": "you",
|
|
"container_id": 1,
|
|
"posture": "sitting at your desk",
|
|
"action": {"verb": "finishing emails"},
|
|
"attention": "the screen",
|
|
})
|
|
_append(conn, kind="activity_change", payload={
|
|
"entity_id": "bot_a",
|
|
"container_id": 1,
|
|
"posture": "sitting at her desk",
|
|
"action": {"verb": "pretending to work"},
|
|
"attention": "you, in glances",
|
|
})
|
|
_append(conn, kind="scene_opened", payload={
|
|
"chat_id": "chat_bot_a",
|
|
"container_id": 1,
|
|
"started_at": "2026-04-26T20:00:00+00:00",
|
|
"participants": ["you", "bot_a"],
|
|
})
|
|
# Close the seeded scene and write a per-POV summary memory so
|
|
# _resolve_previous_scene_summary returns a non-empty string.
|
|
_append(conn, kind="scene_closed", payload={
|
|
"scene_id": 1,
|
|
"ended_at": "2026-04-26T20:30:00+00:00",
|
|
"significance": 2,
|
|
})
|
|
_append(conn, kind="memory_written", payload={
|
|
"owner_id": "bot_a",
|
|
"chat_id": "chat_bot_a",
|
|
"scene_id": 1,
|
|
"pov_summary": prev_scene_blob,
|
|
"witness_you": 1,
|
|
"witness_host": 1,
|
|
"witness_guest": 0,
|
|
"source": "direct",
|
|
"reliability": 1.0,
|
|
"significance": 2,
|
|
})
|
|
project(conn)
|
|
|
|
# Six dialogue turns — last 4 plus 2 older. If older turns are
|
|
# dropped under NICE pressure, the unique markers for turns 0/1
|
|
# disappear; we'll assert they REMAIN to prove dialogue trim
|
|
# didn't fire.
|
|
dialogue = [
|
|
{"speaker": "you", "text": "DLG-OLD-00 hello"},
|
|
{"speaker": "bot_a", "text": "DLG-OLD-01 hi"},
|
|
{"speaker": "you", "text": "DLG-LAST-16 ok"},
|
|
{"speaker": "bot_a", "text": "DLG-LAST-17 sure"},
|
|
{"speaker": "you", "text": "DLG-LAST-18 night"},
|
|
{"speaker": "bot_a", "text": "DLG-LAST-19 indeed"},
|
|
]
|
|
# Four small memories — if "memories beyond top-2" trim fires,
|
|
# MEM-C/MEM-D disappear; we'll assert they REMAIN to prove
|
|
# memories trim didn't fire either.
|
|
memories = ["MEM-A short", "MEM-B short", "MEM-C short", "MEM-D short"]
|
|
|
|
# Soft tuned so the all-NICE config (with the heavy previous
|
|
# scene summary) overflows, but dropping just previous-scene
|
|
# fits comfortably. Hard set high so SHOULD-tier never trims.
|
|
# Soft bumped (was 400) to make room for the larger closing
|
|
# instruction shipped with the asterisk-format spec.
|
|
msgs = assemble_narrative_prompt(
|
|
conn,
|
|
chat_id="chat_bot_a",
|
|
speaker_bot_id="bot_a",
|
|
recent_dialogue=dialogue,
|
|
retrieved_memory_summaries=memories,
|
|
budget_soft=540,
|
|
budget_hard=8000,
|
|
)
|
|
body = msgs[0].content
|
|
# Previous-scene summary was the FIRST NICE drop — its unique
|
|
# marker must be absent.
|
|
assert "PREVSCENE-MARKER" not in body
|
|
# Memories beyond top-2 stayed (proves memories trim did NOT fire).
|
|
assert "MEM-A" in body
|
|
assert "MEM-B" in body
|
|
assert "MEM-C" in body
|
|
assert "MEM-D" in body
|
|
# Older dialogue turns stayed (proves dialogue trim did NOT fire).
|
|
assert "DLG-OLD-00" in body
|
|
assert "DLG-OLD-01" in body
|
|
# Last-4 dialogue turns of course present.
|
|
assert "DLG-LAST-19" in body
|
|
|
|
|
|
def test_assemble_with_tight_budget_drops_guest_activity_first(tmp_path):
|
|
"""Under tight budget MUST blocks survive but SHOULD-tier guest
|
|
activity is dropped first."""
