fix(image): prevent duplicate delivery and replay artifacts

This commit is contained in:
Xubin Ren
2026-05-09 05:45:13 +00:00
parent 4d168c571c
commit 3231aaf9ee
9 changed files with 128 additions and 32 deletions
+31 -12
View File
@@ -2,6 +2,7 @@
import json
import os
import re
import shutil
from contextlib import suppress
from dataclasses import dataclass, field
@@ -21,6 +22,25 @@ from nanobot.utils.helpers import (
)
FILE_MAX_MESSAGES = 2000
_MESSAGE_TIME_PREFIX_RE = re.compile(r"^\[Message Time: [^\]]+\]\n?")
_LOCAL_IMAGE_BREADCRUMB_RE = re.compile(r"^\[image: (?:/|~)[^\]]+\]\s*$")
_TOOL_CALL_ECHO_RE = re.compile(r'^\s*(?:generate_image|message)\([^)]*\)\s*$')
def _sanitize_assistant_replay_text(content: str) -> str:
"""Remove internal replay artifacts that the model may have copied before.
These strings are useful as runtime/session metadata, but when they appear
in assistant examples they become demonstrations for the model to repeat.
"""
content = _MESSAGE_TIME_PREFIX_RE.sub("", content, count=1)
lines = [
line
for line in content.splitlines()
if not _LOCAL_IMAGE_BREADCRUMB_RE.match(line)
and not _TOOL_CALL_ECHO_RE.match(line)
]
return "\n".join(lines).strip()
@dataclass
@@ -41,22 +61,15 @@ class Session:
Annotating *every* assistant turn trains the model (via in-context
demonstrations) to start its own replies with the same
``[Message Time: ...]`` prefix, which leaks metadata back to the user.
We therefore only annotate:
* ``user`` turns — needed so the model can pin the conversation in time.
* proactive deliveries (``_channel_delivery=True``) — cron / heartbeat
assistant pushes that may sit hours away from the next user reply,
and are too infrequent to act as parroting demonstrations.
We therefore only annotate user turns. User-side stamps are enough to
pin adjacent assistant replies for relative-time reasoning, including
proactive messages the user replies to later.
"""
timestamp = message.get("timestamp")
if not timestamp or not isinstance(content, str):
return content
role = message.get("role")
if role == "user":
pass
elif role == "assistant" and message.get("_channel_delivery"):
pass
else:
if role != "user":
return content
return f"[Message Time: {timestamp}]\n{content}"
@@ -105,19 +118,25 @@ class Session:
out: list[dict[str, Any]] = []
for message in sliced:
content = message.get("content", "")
role = message.get("role")
if role == "assistant" and isinstance(content, str):
content = _sanitize_assistant_replay_text(content)
# Synthesize an ``[image: path]`` breadcrumb from the persisted
# ``media`` kwarg so LLM replay still sees *something* where the
# image used to be. Without this, an image-only user turn
# replays as an empty user message — the assistant's reply then
# looks like it's responding to nothing.
media = message.get("media")
if isinstance(media, list) and media and isinstance(content, str):
if role == "user" and isinstance(media, list) and media and isinstance(content, str):
breadcrumbs = "\n".join(
image_placeholder_text(p) for p in media if isinstance(p, str) and p
)
content = f"{content}\n{breadcrumbs}" if content else breadcrumbs
if include_timestamps:
content = self._annotate_message_time(message, content)
if role == "assistant" and isinstance(content, str) and not content.strip():
if not any(key in message for key in ("tool_calls", "reasoning_content", "thinking_blocks")):
continue
entry: dict[str, Any] = {"role": message["role"], "content": content}
for key in ("tool_calls", "tool_call_id", "name", "reasoning_content", "thinking_blocks"):
if key in message: