Grayscale/nanobot
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

review(dream): harden line-age annotation per review feedback

Browse Source 
Follow-up to #3212, fully backward compatible:

- Extract the 14-day staleness threshold as `_STALE_THRESHOLD_DAYS` module
  constant and pass it into the Phase 1 prompt template as
  `{{ stale_threshold_days }}`. The number lived in three places before
  (code threshold, prompt instruction, docstring); now there is one.
- Add `DreamConfig.annotate_line_ages` (default True = current behavior)
  and propagate it through `Dream.__init__` and the gateway wiring in
  cli/commands.py. Gives users a knob to disable the feature without a
  code patch if an LLM reacts poorly to the `← Nd` suffix.
- Harden `_annotate_with_ages` against dirty working trees: when HEAD
  blob line count disagrees with the working-tree content length, skip
  annotation entirely instead of assigning ages to the wrong lines. The
  previous `i >= len(ages)` guard only handled one direction of the
  mismatch.
- Inline-comment the `max_iterations` 10→15 bump with a pointer to
  exp002 so future blame has context.
- Add 4 regression tests: end-to-end `← 30d` reaches prompt, 14/15
  threshold boundary, `annotate_line_ages=False` bypasses git entirely
  (verified via `assert_not_called`), length-mismatch defense, and
  template-var rendering.

Made-with: Cursor
This commit is contained in:
Xubin Ren
2026-04-17 13:45:38 +08:00
committed by Xubin Ren
parent 35f3084c03
commit cc5a666d5d
5 changed files with 132 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files
nanobot/agent/memory.py
+43 -11
View File
@@ -552,6 +552,13 @@ class Consolidator:
# ---------------------------------------------------------------------------
# Single source of truth for the staleness threshold used in _annotate_with_ages
# *and* in the Phase 1 prompt template (passed as `stale_threshold_days`).
# Keep code and prompt aligned — if you bump this, the LLM's instruction string
# updates automatically.
_STALE_THRESHOLD_DAYS = 14
class Dream:
"""Two-phase memory processor: analyze history.jsonl, then edit files via AgentRunner.
@@ -568,6 +575,7 @@ class Dream:
max_batch_size: int = 20,
max_iterations: int = 10,
max_tool_result_chars: int = 16_000,
annotate_line_ages: bool = True,
):
self.store = store
self.provider = provider
@@ -575,6 +583,10 @@ class Dream:
self.max_batch_size = max_batch_size
self.max_iterations = max_iterations
self.max_tool_result_chars = max_tool_result_chars
# Kill switch for the git-blame-based per-line age annotation in Phase 1.
# Default True keeps the #3212 behavior; set False to feed MEMORY.md raw
# (e.g. if a specific LLM reacts poorly to the `← Nd` suffix).
self.annotate_line_ages = annotate_line_ages
self._runner = AgentRunner(provider)
self._tools = self._build_tools()
@@ -635,9 +647,12 @@ class Dream:
def _annotate_with_ages(self, content: str) -> str:
"""Append per-line age suffixes to MEMORY.md content.
Each non-blank line gets a suffix like ``← 30d`` indicating how
many days since it was last modified. Lines ≤14 days old get no
suffix. Returns the original content unchanged if git is unavailable.
Each non-blank line whose age exceeds ``_STALE_THRESHOLD_DAYS`` gets a
suffix like ``← 30d`` indicating days since last modification.
Returns the original content unchanged if git is unavailable,
annotate fails, or the line count doesn't match the age count
(which can happen with an uncommitted working-tree edit — better to
skip annotation than to tag the wrong line).
SOUL.md and USER.md are never annotated.