|
|
db = tmp_path / "t.db"
|
|
apply_migrations(db)
|
|
with open_db(db) as conn:
|
|
_seed_with_guest(conn)
|
|
# Short dialogue so MUST core (speaker identity + edge + last 4
|
|
# turns + closing) sits comfortably under the hard budget while
|
|
# SHOULD-tier additions (guest activity, group node, other edges)
|
|
# would push over.
|
|
dialogue = [
|
|
{"speaker": "you", "text": "line-16 hi there"},
|
|
{"speaker": "bot_a", "text": "line-17 hey"},
|
|
{"speaker": "you", "text": "line-18 quiet night"},
|
|
{"speaker": "bot_a", "text": "line-19 indeed"},
|
|
]
|
|
msgs = assemble_narrative_prompt(
|
|
conn,
|
|
chat_id="chat_bot_a",
|
|
speaker_bot_id="bot_a",
|
|
recent_dialogue=dialogue,
|
|
retrieved_memory_summaries=[],
|
|
# MUST core ~310 tokens; SHOULD additions (guest activity +
|
|
# group node + other edges) push it well over 380. budget_hard
|
|
# is set just above MUST core so SHOULD-tier blocks must be
|
|
# trimmed away.
|
|
# Closing instruction grew with the asterisk-format spec
|
|
# (Phase 4.6 narrative-style fix). Budget bumped enough to
|
|
# accommodate the larger MUST floor while still exercising
|
|
# the SHOULD-tier trim path.
|
|
budget_soft=480,
|
|
budget_hard=510,
|
|
)
|
|
body = msgs[0].content
|
|
# MUST: speaker identity, edge to addressee, last 4 dialogue turns.
|
|
assert "Aria" in body
|
|
assert "YOUR EDGE TO Sam" in body
|
|
for i in range(16, 20):
|
|
assert f"line-{i:02d}" in body
|
|
# Guest activity (SHOULD-tier) must be dropped under tight budget.
|
|
assert "smirking-distinctively" not in body
|
|
# Token budget honoured. Bumped (was 340) for the larger closing
|
|
# instruction that ships the asterisk-format spec.
|
|
import tiktoken
|
|
enc = tiktoken.get_encoding("cl100k_base")
|
|
assert len(enc.encode(body)) <= 510
|
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# Task 60: Active events + open threads in prompt assembly
|
|
# ---------------------------------------------------------------------------
|
|
|
|
|
|
def test_assemble_with_no_events_or_threads_omits_blocks(tmp_path):
|
|
"""Regression: with the basic 2-entity scenario (no events seeded, no
|
|
threads seeded), the assembled prompt must NOT contain the
|
|
``Active events:`` or ``Open threads:`` headers — both blocks are
|
|
omit-when-empty."""
|
|
db = tmp_path / "t.db"
|
|
apply_migrations(db)
|
|
with open_db(db) as conn:
|
|
_seed_basic(conn)
|
|
msgs = assemble_narrative_prompt(
|
|
conn,
|
|
chat_id="chat_bot_a",
|
|
speaker_bot_id="bot_a",
|
|
recent_dialogue=[],
|
|
retrieved_memory_summaries=[],
|
|
)
|
|
body = msgs[0].content
|
|
assert "Active events:" not in body
|
|
assert "Open threads:" not in body
|
|
|
|
|
|
def test_assemble_with_active_events_renders_block(tmp_path):
|
|
"""Seed a planned event then transition it to active; the assembled
|
|
prompt should render the ``Active events:`` block listing the event
|
|
by kind."""
|
|
db = tmp_path / "t.db"
|
|
apply_migrations(db)
|
|
with open_db(db) as conn:
|
|
_seed_basic(conn)
