dohertj2/chat
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

chore: document NICE trim order rationale (T71.3)

Browse Source 
T18 review (Phase 1) noted the NICE-tier trim drops previous-scene
FIRST while §6.3 spec lists previous-scene LAST in the NICE tier
group. Decision: keep the existing greedy order (previous-scene
first), and document why.

Rationale (now in code at the trim ladder):
  1. Cheapest-impact-first — a per-POV previous-scene summary loses
     less narrative continuity than the older dialogue turns or
     memory hits it competes with.
  2. Greedy lookahead is more expensive than the marginal narrative
     loss. Dropping previous-scene typically clears the soft-budget
     slack in one step.

Test added: test_nice_trim_order_documented pins the observed order
(previous-scene -> memories -> dialogue) so a future refactor can't
silently invert it. Sized so that all-NICE config overflows soft but
dropping just previous-scene fits — proves memories and older
dialogue turns survive while previous-scene is the FIRST drop.
This commit is contained in:
Joseph Doherty
2026-04-26 17:16:02 -04:00
parent afd1a50958
commit 73bb8c1f17
2 changed files with 165 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
chat/services/prompt.py
+20
View File
@@ -611,6 +611,26 @@ def assemble_narrative_prompt(
# Drop NICE in order: previous scene → memories beyond top-2 →
# older dialogue turns (collapse to 4).
#
# T71.3 — order rationale: the §6.3 spec lists NICE-tier members
# with previous-scene LAST, which read as a literal trim order
# during T18 review. We deliberately keep the greedy order shown
# here (previous-scene FIRST) for two reasons:
#
# 1. Cheapest-impact-first: a per-POV previous-scene summary is
# a single short paragraph that loses very little narrative
# continuity when dropped, while the older dialogue turns it
# is competing with carry the speaker's last few beats — those
# ground the next response far more concretely.
# 2. Greedy lookahead is more expensive than the marginal
# narrative loss. Dropping previous-scene typically clears
# the soft-budget slack in one step; trying memories or
# dialogue first would routinely require multiple recompute
# passes through the assembler.
#
# The pin test test_nice_trim_order_documented locks this order so
# a future refactor can't quietly invert it without surfacing the
# decision.
if include_prev:
include_prev = False
body, total = _build(
tests/test_prompt.py
+145
View File
@@ -574,6 +574,151 @@ def test_tight_budget_drops_guest_activity_bullet_first(tmp_path):
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.
msgs = assemble_narrative_prompt(
conn,
chat_id="chat_bot_a",
speaker_bot_id="bot_a",
recent_dialogue=dialogue,
retrieved_memory_summaries=memories,
budget_soft=400,
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."""