"""
file_path = "memory/MEMORY.md"
@@ -651,14 +666,23 @@ class Dream:
had_trailing = content.endswith("\n")
lines = content.splitlines()
# If HEAD-blob line count disagrees with the working-tree content we
# received, ages would be assigned to the wrong lines — skip entirely
# and feed the LLM un-annotated content rather than misleading data.
if len(lines) != len(ages):
logger.debug(
"line_ages length mismatch for {} (lines={}, ages={}); skipping annotation",
file_path, len(lines), len(ages),
)
return content
annotated: list[str] = []
for i, line in enumerate(lines):
if not line.strip() or i >= len(ages):
for line, age in zip(lines, ages):
if not line.strip():
annotated.append(line)
continue
d = ages[i].age_days
if d > 14:
annotated.append(f"{line} \u2190 {d}d")
if age.age_days > _STALE_THRESHOLD_DAYS:
annotated.append(f"{line} \u2190 {age.age_days}d")
else:
annotated.append(line)
result = "\n".join(annotated)
@@ -686,10 +710,14 @@ class Dream:
f"[{e['timestamp']}] {e['content']}" for e in batch
)
# Current file contents + per-line age annotations
# Current file contents + per-line age annotations (MEMORY.md only)
current_date = datetime.now().strftime("%Y-%m-%d")
raw_memory = self.store.read_memory() or "(empty)"
current_memory = self._annotate_with_ages(raw_memory)
current_memory = (
self._annotate_with_ages(raw_memory)
if self.annotate_line_ages
else raw_memory
)
current_soul = self.store.read_soul() or "(empty)"
current_user = self.store.read_user() or "(empty)"
@@ -711,7 +739,11 @@ class Dream:
messages=[
{
"role": "system",
"content": render_template("agent/dream_phase1.md", strip=True),
"content": render_template(
"agent/dream_phase1.md",
strip=True,
stale_threshold_days=_STALE_THRESHOLD_DAYS,
),
},
{"role": "user", "content": phase1_prompt},
],
nanobot/cli/commands.py
+1
View File
@@ -871,6 +871,7 @@ def gateway(
agent.dream.model = dream_cfg.model_override
agent.dream.max_batch_size = dream_cfg.max_batch_size
agent.dream.max_iterations = dream_cfg.max_iterations
agent.dream.annotate_line_ages = dream_cfg.annotate_line_ages
from nanobot.cron.types import CronJob, CronPayload
cron.register_system_job(CronJob(
id="dream",
nanobot/config/schema.py
+5
View File
@@ -43,7 +43,12 @@ class DreamConfig(Base):
validation_alias=AliasChoices("modelOverride", "model", "model_override"),
) # Optional Dream-specific model override
max_batch_size: int = Field(default=20, ge=1) # Max history entries per run
# Bumped from 10 to 15 in #3212 (exp002: +30% dedup, no accuracy loss; >15 plateaus).
max_iterations: int = Field(default=15, ge=1) # Max tool calls per Phase 2
# Per-line git-blame age annotation in Phase 1 prompt (see #3212). Default
# on — set to False to feed MEMORY.md raw if a specific LLM reacts poorly
# to the `← Nd` suffix or you want deterministic, git-independent prompts.
annotate_line_ages: bool = True
def build_schedule(self, timezone: str) -> CronSchedule:
"""Build the runtime schedule, preferring the legacy cron override if present."""
nanobot/templates/agent/dream_phase1.md
+1 -1
View File
@@ -26,7 +26,7 @@ Staleness — MEMORY.md lines may have a ``← Nd`` suffix showing days since la
- Age only indicates when content was last touched, not whether it should be removed
- Use content judgment: user habits/preferences/personality traits are permanent regardless of age
- Only prune content that is objectively outdated: passed events, resolved tracking, superseded approaches
- Lines with ``← Nd`` (N>14) deserve closer review but are NOT automatically removable
- Lines with ``← Nd`` (N>{{ stale_threshold_days }}) deserve closer review but are NOT automatically removable
- When removing: prefer deleting individual items over entire sections
Skill discovery — flag [SKILL] when ALL of these are true:
tests/agent/test_dream.py
+82 -1
View File
@@ -2,11 +2,12 @@
import pytest
from unittest.mock import AsyncMock, MagicMock
from unittest.mock import AsyncMock, MagicMock, patch
from nanobot.agent.memory import Dream, MemoryStore
from nanobot.agent.runner import AgentRunResult
from nanobot.agent.skills import BUILTIN_SKILLS_DIR
from nanobot.utils.gitstore import LineAge
@pytest.fixture
@@ -175,3 +176,83 @@ class TestDreamRun:
user_msg = call_args.kwargs.get("messages", call_args[1].get("messages"))[1]["content"]
assert "## Current MEMORY.md" in user_msg
async def test_phase1_prompt_carries_age_suffix_for_stale_lines(
self, dream, mock_provider, mock_runner, store,
):
"""End-to-end: ages >14d must appear verbatim in the LLM prompt, ages ≤14d must not."""
# MEMORY.md fixture has 2 non-blank lines ("# Memory" and "- Project X active").
# Inject four ages to cover threshold boundaries: >14 suffix, ==14 no suffix, <14 no suffix.
store.write_memory("# Memory\n- Project X active\n- fresh item\n- edge case line")
store.append_history("some event")
mock_provider.chat_with_retry.return_value = MagicMock(content="[SKIP]")
mock_runner.run = AsyncMock(return_value=_make_run_result())
fake_ages = [
LineAge(age_days=30), # "# Memory" → should get ← 30d
LineAge(age_days=20), # "- Project X..." → should get ← 20d
LineAge(age_days=14), # "- fresh item" → ==14, threshold is strictly >14, no suffix
LineAge(age_days=5), # "- edge case..." → no suffix
]
with patch.object(store.git, "line_ages", return_value=fake_ages):
await dream.run()
call_args = mock_provider.chat_with_retry.call_args
user_msg = call_args.kwargs.get("messages", call_args[1].get("messages"))[1]["content"]
memory_section = user_msg.split("## Current MEMORY.md")[1].split("## Current SOUL.md")[0]
assert "\u2190 30d" in memory_section
assert "\u2190 20d" in memory_section
assert "\u2190 14d" not in memory_section
assert "\u2190 5d" not in memory_section
async def test_phase1_skips_annotation_when_disabled(
self, dream, mock_provider, mock_runner, store,
):
"""`annotate_line_ages=False` must bypass the git lookup entirely and keep MEMORY.md raw."""
store.append_history("some event")
mock_provider.chat_with_retry.return_value = MagicMock(content="[SKIP]")
mock_runner.run = AsyncMock(return_value=_make_run_result())
dream.annotate_line_ages = False
# line_ages must be bypassed entirely — verify with a spy rather than a
# raising side_effect, because _annotate_with_ages catches Exception
# (which swallows AssertionError) and would hide an accidental call.
with patch.object(store.git, "line_ages") as mock_line_ages:
await dream.run()
mock_line_ages.assert_not_called()
call_args = mock_provider.chat_with_retry.call_args
user_msg = call_args.kwargs.get("messages", call_args[1].get("messages"))[1]["content"]
assert "\u2190" not in user_msg
async def test_phase1_skips_annotation_on_line_ages_length_mismatch(
self, dream, mock_provider, mock_runner, store,
):
"""If ages length != lines length (dirty working tree), skip annotation instead of mis-tagging."""
# MEMORY.md has 2 non-blank lines but we hand back only 1 age → mismatch.
store.append_history("some event")
mock_provider.chat_with_retry.return_value = MagicMock(content="[SKIP]")
mock_runner.run = AsyncMock(return_value=_make_run_result())
with patch.object(store.git, "line_ages", return_value=[LineAge(age_days=999)]):
await dream.run()
call_args = mock_provider.chat_with_retry.call_args
user_msg = call_args.kwargs.get("messages", call_args[1].get("messages"))[1]["content"]
memory_section = user_msg.split("## Current MEMORY.md")[1].split("## Current SOUL.md")[0]
# No age arrow at all — we refused to annotate rather than tag the wrong line.
assert "\u2190" not in memory_section
async def test_phase1_prompt_uses_threshold_from_template_var(
self, dream, mock_provider, mock_runner, store,
):
"""System prompt should reference the stale-threshold constant, not a hardcoded 14."""
store.append_history("some event")
mock_provider.chat_with_retry.return_value = MagicMock(content="[SKIP]")
mock_runner.run = AsyncMock(return_value=_make_run_result())
await dream.run()
system_msg = mock_provider.chat_with_retry.call_args.kwargs["messages"][0]["content"]
# The template renders with stale_threshold_days=14 → LLM must see "N>14"
assert "N>14" in system_msg