|
|
# event_planned then event_started → status="active". Use
|
|
# append_and_apply because _seed_basic already projected; calling
|
|
# project() again would replay every prior event (and trip
|
|
# UNIQUE constraints on chat_created etc.).
|
|
append_and_apply(conn, kind="event_planned", payload={
|
|
"event_id": "evt_park",
|
|
"chat_id": "chat_bot_a",
|
|
"kind": "date_at_park",
|
|
"props": {"location": "Riverside Park", "vibe": "casual"},
|
|
"planned_for": "2026-04-30T18:00:00+00:00",
|
|
})
|
|
append_and_apply(conn, kind="event_started", payload={
|
|
"event_id": "evt_park",
|
|
"started_at": "2026-04-30T18:05:00+00:00",
|
|
})
|
|
msgs = assemble_narrative_prompt(
|
|
conn,
|
|
chat_id="chat_bot_a",
|
|
speaker_bot_id="bot_a",
|
|
recent_dialogue=[],
|
|
retrieved_memory_summaries=[],
|
|
)
|
|
body = msgs[0].content
|
|
assert "Active events:" in body
|
|
assert "date_at_park" in body
|
|
|
|
|
|
def test_assemble_with_open_thread_renders_block(tmp_path):
|
|
"""Seed a single open thread; the assembled prompt should render the
|
|
``Open threads:`` block listing the thread by title."""
|
|
db = tmp_path / "t.db"
|
|
apply_migrations(db)
|
|
with open_db(db) as conn:
|
|
_seed_basic(conn)
|
|
# _seed_basic already projected; use append_and_apply for the
|
|
# post-seed event so we don't re-trigger UNIQUE constraint
|
|
# collisions on the prior chat_created/etc. events.
|
|
append_and_apply(conn, kind="thread_opened", payload={
|
|
"thread_id": "thr_job",
|
|
"chat_id": "chat_bot_a",
|
|
"title": "Maya's job hunt",
|
|
"summary": "Maya is looking for a new job",
|
|
})
|
|
msgs = assemble_narrative_prompt(
|
|
conn,
|
|
chat_id="chat_bot_a",
|
|
speaker_bot_id="bot_a",
|
|
recent_dialogue=[],
|
|
retrieved_memory_summaries=[],
|
|
)
|
|
body = msgs[0].content
|
|
assert "Open threads:" in body
|
|
assert "Maya's job hunt" in body
|
|
|
|
|
|
def test_witness_role_for_none_host_returns_host():
|
|
assert _witness_role_for("bot_a", None) == "host"
|
|
# Sanity check: existing semantics preserved.
|
|
assert _witness_role_for("bot_a", "bot_a") == "host"
|
|
assert _witness_role_for("bot_a", "bot_b") == "guest"
|
|
|
|
|
|
# ---------------------------------------------------------------------------
|
|
# trim_to_max_beats — caps verbose narrative output to N beats
|
|
# ---------------------------------------------------------------------------
|
|
|
|
|
|
def test_trim_to_max_beats_passthrough_when_under_cap():
|
|
assert trim_to_max_beats("", 3) == ""
|
|
assert trim_to_max_beats("plain text", 3) == "plain text"
|
|
two = "*She nods* okay. *She turns* see you."
|
|
assert trim_to_max_beats(two, 3) == two
|
|
|
|
|
|
def test_trim_to_max_beats_passthrough_at_exactly_cap():
|
|
three = "*A* one. *B* two. *C* three."
|
|
assert trim_to_max_beats(three, 3) == three
|
|
|
|
|
|
def test_trim_to_max_beats_cuts_at_fourth_beat():
|
|
"""Cydonia-style 4-beat output trimmed at the start of the 4th
|
|
asterisk action; trailing whitespace stripped."""
|
|
four = "*A* one. *B* two. *C* three. *D* four."
|
|
assert trim_to_max_beats(four, 3) == "*A* one. *B* two. *C* three."
|
|
|
|
|
|
def test_trim_to_max_beats_handles_runaway_six_beats():
|
|
"""The exact failure mode that motivated this — verbose narrator
|
|
rambling for 6 beats when the prompt asked for 2-3."""
|
|
six = "*A* 1 *B* 2 *C* 3 *D* 4 *E* 5 *F* 6"
|
|
assert trim_to_max_beats(six, 3) == "*A* 1 *B* 2 *C* 3"
|
|
|
|
|
|
def test_trim_to_max_beats_respects_lower_cap():
|
|
four = "*A* one. *B* two. *C* three. *D* four."
|
|
assert trim_to_max_beats(four, 2) == "*A* one. *B* two."
|
|
assert trim_to_max_beats(four, 1) == "*A* one."
|
|
|
|
|
|
def test_trim_to_max_beats_zero_returns_empty():
|
|
assert trim_to_max_beats("*A* one. *B* two.", 0) == ""
|